si_cmd_buffer.c revision b8e80941
1/*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based on si_state.c
6 * Copyright © 2015 Advanced Micro Devices, Inc.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28/* command buffer handling for SI */
29
30#include "radv_private.h"
31#include "radv_shader.h"
32#include "radv_cs.h"
33#include "sid.h"
34#include "gfx9d.h"
35#include "radv_util.h"
36#include "main/macros.h"
37
38static void
39si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
40                                  struct radeon_cmdbuf *cs,
41				  unsigned raster_config,
42				  unsigned raster_config_1)
43{
44	unsigned num_se = MAX2(physical_device->rad_info.max_se, 1);
45	unsigned raster_config_se[4];
46	unsigned se;
47
48	ac_get_harvested_configs(&physical_device->rad_info,
49				 raster_config,
50				 &raster_config_1,
51				 raster_config_se);
52
53	for (se = 0; se < num_se; se++) {
54		/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
55		if (physical_device->rad_info.chip_class < CIK)
56			radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
57					      S_00802C_SE_INDEX(se) |
58					      S_00802C_SH_BROADCAST_WRITES(1) |
59					      S_00802C_INSTANCE_BROADCAST_WRITES(1));
60		else
61			radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
62					       S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
63					       S_030800_INSTANCE_BROADCAST_WRITES(1));
64		radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
65	}
66
67	/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
68	if (physical_device->rad_info.chip_class < CIK)
69		radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
70				      S_00802C_SE_BROADCAST_WRITES(1) |
71				      S_00802C_SH_BROADCAST_WRITES(1) |
72				      S_00802C_INSTANCE_BROADCAST_WRITES(1));
73	else
74		radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
75				       S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
76				       S_030800_INSTANCE_BROADCAST_WRITES(1));
77
78	if (physical_device->rad_info.chip_class >= CIK)
79		radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
80}
81
82void
83si_emit_compute(struct radv_physical_device *physical_device,
84                struct radeon_cmdbuf *cs)
85{
86	radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
87	radeon_emit(cs, 0);
88	radeon_emit(cs, 0);
89	radeon_emit(cs, 0);
90
91	radeon_set_sh_reg_seq(cs, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, 2);
92	/* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
93	radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
94	radeon_emit(cs, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
95
96	if (physical_device->rad_info.chip_class >= CIK) {
97		/* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
98		radeon_set_sh_reg_seq(cs,
99				      R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
100		radeon_emit(cs, S_00B864_SH0_CU_EN(0xffff) |
101			    S_00B864_SH1_CU_EN(0xffff));
102		radeon_emit(cs, S_00B868_SH0_CU_EN(0xffff) |
103			    S_00B868_SH1_CU_EN(0xffff));
104	}
105
106	/* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
107	 * and is now per pipe, so it should be handled in the
108	 * kernel if we want to use something other than the default value,
109	 * which is now 0x22f.
110	 */
111	if (physical_device->rad_info.chip_class <= SI) {
112		/* XXX: This should be:
113		 * (number of compute units) * 4 * (waves per simd) - 1 */
114
115		radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
116		                  0x190 /* Default value */);
117	}
118}
119
120/* 12.4 fixed-point */
121static unsigned radv_pack_float_12p4(float x)
122{
123	return x <= 0    ? 0 :
124	       x >= 4096 ? 0xffff : x * 16;
125}
126
127static void
128si_set_raster_config(struct radv_physical_device *physical_device,
129		     struct radeon_cmdbuf *cs)
130{
131	unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
132	unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
133	unsigned raster_config, raster_config_1;
134
135	ac_get_raster_config(&physical_device->rad_info,
136			     &raster_config,
137			     &raster_config_1, NULL);
138
139	/* Always use the default config when all backends are enabled
140	 * (or when we failed to determine the enabled backends).
141	 */
142	if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
143		radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG,
144				       raster_config);
145		if (physical_device->rad_info.chip_class >= CIK)
146			radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
147					       raster_config_1);
148	} else {
149		si_write_harvested_raster_configs(physical_device, cs,
150						  raster_config,
151						  raster_config_1);
152	}
153}
154
155void
156si_emit_graphics(struct radv_physical_device *physical_device,
157		 struct radeon_cmdbuf *cs)
158{
159	int i;
160
161	/* Only SI can disable CLEAR_STATE for now. */
162	assert(physical_device->has_clear_state ||
163	       physical_device->rad_info.chip_class == SI);
164
165	radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
166	radeon_emit(cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
167	radeon_emit(cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
168
169	if (physical_device->has_clear_state) {
170		radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
171		radeon_emit(cs, 0);
172	}
173
174	if (physical_device->rad_info.chip_class <= VI)
175		si_set_raster_config(physical_device, cs);
176
177	radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
178	if (!physical_device->has_clear_state)
179		radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
180
181	/* FIXME calculate these values somehow ??? */
182	if (physical_device->rad_info.chip_class <= VI) {
183		radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
184		radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
185	}
186
187	if (!physical_device->has_clear_state) {
188		radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
189		radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
190		radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
191	}
192
193	radeon_set_context_reg(cs, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
194	if (!physical_device->has_clear_state)
195		radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
196	if (physical_device->rad_info.chip_class < CIK)
197		radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
198				      S_008A14_CLIP_VTX_REORDER_ENA(1));
199
200	radeon_set_context_reg(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
201	radeon_set_context_reg(cs, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
202
203	if (!physical_device->has_clear_state)
204		radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
205
206	/* CLEAR_STATE doesn't clear these correctly on certain generations.
207	 * I don't know why. Deduced by trial and error.
208	 */
209	if (physical_device->rad_info.chip_class <= CIK) {
210		radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
211		radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL,
212				       S_028204_WINDOW_OFFSET_DISABLE(1));
213		radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL,
214				       S_028240_WINDOW_OFFSET_DISABLE(1));
215		radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
216				       S_028244_BR_X(16384) | S_028244_BR_Y(16384));
217		radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
218		radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
219				       S_028034_BR_X(16384) | S_028034_BR_Y(16384));
220	}
221
222	if (!physical_device->has_clear_state) {
223		for (i = 0; i < 16; i++) {
224			radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
225			radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
226		}
227	}
228
229	if (!physical_device->has_clear_state) {
230		radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
231		radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
232		/* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
233		radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
234		radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
235		radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
236		radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
237		radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
238	}
239
240	radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE,
241			       S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
242			       S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
243
244	if (physical_device->rad_info.chip_class >= GFX9) {
245		radeon_set_uconfig_reg(cs, R_030920_VGT_MAX_VTX_INDX, ~0);
246		radeon_set_uconfig_reg(cs, R_030924_VGT_MIN_VTX_INDX, 0);
247		radeon_set_uconfig_reg(cs, R_030928_VGT_INDX_OFFSET, 0);
248	} else {
249		/* These registers, when written, also overwrite the
250		 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
251		 * them.  It would be an issue if there was another UMD
252		 * changing them.
