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
28#include "compiler/nir/nir_serialize.h"
29#include "tgsi/tgsi_parse.h"
30#include "util/hash_table.h"
31#include "util/crc32.h"
32#include "util/u_async_debug.h"
33#include "util/u_memory.h"
34#include "util/u_prim.h"
35
36#include "util/disk_cache.h"
37#include "util/mesa-sha1.h"
38#include "ac_exp_param.h"
39#include "ac_shader_util.h"
40
41/* SHADER_CACHE */
42
43/**
44 * Return the IR binary in a buffer. For TGSI the first 4 bytes contain its
45 * size as integer.
46 */
47void *si_get_ir_binary(struct si_shader_selector *sel)
48{
49	struct blob blob;
50	unsigned ir_size;
51	void *ir_binary;
52
53	if (sel->tokens) {
54		ir_binary = sel->tokens;
55		ir_size = tgsi_num_tokens(sel->tokens) *
56					  sizeof(struct tgsi_token);
57	} else {
58		assert(sel->nir);
59
60		blob_init(&blob);
61		nir_serialize(&blob, sel->nir);
62		ir_binary = blob.data;
63		ir_size = blob.size;
64	}
65
66	unsigned size = 4 + ir_size + sizeof(sel->so);
67	char *result = (char*)MALLOC(size);
68	if (!result)
69		return NULL;
70
71	*((uint32_t*)result) = size;
72	memcpy(result + 4, ir_binary, ir_size);
73	memcpy(result + 4 + ir_size, &sel->so, sizeof(sel->so));
74
75	if (sel->nir)
76		blob_finish(&blob);
77
78	return result;
79}
80
81/** Copy "data" to "ptr" and return the next dword following copied data. */
82static uint32_t *write_data(uint32_t *ptr, const void *data, unsigned size)
83{
84	/* data may be NULL if size == 0 */
85	if (size)
86		memcpy(ptr, data, size);
87	ptr += DIV_ROUND_UP(size, 4);
88	return ptr;
89}
90
91/** Read data from "ptr". Return the next dword following the data. */
92static uint32_t *read_data(uint32_t *ptr, void *data, unsigned size)
93{
94	memcpy(data, ptr, size);
95	ptr += DIV_ROUND_UP(size, 4);
96	return ptr;
97}
98
99/**
100 * Write the size as uint followed by the data. Return the next dword
101 * following the copied data.
102 */
103static uint32_t *write_chunk(uint32_t *ptr, const void *data, unsigned size)
104{
105	*ptr++ = size;
106	return write_data(ptr, data, size);
107}
108
109/**
110 * Read the size as uint followed by the data. Return both via parameters.
111 * Return the next dword following the data.
112 */
113static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
114{
115	*size = *ptr++;
116	assert(*data == NULL);
117	if (!*size)
118		return ptr;
119	*data = malloc(*size);
120	return read_data(ptr, *data, *size);
121}
122
123/**
124 * Return the shader binary in a buffer. The first 4 bytes contain its size
125 * as integer.
126 */
127static void *si_get_shader_binary(struct si_shader *shader)
128{
129	/* There is always a size of data followed by the data itself. */
130	unsigned relocs_size = shader->binary.reloc_count *
131			       sizeof(shader->binary.relocs[0]);
132	unsigned disasm_size = shader->binary.disasm_string ?
133			       strlen(shader->binary.disasm_string) + 1 : 0;
134	unsigned llvm_ir_size = shader->binary.llvm_ir_string ?
135				strlen(shader->binary.llvm_ir_string) + 1 : 0;
136	unsigned size =
137		4 + /* total size */
138		4 + /* CRC32 of the data below */
139		align(sizeof(shader->config), 4) +
140		align(sizeof(shader->info), 4) +
141		4 + align(shader->binary.code_size, 4) +
142		4 + align(shader->binary.rodata_size, 4) +
143		4 + align(relocs_size, 4) +
144		4 + align(disasm_size, 4) +
145		4 + align(llvm_ir_size, 4);
146	void *buffer = CALLOC(1, size);
147	uint32_t *ptr = (uint32_t*)buffer;
148
149	if (!buffer)
150		return NULL;
151
152	*ptr++ = size;
153	ptr++; /* CRC32 is calculated at the end. */
154
155	ptr = write_data(ptr, &shader->config, sizeof(shader->config));
156	ptr = write_data(ptr, &shader->info, sizeof(shader->info));
157	ptr = write_chunk(ptr, shader->binary.code, shader->binary.code_size);
158	ptr = write_chunk(ptr, shader->binary.rodata, shader->binary.rodata_size);
159	ptr = write_chunk(ptr, shader->binary.relocs, relocs_size);
160	ptr = write_chunk(ptr, shader->binary.disasm_string, disasm_size);
161	ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
162	assert((char *)ptr - (char *)buffer == size);
163
164	/* Compute CRC32. */
165	ptr = (uint32_t*)buffer;
166	ptr++;
167	*ptr = util_hash_crc32(ptr + 1, size - 8);
168
169	return buffer;
170}
171
172static bool si_load_shader_binary(struct si_shader *shader, void *binary)
173{
174	uint32_t *ptr = (uint32_t*)binary;
175	uint32_t size = *ptr++;
176	uint32_t crc32 = *ptr++;
177	unsigned chunk_size;
178
179	if (util_hash_crc32(ptr, size - 8) != crc32) {
180		fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
181		return false;
182	}
183
184	ptr = read_data(ptr, &shader->config, sizeof(shader->config));
185	ptr = read_data(ptr, &shader->info, sizeof(shader->info));
186	ptr = read_chunk(ptr, (void**)&shader->binary.code,
187			 &shader->binary.code_size);
188	ptr = read_chunk(ptr, (void**)&shader->binary.rodata,
189			 &shader->binary.rodata_size);
190	ptr = read_chunk(ptr, (void**)&shader->binary.relocs, &chunk_size);
191	shader->binary.reloc_count = chunk_size / sizeof(shader->binary.relocs[0]);
192	ptr = read_chunk(ptr, (void**)&shader->binary.disasm_string, &chunk_size);
193	ptr = read_chunk(ptr, (void**)&shader->binary.llvm_ir_string, &chunk_size);
194
195	return true;
196}
197
198/**
199 * Insert a shader into the cache. It's assumed the shader is not in the cache.
200 * Use si_shader_cache_load_shader before calling this.
201 *
202 * Returns false on failure, in which case the ir_binary should be freed.
203 */
204bool si_shader_cache_insert_shader(struct si_screen *sscreen, void *ir_binary,
205				   struct si_shader *shader,
206				   bool insert_into_disk_cache)
207{
208	void *hw_binary;
209	struct hash_entry *entry;
210	uint8_t key[CACHE_KEY_SIZE];
211
212	entry = _mesa_hash_table_search(sscreen->shader_cache, ir_binary);
213	if (entry)
214		return false; /* already added */
215
216	hw_binary = si_get_shader_binary(shader);
217	if (!hw_binary)
218		return false;
219
220	if (_mesa_hash_table_insert(sscreen->shader_cache, ir_binary,
221				    hw_binary) == NULL) {
222		FREE(hw_binary);
223		return false;
224	}
225
226	if (sscreen->disk_shader_cache && insert_into_disk_cache) {
227		disk_cache_compute_key(sscreen->disk_shader_cache, ir_binary,
228				       *((uint32_t *)ir_binary), key);
229		disk_cache_put(sscreen->disk_shader_cache, key, hw_binary,
230			       *((uint32_t *) hw_binary), NULL);
231	}
232
233	return true;
234}
235
236bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
237				 struct si_shader *shader)
238{
239	struct hash_entry *entry =
240		_mesa_hash_table_search(sscreen->shader_cache, ir_binary);
241	if (!entry) {
242		if (sscreen->disk_shader_cache) {
243			unsigned char sha1[CACHE_KEY_SIZE];
244			size_t tg_size = *((uint32_t *) ir_binary);
245
246			disk_cache_compute_key(sscreen->disk_shader_cache,
247					       ir_binary, tg_size, sha1);
248
249			size_t binary_size;
250			uint8_t *buffer =
251				disk_cache_get(sscreen->disk_shader_cache,
252					       sha1, &binary_size);
253			if (!buffer)
254				return false;
255
256			if (binary_size < sizeof(uint32_t) ||
257			    *((uint32_t*)buffer) != binary_size) {
258				 /* Something has gone wrong discard the item
259				  * from the cache and rebuild/link from
260				  * source.
261				  */
262				assert(!"Invalid radeonsi shader disk cache "
263				       "item!");
264
265				disk_cache_remove(sscreen->disk_shader_cache,
266						  sha1);
267				free(buffer);
268
269				return false;
270			}
271
272			if (!si_load_shader_binary(shader, buffer)) {
273				free(buffer);
274				return false;
275			}
276			free(buffer);
277
278			if (!si_shader_cache_insert_shader(sscreen, ir_binary,
279							   shader, false))
280				FREE(ir_binary);
281		} else {
282			return false;
283		}
284	} else {
285		if (si_load_shader_binary(shader, entry->data))
286			FREE(ir_binary);
287		else
288			return false;
289	}
290	p_atomic_inc(&sscreen->num_shader_cache_hits);
291	return true;
292}
293
294static uint32_t si_shader_cache_key_hash(const void *key)
295{
296	/* The first dword is the key size. */
297	return util_hash_crc32(key, *(uint32_t*)key);
298}
299
300static bool si_shader_cache_key_equals(const void *a, const void *b)
301{
302	uint32_t *keya = (uint32_t*)a;
303	uint32_t *keyb = (uint32_t*)b;
304
305	/* The first dword is the key size. */
306	if (*keya != *keyb)
307		return false;
308
309	return memcmp(keya, keyb, *keya) == 0;
310}
311
312static void si_destroy_shader_cache_entry(struct hash_entry *entry)
313{
314	FREE((void*)entry->key);
315	FREE(entry->data);
316}
317
318bool si_init_shader_cache(struct si_screen *sscreen)
319{
320	(void) mtx_init(&sscreen->shader_cache_mutex, mtx_plain);
321	sscreen->shader_cache =
322		_mesa_hash_table_create(NULL,
323					si_shader_cache_key_hash,
324					si_shader_cache_key_equals);
325
326	return sscreen->shader_cache != NULL;
327}
328
329void si_destroy_shader_cache(struct si_screen *sscreen)
330{
331	if (sscreen->shader_cache)
332		_mesa_hash_table_destroy(sscreen->shader_cache,
333					 si_destroy_shader_cache_entry);
334	mtx_destroy(&sscreen->shader_cache_mutex);
335}
336
337/* SHADER STATES */
338
339static void si_set_tesseval_regs(struct si_screen *sscreen,
340				 const struct si_shader_selector *tes,
341				 struct si_pm4_state *pm4)
342{
343	const struct tgsi_shader_info *info = &tes->info;
344	unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
345	unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
346	bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
347	bool tes_point_mode = info->properties[TGSI_PROPERTY_TES_POINT_MODE];
348	unsigned type, partitioning, topology, distribution_mode;
349
350	switch (tes_prim_mode) {
351	case PIPE_PRIM_LINES:
352		type = V_028B6C_TESS_ISOLINE;
353		break;
354	case PIPE_PRIM_TRIANGLES:
355		type = V_028B6C_TESS_TRIANGLE;
356		break;
357	case PIPE_PRIM_QUADS:
358		type = V_028B6C_TESS_QUAD;
359		break;
360	default:
361		assert(0);
362		return;
363	}
364
365	switch (tes_spacing) {
366	case PIPE_TESS_SPACING_FRACTIONAL_ODD:
367		partitioning = V_028B6C_PART_FRAC_ODD;
368		break;
369	case PIPE_TESS_SPACING_FRACTIONAL_EVEN:
370		partitioning = V_028B6C_PART_FRAC_EVEN;
371		break;
372	case PIPE_TESS_SPACING_EQUAL:
373		partitioning = V_028B6C_PART_INTEGER;
374		break;
375	default:
376		assert(0);
377		return;
378	}
379
380	if (tes_point_mode)
381		topology = V_028B6C_OUTPUT_POINT;
382	else if (tes_prim_mode == PIPE_PRIM_LINES)
383		topology = V_028B6C_OUTPUT_LINE;
384	else if (tes_vertex_order_cw)
385		/* for some reason, this must be the other way around */
386		topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
387	else
388		topology = V_028B6C_OUTPUT_TRIANGLE_CW;
389
390	if (sscreen->has_distributed_tess) {
391		if (sscreen->info.family == CHIP_FIJI ||
392		    sscreen->info.family >= CHIP_POLARIS10)
393			distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
394		else
395			distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
396	} else
397		distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
398
399	assert(pm4->shader);
400	pm4->shader->vgt_tf_param = S_028B6C_TYPE(type) |
401				    S_028B6C_PARTITIONING(partitioning) |
402				    S_028B6C_TOPOLOGY(topology) |
403				    S_028B6C_DISTRIBUTION_MODE(distribution_mode);
404}
405
406/* Polaris needs different VTX_REUSE_DEPTH settings depending on
407 * whether the "fractional odd" tessellation spacing is used.
408 *
409 * Possible VGT configurations and which state should set the register:
410 *
411 *   Reg set in | VGT shader configuration   | Value
412 * ------------------------------------------------------
413 *     VS as VS | VS                         | 30
414 *     VS as ES | ES -> GS -> VS             | 30
415 *    TES as VS | LS -> HS -> VS             | 14 or 30
416 *    TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
417 *
418 * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
419 */
420static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen,
421					 struct si_shader_selector *sel,
422					 struct si_shader *shader,
423					 struct si_pm4_state *pm4)
424{
425	unsigned type = sel->type;
426
427	if (sscreen->info.family < CHIP_POLARIS10)
428		return;
429
430	/* VS as VS, or VS as ES: */
431	if ((type == PIPE_SHADER_VERTEX &&
432	     (!shader ||
433	      (!shader->key.as_ls && !shader->is_gs_copy_shader))) ||
434	    /* TES as VS, or TES as ES: */
435	    type == PIPE_SHADER_TESS_EVAL) {
436		unsigned vtx_reuse_depth = 30;
437
438		if (type == PIPE_SHADER_TESS_EVAL &&
439		    sel->info.properties[TGSI_PROPERTY_TES_SPACING] ==
440		    PIPE_TESS_SPACING_FRACTIONAL_ODD)
441			vtx_reuse_depth = 14;
442
443		assert(pm4->shader);
444		pm4->shader->vgt_vertex_reuse_block_cntl = vtx_reuse_depth;
445	}
446}
447
448static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
449{
450	if (shader->pm4)
451		si_pm4_clear_state(shader->pm4);
452	else
453		shader->pm4 = CALLOC_STRUCT(si_pm4_state);
454
455	if (shader->pm4) {
456		shader->pm4->shader = shader;
457		return shader->pm4;
458	} else {
459		fprintf(stderr, "radeonsi: Failed to create pm4 state.\n");
460		return NULL;
461	}
462}
463
464static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs)
465{
466	/* Add the pointer to VBO descriptors. */
467	return num_always_on_user_sgprs + 1;
468}
469
470static void si_shader_ls(struct si_screen *sscreen, struct si_shader *shader)
471{
472	struct si_pm4_state *pm4;
473	unsigned vgpr_comp_cnt;
474	uint64_t va;
475
476	assert(sscreen->info.chip_class <= VI);
477
478	pm4 = si_get_shader_pm4_state(shader);
479	if (!pm4)
480		return;
481
482	va = shader->bo->gpu_address;
483	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
484
485	/* We need at least 2 components for LS.
486	 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
487	 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
488	 */
489	vgpr_comp_cnt = shader->info.uses_instanceid ? 2 : 1;
490
491	si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
492	si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, S_00B524_MEM_BASE(va >> 40));
493
494	shader->config.rsrc1 = S_00B528_VGPRS((shader->config.num_vgprs - 1) / 4) |
495			   S_00B528_SGPRS((shader->config.num_sgprs - 1) / 8) |
496		           S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt) |
497			   S_00B528_DX10_CLAMP(1) |
498			   S_00B528_FLOAT_MODE(shader->config.float_mode);
499	shader->config.rsrc2 = S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR)) |
500			   S_00B52C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
501}
502
503static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
504{
505	struct si_pm4_state *pm4;
506	uint64_t va;
507	unsigned ls_vgpr_comp_cnt = 0;
508
509	pm4 = si_get_shader_pm4_state(shader);
510	if (!pm4)
511		return;
512
513	va = shader->bo->gpu_address;
514	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
515
516	if (sscreen->info.chip_class >= GFX9) {
517		si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
518		si_pm4_set_reg(pm4, R_00B414_SPI_SHADER_PGM_HI_LS, S_00B414_MEM_BASE(va >> 40));
519
520		/* We need at least 2 components for LS.
