1/*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25#include "si_build_pm4.h"
26#include "gfx9d.h"
27#include "si_query.h"
28
29#include "util/u_dual_blend.h"
30#include "util/u_format.h"
31#include "util/u_format_s3tc.h"
32#include "util/u_memory.h"
33#include "util/u_resource.h"
34#include "util/u_upload_mgr.h"
35#include "util/fast_idiv_by_const.h"
36
37static unsigned si_map_swizzle(unsigned swizzle)
38{
39	switch (swizzle) {
40	case PIPE_SWIZZLE_Y:
41		return V_008F0C_SQ_SEL_Y;
42	case PIPE_SWIZZLE_Z:
43		return V_008F0C_SQ_SEL_Z;
44	case PIPE_SWIZZLE_W:
45		return V_008F0C_SQ_SEL_W;
46	case PIPE_SWIZZLE_0:
47		return V_008F0C_SQ_SEL_0;
48	case PIPE_SWIZZLE_1:
49		return V_008F0C_SQ_SEL_1;
50	default: /* PIPE_SWIZZLE_X */
51		return V_008F0C_SQ_SEL_X;
52	}
53}
54
55/* 12.4 fixed-point */
56static unsigned si_pack_float_12p4(float x)
57{
58	return x <= 0    ? 0 :
59	       x >= 4096 ? 0xffff : x * 16;
60}
61
62/*
63 * Inferred framebuffer and blender state.
64 *
65 * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
66 * if there is not enough PS outputs.
67 */
68static void si_emit_cb_render_state(struct si_context *sctx)
69{
70	struct radeon_cmdbuf *cs = sctx->gfx_cs;
71	struct si_state_blend *blend = sctx->queued.named.blend;
72	/* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
73	 * but you never know. */
74	uint32_t cb_target_mask = sctx->framebuffer.colorbuf_enabled_4bit;
75	unsigned i;
76
77	if (blend)
78		cb_target_mask &= blend->cb_target_mask;
79
80	/* Avoid a hang that happens when dual source blending is enabled
81	 * but there is not enough color outputs. This is undefined behavior,
82	 * so disable color writes completely.
83	 *
84	 * Reproducible with Unigine Heaven 4.0 and drirc missing.
85	 */
86	if (blend && blend->dual_src_blend &&
87	    sctx->ps_shader.cso &&
88	    (sctx->ps_shader.cso->info.colors_written & 0x3) != 0x3)
89		cb_target_mask = 0;
90
91	/* GFX9: Flush DFSM when CB_TARGET_MASK changes.
92	 * I think we don't have to do anything between IBs.
93	 */
94	if (sctx->screen->dfsm_allowed &&
95	    sctx->last_cb_target_mask != cb_target_mask) {
96		sctx->last_cb_target_mask = cb_target_mask;
97
98		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
99		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
100	}
101
102	unsigned initial_cdw = cs->current.cdw;
103	radeon_opt_set_context_reg(sctx, R_028238_CB_TARGET_MASK,
104				   SI_TRACKED_CB_TARGET_MASK, cb_target_mask);
105
106	if (sctx->chip_class >= VI) {
107		/* DCC MSAA workaround for blending.
108		 * Alternatively, we can set CB_COLORi_DCC_CONTROL.OVERWRITE_-
109		 * COMBINER_DISABLE, but that would be more complicated.
110		 */
111		bool oc_disable = (sctx->chip_class == VI ||
112				   sctx->chip_class == GFX9) &&
113				  blend &&
114				  blend->blend_enable_4bit & cb_target_mask &&
115				  sctx->framebuffer.nr_samples >= 2;
116		unsigned watermark = sctx->framebuffer.dcc_overwrite_combiner_watermark;
117
118		radeon_opt_set_context_reg(
119				sctx, R_028424_CB_DCC_CONTROL,
120				SI_TRACKED_CB_DCC_CONTROL,
121				S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
122				S_028424_OVERWRITE_COMBINER_WATERMARK(watermark) |
123				S_028424_OVERWRITE_COMBINER_DISABLE(oc_disable) |
124				S_028424_DISABLE_CONSTANT_ENCODE_REG(sctx->screen->has_dcc_constant_encode));
125	}
126
127	/* RB+ register settings. */
128	if (sctx->screen->rbplus_allowed) {
129		unsigned spi_shader_col_format =
130			sctx->ps_shader.cso ?
131			sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format : 0;
132		unsigned sx_ps_downconvert = 0;
133		unsigned sx_blend_opt_epsilon = 0;
134		unsigned sx_blend_opt_control = 0;
135
136		for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
137			struct si_surface *surf =
138				(struct si_surface*)sctx->framebuffer.state.cbufs[i];
139			unsigned format, swap, spi_format, colormask;
140			bool has_alpha, has_rgb;
141
142			if (!surf)
143				continue;
144
145			format = G_028C70_FORMAT(surf->cb_color_info);
146			swap = G_028C70_COMP_SWAP(surf->cb_color_info);
147			spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
148			colormask = (cb_target_mask >> (i * 4)) & 0xf;
149
150			/* Set if RGB and A are present. */
151			has_alpha = !G_028C74_FORCE_DST_ALPHA_1(surf->cb_color_attrib);
152
153			if (format == V_028C70_COLOR_8 ||
154			    format == V_028C70_COLOR_16 ||
155			    format == V_028C70_COLOR_32)
156				has_rgb = !has_alpha;
157			else
158				has_rgb = true;
159
160			/* Check the colormask and export format. */
161			if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
162				has_rgb = false;
163			if (!(colormask & PIPE_MASK_A))
164				has_alpha = false;
165
166			if (spi_format == V_028714_SPI_SHADER_ZERO) {
167				has_rgb = false;
168				has_alpha = false;
169			}
170
171			/* Disable value checking for disabled channels. */
172			if (!has_rgb)
173				sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
174			if (!has_alpha)
175				sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
176
177			/* Enable down-conversion for 32bpp and smaller formats. */
178			switch (format) {
179			case V_028C70_COLOR_8:
180			case V_028C70_COLOR_8_8:
181			case V_028C70_COLOR_8_8_8_8:
182				/* For 1 and 2-channel formats, use the superset thereof. */
183				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
184				    spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
185				    spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
186					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
187					sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT << (i * 4);
188				}
189				break;
190
191			case V_028C70_COLOR_5_6_5:
192				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
193					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
194					sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT << (i * 4);
195				}
196				break;
197
198			case V_028C70_COLOR_1_5_5_5:
199				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
200					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
201					sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT << (i * 4);
202				}
203				break;
204
205			case V_028C70_COLOR_4_4_4_4:
206				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
207					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
208					sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT << (i * 4);
209				}
210				break;
211
212			case V_028C70_COLOR_32:
213				if (swap == V_028C70_SWAP_STD &&
214				    spi_format == V_028714_SPI_SHADER_32_R)
215					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
216				else if (swap == V_028C70_SWAP_ALT_REV &&
217					 spi_format == V_028714_SPI_SHADER_32_AR)
218					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
219				break;
220
221			case V_028C70_COLOR_16:
222			case V_028C70_COLOR_16_16:
223				/* For 1-channel formats, use the superset thereof. */
224				if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
225				    spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
226				    spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
227				    spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
228					if (swap == V_028C70_SWAP_STD ||
229					    swap == V_028C70_SWAP_STD_REV)
230						sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
231					else
232						sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
233				}
234				break;
235
236			case V_028C70_COLOR_10_11_11:
237				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
238					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
239					sx_blend_opt_epsilon |= V_028758_11BIT_FORMAT << (i * 4);
240				}
241				break;
242
243			case V_028C70_COLOR_2_10_10_10:
244				if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
245					sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
246					sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT << (i * 4);
247				}
248				break;
249			}
250		}
251
252		/* SX_PS_DOWNCONVERT, SX_BLEND_OPT_EPSILON, SX_BLEND_OPT_CONTROL */
253		radeon_opt_set_context_reg3(sctx, R_028754_SX_PS_DOWNCONVERT,
254					    SI_TRACKED_SX_PS_DOWNCONVERT,
255					    sx_ps_downconvert, sx_blend_opt_epsilon,
256					    sx_blend_opt_control);
257	}
258	if (initial_cdw != cs->current.cdw)
259		sctx->context_roll = true;
260}
261
262/*
263 * Blender functions
264 */
265
266static uint32_t si_translate_blend_function(int blend_func)
267{
268	switch (blend_func) {
269	case PIPE_BLEND_ADD:
270		return V_028780_COMB_DST_PLUS_SRC;
271	case PIPE_BLEND_SUBTRACT:
272		return V_028780_COMB_SRC_MINUS_DST;
273	case PIPE_BLEND_REVERSE_SUBTRACT:
274		return V_028780_COMB_DST_MINUS_SRC;
275	case PIPE_BLEND_MIN:
276		return V_028780_COMB_MIN_DST_SRC;
277	case PIPE_BLEND_MAX:
278		return V_028780_COMB_MAX_DST_SRC;
279	default:
280		PRINT_ERR("Unknown blend function %d\n", blend_func);
281		assert(0);
282		break;
283	}
284	return 0;
285}
286
287static uint32_t si_translate_blend_factor(int blend_fact)
288{
289	switch (blend_fact) {
290	case PIPE_BLENDFACTOR_ONE:
291		return V_028780_BLEND_ONE;
292	case PIPE_BLENDFACTOR_SRC_COLOR:
293		return V_028780_BLEND_SRC_COLOR;
294	case PIPE_BLENDFACTOR_SRC_ALPHA:
295		return V_028780_BLEND_SRC_ALPHA;
296	case PIPE_BLENDFACTOR_DST_ALPHA:
297		return V_028780_BLEND_DST_ALPHA;
298	case PIPE_BLENDFACTOR_DST_COLOR:
299		return V_028780_BLEND_DST_COLOR;
300	case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
301		return V_028780_BLEND_SRC_ALPHA_SATURATE;
302	case PIPE_BLENDFACTOR_CONST_COLOR:
303		return V_028780_BLEND_CONSTANT_COLOR;
304	case PIPE_BLENDFACTOR_CONST_ALPHA:
305		return V_028780_BLEND_CONSTANT_ALPHA;
306	case PIPE_BLENDFACTOR_ZERO:
307		return V_028780_BLEND_ZERO;
308	case PIPE_BLENDFACTOR_INV_SRC_COLOR:
309		return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
310	case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
311		return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
312	case PIPE_BLENDFACTOR_INV_DST_ALPHA:
313		return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
314	case PIPE_BLENDFACTOR_INV_DST_COLOR:
315		return V_028780_BLEND_ONE_MINUS_DST_COLOR;
316	case PIPE_BLENDFACTOR_INV_CONST_COLOR:
317		return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
318	case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
319		return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
320	case PIPE_BLENDFACTOR_SRC1_COLOR:
321		return V_028780_BLEND_SRC1_COLOR;
322	case PIPE_BLENDFACTOR_SRC1_ALPHA:
323		return V_028780_BLEND_SRC1_ALPHA;
324	case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
325		return V_028780_BLEND_INV_SRC1_COLOR;
326	case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
327		return V_028780_BLEND_INV_SRC1_ALPHA;
328	default:
329		PRINT_ERR("Bad blend factor %d not supported!\n", blend_fact);
330		assert(0);
331		break;
332	}
333	return 0;
334}
335
336static uint32_t si_translate_blend_opt_function(int blend_func)
337{
338	switch (blend_func) {
339	case PIPE_BLEND_ADD:
340		return V_028760_OPT_COMB_ADD;
341	case PIPE_BLEND_SUBTRACT:
342		return V_028760_OPT_COMB_SUBTRACT;
343	case PIPE_BLEND_REVERSE_SUBTRACT:
344		return V_028760_OPT_COMB_REVSUBTRACT;
345	case PIPE_BLEND_MIN:
346		return V_028760_OPT_COMB_MIN;
347	case PIPE_BLEND_MAX:
348		return V_028760_OPT_COMB_MAX;
349	default:
350		return V_028760_OPT_COMB_BLEND_DISABLED;
351	}
352}
353
354static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
355{
356	switch (blend_fact) {
357	case PIPE_BLENDFACTOR_ZERO:
358		return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
359	case PIPE_BLENDFACTOR_ONE:
360		return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
361	case PIPE_BLENDFACTOR_SRC_COLOR:
362		return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
363				: V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
364	case PIPE_BLENDFACTOR_INV_SRC_COLOR:
365		return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
366				: V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
367	case PIPE_BLENDFACTOR_SRC_ALPHA:
368		return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
369	case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
370		return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
371	case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
372		return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
373				: V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
374	default:
375		return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
376	}
377}
378
379static void si_blend_check_commutativity(struct si_screen *sscreen,
380					 struct si_state_blend *blend,
381					 enum pipe_blend_func func,
382					 enum pipe_blendfactor src,
383					 enum pipe_blendfactor dst,
384					 unsigned chanmask)
385{
386	/* Src factor is allowed when it does not depend on Dst */
387	static const uint32_t src_allowed =
388		(1u << PIPE_BLENDFACTOR_ONE) |
389		(1u << PIPE_BLENDFACTOR_SRC_COLOR) |
390		(1u << PIPE_BLENDFACTOR_SRC_ALPHA) |
391		(1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
392		(1u << PIPE_BLENDFACTOR_CONST_COLOR) |
393		(1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
394		(1u << PIPE_BLENDFACTOR_SRC1_COLOR) |
395		(1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
396		(1u << PIPE_BLENDFACTOR_ZERO) |
397		(1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
398		(1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) |
399		(1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
400		(1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) |
401		(1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
402		(1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
403
404	if (dst == PIPE_BLENDFACTOR_ONE &&
405	    (src_allowed & (1u << src))) {
406		/* Addition is commutative, but floating point addition isn't
407		 * associative: subtle changes can be introduced via different
408		 * rounding.
409		 *
410		 * Out-of-order is also non-deterministic, which means that
411		 * this breaks OpenGL invariance requirements. So only enable
412		 * out-of-order additive blending if explicitly allowed by a
413		 * setting.
414		 */
415		if (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN ||
416		    (func == PIPE_BLEND_ADD && sscreen->commutative_blend_add))
417			blend->commutative_4bit |= chanmask;
418	}
419}
420
421/**
422 * Get rid of DST in the blend factors by commuting the operands:
423 *    func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
424 */
425static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
426				unsigned *dst_factor, unsigned expected_dst,
427				unsigned replacement_src)
428{
429	if (*src_factor == expected_dst &&
430	    *dst_factor == PIPE_BLENDFACTOR_ZERO) {
431		*src_factor = PIPE_BLENDFACTOR_ZERO;
432		*dst_factor = replacement_src;
433
434		/* Commuting the operands requires reversing subtractions. */
435		if (*func == PIPE_BLEND_SUBTRACT)
436			*func = PIPE_BLEND_REVERSE_SUBTRACT;
437		else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
438			*func = PIPE_BLEND_SUBTRACT;
439	}
440}
441
442static bool si_blend_factor_uses_dst(unsigned factor)
443{
444	return factor == PIPE_BLENDFACTOR_DST_COLOR ||
445		factor == PIPE_BLENDFACTOR_DST_ALPHA ||
446		factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
447		factor == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
448		factor == PIPE_BLENDFACTOR_INV_DST_COLOR;
449}
450
451static void *si_create_blend_state_mode(struct pipe_context *ctx,
452					const struct pipe_blend_state *state,
453					unsigned mode)
454{
455	struct si_context *sctx = (struct si_context*)ctx;
456	struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
457	struct si_pm4_state *pm4 = &blend->pm4;
458	uint32_t sx_mrt_blend_opt[8] = {0};
459	uint32_t color_control = 0;
460
461	if (!blend)
462		return NULL;
463
464	blend->alpha_to_coverage = state->alpha_to_coverage;
465	blend->alpha_to_one = state->alpha_to_one;
466	blend->dual_src_blend = util_blend_state_is_dual(state, 0);
467	blend->logicop_enable = state->logicop_enable;
468
469	if (state->logicop_enable) {
470		color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
471	} else {
472		color_control |= S_028808_ROP3(0xcc);
473	}
474
475	si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
476		       S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
477		       S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
478		       S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
479		       S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
480		       S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
481		       S_028B70_OFFSET_ROUND(1));
482
483	if (state->alpha_to_coverage)
484		blend->need_src_alpha_4bit |= 0xf;
485
486	blend->cb_target_mask = 0;
487	blend->cb_target_enabled_4bit = 0;
488
489	for (int i = 0; i < 8; i++) {
490		/* state->rt entries > 0 only written if independent blending */
491		const int j = state->independent_blend_enable ? i : 0;
492
493		unsigned eqRGB = state->rt[j].rgb_func;
494		unsigned srcRGB = state->rt[j].rgb_src_factor;
495		unsigned dstRGB = state->rt[j].rgb_dst_factor;
496		unsigned eqA = state->rt[j].alpha_func;
497		unsigned srcA = state->rt[j].alpha_src_factor;
498		unsigned dstA = state->rt[j].alpha_dst_factor;
499
500		unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
501		unsigned blend_cntl = 0;
502
503		sx_mrt_blend_opt[i] =
504			S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
505			S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
506
507		/* Only set dual source blending for MRT0 to avoid a hang. */
508		if (i >= 1 && blend->dual_src_blend) {
509			/* Vulkan does this for dual source blending. */
510			if (i == 1)
511				blend_cntl |= S_028780_ENABLE(1);
512
513			si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
514			continue;
515		}
516
517		/* Only addition and subtraction equations are supported with
518		 * dual source blending.
519		 */
520		if (blend->dual_src_blend &&
521		    (eqRGB == PIPE_BLEND_MIN || eqRGB == PIPE_BLEND_MAX ||
522		     eqA == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MAX)) {
523			assert(!"Unsupported equation for dual source blending");
524			si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
525			continue;
526		}
527
528		/* cb_render_state will disable unused ones */
529		blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
530		if (state->rt[j].colormask)
531			blend->cb_target_enabled_4bit |= 0xf << (4 * i);
532
533		if (!state->rt[j].colormask || !state->rt[j].blend_enable) {
534			si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
535			continue;
536		}
537
538		si_blend_check_commutativity(sctx->screen, blend,
539					     eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
540		si_blend_check_commutativity(sctx->screen, blend,
541					     eqA, srcA, dstA, 0x8 << (4 * i));
542
543		/* Blending optimizations for RB+.
544		 * These transformations don't change the behavior.
545		 *
546		 * First, get rid of DST in the blend factors:
547		 *    func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
548		 */
549		si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
550				    PIPE_BLENDFACTOR_DST_COLOR,
551				    PIPE_BLENDFACTOR_SRC_COLOR);
552		si_blend_remove_dst(&eqA, &srcA, &dstA,
553				    PIPE_BLENDFACTOR_DST_COLOR,
554				    PIPE_BLENDFACTOR_SRC_COLOR);
555		si_blend_remove_dst(&eqA, &srcA, &dstA,
556				    PIPE_BLENDFACTOR_DST_ALPHA,
557				    PIPE_BLENDFACTOR_SRC_ALPHA);
558
559		/* Look up the ideal settings from tables. */
560		srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
561		dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
562		srcA_opt = si_translate_blend_opt_factor(srcA, true);
563		dstA_opt = si_translate_blend_opt_factor(dstA, true);
564
565		/* Handle interdependencies. */
566		if (si_blend_factor_uses_dst(srcRGB))
567			dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
568		if (si_blend_factor_uses_dst(srcA))
569			dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
570
571		if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
572		    (dstRGB == PIPE_BLENDFACTOR_ZERO ||
573		     dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
574		     dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
575			dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
576
577		/* Set the final value. */
578		sx_mrt_blend_opt[i] =
579			S_028760_COLOR_SRC_OPT(srcRGB_opt) |
580			S_028760_COLOR_DST_OPT(dstRGB_opt) |
581			S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
582			S_028760_ALPHA_SRC_OPT(srcA_opt) |
583			S_028760_ALPHA_DST_OPT(dstA_opt) |
584			S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
585
586		/* Set blend state. */
587		blend_cntl |= S_028780_ENABLE(1);
588		blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
589		blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
590		blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
591
592		if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
593			blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
594			blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
595			blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
596			blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
597		}
598		si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
599
600		blend->blend_enable_4bit |= 0xfu << (i * 4);
601
602		/* This is only important for formats without alpha. */
603		if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
604		    dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
605		    srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
606		    dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
607		    srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
608		    dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
609			blend->need_src_alpha_4bit |= 0xfu << (i * 4);
610	}
611
612	if (blend->cb_target_mask) {
613		color_control |= S_028808_MODE(mode);
614	} else {
615		color_control |= S_028808_MODE(V_028808_CB_DISABLE);
616	}
617
618	if (sctx->screen->rbplus_allowed) {
619		/* Disable RB+ blend optimizations for dual source blending.
620		 * Vulkan does this.
621		 */
622		if (blend->dual_src_blend) {
623			for (int i = 0; i < 8; i++) {
624				sx_mrt_blend_opt[i] =
625					S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
626					S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
627			}
628		}
629
630		for (int i = 0; i < 8; i++)
631			si_pm4_set_reg(pm4, R_028760_SX_MRT0_BLEND_OPT + i * 4,
632				       sx_mrt_blend_opt[i]);
633
634		/* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
635		if (blend->dual_src_blend || state->logicop_enable ||
636		    mode == V_028808_CB_RESOLVE)
637			color_control |= S_028808_DISABLE_DUAL_QUAD(1);
638	}
639
640	si_pm4_set_reg(pm4, R_028808_CB_COLOR_CONTROL, color_control);
641	return blend;
642}
643
644static void *si_create_blend_state(struct pipe_context *ctx,
645				   const struct pipe_blend_state *state)
646{
647	return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
648}
649
650static void si_bind_blend_state(struct pipe_context *ctx, void *state)
651{
652	struct si_context *sctx = (struct si_context *)ctx;
653	struct si_state_blend *old_blend = sctx->queued.named.blend;
654	struct si_state_blend *blend = (struct si_state_blend *)state;
655
656	if (!state)
657		return;
658
659	si_pm4_bind_state(sctx, blend, state);
660
661	if (!old_blend ||
662	    old_blend->cb_target_mask != blend->cb_target_mask ||
663	    old_blend->dual_src_blend != blend->dual_src_blend ||
664	    (old_blend->blend_enable_4bit != blend->blend_enable_4bit &&
665	     sctx->framebuffer.nr_samples >= 2 &&
666	     sctx->screen->dcc_msaa_allowed))
667		si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
668
669	if (!old_blend ||
670	    old_blend->cb_target_mask != blend->cb_target_mask ||
671	    old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
672	    old_blend->alpha_to_one != blend->alpha_to_one ||
673	    old_blend->dual_src_blend != blend->dual_src_blend ||
674	    old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
675	    old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
676		sctx->do_update_shaders = true;
677
678	if (sctx->screen->dpbb_allowed &&
679	    (!old_blend ||
680	     old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
681	     old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
682	     old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
683		si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
684
685	if (sctx->screen->has_out_of_order_rast &&
686	    (!old_blend ||
687	     (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
688	      old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
689	      old_blend->commutative_4bit != blend->commutative_4bit ||
690	      old_blend->logicop_enable != blend->logicop_enable)))
691		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
692}
693
694static void si_delete_blend_state(struct pipe_context *ctx, void *state)
695{
696	struct si_context *sctx = (struct si_context *)ctx;
697	si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
698}
699
700static void si_set_blend_color(struct pipe_context *ctx,
701			       const struct pipe_blend_color *state)
702{
703	struct si_context *sctx = (struct si_context *)ctx;
704	static const struct pipe_blend_color zeros;
705
706	sctx->blend_color.state = *state;
707	sctx->blend_color.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
708	si_mark_atom_dirty(sctx, &sctx->atoms.s.blend_color);
709}
710
711static void si_emit_blend_color(struct si_context *sctx)
712{
713	struct radeon_cmdbuf *cs = sctx->gfx_cs;
714
715	radeon_set_context_reg_seq(cs, R_028414_CB_BLEND_RED, 4);
716	radeon_emit_array(cs, (uint32_t*)sctx->blend_color.state.color, 4);
717}
718
719/*
720 * Clipping
721 */
722
723static void si_set_clip_state(struct pipe_context *ctx,
724			      const struct pipe_clip_state *state)
725{
726	struct si_context *sctx = (struct si_context *)ctx;
727	struct pipe_constant_buffer cb;
728	static const struct pipe_clip_state zeros;
729
730	if (memcmp(&sctx->clip_state.state, state, sizeof(*state)) == 0)
731		return;
732
733	sctx->clip_state.state = *state;
734	sctx->clip_state.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
735	si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_state);
736
737	cb.buffer = NULL;
738	cb.user_buffer = state->ucp;
739	cb.buffer_offset = 0;
740	cb.buffer_size = 4*4*8;
741	si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
742	pipe_resource_reference(&cb.buffer, NULL);
743}
744
745static void si_emit_clip_state(struct si_context *sctx)
746{
747	struct radeon_cmdbuf *cs = sctx->gfx_cs;
748
749	radeon_set_context_reg_seq(cs, R_0285BC_PA_CL_UCP_0_X, 6*4);
750	radeon_emit_array(cs, (uint32_t*)sctx->clip_state.state.ucp, 6*4);
751}
752
753static void si_emit_clip_regs(struct si_context *sctx)
754{
755	struct si_shader *vs = si_get_vs_state(sctx);
756	struct si_shader_selector *vs_sel = vs->selector;
757	struct tgsi_shader_info *info = &vs_sel->info;
758	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
759	unsigned window_space =
760	   info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
761	unsigned clipdist_mask = vs_sel->clipdist_mask;
762	unsigned ucp_mask = clipdist_mask ? 0 : rs->clip_plane_enable & SIX_BITS;
763	unsigned culldist_mask = vs_sel->culldist_mask;
764	unsigned total_mask;
765
766	if (vs->key.opt.clip_disable) {
767		assert(!info->culldist_writemask);
768		clipdist_mask = 0;
769		culldist_mask = 0;
770	}
771	total_mask = clipdist_mask | culldist_mask;
772
773	/* Clip distances on points have no effect, so need to be implemented
774	 * as cull distances. This applies for the clipvertex case as well.
