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 "common/gen_l3_config.h"
32 #include "anv_private.h"
33 #include "compiler/brw_nir.h"
35 #include "spirv/nir_spirv.h"
37 /* Needed for SWIZZLE macros */
38 #include "program/prog_instruction.h"
42 VkResult
anv_CreateShaderModule(
44 const VkShaderModuleCreateInfo
* pCreateInfo
,
45 const VkAllocationCallbacks
* pAllocator
,
46 VkShaderModule
* pShaderModule
)
48 ANV_FROM_HANDLE(anv_device
, device
, _device
);
49 struct anv_shader_module
*module
;
51 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
52 assert(pCreateInfo
->flags
== 0);
54 module
= vk_alloc2(&device
->alloc
, pAllocator
,
55 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
56 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
58 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
60 module
->size
= pCreateInfo
->codeSize
;
61 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
63 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
65 *pShaderModule
= anv_shader_module_to_handle(module
);
70 void anv_DestroyShaderModule(
72 VkShaderModule _module
,
73 const VkAllocationCallbacks
* pAllocator
)
75 ANV_FROM_HANDLE(anv_device
, device
, _device
);
76 ANV_FROM_HANDLE(anv_shader_module
, module
, _module
);
81 vk_free2(&device
->alloc
, pAllocator
, module
);
84 #define SPIR_V_MAGIC_NUMBER 0x07230203
86 /* Eventually, this will become part of anv_CreateShader. Unfortunately,
87 * we can't do that yet because we don't have the ability to copy nir.
90 anv_shader_compile_to_nir(struct anv_pipeline
*pipeline
,
91 struct anv_shader_module
*module
,
92 const char *entrypoint_name
,
93 gl_shader_stage stage
,
94 const VkSpecializationInfo
*spec_info
)
96 const struct anv_device
*device
= pipeline
->device
;
98 const struct brw_compiler
*compiler
=
99 device
->instance
->physicalDevice
.compiler
;
100 const nir_shader_compiler_options
*nir_options
=
101 compiler
->glsl_compiler_options
[stage
].NirOptions
;
103 uint32_t *spirv
= (uint32_t *) module
->data
;
104 assert(spirv
[0] == SPIR_V_MAGIC_NUMBER
);
105 assert(module
->size
% 4 == 0);
107 uint32_t num_spec_entries
= 0;
108 struct nir_spirv_specialization
*spec_entries
= NULL
;
109 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
110 num_spec_entries
= spec_info
->mapEntryCount
;
111 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
112 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
113 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
114 const void *data
= spec_info
->pData
+ entry
.offset
;
115 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
117 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
118 if (spec_info
->dataSize
== 8)
119 spec_entries
[i
].data64
= *(const uint64_t *)data
;
121 spec_entries
[i
].data32
= *(const uint32_t *)data
;
125 const struct nir_spirv_supported_extensions supported_ext
= {
126 .float64
= device
->instance
->physicalDevice
.info
.gen
>= 8,
127 .int64
= device
->instance
->physicalDevice
.info
.gen
>= 8,
128 .tessellation
= true,
129 .draw_parameters
= true,
130 .image_write_without_format
= true,
134 nir_function
*entry_point
=
135 spirv_to_nir(spirv
, module
->size
/ 4,
136 spec_entries
, num_spec_entries
,
137 stage
, entrypoint_name
, &supported_ext
, nir_options
);
138 nir_shader
*nir
= entry_point
->shader
;
139 assert(nir
->stage
== stage
);
140 nir_validate_shader(nir
);
144 /* We have to lower away local constant initializers right before we
145 * inline functions. That way they get properly initialized at the top
146 * of the function and not at the top of its caller.
148 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_local
);
149 NIR_PASS_V(nir
, nir_lower_returns
);
150 NIR_PASS_V(nir
, nir_inline_functions
);
152 /* Pick off the single entrypoint that we want */
153 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
154 if (func
!= entry_point
)
155 exec_node_remove(&func
->node
);
157 assert(exec_list_length(&nir
->functions
) == 1);
158 entry_point
->name
= ralloc_strdup(entry_point
, "main");
160 NIR_PASS_V(nir
, nir_remove_dead_variables
,
161 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
);
163 if (stage
== MESA_SHADER_FRAGMENT
)
164 NIR_PASS_V(nir
, nir_lower_wpos_center
, pipeline
->sample_shading_enable
);
166 /* Now that we've deleted all but the main function, we can go ahead and
167 * lower the rest of the constant initializers.
169 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
170 NIR_PASS_V(nir
, nir_propagate_invariant
);
171 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
172 entry_point
->impl
, true, false);
173 NIR_PASS_V(nir
, nir_lower_system_values
);
175 /* Vulkan uses the separate-shader linking model */
176 nir
->info
->separate_shader
= true;
178 nir
= brw_preprocess_nir(compiler
, nir
);
180 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
182 if (stage
== MESA_SHADER_FRAGMENT
)
183 NIR_PASS_V(nir
, anv_nir_lower_input_attachments
);
188 void anv_DestroyPipeline(
190 VkPipeline _pipeline
,
191 const VkAllocationCallbacks
* pAllocator
)
193 ANV_FROM_HANDLE(anv_device
, device
, _device
);
194 ANV_FROM_HANDLE(anv_pipeline
, pipeline
, _pipeline
);
199 anv_reloc_list_finish(&pipeline
->batch_relocs
,
200 pAllocator
? pAllocator
: &device
->alloc
);
201 if (pipeline
->blend_state
.map
)
202 anv_state_pool_free(&device
->dynamic_state_pool
, pipeline
->blend_state
);
204 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
205 if (pipeline
->shaders
[s
])
206 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
209 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
212 static const uint32_t vk_to_gen_primitive_type
[] = {
213 [VK_PRIMITIVE_TOPOLOGY_POINT_LIST
] = _3DPRIM_POINTLIST
,
214 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST
] = _3DPRIM_LINELIST
,
215 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
216 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
] = _3DPRIM_TRILIST
,
217 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
218 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
219 [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
220 [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
221 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
222 [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
226 populate_sampler_prog_key(const struct gen_device_info
*devinfo
,
227 struct brw_sampler_prog_key_data
*key
)
229 /* Almost all multisampled textures are compressed. The only time when we
230 * don't compress a multisampled texture is for 16x MSAA with a surface
231 * width greater than 8k which is a bit of an edge case. Since the sampler
232 * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
233 * to tell the compiler to always assume compression.