521		 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
522		 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
523		 */
524		ls_vgpr_comp_cnt = shader->info.uses_instanceid ? 2 : 1;
525
526		unsigned num_user_sgprs =
527			si_get_num_vs_user_sgprs(GFX9_TCS_NUM_USER_SGPR);
528
529		shader->config.rsrc2 =
530			S_00B42C_USER_SGPR(num_user_sgprs) |
531			S_00B42C_USER_SGPR_MSB(num_user_sgprs >> 5) |
532			S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
533	} else {
534		si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
535		si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, S_00B424_MEM_BASE(va >> 40));
536
537		shader->config.rsrc2 =
538			S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR) |
539			S_00B42C_OC_LDS_EN(1) |
540			S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
541	}
542
543	si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
544		       S_00B428_VGPRS((shader->config.num_vgprs - 1) / 4) |
545		       S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) |
546		       S_00B428_DX10_CLAMP(1) |
547		       S_00B428_FLOAT_MODE(shader->config.float_mode) |
548		       S_00B428_LS_VGPR_COMP_CNT(ls_vgpr_comp_cnt));
549
550	if (sscreen->info.chip_class <= VI) {
551		si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
552			       shader->config.rsrc2);
553	}
554}
555
556static void si_emit_shader_es(struct si_context *sctx)
557{
558	struct si_shader *shader = sctx->queued.named.es->shader;
559	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
560
561	if (!shader)
562		return;
563
564	radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
565				   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
566				   shader->selector->esgs_itemsize / 4);
567
568	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
569		radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
570					   SI_TRACKED_VGT_TF_PARAM,
571					   shader->vgt_tf_param);
572
573	if (shader->vgt_vertex_reuse_block_cntl)
574		radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
575					   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
576					   shader->vgt_vertex_reuse_block_cntl);
577
578	if (initial_cdw != sctx->gfx_cs->current.cdw)
579		sctx->context_roll = true;
580}
581
582static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
583{
584	struct si_pm4_state *pm4;
585	unsigned num_user_sgprs;
586	unsigned vgpr_comp_cnt;
587	uint64_t va;
588	unsigned oc_lds_en;
589
590	assert(sscreen->info.chip_class <= VI);
591
592	pm4 = si_get_shader_pm4_state(shader);
593	if (!pm4)
594		return;
595
596	pm4->atom.emit = si_emit_shader_es;
597	va = shader->bo->gpu_address;
598	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
599
600	if (shader->selector->type == PIPE_SHADER_VERTEX) {
601		/* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
602		vgpr_comp_cnt = shader->info.uses_instanceid ? 1 : 0;
603		num_user_sgprs = si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR);
604	} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
605		vgpr_comp_cnt = shader->selector->info.uses_primid ? 3 : 2;
606		num_user_sgprs = SI_TES_NUM_USER_SGPR;
607	} else
608		unreachable("invalid shader selector type");
609
610	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
611
612	si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
613	si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
614	si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
615		       S_00B328_VGPRS((shader->config.num_vgprs - 1) / 4) |
616		       S_00B328_SGPRS((shader->config.num_sgprs - 1) / 8) |
617		       S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt) |
618		       S_00B328_DX10_CLAMP(1) |
619		       S_00B328_FLOAT_MODE(shader->config.float_mode));
620	si_pm4_set_reg(pm4, R_00B32C_SPI_SHADER_PGM_RSRC2_ES,
621		       S_00B32C_USER_SGPR(num_user_sgprs) |
622		       S_00B32C_OC_LDS_EN(oc_lds_en) |
623		       S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
624
625	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
626		si_set_tesseval_regs(sscreen, shader->selector, pm4);
627
628	polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
629}
630
631static unsigned si_conv_prim_to_gs_out(unsigned mode)
632{
633	static const int prim_conv[] = {
634		[PIPE_PRIM_POINTS]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
635		[PIPE_PRIM_LINES]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
636		[PIPE_PRIM_LINE_LOOP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
637		[PIPE_PRIM_LINE_STRIP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
638		[PIPE_PRIM_TRIANGLES]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
639		[PIPE_PRIM_TRIANGLE_STRIP]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
640		[PIPE_PRIM_TRIANGLE_FAN]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
641		[PIPE_PRIM_QUADS]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
642		[PIPE_PRIM_QUAD_STRIP]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
643		[PIPE_PRIM_POLYGON]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
644		[PIPE_PRIM_LINES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
645		[PIPE_PRIM_LINE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
646		[PIPE_PRIM_TRIANGLES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
647		[PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
648		[PIPE_PRIM_PATCHES]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
649	};
650	assert(mode < ARRAY_SIZE(prim_conv));
651
652	return prim_conv[mode];
653}
654
655struct gfx9_gs_info {
656	unsigned es_verts_per_subgroup;
657	unsigned gs_prims_per_subgroup;
658	unsigned gs_inst_prims_in_subgroup;
659	unsigned max_prims_per_subgroup;
660	unsigned lds_size;
661};
662
663static void gfx9_get_gs_info(struct si_shader_selector *es,
664				   struct si_shader_selector *gs,
665				   struct gfx9_gs_info *out)
666{
667	unsigned gs_num_invocations = MAX2(gs->gs_num_invocations, 1);
668	unsigned input_prim = gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
669	bool uses_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
670			      input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
671
672	/* All these are in dwords: */
673	/* We can't allow using the whole LDS, because GS waves compete with
674	 * other shader stages for LDS space. */
675	const unsigned max_lds_size = 8 * 1024;
676	const unsigned esgs_itemsize = es->esgs_itemsize / 4;
677	unsigned esgs_lds_size;
678
679	/* All these are per subgroup: */
680	const unsigned max_out_prims = 32 * 1024;
681	const unsigned max_es_verts = 255;
682	const unsigned ideal_gs_prims = 64;
683	unsigned max_gs_prims, gs_prims;
684	unsigned min_es_verts, es_verts, worst_case_es_verts;
685
686	if (uses_adjacency || gs_num_invocations > 1)
687		max_gs_prims = 127 / gs_num_invocations;
688	else
689		max_gs_prims = 255;
690
691	/* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
692	 * Make sure we don't go over the maximum value.
693	 */
694	if (gs->gs_max_out_vertices > 0) {
695		max_gs_prims = MIN2(max_gs_prims,
696				    max_out_prims /
697				    (gs->gs_max_out_vertices * gs_num_invocations));
698	}
699	assert(max_gs_prims > 0);
700
701	/* If the primitive has adjacency, halve the number of vertices
702	 * that will be reused in multiple primitives.
703	 */
704	min_es_verts = gs->gs_input_verts_per_prim / (uses_adjacency ? 2 : 1);
705
706	gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
707	worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
708
709	/* Compute ESGS LDS size based on the worst case number of ES vertices
710	 * needed to create the target number of GS prims per subgroup.
711	 */
712	esgs_lds_size = esgs_itemsize * worst_case_es_verts;
713
714	/* If total LDS usage is too big, refactor partitions based on ratio
715	 * of ESGS item sizes.
716	 */
717	if (esgs_lds_size > max_lds_size) {
718		/* Our target GS Prims Per Subgroup was too large. Calculate
719		 * the maximum number of GS Prims Per Subgroup that will fit
720		 * into LDS, capped by the maximum that the hardware can support.
721		 */
722		gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
723				max_gs_prims);
724		assert(gs_prims > 0);
725		worst_case_es_verts = MIN2(min_es_verts * gs_prims,
726					   max_es_verts);
727
728		esgs_lds_size = esgs_itemsize * worst_case_es_verts;
729		assert(esgs_lds_size <= max_lds_size);
730	}
731
732	/* Now calculate remaining ESGS information. */
733	if (esgs_lds_size)
734		es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
735	else
736		es_verts = max_es_verts;
737
738	/* Vertices for adjacency primitives are not always reused, so restore
739	 * it for ES_VERTS_PER_SUBGRP.
740	 */
741	min_es_verts = gs->gs_input_verts_per_prim;
742
743	/* For normal primitives, the VGT only checks if they are past the ES
744	 * verts per subgroup after allocating a full GS primitive and if they
745	 * are, kick off a new subgroup.  But if those additional ES verts are
746	 * unique (e.g. not reused) we need to make sure there is enough LDS
747	 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
748	 */
749	es_verts -= min_es_verts - 1;
750
751	out->es_verts_per_subgroup = es_verts;
752	out->gs_prims_per_subgroup = gs_prims;
753	out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
754	out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup *
755				      gs->gs_max_out_vertices;
756	out->lds_size = align(esgs_lds_size, 128) / 128;
757
758	assert(out->max_prims_per_subgroup <= max_out_prims);
759}
760
761static void si_emit_shader_gs(struct si_context *sctx)
762{
763	struct si_shader *shader = sctx->queued.named.gs->shader;
764	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
765
766	if (!shader)
767		return;
768
769	/* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
770	 * R_028A68_VGT_GSVS_RING_OFFSET_3, R_028A6C_VGT_GS_OUT_PRIM_TYPE */
771	radeon_opt_set_context_reg4(sctx, R_028A60_VGT_GSVS_RING_OFFSET_1,
772				    SI_TRACKED_VGT_GSVS_RING_OFFSET_1,
773				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_1,
774				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_2,
775				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_3,
776				    shader->ctx_reg.gs.vgt_gs_out_prim_type);
777
778
779	/* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
780	radeon_opt_set_context_reg(sctx, R_028AB0_VGT_GSVS_RING_ITEMSIZE,
781				   SI_TRACKED_VGT_GSVS_RING_ITEMSIZE,
782				   shader->ctx_reg.gs.vgt_gsvs_ring_itemsize);
783
784	/* R_028B38_VGT_GS_MAX_VERT_OUT */
785	radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
786				   SI_TRACKED_VGT_GS_MAX_VERT_OUT,
787				   shader->ctx_reg.gs.vgt_gs_max_vert_out);
788
789	/* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
790	 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
791	radeon_opt_set_context_reg4(sctx, R_028B5C_VGT_GS_VERT_ITEMSIZE,
792				    SI_TRACKED_VGT_GS_VERT_ITEMSIZE,
793				    shader->ctx_reg.gs.vgt_gs_vert_itemsize,
794				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_1,
795				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_2,
796				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_3);
797
798	/* R_028B90_VGT_GS_INSTANCE_CNT */
799	radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
800				   SI_TRACKED_VGT_GS_INSTANCE_CNT,
801				   shader->ctx_reg.gs.vgt_gs_instance_cnt);
802
803	if (sctx->chip_class >= GFX9) {
804		/* R_028A44_VGT_GS_ONCHIP_CNTL */
805		radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
806					   SI_TRACKED_VGT_GS_ONCHIP_CNTL,
807					   shader->ctx_reg.gs.vgt_gs_onchip_cntl);
808		/* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
809		radeon_opt_set_context_reg(sctx, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
810					   SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
811					   shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup);
812		/* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
813		radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
814					   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
815					   shader->ctx_reg.gs.vgt_esgs_ring_itemsize);
816
817		if (shader->key.part.gs.es->type == PIPE_SHADER_TESS_EVAL)
818			radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
819						   SI_TRACKED_VGT_TF_PARAM,
820						   shader->vgt_tf_param);
821		if (shader->vgt_vertex_reuse_block_cntl)
822			radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
823						   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
824						   shader->vgt_vertex_reuse_block_cntl);
825	}
826
827	if (initial_cdw != sctx->gfx_cs->current.cdw)
828		sctx->context_roll = true;
829}
830
831static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
832{
833	struct si_shader_selector *sel = shader->selector;
834	const ubyte *num_components = sel->info.num_stream_output_components;
835	unsigned gs_num_invocations = sel->gs_num_invocations;
836	struct si_pm4_state *pm4;
837	uint64_t va;
838	unsigned max_stream = sel->max_gs_stream;
839	unsigned offset;
840
841	pm4 = si_get_shader_pm4_state(shader);
842	if (!pm4)
843		return;
844
845	pm4->atom.emit = si_emit_shader_gs;
846
847	offset = num_components[0] * sel->gs_max_out_vertices;
848	shader->ctx_reg.gs.vgt_gsvs_ring_offset_1 = offset;
849
850	if (max_stream >= 1)
851		offset += num_components[1] * sel->gs_max_out_vertices;
852	shader->ctx_reg.gs.vgt_gsvs_ring_offset_2 = offset;
853
854	if (max_stream >= 2)
855		offset += num_components[2] * sel->gs_max_out_vertices;
856	shader->ctx_reg.gs.vgt_gsvs_ring_offset_3 = offset;
857
858	shader->ctx_reg.gs.vgt_gs_out_prim_type =
859		si_conv_prim_to_gs_out(sel->gs_output_prim);
860
861	if (max_stream >= 3)
862		offset += num_components[3] * sel->gs_max_out_vertices;
863	shader->ctx_reg.gs.vgt_gsvs_ring_itemsize = offset;
864
865	/* The GSVS_RING_ITEMSIZE register takes 15 bits */
866	assert(offset < (1 << 15));
867
868	shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->gs_max_out_vertices;
869
870	shader->ctx_reg.gs.vgt_gs_vert_itemsize = num_components[0];
871	shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 1) ? num_components[1] : 0;
872	shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 2) ? num_components[2] : 0;
873	shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 3) ? num_components[3] : 0;
874
875	shader->ctx_reg.gs.vgt_gs_instance_cnt = S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
876						 S_028B90_ENABLE(gs_num_invocations > 0);
877
878	va = shader->bo->gpu_address;
879	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
880
881	if (sscreen->info.chip_class >= GFX9) {
882		unsigned input_prim = sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
883		unsigned es_type = shader->key.part.gs.es->type;
884		unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
885		struct gfx9_gs_info gs_info;
886
887		if (es_type == PIPE_SHADER_VERTEX)
888			/* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
889			es_vgpr_comp_cnt = shader->info.uses_instanceid ? 1 : 0;
890		else if (es_type == PIPE_SHADER_TESS_EVAL)
891			es_vgpr_comp_cnt = shader->key.part.gs.es->info.uses_primid ? 3 : 2;
892		else
893			unreachable("invalid shader selector type");
894
895		/* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
896		 * VGPR[0:4] are always loaded.
897		 */
898		if (sel->info.uses_invocationid)
899			gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
900		else if (sel->info.uses_primid)
901			gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
902		else if (input_prim >= PIPE_PRIM_TRIANGLES)
903			gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
904		else
905			gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
906
907		unsigned num_user_sgprs;
908		if (es_type == PIPE_SHADER_VERTEX)
909			num_user_sgprs = si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR);
910		else
911			num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
912
913		gfx9_get_gs_info(shader->key.part.gs.es, sel, &gs_info);
914
915		si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
916		si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, S_00B214_MEM_BASE(va >> 40));
917
918		si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
919			       S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
920			       S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
921			       S_00B228_DX10_CLAMP(1) |
922			       S_00B228_FLOAT_MODE(shader->config.float_mode) |
923			       S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
924		si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
925			       S_00B22C_USER_SGPR(num_user_sgprs) |
926			       S_00B22C_USER_SGPR_MSB(num_user_sgprs >> 5) |
927			       S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
928			       S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
929			       S_00B22C_LDS_SIZE(gs_info.lds_size) |
930			       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
931
932		shader->ctx_reg.gs.vgt_gs_onchip_cntl =
933			S_028A44_ES_VERTS_PER_SUBGRP(gs_info.es_verts_per_subgroup) |
934			S_028A44_GS_PRIMS_PER_SUBGRP(gs_info.gs_prims_per_subgroup) |
935			S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_info.gs_inst_prims_in_subgroup);
936		shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup =
937			S_028A94_MAX_PRIMS_PER_SUBGROUP(gs_info.max_prims_per_subgroup);
938		shader->ctx_reg.gs.vgt_esgs_ring_itemsize =
939			shader->key.part.gs.es->esgs_itemsize / 4;
940
941		if (es_type == PIPE_SHADER_TESS_EVAL)
942			si_set_tesseval_regs(sscreen, shader->key.part.gs.es, pm4);
943
944		polaris_set_vgt_vertex_reuse(sscreen, shader->key.part.gs.es,
945					     NULL, pm4);
946	} else {
947		si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
948		si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, S_00B224_MEM_BASE(va >> 40));
949
950		si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
951			       S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
952			       S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
953			       S_00B228_DX10_CLAMP(1) |
954			       S_00B228_FLOAT_MODE(shader->config.float_mode));
955		si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
956			       S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
957			       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
958	}
959}
960
961static void si_emit_shader_vs(struct si_context *sctx)
962{
963	struct si_shader *shader = sctx->queued.named.vs->shader;
964	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
965
966	if (!shader)
967		return;
968
969	radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE,
970				   SI_TRACKED_VGT_GS_MODE,
971				   shader->ctx_reg.vs.vgt_gs_mode);
972	radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
973				   SI_TRACKED_VGT_PRIMITIVEID_EN,
974				   shader->ctx_reg.vs.vgt_primitiveid_en);
975
976	if (sctx->chip_class <= VI) {
977		radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
978					   SI_TRACKED_VGT_REUSE_OFF,
979					   shader->ctx_reg.vs.vgt_reuse_off);
980	}
981
982	radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
983				   SI_TRACKED_SPI_VS_OUT_CONFIG,
984				   shader->ctx_reg.vs.spi_vs_out_config);
985
986	radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
987				   SI_TRACKED_SPI_SHADER_POS_FORMAT,
988				   shader->ctx_reg.vs.spi_shader_pos_format);
989
990	radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
991				   SI_TRACKED_PA_CL_VTE_CNTL,
992				   shader->ctx_reg.vs.pa_cl_vte_cntl);
993
994	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
995		radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
996					   SI_TRACKED_VGT_TF_PARAM,
997					   shader->vgt_tf_param);
998
999	if (shader->vgt_vertex_reuse_block_cntl)
1000		radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
1001					   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
1002					   shader->vgt_vertex_reuse_block_cntl);
1003
1004	if (initial_cdw != sctx->gfx_cs->current.cdw)
1005		sctx->context_roll = true;
1006}
1007
1008/**
1009 * Compute the state for \p shader, which will run as a vertex shader on the
1010 * hardware.
