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 "ac_exp_param.h"
26#include "ac_shader_util.h"
27#include "compiler/nir/nir_serialize.h"
28#include "nir/tgsi_to_nir.h"
29#include "si_build_pm4.h"
30#include "sid.h"
31#include "util/crc32.h"
32#include "util/disk_cache.h"
33#include "util/hash_table.h"
34#include "util/mesa-sha1.h"
35#include "util/u_async_debug.h"
36#include "util/u_memory.h"
37#include "util/u_prim.h"
38#include "tgsi/tgsi_from_mesa.h"
39
40/* SHADER_CACHE */
41
42/**
43 * Return the IR key for the shader cache.
44 */
45void si_get_ir_cache_key(struct si_shader_selector *sel, bool ngg, bool es,
46                         unsigned char ir_sha1_cache_key[20])
47{
48   struct blob blob = {};
49   unsigned ir_size;
50   void *ir_binary;
51
52   if (sel->nir_binary) {
53      ir_binary = sel->nir_binary;
54      ir_size = sel->nir_size;
55   } else {
56      assert(sel->nir);
57
58      blob_init(&blob);
59      nir_serialize(&blob, sel->nir, true);
60      ir_binary = blob.data;
61      ir_size = blob.size;
62   }
63
64   /* These settings affect the compilation, but they are not derived
65    * from the input shader IR.
66    */
67   unsigned shader_variant_flags = 0;
68
69   if (ngg)
70      shader_variant_flags |= 1 << 0;
71   if (sel->nir)
72      shader_variant_flags |= 1 << 1;
73   if (si_get_wave_size(sel->screen, sel->info.stage, ngg, es) == 32)
74      shader_variant_flags |= 1 << 2;
75   if (sel->info.stage == MESA_SHADER_FRAGMENT &&
76       /* Derivatives imply helper invocations so check for needs_quad_helper_invocations. */
77       sel->info.base.fs.needs_quad_helper_invocations &&
78       sel->info.base.fs.uses_discard &&
79       sel->screen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL))
80      shader_variant_flags |= 1 << 3;
81   /* use_ngg_culling disables NGG passthrough for non-culling shaders to reduce context
82    * rolls, which can be changed with AMD_DEBUG=nonggc or AMD_DEBUG=nggc.
83    */
84   if (sel->screen->use_ngg_culling)
85      shader_variant_flags |= 1 << 4;
86
87   /* bit gap */
88
89   if (sel->screen->options.no_infinite_interp)
90      shader_variant_flags |= 1 << 7;
91   if (sel->screen->options.clamp_div_by_zero)
92      shader_variant_flags |= 1 << 8;
93   if (sel->screen->debug_flags & DBG(GISEL))
94      shader_variant_flags |= 1 << 9;
95   if ((sel->info.stage == MESA_SHADER_VERTEX ||
96        sel->info.stage == MESA_SHADER_TESS_EVAL ||
97        sel->info.stage == MESA_SHADER_GEOMETRY) &&
98       !es &&
99       sel->screen->options.vrs2x2)
100      shader_variant_flags |= 1 << 10;
101   if (sel->screen->options.inline_uniforms)
102      shader_variant_flags |= 1 << 11;
103
104   struct mesa_sha1 ctx;
105   _mesa_sha1_init(&ctx);
106   _mesa_sha1_update(&ctx, &shader_variant_flags, 4);
107   _mesa_sha1_update(&ctx, ir_binary, ir_size);
108   if (sel->info.stage == MESA_SHADER_VERTEX || sel->info.stage == MESA_SHADER_TESS_EVAL ||
109       sel->info.stage == MESA_SHADER_GEOMETRY)
110      _mesa_sha1_update(&ctx, &sel->so, sizeof(sel->so));
111   _mesa_sha1_final(&ctx, ir_sha1_cache_key);
112
113   if (ir_binary == blob.data)
114      blob_finish(&blob);
115}
116
117/** Copy "data" to "ptr" and return the next dword following copied data. */
118static uint32_t *write_data(uint32_t *ptr, const void *data, unsigned size)
119{
120   /* data may be NULL if size == 0 */
121   if (size)
122      memcpy(ptr, data, size);
123   ptr += DIV_ROUND_UP(size, 4);
124   return ptr;
125}
126
127/** Read data from "ptr". Return the next dword following the data. */
128static uint32_t *read_data(uint32_t *ptr, void *data, unsigned size)
129{
130   memcpy(data, ptr, size);
131   ptr += DIV_ROUND_UP(size, 4);
132   return ptr;
133}
134
135/**
136 * Write the size as uint followed by the data. Return the next dword
137 * following the copied data.
138 */
139static uint32_t *write_chunk(uint32_t *ptr, const void *data, unsigned size)
140{
141   *ptr++ = size;
142   return write_data(ptr, data, size);
143}
144
145/**
146 * Read the size as uint followed by the data. Return both via parameters.
147 * Return the next dword following the data.
148 */
149static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
150{
151   *size = *ptr++;
152   assert(*data == NULL);
153   if (!*size)
154      return ptr;
155   *data = malloc(*size);
156   return read_data(ptr, *data, *size);
157}
158
159/**
160 * Return the shader binary in a buffer. The first 4 bytes contain its size
161 * as integer.
162 */
163static void *si_get_shader_binary(struct si_shader *shader)
164{
165   /* There is always a size of data followed by the data itself. */
166   unsigned llvm_ir_size =
167      shader->binary.llvm_ir_string ? strlen(shader->binary.llvm_ir_string) + 1 : 0;
168
169   /* Refuse to allocate overly large buffers and guard against integer
170    * overflow. */
171   if (shader->binary.elf_size > UINT_MAX / 4 || llvm_ir_size > UINT_MAX / 4)
172      return NULL;
173
174   unsigned size = 4 + /* total size */
175                   4 + /* CRC32 of the data below */
176                   align(sizeof(shader->config), 4) + align(sizeof(shader->info), 4) + 4 +
177                   align(shader->binary.elf_size, 4) + 4 + align(llvm_ir_size, 4);
178   void *buffer = CALLOC(1, size);
179   uint32_t *ptr = (uint32_t *)buffer;
180
181   if (!buffer)
182      return NULL;
183
184   *ptr++ = size;
185   ptr++; /* CRC32 is calculated at the end. */
186
187   ptr = write_data(ptr, &shader->config, sizeof(shader->config));
188   ptr = write_data(ptr, &shader->info, sizeof(shader->info));
189   ptr = write_chunk(ptr, shader->binary.elf_buffer, shader->binary.elf_size);
190   ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
191   assert((char *)ptr - (char *)buffer == size);
192
193   /* Compute CRC32. */
194   ptr = (uint32_t *)buffer;
195   ptr++;
196   *ptr = util_hash_crc32(ptr + 1, size - 8);
197
198   return buffer;
199}
200
201static bool si_load_shader_binary(struct si_shader *shader, void *binary)
202{
203   uint32_t *ptr = (uint32_t *)binary;
204   uint32_t size = *ptr++;
205   uint32_t crc32 = *ptr++;
206   unsigned chunk_size;
207   unsigned elf_size;
208
209   if (util_hash_crc32(ptr, size - 8) != crc32) {
210      fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
211      return false;
212   }
213
214   ptr = read_data(ptr, &shader->config, sizeof(shader->config));
215   ptr = read_data(ptr, &shader->info, sizeof(shader->info));
216   ptr = read_chunk(ptr, (void **)&shader->binary.elf_buffer, &elf_size);
217   shader->binary.elf_size = elf_size;
218   ptr = read_chunk(ptr, (void **)&shader->binary.llvm_ir_string, &chunk_size);
219
220   return true;
221}
222
223/**
224 * Insert a shader into the cache. It's assumed the shader is not in the cache.
225 * Use si_shader_cache_load_shader before calling this.
226 */
227void si_shader_cache_insert_shader(struct si_screen *sscreen, unsigned char ir_sha1_cache_key[20],
228                                   struct si_shader *shader, bool insert_into_disk_cache)
229{
230   void *hw_binary;
231   struct hash_entry *entry;
232   uint8_t key[CACHE_KEY_SIZE];
233   bool memory_cache_full = sscreen->shader_cache_size >= sscreen->shader_cache_max_size;
234
235   if (!insert_into_disk_cache && memory_cache_full)
236      return;
237
238   entry = _mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
239   if (entry)
240      return; /* already added */
241
242   hw_binary = si_get_shader_binary(shader);
243   if (!hw_binary)
244      return;
245
246   if (!memory_cache_full) {
247      if (_mesa_hash_table_insert(sscreen->shader_cache,
248                                  mem_dup(ir_sha1_cache_key, 20),
249                                  hw_binary) == NULL) {
250          FREE(hw_binary);
251          return;
252      }
253      /* The size is stored at the start of the binary */
254      sscreen->shader_cache_size += *(uint32_t*)hw_binary;
255   }
256
257   if (sscreen->disk_shader_cache && insert_into_disk_cache) {
258      disk_cache_compute_key(sscreen->disk_shader_cache, ir_sha1_cache_key, 20, key);
259      disk_cache_put(sscreen->disk_shader_cache, key, hw_binary, *((uint32_t *)hw_binary), NULL);
260   }
261
262   if (memory_cache_full)
263      FREE(hw_binary);
264}
265
266bool si_shader_cache_load_shader(struct si_screen *sscreen, unsigned char ir_sha1_cache_key[20],
267                                 struct si_shader *shader)
268{
269   struct hash_entry *entry = _mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
270
271   if (entry) {
272      if (si_load_shader_binary(shader, entry->data)) {
273         p_atomic_inc(&sscreen->num_memory_shader_cache_hits);
274         return true;
275      }
276   }
277   p_atomic_inc(&sscreen->num_memory_shader_cache_misses);
278
279   if (!sscreen->disk_shader_cache)
280      return false;
281
282   unsigned char sha1[CACHE_KEY_SIZE];
283   disk_cache_compute_key(sscreen->disk_shader_cache, ir_sha1_cache_key, 20, sha1);
284
285   size_t binary_size;
286   uint8_t *buffer = disk_cache_get(sscreen->disk_shader_cache, sha1, &binary_size);
287   if (buffer) {
288      if (binary_size >= sizeof(uint32_t) && *((uint32_t *)buffer) == binary_size) {
289         if (si_load_shader_binary(shader, buffer)) {
290            free(buffer);
291            si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key, shader, false);
292            p_atomic_inc(&sscreen->num_disk_shader_cache_hits);
293            return true;
294         }
295      } else {
296         /* Something has gone wrong discard the item from the cache and
297          * rebuild/link from source.
298          */
299         assert(!"Invalid radeonsi shader disk cache item!");
300         disk_cache_remove(sscreen->disk_shader_cache, sha1);
301      }
302   }
303
304   free(buffer);
305   p_atomic_inc(&sscreen->num_disk_shader_cache_misses);
306   return false;
307}
308
309static uint32_t si_shader_cache_key_hash(const void *key)
310{
311   /* Take the first dword of SHA1. */
312   return *(uint32_t *)key;
313}
314
315static bool si_shader_cache_key_equals(const void *a, const void *b)
316{
317   /* Compare SHA1s. */
318   return memcmp(a, b, 20) == 0;
319}
320
321static void si_destroy_shader_cache_entry(struct hash_entry *entry)
322{
323   FREE((void *)entry->key);
324   FREE(entry->data);
325}
326
327bool si_init_shader_cache(struct si_screen *sscreen)
328{
329   (void)simple_mtx_init(&sscreen->shader_cache_mutex, mtx_plain);
330   sscreen->shader_cache =
331      _mesa_hash_table_create(NULL, si_shader_cache_key_hash, si_shader_cache_key_equals);
332   sscreen->shader_cache_size = 0;
333   /* Maximum size: 64MB on 32 bits, 1GB else */
334   sscreen->shader_cache_max_size = ((sizeof(void *) == 4) ? 64 : 1024) * 1024 * 1024;
335
336   return sscreen->shader_cache != NULL;
337}
338
339void si_destroy_shader_cache(struct si_screen *sscreen)
340{
341   if (sscreen->shader_cache)
342      _mesa_hash_table_destroy(sscreen->shader_cache, si_destroy_shader_cache_entry);
343   simple_mtx_destroy(&sscreen->shader_cache_mutex);
344}
345
346/* SHADER STATES */
347
348bool si_shader_mem_ordered(struct si_shader *shader)
349{
350   if (shader->selector->screen->info.chip_class < GFX10)
351      return false;
352
353   const struct si_shader_info *info = &shader->selector->info;
354   const struct si_shader_info *prev_info =
355      shader->previous_stage_sel ? &shader->previous_stage_sel->info : NULL;
356
357   bool sampler_or_bvh = info->uses_vmem_return_type_sampler_or_bvh;
358   bool other = info->uses_vmem_return_type_other ||
359                info->uses_indirect_descriptor ||
360                shader->config.scratch_bytes_per_wave ||
361                (info->stage == MESA_SHADER_FRAGMENT &&
362                 (info->base.fs.uses_fbfetch_output ||
363                  shader->key.part.ps.prolog.poly_stipple));
364
365   if (prev_info) {
366      sampler_or_bvh |= prev_info->uses_vmem_return_type_sampler_or_bvh;
367      other |= prev_info->uses_vmem_return_type_other ||
368               prev_info->uses_indirect_descriptor;
369   }
370
371   /* Return true if both types of VMEM that return something are used. */
372   return sampler_or_bvh && other;
373}
374
375static void si_set_tesseval_regs(struct si_screen *sscreen, const struct si_shader_selector *tes,
376                                 struct si_shader *shader)
377{
378   const struct si_shader_info *info = &tes->info;
379   unsigned tes_prim_mode = info->base.tess.primitive_mode;
380   unsigned tes_spacing = info->base.tess.spacing;
381   bool tes_vertex_order_cw = !info->base.tess.ccw;
382   bool tes_point_mode = info->base.tess.point_mode;
383   unsigned type, partitioning, topology, distribution_mode;
384
385   switch (tes_prim_mode) {
386   case GL_LINES:
387      type = V_028B6C_TESS_ISOLINE;
388      break;
389   case GL_TRIANGLES:
390      type = V_028B6C_TESS_TRIANGLE;
391      break;
392   case GL_QUADS:
393      type = V_028B6C_TESS_QUAD;
394      break;
395   default:
396      assert(0);
397      return;
398   }
399
400   switch (tes_spacing) {
401   case TESS_SPACING_FRACTIONAL_ODD:
402      partitioning = V_028B6C_PART_FRAC_ODD;
403      break;
404   case TESS_SPACING_FRACTIONAL_EVEN:
405      partitioning = V_028B6C_PART_FRAC_EVEN;
406      break;
407   case TESS_SPACING_EQUAL:
408      partitioning = V_028B6C_PART_INTEGER;
409      break;
410   default:
411      assert(0);
412      return;
413   }
414
415   if (tes_point_mode)
416      topology = V_028B6C_OUTPUT_POINT;
417   else if (tes_prim_mode == GL_LINES)
418      topology = V_028B6C_OUTPUT_LINE;
419   else if (tes_vertex_order_cw)
420      /* for some reason, this must be the other way around */
421      topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
422   else
423      topology = V_028B6C_OUTPUT_TRIANGLE_CW;
424
425   if (sscreen->info.has_distributed_tess) {
426      if (sscreen->info.family == CHIP_FIJI || sscreen->info.family >= CHIP_POLARIS10)
427         distribution_mode = V_028B6C_TRAPEZOIDS;
428      else
429         distribution_mode = V_028B6C_DONUTS;
430   } else
431      distribution_mode = V_028B6C_NO_DIST;
432
433   shader->vgt_tf_param = S_028B6C_TYPE(type) | S_028B6C_PARTITIONING(partitioning) |
434                          S_028B6C_TOPOLOGY(topology) |
435                          S_028B6C_DISTRIBUTION_MODE(distribution_mode);
436}
437
438/* Polaris needs different VTX_REUSE_DEPTH settings depending on
439 * whether the "fractional odd" tessellation spacing is used.
440 *
441 * Possible VGT configurations and which state should set the register:
442 *
443 *   Reg set in | VGT shader configuration   | Value
444 * ------------------------------------------------------
445 *     VS as VS | VS                         | 30
446 *     VS as ES | ES -> GS -> VS             | 30
447 *    TES as VS | LS -> HS -> VS             | 14 or 30
448 *    TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
449 */
450static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen, struct si_shader_selector *sel,
451                                         struct si_shader *shader)
452{
453   if (sscreen->info.family < CHIP_POLARIS10 || sscreen->info.chip_class >= GFX10)
454      return;
455
456   /* VS as VS, or VS as ES: */
457   if ((sel->info.stage == MESA_SHADER_VERTEX &&
458        (!shader->key.as_ls && !shader->is_gs_copy_shader)) ||
459       /* TES as VS, or TES as ES: */
460       sel->info.stage == MESA_SHADER_TESS_EVAL) {
461      unsigned vtx_reuse_depth = 30;
462
463      if (sel->info.stage == MESA_SHADER_TESS_EVAL &&
464          sel->info.base.tess.spacing == TESS_SPACING_FRACTIONAL_ODD)
465         vtx_reuse_depth = 14;
466
467      shader->vgt_vertex_reuse_block_cntl = vtx_reuse_depth;
468   }
469}
470
471static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
472{
473   si_pm4_clear_state(&shader->pm4);
474   shader->pm4.is_shader = true;
475   return &shader->pm4;
476}
477
478static unsigned si_get_num_vs_user_sgprs(struct si_shader *shader,
479                                         unsigned num_always_on_user_sgprs)
480{
481   struct si_shader_selector *vs =
482      shader->previous_stage_sel ? shader->previous_stage_sel : shader->selector;
483   unsigned num_vbos_in_user_sgprs = vs->num_vbos_in_user_sgprs;
484
485   /* 1 SGPR is reserved for the vertex buffer pointer. */
486   assert(num_always_on_user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST - 1);
487
488   if (num_vbos_in_user_sgprs)
489      return SI_SGPR_VS_VB_DESCRIPTOR_FIRST + num_vbos_in_user_sgprs * 4;
490
491   /* Add the pointer to VBO descriptors. */
492   return num_always_on_user_sgprs + 1;
493}
494
495/* Return VGPR_COMP_CNT for the API vertex shader. This can be hw LS, LSHS, ES, ESGS, VS. */
496static unsigned si_get_vs_vgpr_comp_cnt(struct si_screen *sscreen, struct si_shader *shader,
497                                        bool legacy_vs_prim_id)
498{
499   assert(shader->selector->info.stage == MESA_SHADER_VERTEX ||
500          (shader->previous_stage_sel && shader->previous_stage_sel->info.stage == MESA_SHADER_VERTEX));
501
502   /* GFX6-9   LS    (VertexID, RelAutoIndex,           InstanceID / StepRate0, InstanceID)
503    * GFX6-9   ES,VS (VertexID, InstanceID / StepRate0, VSPrimID,               InstanceID)
504    * GFX10    LS    (VertexID, RelAutoIndex,           UserVGPR1,              UserVGPR2 or InstanceID)
505    * GFX10    ES,VS (VertexID, UserVGPR1,              UserVGPR2 or VSPrimID,  UserVGPR3 or InstanceID)
506    */
507   bool is_ls = shader->selector->info.stage == MESA_SHADER_TESS_CTRL || shader->key.as_ls;
508   unsigned max = 0;
509
510   if (shader->info.uses_instanceid) {
511      if (sscreen->info.chip_class >= GFX10)
512         max = MAX2(max, 3);
513      else if (is_ls)
514         max = MAX2(max, 2); /* use (InstanceID / StepRate0) because StepRate0 == 1 */
515      else
516         max = MAX2(max, 1); /* use (InstanceID / StepRate0) because StepRate0 == 1 */
517   }
518
519   if (legacy_vs_prim_id)
520      max = MAX2(max, 2); /* VSPrimID */
521
522   if (is_ls)
523      max = MAX2(max, 1); /* RelAutoIndex */
524
525   return max;
526}
527
528static void si_shader_ls(struct si_screen *sscreen, struct si_shader *shader)
529{
530   struct si_pm4_state *pm4;
531   uint64_t va;
532
533   assert(sscreen->info.chip_class <= GFX8);
534
535   pm4 = si_get_shader_pm4_state(shader);
536   if (!pm4)
537      return;
538
539   va = shader->bo->gpu_address;
540   si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
541
542   shader->config.rsrc1 = S_00B528_VGPRS((shader->config.num_vgprs - 1) / 4) |
543                          S_00B528_SGPRS((shader->config.num_sgprs - 1) / 8) |
544                          S_00B528_VGPR_COMP_CNT(si_get_vs_vgpr_comp_cnt(sscreen, shader, false)) |
545                          S_00B528_DX10_CLAMP(1) | S_00B528_FLOAT_MODE(shader->config.float_mode);
546   shader->config.rsrc2 =
547      S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR)) |
548      S_00B52C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
549}
550
551static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
552{
553   struct si_pm4_state *pm4;
554   uint64_t va;
555
556   pm4 = si_get_shader_pm4_state(shader);
557   if (!pm4)
558      return;
559
560   va = shader->bo->gpu_address;
561
562   if (sscreen->info.chip_class >= GFX9) {
563      if (sscreen->info.chip_class >= GFX10) {
564         si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
565      } else {
566         si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
567      }
568
569      unsigned num_user_sgprs = si_get_num_vs_user_sgprs(shader, GFX9_TCS_NUM_USER_SGPR);
570
571      shader->config.rsrc2 = S_00B42C_USER_SGPR(num_user_sgprs) |
572                             S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
573
574      if (sscreen->info.chip_class >= GFX10)
575         shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
576      else
577         shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
578   } else {
579      si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
580      si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS,
581                     S_00B424_MEM_BASE(sscreen->info.address32_hi >> 8));
582
583      shader->config.rsrc2 = S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR) | S_00B42C_OC_LDS_EN(1) |
584                             S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
585   }
586
587   si_pm4_set_reg(
588      pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
589      S_00B428_VGPRS((shader->config.num_vgprs - 1) / (sscreen->ge_wave_size == 32 ? 8 : 4)) |
590         (sscreen->info.chip_class <= GFX9 ? S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8)
591                                           : 0) |
592         S_00B428_DX10_CLAMP(1) | S_00B428_MEM_ORDERED(si_shader_mem_ordered(shader)) |
593         S_00B428_WGP_MODE(sscreen->info.chip_class >= GFX10) |
594         S_00B428_FLOAT_MODE(shader->config.float_mode) |
595         S_00B428_LS_VGPR_COMP_CNT(sscreen->info.chip_class >= GFX9
596                                      ? si_get_vs_vgpr_comp_cnt(sscreen, shader, false)
597                                      : 0));
598
599   if (sscreen->info.chip_class <= GFX8) {
600      si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS, shader->config.rsrc2);
601   }
602}
603
604static void si_emit_shader_es(struct si_context *sctx)
605{
606   struct si_shader *shader = sctx->queued.named.es;
607   if (!shader)
608      return;
609
610   radeon_begin(&sctx->gfx_cs);
611   radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
612                              SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
613                              shader->selector->esgs_itemsize / 4);
614
615   if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL)
616      radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM, SI_TRACKED_VGT_TF_PARAM,
617                                 shader->vgt_tf_param);
618
619   if (shader->vgt_vertex_reuse_block_cntl)
620      radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
621                                 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
622                                 shader->vgt_vertex_reuse_block_cntl);
623   radeon_end_update_context_roll(sctx);
624}
625
626static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
627{
628   struct si_pm4_state *pm4;
629   unsigned num_user_sgprs;
630   unsigned vgpr_comp_cnt;
631   uint64_t va;
632   unsigned oc_lds_en;
633
634   assert(sscreen->info.chip_class <= GFX8);
635
636   pm4 = si_get_shader_pm4_state(shader);
637   if (!pm4)
638      return;
639
640   pm4->atom.emit = si_emit_shader_es;
641   va = shader->bo->gpu_address;
642
643   if (shader->selector->info.stage == MESA_SHADER_VERTEX) {
644      vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
645      num_user_sgprs = si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR);
646   } else if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL) {
647      vgpr_comp_cnt = shader->selector->info.uses_primid ? 3 : 2;
648      num_user_sgprs = SI_TES_NUM_USER_SGPR;
649   } else
650      unreachable("invalid shader selector type");
651
652   oc_lds_en = shader->selector->info.stage == MESA_SHADER_TESS_EVAL ? 1 : 0;
653
654   si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
655   si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES,
656                  S_00B324_MEM_BASE(sscreen->info.address32_hi >> 8));
657   si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
658                  S_00B328_VGPRS((shader->config.num_vgprs - 1) / 4) |
659                     S_00B328_SGPRS((shader->config.num_sgprs - 1) / 8) |
660                     S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt) | S_00B328_DX10_CLAMP(1) |
661                     S_00B328_FLOAT_MODE(shader->config.float_mode));
662   si_pm4_set_reg(pm4, R_00B32C_SPI_SHADER_PGM_RSRC2_ES,
663                  S_00B32C_USER_SGPR(num_user_sgprs) | S_00B32C_OC_LDS_EN(oc_lds_en) |
664                     S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
665
666   if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL)
667      si_set_tesseval_regs(sscreen, shader->selector, shader);
668
669   polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader);
670}
671
672void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
673                      struct gfx9_gs_info *out)
674{
675   unsigned gs_num_invocations = MAX2(gs->info.base.gs.invocations, 1);
676   unsigned input_prim = gs->info.base.gs.input_primitive;
677   bool uses_adjacency =
678      input_prim >= PIPE_PRIM_LINES_ADJACENCY && input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
679
680   /* All these are in dwords: */
681   /* We can't allow using the whole LDS, because GS waves compete with
682    * other shader stages for LDS space. */
683   const unsigned max_lds_size = 8 * 1024;
684   const unsigned esgs_itemsize = es->esgs_itemsize / 4;
685   unsigned esgs_lds_size;
686
687   /* All these are per subgroup: */
688   const unsigned max_out_prims = 32 * 1024;
689   const unsigned max_es_verts = 255;
690   const unsigned ideal_gs_prims = 64;
691   unsigned max_gs_prims, gs_prims;
692   unsigned min_es_verts, es_verts, worst_case_es_verts;
693
694   if (uses_adjacency || gs_num_invocations > 1)
695      max_gs_prims = 127 / gs_num_invocations;
696   else
697      max_gs_prims = 255;
698
699   /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
700    * Make sure we don't go over the maximum value.
