2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include "util/mesa-sha1.h"
31 #include "util/os_time.h"
32 #include "common/gen_l3_config.h"
33 #include "common/gen_disasm.h"
34 #include "anv_private.h"
35 #include "compiler/brw_nir.h"
37 #include "nir/nir_xfb_info.h"
38 #include "spirv/nir_spirv.h"
41 /* Needed for SWIZZLE macros */
42 #include "program/prog_instruction.h"
46 VkResult
anv_CreateShaderModule(
48 const VkShaderModuleCreateInfo
* pCreateInfo
,
49 const VkAllocationCallbacks
* pAllocator
,
50 VkShaderModule
* pShaderModule
)
52 ANV_FROM_HANDLE(anv_device
, device
, _device
);
53 struct anv_shader_module
*module
;
55 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
56 assert(pCreateInfo
->flags
== 0);
58 module
= vk_alloc2(&device
->alloc
, pAllocator
,
59 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
60 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
62 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
64 module
->size
= pCreateInfo
->codeSize
;
65 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
67 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
69 *pShaderModule
= anv_shader_module_to_handle(module
);
74 void anv_DestroyShaderModule(
76 VkShaderModule _module
,
77 const VkAllocationCallbacks
* pAllocator
)
79 ANV_FROM_HANDLE(anv_device
, device
, _device
);
80 ANV_FROM_HANDLE(anv_shader_module
, module
, _module
);
85 vk_free2(&device
->alloc
, pAllocator
, module
);
88 #define SPIR_V_MAGIC_NUMBER 0x07230203
90 static const uint64_t stage_to_debug
[] = {
91 [MESA_SHADER_VERTEX
] = DEBUG_VS
,
92 [MESA_SHADER_TESS_CTRL
] = DEBUG_TCS
,
93 [MESA_SHADER_TESS_EVAL
] = DEBUG_TES
,
94 [MESA_SHADER_GEOMETRY
] = DEBUG_GS
,
95 [MESA_SHADER_FRAGMENT
] = DEBUG_WM
,
96 [MESA_SHADER_COMPUTE
] = DEBUG_CS
,
99 struct anv_spirv_debug_data
{
100 struct anv_device
*device
;
101 const struct anv_shader_module
*module
;
104 static void anv_spirv_nir_debug(void *private_data
,
105 enum nir_spirv_debug_level level
,
109 struct anv_spirv_debug_data
*debug_data
= private_data
;
110 static const VkDebugReportFlagsEXT vk_flags
[] = {
111 [NIR_SPIRV_DEBUG_LEVEL_INFO
] = VK_DEBUG_REPORT_INFORMATION_BIT_EXT
,
112 [NIR_SPIRV_DEBUG_LEVEL_WARNING
] = VK_DEBUG_REPORT_WARNING_BIT_EXT
,
113 [NIR_SPIRV_DEBUG_LEVEL_ERROR
] = VK_DEBUG_REPORT_ERROR_BIT_EXT
,
117 snprintf(buffer
, sizeof(buffer
), "SPIR-V offset %lu: %s", (unsigned long) spirv_offset
, message
);
119 vk_debug_report(&debug_data
->device
->instance
->debug_report_callbacks
,
121 VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT
,
122 (uint64_t) (uintptr_t) debug_data
->module
,
123 0, 0, "anv", buffer
);
126 /* Eventually, this will become part of anv_CreateShader. Unfortunately,
127 * we can't do that yet because we don't have the ability to copy nir.
130 anv_shader_compile_to_nir(struct anv_device
*device
,
132 const struct anv_shader_module
*module
,
133 const char *entrypoint_name
,
134 gl_shader_stage stage
,
135 const VkSpecializationInfo
*spec_info
)
137 const struct anv_physical_device
*pdevice
=
138 &device
->instance
->physicalDevice
;
139 const struct brw_compiler
*compiler
= pdevice
->compiler
;
140 const nir_shader_compiler_options
*nir_options
=
141 compiler
->glsl_compiler_options
[stage
].NirOptions
;
143 uint32_t *spirv
= (uint32_t *) module
->data
;
144 assert(spirv
[0] == SPIR_V_MAGIC_NUMBER
);
145 assert(module
->size
% 4 == 0);
147 uint32_t num_spec_entries
= 0;
148 struct nir_spirv_specialization
*spec_entries
= NULL
;
149 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
150 num_spec_entries
= spec_info
->mapEntryCount
;
151 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
152 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
153 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
154 const void *data
= spec_info
->pData
+ entry
.offset
;
155 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
157 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
158 if (spec_info
->dataSize
== 8)
159 spec_entries
[i
].data64
= *(const uint64_t *)data
;
161 spec_entries
[i
].data32
= *(const uint32_t *)data
;
165 struct anv_spirv_debug_data spirv_debug_data
= {
169 struct spirv_to_nir_options spirv_options
= {
170 .frag_coord_is_sysval
= true,
172 .demote_to_helper_invocation
= true,
173 .derivative_group
= true,
174 .descriptor_array_dynamic_indexing
= true,
175 .descriptor_array_non_uniform_indexing
= true,
176 .descriptor_indexing
= true,
177 .device_group
= true,
178 .draw_parameters
= true,
179 .float16
= pdevice
->info
.gen
>= 8,
180 .float64
= pdevice
->info
.gen
>= 8,
181 .fragment_shader_sample_interlock
= pdevice
->info
.gen
>= 9,
182 .fragment_shader_pixel_interlock
= pdevice
->info
.gen
>= 9,
183 .geometry_streams
= true,
184 .image_write_without_format
= true,
185 .int8
= pdevice
->info
.gen
>= 8,
186 .int16
= pdevice
->info
.gen
>= 8,
187 .int64
= pdevice
->info
.gen
>= 8,
188 .int64_atomics
= pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
,
191 .physical_storage_buffer_address
= pdevice
->has_a64_buffer_access
,
192 .post_depth_coverage
= pdevice
->info
.gen
>= 9,
193 .runtime_descriptor_array
= true,
194 .float_controls
= pdevice
->info
.gen
>= 8,
195 .shader_clock
= true,
196 .shader_viewport_index_layer
= true,
197 .stencil_export
= pdevice
->info
.gen
>= 9,
198 .storage_8bit
= pdevice
->info
.gen
>= 8,
199 .storage_16bit
= pdevice
->info
.gen
>= 8,
200 .subgroup_arithmetic
= true,
201 .subgroup_basic
= true,
202 .subgroup_ballot
= true,
203 .subgroup_quad
= true,
204 .subgroup_shuffle
= true,
205 .subgroup_vote
= true,
206 .tessellation
= true,
207 .transform_feedback
= pdevice
->info
.gen
>= 8,
208 .variable_pointers
= true,
210 .ubo_addr_format
= nir_address_format_32bit_index_offset
,
212 anv_nir_ssbo_addr_format(pdevice
, device
->robust_buffer_access
),
213 .phys_ssbo_addr_format
= nir_address_format_64bit_global
,
214 .push_const_addr_format
= nir_address_format_logical
,
216 /* TODO: Consider changing this to an address format that has the NULL
217 * pointer equals to 0. That might be a better format to play nice
218 * with certain code / code generators.
220 .shared_addr_format
= nir_address_format_32bit_offset
,
222 .func
= anv_spirv_nir_debug
,
223 .private_data
= &spirv_debug_data
,
229 spirv_to_nir(spirv
, module
->size
/ 4,
230 spec_entries
, num_spec_entries
,
231 stage
, entrypoint_name
, &spirv_options
, nir_options
);
232 assert(nir
->info
.stage
== stage
);
233 nir_validate_shader(nir
, "after spirv_to_nir");
234 ralloc_steal(mem_ctx
, nir
);
238 if (unlikely(INTEL_DEBUG
& stage_to_debug
[stage
])) {
239 fprintf(stderr
, "NIR (from SPIR-V) for %s shader:\n",
240 gl_shader_stage_name(stage
));
241 nir_print_shader(nir
, stderr
);
244 /* We have to lower away local constant initializers right before we
245 * inline functions. That way they get properly initialized at the top
246 * of the function and not at the top of its caller.
248 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
249 NIR_PASS_V(nir
, nir_lower_returns
);
250 NIR_PASS_V(nir
, nir_inline_functions
);
251 NIR_PASS_V(nir
, nir_opt_deref
);
253 /* Pick off the single entrypoint that we want */
254 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
255 if (!func
->is_entrypoint
)
256 exec_node_remove(&func
->node
);
258 assert(exec_list_length(&nir
->functions
) == 1);
260 /* Now that we've deleted all but the main function, we can go ahead and
261 * lower the rest of the constant initializers. We do this here so that
262 * nir_remove_dead_variables and split_per_member_structs below see the
263 * corresponding stores.
265 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
267 /* Split member structs. We do this before lower_io_to_temporaries so that
268 * it doesn't lower system values to temporaries by accident.