253		 */
254		radeon_set_context_reg(cs, R_028400_VGT_MAX_VTX_INDX, ~0);
255		radeon_set_context_reg(cs, R_028404_VGT_MIN_VTX_INDX, 0);
256		radeon_set_context_reg(cs, R_028408_VGT_INDX_OFFSET, 0);
257	}
258
259	if (physical_device->rad_info.chip_class >= CIK) {
260		if (physical_device->rad_info.chip_class >= GFX9) {
261			radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
262					  S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
263		} else {
264			radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
265					  S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
266			radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
267					  S_00B41C_WAVE_LIMIT(0x3F));
268			radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
269					  S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
270			/* If this is 0, Bonaire can hang even if GS isn't being used.
271			 * Other chips are unaffected. These are suboptimal values,
272			 * but we don't use on-chip GS.
273			 */
274			radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL,
275					       S_028A44_ES_VERTS_PER_SUBGRP(64) |
276					       S_028A44_GS_PRIMS_PER_SUBGRP(4));
277		}
278		radeon_set_sh_reg(cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
279				  S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
280
281		if (physical_device->rad_info.num_good_cu_per_sh <= 4) {
282			/* Too few available compute units per SH. Disallowing
283			 * VS to run on CU0 could hurt us more than late VS
284			 * allocation would help.
285			 *
286			 * LATE_ALLOC_VS = 2 is the highest safe number.
287			 */
288			radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
289					  S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
290			radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
291		} else {
292			/* Set LATE_ALLOC_VS == 31. It should be less than
293			 * the number of scratch waves. Limitations:
294			 * - VS can't execute on CU0.
295			 * - If HS writes outputs to LDS, LS can't execute on CU0.
296			 */
297			radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
298					  S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
299			radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
300		}
301
302		radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
303				  S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
304	}
305
306	if (physical_device->rad_info.chip_class >= VI) {
307		uint32_t vgt_tess_distribution;
308
309		vgt_tess_distribution = S_028B50_ACCUM_ISOLINE(32) |
310			S_028B50_ACCUM_TRI(11) |
311			S_028B50_ACCUM_QUAD(11) |
312			S_028B50_DONUT_SPLIT(16);
313
314		if (physical_device->rad_info.family == CHIP_FIJI ||
315		    physical_device->rad_info.family >= CHIP_POLARIS10)
316			vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
317
318		radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
319				       vgt_tess_distribution);
320	} else if (!physical_device->has_clear_state) {
321		radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
322		radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
323	}
324
325	if (physical_device->rad_info.chip_class >= GFX9) {
326		unsigned num_se = physical_device->rad_info.max_se;
327		unsigned pc_lines = 0;
328
329		switch (physical_device->rad_info.family) {
330		case CHIP_VEGA10:
331		case CHIP_VEGA12:
332		case CHIP_VEGA20:
333			pc_lines = 4096;
334			break;
335		case CHIP_RAVEN:
336		case CHIP_RAVEN2:
337			pc_lines = 1024;
338			break;
339		default:
340			assert(0);
341		}
342
343		radeon_set_context_reg(cs, R_028C48_PA_SC_BINNER_CNTL_1,
344				       S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
345				       S_028C48_MAX_PRIM_PER_BATCH(1023));
346		radeon_set_context_reg(cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
347				       S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
348		radeon_set_uconfig_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
349	}
350
351	unsigned tmp = (unsigned)(1.0 * 8.0);
352	radeon_set_context_reg_seq(cs, R_028A00_PA_SU_POINT_SIZE, 1);
353	radeon_emit(cs, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
354	radeon_set_context_reg_seq(cs, R_028A04_PA_SU_POINT_MINMAX, 1);
355	radeon_emit(cs, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
356		    S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
357
358	if (!physical_device->has_clear_state) {
359		radeon_set_context_reg(cs, R_028004_DB_COUNT_CONTROL,
360				       S_028004_ZPASS_INCREMENT_DISABLE(1));
361	}
362
363	/* Enable the Polaris small primitive filter control.
364	 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
365	 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
366	 * and AMDVLK doesn't have a workaround as well.
367	 */
368	if (physical_device->rad_info.family >= CHIP_POLARIS10) {
369		unsigned small_prim_filter_cntl =
370			S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
371			/* Workaround for a hw line bug. */
372			S_028830_LINE_FILTER_DISABLE(physical_device->rad_info.family <= CHIP_POLARIS12);
373
374		radeon_set_context_reg(cs, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
375				       small_prim_filter_cntl);
376	}
377
378	si_emit_compute(physical_device, cs);
379}
380
381void
382cik_create_gfx_config(struct radv_device *device)
383{
384	struct radeon_cmdbuf *cs = device->ws->cs_create(device->ws, RING_GFX);
385	if (!cs)
386		return;
387
388	si_emit_graphics(device->physical_device, cs);
389
390	while (cs->cdw & 7) {
391		if (device->physical_device->rad_info.gfx_ib_pad_with_type2)
392			radeon_emit(cs, 0x80000000);
393		else
394			radeon_emit(cs, 0xffff1000);
395	}
396
397	device->gfx_init = device->ws->buffer_create(device->ws,
398						     cs->cdw * 4, 4096,
399						     RADEON_DOMAIN_GTT,
400						     RADEON_FLAG_CPU_ACCESS|
401						     RADEON_FLAG_NO_INTERPROCESS_SHARING |
402						     RADEON_FLAG_READ_ONLY,
403						     RADV_BO_PRIORITY_CS);
404	if (!device->gfx_init)
405		goto fail;
406
407	void *map = device->ws->buffer_map(device->gfx_init);
408	if (!map) {
409		device->ws->buffer_destroy(device->gfx_init);
410		device->gfx_init = NULL;
411		goto fail;
412	}
413	memcpy(map, cs->buf, cs->cdw * 4);
414
415	device->ws->buffer_unmap(device->gfx_init);
416	device->gfx_init_size_dw = cs->cdw;
417fail:
418	device->ws->cs_destroy(cs);
419}