701    */
702   if (gs->info.base.gs.vertices_out > 0) {
703      max_gs_prims =
704         MIN2(max_gs_prims, max_out_prims / (gs->info.base.gs.vertices_out * gs_num_invocations));
705   }
706   assert(max_gs_prims > 0);
707
708   /* If the primitive has adjacency, halve the number of vertices
709    * that will be reused in multiple primitives.
710    */
711   min_es_verts = gs->gs_input_verts_per_prim / (uses_adjacency ? 2 : 1);
712
713   gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
714   worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
715
716   /* Compute ESGS LDS size based on the worst case number of ES vertices
717    * needed to create the target number of GS prims per subgroup.
718    */
719   esgs_lds_size = esgs_itemsize * worst_case_es_verts;
720
721   /* If total LDS usage is too big, refactor partitions based on ratio
722    * of ESGS item sizes.
723    */
724   if (esgs_lds_size > max_lds_size) {
725      /* Our target GS Prims Per Subgroup was too large. Calculate
726       * the maximum number of GS Prims Per Subgroup that will fit
727       * into LDS, capped by the maximum that the hardware can support.
728       */
729      gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)), max_gs_prims);
730      assert(gs_prims > 0);
731      worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
732
733      esgs_lds_size = esgs_itemsize * worst_case_es_verts;
734      assert(esgs_lds_size <= max_lds_size);
735   }
736
737   /* Now calculate remaining ESGS information. */
738   if (esgs_lds_size)
739      es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
740   else
741      es_verts = max_es_verts;
742
743   /* Vertices for adjacency primitives are not always reused, so restore
744    * it for ES_VERTS_PER_SUBGRP.
745    */
746   min_es_verts = gs->gs_input_verts_per_prim;
747
748   /* For normal primitives, the VGT only checks if they are past the ES
749    * verts per subgroup after allocating a full GS primitive and if they
750    * are, kick off a new subgroup.  But if those additional ES verts are
751    * unique (e.g. not reused) we need to make sure there is enough LDS
752    * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
753    */
754   es_verts -= min_es_verts - 1;
755
756   out->es_verts_per_subgroup = es_verts;
757   out->gs_prims_per_subgroup = gs_prims;
758   out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
759   out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup * gs->info.base.gs.vertices_out;
760   out->esgs_ring_size = esgs_lds_size;
761
762   assert(out->max_prims_per_subgroup <= max_out_prims);
763}
764
765static void si_emit_shader_gs(struct si_context *sctx)
766{
767   struct si_shader *shader = sctx->queued.named.gs;
768   if (!shader)
769      return;
770
771   radeon_begin(&sctx->gfx_cs);
772
773   /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
774    * R_028A68_VGT_GSVS_RING_OFFSET_3 */
775   radeon_opt_set_context_reg3(
776      sctx, R_028A60_VGT_GSVS_RING_OFFSET_1, SI_TRACKED_VGT_GSVS_RING_OFFSET_1,
777      shader->ctx_reg.gs.vgt_gsvs_ring_offset_1, shader->ctx_reg.gs.vgt_gsvs_ring_offset_2,
778      shader->ctx_reg.gs.vgt_gsvs_ring_offset_3);
779
780   /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
781   radeon_opt_set_context_reg(sctx, R_028AB0_VGT_GSVS_RING_ITEMSIZE,
782                              SI_TRACKED_VGT_GSVS_RING_ITEMSIZE,
783                              shader->ctx_reg.gs.vgt_gsvs_ring_itemsize);
784
785   /* R_028B38_VGT_GS_MAX_VERT_OUT */
786   radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT, SI_TRACKED_VGT_GS_MAX_VERT_OUT,
787                              shader->ctx_reg.gs.vgt_gs_max_vert_out);
788
789   /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
790    * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
791   radeon_opt_set_context_reg4(
792      sctx, R_028B5C_VGT_GS_VERT_ITEMSIZE, SI_TRACKED_VGT_GS_VERT_ITEMSIZE,
793      shader->ctx_reg.gs.vgt_gs_vert_itemsize, shader->ctx_reg.gs.vgt_gs_vert_itemsize_1,
794      shader->ctx_reg.gs.vgt_gs_vert_itemsize_2, shader->ctx_reg.gs.vgt_gs_vert_itemsize_3);
795
796   /* R_028B90_VGT_GS_INSTANCE_CNT */
797   radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT, 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, SI_TRACKED_VGT_GS_ONCHIP_CNTL,
803                                 shader->ctx_reg.gs.vgt_gs_onchip_cntl);
804      /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
805      radeon_opt_set_context_reg(sctx, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
806                                 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
807                                 shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup);
808      /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
809      radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
810                                 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
811                                 shader->ctx_reg.gs.vgt_esgs_ring_itemsize);
812
813      if (shader->key.part.gs.es->info.stage == MESA_SHADER_TESS_EVAL)
814         radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM, SI_TRACKED_VGT_TF_PARAM,
815                                    shader->vgt_tf_param);
816      if (shader->vgt_vertex_reuse_block_cntl)
817         radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
818                                    SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
819                                    shader->vgt_vertex_reuse_block_cntl);
820   }
821   radeon_end_update_context_roll(sctx);
822
823   /* These don't cause any context rolls. */
824   radeon_begin_again(&sctx->gfx_cs);
825   if (sctx->chip_class >= GFX7) {
826      radeon_opt_set_sh_reg(sctx, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
827                            SI_TRACKED_SPI_SHADER_PGM_RSRC3_GS,
828                            shader->ctx_reg.gs.spi_shader_pgm_rsrc3_gs);
829   }
830   if (sctx->chip_class >= GFX10) {
831      radeon_opt_set_sh_reg(sctx, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
832                            SI_TRACKED_SPI_SHADER_PGM_RSRC4_GS,
833                            shader->ctx_reg.gs.spi_shader_pgm_rsrc4_gs);
834   }
835   radeon_end();
836}
837
838static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
839{
840   struct si_shader_selector *sel = shader->selector;
841   const ubyte *num_components = sel->info.num_stream_output_components;
842   unsigned gs_num_invocations = sel->info.base.gs.invocations;
843   struct si_pm4_state *pm4;
844   uint64_t va;
845   unsigned max_stream = util_last_bit(sel->info.base.gs.active_stream_mask);
846   unsigned offset;
847
848   pm4 = si_get_shader_pm4_state(shader);
849   if (!pm4)
850      return;
851
852   pm4->atom.emit = si_emit_shader_gs;
853
854   offset = num_components[0] * sel->info.base.gs.vertices_out;
855   shader->ctx_reg.gs.vgt_gsvs_ring_offset_1 = offset;
856
857   if (max_stream >= 2)
858      offset += num_components[1] * sel->info.base.gs.vertices_out;
859   shader->ctx_reg.gs.vgt_gsvs_ring_offset_2 = offset;
860
861   if (max_stream >= 3)
862      offset += num_components[2] * sel->info.base.gs.vertices_out;
863   shader->ctx_reg.gs.vgt_gsvs_ring_offset_3 = offset;
864
865   if (max_stream >= 4)
866      offset += num_components[3] * sel->info.base.gs.vertices_out;
867   shader->ctx_reg.gs.vgt_gsvs_ring_itemsize = offset;
868
869   /* The GSVS_RING_ITEMSIZE register takes 15 bits */
870   assert(offset < (1 << 15));
871
872   shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->info.base.gs.vertices_out;
873
874   shader->ctx_reg.gs.vgt_gs_vert_itemsize = num_components[0];
875   shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 2) ? num_components[1] : 0;
876   shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 3) ? num_components[2] : 0;
877   shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 4) ? num_components[3] : 0;
878
879   shader->ctx_reg.gs.vgt_gs_instance_cnt =
880      S_028B90_CNT(MIN2(gs_num_invocations, 127)) | S_028B90_ENABLE(gs_num_invocations > 0);
881
882   /* Copy over fields from the GS copy shader to make them easily accessible from GS. */
883   shader->pa_cl_vs_out_cntl = sel->gs_copy_shader->pa_cl_vs_out_cntl;
884
885   va = shader->bo->gpu_address;
886
887   if (sscreen->info.chip_class >= GFX9) {
888      unsigned input_prim = sel->info.base.gs.input_primitive;
889      gl_shader_stage es_stage = shader->key.part.gs.es->info.stage;
890      unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
891
892      if (es_stage == MESA_SHADER_VERTEX) {
893         es_vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
894      } else if (es_stage == MESA_SHADER_TESS_EVAL)
895         es_vgpr_comp_cnt = shader->key.part.gs.es->info.uses_primid ? 3 : 2;
896      else
897         unreachable("invalid shader selector type");
898
899      /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
900       * VGPR[0:4] are always loaded.
901       */
902      if (sel->info.uses_invocationid)
903         gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
904      else if (sel->info.uses_primid)
905         gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
906      else if (input_prim >= PIPE_PRIM_TRIANGLES)
907         gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
908      else
909         gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
910
911      unsigned num_user_sgprs;
912      if (es_stage == MESA_SHADER_VERTEX)
913         num_user_sgprs = si_get_num_vs_user_sgprs(shader, GFX9_VSGS_NUM_USER_SGPR);
914      else
915         num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
916
917      if (sscreen->info.chip_class >= GFX10) {
918         si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
919      } else {
920         si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
921      }
922
923      uint32_t rsrc1 = S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) | S_00B228_DX10_CLAMP(1) |
924                       S_00B228_MEM_ORDERED(si_shader_mem_ordered(shader)) |
925                       S_00B228_WGP_MODE(sscreen->info.chip_class >= GFX10) |
926                       S_00B228_FLOAT_MODE(shader->config.float_mode) |
927                       S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt);
928      uint32_t rsrc2 = S_00B22C_USER_SGPR(num_user_sgprs) |
929                       S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
930                       S_00B22C_OC_LDS_EN(es_stage == MESA_SHADER_TESS_EVAL) |
931                       S_00B22C_LDS_SIZE(shader->config.lds_size) |
932                       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
933
934      if (sscreen->info.chip_class >= GFX10) {
935         rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
936      } else {
937         rsrc1 |= S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8);
938         rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
939      }
940
941      si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS, rsrc1);
942      si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS, rsrc2);
943
944      shader->ctx_reg.gs.spi_shader_pgm_rsrc3_gs = S_00B21C_CU_EN(0xffff) |
945                                                   S_00B21C_WAVE_LIMIT(0x3F);
946      shader->ctx_reg.gs.spi_shader_pgm_rsrc4_gs =
947         S_00B204_CU_EN(0xffff) | S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(0);
948
949      shader->ctx_reg.gs.vgt_gs_onchip_cntl =
950         S_028A44_ES_VERTS_PER_SUBGRP(shader->gs_info.es_verts_per_subgroup) |
951         S_028A44_GS_PRIMS_PER_SUBGRP(shader->gs_info.gs_prims_per_subgroup) |
952         S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->gs_info.gs_inst_prims_in_subgroup);
953      shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup =
954         S_028A94_MAX_PRIMS_PER_SUBGROUP(shader->gs_info.max_prims_per_subgroup);
955      shader->ctx_reg.gs.vgt_esgs_ring_itemsize = shader->key.part.gs.es->esgs_itemsize / 4;
956
957      if (es_stage == MESA_SHADER_TESS_EVAL)
958         si_set_tesseval_regs(sscreen, shader->key.part.gs.es, shader);
959
960      polaris_set_vgt_vertex_reuse(sscreen, shader->key.part.gs.es, shader);
961   } else {
962      shader->ctx_reg.gs.spi_shader_pgm_rsrc3_gs = S_00B21C_CU_EN(0xffff) |
963                                                   S_00B21C_WAVE_LIMIT(0x3F);
964
965      si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
966      si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS,
967                     S_00B224_MEM_BASE(sscreen->info.address32_hi >> 8));
968
969      si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
970                     S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
971                        S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
972                        S_00B228_DX10_CLAMP(1) | S_00B228_FLOAT_MODE(shader->config.float_mode));
973      si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
974                     S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
975                        S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
976   }
977}
978
979bool gfx10_is_ngg_passthrough(struct si_shader *shader)
980{
981   struct si_shader_selector *sel = shader->selector;
982
983   /* Never use NGG passthrough if culling is possible even when it's not used by this shader,
984    * so that we don't get context rolls when enabling and disabling NGG passthrough.
985    */
986   if (sel->screen->use_ngg_culling)
987      return false;
988
989   /* The definition of NGG passthrough is:
990    * - user GS is turned off (no amplification, no GS instancing, and no culling)
991    * - VGT_ESGS_RING_ITEMSIZE is ignored (behaving as if it was equal to 1)
992    * - vertex indices are packed into 1 VGPR
993    * - Dimgrey and later chips can optionally skip the gs_alloc_req message
994    *
995    * NGG passthrough still allows the use of LDS.
996    */
997   return sel->info.stage != MESA_SHADER_GEOMETRY && !shader->key.opt.ngg_culling;
998}
999
1000/* Common tail code for NGG primitive shaders. */
1001static void gfx10_emit_shader_ngg_tail(struct si_context *sctx, struct si_shader *shader)
1002{
1003   radeon_begin(&sctx->gfx_cs);
1004   radeon_opt_set_context_reg(sctx, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
1005                              SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP,
1006                              shader->ctx_reg.ngg.ge_max_output_per_subgroup);
1007   radeon_opt_set_context_reg(sctx, R_028B4C_GE_NGG_SUBGRP_CNTL, SI_TRACKED_GE_NGG_SUBGRP_CNTL,
1008                              shader->ctx_reg.ngg.ge_ngg_subgrp_cntl);
1009   radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN, SI_TRACKED_VGT_PRIMITIVEID_EN,
1010                              shader->ctx_reg.ngg.vgt_primitiveid_en);
1011   radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL, SI_TRACKED_VGT_GS_ONCHIP_CNTL,
1012                              shader->ctx_reg.ngg.vgt_gs_onchip_cntl);
1013   radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT, SI_TRACKED_VGT_GS_INSTANCE_CNT,
1014                              shader->ctx_reg.ngg.vgt_gs_instance_cnt);
1015   radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
1016                              SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
1017                              shader->ctx_reg.ngg.vgt_esgs_ring_itemsize);
1018   radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG, SI_TRACKED_SPI_VS_OUT_CONFIG,
1019                              shader->ctx_reg.ngg.spi_vs_out_config);
1020   radeon_opt_set_context_reg2(
1021      sctx, R_028708_SPI_SHADER_IDX_FORMAT, SI_TRACKED_SPI_SHADER_IDX_FORMAT,
1022      shader->ctx_reg.ngg.spi_shader_idx_format, shader->ctx_reg.ngg.spi_shader_pos_format);
1023   radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL, SI_TRACKED_PA_CL_VTE_CNTL,
1024                              shader->ctx_reg.ngg.pa_cl_vte_cntl);
1025   radeon_opt_set_context_reg(sctx, R_028838_PA_CL_NGG_CNTL, SI_TRACKED_PA_CL_NGG_CNTL,
1026                              shader->ctx_reg.ngg.pa_cl_ngg_cntl);
1027
1028   radeon_end_update_context_roll(sctx);
1029
1030   /* These don't cause a context roll. */
1031   radeon_begin_again(&sctx->gfx_cs);
1032   radeon_opt_set_uconfig_reg(sctx, R_030980_GE_PC_ALLOC, SI_TRACKED_GE_PC_ALLOC,
1033                              shader->ctx_reg.ngg.ge_pc_alloc);
1034   radeon_opt_set_sh_reg(sctx, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
1035                         SI_TRACKED_SPI_SHADER_PGM_RSRC3_GS,
1036                         shader->ctx_reg.ngg.spi_shader_pgm_rsrc3_gs);
1037   radeon_opt_set_sh_reg(sctx, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
1038                         SI_TRACKED_SPI_SHADER_PGM_RSRC4_GS,
1039                         shader->ctx_reg.ngg.spi_shader_pgm_rsrc4_gs);
1040   radeon_end();
1041}
1042
1043static void gfx10_emit_shader_ngg_notess_nogs(struct si_context *sctx)
1044{
1045   struct si_shader *shader = sctx->queued.named.gs;
1046   if (!shader)
1047      return;
1048
1049   gfx10_emit_shader_ngg_tail(sctx, shader);
1050}
1051
1052static void gfx10_emit_shader_ngg_tess_nogs(struct si_context *sctx)
1053{
1054   struct si_shader *shader = sctx->queued.named.gs;
1055   if (!shader)
1056      return;
1057
1058   radeon_begin(&sctx->gfx_cs);
1059   radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM, SI_TRACKED_VGT_TF_PARAM,
1060                              shader->vgt_tf_param);
1061   radeon_end_update_context_roll(sctx);
1062
1063   gfx10_emit_shader_ngg_tail(sctx, shader);
1064}
1065
1066static void gfx10_emit_shader_ngg_notess_gs(struct si_context *sctx)
1067{
1068   struct si_shader *shader = sctx->queued.named.gs;
1069   if (!shader)
1070      return;
1071
1072   radeon_begin(&sctx->gfx_cs);
1073   radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT, SI_TRACKED_VGT_GS_MAX_VERT_OUT,
1074                              shader->ctx_reg.ngg.vgt_gs_max_vert_out);
1075   radeon_end_update_context_roll(sctx);
1076
1077   gfx10_emit_shader_ngg_tail(sctx, shader);
1078}
1079
1080static void gfx10_emit_shader_ngg_tess_gs(struct si_context *sctx)
1081{
1082   struct si_shader *shader = sctx->queued.named.gs;
1083
1084   if (!shader)
1085      return;
1086
1087   radeon_begin(&sctx->gfx_cs);
1088   radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT, 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, SI_TRACKED_VGT_TF_PARAM,
1091                              shader->vgt_tf_param);
1092   radeon_end_update_context_roll(sctx);
1093
1094   gfx10_emit_shader_ngg_tail(sctx, shader);
1095}
1096
1097unsigned si_get_input_prim(const struct si_shader_selector *gs, const struct si_shader_key *key)
1098{
1099   if (gs->info.stage == MESA_SHADER_GEOMETRY)
1100      return gs->info.base.gs.input_primitive;
1101
1102   if (gs->info.stage == MESA_SHADER_TESS_EVAL) {
1103      if (gs->info.base.tess.point_mode)
1104         return PIPE_PRIM_POINTS;
1105      if (gs->info.base.tess.primitive_mode == GL_LINES)
1106         return PIPE_PRIM_LINES;
1107      return PIPE_PRIM_TRIANGLES;
1108   }
1109
1110   if (key->opt.ngg_culling & SI_NGG_CULL_LINES)
1111      return PIPE_PRIM_LINES;
1112
1113   return PIPE_PRIM_TRIANGLES; /* worst case for all callers */
1114}
1115
1116static unsigned si_get_vs_out_cntl(const struct si_shader_selector *sel,
1117                                   const struct si_shader *shader, bool ngg)
1118{
1119   /* Clip distances can be killed, but cull distances can't. */
1120   unsigned clipcull_mask = (sel->clipdist_mask & ~shader->key.opt.kill_clip_distances) |
1121                            sel->culldist_mask;
1122   bool writes_psize = sel->info.writes_psize && !shader->key.opt.kill_pointsize;
1123   bool misc_vec_ena = writes_psize || (sel->info.writes_edgeflag && !ngg) ||
1124                       sel->screen->options.vrs2x2 ||
1125                       sel->info.writes_layer || sel->info.writes_viewport_index;
1126
1127   return S_02881C_VS_OUT_CCDIST0_VEC_ENA((clipcull_mask & 0x0F) != 0) |
1128          S_02881C_VS_OUT_CCDIST1_VEC_ENA((clipcull_mask & 0xF0) != 0) |
1129          S_02881C_USE_VTX_POINT_SIZE(writes_psize) |
1130          S_02881C_USE_VTX_EDGE_FLAG(sel->info.writes_edgeflag && !ngg) |
1131          S_02881C_USE_VTX_VRS_RATE(sel->screen->options.vrs2x2) |
1132          S_02881C_USE_VTX_RENDER_TARGET_INDX(sel->info.writes_layer) |
1133          S_02881C_USE_VTX_VIEWPORT_INDX(sel->info.writes_viewport_index) |
1134          S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
1135          S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena);
1136}
1137
1138/**
1139 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1140 * in NGG mode.
