36dbfe9df6f108925ee8524f47b71a4068017ca5
[mesa.git] / src / gallium / drivers / radeonsi / si_state_shaders.c
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 "sid.h"
27
28 #include "compiler/nir/nir_serialize.h"
29 #include "nir/tgsi_to_nir.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 key for the shader cache.
45 */
46 void si_get_ir_cache_key(struct si_shader_selector *sel, bool ngg, bool es,
47 unsigned char ir_sha1_cache_key[20])
48 {
49 struct blob blob = {};
50 unsigned ir_size;
51 void *ir_binary;
52
53 if (sel->nir_binary) {
54 ir_binary = sel->nir_binary;
55 ir_size = sel->nir_size;
56 } else {
57 assert(sel->nir);
58
59 blob_init(&blob);
60 nir_serialize(&blob, sel->nir, true);
61 ir_binary = blob.data;
62 ir_size = blob.size;
63 }
64
65 /* These settings affect the compilation, but they are not derived
66 * from the input shader IR.
67 */
68 unsigned shader_variant_flags = 0;
69
70 if (ngg)
71 shader_variant_flags |= 1 << 0;
72 if (sel->nir)
73 shader_variant_flags |= 1 << 1;
74 if (si_get_wave_size(sel->screen, sel->type, ngg, es) == 32)
75 shader_variant_flags |= 1 << 2;
76 if (sel->force_correct_derivs_after_kill)
77 shader_variant_flags |= 1 << 3;
78
79 struct mesa_sha1 ctx;
80 _mesa_sha1_init(&ctx);
81 _mesa_sha1_update(&ctx, &shader_variant_flags, 4);
82 _mesa_sha1_update(&ctx, ir_binary, ir_size);
83 if (sel->type == PIPE_SHADER_VERTEX ||
84 sel->type == PIPE_SHADER_TESS_EVAL ||
85 sel->type == PIPE_SHADER_GEOMETRY)
86 _mesa_sha1_update(&ctx, &sel->so, sizeof(sel->so));
87 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
88
89 if (ir_binary == blob.data)
90 blob_finish(&blob);
91 }
92
93 /** Copy "data" to "ptr" and return the next dword following copied data. */
94 static uint32_t *write_data(uint32_t *ptr, const void *data, unsigned size)
95 {
96 /* data may be NULL if size == 0 */
97 if (size)
98 memcpy(ptr, data, size);
99 ptr += DIV_ROUND_UP(size, 4);
100 return ptr;
101 }
102
103 /** Read data from "ptr". Return the next dword following the data. */
104 static uint32_t *read_data(uint32_t *ptr, void *data, unsigned size)
105 {
106 memcpy(data, ptr, size);
107 ptr += DIV_ROUND_UP(size, 4);
108 return ptr;
109 }
110
111 /**
112 * Write the size as uint followed by the data. Return the next dword
113 * following the copied data.
114 */
115 static uint32_t *write_chunk(uint32_t *ptr, const void *data, unsigned size)
116 {
117 *ptr++ = size;
118 return write_data(ptr, data, size);
119 }
120
121 /**
122 * Read the size as uint followed by the data. Return both via parameters.
123 * Return the next dword following the data.
124 */
125 static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
126 {
127 *size = *ptr++;
128 assert(*data == NULL);
129 if (!*size)
130 return ptr;
131 *data = malloc(*size);
132 return read_data(ptr, *data, *size);
133 }
134
135 /**
136 * Return the shader binary in a buffer. The first 4 bytes contain its size
137 * as integer.
138 */
139 static void *si_get_shader_binary(struct si_shader *shader)
140 {
141 /* There is always a size of data followed by the data itself. */
142 unsigned llvm_ir_size = shader->binary.llvm_ir_string ?
143 strlen(shader->binary.llvm_ir_string) + 1 : 0;
144
145 /* Refuse to allocate overly large buffers and guard against integer
146 * overflow. */
147 if (shader->binary.elf_size > UINT_MAX / 4 ||
148 llvm_ir_size > UINT_MAX / 4)
149 return NULL;
150
151 unsigned size =
152 4 + /* total size */
153 4 + /* CRC32 of the data below */
154 align(sizeof(shader->config), 4) +
155 align(sizeof(shader->info), 4) +
156 4 + align(shader->binary.elf_size, 4) +
157 4 + align(llvm_ir_size, 4);
158 void *buffer = CALLOC(1, size);
159 uint32_t *ptr = (uint32_t*)buffer;
160
161 if (!buffer)
162 return NULL;
163
164 *ptr++ = size;
165 ptr++; /* CRC32 is calculated at the end. */
166
167 ptr = write_data(ptr, &shader->config, sizeof(shader->config));
168 ptr = write_data(ptr, &shader->info, sizeof(shader->info));
169 ptr = write_chunk(ptr, shader->binary.elf_buffer, shader->binary.elf_size);
170 ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
171 assert((char *)ptr - (char *)buffer == size);
172
173 /* Compute CRC32. */
174 ptr = (uint32_t*)buffer;
175 ptr++;
176 *ptr = util_hash_crc32(ptr + 1, size - 8);
177
178 return buffer;
179 }
180
181 static bool si_load_shader_binary(struct si_shader *shader, void *binary)
182 {
183 uint32_t *ptr = (uint32_t*)binary;
184 uint32_t size = *ptr++;
185 uint32_t crc32 = *ptr++;
186 unsigned chunk_size;
187 unsigned elf_size;
188
189 if (util_hash_crc32(ptr, size - 8) != crc32) {
190 fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
191 return false;
192 }
193
194 ptr = read_data(ptr, &shader->config, sizeof(shader->config));
195 ptr = read_data(ptr, &shader->info, sizeof(shader->info));
196 ptr = read_chunk(ptr, (void**)&shader->binary.elf_buffer,
197 &elf_size);
198 shader->binary.elf_size = elf_size;
199 ptr = read_chunk(ptr, (void**)&shader->binary.llvm_ir_string, &chunk_size);
200
201 return true;
202 }
203
204 /**
205 * Insert a shader into the cache. It's assumed the shader is not in the cache.
206 * Use si_shader_cache_load_shader before calling this.
207 */
208 void si_shader_cache_insert_shader(struct si_screen *sscreen,
209 unsigned char ir_sha1_cache_key[20],
210 struct si_shader *shader,
211 bool insert_into_disk_cache)
212 {
213 void *hw_binary;
214 struct hash_entry *entry;
215 uint8_t key[CACHE_KEY_SIZE];
216
217 entry = _mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
218 if (entry)
219 return; /* already added */
220
221 hw_binary = si_get_shader_binary(shader);
222 if (!hw_binary)
223 return;
224
225 if (_mesa_hash_table_insert(sscreen->shader_cache,
226 mem_dup(ir_sha1_cache_key, 20),
227 hw_binary) == NULL) {
228 FREE(hw_binary);
229 return;
230 }
231
232 if (sscreen->disk_shader_cache && insert_into_disk_cache) {
233 disk_cache_compute_key(sscreen->disk_shader_cache,
234 ir_sha1_cache_key, 20, key);
235 disk_cache_put(sscreen->disk_shader_cache, key, hw_binary,
236 *((uint32_t *) hw_binary), NULL);
237 }
238 }
239
240 bool si_shader_cache_load_shader(struct si_screen *sscreen,
241 unsigned char ir_sha1_cache_key[20],
242 struct si_shader *shader)
243 {
244 struct hash_entry *entry =
245 _mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
246 if (!entry) {
247 if (sscreen->disk_shader_cache) {
248 unsigned char sha1[CACHE_KEY_SIZE];
249
250 disk_cache_compute_key(sscreen->disk_shader_cache,
251 ir_sha1_cache_key, 20, sha1);
252
253 size_t binary_size;
254 uint8_t *buffer =
255 disk_cache_get(sscreen->disk_shader_cache,
256 sha1, &binary_size);
257 if (!buffer)
258 return false;
259
260 if (binary_size < sizeof(uint32_t) ||
261 *((uint32_t*)buffer) != binary_size) {
262 /* Something has gone wrong discard the item
263 * from the cache and rebuild/link from
264 * source.
265 */
266 assert(!"Invalid radeonsi shader disk cache "
267 "item!");
268
269 disk_cache_remove(sscreen->disk_shader_cache,
270 sha1);
271 free(buffer);
272
273 return false;
274 }
275
276 if (!si_load_shader_binary(shader, buffer)) {
277 free(buffer);
278 return false;
279 }
280 free(buffer);
281
282 si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key,
283 shader, false);
284 } else {
285 return false;
286 }
287 } else {
288 if (!si_load_shader_binary(shader, entry->data))
289 return false;
290 }
291 p_atomic_inc(&sscreen->num_shader_cache_hits);
292 return true;
293 }
294
295 static uint32_t si_shader_cache_key_hash(const void *key)
296 {
297 /* Take the first dword of SHA1. */
298 return *(uint32_t*)key;
299 }
300
301 static bool si_shader_cache_key_equals(const void *a, const void *b)
302 {
303 /* Compare SHA1s. */
304 return memcmp(a, b, 20) == 0;
305 }
306
307 static void si_destroy_shader_cache_entry(struct hash_entry *entry)
308 {
309 FREE((void*)entry->key);
310 FREE(entry->data);
311 }
312
313 bool si_init_shader_cache(struct si_screen *sscreen)
314 {
315 (void) simple_mtx_init(&sscreen->shader_cache_mutex, mtx_plain);
316 sscreen->shader_cache =
317 _mesa_hash_table_create(NULL,
318 si_shader_cache_key_hash,
319 si_shader_cache_key_equals);
320
321 return sscreen->shader_cache != NULL;
322 }
323
324 void si_destroy_shader_cache(struct si_screen *sscreen)
325 {
326 if (sscreen->shader_cache)
327 _mesa_hash_table_destroy(sscreen->shader_cache,
328 si_destroy_shader_cache_entry);
329 simple_mtx_destroy(&sscreen->shader_cache_mutex);
330 }
331
332 /* SHADER STATES */
333
334 static void si_set_tesseval_regs(struct si_screen *sscreen,
335 const struct si_shader_selector *tes,
336 struct si_pm4_state *pm4)
337 {
338 const struct si_shader_info *info = &tes->info;
339 unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
340 unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
341 bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
342 bool tes_point_mode = info->properties[TGSI_PROPERTY_TES_POINT_MODE];
343 unsigned type, partitioning, topology, distribution_mode;
344
345 switch (tes_prim_mode) {
346 case PIPE_PRIM_LINES:
347 type = V_028B6C_TESS_ISOLINE;
348 break;
349 case PIPE_PRIM_TRIANGLES:
350 type = V_028B6C_TESS_TRIANGLE;
351 break;
352 case PIPE_PRIM_QUADS:
353 type = V_028B6C_TESS_QUAD;
354 break;
355 default:
356 assert(0);
357 return;
358 }
359
360 switch (tes_spacing) {
361 case PIPE_TESS_SPACING_FRACTIONAL_ODD:
362 partitioning = V_028B6C_PART_FRAC_ODD;
363 break;
364 case PIPE_TESS_SPACING_FRACTIONAL_EVEN:
365 partitioning = V_028B6C_PART_FRAC_EVEN;
366 break;
367 case PIPE_TESS_SPACING_EQUAL:
368 partitioning = V_028B6C_PART_INTEGER;
369 break;
370 default:
371 assert(0);
372 return;
373 }
374
375 if (tes_point_mode)
376 topology = V_028B6C_OUTPUT_POINT;
377 else if (tes_prim_mode == PIPE_PRIM_LINES)
378 topology = V_028B6C_OUTPUT_LINE;
379 else if (tes_vertex_order_cw)
380 /* for some reason, this must be the other way around */
381 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
382 else
383 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
384
385 if (sscreen->info.has_distributed_tess) {
386 if (sscreen->info.family == CHIP_FIJI ||
387 sscreen->info.family >= CHIP_POLARIS10)
388 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
389 else
390 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
391 } else
392 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
393
394 assert(pm4->shader);
395 pm4->shader->vgt_tf_param = S_028B6C_TYPE(type) |
396 S_028B6C_PARTITIONING(partitioning) |
397 S_028B6C_TOPOLOGY(topology) |
398 S_028B6C_DISTRIBUTION_MODE(distribution_mode);
399 }
400
401 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
402 * whether the "fractional odd" tessellation spacing is used.
403 *
404 * Possible VGT configurations and which state should set the register:
405 *
406 * Reg set in | VGT shader configuration | Value
407 * ------------------------------------------------------
408 * VS as VS | VS | 30
409 * VS as ES | ES -> GS -> VS | 30
410 * TES as VS | LS -> HS -> VS | 14 or 30
411 * TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
412 *
413 * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
414 */
415 static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen,
416 struct si_shader_selector *sel,
417 struct si_shader *shader,
418 struct si_pm4_state *pm4)
419 {
420 unsigned type = sel->type;
421
422 if (sscreen->info.family < CHIP_POLARIS10 ||
423 sscreen->info.chip_class >= GFX10)
424 return;
425
426 /* VS as VS, or VS as ES: */
427 if ((type == PIPE_SHADER_VERTEX &&
428 (!shader ||
429 (!shader->key.as_ls && !shader->is_gs_copy_shader))) ||
430 /* TES as VS, or TES as ES: */
431 type == PIPE_SHADER_TESS_EVAL) {
432 unsigned vtx_reuse_depth = 30;
433
434 if (type == PIPE_SHADER_TESS_EVAL &&
435 sel->info.properties[TGSI_PROPERTY_TES_SPACING] ==
436 PIPE_TESS_SPACING_FRACTIONAL_ODD)
437 vtx_reuse_depth = 14;
438
439 assert(pm4->shader);
440 pm4->shader->vgt_vertex_reuse_block_cntl = vtx_reuse_depth;
441 }
442 }
443
444 static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
445 {
446 if (shader->pm4)
447 si_pm4_clear_state(shader->pm4);
448 else
449 shader->pm4 = CALLOC_STRUCT(si_pm4_state);
450
451 if (shader->pm4) {
452 shader->pm4->shader = shader;
453 return shader->pm4;
454 } else {
455 fprintf(stderr, "radeonsi: Failed to create pm4 state.\n");
456 return NULL;
457 }
458 }
459
460 static unsigned si_get_num_vs_user_sgprs(struct si_shader *shader,
461 unsigned num_always_on_user_sgprs)
462 {
463 struct si_shader_selector *vs = shader->previous_stage_sel ?
464 shader->previous_stage_sel : shader->selector;
465 unsigned num_vbos_in_user_sgprs = vs->num_vbos_in_user_sgprs;
466
467 /* 1 SGPR is reserved for the vertex buffer pointer. */
468 assert(num_always_on_user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST - 1);
469
470 if (num_vbos_in_user_sgprs)
471 return SI_SGPR_VS_VB_DESCRIPTOR_FIRST + num_vbos_in_user_sgprs * 4;
472
473 /* Add the pointer to VBO descriptors. */
474 return num_always_on_user_sgprs + 1;
475 }
476
477 /* Return VGPR_COMP_CNT for the API vertex shader. This can be hw LS, LSHS, ES, ESGS, VS. */
478 static unsigned si_get_vs_vgpr_comp_cnt(struct si_screen *sscreen,
479 struct si_shader *shader, bool legacy_vs_prim_id)
480 {
481 assert(shader->selector->type == PIPE_SHADER_VERTEX ||
482 (shader->previous_stage_sel &&
483 shader->previous_stage_sel->type == PIPE_SHADER_VERTEX));
484
485 /* GFX6-9 LS (VertexID, RelAutoindex, InstanceID / StepRate0(==1), ...).
486 * GFX6-9 ES,VS (VertexID, InstanceID / StepRate0(==1), VSPrimID, ...)