1011 *
1012 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1013 * is the copy shader.
1014 */
1015static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
1016                         struct si_shader_selector *gs)
1017{
1018	const struct tgsi_shader_info *info = &shader->selector->info;
1019	struct si_pm4_state *pm4;
1020	unsigned num_user_sgprs, vgpr_comp_cnt;
1021	uint64_t va;
1022	unsigned nparams, oc_lds_en;
1023	unsigned window_space =
1024		info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
1025	bool enable_prim_id = shader->key.mono.u.vs_export_prim_id || info->uses_primid;
1026
1027	pm4 = si_get_shader_pm4_state(shader);
1028	if (!pm4)
1029		return;
1030
1031	pm4->atom.emit = si_emit_shader_vs;
1032
1033	/* We always write VGT_GS_MODE in the VS state, because every switch
1034	 * between different shader pipelines involving a different GS or no
1035	 * GS at all involves a switch of the VS (different GS use different
1036	 * copy shaders). On the other hand, when the API switches from a GS to
1037	 * no GS and then back to the same GS used originally, the GS state is
1038	 * not sent again.
1039	 */
1040	if (!gs) {
1041		unsigned mode = V_028A40_GS_OFF;
1042
1043		/* PrimID needs GS scenario A. */
1044		if (enable_prim_id)
1045			mode = V_028A40_GS_SCENARIO_A;
1046
1047		shader->ctx_reg.vs.vgt_gs_mode = S_028A40_MODE(mode);
1048		shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
1049	} else {
1050		shader->ctx_reg.vs.vgt_gs_mode = ac_vgt_gs_mode(gs->gs_max_out_vertices,
1051								sscreen->info.chip_class);
1052		shader->ctx_reg.vs.vgt_primitiveid_en = 0;
1053	}
1054
1055	if (sscreen->info.chip_class <= VI) {
1056		/* Reuse needs to be set off if we write oViewport. */
1057		shader->ctx_reg.vs.vgt_reuse_off =
1058				S_028AB4_REUSE_OFF(info->writes_viewport_index);
1059	}
1060
1061	va = shader->bo->gpu_address;
1062	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
1063
1064	if (gs) {
1065		vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
1066		num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
1067	} else if (shader->selector->type == PIPE_SHADER_VERTEX) {
1068		/* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
1069		 * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
1070		 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
1071		 */
1072		vgpr_comp_cnt = enable_prim_id ? 2 : (shader->info.uses_instanceid ? 1 : 0);
1073
1074		if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
1075			num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
1076					 info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
1077		} else {
1078			num_user_sgprs = si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR);
1079		}
1080	} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
1081		vgpr_comp_cnt = enable_prim_id ? 3 : 2;
1082		num_user_sgprs = SI_TES_NUM_USER_SGPR;
1083	} else
1084		unreachable("invalid shader selector type");
1085
1086	/* VS is required to export at least one param. */
1087	nparams = MAX2(shader->info.nr_param_exports, 1);
1088	shader->ctx_reg.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
1089
1090	shader->ctx_reg.vs.spi_shader_pos_format =
1091			S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
1092			S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
1093						    V_02870C_SPI_SHADER_4COMP :
1094						    V_02870C_SPI_SHADER_NONE) |
1095			S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
1096						    V_02870C_SPI_SHADER_4COMP :
1097						    V_02870C_SPI_SHADER_NONE) |
1098			S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
1099						    V_02870C_SPI_SHADER_4COMP :
1100						    V_02870C_SPI_SHADER_NONE);
1101
1102	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
1103
1104	si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
1105	si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, S_00B124_MEM_BASE(va >> 40));
1106	si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS,
1107		       S_00B128_VGPRS((shader->config.num_vgprs - 1) / 4) |
1108		       S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8) |
1109		       S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
1110		       S_00B128_DX10_CLAMP(1) |
1111		       S_00B128_FLOAT_MODE(shader->config.float_mode));
1112	si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS,
1113		       S_00B12C_USER_SGPR(num_user_sgprs) |
1114		       S_00B12C_OC_LDS_EN(oc_lds_en) |
1115		       S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
1116		       S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
1117		       S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
1118		       S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
1119		       S_00B12C_SO_EN(!!shader->selector->so.num_outputs) |
1120		       S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
1121
1122	if (window_space)
1123		shader->ctx_reg.vs.pa_cl_vte_cntl =
1124				S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1125	else
1126		shader->ctx_reg.vs.pa_cl_vte_cntl =
1127				S_028818_VTX_W0_FMT(1) |
1128				S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1129				S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1130				S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1131
1132	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
1133		si_set_tesseval_regs(sscreen, shader->selector, pm4);
1134
1135	polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
1136}
1137
1138static unsigned si_get_ps_num_interp(struct si_shader *ps)
1139{
1140	struct tgsi_shader_info *info = &ps->selector->info;
1141	unsigned num_colors = !!(info->colors_read & 0x0f) +
1142			      !!(info->colors_read & 0xf0);
1143	unsigned num_interp = ps->selector->info.num_inputs +
1144			      (ps->key.part.ps.prolog.color_two_side ? num_colors : 0);
1145
1146	assert(num_interp <= 32);
1147	return MIN2(num_interp, 32);
1148}
1149
1150static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
1151{
1152	unsigned value = shader->key.part.ps.epilog.spi_shader_col_format;
1153	unsigned i, num_targets = (util_last_bit(value) + 3) / 4;
1154
1155	/* If the i-th target format is set, all previous target formats must
1156	 * be non-zero to avoid hangs.
1157	 */
1158	for (i = 0; i < num_targets; i++)
1159		if (!(value & (0xf << (i * 4))))
1160			value |= V_028714_SPI_SHADER_32_R << (i * 4);
1161
1162	return value;
1163}
1164
1165static void si_emit_shader_ps(struct si_context *sctx)
1166{
1167	struct si_shader *shader = sctx->queued.named.ps->shader;
1168	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1169
1170	if (!shader)
1171		return;
1172
1173	/* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1174	radeon_opt_set_context_reg2(sctx, R_0286CC_SPI_PS_INPUT_ENA,
1175				    SI_TRACKED_SPI_PS_INPUT_ENA,
1176				    shader->ctx_reg.ps.spi_ps_input_ena,
1177				    shader->ctx_reg.ps.spi_ps_input_addr);
1178
1179	radeon_opt_set_context_reg(sctx, R_0286E0_SPI_BARYC_CNTL,
1180				   SI_TRACKED_SPI_BARYC_CNTL,
1181				   shader->ctx_reg.ps.spi_baryc_cntl);
1182	radeon_opt_set_context_reg(sctx, R_0286D8_SPI_PS_IN_CONTROL,
1183				   SI_TRACKED_SPI_PS_IN_CONTROL,
1184				   shader->ctx_reg.ps.spi_ps_in_control);
1185
1186	/* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1187	radeon_opt_set_context_reg2(sctx, R_028710_SPI_SHADER_Z_FORMAT,
1188				    SI_TRACKED_SPI_SHADER_Z_FORMAT,
1189				    shader->ctx_reg.ps.spi_shader_z_format,
1190				    shader->ctx_reg.ps.spi_shader_col_format);
1191
1192	radeon_opt_set_context_reg(sctx, R_02823C_CB_SHADER_MASK,
1193				   SI_TRACKED_CB_SHADER_MASK,
1194				   shader->ctx_reg.ps.cb_shader_mask);
1195
1196	if (initial_cdw != sctx->gfx_cs->current.cdw)
1197		sctx->context_roll = true;
1198}
1199
1200static void si_shader_ps(struct si_shader *shader)
1201{
1202	struct tgsi_shader_info *info = &shader->selector->info;
1203	struct si_pm4_state *pm4;
1204	unsigned spi_ps_in_control, spi_shader_col_format, cb_shader_mask;
1205	unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
1206	uint64_t va;
1207	unsigned input_ena = shader->config.spi_ps_input_ena;
1208
1209	/* we need to enable at least one of them, otherwise we hang the GPU */
1210	assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1211	       G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1212	       G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1213	       G_0286CC_PERSP_PULL_MODEL_ENA(input_ena) ||
1214	       G_0286CC_LINEAR_SAMPLE_ENA(input_ena) ||
1215	       G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1216	       G_0286CC_LINEAR_CENTROID_ENA(input_ena) ||
1217	       G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena));
1218	/* POS_W_FLOAT_ENA requires one of the perspective weights. */
1219	assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena) ||
1220	       G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1221	       G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1222	       G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1223	       G_0286CC_PERSP_PULL_MODEL_ENA(input_ena));
1224
1225	/* Validate interpolation optimization flags (read as implications). */
1226	assert(!shader->key.part.ps.prolog.bc_optimize_for_persp ||
1227	       (G_0286CC_PERSP_CENTER_ENA(input_ena) &&
1228		G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1229	assert(!shader->key.part.ps.prolog.bc_optimize_for_linear ||
1230	       (G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
1231		G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1232	assert(!shader->key.part.ps.prolog.force_persp_center_interp ||
1233	       (!G_0286CC_PERSP_SAMPLE_ENA(input_ena) &&
1234		!G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1235	assert(!shader->key.part.ps.prolog.force_linear_center_interp ||
1236	       (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena) &&
1237		!G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1238	assert(!shader->key.part.ps.prolog.force_persp_sample_interp ||
1239	       (!G_0286CC_PERSP_CENTER_ENA(input_ena) &&
1240		!G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1241	assert(!shader->key.part.ps.prolog.force_linear_sample_interp ||
1242	       (!G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
1243		!G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1244
1245	/* Validate cases when the optimizations are off (read as implications). */
1246	assert(shader->key.part.ps.prolog.bc_optimize_for_persp ||
1247	       !G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1248	       !G_0286CC_PERSP_CENTROID_ENA(input_ena));
1249	assert(shader->key.part.ps.prolog.bc_optimize_for_linear ||
1250	       !G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1251	       !G_0286CC_LINEAR_CENTROID_ENA(input_ena));
1252
1253	pm4 = si_get_shader_pm4_state(shader);
1254	if (!pm4)
1255		return;
1256
1257	pm4->atom.emit = si_emit_shader_ps;
1258
1259	/* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1260	 * Possible vaules:
1261	 * 0 -> Position = pixel center
1262	 * 1 -> Position = pixel centroid
1263	 * 2 -> Position = at sample position
1264	 *
1265	 * From GLSL 4.5 specification, section 7.1:
1266	 *   "The variable gl_FragCoord is available as an input variable from
1267	 *    within fragment shaders and it holds the window relative coordinates
1268	 *    (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1269	 *    value can be for any location within the pixel, or one of the
1270	 *    fragment samples. The use of centroid does not further restrict
1271	 *    this value to be inside the current primitive."
1272	 *
1273	 * Meaning that centroid has no effect and we can return anything within
1274	 * the pixel. Thus, return the value at sample position, because that's
1275	 * the most accurate one shaders can get.
1276	 */
1277	spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1278
1279	if (info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1280	    TGSI_FS_COORD_PIXEL_CENTER_INTEGER)
1281		spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
1282
1283	spi_shader_col_format = si_get_spi_shader_col_format(shader);
1284	cb_shader_mask = ac_get_cb_shader_mask(spi_shader_col_format);
1285
1286	/* Ensure that some export memory is always allocated, for two reasons:
1287	 *
1288	 * 1) Correctness: The hardware ignores the EXEC mask if no export
1289	 *    memory is allocated, so KILL and alpha test do not work correctly
1290	 *    without this.
1291	 * 2) Performance: Every shader needs at least a NULL export, even when
1292	 *    it writes no color/depth output. The NULL export instruction
1293	 *    stalls without this setting.
1294	 *
1295	 * Don't add this to CB_SHADER_MASK.
1296	 */
1297	if (!spi_shader_col_format &&
1298	    !info->writes_z && !info->writes_stencil && !info->writes_samplemask)
1299		spi_shader_col_format = V_028714_SPI_SHADER_32_R;
1300
1301	shader->ctx_reg.ps.spi_ps_input_ena = input_ena;
1302	shader->ctx_reg.ps.spi_ps_input_addr = shader->config.spi_ps_input_addr;
1303
1304	/* Set interpolation controls. */
1305	spi_ps_in_control = S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader));
1306
1307	shader->ctx_reg.ps.spi_baryc_cntl = spi_baryc_cntl;
1308	shader->ctx_reg.ps.spi_ps_in_control = spi_ps_in_control;
1309	shader->ctx_reg.ps.spi_shader_z_format =
1310			ac_get_spi_shader_z_format(info->writes_z,
1311						   info->writes_stencil,
1312						   info->writes_samplemask);
1313	shader->ctx_reg.ps.spi_shader_col_format = spi_shader_col_format;
1314	shader->ctx_reg.ps.cb_shader_mask = cb_shader_mask;
1315
1316	va = shader->bo->gpu_address;
1317	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
1318	si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
1319	si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS, S_00B024_MEM_BASE(va >> 40));
1320
1321	si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS,
1322		       S_00B028_VGPRS((shader->config.num_vgprs - 1) / 4) |
1323		       S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8) |
1324		       S_00B028_DX10_CLAMP(1) |
1325		       S_00B028_FLOAT_MODE(shader->config.float_mode));
1326	si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
1327		       S_00B02C_EXTRA_LDS_SIZE(shader->config.lds_size) |
1328		       S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR) |
1329		       S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
1330}
1331
1332static void si_shader_init_pm4_state(struct si_screen *sscreen,
1333                                     struct si_shader *shader)
1334{
1335	switch (shader->selector->type) {
1336	case PIPE_SHADER_VERTEX:
1337		if (shader->key.as_ls)
1338			si_shader_ls(sscreen, shader);
1339		else if (shader->key.as_es)
1340			si_shader_es(sscreen, shader);
1341		else
1342			si_shader_vs(sscreen, shader, NULL);
1343		break;
1344	case PIPE_SHADER_TESS_CTRL:
1345		si_shader_hs(sscreen, shader);
1346		break;
1347	case PIPE_SHADER_TESS_EVAL:
1348		if (shader->key.as_es)
1349			si_shader_es(sscreen, shader);
1350		else
1351			si_shader_vs(sscreen, shader, NULL);
1352		break;
1353	case PIPE_SHADER_GEOMETRY:
1354		si_shader_gs(sscreen, shader);
1355		break;
1356	case PIPE_SHADER_FRAGMENT:
1357		si_shader_ps(shader);
1358		break;
1359	default:
1360		assert(0);
1361	}
1362}
1363
1364static unsigned si_get_alpha_test_func(struct si_context *sctx)
1365{
1366	/* Alpha-test should be disabled if colorbuffer 0 is integer. */
1367	if (sctx->queued.named.dsa)
1368		return sctx->queued.named.dsa->alpha_func;
1369
1370	return PIPE_FUNC_ALWAYS;
1371}
1372
1373static void si_shader_selector_key_vs(struct si_context *sctx,
1374				      struct si_shader_selector *vs,
1375				      struct si_shader_key *key,
1376				      struct si_vs_prolog_bits *prolog_key)
1377{
1378	if (!sctx->vertex_elements ||
1379	    vs->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS])
1380		return;
1381
1382	prolog_key->instance_divisor_is_one =
1383		sctx->vertex_elements->instance_divisor_is_one;
1384	prolog_key->instance_divisor_is_fetched =
1385		sctx->vertex_elements->instance_divisor_is_fetched;
1386
1387	/* Prefer a monolithic shader to allow scheduling divisions around
1388	 * VBO loads. */
1389	if (prolog_key->instance_divisor_is_fetched)
1390		key->opt.prefer_mono = 1;
1391
1392	unsigned count = MIN2(vs->info.num_inputs,
1393			      sctx->vertex_elements->count);
1394	memcpy(key->mono.vs_fix_fetch, sctx->vertex_elements->fix_fetch, count);
1395}
1396
1397static void si_shader_selector_key_hw_vs(struct si_context *sctx,
1398					 struct si_shader_selector *vs,
1399					 struct si_shader_key *key)
1400{
1401	struct si_shader_selector *ps = sctx->ps_shader.cso;
1402
1403	key->opt.clip_disable =
1404		sctx->queued.named.rasterizer->clip_plane_enable == 0 &&
1405		(vs->info.clipdist_writemask ||
1406		 vs->info.writes_clipvertex) &&
1407		!vs->info.culldist_writemask;
1408
1409	/* Find out if PS is disabled. */
1410	bool ps_disabled = true;
1411	if (ps) {
1412		const struct si_state_blend *blend = sctx->queued.named.blend;
1413		bool alpha_to_coverage = blend && blend->alpha_to_coverage;
1414		bool ps_modifies_zs = ps->info.uses_kill ||
1415				      ps->info.writes_z ||
1416				      ps->info.writes_stencil ||
1417				      ps->info.writes_samplemask ||
1418				      alpha_to_coverage ||
1419				      si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS;
1420		unsigned ps_colormask = si_get_total_colormask(sctx);
1421
1422		ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
1423			      (!ps_colormask &&
1424			       !ps_modifies_zs &&
1425			       !ps->info.writes_memory);
1426	}
1427
1428	/* Find out which VS outputs aren't used by the PS. */
1429	uint64_t outputs_written = vs->outputs_written_before_ps;
1430	uint64_t inputs_read = 0;
1431
1432	/* Ignore outputs that are not passed from VS to PS. */
1433	outputs_written &= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION, 0, true)) |
1434			     (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE, 0, true)) |
1435			     (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX, 0, true)));
1436
1437	if (!ps_disabled) {
1438		inputs_read = ps->inputs_read;
1439	}
1440
1441	uint64_t linked = outputs_written & inputs_read;
1442
1443	key->opt.kill_outputs = ~linked & outputs_written;
1444}
1445
1446/* Compute the key for the hw shader variant */
1447static inline void si_shader_selector_key(struct pipe_context *ctx,
1448					  struct si_shader_selector *sel,
1449					  struct si_shader_key *key)
1450{
1451	struct si_context *sctx = (struct si_context *)ctx;
1452
1453	memset(key, 0, sizeof(*key));
1454
1455	switch (sel->type) {
1456	case PIPE_SHADER_VERTEX:
1457		si_shader_selector_key_vs(sctx, sel, key, &key->part.vs.prolog);
1458
1459		if (sctx->tes_shader.cso)
1460			key->as_ls = 1;
1461		else if (sctx->gs_shader.cso)
1462			key->as_es = 1;
1463		else {
1464			si_shader_selector_key_hw_vs(sctx, sel, key);
1465
1466			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
1467				key->mono.u.vs_export_prim_id = 1;
1468		}
1469		break;
1470	case PIPE_SHADER_TESS_CTRL:
1471		if (sctx->chip_class >= GFX9) {
1472			si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
1473						  key, &key->part.tcs.ls_prolog);
1474			key->part.tcs.ls = sctx->vs_shader.cso;
1475
1476			/* When the LS VGPR fix is needed, monolithic shaders
1477			 * can:
1478			 *  - avoid initializing EXEC in both the LS prolog
1479			 *    and the LS main part when !vs_needs_prolog
1480			 *  - remove the fixup for unused input VGPRs
1481			 */
1482			key->part.tcs.ls_prolog.ls_vgpr_fix = sctx->ls_vgpr_fix;
1483
1484			/* The LS output / HS input layout can be communicated
1485			 * directly instead of via user SGPRs for merged LS-HS.