235 key
->compressed_multisample_layout_mask
= ~0;
237 /* SkyLake added support for 16x MSAA. With this came a new message for
238 * reading from a 16x MSAA surface with compression. The new message was
239 * needed because now the MCS data is 64 bits instead of 32 or lower as is
240 * the case for 8x, 4x, and 2x. The key->msaa_16 bit-field controls which
241 * message we use. Fortunately, the 16x message works for 8x, 4x, and 2x
242 * so we can just use it unconditionally. This may not be quite as
243 * efficient but it saves us from recompiling.
245 if (devinfo
->gen
>= 9)
248 /* XXX: Handle texture swizzle on HSW- */
249 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
250 /* Assume color sampler, no swizzling. (Works for BDW+) */
251 key
->swizzles
[i
] = SWIZZLE_XYZW
;
256 populate_vs_prog_key(const struct gen_device_info
*devinfo
,
257 struct brw_vs_prog_key
*key
)
259 memset(key
, 0, sizeof(*key
));
261 populate_sampler_prog_key(devinfo
, &key
->tex
);
263 /* XXX: Handle vertex input work-arounds */
265 /* XXX: Handle sampler_prog_key */
269 populate_gs_prog_key(const struct gen_device_info
*devinfo
,
270 struct brw_gs_prog_key
*key
)
272 memset(key
, 0, sizeof(*key
));
274 populate_sampler_prog_key(devinfo
, &key
->tex
);
278 populate_wm_prog_key(const struct anv_pipeline
*pipeline
,
279 const VkGraphicsPipelineCreateInfo
*info
,
280 struct brw_wm_prog_key
*key
)
282 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
284 memset(key
, 0, sizeof(*key
));
286 populate_sampler_prog_key(devinfo
, &key
->tex
);
288 /* TODO: we could set this to 0 based on the information in nir_shader, but
289 * this function is called before spirv_to_nir. */
290 const struct brw_vue_map
*vue_map
=
291 &anv_pipeline_get_last_vue_prog_data(pipeline
)->vue_map
;
292 key
->input_slots_valid
= vue_map
->slots_valid
;
294 /* Vulkan doesn't specify a default */
295 key
->high_quality_derivatives
= false;
297 /* XXX Vulkan doesn't appear to specify */
298 key
->clamp_fragment_color
= false;
300 key
->nr_color_regions
= pipeline
->subpass
->color_count
;
302 key
->replicate_alpha
= key
->nr_color_regions
> 1 &&
303 info
->pMultisampleState
&&
304 info
->pMultisampleState
->alphaToCoverageEnable
;
306 if (info
->pMultisampleState
) {
307 /* We should probably pull this out of the shader, but it's fairly
308 * harmless to compute it and then let dead-code take care of it.
310 if (info
->pMultisampleState
->rasterizationSamples
> 1) {
311 key
->persample_interp
=
312 (info
->pMultisampleState
->minSampleShading
*
313 info
->pMultisampleState
->rasterizationSamples
) > 1;
314 key
->multisample_fbo
= true;
317 key
->frag_coord_adds_sample_pos
=
318 info
->pMultisampleState
->sampleShadingEnable
;
323 populate_cs_prog_key(const struct gen_device_info
*devinfo
,
324 struct brw_cs_prog_key
*key
)
326 memset(key
, 0, sizeof(*key
));
328 populate_sampler_prog_key(devinfo
, &key
->tex
);
332 anv_pipeline_hash_shader(struct anv_pipeline
*pipeline
,
333 struct anv_shader_module
*module
,
334 const char *entrypoint
,
335 gl_shader_stage stage
,
336 const VkSpecializationInfo
*spec_info
,
337 const void *key
, size_t key_size
,
338 unsigned char *sha1_out
)
340 struct mesa_sha1 ctx
;
342 _mesa_sha1_init(&ctx
);
343 if (stage
!= MESA_SHADER_COMPUTE
) {
344 _mesa_sha1_update(&ctx
, &pipeline
->subpass
->view_mask
,
345 sizeof(pipeline
->subpass
->view_mask
));
347 if (pipeline
->layout
) {
348 _mesa_sha1_update(&ctx
, pipeline
->layout
->sha1
,
349 sizeof(pipeline
->layout
->sha1
));
351 _mesa_sha1_update(&ctx
, module
->sha1
, sizeof(module
->sha1
));
352 _mesa_sha1_update(&ctx
, entrypoint
, strlen(entrypoint
));
353 _mesa_sha1_update(&ctx
, &stage
, sizeof(stage
));
355 _mesa_sha1_update(&ctx
, spec_info
->pMapEntries
,
356 spec_info
->mapEntryCount
* sizeof(*spec_info
->pMapEntries
));
357 _mesa_sha1_update(&ctx
, spec_info
->pData
, spec_info
->dataSize
);
359 _mesa_sha1_update(&ctx
, key
, key_size
);
360 _mesa_sha1_final(&ctx
, sha1_out
);
364 anv_pipeline_compile(struct anv_pipeline
*pipeline
,
365 struct anv_shader_module
*module
,
366 const char *entrypoint
,
367 gl_shader_stage stage
,
368 const VkSpecializationInfo
*spec_info
,
369 struct brw_stage_prog_data
*prog_data
,
370 struct anv_pipeline_bind_map
*map
)
372 nir_shader
*nir
= anv_shader_compile_to_nir(pipeline
,
373 module
, entrypoint
, stage
,
378 NIR_PASS_V(nir
, anv_nir_lower_push_constants
);
380 if (stage
!= MESA_SHADER_COMPUTE
)
381 NIR_PASS_V(nir
, anv_nir_lower_multiview
, pipeline
->subpass
->view_mask
);
383 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
385 /* Figure out the number of parameters */
386 prog_data
->nr_params
= 0;
388 if (nir
->num_uniforms
> 0) {
389 /* If the shader uses any push constants at all, we'll just give
390 * them the maximum possible number
392 assert(nir
->num_uniforms
<= MAX_PUSH_CONSTANTS_SIZE
);
393 prog_data
->nr_params
+= MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float);
396 if (nir
->info
->num_images
> 0) {
397 prog_data
->nr_params
+= nir
->info
->num_images
* BRW_IMAGE_PARAM_SIZE
;
398 pipeline
->needs_data_cache
= true;
401 if (stage
== MESA_SHADER_COMPUTE
)
402 ((struct brw_cs_prog_data
*)prog_data
)->thread_local_id_index
=
403 prog_data
->nr_params
++; /* The CS Thread ID uniform */
405 if (nir
->info
->num_ssbos
> 0)
406 pipeline
->needs_data_cache
= true;
408 if (prog_data
->nr_params
> 0) {
409 /* XXX: I think we're leaking this */
410 prog_data
->param
= (const union gl_constant_value
**)
411 malloc(prog_data
->nr_params
* sizeof(union gl_constant_value
*));
413 /* We now set the param values to be offsets into a
414 * anv_push_constant_data structure. Since the compiler doesn't
415 * actually dereference any of the gl_constant_value pointers in the
416 * params array, it doesn't really matter what we put here.