1141 */
1142static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader)
1143{
1144   const struct si_shader_selector *gs_sel = shader->selector;
1145   const struct si_shader_info *gs_info = &gs_sel->info;
1146   const gl_shader_stage gs_stage = shader->selector->info.stage;
1147   const struct si_shader_selector *es_sel =
1148      shader->previous_stage_sel ? shader->previous_stage_sel : shader->selector;
1149   const struct si_shader_info *es_info = &es_sel->info;
1150   const gl_shader_stage es_stage = es_sel->info.stage;
1151   unsigned num_user_sgprs;
1152   unsigned nparams, es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
1153   uint64_t va;
1154   bool window_space = gs_info->stage == MESA_SHADER_VERTEX ?
1155                          gs_info->base.vs.window_space_position : 0;
1156   bool es_enable_prim_id = shader->key.mono.u.vs_export_prim_id || es_info->uses_primid;
1157   unsigned gs_num_invocations = MAX2(gs_sel->info.base.gs.invocations, 1);
1158   unsigned input_prim = si_get_input_prim(gs_sel, &shader->key);
1159   bool break_wave_at_eoi = false;
1160   struct si_pm4_state *pm4 = si_get_shader_pm4_state(shader);
1161   if (!pm4)
1162      return;
1163
1164   if (es_stage == MESA_SHADER_TESS_EVAL) {
1165      pm4->atom.emit = gs_stage == MESA_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_tess_gs
1166                                                       : gfx10_emit_shader_ngg_tess_nogs;
1167   } else {
1168      pm4->atom.emit = gs_stage == MESA_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_notess_gs
1169                                                       : gfx10_emit_shader_ngg_notess_nogs;
1170   }
1171
1172   va = shader->bo->gpu_address;
1173
1174   if (es_stage == MESA_SHADER_VERTEX) {
1175      es_vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, false);
1176
1177      if (es_info->base.vs.blit_sgprs_amd) {
1178         num_user_sgprs =
1179            SI_SGPR_VS_BLIT_DATA + es_info->base.vs.blit_sgprs_amd;
1180      } else {
1181         num_user_sgprs = si_get_num_vs_user_sgprs(shader, GFX9_VSGS_NUM_USER_SGPR);
1182      }
1183   } else {
1184      assert(es_stage == MESA_SHADER_TESS_EVAL);
1185      es_vgpr_comp_cnt = es_enable_prim_id ? 3 : 2;
1186      num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
1187
1188      if (es_enable_prim_id || gs_info->uses_primid)
1189         break_wave_at_eoi = true;
1190   }
1191
1192   /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1193    * VGPR[0:4] are always loaded.
1194    *
1195    * Vertex shaders always need to load VGPR3, because they need to
1196    * pass edge flags for decomposed primitives (such as quads) to the PA
1197    * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1198    */
1199   if (gs_info->uses_invocationid ||
1200       (gfx10_edgeflags_have_effect(shader) && !gfx10_is_ngg_passthrough(shader)))
1201      gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID, edge flags. */
1202   else if ((gs_stage == MESA_SHADER_GEOMETRY && gs_info->uses_primid) ||
1203            (gs_stage == MESA_SHADER_VERTEX && shader->key.mono.u.vs_export_prim_id))
1204      gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
1205   else if (input_prim >= PIPE_PRIM_TRIANGLES && !gfx10_is_ngg_passthrough(shader))
1206      gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
1207   else
1208      gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
1209
1210   unsigned wave_size = si_get_shader_wave_size(shader);
1211   unsigned late_alloc_wave64, cu_mask;
1212
1213   ac_compute_late_alloc(&sscreen->info, true, shader->key.opt.ngg_culling,
1214                         shader->config.scratch_bytes_per_wave > 0,
1215                         &late_alloc_wave64, &cu_mask);
1216
1217   si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
1218   si_pm4_set_reg(
1219      pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
1220      S_00B228_VGPRS((shader->config.num_vgprs - 1) / (wave_size == 32 ? 8 : 4)) |
1221         S_00B228_FLOAT_MODE(shader->config.float_mode) | S_00B228_DX10_CLAMP(1) |
1222         S_00B228_MEM_ORDERED(si_shader_mem_ordered(shader)) |
1223         /* Disable the WGP mode on gfx10.3 because it can hang. (it happened on VanGogh)
1224          * Let's disable it on all chips that disable exactly 1 CU per SA for GS. */
1225         S_00B228_WGP_MODE(sscreen->info.chip_class == GFX10) |
1226         S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
1227   si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
1228                  S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0) |
1229                     S_00B22C_USER_SGPR(num_user_sgprs) |
1230                     S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
1231                     S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5) |
1232                     S_00B22C_OC_LDS_EN(es_stage == MESA_SHADER_TESS_EVAL) |
1233                     S_00B22C_LDS_SIZE(shader->config.lds_size));
1234
1235   shader->ctx_reg.ngg.spi_shader_pgm_rsrc3_gs = S_00B21C_CU_EN(cu_mask) |
1236                                                 S_00B21C_WAVE_LIMIT(0x3F);
1237   shader->ctx_reg.ngg.spi_shader_pgm_rsrc4_gs =
1238      S_00B204_CU_EN(0xffff) | S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64);
1239
1240   nparams = MAX2(shader->info.nr_param_exports, 1);
1241   shader->ctx_reg.ngg.spi_vs_out_config =
1242      S_0286C4_VS_EXPORT_COUNT(nparams - 1) |
1243      S_0286C4_NO_PC_EXPORT(shader->info.nr_param_exports == 0);
1244
1245   shader->ctx_reg.ngg.spi_shader_idx_format =
1246      S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP);
1247   shader->ctx_reg.ngg.spi_shader_pos_format =
1248      S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
1249      S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ? V_02870C_SPI_SHADER_4COMP
1250                                                                  : V_02870C_SPI_SHADER_NONE) |
1251      S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ? V_02870C_SPI_SHADER_4COMP
1252                                                                  : V_02870C_SPI_SHADER_NONE) |
1253      S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ? V_02870C_SPI_SHADER_4COMP
1254                                                                  : V_02870C_SPI_SHADER_NONE);
1255
1256   shader->ctx_reg.ngg.vgt_primitiveid_en =
1257      S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
1258      S_028A84_NGG_DISABLE_PROVOK_REUSE(shader->key.mono.u.vs_export_prim_id ||
1259                                        gs_sel->info.writes_primid);
1260
1261   if (gs_stage == MESA_SHADER_GEOMETRY) {
1262      shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = es_sel->esgs_itemsize / 4;
1263      shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->info.base.gs.vertices_out;
1264   } else {
1265      shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = 1;
1266   }
1267
1268   if (es_stage == MESA_SHADER_TESS_EVAL)
1269      si_set_tesseval_regs(sscreen, es_sel, shader);
1270
1271   shader->ctx_reg.ngg.vgt_gs_onchip_cntl =
1272      S_028A44_ES_VERTS_PER_SUBGRP(shader->ngg.hw_max_esverts) |
1273      S_028A44_GS_PRIMS_PER_SUBGRP(shader->ngg.max_gsprims) |
1274      S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->ngg.max_gsprims * gs_num_invocations);
1275   shader->ctx_reg.ngg.ge_max_output_per_subgroup =
1276      S_0287FC_MAX_VERTS_PER_SUBGROUP(shader->ngg.max_out_verts);
1277   shader->ctx_reg.ngg.ge_ngg_subgrp_cntl = S_028B4C_PRIM_AMP_FACTOR(shader->ngg.prim_amp_factor) |
1278                                            S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1279   shader->ctx_reg.ngg.vgt_gs_instance_cnt =
1280      S_028B90_CNT(gs_num_invocations) | S_028B90_ENABLE(gs_num_invocations > 1) |
1281      S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(shader->ngg.max_vert_out_per_gs_instance);
1282
1283   /* Output hw-generated edge flags if needed and pass them via the prim
1284    * export to prevent drawing lines on internal edges of decomposed
1285    * primitives (such as quads) with polygon mode = lines.
1286    */
1287   shader->ctx_reg.ngg.pa_cl_ngg_cntl =
1288      S_028838_INDEX_BUF_EDGE_FLAG_ENA(gfx10_edgeflags_have_effect(shader)) |
1289      /* Reuse for NGG. */
1290      S_028838_VERTEX_REUSE_DEPTH(sscreen->info.chip_class >= GFX10_3 ? 30 : 0);
1291   shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, shader, true);
1292
1293   /* Oversubscribe PC. This improves performance when there are too many varyings. */
1294   unsigned oversub_pc_factor = 1;
1295
1296   if (shader->key.opt.ngg_culling) {
1297      /* Be more aggressive with NGG culling. */
1298      if (shader->info.nr_param_exports > 4)
1299         oversub_pc_factor = 4;
1300      else if (shader->info.nr_param_exports > 2)
1301         oversub_pc_factor = 3;
1302      else
1303         oversub_pc_factor = 2;
1304   }
1305
1306   unsigned oversub_pc_lines =
1307      late_alloc_wave64 ? (sscreen->info.pc_lines / 4) * oversub_pc_factor : 0;
1308   shader->ctx_reg.ngg.ge_pc_alloc = S_030980_OVERSUB_EN(oversub_pc_lines > 0) |
1309                                     S_030980_NUM_PC_LINES(oversub_pc_lines - 1);
1310
1311   shader->ge_cntl = S_03096C_PRIM_GRP_SIZE(shader->ngg.max_gsprims) |
1312                     S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts) |
1313                     S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
1314
1315   /* On gfx10, the GE only checks against the maximum number of ES verts after
1316    * allocating a full GS primitive. So we need to ensure that whenever
1317    * this check passes, there is enough space for a full primitive without
1318    * vertex reuse. VERT_GRP_SIZE=256 doesn't need this. We should always get 256
1319    * if we have enough LDS.
1320    *
1321    * Tessellation is unaffected because it always sets GE_CNTL.VERT_GRP_SIZE = 0.
1322    */
1323   if ((sscreen->info.chip_class == GFX10) &&
1324       (es_stage == MESA_SHADER_VERTEX || gs_stage == MESA_SHADER_VERTEX) && /* = no tess */
1325       shader->ngg.hw_max_esverts != 256 &&
1326       shader->ngg.hw_max_esverts > 5) {
1327      /* This could be based on the input primitive type. 5 is the worst case
1328       * for primitive types with adjacency.
1329       */
1330      shader->ge_cntl &= C_03096C_VERT_GRP_SIZE;
1331      shader->ge_cntl |= S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts - 5);
1332   }
1333
1334   if (window_space) {
1335      shader->ctx_reg.ngg.pa_cl_vte_cntl = S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1336   } else {
1337      shader->ctx_reg.ngg.pa_cl_vte_cntl =
1338         S_028818_VTX_W0_FMT(1) | S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1339         S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1340         S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1341   }
1342
1343   shader->ctx_reg.ngg.vgt_stages.u.ngg = 1;
1344   shader->ctx_reg.ngg.vgt_stages.u.streamout = gs_sel->so.num_outputs;
1345   shader->ctx_reg.ngg.vgt_stages.u.ngg_passthrough = gfx10_is_ngg_passthrough(shader);
1346}
1347
1348static void si_emit_shader_vs(struct si_context *sctx)
1349{
1350   struct si_shader *shader = sctx->queued.named.vs;
1351   if (!shader)
1352      return;
1353
1354   radeon_begin(&sctx->gfx_cs);
1355   radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE, SI_TRACKED_VGT_GS_MODE,
1356                              shader->ctx_reg.vs.vgt_gs_mode);
1357   radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN, SI_TRACKED_VGT_PRIMITIVEID_EN,
1358                              shader->ctx_reg.vs.vgt_primitiveid_en);
1359
1360   if (sctx->chip_class <= GFX8) {
1361      radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF, SI_TRACKED_VGT_REUSE_OFF,
1362                                 shader->ctx_reg.vs.vgt_reuse_off);
1363   }
1364
1365   radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG, SI_TRACKED_SPI_VS_OUT_CONFIG,
1366                              shader->ctx_reg.vs.spi_vs_out_config);
1367
1368   radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
1369                              SI_TRACKED_SPI_SHADER_POS_FORMAT,
1370                              shader->ctx_reg.vs.spi_shader_pos_format);
1371
1372   radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL, SI_TRACKED_PA_CL_VTE_CNTL,
1373                              shader->ctx_reg.vs.pa_cl_vte_cntl);
1374
1375   if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL)
1376      radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM, SI_TRACKED_VGT_TF_PARAM,
1377                                 shader->vgt_tf_param);
1378
1379   if (shader->vgt_vertex_reuse_block_cntl)
1380      radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
1381                                 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
1382                                 shader->vgt_vertex_reuse_block_cntl);
1383
1384   /* Required programming for tessellation. (legacy pipeline only) */
1385   if (sctx->chip_class >= GFX10 && shader->selector->info.stage == MESA_SHADER_TESS_EVAL) {
1386      radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
1387                                 SI_TRACKED_VGT_GS_ONCHIP_CNTL,
1388                                 S_028A44_ES_VERTS_PER_SUBGRP(250) |
1389                                 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
1390                                 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
1391   }
1392
1393   radeon_end_update_context_roll(sctx);
1394
1395   /* GE_PC_ALLOC is not a context register, so it doesn't cause a context roll. */
1396   if (sctx->chip_class >= GFX10) {
1397      radeon_begin_again(&sctx->gfx_cs);
1398      radeon_opt_set_uconfig_reg(sctx, R_030980_GE_PC_ALLOC, SI_TRACKED_GE_PC_ALLOC,
1399                                 shader->ctx_reg.vs.ge_pc_alloc);
1400      radeon_end();
1401   }
1402}
1403
1404/**
1405 * Compute the state for \p shader, which will run as a vertex shader on the
1406 * hardware.
1407 *
1408 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1409 * is the copy shader.
1410 */
1411static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
1412                         struct si_shader_selector *gs)
1413{
1414   const struct si_shader_info *info = &shader->selector->info;
1415   struct si_pm4_state *pm4;
1416   unsigned num_user_sgprs, vgpr_comp_cnt;
1417   uint64_t va;
1418   unsigned nparams, oc_lds_en;
1419   bool window_space = info->stage == MESA_SHADER_VERTEX ?
1420                          info->base.vs.window_space_position : 0;
1421   bool enable_prim_id = shader->key.mono.u.vs_export_prim_id || info->uses_primid;
1422
1423   pm4 = si_get_shader_pm4_state(shader);
1424   if (!pm4)
1425      return;
1426
1427   pm4->atom.emit = si_emit_shader_vs;
1428
1429   /* We always write VGT_GS_MODE in the VS state, because every switch
1430    * between different shader pipelines involving a different GS or no
1431    * GS at all involves a switch of the VS (different GS use different
1432    * copy shaders). On the other hand, when the API switches from a GS to
1433    * no GS and then back to the same GS used originally, the GS state is
1434    * not sent again.
1435    */
1436   if (!gs) {
1437      unsigned mode = V_028A40_GS_OFF;
1438
1439      /* PrimID needs GS scenario A. */
1440      if (enable_prim_id)
1441         mode = V_028A40_GS_SCENARIO_A;
1442
1443      shader->ctx_reg.vs.vgt_gs_mode = S_028A40_MODE(mode);
1444      shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
1445   } else {
1446      shader->ctx_reg.vs.vgt_gs_mode =
1447         ac_vgt_gs_mode(gs->info.base.gs.vertices_out, sscreen->info.chip_class);
1448      shader->ctx_reg.vs.vgt_primitiveid_en = 0;
1449   }
1450
1451   if (sscreen->info.chip_class <= GFX8) {
1452      /* Reuse needs to be set off if we write oViewport. */
1453      shader->ctx_reg.vs.vgt_reuse_off = S_028AB4_REUSE_OFF(info->writes_viewport_index);
1454   }
1455
1456   va = shader->bo->gpu_address;
1457
1458   if (gs) {
1459      vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
1460      num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
1461   } else if (shader->selector->info.stage == MESA_SHADER_VERTEX) {
1462      vgpr_comp_cnt = si_get_vs_vgpr_comp_cnt(sscreen, shader, enable_prim_id);
1463
1464      if (info->base.vs.blit_sgprs_amd) {
1465         num_user_sgprs = SI_SGPR_VS_BLIT_DATA + info->base.vs.blit_sgprs_amd;
1466      } else {
1467         num_user_sgprs = si_get_num_vs_user_sgprs(shader, SI_VS_NUM_USER_SGPR);
1468      }
1469   } else if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL) {
1470      vgpr_comp_cnt = enable_prim_id ? 3 : 2;
1471      num_user_sgprs = SI_TES_NUM_USER_SGPR;
1472   } else
1473      unreachable("invalid shader selector type");
1474
1475   /* VS is required to export at least one param. */
1476   nparams = MAX2(shader->info.nr_param_exports, 1);
1477   shader->ctx_reg.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
1478
1479   if (sscreen->info.chip_class >= GFX10) {
1480      shader->ctx_reg.vs.spi_vs_out_config |=
1481         S_0286C4_NO_PC_EXPORT(shader->info.nr_param_exports == 0);
1482   }
1483
1484   shader->ctx_reg.vs.spi_shader_pos_format =
1485      S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
1486      S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ? V_02870C_SPI_SHADER_4COMP
1487                                                                  : V_02870C_SPI_SHADER_NONE) |
1488      S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ? V_02870C_SPI_SHADER_4COMP
1489                                                                  : V_02870C_SPI_SHADER_NONE) |
1490      S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ? V_02870C_SPI_SHADER_4COMP
1491                                                                  : V_02870C_SPI_SHADER_NONE);
1492   unsigned late_alloc_wave64, cu_mask;
1493   ac_compute_late_alloc(&sscreen->info, false, false,
1494                         shader->config.scratch_bytes_per_wave > 0,
1495                         &late_alloc_wave64, &cu_mask);
1496
1497   shader->ctx_reg.vs.ge_pc_alloc = S_030980_OVERSUB_EN(late_alloc_wave64 > 0) |
1498                                    S_030980_NUM_PC_LINES(sscreen->info.pc_lines / 4 - 1);
1499   shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, shader, false);
1500
1501   oc_lds_en = shader->selector->info.stage == MESA_SHADER_TESS_EVAL ? 1 : 0;
1502
1503   if (sscreen->info.chip_class >= GFX7) {
1504      si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
1505                     S_00B118_CU_EN(cu_mask) | S_00B118_WAVE_LIMIT(0x3F));
1506      si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(late_alloc_wave64));
1507   }
1508
1509   si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
1510   si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS,
1511                  S_00B124_MEM_BASE(sscreen->info.address32_hi >> 8));
1512
1513   uint32_t rsrc1 =
1514      S_00B128_VGPRS((shader->config.num_vgprs - 1) / (sscreen->ge_wave_size == 32 ? 8 : 4)) |
1515      S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) | S_00B128_DX10_CLAMP(1) |
1516      S_00B128_MEM_ORDERED(si_shader_mem_ordered(shader)) |
1517      S_00B128_FLOAT_MODE(shader->config.float_mode);
1518   uint32_t rsrc2 = S_00B12C_USER_SGPR(num_user_sgprs) | S_00B12C_OC_LDS_EN(oc_lds_en) |
1519                    S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
1520
1521   if (sscreen->info.chip_class >= GFX10)
1522      rsrc2 |= S_00B12C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
1523   else if (sscreen->info.chip_class == GFX9)
1524      rsrc2 |= S_00B12C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
1525
1526   if (sscreen->info.chip_class <= GFX9)
1527      rsrc1 |= S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8);
1528
1529   if (!sscreen->use_ngg_streamout) {
1530      rsrc2 |= S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
1531               S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
1532               S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
1533               S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
1534               S_00B12C_SO_EN(!!shader->selector->so.num_outputs);
1535   }
1536
1537   si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS, rsrc1);
1538   si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS, rsrc2);
1539
1540   if (window_space)
1541      shader->ctx_reg.vs.pa_cl_vte_cntl = S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1542   else
1543      shader->ctx_reg.vs.pa_cl_vte_cntl =
1544         S_028818_VTX_W0_FMT(1) | S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1545         S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1546         S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1547
1548   if (shader->selector->info.stage == MESA_SHADER_TESS_EVAL)
1549      si_set_tesseval_regs(sscreen, shader->selector, shader);
1550
1551   polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader);
1552}
1553
1554static unsigned si_get_ps_num_interp(struct si_shader *ps)
1555{
1556   struct si_shader_info *info = &ps->selector->info;
1557   unsigned num_colors = !!(info->colors_read & 0x0f) + !!(info->colors_read & 0xf0);
1558   unsigned num_interp =
1559      ps->selector->info.num_inputs + (ps->key.part.ps.prolog.color_two_side ? num_colors : 0);
1560
1561   assert(num_interp <= 32);
1562   return MIN2(num_interp, 32);
1563}
1564
1565static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
1566{
1567   unsigned spi_shader_col_format = shader->key.part.ps.epilog.spi_shader_col_format;
1568   unsigned value = 0, num_mrts = 0;
1569   unsigned i, num_targets = (util_last_bit(spi_shader_col_format) + 3) / 4;
1570
1571   /* Remove holes in spi_shader_col_format. */
1572   for (i = 0; i < num_targets; i++) {
1573      unsigned spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
1574
1575      if (spi_format) {
1576         value |= spi_format << (num_mrts * 4);
1577         num_mrts++;
1578      }
1579   }
1580
1581   return value;
1582}
1583
1584static void si_emit_shader_ps(struct si_context *sctx)
1585{
1586   struct si_shader *shader = sctx->queued.named.ps;
1587   if (!shader)
1588      return;
1589
1590   radeon_begin(&sctx->gfx_cs);
1591   /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1592   radeon_opt_set_context_reg2(sctx, R_0286CC_SPI_PS_INPUT_ENA, SI_TRACKED_SPI_PS_INPUT_ENA,
1593                               shader->ctx_reg.ps.spi_ps_input_ena,
1594                               shader->ctx_reg.ps.spi_ps_input_addr);
1595
1596   radeon_opt_set_context_reg(sctx, R_0286E0_SPI_BARYC_CNTL, SI_TRACKED_SPI_BARYC_CNTL,
1597                              shader->ctx_reg.ps.spi_baryc_cntl);
1598   radeon_opt_set_context_reg(sctx, R_0286D8_SPI_PS_IN_CONTROL, SI_TRACKED_SPI_PS_IN_CONTROL,
1599                              shader->ctx_reg.ps.spi_ps_in_control);
1600
1601   /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1602   radeon_opt_set_context_reg2(sctx, R_028710_SPI_SHADER_Z_FORMAT, SI_TRACKED_SPI_SHADER_Z_FORMAT,
1603                               shader->ctx_reg.ps.spi_shader_z_format,
1604                               shader->ctx_reg.ps.spi_shader_col_format);
1605
1606   radeon_opt_set_context_reg(sctx, R_02823C_CB_SHADER_MASK, SI_TRACKED_CB_SHADER_MASK,
1607                              shader->ctx_reg.ps.cb_shader_mask);
1608   radeon_end_update_context_roll(sctx);
1609}
1610
1611static void si_shader_ps(struct si_screen *sscreen, struct si_shader *shader)
1612{
1613   struct si_shader_info *info = &shader->selector->info;
1614   struct si_pm4_state *pm4;
1615   unsigned spi_ps_in_control, spi_shader_col_format, cb_shader_mask;
1616   unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
1617   uint64_t va;
1618   unsigned input_ena = shader->config.spi_ps_input_ena;
1619
1620   /* we need to enable at least one of them, otherwise we hang the GPU */
1621   assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) || G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1622          G_0286CC_PERSP_CENTROID_ENA(input_ena) || G_0286CC_PERSP_PULL_MODEL_ENA(input_ena) ||
1623          G_0286CC_LINEAR_SAMPLE_ENA(input_ena) || G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1624          G_0286CC_LINEAR_CENTROID_ENA(input_ena) || G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena));
1625   /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1626   assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena) || G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1627          G_0286CC_PERSP_CENTER_ENA(input_ena) || G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1628          G_0286CC_PERSP_PULL_MODEL_ENA(input_ena));
1629
1630   /* Validate interpolation optimization flags (read as implications). */
1631   assert(!shader->key.part.ps.prolog.bc_optimize_for_persp ||
1632          (G_0286CC_PERSP_CENTER_ENA(input_ena) && G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1633   assert(!shader->key.part.ps.prolog.bc_optimize_for_linear ||
1634          (G_0286CC_LINEAR_CENTER_ENA(input_ena) && G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1635   assert(!shader->key.part.ps.prolog.force_persp_center_interp ||
1636          (!G_0286CC_PERSP_SAMPLE_ENA(input_ena) && !G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1637   assert(!shader->key.part.ps.prolog.force_linear_center_interp ||
1638          (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena) && !G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1639   assert(!shader->key.part.ps.prolog.force_persp_sample_interp ||
1640          (!G_0286CC_PERSP_CENTER_ENA(input_ena) && !G_0286CC_PERSP_CENTROID_ENA(input_ena)));
1641   assert(!shader->key.part.ps.prolog.force_linear_sample_interp ||
1642          (!G_0286CC_LINEAR_CENTER_ENA(input_ena) && !G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
1643
1644   /* Validate cases when the optimizations are off (read as implications). */
1645   assert(shader->key.part.ps.prolog.bc_optimize_for_persp ||
1646          !G_0286CC_PERSP_CENTER_ENA(input_ena) || !G_0286CC_PERSP_CENTROID_ENA(input_ena));
1647   assert(shader->key.part.ps.prolog.bc_optimize_for_linear ||
1648          !G_0286CC_LINEAR_CENTER_ENA(input_ena) || !G_0286CC_LINEAR_CENTROID_ENA(input_ena));
1649
1650   pm4 = si_get_shader_pm4_state(shader);
1651   if (!pm4)
1652      return;
1653
1654   /* If multiple state sets are allowed to be in a bin, break the batch on a new PS. */
1655   if (sscreen->dpbb_allowed &&
1656       (sscreen->pbb_context_states_per_bin > 1 ||
1657        sscreen->pbb_persistent_states_per_bin > 1)) {
1658      si_pm4_cmd_add(pm4, PKT3(PKT3_EVENT_WRITE, 0, 0));
1659      si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
1660   }
1661
1662   pm4->atom.emit = si_emit_shader_ps;
1663
1664   /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1665    * Possible vaules:
1666    * 0 -> Position = pixel center
1667    * 1 -> Position = pixel centroid
1668    * 2 -> Position = at sample position
1669    *
1670    * From GLSL 4.5 specification, section 7.1:
1671    *   "The variable gl_FragCoord is available as an input variable from
1672    *    within fragment shaders and it holds the window relative coordinates
1673    *    (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1674    *    value can be for any location within the pixel, or one of the
1675    *    fragment samples. The use of centroid does not further restrict
1676    *    this value to be inside the current primitive."
