2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "util/mesa-sha1.h"
29 #include "util/u_atomic.h"
30 #include "radv_debug.h"
31 #include "radv_private.h"
32 #include "radv_shader.h"
34 #include "nir/nir_builder.h"
35 #include "spirv/nir_spirv.h"
37 #include <llvm-c/Core.h>
38 #include <llvm-c/TargetMachine.h>
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "ac_nir_to_llvm.h"
45 #include "vk_format.h"
46 #include "util/debug.h"
47 #include "ac_exp_param.h"
49 #include "util/string_buffer.h"
51 static const struct nir_shader_compiler_options nir_options
= {
52 .vertex_id_zero_based
= true,
56 .lower_device_index_to_zero
= true,
60 .lower_pack_snorm_2x16
= true,
61 .lower_pack_snorm_4x8
= true,
62 .lower_pack_unorm_2x16
= true,
63 .lower_pack_unorm_4x8
= true,
64 .lower_unpack_snorm_2x16
= true,
65 .lower_unpack_snorm_4x8
= true,
66 .lower_unpack_unorm_2x16
= true,
67 .lower_unpack_unorm_4x8
= true,
68 .lower_extract_byte
= true,
69 .lower_extract_word
= true,
72 .vs_inputs_dual_locations
= true,
73 .max_unroll_iterations
= 32
76 VkResult
radv_CreateShaderModule(
78 const VkShaderModuleCreateInfo
* pCreateInfo
,
79 const VkAllocationCallbacks
* pAllocator
,
80 VkShaderModule
* pShaderModule
)
82 RADV_FROM_HANDLE(radv_device
, device
, _device
);
83 struct radv_shader_module
*module
;
85 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
86 assert(pCreateInfo
->flags
== 0);
88 module
= vk_alloc2(&device
->alloc
, pAllocator
,
89 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
90 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
92 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
95 module
->size
= pCreateInfo
->codeSize
;
96 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
98 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
100 *pShaderModule
= radv_shader_module_to_handle(module
);
105 void radv_DestroyShaderModule(
107 VkShaderModule _module
,
108 const VkAllocationCallbacks
* pAllocator
)
110 RADV_FROM_HANDLE(radv_device
, device
, _device
);
111 RADV_FROM_HANDLE(radv_shader_module
, module
, _module
);
116 vk_free2(&device
->alloc
, pAllocator
, module
);
120 radv_optimize_nir(struct nir_shader
*shader
, bool optimize_conservatively
)
127 NIR_PASS_V(shader
, nir_lower_vars_to_ssa
);
128 NIR_PASS_V(shader
, nir_lower_pack
);
129 NIR_PASS_V(shader
, nir_lower_alu_to_scalar
);
130 NIR_PASS_V(shader
, nir_lower_phis_to_scalar
);
132 NIR_PASS(progress
, shader
, nir_copy_prop
);
133 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
134 NIR_PASS(progress
, shader
, nir_opt_dce
);
135 if (nir_opt_trivial_continues(shader
)) {
137 NIR_PASS(progress
, shader
, nir_copy_prop
);
138 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
139 NIR_PASS(progress
, shader
, nir_opt_dce
);
141 NIR_PASS(progress
, shader
, nir_opt_if
);
142 NIR_PASS(progress
, shader
, nir_opt_dead_cf
);
143 NIR_PASS(progress
, shader
, nir_opt_cse
);
144 NIR_PASS(progress
, shader
, nir_opt_peephole_select
, 8);
145 NIR_PASS(progress
, shader
, nir_opt_algebraic
);
146 NIR_PASS(progress
, shader
, nir_opt_constant_folding
);
147 NIR_PASS(progress
, shader
, nir_opt_undef
);
148 NIR_PASS(progress
, shader
, nir_opt_conditional_discard
);
149 if (shader
->options
->max_unroll_iterations
) {
150 NIR_PASS(progress
, shader
, nir_opt_loop_unroll
, 0);
152 } while (progress
&& !optimize_conservatively
);
154 NIR_PASS(progress
, shader
, nir_opt_shrink_load
);
155 NIR_PASS(progress
, shader
, nir_opt_move_load_ubo
);
159 radv_shader_compile_to_nir(struct radv_device
*device
,
160 struct radv_shader_module
*module
,
161 const char *entrypoint_name
,
162 gl_shader_stage stage
,
163 const VkSpecializationInfo
*spec_info
,
164 const VkPipelineCreateFlags flags
)
167 nir_function
*entry_point
;
169 /* Some things such as our meta clear/blit code will give us a NIR
170 * shader directly. In that case, we just ignore the SPIR-V entirely
171 * and just use the NIR shader */
173 nir
->options
= &nir_options
;
174 nir_validate_shader(nir
);
176 assert(exec_list_length(&nir
->functions
) == 1);
177 struct exec_node
*node