487 * GFX10 LS (VertexID, RelAutoindex, UserVGPR1, InstanceID).
488 * GFX10 ES,VS (VertexID, UserVGPR0, UserVGPR1 or VSPrimID, UserVGPR2 or InstanceID)
489 */
490 bool is_ls = shader->selector->type == PIPE_SHADER_TESS_CTRL || shader->key.as_ls;
491
492 if (sscreen->info.chip_class >= GFX10 && shader->info.uses_instanceid)
493 return 3;
494 else if ((is_ls && shader->info.uses_instanceid) || legacy_vs_prim_id)
495 return 2;
496 else if (is_ls || shader->info.uses_instanceid)
497 return 1;
498 else
499 return 0;
500 }
501
502 static void si_shader_ls(struct si_screen *sscreen, struct si_shader *shader)
503 {
504 struct si_pm4_state *pm4;
505 uint64_t va;
506
507 assert(sscreen->info.chip_class <= GFX8);
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 si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
517 si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, S_00B524_MEM_BASE(va >> 40));
518
519 shader->config.rsrc1 = S_00B528_VGPRS((shader->config.num_vgprs - 1) / 4) |
520 S_00B528_SGPRS((shader->config.num_sgprs - 1) / 8) |
521 S_00B528_VGPR_COMP_CNT(si_get_vs_vgpr_comp_cnt(sscreen, shader, false)) |
522 S_00B528_DX10_CLAMP(1) |
523 S_00B528_FLOAT_MODE(shader->config.float_mode);
524 shader->config.rsrc2 = S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR)) |
525 S_00B52C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
526 }
527
528 static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
529 {
530 struct si_pm4_state *pm4;
531 uint64_t va;
532
533 pm4 = si_get_shader_pm4_state(shader);
534 if (!pm4)
535 return;
536
537 va = shader->bo->gpu_address;
538 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
539
540 if (sscreen->info.chip_class >= GFX9) {
541 if (sscreen->info.chip_class >= GFX10) {
542 si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
543 si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, S_00B524_MEM_BASE(va >> 40));
544 } else {
545 si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
546 si_pm4_set_reg(pm4, R_00B414_SPI_SHADER_PGM_HI_LS, S_00B414_MEM_BASE(va >> 40));
547 }
548
549 unsigned num_user_sgprs =
550 si_get_num_vs_user_sgprs(shader, GFX9_TCS_NUM_USER_SGPR);
551
552 shader->config.rsrc2 =
553 S_00B42C_USER_SGPR(num_user_sgprs) |
554 S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
555
556 if (sscreen->info.chip_class >= GFX10)
557 shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
558 else
559 shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
560 } else {
561 si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
562 si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, S_00B424_MEM_BASE(va >> 40));
563
564 shader->config.rsrc2 =
565 S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR) |
566 S_00B42C_OC_LDS_EN(1) |
567 S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
568 }
569
570 si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
571 S_00B428_VGPRS((shader->config.num_vgprs - 1) /
572 (sscreen->ge_wave_size == 32 ? 8 : 4)) |
573 (sscreen->info.chip_class <= GFX9 ?
574 S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) : 0) |
575 S_00B428_DX10_CLAMP(1) |
576 S_00B428_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
577 S_00B428_WGP_MODE(sscreen->info.chip_class >= GFX10) |
578 S_00B428_FLOAT_MODE(shader->config.float_mode) |
579 S_00B428_LS_VGPR_COMP_CNT(sscreen->info.chip_class >= GFX9 ?
580 si_get_vs_vgpr_comp_cnt(sscreen, shader, false) : 0));
581
582 if (sscreen->info.chip_class <= GFX8) {
583 si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
584 shader->config.rsrc2);
585 }
586 }
587
588 static void si_emit_shader_es(struct si_context *sctx)
589 {
590 struct si_shader *shader = sctx->queued.named.es->shader;
591 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
592
593 if (!shader)
594 return;
595
596 radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
597 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
598 shader->selector->esgs_itemsize / 4);
599
600 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
601 radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
602 SI_TRACKED_VGT_TF_PARAM,
603 shader->vgt_tf_param);
604
605 if (shader->vgt_vertex_reuse_block_cntl)
606 radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
607 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
608 shader->vgt_vertex_reuse_block_cntl);
609
610 if (initial_cdw != sctx->gfx_cs->current.cdw)
611 sctx->context_roll = true;
612 }
613
614 static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
615 {
616 struct si_pm4_state *pm4;
617 unsigned num_user_sgprs;
618 unsigned vgpr_comp_cnt;
619 uint64_t va;
620 unsigned oc_lds_en;
621
622 assert(sscreen->info.chip_class <= GFX8);
623
624 pm4 = si_get_shader_pm4_state(shader);
625 if (!pm4)
626 return;
627
628 pm4->atom.emit = si_emit_shader_es;
629 va = shader->bo->gpu_address;
630 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
631
632 if (shader->selector->type == PIPE_SHADER_VERTEX) {
633 vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
634 num_user_sgprs = si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR);
635 } else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
636 vgpr_comp_cnt = shader->selector->info.uses_primid ? 3 : 2;
637 num_user_sgprs = SI_TES_NUM_USER_SGPR;
638 } else
639 unreachable("invalid shader selector type");
640
641 oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
642
643 si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
644 si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
645 si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
646 S_00B328_VGPRS((shader->config.num_vgprs - 1) / 4) |
647 S_00B328_SGPRS((shader->config.num_sgprs - 1) / 8) |
648 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt) |
649 S_00B328_DX10_CLAMP(1) |
650 S_00B328_FLOAT_MODE(shader->config.float_mode));
651 si_pm4_set_reg(pm4, R_00B32C_SPI_SHADER_PGM_RSRC2_ES,
652 S_00B32C_USER_SGPR(num_user_sgprs) |
653 S_00B32C_OC_LDS_EN(oc_lds_en) |
654 S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
655
656 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
657 si_set_tesseval_regs(sscreen, shader->selector, pm4);
658
659 polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
660 }
661
662 void gfx9_get_gs_info(struct si_shader_selector *es,
663 struct si_shader_selector *gs,
664 struct gfx9_gs_info *out)
665 {
666 unsigned gs_num_invocations = MAX2(gs->gs_num_invocations, 1);
667 unsigned input_prim = gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
668 bool uses_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
669 input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
670
671 /* All these are in dwords: */
672 /* We can't allow using the whole LDS, because GS waves compete with
673 * other shader stages for LDS space. */
674 const unsigned max_lds_size = 8 * 1024;
675 const unsigned esgs_itemsize = es->esgs_itemsize / 4;
676 unsigned esgs_lds_size;
677
678 /* All these are per subgroup: */
679 const unsigned max_out_prims = 32 * 1024;
680 const unsigned max_es_verts = 255;
681 const unsigned ideal_gs_prims = 64;
682 unsigned max_gs_prims, gs_prims;
683 unsigned min_es_verts, es_verts, worst_case_es_verts;
684
685 if (uses_adjacency || gs_num_invocations > 1)
686 max_gs_prims = 127 / gs_num_invocations;
687 else
688 max_gs_prims = 255;
689
690 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
691 * Make sure we don't go over the maximum value.
692 */
693 if (gs->gs_max_out_vertices > 0) {
694 max_gs_prims = MIN2(max_gs_prims,
695 max_out_prims /
696 (gs->gs_max_out_vertices * gs_num_invocations));
697 }
698 assert(max_gs_prims > 0);
699
700 /* If the primitive has adjacency, halve the number of vertices
701 * that will be reused in multiple primitives.
702 */
703 min_es_verts = gs->gs_input_verts_per_prim / (uses_adjacency ? 2 : 1);
704
705 gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
706 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
707
708 /* Compute ESGS LDS size based on the worst case number of ES vertices
709 * needed to create the target number of GS prims per subgroup.
710 */
711 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
712
713 /* If total LDS usage is too big, refactor partitions based on ratio
714 * of ESGS item sizes.
715 */
716 if (esgs_lds_size > max_lds_size) {
717 /* Our target GS Prims Per Subgroup was too large. Calculate
718 * the maximum number of GS Prims Per Subgroup that will fit
719 * into LDS, capped by the maximum that the hardware can support.
720 */
721 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
722 max_gs_prims);
723 assert(gs_prims > 0);
724 worst_case_es_verts = MIN2(min_es_verts * gs_prims,
725 max_es_verts);
726
727 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
728 assert(esgs_lds_size <= max_lds_size);
729 }
730
731 /* Now calculate remaining ESGS information. */
732 if (esgs_lds_size)
733 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
734 else
735 es_verts = max_es_verts;
736
737 /* Vertices for adjacency primitives are not always reused, so restore
738 * it for ES_VERTS_PER_SUBGRP.
739 */
740 min_es_verts = gs->gs_input_verts_per_prim;
741
742 /* For normal primitives, the VGT only checks if they are past the ES
743 * verts per subgroup after allocating a full GS primitive and if they
744 * are, kick off a new subgroup. But if those additional ES verts are
745 * unique (e.g. not reused) we need to make sure there is enough LDS
746 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
747 */
748 es_verts -= min_es_verts - 1;
749
750 out->es_verts_per_subgroup = es_verts;
751 out->gs_prims_per_subgroup = gs_prims;
752 out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
753 out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup *
754 gs->gs_max_out_vertices;
755 out->esgs_ring_size = 4 * esgs_lds_size;
756
757 assert(out->max_prims_per_subgroup <= max_out_prims);
758 }
759
760 static void si_emit_shader_gs(struct si_context *sctx)
761 {
762 struct si_shader *shader = sctx->queued.named.gs->shader;
763 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
764
765 if (!shader)
766 return;
767
768 /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
769 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
770 radeon_opt_set_context_reg3(sctx, R_028A60_VGT_GSVS_RING_OFFSET_1,
771 SI_TRACKED_VGT_GSVS_RING_OFFSET_1,
772 shader->ctx_reg.gs.vgt_gsvs_ring_offset_1,
773 shader->ctx_reg.gs.vgt_gsvs_ring_offset_2,
774 shader->ctx_reg.gs.vgt_gsvs_ring_offset_3);
775
776 /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
777 radeon_opt_set_context_reg(sctx, R_028AB0_VGT_GSVS_RING_ITEMSIZE,
778 SI_TRACKED_VGT_GSVS_RING_ITEMSIZE,
779 shader->ctx_reg.gs.vgt_gsvs_ring_itemsize);
780
781 /* R_028B38_VGT_GS_MAX_VERT_OUT */
782 radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
783 SI_TRACKED_VGT_GS_MAX_VERT_OUT,
784 shader->ctx_reg.gs.vgt_gs_max_vert_out);
785
786 /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
787 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
788 radeon_opt_set_context_reg4(sctx, R_028B5C_VGT_GS_VERT_ITEMSIZE,
789 SI_TRACKED_VGT_GS_VERT_ITEMSIZE,
790 shader->ctx_reg.gs.vgt_gs_vert_itemsize,
791 shader->ctx_reg.gs.vgt_gs_vert_itemsize_1,
792 shader->ctx_reg.gs.vgt_gs_vert_itemsize_2,
793 shader->ctx_reg.gs.vgt_gs_vert_itemsize_3);
794
795 /* R_028B90_VGT_GS_INSTANCE_CNT */
796 radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
797 SI_TRACKED_VGT_GS_INSTANCE_CNT,
798 shader->ctx_reg.gs.vgt_gs_instance_cnt);
799
800 if (sctx->chip_class >= GFX9) {
801 /* R_028A44_VGT_GS_ONCHIP_CNTL */
802 radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
803 SI_TRACKED_VGT_GS_ONCHIP_CNTL,
804 shader->ctx_reg.gs.vgt_gs_onchip_cntl);
805 /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
806 radeon_opt_set_context_reg(sctx, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
807 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
808 shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup);
809 /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
810 radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
811 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
812 shader->ctx_reg.gs.vgt_esgs_ring_itemsize);
813
814 if (shader->key.part.gs.es->type == PIPE_SHADER_TESS_EVAL)
815 radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
816 SI_TRACKED_VGT_TF_PARAM,
817 shader->vgt_tf_param);
818 if (shader->vgt_vertex_reuse_block_cntl)
819 radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
820 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
821 shader->vgt_vertex_reuse_block_cntl);
822 }
823
824 if (initial_cdw != sctx->gfx_cs->current.cdw)
825 sctx->context_roll = true;
826 }
827
828 static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
829 {
830 struct si_shader_selector *sel = shader->selector;
831 const ubyte *num_components = sel->info.num_stream_output_components;
832 unsigned gs_num_invocations = sel->gs_num_invocations;
833 struct si_pm4_state *pm4;
834 uint64_t va;
835 unsigned max_stream = sel->max_gs_stream;
836 unsigned offset;
837
838 pm4 = si_get_shader_pm4_state(shader);
839 if (!pm4)
840 return;
841
842 pm4->atom.emit = si_emit_shader_gs;
843
844 offset = num_components[0] * sel->gs_max_out_vertices;
845 shader->ctx_reg.gs.vgt_gsvs_ring_offset_1 = offset;
846
847 if (max_stream >= 1)
848 offset += num_components[1] * sel->gs_max_out_vertices;
849 shader->ctx_reg.gs.vgt_gsvs_ring_offset_2 = offset;
850
851 if (max_stream >= 2)
852 offset += num_components[2] * sel->gs_max_out_vertices;
853 shader->ctx_reg.gs.vgt_gsvs_ring_offset_3 = offset;
854
855 if (max_stream >= 3)
856 offset += num_components[3] * sel->gs_max_out_vertices;
857 shader->ctx_reg.gs.vgt_gsvs_ring_itemsize = offset;
858
859 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
860 assert(offset < (1 << 15));
861
862 shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->gs_max_out_vertices;
863
864 shader->ctx_reg.gs.vgt_gs_vert_itemsize = num_components[0];
865 shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 1) ? num_components[1] : 0;
866 shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 2) ? num_components[2] : 0;
867 shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 3) ? num_components[3] : 0;
868
869 shader->ctx_reg.gs.vgt_gs_instance_cnt = S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
870 S_028B90_ENABLE(gs_num_invocations > 0);
871
872 va = shader->bo->gpu_address;
873 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
874
875 if (sscreen->info.chip_class >= GFX9) {
876 unsigned input_prim = sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
877 unsigned es_type = shader->key.part.gs.es->type;
878 unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
879
880 if (es_type == PIPE_SHADER_VERTEX) {
881 es_vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
882 } else if (es_type == PIPE_SHADER_TESS_EVAL)
883 es_vgpr_comp_cnt = shader->key.part.gs.es->info.uses_primid ? 3 : 2;
884 else
885 unreachable("invalid shader selector type");
886
887 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
888 * VGPR[0:4] are always loaded.