418 struct anv_push_constants
*null_data
= NULL
;
419 if (nir
->num_uniforms
> 0) {
420 /* Fill out the push constants section of the param array */
421 for (unsigned i
= 0; i
< MAX_PUSH_CONSTANTS_SIZE
/ sizeof(float); i
++)
422 prog_data
->param
[i
] = (const union gl_constant_value
*)
423 &null_data
->client_data
[i
* sizeof(float)];
427 /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
428 if (pipeline
->layout
)
429 anv_nir_apply_pipeline_layout(pipeline
, nir
, prog_data
, map
);
431 /* nir_lower_io will only handle the push constants; we need to set this
432 * to the full number of possible uniforms.
434 nir
->num_uniforms
= prog_data
->nr_params
* 4;
440 anv_fill_binding_table(struct brw_stage_prog_data
*prog_data
, unsigned bias
)
442 prog_data
->binding_table
.size_bytes
= 0;
443 prog_data
->binding_table
.texture_start
= bias
;
444 prog_data
->binding_table
.gather_texture_start
= bias
;
445 prog_data
->binding_table
.ubo_start
= bias
;
446 prog_data
->binding_table
.ssbo_start
= bias
;
447 prog_data
->binding_table
.image_start
= bias
;
450 static struct anv_shader_bin
*
451 anv_pipeline_upload_kernel(struct anv_pipeline
*pipeline
,
452 struct anv_pipeline_cache
*cache
,
453 const void *key_data
, uint32_t key_size
,
454 const void *kernel_data
, uint32_t kernel_size
,
455 const struct brw_stage_prog_data
*prog_data
,
456 uint32_t prog_data_size
,
457 const struct anv_pipeline_bind_map
*bind_map
)
460 return anv_pipeline_cache_upload_kernel(cache
, key_data
, key_size
,
461 kernel_data
, kernel_size
,
462 prog_data
, prog_data_size
,
465 return anv_shader_bin_create(pipeline
->device
, key_data
, key_size
,
466 kernel_data
, kernel_size
,
467 prog_data
, prog_data_size
,
468 prog_data
->param
, bind_map
);
474 anv_pipeline_add_compiled_stage(struct anv_pipeline
*pipeline
,
475 gl_shader_stage stage
,
476 struct anv_shader_bin
*shader
)
478 pipeline
->shaders
[stage
] = shader
;
479 pipeline
->active_stages
|= mesa_to_vk_shader_stage(stage
);
483 anv_pipeline_compile_vs(struct anv_pipeline
*pipeline
,
484 struct anv_pipeline_cache
*cache
,
485 const VkGraphicsPipelineCreateInfo
*info
,
486 struct anv_shader_module
*module
,
487 const char *entrypoint
,
488 const VkSpecializationInfo
*spec_info
)
490 const struct brw_compiler
*compiler
=
491 pipeline
->device
->instance
->physicalDevice
.compiler
;
492 struct anv_pipeline_bind_map map
;
493 struct brw_vs_prog_key key
;
494 struct anv_shader_bin
*bin
= NULL
;
495 unsigned char sha1
[20];
497 populate_vs_prog_key(&pipeline
->device
->info
, &key
);
500 anv_pipeline_hash_shader(pipeline
, module
, entrypoint
,
501 MESA_SHADER_VERTEX
, spec_info
,
502 &key
, sizeof(key
), sha1
);
503 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
507 struct brw_vs_prog_data prog_data
= { 0, };
508 struct anv_pipeline_binding surface_to_descriptor
[256];
509 struct anv_pipeline_binding sampler_to_descriptor
[256];
511 map
= (struct anv_pipeline_bind_map
) {
512 .surface_to_descriptor
= surface_to_descriptor
,
513 .sampler_to_descriptor
= sampler_to_descriptor
516 nir_shader
*nir
= anv_pipeline_compile(pipeline
, module
, entrypoint
,
517 MESA_SHADER_VERTEX
, spec_info
,
518 &prog_data
.base
.base
, &map
);
520 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
522 anv_fill_binding_table(&prog_data
.base
.base
, 0);
524 void *mem_ctx
= ralloc_context(NULL
);
526 ralloc_steal(mem_ctx
, nir
);
528 prog_data
.inputs_read
= nir
->info
->inputs_read
;
529 prog_data
.double_inputs_read
= nir
->info
->double_inputs_read
;
531 brw_compute_vue_map(&pipeline
->device
->info
,
532 &prog_data
.base
.vue_map
,
533 nir
->info
->outputs_written
,
534 nir
->info
->separate_shader
);
537 const unsigned *shader_code
=
538 brw_compile_vs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
539 NULL
, false, -1, &code_size
, NULL
);
540 if (shader_code
== NULL
) {
541 ralloc_free(mem_ctx
);
542 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
545 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
546 shader_code
, code_size
,
547 &prog_data
.base
.base
, sizeof(prog_data
),
550 ralloc_free(mem_ctx
);
551 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
554 ralloc_free(mem_ctx
);
557 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_VERTEX
, bin
);
563 merge_tess_info(struct shader_info
*tes_info
,
564 const struct shader_info
*tcs_info
)
566 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
568 * "PointMode. Controls generation of points rather than triangles
569 * or lines. This functionality defaults to disabled, and is
570 * enabled if either shader stage includes the execution mode.