1677    *
1678    * Meaning that centroid has no effect and we can return anything within
1679    * the pixel. Thus, return the value at sample position, because that's
1680    * the most accurate one shaders can get.
1681    */
1682   spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1683
1684   if (info->base.fs.pixel_center_integer)
1685      spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
1686
1687   spi_shader_col_format = si_get_spi_shader_col_format(shader);
1688   cb_shader_mask = ac_get_cb_shader_mask(shader->key.part.ps.epilog.spi_shader_col_format);
1689
1690   /* Ensure that some export memory is always allocated, for two reasons:
1691    *
1692    * 1) Correctness: The hardware ignores the EXEC mask if no export
1693    *    memory is allocated, so KILL and alpha test do not work correctly
1694    *    without this.
1695    * 2) Performance: Every shader needs at least a NULL export, even when
1696    *    it writes no color/depth output. The NULL export instruction
1697    *    stalls without this setting.
1698    *
1699    * Don't add this to CB_SHADER_MASK.
1700    *
1701    * GFX10 supports pixel shaders without exports by setting both
1702    * the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
1703    * instructions if any are present.
1704    */
1705   if ((sscreen->info.chip_class <= GFX9 || info->base.fs.uses_discard ||
1706        shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS) &&
1707       !spi_shader_col_format && !info->writes_z && !info->writes_stencil &&
1708       !info->writes_samplemask)
1709      spi_shader_col_format = V_028714_SPI_SHADER_32_R;
1710
1711   shader->ctx_reg.ps.spi_ps_input_ena = input_ena;
1712   shader->ctx_reg.ps.spi_ps_input_addr = shader->config.spi_ps_input_addr;
1713
1714   unsigned num_interp = si_get_ps_num_interp(shader);
1715
1716   /* Set interpolation controls. */
1717   spi_ps_in_control = S_0286D8_NUM_INTERP(num_interp) |
1718                       S_0286D8_PS_W32_EN(sscreen->ps_wave_size == 32);
1719
1720   shader->ctx_reg.ps.num_interp = num_interp;
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, info->writes_stencil, info->writes_samplemask);
1725   shader->ctx_reg.ps.spi_shader_col_format = spi_shader_col_format;
1726   shader->ctx_reg.ps.cb_shader_mask = cb_shader_mask;
1727
1728   va = shader->bo->gpu_address;
1729   si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
1730   si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS,
1731                  S_00B024_MEM_BASE(sscreen->info.address32_hi >> 8));
1732
1733   uint32_t rsrc1 =
1734      S_00B028_VGPRS((shader->config.num_vgprs - 1) / (sscreen->ps_wave_size == 32 ? 8 : 4)) |
1735      S_00B028_DX10_CLAMP(1) | S_00B028_MEM_ORDERED(si_shader_mem_ordered(shader)) |
1736      S_00B028_FLOAT_MODE(shader->config.float_mode);
1737
1738   if (sscreen->info.chip_class < GFX10) {
1739      rsrc1 |= S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8);
1740   }
1741
1742   si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS, rsrc1);
1743   si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
1744                  S_00B02C_EXTRA_LDS_SIZE(shader->config.lds_size) |
1745                     S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR) |
1746                     S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
1747}
1748
1749static void si_shader_init_pm4_state(struct si_screen *sscreen, struct si_shader *shader)
1750{
1751   switch (shader->selector->info.stage) {
1752   case MESA_SHADER_VERTEX:
1753      if (shader->key.as_ls)
1754         si_shader_ls(sscreen, shader);
1755      else if (shader->key.as_es)
1756         si_shader_es(sscreen, shader);
1757      else if (shader->key.as_ngg)
1758         gfx10_shader_ngg(sscreen, shader);
1759      else
1760         si_shader_vs(sscreen, shader, NULL);
1761      break;
1762   case MESA_SHADER_TESS_CTRL:
1763      si_shader_hs(sscreen, shader);
1764      break;
1765   case MESA_SHADER_TESS_EVAL:
1766      if (shader->key.as_es)
1767         si_shader_es(sscreen, shader);
1768      else if (shader->key.as_ngg)
1769         gfx10_shader_ngg(sscreen, shader);
1770      else
1771         si_shader_vs(sscreen, shader, NULL);
1772      break;
1773   case MESA_SHADER_GEOMETRY:
1774      if (shader->key.as_ngg)
1775         gfx10_shader_ngg(sscreen, shader);
1776      else
1777         si_shader_gs(sscreen, shader);
1778      break;
1779   case MESA_SHADER_FRAGMENT:
1780      si_shader_ps(sscreen, shader);
1781      break;
1782   default:
1783      assert(0);
1784   }
1785}
1786
1787static void si_clear_vs_key_inputs(struct si_context *sctx, struct si_shader_key *key,
1788                                   struct si_vs_prolog_bits *prolog_key)
1789{
1790   prolog_key->instance_divisor_is_one = 0;
1791   prolog_key->instance_divisor_is_fetched = 0;
1792   key->mono.vs_fetch_opencode = 0;
1793   memset(key->mono.vs_fix_fetch, 0, sizeof(key->mono.vs_fix_fetch));
1794}
1795
1796void si_vs_key_update_inputs(struct si_context *sctx)
1797{
1798   struct si_shader_selector *vs = sctx->shader.vs.cso;
1799   struct si_vertex_elements *elts = sctx->vertex_elements;
1800   struct si_shader_key *key = &sctx->shader.vs.key;
1801
1802   if (!vs)
1803      return;
1804
1805   if (vs->info.base.vs.blit_sgprs_amd) {
1806      si_clear_vs_key_inputs(sctx, key, &key->part.vs.prolog);
1807      key->opt.prefer_mono = 0;
1808      sctx->uses_nontrivial_vs_prolog = false;
1809      return;
1810   }
1811
1812   bool uses_nontrivial_vs_prolog = false;
1813
1814   if (elts->instance_divisor_is_one || elts->instance_divisor_is_fetched)
1815      uses_nontrivial_vs_prolog = true;
1816
1817   key->part.vs.prolog.instance_divisor_is_one = elts->instance_divisor_is_one;
1818   key->part.vs.prolog.instance_divisor_is_fetched = elts->instance_divisor_is_fetched;
1819   key->opt.prefer_mono = elts->instance_divisor_is_fetched;
1820
1821   unsigned count_mask = (1 << vs->info.num_inputs) - 1;
1822   unsigned fix = elts->fix_fetch_always & count_mask;
1823   unsigned opencode = elts->fix_fetch_opencode & count_mask;
1824
1825   if (sctx->vertex_buffer_unaligned & elts->vb_alignment_check_mask) {
1826      uint32_t mask = elts->fix_fetch_unaligned & count_mask;
1827      while (mask) {
1828         unsigned i = u_bit_scan(&mask);
1829         unsigned log_hw_load_size = 1 + ((elts->hw_load_is_dword >> i) & 1);
1830         unsigned vbidx = elts->vertex_buffer_index[i];
1831         struct pipe_vertex_buffer *vb = &sctx->vertex_buffer[vbidx];
1832         unsigned align_mask = (1 << log_hw_load_size) - 1;
1833         if (vb->buffer_offset & align_mask || vb->stride & align_mask) {
1834            fix |= 1 << i;
1835            opencode |= 1 << i;
1836         }
1837      }
1838   }
1839
1840   memset(key->mono.vs_fix_fetch, 0, sizeof(key->mono.vs_fix_fetch));
1841
1842   while (fix) {
1843      unsigned i = u_bit_scan(&fix);
1844      uint8_t fix_fetch = elts->fix_fetch[i];
1845
1846      key->mono.vs_fix_fetch[i].bits = fix_fetch;
1847      if (fix_fetch)
1848         uses_nontrivial_vs_prolog = true;
1849   }
1850   key->mono.vs_fetch_opencode = opencode;
1851   if (opencode)
1852      uses_nontrivial_vs_prolog = true;
1853
1854   sctx->uses_nontrivial_vs_prolog = uses_nontrivial_vs_prolog;
1855
1856   /* draw_vertex_state (display lists) requires a trivial VS prolog that ignores
1857    * the current vertex buffers and vertex elements.
1858    *
1859    * We just computed the prolog key because we needed to set uses_nontrivial_vs_prolog,
1860    * so that we know whether the VS prolog should be updated when we switch from
1861    * draw_vertex_state to draw_vbo. Now clear the VS prolog for draw_vertex_state.
1862    * This should happen rarely because the VS prolog should be trivial in most
1863    * cases.
1864    */
1865   if (uses_nontrivial_vs_prolog && sctx->force_trivial_vs_prolog)
1866      si_clear_vs_key_inputs(sctx, key, &key->part.vs.prolog);
1867}
1868
1869void si_get_vs_key_inputs(struct si_context *sctx, struct si_shader_key *key,
1870                          struct si_vs_prolog_bits *prolog_key)
1871{
1872   prolog_key->instance_divisor_is_one = sctx->shader.vs.key.part.vs.prolog.instance_divisor_is_one;
1873   prolog_key->instance_divisor_is_fetched = sctx->shader.vs.key.part.vs.prolog.instance_divisor_is_fetched;
1874
1875   key->mono.vs_fetch_opencode = sctx->shader.vs.key.mono.vs_fetch_opencode;
1876   memcpy(key->mono.vs_fix_fetch, sctx->shader.vs.key.mono.vs_fix_fetch,
1877          sizeof(key->mono.vs_fix_fetch));
1878}
1879
1880void si_update_ps_inputs_read_or_disabled(struct si_context *sctx)
1881{
1882   struct si_shader_selector *ps = sctx->shader.ps.cso;
1883
1884   /* Find out if PS is disabled. */
1885   bool ps_disabled = true;
1886   if (ps) {
1887      bool ps_modifies_zs = ps->info.base.fs.uses_discard || ps->info.writes_z || ps->info.writes_stencil ||
1888                            ps->info.writes_samplemask ||
1889                            sctx->queued.named.blend->alpha_to_coverage ||
1890                            sctx->queued.named.dsa->alpha_func != PIPE_FUNC_ALWAYS;
1891      unsigned ps_colormask = si_get_total_colormask(sctx);
1892
1893      ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
1894                    (!ps_colormask && !ps_modifies_zs && !ps->info.base.writes_memory);
1895   }
1896
1897   sctx->ps_inputs_read_or_disabled = ps_disabled ? 0 : ps->inputs_read;
1898}
1899
1900static void si_get_vs_key_outputs(struct si_context *sctx, struct si_shader_selector *vs,
1901                                  struct si_shader_key *key)
1902{
1903
1904   key->opt.kill_clip_distances = vs->clipdist_mask & ~sctx->queued.named.rasterizer->clip_plane_enable;
1905
1906   /* Find out which VS outputs aren't used by the PS. */
1907   uint64_t outputs_written = vs->outputs_written_before_ps;
1908   uint64_t linked = outputs_written & sctx->ps_inputs_read_or_disabled;
1909
1910   key->opt.kill_outputs = ~linked & outputs_written;
1911
1912   if (vs->info.stage != MESA_SHADER_GEOMETRY) {
1913      key->opt.ngg_culling = sctx->ngg_culling;
1914      key->mono.u.vs_export_prim_id = sctx->shader.ps.cso && sctx->shader.ps.cso->info.uses_primid;
1915   } else {
1916      key->opt.ngg_culling = 0;
1917      key->mono.u.vs_export_prim_id = 0;
1918   }
1919
1920   key->opt.kill_pointsize = vs->info.writes_psize &&
1921                             sctx->current_rast_prim != PIPE_PRIM_POINTS &&
1922                             !sctx->queued.named.rasterizer->polygon_mode_is_points;
1923}
1924
1925static void si_clear_vs_key_outputs(struct si_context *sctx, struct si_shader_selector *vs,
1926                                    struct si_shader_key *key)
1927{
1928   key->opt.kill_clip_distances = 0;
1929   key->opt.kill_outputs = 0;
1930   key->opt.ngg_culling = 0;
1931   key->mono.u.vs_export_prim_id = 0;
1932   key->opt.kill_pointsize = 0;
1933}
1934
1935void si_ps_key_update_framebuffer(struct si_context *sctx)
1936{
1937   struct si_shader_selector *sel = sctx->shader.ps.cso;
1938   struct si_shader_key *key = &sctx->shader.ps.key;
1939
1940   if (!sel)
1941      return;
1942
1943   if (sel->info.color0_writes_all_cbufs &&
1944       sel->info.colors_written == 0x1)
1945      key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
1946   else
1947      key->part.ps.epilog.last_cbuf = 0;
1948
1949   /* ps_uses_fbfetch is true only if the color buffer is bound. */
1950   if (sctx->ps_uses_fbfetch) {
1951      struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
1952      struct pipe_resource *tex = cb0->texture;
1953
1954      /* 1D textures are allocated and used as 2D on GFX9. */
1955      key->mono.u.ps.fbfetch_msaa = sctx->framebuffer.nr_samples > 1;
1956      key->mono.u.ps.fbfetch_is_1D =
1957         sctx->chip_class != GFX9 &&
1958         (tex->target == PIPE_TEXTURE_1D || tex->target == PIPE_TEXTURE_1D_ARRAY);
1959      key->mono.u.ps.fbfetch_layered =
1960         tex->target == PIPE_TEXTURE_1D_ARRAY || tex->target == PIPE_TEXTURE_2D_ARRAY ||
1961         tex->target == PIPE_TEXTURE_CUBE || tex->target == PIPE_TEXTURE_CUBE_ARRAY ||
1962         tex->target == PIPE_TEXTURE_3D;
1963   } else {
1964      key->mono.u.ps.fbfetch_msaa = 0;
1965      key->mono.u.ps.fbfetch_is_1D = 0;
1966      key->mono.u.ps.fbfetch_layered = 0;
1967   }
1968}
1969
1970void si_ps_key_update_framebuffer_blend(struct si_context *sctx)
1971{
1972   struct si_shader_selector *sel = sctx->shader.ps.cso;
1973   struct si_shader_key *key = &sctx->shader.ps.key;
1974   struct si_state_blend *blend = sctx->queued.named.blend;
1975
1976   if (!sel)
1977      return;
1978
1979   /* Select the shader color format based on whether
1980    * blending or alpha are needed.
1981    */
1982   key->part.ps.epilog.spi_shader_col_format =
1983      (blend->blend_enable_4bit & blend->need_src_alpha_4bit &
1984       sctx->framebuffer.spi_shader_col_format_blend_alpha) |
1985      (blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
1986       sctx->framebuffer.spi_shader_col_format_blend) |
1987      (~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
1988       sctx->framebuffer.spi_shader_col_format_alpha) |
1989      (~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
1990       sctx->framebuffer.spi_shader_col_format);
1991   key->part.ps.epilog.spi_shader_col_format &= blend->cb_target_enabled_4bit;
1992
1993   /* The output for dual source blending should have
1994    * the same format as the first output.
1995    */
1996   if (blend->dual_src_blend) {
1997      key->part.ps.epilog.spi_shader_col_format |=
1998         (key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
1999   }
2000
2001   /* If alpha-to-coverage is enabled, we have to export alpha
2002    * even if there is no color buffer.
2003    */
2004   if (!(key->part.ps.epilog.spi_shader_col_format & 0xf) && blend->alpha_to_coverage)
2005      key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
2006
2007   /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
2008    * to the range supported by the type if a channel has less
2009    * than 16 bits and the export format is 16_ABGR.
2010    */
2011   if (sctx->chip_class <= GFX7 && sctx->family != CHIP_HAWAII) {
2012      key->part.ps.epilog.color_is_int8 = sctx->framebuffer.color_is_int8;
2013      key->part.ps.epilog.color_is_int10 = sctx->framebuffer.color_is_int10;
2014   }
2015
2016   /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
2017   if (!key->part.ps.epilog.last_cbuf) {
2018      key->part.ps.epilog.spi_shader_col_format &= sel->colors_written_4bit;
2019      key->part.ps.epilog.color_is_int8 &= sel->info.colors_written;
2020      key->part.ps.epilog.color_is_int10 &= sel->info.colors_written;
2021   }
2022
2023   /* Eliminate shader code computing output values that are unused.
2024    * This enables dead code elimination between shader parts.
2025    * Check if any output is eliminated.
