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,
59 .lower_pack_snorm_2x16
= true,
60 .lower_pack_snorm_4x8
= true,
61 .lower_pack_unorm_2x16
= true,
62 .lower_pack_unorm_4x8
= true,
63 .lower_unpack_snorm_2x16
= true,
64 .lower_unpack_snorm_4x8
= true,
65 .lower_unpack_unorm_2x16
= true,
66 .lower_unpack_unorm_4x8
= true,
67 .lower_extract_byte
= true,
68 .lower_extract_word
= true,
70 .vs_inputs_dual_locations
= true,
71 .max_unroll_iterations
= 32
74 VkResult
radv_CreateShaderModule(
76 const VkShaderModuleCreateInfo
* pCreateInfo
,
77 const VkAllocationCallbacks
* pAllocator
,
78 VkShaderModule
* pShaderModule
)
80 RADV_FROM_HANDLE(radv_device
, device
, _device
);
81 struct radv_shader_module
*module
;
83 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
84 assert(pCreateInfo
->flags
== 0);
86 module
= vk_alloc2(&device
->alloc
, pAllocator
,
87 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
88 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
90 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
93 module
->size
= pCreateInfo
->codeSize
;
94 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
96 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
98 *pShaderModule
= radv_shader_module_to_handle(module
);
103 void radv_DestroyShaderModule(
105 VkShaderModule _module
,
106 const VkAllocationCallbacks
* pAllocator
)
108 RADV_FROM_HANDLE(radv_device
, device
, _device
);
109 RADV_FROM_HANDLE(radv_shader_module
, module
, _module
);
114 vk_free2(&device
->alloc
, pAllocator
, module
);
118 radv_optimize_nir(struct nir_shader
*shader
)
125 NIR_PASS_V(shader
, nir_lower_vars_to_ssa
);
126 NIR_PASS_V(shader
, nir_lower_64bit_pack
);
127 NIR_PASS_V(shader
, nir_lower_alu_to_scalar
);
128 NIR_PASS_V(shader
, nir_lower_phis_to_scalar
);
130 NIR_PASS(progress
, shader
, nir_copy_prop
);
131 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
132 NIR_PASS(progress
, shader
, nir_opt_dce
);
133 if (nir_opt_trivial_continues(shader
)) {
135 NIR_PASS(progress
, shader
, nir_copy_prop
);
136 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
137 NIR_PASS(progress
, shader
, nir_opt_dce
);
139 NIR_PASS(progress
, shader
, nir_opt_if
);
140 NIR_PASS(progress
, shader
, nir_opt_dead_cf
);
141 NIR_PASS(progress
, shader
, nir_opt_cse
);
142 NIR_PASS(progress
, shader
, nir_opt_peephole_select
, 8);
143 NIR_PASS(progress
, shader
, nir_opt_algebraic
);
144 NIR_PASS(progress
, shader
, nir_opt_constant_folding
);
145 NIR_PASS(progress
, shader
, nir_opt_undef
);
146 NIR_PASS(progress
, shader
, nir_opt_conditional_discard
);
147 if (shader
->options
->max_unroll_iterations
) {
148 NIR_PASS(progress
, shader
, nir_opt_loop_unroll
, 0);
154 radv_shader_compile_to_nir(struct radv_device
*device
,
155 struct radv_shader_module
*module
,
156 const char *entrypoint_name
,
157 gl_shader_stage stage
,
158 const VkSpecializationInfo
*spec_info
)
160 if (strcmp(entrypoint_name
, "main") != 0) {
161 radv_finishme("Multiple shaders per module not really supported");
165 nir_function
*entry_point
;
167 /* Some things such as our meta clear/blit code will give us a NIR
168 * shader directly. In that case, we just ignore the SPIR-V entirely
169 * and just use the NIR shader */
171 nir
->options
= &nir_options
;
172 nir_validate_shader(nir
);
174 assert(exec_list_length(&nir
->functions
) == 1);
175 struct exec_node
*node
= exec_list_get_head(&nir
->functions
);
176 entry_point
= exec_node_data(nir_function
, node
, node
);
178 uint32_t *spirv
= (uint32_t *) module
->data
;
179 assert(module