= exec_list_get_head(&nir
->functions
);
178 entry_point
= exec_node_data(nir_function
, node
, node
);
180 uint32_t *spirv
= (uint32_t *) module
->data
;
181 assert(module
->size
% 4 == 0);
183 if (device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SPIRV
)
184 radv_print_spirv(spirv
, module
->size
, stderr
);
186 uint32_t num_spec_entries
= 0;
187 struct nir_spirv_specialization
*spec_entries
= NULL
;
188 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
189 num_spec_entries
= spec_info
->mapEntryCount
;
190 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
191 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
192 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
193 const void *data
= spec_info
->pData
+ entry
.offset
;
194 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
196 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
197 if (spec_info
->dataSize
== 8)
198 spec_entries
[i
].data64
= *(const uint64_t *)data
;
200 spec_entries
[i
].data32
= *(const uint32_t *)data
;
203 const struct spirv_to_nir_options spirv_options
= {
205 .device_group
= true,
206 .draw_parameters
= true,
208 .image_read_without_format
= true,
209 .image_write_without_format
= true,
210 .tessellation
= true,
213 .subgroup_ballot
= true,
214 .subgroup_basic
= true,
215 .subgroup_quad
= true,
216 .subgroup_shuffle
= true,
217 .subgroup_vote
= true,
218 .variable_pointers
= true,
220 .trinary_minmax
= true,
221 .shader_viewport_index_layer
= true,
222 .descriptor_array_dynamic_indexing
= true,
223 .runtime_descriptor_array
= true,
226 entry_point
= spirv_to_nir(spirv
, module
->size
/ 4,
227 spec_entries
, num_spec_entries
,
228 stage
, entrypoint_name
,
229 &spirv_options
, &nir_options
);
230 nir
= entry_point
->shader
;
231 assert(nir
->info
.stage
== stage
);
232 nir_validate_shader(nir
);
236 /* We have to lower away local constant initializers right before we
237 * inline functions. That way they get properly initialized at the top
238 * of the function and not at the top of its caller.
240 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_local
);
241 NIR_PASS_V(nir
, nir_lower_returns
);
242 NIR_PASS_V(nir
, nir_inline_functions
);
244 /* Pick off the single entrypoint that we want */
245 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
246 if (func
!= entry_point
)
247 exec_node_remove(&func
->node
);
249 assert(exec_list_length(&nir
->functions
) == 1);
250 entry_point
->name
= ralloc_strdup(entry_point
, "main");
252 /* Make sure we lower constant initializers on output variables so that
253 * nir_remove_dead_variables below sees the corresponding stores
255 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_shader_out
);
257 NIR_PASS_V(nir
, nir_remove_dead_variables
,
258 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
260 /* Now that we've deleted all but the main function, we can go ahead and
261 * lower the rest of the constant initializers.
263 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
264 NIR_PASS_V(nir
, nir_lower_system_values
);
265 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
268 /* Vulkan uses the separate-shader linking model */
269 nir
->info
.separate_shader
= true;
271 nir_shader_gather_info(nir
, entry_point
->impl
);
273 static const nir_lower_tex_options tex_options
= {
277 nir_lower_tex(nir
, &tex_options
);
279 nir_lower_vars_to_ssa(nir
);
280 nir_lower_var_copies(nir
);
281 nir_lower_global_vars_to_local(nir
);
282 nir_remove_dead_variables(nir
, nir_var_local
);
283 nir_lower_subgroups(nir
, &(struct nir_lower_subgroups_options
) {
285 .ballot_bit_size
= 64,
286 .lower_to_scalar
= 1,
287 .lower_subgroup_masks
= 1,
289 .lower_shuffle_to_32bit
= 1,
290 .lower_vote_eq_to_ballot
= 1,
293 if (!(flags
& VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
))
294 radv_optimize_nir(nir
, false);
296 /* Indirect lowering must be called after the radv_optimize_nir() loop
297 * has been called at least once. Otherwise indirect lowering can
298 * bloat the instruction count of the loop and cause it to be
299 * considered too large for unrolling.