2026    */
2027   if (sel->colors_written_4bit &
2028       ~(sctx->framebuffer.colorbuf_enabled_4bit & blend->cb_target_enabled_4bit))
2029      key->opt.prefer_mono = 1;
2030   else
2031      key->opt.prefer_mono = 0;
2032}
2033
2034void si_ps_key_update_blend_rasterizer(struct si_context *sctx)
2035{
2036   struct si_shader_key *key = &sctx->shader.ps.key;
2037   struct si_state_blend *blend = sctx->queued.named.blend;
2038   struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2039
2040   key->part.ps.epilog.alpha_to_one = blend->alpha_to_one && rs->multisample_enable;
2041}
2042
2043void si_ps_key_update_rasterizer(struct si_context *sctx)
2044{
2045   struct si_shader_selector *sel = sctx->shader.ps.cso;
2046   struct si_shader_key *key = &sctx->shader.ps.key;
2047   struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2048
2049   if (!sel)
2050      return;
2051
2052   key->part.ps.prolog.color_two_side = rs->two_side && sel->info.colors_read;
2053   key->part.ps.prolog.flatshade_colors = rs->flatshade && sel->info.uses_interp_color;
2054   key->part.ps.epilog.clamp_color = rs->clamp_fragment_color;
2055}
2056
2057void si_ps_key_update_dsa(struct si_context *sctx)
2058{
2059   struct si_shader_key *key = &sctx->shader.ps.key;
2060
2061   key->part.ps.epilog.alpha_func = sctx->queued.named.dsa->alpha_func;
2062}
2063
2064static void si_ps_key_update_primtype_shader_rasterizer_framebuffer(struct si_context *sctx)
2065{
2066   struct si_shader_key *key = &sctx->shader.ps.key;
2067   struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2068
2069   bool is_poly = !util_prim_is_points_or_lines(sctx->current_rast_prim);
2070   bool is_line = util_prim_is_lines(sctx->current_rast_prim);
2071
2072   key->part.ps.prolog.poly_stipple = rs->poly_stipple_enable && is_poly;
2073   key->part.ps.epilog.poly_line_smoothing =
2074      ((is_poly && rs->poly_smooth) || (is_line && rs->line_smooth)) &&
2075      sctx->framebuffer.nr_samples <= 1;
2076}
2077
2078void si_ps_key_update_sample_shading(struct si_context *sctx)
2079{
2080   struct si_shader_selector *sel = sctx->shader.ps.cso;
2081   struct si_shader_key *key = &sctx->shader.ps.key;
2082
2083   if (!sel)
2084      return;
2085
2086   if (sctx->ps_iter_samples > 1 && sel->info.reads_samplemask)
2087      key->part.ps.prolog.samplemask_log_ps_iter = util_logbase2(sctx->ps_iter_samples);
2088   else
2089      key->part.ps.prolog.samplemask_log_ps_iter = 0;
2090}
2091
2092void si_ps_key_update_framebuffer_rasterizer_sample_shading(struct si_context *sctx)
2093{
2094   struct si_shader_selector *sel = sctx->shader.ps.cso;
2095   struct si_shader_key *key = &sctx->shader.ps.key;
2096   struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2097
2098   if (!sel)
2099      return;
2100
2101   bool uses_persp_center = sel->info.uses_persp_center ||
2102                            (!rs->flatshade && sel->info.uses_persp_center_color);
2103   bool uses_persp_centroid = sel->info.uses_persp_centroid ||
2104                              (!rs->flatshade && sel->info.uses_persp_centroid_color);
2105   bool uses_persp_sample = sel->info.uses_persp_sample ||
2106                            (!rs->flatshade && sel->info.uses_persp_sample_color);
2107
2108   if (rs->force_persample_interp && rs->multisample_enable &&
2109       sctx->framebuffer.nr_samples > 1 && sctx->ps_iter_samples > 1) {
2110      key->part.ps.prolog.force_persp_sample_interp =
2111         uses_persp_center || uses_persp_centroid;
2112
2113      key->part.ps.prolog.force_linear_sample_interp =
2114         sel->info.uses_linear_center || sel->info.uses_linear_centroid;
2115
2116      key->part.ps.prolog.force_persp_center_interp = 0;
2117      key->part.ps.prolog.force_linear_center_interp = 0;
2118      key->part.ps.prolog.bc_optimize_for_persp = 0;
2119      key->part.ps.prolog.bc_optimize_for_linear = 0;
2120      key->mono.u.ps.interpolate_at_sample_force_center = 0;
2121   } else if (rs->multisample_enable && sctx->framebuffer.nr_samples > 1) {
2122      key->part.ps.prolog.force_persp_sample_interp = 0;
2123      key->part.ps.prolog.force_linear_sample_interp = 0;
2124      key->part.ps.prolog.force_persp_center_interp = 0;
2125      key->part.ps.prolog.force_linear_center_interp = 0;
2126      key->part.ps.prolog.bc_optimize_for_persp =
2127         uses_persp_center && uses_persp_centroid;
2128      key->part.ps.prolog.bc_optimize_for_linear =
2129         sel->info.uses_linear_center && sel->info.uses_linear_centroid;
2130      key->mono.u.ps.interpolate_at_sample_force_center = 0;
2131   } else {
2132      key->part.ps.prolog.force_persp_sample_interp = 0;
2133      key->part.ps.prolog.force_linear_sample_interp = 0;
2134
2135      /* Make sure SPI doesn't compute more than 1 pair
2136       * of (i,j), which is the optimization here. */
2137      key->part.ps.prolog.force_persp_center_interp = uses_persp_center +
2138                                                      uses_persp_centroid +
2139                                                      uses_persp_sample > 1;
2140
2141      key->part.ps.prolog.force_linear_center_interp = sel->info.uses_linear_center +
2142                                                       sel->info.uses_linear_centroid +
2143                                                       sel->info.uses_linear_sample > 1;
2144      key->part.ps.prolog.bc_optimize_for_persp = 0;
2145      key->part.ps.prolog.bc_optimize_for_linear = 0;
2146      key->mono.u.ps.interpolate_at_sample_force_center = sel->info.uses_interp_at_sample;
2147   }
2148}
2149
2150/* Compute the key for the hw shader variant */
2151static inline void si_shader_selector_key(struct pipe_context *ctx, struct si_shader_selector *sel,
2152                                          struct si_shader_key *key)
2153{
2154   struct si_context *sctx = (struct si_context *)ctx;
2155
2156   switch (sel->info.stage) {
2157   case MESA_SHADER_VERTEX:
2158      if (!sctx->shader.tes.cso && !sctx->shader.gs.cso)
2159         si_get_vs_key_outputs(sctx, sel, key);
2160      else
2161         si_clear_vs_key_outputs(sctx, sel, key);
2162      break;
2163   case MESA_SHADER_TESS_CTRL:
2164      if (sctx->chip_class >= GFX9) {
2165         si_get_vs_key_inputs(sctx, key, &key->part.tcs.ls_prolog);
2166         key->part.tcs.ls = sctx->shader.vs.cso;
2167      }
2168      break;
2169   case MESA_SHADER_TESS_EVAL:
2170      if (!sctx->shader.gs.cso)
2171         si_get_vs_key_outputs(sctx, sel, key);
2172      else
2173         si_clear_vs_key_outputs(sctx, sel, key);
2174      break;
2175   case MESA_SHADER_GEOMETRY:
2176      if (sctx->chip_class >= GFX9) {
2177         if (sctx->shader.tes.cso) {
2178            si_clear_vs_key_inputs(sctx, key, &key->part.gs.vs_prolog);
2179            key->part.gs.es = sctx->shader.tes.cso;
2180         } else {
2181            si_get_vs_key_inputs(sctx, key, &key->part.gs.vs_prolog);
2182            key->part.gs.es = sctx->shader.vs.cso;
2183         }
2184
2185         /* Only NGG can eliminate GS outputs, because the code is shared with VS. */
2186         if (sctx->ngg)
2187            si_get_vs_key_outputs(sctx, sel, key);
2188         else
2189            si_clear_vs_key_outputs(sctx, sel, key);
2190      }
2191      break;
2192   case MESA_SHADER_FRAGMENT:
2193      si_ps_key_update_primtype_shader_rasterizer_framebuffer(sctx);
2194      break;
2195   default:
2196      assert(0);
2197   }
2198}
2199
2200static void si_build_shader_variant(struct si_shader *shader, int thread_index, bool low_priority)
2201{
2202   struct si_shader_selector *sel = shader->selector;
2203   struct si_screen *sscreen = sel->screen;
2204   struct ac_llvm_compiler *compiler;
2205   struct pipe_debug_callback *debug = &shader->compiler_ctx_state.debug;
2206
2207   if (thread_index >= 0) {
2208      if (low_priority) {
2209         assert(thread_index < ARRAY_SIZE(sscreen->compiler_lowp));
2210         compiler = &sscreen->compiler_lowp[thread_index];
2211      } else {
2212         assert(thread_index < ARRAY_SIZE(sscreen->compiler));
2213         compiler = &sscreen->compiler[thread_index];
2214      }
2215      if (!debug->async)
2216         debug = NULL;
2217   } else {
2218      assert(!low_priority);
2219      compiler = shader->compiler_ctx_state.compiler;
2220   }
2221
2222   if (!compiler->passes)
2223      si_init_compiler(sscreen, compiler);
2224
2225   if (unlikely(!si_create_shader_variant(sscreen, compiler, shader, debug))) {
2226      PRINT_ERR("Failed to build shader variant (type=%u)\n", sel->info.stage);
2227      shader->compilation_failed = true;
2228      return;
2229   }
2230
2231   if (shader->compiler_ctx_state.is_debug_context) {
2232      FILE *f = open_memstream(&shader->shader_log, &shader->shader_log_size);
2233      if (f) {
2234         si_shader_dump(sscreen, shader, NULL, f, false);
2235         fclose(f);
2236      }
2237   }
2238
2239   si_shader_init_pm4_state(sscreen, shader);
2240}
2241
2242static void si_build_shader_variant_low_priority(void *job, void *gdata, int thread_index)
2243{
2244   struct si_shader *shader = (struct si_shader *)job;
2245
2246   assert(thread_index >= 0);
2247
2248   si_build_shader_variant(shader, thread_index, true);
2249}
2250
2251static const struct si_shader_key zeroed;
2252
2253static bool si_check_missing_main_part(struct si_screen *sscreen, struct si_shader_selector *sel,
2254                                       struct si_compiler_ctx_state *compiler_state,
2255                                       const struct si_shader_key *key)
2256{
2257   struct si_shader **mainp = si_get_main_shader_part(sel, key);
2258
2259   if (!*mainp) {
2260      struct si_shader *main_part = CALLOC_STRUCT(si_shader);
2261
2262      if (!main_part)
2263         return false;
2264
2265      /* We can leave the fence as permanently signaled because the
2266       * main part becomes visible globally only after it has been
2267       * compiled. */
2268      util_queue_fence_init(&main_part->ready);
2269
2270      main_part->selector = sel;
2271      main_part->key.as_es = key->as_es;
2272      main_part->key.as_ls = key->as_ls;
2273      main_part->key.as_ngg = key->as_ngg;
2274      main_part->is_monolithic = false;
2275
2276      if (!si_compile_shader(sscreen, compiler_state->compiler, main_part,
2277                             &compiler_state->debug)) {
2278         FREE(main_part);
2279         return false;
2280      }
2281      *mainp = main_part;
2282   }
2283   return true;
2284}
2285
2286/* A helper to copy *key to *local_key and return local_key. */
2287static const struct si_shader_key *
2288use_local_key_copy(const struct si_shader_key *key, struct si_shader_key *local_key)
2289{
2290   if (key != local_key)
2291      memcpy(local_key, key, sizeof(*key));
2292
2293   return local_key;
2294}
2295
2296/**
2297 * Select a shader variant according to the shader key.
2298 *
2299 * \param optimized_or_none  If the key describes an optimized shader variant and
2300 *                           the compilation isn't finished, don't select any
2301 *                           shader and return an error.
2302 */
2303int si_shader_select_with_key(struct si_context *sctx, struct si_shader_ctx_state *state,
2304                              const struct si_shader_key *key, int thread_index,
2305                              bool optimized_or_none)
2306{
2307   struct si_screen *sscreen = sctx->screen;
2308   struct si_shader_selector *sel = state->cso;
2309   struct si_shader_selector *previous_stage_sel = NULL;
2310   struct si_shader *current = state->current;
2311   struct si_shader *iter, *shader = NULL;
2312   /* si_shader_select_with_key must not modify 'key' because it would affect future shaders.
2313    * If we need to modify it for this specific shader (eg: to disable optimizations), we
2314    * use a copy.
2315    */
2316   struct si_shader_key local_key;
2317
2318   if (unlikely(sscreen->debug_flags & DBG(NO_OPT_VARIANT))) {
2319      /* Disable shader variant optimizations. */
2320      key = use_local_key_copy(key, &local_key);
2321      memset(&local_key.opt, 0, sizeof(key->opt));
2322   }
2323
2324again:
2325   /* Check if we don't need to change anything.
2326    * This path is also used for most shaders that don't need multiple
2327    * variants, it will cost just a computation of the key and this
2328    * test. */
2329   if (likely(current && memcmp(&current->key, key, sizeof(*key)) == 0)) {
2330      if (unlikely(!util_queue_fence_is_signalled(&current->ready))) {
2331         if (current->is_optimized) {
2332            if (optimized_or_none)
2333               return -1;
2334
2335            key = use_local_key_copy(key, &local_key);
2336            memset(&local_key.opt, 0, sizeof(key->opt));
2337            goto current_not_ready;
2338         }
2339
2340         util_queue_fence_wait(&current->ready);
2341      }
2342
2343      return current->compilation_failed ? -1 : 0;
2344   }
2345current_not_ready:
2346
2347   /* This must be done before the mutex is locked, because async GS
2348    * compilation calls this function too, and therefore must enter
2349    * the mutex first.
2350    *
2351    * Only wait if we are in a draw call. Don't wait if we are
2352    * in a compiler thread.
2353    */
2354   if (thread_index < 0)
2355      util_queue_fence_wait(&sel->ready);
2356
2357   simple_mtx_lock(&sel->mutex);
2358
2359   /* Compute the size of the key without the uniform values. */
2360   size_t s = (void*)&key->opt.inlined_uniform_values - (void*)key;
2361   int variant_count = 0;
2362   const int max_inline_uniforms_variants = 5;
2363
2364   /* Find the shader variant. */
2365   for (iter = sel->first_variant; iter; iter = iter->next_variant) {
2366      if (memcmp(&iter->key, key, s) == 0) {
2367         /* Check the inlined uniform values separatly, and count
2368          * the number of variants based on them.
2369          */
2370         if (key->opt.inline_uniforms &&
2371             memcmp(iter->key.opt.inlined_uniform_values,
2372                    key->opt.inlined_uniform_values,
2373                    MAX_INLINABLE_UNIFORMS * 4) != 0) {
2374            if (variant_count++ > max_inline_uniforms_variants) {
2375               key = use_local_key_copy(key, &local_key);
2376               /* Too many variants. Disable inlining for this shader. */
2377               local_key.opt.inline_uniforms = 0;
2378               memset(local_key.opt.inlined_uniform_values, 0, MAX_INLINABLE_UNIFORMS * 4);
2379               simple_mtx_unlock(&sel->mutex);
2380               goto again;
2381            }
2382            continue;
2383         }
2384
2385         simple_mtx_unlock(&sel->mutex);
2386
2387         if (unlikely(!util_queue_fence_is_signalled(&iter->ready))) {
2388            /* If it's an optimized shader and its compilation has
2389             * been started but isn't done, use the unoptimized
2390             * shader so as not to cause a stall due to compilation.
2391             */
2392            if (iter->is_optimized) {
2393               if (optimized_or_none)
2394                  return -1;
2395
2396               key = use_local_key_copy(key, &local_key);
2397               memset(&local_key.opt, 0, sizeof(key->opt));
2398               goto again;
2399            }
2400
2401            util_queue_fence_wait(&iter->ready);
2402         }
2403
2404         if (iter->compilation_failed) {
2405            return -1; /* skip the draw call */
2406         }
2407
2408         state->current = iter;
2409         return 0;
2410      }
2411   }
2412
2413   /* Build a new shader. */
2414   shader = CALLOC_STRUCT(si_shader);
2415   if (!shader) {
2416      simple_mtx_unlock(&sel->mutex);
2417      return -ENOMEM;
2418   }
2419
2420   util_queue_fence_init(&shader->ready);
2421
2422   if (!sctx->compiler.passes)
2423      si_init_compiler(sctx->screen, &sctx->compiler);
2424
2425   shader->selector = sel;
2426   shader->key = *key;
2427   shader->compiler_ctx_state.compiler = &sctx->compiler;
2428   shader->compiler_ctx_state.debug = sctx->debug;
2429   shader->compiler_ctx_state.is_debug_context = sctx->is_debug;
2430
2431   /* If this is a merged shader, get the first shader's selector. */
2432   if (sscreen->info.chip_class >= GFX9) {
2433      if (sel->info.stage == MESA_SHADER_TESS_CTRL)
2434         previous_stage_sel = key->part.tcs.ls;
2435      else if (sel->info.stage == MESA_SHADER_GEOMETRY)
2436         previous_stage_sel = key->part.gs.es;
2437
2438      /* We need to wait for the previous shader. */
2439      if (previous_stage_sel && thread_index < 0)
2440         util_queue_fence_wait(&previous_stage_sel->ready);
2441   }
2442
2443   bool is_pure_monolithic =
2444      sscreen->use_monolithic_shaders || memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
2445
2446   /* Compile the main shader part if it doesn't exist. This can happen
2447    * if the initial guess was wrong.
2448    */
2449   if (!is_pure_monolithic) {
2450      bool ok = true;
2451
2452      /* Make sure the main shader part is present. This is needed
2453       * for shaders that can be compiled as VS, LS, or ES, and only
2454       * one of them is compiled at creation.
2455       *
2456       * It is also needed for GS, which can be compiled as non-NGG
2457       * and NGG.
2458       *
2459       * For merged shaders, check that the starting shader's main
2460       * part is present.
2461       */
2462      if (previous_stage_sel) {
2463         struct si_shader_key shader1_key = zeroed;
2464
2465         if (sel->info.stage == MESA_SHADER_TESS_CTRL) {
2466            shader1_key.as_ls = 1;
2467         } else if (sel->info.stage == MESA_SHADER_GEOMETRY) {
2468            shader1_key.as_es = 1;
2469            shader1_key.as_ngg = key->as_ngg; /* for Wave32 vs Wave64 */
2470         } else {
2471            assert(0);
2472         }
2473
2474         simple_mtx_lock(&previous_stage_sel->mutex);
2475         ok = si_check_missing_main_part(sscreen, previous_stage_sel, &shader->compiler_ctx_state,
2476                                         &shader1_key);
2477         simple_mtx_unlock(&previous_stage_sel->mutex);
2478      }
2479
2480      if (ok) {
2481         ok = si_check_missing_main_part(sscreen, sel, &shader->compiler_ctx_state, key);
2482      }
2483
2484      if (!ok) {
2485         FREE(shader);
2486         simple_mtx_unlock(&sel->mutex);
2487         return -ENOMEM; /* skip the draw call */
2488      }
2489   }
2490
2491   /* Keep the reference to the 1st shader of merged shaders, so that
2492    * Gallium can't destroy it before we destroy the 2nd shader.
2493    *
2494    * Set sctx = NULL, because it's unused if we're not releasing
2495    * the shader, and we don't have any sctx here.
2496    */
2497   si_shader_selector_reference(NULL, &shader->previous_stage_sel, previous_stage_sel);
2498
2499   /* Monolithic-only shaders don't make a distinction between optimized
2500    * and unoptimized. */
2501   shader->is_monolithic =
2502      is_pure_monolithic || memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
2503
2504   shader->is_optimized = !is_pure_monolithic &&
2505                          memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
2506
2507   /* If it's an optimized shader, compile it asynchronously. */
2508   if (shader->is_optimized && thread_index < 0) {
2509      /* Compile it asynchronously. */
2510      util_queue_add_job(&sscreen->shader_compiler_queue_low_priority, shader, &shader->ready,
2511                         si_build_shader_variant_low_priority, NULL, 0);
2512
2513      /* Add only after the ready fence was reset, to guard against a
2514       * race with si_bind_XX_shader. */
2515      if (!sel->last_variant) {
2516         sel->first_variant = shader;
2517         sel->last_variant = shader;
2518      } else {
2519         sel->last_variant->next_variant = shader;
2520         sel->last_variant = shader;
2521      }
2522
2523      /* Use the default (unoptimized) shader for now. */
2524      key = use_local_key_copy(key, &local_key);
2525      memset(&local_key.opt, 0, sizeof(key->opt));
2526      simple_mtx_unlock(&sel->mutex);
2527
2528      if (sscreen->options.sync_compile)
2529         util_queue_fence_wait(&shader->ready);
2530
2531      if (optimized_or_none)
2532         return -1;
2533      goto again;
2534   }
2535
2536   /* Reset the fence before adding to the variant list. */
2537   util_queue_fence_reset(&shader->ready);
2538
2539   if (!sel->last_variant) {
2540      sel->first_variant = shader;
2541      sel->last_variant = shader;
2542   } else {
2543      sel->last_variant->next_variant = shader;
2544      sel->last_variant = shader;
2545   }
2546
2547   simple_mtx_unlock(&sel->mutex);
2548
2549   assert(!shader->is_optimized);
2550   si_build_shader_variant(shader, thread_index, false);
2551
2552   util_queue_fence_signal(&shader->ready);
2553
2554   if (!shader->compilation_failed)
2555      state->current = shader;
2556
2557   return shader->compilation_failed ? -1 : 0;
2558}
2559
2560int si_shader_select(struct pipe_context *ctx, struct si_shader_ctx_state *state)
2561{
2562   struct si_context *sctx = (struct si_context *)ctx;
2563
2564   si_shader_selector_key(ctx, state->cso, &state->key);
2565   return si_shader_select_with_key(sctx, state, &state->key, -1, false);
2566}
2567
2568static void si_parse_next_shader_property(const struct si_shader_info *info, bool streamout,
2569                                          struct si_shader_key *key)
2570{
2571   gl_shader_stage next_shader = info->base.next_stage;
2572
2573   switch (info->stage) {
2574   case MESA_SHADER_VERTEX:
2575      switch (next_shader) {
2576      case MESA_SHADER_GEOMETRY:
2577         key->as_es = 1;
2578         break;
2579      case MESA_SHADER_TESS_CTRL:
2580      case MESA_SHADER_TESS_EVAL:
2581         key->as_ls = 1;
2582         break;
2583      default:
2584         /* If POSITION isn't written, it can only be a HW VS
2585          * if streamout is used. If streamout isn't used,
2586          * assume that it's a HW LS. (the next shader is TCS)
2587          * This heuristic is needed for separate shader objects.
2588          */
2589         if (!info->writes_position && !streamout)
2590            key->as_ls = 1;
2591      }
2592      break;
2593
2594   case MESA_SHADER_TESS_EVAL:
2595      if (next_shader == MESA_SHADER_GEOMETRY || !info->writes_position)
2596         key->as_es = 1;
2597      break;
2598
2599   default:;
2600   }
2601}
2602
2603/**
2604 * Compile the main shader part or the monolithic shader as part of
2605 * si_shader_selector initialization. Since it can be done asynchronously,
2606 * there is no way to report compile failures to applications.