775	 *
776	 * Setting this for primitives other than points should have no adverse
777	 * effects.
778	 */
779	clipdist_mask &= rs->clip_plane_enable;
780	culldist_mask |= clipdist_mask;
781
782	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
783	radeon_opt_set_context_reg(sctx, R_02881C_PA_CL_VS_OUT_CNTL,
784		SI_TRACKED_PA_CL_VS_OUT_CNTL,
785		vs_sel->pa_cl_vs_out_cntl |
786		S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0F) != 0) |
787		S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xF0) != 0) |
788		clipdist_mask | (culldist_mask << 8));
789	radeon_opt_set_context_reg(sctx, R_028810_PA_CL_CLIP_CNTL,
790		SI_TRACKED_PA_CL_CLIP_CNTL,
791		rs->pa_cl_clip_cntl |
792		ucp_mask |
793		S_028810_CLIP_DISABLE(window_space));
794
795	if (initial_cdw != sctx->gfx_cs->current.cdw)
796		sctx->context_roll = true;
797}
798
799/*
800 * inferred state between framebuffer and rasterizer
801 */
802static void si_update_poly_offset_state(struct si_context *sctx)
803{
804	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
805
806	if (!rs || !rs->uses_poly_offset || !sctx->framebuffer.state.zsbuf) {
807		si_pm4_bind_state(sctx, poly_offset, NULL);
808		return;
809	}
810
811	/* Use the user format, not db_render_format, so that the polygon
812	 * offset behaves as expected by applications.
813	 */
814	switch (sctx->framebuffer.state.zsbuf->texture->format) {
815	case PIPE_FORMAT_Z16_UNORM:
816		si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[0]);
817		break;
818	default: /* 24-bit */
819		si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[1]);
820		break;
821	case PIPE_FORMAT_Z32_FLOAT:
822	case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
823		si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[2]);
824		break;
825	}
826}
827
828/*
829 * Rasterizer
830 */
831
832static uint32_t si_translate_fill(uint32_t func)
833{
834	switch(func) {
835	case PIPE_POLYGON_MODE_FILL:
836		return V_028814_X_DRAW_TRIANGLES;
837	case PIPE_POLYGON_MODE_LINE:
838		return V_028814_X_DRAW_LINES;
839	case PIPE_POLYGON_MODE_POINT:
840		return V_028814_X_DRAW_POINTS;
841	default:
842		assert(0);
843		return V_028814_X_DRAW_POINTS;
844	}
845}
846
847static void *si_create_rs_state(struct pipe_context *ctx,
848				const struct pipe_rasterizer_state *state)
849{
850	struct si_screen *sscreen = ((struct si_context *)ctx)->screen;
851	struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
852	struct si_pm4_state *pm4 = &rs->pm4;
853	unsigned tmp, i;
854	float psize_min, psize_max;
855
856	if (!rs) {
857		return NULL;
858	}
859
860	rs->scissor_enable = state->scissor;
861	rs->clip_halfz = state->clip_halfz;
862	rs->two_side = state->light_twoside;
863	rs->multisample_enable = state->multisample;
864	rs->force_persample_interp = state->force_persample_interp;
865	rs->clip_plane_enable = state->clip_plane_enable;
866	rs->half_pixel_center = state->half_pixel_center;
867	rs->line_stipple_enable = state->line_stipple_enable;
868	rs->poly_stipple_enable = state->poly_stipple_enable;
869	rs->line_smooth = state->line_smooth;
870	rs->line_width = state->line_width;
871	rs->poly_smooth = state->poly_smooth;
872	rs->uses_poly_offset = state->offset_point || state->offset_line ||
873			       state->offset_tri;
874	rs->clamp_fragment_color = state->clamp_fragment_color;
875	rs->clamp_vertex_color = state->clamp_vertex_color;
876	rs->flatshade = state->flatshade;
877	rs->sprite_coord_enable = state->sprite_coord_enable;
878	rs->rasterizer_discard = state->rasterizer_discard;
879	rs->pa_sc_line_stipple = state->line_stipple_enable ?
880				S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
881				S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
882	rs->pa_cl_clip_cntl =
883		S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
884		S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip_near) |
885		S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip_far) |
886		S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
887		S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
888
889	si_pm4_set_reg(pm4, R_0286D4_SPI_INTERP_CONTROL_0,
890		S_0286D4_FLAT_SHADE_ENA(1) |
891		S_0286D4_PNT_SPRITE_ENA(state->point_quad_rasterization) |
892		S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
893		S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
894		S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
895		S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
896		S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT));
897
898	/* point size 12.4 fixed point */
899	tmp = (unsigned)(state->point_size * 8.0);
900	si_pm4_set_reg(pm4, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
901
902	if (state->point_size_per_vertex) {
903		psize_min = util_get_min_point_size(state);
904		psize_max = SI_MAX_POINT_SIZE;
905	} else {
906		/* Force the point size to be as if the vertex output was disabled. */
907		psize_min = state->point_size;
908		psize_max = state->point_size;
909	}
910	rs->max_point_size = psize_max;
911
912	/* Divide by two, because 0.5 = 1 pixel. */
913	si_pm4_set_reg(pm4, R_028A04_PA_SU_POINT_MINMAX,
914			S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min/2)) |
915			S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max/2)));
916
917	si_pm4_set_reg(pm4, R_028A08_PA_SU_LINE_CNTL,
918		       S_028A08_WIDTH(si_pack_float_12p4(state->line_width/2)));
919	si_pm4_set_reg(pm4, R_028A48_PA_SC_MODE_CNTL_0,
920		       S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
921		       S_028A48_MSAA_ENABLE(state->multisample ||
922					    state->poly_smooth ||
923					    state->line_smooth) |
924		       S_028A48_VPORT_SCISSOR_ENABLE(1) |
925		       S_028A48_ALTERNATE_RBS_PER_TILE(sscreen->info.chip_class >= GFX9));
926
927	si_pm4_set_reg(pm4, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
928	si_pm4_set_reg(pm4, R_028814_PA_SU_SC_MODE_CNTL,
929		S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
930		S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
931		S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
932		S_028814_FACE(!state->front_ccw) |
933		S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
934		S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
935		S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
936		S_028814_POLY_MODE(state->fill_front != PIPE_POLYGON_MODE_FILL ||
937				   state->fill_back != PIPE_POLYGON_MODE_FILL) |
938		S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
939		S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)));
940
941	if (!rs->uses_poly_offset)
942		return rs;
943
944	rs->pm4_poly_offset = CALLOC(3, sizeof(struct si_pm4_state));
945	if (!rs->pm4_poly_offset) {
946		FREE(rs);
947		return NULL;
948	}
949
950	/* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
951	for (i = 0; i < 3; i++) {
952		struct si_pm4_state *pm4 = &rs->pm4_poly_offset[i];
953		float offset_units = state->offset_units;
954		float offset_scale = state->offset_scale * 16.0f;
955		uint32_t pa_su_poly_offset_db_fmt_cntl = 0;
956
957		if (!state->offset_units_unscaled) {
958			switch (i) {
959			case 0: /* 16-bit zbuffer */
960				offset_units *= 4.0f;
961				pa_su_poly_offset_db_fmt_cntl =
962					S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
963				break;
964			case 1: /* 24-bit zbuffer */
965				offset_units *= 2.0f;
966				pa_su_poly_offset_db_fmt_cntl =
967					S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
968				break;
969			case 2: /* 32-bit zbuffer */
970				offset_units *= 1.0f;
971				pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
972								S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
973				break;
974			}
975		}
976
977		si_pm4_set_reg(pm4, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
978			       fui(offset_scale));
979		si_pm4_set_reg(pm4, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
980			       fui(offset_units));
981		si_pm4_set_reg(pm4, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
982			       fui(offset_scale));
983		si_pm4_set_reg(pm4, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
984			       fui(offset_units));
985		si_pm4_set_reg(pm4, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
986			       pa_su_poly_offset_db_fmt_cntl);
987	}
988
989	return rs;
990}
991
992static void si_bind_rs_state(struct pipe_context *ctx, void *state)
993{
994	struct si_context *sctx = (struct si_context *)ctx;
995	struct si_state_rasterizer *old_rs =
996		(struct si_state_rasterizer*)sctx->queued.named.rasterizer;
997	struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
998
999	if (!state)
1000		return;
1001
1002	if (!old_rs || old_rs->multisample_enable != rs->multisample_enable) {
1003		si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1004
1005		/* Update the small primitive filter workaround if necessary. */
1006		if (sctx->screen->has_msaa_sample_loc_bug &&
1007		    sctx->framebuffer.nr_samples > 1)
1008			si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
1009	}
1010
1011	sctx->current_vs_state &= C_VS_STATE_CLAMP_VERTEX_COLOR;
1012	sctx->current_vs_state |= S_VS_STATE_CLAMP_VERTEX_COLOR(rs->clamp_vertex_color);
1013
1014	si_pm4_bind_state(sctx, rasterizer, rs);
1015	si_update_poly_offset_state(sctx);
1016
1017	if (!old_rs ||
1018	    old_rs->scissor_enable != rs->scissor_enable)
1019		si_mark_atom_dirty(sctx, &sctx->atoms.s.scissors);
1020
1021	if (!old_rs ||
1022	    old_rs->line_width != rs->line_width ||
1023	    old_rs->max_point_size != rs->max_point_size ||
1024	    old_rs->half_pixel_center != rs->half_pixel_center)
1025		si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
1026
1027	if (!old_rs ||
1028	    old_rs->clip_halfz != rs->clip_halfz)
1029		si_mark_atom_dirty(sctx, &sctx->atoms.s.viewports);
1030
1031	if (!old_rs ||
1032	    old_rs->clip_plane_enable != rs->clip_plane_enable ||
1033	    old_rs->pa_cl_clip_cntl != rs->pa_cl_clip_cntl)
1034		si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
1035
1036	sctx->ia_multi_vgt_param_key.u.line_stipple_enabled =
1037		rs->line_stipple_enable;
1038
1039	if (!old_rs ||
1040	    old_rs->clip_plane_enable != rs->clip_plane_enable ||
1041	    old_rs->rasterizer_discard != rs->rasterizer_discard ||
1042	    old_rs->sprite_coord_enable != rs->sprite_coord_enable ||
1043	    old_rs->flatshade != rs->flatshade ||
1044	    old_rs->two_side != rs->two_side ||
1045	    old_rs->multisample_enable != rs->multisample_enable ||
1046	    old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1047	    old_rs->poly_smooth != rs->poly_smooth ||
1048	    old_rs->line_smooth != rs->line_smooth ||
1049	    old_rs->clamp_fragment_color != rs->clamp_fragment_color ||
1050	    old_rs->force_persample_interp != rs->force_persample_interp)
1051		sctx->do_update_shaders = true;
1052}
1053
1054static void si_delete_rs_state(struct pipe_context *ctx, void *state)
1055{
1056	struct si_context *sctx = (struct si_context *)ctx;
1057	struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1058
1059	if (sctx->queued.named.rasterizer == state)
1060		si_pm4_bind_state(sctx, poly_offset, NULL);
1061
1062	FREE(rs->pm4_poly_offset);
1063	si_pm4_delete_state(sctx, rasterizer, rs);
1064}
1065
1066/*
1067 * infeered state between dsa and stencil ref
1068 */
1069static void si_emit_stencil_ref(struct si_context *sctx)
1070{
1071	struct radeon_cmdbuf *cs = sctx->gfx_cs;
1072	struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
1073	struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
1074
1075	radeon_set_context_reg_seq(cs, R_028430_DB_STENCILREFMASK, 2);
1076	radeon_emit(cs, S_028430_STENCILTESTVAL(ref->ref_value[0]) |
1077			S_028430_STENCILMASK(dsa->valuemask[0]) |
1078			S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
1079			S_028430_STENCILOPVAL(1));
1080	radeon_emit(cs, S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
1081			S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
1082			S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
1083			S_028434_STENCILOPVAL_BF(1));
1084}
1085
1086static void si_set_stencil_ref(struct pipe_context *ctx,
1087			       const struct pipe_stencil_ref *state)
1088{
1089        struct si_context *sctx = (struct si_context *)ctx;
1090
1091	if (memcmp(&sctx->stencil_ref.state, state, sizeof(*state)) == 0)
1092		return;
1093
1094	sctx->stencil_ref.state = *state;
1095	si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1096}
1097
1098
1099/*
1100 * DSA
1101 */
1102
1103static uint32_t si_translate_stencil_op(int s_op)
1104{
1105	switch (s_op) {
1106	case PIPE_STENCIL_OP_KEEP:
1107		return V_02842C_STENCIL_KEEP;
1108	case PIPE_STENCIL_OP_ZERO:
1109		return V_02842C_STENCIL_ZERO;
1110	case PIPE_STENCIL_OP_REPLACE:
1111		return V_02842C_STENCIL_REPLACE_TEST;
1112	case PIPE_STENCIL_OP_INCR:
1113		return V_02842C_STENCIL_ADD_CLAMP;
1114	case PIPE_STENCIL_OP_DECR:
1115		return V_02842C_STENCIL_SUB_CLAMP;
1116	case PIPE_STENCIL_OP_INCR_WRAP:
1117		return V_02842C_STENCIL_ADD_WRAP;
1118	case PIPE_STENCIL_OP_DECR_WRAP:
1119		return V_02842C_STENCIL_SUB_WRAP;
1120	case PIPE_STENCIL_OP_INVERT:
1121		return V_02842C_STENCIL_INVERT;
1122	default:
1123		PRINT_ERR("Unknown stencil op %d", s_op);
1124		assert(0);
1125		break;
1126	}
1127	return 0;
1128}
1129
1130static bool si_dsa_writes_stencil(const struct pipe_stencil_state *s)
1131{
1132	return s->enabled && s->writemask &&
1133	       (s->fail_op  != PIPE_STENCIL_OP_KEEP ||
1134		s->zfail_op != PIPE_STENCIL_OP_KEEP ||
1135		s->zpass_op != PIPE_STENCIL_OP_KEEP);
1136}
1137
1138static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
1139{
1140	/* REPLACE is normally order invariant, except when the stencil
1141	 * reference value is written by the fragment shader. Tracking this
1142	 * interaction does not seem worth the effort, so be conservative. */
1143	return op != PIPE_STENCIL_OP_INCR &&
1144	       op != PIPE_STENCIL_OP_DECR &&
1145	       op != PIPE_STENCIL_OP_REPLACE;
1146}
1147
1148/* Compute whether, assuming Z writes are disabled, this stencil state is order
1149 * invariant in the sense that the set of passing fragments as well as the
1150 * final stencil buffer result does not depend on the order of fragments. */
1151static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
1152{
1153	return !state->enabled || !state->writemask ||
1154	       /* The following assumes that Z writes are disabled. */
1155	       (state->func == PIPE_FUNC_ALWAYS &&
1156	        si_order_invariant_stencil_op(state->zpass_op) &&
1157	        si_order_invariant_stencil_op(state->zfail_op)) ||
1158	       (state->func == PIPE_FUNC_NEVER &&
1159	        si_order_invariant_stencil_op(state->fail_op));
1160}
1161
1162static void *si_create_dsa_state(struct pipe_context *ctx,
1163				 const struct pipe_depth_stencil_alpha_state *state)
1164{
1165	struct si_context *sctx = (struct si_context *)ctx;
1166	struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
1167	struct si_pm4_state *pm4 = &dsa->pm4;
1168	unsigned db_depth_control;
1169	uint32_t db_stencil_control = 0;
1170
1171	if (!dsa) {
1172		return NULL;
1173	}
1174
1175	dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
1176	dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
1177	dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
1178	dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
1179
1180	db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
1181		S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
1182		S_028800_ZFUNC(state->depth.func) |
1183		S_028800_DEPTH_BOUNDS_ENABLE(state->depth.bounds_test);
1184
1185	/* stencil */
1186	if (state->stencil[0].enabled) {
1187		db_depth_control |= S_028800_STENCIL_ENABLE(1);
1188		db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
1189		db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
1190		db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
1191		db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
1192
1193		if (state->stencil[1].enabled) {
1194			db_depth_control |= S_028800_BACKFACE_ENABLE(1);
1195			db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
1196			db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
1197			db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
1198			db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
1199		}
1200	}
1201
1202	/* alpha */
1203	if (state->alpha.enabled) {
1204		dsa->alpha_func = state->alpha.func;
1205
1206		si_pm4_set_reg(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
1207		               SI_SGPR_ALPHA_REF * 4, fui(state->alpha.ref_value));
1208	} else {
1209		dsa->alpha_func = PIPE_FUNC_ALWAYS;
1210	}
1211
1212	si_pm4_set_reg(pm4, R_028800_DB_DEPTH_CONTROL, db_depth_control);
1213	if (state->stencil[0].enabled)
1214		si_pm4_set_reg(pm4, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
1215	if (state->depth.bounds_test) {
1216		si_pm4_set_reg(pm4, R_028020_DB_DEPTH_BOUNDS_MIN, fui(state->depth.bounds_min));
1217		si_pm4_set_reg(pm4, R_028024_DB_DEPTH_BOUNDS_MAX, fui(state->depth.bounds_max));
1218	}
1219
1220	dsa->depth_enabled = state->depth.enabled;
1221	dsa->depth_write_enabled = state->depth.enabled &&
1222				   state->depth.writemask;
1223	dsa->stencil_enabled = state->stencil[0].enabled;
1224	dsa->stencil_write_enabled = state->stencil[0].enabled &&
1225				     (si_dsa_writes_stencil(&state->stencil[0]) ||
1226				      si_dsa_writes_stencil(&state->stencil[1]));
1227	dsa->db_can_write = dsa->depth_write_enabled ||
1228			    dsa->stencil_write_enabled;
1229
1230	bool zfunc_is_ordered =
1231		state->depth.func == PIPE_FUNC_NEVER ||
1232		state->depth.func == PIPE_FUNC_LESS ||
1233		state->depth.func == PIPE_FUNC_LEQUAL ||
1234		state->depth.func == PIPE_FUNC_GREATER ||
1235		state->depth.func == PIPE_FUNC_GEQUAL;
1236
1237	bool nozwrite_and_order_invariant_stencil =
1238		!dsa->db_can_write ||
1239		(!dsa->depth_write_enabled &&
1240		 si_order_invariant_stencil_state(&state->stencil[0]) &&
1241		 si_order_invariant_stencil_state(&state->stencil[1]));
1242
1243	dsa->order_invariance[1].zs =
1244		nozwrite_and_order_invariant_stencil ||
1245		(!dsa->stencil_write_enabled && zfunc_is_ordered);
1246	dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
1247
1248	dsa->order_invariance[1].pass_set =
1249		nozwrite_and_order_invariant_stencil ||
1250		(!dsa->stencil_write_enabled &&
1251		 (state->depth.func == PIPE_FUNC_ALWAYS ||
1252		  state->depth.func == PIPE_FUNC_NEVER));
1253	dsa->order_invariance[0].pass_set =
1254		!dsa->depth_write_enabled ||
1255		(state->depth.func == PIPE_FUNC_ALWAYS ||
1256		 state->depth.func == PIPE_FUNC_NEVER);
1257
1258	dsa->order_invariance[1].pass_last =
1259		sctx->screen->assume_no_z_fights &&
1260		!dsa->stencil_write_enabled &&
1261		dsa->depth_write_enabled && zfunc_is_ordered;
1262	dsa->order_invariance[0].pass_last =
1263		sctx->screen->assume_no_z_fights &&
1264		dsa->depth_write_enabled && zfunc_is_ordered;
1265
1266	return dsa;
1267}
1268
1269static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1270{
1271        struct si_context *sctx = (struct si_context *)ctx;
1272	struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
1273        struct si_state_dsa *dsa = state;
1274
1275        if (!state)
1276                return;
1277
1278	si_pm4_bind_state(sctx, dsa, dsa);
1279
1280	if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1281		   sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1282		sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1283		si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1284	}
1285
1286	if (!old_dsa || old_dsa->alpha_func != dsa->alpha_func)
1287		sctx->do_update_shaders = true;
1288
1289	if (sctx->screen->dpbb_allowed &&
1290	    (!old_dsa ||
1291	     (old_dsa->depth_enabled != dsa->depth_enabled ||
1292	      old_dsa->stencil_enabled != dsa->stencil_enabled ||
1293	      old_dsa->db_can_write != dsa->db_can_write)))
1294		si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1295
1296	if (sctx->screen->has_out_of_order_rast &&
1297	    (!old_dsa ||
1298	     memcmp(old_dsa->order_invariance, dsa->order_invariance,
1299		    sizeof(old_dsa->order_invariance))))
1300		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1301}
1302
1303static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1304{
1305	struct si_context *sctx = (struct si_context *)ctx;
1306	si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
1307}
1308
1309static void *si_create_db_flush_dsa(struct si_context *sctx)
1310{
1311	struct pipe_depth_stencil_alpha_state dsa = {};
1312
1313	return sctx->b.create_depth_stencil_alpha_state(&sctx->b, &dsa);
1314}
1315
1316/* DB RENDER STATE */
1317
1318static void si_set_active_query_state(struct pipe_context *ctx, boolean enable)
1319{
1320	struct si_context *sctx = (struct si_context*)ctx;
1321
1322	/* Pipeline stat & streamout queries. */
1323	if (enable) {
1324		sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
1325		sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
1326	} else {
1327		sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
1328		sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
1329	}
1330
1331	/* Occlusion queries. */
1332	if (sctx->occlusion_queries_disabled != !enable) {
1333		sctx->occlusion_queries_disabled = !enable;
1334		si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1335	}
1336}
1337
1338void si_set_occlusion_query_state(struct si_context *sctx,
1339				  bool old_perfect_enable)
1340{
1341	si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1342
1343	bool perfect_enable = sctx->num_perfect_occlusion_queries != 0;
1344
1345	if (perfect_enable != old_perfect_enable)
1346		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1347}
1348
1349void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1350{
1351	st->saved_compute = sctx->cs_shader_state.program;
1352
1353	si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1354	si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
1355
1356	st->saved_ssbo_writable_mask = 0;
1357
1358	for (unsigned i = 0; i < 3; i++) {
1359		if (sctx->const_and_shader_buffers[PIPE_SHADER_COMPUTE].writable_mask &
1360		    (1u << si_get_shaderbuf_slot(i)))
1361			st->saved_ssbo_writable_mask |= 1 << i;
1362	}
1363}
1364
1365static void si_emit_db_render_state(struct si_context *sctx)
1366{
1367	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1368	unsigned db_shader_control, db_render_control, db_count_control;
1369	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1370
1371	/* DB_RENDER_CONTROL */
1372	if (sctx->dbcb_depth_copy_enabled ||
1373	    sctx->dbcb_stencil_copy_enabled) {
1374		db_render_control =
1375			S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1376			S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1377			S_028000_COPY_CENTROID(1) |
1378			S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample);
1379	} else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1380		db_render_control =
1381			S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1382			S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace);
1383	} else {
1384		db_render_control =
1385			S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1386			S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear);
1387	}
1388
1389	/* DB_COUNT_CONTROL (occlusion queries) */
1390	if (sctx->num_occlusion_queries > 0 &&
1391	    !sctx->occlusion_queries_disabled) {
1392		bool perfect = sctx->num_perfect_occlusion_queries > 0;
1393
1394		if (sctx->chip_class >= CIK) {
1395			unsigned log_sample_rate = sctx->framebuffer.log_samples;
1396
1397			/* Stoney doesn't increment occlusion query counters
1398			 * if the sample rate is 16x. Use 8x sample rate instead.