301 ac_lower_indirect_derefs(nir
, device
->physical_device
->rad_info
.chip_class
);
302 radv_optimize_nir(nir
, flags
& VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
);
308 radv_alloc_shader_memory(struct radv_device
*device
,
309 struct radv_shader_variant
*shader
)
311 mtx_lock(&device
->shader_slab_mutex
);
312 list_for_each_entry(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
314 list_for_each_entry(struct radv_shader_variant
, s
, &slab
->shaders
, slab_list
) {
315 if (s
->bo_offset
- offset
>= shader
->code_size
) {
316 shader
->bo
= slab
->bo
;
317 shader
->bo_offset
= offset
;
318 list_addtail(&shader
->slab_list
, &s
->slab_list
);
319 mtx_unlock(&device
->shader_slab_mutex
);
320 return slab
->ptr
+ offset
;
322 offset
= align_u64(s
->bo_offset
+ s
->code_size
, 256);
324 if (slab
->size
- offset
>= shader
->code_size
) {
325 shader
->bo
= slab
->bo
;
326 shader
->bo_offset
= offset
;
327 list_addtail(&shader
->slab_list
, &slab
->shaders
);
328 mtx_unlock(&device
->shader_slab_mutex
);
329 return slab
->ptr
+ offset
;
333 mtx_unlock(&device
->shader_slab_mutex
);
334 struct radv_shader_slab
*slab
= calloc(1, sizeof(struct radv_shader_slab
));
336 slab
->size
= 256 * 1024;
337 slab
->bo
= device
->ws
->buffer_create(device
->ws
, slab
->size
, 256,
339 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
340 device
->physical_device
->cpdma_prefetch_writes_memory
?
341 0 : RADEON_FLAG_READ_ONLY
);
342 slab
->ptr
= (char*)device
->ws
->buffer_map(slab
->bo
);
343 list_inithead(&slab
->shaders
);
345 mtx_lock(&device
->shader_slab_mutex
);
346 list_add(&slab
->slabs
, &device
->shader_slabs
);
348 shader
->bo
= slab
->bo
;
349 shader
->bo_offset
= 0;
350 list_add(&shader
->slab_list
, &slab
->shaders
);
351 mtx_unlock(&device
->shader_slab_mutex
);
356 radv_destroy_shader_slabs(struct radv_device
*device
)
358 list_for_each_entry_safe(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
359 device
->ws
->buffer_destroy(slab
->bo
);
362 mtx_destroy(&device
->shader_slab_mutex
);
366 radv_fill_shader_variant(struct radv_device
*device
,
367 struct radv_shader_variant
*variant
,
368 struct ac_shader_binary
*binary
,
369 gl_shader_stage stage
)
371 bool scratch_enabled
= variant
->config
.scratch_bytes_per_wave
> 0;
372 struct radv_shader_info
*info
= &variant
->info
.info
;
373 unsigned vgpr_comp_cnt
= 0;
375 variant
->code_size
= binary
->code_size
;
376 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
377 S_00B12C_SCRATCH_EN(scratch_enabled
);
379 variant
->rsrc1
= S_00B848_VGPRS((variant
->config
.num_vgprs
- 1) / 4) |
380 S_00B848_SGPRS((variant
->config
.num_sgprs
- 1) / 8) |
381 S_00B848_DX10_CLAMP(1) |
382 S_00B848_FLOAT_MODE(variant
->config
.float_mode
);
385 case MESA_SHADER_TESS_EVAL
:
387 variant
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
389 case MESA_SHADER_TESS_CTRL
:
390 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
391 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
393 variant
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
396 case MESA_SHADER_VERTEX
:
397 case MESA_SHADER_GEOMETRY
:
398 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
400 case MESA_SHADER_FRAGMENT
:
402 case MESA_SHADER_COMPUTE
:
404 S_00B84C_TGID_X_EN(info
->cs
.uses_block_id
[0]) |
405 S_00B84C_TGID_Y_EN(info
->cs
.uses_block_id
[1]) |
406 S_00B84C_TGID_Z_EN(info
->cs
.uses_block_id
[2]) |
407 S_00B84C_TIDIG_COMP_CNT(info
->cs
.uses_thread_id
[2] ? 2 :
408 info
->cs
.uses_thread_id
[1] ? 1 : 0) |
409 S_00B84C_TG_SIZE_EN(info
->cs
.uses_local_invocation_idx
) |
410 S_00B84C_LDS_SIZE(variant
->config
.lds_size
);
413 unreachable("unsupported shader type");
417 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
418 stage
== MESA_SHADER_GEOMETRY
) {
419 unsigned es_type
= variant
->info
.gs
.es_type
;
420 unsigned gs_vgpr_comp_cnt
, es_vgpr_comp_cnt
;
422 if (es_type
== MESA_SHADER_VERTEX
) {
423 es_vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
424 } else if (es_type
== MESA_SHADER_TESS_EVAL
) {
425 es_vgpr_comp_cnt
= 3;
427 unreachable("invalid shader ES type");
430 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
431 * VGPR[0:4] are always loaded.