1486			 * The LS VGPR fix prefers this too.
1487			 */
1488			key->opt.prefer_mono = 1;
1489		}
1490
1491		key->part.tcs.epilog.prim_mode =
1492			sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
1493		key->part.tcs.epilog.invoc0_tess_factors_are_def =
1494			sel->tcs_info.tessfactors_are_def_in_all_invocs;
1495		key->part.tcs.epilog.tes_reads_tess_factors =
1496			sctx->tes_shader.cso->info.reads_tess_factors;
1497
1498		if (sel == sctx->fixed_func_tcs_shader.cso)
1499			key->mono.u.ff_tcs_inputs_to_copy = sctx->vs_shader.cso->outputs_written;
1500		break;
1501	case PIPE_SHADER_TESS_EVAL:
1502		if (sctx->gs_shader.cso)
1503			key->as_es = 1;
1504		else {
1505			si_shader_selector_key_hw_vs(sctx, sel, key);
1506
1507			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
1508				key->mono.u.vs_export_prim_id = 1;
1509		}
1510		break;
1511	case PIPE_SHADER_GEOMETRY:
1512		if (sctx->chip_class >= GFX9) {
1513			if (sctx->tes_shader.cso) {
1514				key->part.gs.es = sctx->tes_shader.cso;
1515			} else {
1516				si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
1517							  key, &key->part.gs.vs_prolog);
1518				key->part.gs.es = sctx->vs_shader.cso;
1519				key->part.gs.prolog.gfx9_prev_is_vs = 1;
1520			}
1521
1522			/* Merged ES-GS can have unbalanced wave usage.
1523			 *
1524			 * ES threads are per-vertex, while GS threads are
1525			 * per-primitive. So without any amplification, there
1526			 * are fewer GS threads than ES threads, which can result
1527			 * in empty (no-op) GS waves. With too much amplification,
1528			 * there are more GS threads than ES threads, which
1529			 * can result in empty (no-op) ES waves.
1530			 *
1531			 * Non-monolithic shaders are implemented by setting EXEC
1532			 * at the beginning of shader parts, and don't jump to
1533			 * the end if EXEC is 0.
1534			 *
1535			 * Monolithic shaders use conditional blocks, so they can
1536			 * jump and skip empty waves of ES or GS. So set this to
1537			 * always use optimized variants, which are monolithic.
1538			 */
1539			key->opt.prefer_mono = 1;
1540		}
1541		key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
1542		break;
1543	case PIPE_SHADER_FRAGMENT: {
1544		struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1545		struct si_state_blend *blend = sctx->queued.named.blend;
1546
1547		if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1548		    sel->info.colors_written == 0x1)
1549			key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
1550
1551		if (blend) {
1552			/* Select the shader color format based on whether
1553			 * blending or alpha are needed.
1554			 */
1555			key->part.ps.epilog.spi_shader_col_format =
1556				(blend->blend_enable_4bit & blend->need_src_alpha_4bit &
1557				 sctx->framebuffer.spi_shader_col_format_blend_alpha) |
1558				(blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
1559				 sctx->framebuffer.spi_shader_col_format_blend) |
1560				(~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
1561				 sctx->framebuffer.spi_shader_col_format_alpha) |
1562				(~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
1563				 sctx->framebuffer.spi_shader_col_format);
1564			key->part.ps.epilog.spi_shader_col_format &= blend->cb_target_enabled_4bit;
1565
1566			/* The output for dual source blending should have
1567			 * the same format as the first output.
1568			 */
1569			if (blend->dual_src_blend)
1570				key->part.ps.epilog.spi_shader_col_format |=
1571					(key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
1572		} else
1573			key->part.ps.epilog.spi_shader_col_format = sctx->framebuffer.spi_shader_col_format;
1574
1575		/* If alpha-to-coverage is enabled, we have to export alpha
1576		 * even if there is no color buffer.
1577		 */
1578		if (!(key->part.ps.epilog.spi_shader_col_format & 0xf) &&
1579		    blend && blend->alpha_to_coverage)
1580			key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
1581
1582		/* On SI and CIK except Hawaii, the CB doesn't clamp outputs
1583		 * to the range supported by the type if a channel has less
1584		 * than 16 bits and the export format is 16_ABGR.
1585		 */
1586		if (sctx->chip_class <= CIK && sctx->family != CHIP_HAWAII) {
1587			key->part.ps.epilog.color_is_int8 = sctx->framebuffer.color_is_int8;
1588			key->part.ps.epilog.color_is_int10 = sctx->framebuffer.color_is_int10;
1589		}
1590
1591		/* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1592		if (!key->part.ps.epilog.last_cbuf) {
1593			key->part.ps.epilog.spi_shader_col_format &= sel->colors_written_4bit;
1594			key->part.ps.epilog.color_is_int8 &= sel->info.colors_written;
1595			key->part.ps.epilog.color_is_int10 &= sel->info.colors_written;
1596		}
1597
1598		bool is_poly = !util_prim_is_points_or_lines(sctx->current_rast_prim);
1599		bool is_line = util_prim_is_lines(sctx->current_rast_prim);
1600
1601		key->part.ps.prolog.color_two_side = rs->two_side && sel->info.colors_read;
1602		key->part.ps.prolog.flatshade_colors = rs->flatshade && sel->info.colors_read;
1603
1604		if (sctx->queued.named.blend) {
1605			key->part.ps.epilog.alpha_to_one = sctx->queued.named.blend->alpha_to_one &&
1606							   rs->multisample_enable;
1607		}
1608
1609		key->part.ps.prolog.poly_stipple = rs->poly_stipple_enable && is_poly;
1610		key->part.ps.epilog.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
1611							   (is_line && rs->line_smooth)) &&
1612							  sctx->framebuffer.nr_samples <= 1;
1613		key->part.ps.epilog.clamp_color = rs->clamp_fragment_color;
1614
1615		if (sctx->ps_iter_samples > 1 &&
1616		    sel->info.reads_samplemask) {
1617			key->part.ps.prolog.samplemask_log_ps_iter =
1618				util_logbase2(sctx->ps_iter_samples);
1619		}
1620
1621		if (rs->force_persample_interp &&
1622		    rs->multisample_enable &&
1623		    sctx->framebuffer.nr_samples > 1 &&
1624		    sctx->ps_iter_samples > 1) {
1625			key->part.ps.prolog.force_persp_sample_interp =
1626				sel->info.uses_persp_center ||
1627				sel->info.uses_persp_centroid;
1628
1629			key->part.ps.prolog.force_linear_sample_interp =
1630				sel->info.uses_linear_center ||
1631				sel->info.uses_linear_centroid;
1632		} else if (rs->multisample_enable &&
1633			   sctx->framebuffer.nr_samples > 1) {
1634			key->part.ps.prolog.bc_optimize_for_persp =
1635				sel->info.uses_persp_center &&
1636				sel->info.uses_persp_centroid;
1637			key->part.ps.prolog.bc_optimize_for_linear =
1638				sel->info.uses_linear_center &&
1639				sel->info.uses_linear_centroid;
1640		} else {
1641			/* Make sure SPI doesn't compute more than 1 pair
1642			 * of (i,j), which is the optimization here. */
1643			key->part.ps.prolog.force_persp_center_interp =
1644				sel->info.uses_persp_center +
1645				sel->info.uses_persp_centroid +
1646				sel->info.uses_persp_sample > 1;
1647
1648			key->part.ps.prolog.force_linear_center_interp =
1649				sel->info.uses_linear_center +
1650				sel->info.uses_linear_centroid +
1651				sel->info.uses_linear_sample > 1;
1652
1653			if (sel->info.opcode_count[TGSI_OPCODE_INTERP_SAMPLE])
1654				key->mono.u.ps.interpolate_at_sample_force_center = 1;
1655		}
1656
1657		key->part.ps.epilog.alpha_func = si_get_alpha_test_func(sctx);
1658
1659		/* ps_uses_fbfetch is true only if the color buffer is bound. */
1660		if (sctx->ps_uses_fbfetch && !sctx->blitter->running) {
1661			struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
1662			struct pipe_resource *tex = cb0->texture;
1663
1664			/* 1D textures are allocated and used as 2D on GFX9. */
1665			key->mono.u.ps.fbfetch_msaa = sctx->framebuffer.nr_samples > 1;
1666			key->mono.u.ps.fbfetch_is_1D = sctx->chip_class != GFX9 &&
1667						       (tex->target == PIPE_TEXTURE_1D ||
1668							tex->target == PIPE_TEXTURE_1D_ARRAY);
1669			key->mono.u.ps.fbfetch_layered = tex->target == PIPE_TEXTURE_1D_ARRAY ||
1670							 tex->target == PIPE_TEXTURE_2D_ARRAY ||
1671							 tex->target == PIPE_TEXTURE_CUBE ||
1672							 tex->target == PIPE_TEXTURE_CUBE_ARRAY ||
1673							 tex->target == PIPE_TEXTURE_3D;
1674		}
1675		break;
1676	}
1677	default:
1678		assert(0);
1679	}
1680
1681	if (unlikely(sctx->screen->debug_flags & DBG(NO_OPT_VARIANT)))
1682		memset(&key->opt, 0, sizeof(key->opt));
1683}
1684
1685static void si_build_shader_variant(struct si_shader *shader,
1686				    int thread_index,
1687				    bool low_priority)
1688{
1689	struct si_shader_selector *sel = shader->selector;
1690	struct si_screen *sscreen = sel->screen;
1691	struct ac_llvm_compiler *compiler;
1692	struct pipe_debug_callback *debug = &shader->compiler_ctx_state.debug;
1693	int r;
1694
1695	if (thread_index >= 0) {
1696		if (low_priority) {
1697			assert(thread_index < ARRAY_SIZE(sscreen->compiler_lowp));
1698			compiler = &sscreen->compiler_lowp[thread_index];
1699		} else {
1700			assert(thread_index < ARRAY_SIZE(sscreen->compiler));
1701			compiler = &sscreen->compiler[thread_index];
1702		}
1703		if (!debug->async)
1704			debug = NULL;
1705	} else {
1706		assert(!low_priority);
1707		compiler = shader->compiler_ctx_state.compiler;
1708	}
1709
1710	r = si_shader_create(sscreen, compiler, shader, debug);
1711	if (unlikely(r)) {
1712		PRINT_ERR("Failed to build shader variant (type=%u) %d\n",
1713			 sel->type, r);
1714		shader->compilation_failed = true;
1715		return;
1716	}
1717
1718	if (shader->compiler_ctx_state.is_debug_context) {
1719		FILE *f = open_memstream(&shader->shader_log,
1720					 &shader->shader_log_size);
1721		if (f) {
1722			si_shader_dump(sscreen, shader, NULL, sel->type, f, false);
1723			fclose(f);
1724		}
1725	}
1726
1727	si_shader_init_pm4_state(sscreen, shader);
1728}
1729
1730static void si_build_shader_variant_low_priority(void *job, int thread_index)
1731{
1732	struct si_shader *shader = (struct si_shader *)job;
1733
1734	assert(thread_index >= 0);
1735
1736	si_build_shader_variant(shader, thread_index, true);
1737}
1738
1739static const struct si_shader_key zeroed;
1740
1741static bool si_check_missing_main_part(struct si_screen *sscreen,
1742				       struct si_shader_selector *sel,
1743				       struct si_compiler_ctx_state *compiler_state,
1744				       struct si_shader_key *key)
1745{
1746	struct si_shader **mainp = si_get_main_shader_part(sel, key);
1747
1748	if (!*mainp) {
1749		struct si_shader *main_part = CALLOC_STRUCT(si_shader);
1750
1751		if (!main_part)
1752			return false;
1753
1754		/* We can leave the fence as permanently signaled because the
1755		 * main part becomes visible globally only after it has been
1756		 * compiled. */
1757		util_queue_fence_init(&main_part->ready);
1758
1759		main_part->selector = sel;
1760		main_part->key.as_es = key->as_es;
1761		main_part->key.as_ls = key->as_ls;
1762		main_part->is_monolithic = false;
1763
1764		if (si_compile_tgsi_shader(sscreen, compiler_state->compiler,
1765					   main_part, &compiler_state->debug) != 0) {
1766			FREE(main_part);
1767			return false;
1768		}
1769		*mainp = main_part;
1770	}
1771	return true;
1772}
1773
1774/* Select the hw shader variant depending on the current state. */
1775static int si_shader_select_with_key(struct si_screen *sscreen,
1776				     struct si_shader_ctx_state *state,
1777				     struct si_compiler_ctx_state *compiler_state,
1778				     struct si_shader_key *key,
1779				     int thread_index)
1780{
1781	struct si_shader_selector *sel = state->cso;
1782	struct si_shader_selector *previous_stage_sel = NULL;
1783	struct si_shader *current = state->current;
1784	struct si_shader *iter, *shader = NULL;
1785
1786again:
1787	/* Check if we don't need to change anything.
1788	 * This path is also used for most shaders that don't need multiple
1789	 * variants, it will cost just a computation of the key and this
1790	 * test. */
1791	if (likely(current &&
1792		   memcmp(&current->key, key, sizeof(*key)) == 0)) {
1793		if (unlikely(!util_queue_fence_is_signalled(&current->ready))) {
1794			if (current->is_optimized) {
1795				memset(&key->opt, 0, sizeof(key->opt));
1796				goto current_not_ready;
1797			}
1798
1799			util_queue_fence_wait(&current->ready);
1800		}
1801
1802		return current->compilation_failed ? -1 : 0;
1803	}
1804current_not_ready:
1805
1806	/* This must be done before the mutex is locked, because async GS
1807	 * compilation calls this function too, and therefore must enter
1808	 * the mutex first.
1809	 *
1810	 * Only wait if we are in a draw call. Don't wait if we are
1811	 * in a compiler thread.
1812	 */
1813	if (thread_index < 0)
1814		util_queue_fence_wait(&sel->ready);
1815
1816	mtx_lock(&sel->mutex);
1817
1818	/* Find the shader variant. */
1819	for (iter = sel->first_variant; iter; iter = iter->next_variant) {
1820		/* Don't check the "current" shader. We checked it above. */
1821		if (current != iter &&
1822		    memcmp(&iter->key, key, sizeof(*key)) == 0) {
1823			mtx_unlock(&sel->mutex);
1824
1825			if (unlikely(!util_queue_fence_is_signalled(&iter->ready))) {
1826				/* If it's an optimized shader and its compilation has
1827				 * been started but isn't done, use the unoptimized
1828				 * shader so as not to cause a stall due to compilation.