270 NIR_PASS_V(nir
, nir_split_var_copies
);
271 NIR_PASS_V(nir
, nir_split_per_member_structs
);
273 NIR_PASS_V(nir
, nir_remove_dead_variables
,
274 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
276 NIR_PASS_V(nir
, nir_propagate_invariant
);
277 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
278 nir_shader_get_entrypoint(nir
), true, false);
280 NIR_PASS_V(nir
, nir_lower_frexp
);
282 /* Vulkan uses the separate-shader linking model */
283 nir
->info
.separate_shader
= true;
285 brw_preprocess_nir(compiler
, nir
, NULL
);
290 void anv_DestroyPipeline(
292 VkPipeline _pipeline
,
293 const VkAllocationCallbacks
* pAllocator
)
295 ANV_FROM_HANDLE(anv_device
, device
, _device
);
296 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, _pipeline
);
301 anv_reloc_list_finish(&pipeline
->batch_relocs
,
302 pAllocator
? pAllocator
: &device
->alloc
);
304 ralloc_free(pipeline
->mem_ctx
);
306 if (pipeline
->blend_state
.map
)
307 anv_state_pool_free(&device
->dynamic_state_pool
, pipeline
->blend_state
);
309 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
310 if (pipeline
->shaders
[s
])
311 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
314 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
317 static const uint32_t vk_to_gen_primitive_type
[] = {
318 [VK_PRIMITIVE_TOPOLOGY_POINT_LIST
] = _3DPRIM_POINTLIST
,
319 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST
] = _3DPRIM_LINELIST
,
320 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
321 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
] = _3DPRIM_TRILIST
,
322 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
323 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
324 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
325 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
326 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
327 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
331 populate_sampler_prog_key(const struct gen_device_info
*devinfo
,
332 struct brw_sampler_prog_key_data
*key
)
334 /* Almost all multisampled textures are compressed. The only time when we
335 * don't compress a multisampled texture is for 16x MSAA with a surface
336 * width greater than 8k which is a bit of an edge case. Since the sampler
337 * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
338 * to tell the compiler to always assume compression.
340 key
->compressed_multisample_layout_mask
= ~0;
342 /* SkyLake added support for 16x MSAA. With this came a new message for
343 * reading from a 16x MSAA surface with compression. The new message was
344 * needed because now the MCS data is 64 bits instead of 32 or lower as is
345 * the case for 8x, 4x, and 2x. The key->msaa_16 bit-field controls which
346 * message we use. Fortunately, the 16x message works for 8x, 4x, and 2x
347 * so we can just use it unconditionally. This may not be quite as
348 * efficient but it saves us from recompiling.
350 if (devinfo
->gen
>= 9)
353 /* XXX: Handle texture swizzle on HSW- */
354 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
355 /* Assume color sampler, no swizzling. (Works for BDW+) */
356 key
->swizzles
[i
] = SWIZZLE_XYZW
;
361 populate_base_prog_key(const struct gen_device_info
*devinfo
,
362 VkPipelineShaderStageCreateFlags flags
,
363 struct brw_base_prog_key
*key
)
365 if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT
)
366 key
->subgroup_size_type
= BRW_SUBGROUP_SIZE_VARYING
;
368 key
->subgroup_size_type
= BRW_SUBGROUP_SIZE_API_CONSTANT
;
370 populate_sampler_prog_key(devinfo
, &key
->tex
);
374 populate_vs_prog_key(const struct gen_device_info
*devinfo
,
375 VkPipelineShaderStageCreateFlags flags
,
376 struct brw_vs_prog_key
*key
)
378 memset(key
, 0, sizeof(*key
));
380 populate_base_prog_key(devinfo
, flags
, &key
->base
);
382 /* XXX: Handle vertex input work-arounds */
384 /* XXX: Handle sampler_prog_key */
388 populate_tcs_prog_key(const struct gen_device_info
*devinfo
,
389 VkPipelineShaderStageCreateFlags flags
,
390 unsigned input_vertices
,
391 struct brw_tcs_prog_key
*key
)
393 memset(key
, 0, sizeof(*key
));
395 populate_base_prog_key(devinfo
, flags
, &key
->base
);
397 key
->input_vertices
= input_vertices
;
401 populate_tes_prog_key(const struct gen_device_info
*devinfo
,
402 VkPipelineShaderStageCreateFlags flags
,
403 struct brw_tes_prog_key
*key
)
405 memset(key
, 0, sizeof(*key
));
407 populate_base_prog_key(devinfo
, flags
, &key
->base
);
411 populate_gs_prog_key(const struct gen_device_info
*devinfo
,
412 VkPipelineShaderStageCreateFlags flags
,
413 struct brw_gs_prog_key
*key
)
415 memset(key
, 0, sizeof(*key
));
417 populate_base_prog_key(devinfo
, flags
, &key
->base
);
421 populate_wm_prog_key(const struct gen_device_info
*devinfo
,
422 VkPipelineShaderStageCreateFlags flags
,
423 const struct anv_subpass
*subpass
,
424 const VkPipelineMultisampleStateCreateInfo
*ms_info
,
425 struct brw_wm_prog_key
*key
)
427 memset(key
, 0, sizeof(*key
));
429 populate_base_prog_key(devinfo
, flags
, &key
->base
);
431 /* We set this to 0 here and set to the actual value before we call
434 key
->input_slots_valid
= 0;
436 /* Vulkan doesn't specify a default */
437 key
->high_quality_derivatives
= false;
439 /* XXX Vulkan doesn't appear to specify */
440 key
->clamp_fragment_color
= false;
442 assert(subpass
->color_count
<= MAX_RTS
);
443 for (uint32_t i
= 0; i
< subpass
->color_count
; i
++) {
444 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
445 key
->color_outputs_valid
|= (1 << i
);
448 key
->nr_color_regions
= util_bitcount(key
->color_outputs_valid
);
450 /* To reduce possible shader recompilations we would need to know if
451 * there is a SampleMask output variable to compute if we should emit
452 * code to workaround the issue that hardware disables alpha to coverage
453 * when there is SampleMask output.
455 key
->alpha_to_coverage
= ms_info
&& ms_info
->alphaToCoverageEnable
;
457 /* Vulkan doesn't support fixed-function alpha test */
458 key
->alpha_test_replicate_alpha
= false;
461 /* We should probably pull this out of the shader, but it's fairly
462 * harmless to compute it and then let dead-code take care of it.
464 if (ms_info
->rasterizationSamples
> 1) {
465 key
->persample_interp
= ms_info
->sampleShadingEnable
&&
466 (ms_info
->minSampleShading
* ms_info
->rasterizationSamples
) > 1;
467 key
->multisample_fbo
= true;
470 key
->frag_coord_adds_sample_pos
= key
->persample_interp
;
475 populate_cs_prog_key(const struct gen_device_info
*devinfo
,
476 VkPipelineShaderStageCreateFlags flags
,
477 const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT
*rss_info
,
478 struct brw_cs_prog_key
*key
)
480 memset(key
, 0, sizeof(*key
));
482 populate_base_prog_key(devinfo
, flags
, &key
->base
);
485 assert(key
->base
.subgroup_size_type
!= BRW_SUBGROUP_SIZE_VARYING
);
487 /* These enum values are expressly chosen to be equal to the subgroup
488 * size that they require.
490 assert(rss_info
->requiredSubgroupSize
== 8 ||
491 rss_info
->requiredSubgroupSize
== 16 ||
492 rss_info
->requiredSubgroupSize
== 32);
493 key
->base
.subgroup_size_type
= rss_info
->requiredSubgroupSize
;
494 } else if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT_EXT
) {
495 /* If the client expressly requests full subgroups and they don't
496 * specify a subgroup size, we need to pick one. If they're requested
497 * varying subgroup sizes, we set it to UNIFORM and let the back-end
498 * compiler pick. Otherwise, we specify the API value of 32.
499 * Performance will likely be terrible in this case but there's nothing
500 * we can do about that. The client should have chosen a size.