572 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
573 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
574 * and OutputVertices, it says:
576 * "One mode must be set in at least one of the tessellation
579 * So, the fields can be set in either the TCS or TES, but they must
580 * agree if set in both. Our backend looks at TES, so bitwise-or in
581 * the values from the TCS.
583 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
584 tes_info
->tess
.tcs_vertices_out
== 0 ||
585 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
586 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
588 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
589 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
590 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
591 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
593 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
594 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
598 anv_pipeline_compile_tcs_tes(struct anv_pipeline
*pipeline
,
599 struct anv_pipeline_cache
*cache
,
600 const VkGraphicsPipelineCreateInfo
*info
,
601 struct anv_shader_module
*tcs_module
,
602 const char *tcs_entrypoint
,
603 const VkSpecializationInfo
*tcs_spec_info
,
604 struct anv_shader_module
*tes_module
,
605 const char *tes_entrypoint
,
606 const VkSpecializationInfo
*tes_spec_info
)
608 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
609 const struct brw_compiler
*compiler
=
610 pipeline
->device
->instance
->physicalDevice
.compiler
;
611 struct anv_pipeline_bind_map tcs_map
;
612 struct anv_pipeline_bind_map tes_map
;
613 struct brw_tcs_prog_key tcs_key
= { 0, };
614 struct brw_tes_prog_key tes_key
= { 0, };
615 struct anv_shader_bin
*tcs_bin
= NULL
;
616 struct anv_shader_bin
*tes_bin
= NULL
;
617 unsigned char tcs_sha1
[40];
618 unsigned char tes_sha1
[40];
620 populate_sampler_prog_key(&pipeline
->device
->info
, &tcs_key
.tex
);
621 populate_sampler_prog_key(&pipeline
->device
->info
, &tes_key
.tex
);
622 tcs_key
.input_vertices
= info
->pTessellationState
->patchControlPoints
;
625 anv_pipeline_hash_shader(pipeline
, tcs_module
, tcs_entrypoint
,
626 MESA_SHADER_TESS_CTRL
, tcs_spec_info
,
627 &tcs_key
, sizeof(tcs_key
), tcs_sha1
);
628 anv_pipeline_hash_shader(pipeline
, tes_module
, tes_entrypoint
,
629 MESA_SHADER_TESS_EVAL
, tes_spec_info
,
630 &tes_key
, sizeof(tes_key
), tes_sha1
);
631 memcpy(&tcs_sha1
[20], tes_sha1
, 20);
632 memcpy(&tes_sha1
[20], tcs_sha1
, 20);
633 tcs_bin
= anv_pipeline_cache_search(cache
, tcs_sha1
, sizeof(tcs_sha1
));
634 tes_bin
= anv_pipeline_cache_search(cache
, tes_sha1
, sizeof(tes_sha1
));
637 if (tcs_bin
== NULL
|| tes_bin
== NULL
) {
638 struct brw_tcs_prog_data tcs_prog_data
= { 0, };
639 struct brw_tes_prog_data tes_prog_data
= { 0, };
640 struct anv_pipeline_binding tcs_surface_to_descriptor
[256];
641 struct anv_pipeline_binding tcs_sampler_to_descriptor
[256];
642 struct anv_pipeline_binding tes_surface_to_descriptor
[256];
643 struct anv_pipeline_binding tes_sampler_to_descriptor
[256];
645 tcs_map
= (struct anv_pipeline_bind_map
) {
646 .surface_to_descriptor
= tcs_surface_to_descriptor
,
647 .sampler_to_descriptor
= tcs_sampler_to_descriptor
649 tes_map
= (struct anv_pipeline_bind_map
) {
650 .surface_to_descriptor
= tes_surface_to_descriptor
,
651 .sampler_to_descriptor
= tes_sampler_to_descriptor
654 nir_shader
*tcs_nir
=
655 anv_pipeline_compile(pipeline
, tcs_module
, tcs_entrypoint
,
656 MESA_SHADER_TESS_CTRL
, tcs_spec_info
,
657 &tcs_prog_data
.base
.base
, &tcs_map
);
658 nir_shader
*tes_nir
=
659 anv_pipeline_compile(pipeline
, tes_module
, tes_entrypoint
,
660 MESA_SHADER_TESS_EVAL
, tes_spec_info
,
661 &tes_prog_data
.base
.base
, &tes_map
);
662 if (tcs_nir
== NULL
|| tes_nir
== NULL
)
663 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
665 nir_lower_tes_patch_vertices(tes_nir
,
666 tcs_nir
->info
->tess
.tcs_vertices_out
);
668 /* Copy TCS info into the TES info */
669 merge_tess_info(tes_nir
->info
, tcs_nir
->info
);
671 anv_fill_binding_table(&tcs_prog_data
.base
.base
, 0);
672 anv_fill_binding_table(&tes_prog_data
.base
.base
, 0);
674 void *mem_ctx
= ralloc_context(NULL
);
676 ralloc_steal(mem_ctx
, tcs_nir
);
677 ralloc_steal(mem_ctx
, tes_nir
);
679 /* Whacking the key after cache lookup is a bit sketchy, but all of
680 * this comes from the SPIR-V, which is part of the hash used for the
681 * pipeline cache. So it should be safe.