1829				 */
1830				if (iter->is_optimized) {
1831					memset(&key->opt, 0, sizeof(key->opt));
1832					goto again;
1833				}
1834
1835				util_queue_fence_wait(&iter->ready);
1836			}
1837
1838			if (iter->compilation_failed) {
1839				return -1; /* skip the draw call */
1840			}
1841
1842			state->current = iter;
1843			return 0;
1844		}
1845	}
1846
1847	/* Build a new shader. */
1848	shader = CALLOC_STRUCT(si_shader);
1849	if (!shader) {
1850		mtx_unlock(&sel->mutex);
1851		return -ENOMEM;
1852	}
1853
1854	util_queue_fence_init(&shader->ready);
1855
1856	shader->selector = sel;
1857	shader->key = *key;
1858	shader->compiler_ctx_state = *compiler_state;
1859
1860	/* If this is a merged shader, get the first shader's selector. */
1861	if (sscreen->info.chip_class >= GFX9) {
1862		if (sel->type == PIPE_SHADER_TESS_CTRL)
1863			previous_stage_sel = key->part.tcs.ls;
1864		else if (sel->type == PIPE_SHADER_GEOMETRY)
1865			previous_stage_sel = key->part.gs.es;
1866
1867		/* We need to wait for the previous shader. */
1868		if (previous_stage_sel && thread_index < 0)
1869			util_queue_fence_wait(&previous_stage_sel->ready);
1870	}
1871
1872	/* Compile the main shader part if it doesn't exist. This can happen
1873	 * if the initial guess was wrong. */
1874	bool is_pure_monolithic =
1875		sscreen->use_monolithic_shaders ||
1876		memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
1877
1878	if (!is_pure_monolithic) {
1879		bool ok;
1880
1881		/* Make sure the main shader part is present. This is needed
1882		 * for shaders that can be compiled as VS, LS, or ES, and only
1883		 * one of them is compiled at creation.
1884		 *
1885		 * For merged shaders, check that the starting shader's main
1886		 * part is present.
1887		 */
1888		if (previous_stage_sel) {
1889			struct si_shader_key shader1_key = zeroed;
1890
1891			if (sel->type == PIPE_SHADER_TESS_CTRL)
1892				shader1_key.as_ls = 1;
1893			else if (sel->type == PIPE_SHADER_GEOMETRY)
1894				shader1_key.as_es = 1;
1895			else
1896				assert(0);
1897
1898			mtx_lock(&previous_stage_sel->mutex);
1899			ok = si_check_missing_main_part(sscreen,
1900							previous_stage_sel,
1901							compiler_state, &shader1_key);
1902			mtx_unlock(&previous_stage_sel->mutex);
1903		} else {
1904			ok = si_check_missing_main_part(sscreen, sel,
1905							compiler_state, key);
1906		}
1907		if (!ok) {
1908			FREE(shader);
1909			mtx_unlock(&sel->mutex);
1910			return -ENOMEM; /* skip the draw call */
1911		}
1912	}
1913
1914	/* Keep the reference to the 1st shader of merged shaders, so that
1915	 * Gallium can't destroy it before we destroy the 2nd shader.
1916	 *
1917	 * Set sctx = NULL, because it's unused if we're not releasing
1918	 * the shader, and we don't have any sctx here.
1919	 */
1920	si_shader_selector_reference(NULL, &shader->previous_stage_sel,
1921				     previous_stage_sel);
1922
1923	/* Monolithic-only shaders don't make a distinction between optimized
1924	 * and unoptimized. */
1925	shader->is_monolithic =
1926		is_pure_monolithic ||
1927		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
1928
1929	shader->is_optimized =
1930		!is_pure_monolithic &&
1931		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
1932
1933	/* If it's an optimized shader, compile it asynchronously. */
1934	if (shader->is_optimized &&
1935	    !is_pure_monolithic &&
1936	    thread_index < 0) {
1937		/* Compile it asynchronously. */
1938		util_queue_add_job(&sscreen->shader_compiler_queue_low_priority,
1939				   shader, &shader->ready,
1940				   si_build_shader_variant_low_priority, NULL);
1941
1942		/* Add only after the ready fence was reset, to guard against a
1943		 * race with si_bind_XX_shader. */
1944		if (!sel->last_variant) {
1945			sel->first_variant = shader;
1946			sel->last_variant = shader;
1947		} else {
1948			sel->last_variant->next_variant = shader;
1949			sel->last_variant = shader;
1950		}
1951
1952		/* Use the default (unoptimized) shader for now. */
1953		memset(&key->opt, 0, sizeof(key->opt));
1954		mtx_unlock(&sel->mutex);
1955
1956		if (sscreen->options.sync_compile)
1957			util_queue_fence_wait(&shader->ready);
1958
1959		goto again;
1960	}
1961
1962	/* Reset the fence before adding to the variant list. */
1963	util_queue_fence_reset(&shader->ready);
1964
1965	if (!sel->last_variant) {
1966		sel->first_variant = shader;
1967		sel->last_variant = shader;
1968	} else {
1969		sel->last_variant->next_variant = shader;
1970		sel->last_variant = shader;
1971	}
1972
1973	mtx_unlock(&sel->mutex);
1974
1975	assert(!shader->is_optimized);
1976	si_build_shader_variant(shader, thread_index, false);
1977
1978	util_queue_fence_signal(&shader->ready);
1979
1980	if (!shader->compilation_failed)
1981		state->current = shader;
1982
1983	return shader->compilation_failed ? -1 : 0;
1984}
1985
1986static int si_shader_select(struct pipe_context *ctx,
1987			    struct si_shader_ctx_state *state,
1988			    struct si_compiler_ctx_state *compiler_state)
1989{
1990	struct si_context *sctx = (struct si_context *)ctx;
1991	struct si_shader_key key;
1992
1993	si_shader_selector_key(ctx, state->cso, &key);
1994	return si_shader_select_with_key(sctx->screen, state, compiler_state,
1995					 &key, -1);
1996}
1997
1998static void si_parse_next_shader_property(const struct tgsi_shader_info *info,
1999					  bool streamout,
2000					  struct si_shader_key *key)
2001{
2002	unsigned next_shader = info->properties[TGSI_PROPERTY_NEXT_SHADER];
2003
2004	switch (info->processor) {
2005	case PIPE_SHADER_VERTEX:
2006		switch (next_shader) {
2007		case PIPE_SHADER_GEOMETRY:
2008			key->as_es = 1;
2009			break;
2010		case PIPE_SHADER_TESS_CTRL:
2011		case PIPE_SHADER_TESS_EVAL:
2012			key->as_ls = 1;
2013			break;
2014		default:
2015			/* If POSITION isn't written, it can only be a HW VS
2016			 * if streamout is used. If streamout isn't used,
2017			 * assume that it's a HW LS. (the next shader is TCS)
2018			 * This heuristic is needed for separate shader objects.
2019			 */
2020			if (!info->writes_position && !streamout)
2021				key->as_ls = 1;
2022		}
2023		break;
2024
2025	case PIPE_SHADER_TESS_EVAL:
2026		if (next_shader == PIPE_SHADER_GEOMETRY ||
2027		    !info->writes_position)
2028			key->as_es = 1;
2029		break;
2030	}
2031}
2032
2033/**
2034 * Compile the main shader part or the monolithic shader as part of
2035 * si_shader_selector initialization. Since it can be done asynchronously,
2036 * there is no way to report compile failures to applications.
2037 */
2038static void si_init_shader_selector_async(void *job, int thread_index)
2039{
2040	struct si_shader_selector *sel = (struct si_shader_selector *)job;
2041	struct si_screen *sscreen = sel->screen;
2042	struct ac_llvm_compiler *compiler;
2043	struct pipe_debug_callback *debug = &sel->compiler_ctx_state.debug;
2044
2045	assert(!debug->debug_message || debug->async);
2046	assert(thread_index >= 0);
2047	assert(thread_index < ARRAY_SIZE(sscreen->compiler));
2048	compiler = &sscreen->compiler[thread_index];
2049
2050	if (sel->nir)
2051		si_lower_nir(sel);
2052
2053	/* Compile the main shader part for use with a prolog and/or epilog.
2054	 * If this fails, the driver will try to compile a monolithic shader
2055	 * on demand.
2056	 */
2057	if (!sscreen->use_monolithic_shaders) {
2058		struct si_shader *shader = CALLOC_STRUCT(si_shader);
2059		void *ir_binary = NULL;
2060
2061		if (!shader) {
2062			fprintf(stderr, "radeonsi: can't allocate a main shader part\n");
2063			return;
2064		}
2065
2066		/* We can leave the fence signaled because use of the default
2067		 * main part is guarded by the selector's ready fence. */
2068		util_queue_fence_init(&shader->ready);
2069
2070		shader->selector = sel;
2071		shader->is_monolithic = false;
2072		si_parse_next_shader_property(&sel->info,
2073					      sel->so.num_outputs != 0,
2074					      &shader->key);
2075
2076		if (sel->tokens || sel->nir)
2077			ir_binary = si_get_ir_binary(sel);
2078
2079		/* Try to load the shader from the shader cache. */
2080		mtx_lock(&sscreen->shader_cache_mutex);
2081
2082		if (ir_binary &&
2083		    si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
2084			mtx_unlock(&sscreen->shader_cache_mutex);
2085			si_shader_dump_stats_for_shader_db(shader, debug);
2086		} else {
2087			mtx_unlock(&sscreen->shader_cache_mutex);
2088
2089			/* Compile the shader if it hasn't been loaded from the cache. */
2090			if (si_compile_tgsi_shader(sscreen, compiler, shader,
2091						   debug) != 0) {
2092				FREE(shader);
2093				FREE(ir_binary);
2094				fprintf(stderr, "radeonsi: can't compile a main shader part\n");
2095				return;
2096			}
2097
2098			if (ir_binary) {
2099				mtx_lock(&sscreen->shader_cache_mutex);
2100				if (!si_shader_cache_insert_shader(sscreen, ir_binary, shader, true))
2101					FREE(ir_binary);
2102				mtx_unlock(&sscreen->shader_cache_mutex);
2103			}
2104		}
2105
2106		*si_get_main_shader_part(sel, &shader->key) = shader;
2107
2108		/* Unset "outputs_written" flags for outputs converted to
2109		 * DEFAULT_VAL, so that later inter-shader optimizations don't
2110		 * try to eliminate outputs that don't exist in the final
2111		 * shader.
2112		 *
2113		 * This is only done if non-monolithic shaders are enabled.
2114		 */
2115		if ((sel->type == PIPE_SHADER_VERTEX ||
2116		     sel->type == PIPE_SHADER_TESS_EVAL) &&
2117		    !shader->key.as_ls &&
2118		    !shader->key.as_es) {
2119			unsigned i;
2120
2121			for (i = 0; i < sel->info.num_outputs; i++) {
2122				unsigned offset = shader->info.vs_output_param_offset[i];
2123
2124				if (offset <= AC_EXP_PARAM_OFFSET_31)
2125					continue;
2126
2127				unsigned name = sel->info.output_semantic_name[i];
2128				unsigned index = sel->info.output_semantic_index[i];
2129				unsigned id;
2130
2131				switch (name) {
2132				case TGSI_SEMANTIC_GENERIC:
2133					/* don't process indices the function can't handle */
2134					if (index >= SI_MAX_IO_GENERIC)
2135						break;
2136					/* fall through */
2137				default:
2138					id = si_shader_io_get_unique_index(name, index, true);
2139					sel->outputs_written_before_ps &= ~(1ull << id);
2140					break;
2141				case TGSI_SEMANTIC_POSITION: /* ignore these */
2142				case TGSI_SEMANTIC_PSIZE:
2143				case TGSI_SEMANTIC_CLIPVERTEX:
2144				case TGSI_SEMANTIC_EDGEFLAG:
2145					break;
2146				}
2147			}
2148		}
2149	}
2150
2151	/* The GS copy shader is always pre-compiled. */
2152	if (sel->type == PIPE_SHADER_GEOMETRY) {
2153		sel->gs_copy_shader = si_generate_gs_copy_shader(sscreen, compiler, sel, debug);
2154		if (!sel->gs_copy_shader) {
2155			fprintf(stderr, "radeonsi: can't create GS copy shader\n");
2156			return;
2157		}
2158
2159		si_shader_vs(sscreen, sel->gs_copy_shader, sel);
2160	}
2161}
2162
2163void si_schedule_initial_compile(struct si_context *sctx, unsigned processor,
2164				 struct util_queue_fence *ready_fence,
2165				 struct si_compiler_ctx_state *compiler_ctx_state,
2166				 void *job, util_queue_execute_func execute)
2167{
2168	util_queue_fence_init(ready_fence);
2169
2170	struct util_async_debug_callback async_debug;
2171	bool debug =
2172		(sctx->debug.debug_message && !sctx->debug.async) ||
2173		sctx->is_debug ||
2174		si_can_dump_shader(sctx->screen, processor);
2175
2176	if (debug) {
2177		u_async_debug_init(&async_debug);
2178		compiler_ctx_state->debug = async_debug.base;
2179	}
2180
2181	util_queue_add_job(&sctx->screen->shader_compiler_queue, job,
2182			   ready_fence, execute, NULL);
2183
2184	if (debug) {
2185		util_queue_fence_wait(ready_fence);
2186		u_async_debug_drain(&async_debug, &sctx->debug);
2187		u_async_debug_cleanup(&async_debug);
2188	}
2189
2190	if (sctx->screen->options.sync_compile)
2191		util_queue_fence_wait(ready_fence);
2192}
2193
2194/* Return descriptor slot usage masks from the given shader info. */
2195void si_get_active_slot_masks(const struct tgsi_shader_info *info,
2196			      uint32_t *const_and_shader_buffers,
2197			      uint64_t *samplers_and_images)
2198{
2199	unsigned start, num_shaderbufs, num_constbufs, num_images, num_samplers;
2200
2201	num_shaderbufs = util_last_bit(info->shader_buffers_declared);
2202	num_constbufs = util_last_bit(info->const_buffers_declared);
2203	/* two 8-byte images share one 16-byte slot */
2204	num_images = align(util_last_bit(info->images_declared), 2);
2205	num_samplers = util_last_bit(info->samplers_declared);
2206
2207	/* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2208	start = si_get_shaderbuf_slot(num_shaderbufs - 1);
2209	*const_and_shader_buffers =
2210		u_bit_consecutive(start, num_shaderbufs + num_constbufs);
2211
2212	/* The layout is: image[last] ... image[0], sampler[0] ... sampler[last] */
2213	start = si_get_image_slot(num_images - 1) / 2;
2214	*samplers_and_images =
2215		u_bit_consecutive64(start, num_images / 2 + num_samplers);
2216}
2217
2218static void *si_create_shader_selector(struct pipe_context *ctx,
2219				       const struct pipe_shader_state *state)
2220{
2221	struct si_screen *sscreen = (struct si_screen *)ctx->screen;
2222	struct si_context *sctx = (struct si_context*)ctx;
2223	struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
2224	int i;
2225
2226	if (!sel)
2227		return NULL;
2228
2229	pipe_reference_init(&sel->reference, 1);
2230	sel->screen = sscreen;
2231	sel->compiler_ctx_state.debug = sctx->debug;
2232	sel->compiler_ctx_state.is_debug_context = sctx->is_debug;
2233
2234	sel->so = state->stream_output;
2235
2236	if (state->type == PIPE_SHADER_IR_TGSI) {
2237		sel->tokens = tgsi_dup_tokens(state->tokens);
2238		if (!sel->tokens) {
2239			FREE(sel);
2240			return NULL;
2241		}
2242
2243		tgsi_scan_shader(state->tokens, &sel->info);
2244		tgsi_scan_tess_ctrl(state->tokens, &sel->info, &sel->tcs_info);
2245	} else {
2246		assert(state->type == PIPE_SHADER_IR_NIR);
2247
2248		sel->nir = state->ir.nir;
2249
2250		si_nir_opts(sel->nir);
2251		si_nir_scan_shader(sel->nir, &sel->info);
2252		si_nir_scan_tess_ctrl(sel->nir, &sel->tcs_info);
2253	}
2254
2255	sel->type = sel->info.processor;
2256	p_atomic_inc(&sscreen->num_shaders_created);
2257	si_get_active_slot_masks(&sel->info,
2258				 &sel->active_const_and_shader_buffers,
2259				 &sel->active_samplers_and_images);
2260
2261	/* Record which streamout buffers are enabled. */
2262	for (i = 0; i < sel->so.num_outputs; i++) {
2263		sel->enabled_streamout_buffer_mask |=
2264			(1 << sel->so.output[i].output_buffer) <<
2265			(sel->so.output[i].stream * 4);
2266	}
2267
2268	/* The prolog is a no-op if there are no inputs. */
2269	sel->vs_needs_prolog = sel->type == PIPE_SHADER_VERTEX &&
2270			       sel->info.num_inputs &&
2271			       !sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
2272
2273	sel->force_correct_derivs_after_kill =
2274		sel->type == PIPE_SHADER_FRAGMENT &&
2275		sel->info.uses_derivatives &&
2276		sel->info.uses_kill &&
2277		sctx->screen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL);
2278
2279	/* Set which opcode uses which (i,j) pair. */
2280	if (sel->info.uses_persp_opcode_interp_centroid)
2281		sel->info.uses_persp_centroid = true;
2282
2283	if (sel->info.uses_linear_opcode_interp_centroid)
2284		sel->info.uses_linear_centroid = true;
2285
2286	if (sel->info.uses_persp_opcode_interp_offset ||
2287	    sel->info.uses_persp_opcode_interp_sample)
2288		sel->info.uses_persp_center = true;
2289
2290	if (sel->info.uses_linear_opcode_interp_offset ||
2291	    sel->info.uses_linear_opcode_interp_sample)
2292		sel->info.uses_linear_center = true;
2293
2294	switch (sel->type) {
2295	case PIPE_SHADER_GEOMETRY:
2296		sel->gs_output_prim =
2297			sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
2298		sel->gs_max_out_vertices =
2299			sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
2300		sel->gs_num_invocations =
2301			sel->info.properties[TGSI_PROPERTY_GS_INVOCATIONS];
2302		sel->gsvs_vertex_size = sel->info.num_outputs * 16;
2303		sel->max_gsvs_emit_size = sel->gsvs_vertex_size *
2304					  sel->gs_max_out_vertices;
2305
2306		sel->max_gs_stream = 0;
2307		for (i = 0; i < sel->so.num_outputs; i++)
2308			sel->max_gs_stream = MAX2(sel->max_gs_stream,
2309						  sel->so.output[i].stream);
2310
2311		sel->gs_input_verts_per_prim =
2312			u_vertices_per_prim(sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM]);
2313		break;
2314
2315	case PIPE_SHADER_TESS_CTRL:
2316		/* Always reserve space for these. */
2317		sel->patch_outputs_written |=
2318			(1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER, 0)) |
2319			(1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER, 0));
2320		/* fall through */
2321	case PIPE_SHADER_VERTEX:
2322	case PIPE_SHADER_TESS_EVAL:
2323		for (i = 0; i < sel->info.num_outputs; i++) {
2324			unsigned name = sel->info.output_semantic_name[i];
2325			unsigned index = sel->info.output_semantic_index[i];
2326
2327			switch (name) {
2328			case TGSI_SEMANTIC_TESSINNER:
2329			case TGSI_SEMANTIC_TESSOUTER:
2330			case TGSI_SEMANTIC_PATCH:
2331				sel->patch_outputs_written |=
2332					1ull << si_shader_io_get_unique_index_patch(name, index);
2333				break;
2334
2335			case TGSI_SEMANTIC_GENERIC:
2336				/* don't process indices the function can't handle */
2337				if (index >= SI_MAX_IO_GENERIC)
2338					break;
2339				/* fall through */
2340			default:
2341				sel->outputs_written |=
2342					1ull << si_shader_io_get_unique_index(name, index, false);
2343				sel->outputs_written_before_ps |=
2344					1ull << si_shader_io_get_unique_index(name, index, true);
2345				break;
2346			case TGSI_SEMANTIC_EDGEFLAG:
2347				break;
2348			}
2349		}
2350		sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
2351		sel->lshs_vertex_stride = sel->esgs_itemsize;
2352
2353		/* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2354		 * will start on a different bank. (except for the maximum 32*16).