502 if (flags
& VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT_EXT
)
503 key
->base
.subgroup_size_type
= BRW_SUBGROUP_SIZE_UNIFORM
;
505 key
->base
.subgroup_size_type
= BRW_SUBGROUP_SIZE_REQUIRE_32
;
509 struct anv_pipeline_stage
{
510 gl_shader_stage stage
;
512 const struct anv_shader_module
*module
;
513 const char *entrypoint
;
514 const VkSpecializationInfo
*spec_info
;
516 unsigned char shader_sha1
[20];
518 union brw_any_prog_key key
;
521 gl_shader_stage stage
;
522 unsigned char sha1
[20];
527 struct anv_pipeline_binding surface_to_descriptor
[256];
528 struct anv_pipeline_binding sampler_to_descriptor
[256];
529 struct anv_pipeline_bind_map bind_map
;
531 union brw_any_prog_data prog_data
;
534 struct brw_compile_stats stats
[3];
537 VkPipelineCreationFeedbackEXT feedback
;
539 const unsigned *code
;
543 anv_pipeline_hash_shader(const struct anv_shader_module
*module
,
544 const char *entrypoint
,
545 gl_shader_stage stage
,
546 const VkSpecializationInfo
*spec_info
,
547 unsigned char *sha1_out
)
549 struct mesa_sha1 ctx
;
550 _mesa_sha1_init(&ctx
);
552 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
553 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
554 _mesa_sha1_update(&ctx
, &stage
, sizeof(stage
));
556 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
557 spec_info
->mapEntryCount
*
558 sizeof(*spec_info
->pMapEntries
));
559 _mesa_sha1_update(&ctx
, spec_info
->pData
,
560 spec_info
->dataSize
);
563 _mesa_sha1_final(&ctx
, sha1_out
);
567 anv_pipeline_hash_graphics(struct anv_pipeline
*pipeline
,
568 struct anv_pipeline_layout
*layout
,
569 struct anv_pipeline_stage
*stages
,
570 unsigned char *sha1_out
)
572 struct mesa_sha1 ctx
;
573 _mesa_sha1_init(&ctx
);
575 _mesa_sha1_update(&ctx
, &pipeline
->subpass
->view_mask
,
576 sizeof(pipeline
->subpass
->view_mask
));
579 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
581 const bool rba
= pipeline
->device
->robust_buffer_access
;
582 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
584 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
585 if (stages
[s
].entrypoint
) {
586 _mesa_sha1_update(&ctx
, stages
[s
].shader_sha1
,
587 sizeof(stages
[s
].shader_sha1
));
588 _mesa_sha1_update(&ctx
, &stages
[s
].key
, brw_prog_key_size(s
));
592 _mesa_sha1_final(&ctx
, sha1_out
);
596 anv_pipeline_hash_compute(struct anv_pipeline
*pipeline
,
597 struct anv_pipeline_layout
*layout
,
598 struct anv_pipeline_stage
*stage
,
599 unsigned char *sha1_out
)
601 struct mesa_sha1 ctx
;
602 _mesa_sha1_init(&ctx
);
605 _mesa_sha1_update(&ctx
, layout
->sha1
, sizeof(layout
->sha1
));
607 const bool rba
= pipeline
->device
->robust_buffer_access
;
608 _mesa_sha1_update(&ctx
, &rba
, sizeof(rba
));
610 _mesa_sha1_update(&ctx
, stage
->shader_sha1
,
611 sizeof(stage
->shader_sha1
));
612 _mesa_sha1_update(&ctx
, &stage
->key
.cs
, sizeof(stage
->key
.cs
));
614 _mesa_sha1_final(&ctx
, sha1_out
);
618 anv_pipeline_stage_get_nir(struct anv_pipeline
*pipeline
,
619 struct anv_pipeline_cache
*cache
,
621 struct anv_pipeline_stage
*stage
)
623 const struct brw_compiler
*compiler
=
624 pipeline
->device
->instance
->physicalDevice
.compiler
;
625 const nir_shader_compiler_options
*nir_options
=
626 compiler
->glsl_compiler_options
[stage
->stage
].NirOptions
;
629 nir
= anv_device_search_for_nir(pipeline
->device
, cache
,
634 assert(nir
->info
.stage
== stage
->stage
);
638 nir
= anv_shader_compile_to_nir(pipeline
->device
,
645 anv_device_upload_nir(pipeline
->device
, cache
, nir
, stage
->shader_sha1
);
653 anv_pipeline_lower_nir(struct anv_pipeline
*pipeline
,
655 struct anv_pipeline_stage
*stage
,
656 struct anv_pipeline_layout
*layout
)
658 const struct anv_physical_device
*pdevice
=
659 &pipeline
->device
->instance
->physicalDevice
;
660 const struct brw_compiler
*compiler
= pdevice
->compiler
;
662 struct brw_stage_prog_data
*prog_data
= &stage
->prog_data
.base
;
663 nir_shader
*nir
= stage
->nir
;
665 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
666 NIR_PASS_V(nir
, nir_lower_wpos_center
, pipeline
->sample_shading_enable
);
667 NIR_PASS_V(nir
, nir_lower_input_attachments
, true);
670 NIR_PASS_V(nir
, anv_nir_lower_ycbcr_textures
, layout
);
672 NIR_PASS_V(nir
, anv_nir_lower_push_constants
);
674 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
675 NIR_PASS_V(nir
, anv_nir_lower_multiview
, pipeline
->subpass
->view_mask
);
677 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
679 if (nir
->num_uniforms
> 0) {
680 assert(prog_data
->nr_params
== 0);
682 /* If the shader uses any push constants at all, we'll just give
683 * them the maximum possible number
685 assert(nir
->num_uniforms
<= MAX_PUSH_CONSTANTS_SIZE
);
686 nir
->num_uniforms
= MAX_PUSH_CONSTANTS_SIZE
;
687 prog_data
->nr_params
+= MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float);
688 prog_data
->param
= ralloc_array(mem_ctx
, uint32_t, prog_data
->nr_params
);
690 /* We now set the param values to be offsets into a
691 * anv_push_constant_data structure. Since the compiler doesn't
692 * actually dereference any of the gl_constant_value pointers in the
693 * params array, it doesn't really matter what we put here.
695 struct anv_push_constants
*null_data
= NULL
;
696 /* Fill out the push constants section of the param array */
697 for (unsigned i
= 0; i
< MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float); i
++) {
698 prog_data
->param
[i
] = ANV_PARAM_PUSH(
699 (uintptr_t)&null_data
->client_data
[i
* sizeof(float)]);
703 if (nir
->info
.num_ssbos
> 0 || nir
->info
.num_images
> 0)
704 pipeline
->needs_data_cache
= true;
706 NIR_PASS_V(nir
, brw_nir_lower_image_load_store
, compiler
->devinfo
);
708 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_global
,
709 nir_address_format_64bit_global
);
711 /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
713 anv_nir_apply_pipeline_layout(pdevice
,
714 pipeline
->device
->robust_buffer_access
,
715 layout
, nir
, prog_data
,
718 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_ubo
,
719 nir_address_format_32bit_index_offset
);
720 NIR_PASS_V(nir
, nir_lower_explicit_io
, nir_var_mem_ssbo
,
721 anv_nir_ssbo_addr_format(pdevice
,
722 pipeline
->device
->robust_buffer_access
));
724 NIR_PASS_V(nir
, nir_opt_constant_folding
);
726 /* We don't support non-uniform UBOs and non-uniform SSBO access is
727 * handled naturally by falling back to A64 messages.
729 NIR_PASS_V(nir
, nir_lower_non_uniform_access
,
730 nir_lower_non_uniform_texture_access
|
731 nir_lower_non_uniform_image_access
);
734 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
735 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
737 assert(nir
->num_uniforms
== prog_data
->nr_params
* 4);
743 anv_pipeline_link_vs(const struct brw_compiler
*compiler
,
744 struct anv_pipeline_stage
*vs_stage
,
745 struct anv_pipeline_stage
*next_stage
)
748 brw_nir_link_shaders(compiler
, vs_stage
->nir
, next_stage
->nir
);
752 anv_pipeline_compile_vs(const struct brw_compiler
*compiler
,
754 struct anv_device
*device
,
755 struct anv_pipeline_stage
*vs_stage
)
757 brw_compute_vue_map(compiler
->devinfo
,
758 &vs_stage
->prog_data
.vs
.base
.vue_map
,
759 vs_stage
->nir
->info
.outputs_written
,
760 vs_stage
->nir
->info
.separate_shader
);
762 vs_stage
->num_stats
= 1;
763 vs_stage
->code
= brw_compile_vs(compiler
, device
, mem_ctx
,
765 &vs_stage
->prog_data
.vs
,
767 vs_stage
->stats
, NULL
);
771 merge_tess_info(struct shader_info
*tes_info
,
772 const struct shader_info
*tcs_info
)
774 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
776 * "PointMode. Controls generation of points rather than triangles
777 * or lines. This functionality defaults to disabled, and is
778 * enabled if either shader stage includes the execution mode.
780 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
781 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
782 * and OutputVertices, it says:
784 * "One mode must be set in at least one of the tessellation
787 * So, the fields can be set in either the TCS or TES, but they must
788 * agree if set in both. Our backend looks at TES, so bitwise-or in
789 * the values from the TCS.
791 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
792 tes_info
->tess
.tcs_vertices_out
== 0 ||
793 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
794 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
796 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
797 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
798 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
799 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
801 assert(tcs_info
->tess
.primitive_mode
== 0 ||
802 tes_info
->tess
.primitive_mode
== 0 ||
803 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
804 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
805 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
806 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
810 anv_pipeline_link_tcs(const struct brw_compiler
*compiler
,
811 struct anv_pipeline_stage
*tcs_stage
,
812 struct anv_pipeline_stage
*tes_stage
)
814 assert(tes_stage
&& tes_stage
->stage
== MESA_SHADER_TESS_EVAL
);
816 brw_nir_link_shaders(compiler
, tcs_stage
->nir
, tes_stage
->nir
);
818 nir_lower_patch_vertices(tes_stage
->nir
,
819 tcs_stage
->nir
->info
.tess
.tcs_vertices_out
,
822 /* Copy TCS info into the TES info */
823 merge_tess_info(&tes_stage
->nir
->info
, &tcs_stage
->nir
->info
);
825 /* Whacking the key after cache lookup is a bit sketchy, but all of
826 * this comes from the SPIR-V, which is part of the hash used for the
827 * pipeline cache. So it should be safe.