433 if (info
->uses_invocation_id
) {
434 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
435 } else if (info
->uses_prim_id
) {
436 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
437 } else if (variant
->info
.gs
.vertices_in
>= 3) {
438 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
440 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
443 variant
->rsrc1
|= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
444 variant
->rsrc2
|= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
445 S_00B22C_OC_LDS_EN(es_type
== MESA_SHADER_TESS_EVAL
);
446 } else if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
447 stage
== MESA_SHADER_TESS_CTRL
) {
448 variant
->rsrc1
|= S_00B428_LS_VGPR_COMP_CNT(vgpr_comp_cnt
);
450 variant
->rsrc1
|= S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
);
453 void *ptr
= radv_alloc_shader_memory(device
, variant
);
454 memcpy(ptr
, binary
->code
, binary
->code_size
);
457 static struct radv_shader_variant
*
458 shader_variant_create(struct radv_device
*device
,
459 struct radv_shader_module
*module
,
460 struct nir_shader
* const *shaders
,
462 gl_shader_stage stage
,
463 struct radv_nir_compiler_options
*options
,
466 unsigned *code_size_out
)
468 enum radeon_family chip_family
= device
->physical_device
->rad_info
.family
;
469 enum ac_target_machine_options tm_options
= 0;
470 struct radv_shader_variant
*variant
;
471 struct ac_shader_binary binary
;
472 LLVMTargetMachineRef tm
;
474 variant
= calloc(1, sizeof(struct radv_shader_variant
));
478 options
->family
= chip_family
;
479 options
->chip_class
= device
->physical_device
->rad_info
.chip_class
;
480 options
->dump_shader
= radv_can_dump_shader(device
, module
, gs_copy_shader
);
481 options
->dump_preoptir
= options
->dump_shader
&&
482 device
->instance
->debug_flags
& RADV_DEBUG_PREOPTIR
;
483 options
->record_llvm_ir
= device
->keep_shader_info
;
484 options
->tess_offchip_block_dw_size
= device
->tess_offchip_block_dw_size
;
485 options
->address32_hi
= device
->physical_device
->rad_info
.address32_hi
;
487 if (options
->supports_spill
)
488 tm_options
|= AC_TM_SUPPORTS_SPILL
;
489 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
490 tm_options
|= AC_TM_SISCHED
;
491 tm
= ac_create_target_machine(chip_family
, tm_options
, NULL
);
493 if (gs_copy_shader
) {
494 assert(shader_count
== 1);
495 radv_compile_gs_copy_shader(tm
, *shaders
, &binary
,
496 &variant
->config
, &variant
->info
,
499 radv_compile_nir_shader(tm
, &binary
, &variant
->config
,
500 &variant
->info
, shaders
, shader_count
,
504 LLVMDisposeTargetMachine(tm
);
506 radv_fill_shader_variant(device
, variant
, &binary
, stage
);
509 *code_out
= binary
.code
;
510 *code_size_out
= binary
.code_size
;
515 free(binary
.global_symbol_offsets
);
517 variant
->ref_count
= 1;
519 if (device
->keep_shader_info
) {
520 variant
->disasm_string
= binary
.disasm_string
;
521 variant
->llvm_ir_string
= binary
.llvm_ir_string
;
522 if (!gs_copy_shader
&& !module
->nir
) {
523 variant
->nir
= *shaders
;
524 variant
->spirv
= (uint32_t *)module
->data
;
525 variant
->spirv_size
= module
->size
;
528 free(binary
.disasm_string
);
534 struct radv_shader_variant
*
535 radv_shader_variant_create(struct radv_device
*device
,
536 struct radv_shader_module
*module
,
537 struct nir_shader
*const *shaders
,
539 struct radv_pipeline_layout
*layout
,
540 const struct radv_shader_variant_key