683 tcs_key
.tes_primitive_mode
= tes_nir
->info
->tess
.primitive_mode
;
684 tcs_key
.outputs_written
= tcs_nir
->info
->outputs_written
;
685 tcs_key
.patch_outputs_written
= tcs_nir
->info
->patch_outputs_written
;
686 tcs_key
.quads_workaround
=
688 tes_nir
->info
->tess
.primitive_mode
== 7 /* GL_QUADS */ &&
689 tes_nir
->info
->tess
.spacing
== TESS_SPACING_EQUAL
;
691 tes_key
.inputs_read
= tcs_key
.outputs_written
;
692 tes_key
.patch_inputs_read
= tcs_key
.patch_outputs_written
;
695 const int shader_time_index
= -1;
696 const unsigned *shader_code
;
699 brw_compile_tcs(compiler
, NULL
, mem_ctx
, &tcs_key
, &tcs_prog_data
,
700 tcs_nir
, shader_time_index
, &code_size
, NULL
);
701 if (shader_code
== NULL
) {
702 ralloc_free(mem_ctx
);
703 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
706 tcs_bin
= anv_pipeline_upload_kernel(pipeline
, cache
,
707 tcs_sha1
, sizeof(tcs_sha1
),
708 shader_code
, code_size
,
709 &tcs_prog_data
.base
.base
,
710 sizeof(tcs_prog_data
),
713 ralloc_free(mem_ctx
);
714 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
718 brw_compile_tes(compiler
, NULL
, mem_ctx
, &tes_key
,
719 &tcs_prog_data
.base
.vue_map
, &tes_prog_data
, tes_nir
,
720 NULL
, shader_time_index
, &code_size
, NULL
);
721 if (shader_code
== NULL
) {
722 ralloc_free(mem_ctx
);
723 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
726 tes_bin
= anv_pipeline_upload_kernel(pipeline
, cache
,
727 tes_sha1
, sizeof(tes_sha1
),
728 shader_code
, code_size
,
729 &tes_prog_data
.base
.base
,
730 sizeof(tes_prog_data
),
733 ralloc_free(mem_ctx
);
734 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
737 ralloc_free(mem_ctx
);
740 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_TESS_CTRL
, tcs_bin
);
741 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_TESS_EVAL
, tes_bin
);
747 anv_pipeline_compile_gs(struct anv_pipeline
*pipeline
,
748 struct anv_pipeline_cache
*cache
,
749 const VkGraphicsPipelineCreateInfo
*info
,
750 struct anv_shader_module
*module
,
751 const char *entrypoint
,
752 const VkSpecializationInfo
*spec_info
)
754 const struct brw_compiler
*compiler
=
755 pipeline
->device
->instance
->physicalDevice
.compiler
;
756 struct anv_pipeline_bind_map map
;
757 struct brw_gs_prog_key key
;
758 struct anv_shader_bin
*bin
= NULL
;
759 unsigned char sha1
[20];
761 populate_gs_prog_key(&pipeline
->device
->info
, &key
);
764 anv_pipeline_hash_shader(pipeline
, module
, entrypoint
,
765 MESA_SHADER_GEOMETRY
, spec_info
,
766 &key
, sizeof(key
), sha1
);
767 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
771 struct brw_gs_prog_data prog_data
= { 0, };
772 struct anv_pipeline_binding surface_to_descriptor
[256];
773 struct anv_pipeline_binding sampler_to_descriptor
[256];
775 map
= (struct anv_pipeline_bind_map
) {
776 .surface_to_descriptor
= surface_to_descriptor
,
777 .sampler_to_descriptor
= sampler_to_descriptor
780 nir_shader
*nir
= anv_pipeline_compile(pipeline
, module
, entrypoint
,
781 MESA_SHADER_GEOMETRY
, spec_info
,
782 &prog_data
.base
.base
, &map
);
784 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
786 anv_fill_binding_table(&prog_data
.base
.base
, 0);
788 void *mem_ctx
= ralloc_context(NULL
);
790 ralloc_steal(mem_ctx
, nir
);
792 brw_compute_vue_map(&pipeline
->device
->info
,
793 &prog_data
.base
.vue_map
,
794 nir
->info
->outputs_written
,
795 nir
->info
->separate_shader
);
798 const unsigned *shader_code
=
799 brw_compile_gs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
800 NULL
, -1, &code_size
, NULL
);
801 if (shader_code
== NULL
) {
802 ralloc_free(mem_ctx
);
803 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
807 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
808 shader_code
, code_size
,
809 &prog_data
.base
.base
, sizeof(prog_data
),
812 ralloc_free(mem_ctx
);
813 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
816 ralloc_free(mem_ctx
);
819 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_GEOMETRY
, bin
);
825 anv_pipeline_compile_fs(struct anv_pipeline
*pipeline
,
826 struct anv_pipeline_cache
*cache
,
827 const VkGraphicsPipelineCreateInfo
*info
,
828 struct anv_shader_module
*module
,
829 const char *entrypoint
,
830 const VkSpecializationInfo
*spec_info
)
832 const struct brw_compiler
*compiler
=
833 pipeline
->device
->instance
->physicalDevice
.compiler
;
834 struct anv_pipeline_bind_map map
;
835 struct brw_wm_prog_key key
;
836 struct anv_shader_bin
*bin
= NULL
;
837 unsigned char sha1
[20];
839 populate_wm_prog_key(pipeline
, info
, &key
);
842 anv_pipeline_hash_shader(pipeline
, module
, entrypoint
,
843 MESA_SHADER_FRAGMENT
, spec_info
,
844 &key
, sizeof(key
), sha1
);
845 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
849 struct brw_wm_prog_data prog_data
= { 0, };
850 struct anv_pipeline_binding surface_to_descriptor
[256];
851 struct anv_pipeline_binding sampler_to_descriptor