889 */
890 if (sel->info.uses_invocationid)
891 gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
892 else if (sel->info.uses_primid)
893 gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
894 else if (input_prim >= PIPE_PRIM_TRIANGLES)
895 gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
896 else
897 gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
898
899 unsigned num_user_sgprs;
900 if (es_type == PIPE_SHADER_VERTEX)
901 num_user_sgprs = si_get_num_vs_user_sgprs(shader, GFX9_VSGS_NUM_USER_SGPR);
902 else
903 num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
904
905 if (sscreen->info.chip_class >= GFX10) {
906 si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
907 si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
908 } else {
909 si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
910 si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, S_00B214_MEM_BASE(va >> 40));
911 }
912
913 uint32_t rsrc1 =
914 S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
915 S_00B228_DX10_CLAMP(1) |
916 S_00B228_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
917 S_00B228_WGP_MODE(sscreen->info.chip_class >= GFX10) |
918 S_00B228_FLOAT_MODE(shader->config.float_mode) |
919 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt);
920 uint32_t rsrc2 =
921 S_00B22C_USER_SGPR(num_user_sgprs) |
922 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
923 S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
924 S_00B22C_LDS_SIZE(shader->config.lds_size) |
925 S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
926
927 if (sscreen->info.chip_class >= GFX10) {
928 rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
929 } else {
930 rsrc1 |= S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8);
931 rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
932 }
933
934 si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS, rsrc1);
935 si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS, rsrc2);
936
937 shader->ctx_reg.gs.vgt_gs_onchip_cntl =
938 S_028A44_ES_VERTS_PER_SUBGRP(shader->gs_info.es_verts_per_subgroup) |
939 S_028A44_GS_PRIMS_PER_SUBGRP(shader->gs_info.gs_prims_per_subgroup) |
940 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->gs_info.gs_inst_prims_in_subgroup);
941 shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup =
942 S_028A94_MAX_PRIMS_PER_SUBGROUP(shader->gs_info.max_prims_per_subgroup);
943 shader->ctx_reg.gs.vgt_esgs_ring_itemsize =
944 shader->key.part.gs.es->esgs_itemsize / 4;
945
946 if (es_type == PIPE_SHADER_TESS_EVAL)
947 si_set_tesseval_regs(sscreen, shader->key.part.gs.es, pm4);
948
949 polaris_set_vgt_vertex_reuse(sscreen, shader->key.part.gs.es,
950 NULL, pm4);
951 } else {
952 si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
953 si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, S_00B224_MEM_BASE(va >> 40));
954
955 si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
956 S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
957 S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
958 S_00B228_DX10_CLAMP(1) |
959 S_00B228_FLOAT_MODE(shader->config.float_mode));
960 si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
961 S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
962 S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
963 }
964 }
965
966 static void gfx10_emit_ge_pc_alloc(struct si_context *sctx, unsigned value)
967 {
968 enum si_tracked_reg reg = SI_TRACKED_GE_PC_ALLOC;
969
970 if (((sctx->tracked_regs.reg_saved >> reg) & 0x1) != 0x1 ||
971 sctx->tracked_regs.reg_value[reg] != value) {
972 struct radeon_cmdbuf *cs = sctx->gfx_cs;
973
974 if (sctx->family == CHIP_NAVI10 ||
975 sctx->family == CHIP_NAVI12 ||
976 sctx->family == CHIP_NAVI14) {
977 /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
978 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
979 radeon_emit(cs, EVENT_TYPE(V_028A90_SQ_NON_EVENT) | EVENT_INDEX(0));
980 }
981
982 radeon_set_uconfig_reg(cs, R_030980_GE_PC_ALLOC, value);
983
984 sctx->tracked_regs.reg_saved |= 0x1ull << reg;
985 sctx->tracked_regs.reg_value[reg] = value;
986 }
987 }
988
989 /* Common tail code for NGG primitive shaders. */
990 static void gfx10_emit_shader_ngg_tail(struct si_context *sctx,
991 struct si_shader *shader,
992 unsigned initial_cdw)
993 {
994 radeon_opt_set_context_reg(sctx, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
995 SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP,
996 shader->ctx_reg.ngg.ge_max_output_per_subgroup);
997 radeon_opt_set_context_reg(sctx, R_028B4C_GE_NGG_SUBGRP_CNTL,
998 SI_TRACKED_GE_NGG_SUBGRP_CNTL,
999 shader->ctx_reg.ngg.ge_ngg_subgrp_cntl);
1000 radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
1001 SI_TRACKED_VGT_PRIMITIVEID_EN,
1002 shader->ctx_reg.ngg.vgt_primitiveid_en);
1003 radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
1004 SI_TRACKED_VGT_GS_ONCHIP_CNTL,
1005 shader->ctx_reg.ngg.vgt_gs_onchip_cntl);
1006 radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
1007 SI_TRACKED_VGT_GS_INSTANCE_CNT,
1008 shader->ctx_reg.ngg.vgt_gs_instance_cnt);
1009 radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
1010 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
1011 shader->ctx_reg.ngg.vgt_esgs_ring_itemsize);
1012 radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
1013 SI_TRACKED_SPI_VS_OUT_CONFIG,
1014 shader->ctx_reg.ngg.spi_vs_out_config);
1015 radeon_opt_set_context_reg2(sctx, R_028708_SPI_SHADER_IDX_FORMAT,
1016 SI_TRACKED_SPI_SHADER_IDX_FORMAT,
1017 shader->ctx_reg.ngg.spi_shader_idx_format,
1018 shader->ctx_reg.ngg.spi_shader_pos_format);
1019 radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
1020 SI_TRACKED_PA_CL_VTE_CNTL,
1021 shader->ctx_reg.ngg.pa_cl_vte_cntl);
1022 radeon_opt_set_context_reg(sctx, R_028838_PA_CL_NGG_CNTL,
1023 SI_TRACKED_PA_CL_NGG_CNTL,
1024 shader->ctx_reg.ngg.pa_cl_ngg_cntl);
1025
1026 radeon_opt_set_context_reg_rmw(sctx, R_02881C_PA_CL_VS_OUT_CNTL,
1027 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS,
1028 shader->pa_cl_vs_out_cntl,
1029 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK);
1030
1031 if (initial_cdw != sctx->gfx_cs->current.cdw)
1032 sctx->context_roll = true;
1033
1034 /* GE_PC_ALLOC is not a context register, so it doesn't cause a context roll. */
1035 gfx10_emit_ge_pc_alloc(sctx, shader->ctx_reg.ngg.ge_pc_alloc);
1036 }
1037
1038 static void gfx10_emit_shader_ngg_notess_nogs(struct si_context *sctx)
1039 {
1040 struct si_shader *shader = sctx->queued.named.gs->shader;
1041 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1042
1043 if (!shader)
1044 return;
1045
1046 gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
1047 }
1048
1049 static void gfx10_emit_shader_ngg_tess_nogs(struct si_context *sctx)
1050 {
1051 struct si_shader *shader = sctx->queued.named.gs->shader;
1052 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1053
1054 if (!shader)
1055 return;
1056
1057 radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
1058 SI_TRACKED_VGT_TF_PARAM,
1059 shader->vgt_tf_param);
1060
1061 gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
1062 }
1063
1064 static void gfx10_emit_shader_ngg_notess_gs(struct si_context *sctx)
1065 {
1066 struct si_shader *shader = sctx->queued.named.gs->shader;
1067 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1068
1069 if (!shader)
1070 return;
1071
1072 radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
1073 SI_TRACKED_VGT_GS_MAX_VERT_OUT,
1074 shader->ctx_reg.ngg.vgt_gs_max_vert_out);
1075
1076 gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
1077 }
1078
1079 static void gfx10_emit_shader_ngg_tess_gs(struct si_context *sctx)
1080 {
1081 struct si_shader *shader = sctx->queued.named.gs->shader;
1082 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1083
1084 if (!shader)
1085 return;
1086
1087 radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
1088 SI_TRACKED_VGT_GS_MAX_VERT_OUT,
1089 shader->ctx_reg.ngg.vgt_gs_max_vert_out);
1090 radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
1091 SI_TRACKED_VGT_TF_PARAM,
1092 shader->vgt_tf_param);
1093
1094 gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
1095 }
1096
1097 unsigned si_get_input_prim(const struct si_shader_selector *gs)
1098 {
1099 if (gs->type == PIPE_SHADER_GEOMETRY)
1100 return gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
1101
1102 if (gs->type == PIPE_SHADER_TESS_EVAL) {
1103 if (gs->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
1104 return PIPE_PRIM_POINTS;
1105 if (gs->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
1106 return PIPE_PRIM_LINES;
1107 return PIPE_PRIM_TRIANGLES;
1108 }
1109
1110 /* TODO: Set this correctly if the primitive type is set in the shader key. */
1111 return PIPE_PRIM_TRIANGLES; /* worst case for all callers */
1112 }
1113
1114 static unsigned si_get_vs_out_cntl(const struct si_shader_selector *sel, bool ngg)
1115 {
1116 bool misc_vec_ena =
1117 sel->info.writes_psize || (sel->info.writes_edgeflag && !ngg) ||
1118 sel->info.writes_layer || sel->info.writes_viewport_index;
1119 return S_02881C_USE_VTX_POINT_SIZE(sel->info.writes_psize) |
1120 S_02881C_USE_VTX_EDGE_FLAG(sel->info.writes_edgeflag && !ngg) |
1121 S_02881C_USE_VTX_RENDER_TARGET_INDX(sel->info.writes_layer) |
1122 S_02881C_USE_VTX_VIEWPORT_INDX(sel->info.writes_viewport_index) |
1123 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
1124 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena);
1125 }
1126
1127 /**
1128 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1129 * in NGG mode.
1130 */
1131 static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader)
1132 {
1133 const struct si_shader_selector *gs_sel = shader->selector;
1134 const struct si_shader_info *gs_info = &gs_sel->info;
1135 enum pipe_shader_type gs_type = shader->selector->type;
1136 const struct si_shader_selector *es_sel =
1137 shader->previous_stage_sel ? shader->previous_stage_sel : shader->selector;
1138 const struct si_shader_info *es_info = &es_sel->info;
1139 enum pipe_shader_type es_type = es_sel->type;
1140 unsigned num_user_sgprs;
1141 unsigned nparams, es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
1142 uint64_t va;
1143 unsigned window_space =
1144 gs_info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
1145 bool es_enable_prim_id = shader->key.mono.u.vs_export_prim_id || es_info->uses_primid;
1146 unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
1147 unsigned input_prim = si_get_input_prim(gs_sel);
1148 bool break_wave_at_eoi = false;
1149 struct si_pm4_state *pm4 = si_get_shader_pm4_state(shader);
1150 if (!pm4)
1151 return;
1152
1153 if (es_type == PIPE_SHADER_TESS_EVAL) {
1154 pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_tess_gs
1155 : gfx10_emit_shader_ngg_tess_nogs;
1156 } else {
1157 pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_notess_gs
1158 : gfx10_emit_shader_ngg_notess_nogs;
1159 }
1160
1161 va = shader->bo->gpu_address;
1162 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
1163
1164 if (es_type == PIPE_SHADER_VERTEX) {
1165 es_vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
1166
1167 if (es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
1168 num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
1169 es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
1170 } else {
1171 num_user_sgprs = si_get_num_vs_user_sgprs(shader, GFX9_VSGS_NUM_USER_SGPR);
1172 }
1173 } else {
1174 assert(es_type == PIPE_SHADER_TESS_EVAL);
1175 es_vgpr_comp_cnt = es_enable_prim_id ? 3 : 2;
1176 num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
1177
1178 if (es_enable_prim_id || gs_info->uses_primid)
1179 break_wave_at_eoi = true;
1180 }
1181
1182 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1183 * VGPR[0:4] are always loaded.
1184 *
1185 * Vertex shaders always need to load VGPR3, because they need to
1186 * pass edge flags for decomposed primitives (such as quads) to the PA
1187 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1188 */
1189 if (gs_info->uses_invocationid ||
1190 (gs_type == PIPE_SHADER_VERTEX && !gfx10_is_ngg_passthrough(shader)))
1191 gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID, edge flags. */
1192 else if ((gs_type == PIPE_SHADER_GEOMETRY && gs_info->uses_primid) ||
1193 (gs_type == PIPE_SHADER_VERTEX && shader->key.mono.u.vs_export_prim_id))
1194 gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
1195 else if (input_prim >= PIPE_PRIM_TRIANGLES && !gfx10_is_ngg_passthrough(shader))
1196 gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
1197 else
1198 gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
1199
1200 si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
1201 si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, va >> 40);
1202 si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
1203 S_00B228_VGPRS((shader->config.num_vgprs - 1) /
1204 (sscreen->ge_wave_size == 32 ? 8 : 4)) |
1205 S_00B228_FLOAT_MODE(shader->config.float_mode) |
1206 S_00B228_DX10_CLAMP(1) |
1207 S_00B228_MEM_ORDERED(1) |
1208 S_00B228_WGP_MODE(1) |
1209 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
1210 si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
1211 S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0) |
1212 S_00B22C_USER_SGPR(num_user_sgprs) |
1213 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
1214 S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5) |
1215 S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
1216 S_00B22C_LDS_SIZE(shader->config.lds_size));
1217
1218 nparams = MAX2(shader->info.nr_param_exports, 1);
1219 shader->ctx_reg.ngg.spi_vs_out_config =
1220 S_0286C4_VS_EXPORT_COUNT(nparams - 1) |
1221 S_0286C4_NO_PC_EXPORT(shader->info.nr_param_exports == 0);
1222
1223 shader->ctx_reg.ngg.spi_shader_idx_format =
1224 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP);
1225 shader->ctx_reg.ngg.spi_shader_pos_format =
1226 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
1227 S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
1228 V_02870C_SPI_SHADER_4COMP :
1229 V_02870C_SPI_SHADER_NONE) |
1230 S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
1231 V_02870C_SPI_SHADER_4COMP :
1232 V_02870C_SPI_SHADER_NONE) |
1233 S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
1234 V_02870C_SPI_SHADER_4COMP :
1235 V_02870C_SPI_SHADER_NONE);
1236
1237 shader->ctx_reg.ngg.vgt_primitiveid_en =
1238 S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
1239 S_028A84_NGG_DISABLE_PROVOK_REUSE(shader->key.mono.u.vs_export_prim_id ||
1240 gs_sel->info.writes_primid);
1241
1242 if (gs_type == PIPE_SHADER_GEOMETRY) {
1243 shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = es_sel->esgs_itemsize / 4;
1244 shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->gs_max_out_vertices;
1245 } else {
1246 shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = 1;
1247 }
1248
1249 if (es_type == PIPE_SHADER_TESS_EVAL)
1250 si_set_tesseval_regs(sscreen, es_sel, pm4);
1251
1252 shader->ctx_reg.ngg.vgt_gs_onchip_cntl =
1253 S_028A44_ES_VERTS_PER_SUBGRP(shader->ngg.hw_max_esverts) |
1254 S_028A44_GS_PRIMS_PER_SUBGRP(shader->ngg.max_gsprims) |
1255 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->ngg.max_gsprims * gs_num_invocations);
1256 shader->ctx_reg.ngg.ge_max_output_per_subgroup =
1257 S_0287FC_MAX_VERTS_PER_SUBGROUP(shader->ngg.max_out_verts);
1258 shader->ctx_reg.ngg.ge_ngg_subgrp_cntl =
1259 S_028B4C_PRIM_AMP_FACTOR(shader->ngg.prim_amp_factor) |
1260 S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1261 shader->ctx_reg.ngg.vgt_gs_instance_cnt =
1262 S_028B90_CNT(gs_num_invocations) |
1263 S_028B90_ENABLE(gs_num_invocations > 1) |
1264 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
1265 shader->ngg.max_vert_out_per_gs_instance);
1266
1267 /* Always output hw-generated edge flags and pass them via the prim
1268 * export to prevent drawing lines on internal edges of decomposed
1269 * primitives (such as quads) with polygon mode = lines. Only VS needs
1270 * this.
1271 */
1272 shader->ctx_reg.ngg.pa_cl_ngg_cntl =
1273 S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type == PIPE_SHADER_VERTEX);
1274 shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(gs_sel, true);
1275
1276 /* Oversubscribe PC. This improves performance when there are too many varyings. */
1277 float oversub_pc_factor = 0.25;
1278
1279 if (shader->key.opt.ngg_culling) {
1280 /* Be more aggressive with NGG culling. */
1281 if (shader->info.nr_param_exports > 4)
1282 oversub_pc_factor = 1;
1283 else if (shader->info.nr_param_exports > 2)
1284 oversub_pc_factor = 0.75;
1285 else
1286 oversub_pc_factor = 0.5;
1287 }
1288
1289 unsigned oversub_pc_lines = sscreen->info.pc_lines * oversub_pc_factor;
1290 shader->ctx_reg.ngg.ge_pc_alloc = S_030980_OVERSUB_EN(1) |
1291 S_030980_NUM_PC_LINES(oversub_pc_lines - 1);
1292
1293 shader->ge_cntl =
1294 S_03096C_PRIM_GRP_SIZE(shader->ngg.max_gsprims) |
1295 S_03096C_VERT_GRP_SIZE(256) | /* 256 = disable vertex grouping */
1296 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
1297
1298 /* Bug workaround for a possible hang with non-tessellation cases.
