2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "util/mesa-sha1.h"
29 #include "util/u_atomic.h"
30 #include "radv_debug.h"
31 #include "radv_private.h"
32 #include "radv_shader.h"
33 #include "radv_shader_helper.h"
34 #include "radv_shader_args.h"
36 #include "nir/nir_builder.h"
37 #include "spirv/nir_spirv.h"
39 #include <llvm-c/Core.h>
40 #include <llvm-c/TargetMachine.h>
41 #include <llvm-c/Support.h>
44 #include "ac_binary.h"
45 #include "ac_llvm_util.h"
46 #include "ac_nir_to_llvm.h"
48 #include "vk_format.h"
49 #include "util/debug.h"
50 #include "ac_exp_param.h"
52 #include "aco_interface.h"
54 #include "util/string_buffer.h"
56 static const struct nir_shader_compiler_options nir_options_llvm
= {
57 .vertex_id_zero_based
= true,
62 .lower_device_index_to_zero
= true,
66 .lower_bitfield_insert_to_bitfield_select
= true,
67 .lower_bitfield_extract
= true,
69 .lower_pack_snorm_2x16
= true,
70 .lower_pack_snorm_4x8
= true,
71 .lower_pack_unorm_2x16
= true,
72 .lower_pack_unorm_4x8
= true,
73 .lower_unpack_snorm_2x16
= true,
74 .lower_unpack_snorm_4x8
= true,
75 .lower_unpack_unorm_2x16
= true,
76 .lower_unpack_unorm_4x8
= true,
77 .lower_extract_byte
= true,
78 .lower_extract_word
= true,
81 .lower_mul_2x32_64
= true,
83 .max_unroll_iterations
= 32,
84 .use_interpolated_input_intrinsics
= true,
85 /* nir_lower_int64() isn't actually called for the LLVM backend, but
86 * this helps the loop unrolling heuristics. */
87 .lower_int64_options
= nir_lower_imul64
|
88 nir_lower_imul_high64
|
89 nir_lower_imul_2x32_64
|
95 static const struct nir_shader_compiler_options nir_options_aco
= {
96 .vertex_id_zero_based
= true,
100 .lower_flrp64
= true,
101 .lower_device_index_to_zero
= true,
104 .lower_bitfield_insert_to_bitfield_select
= true,
105 .lower_bitfield_extract
= true,
106 .lower_pack_snorm_2x16
= true,
107 .lower_pack_snorm_4x8
= true,
108 .lower_pack_unorm_2x16
= true,
109 .lower_pack_unorm_4x8
= true,
110 .lower_unpack_snorm_2x16
= true,
111 .lower_unpack_snorm_4x8
= true,
112 .lower_unpack_unorm_2x16
= true,
113 .lower_unpack_unorm_4x8
= true,
114 .lower_unpack_half_2x16
= true,
115 .lower_extract_byte
= true,
116 .lower_extract_word
= true,
119 .lower_mul_2x32_64
= true,
120 .lower_rotate
= true,
121 .max_unroll_iterations
= 32,
122 .use_interpolated_input_intrinsics
= true,
123 .lower_int64_options
= nir_lower_imul64
|
124 nir_lower_imul_high64
|
125 nir_lower_imul_2x32_64
|
133 radv_can_dump_shader(struct radv_device
*device
,
134 struct radv_shader_module
*module
,
135 bool is_gs_copy_shader
)
137 if (!(device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SHADERS
))
140 return !module
->nir
||
141 (device
->instance
->debug_flags
& RADV_DEBUG_DUMP_META_SHADERS
);
143 return is_gs_copy_shader
;
147 radv_can_dump_shader_stats(struct radv_device
*device
,
148 struct radv_shader_module
*module
)
150 /* Only dump non-meta shader stats. */
151 return device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SHADER_STATS
&&
152 module
&& !module
->nir
;
155 unsigned shader_io_get_unique_index(gl_varying_slot slot
)
157 /* handle patch indices separate */
158 if (slot
== VARYING_SLOT_TESS_LEVEL_OUTER
)
160 if (slot
== VARYING_SLOT_TESS_LEVEL_INNER
)
162 if (slot
>= VARYING_SLOT_PATCH0
&& slot
<= VARYING_SLOT_TESS_MAX
)
163 return 2 + (slot
- VARYING_SLOT_PATCH0
);
164 if (slot
== VARYING_SLOT_POS
)
166 if (slot
== VARYING_SLOT_PSIZ
)
168 if (slot
== VARYING_SLOT_CLIP_DIST0
)
170 if (slot
== VARYING_SLOT_CLIP_DIST1
)