->size
% 4 == 0);
181 if (device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SPIRV
)
182 radv_print_spirv(spirv
, module
->size
, stderr
);
184 uint32_t num_spec_entries
= 0;
185 struct nir_spirv_specialization
*spec_entries
= NULL
;
186 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
187 num_spec_entries
= spec_info
->mapEntryCount
;
188 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
189 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
190 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
191 const void *data
= spec_info
->pData
+ entry
.offset
;
192 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
194 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
195 if (spec_info
->dataSize
== 8)
196 spec_entries
[i
].data64
= *(const uint64_t *)data
;
198 spec_entries
[i
].data32
= *(const uint32_t *)data
;
201 const struct spirv_to_nir_options spirv_options
= {
203 .draw_parameters
= true,
205 .image_read_without_format
= true,
206 .image_write_without_format
= true,
207 .tessellation
= true,
210 .variable_pointers
= true,
213 entry_point
= spirv_to_nir(spirv
, module
->size
/ 4,
214 spec_entries
, num_spec_entries
,
215 stage
, entrypoint_name
,
216 &spirv_options
, &nir_options
);
217 nir
= entry_point
->shader
;
218 assert(nir
->info
.stage
== stage
);
219 nir_validate_shader(nir
);
223 /* We have to lower away local constant initializers right before we
224 * inline functions. That way they get properly initialized at the top
225 * of the function and not at the top of its caller.
227 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_local
);
228 NIR_PASS_V(nir
, nir_lower_returns
);
229 NIR_PASS_V(nir
, nir_inline_functions
);
231 /* Pick off the single entrypoint that we want */
232 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
233 if (func
!= entry_point
)
234 exec_node_remove(&func
->node
);
236 assert(exec_list_length(&nir
->functions
) == 1);
237 entry_point
->name
= ralloc_strdup(entry_point
, "main");
239 NIR_PASS_V(nir
, nir_remove_dead_variables
,
240 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
242 /* Now that we've deleted all but the main function, we can go ahead and
243 * lower the rest of the constant initializers.
245 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
246 NIR_PASS_V(nir
, nir_lower_system_values
);
247 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
250 /* Vulkan uses the separate-shader linking model */
251 nir
->info
.separate_shader
= true;
253 nir_shader_gather_info(nir
, entry_point
->impl
);
255 /* While it would be nice not to have this flag, we are constrained
256 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
259 bool llvm_has_working_vgpr_indexing
=
260 device
->physical_device
->rad_info
.chip_class
<= VI
;
262 /* TODO: Indirect indexing of GS inputs is unimplemented.
264 * TCS and TES load inputs directly from LDS or offchip memory, so
265 * indirect indexing is trivial.
267 nir_variable_mode indirect_mask
= 0;
268 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
269 (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
&&
270 nir
->info
.stage
!= MESA_SHADER_TESS_EVAL
&&
271 !llvm_has_working_vgpr_indexing
)) {
272 indirect_mask
|= nir_var_shader_in
;
274 if (!llvm_has_working_vgpr_indexing
&&
275 nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
276 indirect_mask
|= nir_var_shader_out
;
278 /* TODO: We shouldn't need to do this, however LLVM isn't currently
279 * smart enough to handle indirects without causing excess spilling
280 * causing the gpu to hang.