1299 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
1300 *
1301 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
1302 */
1303 if ((sscreen->info.family == CHIP_NAVI10 ||
1304 sscreen->info.family == CHIP_NAVI12 ||
1305 sscreen->info.family == CHIP_NAVI14) &&
1306 (es_type == PIPE_SHADER_VERTEX || gs_type == PIPE_SHADER_VERTEX) && /* = no tess */
1307 shader->ngg.hw_max_esverts != 256) {
1308 shader->ge_cntl &= C_03096C_VERT_GRP_SIZE;
1309
1310 if (shader->ngg.hw_max_esverts > 5) {
1311 shader->ge_cntl |=
1312 S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts - 5);
1313 }
1314 }
1315
1316 if (window_space) {
1317 shader->ctx_reg.ngg.pa_cl_vte_cntl =
1318 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1319 } else {
1320 shader->ctx_reg.ngg.pa_cl_vte_cntl =
1321 S_028818_VTX_W0_FMT(1) |
1322 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1323 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1324 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1325 }
1326 }
1327
1328 static void si_emit_shader_vs(struct si_context *sctx)
1329 {
1330 struct si_shader *shader = sctx->queued.named.vs->shader;
1331 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1332
1333 if (!shader)
1334 return;
1335
1336 radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE,
1337 SI_TRACKED_VGT_GS_MODE,
1338 shader->ctx_reg.vs.vgt_gs_mode);
1339 radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
1340 SI_TRACKED_VGT_PRIMITIVEID_EN,
1341 shader->ctx_reg.vs.vgt_primitiveid_en);
1342
1343 if (sctx->chip_class <= GFX8) {
1344 radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
1345 SI_TRACKED_VGT_REUSE_OFF,
1346 shader->ctx_reg.vs.vgt_reuse_off);
1347 }
1348
1349 radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
1350 SI_TRACKED_SPI_VS_OUT_CONFIG,
1351 shader->ctx_reg.vs.spi_vs_out_config);
1352
1353 radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
1354 SI_TRACKED_SPI_SHADER_POS_FORMAT,
1355 shader->ctx_reg.vs.spi_shader_pos_format);
1356
1357 radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
1358 SI_TRACKED_PA_CL_VTE_CNTL,
1359 shader->ctx_reg.vs.pa_cl_vte_cntl);
1360
1361 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
1362 radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
1363 SI_TRACKED_VGT_TF_PARAM,
1364 shader->vgt_tf_param);
1365
1366 if (shader->vgt_vertex_reuse_block_cntl)
1367 radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
1368 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
1369 shader->vgt_vertex_reuse_block_cntl);
1370
1371 /* Required programming for tessellation. (legacy pipeline only) */
1372 if (sctx->chip_class == GFX10 &&
1373 shader->selector->type == PIPE_SHADER_TESS_EVAL) {
1374 radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
1375 SI_TRACKED_VGT_GS_ONCHIP_CNTL,
1376 S_028A44_ES_VERTS_PER_SUBGRP(250) |
1377 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
1378 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
1379 }
1380
1381 if (sctx->chip_class >= GFX10) {
1382 radeon_opt_set_context_reg_rmw(sctx, R_02881C_PA_CL_VS_OUT_CNTL,
1383 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS,
1384 shader->pa_cl_vs_out_cntl,
1385 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK);
1386 }
1387
1388 if (initial_cdw != sctx->gfx_cs->current.cdw)
1389 sctx->context_roll = true;
1390
1391 /* GE_PC_ALLOC is not a context register, so it doesn't cause a context roll. */
1392 if (sctx->chip_class >= GFX10)
1393 gfx10_emit_ge_pc_alloc(sctx, shader->ctx_reg.vs.ge_pc_alloc);
1394 }
1395
1396 /**
1397 * Compute the state for \p shader, which will run as a vertex shader on the
1398 * hardware.
1399 *
1400 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1401 * is the copy shader.
1402 */
1403 static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
1404 struct si_shader_selector *gs)
1405 {
1406 const struct si_shader_info *info = &shader->selector->info;
1407 struct si_pm4_state *pm4;
1408 unsigned num_user_sgprs, vgpr_comp_cnt;
1409 uint64_t va;
1410 unsigned nparams, oc_lds_en;
1411 unsigned window_space =
1412 info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
1413 bool enable_prim_id = shader->key.mono.u.vs_export_prim_id || info->uses_primid;
1414
1415 pm4 = si_get_shader_pm4_state(shader);
1416 if (!pm4)
1417 return;
1418
1419 pm4->atom.emit = si_emit_shader_vs;
1420
1421 /* We always write VGT_GS_MODE in the VS state, because every switch
1422 * between different shader pipelines involving a different GS or no
1423 * GS at all involves a switch of the VS (different GS use different
1424 * copy shaders). On the other hand, when the API switches from a GS to
1425 * no GS and then back to the same GS used originally, the GS state is
1426 * not sent again.
1427 */
1428 if (!gs) {
1429 unsigned mode = V_028A40_GS_OFF;
1430
1431 /* PrimID needs GS scenario A. */
1432 if (enable_prim_id)
1433 mode = V_028A40_GS_SCENARIO_A;
1434
1435 shader->ctx_reg.vs.vgt_gs_mode = S_028A40_MODE(mode);
1436 shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
1437 } else {
1438 shader->ctx_reg.vs.vgt_gs_mode = ac_vgt_gs_mode(gs->gs_max_out_vertices,
1439 sscreen->info.chip_class);
1440 shader->ctx_reg.vs.vgt_primitiveid_en = 0;
1441 }
1442
1443 if (sscreen->info.chip_class <= GFX8) {
1444 /* Reuse needs to be set off if we write oViewport. */
1445 shader->ctx_reg.vs.vgt_reuse_off =
1446 S_028AB4_REUSE_OFF(info->writes_viewport_index);
1447 }
1448
1449 va = shader->bo->gpu_address;
1450 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
1451
1452 if (gs) {
1453 vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
1454 num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
1455 } else if (shader->selector->type == PIPE_SHADER_VERTEX) {
1456 vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, enable_prim_id);
1457
1458 if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
1459 num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
1460 info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
1461 } else {
1462 num_user_sgprs = si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR);
1463 }
1464 } else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
1465 vgpr_comp_cnt = enable_prim_id ? 3 : 2;
1466 num_user_sgprs = SI_TES_NUM_USER_SGPR;
1467 } else
1468 unreachable("invalid shader selector type");
1469
1470 /* VS is required to export at least one param. */
1471 nparams = MAX2(shader->info.nr_param_exports, 1);
1472 shader->ctx_reg.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
1473
1474 if (sscreen->info.chip_class >= GFX10) {
1475 shader->ctx_reg.vs.spi_vs_out_config |=
1476 S_0286C4_NO_PC_EXPORT(shader->info.nr_param_exports == 0);
1477 }
1478
1479 shader->ctx_reg.vs.spi_shader_pos_format =
1480 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
1481 S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
1482 V_02870C_SPI_SHADER_4COMP :
1483 V_02870C_SPI_SHADER_NONE) |
1484 S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
1485 V_02870C_SPI_SHADER_4COMP :
1486 V_02870C_SPI_SHADER_NONE) |
1487 S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
1488 V_02870C_SPI_SHADER_4COMP :
1489 V_02870C_SPI_SHADER_NONE);
1490 shader->ctx_reg.vs.ge_pc_alloc = S_030980_OVERSUB_EN(1) |
1491 S_030980_NUM_PC_LINES(sscreen->info.pc_lines / 4 - 1);
1492 shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, false);
1493
1494 oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
1495
1496 si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
1497 si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, S_00B124_MEM_BASE(va >> 40));
1498
1499 uint32_t rsrc1 = S_00B128_VGPRS((shader->config.num_vgprs - 1) /
1500 (sscreen->ge_wave_size == 32 ? 8 : 4)) |
1501 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
1502 S_00B128_DX10_CLAMP(1) |
1503 S_00B128_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
1504 S_00B128_FLOAT_MODE(shader->config.float_mode);
1505 uint32_t rsrc2 = S_00B12C_USER_SGPR(num_user_sgprs) |
1506 S_00B12C_OC_LDS_EN(oc_lds_en) |
1507 S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
1508
1509 if (sscreen->info.chip_class >= GFX10)
1510 rsrc2 |= S_00B12C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
1511 else if (sscreen->info.chip_class == GFX9)
1512 rsrc2 |= S_00B12C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
1513
1514 if (sscreen->info.chip_class <= GFX9)
1515 rsrc1 |= S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8);
1516
1517 if (!sscreen->use_ngg_streamout) {
1518 rsrc2 |= S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
1519 S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
1520 S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
1521 S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
1522 S_00B12C_SO_EN(!!shader->selector->so.num_outputs);
1523 }
1524
1525 si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS, rsrc1);
1526 si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS, rsrc2);
1527
1528 if (window_space)
1529 shader->ctx_reg.vs.pa_cl_vte_cntl =
1530 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1531 else
1532 shader->ctx_reg.vs.pa_cl_vte_cntl =
1533 S_028818_VTX_W0_FMT(1) |
1534 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1535 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1536 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1537
1538 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
1539 si_set_tesseval_regs(sscreen, shader->selector, pm4);
1540
1541 polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
1542 }
1543
1544 static unsigned si_get_ps_num_interp(struct si_shader *ps)
1545 {
1546 struct si_shader_info *info = &ps->selector->info;
1547 unsigned num_colors = !!(info->colors_read & 0x0f) +
1548 !!(info->colors_read & 0xf0);
1549 unsigned num_interp = ps->selector->info.num_inputs +
1550 (ps->key.part.ps.prolog.color_two_side ? num_colors : 0);
1551
1552 assert(num_interp <= 32);
1553 return MIN2(num_interp, 32);
1554 }
1555
1556 static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
1557 {
1558 unsigned value = shader->key.part.ps.epilog.spi_shader_col_format;
1559 unsigned i, num_targets = (util_last_bit(value) + 3) / 4;
1560
1561 /* If the i-th target format is set, all previous target formats must
1562 * be non-zero to avoid hangs.
1563 */
1564 for (i = 0; i < num_targets; i++)
1565 if (!(value & (0xf << (i * 4))))
1566 value |= V_028714_SPI_SHADER_32_R << (i * 4);
1567
1568 return value;
1569 }
1570
1571 static void si_emit_shader_ps(struct si_context *sctx)
1572 {
1573 struct si_shader *shader = sctx->queued.named.ps->shader;
1574 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1575
1576 if (!shader)
1577 return;
1578
1579 /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1580 radeon_opt_set_context_reg2(sctx, R_0286CC_SPI_PS_INPUT_ENA,
1581 SI_TRACKED_SPI_PS_INPUT_ENA,
1582 shader->ctx_reg.ps.spi_ps_input_ena,
1583 shader->ctx_reg.ps.spi_ps_input_addr);
1584
1585 radeon_opt_set_context_reg(sctx, R_0286E0_SPI_BARYC_CNTL,
1586 SI_TRACKED_SPI_BARYC_CNTL,
1587 shader->ctx_reg.ps.spi_baryc_cntl);
1588 radeon_opt_set_context_reg(sctx, R_0286D8_SPI_PS_IN_CONTROL,
1589 SI_TRACKED_SPI_PS_IN_CONTROL,
1590 shader->ctx_reg.ps.spi_ps_in_control);
1591
1592 /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1593 radeon_opt_set_context_reg2(sctx, R_028710_SPI_SHADER_Z_FORMAT,
1594 SI_TRACKED_SPI_SHADER_Z_FORMAT,
1595 shader->ctx_reg.ps.spi_shader_z_format,
1596 shader->ctx_reg.ps.spi_shader_col_format);
1597
1598 radeon_opt_set_context_reg(sctx, R_02823C_CB_SHADER_MASK,
1599 SI_TRACKED_CB_SHADER_MASK,
1600 shader->ctx_reg.ps.cb_shader_mask);
1601
1602 if (initial_cdw != sctx->gfx_cs->current.cdw)
1603 sctx->context_roll = true;
1604 }
1605
1606 static void si_shader_ps(struct si_screen *sscreen, struct si_shader *shader)
1607 {
1608 struct si_shader_info *info = &shader->selector->info;
1609 struct si_pm4_state *pm4;
1610 unsigned spi_ps_in_control, spi_shader_col_format, cb_shader_mask;
1611 unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
1612 uint64_t va;
1613 unsigned input_ena = shader->config.spi_ps_input_ena;
1614
1615 /* we need to enable at least one of them, otherwise we hang the GPU */
1616 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1617 G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1618 G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1619 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena) ||
1620 G_0286CC_LINEAR_SAMPLE_ENA(input_ena) ||
1621 G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1622 G_0286CC_LINEAR_CENTROID_ENA(input_ena) ||
1623 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena));
1624 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1625 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena) ||
1626 G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1627 G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1628 G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1629 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena));
1630
1631 /* Validate interpolation optimization flags (read as implications). */
1632 assert(!shader->key.part.ps.prolog.bc_optimize_for_persp ||
1633 (G_0286CC_PERSP_CENTER_ENA(input_ena) &&
1634 G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1635 assert(!shader->key.part.ps.prolog.bc_optimize_for_linear ||
1636 (G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
1637 G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1638 assert(!shader->key.part.ps.prolog.force_persp_center_interp ||
1639 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena) &&
1640 !G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1641 assert(!shader->key.part.ps.prolog.force_linear_center_interp ||
1642 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena) &&
1643 !G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1644 assert(!shader->key.part.ps.prolog.force_persp_sample_interp ||
1645 (!G_0286CC_PERSP_CENTER_ENA(input_ena) &&
1646 !G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1647 assert(!shader->key.part.ps.prolog.force_linear_sample_interp ||
1648 (!G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
1649 !G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1650
1651 /* Validate cases when the optimizations are off (read as implications). */
1652 assert(shader->key.part.ps.prolog.bc_optimize_for_persp ||
1653 !G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1654 !G_0286CC_PERSP_CENTROID_ENA(input_ena));
1655 assert(shader->key.part.ps.prolog.bc_optimize_for_linear ||
1656 !G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1657 !G_0286CC_LINEAR_CENTROID_ENA(input_ena));
1658
1659 pm4 = si_get_shader_pm4_state(shader);
1660 if (!pm4)
1661 return;
1662
1663 pm4->atom.emit = si_emit_shader_ps;
1664
1665 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1666 * Possible vaules:
1667 * 0 -> Position = pixel center
1668 * 1 -> Position = pixel centroid
1669 * 2 -> Position = at sample position
1670 *
1671 * From GLSL 4.5 specification, section 7.1:
1672 * "The variable gl_FragCoord is available as an input variable from
1673 * within fragment shaders and it holds the window relative coordinates
1674 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1675 * value can be for any location within the pixel, or one of the
1676 * fragment samples. The use of centroid does not further restrict
1677 * this value to be inside the current primitive."
1678 *
1679 * Meaning that centroid has no effect and we can return anything within
1680 * the pixel. Thus, return the value at sample position, because that's
1681 * the most accurate one shaders can get.