829 tcs_stage
->key
.tcs
.tes_primitive_mode
=
830 tes_stage
->nir
->info
.tess
.primitive_mode
;
831 tcs_stage
->key
.tcs
.quads_workaround
=
832 compiler
->devinfo
->gen
< 9 &&
833 tes_stage
->nir
->info
.tess
.primitive_mode
== 7 /* GL_QUADS */ &&
834 tes_stage
->nir
->info
.tess
.spacing
== TESS_SPACING_EQUAL
;
838 anv_pipeline_compile_tcs(const struct brw_compiler
*compiler
,
840 struct anv_device
*device
,
841 struct anv_pipeline_stage
*tcs_stage
,
842 struct anv_pipeline_stage
*prev_stage
)
844 tcs_stage
->key
.tcs
.outputs_written
=
845 tcs_stage
->nir
->info
.outputs_written
;
846 tcs_stage
->key
.tcs
.patch_outputs_written
=
847 tcs_stage
->nir
->info
.patch_outputs_written
;
849 tcs_stage
->num_stats
= 1;
850 tcs_stage
->code
= brw_compile_tcs(compiler
, device
, mem_ctx
,
852 &tcs_stage
->prog_data
.tcs
,
854 tcs_stage
->stats
, NULL
);
858 anv_pipeline_link_tes(const struct brw_compiler
*compiler
,
859 struct anv_pipeline_stage
*tes_stage
,
860 struct anv_pipeline_stage
*next_stage
)
863 brw_nir_link_shaders(compiler
, tes_stage
->nir
, next_stage
->nir
);
867 anv_pipeline_compile_tes(const struct brw_compiler
*compiler
,
869 struct anv_device
*device
,
870 struct anv_pipeline_stage
*tes_stage
,
871 struct anv_pipeline_stage
*tcs_stage
)
873 tes_stage
->key
.tes
.inputs_read
=
874 tcs_stage
->nir
->info
.outputs_written
;
875 tes_stage
->key
.tes
.patch_inputs_read
=
876 tcs_stage
->nir
->info
.patch_outputs_written
;
878 tes_stage
->num_stats
= 1;
879 tes_stage
->code
= brw_compile_tes(compiler
, device
, mem_ctx
,
881 &tcs_stage
->prog_data
.tcs
.base
.vue_map
,
882 &tes_stage
->prog_data
.tes
,
884 tes_stage
->stats
, NULL
);
888 anv_pipeline_link_gs(const struct brw_compiler
*compiler
,
889 struct anv_pipeline_stage
*gs_stage
,
890 struct anv_pipeline_stage
*next_stage
)
893 brw_nir_link_shaders(compiler
, gs_stage
->nir
, next_stage
->nir
);
897 anv_pipeline_compile_gs(const struct brw_compiler
*compiler
,
899 struct anv_device
*device
,
900 struct anv_pipeline_stage
*gs_stage
,
901 struct anv_pipeline_stage
*prev_stage
)
903 brw_compute_vue_map(compiler
->devinfo
,
904 &gs_stage
->prog_data
.gs
.base
.vue_map
,
905 gs_stage
->nir
->info
.outputs_written
,
906 gs_stage
->nir
->info
.separate_shader
);
908 gs_stage
->num_stats
= 1;
909 gs_stage
->code
= brw_compile_gs(compiler
, device
, mem_ctx
,
911 &gs_stage
->prog_data
.gs
,
912 gs_stage
->nir
, NULL
, -1,
913 gs_stage
->stats
, NULL
);
917 anv_pipeline_link_fs(const struct brw_compiler
*compiler
,
918 struct anv_pipeline_stage
*stage
)
920 unsigned num_rts
= 0;
921 const int max_rt
= FRAG_RESULT_DATA7
- FRAG_RESULT_DATA0
+ 1;
922 struct anv_pipeline_binding rt_bindings
[max_rt
];
923 nir_function_impl
*impl
= nir_shader_get_entrypoint(stage
->nir
);
924 int rt_to_bindings
[max_rt
];
925 memset(rt_to_bindings
, -1, sizeof(rt_to_bindings
));
926 bool rt_used
[max_rt
];
927 memset(rt_used
, 0, sizeof(rt_used
));
929 /* Flag used render targets */
930 nir_foreach_variable_safe(var
, &stage
->nir
->outputs
) {
931 if (var
->data
.location
< FRAG_RESULT_DATA0
)
934 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
939 const unsigned array_len
=
940 glsl_type_is_array(var
->type
) ? glsl_get_length(var
->type
) : 1;
941 assert(rt
+ array_len
<= max_rt
);
944 if (!(stage
->key
.wm
.color_outputs_valid
& BITFIELD_RANGE(rt
, array_len
))) {
945 /* If this is the RT at location 0 and we have alpha to coverage
946 * enabled we will have to create a null RT for it, so mark it as
949 if (rt
> 0 || !stage
->key
.wm
.alpha_to_coverage
)
953 for (unsigned i
= 0; i
< array_len
; i
++)
954 rt_used
[rt
+ i
] = true;
957 /* Set new, compacted, location */
958 for (unsigned i
= 0; i
< max_rt
; i
++) {
962 rt_to_bindings
[i
] = num_rts
;
964 if (stage
->key
.wm
.color_outputs_valid
& (1 << i
)) {
965 rt_bindings
[rt_to_bindings
[i
]] = (struct anv_pipeline_binding
) {
966 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
971 /* Setup a null render target */
972 rt_bindings
[rt_to_bindings
[i
]] = (struct anv_pipeline_binding
) {
973 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
982 bool deleted_output
= false;
983 nir_foreach_variable_safe(var
, &stage
->nir
->outputs
) {
984 if (var
->data
.location
< FRAG_RESULT_DATA0
)
987 const unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
989 if (rt
>= MAX_RTS
|| !rt_used
[rt
]) {
990 /* Unused or out-of-bounds, throw it away, unless it is the first
991 * RT and we have alpha to coverage enabled.
993 deleted_output
= true;
994 var
->data
.mode
= nir_var_function_temp
;
995 exec_node_remove(&var
->node
);
996 exec_list_push_tail(&impl
->locals
, &var
->node
);
1000 /* Give it the new location */
1001 assert(rt_to_bindings
[rt
] != -1);
1002 var
->data
.location
= rt_to_bindings
[rt
] + FRAG_RESULT_DATA0
;
1006 nir_fixup_deref_modes(stage
->nir
);
1009 /* If we have no render targets, we need a null render target */
1010 rt_bindings
[0] = (struct anv_pipeline_binding
) {
1011 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
1013 .index
= UINT32_MAX
,
1018 /* Now that we've determined the actual number of render targets, adjust
1019 * the key accordingly.
1021 stage
->key
.wm
.nr_color_regions
= num_rts
;
1022 stage
->key
.wm
.color_outputs_valid
= (1 << num_rts
) - 1;
1024 assert(num_rts
<= max_rt
);
1025 assert(stage
->bind_map
.surface_count
== 0);
1026 typed_memcpy(stage
->bind_map
.surface_to_descriptor
,
1027 rt_bindings
, num_rts
);
1028 stage
->bind_map
.surface_count
+= num_rts
;
1032 anv_pipeline_compile_fs(const struct brw_compiler
*compiler
,
1034 struct anv_device
*device
,
1035 struct anv_pipeline_stage
*fs_stage
,
1036 struct anv_pipeline_stage
*prev_stage
)
1038 /* TODO: we could set this to 0 based on the information in nir_shader, but
1039 * we need this before we call spirv_to_nir.
1042 fs_stage
->key
.wm
.input_slots_valid
=
1043 prev_stage
->prog_data
.vue
.vue_map
.slots_valid
;
1045 fs_stage
->code
= brw_compile_fs(compiler
, device
, mem_ctx
,
1047 &fs_stage
->prog_data
.wm
,
1048 fs_stage
->nir
, -1, -1, -1,
1050 fs_stage
->stats
, NULL
);
1052 fs_stage
->num_stats
= (uint32_t)fs_stage
->prog_data
.wm
.dispatch_8
+
1053 (uint32_t)fs_stage
->prog_data
.wm
.dispatch_16
+
1054 (uint32_t)fs_stage
->prog_data
.wm
.dispatch_32
;
1056 if (fs_stage
->key
.wm
.nr_color_regions
== 0 &&
1057 !fs_stage
->prog_data
.wm
.has_side_effects
&&
1058 !fs_stage
->prog_data
.wm
.uses_kill
&&
1059 fs_stage
->prog_data
.wm
.computed_depth_mode
== BRW_PSCDEPTH_OFF
&&
1060 !fs_stage
->prog_data
.wm
.computed_stencil
) {
1061 /* This fragment shader has no outputs and no side effects. Go ahead
1062 * and return the code pointer so we don't accidentally think the
1063 * compile failed but zero out prog_data which will set program_size to
1064 * zero and disable the stage.
1066 memset(&fs_stage
->prog_data
, 0, sizeof(fs_stage
->prog_data
));
1071 anv_pipeline_add_executable(struct anv_pipeline
*pipeline
,
1072 struct anv_pipeline_stage
*stage
,
1073 struct brw_compile_stats
*stats
,
1074 uint32_t code_offset
)
1076 char *disasm
= NULL
;
1079 VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
)) {
1080 char *stream_data
= NULL
;
1081 size_t stream_size
= 0;
1082 FILE *stream
= open_memstream(&stream_data
, &stream_size
);
1084 /* Creating this is far cheaper than it looks. It's perfectly fine to
1085 * do it for every binary.
1087 struct gen_disasm
*d
= gen_disasm_create(&pipeline
->device
->info
);
1088 gen_disasm_disassemble(d
, stage
->code
, code_offset
, stream
);
1089 gen_disasm_destroy(d
);
1093 /* Copy it to a ralloc'd thing */
1094 disasm
= ralloc_size(pipeline
->mem_ctx
, stream_size
+ 1);
1095 memcpy(disasm
, stream_data
, stream_size
);
1096 disasm
[stream_size
] = 0;
1101 pipeline
->executables
[pipeline
->num_executables
++] =
1102 (struct anv_pipeline_executable
) {
1103 .stage
= stage
->stage
,
1110 anv_pipeline_add_executables(struct anv_pipeline
*pipeline
,
1111 struct anv_pipeline_stage
*stage
,
1112 struct anv_shader_bin
*bin
)
1114 if (stage
->stage
== MESA_SHADER_FRAGMENT
) {
1115 /* We pull the prog data and stats out of the anv_shader_bin because
1116 * the anv_pipeline_stage may not be fully populated if we successfully
1117 * looked up the shader in a cache.