2355		 */
2356		if (sel->lshs_vertex_stride < 32*16)
2357			sel->lshs_vertex_stride += 4;
2358
2359		/* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2360		 * conflicts, i.e. each vertex will start at a different bank.
2361		 */
2362		if (sctx->chip_class >= GFX9)
2363			sel->esgs_itemsize += 4;
2364
2365		assert(((sel->esgs_itemsize / 4) & C_028AAC_ITEMSIZE) == 0);
2366		break;
2367
2368	case PIPE_SHADER_FRAGMENT:
2369		for (i = 0; i < sel->info.num_inputs; i++) {
2370			unsigned name = sel->info.input_semantic_name[i];
2371			unsigned index = sel->info.input_semantic_index[i];
2372
2373			switch (name) {
2374			case TGSI_SEMANTIC_GENERIC:
2375				/* don't process indices the function can't handle */
2376				if (index >= SI_MAX_IO_GENERIC)
2377					break;
2378				/* fall through */
2379			default:
2380				sel->inputs_read |=
2381					1ull << si_shader_io_get_unique_index(name, index, true);
2382				break;
2383			case TGSI_SEMANTIC_PCOORD: /* ignore this */
2384				break;
2385			}
2386		}
2387
2388		for (i = 0; i < 8; i++)
2389			if (sel->info.colors_written & (1 << i))
2390				sel->colors_written_4bit |= 0xf << (4 * i);
2391
2392		for (i = 0; i < sel->info.num_inputs; i++) {
2393			if (sel->info.input_semantic_name[i] == TGSI_SEMANTIC_COLOR) {
2394				int index = sel->info.input_semantic_index[i];
2395				sel->color_attr_index[index] = i;
2396			}
2397		}
2398		break;
2399	}
2400
2401	/* PA_CL_VS_OUT_CNTL */
2402	bool misc_vec_ena =
2403		sel->info.writes_psize || sel->info.writes_edgeflag ||
2404		sel->info.writes_layer || sel->info.writes_viewport_index;
2405	sel->pa_cl_vs_out_cntl =
2406		S_02881C_USE_VTX_POINT_SIZE(sel->info.writes_psize) |
2407		S_02881C_USE_VTX_EDGE_FLAG(sel->info.writes_edgeflag) |
2408		S_02881C_USE_VTX_RENDER_TARGET_INDX(sel->info.writes_layer) |
2409		S_02881C_USE_VTX_VIEWPORT_INDX(sel->info.writes_viewport_index) |
2410		S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
2411		S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena);
2412	sel->clipdist_mask = sel->info.writes_clipvertex ?
2413				     SIX_BITS : sel->info.clipdist_writemask;
2414	sel->culldist_mask = sel->info.culldist_writemask <<
2415			     sel->info.num_written_clipdistance;
2416
2417	/* DB_SHADER_CONTROL */
2418	sel->db_shader_control =
2419		S_02880C_Z_EXPORT_ENABLE(sel->info.writes_z) |
2420		S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel->info.writes_stencil) |
2421		S_02880C_MASK_EXPORT_ENABLE(sel->info.writes_samplemask) |
2422		S_02880C_KILL_ENABLE(sel->info.uses_kill);
2423
2424	switch (sel->info.properties[TGSI_PROPERTY_FS_DEPTH_LAYOUT]) {
2425	case TGSI_FS_DEPTH_LAYOUT_GREATER:
2426		sel->db_shader_control |=
2427			S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
2428		break;
2429	case TGSI_FS_DEPTH_LAYOUT_LESS:
2430		sel->db_shader_control |=
2431			S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
2432		break;
2433	}
2434
2435	/* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2436	 *
2437	 *   | early Z/S | writes_mem | allow_ReZ? |      Z_ORDER       | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2438	 * --|-----------|------------|------------|--------------------|-------------------|-------------
2439	 * 1a|   false   |   false    |   true     | EarlyZ_Then_ReZ    |         0         |     0
2440	 * 1b|   false   |   false    |   false    | EarlyZ_Then_LateZ  |         0         |     0
2441	 * 2 |   false   |   true     |   n/a      |       LateZ        |         1         |     0
2442	 * 3 |   true    |   false    |   n/a      | EarlyZ_Then_LateZ  |         0         |     0
2443	 * 4 |   true    |   true     |   n/a      | EarlyZ_Then_LateZ  |         0         |     1
2444	 *
2445	 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2446	 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2447	 *
2448	 * Don't use ReZ without profiling !!!
2449	 *
2450	 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2451	 * shaders.
2452	 */
2453	if (sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]) {
2454		/* Cases 3, 4. */
2455		sel->db_shader_control |= S_02880C_DEPTH_BEFORE_SHADER(1) |
2456					  S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z) |
2457					  S_02880C_EXEC_ON_NOOP(sel->info.writes_memory);
2458	} else if (sel->info.writes_memory) {
2459		/* Case 2. */
2460		sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z) |
2461					  S_02880C_EXEC_ON_HIER_FAIL(1);
2462	} else {
2463		/* Case 1. */
2464		sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
2465	}
2466
2467	(void) mtx_init(&sel->mutex, mtx_plain);
2468
2469	si_schedule_initial_compile(sctx, sel->info.processor, &sel->ready,
2470				    &sel->compiler_ctx_state, sel,
2471				    si_init_shader_selector_async);
2472	return sel;
2473}
2474
2475static void si_update_streamout_state(struct si_context *sctx)
2476{
2477	struct si_shader_selector *shader_with_so = si_get_vs(sctx)->cso;
2478
2479	if (!shader_with_so)
2480		return;
2481
2482	sctx->streamout.enabled_stream_buffers_mask =
2483		shader_with_so->enabled_streamout_buffer_mask;
2484	sctx->streamout.stride_in_dw = shader_with_so->so.stride;
2485}
2486
2487static void si_update_clip_regs(struct si_context *sctx,
2488				struct si_shader_selector *old_hw_vs,
2489				struct si_shader *old_hw_vs_variant,
2490				struct si_shader_selector *next_hw_vs,
2491				struct si_shader *next_hw_vs_variant)
2492{
2493	if (next_hw_vs &&
2494	    (!old_hw_vs ||
2495	     old_hw_vs->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] !=
2496	     next_hw_vs->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] ||
2497	     old_hw_vs->pa_cl_vs_out_cntl != next_hw_vs->pa_cl_vs_out_cntl ||
2498	     old_hw_vs->clipdist_mask != next_hw_vs->clipdist_mask ||
2499	     old_hw_vs->culldist_mask != next_hw_vs->culldist_mask ||
2500	     !old_hw_vs_variant ||
2501	     !next_hw_vs_variant ||
2502	     old_hw_vs_variant->key.opt.clip_disable !=
2503	     next_hw_vs_variant->key.opt.clip_disable))
2504		si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
2505}
2506
2507static void si_update_common_shader_state(struct si_context *sctx)
2508{
2509	sctx->uses_bindless_samplers =
2510		si_shader_uses_bindless_samplers(sctx->vs_shader.cso)  ||
2511		si_shader_uses_bindless_samplers(sctx->gs_shader.cso)  ||
2512		si_shader_uses_bindless_samplers(sctx->ps_shader.cso)  ||
2513		si_shader_uses_bindless_samplers(sctx->tcs_shader.cso) ||
2514		si_shader_uses_bindless_samplers(sctx->tes_shader.cso);
2515	sctx->uses_bindless_images =
2516		si_shader_uses_bindless_images(sctx->vs_shader.cso)  ||
2517		si_shader_uses_bindless_images(sctx->gs_shader.cso)  ||
2518		si_shader_uses_bindless_images(sctx->ps_shader.cso)  ||
2519		si_shader_uses_bindless_images(sctx->tcs_shader.cso) ||
2520		si_shader_uses_bindless_images(sctx->tes_shader.cso);
2521	sctx->do_update_shaders = true;
2522}
2523
2524static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
2525{
2526	struct si_context *sctx = (struct si_context *)ctx;
2527	struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
2528	struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
2529	struct si_shader_selector *sel = state;
2530
2531	if (sctx->vs_shader.cso == sel)
2532		return;
2533
2534	sctx->vs_shader.cso = sel;
2535	sctx->vs_shader.current = sel ? sel->first_variant : NULL;
2536	sctx->num_vs_blit_sgprs = sel ? sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS] : 0;
2537
2538	si_update_common_shader_state(sctx);
2539	si_update_vs_viewport_state(sctx);
2540	si_set_active_descriptors_for_shader(sctx, sel);
2541	si_update_streamout_state(sctx);
2542	si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
2543			    si_get_vs(sctx)->cso, si_get_vs_state(sctx));
2544}
2545
2546static void si_update_tess_uses_prim_id(struct si_context *sctx)
2547{
2548	sctx->ia_multi_vgt_param_key.u.tess_uses_prim_id =
2549		(sctx->tes_shader.cso &&
2550		 sctx->tes_shader.cso->info.uses_primid) ||
2551		(sctx->tcs_shader.cso &&
2552		 sctx->tcs_shader.cso->info.uses_primid) ||
2553		(sctx->gs_shader.cso &&
2554		 sctx->gs_shader.cso->info.uses_primid) ||
2555		(sctx->ps_shader.cso && !sctx->gs_shader.cso &&
2556		 sctx->ps_shader.cso->info.uses_primid);
2557}
2558
2559static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
2560{
2561	struct si_context *sctx = (struct si_context *)ctx;
2562	struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
2563	struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
2564	struct si_shader_selector *sel = state;
2565	bool enable_changed = !!sctx->gs_shader.cso != !!sel;
2566
2567	if (sctx->gs_shader.cso == sel)
2568		return;
2569
2570	sctx->gs_shader.cso = sel;
2571	sctx->gs_shader.current = sel ? sel->first_variant : NULL;
2572	sctx->ia_multi_vgt_param_key.u.uses_gs = sel != NULL;
2573
2574	si_update_common_shader_state(sctx);
2575	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
2576
2577	if (enable_changed) {
2578		si_shader_change_notify(sctx);
2579		if (sctx->ia_multi_vgt_param_key.u.uses_tess)
2580			si_update_tess_uses_prim_id(sctx);
2581	}
2582	si_update_vs_viewport_state(sctx);
2583	si_set_active_descriptors_for_shader(sctx, sel);
2584	si_update_streamout_state(sctx);
2585	si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
2586			    si_get_vs(sctx)->cso, si_get_vs_state(sctx));
2587}
2588
2589static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
2590{
2591	struct si_context *sctx = (struct si_context *)ctx;
2592	struct si_shader_selector *sel = state;
2593	bool enable_changed = !!sctx->tcs_shader.cso != !!sel;
2594
2595	if (sctx->tcs_shader.cso == sel)
2596		return;
2597
2598	sctx->tcs_shader.cso = sel;
2599	sctx->tcs_shader.current = sel ? sel->first_variant : NULL;
2600	si_update_tess_uses_prim_id(sctx);
2601
2602	si_update_common_shader_state(sctx);
2603
2604	if (enable_changed)
2605		sctx->last_tcs = NULL; /* invalidate derived tess state */
2606
2607	si_set_active_descriptors_for_shader(sctx, sel);
2608}
2609
2610static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
2611{
2612	struct si_context *sctx = (struct si_context *)ctx;
2613	struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
2614	struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
2615	struct si_shader_selector *sel = state;
2616	bool enable_changed = !!sctx->tes_shader.cso != !!sel;
2617
2618	if (sctx->tes_shader.cso == sel)
2619		return;
2620
2621	sctx->tes_shader.cso = sel;
2622	sctx->tes_shader.current = sel ? sel->first_variant : NULL;
2623	sctx->ia_multi_vgt_param_key.u.uses_tess = sel != NULL;
2624	si_update_tess_uses_prim_id(sctx);
2625
2626	si_update_common_shader_state(sctx);
2627	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
2628
2629	if (enable_changed) {
2630		si_shader_change_notify(sctx);
2631		sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
2632	}
2633	si_update_vs_viewport_state(sctx);
2634	si_set_active_descriptors_for_shader(sctx, sel);
2635	si_update_streamout_state(sctx);
2636	si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
2637			    si_get_vs(sctx)->cso, si_get_vs_state(sctx));
2638}
2639
2640static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
2641{
2642	struct si_context *sctx = (struct si_context *)ctx;
2643	struct si_shader_selector *old_sel = sctx->ps_shader.cso;
2644	struct si_shader_selector *sel = state;
2645
2646	/* skip if supplied shader is one already in use */
2647	if (old_sel == sel)
2648		return;
2649
2650	sctx->ps_shader.cso = sel;
2651	sctx->ps_shader.current = sel ? sel->first_variant : NULL;
2652
2653	si_update_common_shader_state(sctx);
2654	if (sel) {
2655		if (sctx->ia_multi_vgt_param_key.u.uses_tess)
2656			si_update_tess_uses_prim_id(sctx);
2657
2658		if (!old_sel ||
2659		    old_sel->info.colors_written != sel->info.colors_written)
2660			si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
2661
2662		if (sctx->screen->has_out_of_order_rast &&
2663		    (!old_sel ||
2664		     old_sel->info.writes_memory != sel->info.writes_memory ||
2665		     old_sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] !=
2666		     sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]))
2667			si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2668	}
2669	si_set_active_descriptors_for_shader(sctx, sel);
2670	si_update_ps_colorbuf0_slot(sctx);
2671}
2672
2673static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
2674{
2675	if (shader->is_optimized) {
2676		util_queue_drop_job(&sctx->screen->shader_compiler_queue_low_priority,
2677				    &shader->ready);
2678	}
2679
2680	util_queue_fence_destroy(&shader->ready);
2681
2682	if (shader->pm4) {
2683		switch (shader->selector->type) {
2684		case PIPE_SHADER_VERTEX:
2685			if (shader->key.as_ls) {
2686				assert(sctx->chip_class <= VI);
2687				si_pm4_delete_state(sctx, ls, shader->pm4);
2688			} else if (shader->key.as_es) {
2689				assert(sctx->chip_class <= VI);
2690				si_pm4_delete_state(sctx, es, shader->pm4);
2691			} else {
2692				si_pm4_delete_state(sctx, vs, shader->pm4);
2693			}
2694			break;
2695		case PIPE_SHADER_TESS_CTRL:
2696			si_pm4_delete_state(sctx, hs, shader->pm4);
2697			break;
2698		case PIPE_SHADER_TESS_EVAL:
2699			if (shader->key.as_es) {
2700				assert(sctx->chip_class <= VI);
2701				si_pm4_delete_state(sctx, es, shader->pm4);
2702			} else {
2703				si_pm4_delete_state(sctx, vs, shader->pm4);
2704			}
2705			break;
2706		case PIPE_SHADER_GEOMETRY:
2707			if (shader->is_gs_copy_shader)
2708				si_pm4_delete_state(sctx, vs, shader->pm4);
2709			else
2710				si_pm4_delete_state(sctx, gs, shader->pm4);
2711			break;
2712		case PIPE_SHADER_FRAGMENT:
2713			si_pm4_delete_state(sctx, ps, shader->pm4);
2714			break;
2715		}
2716	}
2717
2718	si_shader_selector_reference(sctx, &shader->previous_stage_sel, NULL);
2719	si_shader_destroy(shader);
2720	free(shader);
2721}
2722
2723void si_destroy_shader_selector(struct si_context *sctx,
2724				struct si_shader_selector *sel)
2725{
2726	struct si_shader *p = sel->first_variant, *c;
2727	struct si_shader_ctx_state *current_shader[SI_NUM_SHADERS] = {
2728		[PIPE_SHADER_VERTEX] = &sctx->vs_shader,
2729		[PIPE_SHADER_TESS_CTRL] = &sctx->tcs_shader,
2730		[PIPE_SHADER_TESS_EVAL] = &sctx->tes_shader,
2731		[PIPE_SHADER_GEOMETRY] = &sctx->gs_shader,
2732		[PIPE_SHADER_FRAGMENT] = &sctx->ps_shader,
2733	};
2734
2735	util_queue_drop_job(&sctx->screen->shader_compiler_queue, &sel->ready);
2736
2737	if (current_shader[sel->type]->cso == sel) {
2738		current_shader[sel->type]->cso = NULL;
2739		current_shader[sel->type]->current = NULL;
2740	}
2741
2742	while (p) {
2743		c = p->next_variant;
2744		si_delete_shader(sctx, p);
2745		p = c;
2746	}
2747
2748	if (sel->main_shader_part)
2749		si_delete_shader(sctx, sel->main_shader_part);
2750	if (sel->main_shader_part_ls)
2751		si_delete_shader(sctx, sel->main_shader_part_ls);
2752	if (sel->main_shader_part_es)
2753		si_delete_shader(sctx, sel->main_shader_part_es);
2754	if (sel->gs_copy_shader)
2755		si_delete_shader(sctx, sel->gs_copy_shader);
2756
2757	util_queue_fence_destroy(&sel->ready);
2758	mtx_destroy(&sel->mutex);
2759	free(sel->tokens);
2760	ralloc_free(sel->nir);
2761	free(sel);
2762}
2763
2764static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
2765{
2766	struct si_context *sctx = (struct si_context *)ctx;
2767	struct si_shader_selector *sel = (struct si_shader_selector *)state;
2768
2769	si_shader_selector_reference(sctx, &sel, NULL);
2770}
2771
2772static unsigned si_get_ps_input_cntl(struct si_context *sctx,
2773				     struct si_shader *vs, unsigned name,
2774				     unsigned index, unsigned interpolate)
2775{
2776	struct tgsi_shader_info *vsinfo = &vs->selector->info;
2777	unsigned j, offset, ps_input_cntl = 0;
2778
2779	if (interpolate == TGSI_INTERPOLATE_CONSTANT ||
2780	    (interpolate == TGSI_INTERPOLATE_COLOR && sctx->flatshade))
2781		ps_input_cntl |= S_028644_FLAT_SHADE(1);
2782
2783	if (name == TGSI_SEMANTIC_PCOORD ||
2784	    (name == TGSI_SEMANTIC_TEXCOORD &&
2785	     sctx->sprite_coord_enable & (1 << index))) {
2786		ps_input_cntl |= S_028644_PT_SPRITE_TEX(1);
2787	}
2788
2789	for (j = 0; j < vsinfo->num_outputs; j++) {
2790		if (name == vsinfo->output_semantic_name[j] &&
2791		    index == vsinfo->output_semantic_index[j]) {
2792			offset = vs->info.vs_output_param_offset[j];
2793
2794			if (offset <= AC_EXP_PARAM_OFFSET_31) {
2795				/* The input is loaded from parameter memory. */
2796				ps_input_cntl |= S_028644_OFFSET(offset);
2797			} else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl)) {
2798				if (offset == AC_EXP_PARAM_UNDEFINED) {
2799					/* This can happen with depth-only rendering. */
2800					offset = 0;
2801				} else {
2802					/* The input is a DEFAULT_VAL constant. */
2803					assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
2804					       offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
2805					offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
2806				}
2807
2808				ps_input_cntl = S_028644_OFFSET(0x20) |
2809						S_028644_DEFAULT_VAL(offset);
2810			}
2811			break;
2812		}
2813	}
2814
2815	if (name == TGSI_SEMANTIC_PRIMID)
2816		/* PrimID is written after the last output. */
2817		ps_input_cntl |= S_028644_OFFSET(vs->info.vs_output_param_offset[vsinfo->num_outputs]);
2818	else if (j == vsinfo->num_outputs && !G_028644_PT_SPRITE_TEX(ps_input_cntl)) {
2819		/* No corresponding output found, load defaults into input.