1399			 */
1400			if (sctx->family == CHIP_STONEY)
1401				log_sample_rate = MIN2(log_sample_rate, 3);
1402
1403			db_count_control =
1404				S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1405				S_028004_SAMPLE_RATE(log_sample_rate) |
1406				S_028004_ZPASS_ENABLE(1) |
1407				S_028004_SLICE_EVEN_ENABLE(1) |
1408				S_028004_SLICE_ODD_ENABLE(1);
1409		} else {
1410			db_count_control =
1411				S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1412				S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples);
1413		}
1414	} else {
1415		/* Disable occlusion queries. */
1416		if (sctx->chip_class >= CIK) {
1417			db_count_control = 0;
1418		} else {
1419			db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
1420		}
1421	}
1422
1423	radeon_opt_set_context_reg2(sctx, R_028000_DB_RENDER_CONTROL,
1424				    SI_TRACKED_DB_RENDER_CONTROL, db_render_control,
1425				    db_count_control);
1426
1427	/* DB_RENDER_OVERRIDE2 */
1428	radeon_opt_set_context_reg(sctx,  R_028010_DB_RENDER_OVERRIDE2,
1429		SI_TRACKED_DB_RENDER_OVERRIDE2,
1430		S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1431		S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1432		S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4));
1433
1434	db_shader_control = sctx->ps_db_shader_control;
1435
1436	/* Bug workaround for smoothing (overrasterization) on SI. */
1437	if (sctx->chip_class == SI && sctx->smoothing_enabled) {
1438		db_shader_control &= C_02880C_Z_ORDER;
1439		db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z);
1440	}
1441
1442	/* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1443	if (!rs->multisample_enable)
1444		db_shader_control &= C_02880C_MASK_EXPORT_ENABLE;
1445
1446	if (sctx->screen->has_rbplus &&
1447	    !sctx->screen->rbplus_allowed)
1448		db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1);
1449
1450	radeon_opt_set_context_reg(sctx, R_02880C_DB_SHADER_CONTROL,
1451				   SI_TRACKED_DB_SHADER_CONTROL, db_shader_control);
1452
1453	if (initial_cdw != sctx->gfx_cs->current.cdw)
1454		sctx->context_roll = true;
1455}
1456
1457/*
1458 * format translation
1459 */
1460static uint32_t si_translate_colorformat(enum pipe_format format)
1461{
1462	const struct util_format_description *desc = util_format_description(format);
1463	if (!desc)
1464		return V_028C70_COLOR_INVALID;
1465
1466#define HAS_SIZE(x,y,z,w) \
1467	(desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1468         desc->channel[2].size == (z) && desc->channel[3].size == (w))
1469
1470	if (format == PIPE_FORMAT_R11G11B10_FLOAT) /* isn't plain */
1471		return V_028C70_COLOR_10_11_11;
1472
1473	if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
1474		return V_028C70_COLOR_INVALID;
1475
1476	/* hw cannot support mixed formats (except depth/stencil, since
1477	 * stencil is not written to). */
1478	if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1479		return V_028C70_COLOR_INVALID;
1480
1481	switch (desc->nr_channels) {
1482	case 1:
1483		switch (desc->channel[0].size) {
1484		case 8:
1485			return V_028C70_COLOR_8;
1486		case 16:
1487			return V_028C70_COLOR_16;
1488		case 32:
1489			return V_028C70_COLOR_32;
1490		}
1491		break;
1492	case 2:
1493		if (desc->channel[0].size == desc->channel[1].size) {
1494			switch (desc->channel[0].size) {
1495			case 8:
1496				return V_028C70_COLOR_8_8;
1497			case 16:
1498				return V_028C70_COLOR_16_16;
1499			case 32:
1500				return V_028C70_COLOR_32_32;
1501			}
1502		} else if (HAS_SIZE(8,24,0,0)) {
1503			return V_028C70_COLOR_24_8;
1504		} else if (HAS_SIZE(24,8,0,0)) {
1505			return V_028C70_COLOR_8_24;
1506		}
1507		break;
1508	case 3:
1509		if (HAS_SIZE(5,6,5,0)) {
1510			return V_028C70_COLOR_5_6_5;
1511		} else if (HAS_SIZE(32,8,24,0)) {
1512			return V_028C70_COLOR_X24_8_32_FLOAT;
1513		}
1514		break;
1515	case 4:
1516		if (desc->channel[0].size == desc->channel[1].size &&
1517		    desc->channel[0].size == desc->channel[2].size &&
1518		    desc->channel[0].size == desc->channel[3].size) {
1519			switch (desc->channel[0].size) {
1520			case 4:
1521				return V_028C70_COLOR_4_4_4_4;
1522			case 8:
1523				return V_028C70_COLOR_8_8_8_8;
1524			case 16:
1525				return V_028C70_COLOR_16_16_16_16;
1526			case 32:
1527				return V_028C70_COLOR_32_32_32_32;
1528			}
1529		} else if (HAS_SIZE(5,5,5,1)) {
1530			return V_028C70_COLOR_1_5_5_5;
1531		} else if (HAS_SIZE(1,5,5,5)) {
1532			return V_028C70_COLOR_5_5_5_1;
1533		} else if (HAS_SIZE(10,10,10,2)) {
1534			return V_028C70_COLOR_2_10_10_10;
1535		}
1536		break;
1537	}
1538	return V_028C70_COLOR_INVALID;
1539}
1540
1541static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
1542{
1543	if (SI_BIG_ENDIAN) {
1544		switch(colorformat) {
1545		/* 8-bit buffers. */
1546		case V_028C70_COLOR_8:
1547			return V_028C70_ENDIAN_NONE;
1548
1549		/* 16-bit buffers. */
1550		case V_028C70_COLOR_5_6_5:
1551		case V_028C70_COLOR_1_5_5_5:
1552		case V_028C70_COLOR_4_4_4_4:
1553		case V_028C70_COLOR_16:
1554		case V_028C70_COLOR_8_8:
1555			return V_028C70_ENDIAN_8IN16;
1556
1557		/* 32-bit buffers. */
1558		case V_028C70_COLOR_8_8_8_8:
1559		case V_028C70_COLOR_2_10_10_10:
1560		case V_028C70_COLOR_8_24:
1561		case V_028C70_COLOR_24_8:
1562		case V_028C70_COLOR_16_16:
1563			return V_028C70_ENDIAN_8IN32;
1564
1565		/* 64-bit buffers. */
1566		case V_028C70_COLOR_16_16_16_16:
1567			return V_028C70_ENDIAN_8IN16;
1568
1569		case V_028C70_COLOR_32_32:
1570			return V_028C70_ENDIAN_8IN32;
1571
1572		/* 128-bit buffers. */
1573		case V_028C70_COLOR_32_32_32_32:
1574			return V_028C70_ENDIAN_8IN32;
1575		default:
1576			return V_028C70_ENDIAN_NONE; /* Unsupported. */
1577		}
1578	} else {
1579		return V_028C70_ENDIAN_NONE;
1580	}
1581}
1582
1583static uint32_t si_translate_dbformat(enum pipe_format format)
1584{
1585	switch (format) {
1586	case PIPE_FORMAT_Z16_UNORM:
1587		return V_028040_Z_16;
1588	case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1589	case PIPE_FORMAT_X8Z24_UNORM:
1590	case PIPE_FORMAT_Z24X8_UNORM:
1591	case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1592		return V_028040_Z_24; /* deprecated on SI */
1593	case PIPE_FORMAT_Z32_FLOAT:
1594	case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1595		return V_028040_Z_32_FLOAT;
1596	default:
1597		return V_028040_Z_INVALID;
1598	}
1599}
1600
1601/*
1602 * Texture translation
1603 */
1604
1605static uint32_t si_translate_texformat(struct pipe_screen *screen,
1606				       enum pipe_format format,
1607				       const struct util_format_description *desc,
1608				       int first_non_void)
1609{
1610	struct si_screen *sscreen = (struct si_screen*)screen;
1611	bool uniform = true;
1612	int i;
1613
1614	/* Colorspace (return non-RGB formats directly). */
1615	switch (desc->colorspace) {
1616	/* Depth stencil formats */
1617	case UTIL_FORMAT_COLORSPACE_ZS:
1618		switch (format) {
1619		case PIPE_FORMAT_Z16_UNORM:
1620			return V_008F14_IMG_DATA_FORMAT_16;
1621		case PIPE_FORMAT_X24S8_UINT:
1622		case PIPE_FORMAT_S8X24_UINT:
1623			/*
1624			 * Implemented as an 8_8_8_8 data format to fix texture
1625			 * gathers in stencil sampling. This affects at least
1626			 * GL45-CTS.texture_cube_map_array.sampling on VI.
1627			 */
1628			if (sscreen->info.chip_class <= VI)
1629				return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1630
1631			if (format == PIPE_FORMAT_X24S8_UINT)
1632				return V_008F14_IMG_DATA_FORMAT_8_24;
1633			else
1634				return V_008F14_IMG_DATA_FORMAT_24_8;
1635		case PIPE_FORMAT_Z24X8_UNORM:
1636		case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1637			return V_008F14_IMG_DATA_FORMAT_8_24;
1638		case PIPE_FORMAT_X8Z24_UNORM:
1639		case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1640			return V_008F14_IMG_DATA_FORMAT_24_8;
1641		case PIPE_FORMAT_S8_UINT:
1642			return V_008F14_IMG_DATA_FORMAT_8;
1643		case PIPE_FORMAT_Z32_FLOAT:
1644			return V_008F14_IMG_DATA_FORMAT_32;
1645		case PIPE_FORMAT_X32_S8X24_UINT:
1646		case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1647			return V_008F14_IMG_DATA_FORMAT_X24_8_32;
1648		default:
1649			goto out_unknown;
1650		}
1651
1652	case UTIL_FORMAT_COLORSPACE_YUV:
1653		goto out_unknown; /* TODO */
1654
1655	case UTIL_FORMAT_COLORSPACE_SRGB:
1656		if (desc->nr_channels != 4 && desc->nr_channels != 1)
1657			goto out_unknown;
1658		break;
1659
1660	default:
1661		break;
1662	}
1663
1664	if (desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
1665		if (!sscreen->info.has_format_bc1_through_bc7)
1666			goto out_unknown;
1667
1668		switch (format) {
1669		case PIPE_FORMAT_RGTC1_SNORM:
1670		case PIPE_FORMAT_LATC1_SNORM:
1671		case PIPE_FORMAT_RGTC1_UNORM:
1672		case PIPE_FORMAT_LATC1_UNORM:
1673			return V_008F14_IMG_DATA_FORMAT_BC4;
1674		case PIPE_FORMAT_RGTC2_SNORM:
1675		case PIPE_FORMAT_LATC2_SNORM:
1676		case PIPE_FORMAT_RGTC2_UNORM:
1677		case PIPE_FORMAT_LATC2_UNORM:
1678			return V_008F14_IMG_DATA_FORMAT_BC5;
1679		default:
1680			goto out_unknown;
1681		}
1682	}
1683
1684	if (desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
1685	    (sscreen->info.family == CHIP_STONEY ||
1686	     sscreen->info.family == CHIP_VEGA10 ||
1687	     sscreen->info.family == CHIP_RAVEN)) {
1688		switch (format) {
1689		case PIPE_FORMAT_ETC1_RGB8:
1690		case PIPE_FORMAT_ETC2_RGB8:
1691		case PIPE_FORMAT_ETC2_SRGB8:
1692			return V_008F14_IMG_DATA_FORMAT_ETC2_RGB;
1693		case PIPE_FORMAT_ETC2_RGB8A1:
1694		case PIPE_FORMAT_ETC2_SRGB8A1:
1695			return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1;
1696		case PIPE_FORMAT_ETC2_RGBA8:
1697		case PIPE_FORMAT_ETC2_SRGBA8:
1698			return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA;
1699		case PIPE_FORMAT_ETC2_R11_UNORM:
1700		case PIPE_FORMAT_ETC2_R11_SNORM:
1701			return V_008F14_IMG_DATA_FORMAT_ETC2_R;
1702		case PIPE_FORMAT_ETC2_RG11_UNORM:
1703		case PIPE_FORMAT_ETC2_RG11_SNORM:
1704			return V_008F14_IMG_DATA_FORMAT_ETC2_RG;
1705		default:
1706			goto out_unknown;
1707		}
1708	}
1709
1710	if (desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
1711		if (!sscreen->info.has_format_bc1_through_bc7)
1712			goto out_unknown;
1713
1714		switch (format) {
1715		case PIPE_FORMAT_BPTC_RGBA_UNORM:
1716		case PIPE_FORMAT_BPTC_SRGBA:
1717			return V_008F14_IMG_DATA_FORMAT_BC7;
1718		case PIPE_FORMAT_BPTC_RGB_FLOAT:
1719		case PIPE_FORMAT_BPTC_RGB_UFLOAT:
1720			return V_008F14_IMG_DATA_FORMAT_BC6;
1721		default:
1722			goto out_unknown;
1723		}
1724	}
1725
1726	if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
1727		switch (format) {
1728		case PIPE_FORMAT_R8G8_B8G8_UNORM:
1729		case PIPE_FORMAT_G8R8_B8R8_UNORM:
1730			return V_008F14_IMG_DATA_FORMAT_GB_GR;
1731		case PIPE_FORMAT_G8R8_G8B8_UNORM:
1732		case PIPE_FORMAT_R8G8_R8B8_UNORM:
1733			return V_008F14_IMG_DATA_FORMAT_BG_RG;
1734		default:
1735			goto out_unknown;
1736		}
1737	}
1738
1739	if (desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
1740		if (!sscreen->info.has_format_bc1_through_bc7)
1741			goto out_unknown;
1742
1743		switch (format) {
1744		case PIPE_FORMAT_DXT1_RGB:
1745		case PIPE_FORMAT_DXT1_RGBA:
1746		case PIPE_FORMAT_DXT1_SRGB:
1747		case PIPE_FORMAT_DXT1_SRGBA:
1748			return V_008F14_IMG_DATA_FORMAT_BC1;
1749		case PIPE_FORMAT_DXT3_RGBA:
1750		case PIPE_FORMAT_DXT3_SRGBA:
1751			return V_008F14_IMG_DATA_FORMAT_BC2;
1752		case PIPE_FORMAT_DXT5_RGBA:
1753		case PIPE_FORMAT_DXT5_SRGBA:
1754			return V_008F14_IMG_DATA_FORMAT_BC3;
1755		default:
1756			goto out_unknown;
1757		}
1758	}
1759
1760	if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1761		return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
1762	} else if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
1763		return V_008F14_IMG_DATA_FORMAT_10_11_11;
1764	}
1765
1766	/* R8G8Bx_SNORM - TODO CxV8U8 */
1767
1768	/* hw cannot support mixed formats (except depth/stencil, since only
1769	 * depth is read).*/
1770	if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1771		goto out_unknown;
1772
1773	/* See whether the components are of the same size. */
1774	for (i = 1; i < desc->nr_channels; i++) {
1775		uniform = uniform && desc->channel[0].size == desc->channel[i].size;
1776	}
1777
1778	/* Non-uniform formats. */
1779	if (!uniform) {
1780		switch(desc->nr_channels) {
1781		case 3:
1782			if (desc->channel[0].size == 5 &&
1783			    desc->channel[1].size == 6 &&
1784			    desc->channel[2].size == 5) {
1785				return V_008F14_IMG_DATA_FORMAT_5_6_5;
1786			}
1787			goto out_unknown;
1788		case 4:
1789			if (desc->channel[0].size == 5 &&
1790			    desc->channel[1].size == 5 &&
1791			    desc->channel[2].size == 5 &&
1792			    desc->channel[3].size == 1) {
1793				return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
1794			}
1795			if (desc->channel[0].size == 1 &&
1796			    desc->channel[1].size == 5 &&
1797			    desc->channel[2].size == 5 &&
1798			    desc->channel[3].size == 5) {
1799				return V_008F14_IMG_DATA_FORMAT_5_5_5_1;
1800			}
1801			if (desc->channel[0].size == 10 &&
1802			    desc->channel[1].size == 10 &&
1803			    desc->channel[2].size == 10 &&
1804			    desc->channel[3].size == 2) {
1805				return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
1806			}
1807			goto out_unknown;
1808		}
1809		goto out_unknown;
1810	}
1811
1812	if (first_non_void < 0 || first_non_void > 3)
1813		goto out_unknown;
1814
1815	/* uniform formats */
1816	switch (desc->channel[first_non_void].size) {
1817	case 4:
1818		switch (desc->nr_channels) {
1819#if 0 /* Not supported for render targets */
1820		case 2:
1821			return V_008F14_IMG_DATA_FORMAT_4_4;
1822#endif
1823		case 4:
1824			return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
1825		}
1826		break;
1827	case 8:
1828		switch (desc->nr_channels) {
1829		case 1:
1830			return V_008F14_IMG_DATA_FORMAT_8;
1831		case 2:
1832			return V_008F14_IMG_DATA_FORMAT_8_8;
1833		case 4:
1834			return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1835		}
1836		break;
1837	case 16:
1838		switch (desc->nr_channels) {
1839		case 1:
1840			return V_008F14_IMG_DATA_FORMAT_16;
1841		case 2:
1842			return V_008F14_IMG_DATA_FORMAT_16_16;
1843		case 4:
1844			return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
1845		}
1846		break;
1847	case 32:
1848		switch (desc->nr_channels) {
1849		case 1:
1850			return V_008F14_IMG_DATA_FORMAT_32;
1851		case 2:
1852			return V_008F14_IMG_DATA_FORMAT_32_32;
1853#if 0 /* Not supported for render targets */
1854		case 3:
1855			return V_008F14_IMG_DATA_FORMAT_32_32_32;
1856#endif
1857		case 4:
1858			return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
1859		}
1860	}
1861
1862out_unknown:
1863	return ~0;
1864}
1865
1866static unsigned si_tex_wrap(unsigned wrap)
1867{
1868	switch (wrap) {
1869	default:
1870	case PIPE_TEX_WRAP_REPEAT:
1871		return V_008F30_SQ_TEX_WRAP;
1872	case PIPE_TEX_WRAP_CLAMP:
1873		return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
1874	case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
1875		return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
1876	case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1877		return V_008F30_SQ_TEX_CLAMP_BORDER;
1878	case PIPE_TEX_WRAP_MIRROR_REPEAT:
1879		return V_008F30_SQ_TEX_MIRROR;
1880	case PIPE_TEX_WRAP_MIRROR_CLAMP:
1881		return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
1882	case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
1883		return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
1884	case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1885		return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
1886	}
1887}
1888
1889static unsigned si_tex_mipfilter(unsigned filter)
1890{
1891	switch (filter) {
1892	case PIPE_TEX_MIPFILTER_NEAREST:
1893		return V_008F38_SQ_TEX_Z_FILTER_POINT;
1894	case PIPE_TEX_MIPFILTER_LINEAR:
1895		return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
1896	default:
1897	case PIPE_TEX_MIPFILTER_NONE:
1898		return V_008F38_SQ_TEX_Z_FILTER_NONE;
1899	}
1900}
1901
1902static unsigned si_tex_compare(unsigned compare)
1903{
1904	switch (compare) {
1905	default:
1906	case PIPE_FUNC_NEVER:
1907		return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
1908	case PIPE_FUNC_LESS:
1909		return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
1910	case PIPE_FUNC_EQUAL:
1911		return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
1912	case PIPE_FUNC_LEQUAL:
1913		return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
1914	case PIPE_FUNC_GREATER:
1915		return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
1916	case PIPE_FUNC_NOTEQUAL:
1917		return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
1918	case PIPE_FUNC_GEQUAL:
1919		return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
1920	case PIPE_FUNC_ALWAYS:
1921		return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
1922	}
1923}
1924
1925static unsigned si_tex_dim(struct si_screen *sscreen, struct si_texture *tex,
1926			   unsigned view_target, unsigned nr_samples)
1927{
1928	unsigned res_target = tex->buffer.b.b.target;
1929
1930	if (view_target == PIPE_TEXTURE_CUBE ||
1931	    view_target == PIPE_TEXTURE_CUBE_ARRAY)
1932		res_target = view_target;
1933	/* If interpreting cubemaps as something else, set 2D_ARRAY. */
1934	else if (res_target == PIPE_TEXTURE_CUBE ||
1935		 res_target == PIPE_TEXTURE_CUBE_ARRAY)
1936		res_target = PIPE_TEXTURE_2D_ARRAY;
1937
1938	/* GFX9 allocates 1D textures as 2D. */
1939	if ((res_target == PIPE_TEXTURE_1D ||
1940	     res_target == PIPE_TEXTURE_1D_ARRAY) &&
1941	    sscreen->info.chip_class >= GFX9 &&
1942	    tex->surface.u.gfx9.resource_type == RADEON_RESOURCE_2D) {
1943		if (res_target == PIPE_TEXTURE_1D)
1944			res_target = PIPE_TEXTURE_2D;
1945		else
1946			res_target = PIPE_TEXTURE_2D_ARRAY;
1947	}
1948
1949	switch (res_target) {
1950	default:
1951	case PIPE_TEXTURE_1D:
1952		return V_008F1C_SQ_RSRC_IMG_1D;
1953	case PIPE_TEXTURE_1D_ARRAY:
1954		return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
1955	case PIPE_TEXTURE_2D:
1956	case PIPE_TEXTURE_RECT:
1957		return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA :
1958					V_008F1C_SQ_RSRC_IMG_2D;
1959	case PIPE_TEXTURE_2D_ARRAY:
1960		return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY :
1961					V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
1962	case PIPE_TEXTURE_3D:
1963		return V_008F1C_SQ_RSRC_IMG_3D;
1964	case PIPE_TEXTURE_CUBE:
1965	case PIPE_TEXTURE_CUBE_ARRAY:
1966		return V_008F1C_SQ_RSRC_IMG_CUBE;
1967	}
1968}
1969
1970/*
1971 * Format support testing
1972 */
1973
1974static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
1975{
1976	const struct util_format_description *desc = util_format_description(format);
1977	if (!desc)
1978		return false;
1979
1980	return si_translate_texformat(screen, format, desc,
1981				      util_format_get_first_non_void_channel(format)) != ~0U;
1982}
1983
1984static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
1985					       const struct util_format_description *desc,
1986					       int first_non_void)
1987{
1988	int i;
1989
1990	if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
1991		return V_008F0C_BUF_DATA_FORMAT_10_11_11;
1992
1993	assert(first_non_void >= 0);
1994
1995	if (desc->nr_channels == 4 &&
1996	    desc->channel[0].size == 10 &&
1997	    desc->channel[1].size == 10 &&
1998	    desc->channel[2].size == 10 &&
1999	    desc->channel[3].size == 2)
2000		return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
2001
2002	/* See whether the components are of the same size. */
2003	for (i = 0; i < desc->nr_channels; i++) {
2004		if (desc->channel[first_non_void].size != desc->channel[i].size)
2005			return V_008F0C_BUF_DATA_FORMAT_INVALID;
2006	}
2007
2008	switch (desc->channel[first_non_void].size) {
2009	case 8:
2010		switch (desc->nr_channels) {
2011		case 1:
2012		case 3: /* 3 loads */
2013			return V_008F0C_BUF_DATA_FORMAT_8;
2014		case 2:
2015			return V_008F0C_BUF_DATA_FORMAT_8_8;
2016		case 4:
2017			return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
2018		}
2019		break;
2020	case 16:
2021		switch (desc->nr_channels) {
2022		case 1:
2023		case 3: /* 3 loads */
2024			return V_008F0C_BUF_DATA_FORMAT_16;
2025		case 2:
2026			return V_008F0C_BUF_DATA_FORMAT_16_16;
2027		case 4:
2028			return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
2029		}
2030		break;
2031	case 32:
2032		switch (desc->nr_channels) {
2033		case 1:
2034			return V_008F0C_BUF_DATA_FORMAT_32;
2035		case 2:
2036			return V_008F0C_BUF_DATA_FORMAT_32_32;
2037		case 3:
2038			return V_008F0C_BUF_DATA_FORMAT_32_32_32;
2039		case 4:
2040			return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2041		}
2042		break;
2043	case 64:
2044		/* Legacy double formats. */
2045		switch (desc->nr_channels) {
2046		case 1: /* 1 load */
2047			return V_008F0C_BUF_DATA_FORMAT_32_32;
2048		case 2: /* 1 load */
2049			return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2050		case 3: /* 3 loads */
2051			return V_008F0C_BUF_DATA_FORMAT_32_32;
2052		case 4: /* 2 loads */
2053			return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2054		}
2055		break;
2056	}
2057
2058	return V_008F0C_BUF_DATA_FORMAT_INVALID;
2059}
2060
2061static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
2062					      const struct util_format_description *desc,
2063					      int first_non_void)
2064{
2065	if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
2066		return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2067
2068	assert(first_non_void >= 0);
2069
2070	switch (desc->channel[first_non_void].type) {
2071	case UTIL_FORMAT_TYPE_SIGNED:
2072	case UTIL_FORMAT_TYPE_FIXED:
2073		if (desc->channel[first_non_void].size >= 32 ||
2074		    desc->channel[first_non_void].pure_integer)
2075			return V_008F0C_BUF_NUM_FORMAT_SINT;
2076		else if (desc->channel[first_non_void].normalized)
2077			return V_008F0C_BUF_NUM_FORMAT_SNORM;
2078		else
2079			return V_008F0C_BUF_NUM_FORMAT_SSCALED;
2080		break;
2081	case UTIL_FORMAT_TYPE_UNSIGNED:
2082		if (desc->channel[first_non_void].size >= 32 ||
2083		    desc->channel[first_non_void].pure_integer)
2084			return V_008F0C_BUF_NUM_FORMAT_UINT;
2085		else if (desc->channel[first_non_void].normalized)
2086			return V_008F0C_BUF_NUM_FORMAT_UNORM;
2087		else
2088			return V_008F0C_BUF_NUM_FORMAT_USCALED;
2089		break;
2090	case UTIL_FORMAT_TYPE_FLOAT:
2091	default:
2092		return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2093	}
2094}
2095
2096static unsigned si_is_vertex_format_supported(struct pipe_screen *screen,
2097					      enum pipe_format format,
2098					      unsigned usage)
2099{
2100	const struct util_format_description *desc;
2101	int first_non_void;
2102	unsigned data_format;
2103
2104	assert((usage & ~(PIPE_BIND_SHADER_IMAGE |
2105			  PIPE_BIND_SAMPLER_VIEW |
2106			  PIPE_BIND_VERTEX_BUFFER)) == 0);
2107
2108	desc = util_format_description(format);
2109	if (!desc)
2110		return 0;
2111
2112	/* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2113	 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2114	 * for read-only access (with caveats surrounding bounds checks), but
2115	 * obviously fails for write access which we have to implement for
2116	 * shader images. Luckily, OpenGL doesn't expect this to be supported
2117	 * anyway, and so the only impact is on PBO uploads / downloads, which
2118	 * shouldn't be expected to be fast for GL_RGB anyway.