420
421static void
422get_viewport_xform(const VkViewport *viewport,
423                   float scale[3], float translate[3])
424{
425	float x = viewport->x;
426	float y = viewport->y;
427	float half_width = 0.5f * viewport->width;
428	float half_height = 0.5f * viewport->height;
429	double n = viewport->minDepth;
430	double f = viewport->maxDepth;
431
432	scale[0] = half_width;
433	translate[0] = half_width + x;
434	scale[1] = half_height;
435	translate[1] = half_height + y;
436
437	scale[2] = (f - n);
438	translate[2] = n;
439}
440
441void
442si_write_viewport(struct radeon_cmdbuf *cs, int first_vp,
443                  int count, const VkViewport *viewports)
444{
445	int i;
446
447	assert(count);
448	radeon_set_context_reg_seq(cs, R_02843C_PA_CL_VPORT_XSCALE +
449				   first_vp * 4 * 6, count * 6);
450
451	for (i = 0; i < count; i++) {
452		float scale[3], translate[3];
453
454
455		get_viewport_xform(&viewports[i], scale, translate);
456		radeon_emit(cs, fui(scale[0]));
457		radeon_emit(cs, fui(translate[0]));
458		radeon_emit(cs, fui(scale[1]));
459		radeon_emit(cs, fui(translate[1]));
460		radeon_emit(cs, fui(scale[2]));
461		radeon_emit(cs, fui(translate[2]));
462	}
463
464	radeon_set_context_reg_seq(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 +
465				   first_vp * 4 * 2, count * 2);
466	for (i = 0; i < count; i++) {
467		float zmin = MIN2(viewports[i].minDepth, viewports[i].maxDepth);
468		float zmax = MAX2(viewports[i].minDepth, viewports[i].maxDepth);
469		radeon_emit(cs, fui(zmin));
470		radeon_emit(cs, fui(zmax));
471	}
472}
473
474static VkRect2D si_scissor_from_viewport(const VkViewport *viewport)
475{
476	float scale[3], translate[3];
477	VkRect2D rect;
478
479	get_viewport_xform(viewport, scale, translate);
480
481	rect.offset.x = translate[0] - fabs(scale[0]);
482	rect.offset.y = translate[1] - fabs(scale[1]);
483	rect.extent.width = ceilf(translate[0] + fabs(scale[0])) - rect.offset.x;
484	rect.extent.height = ceilf(translate[1] + fabs(scale[1])) - rect.offset.y;
485
486	return rect;
487}
488
489static VkRect2D si_intersect_scissor(const VkRect2D *a, const VkRect2D *b) {
490	VkRect2D ret;
491	ret.offset.x = MAX2(a->offset.x, b->offset.x);
492	ret.offset.y = MAX2(a->offset.y, b->offset.y);
493	ret.extent.width = MIN2(a->offset.x + a->extent.width,
494	                        b->offset.x + b->extent.width) - ret.offset.x;
495	ret.extent.height = MIN2(a->offset.y + a->extent.height,
496	                         b->offset.y + b->extent.height) - ret.offset.y;
497	return ret;
498}
499
500void
501si_write_scissors(struct radeon_cmdbuf *cs, int first,
502                  int count, const VkRect2D *scissors,
503                  const VkViewport *viewports, bool can_use_guardband)
504{
505	int i;
506	float scale[3], translate[3], guardband_x = INFINITY, guardband_y = INFINITY;
507	const float max_range = 32767.0f;
508	if (!count)
509		return;
510
511	radeon_set_context_reg_seq(cs, R_028250_PA_SC_VPORT_SCISSOR_0_TL + first * 4 * 2, count * 2);
512	for (i = 0; i < count; i++) {
513		VkRect2D viewport_scissor = si_scissor_from_viewport(viewports + i);
514		VkRect2D scissor = si_intersect_scissor(&scissors[i], &viewport_scissor);
515
516		get_viewport_xform(viewports + i, scale, translate);
517		scale[0] = fabsf(scale[0]);
518		scale[1] = fabsf(scale[1]);
519
520		if (scale[0] < 0.5)
521			scale[0] = 0.5;
522		if (scale[1] < 0.5)
523			scale[1] = 0.5;
524
525		guardband_x = MIN2(guardband_x, (max_range - fabsf(translate[0])) / scale[0]);
526		guardband_y = MIN2(guardband_y, (max_range - fabsf(translate[1])) / scale[1]);
527
528		radeon_emit(cs, S_028250_TL_X(scissor.offset.x) |
529			    S_028250_TL_Y(scissor.offset.y) |
530			    S_028250_WINDOW_OFFSET_DISABLE(1));
531		radeon_emit(cs, S_028254_BR_X(scissor.offset.x + scissor.extent.width) |
532			    S_028254_BR_Y(scissor.offset.y + scissor.extent.height));
533	}
534	if (!can_use_guardband) {
535		guardband_x = 1.0;
536		guardband_y = 1.0;
537	}
538
539	radeon_set_context_reg_seq(cs, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ, 4);
540	radeon_emit(cs, fui(guardband_y));
541	radeon_emit(cs, fui(1.0));
542	radeon_emit(cs, fui(guardband_x));
543	radeon_emit(cs, fui(1.0));
544}
545
546static inline unsigned
547radv_prims_for_vertices(struct radv_prim_vertex_count *info, unsigned num)
548{
549	if (num == 0)
550		return 0;
551
552	if (info->incr == 0)
553		return 0;
554
555	if (num < info->min)
556		return 0;
557
558	return 1 + ((num - info->min) / info->incr);
559}
560
561uint32_t
562si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
563			  bool instanced_draw, bool indirect_draw,
564			  bool count_from_stream_output,
565			  uint32_t draw_vertex_count)
566{
567	enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
568	enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
569	struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
570	const unsigned max_primgroup_in_wave = 2;
571	/* SWITCH_ON_EOP(0) is always preferable. */
572	bool wd_switch_on_eop = false;
573	bool ia_switch_on_eop = false;
574	bool ia_switch_on_eoi = false;
575	bool partial_vs_wave = false;
576	bool partial_es_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_es_wave;
577	bool multi_instances_smaller_than_primgroup;
578
579	multi_instances_smaller_than_primgroup = indirect_draw;
580	if (!multi_instances_smaller_than_primgroup && instanced_draw) {
581		uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
582		if (num_prims < cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.primgroup_size)
583			multi_instances_smaller_than_primgroup = true;
584	}
585
586	ia_switch_on_eoi = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.ia_switch_on_eoi;
587	partial_vs_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_vs_wave;
588
589	if (chip_class >= CIK) {
590		wd_switch_on_eop = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.wd_switch_on_eop;
591
592		/* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
593		 * We don't know that for indirect drawing, so treat it as
594		 * always problematic. */
595		if (family == CHIP_HAWAII &&
596		    (instanced_draw || indirect_draw))
597			wd_switch_on_eop = true;
598
599		/* Performance recommendation for 4 SE Gfx7-8 parts if
600		 * instances are smaller than a primgroup.