2820		 * Don't set any other bits.
2821		 * (FLAT_SHADE=1 completely changes behavior) */
2822		ps_input_cntl = S_028644_OFFSET(0x20);
2823		/* D3D 9 behaviour. GL is undefined */
2824		if (name == TGSI_SEMANTIC_COLOR && index == 0)
2825			ps_input_cntl |= S_028644_DEFAULT_VAL(3);
2826	}
2827	return ps_input_cntl;
2828}
2829
2830static void si_emit_spi_map(struct si_context *sctx)
2831{
2832	struct si_shader *ps = sctx->ps_shader.current;
2833	struct si_shader *vs = si_get_vs_state(sctx);
2834	struct tgsi_shader_info *psinfo = ps ? &ps->selector->info : NULL;
2835	unsigned i, num_interp, num_written = 0, bcol_interp[2];
2836	unsigned spi_ps_input_cntl[32];
2837
2838	if (!ps || !ps->selector->info.num_inputs)
2839		return;
2840
2841	num_interp = si_get_ps_num_interp(ps);
2842	assert(num_interp > 0);
2843
2844	for (i = 0; i < psinfo->num_inputs; i++) {
2845		unsigned name = psinfo->input_semantic_name[i];
2846		unsigned index = psinfo->input_semantic_index[i];
2847		unsigned interpolate = psinfo->input_interpolate[i];
2848
2849		spi_ps_input_cntl[num_written++] = si_get_ps_input_cntl(sctx, vs, name,
2850							    index, interpolate);
2851
2852		if (name == TGSI_SEMANTIC_COLOR) {
2853			assert(index < ARRAY_SIZE(bcol_interp));
2854			bcol_interp[index] = interpolate;
2855		}
2856	}
2857
2858	if (ps->key.part.ps.prolog.color_two_side) {
2859		unsigned bcol = TGSI_SEMANTIC_BCOLOR;
2860
2861		for (i = 0; i < 2; i++) {
2862			if (!(psinfo->colors_read & (0xf << (i * 4))))
2863				continue;
2864
2865			spi_ps_input_cntl[num_written++] =
2866			  si_get_ps_input_cntl(sctx, vs, bcol, i, bcol_interp[i]);
2867
2868		}
2869	}
2870	assert(num_interp == num_written);
2871
2872	/* R_028644_SPI_PS_INPUT_CNTL_0 */
2873	/* Dota 2: Only ~16% of SPI map updates set different values. */
2874	/* Talos: Only ~9% of SPI map updates set different values. */
2875	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
2876	radeon_opt_set_context_regn(sctx, R_028644_SPI_PS_INPUT_CNTL_0,
2877				    spi_ps_input_cntl,
2878				    sctx->tracked_regs.spi_ps_input_cntl, num_interp);
2879
2880	if (initial_cdw != sctx->gfx_cs->current.cdw)
2881		sctx->context_roll = true;
2882}
2883
2884/**
2885 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
2886 */
2887static void si_init_config_add_vgt_flush(struct si_context *sctx)
2888{
2889	if (sctx->init_config_has_vgt_flush)
2890		return;
2891
2892	/* Done by Vulkan before VGT_FLUSH. */
2893	si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
2894	si_pm4_cmd_add(sctx->init_config,
2895		       EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
2896	si_pm4_cmd_end(sctx->init_config, false);
2897
2898	/* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
2899	si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
2900	si_pm4_cmd_add(sctx->init_config, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
2901	si_pm4_cmd_end(sctx->init_config, false);
2902	sctx->init_config_has_vgt_flush = true;
2903}
2904
2905/* Initialize state related to ESGS / GSVS ring buffers */
2906static bool si_update_gs_ring_buffers(struct si_context *sctx)
2907{
2908	struct si_shader_selector *es =
2909		sctx->tes_shader.cso ? sctx->tes_shader.cso : sctx->vs_shader.cso;
2910	struct si_shader_selector *gs = sctx->gs_shader.cso;
2911	struct si_pm4_state *pm4;
2912
2913	/* Chip constants. */
2914	unsigned num_se = sctx->screen->info.max_se;
2915	unsigned wave_size = 64;
2916	unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
2917	/* On SI-CI, the value comes from VGT_GS_VERTEX_REUSE = 16.
2918	 * On VI+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
2919	 */
2920	unsigned gs_vertex_reuse = (sctx->chip_class >= VI ? 32 : 16) * num_se;
2921	unsigned alignment = 256 * num_se;
2922	/* The maximum size is 63.999 MB per SE. */
2923	unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
2924
2925	/* Calculate the minimum size. */
2926	unsigned min_esgs_ring_size = align(es->esgs_itemsize * gs_vertex_reuse *
2927					    wave_size, alignment);
2928
2929	/* These are recommended sizes, not minimum sizes. */
2930	unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
2931				  es->esgs_itemsize * gs->gs_input_verts_per_prim;
2932	unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
2933				  gs->max_gsvs_emit_size;
2934
2935	min_esgs_ring_size = align(min_esgs_ring_size, alignment);
2936	esgs_ring_size = align(esgs_ring_size, alignment);
2937	gsvs_ring_size = align(gsvs_ring_size, alignment);
2938
2939	esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
2940	gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
2941
2942	/* Some rings don't have to be allocated if shaders don't use them.
2943	 * (e.g. no varyings between ES and GS or GS and VS)
2944	 *
2945	 * GFX9 doesn't have the ESGS ring.
2946	 */
2947	bool update_esgs = sctx->chip_class <= VI &&
2948			   esgs_ring_size &&
2949			   (!sctx->esgs_ring ||
2950			    sctx->esgs_ring->width0 < esgs_ring_size);
2951	bool update_gsvs = gsvs_ring_size &&
2952			   (!sctx->gsvs_ring ||
2953			    sctx->gsvs_ring->width0 < gsvs_ring_size);
2954
2955	if (!update_esgs && !update_gsvs)
2956		return true;
2957
2958	if (update_esgs) {
2959		pipe_resource_reference(&sctx->esgs_ring, NULL);
2960		sctx->esgs_ring =
2961			pipe_aligned_buffer_create(sctx->b.screen,
2962						   SI_RESOURCE_FLAG_UNMAPPABLE,
2963						   PIPE_USAGE_DEFAULT,
2964						   esgs_ring_size, alignment);
2965		if (!sctx->esgs_ring)
2966			return false;
2967	}
2968
2969	if (update_gsvs) {
2970		pipe_resource_reference(&sctx->gsvs_ring, NULL);
2971		sctx->gsvs_ring =
2972			pipe_aligned_buffer_create(sctx->b.screen,
2973						   SI_RESOURCE_FLAG_UNMAPPABLE,
2974						   PIPE_USAGE_DEFAULT,
2975						   gsvs_ring_size, alignment);
2976		if (!sctx->gsvs_ring)
2977			return false;
2978	}
2979
2980	/* Create the "init_config_gs_rings" state. */
2981	pm4 = CALLOC_STRUCT(si_pm4_state);
2982	if (!pm4)
2983		return false;
2984
2985	if (sctx->chip_class >= CIK) {
2986		if (sctx->esgs_ring) {
2987			assert(sctx->chip_class <= VI);
2988			si_pm4_set_reg(pm4, R_030900_VGT_ESGS_RING_SIZE,
2989				       sctx->esgs_ring->width0 / 256);
2990		}
2991		if (sctx->gsvs_ring)
2992			si_pm4_set_reg(pm4, R_030904_VGT_GSVS_RING_SIZE,
2993				       sctx->gsvs_ring->width0 / 256);
2994	} else {
2995		if (sctx->esgs_ring)
2996			si_pm4_set_reg(pm4, R_0088C8_VGT_ESGS_RING_SIZE,
2997				       sctx->esgs_ring->width0 / 256);
2998		if (sctx->gsvs_ring)
2999			si_pm4_set_reg(pm4, R_0088CC_VGT_GSVS_RING_SIZE,
3000				       sctx->gsvs_ring->width0 / 256);
3001	}
3002
3003	/* Set the state. */
3004	if (sctx->init_config_gs_rings)
3005		si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
3006	sctx->init_config_gs_rings = pm4;
3007
3008	if (!sctx->init_config_has_vgt_flush) {
3009		si_init_config_add_vgt_flush(sctx);
3010		si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
3011	}
3012
3013	/* Flush the context to re-emit both init_config states. */
3014	sctx->initial_gfx_cs_size = 0; /* force flush */
3015	si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
3016
3017	/* Set ring bindings. */
3018	if (sctx->esgs_ring) {
3019		assert(sctx->chip_class <= VI);
3020		si_set_ring_buffer(sctx, SI_ES_RING_ESGS,
3021				   sctx->esgs_ring, 0, sctx->esgs_ring->width0,
3022				   true, true, 4, 64, 0);
3023		si_set_ring_buffer(sctx, SI_GS_RING_ESGS,
3024				   sctx->esgs_ring, 0, sctx->esgs_ring->width0,
3025				   false, false, 0, 0, 0);
3026	}
3027	if (sctx->gsvs_ring) {
3028		si_set_ring_buffer(sctx, SI_RING_GSVS,
3029				   sctx->gsvs_ring, 0, sctx->gsvs_ring->width0,
3030				   false, false, 0, 0, 0);
3031	}
3032
3033	return true;
3034}
3035
3036static void si_shader_lock(struct si_shader *shader)
3037{
3038	mtx_lock(&shader->selector->mutex);
3039	if (shader->previous_stage_sel) {
3040		assert(shader->previous_stage_sel != shader->selector);
3041		mtx_lock(&shader->previous_stage_sel->mutex);
3042	}
3043}
3044
3045static void si_shader_unlock(struct si_shader *shader)
3046{
3047	if (shader->previous_stage_sel)
3048		mtx_unlock(&shader->previous_stage_sel->mutex);
3049	mtx_unlock(&shader->selector->mutex);
3050}
3051
3052/**
3053 * @returns 1 if \p sel has been updated to use a new scratch buffer
3054 *          0 if not
3055 *          < 0 if there was a failure
3056 */
3057static int si_update_scratch_buffer(struct si_context *sctx,
3058				    struct si_shader *shader)
3059{
3060	uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
3061	int r;
3062
3063	if (!shader)
3064		return 0;
3065
3066	/* This shader doesn't need a scratch buffer */
3067	if (shader->config.scratch_bytes_per_wave == 0)
3068		return 0;
3069
3070	/* Prevent race conditions when updating:
3071	 * - si_shader::scratch_bo
3072	 * - si_shader::binary::code
3073	 * - si_shader::previous_stage::binary::code.
3074	 */
3075	si_shader_lock(shader);
3076
3077	/* This shader is already configured to use the current
3078	 * scratch buffer. */
3079	if (shader->scratch_bo == sctx->scratch_buffer) {
3080		si_shader_unlock(shader);
3081		return 0;
3082	}
3083
3084	assert(sctx->scratch_buffer);
3085
3086	if (shader->previous_stage)
3087		si_shader_apply_scratch_relocs(shader->previous_stage, scratch_va);
3088
3089	si_shader_apply_scratch_relocs(shader, scratch_va);
3090
3091	/* Replace the shader bo with a new bo that has the relocs applied. */
3092	r = si_shader_binary_upload(sctx->screen, shader);
3093	if (r) {
3094		si_shader_unlock(shader);
3095		return r;
3096	}
3097
3098	/* Update the shader state to use the new shader bo. */
3099	si_shader_init_pm4_state(sctx->screen, shader);
3100
3101	si_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
3102
3103	si_shader_unlock(shader);
3104	return 1;
3105}
3106
3107static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader *shader)
3108{
3109	return shader ? shader->config.scratch_bytes_per_wave : 0;
3110}
3111
3112static struct si_shader *si_get_tcs_current(struct si_context *sctx)
3113{
3114	if (!sctx->tes_shader.cso)
3115		return NULL; /* tessellation disabled */
3116
3117	return sctx->tcs_shader.cso ? sctx->tcs_shader.current :
3118				      sctx->fixed_func_tcs_shader.current;
3119}
3120
3121static bool si_update_scratch_relocs(struct si_context *sctx)
3122{
3123	struct si_shader *tcs = si_get_tcs_current(sctx);
3124	int r;
3125
3126	/* Update the shaders, so that they are using the latest scratch.
3127	 * The scratch buffer may have been changed since these shaders were
3128	 * last used, so we still need to try to update them, even if they
3129	 * require scratch buffers smaller than the current size.