282 * See the following thread for more details of the problem:
283 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
285 indirect_mask
|= nir_var_local
;
287 nir_lower_indirect_derefs(nir
, indirect_mask
);
289 static const nir_lower_tex_options tex_options
= {
293 nir_lower_tex(nir
, &tex_options
);
295 nir_lower_vars_to_ssa(nir
);
296 nir_lower_var_copies(nir
);
297 nir_lower_global_vars_to_local(nir
);
298 nir_remove_dead_variables(nir
, nir_var_local
);
299 radv_optimize_nir(nir
);
305 radv_alloc_shader_memory(struct radv_device
*device
,
306 struct radv_shader_variant
*shader
)
308 mtx_lock(&device
->shader_slab_mutex
);
309 list_for_each_entry(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
311 list_for_each_entry(struct radv_shader_variant
, s
, &slab
->shaders
, slab_list
) {
312 if (s
->bo_offset
- offset
>= shader
->code_size
) {
313 shader
->bo
= slab
->bo
;
314 shader
->bo_offset
= offset
;
315 list_addtail(&shader
->slab_list
, &s
->slab_list
);
316 mtx_unlock(&device
->shader_slab_mutex
);
317 return slab
->ptr
+ offset
;
319 offset
= align_u64(s
->bo_offset
+ s
->code_size
, 256);
321 if (slab
->size
- offset
>= shader
->code_size
) {
322 shader
->bo
= slab
->bo
;
323 shader
->bo_offset
= offset
;
324 list_addtail(&shader
->slab_list
, &slab
->shaders
);
325 mtx_unlock(&device
->shader_slab_mutex
);
326 return slab
->ptr
+ offset
;
330 mtx_unlock(&device
->shader_slab_mutex
);
331 struct radv_shader_slab
*slab
= calloc(1, sizeof(struct radv_shader_slab
));
333 slab
->size
= 256 * 1024;
334 slab
->bo
= device
->ws
->buffer_create(device
->ws
, slab
->size
, 256,
336 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
337 device
->physical_device
->cpdma_prefetch_writes_memory
?
338 0 : RADEON_FLAG_READ_ONLY
);
339 slab
->ptr
= (char*)device
->ws
->buffer_map(slab
->bo
);
340 list_inithead(&slab
->shaders
);
342 mtx_lock(&device
->shader_slab_mutex
);
343 list_add(&slab
->slabs
, &device
->shader_slabs
);
345 shader
->bo
= slab
->bo
;
346 shader
->bo_offset
= 0;
347 list_add(&shader
->slab_list
, &slab
->shaders
);
348 mtx_unlock(&device
->shader_slab_mutex
);
353 radv_destroy_shader_slabs(struct radv_device
*device
)
355 list_for_each_entry_safe(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
356 device
->ws
->buffer_destroy(slab
->bo
);
359 mtx_destroy(&device
->shader_slab_mutex
);
363 radv_fill_shader_variant(struct radv_device
*device
,
364 struct radv_shader_variant
*variant
,
365 struct ac_shader_binary
*binary
,
366 gl_shader_stage stage
)
368 bool scratch_enabled
= variant
->config
.scratch_bytes_per_wave
> 0;
369 unsigned vgpr_comp_cnt
= 0;
371 if (scratch_enabled
&& !device
->llvm_supports_spill
)
372 radv_finishme("shader scratch support only available with LLVM 4.0");
374 variant
->code_size
= binary
->code_size
;
375 variant
->rsrc2
= S_00B12C_USER_SGPR(variant
->info
.num_user_sgprs
) |
376 S_00B12C_SCRATCH_EN(scratch_enabled
);
378 variant
->rsrc1
= S_00B848_VGPRS((variant
->config
.num_vgprs
- 1) / 4) |
379 S_00B848_SGPRS((variant
->config
.num_sgprs
- 1) / 8) |
380 S_00B848_DX10_CLAMP(1) |
381 S_00B848_FLOAT_MODE(variant
->config
.float_mode
);
384 case MESA_SHADER_TESS_EVAL
:
386 variant
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
388 case MESA_SHADER_TESS_CTRL
:
389 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
)
390 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
392 variant