1119 const struct brw_wm_prog_data
*wm_prog_data
=
1120 (const struct brw_wm_prog_data
*)bin
->prog_data
;
1121 struct brw_compile_stats
*stats
= bin
->stats
;
1123 if (wm_prog_data
->dispatch_8
) {
1124 anv_pipeline_add_executable(pipeline
, stage
, stats
++, 0);
1127 if (wm_prog_data
->dispatch_16
) {
1128 anv_pipeline_add_executable(pipeline
, stage
, stats
++,
1129 wm_prog_data
->prog_offset_16
);
1132 if (wm_prog_data
->dispatch_32
) {
1133 anv_pipeline_add_executable(pipeline
, stage
, stats
++,
1134 wm_prog_data
->prog_offset_32
);
1137 anv_pipeline_add_executable(pipeline
, stage
, bin
->stats
, 0);
1142 anv_pipeline_compile_graphics(struct anv_pipeline
*pipeline
,
1143 struct anv_pipeline_cache
*cache
,
1144 const VkGraphicsPipelineCreateInfo
*info
)
1146 VkPipelineCreationFeedbackEXT pipeline_feedback
= {
1147 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1149 int64_t pipeline_start
= os_time_get_nano();
1151 const struct brw_compiler
*compiler
=
1152 pipeline
->device
->instance
->physicalDevice
.compiler
;
1153 struct anv_pipeline_stage stages
[MESA_SHADER_STAGES
] = {};
1155 pipeline
->active_stages
= 0;
1158 for (uint32_t i
= 0; i
< info
->stageCount
; i
++) {
1159 const VkPipelineShaderStageCreateInfo
*sinfo
= &info
->pStages
[i
];
1160 gl_shader_stage stage
= vk_to_mesa_shader_stage(sinfo
->stage
);
1162 pipeline
->active_stages
|= sinfo
->stage
;
1164 int64_t stage_start
= os_time_get_nano();
1166 stages
[stage
].stage
= stage
;
1167 stages
[stage
].module
= anv_shader_module_from_handle(sinfo
->module
);
1168 stages
[stage
].entrypoint
= sinfo
->pName
;
1169 stages
[stage
].spec_info
= sinfo
->pSpecializationInfo
;
1170 anv_pipeline_hash_shader(stages
[stage
].module
,
1171 stages
[stage
].entrypoint
,
1173 stages
[stage
].spec_info
,
1174 stages
[stage
].shader_sha1
);
1176 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1178 case MESA_SHADER_VERTEX
:
1179 populate_vs_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.vs
);
1181 case MESA_SHADER_TESS_CTRL
:
1182 populate_tcs_prog_key(devinfo
, sinfo
->flags
,
1183 info
->pTessellationState
->patchControlPoints
,
1184 &stages
[stage
].key
.tcs
);
1186 case MESA_SHADER_TESS_EVAL
:
1187 populate_tes_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.tes
);
1189 case MESA_SHADER_GEOMETRY
:
1190 populate_gs_prog_key(devinfo
, sinfo
->flags
, &stages
[stage
].key
.gs
);
1192 case MESA_SHADER_FRAGMENT
:
1193 populate_wm_prog_key(devinfo
, sinfo
->flags
,
1195 info
->pMultisampleState
,
1196 &stages
[stage
].key
.wm
);
1199 unreachable("Invalid graphics shader stage");
1202 stages
[stage
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1203 stages
[stage
].feedback
.flags
|= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
;
1206 if (pipeline
->active_stages
& VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
)
1207 pipeline
->active_stages
|= VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
;
1209 assert(pipeline
->active_stages
& VK_SHADER_STAGE_VERTEX_BIT
);
1211 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1213 unsigned char sha1
[20];
1214 anv_pipeline_hash_graphics(pipeline
, layout
, stages
, sha1
);
1216 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1217 if (!stages
[s
].entrypoint
)
1220 stages
[s
].cache_key
.stage
= s
;
1221 memcpy(stages
[s
].cache_key
.sha1
, sha1
, sizeof(sha1
));
1224 const bool skip_cache_lookup
=
1225 (pipeline
->flags
& VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
);
1227 if (!skip_cache_lookup
) {
1229 unsigned cache_hits
= 0;
1230 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1231 if (!stages
[s
].entrypoint
)
1234 int64_t stage_start
= os_time_get_nano();
1237 struct anv_shader_bin
*bin
=
1238 anv_device_search_for_kernel(pipeline
->device
, cache
,
1239 &stages
[s
].cache_key
,
1240 sizeof(stages
[s
].cache_key
), &cache_hit
);
1243 pipeline
->shaders
[s
] = bin
;
1248 stages
[s
].feedback
.flags
|=
1249 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1251 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1254 if (found
== __builtin_popcount(pipeline
->active_stages
)) {
1255 if (cache_hits
== found
) {
1256 pipeline_feedback
.flags
|=
1257 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1259 /* We found all our shaders in the cache. We're done. */
1260 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1261 if (!stages
[s
].entrypoint
)
1264 anv_pipeline_add_executables(pipeline
, &stages
[s
],
1265 pipeline
->shaders
[s
]);
1268 } else if (found
> 0) {
1269 /* We found some but not all of our shaders. This shouldn't happen
1270 * most of the time but it can if we have a partially populated
1273 assert(found
< __builtin_popcount(pipeline
->active_stages
));
1275 vk_debug_report(&pipeline
->device
->instance
->debug_report_callbacks
,
1276 VK_DEBUG_REPORT_WARNING_BIT_EXT
|
1277 VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT
,
1278 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_CACHE_EXT
,
1279 (uint64_t)(uintptr_t)cache
,
1281 "Found a partial pipeline in the cache. This is "
1282 "most likely caused by an incomplete pipeline cache "
1283 "import or export");
1285 /* We're going to have to recompile anyway, so just throw away our
1286 * references to the shaders in the cache. We'll get them out of the
1287 * cache again as part of the compilation process.
1289 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1290 stages
[s
].feedback
.flags
= 0;
1291 if (pipeline
->shaders
[s
]) {
1292 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1293 pipeline
->shaders
[s
] = NULL
;
1299 void *pipeline_ctx
= ralloc_context(NULL
);
1301 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1302 if (!stages
[s
].entrypoint
)
1305 int64_t stage_start
= os_time_get_nano();
1307 assert(stages
[s
].stage
== s
);
1308 assert(pipeline
->shaders
[s
] == NULL
);
1310 stages
[s
].bind_map
= (struct anv_pipeline_bind_map
) {
1311 .surface_to_descriptor
= stages
[s
].surface_to_descriptor
,
1312 .sampler_to_descriptor
= stages
[s
].sampler_to_descriptor
1315 stages
[s
].nir
= anv_pipeline_stage_get_nir(pipeline
, cache
,
1318 if (stages
[s
].nir
== NULL
) {
1319 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1323 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1326 /* Walk backwards to link */
1327 struct anv_pipeline_stage
*next_stage
= NULL
;
1328 for (int s
= MESA_SHADER_STAGES
- 1; s
>= 0; s
--) {
1329 if (!stages
[s
].entrypoint
)
1333 case MESA_SHADER_VERTEX
:
1334 anv_pipeline_link_vs(compiler
, &stages
[s
], next_stage
);
1336 case MESA_SHADER_TESS_CTRL
:
1337 anv_pipeline_link_tcs(compiler
, &stages
[s
], next_stage
);
1339 case MESA_SHADER_TESS_EVAL
:
1340 anv_pipeline_link_tes(compiler
, &stages
[s
], next_stage
);
1342 case MESA_SHADER_GEOMETRY
:
1343 anv_pipeline_link_gs(compiler
, &stages
[s
], next_stage
);
1345 case MESA_SHADER_FRAGMENT
:
1346 anv_pipeline_link_fs(compiler
, &stages
[s
]);
1349 unreachable("Invalid graphics shader stage");
1352 next_stage
= &stages
[s
];
1355 struct anv_pipeline_stage
*prev_stage
= NULL
;
1356 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1357 if (!stages
[s
].entrypoint
)
1360 int64_t stage_start
= os_time_get_nano();
1362 void *stage_ctx
= ralloc_context(NULL
);
1364 nir_xfb_info
*xfb_info
= NULL
;
1365 if (s
== MESA_SHADER_VERTEX
||
1366 s
== MESA_SHADER_TESS_EVAL
||
1367 s
== MESA_SHADER_GEOMETRY
)
1368 xfb_info
= nir_gather_xfb_info(stages
[s
].nir
, stage_ctx
);
1370 anv_pipeline_lower_nir(pipeline
, stage_ctx
, &stages
[s
], layout
);
1373 case MESA_SHADER_VERTEX
:
1374 anv_pipeline_compile_vs(compiler
, stage_ctx
, pipeline
->device
,
1377 case MESA_SHADER_TESS_CTRL
:
1378 anv_pipeline_compile_tcs(compiler
, stage_ctx
, pipeline
->device
,
1379 &stages
[s
], prev_stage
);
1381 case MESA_SHADER_TESS_EVAL
:
1382 anv_pipeline_compile_tes(compiler
, stage_ctx
, pipeline
->device
,
1383 &stages
[s
], prev_stage
);
1385 case MESA_SHADER_GEOMETRY
:
1386 anv_pipeline_compile_gs(compiler
, stage_ctx
, pipeline
->device
,
1387 &stages
[s
], prev_stage
);
1389 case MESA_SHADER_FRAGMENT
:
1390 anv_pipeline_compile_fs(compiler
, stage_ctx
, pipeline
->device
,
1391 &stages
[s
], prev_stage
);
1394 unreachable("Invalid graphics shader stage");
1396 if (stages
[s
].code
== NULL
) {
1397 ralloc_free(stage_ctx
);
1398 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1402 struct anv_shader_bin
*bin
=
1403 anv_device_upload_kernel(pipeline
->device
, cache
,
1404 &stages
[s
].cache_key
,
1405 sizeof(stages
[s
].cache_key
),
1407 stages
[s
].prog_data
.base
.program_size
,
1408 stages
[s
].nir
->constant_data
,
1409 stages
[s
].nir
->constant_data_size
,
1410 &stages
[s
].prog_data
.base
,
1411 brw_prog_data_size(s
),
1412 stages
[s
].stats
, stages
[s
].num_stats
,
1413 xfb_info
, &stages
[s
].bind_map
);
1415 ralloc_free(stage_ctx
);
1416 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1420 anv_pipeline_add_executables(pipeline
, &stages
[s
], bin
);
1422 pipeline
->shaders
[s
] = bin
;
1423 ralloc_free(stage_ctx
);
1425 stages
[s
].feedback
.duration
+= os_time_get_nano() - stage_start
;
1427 prev_stage
= &stages
[s
];
1430 ralloc_free(pipeline_ctx
);
1434 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
] &&
1435 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->prog_data
->program_size
== 0) {
1436 /* This can happen if we decided to implicitly disable the fragment
1437 * shader. See anv_pipeline_compile_fs().