2119	 */
2120	if (desc->block.bits == 3 * 8 ||
2121	    desc->block.bits == 3 * 16) {
2122		if (usage & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW)) {
2123		    usage &= ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW);
2124			if (!usage)
2125				return 0;
2126		}
2127	}
2128
2129	first_non_void = util_format_get_first_non_void_channel(format);
2130	data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
2131	if (data_format == V_008F0C_BUF_DATA_FORMAT_INVALID)
2132		return 0;
2133
2134	return usage;
2135}
2136
2137static bool si_is_colorbuffer_format_supported(enum pipe_format format)
2138{
2139	return si_translate_colorformat(format) != V_028C70_COLOR_INVALID &&
2140		si_translate_colorswap(format, false) != ~0U;
2141}
2142
2143static bool si_is_zs_format_supported(enum pipe_format format)
2144{
2145	return si_translate_dbformat(format) != V_028040_Z_INVALID;
2146}
2147
2148static boolean si_is_format_supported(struct pipe_screen *screen,
2149				      enum pipe_format format,
2150				      enum pipe_texture_target target,
2151				      unsigned sample_count,
2152				      unsigned storage_sample_count,
2153				      unsigned usage)
2154{
2155	struct si_screen *sscreen = (struct si_screen *)screen;
2156	unsigned retval = 0;
2157
2158	if (target >= PIPE_MAX_TEXTURE_TYPES) {
2159		PRINT_ERR("radeonsi: unsupported texture type %d\n", target);
2160		return false;
2161	}
2162
2163	if (MAX2(1, sample_count) < MAX2(1, storage_sample_count))
2164		return false;
2165
2166	if (sample_count > 1) {
2167		if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
2168			return false;
2169
2170		if (usage & PIPE_BIND_SHADER_IMAGE)
2171			return false;
2172
2173		/* Only power-of-two sample counts are supported. */
2174		if (!util_is_power_of_two_or_zero(sample_count) ||
2175		    !util_is_power_of_two_or_zero(storage_sample_count))
2176			return false;
2177
2178		/* MSAA support without framebuffer attachments. */
2179		if (format == PIPE_FORMAT_NONE && sample_count <= 16)
2180			return true;
2181
2182		if (!sscreen->info.has_eqaa_surface_allocator ||
2183		    util_format_is_depth_or_stencil(format)) {
2184			/* Color without EQAA or depth/stencil. */
2185			if (sample_count > 8 ||
2186			    sample_count != storage_sample_count)
2187				return false;
2188		} else {
2189			/* Color with EQAA. */
2190			if (sample_count > 16 ||
2191			    storage_sample_count > 8)
2192				return false;
2193		}
2194	}
2195
2196	if (usage & (PIPE_BIND_SAMPLER_VIEW |
2197		     PIPE_BIND_SHADER_IMAGE)) {
2198		if (target == PIPE_BUFFER) {
2199			retval |= si_is_vertex_format_supported(
2200				screen, format, usage & (PIPE_BIND_SAMPLER_VIEW |
2201						         PIPE_BIND_SHADER_IMAGE));
2202		} else {
2203			if (si_is_sampler_format_supported(screen, format))
2204				retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
2205						   PIPE_BIND_SHADER_IMAGE);
2206		}
2207	}
2208
2209	if ((usage & (PIPE_BIND_RENDER_TARGET |
2210		      PIPE_BIND_DISPLAY_TARGET |
2211		      PIPE_BIND_SCANOUT |
2212		      PIPE_BIND_SHARED |
2213		      PIPE_BIND_BLENDABLE)) &&
2214	    si_is_colorbuffer_format_supported(format)) {
2215		retval |= usage &
2216			  (PIPE_BIND_RENDER_TARGET |
2217			   PIPE_BIND_DISPLAY_TARGET |
2218			   PIPE_BIND_SCANOUT |
2219			   PIPE_BIND_SHARED);
2220		if (!util_format_is_pure_integer(format) &&
2221		    !util_format_is_depth_or_stencil(format))
2222			retval |= usage & PIPE_BIND_BLENDABLE;
2223	}
2224
2225	if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
2226	    si_is_zs_format_supported(format)) {
2227		retval |= PIPE_BIND_DEPTH_STENCIL;
2228	}
2229
2230	if (usage & PIPE_BIND_VERTEX_BUFFER) {
2231		retval |= si_is_vertex_format_supported(screen, format,
2232							PIPE_BIND_VERTEX_BUFFER);
2233	}
2234
2235	if ((usage & PIPE_BIND_LINEAR) &&
2236	    !util_format_is_compressed(format) &&
2237	    !(usage & PIPE_BIND_DEPTH_STENCIL))
2238		retval |= PIPE_BIND_LINEAR;
2239
2240	return retval == usage;
2241}
2242
2243/*
2244 * framebuffer handling
2245 */
2246
2247static void si_choose_spi_color_formats(struct si_surface *surf,
2248					unsigned format, unsigned swap,
2249					unsigned ntype, bool is_depth)
2250{
2251	/* Alpha is needed for alpha-to-coverage.
2252	 * Blending may be with or without alpha.
2253	 */
2254	unsigned normal = 0; /* most optimal, may not support blending or export alpha */
2255	unsigned alpha = 0; /* exports alpha, but may not support blending */
2256	unsigned blend = 0; /* supports blending, but may not export alpha */
2257	unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
2258
2259	/* Choose the SPI color formats. These are required values for RB+.
2260	 * Other chips have multiple choices, though they are not necessarily better.
2261	 */
2262	switch (format) {
2263	case V_028C70_COLOR_5_6_5:
2264	case V_028C70_COLOR_1_5_5_5:
2265	case V_028C70_COLOR_5_5_5_1:
2266	case V_028C70_COLOR_4_4_4_4:
2267	case V_028C70_COLOR_10_11_11:
2268	case V_028C70_COLOR_11_11_10:
2269	case V_028C70_COLOR_8:
2270	case V_028C70_COLOR_8_8:
2271	case V_028C70_COLOR_8_8_8_8:
2272	case V_028C70_COLOR_10_10_10_2:
2273	case V_028C70_COLOR_2_10_10_10:
2274		if (ntype == V_028C70_NUMBER_UINT)
2275			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2276		else if (ntype == V_028C70_NUMBER_SINT)
2277			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2278		else
2279			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2280		break;
2281
2282	case V_028C70_COLOR_16:
2283	case V_028C70_COLOR_16_16:
2284	case V_028C70_COLOR_16_16_16_16:
2285		if (ntype == V_028C70_NUMBER_UNORM ||
2286		    ntype == V_028C70_NUMBER_SNORM) {
2287			/* UNORM16 and SNORM16 don't support blending */
2288			if (ntype == V_028C70_NUMBER_UNORM)
2289				normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
2290			else
2291				normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
2292
2293			/* Use 32 bits per channel for blending. */
2294			if (format == V_028C70_COLOR_16) {
2295				if (swap == V_028C70_SWAP_STD) { /* R */
2296					blend = V_028714_SPI_SHADER_32_R;
2297					blend_alpha = V_028714_SPI_SHADER_32_AR;
2298				} else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2299					blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2300				else
2301					assert(0);
2302			} else if (format == V_028C70_COLOR_16_16) {
2303				if (swap == V_028C70_SWAP_STD) { /* RG */
2304					blend = V_028714_SPI_SHADER_32_GR;
2305					blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2306				} else if (swap == V_028C70_SWAP_ALT) /* RA */
2307					blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2308				else
2309					assert(0);
2310			} else /* 16_16_16_16 */
2311				blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2312		} else if (ntype == V_028C70_NUMBER_UINT)
2313			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2314		else if (ntype == V_028C70_NUMBER_SINT)
2315			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2316		else if (ntype == V_028C70_NUMBER_FLOAT)
2317			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2318		else
2319			assert(0);
2320		break;
2321
2322	case V_028C70_COLOR_32:
2323		if (swap == V_028C70_SWAP_STD) { /* R */
2324			blend = normal = V_028714_SPI_SHADER_32_R;
2325			alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
2326		} else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2327			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2328		else
2329			assert(0);
2330		break;
2331
2332	case V_028C70_COLOR_32_32:
2333		if (swap == V_028C70_SWAP_STD) { /* RG */
2334			blend = normal = V_028714_SPI_SHADER_32_GR;
2335			alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2336		} else if (swap == V_028C70_SWAP_ALT) /* RA */
2337			alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2338		else
2339			assert(0);
2340		break;
2341
2342	case V_028C70_COLOR_32_32_32_32:
2343	case V_028C70_COLOR_8_24:
2344	case V_028C70_COLOR_24_8:
2345	case V_028C70_COLOR_X24_8_32_FLOAT:
2346		alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2347		break;
2348
2349	default:
2350		assert(0);
2351		return;
2352	}
2353
2354	/* The DB->CB copy needs 32_ABGR. */
2355	if (is_depth)
2356		alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2357
2358	surf->spi_shader_col_format = normal;
2359	surf->spi_shader_col_format_alpha = alpha;
2360	surf->spi_shader_col_format_blend = blend;
2361	surf->spi_shader_col_format_blend_alpha = blend_alpha;
2362}
2363
2364static void si_initialize_color_surface(struct si_context *sctx,
2365					struct si_surface *surf)
2366{
2367	struct si_texture *tex = (struct si_texture*)surf->base.texture;
2368	unsigned color_info, color_attrib;
2369	unsigned format, swap, ntype, endian;
2370	const struct util_format_description *desc;
2371	int firstchan;
2372	unsigned blend_clamp = 0, blend_bypass = 0;
2373
2374	desc = util_format_description(surf->base.format);
2375	for (firstchan = 0; firstchan < 4; firstchan++) {
2376		if (desc->channel[firstchan].type != UTIL_FORMAT_TYPE_VOID) {
2377			break;
2378		}
2379	}
2380	if (firstchan == 4 || desc->channel[firstchan].type == UTIL_FORMAT_TYPE_FLOAT) {
2381		ntype = V_028C70_NUMBER_FLOAT;
2382	} else {
2383		ntype = V_028C70_NUMBER_UNORM;
2384		if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
2385			ntype = V_028C70_NUMBER_SRGB;
2386		else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_SIGNED) {
2387			if (desc->channel[firstchan].pure_integer) {
2388				ntype = V_028C70_NUMBER_SINT;
2389			} else {
2390				assert(desc->channel[firstchan].normalized);
2391				ntype = V_028C70_NUMBER_SNORM;
2392			}
2393		} else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2394			if (desc->channel[firstchan].pure_integer) {
2395				ntype = V_028C70_NUMBER_UINT;
2396			} else {
2397				assert(desc->channel[firstchan].normalized);
2398				ntype = V_028C70_NUMBER_UNORM;
2399			}
2400		}
2401	}
2402
2403	format = si_translate_colorformat(surf->base.format);
2404	if (format == V_028C70_COLOR_INVALID) {
2405		PRINT_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2406	}
2407	assert(format != V_028C70_COLOR_INVALID);
2408	swap = si_translate_colorswap(surf->base.format, false);
2409	endian = si_colorformat_endian_swap(format);
2410
2411	/* blend clamp should be set for all NORM/SRGB types */
2412	if (ntype == V_028C70_NUMBER_UNORM ||
2413	    ntype == V_028C70_NUMBER_SNORM ||
2414	    ntype == V_028C70_NUMBER_SRGB)
2415		blend_clamp = 1;
2416
2417	/* set blend bypass according to docs if SINT/UINT or
2418	   8/24 COLOR variants */
2419	if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
2420	    format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
2421	    format == V_028C70_COLOR_X24_8_32_FLOAT) {
2422		blend_clamp = 0;
2423		blend_bypass = 1;
2424	}
2425
2426	if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) {
2427		if (format == V_028C70_COLOR_8 ||
2428		    format == V_028C70_COLOR_8_8 ||
2429		    format == V_028C70_COLOR_8_8_8_8)
2430			surf->color_is_int8 = true;
2431		else if (format == V_028C70_COLOR_10_10_10_2 ||
2432			 format == V_028C70_COLOR_2_10_10_10)
2433			surf->color_is_int10 = true;
2434	}
2435
2436	color_info = S_028C70_FORMAT(format) |
2437		S_028C70_COMP_SWAP(swap) |
2438		S_028C70_BLEND_CLAMP(blend_clamp) |
2439		S_028C70_BLEND_BYPASS(blend_bypass) |
2440		S_028C70_SIMPLE_FLOAT(1) |
2441		S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM &&
2442				    ntype != V_028C70_NUMBER_SNORM &&
2443				    ntype != V_028C70_NUMBER_SRGB &&
2444				    format != V_028C70_COLOR_8_24 &&
2445				    format != V_028C70_COLOR_24_8) |
2446		S_028C70_NUMBER_TYPE(ntype) |
2447		S_028C70_ENDIAN(endian);
2448
2449	/* Intensity is implemented as Red, so treat it that way. */
2450	color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == PIPE_SWIZZLE_1 ||
2451						  util_format_is_intensity(surf->base.format));
2452
2453	if (tex->buffer.b.b.nr_samples > 1) {
2454		unsigned log_samples = util_logbase2(tex->buffer.b.b.nr_samples);
2455		unsigned log_fragments = util_logbase2(tex->buffer.b.b.nr_storage_samples);
2456
2457		color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
2458				S_028C74_NUM_FRAGMENTS(log_fragments);
2459
2460		if (tex->surface.fmask_size) {
2461			color_info |= S_028C70_COMPRESSION(1);
2462			unsigned fmask_bankh = util_logbase2(tex->surface.u.legacy.fmask.bankh);
2463
2464			if (sctx->chip_class == SI) {
2465				/* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2466				color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
2467			}
2468		}
2469	}
2470
2471	if (sctx->chip_class >= VI) {
2472		unsigned max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_256B;
2473		unsigned min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_32B;
2474
2475		/* amdvlk: [min-compressed-block-size] should be set to 32 for dGPU and
2476		   64 for APU because all of our APUs to date use DIMMs which have
2477		   a request granularity size of 64B while all other chips have a
2478		   32B request size */
2479		if (!sctx->screen->info.has_dedicated_vram)
2480			min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_64B;
2481
2482		if (tex->buffer.b.b.nr_storage_samples > 1) {
2483			if (tex->surface.bpe == 1)
2484				max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
2485			else if (tex->surface.bpe == 2)
2486				max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_128B;
2487		}
2488
2489		surf->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
2490				       S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
2491		                       S_028C78_INDEPENDENT_64B_BLOCKS(1);
2492	}
2493
2494	/* This must be set for fast clear to work without FMASK. */
2495	if (!tex->surface.fmask_size && sctx->chip_class == SI) {
2496		unsigned bankh = util_logbase2(tex->surface.u.legacy.bankh);
2497		color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
2498	}
2499
2500	unsigned color_view = S_028C6C_SLICE_START(surf->base.u.tex.first_layer) |
2501			      S_028C6C_SLICE_MAX(surf->base.u.tex.last_layer);
2502
2503	if (sctx->chip_class >= GFX9) {
2504		unsigned mip0_depth = util_max_layer(&tex->buffer.b.b, 0);
2505
2506		color_view |= S_028C6C_MIP_LEVEL(surf->base.u.tex.level);
2507		color_attrib |= S_028C74_MIP0_DEPTH(mip0_depth) |
2508				S_028C74_RESOURCE_TYPE(tex->surface.u.gfx9.resource_type);
2509		surf->cb_color_attrib2 = S_028C68_MIP0_WIDTH(surf->width0 - 1) |
2510					 S_028C68_MIP0_HEIGHT(surf->height0 - 1) |
2511					 S_028C68_MAX_MIP(tex->buffer.b.b.last_level);
2512	}
2513
2514	surf->cb_color_view = color_view;
2515	surf->cb_color_info = color_info;
2516	surf->cb_color_attrib = color_attrib;
2517
2518	/* Determine pixel shader export format */
2519	si_choose_spi_color_formats(surf, format, swap, ntype, tex->is_depth);
2520
2521	surf->color_initialized = true;
2522}
2523
2524static void si_init_depth_surface(struct si_context *sctx,
2525				  struct si_surface *surf)
2526{
2527	struct si_texture *tex = (struct si_texture*)surf->base.texture;
2528	unsigned level = surf->base.u.tex.level;
2529	unsigned format, stencil_format;
2530	uint32_t z_info, s_info;
2531
2532	format = si_translate_dbformat(tex->db_render_format);
2533	stencil_format = tex->surface.has_stencil ?
2534				 V_028044_STENCIL_8 : V_028044_STENCIL_INVALID;
2535
2536	assert(format != V_028040_Z_INVALID);
2537	if (format == V_028040_Z_INVALID)
2538		PRINT_ERR("Invalid DB format: %d, disabling DB.\n", tex->buffer.b.b.format);
2539
2540	surf->db_depth_view = S_028008_SLICE_START(surf->base.u.tex.first_layer) |
2541			      S_028008_SLICE_MAX(surf->base.u.tex.last_layer);
2542	surf->db_htile_data_base = 0;
2543	surf->db_htile_surface = 0;
2544
2545	if (sctx->chip_class >= GFX9) {
2546		assert(tex->surface.u.gfx9.surf_offset == 0);
2547		surf->db_depth_base = tex->buffer.gpu_address >> 8;
2548		surf->db_stencil_base = (tex->buffer.gpu_address +
2549					 tex->surface.u.gfx9.stencil_offset) >> 8;
2550		z_info = S_028038_FORMAT(format) |
2551			 S_028038_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples)) |
2552			 S_028038_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
2553			 S_028038_MAXMIP(tex->buffer.b.b.last_level);
2554		s_info = S_02803C_FORMAT(stencil_format) |
2555			 S_02803C_SW_MODE(tex->surface.u.gfx9.stencil.swizzle_mode);
2556		surf->db_z_info2 = S_028068_EPITCH(tex->surface.u.gfx9.surf.epitch);
2557		surf->db_stencil_info2 = S_02806C_EPITCH(tex->surface.u.gfx9.stencil.epitch);
2558		surf->db_depth_view |= S_028008_MIPID(level);
2559		surf->db_depth_size = S_02801C_X_MAX(tex->buffer.b.b.width0 - 1) |
2560				      S_02801C_Y_MAX(tex->buffer.b.b.height0 - 1);
2561
2562		if (si_htile_enabled(tex, level)) {
2563			z_info |= S_028038_TILE_SURFACE_ENABLE(1) |
2564				  S_028038_ALLOW_EXPCLEAR(1);
2565
2566			if (tex->tc_compatible_htile) {
2567				unsigned max_zplanes = 4;
2568
2569				if (tex->db_render_format == PIPE_FORMAT_Z16_UNORM &&
2570				    tex->buffer.b.b.nr_samples > 1)
2571					max_zplanes = 2;
2572
2573				z_info |= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes + 1) |
2574					  S_028038_ITERATE_FLUSH(1);
2575				s_info |= S_02803C_ITERATE_FLUSH(1);
2576			}
2577
2578			if (tex->surface.has_stencil) {
2579				/* Stencil buffer workaround ported from the SI-CI-VI code.
2580				 * See that for explanation.
2581				 */
2582				s_info |= S_02803C_ALLOW_EXPCLEAR(tex->buffer.b.b.nr_samples <= 1);
2583			} else {
2584				/* Use all HTILE for depth if there's no stencil. */
2585				s_info |= S_02803C_TILE_STENCIL_DISABLE(1);
2586			}
2587
2588			surf->db_htile_data_base = (tex->buffer.gpu_address +
2589						    tex->htile_offset) >> 8;
2590			surf->db_htile_surface = S_028ABC_FULL_CACHE(1) |
2591						 S_028ABC_PIPE_ALIGNED(tex->surface.u.gfx9.htile.pipe_aligned) |
2592						 S_028ABC_RB_ALIGNED(tex->surface.u.gfx9.htile.rb_aligned);
2593		}
2594	} else {
2595		/* SI-CI-VI */
2596		struct legacy_surf_level *levelinfo = &tex->surface.u.legacy.level[level];
2597
2598		assert(levelinfo->nblk_x % 8 == 0 && levelinfo->nblk_y % 8 == 0);
2599
2600		surf->db_depth_base = (tex->buffer.gpu_address +
2601				       tex->surface.u.legacy.level[level].offset) >> 8;
2602		surf->db_stencil_base = (tex->buffer.gpu_address +
2603					 tex->surface.u.legacy.stencil_level[level].offset) >> 8;
2604
2605		z_info = S_028040_FORMAT(format) |
2606			 S_028040_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples));
2607		s_info = S_028044_FORMAT(stencil_format);
2608		surf->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(!tex->tc_compatible_htile);
2609
2610		if (sctx->chip_class >= CIK) {
2611			struct radeon_info *info = &sctx->screen->info;
2612			unsigned index = tex->surface.u.legacy.tiling_index[level];
2613			unsigned stencil_index = tex->surface.u.legacy.stencil_tiling_index[level];
2614			unsigned macro_index = tex->surface.u.legacy.macro_tile_index;
2615			unsigned tile_mode = info->si_tile_mode_array[index];
2616			unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
2617			unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
2618
2619			surf->db_depth_info |=
2620				S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
2621				S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
2622				S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
2623				S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
2624				S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
2625				S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
2626			z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
2627			s_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
2628		} else {
2629			unsigned tile_mode_index = si_tile_mode_index(tex, level, false);
2630			z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
2631			tile_mode_index = si_tile_mode_index(tex, level, true);
2632			s_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
2633		}
2634
2635		surf->db_depth_size = S_028058_PITCH_TILE_MAX((levelinfo->nblk_x / 8) - 1) |
2636				      S_028058_HEIGHT_TILE_MAX((levelinfo->nblk_y / 8) - 1);
2637		surf->db_depth_slice = S_02805C_SLICE_TILE_MAX((levelinfo->nblk_x *
2638								levelinfo->nblk_y) / 64 - 1);
2639
2640		if (si_htile_enabled(tex, level)) {
2641			z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
2642				  S_028040_ALLOW_EXPCLEAR(1);
2643
2644			if (tex->surface.has_stencil) {
2645				/* Workaround: For a not yet understood reason, the
2646				 * combination of MSAA, fast stencil clear and stencil
2647				 * decompress messes with subsequent stencil buffer
2648				 * uses. Problem was reproduced on Verde, Bonaire,
2649				 * Tonga, and Carrizo.
2650				 *
2651				 * Disabling EXPCLEAR works around the problem.
2652				 *
2653				 * Check piglit's arb_texture_multisample-stencil-clear
2654				 * test if you want to try changing this.
2655				 */
2656				if (tex->buffer.b.b.nr_samples <= 1)
2657					s_info |= S_028044_ALLOW_EXPCLEAR(1);
2658			} else if (!tex->tc_compatible_htile) {
2659				/* Use all of the htile_buffer for depth if there's no stencil.
2660				 * This must not be set when TC-compatible HTILE is enabled
2661				 * due to a hw bug.