1682 */
1683 spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1684
1685 if (info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1686 TGSI_FS_COORD_PIXEL_CENTER_INTEGER)
1687 spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
1688
1689 spi_shader_col_format = si_get_spi_shader_col_format(shader);
1690 cb_shader_mask = ac_get_cb_shader_mask(spi_shader_col_format);
1691
1692 /* Ensure that some export memory is always allocated, for two reasons:
1693 *
1694 * 1) Correctness: The hardware ignores the EXEC mask if no export
1695 * memory is allocated, so KILL and alpha test do not work correctly
1696 * without this.
1697 * 2) Performance: Every shader needs at least a NULL export, even when
1698 * it writes no color/depth output. The NULL export instruction
1699 * stalls without this setting.
1700 *
1701 * Don't add this to CB_SHADER_MASK.
1702 *
1703 * GFX10 supports pixel shaders without exports by setting both
1704 * the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
1705 * instructions if any are present.
1706 */
1707 if ((sscreen->info.chip_class <= GFX9 ||
1708 info->uses_kill ||
1709 shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS) &&
1710 !spi_shader_col_format &&
1711 !info->writes_z && !info->writes_stencil && !info->writes_samplemask)
1712 spi_shader_col_format = V_028714_SPI_SHADER_32_R;
1713
1714 shader->ctx_reg.ps.spi_ps_input_ena = input_ena;
1715 shader->ctx_reg.ps.spi_ps_input_addr = shader->config.spi_ps_input_addr;
1716
1717 /* Set interpolation controls. */
1718 spi_ps_in_control = S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader)) |
1719 S_0286D8_PS_W32_EN(sscreen->ps_wave_size == 32);
1720
1721 shader->ctx_reg.ps.spi_baryc_cntl = spi_baryc_cntl;
1722 shader->ctx_reg.ps.spi_ps_in_control = spi_ps_in_control;
1723 shader->ctx_reg.ps.spi_shader_z_format =
1724 ac_get_spi_shader_z_format(info->writes_z,
1725 info->writes_stencil,
1726 info->writes_samplemask);
1727 shader->ctx_reg.ps.spi_shader_col_format = spi_shader_col_format;
1728 shader->ctx_reg.ps.cb_shader_mask = cb_shader_mask;
1729
1730 va = shader->bo->gpu_address;
1731 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
1732 si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
1733 si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS, S_00B024_MEM_BASE(va >> 40));
1734
1735 uint32_t rsrc1 =
1736 S_00B028_VGPRS((shader->config.num_vgprs - 1) /
1737 (sscreen->ps_wave_size == 32 ? 8 : 4)) |
1738 S_00B028_DX10_CLAMP(1) |
1739 S_00B028_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
1740 S_00B028_FLOAT_MODE(shader->config.float_mode);
1741
1742 if (sscreen->info.chip_class < GFX10) {
1743 rsrc1 |= S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8);
1744 }
1745
1746 si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS, rsrc1);
1747 si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
1748 S_00B02C_EXTRA_LDS_SIZE(shader->config.lds_size) |
1749 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR) |
1750 S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
1751 }
1752
1753 static void si_shader_init_pm4_state(struct si_screen *sscreen,
1754 struct si_shader *shader)
1755 {
1756 switch (shader->selector->type) {
1757 case PIPE_SHADER_VERTEX:
1758 if (shader->key.as_ls)
1759 si_shader_ls(sscreen, shader);
1760 else if (shader->key.as_es)
1761 si_shader_es(sscreen, shader);
1762 else if (shader->key.as_ngg)
1763 gfx10_shader_ngg(sscreen, shader);
1764 else
1765 si_shader_vs(sscreen, shader, NULL);
1766 break;
1767 case PIPE_SHADER_TESS_CTRL:
1768 si_shader_hs(sscreen, shader);
1769 break;
1770 case PIPE_SHADER_TESS_EVAL:
1771 if (shader->key.as_es)
1772 si_shader_es(sscreen, shader);
1773 else if (shader->key.as_ngg)
1774 gfx10_shader_ngg(sscreen, shader);
1775 else
1776 si_shader_vs(sscreen, shader, NULL);
1777 break;
1778 case PIPE_SHADER_GEOMETRY:
1779 if (shader->key.as_ngg)
1780 gfx10_shader_ngg(sscreen, shader);
1781 else
1782 si_shader_gs(sscreen, shader);
1783 break;
1784 case PIPE_SHADER_FRAGMENT:
1785 si_shader_ps(sscreen, shader);
1786 break;
1787 default:
1788 assert(0);
1789 }
1790 }
1791
1792 static unsigned si_get_alpha_test_func(struct si_context *sctx)
1793 {
1794 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
1795 return sctx->queued.named.dsa->alpha_func;
1796 }
1797
1798 void si_shader_selector_key_vs(struct si_context *sctx,
1799 struct si_shader_selector *vs,
1800 struct si_shader_key *key,
1801 struct si_vs_prolog_bits *prolog_key)
1802 {
1803 if (!sctx->vertex_elements ||
1804 vs->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD])
1805 return;
1806
1807 struct si_vertex_elements *elts = sctx->vertex_elements;
1808
1809 prolog_key->instance_divisor_is_one = elts->instance_divisor_is_one;
1810 prolog_key->instance_divisor_is_fetched = elts->instance_divisor_is_fetched;
1811 prolog_key->unpack_instance_id_from_vertex_id =
1812 sctx->prim_discard_cs_instancing;
1813
1814 /* Prefer a monolithic shader to allow scheduling divisions around
1815 * VBO loads. */
1816 if (prolog_key->instance_divisor_is_fetched)
1817 key->opt.prefer_mono = 1;
1818
1819 unsigned count = MIN2(vs->info.num_inputs, elts->count);
1820 unsigned count_mask = (1 << count) - 1;
1821 unsigned fix = elts->fix_fetch_always & count_mask;
1822 unsigned opencode = elts->fix_fetch_opencode & count_mask;
1823
1824 if (sctx->vertex_buffer_unaligned & elts->vb_alignment_check_mask) {
1825 uint32_t mask = elts->fix_fetch_unaligned & count_mask;
1826 while (mask) {
1827 unsigned i = u_bit_scan(&mask);
1828 unsigned log_hw_load_size = 1 + ((elts->hw_load_is_dword >> i) & 1);
1829 unsigned vbidx = elts->vertex_buffer_index[i];
1830 struct pipe_vertex_buffer *vb = &sctx->vertex_buffer[vbidx];
1831 unsigned align_mask = (1 << log_hw_load_size) - 1;
1832 if (vb->buffer_offset & align_mask ||
1833 vb->stride & align_mask) {
1834 fix |= 1 << i;
1835 opencode |= 1 << i;
1836 }
1837 }
1838 }
1839
1840 while (fix) {
1841 unsigned i = u_bit_scan(&fix);
1842 key->mono.vs_fix_fetch[i].bits = elts->fix_fetch[i];
1843 }
1844 key->mono.vs_fetch_opencode = opencode;
1845 }
1846
1847 static void si_shader_selector_key_hw_vs(struct si_context *sctx,
1848 struct si_shader_selector *vs,
1849 struct si_shader_key *key)
1850 {
1851 struct si_shader_selector *ps = sctx->ps_shader.cso;
1852
1853 key->opt.clip_disable =
1854 sctx->queued.named.rasterizer->clip_plane_enable == 0 &&
1855 (vs->info.clipdist_writemask ||
1856 vs->info.writes_clipvertex) &&
1857 !vs->info.culldist_writemask;
1858
1859 /* Find out if PS is disabled. */
1860 bool ps_disabled = true;
1861 if (ps) {
1862 bool ps_modifies_zs = ps->info.uses_kill ||
1863 ps->info.writes_z ||
1864 ps->info.writes_stencil ||
1865 ps->info.writes_samplemask ||
1866 sctx->queued.named.blend->alpha_to_coverage ||
1867 si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS;
1868 unsigned ps_colormask = si_get_total_colormask(sctx);
1869
1870 ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
1871 (!ps_colormask &&
1872 !ps_modifies_zs &&
1873 !ps->info.writes_memory);
1874 }
1875
1876 /* Find out which VS outputs aren't used by the PS. */
1877 uint64_t outputs_written = vs->outputs_written_before_ps;
1878 uint64_t inputs_read = 0;
1879
1880 /* Ignore outputs that are not passed from VS to PS. */
1881 outputs_written &= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION, 0, true)) |
1882 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE, 0, true)) |
1883 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX, 0, true)));
1884
1885 if (!ps_disabled) {
1886 inputs_read = ps->inputs_read;
1887 }
1888
1889 uint64_t linked = outputs_written & inputs_read;
1890
1891 key->opt.kill_outputs = ~linked & outputs_written;
1892 key->opt.ngg_culling = sctx->ngg_culling;
1893 }
1894
1895 /* Compute the key for the hw shader variant */
1896 static inline void si_shader_selector_key(struct pipe_context *ctx,
1897 struct si_shader_selector *sel,
1898 union si_vgt_stages_key stages_key,
1899 struct si_shader_key *key)
1900 {
1901 struct si_context *sctx = (struct si_context *)ctx;
1902
1903 memset(key, 0, sizeof(*key));
1904
1905 switch (sel->type) {
1906 case PIPE_SHADER_VERTEX:
1907 si_shader_selector_key_vs(sctx, sel, key, &key->part.vs.prolog);
1908
1909 if (sctx->tes_shader.cso)
1910 key->as_ls = 1;
1911 else if (sctx->gs_shader.cso) {
1912 key->as_es = 1;
1913 key->as_ngg = stages_key.u.ngg;
1914 } else {
1915 key->as_ngg = stages_key.u.ngg;
1916 si_shader_selector_key_hw_vs(sctx, sel, key);
1917
1918 if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
1919 key->mono.u.vs_export_prim_id = 1;
1920 }
1921 break;
1922 case PIPE_SHADER_TESS_CTRL:
1923 if (sctx->chip_class >= GFX9) {
1924 si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
1925 key, &key->part.tcs.ls_prolog);
1926 key->part.tcs.ls = sctx->vs_shader.cso;
1927
1928 /* When the LS VGPR fix is needed, monolithic shaders
1929 * can:
1930 * - avoid initializing EXEC in both the LS prolog
1931 * and the LS main part when !vs_needs_prolog
1932 * - remove the fixup for unused input VGPRs
1933 */
1934 key->part.tcs.ls_prolog.ls_vgpr_fix = sctx->ls_vgpr_fix;
1935
1936 /* The LS output / HS input layout can be communicated
1937 * directly instead of via user SGPRs for merged LS-HS.
1938 * The LS VGPR fix prefers this too.
1939 */
1940 key->opt.prefer_mono = 1;
1941 }
1942
1943 key->part.tcs.epilog.prim_mode =
1944 sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
1945 key->part.tcs.epilog.invoc0_tess_factors_are_def =
1946 sel->info.tessfactors_are_def_in_all_invocs;
1947 key->part.tcs.epilog.tes_reads_tess_factors =
1948 sctx->tes_shader.cso->info.reads_tess_factors;
1949
1950 if (sel == sctx->fixed_func_tcs_shader.cso)
1951 key->mono.u.ff_tcs_inputs_to_copy = sctx->vs_shader.cso->outputs_written;
1952 break;
1953 case PIPE_SHADER_TESS_EVAL:
1954 key->as_ngg = stages_key.u.ngg;
1955
1956 if (sctx->gs_shader.cso)
1957 key->as_es = 1;
1958 else {
1959 si_shader_selector_key_hw_vs(sctx, sel, key);
1960
1961 if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
1962 key->mono.u.vs_export_prim_id = 1;
1963 }
1964 break;
1965 case PIPE_SHADER_GEOMETRY:
1966 if (sctx->chip_class >= GFX9) {
1967 if (sctx->tes_shader.cso) {
1968 key->part.gs.es = sctx->tes_shader.cso;
1969 } else {
1970 si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
1971 key, &key->part.gs.vs_prolog);
1972 key->part.gs.es = sctx->vs_shader.cso;
1973 key->part.gs.prolog.gfx9_prev_is_vs = 1;
1974 }
1975
1976 key->as_ngg = stages_key.u.ngg;
1977
1978 /* Merged ES-GS can have unbalanced wave usage.
1979 *
1980 * ES threads are per-vertex, while GS threads are
1981 * per-primitive. So without any amplification, there
1982 * are fewer GS threads than ES threads, which can result
1983 * in empty (no-op) GS waves. With too much amplification,
1984 * there are more GS threads than ES threads, which
1985 * can result in empty (no-op) ES waves.
1986 *
1987 * Non-monolithic shaders are implemented by setting EXEC
1988 * at the beginning of shader parts, and don't jump to
1989 * the end if EXEC is 0.
1990 *
1991 * Monolithic shaders use conditional blocks, so they can
1992 * jump and skip empty waves of ES or GS. So set this to
1993 * always use optimized variants, which are monolithic.
1994 */
1995 key->opt.prefer_mono = 1;
1996 }
1997 key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
1998 break;
1999 case PIPE_SHADER_FRAGMENT: {
2000 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2001 struct si_state_blend *blend = sctx->queued.named.blend;
2002
2003 if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
2004 sel->info.colors_written == 0x1)
2005 key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
2006
2007 /* Select the shader color format based on whether
2008 * blending or alpha are needed.
2009 */
2010 key->part.ps.epilog.spi_shader_col_format =
2011 (blend->blend_enable_4bit & blend->need_src_alpha_4bit &
2012 sctx->framebuffer.spi_shader_col_format_blend_alpha) |
2013 (blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
2014 sctx->framebuffer.spi_shader_col_format_blend) |
2015 (~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
2016 sctx->framebuffer.spi_shader_col_format_alpha) |
2017 (~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
2018 sctx->framebuffer.spi_shader_col_format);
2019 key->part.ps.epilog.spi_shader_col_format &= blend->cb_target_enabled_4bit;
2020
2021 /* The output for dual source blending should have
2022 * the same format as the first output.
2023 */
2024 if (blend->dual_src_blend) {
2025 key->part.ps.epilog.spi_shader_col_format |=
2026 (key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
2027 }
2028
2029 /* If alpha-to-coverage is enabled, we have to export alpha
2030 * even if there is no color buffer.
2031 */
2032 if (!(key->part.ps.epilog.spi_shader_col_format & 0xf) &&
2033 blend->alpha_to_coverage)
2034 key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
2035
2036 /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
2037 * to the range supported by the type if a channel has less
2038 * than 16 bits and the export format is 16_ABGR.