1439 anv_shader_bin_unref(pipeline
->device
,
1440 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
1441 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] = NULL
;
1442 pipeline
->active_stages
&= ~VK_SHADER_STAGE_FRAGMENT_BIT
;
1445 pipeline_feedback
.duration
= os_time_get_nano() - pipeline_start
;
1447 const VkPipelineCreationFeedbackCreateInfoEXT
*create_feedback
=
1448 vk_find_struct_const(info
->pNext
, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT
);
1449 if (create_feedback
) {
1450 *create_feedback
->pPipelineCreationFeedback
= pipeline_feedback
;
1452 assert(info
->stageCount
== create_feedback
->pipelineStageCreationFeedbackCount
);
1453 for (uint32_t i
= 0; i
< info
->stageCount
; i
++) {
1454 gl_shader_stage s
= vk_to_mesa_shader_stage(info
->pStages
[i
].stage
);
1455 create_feedback
->pPipelineStageCreationFeedbacks
[i
] = stages
[s
].feedback
;
1462 ralloc_free(pipeline_ctx
);
1464 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1465 if (pipeline
->shaders
[s
])
1466 anv_shader_bin_unref(pipeline
->device
, pipeline
->shaders
[s
]);
1473 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
1475 assert(glsl_type_is_vector_or_scalar(type
));
1477 uint32_t comp_size
= glsl_type_is_boolean(type
)
1478 ? 4 : glsl_get_bit_size(type
) / 8;
1479 unsigned length
= glsl_get_vector_elements(type
);
1480 *size
= comp_size
* length
,
1481 *align
= comp_size
* (length
== 3 ? 4 : length
);
1485 anv_pipeline_compile_cs(struct anv_pipeline
*pipeline
,
1486 struct anv_pipeline_cache
*cache
,
1487 const VkComputePipelineCreateInfo
*info
,
1488 const struct anv_shader_module
*module
,
1489 const char *entrypoint
,
1490 const VkSpecializationInfo
*spec_info
)
1492 VkPipelineCreationFeedbackEXT pipeline_feedback
= {
1493 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1495 int64_t pipeline_start
= os_time_get_nano();
1497 const struct brw_compiler
*compiler
=
1498 pipeline
->device
->instance
->physicalDevice
.compiler
;
1500 struct anv_pipeline_stage stage
= {
1501 .stage
= MESA_SHADER_COMPUTE
,
1503 .entrypoint
= entrypoint
,
1504 .spec_info
= spec_info
,
1506 .stage
= MESA_SHADER_COMPUTE
,
1509 .flags
= VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT
,
1512 anv_pipeline_hash_shader(stage
.module
,
1514 MESA_SHADER_COMPUTE
,
1518 struct anv_shader_bin
*bin
= NULL
;
1520 const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT
*rss_info
=
1521 vk_find_struct_const(info
->stage
.pNext
,
1522 PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT
);
1524 populate_cs_prog_key(&pipeline
->device
->info
, info
->stage
.flags
,
1525 rss_info
, &stage
.key
.cs
);
1527 ANV_FROM_HANDLE(anv_pipeline_layout
, layout
, info
->layout
);
1529 const bool skip_cache_lookup
=
1530 (pipeline
->flags
& VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR
);
1532 anv_pipeline_hash_compute(pipeline
, layout
, &stage
, stage
.cache_key
.sha1
);
1534 bool cache_hit
= false;
1535 if (!skip_cache_lookup
) {
1536 bin
= anv_device_search_for_kernel(pipeline
->device
, cache
,
1538 sizeof(stage
.cache_key
),
1542 void *mem_ctx
= ralloc_context(NULL
);
1544 int64_t stage_start
= os_time_get_nano();
1546 stage
.bind_map
= (struct anv_pipeline_bind_map
) {
1547 .surface_to_descriptor
= stage
.surface_to_descriptor
,
1548 .sampler_to_descriptor
= stage
.sampler_to_descriptor
1551 /* Set up a binding for the gl_NumWorkGroups */
1552 stage
.bind_map
.surface_count
= 1;
1553 stage
.bind_map
.surface_to_descriptor
[0] = (struct anv_pipeline_binding
) {
1554 .set
= ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
,
1557 stage
.nir
= anv_pipeline_stage_get_nir(pipeline
, cache
, mem_ctx
, &stage
);
1558 if (stage
.nir
== NULL
) {
1559 ralloc_free(mem_ctx
);
1560 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1563 anv_pipeline_lower_nir(pipeline
, mem_ctx
, &stage
, layout
);
1565 NIR_PASS_V(stage
.nir
, anv_nir_add_base_work_group_id
,
1566 &stage
.prog_data
.cs
);
1568 NIR_PASS_V(stage
.nir
, nir_lower_vars_to_explicit_types
,
1569 nir_var_mem_shared
, shared_type_info
);
1570 NIR_PASS_V(stage
.nir
, nir_lower_explicit_io
,
1571 nir_var_mem_shared
, nir_address_format_32bit_offset
);
1573 stage
.num_stats
= 1;
1574 stage
.code
= brw_compile_cs(compiler
, pipeline
->device
, mem_ctx
,
1575 &stage
.key
.cs
, &stage
.prog_data
.cs
,
1576 stage
.nir
, -1, stage
.stats
, NULL
);
1577 if (stage
.code
== NULL
) {
1578 ralloc_free(mem_ctx
);
1579 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1582 const unsigned code_size
= stage
.prog_data
.base
.program_size
;
1583 bin
= anv_device_upload_kernel(pipeline
->device
, cache
,
1584 &stage
.cache_key
, sizeof(stage
.cache_key
),
1585 stage
.code
, code_size
,
1586 stage
.nir
->constant_data
,
1587 stage
.nir
->constant_data_size
,
1588 &stage
.prog_data
.base
,
1589 sizeof(stage
.prog_data
.cs
),
1590 stage
.stats
, stage
.num_stats
,
1591 NULL
, &stage
.bind_map
);
1593 ralloc_free(mem_ctx
);
1594 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1597 stage
.feedback
.duration
= os_time_get_nano() - stage_start
;
1600 anv_pipeline_add_executables(pipeline
, &stage
, bin
);
1602 ralloc_free(mem_ctx
);
1605 stage
.feedback
.flags
|=
1606 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1607 pipeline_feedback
.flags
|=
1608 VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT
;
1610 pipeline_feedback
.duration
= os_time_get_nano() - pipeline_start
;
1612 const VkPipelineCreationFeedbackCreateInfoEXT
*create_feedback
=
1613 vk_find_struct_const(info
->pNext
, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT
);
1614 if (create_feedback
) {
1615 *create_feedback
->pPipelineCreationFeedback
= pipeline_feedback
;
1617 assert(create_feedback
->pipelineStageCreationFeedbackCount
== 1);
1618 create_feedback
->pPipelineStageCreationFeedbacks
[0] = stage
.feedback
;
1621 pipeline
->active_stages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1622 pipeline
->shaders
[MESA_SHADER_COMPUTE
] = bin
;
1628 * Copy pipeline state not marked as dynamic.
1629 * Dynamic state is pipeline state which hasn't been provided at pipeline
1630 * creation time, but is dynamically provided afterwards using various
1631 * vkCmdSet* functions.
1633 * The set of state considered "non_dynamic" is determined by the pieces of
1634 * state that have their corresponding VkDynamicState enums omitted from
1635 * VkPipelineDynamicStateCreateInfo::pDynamicStates.
1637 * @param[out] pipeline Destination non_dynamic state.
1638 * @param[in] pCreateInfo Source of non_dynamic state to be copied.
1641 copy_non_dynamic_state(struct anv_pipeline
*pipeline
,
1642 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1644 anv_cmd_dirty_mask_t states
= ANV_CMD_DIRTY_DYNAMIC_ALL
;
1645 struct anv_subpass
*subpass
= pipeline
->subpass
;
1647 pipeline
->dynamic_state
= default_dynamic_state
;
1649 if (pCreateInfo
->pDynamicState
) {
1650 /* Remove all of the states that are marked as dynamic */
1651 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1652 for (uint32_t s
= 0; s
< count
; s
++) {
1653 states
&= ~anv_cmd_dirty_bit_for_vk_dynamic_state(
1654 pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1658 struct anv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1660 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1662 * pViewportState is [...] NULL if the pipeline
1663 * has rasterization disabled.