172 /* 3 is reserved for clip dist as well */
173 if (slot
>= VARYING_SLOT_VAR0
&& slot
<= VARYING_SLOT_VAR31
)
174 return 4 + (slot
- VARYING_SLOT_VAR0
);
175 unreachable("illegal slot in get unique index\n");
178 VkResult
radv_CreateShaderModule(
180 const VkShaderModuleCreateInfo
* pCreateInfo
,
181 const VkAllocationCallbacks
* pAllocator
,
182 VkShaderModule
* pShaderModule
)
184 RADV_FROM_HANDLE(radv_device
, device
, _device
);
185 struct radv_shader_module
*module
;
187 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO
);
188 assert(pCreateInfo
->flags
== 0);
190 module
= vk_alloc2(&device
->alloc
, pAllocator
,
191 sizeof(*module
) + pCreateInfo
->codeSize
, 8,
192 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
194 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
197 module
->size
= pCreateInfo
->codeSize
;
198 memcpy(module
->data
, pCreateInfo
->pCode
, module
->size
);
200 _mesa_sha1_compute(module
->data
, module
->size
, module
->sha1
);
202 *pShaderModule
= radv_shader_module_to_handle(module
);
207 void radv_DestroyShaderModule(
209 VkShaderModule _module
,
210 const VkAllocationCallbacks
* pAllocator
)
212 RADV_FROM_HANDLE(radv_device
, device
, _device
);
213 RADV_FROM_HANDLE(radv_shader_module
, module
, _module
);
218 vk_free2(&device
->alloc
, pAllocator
, module
);
222 radv_optimize_nir(struct nir_shader
*shader
, bool optimize_conservatively
,
226 unsigned lower_flrp
=
227 (shader
->options
->lower_flrp16
? 16 : 0) |
228 (shader
->options
->lower_flrp32
? 32 : 0) |
229 (shader
->options
->lower_flrp64
? 64 : 0);
234 NIR_PASS(progress
, shader
, nir_split_array_vars
, nir_var_function_temp
);
235 NIR_PASS(progress
, shader
, nir_shrink_vec_array_vars
, nir_var_function_temp
);
237 NIR_PASS_V(shader
, nir_lower_vars_to_ssa
);
238 NIR_PASS_V(shader
, nir_lower_pack
);
241 /* Only run this pass in the first call to
242 * radv_optimize_nir. Later calls assume that we've
243 * lowered away any copy_deref instructions and we
244 * don't want to introduce any more.
246 NIR_PASS(progress
, shader
, nir_opt_find_array_copies
);
249 NIR_PASS(progress
, shader
, nir_opt_copy_prop_vars
);
250 NIR_PASS(progress
, shader
, nir_opt_dead_write_vars
);
251 NIR_PASS(progress
, shader
, nir_remove_dead_variables
,
252 nir_var_function_temp
| nir_var_shader_in
| nir_var_shader_out
);
254 NIR_PASS_V(shader
, nir_lower_alu_to_scalar
, NULL
, NULL
);
255 NIR_PASS_V(shader
, nir_lower_phis_to_scalar
);
257 NIR_PASS(progress
, shader
, nir_copy_prop
);
258 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
259 NIR_PASS(progress
, shader
, nir_opt_dce
);
260 if (nir_opt_trivial_continues(shader
)) {
262 NIR_PASS(progress
, shader
, nir_copy_prop
);
263 NIR_PASS(progress
, shader
, nir_opt_remove_phis
);
264 NIR_PASS(progress
, shader
, nir_opt_dce
);
266 NIR_PASS(progress
, shader
, nir_opt_if
, true);
267 NIR_PASS(progress
, shader
, nir_opt_dead_cf
);
268 NIR_PASS(progress
, shader
, nir_opt_cse
);
269 NIR_PASS(progress
, shader
, nir_opt_peephole_select
, 8, true, true);
270 NIR_PASS(progress
, shader
, nir_opt_constant_folding
);
271 NIR_PASS(progress
, shader
, nir_opt_algebraic
);
273 if (lower_flrp
!= 0) {
274 bool lower_flrp_progress
= false;
275 NIR_PASS(lower_flrp_progress
,
279 false /* always_precise */,
280 shader
->options
->lower_ffma
);
281 if (lower_flrp_progress
) {
282 NIR_PASS(progress
, shader
,
283 nir_opt_constant_folding
);
287 /* Nothing should rematerialize any flrps, so we only
288 * need to do this lowering once.