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
394 case MESA_SHADER_VERTEX
:
395 case MESA_SHADER_GEOMETRY
:
396 vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
398 case MESA_SHADER_FRAGMENT
:
400 case MESA_SHADER_COMPUTE
: {
401 struct ac_shader_info
*info
= &variant
->info
.info
;
403 S_00B84C_TGID_X_EN(info
->cs
.uses_block_id
[0]) |
404 S_00B84C_TGID_Y_EN(info
->cs
.uses_block_id
[1]) |
405 S_00B84C_TGID_Z_EN(info
->cs
.uses_block_id
[2]) |
406 S_00B84C_TIDIG_COMP_CNT(info
->cs
.uses_thread_id
[2] ? 2 :
407 info
->cs
.uses_thread_id
[1] ? 1 : 0) |
408 S_00B84C_TG_SIZE_EN(info
->cs
.uses_local_invocation_idx
) |
409 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 struct ac_shader_info
*info
= &variant
->info
.info
;
420 unsigned es_type
= variant
->info
.gs
.es_type
;
421 unsigned gs_vgpr_comp_cnt
, es_vgpr_comp_cnt
;
423 if (es_type
== MESA_SHADER_VERTEX
) {
424 es_vgpr_comp_cnt
= variant
->info
.vs
.vgpr_comp_cnt
;
425 } else if (es_type
== MESA_SHADER_TESS_EVAL
) {
426 es_vgpr_comp_cnt
= 3;
428 unreachable("invalid shader ES type");
431 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
432 * VGPR[0:4] are always loaded.
434 if (info
->uses_invocation_id
)
435 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
436 else if (info
->uses_prim_id
)
437 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
438 else if (variant
->info
.gs
.vertices_in
>= 3)
439 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
441 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
);
452 void *ptr
= radv_alloc_shader_memory(device
, variant
);
453 memcpy(ptr
, binary
->code
, binary
->code_size
);
456 static struct radv_shader_variant
*
457 shader_variant_create(struct radv_device
*device
,
458 struct radv_shader_module
*module
,
459 struct nir_shader
* const *shaders
,
461 gl_shader_stage stage
,
462 struct ac_nir_compiler_options
*options
,
465 unsigned *code_size_out
)
467 enum radeon_family chip_family
= device
->physical_device
->rad_info
.family
;
468 bool dump_shaders
= radv_can_dump_shader(device
, module
);
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_preoptir
= radv_can_dump_shader(device
, module
) &&
481 device
->instance
->debug_flags
& RADV_DEBUG_PREOPTIR
;
483 if (options
->supports_spill
)
484 tm_options
|= AC_TM_SUPPORTS_SPILL
;
485 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
486 tm_options
|= AC_TM_SISCHED
;
487 tm
= ac_create_target_machine(chip_family
, tm_options
);
489 if (gs_copy_shader
) {
490 assert(shader_count
== 1);
491 ac_create_gs_copy_shader(tm
, *shaders
, &binary
, &variant
->config
,
492 &variant
->info
, options
, dump_shaders
);
494 ac_compile_nir_shader(tm
, &binary
, &variant
->config
,
495 &variant
->info
, shaders
, shader_count
, options
,
499 LLVMDisposeTargetMachine(tm
);
501 radv_fill_shader_variant(device
, variant
, &binary
, stage
);
504 *code_out
= binary
.code
;
505 *code_size_out
= binary
.code_size
;
510 free(binary
.global_symbol_offsets
);
512 variant
->ref_count
= 1;
514 if (device
->keep_shader_info
) {
515 variant
->disasm_string
= binary
.disasm_string
;
516 if (!gs_copy_shader
&& !module
->nir
) {
517 variant
->nir
= *shaders
;
518 variant
->spirv
= (uint32_t *)module
->data
;
519 variant
->spirv_size
= module
->size
;
522 free(binary
.disasm_string
);
528 struct radv_shader_variant
*
529 radv_shader_variant_create(struct radv_device
*device
,
530 struct radv_shader_module