2662				 */
2663				s_info |= S_028044_TILE_STENCIL_DISABLE(1);
2664			}
2665
2666			surf->db_htile_data_base = (tex->buffer.gpu_address +
2667						    tex->htile_offset) >> 8;
2668			surf->db_htile_surface = S_028ABC_FULL_CACHE(1);
2669
2670			if (tex->tc_compatible_htile) {
2671				surf->db_htile_surface |= S_028ABC_TC_COMPATIBLE(1);
2672
2673				/* 0 = full compression. N = only compress up to N-1 Z planes. */
2674				if (tex->buffer.b.b.nr_samples <= 1)
2675					z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2676				else if (tex->buffer.b.b.nr_samples <= 4)
2677					z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2678				else
2679					z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2680			}
2681		}
2682	}
2683
2684	surf->db_z_info = z_info;
2685	surf->db_stencil_info = s_info;
2686
2687	surf->depth_initialized = true;
2688}
2689
2690void si_update_fb_dirtiness_after_rendering(struct si_context *sctx)
2691{
2692	if (sctx->decompression_enabled)
2693		return;
2694
2695	if (sctx->framebuffer.state.zsbuf) {
2696		struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
2697		struct si_texture *tex = (struct si_texture *)surf->texture;
2698
2699		tex->dirty_level_mask |= 1 << surf->u.tex.level;
2700
2701		if (tex->surface.has_stencil)
2702			tex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
2703	}
2704
2705	unsigned compressed_cb_mask = sctx->framebuffer.compressed_cb_mask;
2706	while (compressed_cb_mask) {
2707		unsigned i = u_bit_scan(&compressed_cb_mask);
2708		struct pipe_surface *surf = sctx->framebuffer.state.cbufs[i];
2709		struct si_texture *tex = (struct si_texture*)surf->texture;
2710
2711		if (tex->surface.fmask_size)
2712			tex->dirty_level_mask |= 1 << surf->u.tex.level;
2713		if (tex->dcc_gather_statistics)
2714			tex->separate_dcc_dirty = true;
2715	}
2716}
2717
2718static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2719{
2720	for (int i = 0; i < state->nr_cbufs; ++i) {
2721		struct si_surface *surf = NULL;
2722		struct si_texture *tex;
2723
2724		if (!state->cbufs[i])
2725			continue;
2726		surf = (struct si_surface*)state->cbufs[i];
2727		tex = (struct si_texture*)surf->base.texture;
2728
2729		p_atomic_dec(&tex->framebuffers_bound);
2730	}
2731}
2732
2733static void si_set_framebuffer_state(struct pipe_context *ctx,
2734				     const struct pipe_framebuffer_state *state)
2735{
2736	struct si_context *sctx = (struct si_context *)ctx;
2737	struct si_surface *surf = NULL;
2738	struct si_texture *tex;
2739	bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2740	unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2741	unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
2742	bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
2743	bool old_has_stencil =
2744		old_has_zsbuf &&
2745		((struct si_texture*)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
2746	bool unbound = false;
2747	int i;
2748
2749	/* Reject zero-sized framebuffers due to a hw bug on SI that occurs
2750	 * when PA_SU_HARDWARE_SCREEN_OFFSET != 0 and any_scissor.BR_X/Y <= 0.
2751	 * We could implement the full workaround here, but it's a useless case.
2752	 */
2753	if ((!state->width || !state->height) && (state->nr_cbufs || state->zsbuf)) {
2754		unreachable("the framebuffer shouldn't have zero area");
2755		return;
2756	}
2757
2758	si_update_fb_dirtiness_after_rendering(sctx);
2759
2760	for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
2761		if (!sctx->framebuffer.state.cbufs[i])
2762			continue;
2763
2764		tex = (struct si_texture*)sctx->framebuffer.state.cbufs[i]->texture;
2765		if (tex->dcc_gather_statistics)
2766			vi_separate_dcc_stop_query(sctx, tex);
2767	}
2768
2769	/* Disable DCC if the formats are incompatible. */
2770	for (i = 0; i < state->nr_cbufs; i++) {
2771		if (!state->cbufs[i])
2772			continue;
2773
2774		surf = (struct si_surface*)state->cbufs[i];
2775		tex = (struct si_texture*)surf->base.texture;
2776
2777		if (!surf->dcc_incompatible)
2778			continue;
2779
2780		/* Since the DCC decompression calls back into set_framebuffer-
2781		 * _state, we need to unbind the framebuffer, so that
2782		 * vi_separate_dcc_stop_query isn't called twice with the same
2783		 * color buffer.
2784		 */
2785		if (!unbound) {
2786			util_copy_framebuffer_state(&sctx->framebuffer.state, NULL);
2787			unbound = true;
2788		}
2789
2790		if (vi_dcc_enabled(tex, surf->base.u.tex.level))
2791			if (!si_texture_disable_dcc(sctx, tex))
2792				si_decompress_dcc(sctx, tex);
2793
2794		surf->dcc_incompatible = false;
2795	}
2796
2797	/* Only flush TC when changing the framebuffer state, because
2798	 * the only client not using TC that can change textures is
2799	 * the framebuffer.
2800	 *
2801	 * Wait for compute shaders because of possible transitions:
2802	 * - FB write -> shader read
2803	 * - shader write -> FB read
2804	 *
2805	 * DB caches are flushed on demand (using si_decompress_textures).
2806	 *
2807	 * When MSAA is enabled, CB and TC caches are flushed on demand
2808	 * (after FMASK decompression). Shader write -> FB read transitions
2809	 * cannot happen for MSAA textures, because MSAA shader images are
2810	 * not supported.
2811	 *
2812	 * Only flush and wait for CB if there is actually a bound color buffer.
2813	 */
2814	if (sctx->framebuffer.uncompressed_cb_mask) {
2815		si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
2816					   sctx->framebuffer.CB_has_shader_readable_metadata,
2817					   sctx->framebuffer.all_DCC_pipe_aligned);
2818	}
2819
2820	sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
2821
2822	/* u_blitter doesn't invoke depth decompression when it does multiple
2823	 * blits in a row, but the only case when it matters for DB is when
2824	 * doing generate_mipmap. So here we flush DB manually between
2825	 * individual generate_mipmap blits.
2826	 * Note that lower mipmap levels aren't compressed.
2827	 */
2828	if (sctx->generate_mipmap_for_depth) {
2829		si_make_DB_shader_coherent(sctx, 1, false,
2830					   sctx->framebuffer.DB_has_shader_readable_metadata);
2831	} else if (sctx->chip_class == GFX9) {
2832		/* It appears that DB metadata "leaks" in a sequence of:
2833		 *  - depth clear
2834		 *  - DCC decompress for shader image writes (with DB disabled)
2835		 *  - render with DEPTH_BEFORE_SHADER=1
2836		 * Flushing DB metadata works around the problem.
2837		 */
2838		sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB_META;
2839	}
2840
2841	/* Take the maximum of the old and new count. If the new count is lower,
2842	 * dirtying is needed to disable the unbound colorbuffers.
2843	 */
2844	sctx->framebuffer.dirty_cbufs |=
2845		(1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2846	sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2847
2848	si_dec_framebuffer_counters(&sctx->framebuffer.state);
2849	util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2850
2851	sctx->framebuffer.colorbuf_enabled_4bit = 0;
2852	sctx->framebuffer.spi_shader_col_format = 0;
2853	sctx->framebuffer.spi_shader_col_format_alpha = 0;
2854	sctx->framebuffer.spi_shader_col_format_blend = 0;
2855	sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2856	sctx->framebuffer.color_is_int8 = 0;
2857	sctx->framebuffer.color_is_int10 = 0;
2858
2859	sctx->framebuffer.compressed_cb_mask = 0;
2860	sctx->framebuffer.uncompressed_cb_mask = 0;
2861	sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2862	sctx->framebuffer.nr_color_samples = sctx->framebuffer.nr_samples;
2863	sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2864	sctx->framebuffer.any_dst_linear = false;
2865	sctx->framebuffer.CB_has_shader_readable_metadata = false;
2866	sctx->framebuffer.DB_has_shader_readable_metadata = false;
2867	sctx->framebuffer.all_DCC_pipe_aligned = true;
2868	unsigned num_bpp64_colorbufs = 0;
2869
2870	for (i = 0; i < state->nr_cbufs; i++) {
2871		if (!state->cbufs[i])
2872			continue;
2873
2874		surf = (struct si_surface*)state->cbufs[i];
2875		tex = (struct si_texture*)surf->base.texture;
2876
2877		if (!surf->color_initialized) {
2878			si_initialize_color_surface(sctx, surf);
2879		}
2880
2881		sctx->framebuffer.colorbuf_enabled_4bit |= 0xf << (i * 4);
2882		sctx->framebuffer.spi_shader_col_format |=
2883			surf->spi_shader_col_format << (i * 4);
2884		sctx->framebuffer.spi_shader_col_format_alpha |=
2885			surf->spi_shader_col_format_alpha << (i * 4);
2886		sctx->framebuffer.spi_shader_col_format_blend |=
2887			surf->spi_shader_col_format_blend << (i * 4);
2888		sctx->framebuffer.spi_shader_col_format_blend_alpha |=
2889			surf->spi_shader_col_format_blend_alpha << (i * 4);
2890
2891		if (surf->color_is_int8)
2892			sctx->framebuffer.color_is_int8 |= 1 << i;
2893		if (surf->color_is_int10)
2894			sctx->framebuffer.color_is_int10 |= 1 << i;
2895
2896		if (tex->surface.fmask_size)
2897			sctx->framebuffer.compressed_cb_mask |= 1 << i;
2898		else
2899			sctx->framebuffer.uncompressed_cb_mask |= 1 << i;
2900
2901		/* Don't update nr_color_samples for non-AA buffers.
2902		 * (e.g. destination of MSAA resolve)
2903		 */
2904		if (tex->buffer.b.b.nr_samples >= 2 &&
2905		    tex->buffer.b.b.nr_storage_samples < tex->buffer.b.b.nr_samples) {
2906			sctx->framebuffer.nr_color_samples =
2907				MIN2(sctx->framebuffer.nr_color_samples,
2908				     tex->buffer.b.b.nr_storage_samples);
2909			sctx->framebuffer.nr_color_samples =
2910				MAX2(1, sctx->framebuffer.nr_color_samples);
2911		}
2912
2913		if (tex->surface.is_linear)
2914			sctx->framebuffer.any_dst_linear = true;
2915		if (tex->surface.bpe >= 8)
2916			num_bpp64_colorbufs++;
2917
2918		if (vi_dcc_enabled(tex, surf->base.u.tex.level)) {
2919			sctx->framebuffer.CB_has_shader_readable_metadata = true;
2920
2921			if (sctx->chip_class >= GFX9 &&
2922			    !tex->surface.u.gfx9.dcc.pipe_aligned)
2923				sctx->framebuffer.all_DCC_pipe_aligned = false;
2924		}
2925
2926		si_context_add_resource_size(sctx, surf->base.texture);
2927
2928		p_atomic_inc(&tex->framebuffers_bound);
2929
2930		if (tex->dcc_gather_statistics) {
2931			/* Dirty tracking must be enabled for DCC usage analysis. */
2932			sctx->framebuffer.compressed_cb_mask |= 1 << i;
2933			vi_separate_dcc_start_query(sctx, tex);
2934		}
2935	}
2936
2937	/* For optimal DCC performance. */
2938	if (sctx->chip_class == VI)
2939		sctx->framebuffer.dcc_overwrite_combiner_watermark = 4;
2940	else if (num_bpp64_colorbufs >= 5)
2941		sctx->framebuffer.dcc_overwrite_combiner_watermark = 8;
2942	else
2943		sctx->framebuffer.dcc_overwrite_combiner_watermark = 6;
2944
2945	struct si_texture *zstex = NULL;
2946
2947	if (state->zsbuf) {
2948		surf = (struct si_surface*)state->zsbuf;
2949		zstex = (struct si_texture*)surf->base.texture;
2950
2951		if (!surf->depth_initialized) {
2952			si_init_depth_surface(sctx, surf);
2953		}
2954
2955		if (vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level))
2956			sctx->framebuffer.DB_has_shader_readable_metadata = true;
2957
2958		si_context_add_resource_size(sctx, surf->base.texture);
2959	}
2960
2961	si_update_ps_colorbuf0_slot(sctx);
2962	si_update_poly_offset_state(sctx);
2963	si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
2964	si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
2965
2966	if (sctx->screen->dpbb_allowed)
2967		si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
2968
2969	if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
2970		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2971
2972	if (sctx->screen->has_out_of_order_rast &&
2973	    (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
2974	     !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
2975	     (zstex && zstex->surface.has_stencil != old_has_stencil)))
2976		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2977
2978	if (sctx->framebuffer.nr_samples != old_nr_samples) {
2979		struct pipe_constant_buffer constbuf = {0};
2980
2981		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2982		si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
2983
2984		constbuf.buffer = sctx->sample_pos_buffer;
2985
2986		/* Set sample locations as fragment shader constants. */
2987		switch (sctx->framebuffer.nr_samples) {
2988		case 1:
2989			constbuf.buffer_offset = 0;
2990			break;
2991		case 2:
2992			constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x2 -
2993						 (ubyte*)sctx->sample_positions.x1;
2994			break;
2995		case 4:
2996			constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x4 -
2997						 (ubyte*)sctx->sample_positions.x1;
2998			break;
2999		case 8:
3000			constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x8 -
3001						 (ubyte*)sctx->sample_positions.x1;
3002			break;
3003		case 16:
3004			constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x16 -
3005						 (ubyte*)sctx->sample_positions.x1;
3006			break;
3007		default:
3008			PRINT_ERR("Requested an invalid number of samples %i.\n",
3009				 sctx->framebuffer.nr_samples);
3010			assert(0);
3011		}
3012		constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
3013		si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
3014
3015		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
3016	}
3017
3018	sctx->do_update_shaders = true;
3019
3020	if (!sctx->decompression_enabled) {
3021		/* Prevent textures decompression when the framebuffer state
3022		 * changes come from the decompression passes themselves.
3023		 */
3024		sctx->need_check_render_feedback = true;
3025	}
3026}
3027
3028static void si_emit_framebuffer_state(struct si_context *sctx)
3029{
3030	struct radeon_cmdbuf *cs = sctx->gfx_cs;
3031	struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
3032	unsigned i, nr_cbufs = state->nr_cbufs;
3033	struct si_texture *tex = NULL;
3034	struct si_surface *cb = NULL;
3035	unsigned cb_color_info = 0;
3036
3037	/* Colorbuffers. */
3038	for (i = 0; i < nr_cbufs; i++) {
3039		uint64_t cb_color_base, cb_color_fmask, cb_color_cmask, cb_dcc_base;
3040		unsigned cb_color_attrib;
3041
3042		if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
3043			continue;
3044
3045		cb = (struct si_surface*)state->cbufs[i];
3046		if (!cb) {
3047			radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
3048					       S_028C70_FORMAT(V_028C70_COLOR_INVALID));
3049			continue;
3050		}
3051
3052		tex = (struct si_texture *)cb->base.texture;
3053		radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3054				      &tex->buffer, RADEON_USAGE_READWRITE,
3055				      tex->buffer.b.b.nr_samples > 1 ?
3056					      RADEON_PRIO_COLOR_BUFFER_MSAA :
3057					      RADEON_PRIO_COLOR_BUFFER);
3058
3059		if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
3060			radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3061				tex->cmask_buffer, RADEON_USAGE_READWRITE,
3062				RADEON_PRIO_SEPARATE_META);
3063		}
3064
3065		if (tex->dcc_separate_buffer)
3066			radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3067						  tex->dcc_separate_buffer,
3068						  RADEON_USAGE_READWRITE,
3069						  RADEON_PRIO_SEPARATE_META);
3070
3071		/* Compute mutable surface parameters. */
3072		cb_color_base = tex->buffer.gpu_address >> 8;
3073		cb_color_fmask = 0;
3074		cb_color_cmask = tex->cmask_base_address_reg;
3075		cb_dcc_base = 0;
3076		cb_color_info = cb->cb_color_info | tex->cb_color_info;
3077		cb_color_attrib = cb->cb_color_attrib;
3078
3079		if (cb->base.u.tex.level > 0)
3080			cb_color_info &= C_028C70_FAST_CLEAR;
3081
3082		if (tex->surface.fmask_size) {
3083			cb_color_fmask = (tex->buffer.gpu_address + tex->fmask_offset) >> 8;
3084			cb_color_fmask |= tex->surface.fmask_tile_swizzle;
3085		}
3086
3087		/* Set up DCC. */
3088		if (vi_dcc_enabled(tex, cb->base.u.tex.level)) {
3089			bool is_msaa_resolve_dst = state->cbufs[0] &&
3090						   state->cbufs[0]->texture->nr_samples > 1 &&
3091						   state->cbufs[1] == &cb->base &&
3092						   state->cbufs[1]->texture->nr_samples <= 1;
3093
3094			if (!is_msaa_resolve_dst)
3095				cb_color_info |= S_028C70_DCC_ENABLE(1);
3096
3097			cb_dcc_base = ((!tex->dcc_separate_buffer ? tex->buffer.gpu_address : 0) +
3098				       tex->dcc_offset) >> 8;
3099			cb_dcc_base |= tex->surface.tile_swizzle;
3100		}
3101
3102		if (sctx->chip_class >= GFX9) {
3103			struct gfx9_surf_meta_flags meta;
3104
3105			if (tex->dcc_offset)
3106				meta = tex->surface.u.gfx9.dcc;
3107			else
3108				meta = tex->surface.u.gfx9.cmask;
3109
3110			/* Set mutable surface parameters. */
3111			cb_color_base += tex->surface.u.gfx9.surf_offset >> 8;
3112			cb_color_base |= tex->surface.tile_swizzle;
3113			if (!tex->surface.fmask_size)
3114				cb_color_fmask = cb_color_base;
3115			if (cb->base.u.tex.level > 0)
3116				cb_color_cmask = cb_color_base;
3117			cb_color_attrib |= S_028C74_COLOR_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
3118					   S_028C74_FMASK_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
3119					   S_028C74_RB_ALIGNED(meta.rb_aligned) |
3120					   S_028C74_PIPE_ALIGNED(meta.pipe_aligned);
3121
3122			radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 15);
3123			radeon_emit(cs, cb_color_base);		/* CB_COLOR0_BASE */
3124			radeon_emit(cs, S_028C64_BASE_256B(cb_color_base >> 32)); /* CB_COLOR0_BASE_EXT */
3125			radeon_emit(cs, cb->cb_color_attrib2);	/* CB_COLOR0_ATTRIB2 */
3126			radeon_emit(cs, cb->cb_color_view);	/* CB_COLOR0_VIEW */
3127			radeon_emit(cs, cb_color_info);		/* CB_COLOR0_INFO */
3128			radeon_emit(cs, cb_color_attrib);	/* CB_COLOR0_ATTRIB */
3129			radeon_emit(cs, cb->cb_dcc_control);	/* CB_COLOR0_DCC_CONTROL */
3130			radeon_emit(cs, cb_color_cmask);	/* CB_COLOR0_CMASK */
3131			radeon_emit(cs, S_028C80_BASE_256B(cb_color_cmask >> 32)); /* CB_COLOR0_CMASK_BASE_EXT */
3132			radeon_emit(cs, cb_color_fmask);	/* CB_COLOR0_FMASK */
3133			radeon_emit(cs, S_028C88_BASE_256B(cb_color_fmask >> 32)); /* CB_COLOR0_FMASK_BASE_EXT */
3134			radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3135			radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3136			radeon_emit(cs, cb_dcc_base);		/* CB_COLOR0_DCC_BASE */
3137			radeon_emit(cs, S_028C98_BASE_256B(cb_dcc_base >> 32)); /* CB_COLOR0_DCC_BASE_EXT */
3138
3139			radeon_set_context_reg(cs, R_0287A0_CB_MRT0_EPITCH + i * 4,
3140					       S_0287A0_EPITCH(tex->surface.u.gfx9.surf.epitch));
3141		} else {
3142			/* Compute mutable surface parameters (SI-CI-VI). */
3143			const struct legacy_surf_level *level_info =
3144				&tex->surface.u.legacy.level[cb->base.u.tex.level];
3145			unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
3146			unsigned cb_color_pitch, cb_color_slice, cb_color_fmask_slice;
3147
3148			cb_color_base += level_info->offset >> 8;
3149			/* Only macrotiled modes can set tile swizzle. */
3150			if (level_info->mode == RADEON_SURF_MODE_2D)
3151				cb_color_base |= tex->surface.tile_swizzle;
3152
3153			if (!tex->surface.fmask_size)
3154				cb_color_fmask = cb_color_base;
3155			if (cb->base.u.tex.level > 0)
3156				cb_color_cmask = cb_color_base;
3157			if (cb_dcc_base)
3158				cb_dcc_base += level_info->dcc_offset >> 8;
3159
3160			pitch_tile_max = level_info->nblk_x / 8 - 1;
3161			slice_tile_max = level_info->nblk_x *
3162					 level_info->nblk_y / 64 - 1;
3163			tile_mode_index = si_tile_mode_index(tex, cb->base.u.tex.level, false);
3164
3165			cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
3166			cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
3167			cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
3168
3169			if (tex->surface.fmask_size) {
3170				if (sctx->chip_class >= CIK)
3171					cb_color_pitch |= S_028C64_FMASK_TILE_MAX(tex->surface.u.legacy.fmask.pitch_in_pixels / 8 - 1);
3172				cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3173				cb_color_fmask_slice = S_028C88_TILE_MAX(tex->surface.u.legacy.fmask.slice_tile_max);
3174			} else {
3175				/* This must be set for fast clear to work without FMASK. */
3176				if (sctx->chip_class >= CIK)
3177					cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
3178				cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
3179				cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
3180			}
3181
3182			radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C,
3183						   sctx->chip_class >= VI ? 14 : 13);
3184			radeon_emit(cs, cb_color_base);		/* CB_COLOR0_BASE */
3185			radeon_emit(cs, cb_color_pitch);	/* CB_COLOR0_PITCH */
3186			radeon_emit(cs, cb_color_slice);	/* CB_COLOR0_SLICE */
3187			radeon_emit(cs, cb->cb_color_view);	/* CB_COLOR0_VIEW */
3188			radeon_emit(cs, cb_color_info);		/* CB_COLOR0_INFO */
3189			radeon_emit(cs, cb_color_attrib);	/* CB_COLOR0_ATTRIB */
3190			radeon_emit(cs, cb->cb_dcc_control);	/* CB_COLOR0_DCC_CONTROL */
3191			radeon_emit(cs, cb_color_cmask);	/* CB_COLOR0_CMASK */
3192			radeon_emit(cs, tex->surface.u.legacy.cmask_slice_tile_max); /* CB_COLOR0_CMASK_SLICE */
3193			radeon_emit(cs, cb_color_fmask);		/* CB_COLOR0_FMASK */
3194			radeon_emit(cs, cb_color_fmask_slice);		/* CB_COLOR0_FMASK_SLICE */
3195			radeon_emit(cs, tex->color_clear_value[0]);	/* CB_COLOR0_CLEAR_WORD0 */
3196			radeon_emit(cs, tex->color_clear_value[1]);	/* CB_COLOR0_CLEAR_WORD1 */
3197
3198			if (sctx->chip_class >= VI) /* R_028C94_CB_COLOR0_DCC_BASE */
3199				radeon_emit(cs, cb_dcc_base);
3200		}
3201	}
3202	for (; i < 8 ; i++)
3203		if (sctx->framebuffer.dirty_cbufs & (1 << i))
3204			radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
3205
3206	/* ZS buffer. */
3207	if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3208		struct si_surface *zb = (struct si_surface*)state->zsbuf;
3209		struct si_texture *tex = (struct si_texture*)zb->base.texture;
3210
3211		radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3212				      &tex->buffer, RADEON_USAGE_READWRITE,
3213				      zb->base.texture->nr_samples > 1 ?
3214					      RADEON_PRIO_DEPTH_BUFFER_MSAA :
3215					      RADEON_PRIO_DEPTH_BUFFER);
3216
3217		if (sctx->chip_class >= GFX9) {
3218			radeon_set_context_reg_seq(cs, R_028014_DB_HTILE_DATA_BASE, 3);
3219			radeon_emit(cs, zb->db_htile_data_base);	/* DB_HTILE_DATA_BASE */
3220			radeon_emit(cs, S_028018_BASE_HI(zb->db_htile_data_base >> 32)); /* DB_HTILE_DATA_BASE_HI */
3221			radeon_emit(cs, zb->db_depth_size);		/* DB_DEPTH_SIZE */
3222
3223			radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 10);
3224			radeon_emit(cs, zb->db_z_info |			/* DB_Z_INFO */
3225				    S_028038_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3226			radeon_emit(cs, zb->db_stencil_info);		/* DB_STENCIL_INFO */
3227			radeon_emit(cs, zb->db_depth_base);		/* DB_Z_READ_BASE */
3228			radeon_emit(cs, S_028044_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_READ_BASE_HI */
3229			radeon_emit(cs, zb->db_stencil_base);		/* DB_STENCIL_READ_BASE */
3230			radeon_emit(cs, S_02804C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
3231			radeon_emit(cs, zb->db_depth_base);		/* DB_Z_WRITE_BASE */
3232			radeon_emit(cs, S_028054_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_WRITE_BASE_HI */
3233			radeon_emit(cs, zb->db_stencil_base);		/* DB_STENCIL_WRITE_BASE */
3234			radeon_emit(cs, S_02805C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
3235
3236			radeon_set_context_reg_seq(cs, R_028068_DB_Z_INFO2, 2);
3237			radeon_emit(cs, zb->db_z_info2);	/* DB_Z_INFO2 */
3238			radeon_emit(cs, zb->db_stencil_info2);	/* DB_STENCIL_INFO2 */
3239		} else {
3240			radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
3241
3242			radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 9);
3243			radeon_emit(cs, zb->db_depth_info);	/* DB_DEPTH_INFO */
3244			radeon_emit(cs, zb->db_z_info |		/* DB_Z_INFO */
3245				    S_028040_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3246			radeon_emit(cs, zb->db_stencil_info);	/* DB_STENCIL_INFO */
3247			radeon_emit(cs, zb->db_depth_base);	/* DB_Z_READ_BASE */
3248			radeon_emit(cs, zb->db_stencil_base);	/* DB_STENCIL_READ_BASE */
3249			radeon_emit(cs, zb->db_depth_base);	/* DB_Z_WRITE_BASE */
3250			radeon_emit(cs, zb->db_stencil_base);	/* DB_STENCIL_WRITE_BASE */
3251			radeon_emit(cs, zb->db_depth_size);	/* DB_DEPTH_SIZE */
3252			radeon_emit(cs, zb->db_depth_slice);	/* DB_DEPTH_SLICE */
3253		}
3254
3255		radeon_set_context_reg_seq(cs, R_028028_DB_STENCIL_CLEAR, 2);
3256		radeon_emit(cs, tex->stencil_clear_value); /* R_028028_DB_STENCIL_CLEAR */
3257		radeon_emit(cs, fui(tex->depth_clear_value)); /* R_02802C_DB_DEPTH_CLEAR */
3258
3259		radeon_set_context_reg(cs, R_028008_DB_DEPTH_VIEW, zb->db_depth_view);
3260		radeon_set_context_reg(cs, R_028ABC_DB_HTILE_SURFACE, zb->db_htile_surface);
3261	} else if (sctx->framebuffer.dirty_zsbuf) {
3262		if (sctx->chip_class >= GFX9)
3263			radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 2);
3264		else
3265			radeon_set_context_reg_seq(cs, R_028040_DB_Z_INFO, 2);
3266
3267		radeon_emit(cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
3268		radeon_emit(cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
3269	}
3270
3271	/* Framebuffer dimensions. */
3272        /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3273	radeon_set_context_reg(cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
3274			       S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3275
3276	if (sctx->screen->dfsm_allowed) {
3277		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3278		radeon_emit(cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
3279	}
3280
3281	sctx->framebuffer.dirty_cbufs = 0;
3282	sctx->framebuffer.dirty_zsbuf = false;
3283}
3284
3285static void si_emit_msaa_sample_locs(struct si_context *sctx)
3286{
3287	struct radeon_cmdbuf *cs = sctx->gfx_cs;
3288	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3289	unsigned nr_samples = sctx->framebuffer.nr_samples;
3290	bool has_msaa_sample_loc_bug = sctx->screen->has_msaa_sample_loc_bug;
3291
3292	/* Smoothing (only possible with nr_samples == 1) uses the same
3293	 * sample locations as the MSAA it simulates.