293 NIR_PASS(progress
, shader
, nir_opt_undef
);
294 if (shader
->options
->max_unroll_iterations
) {
295 NIR_PASS(progress
, shader
, nir_opt_loop_unroll
, 0);
297 } while (progress
&& !optimize_conservatively
);
299 NIR_PASS(progress
, shader
, nir_opt_conditional_discard
);
300 NIR_PASS(progress
, shader
, nir_opt_shrink_load
);
301 NIR_PASS(progress
, shader
, nir_opt_move
, nir_move_load_ubo
);
305 shared_var_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
307 assert(glsl_type_is_vector_or_scalar(type
));
309 uint32_t comp_size
= glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
310 unsigned length
= glsl_get_vector_elements(type
);
311 *size
= comp_size
* length
,
316 radv_shader_compile_to_nir(struct radv_device
*device
,
317 struct radv_shader_module
*module
,
318 const char *entrypoint_name
,
319 gl_shader_stage stage
,
320 const VkSpecializationInfo
*spec_info
,
321 const VkPipelineCreateFlags flags
,
322 const struct radv_pipeline_layout
*layout
,
326 const nir_shader_compiler_options
*nir_options
= use_aco
? &nir_options_aco
:
329 /* Some things such as our meta clear/blit code will give us a NIR
330 * shader directly. In that case, we just ignore the SPIR-V entirely
331 * and just use the NIR shader */
333 nir
->options
= nir_options
;
334 nir_validate_shader(nir
, "in internal shader");
336 assert(exec_list_length(&nir
->functions
) == 1);
338 uint32_t *spirv
= (uint32_t *) module
->data
;
339 assert(module
->size
% 4 == 0);
341 if (device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SPIRV
)
342 radv_print_spirv(module
->data
, module
->size
, stderr
);
344 uint32_t num_spec_entries
= 0;
345 struct nir_spirv_specialization
*spec_entries
= NULL
;
346 if (spec_info
&& spec_info
->mapEntryCount
> 0) {
347 num_spec_entries
= spec_info
->mapEntryCount
;
348 spec_entries
= malloc(num_spec_entries
* sizeof(*spec_entries
));
349 for (uint32_t i
= 0; i
< num_spec_entries
; i
++) {
350 VkSpecializationMapEntry entry
= spec_info
->pMapEntries
[i
];
351 const void *data
= spec_info
->pData
+ entry
.offset
;
352 assert(data
+ entry
.size
<= spec_info
->pData
+ spec_info
->dataSize
);
354 spec_entries
[i
].id
= spec_info
->pMapEntries
[i
].constantID
;
355 if (spec_info
->dataSize
== 8)
356 spec_entries
[i
].data64
= *(const uint64_t *)data
;
358 spec_entries
[i
].data32
= *(const uint32_t *)data
;
361 const struct spirv_to_nir_options spirv_options
= {
362 .lower_ubo_ssbo_access_to_offsets
= true,
364 .amd_gcn_shader
= true,
365 .amd_image_read_write_lod
= true,
366 .amd_shader_ballot
= device
->physical_device
->use_shader_ballot
,
367 .amd_trinary_minmax
= true,
368 .demote_to_helper_invocation
= device
->physical_device
->use_aco
,
369 .derivative_group
= true,
370 .descriptor_array_dynamic_indexing
= true,
371 .descriptor_array_non_uniform_indexing
= true,
372 .descriptor_indexing