601		 * Assume indirect draws always use small instances.
602		 * This is needed for good VS wave utilization.
603		 */
604		if (chip_class <= VI &&
605		    info->max_se == 4 &&
606		    multi_instances_smaller_than_primgroup)
607			wd_switch_on_eop = true;
608
609		/* Required on CIK and later. */
610		if (info->max_se > 2 && !wd_switch_on_eop)
611			ia_switch_on_eoi = true;
612
613		/* Required by Hawaii and, for some special cases, by VI. */
614		if (ia_switch_on_eoi &&
615		    (family == CHIP_HAWAII ||
616		     (chip_class == VI &&
617		      /* max primgroup in wave is always 2 - leave this for documentation */
618		      (radv_pipeline_has_gs(cmd_buffer->state.pipeline) || max_primgroup_in_wave != 2))))
619			partial_vs_wave = true;
620
621		/* Instancing bug on Bonaire. */
622		if (family == CHIP_BONAIRE && ia_switch_on_eoi &&
623		    (instanced_draw || indirect_draw))
624			partial_vs_wave = true;
625
626		/* Hardware requirement when drawing primitives from a stream
627		 * output buffer.
628		 */
629		if (count_from_stream_output)
630			wd_switch_on_eop = true;
631
632		/* If the WD switch is false, the IA switch must be false too. */
633		assert(wd_switch_on_eop || !ia_switch_on_eop);
634	}
635	/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
636	if (chip_class <= VI && ia_switch_on_eoi)
637		partial_es_wave = true;
638
639	if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
640		/* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
641		 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
642		 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
643		 */
644		if (family == CHIP_HAWAII && ia_switch_on_eoi) {
645			bool set_vgt_flush = indirect_draw;
646			if (!set_vgt_flush && instanced_draw) {
647				uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
648				if (num_prims <= 1)
649					set_vgt_flush = true;
650			}
651			if (set_vgt_flush)
652				cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VGT_FLUSH;
653		}
654	}
655
656	return cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.base |
657		S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
658		S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
659		S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
660		S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
661		S_028AA8_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0);
662
663}
664
665void si_cs_emit_write_event_eop(struct radeon_cmdbuf *cs,
666				enum chip_class chip_class,
667				bool is_mec,
668				unsigned event, unsigned event_flags,
669				unsigned data_sel,
670				uint64_t va,
671				uint32_t new_fence,
672				uint64_t gfx9_eop_bug_va)
673{
674	unsigned op = EVENT_TYPE(event) |
675		EVENT_INDEX(5) |
676		event_flags;
677	unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
678	unsigned sel = EOP_DATA_SEL(data_sel);
679
680	/* Wait for write confirmation before writing data, but don't send
681	 * an interrupt. */
682	if (data_sel != EOP_DATA_SEL_DISCARD)
683		sel |= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM);
684
685	if (chip_class >= GFX9 || is_gfx8_mec) {
686		/* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
687		 * counters) must immediately precede every timestamp event to
688		 * prevent a GPU hang on GFX9.
689		 */
690		if (chip_class == GFX9 && !is_mec) {
691			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
692			radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
693			radeon_emit(cs, gfx9_eop_bug_va);
694			radeon_emit(cs, gfx9_eop_bug_va >> 32);
695		}
696
697		radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, false));
698		radeon_emit(cs, op);
699		radeon_emit(cs, sel);
700		radeon_emit(cs, va);            /* address lo */
701		radeon_emit(cs, va >> 32);      /* address hi */
702		radeon_emit(cs, new_fence);     /* immediate data lo */
703		radeon_emit(cs, 0); /* immediate data hi */
704		if (!is_gfx8_mec)
705			radeon_emit(cs, 0); /* unused */
706	} else {
707		if (chip_class == CIK ||
708		    chip_class == VI) {
709			/* Two EOP events are required to make all engines go idle
710			 * (and optional cache flushes executed) before the timestamp
711			 * is written.
712			 */
713			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
714			radeon_emit(cs, op);
715			radeon_emit(cs, va);
716			radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
717			radeon_emit(cs, 0); /* immediate data */
718			radeon_emit(cs, 0); /* unused */
719		}
720
721		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
722		radeon_emit(cs, op);
723		radeon_emit(cs, va);
724		radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
725		radeon_emit(cs, new_fence); /* immediate data */
726		radeon_emit(cs, 0); /* unused */
727	}
728}
729
730void
731radv_cp_wait_mem(struct radeon_cmdbuf *cs, uint32_t op, uint64_t va,
732		 uint32_t ref, uint32_t mask)
733{
734	assert(op == WAIT_REG_MEM_EQUAL ||
735	       op == WAIT_REG_MEM_NOT_EQUAL ||
736	       op == WAIT_REG_MEM_GREATER_OR_EQUAL);
737
738	radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, false));
739	radeon_emit(cs, op | WAIT_REG_MEM_MEM_SPACE(1));
740	radeon_emit(cs, va);
741	radeon_emit(cs, va >> 32);
742	radeon_emit(cs, ref); /* reference value */
743	radeon_emit(cs, mask); /* mask */
744	radeon_emit(cs, 4); /* poll interval */
745}
746
747static void
748si_emit_acquire_mem(struct radeon_cmdbuf *cs,
749                    bool is_mec,
750		    bool is_gfx9,
751                    unsigned cp_coher_cntl)
752{
753	if (is_mec || is_gfx9) {
754		uint32_t hi_val = is_gfx9 ? 0xffffff : 0xff;
755		radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, false) |
756		                            PKT3_SHADER_TYPE_S(is_mec));
757		radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
758		radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
759		radeon_emit(cs, hi_val);          /* CP_COHER_SIZE_HI */
760		radeon_emit(cs, 0);               /* CP_COHER_BASE */
761		radeon_emit(cs, 0);               /* CP_COHER_BASE_HI */
762		radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
763	} else {
764		/* ACQUIRE_MEM is only required on a compute ring. */
765		radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, false));
766		radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
767		radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
768		radeon_emit(cs, 0);               /* CP_COHER_BASE */
769		radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
770	}
771}
772
773void
774si_cs_emit_cache_flush(struct radeon_cmdbuf *cs,
775                       enum chip_class chip_class,
776		       uint32_t *flush_cnt,
777		       uint64_t flush_va,
778                       bool is_mec,
779                       enum radv_cmd_flush_bits flush_bits,
780		       uint64_t gfx9_eop_bug_va)
781{
782	unsigned cp_coher_cntl = 0;
783	uint32_t flush_cb_db = flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
784					     RADV_CMD_FLAG_FLUSH_AND_INV_DB);
785
786	if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
787		cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
788	if (flush_bits & RADV_CMD_FLAG_INV_SMEM_L1)
789		cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
790
791	if (chip_class <= VI) {
792		if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
793			cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
794				S_0085F0_CB0_DEST_BASE_ENA(1) |
795				S_0085F0_CB1_DEST_BASE_ENA(1) |
796				S_0085F0_CB2_DEST_BASE_ENA(1) |
797				S_0085F0_CB3_DEST_BASE_ENA(1) |
798				S_0085F0_CB4_DEST_BASE_ENA(1) |
799				S_0085F0_CB5_DEST_BASE_ENA(1) |
800				S_0085F0_CB6_DEST_BASE_ENA(1) |
801				S_0085F0_CB7_DEST_BASE_ENA(1);
802
803			/* Necessary for DCC */
804			if (chip_class >= VI) {
805				si_cs_emit_write_event_eop(cs,
806							   chip_class,
807							   is_mec,
808							   V_028A90_FLUSH_AND_INV_CB_DATA_TS,
809							   0,
810							   EOP_DATA_SEL_DISCARD,
811							   0, 0,
812							   gfx9_eop_bug_va);
813			}
814		}
815		if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
816			cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
817				S_0085F0_DB_DEST_BASE_ENA(1);
818		}
819	}
820
821	if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) {
822		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
823		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
824	}
825
826	if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) {
827		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
828		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
829	}
830
831	if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
832		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
833		radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
834	} else if (flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
835		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
836		radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
837	}
838
839	if (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
840		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
841		radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
842	}
843
844	if (chip_class >= GFX9 && flush_cb_db) {
845		unsigned cb_db_event, tc_flags;
846
847		/* Set the CB/DB flush event. */
848		cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
849
850		/* These are the only allowed combinations. If you need to
851		 * do multiple operations at once, do them separately.