2607 */
2608static void si_init_shader_selector_async(void *job, void *gdata, int thread_index)
2609{
2610   struct si_shader_selector *sel = (struct si_shader_selector *)job;
2611   struct si_screen *sscreen = sel->screen;
2612   struct ac_llvm_compiler *compiler;
2613   struct pipe_debug_callback *debug = &sel->compiler_ctx_state.debug;
2614
2615   assert(!debug->debug_message || debug->async);
2616   assert(thread_index >= 0);
2617   assert(thread_index < ARRAY_SIZE(sscreen->compiler));
2618   compiler = &sscreen->compiler[thread_index];
2619
2620   if (!compiler->passes)
2621      si_init_compiler(sscreen, compiler);
2622
2623   /* The GS copy shader is always pre-compiled. */
2624   if (sel->info.stage == MESA_SHADER_GEOMETRY &&
2625       (!sscreen->use_ngg || !sscreen->use_ngg_streamout || /* also for PRIMITIVES_GENERATED */
2626        sel->tess_turns_off_ngg)) {
2627      sel->gs_copy_shader = si_generate_gs_copy_shader(sscreen, compiler, sel, debug);
2628      if (!sel->gs_copy_shader) {
2629         fprintf(stderr, "radeonsi: can't create GS copy shader\n");
2630         return;
2631      }
2632
2633      si_shader_vs(sscreen, sel->gs_copy_shader, sel);
2634   }
2635
2636   /* Serialize NIR to save memory. Monolithic shader variants
2637    * have to deserialize NIR before compilation.
2638    */
2639   if (sel->nir) {
2640      struct blob blob;
2641      size_t size;
2642
2643      blob_init(&blob);
2644      /* true = remove optional debugging data to increase
2645       * the likehood of getting more shader cache hits.
2646       * It also drops variable names, so we'll save more memory.
2647       */
2648      nir_serialize(&blob, sel->nir, true);
2649      blob_finish_get_buffer(&blob, &sel->nir_binary, &size);
2650      sel->nir_size = size;
2651   }
2652
2653   /* Compile the main shader part for use with a prolog and/or epilog.
2654    * If this fails, the driver will try to compile a monolithic shader
2655    * on demand.
2656    */
2657   if (!sscreen->use_monolithic_shaders) {
2658      struct si_shader *shader = CALLOC_STRUCT(si_shader);
2659      unsigned char ir_sha1_cache_key[20];
2660
2661      if (!shader) {
2662         fprintf(stderr, "radeonsi: can't allocate a main shader part\n");
2663         return;
2664      }
2665
2666      /* We can leave the fence signaled because use of the default
2667       * main part is guarded by the selector's ready fence. */
2668      util_queue_fence_init(&shader->ready);
2669
2670      shader->selector = sel;
2671      shader->is_monolithic = false;
2672      si_parse_next_shader_property(&sel->info, sel->so.num_outputs != 0, &shader->key);
2673
2674      if (sscreen->use_ngg && (!sel->so.num_outputs || sscreen->use_ngg_streamout) &&
2675          ((sel->info.stage == MESA_SHADER_VERTEX && !shader->key.as_ls) ||
2676           sel->info.stage == MESA_SHADER_TESS_EVAL || sel->info.stage == MESA_SHADER_GEOMETRY))
2677         shader->key.as_ngg = 1;
2678
2679      if (sel->nir) {
2680         si_get_ir_cache_key(sel, shader->key.as_ngg, shader->key.as_es, ir_sha1_cache_key);
2681      }
2682
2683      /* Try to load the shader from the shader cache. */
2684      simple_mtx_lock(&sscreen->shader_cache_mutex);
2685
2686      if (si_shader_cache_load_shader(sscreen, ir_sha1_cache_key, shader)) {
2687         simple_mtx_unlock(&sscreen->shader_cache_mutex);
2688         si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
2689      } else {
2690         simple_mtx_unlock(&sscreen->shader_cache_mutex);
2691
2692         /* Compile the shader if it hasn't been loaded from the cache. */
2693         if (!si_compile_shader(sscreen, compiler, shader, debug)) {
2694            FREE(shader);
2695            fprintf(stderr, "radeonsi: can't compile a main shader part\n");
2696            return;
2697         }
2698
2699         simple_mtx_lock(&sscreen->shader_cache_mutex);
2700         si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key, shader, true);
2701         simple_mtx_unlock(&sscreen->shader_cache_mutex);
2702      }
2703
2704      *si_get_main_shader_part(sel, &shader->key) = shader;
2705
2706      /* Unset "outputs_written" flags for outputs converted to
2707       * DEFAULT_VAL, so that later inter-shader optimizations don't
2708       * try to eliminate outputs that don't exist in the final
2709       * shader.
2710       *
2711       * This is only done if non-monolithic shaders are enabled.
2712       */
2713      if ((sel->info.stage == MESA_SHADER_VERTEX ||
2714           sel->info.stage == MESA_SHADER_TESS_EVAL ||
2715           sel->info.stage == MESA_SHADER_GEOMETRY) &&
2716          !shader->key.as_ls && !shader->key.as_es) {
2717         unsigned i;
2718
2719         for (i = 0; i < sel->info.num_outputs; i++) {
2720            unsigned semantic = sel->info.output_semantic[i];
2721            unsigned ps_input_cntl = shader->info.vs_output_ps_input_cntl[semantic];
2722
2723            /* OFFSET=0x20 means DEFAULT_VAL, which means VS doesn't export it. */
2724            if (G_028644_OFFSET(ps_input_cntl) != 0x20)
2725               continue;
2726
2727            unsigned id;
2728
2729            /* Remove the output from the mask. */
2730            if ((semantic <= VARYING_SLOT_VAR31 || semantic >= VARYING_SLOT_VAR0_16BIT) &&
2731                semantic != VARYING_SLOT_POS &&
2732                semantic != VARYING_SLOT_PSIZ &&
2733                semantic != VARYING_SLOT_CLIP_VERTEX &&
2734                semantic != VARYING_SLOT_EDGE) {
2735               id = si_shader_io_get_unique_index(semantic, true);
2736               sel->outputs_written_before_ps &= ~(1ull << id);
2737            }
2738         }
2739      }
2740   }
2741
2742   /* Free NIR. We only keep serialized NIR after this point. */
2743   if (sel->nir) {
2744      ralloc_free(sel->nir);
2745      sel->nir = NULL;
2746   }
2747}
2748
2749void si_schedule_initial_compile(struct si_context *sctx, gl_shader_stage stage,
2750                                 struct util_queue_fence *ready_fence,
2751                                 struct si_compiler_ctx_state *compiler_ctx_state, void *job,
2752                                 util_queue_execute_func execute)
2753{
2754   util_queue_fence_init(ready_fence);
2755
2756   struct util_async_debug_callback async_debug;
2757   bool debug = (sctx->debug.debug_message && !sctx->debug.async) || sctx->is_debug ||
2758                si_can_dump_shader(sctx->screen, stage);
2759
2760   if (debug) {
2761      u_async_debug_init(&async_debug);
2762      compiler_ctx_state->debug = async_debug.base;
2763   }
2764
2765   util_queue_add_job(&sctx->screen->shader_compiler_queue, job, ready_fence, execute, NULL, 0);
2766
2767   if (debug) {
2768      util_queue_fence_wait(ready_fence);
2769      u_async_debug_drain(&async_debug, &sctx->debug);
2770      u_async_debug_cleanup(&async_debug);
2771   }
2772
2773   if (sctx->screen->options.sync_compile)
2774      util_queue_fence_wait(ready_fence);
2775}
2776
2777/* Return descriptor slot usage masks from the given shader info. */
2778void si_get_active_slot_masks(const struct si_shader_info *info, uint64_t *const_and_shader_buffers,
2779                              uint64_t *samplers_and_images)
2780{
2781   unsigned start, num_shaderbufs, num_constbufs, num_images, num_msaa_images, num_samplers;
2782
2783   num_shaderbufs = info->base.num_ssbos;
2784   num_constbufs = info->base.num_ubos;
2785   /* two 8-byte images share one 16-byte slot */
2786   num_images = align(info->base.num_images, 2);
2787   num_msaa_images = align(util_last_bit(info->base.msaa_images), 2);
2788   num_samplers = BITSET_LAST_BIT(info->base.textures_used);
2789
2790   /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2791   start = si_get_shaderbuf_slot(num_shaderbufs - 1);
2792   *const_and_shader_buffers = u_bit_consecutive64(start, num_shaderbufs + num_constbufs);
2793
2794   /* The layout is:
2795    *   - fmask[last] ... fmask[0]     go to [15-last .. 15]
2796    *   - image[last] ... image[0]     go to [31-last .. 31]
2797    *   - sampler[0] ... sampler[last] go to [32 .. 32+last*2]
2798    *
2799    * FMASKs for images are placed separately, because MSAA images are rare,
2800    * and so we can benefit from a better cache hit rate if we keep image
2801    * descriptors together.
2802    */
2803   if (num_msaa_images)
2804      num_images = SI_NUM_IMAGES + num_msaa_images; /* add FMASK descriptors */
2805
2806   start = si_get_image_slot(num_images - 1) / 2;
2807   *samplers_and_images = u_bit_consecutive64(start, num_images / 2 + num_samplers);
2808}
2809
2810static void *si_create_shader_selector(struct pipe_context *ctx,
2811                                       const struct pipe_shader_state *state)
2812{
2813   struct si_screen *sscreen = (struct si_screen *)ctx->screen;
2814   struct si_context *sctx = (struct si_context *)ctx;
2815   struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
2816   int i;
2817
2818   if (!sel)
2819      return NULL;
2820
2821   sel->screen = sscreen;
2822   sel->compiler_ctx_state.debug = sctx->debug;
2823   sel->compiler_ctx_state.is_debug_context = sctx->is_debug;
2824
2825   sel->so = state->stream_output;
2826
2827   if (state->type == PIPE_SHADER_IR_TGSI) {
2828      sel->nir = tgsi_to_nir(state->tokens, ctx->screen, true);
2829   } else {
2830      assert(state->type == PIPE_SHADER_IR_NIR);
2831      sel->nir = state->ir.nir;
2832   }
2833
2834   si_nir_scan_shader(sel->nir, &sel->info);
2835
2836   const enum pipe_shader_type type = pipe_shader_type_from_mesa(sel->info.stage);
2837   sel->pipe_shader_type = type;
2838   sel->const_and_shader_buf_descriptors_index =
2839      si_const_and_shader_buffer_descriptors_idx(type);
2840   sel->sampler_and_images_descriptors_index =
2841      si_sampler_and_image_descriptors_idx(type);
2842
2843   p_atomic_inc(&sscreen->num_shaders_created);
2844   si_get_active_slot_masks(&sel->info, &sel->active_const_and_shader_buffers,
2845                            &sel->active_samplers_and_images);
2846
2847   /* Record which streamout buffers are enabled. */
2848   for (i = 0; i < sel->so.num_outputs; i++) {
2849      sel->enabled_streamout_buffer_mask |= (1 << sel->so.output[i].output_buffer)
2850                                            << (sel->so.output[i].stream * 4);
2851   }
2852
2853   sel->num_vs_inputs =
2854      sel->info.stage == MESA_SHADER_VERTEX && !sel->info.base.vs.blit_sgprs_amd
2855         ? sel->info.num_inputs
2856         : 0;
2857   unsigned num_vbos_in_sgprs = si_num_vbos_in_user_sgprs_inline(sscreen->info.chip_class);
2858   sel->num_vbos_in_user_sgprs = MIN2(sel->num_vs_inputs, num_vbos_in_sgprs);
2859
2860   /* The prolog is a no-op if there are no inputs. */
2861   sel->vs_needs_prolog = sel->info.stage == MESA_SHADER_VERTEX && sel->info.num_inputs &&
2862                          !sel->info.base.vs.blit_sgprs_amd;
2863
2864   if (sel->info.stage == MESA_SHADER_VERTEX ||
2865       sel->info.stage == MESA_SHADER_TESS_CTRL ||
2866       sel->info.stage == MESA_SHADER_TESS_EVAL ||
2867       sel->info.stage == MESA_SHADER_GEOMETRY) {
2868      if (sel->info.stage == MESA_SHADER_TESS_CTRL) {
2869         /* Always reserve space for these. */
2870         sel->patch_outputs_written |=
2871            (1ull << si_shader_io_get_unique_index_patch(VARYING_SLOT_TESS_LEVEL_INNER)) |
2872            (1ull << si_shader_io_get_unique_index_patch(VARYING_SLOT_TESS_LEVEL_OUTER));
2873      }
2874      for (i = 0; i < sel->info.num_outputs; i++) {
2875         unsigned semantic = sel->info.output_semantic[i];
2876
2877         if (semantic == VARYING_SLOT_TESS_LEVEL_INNER ||
2878             semantic == VARYING_SLOT_TESS_LEVEL_OUTER ||
2879             (semantic >= VARYING_SLOT_PATCH0 && semantic < VARYING_SLOT_TESS_MAX)) {
2880            sel->patch_outputs_written |= 1ull << si_shader_io_get_unique_index_patch(semantic);
2881         } else if ((semantic <= VARYING_SLOT_VAR31 || semantic >= VARYING_SLOT_VAR0_16BIT) &&
2882                    semantic != VARYING_SLOT_EDGE) {
2883            sel->outputs_written |= 1ull << si_shader_io_get_unique_index(semantic, false);
2884
2885            /* Ignore outputs that are not passed from VS to PS. */
2886            if (semantic != VARYING_SLOT_POS &&
2887                semantic != VARYING_SLOT_PSIZ &&
2888                semantic != VARYING_SLOT_CLIP_VERTEX) {
2889               sel->outputs_written_before_ps |= 1ull
2890                                                 << si_shader_io_get_unique_index(semantic, true);
2891            }
2892         }
2893      }
2894   }
2895
2896   switch (sel->info.stage) {
2897   case MESA_SHADER_GEOMETRY:
2898      /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
2899      sel->rast_prim = sel->info.base.gs.output_primitive;
2900      if (util_rast_prim_is_triangles(sel->rast_prim))
2901         sel->rast_prim = PIPE_PRIM_TRIANGLES;
2902
2903      sel->gsvs_vertex_size = sel->info.num_outputs * 16;
2904      sel->max_gsvs_emit_size = sel->gsvs_vertex_size * sel->info.base.gs.vertices_out;
2905      sel->gs_input_verts_per_prim =
2906         u_vertices_per_prim(sel->info.base.gs.input_primitive);
2907
2908      /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation so
2909       * we can't split workgroups. Disable ngg if any of the following conditions is true:
2910       * - num_invocations * gs.vertices_out > 256
2911       * - LDS usage is too high
2912       */
2913      sel->tess_turns_off_ngg = sscreen->info.chip_class >= GFX10 &&
2914                                (sel->info.base.gs.invocations * sel->info.base.gs.vertices_out > 256 ||
2915                                 sel->info.base.gs.invocations * sel->info.base.gs.vertices_out *
2916                                 (sel->info.num_outputs * 4 + 1) > 6500 /* max dw per GS primitive */);
2917      break;
2918
2919   case MESA_SHADER_VERTEX:
2920   case MESA_SHADER_TESS_CTRL:
2921   case MESA_SHADER_TESS_EVAL:
2922      sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
2923      sel->lshs_vertex_stride = sel->esgs_itemsize;
2924
2925      /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2926       * will start on a different bank. (except for the maximum 32*16).
2927       */
2928      if (sel->lshs_vertex_stride < 32 * 16)
2929         sel->lshs_vertex_stride += 4;
2930
2931      /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2932       * conflicts, i.e. each vertex will start at a different bank.
2933       */
2934      if (sctx->chip_class >= GFX9)
2935         sel->esgs_itemsize += 4;
2936
2937      assert(((sel->esgs_itemsize / 4) & C_028AAC_ITEMSIZE) == 0);
2938
2939      sel->tcs_vgpr_only_inputs = ~sel->info.base.tess.tcs_cross_invocation_inputs_read &
2940                                  ~sel->info.base.inputs_read_indirectly &
2941                                  sel->info.base.inputs_read;
2942
2943      /* Only for TES: */
2944      if (sel->info.stage == MESA_SHADER_TESS_EVAL) {
2945         if (sel->info.base.tess.point_mode)
2946            sel->rast_prim = PIPE_PRIM_POINTS;
2947         else if (sel->info.base.tess.primitive_mode == GL_LINES)
2948            sel->rast_prim = PIPE_PRIM_LINE_STRIP;
2949         else
2950            sel->rast_prim = PIPE_PRIM_TRIANGLES;
2951      } else {
2952         sel->rast_prim = PIPE_PRIM_TRIANGLES;
2953      }
2954      break;
2955
2956   case MESA_SHADER_FRAGMENT:
2957      for (i = 0; i < sel->info.num_inputs; i++) {
2958         unsigned semantic = sel->info.input[i].semantic;
2959
2960         if ((semantic <= VARYING_SLOT_VAR31 || semantic >= VARYING_SLOT_VAR0_16BIT) &&
2961             semantic != VARYING_SLOT_PNTC) {
2962            sel->inputs_read |= 1ull << si_shader_io_get_unique_index(semantic, true);
2963         }
2964      }
2965
2966      for (i = 0; i < 8; i++)
2967         if (sel->info.colors_written & (1 << i))
2968            sel->colors_written_4bit |= 0xf << (4 * i);
2969
2970      for (i = 0; i < sel->info.num_inputs; i++) {
2971         if (sel->info.input[i].semantic == VARYING_SLOT_COL0)
2972            sel->color_attr_index[0] = i;
2973         else if (sel->info.input[i].semantic == VARYING_SLOT_COL1)
2974            sel->color_attr_index[1] = i;
2975      }
2976      break;
2977   default:;
2978   }
2979
2980   bool ngg_culling_allowed =
2981      sscreen->info.chip_class >= GFX10 &&
2982      sscreen->use_ngg_culling &&
2983      (sel->info.stage == MESA_SHADER_VERTEX ||
2984       sel->info.stage == MESA_SHADER_TESS_EVAL) &&
2985      sel->info.writes_position &&
2986      !sel->info.writes_viewport_index && /* cull only against viewport 0 */
2987      !sel->info.base.writes_memory && !sel->so.num_outputs &&
2988      (sel->info.stage != MESA_SHADER_VERTEX ||
2989       (!sel->info.base.vs.blit_sgprs_amd &&
2990        !sel->info.base.vs.window_space_position));
2991
2992   sel->ngg_cull_vert_threshold = UINT_MAX; /* disabled (changed below) */
2993
2994   if (ngg_culling_allowed) {
2995      if (sel->info.stage == MESA_SHADER_VERTEX) {
2996         if (sscreen->debug_flags & DBG(ALWAYS_NGG_CULLING_ALL))
2997            sel->ngg_cull_vert_threshold = 0; /* always enabled */
2998         else if (sscreen->options.shader_culling ||
2999                  sscreen->info.chip_class == GFX10_3 ||
3000                  (sscreen->info.chip_class == GFX10 &&
3001                   sscreen->info.is_pro_graphics)) {
3002            sel->ngg_cull_vert_threshold = 128;
3003         }
3004      } else if (sel->info.stage == MESA_SHADER_TESS_EVAL) {
3005         if (sel->rast_prim != PIPE_PRIM_POINTS &&
3006             (sscreen->debug_flags & DBG(ALWAYS_NGG_CULLING_ALL) ||
3007              sscreen->debug_flags & DBG(ALWAYS_NGG_CULLING_TESS) ||
3008              sscreen->info.chip_class == GFX10_3))
3009            sel->ngg_cull_vert_threshold = 0; /* always enabled */
3010      }
3011   }
3012
3013   sel->clipdist_mask = sel->info.writes_clipvertex ? SIX_BITS :
3014                           u_bit_consecutive(0, sel->info.base.clip_distance_array_size);
3015   sel->culldist_mask = u_bit_consecutive(0, sel->info.base.cull_distance_array_size) <<
3016                        sel->info.base.clip_distance_array_size;
3017
3018   /* DB_SHADER_CONTROL */
3019   sel->db_shader_control = S_02880C_Z_EXPORT_ENABLE(sel->info.writes_z) |
3020                            S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel->info.writes_stencil) |
3021                            S_02880C_MASK_EXPORT_ENABLE(sel->info.writes_samplemask) |
3022                            S_02880C_KILL_ENABLE(sel->info.base.fs.uses_discard);
3023
3024   if (sel->info.stage == MESA_SHADER_FRAGMENT) {
3025      switch (sel->info.base.fs.depth_layout) {
3026      case FRAG_DEPTH_LAYOUT_GREATER:
3027         sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
3028         break;
3029      case FRAG_DEPTH_LAYOUT_LESS:
3030         sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
3031         break;
3032      default:;
3033      }
3034
3035      /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
3036       *
3037       *   | early Z/S | writes_mem | allow_ReZ? |      Z_ORDER       | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
3038       * --|-----------|------------|------------|--------------------|-------------------|-------------
3039       * 1a|   false   |   false    |   true     | EarlyZ_Then_ReZ    |         0         |     0
3040       * 1b|   false   |   false    |   false    | EarlyZ_Then_LateZ  |         0         |     0
3041       * 2 |   false   |   true     |   n/a      |       LateZ        |         1         |     0
3042       * 3 |   true    |   false    |   n/a      | EarlyZ_Then_LateZ  |         0         |     0
3043       * 4 |   true    |   true     |   n/a      | EarlyZ_Then_LateZ  |         0         |     1
3044       *
3045       * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
3046       * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
3047       *
3048       * Don't use ReZ without profiling !!!
3049       *
3050       * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
3051       * shaders.