*key
,
542 unsigned *code_size_out
)
544 struct radv_nir_compiler_options options
= {0};
546 options
.layout
= layout
;
550 options
.unsafe_math
= !!(device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
);
551 options
.supports_spill
= true;
553 return shader_variant_create(device
, module
, shaders
, shader_count
, shaders
[shader_count
- 1]->info
.stage
,
554 &options
, false, code_out
, code_size_out
);
557 struct radv_shader_variant
*
558 radv_create_gs_copy_shader(struct radv_device
*device
,
559 struct nir_shader
*shader
,
561 unsigned *code_size_out
,
564 struct radv_nir_compiler_options options
= {0};
566 options
.key
.has_multiview_view_index
= multiview
;
568 return shader_variant_create(device
, NULL
, &shader
, 1, MESA_SHADER_VERTEX
,
569 &options
, true, code_out
, code_size_out
);
573 radv_shader_variant_destroy(struct radv_device
*device
,
574 struct radv_shader_variant
*variant
)
576 if (!p_atomic_dec_zero(&variant
->ref_count
))
579 mtx_lock(&device
->shader_slab_mutex
);
580 list_del(&variant
->slab_list
);
581 mtx_unlock(&device
->shader_slab_mutex
);
583 ralloc_free(variant
->nir
);
584 free(variant
->disasm_string
);
585 free(variant
->llvm_ir_string
);
590 radv_get_shader_name(struct radv_shader_variant
*var
, gl_shader_stage stage
)
593 case MESA_SHADER_VERTEX
: return var
->info
.vs
.as_ls
? "Vertex Shader as LS" : var
->info
.vs
.as_es
? "Vertex Shader as ES" : "Vertex Shader as VS";
594 case MESA_SHADER_GEOMETRY
: return "Geometry Shader";
595 case MESA_SHADER_FRAGMENT
: return "Pixel Shader";
596 case MESA_SHADER_COMPUTE
: return "Compute Shader";
597 case MESA_SHADER_TESS_CTRL
: return "Tessellation Control Shader";
598 case MESA_SHADER_TESS_EVAL
: return var
->info
.tes
.as_es
? "Tessellation Evaluation Shader as ES" : "Tessellation Evaluation Shader as VS";
600 return "Unknown shader";
605 generate_shader_stats(struct radv_device
*device
,
606 struct radv_shader_variant
*variant
,
607 gl_shader_stage stage
,
608 struct _mesa_string_buffer
*buf
)
610 unsigned lds_increment
= device
->physical_device
->rad_info
.chip_class
>= CIK
? 512 : 256;
611 struct ac_shader_config
*conf
;
612 unsigned max_simd_waves
;
613 unsigned lds_per_wave
= 0;
615 max_simd_waves
= ac_get_max_simd_waves(device
->physical_device
->rad_info
.family
);
617 conf
= &variant
->config
;
619 if (stage
== MESA_SHADER_FRAGMENT
) {
620 lds_per_wave
= conf
->lds_size
* lds_increment
+
621 align(variant
->info
.fs
.num_interp
* 48,
628 radv_get_num_physical_sgprs(device
->physical_device
) / conf
->num_sgprs
);
633 RADV_NUM_PHYSICAL_VGPRS
/ conf
->num_vgprs
);
635 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
639 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
641 if (stage
== MESA_SHADER_FRAGMENT
) {
642 _mesa_string_buffer_printf(buf
, "*** SHADER CONFIG ***\n"
643 "SPI_PS_INPUT_ADDR = 0x%04x\n"
644 "SPI_PS_INPUT_ENA = 0x%04x\n",
645 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
648 _mesa_string_buffer_printf(buf
, "*** SHADER STATS ***\n"
651 "Spilled SGPRs: %d\n"
652 "Spilled VGPRs: %d\n"
653 "PrivMem VGPRS: %d\n"
654 "Code Size: %d bytes\n"
656 "Scratch: %d bytes per wave\n"
658 "********************\n\n\n",
659 conf
->num_sgprs
, conf
->num_vgprs
,
660 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
661 variant
->info
.private_mem_vgprs
, variant
->code_size
,
662 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