852		 * All operations that invalidate L2 also seem to invalidate
853		 * metadata. Volatile (VOL) and WC flushes are not listed here.
854		 *
855		 * TC    | TC_WB         = writeback & invalidate L2 & L1
856		 * TC    | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
857		 *         TC_WB | TC_NC = writeback L2 for MTYPE == NC
858		 * TC            | TC_NC = invalidate L2 for MTYPE == NC
859		 * TC    | TC_MD         = writeback & invalidate L2 metadata (DCC, etc.)
860		 * TCL1                  = invalidate L1
861		 */
862		tc_flags = EVENT_TC_ACTION_ENA |
863		           EVENT_TC_MD_ACTION_ENA;
864
865		/* Ideally flush TC together with CB/DB. */
866		if (flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) {
867			/* Writeback and invalidate everything in L2 & L1. */
868			tc_flags = EVENT_TC_ACTION_ENA |
869			           EVENT_TC_WB_ACTION_ENA;
870
871
872			/* Clear the flags. */
873		        flush_bits &= ~(RADV_CMD_FLAG_INV_GLOBAL_L2 |
874					 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 |
875					 RADV_CMD_FLAG_INV_VMEM_L1);
876		}
877		assert(flush_cnt);
878		(*flush_cnt)++;
879
880		si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event, tc_flags,
881					   EOP_DATA_SEL_VALUE_32BIT,
882					   flush_va, *flush_cnt,
883					   gfx9_eop_bug_va);
884		radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL, flush_va,
885				 *flush_cnt, 0xffffffff);
886	}
887
888	/* VGT state sync */
889	if (flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
890		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
891		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
892	}
893
894	/* VGT streamout state sync */
895	if (flush_bits & RADV_CMD_FLAG_VGT_STREAMOUT_SYNC) {
896		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
897		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
898	}
899
900	/* Make sure ME is idle (it executes most packets) before continuing.
901	 * This prevents read-after-write hazards between PFP and ME.
902	 */
903	if ((cp_coher_cntl ||
904	     (flush_bits & (RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
905			    RADV_CMD_FLAG_INV_VMEM_L1 |
906			    RADV_CMD_FLAG_INV_GLOBAL_L2 |
907			    RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) &&
908	    !is_mec) {
909		radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
910		radeon_emit(cs, 0);
911	}
912
913	if ((flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) ||
914	    (chip_class <= CIK && (flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) {
915		si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9,
916				    cp_coher_cntl |
917				    S_0085F0_TC_ACTION_ENA(1) |
918				    S_0085F0_TCL1_ACTION_ENA(1) |
919				    S_0301F0_TC_WB_ACTION_ENA(chip_class >= VI));
920		cp_coher_cntl = 0;
921	} else {
922		if(flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2) {
923			/* WB = write-back
924			 * NC = apply to non-coherent MTYPEs
925			 *      (i.e. MTYPE <= 1, which is what we use everywhere)
926			 *
927			 * WB doesn't work without NC.
928			 */
929			si_emit_acquire_mem(cs, is_mec,
930					    chip_class >= GFX9,
931					    cp_coher_cntl |
932					    S_0301F0_TC_WB_ACTION_ENA(1) |
933					    S_0301F0_TC_NC_ACTION_ENA(1));
934			cp_coher_cntl = 0;
935		}
936		if (flush_bits & RADV_CMD_FLAG_INV_VMEM_L1) {
937			si_emit_acquire_mem(cs, is_mec,
938					    chip_class >= GFX9,
939					    cp_coher_cntl |
940					    S_0085F0_TCL1_ACTION_ENA(1));
941			cp_coher_cntl = 0;
942		}
943	}
944
945	/* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
946	 * Therefore, it should be last. Done in PFP.
947	 */
948	if (cp_coher_cntl)
949		si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9, cp_coher_cntl);
950
951	if (flush_bits & RADV_CMD_FLAG_START_PIPELINE_STATS) {
952		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
953		radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
954			        EVENT_INDEX(0));
955	} else if (flush_bits & RADV_CMD_FLAG_STOP_PIPELINE_STATS) {
956		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
957		radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
958			        EVENT_INDEX(0));
959	}
960}
961
962void
963si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
964{
965	bool is_compute = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
966
967	if (is_compute)
968		cmd_buffer->state.flush_bits &= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB |
969	                                          RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
970	                                          RADV_CMD_FLAG_FLUSH_AND_INV_DB |
971	                                          RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
972	                                          RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
973	                                          RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
974	                                          RADV_CMD_FLAG_VGT_FLUSH |
975						  RADV_CMD_FLAG_START_PIPELINE_STATS |
976						  RADV_CMD_FLAG_STOP_PIPELINE_STATS);
977
978	if (!cmd_buffer->state.flush_bits)
979		return;
980
981	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 128);
982
983	si_cs_emit_cache_flush(cmd_buffer->cs,
984	                       cmd_buffer->device->physical_device->rad_info.chip_class,
985			       &cmd_buffer->gfx9_fence_idx,
986			       cmd_buffer->gfx9_fence_va,
987	                       radv_cmd_buffer_uses_mec(cmd_buffer),
988	                       cmd_buffer->state.flush_bits,
989			       cmd_buffer->gfx9_eop_bug_va);
990
991
992	if (unlikely(cmd_buffer->device->trace_bo))
993		radv_cmd_buffer_trace_emit(cmd_buffer);
994
995	cmd_buffer->state.flush_bits = 0;
996
997	/* If the driver used a compute shader for resetting a query pool, it
998	 * should be finished at this point.