[256];
853 map
= (struct anv_pipeline_bind_map
) {
854 .surface_to_descriptor
= surface_to_descriptor
+ 8,
855 .sampler_to_descriptor
= sampler_to_descriptor
858 nir_shader
*nir
= anv_pipeline_compile(pipeline
, module
, entrypoint
,
859 MESA_SHADER_FRAGMENT
, spec_info
,
860 &prog_data
.base
, &map
);
862 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
864 unsigned num_rts
= 0;
865 struct anv_pipeline_binding rt_bindings
[8];
866 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
867 nir_foreach_variable_safe(var
, &nir
->outputs
) {
868 if (var
->data
.location
< FRAG_RESULT_DATA0
)
871 unsigned rt
= var
->data
.location
- FRAG_RESULT_DATA0
;
872 if (rt
>= key
.nr_color_regions
) {
873 /* Out-of-bounds, throw it away */
874 var
->data
.mode
= nir_var_local
;
875 exec_node_remove(&var
->node
);
876 exec_list_push_tail(&impl
->locals
, &var
->node
);
880 /* Give it a new, compacted, location */
881 var
->data
.location
= FRAG_RESULT_DATA0
+ num_rts
;
884 glsl_type_is_array(var
->type
) ? glsl_get_length(var
->type
) : 1;
885 assert(num_rts
+ array_len
<= 8);
887 for (unsigned i
= 0; i
< array_len
; i
++) {
888 rt_bindings
[num_rts
+ i
] = (struct anv_pipeline_binding
) {
889 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
895 num_rts
+= array_len
;
899 /* If we have no render targets, we need a null render target */
900 rt_bindings
[0] = (struct anv_pipeline_binding
) {
901 .set
= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
,
908 assert(num_rts
<= 8);
909 map
.surface_to_descriptor
-= num_rts
;
910 map
.surface_count
+= num_rts
;
911 assert(map
.surface_count
<= 256);
912 memcpy(map
.surface_to_descriptor
, rt_bindings
,
913 num_rts
* sizeof(*rt_bindings
));
915 anv_fill_binding_table(&prog_data
.base
, num_rts
);
917 void *mem_ctx
= ralloc_context(NULL
);
919 ralloc_steal(mem_ctx
, nir
);
922 const unsigned *shader_code
=
923 brw_compile_fs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
924 NULL
, -1, -1, true, false, NULL
, &code_size
, NULL
);
925 if (shader_code
== NULL
) {
926 ralloc_free(mem_ctx
);
927 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
930 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
931 shader_code
, code_size
,
932 &prog_data
.base
, sizeof(prog_data
),
935 ralloc_free(mem_ctx
);
936 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
939 ralloc_free(mem_ctx
);
942 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_FRAGMENT
, bin
);
948 anv_pipeline_compile_cs(struct anv_pipeline
*pipeline
,
949 struct anv_pipeline_cache
*cache
,
950 const VkComputePipelineCreateInfo
*info
,
951 struct anv_shader_module
*module
,
952 const char *entrypoint
,
953 const VkSpecializationInfo
*spec_info
)
955 const struct brw_compiler
*compiler
=
956 pipeline
->device
->instance
->physicalDevice
.compiler
;
957 struct anv_pipeline_bind_map map
;
958 struct brw_cs_prog_key key
;
959 struct anv_shader_bin
*bin
= NULL
;
960 unsigned char sha1
[20];
962 populate_cs_prog_key(&pipeline
->device
->info
, &key
);
965 anv_pipeline_hash_shader(pipeline
, module
, entrypoint
,
966 MESA_SHADER_COMPUTE
, spec_info
,
967 &key
, sizeof(key
), sha1
);
968 bin
= anv_pipeline_cache_search(cache
, sha1
, 20);
972 struct brw_cs_prog_data prog_data
= { 0, };
973 struct anv_pipeline_binding surface_to_descriptor
[256];
974 struct anv_pipeline_binding sampler_to_descriptor
[256];
976 map
= (struct anv_pipeline_bind_map
) {
977 .surface_to_descriptor
= surface_to_descriptor
,
978 .sampler_to_descriptor
= sampler_to_descriptor
981 nir_shader
*nir
= anv_pipeline_compile(pipeline
, module
, entrypoint
,
982 MESA_SHADER_COMPUTE
, spec_info
,
983 &prog_data
.base
, &map
);
985 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
987 anv_fill_binding_table(&prog_data
.base
, 1);
989 void *mem_ctx
= ralloc_context(NULL
);
991 ralloc_steal(mem_ctx
, nir
);
994 const unsigned *shader_code
=
995 brw_compile_cs(compiler
, NULL
, mem_ctx
, &key
, &prog_data
, nir
,
996 -1, &code_size
, NULL
);
997 if (shader_code
== NULL
) {
998 ralloc_free(mem_ctx
);
999 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1002 bin
= anv_pipeline_upload_kernel(pipeline
, cache
, sha1
, 20,
1003 shader_code
, code_size
,
1004 &prog_data
.base
, sizeof(prog_data
),
1007 ralloc_free(mem_ctx
);
1008 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1011 ralloc_free(mem_ctx
);
1014 anv_pipeline_add_compiled_stage(pipeline
, MESA_SHADER_COMPUTE
, bin
);
1020 * Copy pipeline state not marked as dynamic.
1021 * Dynamic state is pipeline state which hasn't been provided at pipeline
1022 * creation time, but is dynamically provided afterwards using various
1023 * vkCmdSet* functions.
1025 * The set of state considered "non_dynamic" is determined by the pieces of
1026 * state that have their corresponding VkDynamicState enums omitted from
1027 * VkPipelineDynamicStateCreateInfo::pDynamicStates.