2039 */
2040 if (sctx->chip_class <= GFX7 && sctx->family != CHIP_HAWAII) {
2041 key->part.ps.epilog.color_is_int8 = sctx->framebuffer.color_is_int8;
2042 key->part.ps.epilog.color_is_int10 = sctx->framebuffer.color_is_int10;
2043 }
2044
2045 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
2046 if (!key->part.ps.epilog.last_cbuf) {
2047 key->part.ps.epilog.spi_shader_col_format &= sel->colors_written_4bit;
2048 key->part.ps.epilog.color_is_int8 &= sel->info.colors_written;
2049 key->part.ps.epilog.color_is_int10 &= sel->info.colors_written;
2050 }
2051
2052 bool is_poly = !util_prim_is_points_or_lines(sctx->current_rast_prim);
2053 bool is_line = util_prim_is_lines(sctx->current_rast_prim);
2054
2055 key->part.ps.prolog.color_two_side = rs->two_side && sel->info.colors_read;
2056 key->part.ps.prolog.flatshade_colors = rs->flatshade && sel->info.colors_read;
2057
2058 key->part.ps.epilog.alpha_to_one = blend->alpha_to_one &&
2059 rs->multisample_enable;
2060
2061 key->part.ps.prolog.poly_stipple = rs->poly_stipple_enable && is_poly;
2062 key->part.ps.epilog.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
2063 (is_line && rs->line_smooth)) &&
2064 sctx->framebuffer.nr_samples <= 1;
2065 key->part.ps.epilog.clamp_color = rs->clamp_fragment_color;
2066
2067 if (sctx->ps_iter_samples > 1 &&
2068 sel->info.reads_samplemask) {
2069 key->part.ps.prolog.samplemask_log_ps_iter =
2070 util_logbase2(sctx->ps_iter_samples);
2071 }
2072
2073 if (rs->force_persample_interp &&
2074 rs->multisample_enable &&
2075 sctx->framebuffer.nr_samples > 1 &&
2076 sctx->ps_iter_samples > 1) {
2077 key->part.ps.prolog.force_persp_sample_interp =
2078 sel->info.uses_persp_center ||
2079 sel->info.uses_persp_centroid;
2080
2081 key->part.ps.prolog.force_linear_sample_interp =
2082 sel->info.uses_linear_center ||
2083 sel->info.uses_linear_centroid;
2084 } else if (rs->multisample_enable &&
2085 sctx->framebuffer.nr_samples > 1) {
2086 key->part.ps.prolog.bc_optimize_for_persp =
2087 sel->info.uses_persp_center &&
2088 sel->info.uses_persp_centroid;
2089 key->part.ps.prolog.bc_optimize_for_linear =
2090 sel->info.uses_linear_center &&
2091 sel->info.uses_linear_centroid;
2092 } else {
2093 /* Make sure SPI doesn't compute more than 1 pair
2094 * of (i,j), which is the optimization here. */
2095 key->part.ps.prolog.force_persp_center_interp =
2096 sel->info.uses_persp_center +
2097 sel->info.uses_persp_centroid +
2098 sel->info.uses_persp_sample > 1;
2099
2100 key->part.ps.prolog.force_linear_center_interp =
2101 sel->info.uses_linear_center +
2102 sel->info.uses_linear_centroid +
2103 sel->info.uses_linear_sample > 1;
2104
2105 if (sel->info.uses_persp_opcode_interp_sample ||
2106 sel->info.uses_linear_opcode_interp_sample)
2107 key->mono.u.ps.interpolate_at_sample_force_center = 1;
2108 }
2109
2110 key->part.ps.epilog.alpha_func = si_get_alpha_test_func(sctx);
2111
2112 /* ps_uses_fbfetch is true only if the color buffer is bound. */
2113 if (sctx->ps_uses_fbfetch && !sctx->blitter->running) {
2114 struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
2115 struct pipe_resource *tex = cb0->texture;
2116
2117 /* 1D textures are allocated and used as 2D on GFX9. */
2118 key->mono.u.ps.fbfetch_msaa = sctx->framebuffer.nr_samples > 1;
2119 key->mono.u.ps.fbfetch_is_1D = sctx->chip_class != GFX9 &&
2120 (tex->target == PIPE_TEXTURE_1D ||
2121 tex->target == PIPE_TEXTURE_1D_ARRAY);
2122 key->mono.u.ps.fbfetch_layered = tex->target == PIPE_TEXTURE_1D_ARRAY ||
2123 tex->target == PIPE_TEXTURE_2D_ARRAY ||
2124 tex->target == PIPE_TEXTURE_CUBE ||
2125 tex->target == PIPE_TEXTURE_CUBE_ARRAY ||
2126 tex->target == PIPE_TEXTURE_3D;
2127 }
2128 break;
2129 }
2130 default:
2131 assert(0);
2132 }
2133
2134 if (unlikely(sctx->screen->debug_flags & DBG(NO_OPT_VARIANT)))
2135 memset(&key->opt, 0, sizeof(key->opt));
2136 }
2137
2138 static void si_build_shader_variant(struct si_shader *shader,
2139 int thread_index,
2140 bool low_priority)
2141 {
2142 struct si_shader_selector *sel = shader->selector;
2143 struct si_screen *sscreen = sel->screen;
2144 struct ac_llvm_compiler *compiler;
2145 struct pipe_debug_callback *debug = &shader->compiler_ctx_state.debug;
2146
2147 if (thread_index >= 0) {
2148 if (low_priority) {
2149 assert(thread_index < ARRAY_SIZE(sscreen->compiler_lowp));
2150 compiler = &sscreen->compiler_lowp[thread_index];
2151 } else {
2152 assert(thread_index < ARRAY_SIZE(sscreen->compiler));
2153 compiler = &sscreen->compiler[thread_index];
2154 }
2155 if (!debug->async)
2156 debug = NULL;
2157 } else {
2158 assert(!low_priority);
2159 compiler = shader->compiler_ctx_state.compiler;
2160 }
2161
2162 if (!compiler->passes)
2163 si_init_compiler(sscreen, compiler);
2164
2165 if (unlikely(!si_create_shader_variant(sscreen, compiler, shader, debug))) {
2166 PRINT_ERR("Failed to build shader variant (type=%u)\n",
2167 sel->type);
2168 shader->compilation_failed = true;
2169 return;
2170 }
2171
2172 if (shader->compiler_ctx_state.is_debug_context) {
2173 FILE *f = open_memstream(&shader->shader_log,
2174 &shader->shader_log_size);
2175 if (f) {
2176 si_shader_dump(sscreen, shader, NULL, f, false);
2177 fclose(f);
2178 }
2179 }
2180
2181 si_shader_init_pm4_state(sscreen, shader);
2182 }
2183
2184 static void si_build_shader_variant_low_priority(void *job, int thread_index)
2185 {
2186 struct si_shader *shader = (struct si_shader *)job;
2187
2188 assert(thread_index >= 0);
2189
2190 si_build_shader_variant(shader, thread_index, true);
2191 }
2192
2193 static const struct si_shader_key zeroed;
2194
2195 static bool si_check_missing_main_part(struct si_screen *sscreen,
2196 struct si_shader_selector *sel,
2197 struct si_compiler_ctx_state *compiler_state,
2198 struct si_shader_key *key)
2199 {
2200 struct si_shader **mainp = si_get_main_shader_part(sel, key);
2201
2202 if (!*mainp) {
2203 struct si_shader *main_part = CALLOC_STRUCT(si_shader);
2204
2205 if (!main_part)
2206 return false;
2207
2208 /* We can leave the fence as permanently signaled because the
2209 * main part becomes visible globally only after it has been
2210 * compiled. */
2211 util_queue_fence_init(&main_part->ready);
2212
2213 main_part->selector = sel;
2214 main_part->key.as_es = key->as_es;
2215 main_part->key.as_ls = key->as_ls;
2216 main_part->key.as_ngg = key->as_ngg;
2217 main_part->is_monolithic = false;
2218
2219 if (si_compile_shader(sscreen, compiler_state->compiler,
2220 main_part, &compiler_state->debug) != 0) {
2221 FREE(main_part);
2222 return false;
2223 }
2224 *mainp = main_part;
2225 }
2226 return true;
2227 }
2228
2229 /**
2230 * Select a shader variant according to the shader key.
2231 *
2232 * \param optimized_or_none If the key describes an optimized shader variant and
2233 * the compilation isn't finished, don't select any
2234 * shader and return an error.
2235 */
2236 int si_shader_select_with_key(struct si_screen *sscreen,
2237 struct si_shader_ctx_state *state,
2238 struct si_compiler_ctx_state *compiler_state,
2239 struct si_shader_key *key,
2240 int thread_index,
2241 bool optimized_or_none)
2242 {
2243 struct si_shader_selector *sel = state->cso;
2244 struct si_shader_selector *previous_stage_sel = NULL;
2245 struct si_shader *current = state->current;
2246 struct si_shader *iter, *shader = NULL;
2247
2248 again:
2249 /* Check if we don't need to change anything.
2250 * This path is also used for most shaders that don't need multiple
2251 * variants, it will cost just a computation of the key and this
2252 * test. */
2253 if (likely(current &&
2254 memcmp(&current->key, key, sizeof(*key)) == 0)) {
2255 if (unlikely(!util_queue_fence_is_signalled(&current->ready))) {
2256 if (current->is_optimized) {
2257 if (optimized_or_none)
2258 return -1;
2259
2260 memset(&key->opt, 0, sizeof(key->opt));
2261 goto current_not_ready;
2262 }
2263
2264 util_queue_fence_wait(&current->ready);
2265 }
2266
2267 return current->compilation_failed ? -1 : 0;
2268 }
2269 current_not_ready:
2270
2271 /* This must be done before the mutex is locked, because async GS
2272 * compilation calls this function too, and therefore must enter
2273 * the mutex first.
2274 *
2275 * Only wait if we are in a draw call. Don't wait if we are
2276 * in a compiler thread.
2277 */
2278 if (thread_index < 0)
2279 util_queue_fence_wait(&sel->ready);
2280
2281 simple_mtx_lock(&sel->mutex);
2282
2283 /* Find the shader variant. */
2284 for (iter = sel->first_variant; iter; iter = iter->next_variant) {
2285 /* Don't check the "current" shader. We checked it above. */
2286 if (current != iter &&
2287 memcmp(&iter->key, key, sizeof(*key)) == 0) {
2288 simple_mtx_unlock(&sel->mutex);
2289
2290 if (unlikely(!util_queue_fence_is_signalled(&iter->ready))) {
2291 /* If it's an optimized shader and its compilation has
2292 * been started but isn't done, use the unoptimized
2293 * shader so as not to cause a stall due to compilation.
2294 */
2295 if (iter->is_optimized) {
2296 if (optimized_or_none)
2297 return -1;
2298 memset(&key->opt, 0, sizeof(key->opt));
2299 goto again;
2300 }
2301
2302 util_queue_fence_wait(&iter->ready);
2303 }
2304
2305 if (iter->compilation_failed) {
2306 return -1; /* skip the draw call */
2307 }
2308
2309 state->current = iter;
2310 return 0;
2311 }
2312 }
2313
2314 /* Build a new shader. */
2315 shader = CALLOC_STRUCT(si_shader);
2316 if (!shader) {
2317 simple_mtx_unlock(&sel->mutex);
2318 return -ENOMEM;
2319 }
2320
2321 util_queue_fence_init(&shader->ready);
2322
2323 shader->selector = sel;
2324 shader->key = *key;
2325 shader->compiler_ctx_state = *compiler_state;
2326
2327 /* If this is a merged shader, get the first shader's selector. */
2328 if (sscreen->info.chip_class >= GFX9) {
2329 if (sel->type == PIPE_SHADER_TESS_CTRL)
2330 previous_stage_sel = key->part.tcs.ls;
2331 else if (sel->type == PIPE_SHADER_GEOMETRY)
2332 previous_stage_sel = key->part.gs.es;
2333
2334 /* We need to wait for the previous shader. */
2335 if (previous_stage_sel && thread_index < 0)
2336 util_queue_fence_wait(&previous_stage_sel->ready);
2337 }
2338
2339 bool is_pure_monolithic =
2340 sscreen->use_monolithic_shaders ||
2341 memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
2342
2343 /* Compile the main shader part if it doesn't exist. This can happen
2344 * if the initial guess was wrong.
2345 *
2346 * The prim discard CS doesn't need the main shader part.
2347 */
2348 if (!is_pure_monolithic &&
2349 !key->opt.vs_as_prim_discard_cs) {
2350 bool ok = true;
2351
2352 /* Make sure the main shader part is present. This is needed
2353 * for shaders that can be compiled as VS, LS, or ES, and only
2354 * one of them is compiled at creation.
2355 *
2356 * It is also needed for GS, which can be compiled as non-NGG
2357 * and NGG.
2358 *
2359 * For merged shaders, check that the starting shader's main
2360 * part is present.
2361 */
2362 if (previous_stage_sel) {
2363 struct si_shader_key shader1_key = zeroed;
2364
2365 if (sel->type == PIPE_SHADER_TESS_CTRL) {
2366 shader1_key.as_ls = 1;
2367 } else if (sel->type == PIPE_SHADER_GEOMETRY) {
2368 shader1_key.as_es = 1;
2369 shader1_key.as_ngg = key->as_ngg; /* for Wave32 vs Wave64 */
2370 } else {
2371 assert(0);
2372 }
2373
2374 simple_mtx_lock(&previous_stage_sel->mutex);
2375 ok = si_check_missing_main_part(sscreen,
2376 previous_stage_sel,
2377 compiler_state, &shader1_key);
2378 simple_mtx_unlock(&previous_stage_sel->mutex);
2379 }
2380
2381 if (ok) {
2382 ok = si_check_missing_main_part(sscreen, sel,
2383 compiler_state, key);
2384 }
2385
2386 if (!ok) {
2387 FREE(shader);
2388 simple_mtx_unlock(&sel->mutex);
2389 return -ENOMEM; /* skip the draw call */
2390 }
2391 }
2392
2393 /* Keep the reference to the 1st shader of merged shaders, so that
2394 * Gallium can't destroy it before we destroy the 2nd shader.
2395 *
2396 * Set sctx = NULL, because it's unused if we're not releasing
2397 * the shader, and we don't have any sctx here.
2398 */
2399 si_shader_selector_reference(NULL, &shader->previous_stage_sel,
2400 previous_stage_sel);
2401
2402 /* Monolithic-only shaders don't make a distinction between optimized
2403 * and unoptimized. */
2404 shader->is_monolithic =
2405 is_pure_monolithic ||
2406 memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
2407
2408 /* The prim discard CS is always optimized. */
2409 shader->is_optimized =
2410 (!is_pure_monolithic || key->opt.vs_as_prim_discard_cs) &&
2411 memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
2412
2413 /* If it's an optimized shader, compile it asynchronously. */
2414 if (shader->is_optimized && thread_index < 0) {
2415 /* Compile it asynchronously. */
2416 util_queue_add_job(&sscreen->shader_compiler_queue_low_priority,
2417 shader, &shader->ready,
2418 si_build_shader_variant_low_priority, NULL,
2419 0);
2420
2421 /* Add only after the ready fence was reset, to guard against a
2422 * race with si_bind_XX_shader. */
2423 if (!sel->last_variant) {
2424 sel->first_variant = shader;
2425 sel->last_variant = shader;
2426 } else {
2427 sel->last_variant->next_variant = shader;
2428 sel->last_variant = shader;
2429 }
2430
2431 /* Use the default (unoptimized) shader for now. */
2432 memset(&key->opt, 0, sizeof(key->opt));
2433 simple_mtx_unlock(&sel->mutex);
2434
2435 if (sscreen->options.sync_compile)
2436 util_queue_fence_wait(&shader->ready);
2437
2438 if (optimized_or_none)
2439 return -1;
2440 goto again;
2441 }
2442
2443 /* Reset the fence before adding to the variant list. */
2444 util_queue_fence_reset(&shader->ready);
2445
2446 if (!sel->last_variant) {
2447 sel->first_variant = shader;
2448 sel->last_variant = shader;
2449 } else {
2450 sel->last_variant->next_variant = shader;
2451 sel->last_variant = shader;
2452 }
2453
2454 simple_mtx_unlock(&sel->mutex);
2455
2456 assert(!shader->is_optimized);
2457 si_build_shader_variant(shader, thread_index, false);
2458
2459 util_queue_fence_signal(&shader->ready);
2460
2461 if (!shader->compilation_failed)
2462 state->current = shader;
2463
2464 return shader->compilation_failed ? -1 : 0;
2465 }
2466
2467 static int si_shader_select(struct pipe_context *ctx,
2468 struct si_shader_ctx_state *state,
2469 union si_vgt_stages_key stages_key,
2470 struct si_compiler_ctx_state *compiler_state)
2471 {
2472 struct si_context *sctx = (struct si_context *)ctx;
2473 struct si_shader_key key;
2474
2475 si_shader_selector_key(ctx, state->cso, stages_key, &key);
2476 return si_shader_select_with_key(sctx->screen, state, compiler_state,
2477 &key, -1, false);
2478 }
2479
2480 static void si_parse_next_shader_property(const struct si_shader_info *info,
2481 bool streamout,
2482 struct si_shader_key *key)
2483 {
2484 unsigned next_shader = info->properties[TGSI_PROPERTY_NEXT_SHADER];
2485
2486 switch (info->processor) {
2487 case PIPE_SHADER_VERTEX:
2488 switch (next_shader) {
2489 case PIPE_SHADER_GEOMETRY:
2490 key->as_es = 1;
2491 break;
2492 case PIPE_SHADER_TESS_CTRL:
2493 case PIPE_SHADER_TESS_EVAL:
2494 key->as_ls = 1;
2495 break;
2496 default:
2497 /* If POSITION isn't written, it can only be a HW VS
2498 * if streamout is used. If streamout isn't used,
2499 * assume that it's a HW LS. (the next shader is TCS)
2500 * This heuristic is needed for separate shader objects.