999	 */
1000	cmd_buffer->pending_reset_query = false;
1001}
1002
1003/* sets the CP predication state using a boolean stored at va */
1004void
1005si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer,
1006			      bool draw_visible, uint64_t va)
1007{
1008	uint32_t op = 0;
1009
1010	if (va) {
1011		op = PRED_OP(PREDICATION_OP_BOOL64);
1012
1013		/* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1014		 * zero, all rendering commands are discarded. Otherwise, they
1015		 * are discarded if the value is non zero.
1016		 */
1017		op |= draw_visible ? PREDICATION_DRAW_VISIBLE :
1018				     PREDICATION_DRAW_NOT_VISIBLE;
1019	}
1020	if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1021		radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
1022		radeon_emit(cmd_buffer->cs, op);
1023		radeon_emit(cmd_buffer->cs, va);
1024		radeon_emit(cmd_buffer->cs, va >> 32);
1025	} else {
1026		radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
1027		radeon_emit(cmd_buffer->cs, va);
1028		radeon_emit(cmd_buffer->cs, op | ((va >> 32) & 0xFF));
1029	}
1030}
1031
1032/* Set this if you want the 3D engine to wait until CP DMA is done.
1033 * It should be set on the last CP DMA packet. */
1034#define CP_DMA_SYNC	(1 << 0)
1035
1036/* Set this if the source data was used as a destination in a previous CP DMA
1037 * packet. It's for preventing a read-after-write (RAW) hazard between two
1038 * CP DMA packets. */
1039#define CP_DMA_RAW_WAIT	(1 << 1)
1040#define CP_DMA_USE_L2	(1 << 2)
1041#define CP_DMA_CLEAR	(1 << 3)
1042
1043/* Alignment for optimal performance. */
1044#define SI_CPDMA_ALIGNMENT	32
1045
1046/* The max number of bytes that can be copied per packet. */
1047static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer *cmd_buffer)
1048{
1049	unsigned max = cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 ?
1050			       S_414_BYTE_COUNT_GFX9(~0u) :
1051			       S_414_BYTE_COUNT_GFX6(~0u);
1052
1053	/* make it aligned for optimal performance */
1054	return max & ~(SI_CPDMA_ALIGNMENT - 1);
1055}
1056
1057/* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1058 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1059 * clear value.
1060 */
1061static void si_emit_cp_dma(struct radv_cmd_buffer *cmd_buffer,
1062			   uint64_t dst_va, uint64_t src_va,
1063			   unsigned size, unsigned flags)
1064{
1065	struct radeon_cmdbuf *cs = cmd_buffer->cs;
1066	uint32_t header = 0, command = 0;
1067
1068	assert(size <= cp_dma_max_byte_count(cmd_buffer));
1069
1070	radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
1071	if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1072		command |= S_414_BYTE_COUNT_GFX9(size);
1073	else
1074		command |= S_414_BYTE_COUNT_GFX6(size);
1075
1076	/* Sync flags. */
1077	if (flags & CP_DMA_SYNC)
1078		header |= S_411_CP_SYNC(1);
1079	else {
1080		if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1081			command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1082		else
1083			command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1084	}
1085
1086	if (flags & CP_DMA_RAW_WAIT)
1087		command |= S_414_RAW_WAIT(1);
1088
1089	/* Src and dst flags. */
1090	if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 &&
1091	    !(flags & CP_DMA_CLEAR) &&
1092	    src_va == dst_va)
1093		header |= S_411_DST_SEL(V_411_NOWHERE); /* prefetch only */
1094	else if (flags & CP_DMA_USE_L2)
1095		header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
1096
1097	if (flags & CP_DMA_CLEAR)
1098		header |= S_411_SRC_SEL(V_411_DATA);
1099	else if (flags & CP_DMA_USE_L2)
1100		header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
1101
1102	if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1103		radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, cmd_buffer->state.predicating));
1104		radeon_emit(cs, header);
1105		radeon_emit(cs, src_va);		/* SRC_ADDR_LO [31:0] */
1106		radeon_emit(cs, src_va >> 32);		/* SRC_ADDR_HI [31:0] */
1107		radeon_emit(cs, dst_va);		/* DST_ADDR_LO [31:0] */
1108		radeon_emit(cs, dst_va >> 32);		/* DST_ADDR_HI [31:0] */
1109		radeon_emit(cs, command);
1110	} else {
1111		assert(!(flags & CP_DMA_USE_L2));
1112		header |= S_411_SRC_ADDR_HI(src_va >> 32);
1113		radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, cmd_buffer->state.predicating));
1114		radeon_emit(cs, src_va);			/* SRC_ADDR_LO [31:0] */
1115		radeon_emit(cs, header);			/* SRC_ADDR_HI [15:0] + flags. */
1116		radeon_emit(cs, dst_va);			/* DST_ADDR_LO [31:0] */
1117		radeon_emit(cs, (dst_va >> 32) & 0xffff);	/* DST_ADDR_HI [15:0] */
1118		radeon_emit(cs, command);
1119	}
1120
1121	/* CP DMA is executed in ME, but index buffers are read by PFP.
1122	 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1123	 * indices. If we wanted to execute CP DMA in PFP, this packet
1124	 * should precede it.
1125	 */
1126	if (flags & CP_DMA_SYNC) {
1127		if (cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL) {
1128			radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
1129			radeon_emit(cs, 0);
1130		}
1131
1132		/* CP will see the sync flag and wait for all DMAs to complete. */
1133		cmd_buffer->state.dma_is_busy = false;
1134	}
1135
1136	if (unlikely(cmd_buffer->device->trace_bo))
1137		radv_cmd_buffer_trace_emit(cmd_buffer);
1138}
1139
1140void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1141                        unsigned size)
1142{
1143	uint64_t aligned_va = va & ~(SI_CPDMA_ALIGNMENT - 1);
1144	uint64_t aligned_size = ((va + size + SI_CPDMA_ALIGNMENT -1) & ~(SI_CPDMA_ALIGNMENT - 1)) - aligned_va;
1145
1146	si_emit_cp_dma(cmd_buffer, aligned_va, aligned_va,
1147		       aligned_size, CP_DMA_USE_L2);
1148}
1149
1150static void si_cp_dma_prepare(struct radv_cmd_buffer *cmd_buffer, uint64_t byte_count,
1151			      uint64_t remaining_size, unsigned *flags)
1152{
1153
1154	/* Flush the caches for the first copy only.