= true,
373 .device_group
= true,
374 .draw_parameters
= true,
375 .float_controls
= true,
376 .float16
= !device
->physical_device
->use_aco
,
378 .geometry_streams
= true,
379 .image_ms_array
= true,
380 .image_read_without_format
= true,
381 .image_write_without_format
= true,
382 .int8
= !device
->physical_device
->use_aco
,
383 .int16
= !device
->physical_device
->use_aco
,
385 .int64_atomics
= true,
387 .physical_storage_buffer_address
= true,
388 .post_depth_coverage
= true,
389 .runtime_descriptor_array
= true,
390 .shader_clock
= true,
391 .shader_viewport_index_layer
= true,
392 .stencil_export
= true,
393 .storage_8bit
= !device
->physical_device
->use_aco
,
394 .storage_16bit
= !device
->physical_device
->use_aco
,
395 .storage_image_ms
= true,
396 .subgroup_arithmetic
= true,
397 .subgroup_ballot
= true,
398 .subgroup_basic
= true,
399 .subgroup_quad
= true,
400 .subgroup_shuffle
= true,
401 .subgroup_vote
= true,
402 .tessellation
= true,
403 .transform_feedback
= true,
404 .variable_pointers
= true,
406 .ubo_addr_format
= nir_address_format_32bit_index_offset
,
407 .ssbo_addr_format
= nir_address_format_32bit_index_offset
,
408 .phys_ssbo_addr_format
= nir_address_format_64bit_global
,
409 .push_const_addr_format
= nir_address_format_logical
,
410 .shared_addr_format
= nir_address_format_32bit_offset
,
411 .frag_coord_is_sysval
= true,
413 nir
= spirv_to_nir(spirv
, module
->size
/ 4,
414 spec_entries
, num_spec_entries
,
415 stage
, entrypoint_name
,
416 &spirv_options
, nir_options
);
417 assert(nir
->info
.stage
== stage
);
418 nir_validate_shader(nir
, "after spirv_to_nir");
422 /* We have to lower away local constant initializers right before we
423 * inline functions. That way they get properly initialized at the top
424 * of the function and not at the top of its caller.
426 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
427 NIR_PASS_V(nir
, nir_lower_returns
);
428 NIR_PASS_V(nir
, nir_inline_functions
);
429 NIR_PASS_V(nir
, nir_opt_deref
);
431 /* Pick off the single entrypoint that we want */
432 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
433 if (func
->is_entrypoint
)
434 func
->name
= ralloc_strdup(func
, "main");
436 exec_node_remove(&func
->node
);
438 assert(exec_list_length(&nir
->functions
) == 1);
440 /* Make sure we lower constant initializers on output variables so that
441 * nir_remove_dead_variables below sees the corresponding stores
443 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_shader_out
);
445 /* Now that we've deleted all but the main function, we can go ahead and
446 * lower the rest of the constant initializers.
448 NIR_PASS_V(nir
, nir_lower_constant_initializers
, ~0);
450 /* Split member structs. We do this before lower_io_to_temporaries so that
451 * it doesn't lower system values to temporaries by accident.