3294	 */
3295	if (nr_samples <= 1 && sctx->smoothing_enabled)
3296		nr_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3297
3298	/* On Polaris, the small primitive filter uses the sample locations
3299	 * even when MSAA is off, so we need to make sure they're set to 0.
3300	 */
3301	if ((nr_samples >= 2 || has_msaa_sample_loc_bug) &&
3302	    nr_samples != sctx->sample_locs_num_samples) {
3303		sctx->sample_locs_num_samples = nr_samples;
3304		si_emit_sample_locations(cs, nr_samples);
3305	}
3306
3307	if (sctx->family >= CHIP_POLARIS10) {
3308		unsigned small_prim_filter_cntl =
3309			S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3310			/* line bug */
3311			S_028830_LINE_FILTER_DISABLE(sctx->family <= CHIP_POLARIS12);
3312
3313		/* The alternative of setting sample locations to 0 would
3314		 * require a DB flush to avoid Z errors, see
3315		 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3316		 */
3317		if (has_msaa_sample_loc_bug &&
3318		    sctx->framebuffer.nr_samples > 1 &&
3319		    !rs->multisample_enable)
3320			small_prim_filter_cntl &= C_028830_SMALL_PRIM_FILTER_ENABLE;
3321
3322		radeon_opt_set_context_reg(sctx,
3323					   R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
3324					   SI_TRACKED_PA_SU_SMALL_PRIM_FILTER_CNTL,
3325					   small_prim_filter_cntl);
3326	}
3327
3328	/* The exclusion bits can be set to improve rasterization efficiency
3329	 * if no sample lies on the pixel boundary (-8 sample offset).
3330	 */
3331	bool exclusion = sctx->chip_class >= CIK &&
3332			 (!rs->multisample_enable || nr_samples != 16);
3333	radeon_opt_set_context_reg(sctx, R_02882C_PA_SU_PRIM_FILTER_CNTL,
3334				   SI_TRACKED_PA_SU_PRIM_FILTER_CNTL,
3335				   S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
3336				   S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
3337}
3338
3339static bool si_out_of_order_rasterization(struct si_context *sctx)
3340{
3341	struct si_state_blend *blend = sctx->queued.named.blend;
3342	struct si_state_dsa *dsa = sctx->queued.named.dsa;
3343
3344	if (!sctx->screen->has_out_of_order_rast)
3345		return false;
3346
3347	unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
3348
3349	if (blend) {
3350		colormask &= blend->cb_target_enabled_4bit;
3351	} else {
3352		colormask = 0;
3353	}
3354
3355	/* Conservative: No logic op. */
3356	if (colormask && blend->logicop_enable)
3357		return false;
3358
3359	struct si_dsa_order_invariance dsa_order_invariant = {
3360		.zs = true, .pass_set = true, .pass_last = false
3361	};
3362
3363	if (sctx->framebuffer.state.zsbuf) {
3364		struct si_texture *zstex =
3365			(struct si_texture*)sctx->framebuffer.state.zsbuf->texture;
3366		bool has_stencil = zstex->surface.has_stencil;
3367		dsa_order_invariant = dsa->order_invariance[has_stencil];
3368		if (!dsa_order_invariant.zs)
3369			return false;
3370
3371		/* The set of PS invocations is always order invariant,
3372		 * except when early Z/S tests are requested. */
3373		if (sctx->ps_shader.cso &&
3374		    sctx->ps_shader.cso->info.writes_memory &&
3375		    sctx->ps_shader.cso->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] &&
3376		    !dsa_order_invariant.pass_set)
3377			return false;
3378
3379		if (sctx->num_perfect_occlusion_queries != 0 &&
3380		    !dsa_order_invariant.pass_set)
3381			return false;
3382	}
3383
3384	if (!colormask)
3385		return true;
3386
3387	unsigned blendmask = colormask & blend->blend_enable_4bit;
3388
3389	if (blendmask) {
3390		/* Only commutative blending. */
3391		if (blendmask & ~blend->commutative_4bit)
3392			return false;
3393
3394		if (!dsa_order_invariant.pass_set)
3395			return false;
3396	}
3397
3398	if (colormask & ~blendmask) {
3399		if (!dsa_order_invariant.pass_last)
3400			return false;
3401	}
3402
3403	return true;
3404}
3405
3406static void si_emit_msaa_config(struct si_context *sctx)
3407{
3408	struct radeon_cmdbuf *cs = sctx->gfx_cs;
3409	unsigned num_tile_pipes = sctx->screen->info.num_tile_pipes;
3410	/* 33% faster rendering to linear color buffers */
3411	bool dst_is_linear = sctx->framebuffer.any_dst_linear;
3412	bool out_of_order_rast = si_out_of_order_rasterization(sctx);
3413	unsigned sc_mode_cntl_1 =
3414		S_028A4C_WALK_SIZE(dst_is_linear) |
3415		S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
3416		S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
3417		S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
3418		S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3419		/* always 1: */
3420		S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3421		S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3422		S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3423		S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3424		S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3425		S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3426	unsigned db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
3427			   S_028804_INCOHERENT_EQAA_READS(1) |
3428			   S_028804_INTERPOLATE_COMP_Z(1) |
3429			   S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
3430	unsigned coverage_samples, color_samples, z_samples;
3431
3432	/* S: Coverage samples (up to 16x):
3433	 * - Scan conversion samples (PA_SC_AA_CONFIG.MSAA_NUM_SAMPLES)
3434	 * - CB FMASK samples (CB_COLORi_ATTRIB.NUM_SAMPLES)
3435	 *
3436	 * Z: Z/S samples (up to 8x, must be <= coverage samples and >= color samples):
3437	 * - Value seen by DB (DB_Z_INFO.NUM_SAMPLES)
3438	 * - Value seen by CB, must be correct even if Z/S is unbound (DB_EQAA.MAX_ANCHOR_SAMPLES)
3439	 * # Missing samples are derived from Z planes if Z is compressed (up to 16x quality), or
3440	 * # from the closest defined sample if Z is uncompressed (same quality as the number of
3441	 * # Z samples).
3442	 *
3443	 * F: Color samples (up to 8x, must be <= coverage samples):
3444	 * - CB color samples (CB_COLORi_ATTRIB.NUM_FRAGMENTS)
3445	 * - PS iter samples (DB_EQAA.PS_ITER_SAMPLES)
3446	 *
3447	 * Can be anything between coverage and color samples:
3448	 * - SampleMaskIn samples (PA_SC_AA_CONFIG.MSAA_EXPOSED_SAMPLES)
3449	 * - SampleMaskOut samples (DB_EQAA.MASK_EXPORT_NUM_SAMPLES)
3450	 * - Alpha-to-coverage samples (DB_EQAA.ALPHA_TO_MASK_NUM_SAMPLES)
3451	 * - Occlusion query samples (DB_COUNT_CONTROL.SAMPLE_RATE)
3452	 * # All are currently set the same as coverage samples.
3453	 *
3454	 * If color samples < coverage samples, FMASK has a higher bpp to store an "unknown"
3455	 * flag for undefined color samples. A shader-based resolve must handle unknowns
3456	 * or mask them out with AND. Unknowns can also be guessed from neighbors via
3457	 * an edge-detect shader-based resolve, which is required to make "color samples = 1"
3458	 * useful. The CB resolve always drops unknowns.
3459	 *
3460	 * Sensible AA configurations:
3461	 *   EQAA 16s 8z 8f - might look the same as 16x MSAA if Z is compressed
3462	 *   EQAA 16s 8z 4f - might look the same as 16x MSAA if Z is compressed
3463	 *   EQAA 16s 4z 4f - might look the same as 16x MSAA if Z is compressed
3464	 *   EQAA  8s 8z 8f = 8x MSAA
3465	 *   EQAA  8s 8z 4f - might look the same as 8x MSAA
3466	 *   EQAA  8s 8z 2f - might look the same as 8x MSAA with low-density geometry
3467	 *   EQAA  8s 4z 4f - might look the same as 8x MSAA if Z is compressed
3468	 *   EQAA  8s 4z 2f - might look the same as 8x MSAA with low-density geometry if Z is compressed
3469	 *   EQAA  4s 4z 4f = 4x MSAA
3470	 *   EQAA  4s 4z 2f - might look the same as 4x MSAA with low-density geometry
3471	 *   EQAA  2s 2z 2f = 2x MSAA
3472	 */
3473	if (sctx->framebuffer.nr_samples > 1) {
3474		coverage_samples = sctx->framebuffer.nr_samples;
3475		color_samples = sctx->framebuffer.nr_color_samples;
3476
3477		if (sctx->framebuffer.state.zsbuf) {
3478			z_samples = sctx->framebuffer.state.zsbuf->texture->nr_samples;
3479			z_samples = MAX2(1, z_samples);
3480		} else {
3481			z_samples = coverage_samples;
3482		}
3483	} else if (sctx->smoothing_enabled) {
3484		coverage_samples = color_samples = z_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3485	} else {
3486		coverage_samples = color_samples = z_samples = 1;
3487	}
3488
3489	/* Required by OpenGL line rasterization.
3490	 *
3491	 * TODO: We should also enable perpendicular endcaps for AA lines,
3492	 *       but that requires implementing line stippling in the pixel
3493	 *       shader. SC can only do line stippling with axis-aligned
3494	 *       endcaps.
3495	 */
3496	unsigned sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
3497	unsigned sc_aa_config = 0;
3498
3499	if (coverage_samples > 1) {
3500		/* distance from the pixel center, indexed by log2(nr_samples) */
3501		static unsigned max_dist[] = {
3502			0, /* unused */
3503			4, /* 2x MSAA */
3504			6, /* 4x MSAA */
3505			7, /* 8x MSAA */
3506			8, /* 16x MSAA */
3507		};
3508		unsigned log_samples = util_logbase2(coverage_samples);
3509		unsigned log_z_samples = util_logbase2(z_samples);
3510		unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
3511		unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
3512
3513		sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1);
3514		sc_aa_config = S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
3515			       S_028BE0_MAX_SAMPLE_DIST(max_dist[log_samples]) |
3516			       S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples);
3517
3518		if (sctx->framebuffer.nr_samples > 1) {
3519			db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_z_samples) |
3520				   S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
3521				   S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3522				   S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
3523			sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
3524		} else if (sctx->smoothing_enabled) {
3525			db_eqaa |= S_028804_OVERRASTERIZATION_AMOUNT(log_samples);
3526		}
3527	}
3528
3529	unsigned initial_cdw = cs->current.cdw;
3530
3531	/* R_028BDC_PA_SC_LINE_CNTL, R_028BE0_PA_SC_AA_CONFIG */
3532	radeon_opt_set_context_reg2(sctx, R_028BDC_PA_SC_LINE_CNTL,
3533				    SI_TRACKED_PA_SC_LINE_CNTL, sc_line_cntl,
3534				    sc_aa_config);
3535	/* R_028804_DB_EQAA */
3536	radeon_opt_set_context_reg(sctx, R_028804_DB_EQAA, SI_TRACKED_DB_EQAA,
3537				   db_eqaa);
3538	/* R_028A4C_PA_SC_MODE_CNTL_1 */
3539	radeon_opt_set_context_reg(sctx, R_028A4C_PA_SC_MODE_CNTL_1,
3540				   SI_TRACKED_PA_SC_MODE_CNTL_1, sc_mode_cntl_1);
3541
3542	if (initial_cdw != cs->current.cdw) {
3543		sctx->context_roll = true;
3544
3545		/* GFX9: Flush DFSM when the AA mode changes. */
3546		if (sctx->screen->dfsm_allowed) {
3547			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3548			radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3549		}
3550	}
3551}
3552
3553void si_update_ps_iter_samples(struct si_context *sctx)
3554{
3555	if (sctx->framebuffer.nr_samples > 1)
3556		si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3557	if (sctx->screen->dpbb_allowed)
3558		si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3559}
3560
3561static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
3562{
3563	struct si_context *sctx = (struct si_context *)ctx;
3564
3565	/* The hardware can only do sample shading with 2^n samples. */
3566	min_samples = util_next_power_of_two(min_samples);
3567
3568	if (sctx->ps_iter_samples == min_samples)
3569		return;
3570
3571	sctx->ps_iter_samples = min_samples;
3572	sctx->do_update_shaders = true;
3573
3574	si_update_ps_iter_samples(sctx);
3575}
3576
3577/*
3578 * Samplers
3579 */
3580
3581/**
3582 * Build the sampler view descriptor for a buffer texture.
3583 * @param state 256-bit descriptor; only the high 128 bits are filled in
3584 */
3585void
3586si_make_buffer_descriptor(struct si_screen *screen, struct si_resource *buf,
3587			  enum pipe_format format,
3588			  unsigned offset, unsigned size,
3589			  uint32_t *state)
3590{
3591	const struct util_format_description *desc;
3592	int first_non_void;
3593	unsigned stride;
3594	unsigned num_records;
3595	unsigned num_format, data_format;
3596
3597	desc = util_format_description(format);
3598	first_non_void = util_format_get_first_non_void_channel(format);
3599	stride = desc->block.bits / 8;
3600	num_format = si_translate_buffer_numformat(&screen->b, desc, first_non_void);
3601	data_format = si_translate_buffer_dataformat(&screen->b, desc, first_non_void);
3602
3603	num_records = size / stride;
3604	num_records = MIN2(num_records, (buf->b.b.width0 - offset) / stride);
3605
3606	/* The NUM_RECORDS field has a different meaning depending on the chip,
3607	 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3608	 *
3609	 * SI-CIK:
3610	 * - If STRIDE == 0, it's in byte units.
3611	 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3612	 *
3613	 * VI:
3614	 * - For SMEM and STRIDE == 0, it's in byte units.
3615	 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3616	 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3617	 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3618	 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3619	 *       ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3620	 *       using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3621	 *       That way the same descriptor can be used by both SMEM and VMEM.
3622	 *
3623	 * GFX9:
3624	 * - For SMEM and STRIDE == 0, it's in byte units.
3625	 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3626	 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3627	 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3628	 */
3629	if (screen->info.chip_class >= GFX9 && HAVE_LLVM < 0x0800)
3630		/* When vindex == 0, LLVM < 8.0 sets IDXEN = 0, thus changing units
3631		 * from STRIDE to bytes. This works around it by setting
3632		 * NUM_RECORDS to at least the size of one element, so that
3633		 * the first element is readable when IDXEN == 0.
3634		 */
3635		num_records = num_records ? MAX2(num_records, stride) : 0;
3636	else if (screen->info.chip_class == VI)
3637		num_records *= stride;
3638
3639	state[4] = 0;
3640	state[5] = S_008F04_STRIDE(stride);
3641	state[6] = num_records;
3642	state[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
3643		   S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
3644		   S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
3645		   S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
3646		   S_008F0C_NUM_FORMAT(num_format) |
3647		   S_008F0C_DATA_FORMAT(data_format);
3648}
3649
3650static unsigned gfx9_border_color_swizzle(const unsigned char swizzle[4])
3651{
3652	unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3653
3654	if (swizzle[3] == PIPE_SWIZZLE_X) {
3655		/* For the pre-defined border color values (white, opaque
3656		 * black, transparent black), the only thing that matters is
3657		 * that the alpha channel winds up in the correct place
3658		 * (because the RGB channels are all the same) so either of
3659		 * these enumerations will work.
3660		 */
3661		if (swizzle[2] == PIPE_SWIZZLE_Y)
3662			bc_swizzle = V_008F20_BC_SWIZZLE_WZYX;
3663		else
3664			bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ;
3665	} else if (swizzle[0] == PIPE_SWIZZLE_X) {
3666		if (swizzle[1] == PIPE_SWIZZLE_Y)
3667			bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3668		else
3669			bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ;
3670	} else if (swizzle[1] == PIPE_SWIZZLE_X) {
3671		bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ;
3672	} else if (swizzle[2] == PIPE_SWIZZLE_X) {
3673		bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW;
3674	}
3675
3676	return bc_swizzle;
3677}
3678
3679/**
3680 * Build the sampler view descriptor for a texture.
3681 */
3682void
3683si_make_texture_descriptor(struct si_screen *screen,
3684			   struct si_texture *tex,
3685			   bool sampler,
3686			   enum pipe_texture_target target,
3687			   enum pipe_format pipe_format,
3688			   const unsigned char state_swizzle[4],
3689			   unsigned first_level, unsigned last_level,
3690			   unsigned first_layer, unsigned last_layer,
3691			   unsigned width, unsigned height, unsigned depth,
3692			   uint32_t *state,
3693			   uint32_t *fmask_state)
3694{
3695	struct pipe_resource *res = &tex->buffer.b.b;
3696	const struct util_format_description *desc;
3697	unsigned char swizzle[4];
3698	int first_non_void;
3699	unsigned num_format, data_format, type, num_samples;
3700	uint64_t va;
3701
3702	desc = util_format_description(pipe_format);
3703
3704	num_samples = desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS ?
3705			MAX2(1, res->nr_samples) :
3706			MAX2(1, res->nr_storage_samples);
3707
3708	if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3709		const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3710		const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3711		const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3712
3713		switch (pipe_format) {
3714		case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3715		case PIPE_FORMAT_X32_S8X24_UINT:
3716		case PIPE_FORMAT_X8Z24_UNORM:
3717			util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3718			break;
3719		case PIPE_FORMAT_X24S8_UINT:
3720			/*
3721			 * X24S8 is implemented as an 8_8_8_8 data format, to
3722			 * fix texture gathers. This affects at least
3723			 * GL45-CTS.texture_cube_map_array.sampling on VI.
3724			 */
3725			if (screen->info.chip_class <= VI)
3726				util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3727			else
3728				util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3729			break;
3730		default:
3731			util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3732		}
3733	} else {
3734		util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3735	}
3736
3737	first_non_void = util_format_get_first_non_void_channel(pipe_format);
3738
3739	switch (pipe_format) {
3740	case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3741		num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3742		break;
3743	default:
3744		if (first_non_void < 0) {
3745			if (util_format_is_compressed(pipe_format)) {
3746				switch (pipe_format) {
3747				case PIPE_FORMAT_DXT1_SRGB:
3748				case PIPE_FORMAT_DXT1_SRGBA:
3749				case PIPE_FORMAT_DXT3_SRGBA:
3750				case PIPE_FORMAT_DXT5_SRGBA:
3751				case PIPE_FORMAT_BPTC_SRGBA:
3752				case PIPE_FORMAT_ETC2_SRGB8:
3753				case PIPE_FORMAT_ETC2_SRGB8A1:
3754				case PIPE_FORMAT_ETC2_SRGBA8:
3755					num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3756					break;
3757				case PIPE_FORMAT_RGTC1_SNORM:
3758				case PIPE_FORMAT_LATC1_SNORM:
3759				case PIPE_FORMAT_RGTC2_SNORM:
3760				case PIPE_FORMAT_LATC2_SNORM:
3761				case PIPE_FORMAT_ETC2_R11_SNORM:
3762				case PIPE_FORMAT_ETC2_RG11_SNORM:
3763				/* implies float, so use SNORM/UNORM to determine
3764				   whether data is signed or not */
3765				case PIPE_FORMAT_BPTC_RGB_FLOAT:
3766					num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3767					break;
3768				default:
3769					num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3770					break;
3771				}
3772			} else if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
3773				num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3774			} else {
3775				num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3776			}
3777		} else if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
3778			num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3779		} else {
3780			num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3781
3782			switch (desc->channel[first_non_void].type) {
3783			case UTIL_FORMAT_TYPE_FLOAT:
3784				num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3785				break;
3786			case UTIL_FORMAT_TYPE_SIGNED:
3787				if (desc->channel[first_non_void].normalized)
3788					num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3789				else if (desc->channel[first_non_void].pure_integer)
3790					num_format = V_008F14_IMG_NUM_FORMAT_SINT;
3791				else
3792					num_format = V_008F14_IMG_NUM_FORMAT_SSCALED;
3793				break;
3794			case UTIL_FORMAT_TYPE_UNSIGNED:
3795				if (desc->channel[first_non_void].normalized)
3796					num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3797				else if (desc->channel[first_non_void].pure_integer)
3798					num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3799				else
3800					num_format = V_008F14_IMG_NUM_FORMAT_USCALED;
3801			}
3802		}
3803	}
3804
3805	data_format = si_translate_texformat(&screen->b, pipe_format, desc, first_non_void);
3806	if (data_format == ~0) {
3807		data_format = 0;
3808	}
3809
3810	/* S8 with Z32 HTILE needs a special format. */
3811	if (screen->info.chip_class >= GFX9 &&
3812	    pipe_format == PIPE_FORMAT_S8_UINT &&
3813	    tex->tc_compatible_htile)
3814		data_format = V_008F14_IMG_DATA_FORMAT_S8_32;
3815
3816	if (!sampler &&
3817	    (res->target == PIPE_TEXTURE_CUBE ||
3818	     res->target == PIPE_TEXTURE_CUBE_ARRAY ||
3819	     (screen->info.chip_class <= VI &&
3820	      res->target == PIPE_TEXTURE_3D))) {
3821		/* For the purpose of shader images, treat cube maps and 3D
3822		 * textures as 2D arrays. For 3D textures, the address
3823		 * calculations for mipmaps are different, so we rely on the
3824		 * caller to effectively disable mipmaps.
3825		 */
3826		type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3827
3828		assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
3829	} else {
3830		type = si_tex_dim(screen, tex, target, num_samples);
3831	}
3832
3833	if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3834	        height = 1;
3835		depth = res->array_size;
3836	} else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
3837		   type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3838		if (sampler || res->target != PIPE_TEXTURE_3D)
3839			depth = res->array_size;
3840	} else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3841		depth = res->array_size / 6;
3842
3843	state[0] = 0;
3844	state[1] = (S_008F14_DATA_FORMAT_GFX6(data_format) |
3845		    S_008F14_NUM_FORMAT_GFX6(num_format));
3846	state[2] = (S_008F18_WIDTH(width - 1) |
3847		    S_008F18_HEIGHT(height - 1) |
3848		    S_008F18_PERF_MOD(4));
3849	state[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
3850		    S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
3851		    S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
3852		    S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
3853		    S_008F1C_BASE_LEVEL(num_samples > 1 ? 0 : first_level) |
3854		    S_008F1C_LAST_LEVEL(num_samples > 1 ?
3855					util_logbase2(num_samples) :
3856					last_level) |
3857		    S_008F1C_TYPE(type));
3858	state[4] = 0;
3859	state[5] = S_008F24_BASE_ARRAY(first_layer);
3860	state[6] = 0;
3861	state[7] = 0;
3862
3863	if (screen->info.chip_class >= GFX9) {
3864		unsigned bc_swizzle = gfx9_border_color_swizzle(desc->swizzle);
3865
3866		/* Depth is the the last accessible layer on Gfx9.
3867		 * The hw doesn't need to know the total number of layers.
3868		 */
3869		if (type == V_008F1C_SQ_RSRC_IMG_3D)
3870			state[4] |= S_008F20_DEPTH(depth - 1);
3871		else
3872			state[4] |= S_008F20_DEPTH(last_layer);
3873
3874		state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle);
3875		state[5] |= S_008F24_MAX_MIP(num_samples > 1 ?
3876					     util_logbase2(num_samples) :
3877					     tex->buffer.b.b.last_level);
3878	} else {
3879		state[3] |= S_008F1C_POW2_PAD(res->last_level > 0);
3880		state[4] |= S_008F20_DEPTH(depth - 1);
3881		state[5] |= S_008F24_LAST_ARRAY(last_layer);
3882	}
3883
3884	if (tex->dcc_offset) {
3885		state[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(pipe_format));
3886	} else {
3887		/* The last dword is unused by hw. The shader uses it to clear
3888		 * bits in the first dword of sampler state.