1155	 * Also wait for the previous CP DMA operations.
1156	 */
1157	if (cmd_buffer->state.flush_bits) {
1158		si_emit_cache_flush(cmd_buffer);
1159		*flags |= CP_DMA_RAW_WAIT;
1160	}
1161
1162	/* Do the synchronization after the last dma, so that all data
1163	 * is written to memory.
1164	 */
1165	if (byte_count == remaining_size)
1166		*flags |= CP_DMA_SYNC;
1167}
1168
1169static void si_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigned size)
1170{
1171	uint64_t va;
1172	uint32_t offset;
1173	unsigned dma_flags = 0;
1174	unsigned buf_size = SI_CPDMA_ALIGNMENT * 2;
1175	void *ptr;
1176
1177	assert(size < SI_CPDMA_ALIGNMENT);
1178
1179	radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, SI_CPDMA_ALIGNMENT,  &offset, &ptr);
1180
1181	va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1182	va += offset;
1183
1184	si_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
1185
1186	si_emit_cp_dma(cmd_buffer, va, va + SI_CPDMA_ALIGNMENT, size,
1187		       dma_flags);
1188}
1189
1190void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
1191			   uint64_t src_va, uint64_t dest_va,
1192			   uint64_t size)
1193{
1194	uint64_t main_src_va, main_dest_va;
1195	uint64_t skipped_size = 0, realign_size = 0;
1196
1197	/* Assume that we are not going to sync after the last DMA operation. */
1198	cmd_buffer->state.dma_is_busy = true;
1199
1200	if (cmd_buffer->device->physical_device->rad_info.family <= CHIP_CARRIZO ||
1201	    cmd_buffer->device->physical_device->rad_info.family == CHIP_STONEY) {
1202		/* If the size is not aligned, we must add a dummy copy at the end
1203		 * just to align the internal counter. Otherwise, the DMA engine
1204		 * would slow down by an order of magnitude for following copies.
1205		 */
1206		if (size % SI_CPDMA_ALIGNMENT)
1207			realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
1208
1209		/* If the copy begins unaligned, we must start copying from the next
1210		 * aligned block and the skipped part should be copied after everything
1211		 * else has been copied. Only the src alignment matters, not dst.
1212		 */
1213		if (src_va % SI_CPDMA_ALIGNMENT) {
1214			skipped_size = SI_CPDMA_ALIGNMENT - (src_va % SI_CPDMA_ALIGNMENT);
1215			/* The main part will be skipped if the size is too small. */
1216			skipped_size = MIN2(skipped_size, size);
1217			size -= skipped_size;
1218		}
1219	}
1220	main_src_va = src_va + skipped_size;
1221	main_dest_va = dest_va + skipped_size;
1222
1223	while (size) {
1224		unsigned dma_flags = 0;
1225		unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1226
1227		si_cp_dma_prepare(cmd_buffer, byte_count,
1228				  size + skipped_size + realign_size,
1229				  &dma_flags);
1230
1231		dma_flags &= ~CP_DMA_SYNC;
1232
1233		si_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va,
1234			       byte_count, dma_flags);
1235
1236		size -= byte_count;
1237		main_src_va += byte_count;
1238		main_dest_va += byte_count;
1239	}
1240
1241	if (skipped_size) {
1242		unsigned dma_flags = 0;
1243
1244		si_cp_dma_prepare(cmd_buffer, skipped_size,
1245				  size + skipped_size + realign_size,
1246				  &dma_flags);
1247
1248		si_emit_cp_dma(cmd_buffer, dest_va, src_va,
1249			       skipped_size, dma_flags);
1250	}
1251	if (realign_size)
1252		si_cp_dma_realign_engine(cmd_buffer, realign_size);
1253}
1254
1255void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1256			    uint64_t size, unsigned value)
1257{
1258
1259	if (!size)
1260		return;
1261
1262	assert(va % 4 == 0 && size % 4 == 0);
1263
1264	/* Assume that we are not going to sync after the last DMA operation. */
1265	cmd_buffer->state.dma_is_busy = true;
1266
1267	while (size) {
1268		unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1269		unsigned dma_flags = CP_DMA_CLEAR;
1270
1271		si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
1272
1273		/* Emit the clear packet. */
1274		si_emit_cp_dma(cmd_buffer, va, value, byte_count,
1275			       dma_flags);
1276
1277		size -= byte_count;
1278		va += byte_count;
1279	}
1280}
1281
1282void si_cp_dma_wait_for_idle(struct radv_cmd_buffer *cmd_buffer)
1283{
1284	if (cmd_buffer->device->physical_device->rad_info.chip_class < CIK)
1285		return;
1286
1287	if (!cmd_buffer->state.dma_is_busy)
1288		return;
1289
1290	/* Issue a dummy DMA that copies zero bytes.
1291	 *
1292	 * The DMA engine will see that there's no work to do and skip this
1293	 * DMA request, however, the CP will see the sync flag and still wait
1294	 * for all DMAs to complete.