1665 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1666 assert(pCreateInfo
->pViewportState
);
1668 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1669 if (states
& ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
) {
1670 typed_memcpy(dynamic
->viewport
.viewports
,
1671 pCreateInfo
->pViewportState
->pViewports
,
1672 pCreateInfo
->pViewportState
->viewportCount
);
1675 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1676 if (states
& ANV_CMD_DIRTY_DYNAMIC_SCISSOR
) {
1677 typed_memcpy(dynamic
->scissor
.scissors
,
1678 pCreateInfo
->pViewportState
->pScissors
,
1679 pCreateInfo
->pViewportState
->scissorCount
);
1683 if (states
& ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
) {
1684 assert(pCreateInfo
->pRasterizationState
);
1685 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1688 if (states
& ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
) {
1689 assert(pCreateInfo
->pRasterizationState
);
1690 dynamic
->depth_bias
.bias
=
1691 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1692 dynamic
->depth_bias
.clamp
=
1693 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1694 dynamic
->depth_bias
.slope
=
1695 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1698 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1700 * pColorBlendState is [...] NULL if the pipeline has rasterization
1701 * disabled or if the subpass of the render pass the pipeline is
1702 * created against does not use any color attachments.
1704 bool uses_color_att
= false;
1705 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1706 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1707 uses_color_att
= true;
1712 if (uses_color_att
&&
1713 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1714 assert(pCreateInfo
->pColorBlendState
);
1716 if (states
& ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
)
1717 typed_memcpy(dynamic
->blend_constants
,
1718 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1721 /* If there is no depthstencil attachment, then don't read
1722 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1723 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1724 * no need to override the depthstencil defaults in
1725 * anv_pipeline::dynamic_state when there is no depthstencil attachment.
1727 * Section 9.2 of the Vulkan 1.0.15 spec says:
1729 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1730 * disabled or if the subpass of the render pass the pipeline is created
1731 * against does not use a depth/stencil attachment.
1733 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1734 subpass
->depth_stencil_attachment
) {
1735 assert(pCreateInfo
->pDepthStencilState
);
1737 if (states
& ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
) {
1738 dynamic
->depth_bounds
.min
=
1739 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1740 dynamic
->depth_bounds
.max
=
1741 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1744 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
) {
1745 dynamic
->stencil_compare_mask
.front
=
1746 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1747 dynamic
->stencil_compare_mask
.back
=
1748 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1751 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
) {
1752 dynamic
->stencil_write_mask
.front
=
1753 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1754 dynamic
->stencil_write_mask
.back
=
1755 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1758 if (states
& ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
) {
1759 dynamic
->stencil_reference
.front
=
1760 pCreateInfo
->pDepthStencilState
->front
.reference
;
1761 dynamic
->stencil_reference
.back
=
1762 pCreateInfo
->pDepthStencilState
->back
.reference
;
1766 const VkPipelineRasterizationLineStateCreateInfoEXT
*line_state
=
1767 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
,
1768 PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT
);
1770 if (states
& ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE
) {
1771 dynamic
->line_stipple
.factor
= line_state
->lineStippleFactor
;
1772 dynamic
->line_stipple
.pattern
= line_state
->lineStipplePattern
;
1776 pipeline
->dynamic_state_mask
= states
;
1780 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo
*info
)
1783 struct anv_render_pass
*renderpass
= NULL
;
1784 struct anv_subpass
*subpass
= NULL
;
1786 /* Assert that all required members of VkGraphicsPipelineCreateInfo are
1787 * present. See the Vulkan 1.0.28 spec, Section 9.2 Graphics Pipelines.
1789 assert(info
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1791 renderpass
= anv_render_pass_from_handle(info
->renderPass
);
1794 assert(info
->subpass
< renderpass
->subpass_count
);
1795 subpass
= &renderpass
->subpasses
[info
->subpass
];
1797 assert(info
->stageCount
>= 1);
1798 assert(info
->pVertexInputState
);
1799 assert(info
->pInputAssemblyState
);
1800 assert(info
->pRasterizationState
);
1801 if (!info
->pRasterizationState
->rasterizerDiscardEnable
) {
1802 assert(info
->pViewportState
);
1803 assert(info
->pMultisampleState
);
1805 if (subpass
&& subpass
->depth_stencil_attachment
)
1806 assert(info
->pDepthStencilState
);
1808 if (subpass
&& subpass
->color_count
> 0) {
1809 bool all_color_unused
= true;
1810 for (int i
= 0; i
< subpass
->color_count
; i
++) {
1811 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
1812 all_color_unused
= false;
1814 /* pColorBlendState is ignored if the pipeline has rasterization
1815 * disabled or if the subpass of the render pass the pipeline is
1816 * created against does not use any color attachments.
1818 assert(info
->pColorBlendState
|| all_color_unused
);
1822 for (uint32_t i
= 0; i
< info
->stageCount
; ++i
) {
1823 switch (info
->pStages
[i
].stage
) {
1824 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
:
1825 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
:
1826 assert(info
->pTessellationState
);
1836 * Calculate the desired L3 partitioning based on the current state of the
1837 * pipeline. For now this simply returns the conservative defaults calculated
1838 * by get_default_l3_weights(), but we could probably do better by gathering
1839 * more statistics from the pipeline state (e.g. guess of expected URB usage
1840 * and bound surfaces), or by using feed-back from performance counters.
1843 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
)
1845 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1847 const struct gen_l3_weights w
=
1848 gen_get_default_l3_weights(devinfo
, pipeline
->needs_data_cache
, needs_slm
);
1850 pipeline
->urb
.l3_config
= gen_get_l3_config(devinfo
, w
);
1851 pipeline
->urb
.total_size
=
1852 gen_get_l3_config_urb_size(devinfo
, pipeline
->urb
.l3_config
);
1856 anv_pipeline_init(struct anv_pipeline
*pipeline
,
1857 struct anv_device
*device
,
1858 struct anv_pipeline_cache
*cache
,
1859 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1860 const VkAllocationCallbacks
*alloc
)
1864 anv_pipeline_validate_create_info(pCreateInfo
);
1867 alloc
= &device
->alloc
;
1869 pipeline
->device
= device
;
1871 ANV_FROM_HANDLE(anv_render_pass
, render_pass
, pCreateInfo
->renderPass
);
1872 assert(pCreateInfo
->subpass
< render_pass
->subpass_count
);
1873 pipeline
->subpass
= &render_pass
->subpasses
[pCreateInfo
->subpass
];
1875 result
= anv_reloc_list_init(&pipeline
->batch_relocs
, alloc
);
1876 if (result
!= VK_SUCCESS
)
1879 pipeline
->batch
.alloc
= alloc
;
1880 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1881 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1882 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1883 pipeline
->batch
.status
= VK_SUCCESS
;
1885 pipeline
->mem_ctx
= ralloc_context(NULL
);
1886 pipeline
->flags
= pCreateInfo
->flags
;
1888 copy_non_dynamic_state(pipeline
, pCreateInfo
);
1889 pipeline
->depth_clamp_enable
= pCreateInfo
->pRasterizationState
&&
1890 pCreateInfo
->pRasterizationState
->depthClampEnable
;
1892 /* Previously we enabled depth clipping when !depthClampEnable.
1893 * DepthClipStateCreateInfo now makes depth clipping explicit so if the
1894 * clipping info is available, use its enable value to determine clipping,
1895 * otherwise fallback to the previous !depthClampEnable logic.
1897 const VkPipelineRasterizationDepthClipStateCreateInfoEXT
*clip_info
=
1898 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
,
1899 PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT
);
1900 pipeline
->depth_clip_enable
= clip_info
? clip_info
->depthClipEnable
: !pipeline
->depth_clamp_enable
;
1902 pipeline
->sample_shading_enable
= pCreateInfo
->pMultisampleState
&&
1903 pCreateInfo
->pMultisampleState
->sampleShadingEnable
;
1905 pipeline
->needs_data_cache
= false;
1907 /* When we free the pipeline, we detect stages based on the NULL status
1908 * of various prog_data pointers. Make them NULL by default.
1910 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1911 pipeline
->num_executables
= 0;
1913 result
= anv_pipeline_compile_graphics(pipeline
, cache
, pCreateInfo
);
1914 if (result
!= VK_SUCCESS
) {
1915 ralloc_free(pipeline
->mem_ctx
);
1916 anv_reloc_list_finish(&pipeline
->batch_relocs
, alloc
);
1920 assert(pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1922 anv_pipeline_setup_l3_config(pipeline
, false);
1924 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1925 pCreateInfo
->pVertexInputState
;
1927 const uint64_t inputs_read
= get_vs_prog_data(pipeline
)->inputs_read
;
1929 pipeline
->vb_used
= 0;
1930 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1931 const VkVertexInputAttributeDescription
*desc
=
1932 &vi_info
->pVertexAttributeDescriptions
[i
];
1934 if (inputs_read
& (1ull << (VERT_ATTRIB_GENERIC0
+ desc
->location
)))
1935 pipeline
->vb_used
|= 1 << desc
->binding
;
1938 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1939 const VkVertexInputBindingDescription
*desc
=
1940 &vi_info
->pVertexBindingDescriptions
[i
];
1942 pipeline
->vb
[desc
->binding
].stride
= desc
->stride
;
1944 /* Step rate is programmed per vertex element (attribute), not
1945 * binding. Set up a map of which bindings step per instance, for
1946 * reference by vertex element setup. */
1947 switch (desc
->inputRate
) {
1949 case VK_VERTEX_INPUT_RATE_VERTEX
:
1950 pipeline
->vb
[desc
->binding
].instanced
= false;
1952 case VK_VERTEX_INPUT_RATE_INSTANCE
:
1953 pipeline
->vb
[desc
->binding
].instanced
= true;
1957 pipeline
->vb
[desc
->binding
].instance_divisor
= 1;
1960 const VkPipelineVertexInputDivisorStateCreateInfoEXT
*vi_div_state
=
1961 vk_find_struct_const(vi_info
->pNext
,
1962 PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT
);
1964 for (uint32_t i
= 0; i
< vi_div_state
->vertexBindingDivisorCount
; i
++) {
1965 const VkVertexInputBindingDivisorDescriptionEXT
*desc
=
1966 &vi_div_state
->pVertexBindingDivisors
[i
];
1968 pipeline
->vb
[desc
->binding
].instance_divisor
= desc
->divisor
;
1972 /* Our implementation of VK_KHR_multiview uses instancing to draw the
1973 * different views. If the client asks for instancing, we need to multiply
1974 * the instance divisor by the number of views ensure that we repeat the
1975 * client's per-instance data once for each view.