2501 */
2502 if (!info->writes_position && !streamout)
2503 key->as_ls = 1;
2504 }
2505 break;
2506
2507 case PIPE_SHADER_TESS_EVAL:
2508 if (next_shader == PIPE_SHADER_GEOMETRY ||
2509 !info->writes_position)
2510 key->as_es = 1;
2511 break;
2512 }
2513 }
2514
2515 /**
2516 * Compile the main shader part or the monolithic shader as part of
2517 * si_shader_selector initialization. Since it can be done asynchronously,
2518 * there is no way to report compile failures to applications.
2519 */
2520 static void si_init_shader_selector_async(void *job, int thread_index)
2521 {
2522 struct si_shader_selector *sel = (struct si_shader_selector *)job;
2523 struct si_screen *sscreen = sel->screen;
2524 struct ac_llvm_compiler *compiler;
2525 struct pipe_debug_callback *debug = &sel->compiler_ctx_state.debug;
2526
2527 assert(!debug->debug_message || debug->async);
2528 assert(thread_index >= 0);
2529 assert(thread_index < ARRAY_SIZE(sscreen->compiler));
2530 compiler = &sscreen->compiler[thread_index];
2531
2532 if (!compiler->passes)
2533 si_init_compiler(sscreen, compiler);
2534
2535 /* Serialize NIR to save memory. Monolithic shader variants
2536 * have to deserialize NIR before compilation.
2537 */
2538 if (sel->nir) {
2539 struct blob blob;
2540 size_t size;
2541
2542 blob_init(&blob);
2543 /* true = remove optional debugging data to increase
2544 * the likehood of getting more shader cache hits.
2545 * It also drops variable names, so we'll save more memory.
2546 */
2547 nir_serialize(&blob, sel->nir, true);
2548 blob_finish_get_buffer(&blob, &sel->nir_binary, &size);
2549 sel->nir_size = size;
2550 }
2551
2552 /* Compile the main shader part for use with a prolog and/or epilog.
2553 * If this fails, the driver will try to compile a monolithic shader
2554 * on demand.
2555 */
2556 if (!sscreen->use_monolithic_shaders) {
2557 struct si_shader *shader = CALLOC_STRUCT(si_shader);
2558 unsigned char ir_sha1_cache_key[20];
2559
2560 if (!shader) {
2561 fprintf(stderr, "radeonsi: can't allocate a main shader part\n");
2562 return;
2563 }
2564
2565 /* We can leave the fence signaled because use of the default
2566 * main part is guarded by the selector's ready fence. */
2567 util_queue_fence_init(&shader->ready);
2568
2569 shader->selector = sel;
2570 shader->is_monolithic = false;
2571 si_parse_next_shader_property(&sel->info,
2572 sel->so.num_outputs != 0,
2573 &shader->key);
2574
2575 if (sscreen->use_ngg &&
2576 (!sel->so.num_outputs || sscreen->use_ngg_streamout) &&
2577 ((sel->type == PIPE_SHADER_VERTEX && !shader->key.as_ls) ||
2578 sel->type == PIPE_SHADER_TESS_EVAL ||
2579 sel->type == PIPE_SHADER_GEOMETRY))
2580 shader->key.as_ngg = 1;
2581
2582 if (sel->nir) {
2583 si_get_ir_cache_key(sel, shader->key.as_ngg,
2584 shader->key.as_es, ir_sha1_cache_key);
2585 }
2586
2587 /* Try to load the shader from the shader cache. */
2588 simple_mtx_lock(&sscreen->shader_cache_mutex);
2589
2590 if (si_shader_cache_load_shader(sscreen, ir_sha1_cache_key, shader)) {
2591 simple_mtx_unlock(&sscreen->shader_cache_mutex);
2592 si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
2593 } else {
2594 simple_mtx_unlock(&sscreen->shader_cache_mutex);
2595
2596 /* Compile the shader if it hasn't been loaded from the cache. */
2597 if (si_compile_shader(sscreen, compiler, shader,
2598 debug) != 0) {
2599 FREE(shader);
2600 fprintf(stderr, "radeonsi: can't compile a main shader part\n");
2601 return;
2602 }
2603
2604 simple_mtx_lock(&sscreen->shader_cache_mutex);
2605 si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key,
2606 shader, true);
2607 simple_mtx_unlock(&sscreen->shader_cache_mutex);
2608 }
2609
2610 *si_get_main_shader_part(sel, &shader->key) = shader;
2611
2612 /* Unset "outputs_written" flags for outputs converted to
2613 * DEFAULT_VAL, so that later inter-shader optimizations don't
2614 * try to eliminate outputs that don't exist in the final
2615 * shader.
2616 *
2617 * This is only done if non-monolithic shaders are enabled.
2618 */
2619 if ((sel->type == PIPE_SHADER_VERTEX ||
2620 sel->type == PIPE_SHADER_TESS_EVAL) &&
2621 !shader->key.as_ls &&
2622 !shader->key.as_es) {
2623 unsigned i;
2624
2625 for (i = 0; i < sel->info.num_outputs; i++) {
2626 unsigned offset = shader->info.vs_output_param_offset[i];
2627
2628 if (offset <= AC_EXP_PARAM_OFFSET_31)
2629 continue;
2630
2631 unsigned name = sel->info.output_semantic_name[i];
2632 unsigned index = sel->info.output_semantic_index[i];
2633 unsigned id;
2634
2635 switch (name) {
2636 case TGSI_SEMANTIC_GENERIC:
2637 /* don't process indices the function can't handle */
2638 if (index >= SI_MAX_IO_GENERIC)
2639 break;
2640 /* fall through */
2641 default:
2642 id = si_shader_io_get_unique_index(name, index, true);
2643 sel->outputs_written_before_ps &= ~(1ull << id);
2644 break;
2645 case TGSI_SEMANTIC_POSITION: /* ignore these */
2646 case TGSI_SEMANTIC_PSIZE:
2647 case TGSI_SEMANTIC_CLIPVERTEX:
2648 case TGSI_SEMANTIC_EDGEFLAG:
2649 break;
2650 }
2651 }
2652 }
2653 }
2654
2655 /* The GS copy shader is always pre-compiled. */
2656 if (sel->type == PIPE_SHADER_GEOMETRY &&
2657 (!sscreen->use_ngg ||
2658 !sscreen->use_ngg_streamout || /* also for PRIMITIVES_GENERATED */
2659 sel->tess_turns_off_ngg)) {
2660 sel->gs_copy_shader = si_generate_gs_copy_shader(sscreen, compiler, sel, debug);
2661 if (!sel->gs_copy_shader) {
2662 fprintf(stderr, "radeonsi: can't create GS copy shader\n");
2663 return;
2664 }
2665
2666 si_shader_vs(sscreen, sel->gs_copy_shader, sel);
2667 }
2668
2669 /* Free NIR. We only keep serialized NIR after this point. */
2670 if (sel->nir) {
2671 ralloc_free(sel->nir);
2672 sel->nir = NULL;
2673 }
2674 }
2675
2676 void si_schedule_initial_compile(struct si_context *sctx, unsigned processor,
2677 struct util_queue_fence *ready_fence,
2678 struct si_compiler_ctx_state *compiler_ctx_state,
2679 void *job, util_queue_execute_func execute)
2680 {
2681 util_queue_fence_init(ready_fence);
2682
2683 struct util_async_debug_callback async_debug;
2684 bool debug =
2685 (sctx->debug.debug_message && !sctx->debug.async) ||
2686 sctx->is_debug ||
2687 si_can_dump_shader(sctx->screen, processor);
2688
2689 if (debug) {
2690 u_async_debug_init(&async_debug);
2691 compiler_ctx_state->debug = async_debug.base;
2692 }
2693
2694 util_queue_add_job(&sctx->screen->shader_compiler_queue, job,
2695 ready_fence, execute, NULL, 0);
2696
2697 if (debug) {
2698 util_queue_fence_wait(ready_fence);
2699 u_async_debug_drain(&async_debug, &sctx->debug);
2700 u_async_debug_cleanup(&async_debug);
2701 }
2702
2703 if (sctx->screen->options.sync_compile)
2704 util_queue_fence_wait(ready_fence);
2705 }
2706
2707 /* Return descriptor slot usage masks from the given shader info. */
2708 void si_get_active_slot_masks(const struct si_shader_info *info,
2709 uint32_t *const_and_shader_buffers,
2710 uint64_t *samplers_and_images)
2711 {
2712 unsigned start, num_shaderbufs, num_constbufs, num_images, num_msaa_images, num_samplers;
2713
2714 num_shaderbufs = util_last_bit(info->shader_buffers_declared);
2715 num_constbufs = util_last_bit(info->const_buffers_declared);
2716 /* two 8-byte images share one 16-byte slot */
2717 num_images = align(util_last_bit(info->images_declared), 2);
2718 num_msaa_images = align(util_last_bit(info->msaa_images_declared), 2);
2719 num_samplers = util_last_bit(info->samplers_declared);
2720
2721 /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2722 start = si_get_shaderbuf_slot(num_shaderbufs - 1);
2723 *const_and_shader_buffers =
2724 u_bit_consecutive(start, num_shaderbufs + num_constbufs);
2725
2726 /* The layout is:
2727 * - fmask[last] ... fmask[0] go to [15-last .. 15]
2728 * - image[last] ... image[0] go to [31-last .. 31]
2729 * - sampler[0] ... sampler[last] go to [32 .. 32+last*2]
2730 *
2731 * FMASKs for images are placed separately, because MSAA images are rare,
2732 * and so we can benefit from a better cache hit rate if we keep image
2733 * descriptors together.
2734 */
2735 if (num_msaa_images)
2736 num_images = SI_NUM_IMAGES + num_msaa_images; /* add FMASK descriptors */
2737
2738 start = si_get_image_slot(num_images - 1) / 2;
2739 *samplers_and_images =
2740 u_bit_consecutive64(start, num_images / 2 + num_samplers);
2741 }
2742
2743 static void *si_create_shader_selector(struct pipe_context *ctx,
2744 const struct pipe_shader_state *state)
2745 {
2746 struct si_screen *sscreen = (struct si_screen *)ctx->screen;
2747 struct si_context *sctx = (struct si_context*)ctx;
2748 struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
2749 int i;
2750
2751 if (!sel)
2752 return NULL;
2753
2754 pipe_reference_init(&sel->reference, 1);
2755 sel->screen = sscreen;
2756 sel->compiler_ctx_state.debug = sctx->debug;
2757 sel->compiler_ctx_state.is_debug_context = sctx->is_debug;
2758
2759 sel->so = state->stream_output;
2760
2761 if (state->type == PIPE_SHADER_IR_TGSI) {
2762 sel->nir = tgsi_to_nir(state->tokens, ctx->screen);
2763 } else {
2764 assert(state->type == PIPE_SHADER_IR_NIR);
2765 sel->nir = state->ir.nir;
2766 }
2767
2768 si_nir_scan_shader(sel->nir, &sel->info);
2769 si_nir_adjust_driver_locations(sel->nir);
2770
2771 sel->type = sel->info.processor;
2772 p_atomic_inc(&sscreen->num_shaders_created);
2773 si_get_active_slot_masks(&sel->info,
2774 &sel->active_const_and_shader_buffers,
2775 &sel->active_samplers_and_images);
2776
2777 /* Record which streamout buffers are enabled. */
2778 for (i = 0; i < sel->so.num_outputs; i++) {
2779 sel->enabled_streamout_buffer_mask |=
2780 (1 << sel->so.output[i].output_buffer) <<
2781 (sel->so.output[i].stream * 4);
2782 }
2783
2784 sel->num_vs_inputs = sel->type == PIPE_SHADER_VERTEX &&
2785 !sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD] ?
2786 sel->info.num_inputs : 0;
2787 sel->num_vbos_in_user_sgprs =
2788 MIN2(sel->num_vs_inputs, sscreen->num_vbos_in_user_sgprs);
2789
2790 /* The prolog is a no-op if there are no inputs. */
2791 sel->vs_needs_prolog = sel->type == PIPE_SHADER_VERTEX &&
2792 sel->info.num_inputs &&
2793 !sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
2794
2795 sel->force_correct_derivs_after_kill =
2796 sel->type == PIPE_SHADER_FRAGMENT &&
2797 sel->info.uses_derivatives &&
2798 sel->info.uses_kill &&
2799 sctx->screen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL);
2800
2801 sel->prim_discard_cs_allowed =
2802 sel->type == PIPE_SHADER_VERTEX &&
2803 !sel->info.uses_bindless_images &&
2804 !sel->info.uses_bindless_samplers &&
2805 !sel->info.writes_memory &&
2806 !sel->info.writes_viewport_index &&
2807 !sel->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] &&
2808 !sel->so.num_outputs;
2809
2810 switch (sel->type) {
2811 case PIPE_SHADER_GEOMETRY:
2812 sel->gs_output_prim =
2813 sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
2814
2815 /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
2816 sel->rast_prim = sel->gs_output_prim;
2817 if (util_rast_prim_is_triangles(sel->rast_prim))
2818 sel->rast_prim = PIPE_PRIM_TRIANGLES;
2819
2820 sel->gs_max_out_vertices =
2821 sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
2822 sel->gs_num_invocations =
2823 sel->info.properties[TGSI_PROPERTY_GS_INVOCATIONS];
2824 sel->gsvs_vertex_size = sel->info.num_outputs * 16;
2825 sel->max_gsvs_emit_size = sel->gsvs_vertex_size *
2826 sel->gs_max_out_vertices;
2827
2828 sel->max_gs_stream = 0;
2829 for (i = 0; i < sel->so.num_outputs; i++)
2830 sel->max_gs_stream = MAX2(sel->max_gs_stream,
2831 sel->so.output[i].stream);
2832
2833 sel->gs_input_verts_per_prim =
2834 u_vertices_per_prim(sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM]);
2835
2836 /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
2837 sel->tess_turns_off_ngg =
2838 sscreen->info.chip_class == GFX10 &&
2839 sel->gs_num_invocations * sel->gs_max_out_vertices > 256;
2840 break;
2841
2842 case PIPE_SHADER_TESS_CTRL:
2843 /* Always reserve space for these. */
2844 sel->patch_outputs_written |=
2845 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER, 0)) |
2846 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER, 0));
2847 /* fall through */
2848 case PIPE_SHADER_VERTEX:
2849 case PIPE_SHADER_TESS_EVAL:
2850 for (i = 0; i < sel->info.num_outputs; i++) {
2851 unsigned name = sel->info.output_semantic_name[i];
2852 unsigned index = sel->info.output_semantic_index[i];
2853
2854 switch (name) {
2855 case TGSI_SEMANTIC_TESSINNER:
2856 case TGSI_SEMANTIC_TESSOUTER:
2857 case TGSI_SEMANTIC_PATCH:
2858 sel->patch_outputs_written |=
2859 1ull << si_shader_io_get_unique_index_patch(name, index);
2860 break;
2861
2862 case TGSI_SEMANTIC_GENERIC:
2863 /* don't process indices the function can't handle */
2864 if (index >= SI_MAX_IO_GENERIC)
2865 break;
2866 /* fall through */
2867 default:
2868 sel->outputs_written |=
2869 1ull << si_shader_io_get_unique_index(name, index, false);
2870 sel->outputs_written_before_ps |=
2871 1ull << si_shader_io_get_unique_index(name, index, true);
2872 break;
2873 case TGSI_SEMANTIC_EDGEFLAG:
2874 break;
2875 }
2876 }
2877 sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
2878 sel->lshs_vertex_stride = sel->esgs_itemsize;
2879
2880 /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2881 * will start on a different bank. (except for the maximum 32*16).