*module
,
531 struct nir_shader
*const *shaders
,
533 struct radv_pipeline_layout
*layout
,
534 const struct ac_shader_variant_key
*key
,
536 unsigned *code_size_out
)
538 struct ac_nir_compiler_options options
= {0};
540 options
.layout
= layout
;
544 options
.unsafe_math
= !!(device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
);
545 options
.supports_spill
= device
->llvm_supports_spill
;
547 return shader_variant_create(device
, module
, shaders
, shader_count
, shaders
[shader_count
- 1]->info
.stage
,
548 &options
, false, code_out
, code_size_out
);
551 struct radv_shader_variant
*
552 radv_create_gs_copy_shader(struct radv_device
*device
,
553 struct nir_shader
*shader
,
555 unsigned *code_size_out
,
558 struct ac_nir_compiler_options options
= {0};
560 options
.key
.has_multiview_view_index
= multiview
;
562 return shader_variant_create(device
, NULL
, &shader
, 1, MESA_SHADER_VERTEX
,
563 &options
, true, code_out
, code_size_out
);
567 radv_shader_variant_destroy(struct radv_device
*device
,
568 struct radv_shader_variant
*variant
)
570 if (!p_atomic_dec_zero(&variant
->ref_count
))
573 mtx_lock(&device
->shader_slab_mutex
);
574 list_del(&variant
->slab_list
);
575 mtx_unlock(&device
->shader_slab_mutex
);
577 ralloc_free(variant
->nir
);
578 free(variant
->disasm_string
);
583 radv_get_shader_name(struct radv_shader_variant
*var
, gl_shader_stage stage
)
586 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";
587 case MESA_SHADER_GEOMETRY
: return "Geometry Shader";
588 case MESA_SHADER_FRAGMENT
: return "Pixel Shader";
589 case MESA_SHADER_COMPUTE
: return "Compute Shader";
590 case MESA_SHADER_TESS_CTRL
: return "Tessellation Control Shader";
591 case MESA_SHADER_TESS_EVAL
: return var
->info
.tes
.as_es
? "Tessellation Evaluation Shader as ES" : "Tessellation Evaluation Shader as VS";
593 return "Unknown shader";
598 get_total_sgprs(struct radv_device
*device
)
600 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
607 generate_shader_stats(struct radv_device
*device
,
608 struct radv_shader_variant
*variant
,
609 gl_shader_stage stage
,
610 struct _mesa_string_buffer
*buf
)
612 unsigned lds_increment
= device
->physical_device
->rad_info
.chip_class
>= CIK
? 512 : 256;
613 struct ac_shader_config
*conf
;
614 unsigned max_simd_waves
;
615 unsigned lds_per_wave
= 0;
617 switch (device
->physical_device
->rad_info
.family
) {
618 /* These always have 8 waves: */
628 conf
= &variant
->config
;
630 if (stage
== MESA_SHADER_FRAGMENT
) {
631 lds_per_wave
= conf
->lds_size
* lds_increment
+
632 align(variant
->info
.fs
.num_interp
* 48,
637 max_simd_waves
= MIN2(max_simd_waves
, get_total_sgprs(device
) / conf
->num_sgprs
);
640 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
642 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
646 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
648 if (stage
== MESA_SHADER_FRAGMENT
) {
649 _mesa_string_buffer_printf(buf
, "*** SHADER CONFIG ***\n"
650 "SPI_PS_INPUT_ADDR = 0x%04x\n"
651 "SPI_PS_INPUT_ENA = 0x%04x\n",
652 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
655 _mesa_string_buffer_printf(buf
, "*** SHADER STATS ***\n"
658 "Spilled SGPRs: %d\n"
659 "Spilled VGPRs: %d\n"
660 "Code Size: %d bytes\n"
662 "Scratch: %d bytes per wave\n"
664 "********************\n\n\n",
665 conf
->num_sgprs
, conf
->num_vgprs
,
666 conf
->spilled_sgprs
, conf
->spilled_vgprs
, variant