1977 if (pipeline
->subpass
->view_mask
) {
1978 const uint32_t view_count
= anv_subpass_view_count(pipeline
->subpass
);
1979 for (uint32_t vb
= 0; vb
< MAX_VBS
; vb
++) {
1980 if (pipeline
->vb
[vb
].instanced
)
1981 pipeline
->vb
[vb
].instance_divisor
*= view_count
;
1985 const VkPipelineInputAssemblyStateCreateInfo
*ia_info
=
1986 pCreateInfo
->pInputAssemblyState
;
1987 const VkPipelineTessellationStateCreateInfo
*tess_info
=
1988 pCreateInfo
->pTessellationState
;
1989 pipeline
->primitive_restart
= ia_info
->primitiveRestartEnable
;
1991 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
1992 pipeline
->topology
= _3DPRIM_PATCHLIST(tess_info
->patchControlPoints
);
1994 pipeline
->topology
= vk_to_gen_primitive_type
[ia_info
->topology
];
1999 #define WRITE_STR(field, ...) ({ \
2000 memset(field, 0, sizeof(field)); \
2001 UNUSED int i = snprintf(field, sizeof(field), __VA_ARGS__); \
2002 assert(i > 0 && i < sizeof(field)); \
2005 VkResult
anv_GetPipelineExecutablePropertiesKHR(
2007 const VkPipelineInfoKHR
* pPipelineInfo
,
2008 uint32_t* pExecutableCount
,
2009 VkPipelineExecutablePropertiesKHR
* pProperties
)
2011 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pPipelineInfo
->pipeline
);
2012 VK_OUTARRAY_MAKE(out
, pProperties
, pExecutableCount
);
2014 for (uint32_t i
= 0; i
< pipeline
->num_executables
; i
++) {
2015 vk_outarray_append(&out
, props
) {
2016 gl_shader_stage stage
= pipeline
->executables
[i
].stage
;
2017 props
->stages
= mesa_to_vk_shader_stage(stage
);
2019 unsigned simd_width
= pipeline
->executables
[i
].stats
.dispatch_width
;
2020 if (stage
== MESA_SHADER_FRAGMENT
) {
2021 WRITE_STR(props
->name
, "%s%d %s",
2022 simd_width
? "SIMD" : "vec",
2023 simd_width
? simd_width
: 4,
2024 _mesa_shader_stage_to_string(stage
));
2026 WRITE_STR(props
->name
, "%s", _mesa_shader_stage_to_string(stage
));
2028 WRITE_STR(props
->description
, "%s%d %s shader",
2029 simd_width
? "SIMD" : "vec",
2030 simd_width
? simd_width
: 4,
2031 _mesa_shader_stage_to_string(stage
));
2033 /* The compiler gives us a dispatch width of 0 for vec4 but Vulkan
2034 * wants a subgroup size of 1.
2036 props
->subgroupSize
= MAX2(simd_width
, 1);
2040 return vk_outarray_status(&out
);
2043 VkResult
anv_GetPipelineExecutableStatisticsKHR(
2045 const VkPipelineExecutableInfoKHR
* pExecutableInfo
,
2046 uint32_t* pStatisticCount
,
2047 VkPipelineExecutableStatisticKHR
* pStatistics
)
2049 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pExecutableInfo
->pipeline
);
2050 VK_OUTARRAY_MAKE(out
, pStatistics
, pStatisticCount
);
2052 assert(pExecutableInfo
->executableIndex
< pipeline
->num_executables
);
2053 const struct anv_pipeline_executable
*exe
=
2054 &pipeline
->executables
[pExecutableInfo
->executableIndex
];
2055 const struct brw_stage_prog_data
*prog_data
=
2056 pipeline
->shaders
[exe
->stage
]->prog_data
;
2058 vk_outarray_append(&out
, stat
) {
2059 WRITE_STR(stat
->name
, "Instruction Count");
2060 WRITE_STR(stat
->description
,
2061 "Number of GEN instructions in the final generated "
2062 "shader executable.");
2063 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2064 stat
->value
.u64
= exe
->stats
.instructions
;
2067 vk_outarray_append(&out
, stat
) {
2068 WRITE_STR(stat
->name
, "Loop Count");
2069 WRITE_STR(stat
->description
,
2070 "Number of loops (not unrolled) in the final generated "
2071 "shader executable.");
2072 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2073 stat
->value
.u64
= exe
->stats
.loops
;
2076 vk_outarray_append(&out
, stat
) {
2077 WRITE_STR(stat
->name
, "Cycle Count");
2078 WRITE_STR(stat
->description
,
2079 "Estimate of the number of EU cycles required to execute "
2080 "the final generated executable. This is an estimate only "
2081 "and may vary greatly from actual run-time performance.");
2082 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2083 stat
->value
.u64
= exe
->stats
.cycles
;
2086 vk_outarray_append(&out
, stat
) {
2087 WRITE_STR(stat
->name
, "Spill Count");
2088 WRITE_STR(stat
->description
,
2089 "Number of scratch spill operations. This gives a rough "
2090 "estimate of the cost incurred due to spilling temporary "
2091 "values to memory. If this is non-zero, you may want to "
2092 "adjust your shader to reduce register pressure.");
2093 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2094 stat
->value
.u64
= exe
->stats
.spills
;
2097 vk_outarray_append(&out
, stat
) {
2098 WRITE_STR(stat
->name
, "Fill Count");
2099 WRITE_STR(stat
->description
,
2100 "Number of scratch fill operations. This gives a rough "
2101 "estimate of the cost incurred due to spilling temporary "
2102 "values to memory. If this is non-zero, you may want to "
2103 "adjust your shader to reduce register pressure.");
2104 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2105 stat
->value
.u64
= exe
->stats
.fills
;
2108 vk_outarray_append(&out
, stat
) {
2109 WRITE_STR(stat
->name
, "Scratch Memory Size");
2110 WRITE_STR(stat
->description
,
2111 "Number of bytes of scratch memory required by the "
2112 "generated shader executable. If this is non-zero, you "
2113 "may want to adjust your shader to reduce register "
2115 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2116 stat
->value
.u64
= prog_data
->total_scratch
;
2119 if (exe
->stage
== MESA_SHADER_COMPUTE
) {
2120 vk_outarray_append(&out
, stat
) {
2121 WRITE_STR(stat
->name
, "Workgroup Memory Size");
2122 WRITE_STR(stat
->description
,
2123 "Number of bytes of workgroup shared memory used by this "
2124 "compute shader including any padding.");
2125 stat
->format
= VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR
;
2126 stat
->value
.u64
= prog_data
->total_scratch
;
2130 return vk_outarray_status(&out
);
2134 write_ir_text(VkPipelineExecutableInternalRepresentationKHR
* ir
,
2137 ir
->isText
= VK_TRUE
;
2139 size_t data_len
= strlen(data
) + 1;
2141 if (ir
->pData
== NULL
) {
2142 ir
->dataSize
= data_len
;
2146 strncpy(ir
->pData
, data
, ir
->dataSize
);
2147 if (ir
->dataSize
< data_len
)
2150 ir
->dataSize
= data_len
;
2154 VkResult
anv_GetPipelineExecutableInternalRepresentationsKHR(
2156 const VkPipelineExecutableInfoKHR
* pExecutableInfo
,
2157 uint32_t* pInternalRepresentationCount
,
2158 VkPipelineExecutableInternalRepresentationKHR
* pInternalRepresentations
)
2160 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, pExecutableInfo
->pipeline
);
2161 VK_OUTARRAY_MAKE(out
, pInternalRepresentations
,
2162 pInternalRepresentationCount
);
2163 bool incomplete_text
= false;
2165 assert(pExecutableInfo
->executableIndex
< pipeline
->num_executables
);
2166 const struct anv_pipeline_executable
*exe
=
2167 &pipeline
->executables
[pExecutableInfo
->executableIndex
];
2170 vk_outarray_append(&out
, ir
) {
2171 WRITE_STR(ir
->name
, "GEN Assembly");
2172 WRITE_STR(ir
->description
,
2173 "Final GEN assembly for the generated shader binary");
2175 if (!write_ir_text(ir
, exe
->disasm
))
2176 incomplete_text
= true;
2180 return incomplete_text
? VK_INCOMPLETE
: vk_outarray_status(&out
);