1295	 */
1296	si_emit_cp_dma(cmd_buffer, 0, 0, 0, CP_DMA_SYNC);
1297
1298	cmd_buffer->state.dma_is_busy = false;
1299}
1300
1301/* For MSAA sample positions. */
1302#define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y)  \
1303	(((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) |		   \
1304	(((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) |	   \
1305	(((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) |	   \
1306	 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1307
1308
1309/* 2xMSAA
1310 * There are two locations (4, 4), (-4, -4). */
1311const uint32_t eg_sample_locs_2x[4] = {
1312	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1313	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1314	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1315	FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1316};
1317const unsigned eg_max_dist_2x = 4;
1318/* 4xMSAA
1319 * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
1320const uint32_t eg_sample_locs_4x[4] = {
1321	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1322	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1323	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1324	FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1325};
1326const unsigned eg_max_dist_4x = 6;
1327
1328/* Cayman 8xMSAA */
1329static const uint32_t cm_sample_locs_8x[] = {
1330	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
1331	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
1332	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
1333	FILL_SREG( 1, -3, -1,  3, 5,  1, -3, -5),
1334	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
1335	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
1336	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
1337	FILL_SREG(-5,  5, -7, -1, 3,  7,  7, -7),
1338};
1339static const unsigned cm_max_dist_8x = 8;
1340/* Cayman 16xMSAA */
1341static const uint32_t cm_sample_locs_16x[] = {
1342	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
1343	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
1344	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
1345	FILL_SREG( 1,  1, -1, -3, -3,  2,  4, -1),
1346	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
1347	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
1348	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
1349	FILL_SREG(-5, -2,  2,  5,  5,  3,  3, -5),
1350	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
1351	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
1352	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
1353	FILL_SREG(-2,  6,  0, -7, -4, -6, -6,  4),
1354	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
1355	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
1356	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
1357	FILL_SREG(-8,  0,  7, -4,  6,  7, -7, -8),
1358};
1359static const unsigned cm_max_dist_16x = 8;
1360
1361unsigned radv_cayman_get_maxdist(int log_samples)
1362{
1363	unsigned max_dist[] = {
1364		0,
1365		eg_max_dist_2x,
1366		eg_max_dist_4x,
1367		cm_max_dist_8x,
1368		cm_max_dist_16x
1369	};
1370	return max_dist[log_samples];
1371}
1372
1373void radv_cayman_emit_msaa_sample_locs(struct radeon_cmdbuf *cs, int nr_samples)
1374{
1375	switch (nr_samples) {
1376	default:
1377	case 1:
1378		radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 0);
1379		radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, 0);
1380		radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, 0);
1381		radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, 0);
1382		break;
1383	case 2:
1384		radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_2x[0]);
1385		radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_2x[1]);
1386		radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_2x[2]);
1387		radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_2x[3]);
1388		break;
1389	case 4:
1390		radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_4x[0]);
1391		radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_4x[1]);
1392		radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_4x[2]);
1393		radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_4x[3]);
1394		break;
1395	case 8:
1396		radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
1397		radeon_emit(cs, cm_sample_locs_8x[0]);
1398		radeon_emit(cs, cm_sample_locs_8x[4]);
1399		radeon_emit(cs, 0);
1400		radeon_emit(cs, 0);
1401		radeon_emit(cs, cm_sample_locs_8x[1]);
1402		radeon_emit(cs, cm_sample_locs_8x[5]);
1403		radeon_emit(cs, 0);
1404		radeon_emit(cs, 0);
1405		radeon_emit(cs, cm_sample_locs_8x[2]);
1406		radeon_emit(cs, cm_sample_locs_8x[6]);
1407		radeon_emit(cs, 0);
1408		radeon_emit(cs, 0);
1409		radeon_emit(cs, cm_sample_locs_8x[3]);
1410		radeon_emit(cs, cm_sample_locs_8x[7]);
1411		break;
1412	case 16:
1413		radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 16);
1414		radeon_emit(cs, cm_sample_locs_16x[0]);
1415		radeon_emit(cs, cm_sample_locs_16x[4]);
1416		radeon_emit(cs, cm_sample_locs_16x[8]);
1417		radeon_emit(cs, cm_sample_locs_16x[12]);
1418		radeon_emit(cs, cm_sample_locs_16x[1]);
1419		radeon_emit(cs, cm_sample_locs_16x[5]);
1420		radeon_emit(cs, cm_sample_locs_16x[9]);
1421		radeon_emit(cs, cm_sample_locs_16x[13]);
1422		radeon_emit(cs, cm_sample_locs_16x[2]);
1423		radeon_emit(cs, cm_sample_locs_16x[6]);
1424		radeon_emit(cs, cm_sample_locs_16x[10]);
1425		radeon_emit(cs, cm_sample_locs_16x[14]);
1426		radeon_emit(cs, cm_sample_locs_16x[3]);
1427		radeon_emit(cs, cm_sample_locs_16x[7]);
1428		radeon_emit(cs, cm_sample_locs_16x[11]);
1429		radeon_emit(cs, cm_sample_locs_16x[15]);
1430		break;
1431	}
1432}
1433
1434static void radv_cayman_get_sample_position(struct radv_device *device,
1435					    unsigned sample_count,
1436					    unsigned sample_index, float *out_value)
1437{
1438	int offset, index;
1439	struct {
1440		int idx:4;
1441	} val;
1442	switch (sample_count) {
1443	case 1:
1444	default:
1445		out_value[0] = out_value[1] = 0.5;
1446		break;
1447	case 2:
1448		offset = 4 * (sample_index * 2);
1449		val.idx = (eg_sample_locs_2x[0] >> offset) & 0xf;
1450		out_value[0] = (float)(val.idx + 8) / 16.0f;
1451		val.idx = (eg_sample_locs_2x[0] >> (offset + 4)) & 0xf;
1452		out_value[1] = (float)(val.idx + 8) / 16.0f;
1453		break;
1454	case 4:
1455		offset = 4 * (sample_index * 2);
1456		val.idx = (eg_sample_locs_4x[0] >> offset) & 0xf;
1457		out_value[0] = (float)(val.idx + 8) / 16.0f;
1458		val.idx = (eg_sample_locs_4x[0] >> (offset + 4)) & 0xf;
1459		out_value[1] = (float)(val.idx + 8) / 16.0f;
1460		break;
1461	case 8:
1462		offset = 4 * (sample_index % 4 * 2);
1463		index = (sample_index / 4) * 4;
1464		val.idx = (cm_sample_locs_8x[index] >> offset) & 0xf;
1465		out_value[0] = (float)(val.idx + 8) / 16.0f;
1466		val.idx = (cm_sample_locs_8x[index] >> (offset + 4)) & 0xf;
1467		out_value[1] = (float)(val.idx + 8) / 16.0f;
1468		break;
1469	case 16:
1470		offset = 4 * (sample_index % 4 * 2);
1471		index = (sample_index / 4) * 4;
1472		val.idx = (cm_sample_locs_16x[index] >> offset) & 0xf;
1473		out_value[0] = (float)(val.idx + 8) / 16.0f;
1474		val.idx = (cm_sample_locs_16x[index] >> (offset + 4)) & 0xf;
1475		out_value[1] = (float)(val.idx + 8) / 16.0f;
1476		break;
1477	}
1478}
1479
1480void radv_device_init_msaa(struct radv_device *device)
1481{
1482	int i;
1483	radv_cayman_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
1484
1485	for (i = 0; i < 2; i++)
1486		radv_cayman_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
1487	for (i = 0; i < 4; i++)
1488		radv_cayman_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
1489	for (i = 0; i < 8; i++)
1490		radv_cayman_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
1491	for (i = 0; i < 16; i++)
1492		radv_cayman_get_sample_position(device, 16, i, device->sample_locations_16x[i]);
1493}
1494