->code_size
,
667 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
672 radv_shader_dump_stats(struct radv_device
*device
,
673 struct radv_shader_variant
*variant
,
674 gl_shader_stage stage
,
677 struct _mesa_string_buffer
*buf
= _mesa_string_buffer_create(NULL
, 256);
679 generate_shader_stats(device
, variant
, stage
, buf
);
681 fprintf(file
, "\n%s:\n", radv_get_shader_name(variant
, stage
));
682 fprintf(file
, "%s", buf
->buf
);
684 _mesa_string_buffer_destroy(buf
);
688 radv_GetShaderInfoAMD(VkDevice _device
,
689 VkPipeline _pipeline
,
690 VkShaderStageFlagBits shaderStage
,
691 VkShaderInfoTypeAMD infoType
,
695 RADV_FROM_HANDLE(radv_device
, device
, _device
);
696 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
697 gl_shader_stage stage
= vk_to_mesa_shader_stage(shaderStage
);
698 struct radv_shader_variant
*variant
= pipeline
->shaders
[stage
];
699 struct _mesa_string_buffer
*buf
;
700 VkResult result
= VK_SUCCESS
;
702 /* Spec doesn't indicate what to do if the stage is invalid, so just
703 * return no info for this. */
705 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
708 case VK_SHADER_INFO_TYPE_STATISTICS_AMD
:
710 *pInfoSize
= sizeof(VkShaderStatisticsInfoAMD
);
712 unsigned lds_multiplier
= device
->physical_device
->rad_info
.chip_class
>= CIK
? 512 : 256;
713 struct ac_shader_config
*conf
= &variant
->config
;
715 VkShaderStatisticsInfoAMD statistics
= {};
716 statistics
.shaderStageMask
= shaderStage
;
717 statistics
.numPhysicalVgprs
= 256;
718 statistics
.numPhysicalSgprs
= get_total_sgprs(device
);
719 statistics
.numAvailableSgprs
= statistics
.numPhysicalSgprs
;
721 if (stage
== MESA_SHADER_COMPUTE
) {
722 unsigned *local_size
= variant
->nir
->info
.cs
.local_size
;
723 unsigned workgroup_size
= local_size
[0] * local_size
[1] * local_size
[2];
725 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
/
726 ceil(workgroup_size
/ statistics
.numPhysicalVgprs
);
728 statistics
.computeWorkGroupSize
[0] = local_size
[0];
729 statistics
.computeWorkGroupSize
[1] = local_size
[1];
730 statistics
.computeWorkGroupSize
[2] = local_size
[2];
732 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
;
735 statistics
.resourceUsage
.numUsedVgprs
= conf
->num_vgprs
;
736 statistics
.resourceUsage
.numUsedSgprs
= conf
->num_sgprs
;
737 statistics
.resourceUsage
.ldsSizePerLocalWorkGroup
= 32768;
738 statistics
.resourceUsage
.ldsUsageSizeInBytes
= conf
->lds_size
* lds_multiplier
;
739 statistics
.resourceUsage
.scratchMemUsageInBytes
= conf
->scratch_bytes_per_wave
;
741 size_t size
= *pInfoSize
;
742 *pInfoSize
= sizeof(statistics
);
744 memcpy(pInfo
, &statistics
, MIN2(size
, *pInfoSize
));
746 if (size
< *pInfoSize
)
747 result
= VK_INCOMPLETE
;
751 case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD
:
752 buf
= _mesa_string_buffer_create(NULL
, 1024);
754 _mesa_string_buffer_printf(buf
, "%s:\n", radv_get_shader_name(variant
, stage
));
755 _mesa_string_buffer_printf(buf
, "%s\n\n", variant
->disasm_string
);
756 generate_shader_stats(device
, variant
, stage
, buf
);
758 /* Need to include the null terminator. */
759 size_t length
= buf
->length
+ 1;
764 size_t size
= *pInfoSize
;
767 memcpy(pInfo
, buf
->buf
, MIN2(size
, length
));
770 result
= VK_INCOMPLETE
;
773 _mesa_string_buffer_destroy(buf
);
776 /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
777 result
= VK_ERROR_FEATURE_NOT_PRESENT
;