3052       */
3053      if (sel->info.base.fs.early_fragment_tests) {
3054         /* Cases 3, 4. */
3055         sel->db_shader_control |= S_02880C_DEPTH_BEFORE_SHADER(1) |
3056                                   S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z) |
3057                                   S_02880C_EXEC_ON_NOOP(sel->info.base.writes_memory);
3058      } else if (sel->info.base.writes_memory) {
3059         /* Case 2. */
3060         sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z) | S_02880C_EXEC_ON_HIER_FAIL(1);
3061      } else {
3062         /* Case 1. */
3063         sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
3064      }
3065
3066      if (sel->info.base.fs.post_depth_coverage)
3067         sel->db_shader_control |= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
3068   }
3069
3070   (void)simple_mtx_init(&sel->mutex, mtx_plain);
3071
3072   si_schedule_initial_compile(sctx, sel->info.stage, &sel->ready, &sel->compiler_ctx_state,
3073                               sel, si_init_shader_selector_async);
3074   return sel;
3075}
3076
3077static void *si_create_shader(struct pipe_context *ctx, const struct pipe_shader_state *state)
3078{
3079   struct si_context *sctx = (struct si_context *)ctx;
3080   struct si_screen *sscreen = (struct si_screen *)ctx->screen;
3081   bool cache_hit;
3082   struct si_shader_selector *sel = (struct si_shader_selector *)util_live_shader_cache_get(
3083      ctx, &sscreen->live_shader_cache, state, &cache_hit);
3084
3085   if (sel && cache_hit && sctx->debug.debug_message) {
3086      if (sel->main_shader_part)
3087         si_shader_dump_stats_for_shader_db(sscreen, sel->main_shader_part, &sctx->debug);
3088      if (sel->main_shader_part_ls)
3089         si_shader_dump_stats_for_shader_db(sscreen, sel->main_shader_part_ls, &sctx->debug);
3090      if (sel->main_shader_part_es)
3091         si_shader_dump_stats_for_shader_db(sscreen, sel->main_shader_part_es, &sctx->debug);
3092      if (sel->main_shader_part_ngg)
3093         si_shader_dump_stats_for_shader_db(sscreen, sel->main_shader_part_ngg, &sctx->debug);
3094      if (sel->main_shader_part_ngg_es)
3095         si_shader_dump_stats_for_shader_db(sscreen, sel->main_shader_part_ngg_es, &sctx->debug);
3096   }
3097   return sel;
3098}
3099
3100static void si_update_streamout_state(struct si_context *sctx)
3101{
3102   struct si_shader_selector *shader_with_so = si_get_vs(sctx)->cso;
3103
3104   if (!shader_with_so)
3105      return;
3106
3107   sctx->streamout.enabled_stream_buffers_mask = shader_with_so->enabled_streamout_buffer_mask;
3108   sctx->streamout.stride_in_dw = shader_with_so->so.stride;
3109}
3110
3111static void si_update_clip_regs(struct si_context *sctx, struct si_shader_selector *old_hw_vs,
3112                                struct si_shader *old_hw_vs_variant,
3113                                struct si_shader_selector *next_hw_vs,
3114                                struct si_shader *next_hw_vs_variant)
3115{
3116   if (next_hw_vs &&
3117       (!old_hw_vs ||
3118        (old_hw_vs->info.stage == MESA_SHADER_VERTEX && old_hw_vs->info.base.vs.window_space_position) !=
3119        (next_hw_vs->info.stage == MESA_SHADER_VERTEX && next_hw_vs->info.base.vs.window_space_position) ||
3120        old_hw_vs->clipdist_mask != next_hw_vs->clipdist_mask ||
3121        old_hw_vs->culldist_mask != next_hw_vs->culldist_mask || !old_hw_vs_variant ||
3122        !next_hw_vs_variant ||
3123        old_hw_vs_variant->pa_cl_vs_out_cntl != next_hw_vs_variant->pa_cl_vs_out_cntl))
3124      si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
3125}
3126
3127static void si_update_rasterized_prim(struct si_context *sctx)
3128{
3129   enum pipe_prim_type rast_prim;
3130
3131   if (sctx->shader.gs.cso) {
3132      /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
3133      rast_prim = sctx->shader.gs.cso->rast_prim;
3134   } else if (sctx->shader.tes.cso) {
3135      /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
3136      rast_prim = sctx->shader.tes.cso->rast_prim;
3137   } else {
3138      /* Determined by draw calls. */
3139      return;
3140   }
3141
3142   if (rast_prim != sctx->current_rast_prim) {
3143      if (util_prim_is_points_or_lines(sctx->current_rast_prim) !=
3144          util_prim_is_points_or_lines(rast_prim))
3145         si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
3146
3147      sctx->current_rast_prim = rast_prim;
3148   }
3149}
3150
3151static void si_update_common_shader_state(struct si_context *sctx, struct si_shader_selector *sel,
3152                                          enum pipe_shader_type type)
3153{
3154   si_set_active_descriptors_for_shader(sctx, sel);
3155
3156   sctx->uses_bindless_samplers = si_shader_uses_bindless_samplers(sctx->shader.vs.cso) ||
3157                                  si_shader_uses_bindless_samplers(sctx->shader.gs.cso) ||
3158                                  si_shader_uses_bindless_samplers(sctx->shader.ps.cso) ||
3159                                  si_shader_uses_bindless_samplers(sctx->shader.tcs.cso) ||
3160                                  si_shader_uses_bindless_samplers(sctx->shader.tes.cso);
3161   sctx->uses_bindless_images = si_shader_uses_bindless_images(sctx->shader.vs.cso) ||
3162                                si_shader_uses_bindless_images(sctx->shader.gs.cso) ||
3163                                si_shader_uses_bindless_images(sctx->shader.ps.cso) ||
3164                                si_shader_uses_bindless_images(sctx->shader.tcs.cso) ||
3165                                si_shader_uses_bindless_images(sctx->shader.tes.cso);
3166
3167   if (type == PIPE_SHADER_VERTEX || type == PIPE_SHADER_TESS_EVAL || type == PIPE_SHADER_GEOMETRY)
3168      sctx->ngg_culling = 0; /* this will be enabled on the first draw if needed */
3169
3170   si_invalidate_inlinable_uniforms(sctx, type);
3171   sctx->do_update_shaders = true;
3172}
3173
3174static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
3175{
3176   struct si_context *sctx = (struct si_context *)ctx;
3177   struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3178   struct si_shader *old_hw_vs_variant = si_get_vs(sctx)->current;
3179   struct si_shader_selector *sel = state;
3180
3181   if (sctx->shader.vs.cso == sel)
3182      return;
3183
3184   sctx->shader.vs.cso = sel;
3185   sctx->shader.vs.current = sel ? sel->first_variant : NULL;
3186   sctx->num_vs_blit_sgprs = sel ? sel->info.base.vs.blit_sgprs_amd : 0;
3187   sctx->vs_uses_draw_id = sel ? sel->info.uses_drawid : false;
3188   sctx->fixed_func_tcs_shader.key.mono.u.ff_tcs_inputs_to_copy = sel ? sel->outputs_written : 0;
3189
3190   if (si_update_ngg(sctx))
3191      si_shader_change_notify(sctx);
3192
3193   si_update_common_shader_state(sctx, sel, PIPE_SHADER_VERTEX);
3194   si_select_draw_vbo(sctx);
3195   si_update_vs_viewport_state(sctx);
3196   si_update_streamout_state(sctx);
3197   si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant, si_get_vs(sctx)->cso,
3198                       si_get_vs(sctx)->current);
3199   si_update_rasterized_prim(sctx);
3200   si_vs_key_update_inputs(sctx);
3201}
3202
3203static void si_update_tess_uses_prim_id(struct si_context *sctx)
3204{
3205   sctx->ia_multi_vgt_param_key.u.tess_uses_prim_id =
3206      (sctx->shader.tes.cso && sctx->shader.tes.cso->info.uses_primid) ||
3207      (sctx->shader.tcs.cso && sctx->shader.tcs.cso->info.uses_primid) ||
3208      (sctx->shader.gs.cso && sctx->shader.gs.cso->info.uses_primid) ||
3209      (sctx->shader.ps.cso && !sctx->shader.gs.cso && sctx->shader.ps.cso->info.uses_primid);
3210}
3211
3212bool si_update_ngg(struct si_context *sctx)
3213{
3214   if (!sctx->screen->use_ngg) {
3215      assert(!sctx->ngg);
3216      return false;
3217   }
3218
3219   bool new_ngg = true;
3220
3221   if (sctx->shader.gs.cso && sctx->shader.tes.cso && sctx->shader.gs.cso->tess_turns_off_ngg) {
3222      new_ngg = false;
3223   } else if (!sctx->screen->use_ngg_streamout) {
3224      struct si_shader_selector *last = si_get_vs(sctx)->cso;
3225
3226      if ((last && last->so.num_outputs) || sctx->streamout.prims_gen_query_enabled)
3227         new_ngg = false;
3228   }
3229
3230   if (new_ngg != sctx->ngg) {
3231      /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
3232       * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
3233       * pointers are set.
3234       */
3235      if (sctx->screen->info.has_vgt_flush_ngg_legacy_bug && !new_ngg) {
3236         sctx->flags |= SI_CONTEXT_VGT_FLUSH;
3237         if (sctx->chip_class == GFX10) {
3238            /* Workaround for https://gitlab.freedesktop.org/mesa/mesa/-/issues/2941 */
3239            si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
3240         }
3241      }
3242
3243      sctx->ngg = new_ngg;
3244      sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3245      si_select_draw_vbo(sctx);
3246      return true;
3247   }
3248   return false;
3249}
3250
3251static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
3252{
3253   struct si_context *sctx = (struct si_context *)ctx;
3254   struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3255   struct si_shader *old_hw_vs_variant = si_get_vs(sctx)->current;
3256   struct si_shader_selector *sel = state;
3257   bool enable_changed = !!sctx->shader.gs.cso != !!sel;
3258   bool ngg_changed;
3259
3260   if (sctx->shader.gs.cso == sel)
3261      return;
3262
3263   sctx->shader.gs.cso = sel;
3264   sctx->shader.gs.current = sel ? sel->first_variant : NULL;
3265   sctx->ia_multi_vgt_param_key.u.uses_gs = sel != NULL;
3266
3267   si_update_common_shader_state(sctx, sel, PIPE_SHADER_GEOMETRY);
3268   si_select_draw_vbo(sctx);
3269   sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3270
3271   ngg_changed = si_update_ngg(sctx);
3272   if (ngg_changed || enable_changed)
3273      si_shader_change_notify(sctx);
3274   if (enable_changed) {
3275      if (sctx->ia_multi_vgt_param_key.u.uses_tess)
3276         si_update_tess_uses_prim_id(sctx);
3277   }
3278   si_update_vs_viewport_state(sctx);
3279   si_update_streamout_state(sctx);
3280   si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant, si_get_vs(sctx)->cso,
3281                       si_get_vs(sctx)->current);
3282   si_update_rasterized_prim(sctx);
3283}
3284
3285static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
3286{
3287   struct si_context *sctx = (struct si_context *)ctx;
3288   struct si_shader_selector *sel = state;
3289   bool enable_changed = !!sctx->shader.tcs.cso != !!sel;
3290
3291   if (sctx->shader.tcs.cso == sel)
3292      return;
3293
3294   sctx->shader.tcs.cso = sel;
3295   sctx->shader.tcs.current = sel ? sel->first_variant : NULL;
3296   sctx->shader.tcs.key.part.tcs.epilog.invoc0_tess_factors_are_def =
3297      sel ? sel->info.tessfactors_are_def_in_all_invocs : 0;
3298   si_update_tess_uses_prim_id(sctx);
3299
3300   si_update_common_shader_state(sctx, sel, PIPE_SHADER_TESS_CTRL);
3301
3302   if (enable_changed)
3303      sctx->last_tcs = NULL; /* invalidate derived tess state */
3304}
3305
3306static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
3307{
3308   struct si_context *sctx = (struct si_context *)ctx;
3309   struct si_shader_selector *old_hw_vs = si_get_vs(sctx)->cso;
3310   struct si_shader *old_hw_vs_variant = si_get_vs(sctx)->current;
3311   struct si_shader_selector *sel = state;
3312   bool enable_changed = !!sctx->shader.tes.cso != !!sel;
3313
3314   if (sctx->shader.tes.cso == sel)
3315      return;
3316
3317   sctx->shader.tes.cso = sel;
3318   sctx->shader.tes.current = sel ? sel->first_variant : NULL;
3319   sctx->ia_multi_vgt_param_key.u.uses_tess = sel != NULL;
3320   si_update_tess_uses_prim_id(sctx);
3321
3322   sctx->shader.tcs.key.part.tcs.epilog.prim_mode =
3323   sctx->fixed_func_tcs_shader.key.part.tcs.epilog.prim_mode =
3324      sel ? sel->info.base.tess.primitive_mode : 0;
3325
3326   sctx->shader.tcs.key.part.tcs.epilog.tes_reads_tess_factors =
3327   sctx->fixed_func_tcs_shader.key.part.tcs.epilog.tes_reads_tess_factors =
3328      sel ? sel->info.reads_tess_factors : 0;
3329
3330   si_update_common_shader_state(sctx, sel, PIPE_SHADER_TESS_EVAL);
3331   si_select_draw_vbo(sctx);
3332   sctx->last_gs_out_prim = -1; /* reset this so that it gets updated */
3333
3334   bool ngg_changed = si_update_ngg(sctx);
3335   if (ngg_changed || enable_changed)
3336      si_shader_change_notify(sctx);
3337   if (enable_changed)
3338      sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
3339   si_update_vs_viewport_state(sctx);
3340   si_update_streamout_state(sctx);
3341   si_update_clip_regs(sctx, old_hw_vs, old_hw_vs_variant, si_get_vs(sctx)->cso,
3342                       si_get_vs(sctx)->current);
3343   si_update_rasterized_prim(sctx);
3344}
3345
3346void si_update_ps_kill_enable(struct si_context *sctx)
3347{
3348   if (!sctx->shader.ps.cso)
3349      return;
3350
3351   unsigned db_shader_control = sctx->shader.ps.cso->db_shader_control |
3352                                S_02880C_KILL_ENABLE(sctx->queued.named.dsa->alpha_func != PIPE_FUNC_ALWAYS);
3353
3354   if (sctx->ps_db_shader_control != db_shader_control) {
3355      sctx->ps_db_shader_control = db_shader_control;
3356      si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
3357      if (sctx->screen->dpbb_allowed)
3358         si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3359   }
3360}
3361
3362void si_update_vrs_flat_shading(struct si_context *sctx)
3363{
3364   if (sctx->chip_class >= GFX10_3 && sctx->shader.ps.cso) {
3365      struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3366      struct si_shader_info *info = &sctx->shader.ps.cso->info;
3367      bool allow_flat_shading = info->allow_flat_shading;
3368
3369      if (allow_flat_shading &&
3370          (rs->line_smooth || rs->poly_smooth || rs->poly_stipple_enable ||
3371           (!rs->flatshade && info->uses_interp_color)))
3372         allow_flat_shading = false;
3373
3374      if (sctx->allow_flat_shading != allow_flat_shading) {
3375         sctx->allow_flat_shading = allow_flat_shading;
3376         si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
3377      }
3378   }
3379}
3380
3381static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
3382{
3383   struct si_context *sctx = (struct si_context *)ctx;
3384   struct si_shader_selector *old_sel = sctx->shader.ps.cso;
3385   struct si_shader_selector *sel = state;
3386
3387   /* skip if supplied shader is one already in use */
3388   if (old_sel == sel)
3389      return;
3390
3391   sctx->shader.ps.cso = sel;
3392   sctx->shader.ps.current = sel ? sel->first_variant : NULL;
3393
3394   si_update_common_shader_state(sctx, sel, PIPE_SHADER_FRAGMENT);
3395   if (sel) {
3396      if (sctx->ia_multi_vgt_param_key.u.uses_tess)
3397         si_update_tess_uses_prim_id(sctx);
3398
3399      if (!old_sel || old_sel->info.colors_written != sel->info.colors_written)
3400         si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
3401
3402      if (sctx->screen->has_out_of_order_rast &&
3403          (!old_sel || old_sel->info.base.writes_memory != sel->info.base.writes_memory ||
3404           old_sel->info.base.fs.early_fragment_tests !=
3405              sel->info.base.fs.early_fragment_tests))
3406         si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3407   }
3408   si_update_ps_colorbuf0_slot(sctx);
3409
3410   si_ps_key_update_framebuffer(sctx);
3411   si_ps_key_update_framebuffer_blend(sctx);
3412   si_ps_key_update_blend_rasterizer(sctx);
3413   si_ps_key_update_rasterizer(sctx);
3414   si_ps_key_update_dsa(sctx);
3415   si_ps_key_update_sample_shading(sctx);
3416   si_ps_key_update_framebuffer_rasterizer_sample_shading(sctx);
3417   si_update_ps_inputs_read_or_disabled(sctx);
3418   si_update_ps_kill_enable(sctx);
3419   si_update_vrs_flat_shading(sctx);
3420}
3421
3422static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
3423{
3424   if (shader->is_optimized) {
3425      util_queue_drop_job(&sctx->screen->shader_compiler_queue_low_priority, &shader->ready);
3426   }
3427
3428   util_queue_fence_destroy(&shader->ready);
3429
3430   /* If destroyed shaders were not unbound, the next compiled
3431    * shader variant could get the same pointer address and so
3432    * binding it to the same shader stage would be considered
3433    * a no-op, causing random behavior.
3434    */
3435   int state_index = -1;
3436
3437   switch (shader->selector->info.stage) {
3438   case MESA_SHADER_VERTEX:
3439      if (shader->key.as_ls) {
3440         if (sctx->chip_class <= GFX8)
3441            state_index = SI_STATE_IDX(ls);
3442      } else if (shader->key.as_es) {
3443         if (sctx->chip_class <= GFX8)
3444            state_index = SI_STATE_IDX(es);
3445      } else if (shader->key.as_ngg) {
3446         state_index = SI_STATE_IDX(gs);
3447      } else {
3448         state_index = SI_STATE_IDX(vs);
3449      }
3450      break;
3451   case MESA_SHADER_TESS_CTRL:
3452      state_index = SI_STATE_IDX(hs);
3453      break;
3454   case MESA_SHADER_TESS_EVAL:
3455      if (shader->key.as_es) {
3456         if (sctx->chip_class <= GFX8)
3457            state_index = SI_STATE_IDX(es);
3458      } else if (shader->key.as_ngg) {
3459         state_index = SI_STATE_IDX(gs);
3460      } else {
3461         state_index = SI_STATE_IDX(vs);
3462      }
3463      break;
3464   case MESA_SHADER_GEOMETRY:
3465      if (shader->is_gs_copy_shader)
3466         state_index = SI_STATE_IDX(vs);
3467      else
3468         state_index = SI_STATE_IDX(gs);
3469      break;
3470   case MESA_SHADER_FRAGMENT:
3471      state_index = SI_STATE_IDX(ps);
3472      break;
3473   default:;
3474   }
3475
3476   si_shader_selector_reference(sctx, &shader->previous_stage_sel, NULL);
3477   si_shader_destroy(shader);
3478   si_pm4_free_state(sctx, &shader->pm4, state_index);
3479}
3480
3481static void si_destroy_shader_selector(struct pipe_context *ctx, void *cso)
3482{
3483   struct si_context *sctx = (struct si_context *)ctx;
3484   struct si_shader_selector *sel = (struct si_shader_selector *)cso;
3485   struct si_shader *p = sel->first_variant, *c;
3486   enum pipe_shader_type type = pipe_shader_type_from_mesa(sel->info.stage);
3487
3488   util_queue_drop_job(&sctx->screen->shader_compiler_queue, &sel->ready);
3489
3490   if (sctx->shaders[type].cso == sel) {
3491      sctx->shaders[type].cso = NULL;
3492      sctx->shaders[type].current = NULL;
3493   }
3494
3495   while (p) {
3496      c = p->next_variant;
3497      si_delete_shader(sctx, p);
3498      p = c;
3499   }
3500
3501   if (sel->main_shader_part)
3502      si_delete_shader(sctx, sel->main_shader_part);
3503   if (sel->main_shader_part_ls)
3504      si_delete_shader(sctx, sel->main_shader_part_ls);
3505   if (sel->main_shader_part_es)
3506      si_delete_shader(sctx, sel->main_shader_part_es);
3507   if (sel->main_shader_part_ngg)
3508      si_delete_shader(sctx, sel->main_shader_part_ngg);
3509   if (sel->gs_copy_shader)
3510      si_delete_shader(sctx, sel->gs_copy_shader);
3511
3512   util_queue_fence_destroy(&sel->ready);
3513   simple_mtx_destroy(&sel->mutex);
3514   ralloc_free(sel->nir);
3515   free(sel->nir_binary);
3516   free(sel);
3517}
3518
3519static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
3520{
3521   struct si_context *sctx = (struct si_context *)ctx;
3522   struct si_shader_selector *sel = (struct si_shader_selector *)state;
3523
3524   si_shader_selector_reference(sctx, &sel, NULL);
3525}
3526
3527/**
3528 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3529 */
3530static void si_cs_preamble_add_vgt_flush(struct si_context *sctx)
3531{
3532   /* We shouldn't get here if registers are shadowed. */
3533   assert(!sctx->shadowed_regs);
3534
3535   if (sctx->cs_preamble_has_vgt_flush)
3536      return;
3537
3538   /* Done by Vulkan before VGT_FLUSH. */
3539   si_pm4_cmd_add(sctx->cs_preamble_state, PKT3(PKT3_EVENT_WRITE, 0, 0));
3540   si_pm4_cmd_add(sctx->cs_preamble_state, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
3541
3542   /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3543   si_pm4_cmd_add(sctx->cs_preamble_state, PKT3(PKT3_EVENT_WRITE, 0, 0));
3544   si_pm4_cmd_add(sctx->cs_preamble_state, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
3545   sctx->cs_preamble_has_vgt_flush = true;
3546}
3547
3548/**
3549 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3550 */
3551static void si_emit_vgt_flush(struct radeon_cmdbuf *cs)
3552{
3553   radeon_begin(cs);
3554
3555   /* This is required before VGT_FLUSH. */
3556   radeon_emit(PKT3(PKT3_EVENT_WRITE, 0, 0));
3557   radeon_emit(EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
3558
3559   /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3560   radeon_emit(PKT3(PKT3_EVENT_WRITE, 0, 0));
3561   radeon_emit(EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
3562   radeon_end();
3563}
3564
3565/* Initialize state related to ESGS / GSVS ring buffers */
3566bool si_update_gs_ring_buffers(struct si_context *sctx)
3567{
3568   struct si_shader_selector *es =
3569      sctx->shader.tes.cso ? sctx->shader.tes.cso : sctx->shader.vs.cso;
3570   struct si_shader_selector *gs = sctx->shader.gs.cso;
3571   struct si_pm4_state *pm4;
3572
3573   /* Chip constants. */
3574   unsigned num_se = sctx->screen->info.max_se;
3575   unsigned wave_size = 64;
3576   unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
3577   /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
3578    * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
3579    */
3580   unsigned gs_vertex_reuse = (sctx->chip_class >= GFX8 ? 32 : 16) * num_se;
3581   unsigned alignment = 256 * num_se;
3582   /* The maximum size is 63.999 MB per SE. */
3583   unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
3584
3585   /* Calculate the minimum size. */
3586   unsigned min_esgs_ring_size = align(es->esgs_itemsize * gs_vertex_reuse * wave_size, alignment);
3587
3588   /* These are recommended sizes, not minimum sizes. */
3589   unsigned esgs_ring_size =
3590      max_gs_waves * 2 * wave_size * es->esgs_itemsize * gs->gs_input_verts_per_prim;
3591   unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size * gs->max_gsvs_emit_size;
3592
3593   min_esgs_ring_size = align(min_esgs_ring_size, alignment);
3594   esgs_ring_size = align(esgs_ring_size, alignment);
3595   gsvs_ring_size = align(gsvs_ring_size, alignment);
3596
3597   esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
3598   gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
3599
3600   /* Some rings don't have to be allocated if shaders don't use them.