2882 */
2883 if (sel->lshs_vertex_stride < 32*16)
2884 sel->lshs_vertex_stride += 4;
2885
2886 /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2887 * conflicts, i.e. each vertex will start at a different bank.
2888 */
2889 if (sctx->chip_class >= GFX9)
2890 sel->esgs_itemsize += 4;
2891
2892 assert(((sel->esgs_itemsize / 4) & C_028AAC_ITEMSIZE) == 0);
2893
2894 /* Only for TES: */
2895 if (sel->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
2896 sel->rast_prim = PIPE_PRIM_POINTS;
2897 else if (sel->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
2898 sel->rast_prim = PIPE_PRIM_LINE_STRIP;
2899 else
2900 sel->rast_prim = PIPE_PRIM_TRIANGLES;
2901 break;
2902
2903 case PIPE_SHADER_FRAGMENT:
2904 for (i = 0; i < sel->info.num_inputs; i++) {
2905 unsigned name = sel->info.input_semantic_name[i];
2906 unsigned index = sel->info.input_semantic_index[i];
2907
2908 switch (name) {
2909 case TGSI_SEMANTIC_GENERIC:
2910 /* don't process indices the function can't handle */
2911 if (index >= SI_MAX_IO_GENERIC)
2912 break;
2913 /* fall through */
2914 default:
2915 sel->inputs_read |=
2916 1ull << si_shader_io_get_unique_index(name, index, true);
2917 break;
2918 case TGSI_SEMANTIC_PCOORD: /* ignore this */
2919 break;
2920 }
2921 }
2922
2923 for (i = 0; i < 8; i++)
2924 if (sel->info.colors_written & (1 << i))
2925 sel->colors_written_4bit |= 0xf << (4 * i);
2926
2927 for (i = 0; i < sel->info.num_inputs; i++) {
2928 if (sel->info.input_semantic_name[i] == TGSI_SEMANTIC_COLOR) {
2929 int index = sel->info.input_semantic_index[i];
2930 sel->color_attr_index[index] = i;
2931 }
2932 }
2933 break;
2934 default:;
2935 }
2936
2937 sel->ngg_culling_allowed =
2938 sscreen->info.chip_class == GFX10 &&
2939 sscreen->info.has_dedicated_vram &&
2940 sscreen->use_ngg_culling &&
2941 /* Disallow TES by default, because TessMark results are mixed. */
2942 (sel->type == PIPE_SHADER_VERTEX ||
2943 (sscreen->always_use_ngg_culling && sel->type == PIPE_SHADER_TESS_EVAL)) &&
2944 sel->info.writes_position &&
2945 !sel->info.writes_viewport_index && /* cull only against viewport 0 */
2946 !sel->info.writes_memory &&
2947 !sel->so.num_outputs &&
2948 !sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD] &&
2949 !sel->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
2950
2951 /* PA_CL_VS_OUT_CNTL */
2952 if (sctx->chip_class <= GFX9)
2953 sel->pa_cl_vs_out_cntl = si_get_vs_out_cntl(sel, false);
2954
2955 sel->clipdist_mask = sel->info.writes_clipvertex ?
2956 SIX_BITS : sel->info.clipdist_writemask;
2957 sel->culldist_mask = sel->info.culldist_writemask <<
2958 sel->info.num_written_clipdistance;
2959
2960 /* DB_SHADER_CONTROL */
2961 sel->db_shader_control =
2962 S_02880C_Z_EXPORT_ENABLE(sel->info.writes_z) |
2963 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel->info.writes_stencil) |
2964 S_02880C_MASK_EXPORT_ENABLE(sel->info.writes_samplemask) |
2965 S_02880C_KILL_ENABLE(sel->info.uses_kill);
2966
2967 switch (sel->info.properties[TGSI_PROPERTY_FS_DEPTH_LAYOUT]) {
2968 case TGSI_FS_DEPTH_LAYOUT_GREATER:
2969 sel->db_shader_control |=
2970 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
2971 break;
2972 case TGSI_FS_DEPTH_LAYOUT_LESS:
2973 sel->db_shader_control |=
2974 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
2975 break;
2976 }
2977
2978 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2979 *
2980 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2981 * --|-----------|------------|------------|--------------------|-------------------|-------------
2982 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
2983 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
2984 * 2 | false | true | n/a | LateZ | 1 | 0
2985 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
2986 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
2987 *
2988 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2989 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2990 *
2991 * Don't use ReZ without profiling !!!
2992 *
2993 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2994 * shaders.
2995 */
2996 if (sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]) {
2997 /* Cases 3, 4. */
2998 sel->db_shader_control |= S_02880C_DEPTH_BEFORE_SHADER(1) |
2999 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z) |
3000 S_02880C_EXEC_ON_NOOP(sel->info.writes_memory);
3001 } else if (sel->info.writes_memory) {
3002 /* Case 2. */
3003 sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z) |
3004 S_02880C_EXEC_ON_HIER_FAIL(1);
3005 } else {
3006 /* Case 1. */
3007 sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
3008 }
3009
3010 if (sel->info.properties[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE])
3011 sel->db_shader_control |= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
3012
3013 (void) simple_mtx_init(&sel->mutex, mtx_plain);
3014
3015 si_schedule_initial_compile(sctx, sel->info.processor, &sel->ready,
3016 &sel->compiler_ctx_state, sel,
3017 si_init_shader_selector_async);
3018 return sel;
3019 }
3020
3021 static void si_update_streamout_state(struct si_context *sctx)
3022 {
3023 struct si_shader_selector *shader_with_so = si_get_vs(sctx)->cso;
3024
3025 if (!shader_with_so)
3026 return;
3027
3028 sctx->streamout.enabled_stream_buffers_mask =
3029 shader_with_so->enabled_streamout_buffer_mask;
3030 sctx->streamout.stride_in_dw = shader_with_so->so.stride;
3031 }
3032
3033 static void si_update_clip_regs(struct si_context *sctx,
3034 struct si_shader_selector *old_hw_vs,
3035 struct si_shader *old_hw_vs_variant,
3036 struct si_shader_selector *next_hw_vs,
3037 struct si_shader *next_hw_vs_variant)
3038 {
3039 if (next_hw_vs &&
3040 (!old_hw_vs ||
3041 old_hw_vs->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] !=
3042 next_hw_vs->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] ||
3043 old_hw_vs->pa_cl_vs_out_cntl != next_hw_vs->pa_cl_vs_out_cntl ||
3044 old_hw_vs->clipdist_mask != next_hw_vs->clipdist_mask ||
3045 old_hw_vs->culldist_mask != next_hw_vs->culldist_mask ||
3046 !old_hw_vs_variant ||
3047 !next_hw_vs_variant ||
3048 old_hw_vs_variant->key.opt.clip_disable !=
3049 next_hw_vs_variant->key.opt.clip_disable))
3050 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
3051 }
3052
3053 static void si_update_common_shader_state(struct si_context *sctx)
3054 {
3055 sctx->uses_bindless_samplers =
3056 si_shader_uses_bindless_samplers(sctx->vs_shader.cso) ||
3057 si_shader_uses_bindless_samplers(sctx->gs_shader.cso) ||
3058 si_shader_uses_bindless_samplers(sctx->ps_shader.cso) ||
3059 si_shader_uses_bindless_samplers(sctx->tcs_shader.cso) ||
3060 si_shader_uses_bindless_samplers(sctx->tes_shader.cso);
3061 sctx->uses_bindless_images =
3062 si_shader_uses_bindless_images(sctx->vs_shader.cso) ||
3063 si_shader_uses_bindless_images(sctx->gs_shader.cso) ||
3064 si_shader_uses_bindless_images(sctx->ps_shader.cso) ||
3065 si_shader_uses_bindless_images(sctx->tcs_shader.cso) ||
3066 si_shader_uses_bindless_images(sctx->tes_shader.cso);
3067 sctx->do_update_shaders = true;
3068 }
3069
3070 static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
3071 {
3072 struct si_context *sctx = (struct si_context *)ctx;
3073 struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3074 struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
3075 struct si_shader_selector *sel = state;
3076
3077 if (sctx->vs_shader.cso == sel)
3078 return;
3079
3080 sctx->vs_shader.cso = sel;
3081 sctx->vs_shader.current = sel ? sel->first_variant : NULL;
3082 sctx->num_vs_blit_sgprs = sel ? sel->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD] : 0;
3083
3084 if (si_update_ngg(sctx))
3085 si_shader_change_notify(sctx);
3086
3087 si_update_common_shader_state(sctx);
3088 si_update_vs_viewport_state(sctx);
3089 si_set_active_descriptors_for_shader(sctx, sel);
3090 si_update_streamout_state(sctx);
3091 si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
3092 si_get_vs(sctx)->cso, si_get_vs_state(sctx));
3093 }
3094
3095 static void si_update_tess_uses_prim_id(struct si_context *sctx)
3096 {
3097 sctx->ia_multi_vgt_param_key.u.tess_uses_prim_id =
3098 (sctx->tes_shader.cso &&
3099 sctx->tes_shader.cso->info.uses_primid) ||
3100 (sctx->tcs_shader.cso &&
3101 sctx->tcs_shader.cso->info.uses_primid) ||
3102 (sctx->gs_shader.cso &&
3103 sctx->gs_shader.cso->info.uses_primid) ||
3104 (sctx->ps_shader.cso && !sctx->gs_shader.cso &&
3105 sctx->ps_shader.cso->info.uses_primid);
3106 }
3107
3108 bool si_update_ngg(struct si_context *sctx)
3109 {
3110 if (!sctx->screen->use_ngg) {
3111 assert(!sctx->ngg);
3112 return false;
3113 }
3114
3115 bool new_ngg = true;
3116
3117 if (sctx->gs_shader.cso && sctx->tes_shader.cso &&
3118 sctx->gs_shader.cso->tess_turns_off_ngg) {
3119 new_ngg = false;
3120 } else if (!sctx->screen->use_ngg_streamout) {
3121 struct si_shader_selector *last = si_get_vs(sctx)->cso;
3122
3123 if ((last && last->so.num_outputs) ||
3124 sctx->streamout.prims_gen_query_enabled)
3125 new_ngg = false;
3126 }
3127
3128 if (new_ngg != sctx->ngg) {
3129 /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
3130 * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
3131 * pointers are set.
3132 */
3133 if ((sctx->family == CHIP_NAVI10 ||
3134 sctx->family == CHIP_NAVI12 ||
3135 sctx->family == CHIP_NAVI14) &&
3136 !new_ngg)
3137 sctx->flags |= SI_CONTEXT_VGT_FLUSH;
3138
3139 sctx->ngg = new_ngg;
3140 sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3141 return true;
3142 }
3143 return false;
3144 }
3145
3146 static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
3147 {
3148 struct si_context *sctx = (struct si_context *)ctx;
3149 struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3150 struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
3151 struct si_shader_selector *sel = state;
3152 bool enable_changed = !!sctx->gs_shader.cso != !!sel;
3153 bool ngg_changed;
3154
3155 if (sctx->gs_shader.cso == sel)
3156 return;
3157
3158 sctx->gs_shader.cso = sel;
3159 sctx->gs_shader.current = sel ? sel->first_variant : NULL;
3160 sctx->ia_multi_vgt_param_key.u.uses_gs = sel != NULL;
3161
3162 si_update_common_shader_state(sctx);
3163 sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3164
3165 ngg_changed = si_update_ngg(sctx);
3166 if (ngg_changed || enable_changed)
3167 si_shader_change_notify(sctx);
3168 if (enable_changed) {
3169 if (sctx->ia_multi_vgt_param_key.u.uses_tess)
3170 si_update_tess_uses_prim_id(sctx);
3171 }
3172 si_update_vs_viewport_state(sctx);
3173 si_set_active_descriptors_for_shader(sctx, sel);
3174 si_update_streamout_state(sctx);
3175 si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
3176 si_get_vs(sctx)->cso, si_get_vs_state(sctx));
3177 }
3178
3179 static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
3180 {
3181 struct si_context *sctx = (struct si_context *)ctx;
3182 struct si_shader_selector *sel = state;
3183 bool enable_changed = !!sctx->tcs_shader.cso != !!sel;
3184
3185 if (sctx->tcs_shader.cso == sel)
3186 return;
3187
3188 sctx->tcs_shader.cso = sel;
3189 sctx->tcs_shader.current = sel ? sel->first_variant : NULL;
3190 si_update_tess_uses_prim_id(sctx);
3191
3192 si_update_common_shader_state(sctx);
3193
3194 if (enable_changed)
3195 sctx->last_tcs = NULL; /* invalidate derived tess state */
3196
3197 si_set_active_descriptors_for_shader(sctx, sel);
3198 }
3199
3200 static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
3201 {
3202 struct si_context *sctx = (struct si_context *)ctx;
3203 struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3204 struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
3205 struct si_shader_selector *sel = state;
3206 bool enable_changed = !!sctx->tes_shader.cso != !!sel;
3207
3208 if (sctx->tes_shader.cso == sel)
3209 return;
3210
3211 sctx->tes_shader.cso = sel;
3212 sctx->tes_shader.current = sel ? sel->first_variant : NULL;
3213 sctx->ia_multi_vgt_param_key.u.uses_tess = sel != NULL;
3214 si_update_tess_uses_prim_id(sctx);
3215
3216 si_update_common_shader_state(sctx);
3217 sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3218
3219 bool ngg_changed = si_update_ngg(sctx);
3220 if (ngg_changed || enable_changed)
3221 si_shader_change_notify(sctx);
3222 if (enable_changed)
3223 sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
3224 si_update_vs_viewport_state(sctx);
3225 si_set_active_descriptors_for_shader(sctx, sel);
3226 si_update_streamout_state(sctx);
3227 si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant,
3228 si_get_vs(sctx)->cso, si_get_vs_state(sctx));
3229 }
3230
3231 static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
3232 {
3233 struct si_context *sctx = (struct si_context *)ctx;
3234 struct si_shader_selector *old_sel = sctx->ps_shader.cso;
3235 struct si_shader_selector *sel = state;
3236
3237 /* skip if supplied shader is one already in use */
3238 if (old_sel == sel)
3239 return;
3240
3241 sctx->ps_shader.cso = sel;
3242 sctx->ps_shader.current = sel ? sel->first_variant : NULL;
3243
3244 si_update_common_shader_state(sctx);
3245 if (sel) {
3246 if (sctx->ia_multi_vgt_param_key.u.uses_tess)
3247 si_update_tess_uses_prim_id(sctx);
3248
3249 if (!old_sel ||
3250 old_sel->info.colors_written != sel->info.colors_written)
3251 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
3252
3253 if (sctx->screen->has_out_of_order_rast &&
3254 (!old_sel ||
3255 old_sel->info.writes_memory != sel->info.writes_memory ||
3256 old_sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] !=
3257 sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]))
3258 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3259 }
3260 si_set_active_descriptors_for_shader(sctx, sel);
3261 si_update_ps_colorbuf0_slot(sctx);
3262 }
3263
3264 static void si_delete_sha