3889		 */
3890		if (screen->info.chip_class <= CIK && res->nr_samples <= 1) {
3891			if (first_level == last_level)
3892				state[7] = C_008F30_MAX_ANISO_RATIO;
3893			else
3894				state[7] = 0xffffffff;
3895		}
3896	}
3897
3898	/* Initialize the sampler view for FMASK. */
3899	if (tex->surface.fmask_size) {
3900		uint32_t data_format, num_format;
3901
3902		va = tex->buffer.gpu_address + tex->fmask_offset;
3903
3904#define FMASK(s,f) (((unsigned)(MAX2(1, s)) * 16) + (MAX2(1, f)))
3905		if (screen->info.chip_class >= GFX9) {
3906			data_format = V_008F14_IMG_DATA_FORMAT_FMASK;
3907			switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
3908			case FMASK(2,1):
3909				num_format = V_008F14_IMG_FMASK_8_2_1;
3910				break;
3911			case FMASK(2,2):
3912				num_format = V_008F14_IMG_FMASK_8_2_2;
3913				break;
3914			case FMASK(4,1):
3915				num_format = V_008F14_IMG_FMASK_8_4_1;
3916				break;
3917			case FMASK(4,2):
3918				num_format = V_008F14_IMG_FMASK_8_4_2;
3919				break;
3920			case FMASK(4,4):
3921				num_format = V_008F14_IMG_FMASK_8_4_4;
3922				break;
3923			case FMASK(8,1):
3924				num_format = V_008F14_IMG_FMASK_8_8_1;
3925				break;
3926			case FMASK(8,2):
3927				num_format = V_008F14_IMG_FMASK_16_8_2;
3928				break;
3929			case FMASK(8,4):
3930				num_format = V_008F14_IMG_FMASK_32_8_4;
3931				break;
3932			case FMASK(8,8):
3933				num_format = V_008F14_IMG_FMASK_32_8_8;
3934				break;
3935			case FMASK(16,1):
3936				num_format = V_008F14_IMG_FMASK_16_16_1;
3937				break;
3938			case FMASK(16,2):
3939				num_format = V_008F14_IMG_FMASK_32_16_2;
3940				break;
3941			case FMASK(16,4):
3942				num_format = V_008F14_IMG_FMASK_64_16_4;
3943				break;
3944			case FMASK(16,8):
3945				num_format = V_008F14_IMG_FMASK_64_16_8;
3946				break;
3947			default:
3948				unreachable("invalid nr_samples");
3949			}
3950		} else {
3951			switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
3952			case FMASK(2,1):
3953				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F1;
3954				break;
3955			case FMASK(2,2):
3956				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
3957				break;
3958			case FMASK(4,1):
3959				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F1;
3960				break;
3961			case FMASK(4,2):
3962				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F2;
3963				break;
3964			case FMASK(4,4):
3965				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
3966				break;
3967			case FMASK(8,1):
3968				data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S8_F1;
3969				break;
3970			case FMASK(8,2):
3971				data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S8_F2;
3972				break;
3973			case FMASK(8,4):
3974				data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F4;
3975				break;
3976			case FMASK(8,8):
3977				data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
3978				break;
3979			case FMASK(16,1):
3980				data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S16_F1;
3981				break;
3982			case FMASK(16,2):
3983				data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S16_F2;
3984				break;
3985			case FMASK(16,4):
3986				data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F4;
3987				break;
3988			case FMASK(16,8):
3989				data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F8;
3990				break;
3991			default:
3992				unreachable("invalid nr_samples");
3993			}
3994			num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3995		}
3996#undef FMASK
3997
3998		fmask_state[0] = (va >> 8) | tex->surface.fmask_tile_swizzle;
3999		fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
4000				 S_008F14_DATA_FORMAT_GFX6(data_format) |
4001				 S_008F14_NUM_FORMAT_GFX6(num_format);
4002		fmask_state[2] = S_008F18_WIDTH(width - 1) |
4003				 S_008F18_HEIGHT(height - 1);
4004		fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
4005				 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
4006				 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
4007				 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
4008				 S_008F1C_TYPE(si_tex_dim(screen, tex, target, 0));
4009		fmask_state[4] = 0;
4010		fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
4011		fmask_state[6] = 0;
4012		fmask_state[7] = 0;
4013
4014		if (screen->info.chip_class >= GFX9) {
4015			fmask_state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode);
4016			fmask_state[4] |= S_008F20_DEPTH(last_layer) |
4017					  S_008F20_PITCH_GFX9(tex->surface.u.gfx9.fmask.epitch);
4018			fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned) |
4019					  S_008F24_META_RB_ALIGNED(tex->surface.u.gfx9.cmask.rb_aligned);
4020		} else {
4021			fmask_state[3] |= S_008F1C_TILING_INDEX(tex->surface.u.legacy.fmask.tiling_index);
4022			fmask_state[4] |= S_008F20_DEPTH(depth - 1) |
4023					  S_008F20_PITCH_GFX6(tex->surface.u.legacy.fmask.pitch_in_pixels - 1);
4024			fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer);
4025		}
4026	}
4027}
4028
4029/**
4030 * Create a sampler view.
4031 *
4032 * @param ctx		context
4033 * @param texture	texture
4034 * @param state		sampler view template
4035 * @param width0	width0 override (for compressed textures as int)
4036 * @param height0	height0 override (for compressed textures as int)
4037 * @param force_level   set the base address to the level (for compressed textures)
4038 */
4039struct pipe_sampler_view *
4040si_create_sampler_view_custom(struct pipe_context *ctx,
4041			      struct pipe_resource *texture,
4042			      const struct pipe_sampler_view *state,
4043			      unsigned width0, unsigned height0,
4044			      unsigned force_level)
4045{
4046	struct si_context *sctx = (struct si_context*)ctx;
4047	struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
4048	struct si_texture *tex = (struct si_texture*)texture;
4049	unsigned base_level, first_level, last_level;
4050	unsigned char state_swizzle[4];
4051	unsigned height, depth, width;
4052	unsigned last_layer = state->u.tex.last_layer;
4053	enum pipe_format pipe_format;
4054	const struct legacy_surf_level *surflevel;
4055
4056	if (!view)
4057		return NULL;
4058
4059	/* initialize base object */
4060	view->base = *state;
4061	view->base.texture = NULL;
4062	view->base.reference.count = 1;
4063	view->base.context = ctx;
4064
4065	assert(texture);
4066	pipe_resource_reference(&view->base.texture, texture);
4067
4068	if (state->format == PIPE_FORMAT_X24S8_UINT ||
4069	    state->format == PIPE_FORMAT_S8X24_UINT ||
4070	    state->format == PIPE_FORMAT_X32_S8X24_UINT ||
4071	    state->format == PIPE_FORMAT_S8_UINT)
4072		view->is_stencil_sampler = true;
4073
4074	/* Buffer resource. */
4075	if (texture->target == PIPE_BUFFER) {
4076		si_make_buffer_descriptor(sctx->screen,
4077					  si_resource(texture),
4078					  state->format,
4079					  state->u.buf.offset,
4080					  state->u.buf.size,
4081					  view->state);
4082		return &view->base;
4083	}
4084
4085	state_swizzle[0] = state->swizzle_r;
4086	state_swizzle[1] = state->swizzle_g;
4087	state_swizzle[2] = state->swizzle_b;
4088	state_swizzle[3] = state->swizzle_a;
4089
4090	base_level = 0;
4091	first_level = state->u.tex.first_level;
4092	last_level = state->u.tex.last_level;
4093	width = width0;
4094	height = height0;
4095	depth = texture->depth0;
4096
4097	if (sctx->chip_class <= VI && force_level) {
4098		assert(force_level == first_level &&
4099		       force_level == last_level);
4100		base_level = force_level;
4101		first_level = 0;
4102		last_level = 0;
4103		width = u_minify(width, force_level);
4104		height = u_minify(height, force_level);
4105		depth = u_minify(depth, force_level);
4106	}
4107
4108	/* This is not needed if state trackers set last_layer correctly. */
4109	if (state->target == PIPE_TEXTURE_1D ||
4110	    state->target == PIPE_TEXTURE_2D ||
4111	    state->target == PIPE_TEXTURE_RECT ||
4112	    state->target == PIPE_TEXTURE_CUBE)
4113		last_layer = state->u.tex.first_layer;
4114
4115	/* Texturing with separate depth and stencil. */
4116	pipe_format = state->format;
4117
4118	/* Depth/stencil texturing sometimes needs separate texture. */
4119	if (tex->is_depth && !si_can_sample_zs(tex, view->is_stencil_sampler)) {
4120		if (!tex->flushed_depth_texture &&
4121		    !si_init_flushed_depth_texture(ctx, texture, NULL)) {
4122			pipe_resource_reference(&view->base.texture, NULL);
4123			FREE(view);
4124			return NULL;
4125		}
4126
4127		assert(tex->flushed_depth_texture);
4128
4129		/* Override format for the case where the flushed texture
4130		 * contains only Z or only S.
4131		 */
4132		if (tex->flushed_depth_texture->buffer.b.b.format != tex->buffer.b.b.format)
4133			pipe_format = tex->flushed_depth_texture->buffer.b.b.format;
4134
4135		tex = tex->flushed_depth_texture;
4136	}
4137
4138	surflevel = tex->surface.u.legacy.level;
4139
4140	if (tex->db_compatible) {
4141		if (!view->is_stencil_sampler)
4142			pipe_format = tex->db_render_format;
4143
4144		switch (pipe_format) {
4145		case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
4146			pipe_format = PIPE_FORMAT_Z32_FLOAT;
4147			break;
4148		case PIPE_FORMAT_X8Z24_UNORM:
4149		case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4150			/* Z24 is always stored like this for DB
4151			 * compatibility.
4152			 */
4153			pipe_format = PIPE_FORMAT_Z24X8_UNORM;
4154			break;
4155		case PIPE_FORMAT_X24S8_UINT:
4156		case PIPE_FORMAT_S8X24_UINT:
4157		case PIPE_FORMAT_X32_S8X24_UINT:
4158			pipe_format = PIPE_FORMAT_S8_UINT;
4159			surflevel = tex->surface.u.legacy.stencil_level;
4160			break;
4161		default:;
4162		}
4163	}
4164
4165	view->dcc_incompatible =
4166		vi_dcc_formats_are_incompatible(texture,
4167						state->u.tex.first_level,
4168						state->format);
4169
4170	si_make_texture_descriptor(sctx->screen, tex, true,
4171				   state->target, pipe_format, state_swizzle,
4172				   first_level, last_level,
4173				   state->u.tex.first_layer, last_layer,
4174				   width, height, depth,
4175				   view->state, view->fmask_state);
4176
4177	unsigned num_format = G_008F14_NUM_FORMAT_GFX6(view->state[1]);
4178	view->is_integer =
4179		num_format == V_008F14_IMG_NUM_FORMAT_USCALED ||
4180		num_format == V_008F14_IMG_NUM_FORMAT_SSCALED ||
4181		num_format == V_008F14_IMG_NUM_FORMAT_UINT ||
4182		num_format == V_008F14_IMG_NUM_FORMAT_SINT;
4183	view->base_level_info = &surflevel[base_level];
4184	view->base_level = base_level;
4185	view->block_width = util_format_get_blockwidth(pipe_format);
4186	return &view->base;
4187}
4188
4189static struct pipe_sampler_view *
4190si_create_sampler_view(struct pipe_context *ctx,
4191		       struct pipe_resource *texture,
4192		       const struct pipe_sampler_view *state)
4193{
4194	return si_create_sampler_view_custom(ctx, texture, state,
4195					     texture ? texture->width0 : 0,
4196					     texture ? texture->height0 : 0, 0);
4197}
4198
4199static void si_sampler_view_destroy(struct pipe_context *ctx,
4200				    struct pipe_sampler_view *state)
4201{
4202	struct si_sampler_view *view = (struct si_sampler_view *)state;
4203
4204	pipe_resource_reference(&state->texture, NULL);
4205	FREE(view);
4206}
4207
4208static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
4209{
4210	return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
4211	       wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
4212	       (linear_filter &&
4213	        (wrap == PIPE_TEX_WRAP_CLAMP ||
4214		 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
4215}
4216
4217static uint32_t si_translate_border_color(struct si_context *sctx,
4218					  const struct pipe_sampler_state *state,
4219					  const union pipe_color_union *color,
4220					  bool is_integer)
4221{
4222	bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
4223			     state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
4224
4225	if (!wrap_mode_uses_border_color(state->wrap_s, linear_filter) &&
4226	    !wrap_mode_uses_border_color(state->wrap_t, linear_filter) &&
4227	    !wrap_mode_uses_border_color(state->wrap_r, linear_filter))
4228		return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4229
4230#define simple_border_types(elt) \
4231do { \
4232	if (color->elt[0] == 0 && color->elt[1] == 0 &&                         \
4233	    color->elt[2] == 0 && color->elt[3] == 0)                           \
4234		return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
4235	if (color->elt[0] == 0 && color->elt[1] == 0 &&                         \
4236	    color->elt[2] == 0 && color->elt[3] == 1)                           \
4237		return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
4238	if (color->elt[0] == 1 && color->elt[1] == 1 &&                         \
4239	    color->elt[2] == 1 && color->elt[3] == 1)                           \
4240		return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
4241} while (false)
4242
4243	if (is_integer)
4244		simple_border_types(ui);
4245	else
4246		simple_border_types(f);
4247
4248#undef simple_border_types
4249
4250	int i;
4251
4252	/* Check if the border has been uploaded already. */
4253	for (i = 0; i < sctx->border_color_count; i++)
4254		if (memcmp(&sctx->border_color_table[i], color,
4255			   sizeof(*color)) == 0)
4256			break;
4257
4258	if (i >= SI_MAX_BORDER_COLORS) {
4259		/* Getting 4096 unique border colors is very unlikely. */
4260		fprintf(stderr, "radeonsi: The border color table is full. "
4261			"Any new border colors will be just black. "
4262			"Please file a bug.\n");
4263		return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4264	}
4265
4266	if (i == sctx->border_color_count) {
4267		/* Upload a new border color. */
4268		memcpy(&sctx->border_color_table[i], color,
4269		       sizeof(*color));
4270		util_memcpy_cpu_to_le32(&sctx->border_color_map[i],
4271					color, sizeof(*color));
4272		sctx->border_color_count++;
4273	}
4274
4275	return S_008F3C_BORDER_COLOR_PTR(i) |
4276	       S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER);
4277}
4278
4279static inline int S_FIXED(float value, unsigned frac_bits)
4280{
4281	return value * (1 << frac_bits);
4282}
4283
4284static inline unsigned si_tex_filter(unsigned filter, unsigned max_aniso)
4285{
4286	if (filter == PIPE_TEX_FILTER_LINEAR)
4287		return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
4288				     : V_008F38_SQ_TEX_XY_FILTER_BILINEAR;
4289	else
4290		return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
4291				     : V_008F38_SQ_TEX_XY_FILTER_POINT;
4292}
4293
4294static inline unsigned si_tex_aniso_filter(unsigned filter)
4295{
4296	if (filter < 2)
4297		return 0;
4298	if (filter < 4)
4299		return 1;
4300	if (filter < 8)
4301		return 2;
4302	if (filter < 16)
4303		return 3;
4304	return 4;
4305}
4306
4307static void *si_create_sampler_state(struct pipe_context *ctx,
4308				     const struct pipe_sampler_state *state)
4309{
4310	struct si_context *sctx = (struct si_context *)ctx;
4311	struct si_screen *sscreen = sctx->screen;
4312	struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
4313	unsigned max_aniso = sscreen->force_aniso >= 0 ? sscreen->force_aniso
4314						       : state->max_anisotropy;
4315	unsigned max_aniso_ratio = si_tex_aniso_filter(max_aniso);
4316	union pipe_color_union clamped_border_color;
4317
4318	if (!rstate) {
4319		return NULL;
4320	}
4321
4322#ifdef DEBUG
4323	rstate->magic = SI_SAMPLER_STATE_MAGIC;
4324#endif
4325	rstate->val[0] = (S_008F30_CLAMP_X(si_tex_wrap(state->wrap_s)) |
4326			  S_008F30_CLAMP_Y(si_tex_wrap(state->wrap_t)) |
4327			  S_008F30_CLAMP_Z(si_tex_wrap(state->wrap_r)) |
4328			  S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
4329			  S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state->compare_func)) |
4330			  S_008F30_FORCE_UNNORMALIZED(!state->normalized_coords) |
4331			  S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
4332			  S_008F30_ANISO_BIAS(max_aniso_ratio) |
4333			  S_008F30_DISABLE_CUBE_WRAP(!state->seamless_cube_map) |
4334			  S_008F30_COMPAT_MODE(sctx->chip_class >= VI));
4335	rstate->val[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
4336			  S_008F34_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)) |
4337			  S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
4338	rstate->val[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
4339			  S_008F38_XY_MAG_FILTER(si_tex_filter(state->mag_img_filter, max_aniso)) |
4340			  S_008F38_XY_MIN_FILTER(si_tex_filter(state->min_img_filter, max_aniso)) |
4341			  S_008F38_MIP_FILTER(si_tex_mipfilter(state->min_mip_filter)) |
4342			  S_008F38_MIP_POINT_PRECLAMP(0) |
4343			  S_008F38_DISABLE_LSB_CEIL(sctx->chip_class <= VI) |
4344			  S_008F38_FILTER_PREC_FIX(1) |
4345			  S_008F38_ANISO_OVERRIDE(sctx->chip_class >= VI));
4346	rstate->val[3] = si_translate_border_color(sctx, state, &state->border_color, false);
4347
4348	/* Create sampler resource for integer textures. */
4349	memcpy(rstate->integer_val, rstate->val, sizeof(rstate->val));
4350	rstate->integer_val[3] = si_translate_border_color(sctx, state, &state->border_color, true);
4351
4352	/* Create sampler resource for upgraded depth textures. */
4353	memcpy(rstate->upgraded_depth_val, rstate->val, sizeof(rstate->val));
4354
4355	for (unsigned i = 0; i < 4; ++i) {
4356		/* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4357		 * when the border color is 1.0. */
4358		clamped_border_color.f[i] = CLAMP(state->border_color.f[0], 0, 1);
4359	}
4360
4361	if (memcmp(&state->border_color, &clamped_border_color, sizeof(clamped_border_color)) == 0)
4362		rstate->upgraded_depth_val[3] |= S_008F3C_UPGRADED_DEPTH(1);
4363	else
4364		rstate->upgraded_depth_val[3] =
4365			si_translate_border_color(sctx, state, &clamped_border_color, false) |
4366			S_008F3C_UPGRADED_DEPTH(1);
4367
4368	return rstate;
4369}
4370
4371static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
4372{
4373	struct si_context *sctx = (struct si_context *)ctx;
4374
4375	if (sctx->sample_mask == (uint16_t)sample_mask)
4376		return;
4377
4378	sctx->sample_mask = sample_mask;
4379	si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_mask);
4380}
4381
4382static void si_emit_sample_mask(struct si_context *sctx)
4383{
4384	struct radeon_cmdbuf *cs = sctx->gfx_cs;
4385	unsigned mask = sctx->sample_mask;
4386
4387	/* Needed for line and polygon smoothing as well as for the Polaris
4388	 * small primitive filter. We expect the state tracker to take care of
4389	 * this for us.
4390	 */
4391	assert(mask == 0xffff || sctx->framebuffer.nr_samples > 1 ||
4392	       (mask & 1 && sctx->blitter->running));
4393
4394	radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
4395	radeon_emit(cs, mask | (mask << 16));
4396	radeon_emit(cs, mask | (mask << 16));
4397}
4398
4399static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
4400{
4401#ifdef DEBUG
4402	struct si_sampler_state *s = state;
4403
4404	assert(s->magic == SI_SAMPLER_STATE_MAGIC);
4405	s->magic = 0;
4406#endif
4407	free(state);
4408}
4409
4410/*
4411 * Vertex elements & buffers
4412 */
4413
4414struct util_fast_udiv_info32 {
4415   unsigned multiplier; /* the "magic number" multiplier */
4416   unsigned pre_shift; /* shift for the dividend before multiplying */
4417   unsigned post_shift; /* shift for the dividend after multiplying */
4418   int increment; /* 0 or 1; if set then increment the numerator, using one of
4419                     the two strategies */
4420};
4421
4422static struct util_fast_udiv_info32
4423util_compute_fast_udiv_info32(uint32_t D, unsigned num_bits)
4424{
4425	struct util_fast_udiv_info info =
4426		util_compute_fast_udiv_info(D, num_bits, 32);
4427
4428	struct util_fast_udiv_info32 result = {
4429		info.multiplier,
4430		info.pre_shift,
4431		info.post_shift,
4432		info.increment,
4433	};
4434	return result;
4435}
4436
4437static void *si_create_vertex_elements(struct pipe_context *ctx,
4438				       unsigned count,
4439				       const struct pipe_vertex_element *elements)
4440{
4441	struct si_screen *sscreen = (struct si_screen*)ctx->screen;
4442	struct si_vertex_elements *v = CALLOC_STRUCT(si_vertex_elements);
4443	bool used[SI_NUM_VERTEX_BUFFERS] = {};
4444	struct util_fast_udiv_info32 divisor_factors[SI_MAX_ATTRIBS] = {};
4445	STATIC_ASSERT(sizeof(struct util_fast_udiv_info32) == 16);
4446	STATIC_ASSERT(sizeof(divisor_factors[0].multiplier) == 4);
4447	STATIC_ASSERT(sizeof(divisor_factors[0].pre_shift) == 4);
4448	STATIC_ASSERT(sizeof(divisor_factors[0].post_shift) == 4);
4449	STATIC_ASSERT(sizeof(divisor_factors[0].increment) == 4);
4450	int i;
4451
4452	assert(count <= SI_MAX_ATTRIBS);
4453	if (!v)
4454		return NULL;
4455
4456	v->count = count;
4457	v->desc_list_byte_size = align(count * 16, SI_CPDMA_ALIGNMENT);
4458
4459	for (i = 0; i < count; ++i) {
4460		const struct util_format_description *desc;
4461		const struct util_format_channel_description *channel;
4462		unsigned data_format, num_format;
4463		int first_non_void;
4464		unsigned vbo_index = elements[i].vertex_buffer_index;
4465		unsigned char swizzle[4];
4466
4467		if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
4468			FREE(v);
4469			return NULL;
4470		}
4471
4472		unsigned instance_divisor = elements[i].instance_divisor;
4473		if (instance_divisor) {
4474			v->uses_instance_divisors = true;
4475
4476			if (instance_divisor == 1) {
4477				v->instance_divisor_is_one |= 1u << i;
4478			} else {
4479				v->instance_divisor_is_fetched |= 1u << i;
4480				divisor_factors[i] =
4481					util_compute_fast_udiv_info32(instance_divisor, 32);
4482			}
4483		}
4484
4485		if (!used[vbo_index]) {
4486			v->first_vb_use_mask |= 1 << i;
4487			used[vbo_index] = true;
4488		}
4489
4490		desc = util_format_description(elements[i].src_format);
4491		first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
4492		data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
4493		num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
4494		channel = first_non_void >= 0 ? &desc->channel[first_non_void] : NULL;
4495		memcpy(swizzle, desc->swizzle, sizeof(swizzle));
4496
4497		v->format_size[i] = desc->block.bits / 8;
4498		v->src_offset[i] = elements[i].src_offset;
4499		v->vertex_buffer_index[i] = vbo_index;
4500
4501		/* The hardware always treats the 2-bit alpha channel as
4502		 * unsigned, so a shader workaround is needed. The affected
4503		 * chips are VI and older except Stoney (GFX8.1).