3601    * (e.g. no varyings between ES and GS or GS and VS)
3602    *
3603    * GFX9 doesn't have the ESGS ring.
3604    */
3605   bool update_esgs = sctx->chip_class <= GFX8 && esgs_ring_size &&
3606                      (!sctx->esgs_ring || sctx->esgs_ring->width0 < esgs_ring_size);
3607   bool update_gsvs =
3608      gsvs_ring_size && (!sctx->gsvs_ring || sctx->gsvs_ring->width0 < gsvs_ring_size);
3609
3610   if (!update_esgs && !update_gsvs)
3611      return true;
3612
3613   if (update_esgs) {
3614      pipe_resource_reference(&sctx->esgs_ring, NULL);
3615      sctx->esgs_ring =
3616         pipe_aligned_buffer_create(sctx->b.screen,
3617                                    SI_RESOURCE_FLAG_UNMAPPABLE | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
3618                                    PIPE_USAGE_DEFAULT,
3619                                    esgs_ring_size, sctx->screen->info.pte_fragment_size);
3620      if (!sctx->esgs_ring)
3621         return false;
3622   }
3623
3624   if (update_gsvs) {
3625      pipe_resource_reference(&sctx->gsvs_ring, NULL);
3626      sctx->gsvs_ring =
3627         pipe_aligned_buffer_create(sctx->b.screen,
3628                                    SI_RESOURCE_FLAG_UNMAPPABLE | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
3629                                    PIPE_USAGE_DEFAULT,
3630                                    gsvs_ring_size, sctx->screen->info.pte_fragment_size);
3631      if (!sctx->gsvs_ring)
3632         return false;
3633   }
3634
3635   /* Set ring bindings. */
3636   if (sctx->esgs_ring) {
3637      assert(sctx->chip_class <= GFX8);
3638      si_set_ring_buffer(sctx, SI_ES_RING_ESGS, sctx->esgs_ring, 0, sctx->esgs_ring->width0, true,
3639                         true, 4, 64, 0);
3640      si_set_ring_buffer(sctx, SI_GS_RING_ESGS, sctx->esgs_ring, 0, sctx->esgs_ring->width0, false,
3641                         false, 0, 0, 0);
3642   }
3643   if (sctx->gsvs_ring) {
3644      si_set_ring_buffer(sctx, SI_RING_GSVS, sctx->gsvs_ring, 0, sctx->gsvs_ring->width0, false,
3645                         false, 0, 0, 0);
3646   }
3647
3648   if (sctx->shadowed_regs) {
3649      /* These registers will be shadowed, so set them only once. */
3650      struct radeon_cmdbuf *cs = &sctx->gfx_cs;
3651
3652      assert(sctx->chip_class >= GFX7);
3653
3654      si_emit_vgt_flush(cs);
3655
3656      radeon_begin(cs);
3657
3658      /* Set the GS registers. */
3659      if (sctx->esgs_ring) {
3660         assert(sctx->chip_class <= GFX8);
3661         radeon_set_uconfig_reg(R_030900_VGT_ESGS_RING_SIZE,
3662                                sctx->esgs_ring->width0 / 256);
3663      }
3664      if (sctx->gsvs_ring) {
3665         radeon_set_uconfig_reg(R_030904_VGT_GSVS_RING_SIZE,
3666                                sctx->gsvs_ring->width0 / 256);
3667      }
3668      radeon_end();
3669      return true;
3670   }
3671
3672   /* The codepath without register shadowing. */
3673   /* Create the "cs_preamble_gs_rings" state. */
3674   pm4 = CALLOC_STRUCT(si_pm4_state);
3675   if (!pm4)
3676      return false;
3677
3678   if (sctx->chip_class >= GFX7) {
3679      if (sctx->esgs_ring) {
3680         assert(sctx->chip_class <= GFX8);
3681         si_pm4_set_reg(pm4, R_030900_VGT_ESGS_RING_SIZE, sctx->esgs_ring->width0 / 256);
3682      }
3683      if (sctx->gsvs_ring)
3684         si_pm4_set_reg(pm4, R_030904_VGT_GSVS_RING_SIZE, sctx->gsvs_ring->width0 / 256);
3685   } else {
3686      if (sctx->esgs_ring)
3687         si_pm4_set_reg(pm4, R_0088C8_VGT_ESGS_RING_SIZE, sctx->esgs_ring->width0 / 256);
3688      if (sctx->gsvs_ring)
3689         si_pm4_set_reg(pm4, R_0088CC_VGT_GSVS_RING_SIZE, sctx->gsvs_ring->width0 / 256);
3690   }
3691
3692   /* Set the state. */
3693   if (sctx->cs_preamble_gs_rings)
3694      si_pm4_free_state(sctx, sctx->cs_preamble_gs_rings, ~0);
3695   sctx->cs_preamble_gs_rings = pm4;
3696
3697   si_cs_preamble_add_vgt_flush(sctx);
3698
3699   /* Flush the context to re-emit both cs_preamble states. */
3700   sctx->initial_gfx_cs_size = 0; /* force flush */
3701   si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
3702
3703   return true;
3704}
3705
3706static void si_shader_lock(struct si_shader *shader)
3707{
3708   simple_mtx_lock(&shader->selector->mutex);
3709   if (shader->previous_stage_sel) {
3710      assert(shader->previous_stage_sel != shader->selector);
3711      simple_mtx_lock(&shader->previous_stage_sel->mutex);
3712   }
3713}
3714
3715static void si_shader_unlock(struct si_shader *shader)
3716{
3717   if (shader->previous_stage_sel)
3718      simple_mtx_unlock(&shader->previous_stage_sel->mutex);
3719   simple_mtx_unlock(&shader->selector->mutex);
3720}
3721
3722/**
3723 * @returns 1 if \p sel has been updated to use a new scratch buffer
3724 *          0 if not
3725 *          < 0 if there was a failure
3726 */
3727static int si_update_scratch_buffer(struct si_context *sctx, struct si_shader *shader)
3728{
3729   uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
3730
3731   if (!shader)
3732      return 0;
3733
3734   /* This shader doesn't need a scratch buffer */
3735   if (shader->config.scratch_bytes_per_wave == 0)
3736      return 0;
3737
3738   /* Prevent race conditions when updating:
3739    * - si_shader::scratch_bo
3740    * - si_shader::binary::code
3741    * - si_shader::previous_stage::binary::code.
3742    */
3743   si_shader_lock(shader);
3744
3745   /* This shader is already configured to use the current
3746    * scratch buffer. */
3747   if (shader->scratch_bo == sctx->scratch_buffer) {
3748      si_shader_unlock(shader);
3749      return 0;
3750   }
3751
3752   assert(sctx->scratch_buffer);
3753
3754   /* Replace the shader bo with a new bo that has the relocs applied. */
3755   if (!si_shader_binary_upload(sctx->screen, shader, scratch_va)) {
3756      si_shader_unlock(shader);
3757      return -1;
3758   }
3759
3760   /* Update the shader state to use the new shader bo. */
3761   si_shader_init_pm4_state(sctx->screen, shader);
3762
3763   si_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
3764
3765   si_shader_unlock(shader);
3766   return 1;
3767}
3768
3769static struct si_shader *si_get_tcs_current(struct si_context *sctx)
3770{
3771   if (!sctx->shader.tes.cso)
3772      return NULL; /* tessellation disabled */
3773
3774   return sctx->shader.tcs.cso ? sctx->shader.tcs.current : sctx->fixed_func_tcs_shader.current;
3775}
3776
3777static bool si_update_scratch_relocs(struct si_context *sctx)
3778{
3779   struct si_shader *tcs = si_get_tcs_current(sctx);
3780   int r;
3781
3782   /* Update the shaders, so that they are using the latest scratch.
3783    * The scratch buffer may have been changed since these shaders were
3784    * last used, so we still need to try to update them, even if they
3785    * require scratch buffers smaller than the current size.
3786    */
3787   r = si_update_scratch_buffer(sctx, sctx->shader.ps.current);
3788   if (r < 0)
3789      return false;
3790   if (r == 1)
3791      si_pm4_bind_state(sctx, ps, sctx->shader.ps.current);
3792
3793   r = si_update_scratch_buffer(sctx, sctx->shader.gs.current);
3794   if (r < 0)
3795      return false;
3796   if (r == 1)
3797      si_pm4_bind_state(sctx, gs, sctx->shader.gs.current);
3798
3799   r = si_update_scratch_buffer(sctx, tcs);
3800   if (r < 0)
3801      return false;
3802   if (r == 1)
3803      si_pm4_bind_state(sctx, hs, tcs);
3804
3805   /* VS can be bound as LS, ES, or VS. */
3806   r = si_update_scratch_buffer(sctx, sctx->shader.vs.current);
3807   if (r < 0)
3808      return false;
3809   if (r == 1) {
3810      if (sctx->shader.vs.current->key.as_ls)
3811         si_pm4_bind_state(sctx, ls, sctx->shader.vs.current);
3812      else if (sctx->shader.vs.current->key.as_es)
3813         si_pm4_bind_state(sctx, es, sctx->shader.vs.current);
3814      else if (sctx->shader.vs.current->key.as_ngg)
3815         si_pm4_bind_state(sctx, gs, sctx->shader.vs.current);
3816      else
3817         si_pm4_bind_state(sctx, vs, sctx->shader.vs.current);
3818   }
3819
3820   /* TES can be bound as ES or VS. */
3821   r = si_update_scratch_buffer(sctx, sctx->shader.tes.current);
3822   if (r < 0)
3823      return false;
3824   if (r == 1) {
3825      if (sctx->shader.tes.current->key.as_es)
3826         si_pm4_bind_state(sctx, es, sctx->shader.tes.current);
3827      else if (sctx->shader.tes.current->key.as_ngg)
3828         si_pm4_bind_state(sctx, gs, sctx->shader.tes.current);
3829      else
3830         si_pm4_bind_state(sctx, vs, sctx->shader.tes.current);
3831   }
3832
3833   return true;
3834}
3835
3836bool si_update_spi_tmpring_size(struct si_context *sctx, unsigned bytes)
3837{
3838   /* SPI_TMPRING_SIZE.WAVESIZE must be constant for each scratch buffer.
3839    * There are 2 cases to handle:
3840    *
3841    * - If the current needed size is less than the maximum seen size,
3842    *   use the maximum seen size, so that WAVESIZE remains the same.
3843    *
3844    * - If the current needed size is greater than the maximum seen size,
3845    *   the scratch buffer is reallocated, so we can increase WAVESIZE.
3846    *
3847    * Shaders that set SCRATCH_EN=0 don't allocate scratch space.
3848    * Otherwise, the number of waves that can use scratch is
3849    * SPI_TMPRING_SIZE.WAVES.
3850    */
3851   sctx->max_seen_scratch_bytes_per_wave = MAX2(sctx->max_seen_scratch_bytes_per_wave, bytes);
3852
3853   unsigned scratch_needed_size = sctx->max_seen_scratch_bytes_per_wave * sctx->scratch_waves;
3854   unsigned spi_tmpring_size;
3855
3856   if (scratch_needed_size > 0) {
3857      if (!sctx->scratch_buffer || scratch_needed_size > sctx->scratch_buffer->b.b.width0) {
3858         /* Create a bigger scratch buffer */
3859         si_resource_reference(&sctx->scratch_buffer, NULL);
3860
3861         sctx->scratch_buffer = si_aligned_buffer_create(
3862            &sctx->screen->b,
3863            SI_RESOURCE_FLAG_UNMAPPABLE | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
3864            PIPE_USAGE_DEFAULT, scratch_needed_size,
3865            sctx->screen->info.pte_fragment_size);
3866         if (!sctx->scratch_buffer)
3867            return false;
3868
3869         si_context_add_resource_size(sctx, &sctx->scratch_buffer->b.b);
3870      }
3871
3872      if (!si_update_scratch_relocs(sctx))
3873         return false;
3874   }
3875
3876   /* The LLVM shader backend should be reporting aligned scratch_sizes. */
3877   assert((scratch_needed_size & ~0x3FF) == scratch_needed_size &&
3878          "scratch size should already be aligned correctly.");
3879
3880   spi_tmpring_size = S_0286E8_WAVES(sctx->scratch_waves) |
3881                      S_0286E8_WAVESIZE(sctx->max_seen_scratch_bytes_per_wave >> 10);
3882   if (spi_tmpring_size != sctx->spi_tmpring_size) {
3883      sctx->spi_tmpring_size = spi_tmpring_size;
3884      si_mark_atom_dirty(sctx, &sctx->atoms.s.scratch_state);
3885   }
3886   return true;
3887}
3888
3889void si_init_tess_factor_ring(struct si_context *sctx)
3890{
3891   assert(!sctx->tess_rings);
3892   assert(((sctx->screen->tess_factor_ring_size / 4) & C_030938_SIZE) == 0);
3893
3894   /* The address must be aligned to 2^19, because the shader only
3895    * receives the high 13 bits.
3896    */
3897   sctx->tess_rings = pipe_aligned_buffer_create(
3898      sctx->b.screen, SI_RESOURCE_FLAG_32BIT | SI_RESOURCE_FLAG_DRIVER_INTERNAL, PIPE_USAGE_DEFAULT,
3899      sctx->screen->tess_offchip_ring_size + sctx->screen->tess_factor_ring_size, 1 << 19);
3900   if (!sctx->tess_rings)
3901      return;
3902
3903   if (sctx->screen->info.has_tmz_support) {
3904      sctx->tess_rings_tmz = pipe_aligned_buffer_create(
3905         sctx->b.screen,
3906         PIPE_RESOURCE_FLAG_ENCRYPTED | SI_RESOURCE_FLAG_32BIT | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
3907         PIPE_USAGE_DEFAULT,
3908         sctx->screen->tess_offchip_ring_size + sctx->screen->tess_factor_ring_size, 1 << 19);
3909   }
3910
3911   uint64_t factor_va =
3912      si_resource(sctx->tess_rings)->gpu_address + sctx->screen->tess_offchip_ring_size;
3913
3914   if (sctx->shadowed_regs) {
3915      /* These registers will be shadowed, so set them only once. */
3916      /* TODO: tmz + shadowed_regs support */
3917      struct radeon_cmdbuf *cs = &sctx->gfx_cs;
3918
3919      assert(sctx->chip_class >= GFX7);
3920
3921      radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, si_resource(sctx->tess_rings),
3922                                RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
3923      si_emit_vgt_flush(cs);
3924
3925      /* Set tessellation registers. */
3926      radeon_begin(cs);
3927      radeon_set_uconfig_reg(R_030938_VGT_TF_RING_SIZE,
3928                             S_030938_SIZE(sctx->screen->tess_factor_ring_size / 4));
3929      radeon_set_uconfig_reg(R_030940_VGT_TF_MEMORY_BASE, factor_va >> 8);
3930      if (sctx->chip_class >= GFX10) {
3931         radeon_set_uconfig_reg(R_030984_VGT_TF_MEMORY_BASE_HI_UMD,
3932                                S_030984_BASE_HI(factor_va >> 40));
3933      } else if (sctx->chip_class == GFX9) {
3934         radeon_set_uconfig_reg(R_030944_VGT_TF_MEMORY_BASE_HI,
3935                                S_030944_BASE_HI(factor_va >> 40));
3936      }
3937      radeon_set_uconfig_reg(R_03093C_VGT_HS_OFFCHIP_PARAM,
3938                             sctx->screen->vgt_hs_offchip_param);
3939      radeon_end();
3940      return;
3941   }
3942
3943   /* The codepath without register shadowing. */
3944   si_cs_preamble_add_vgt_flush(sctx);
3945
3946   /* Append these registers to the init config state. */
3947   if (sctx->chip_class >= GFX7) {
3948      si_pm4_set_reg(sctx->cs_preamble_state, R_030938_VGT_TF_RING_SIZE,
3949                     S_030938_SIZE(sctx->screen->tess_factor_ring_size / 4));
3950      si_pm4_set_reg(sctx->cs_preamble_state, R_030940_VGT_TF_MEMORY_BASE, factor_va >> 8);
3951      if (sctx->chip_class >= GFX10)
3952         si_pm4_set_reg(sctx->cs_preamble_state, R_030984_VGT_TF_MEMORY_BASE_HI_UMD,
3953                        S_030984_BASE_HI(factor_va >> 40));
3954      else if (sctx->chip_class == GFX9)
3955         si_pm4_set_reg(sctx->cs_preamble_state, R_030944_VGT_TF_MEMORY_BASE_HI,
3956                        S_030944_BASE_HI(factor_va >> 40));
3957      si_pm4_set_reg(sctx->cs_preamble_state, R_03093C_VGT_HS_OFFCHIP_PARAM,
3958                     sctx->screen->vgt_hs_offchip_param);
3959   } else {
3960      struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
3961
3962      si_pm4_set_reg(pm4, R_008988_VGT_TF_RING_SIZE,
3963                     S_008988_SIZE(sctx->screen->tess_factor_ring_size / 4));
3964      si_pm4_set_reg(pm4, R_0089B8_VGT_TF_MEMORY_BASE, factor_va >> 8);
3965      si_pm4_set_reg(pm4, R_0089B0_VGT_HS_OFFCHIP_PARAM,
3966                     sctx->screen->vgt_hs_offchip_param);
3967      sctx->cs_preamble_tess_rings = pm4;
3968
3969      if (sctx->screen->info.has_tmz_support) {
3970         pm4 = CALLOC_STRUCT(si_pm4_state);
3971         uint64_t factor_va_tmz =
3972            si_resource(sctx->tess_rings_tmz)->gpu_address + sctx->screen->tess_offchip_ring_size;
3973         si_pm4_set_reg(pm4, R_008988_VGT_TF_RING_SIZE,
3974                     S_008988_SIZE(sctx->screen->tess_factor_ring_size / 4));
3975         si_pm4_set_reg(pm4, R_0089B8_VGT_TF_MEMORY_BASE, factor_va_tmz >> 8);
3976         si_pm4_set_reg(pm4, R_0089B0_VGT_HS_OFFCHIP_PARAM,
3977                        sctx->screen->vgt_hs_offchip_param);
3978         sctx->cs_preamble_tess_rings_tmz = pm4;
3979      }
3980   }
3981
3982   /* Flush the context to re-emit the cs_preamble state.
3983    * This is done only once in a lifetime of a context.
3984    */
3985   sctx->initial_gfx_cs_size = 0; /* force flush */
3986   si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
3987}
3988
3989struct si_pm4_state *si_build_vgt_shader_config(struct si_screen *screen, union si_vgt_stages_key key)
3990{
3991   struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
3992   uint32_t stages = 0;
3993
3994   if (key.u.tess) {
3995      stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) | S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3996
3997      if (key.u.gs)
3998         stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) | S_028B54_GS_EN(1);
3999      else if (key.u.ngg)
4000         stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS);
4001      else
4002         stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
4003   } else if (key.u.gs) {
4004      stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) | S_028B54_GS_EN(1);
4005   } else if (key.u.ngg) {
4006      stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL);
4007   }
4008
4009   if (key.u.ngg) {
4010      stages |= S_028B54_PRIMGEN_EN(1) |
4011                S_028B54_NGG_WAVE_ID_EN(key.u.streamout) |
4012                S_028B54_PRIMGEN_PASSTHRU_EN(key.u.ngg_passthrough) |
4013                S_028B54_PRIMGEN_PASSTHRU_NO_MSG(key.u.ngg_passthrough &&
4014                                                 screen->info.family >= CHIP_DIMGREY_CAVEFISH);
4015   } else if (key.u.gs)
4016      stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
4017
4018   if (screen->info.chip_class >= GFX9)
4019      stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
4020
4021   if (screen->info.chip_class >= GFX10 && screen->ge_wave_size == 32) {
4022      stages |= S_028B54_HS_W32_EN(1) |
4023                S_028B54_GS_W32_EN(key.u.ngg) | /* legacy GS only supports Wave64 */
4024                S_028B54_VS_W32_EN(1);
4025   }
4026
4027   si_pm4_set_reg(pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
4028   return pm4;
4029}
4030
4031static void si_emit_scratch_state(struct si_context *sctx)
4032{
4033   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
4034
4035   radeon_begin(cs);
4036   radeon_set_context_reg(R_0286E8_SPI_TMPRING_SIZE, sctx->spi_tmpring_size);
4037   radeon_end();
4038
4039   if (sctx->scratch_buffer) {
4040      radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, sctx->scratch_buffer, RADEON_USAGE_READWRITE,
4041                                RADEON_PRIO_SCRATCH_BUFFER);
4042   }
4043}
4044
4045void si_init_screen_live_shader_cache(struct si_screen *sscreen)
4046{
4047   util_live_shader_cache_init(&sscreen->live_shader_cache, si_create_shader_selector,
4048                               si_destroy_shader_selector);
4049}
4050
4051void si_init_shader_functions(struct si_context *sctx)
4052{
4053   sctx->atoms.s.scratch_state.emit = si_emit_scratch_state;
4054
4055   sctx->b.create_vs_state = si_create_shader;
4056   sctx->b.create_tcs_state = si_create_shader;
4057   sctx->b.create_tes_state = si_create_shader;
4058   sctx->b.create_gs_state = si_create_shader;
4059   sctx->b.create_fs_state = si_create_shader;
4060
4061   sctx->b.bind_vs_state = si_bind_vs_shader;
4062   sctx->b.bind_tcs_state = si_bind_tcs_shader;
4063   sctx->b.bind_tes_state = si_bind_tes_shader;
4064   sctx->b.bind_gs_state = si_bind_gs_shader;
4065   sctx->b.bind_fs_state = si_bind_ps_shader;
4066
4067   sctx->b.delete_vs_state = si_delete_shader_selector;
4068   sctx->b.delete_tcs_state = si_delete_shader_selector;
4069   sctx->b.delete_tes_state = si_delete_shader_selector;
4070   sctx->b.delete_gs_state = si_delete_shader_selector;
4071   sctx->b.delete_fs_state = si_delete_shader_selector;
4072}
4073