667 radv_shader_dump_stats(struct radv_device
*device
,
668 struct radv_shader_variant
*variant
,
669 gl_shader_stage stage
,
672 struct _mesa_string_buffer
*buf
= _mesa_string_buffer_create(NULL
, 256);
674 generate_shader_stats(device
, variant
, stage
, buf
);
676 fprintf(file
, "\n%s:\n", radv_get_shader_name(variant
, stage
));
677 fprintf(file
, "%s", buf
->buf
);
679 _mesa_string_buffer_destroy(buf
);
683 radv_GetShaderInfoAMD(VkDevice _device
,
684 VkPipeline _pipeline
,
685 VkShaderStageFlagBits shaderStage
,
686 VkShaderInfoTypeAMD infoType
,
690 RADV_FROM_HANDLE(radv_device
, device
, _device
);
691 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
692 gl_shader_stage stage
= vk_to_mesa_shader_stage(shaderStage
);
693 struct radv_shader_variant
*variant
= pipeline
->shaders
[stage
];
694 struct _mesa_string_buffer
*buf
;
695 VkResult result
= VK_SUCCESS
;
697 /* Spec doesn't indicate what to do if the stage is invalid, so just
698 * return no info for this. */
700 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
703 case VK_SHADER_INFO_TYPE_STATISTICS_AMD
:
705 *pInfoSize
= sizeof(VkShaderStatisticsInfoAMD
);
707 unsigned lds_multiplier
= device
->physical_device
->rad_info
.chip_class
>= CIK
? 512 : 256;
708 struct ac_shader_config
*conf
= &variant
->config
;
710 VkShaderStatisticsInfoAMD statistics
= {};
711 statistics
.shaderStageMask
= shaderStage
;
712 statistics
.numPhysicalVgprs
= RADV_NUM_PHYSICAL_VGPRS
;
713 statistics
.numPhysicalSgprs
= radv_get_num_physical_sgprs(device
->physical_device
);
714 statistics
.numAvailableSgprs
= statistics
.numPhysicalSgprs
;
716 if (stage
== MESA_SHADER_COMPUTE
) {
717 unsigned *local_size
= variant
->nir
->info
.cs
.local_size
;
718 unsigned workgroup_size
= local_size
[0] * local_size
[1] * local_size
[2];
720 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
/
721 ceil(workgroup_size
/ statistics
.numPhysicalVgprs
);
723 statistics
.computeWorkGroupSize
[0] = local_size
[0];
724 statistics
.computeWorkGroupSize
[1] = local_size
[1];
725 statistics
.computeWorkGroupSize
[2] = local_size
[2];
727 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
;
730 statistics
.resourceUsage
.numUsedVgprs
= conf
->num_vgprs
;
731 statistics
.resourceUsage
.numUsedSgprs
= conf
->num_sgprs
;
732 statistics
.resourceUsage
.ldsSizePerLocalWorkGroup
= 32768;
733 statistics
.resourceUsage
.ldsUsageSizeInBytes
= conf
->lds_size
* lds_multiplier
;
734 statistics
.resourceUsage
.scratchMemUsageInBytes
= conf
->scratch_bytes_per_wave
;
736 size_t size
= *pInfoSize
;
737 *pInfoSize
= sizeof(statistics
);
739 memcpy(pInfo
, &statistics
, MIN2(size
, *pInfoSize
));
741 if (size
< *pInfoSize
)
742 result
= VK_INCOMPLETE
;
746 case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD
:
747 buf
= _mesa_string_buffer_create(NULL
, 1024);
749 _mesa_string_buffer_printf(buf
, "%s:\n", radv_get_shader_name(variant
, stage
));
750 _mesa_string_buffer_printf(buf
, "%s\n\n", variant
->disasm_string
);
751 generate_shader_stats(device
, variant
, stage
, buf
);
753 /* Need to include the null terminator. */
754 size_t length
= buf
->length
+ 1;
759 size_t size
= *pInfoSize
;
762 memcpy(pInfo
, buf
->buf
, MIN2(size
, length
));
765 result
= VK_INCOMPLETE
;
768 _mesa_string_buffer_destroy(buf
);
771 /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
772 result
= VK_ERROR_FEATURE_NOT_PRESENT
;