1029 * @param[out] pipeline Destination non_dynamic state.
1030 * @param[in] pCreateInfo Source of non_dynamic state to be copied.
1033 copy_non_dynamic_state(struct anv_pipeline
*pipeline
,
1034 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1036 anv_cmd_dirty_mask_t states
= ANV_CMD_DIRTY_DYNAMIC_ALL
;
1037 struct anv_subpass
*subpass
= pipeline
->subpass
;
1039 pipeline
->dynamic_state
= default_dynamic_state
;
1041 if (pCreateInfo
->pDynamicState
) {
1042 /* Remove all of the states that are marked as dynamic */
1043 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1044 for (uint32_t s
= 0; s
< count
; s
++)
1045 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1048 struct anv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1050 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1052 * pViewportState is [...] NULL if the pipeline
1053 * has rasterization disabled.
1055 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1056 assert(pCreateInfo
->pViewportState
);
1058 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1059 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1060 typed_memcpy(dynamic
->viewport
.viewports
,
1061 pCreateInfo
->pViewportState
->pViewports
,
1062 pCreateInfo
->pViewportState
->viewportCount
);
1065 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1066 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1067 typed_memcpy(dynamic
->scissor
.scissors
,
1068 pCreateInfo
->pViewportState
->pScissors
,
1069 pCreateInfo
->pViewportState
->scissorCount
);
1073 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1074 assert(pCreateInfo
->pRasterizationState
);
1075 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1078 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1079 assert(pCreateInfo
->pRasterizationState
);
1080 dynamic
->depth_bias
.bias
=
1081 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1082 dynamic
->depth_bias
.clamp
=
1083 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1084 dynamic
->depth_bias
.slope
=
1085 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1088 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1090 * pColorBlendState is [...] NULL if the pipeline has rasterization
1091 * disabled or if the subpass of the render pass the pipeline is
1092 * created against does not use any color attachments.
1094 bool uses_color_att
= false;
1095 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1096 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1097 uses_color_att
= true;
1102 if (uses_color_att
&&
1103 !pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1104 assert(pCreateInfo
->pColorBlendState
);
1106 if (states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
))
1107 typed_memcpy(dynamic
->blend_constants
,
1108 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1111 /* If there is no depthstencil attachment, then don't read
1112 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1113 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1114 * no need to override the depthstencil defaults in
1115 * anv_pipeline::dynamic_state when there is no depthstencil attachment.
1117 * Section 9.2 of the Vulkan 1.0.15 spec says:
1119 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1120 * disabled or if the subpass of the render pass the pipeline is created
1121 * against does not use a depth/stencil attachment.
1123 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1124 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1125 assert(pCreateInfo
->pDepthStencilState
);
1127 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1128 dynamic
->depth_bounds
.min
=
1129 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1130 dynamic
->depth_bounds
.max
=
1131 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1134 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1135 dynamic
->stencil_compare_mask
.front
=
1136 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1137 dynamic
->stencil_compare_mask
.back
=
1138 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1141 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1142 dynamic
->stencil_write_mask
.front
=
1143 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1144 dynamic
->stencil_write_mask
.back
=
1145 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1148 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1149 dynamic
->stencil_reference
.front
=
1150 pCreateInfo
->pDepthStencilState
->front
.reference
;
1151 dynamic
->stencil_reference
.back
=
1152 pCreateInfo
->pDepthStencilState
->back
.reference
;
1156 pipeline
->dynamic_state_mask
= states
;
1160 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo
*info
)
1163 struct anv_render_pass
*renderpass
= NULL
;
1164 struct anv_subpass
*subpass
= NULL
;
1166 /* Assert that all required members of VkGraphicsPipelineCreateInfo are
1167 * present. See the Vulkan 1.0.28 spec, Section 9.2 Graphics Pipelines.
1169 assert(info
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1171 renderpass
= anv_render_pass_from_handle(info
->renderPass
);
1174 assert(info
->subpass
< renderpass
->subpass_count
);
1175 subpass
= &renderpass
->subpasses
[info
->subpass
];
1177 assert(info
->stageCount
>= 1);
1178 assert(info
->pVertexInputState
);
1179 assert(info
->pInputAssemblyState
);
1180 assert(info
->pRasterizationState
);
1181 if (!info
->pRasterizationState
->rasterizerDiscardEnable
) {
1182 assert(info
->pViewportState
);
1183 assert(info
->pMultisampleState
);
1185 if (subpass
&& subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
)
1186 assert(info
->pDepthStencilState
);
1188 if (subpass
&& subpass
->color_count
> 0)
1189 assert(info
->pColorBlendState
);
1192 for (uint32_t i
= 0; i
< info
->stageCount
; ++i
) {
1193 switch (info
->pStages
[i
].stage
) {
1194 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT
:
1195 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
:
1196 assert(info
->pTessellationState
);
1206 * Calculate the desired L3 partitioning based on the current state of the
1207 * pipeline. For now this simply returns the conservative defaults calculated
1208 * by get_default_l3_weights(), but we could probably do better by gathering
1209 * more statistics from the pipeline state (e.g. guess of expected URB usage
1210 * and bound surfaces), or by using feed-back from performance counters.