3130	 */
3131	r = si_update_scratch_buffer(sctx, sctx->ps_shader.current);
3132	if (r < 0)
3133		return false;
3134	if (r == 1)
3135		si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
3136
3137	r = si_update_scratch_buffer(sctx, sctx->gs_shader.current);
3138	if (r < 0)
3139		return false;
3140	if (r == 1)
3141		si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
3142
3143	r = si_update_scratch_buffer(sctx, tcs);
3144	if (r < 0)
3145		return false;
3146	if (r == 1)
3147		si_pm4_bind_state(sctx, hs, tcs->pm4);
3148
3149	/* VS can be bound as LS, ES, or VS. */
3150	r = si_update_scratch_buffer(sctx, sctx->vs_shader.current);
3151	if (r < 0)
3152		return false;
3153	if (r == 1) {
3154		if (sctx->tes_shader.current)
3155			si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
3156		else if (sctx->gs_shader.current)
3157			si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
3158		else
3159			si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
3160	}
3161
3162	/* TES can be bound as ES or VS. */
3163	r = si_update_scratch_buffer(sctx, sctx->tes_shader.current);
3164	if (r < 0)
3165		return false;
3166	if (r == 1) {
3167		if (sctx->gs_shader.current)
3168			si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
3169		else
3170			si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
3171	}
3172
3173	return true;
3174}
3175
3176static bool si_update_spi_tmpring_size(struct si_context *sctx)
3177{
3178	/* SPI_TMPRING_SIZE.WAVESIZE must be constant for each scratch buffer.
3179	 * There are 2 cases to handle:
3180	 *
3181	 * - If the current needed size is less than the maximum seen size,
3182	 *   use the maximum seen size, so that WAVESIZE remains the same.
3183	 *
3184	 * - If the current needed size is greater than the maximum seen size,
3185	 *   the scratch buffer is reallocated, so we can increase WAVESIZE.
3186	 *
3187	 * Shaders that set SCRATCH_EN=0 don't allocate scratch space.
3188	 * Otherwise, the number of waves that can use scratch is
3189	 * SPI_TMPRING_SIZE.WAVES.
3190	 */
3191	unsigned bytes = 0;
3192
3193	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->ps_shader.current));
3194	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->gs_shader.current));
3195	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->vs_shader.current));
3196
3197	if (sctx->tes_shader.cso) {
3198		bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tes_shader.current));
3199		bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(si_get_tcs_current(sctx)));
3200	}
3201
3202	sctx->max_seen_scratch_bytes_per_wave =
3203		MAX2(sctx->max_seen_scratch_bytes_per_wave, bytes);
3204
3205	unsigned scratch_needed_size =
3206		sctx->max_seen_scratch_bytes_per_wave * sctx->scratch_waves;
3207	unsigned spi_tmpring_size;
3208
3209	if (scratch_needed_size > 0) {
3210		if (!sctx->scratch_buffer ||
3211		    scratch_needed_size > sctx->scratch_buffer->b.b.width0) {
3212			/* Create a bigger scratch buffer */
3213			si_resource_reference(&sctx->scratch_buffer, NULL);
3214
3215			sctx->scratch_buffer =
3216				si_aligned_buffer_create(&sctx->screen->b,
3217							   SI_RESOURCE_FLAG_UNMAPPABLE,
3218							   PIPE_USAGE_DEFAULT,
3219							   scratch_needed_size, 256);
3220			if (!sctx->scratch_buffer)
3221				return false;
3222
3223			si_mark_atom_dirty(sctx, &sctx->atoms.s.scratch_state);
3224			si_context_add_resource_size(sctx,
3225						     &sctx->scratch_buffer->b.b);
3226		}
3227
3228		if (!si_update_scratch_relocs(sctx))
3229			return false;
3230	}
3231
3232	/* The LLVM shader backend should be reporting aligned scratch_sizes. */
3233	assert((scratch_needed_size & ~0x3FF) == scratch_needed_size &&
3234		"scratch size should already be aligned correctly.");
3235
3236	spi_tmpring_size = S_0286E8_WAVES(sctx->scratch_waves) |
3237			   S_0286E8_WAVESIZE(sctx->max_seen_scratch_bytes_per_wave >> 10);
3238	if (spi_tmpring_size != sctx->spi_tmpring_size) {
3239		sctx->spi_tmpring_size = spi_tmpring_size;
3240		si_mark_atom_dirty(sctx, &sctx->atoms.s.scratch_state);
3241	}
3242	return true;
3243}
3244
3245static void si_init_tess_factor_ring(struct si_context *sctx)
3246{
3247	assert(!sctx->tess_rings);
3248
3249	/* The address must be aligned to 2^19, because the shader only
3250	 * receives the high 13 bits.
3251	 */
3252	sctx->tess_rings = pipe_aligned_buffer_create(sctx->b.screen,
3253						    SI_RESOURCE_FLAG_32BIT,
3254						    PIPE_USAGE_DEFAULT,
3255						    sctx->screen->tess_offchip_ring_size +
3256						    sctx->screen->tess_factor_ring_size,
3257						    1 << 19);
3258	if (!sctx->tess_rings)
3259		return;
3260
3261	si_init_config_add_vgt_flush(sctx);
3262
3263	si_pm4_add_bo(sctx->init_config, si_resource(sctx->tess_rings),
3264		      RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
3265
3266	uint64_t factor_va = si_resource(sctx->tess_rings)->gpu_address +
3267			     sctx->screen->tess_offchip_ring_size;
3268
3269	/* Append these registers to the init config state. */
3270	if (sctx->chip_class >= CIK) {
3271		si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
3272			       S_030938_SIZE(sctx->screen->tess_factor_ring_size / 4));
3273		si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
3274			       factor_va >> 8);
3275		if (sctx->chip_class >= GFX9)
3276			si_pm4_set_reg(sctx->init_config, R_030944_VGT_TF_MEMORY_BASE_HI,
3277				       S_030944_BASE_HI(factor_va >> 40));
3278		si_pm4_set_reg(sctx->init_config, R_03093C_VGT_HS_OFFCHIP_PARAM,
3279			       sctx->screen->vgt_hs_offchip_param);
3280	} else {
3281		si_pm4_set_reg(sctx->init_config, R_008988_VGT_TF_RING_SIZE,
3282			       S_008988_SIZE(sctx->screen->tess_factor_ring_size / 4));
3283		si_pm4_set_reg(sctx->init_config, R_0089B8_VGT_TF_MEMORY_BASE,
3284			       factor_va >> 8);
3285		si_pm4_set_reg(sctx->init_config, R_0089B0_VGT_HS_OFFCHIP_PARAM,
3286		               sctx->screen->vgt_hs_offchip_param);
3287	}
3288
3289	/* Flush the context to re-emit the init_config state.
3290	 * This is done only once in a lifetime of a context.
3291	 */
3292	si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
3293	sctx->initial_gfx_cs_size = 0; /* force flush */
3294	si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
3295}
3296
3297static void si_update_vgt_shader_config(struct si_context *sctx)
3298{
3299	/* Calculate the index of the config.
3300	 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
3301	unsigned index = 2*!!sctx->tes_shader.cso + !!sctx->gs_shader.cso;
3302	struct si_pm4_state **pm4 = &sctx->vgt_shader_config[index];
3303
3304	if (!*pm4) {
3305		uint32_t stages = 0;
3306
3307		*pm4 = CALLOC_STRUCT(si_pm4_state);
3308
3309		if (sctx->tes_shader.cso) {
3310			stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
3311				  S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3312
3313			if (sctx->gs_shader.cso)
3314				stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
3315					  S_028B54_GS_EN(1) |
3316				          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3317			else
3318				stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
3319		} else if (sctx->gs_shader.cso) {
3320			stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
3321				  S_028B54_GS_EN(1) |
3322			          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
3323		}
3324
3325		if (sctx->chip_class >= GFX9)
3326			stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3327
3328		si_pm4_set_reg(*pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
3329	}
3330	si_pm4_bind_state(sctx, vgt_shader_config, *pm4);
3331}
3332
3333bool si_update_shaders(struct si_context *sctx)
3334{
3335	struct pipe_context *ctx = (struct pipe_context*)sctx;
3336	struct si_compiler_ctx_state compiler_state;
3337	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3338	struct si_shader *old_vs = si_get_vs_state(sctx);
3339	bool old_clip_disable = old_vs ? old_vs->key.opt.clip_disable : false;
3340	struct si_shader *old_ps = sctx->ps_shader.current;
3341	unsigned old_spi_shader_col_format =
3342		old_ps ? old_ps->key.part.ps.epilog.spi_shader_col_format : 0;
3343	int r;
3344
3345	compiler_state.compiler = &sctx->compiler;
3346	compiler_state.debug = sctx->debug;
3347	compiler_state.is_debug_context = sctx->is_debug;
3348
3349	/* Update stages before GS. */
3350	if (sctx->tes_shader.cso) {
3351		if (!sctx->tess_rings) {
3352			si_init_tess_factor_ring(sctx);
3353			if (!sctx->tess_rings)
3354				return false;
3355		}
3356
3357		/* VS as LS */
3358		if (sctx->chip_class <= VI) {
3359			r = si_shader_select(ctx, &sctx->vs_shader,
3360					     &compiler_state);
3361			if (r)
3362				return false;
3363			si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
3364		}
3365
3366		if (sctx->tcs_shader.cso) {
3367			r = si_shader_select(ctx, &sctx->tcs_shader,
3368					     &compiler_state);
3369			if (r)
3370				return false;
3371			si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
3372		} else {
3373			if (!sctx->fixed_func_tcs_shader.cso) {
3374				sctx->fixed_func_tcs_shader.cso =
3375					si_create_fixed_func_tcs(sctx);
3376				if (!sctx->fixed_func_tcs_shader.cso)
3377					return false;
3378			}
3379
3380			r = si_shader_select(ctx, &sctx->fixed_func_tcs_shader,
3381					     &compiler_state);
3382			if (r)
3383				return false;
3384			si_pm4_bind_state(sctx, hs,
3385					  sctx->fixed_func_tcs_shader.current->pm4);
3386		}
3387
3388		if (sctx->gs_shader.cso) {
3389			/* TES as ES */
3390			if (sctx->chip_class <= VI) {
3391				r = si_shader_select(ctx, &sctx->tes_shader,
3392						     &compiler_state);
3393				if (r)
3394					return false;
3395				si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
3396			}
3397		} else {
3398			/* TES as VS */
3399			r = si_shader_select(ctx, &sctx->tes_shader,
3400					     &compiler_state);
3401			if (r)
3402				return false;
3403			si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
3404		}
3405	} else if (sctx->gs_shader.cso) {
3406		if (sctx->chip_class <= VI) {
3407			/* VS as ES */
3408			r = si_shader_select(ctx, &sctx->vs_shader,
3409					     &compiler_state);
3410			if (r)
3411				return false;
3412			si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
3413
3414			si_pm4_bind_state(sctx, ls, NULL);
3415			si_pm4_bind_state(sctx, hs, NULL);
3416		}
3417	} else {
3418		/* VS as VS */
3419		r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state);
3420		if (r)
3421			return false;
3422		si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
3423		si_pm4_bind_state(sctx, ls, NULL);
3424		si_pm4_bind_state(sctx, hs, NULL);
3425	}
3426
3427	/* Update GS. */
3428	if (sctx->gs_shader.cso) {
3429		r = si_shader_select(ctx, &sctx->gs_shader, &compiler_state);
3430		if (r)
3431			return false;
3432		si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
3433		si_pm4_bind_state(sctx, vs, sctx->gs_shader.cso->gs_copy_shader->pm4);
3434
3435		if (!si_update_gs_ring_buffers(sctx))
3436			return false;
3437	} else {
3438		si_pm4_bind_state(sctx, gs, NULL);
3439		if (sctx->chip_class <= VI)
3440			si_pm4_bind_state(sctx, es, NULL);
3441	}
3442
3443	si_update_vgt_shader_config(sctx);
3444
3445	if (old_clip_disable != si_get_vs_state(sctx)->key.opt.clip_disable)
3446		si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
3447
3448	if (sctx->ps_shader.cso) {
3449		unsigned db_shader_control;
3450
3451		r = si_shader_select(ctx, &sctx->ps_shader, &compiler_state);
3452		if (r)
3453			return false;
3454		si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
3455
3456		db_shader_control =
3457			sctx->ps_shader.cso->db_shader_control |
3458			S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS);
3459
3460		if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
3461		    sctx->sprite_coord_enable != rs->sprite_coord_enable ||
3462		    sctx->flatshade != rs->flatshade) {
3463			sctx->sprite_coord_enable = rs->sprite_coord_enable;
3464			sctx->flatshade = rs->flatshade;
3465			si_mark_atom_dirty(sctx, &sctx->atoms.s.spi_map);
3466		}
3467
3468		if (sctx->screen->rbplus_allowed &&
3469		    si_pm4_state_changed(sctx, ps) &&
3470		    (!old_ps ||
3471		     old_spi_shader_col_format !=
3472		     sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format))
3473			si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
3474
3475		if (sctx->ps_db_shader_control != db_shader_control) {
3476			sctx->ps_db_shader_control = db_shader_control;
3477			si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
3478			if (sctx->screen->dpbb_allowed)
3479				si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3480		}
3481
3482		if (sctx->smoothing_enabled != sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing) {
3483			sctx->smoothing_enabled = sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing;
3484			si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3485
3486			if (sctx->chip_class == SI)
3487				si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
3488
3489			if (sctx->framebuffer.nr_samples <= 1)
3490				si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
3491		}
3492	}
3493
3494	if (si_pm4_state_enabled_and_changed(sctx, ls) ||
3495	    si_pm4_state_enabled_and_changed(sctx, hs) ||
3496	    si_pm4_state_enabled_and_changed(sctx, es) ||
3497	    si_pm4_state_enabled_and_changed(sctx, gs) ||
3498	    si_pm4_state_enabled_and_changed(sctx, vs) ||
3499	    si_pm4_state_enabled_and_changed(sctx, ps)) {
3500		if (!si_update_spi_tmpring_size(sctx))
3501			return false;
3502	}
3503
3504	if (sctx->chip_class >= CIK) {
3505		if (si_pm4_state_enabled_and_changed(sctx, ls))
3506			sctx->prefetch_L2_mask |= SI_PREFETCH_LS;
3507		else if (!sctx->queued.named.ls)
3508			sctx->prefetch_L2_mask &= ~SI_PREFETCH_LS;
3509
3510		if (si_pm4_state_enabled_and_changed(sctx, hs))
3511			sctx->prefetch_L2_mask |= SI_PREFETCH_HS;
3512		else if (!sctx->queued.named.hs)
3513			sctx->prefetch_L2_mask &= ~SI_PREFETCH_HS;
3514
3515		if (si_pm4_state_enabled_and_changed(sctx, es))
3516			sctx->prefetch_L2_mask |= SI_PREFETCH_ES;
3517		else if (!sctx->queued.named.es)
3518			sctx->prefetch_L2_mask &= ~SI_PREFETCH_ES;
3519
3520		if (si_pm4_state_enabled_and_changed(sctx, gs))
3521			sctx->prefetch_L2_mask |= SI_PREFETCH_GS;
3522		else if (!sctx->queued.named.gs)
3523			sctx->prefetch_L2_mask &= ~SI_PREFETCH_GS;
3524
3525		if (si_pm4_state_enabled_and_changed(sctx, vs))
3526			sctx->prefetch_L2_mask |= SI_PREFETCH_VS;
3527		else if (!sctx->queued.named.vs)
3528			sctx->prefetch_L2_mask &= ~SI_PREFETCH_VS;
3529
3530		if (si_pm4_state_enabled_and_changed(sctx, ps))
3531			sctx->prefetch_L2_mask |= SI_PREFETCH_PS;
3532		else if (!sctx->queued.named.ps)
3533			sctx->prefetch_L2_mask &= ~SI_PREFETCH_PS;
3534	}
3535
3536	sctx->do_update_shaders = false;
3537	return true;
3538}
3539
3540static void si_emit_scratch_state(struct si_context *sctx)
3541{
3542	struct radeon_cmdbuf *cs = sctx->gfx_cs;
3543
3544	radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
3545			       sctx->spi_tmpring_size);
3546
3547	if (sctx->scratch_buffer) {
3548		radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3549				      sctx->scratch_buffer, RADEON_USAGE_READWRITE,
3550				      RADEON_PRIO_SCRATCH_BUFFER);
3551	}
3552}
3553
3554void si_init_shader_functions(struct si_context *sctx)
3555{
3556	sctx->atoms.s.spi_map.emit = si_emit_spi_map;
3557	sctx->atoms.s.scratch_state.emit = si_emit_scratch_state;
3558
3559	sctx->b.create_vs_state = si_create_shader_selector;
3560	sctx->b.create_tcs_state = si_create_shader_selector;
3561	sctx->b.create_tes_state = si_create_shader_selector;
3562	sctx->b.create_gs_state = si_create_shader_selector;
3563	sctx->b.create_fs_state = si_create_shader_selector;
3564
3565	sctx->b.bind_vs_state = si_bind_vs_shader;
3566	sctx->b.bind_tcs_state = si_bind_tcs_shader;
3567	sctx->b.bind_tes_state = si_bind_tes_shader;
3568	sctx->b.bind_gs_state = si_bind_gs_shader;
3569	sctx->b.bind_fs_state = si_bind_ps_shader;
3570
3571	sctx->b.delete_vs_state = si_delete_shader_selector;
3572	sctx->b.delete_tcs_state = si_delete_shader_selector;
3573	sctx->b.delete_tes_state = si_delete_shader_selector;
3574	sctx->b.delete_gs_state = si_delete_shader_selector;
3575	sctx->b.delete_fs_state = si_delete_shader_selector;
3576}
3577