4504		 */
4505		if (data_format == V_008F0C_BUF_DATA_FORMAT_2_10_10_10 &&
4506		    sscreen->info.chip_class <= VI &&
4507		    sscreen->info.family != CHIP_STONEY) {
4508			if (num_format == V_008F0C_BUF_NUM_FORMAT_SNORM) {
4509				v->fix_fetch[i] = SI_FIX_FETCH_A2_SNORM;
4510			} else if (num_format == V_008F0C_BUF_NUM_FORMAT_SSCALED) {
4511				v->fix_fetch[i] = SI_FIX_FETCH_A2_SSCALED;
4512			} else if (num_format == V_008F0C_BUF_NUM_FORMAT_SINT) {
4513				/* This isn't actually used in OpenGL. */
4514				v->fix_fetch[i] = SI_FIX_FETCH_A2_SINT;
4515			}
4516		} else if (channel && channel->type == UTIL_FORMAT_TYPE_FIXED) {
4517			if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4518				v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_FIXED;
4519			else
4520				v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_FIXED;
4521		} else if (channel && channel->size == 32 && !channel->pure_integer) {
4522			if (channel->type == UTIL_FORMAT_TYPE_SIGNED) {
4523				if (channel->normalized) {
4524					if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4525						v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_SNORM;
4526					else
4527						v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SNORM;
4528				} else {
4529					v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SSCALED;
4530				}
4531			} else if (channel->type == UTIL_FORMAT_TYPE_UNSIGNED) {
4532				if (channel->normalized) {
4533					if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4534						v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_UNORM;
4535					else
4536						v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_UNORM;
4537				} else {
4538					v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_USCALED;
4539				}
4540			}
4541		} else if (channel && channel->size == 64 &&
4542			   channel->type == UTIL_FORMAT_TYPE_FLOAT) {
4543			switch (desc->nr_channels) {
4544			case 1:
4545			case 2:
4546				v->fix_fetch[i] = SI_FIX_FETCH_RG_64_FLOAT;
4547				swizzle[0] = PIPE_SWIZZLE_X;
4548				swizzle[1] = PIPE_SWIZZLE_Y;
4549				swizzle[2] = desc->nr_channels == 2 ? PIPE_SWIZZLE_Z : PIPE_SWIZZLE_0;
4550				swizzle[3] = desc->nr_channels == 2 ? PIPE_SWIZZLE_W : PIPE_SWIZZLE_0;
4551				break;
4552			case 3:
4553				v->fix_fetch[i] = SI_FIX_FETCH_RGB_64_FLOAT;
4554				swizzle[0] = PIPE_SWIZZLE_X; /* 3 loads */
4555				swizzle[1] = PIPE_SWIZZLE_Y;
4556				swizzle[2] = PIPE_SWIZZLE_0;
4557				swizzle[3] = PIPE_SWIZZLE_0;
4558				break;
4559			case 4:
4560				v->fix_fetch[i] = SI_FIX_FETCH_RGBA_64_FLOAT;
4561				swizzle[0] = PIPE_SWIZZLE_X; /* 2 loads */
4562				swizzle[1] = PIPE_SWIZZLE_Y;
4563				swizzle[2] = PIPE_SWIZZLE_Z;
4564				swizzle[3] = PIPE_SWIZZLE_W;
4565				break;
4566			default:
4567				assert(0);
4568			}
4569		} else if (channel && desc->nr_channels == 3) {
4570			assert(desc->swizzle[0] == PIPE_SWIZZLE_X);
4571
4572			if (channel->size == 8) {
4573				if (channel->pure_integer)
4574					v->fix_fetch[i] = SI_FIX_FETCH_RGB_8_INT;
4575				else
4576					v->fix_fetch[i] = SI_FIX_FETCH_RGB_8;
4577			} else if (channel->size == 16) {
4578				if (channel->pure_integer)
4579					v->fix_fetch[i] = SI_FIX_FETCH_RGB_16_INT;
4580				else
4581					v->fix_fetch[i] = SI_FIX_FETCH_RGB_16;
4582			}
4583		}
4584
4585		v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
4586				   S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
4587				   S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
4588				   S_008F0C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
4589				   S_008F0C_NUM_FORMAT(num_format) |
4590				   S_008F0C_DATA_FORMAT(data_format);
4591	}
4592
4593	if (v->instance_divisor_is_fetched) {
4594		unsigned num_divisors = util_last_bit(v->instance_divisor_is_fetched);
4595
4596		v->instance_divisor_factor_buffer =
4597			(struct si_resource*)
4598			pipe_buffer_create(&sscreen->b, 0, PIPE_USAGE_DEFAULT,
4599					   num_divisors * sizeof(divisor_factors[0]));
4600		if (!v->instance_divisor_factor_buffer) {
4601			FREE(v);
4602			return NULL;
4603		}
4604		void *map = sscreen->ws->buffer_map(v->instance_divisor_factor_buffer->buf,
4605						    NULL, PIPE_TRANSFER_WRITE);
4606		memcpy(map , divisor_factors, num_divisors * sizeof(divisor_factors[0]));
4607	}
4608	return v;
4609}
4610
4611static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
4612{
4613	struct si_context *sctx = (struct si_context *)ctx;
4614	struct si_vertex_elements *old = sctx->vertex_elements;
4615	struct si_vertex_elements *v = (struct si_vertex_elements*)state;
4616
4617	sctx->vertex_elements = v;
4618	sctx->vertex_buffers_dirty = true;
4619
4620	if (v &&
4621	    (!old ||
4622	     old->count != v->count ||
4623	     old->uses_instance_divisors != v->uses_instance_divisors ||
4624	     v->uses_instance_divisors || /* we don't check which divisors changed */
4625	     memcmp(old->fix_fetch, v->fix_fetch, sizeof(v->fix_fetch[0]) * v->count)))
4626		sctx->do_update_shaders = true;
4627
4628	if (v && v->instance_divisor_is_fetched) {
4629		struct pipe_constant_buffer cb;
4630
4631		cb.buffer = &v->instance_divisor_factor_buffer->b.b;
4632		cb.user_buffer = NULL;
4633		cb.buffer_offset = 0;
4634		cb.buffer_size = 0xffffffff;
4635		si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS, &cb);
4636	}
4637}
4638
4639static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
4640{
4641	struct si_context *sctx = (struct si_context *)ctx;
4642	struct si_vertex_elements *v = (struct si_vertex_elements*)state;
4643
4644	if (sctx->vertex_elements == state)
4645		sctx->vertex_elements = NULL;
4646	si_resource_reference(&v->instance_divisor_factor_buffer, NULL);
4647	FREE(state);
4648}
4649
4650static void si_set_vertex_buffers(struct pipe_context *ctx,
4651				  unsigned start_slot, unsigned count,
4652				  const struct pipe_vertex_buffer *buffers)
4653{
4654	struct si_context *sctx = (struct si_context *)ctx;
4655	struct pipe_vertex_buffer *dst = sctx->vertex_buffer + start_slot;
4656	int i;
4657
4658	assert(start_slot + count <= ARRAY_SIZE(sctx->vertex_buffer));
4659
4660	if (buffers) {
4661		for (i = 0; i < count; i++) {
4662			const struct pipe_vertex_buffer *src = buffers + i;
4663			struct pipe_vertex_buffer *dsti = dst + i;
4664			struct pipe_resource *buf = src->buffer.resource;
4665
4666			pipe_resource_reference(&dsti->buffer.resource, buf);
4667			dsti->buffer_offset = src->buffer_offset;
4668			dsti->stride = src->stride;
4669			si_context_add_resource_size(sctx, buf);
4670			if (buf)
4671				si_resource(buf)->bind_history |= PIPE_BIND_VERTEX_BUFFER;
4672		}
4673	} else {
4674		for (i = 0; i < count; i++) {
4675			pipe_resource_reference(&dst[i].buffer.resource, NULL);
4676		}
4677	}
4678	sctx->vertex_buffers_dirty = true;
4679}
4680
4681/*
4682 * Misc
4683 */
4684
4685static void si_set_tess_state(struct pipe_context *ctx,
4686			      const float default_outer_level[4],
4687			      const float default_inner_level[2])
4688{
4689	struct si_context *sctx = (struct si_context *)ctx;
4690	struct pipe_constant_buffer cb;
4691	float array[8];
4692
4693	memcpy(array, default_outer_level, sizeof(float) * 4);
4694	memcpy(array+4, default_inner_level, sizeof(float) * 2);
4695
4696	cb.buffer = NULL;
4697	cb.user_buffer = NULL;
4698	cb.buffer_size = sizeof(array);
4699
4700	si_upload_const_buffer(sctx, (struct si_resource**)&cb.buffer,
4701			       (void*)array, sizeof(array),
4702			       &cb.buffer_offset);
4703
4704	si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
4705	pipe_resource_reference(&cb.buffer, NULL);
4706}
4707
4708static void si_texture_barrier(struct pipe_context *ctx, unsigned flags)
4709{
4710	struct si_context *sctx = (struct si_context *)ctx;
4711
4712	si_update_fb_dirtiness_after_rendering(sctx);
4713
4714	/* Multisample surfaces are flushed in si_decompress_textures. */
4715	if (sctx->framebuffer.uncompressed_cb_mask) {
4716		si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
4717					   sctx->framebuffer.CB_has_shader_readable_metadata,
4718					   sctx->framebuffer.all_DCC_pipe_aligned);
4719	}
4720}
4721
4722/* This only ensures coherency for shader image/buffer stores. */
4723static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
4724{
4725	struct si_context *sctx = (struct si_context *)ctx;
4726
4727	if (!(flags & ~PIPE_BARRIER_UPDATE))
4728		return;
4729
4730	/* Subsequent commands must wait for all shader invocations to
4731	 * complete. */
4732	sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
4733	                 SI_CONTEXT_CS_PARTIAL_FLUSH;
4734
4735	if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
4736		sctx->flags |= SI_CONTEXT_INV_SMEM_L1 |
4737				 SI_CONTEXT_INV_VMEM_L1;
4738
4739	if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
4740		     PIPE_BARRIER_SHADER_BUFFER |
4741		     PIPE_BARRIER_TEXTURE |
4742		     PIPE_BARRIER_IMAGE |
4743		     PIPE_BARRIER_STREAMOUT_BUFFER |
4744		     PIPE_BARRIER_GLOBAL_BUFFER)) {
4745		/* As far as I can tell, L1 contents are written back to L2
4746		 * automatically at end of shader, but the contents of other
4747		 * L1 caches might still be stale. */
4748		sctx->flags |= SI_CONTEXT_INV_VMEM_L1;
4749	}
4750
4751	if (flags & PIPE_BARRIER_INDEX_BUFFER) {
4752		/* Indices are read through TC L2 since VI.
4753		 * L1 isn't used.
4754		 */
4755		if (sctx->screen->info.chip_class <= CIK)
4756			sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4757	}
4758
4759	/* MSAA color, any depth and any stencil are flushed in
4760	 * si_decompress_textures when needed.
4761	 */
4762	if (flags & PIPE_BARRIER_FRAMEBUFFER &&
4763	    sctx->framebuffer.uncompressed_cb_mask) {
4764		sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
4765
4766		if (sctx->chip_class <= VI)
4767			sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4768	}
4769
4770	/* Indirect buffers use TC L2 on GFX9, but not older hw. */
4771	if (sctx->screen->info.chip_class <= VI &&
4772	    flags & PIPE_BARRIER_INDIRECT_BUFFER)
4773		sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4774}
4775
4776static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
4777{
4778	struct pipe_blend_state blend;
4779
4780	memset(&blend, 0, sizeof(blend));
4781	blend.independent_blend_enable = true;
4782	blend.rt[0].colormask = 0xf;
4783	return si_create_blend_state_mode(&sctx->b, &blend, mode);
4784}
4785
4786static void si_init_config(struct si_context *sctx);
4787
4788void si_init_state_compute_functions(struct si_context *sctx)
4789{
4790	sctx->b.create_sampler_state = si_create_sampler_state;
4791	sctx->b.delete_sampler_state = si_delete_sampler_state;
4792	sctx->b.create_sampler_view = si_create_sampler_view;
4793	sctx->b.sampler_view_destroy = si_sampler_view_destroy;
4794	sctx->b.memory_barrier = si_memory_barrier;
4795}
4796
4797void si_init_state_functions(struct si_context *sctx)
4798{
4799	sctx->atoms.s.framebuffer.emit = si_emit_framebuffer_state;
4800	sctx->atoms.s.msaa_sample_locs.emit = si_emit_msaa_sample_locs;
4801	sctx->atoms.s.db_render_state.emit = si_emit_db_render_state;
4802	sctx->atoms.s.dpbb_state.emit = si_emit_dpbb_state;
4803	sctx->atoms.s.msaa_config.emit = si_emit_msaa_config;
4804	sctx->atoms.s.sample_mask.emit = si_emit_sample_mask;
4805	sctx->atoms.s.cb_render_state.emit = si_emit_cb_render_state;
4806	sctx->atoms.s.blend_color.emit = si_emit_blend_color;
4807	sctx->atoms.s.clip_regs.emit = si_emit_clip_regs;
4808	sctx->atoms.s.clip_state.emit = si_emit_clip_state;
4809	sctx->atoms.s.stencil_ref.emit = si_emit_stencil_ref;
4810
4811	sctx->b.create_blend_state = si_create_blend_state;
4812	sctx->b.bind_blend_state = si_bind_blend_state;
4813	sctx->b.delete_blend_state = si_delete_blend_state;
4814	sctx->b.set_blend_color = si_set_blend_color;
4815
4816	sctx->b.create_rasterizer_state = si_create_rs_state;
4817	sctx->b.bind_rasterizer_state = si_bind_rs_state;
4818	sctx->b.delete_rasterizer_state = si_delete_rs_state;
4819
4820	sctx->b.create_depth_stencil_alpha_state = si_create_dsa_state;
4821	sctx->b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
4822	sctx->b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
4823
4824	sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
4825	sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
4826	sctx->custom_blend_fmask_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
4827	sctx->custom_blend_eliminate_fastclear = si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
4828	sctx->custom_blend_dcc_decompress = si_create_blend_custom(sctx, V_028808_CB_DCC_DECOMPRESS);
4829
4830	sctx->b.set_clip_state = si_set_clip_state;
4831	sctx->b.set_stencil_ref = si_set_stencil_ref;
4832
4833	sctx->b.set_framebuffer_state = si_set_framebuffer_state;
4834
4835	sctx->b.set_sample_mask = si_set_sample_mask;
4836
4837	sctx->b.create_vertex_elements_state = si_create_vertex_elements;
4838	sctx->b.bind_vertex_elements_state = si_bind_vertex_elements;
4839	sctx->b.delete_vertex_elements_state = si_delete_vertex_element;
4840	sctx->b.set_vertex_buffers = si_set_vertex_buffers;
4841
4842	sctx->b.texture_barrier = si_texture_barrier;
4843	sctx->b.set_min_samples = si_set_min_samples;
4844	sctx->b.set_tess_state = si_set_tess_state;
4845
4846	sctx->b.set_active_query_state = si_set_active_query_state;
4847
4848	si_init_config(sctx);
4849}
4850
4851void si_init_screen_state_functions(struct si_screen *sscreen)
4852{
4853	sscreen->b.is_format_supported = si_is_format_supported;
4854}
4855
4856static void si_set_grbm_gfx_index(struct si_context *sctx,
4857				  struct si_pm4_state *pm4,  unsigned value)
4858{
4859	unsigned reg = sctx->chip_class >= CIK ? R_030800_GRBM_GFX_INDEX :
4860						   R_00802C_GRBM_GFX_INDEX;
4861	si_pm4_set_reg(pm4, reg, value);
4862}
4863
4864static void si_set_grbm_gfx_index_se(struct si_context *sctx,
4865				     struct si_pm4_state *pm4, unsigned se)
4866{
4867	assert(se == ~0 || se < sctx->screen->info.max_se);
4868	si_set_grbm_gfx_index(sctx, pm4,
4869			      (se == ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
4870					  S_030800_SE_INDEX(se)) |
4871			      S_030800_SH_BROADCAST_WRITES(1) |
4872			      S_030800_INSTANCE_BROADCAST_WRITES(1));
4873}
4874
4875static void
4876si_write_harvested_raster_configs(struct si_context *sctx,
4877				  struct si_pm4_state *pm4,
4878				  unsigned raster_config,
4879				  unsigned raster_config_1)
4880{
4881	unsigned num_se = MAX2(sctx->screen->info.max_se, 1);
4882	unsigned raster_config_se[4];
4883	unsigned se;
4884
4885	ac_get_harvested_configs(&sctx->screen->info,
4886				 raster_config,
4887				 &raster_config_1,
4888				 raster_config_se);
4889
4890	for (se = 0; se < num_se; se++) {
4891		si_set_grbm_gfx_index_se(sctx, pm4, se);
4892		si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
4893	}
4894	si_set_grbm_gfx_index(sctx, pm4, ~0);
4895
4896	if (sctx->chip_class >= CIK) {
4897		si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
4898	}
4899}
4900
4901static void si_set_raster_config(struct si_context *sctx, struct si_pm4_state *pm4)
4902{
4903	struct si_screen *sscreen = sctx->screen;
4904	unsigned num_rb = MIN2(sscreen->info.num_render_backends, 16);
4905	unsigned rb_mask = sscreen->info.enabled_rb_mask;
4906	unsigned raster_config = sscreen->pa_sc_raster_config;
4907	unsigned raster_config_1 = sscreen->pa_sc_raster_config_1;
4908
4909	if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
4910		/* Always use the default config when all backends are enabled
4911		 * (or when we failed to determine the enabled backends).
4912		 */
4913		si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG,
4914			       raster_config);
4915		if (sctx->chip_class >= CIK)
4916			si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1,
4917				       raster_config_1);
4918	} else {
4919		si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
4920	}
4921}
4922
4923static void si_init_config(struct si_context *sctx)
4924{
4925	struct si_screen *sscreen = sctx->screen;
4926	uint64_t border_color_va = sctx->border_color_buffer->gpu_address;
4927	bool has_clear_state = sscreen->has_clear_state;
4928	struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
4929
4930       /* SI, radeon kernel disabled CLEAR_STATE. */
4931       assert(has_clear_state || sscreen->info.chip_class == SI ||
4932              sscreen->info.drm_major != 3);
4933
4934	if (!pm4)
4935		return;
4936
4937	si_pm4_cmd_begin(pm4, PKT3_CONTEXT_CONTROL);
4938	si_pm4_cmd_add(pm4, CONTEXT_CONTROL_LOAD_ENABLE(1));
4939	si_pm4_cmd_add(pm4, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4940	si_pm4_cmd_end(pm4, false);
4941
4942	if (has_clear_state) {
4943		si_pm4_cmd_begin(pm4, PKT3_CLEAR_STATE);
4944		si_pm4_cmd_add(pm4, 0);
4945		si_pm4_cmd_end(pm4, false);
4946	}
4947
4948	if (sctx->chip_class <= VI)
4949		si_set_raster_config(sctx, pm4);
4950
4951	si_pm4_set_reg(pm4, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
4952	if (!has_clear_state)
4953		si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
4954
4955	/* FIXME calculate these values somehow ??? */
4956	if (sctx->chip_class <= VI) {
4957		si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
4958		si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
4959	}
4960
4961	if (!has_clear_state) {
4962		si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
4963		si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
4964		si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
4965	}
4966
4967	si_pm4_set_reg(pm4, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
4968	if (!has_clear_state)
4969		si_pm4_set_reg(pm4, R_028AB8_VGT_VTX_CNT_EN, 0x0);
4970	if (sctx->chip_class < CIK)
4971		si_pm4_set_reg(pm4, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
4972			       S_008A14_CLIP_VTX_REORDER_ENA(1));
4973
4974	/* CLEAR_STATE doesn't clear these correctly on certain generations.
4975	 * I don't know why. Deduced by trial and error.
4976	 */
4977	if (sctx->chip_class <= CIK) {
4978		si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
4979		si_pm4_set_reg(pm4, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
4980		si_pm4_set_reg(pm4, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
4981		si_pm4_set_reg(pm4, R_028244_PA_SC_GENERIC_SCISSOR_BR,
4982			       S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4983		si_pm4_set_reg(pm4, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
4984		si_pm4_set_reg(pm4, R_028034_PA_SC_SCREEN_SCISSOR_BR,
4985			       S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4986	}
4987
4988	if (!has_clear_state) {
4989		si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE,
4990			       S_028230_ER_TRI(0xA) |
4991			       S_028230_ER_POINT(0xA) |
4992			       S_028230_ER_RECT(0xA) |
4993			       /* Required by DX10_DIAMOND_TEST_ENA: */
4994			       S_028230_ER_LINE_LR(0x1A) |
4995			       S_028230_ER_LINE_RL(0x26) |
4996			       S_028230_ER_LINE_TB(0xA) |
4997			       S_028230_ER_LINE_BT(0xA));
4998		si_pm4_set_reg(pm4, R_028820_PA_CL_NANINF_CNTL, 0);
4999		si_pm4_set_reg(pm4, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
5000		si_pm4_set_reg(pm4, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
5001		si_pm4_set_reg(pm4, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
5002		si_pm4_set_reg(pm4, R_02800C_DB_RENDER_OVERRIDE, 0);
5003	}
5004
5005	if (sctx->chip_class >= GFX9) {
5006		si_pm4_set_reg(pm4, R_030920_VGT_MAX_VTX_INDX, ~0);
5007		si_pm4_set_reg(pm4, R_030924_VGT_MIN_VTX_INDX, 0);
5008		si_pm4_set_reg(pm4, R_030928_VGT_INDX_OFFSET, 0);
5009	} else {
5010		/* These registers, when written, also overwrite the CLEAR_STATE
5011		 * context, so we can't rely on CLEAR_STATE setting them.
5012		 * It would be an issue if there was another UMD changing them.
5013		 */
5014		si_pm4_set_reg(pm4, R_028400_VGT_MAX_VTX_INDX, ~0);
5015		si_pm4_set_reg(pm4, R_028404_VGT_MIN_VTX_INDX, 0);
5016		si_pm4_set_reg(pm4, R_028408_VGT_INDX_OFFSET, 0);
5017	}
5018
5019	if (sctx->chip_class >= CIK) {
5020		if (sctx->chip_class >= GFX9) {
5021			si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
5022				       S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
5023		} else {
5024			si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
5025				       S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
5026			si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
5027				       S_00B41C_WAVE_LIMIT(0x3F));
5028			si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
5029				       S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
5030
5031			/* If this is 0, Bonaire can hang even if GS isn't being used.
5032			 * Other chips are unaffected. These are suboptimal values,
5033			 * but we don't use on-chip GS.
5034			 */
5035			si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
5036				       S_028A44_ES_VERTS_PER_SUBGRP(64) |
5037				       S_028A44_GS_PRIMS_PER_SUBGRP(4));
5038		}
5039		si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
5040			       S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
5041
5042		/* Compute LATE_ALLOC_VS.LIMIT. */
5043		unsigned num_cu_per_sh = sscreen->info.num_good_cu_per_sh;
5044		unsigned late_alloc_limit; /* The limit is per SH. */
5045
5046		if (sctx->family == CHIP_KABINI) {
5047			late_alloc_limit = 0; /* Potential hang on Kabini. */
5048		} else if (num_cu_per_sh <= 4) {
5049			/* Too few available compute units per SH. Disallowing
5050			 * VS to run on one CU could hurt us more than late VS
5051			 * allocation would help.
5052			 *
5053			 * 2 is the highest safe number that allows us to keep
5054			 * all CUs enabled.
5055			 */
5056			late_alloc_limit = 2;
5057		} else {
5058			/* This is a good initial value, allowing 1 late_alloc
5059			 * wave per SIMD on num_cu - 2.
5060			 */
5061			late_alloc_limit = (num_cu_per_sh - 2) * 4;
5062
5063			/* The limit is 0-based, so 0 means 1. */
5064			assert(late_alloc_limit > 0 && late_alloc_limit <= 64);
5065			late_alloc_limit -= 1;
5066		}
5067
5068		/* VS can't execute on one CU if the limit is > 2. */
5069		si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
5070			       S_00B118_CU_EN(late_alloc_limit > 2 ? 0xfffe : 0xffff) |
5071			       S_00B118_WAVE_LIMIT(0x3F));
5072		si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
5073			       S_00B11C_LIMIT(late_alloc_limit));
5074		si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
5075			       S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
5076	}
5077
5078	if (sctx->chip_class >= VI) {
5079		unsigned vgt_tess_distribution;
5080
5081		vgt_tess_distribution =
5082			S_028B50_ACCUM_ISOLINE(32) |
5083			S_028B50_ACCUM_TRI(11) |
5084			S_028B50_ACCUM_QUAD(11) |
5085			S_028B50_DONUT_SPLIT(16);
5086
5087		/* Testing with Unigine Heaven extreme tesselation yielded best results
5088		 * with TRAP_SPLIT = 3.
5089		 */
5090		if (sctx->family == CHIP_FIJI ||
5091		    sctx->family >= CHIP_POLARIS10)
5092			vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
5093
5094		si_pm4_set_reg(pm4, R_028B50_VGT_TESS_DISTRIBUTION, vgt_tess_distribution);
5095	} else if (!has_clear_state) {
5096		si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
5097		si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
5098	}
5099
5100	si_pm4_set_reg(pm4, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
5101	if (sctx->chip_class >= CIK) {
5102		si_pm4_set_reg(pm4, R_028084_TA_BC_BASE_ADDR_HI,
5103			       S_028084_ADDRESS(border_color_va >> 40));
5104	}
5105	si_pm4_add_bo(pm4, sctx->border_color_buffer, RADEON_USAGE_READ,
5106		      RADEON_PRIO_BORDER_COLORS);
5107
5108	if (sctx->chip_class >= GFX9) {
5109		unsigned num_se = sscreen->info.max_se;
5110		unsigned pc_lines = 0;
5111
5112		switch (sctx->family) {
5113		case CHIP_VEGA10:
5114		case CHIP_VEGA12:
5115		case CHIP_VEGA20:
5116			pc_lines = 4096;
5117			break;
5118		case CHIP_RAVEN:
5119		case CHIP_RAVEN2:
5120			pc_lines = 1024;
5121			break;
5122		default:
5123			assert(0);
5124		}
5125
5126		si_pm4_set_reg(pm4, R_028C48_PA_SC_BINNER_CNTL_1,
5127			       S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
5128			       S_028C48_MAX_PRIM_PER_BATCH(1023));
5129		si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
5130			       S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5131		si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
5132	}
5133
5134	si_pm4_upload_indirect_buffer(sctx, pm4);
5135	sctx->init_config = pm4;
5136}
5137