1213 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
)
1215 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1217 const struct gen_l3_weights w
=
1218 gen_get_default_l3_weights(devinfo
, pipeline
->needs_data_cache
, needs_slm
);
1220 pipeline
->urb
.l3_config
= gen_get_l3_config(devinfo
, w
);
1221 pipeline
->urb
.total_size
=
1222 gen_get_l3_config_urb_size(devinfo
, pipeline
->urb
.l3_config
);
1226 anv_pipeline_init(struct anv_pipeline
*pipeline
,
1227 struct anv_device
*device
,
1228 struct anv_pipeline_cache
*cache
,
1229 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1230 const VkAllocationCallbacks
*alloc
)
1234 anv_pipeline_validate_create_info(pCreateInfo
);
1237 alloc
= &device
->alloc
;
1239 pipeline
->device
= device
;
1241 ANV_FROM_HANDLE(anv_render_pass
, render_pass
, pCreateInfo
->renderPass
);
1242 assert(pCreateInfo
->subpass
< render_pass
->subpass_count
);
1243 pipeline
->subpass
= &render_pass
->subpasses
[pCreateInfo
->subpass
];
1245 pipeline
->layout
= anv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1247 result
= anv_reloc_list_init(&pipeline
->batch_relocs
, alloc
);
1248 if (result
!= VK_SUCCESS
)
1251 pipeline
->batch
.alloc
= alloc
;
1252 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1253 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1254 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1255 pipeline
->batch
.status
= VK_SUCCESS
;
1257 copy_non_dynamic_state(pipeline
, pCreateInfo
);
1258 pipeline
->depth_clamp_enable
= pCreateInfo
->pRasterizationState
&&
1259 pCreateInfo
->pRasterizationState
->depthClampEnable
;
1261 pipeline
->sample_shading_enable
= pCreateInfo
->pMultisampleState
&&
1262 pCreateInfo
->pMultisampleState
->sampleShadingEnable
;
1264 pipeline
->needs_data_cache
= false;
1266 /* When we free the pipeline, we detect stages based on the NULL status
1267 * of various prog_data pointers. Make them NULL by default.
1269 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1271 pipeline
->active_stages
= 0;
1273 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
1274 struct anv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1275 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
1276 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
1277 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
1278 modules
[stage
] = anv_shader_module_from_handle(pStages
[stage
]->module
);
1281 if (modules
[MESA_SHADER_VERTEX
]) {
1282 result
= anv_pipeline_compile_vs(pipeline
, cache
, pCreateInfo
,
1283 modules
[MESA_SHADER_VERTEX
],
1284 pStages
[MESA_SHADER_VERTEX
]->pName
,
1285 pStages
[MESA_SHADER_VERTEX
]->pSpecializationInfo
);
1286 if (result
!= VK_SUCCESS
)
1290 if (modules
[MESA_SHADER_TESS_EVAL
]) {
1291 anv_pipeline_compile_tcs_tes(pipeline
, cache
, pCreateInfo
,
1292 modules
[MESA_SHADER_TESS_CTRL
],
1293 pStages
[MESA_SHADER_TESS_CTRL
]->pName
,
1294 pStages
[MESA_SHADER_TESS_CTRL
]->pSpecializationInfo
,
1295 modules
[MESA_SHADER_TESS_EVAL
],
1296 pStages
[MESA_SHADER_TESS_EVAL
]->pName
,
1297 pStages
[MESA_SHADER_TESS_EVAL
]->pSpecializationInfo
);
1300 if (modules
[MESA_SHADER_GEOMETRY
]) {
1301 result
= anv_pipeline_compile_gs(pipeline
, cache
, pCreateInfo
,
1302 modules
[MESA_SHADER_GEOMETRY
],
1303 pStages
[MESA_SHADER_GEOMETRY
]->pName
,
1304 pStages
[MESA_SHADER_GEOMETRY
]->pSpecializationInfo
);
1305 if (result
!= VK_SUCCESS
)
1309 if (modules
[MESA_SHADER_FRAGMENT
]) {
1310 result
= anv_pipeline_compile_fs(pipeline
, cache
, pCreateInfo
,
1311 modules
[MESA_SHADER_FRAGMENT
],
1312 pStages
[MESA_SHADER_FRAGMENT
]->pName
,
1313 pStages
[MESA_SHADER_FRAGMENT
]->pSpecializationInfo
);
1314 if (result
!= VK_SUCCESS
)
1318 assert(pipeline
->active_stages
& VK_SHADER_STAGE_VERTEX_BIT
);
1320 anv_pipeline_setup_l3_config(pipeline
, false);
1322 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
1323 pCreateInfo
->pVertexInputState
;
1325 const uint64_t inputs_read
= get_vs_prog_data(pipeline
)->inputs_read
;
1327 pipeline
->vb_used
= 0;
1328 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
1329 const VkVertexInputAttributeDescription
*desc
=
1330 &vi_info
->pVertexAttributeDescriptions
[i
];
1332 if (inputs_read
& (1 << (VERT_ATTRIB_GENERIC0
+ desc
->location
)))
1333 pipeline
->vb_used
|= 1 << desc
->binding
;
1336 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
1337 const VkVertexInputBindingDescription
*desc
=
1338 &vi_info
->pVertexBindingDescriptions
[i
];
1340 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
1342 /* Step rate is programmed per vertex element (attribute), not
1343 * binding. Set up a map of which bindings step per instance, for
1344 * reference by vertex element setup. */
1345 switch (desc
->inputRate
) {
1347 case VK_VERTEX_INPUT_RATE_VERTEX
:
1348 pipeline
->instancing_enable
[desc
->binding
] = false;
1350 case VK_VERTEX_INPUT_RATE_INSTANCE
:
1351 pipeline
->instancing_enable
[desc
->binding
] = true;
1356 const VkPipelineInputAssemblyStateCreateInfo
*ia_info
=
1357 pCreateInfo
->pInputAssemblyState
;
1358 const VkPipelineTessellationStateCreateInfo
*tess_info
=
1359 pCreateInfo
->pTessellationState
;
1360 pipeline
->primitive_restart
= ia_info
->primitiveRestartEnable
;
1362 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
1363 pipeline
->topology
= _3DPRIM_PATCHLIST(tess_info
->patchControlPoints
);
1365 pipeline
->topology
= vk_to_gen_primitive_type
[ia_info
->topology
];
1370 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1371 if (pipeline
->shaders
[s
])
1372 anv_shader_bin_unref(device
, pipeline
->shaders
[s
]);
1375 anv_reloc_list_finish(&pipeline
->batch_relocs
, alloc
);