453 NIR_PASS_V(nir
, nir_split_var_copies
);
454 NIR_PASS_V(nir
, nir_split_per_member_structs
);
456 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&& use_aco
)
457 NIR_PASS_V(nir
, nir_lower_io_to_vector
, nir_var_shader_out
);
458 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
)
459 NIR_PASS_V(nir
, nir_lower_input_attachments
, true);
461 NIR_PASS_V(nir
, nir_remove_dead_variables
,
462 nir_var_shader_in
| nir_var_shader_out
| nir_var_system_value
| nir_var_mem_shared
);
464 NIR_PASS_V(nir
, nir_propagate_invariant
);
466 NIR_PASS_V(nir
, nir_lower_system_values
);
467 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
468 NIR_PASS_V(nir
, radv_nir_lower_ycbcr_textures
, layout
);
471 /* Vulkan uses the separate-shader linking model */
472 nir
->info
.separate_shader
= true;
474 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
476 static const nir_lower_tex_options tex_options
= {
478 .lower_tg4_offsets
= true,
481 nir_lower_tex(nir
, &tex_options
);
483 nir_lower_vars_to_ssa(nir
);
485 if (nir
->info
.stage
== MESA_SHADER_VERTEX
||
486 nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
487 nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
488 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
489 nir_shader_get_entrypoint(nir
), true, true);
490 } else if (nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
491 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
492 nir_shader_get_entrypoint(nir
), true, false);
495 nir_split_var_copies(nir
);
497 nir_lower_global_vars_to_local(nir
);
498 nir_remove_dead_variables(nir
, nir_var_function_temp
);
499 nir_lower_subgroups(nir
, &(struct nir_lower_subgroups_options
) {
501 .ballot_bit_size
= 64,
502 .lower_to_scalar
= 1,
503 .lower_subgroup_masks
= 1,
505 .lower_shuffle_to_32bit
= 1,
506 .lower_vote_eq_to_ballot
= 1,
509 nir_lower_load_const_to_scalar(nir
);
511 if (!(flags
& VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
))
512 radv_optimize_nir(nir
, false, true);
514 /* We call nir_lower_var_copies() after the first radv_optimize_nir()
515 * to remove any copies introduced by nir_opt_find_array_copies().
517 nir_lower_var_copies(nir
);
519 /* Lower deref operations for compute shared memory. */
520 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
521 NIR_PASS_V(nir
, nir_lower_vars_to_explicit_types
,
522 nir_var_mem_shared
, shared_var_info
);
523 NIR_PASS_V(nir
, nir_lower_explicit_io
,
524 nir_var_mem_shared
, nir_address_format_32bit_offset
);
527 /* Lower large variables that are always constant with load_constant
528 * intrinsics, which get turned into PC-relative loads from a data
529 * section next to the shader.
531 NIR_PASS_V(nir
, nir_opt_large_constants
,
532 glsl_get_natural_size_align_bytes
, 16);
534 /* Indirect lowering must be called after the radv_optimize_nir() loop
535 * has been called at least once. Otherwise indirect lowering can
536 * bloat the instruction count of the loop and cause it to be
537 * considered too large for unrolling.
539 ac_lower_indirect_derefs(nir
, device
->physical_device
->rad_info
.chip_class
);
540 radv_optimize_nir(nir
, flags
& VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
, false);
546 type_size_vec4(const struct glsl_type
*type
, bool bindless
)
548 return glsl_count_attribute_slots(type
, false);
551 static nir_variable
*
552 find_layer_in_var(nir_shader
*nir
)
554 nir_foreach_variable(var
, &nir
->inputs
) {
555 if (var
->data
.location
== VARYING_SLOT_LAYER
) {
561 nir_variable_create(nir
, nir_var_shader_in
, glsl_int_type(), "layer id");
562 var
->data
.location
= VARYING_SLOT_LAYER
;
563 var
->data
.interpolation
= INTERP_MODE_FLAT
;
567 /* We use layered rendering to implement multiview, which means we need to map
568 * view_index to gl_Layer. The attachment lowering also uses needs to know the
569 * layer so that it can sample from the correct layer. The code generates a
570 * load from the layer_id sysval, but since we don't have a way to get at this
571 * information from the fragment shader, we also need to lower this to the
572 * gl_Layer varying. This pass lowers both to a varying load from the LAYER
573 * slot, before lowering io, so that nir_assign_var_locations() will give the
574 * LAYER varying the correct driver_location.
578 lower_view_index(nir_shader
*nir
)
580 bool progress
= false;
581 nir_function_impl
*entry
= nir_shader_get_entrypoint(nir
);
583 nir_builder_init(&b
, entry
);
585 nir_variable
*layer
= NULL
;
586 nir_foreach_block(block
, entry
) {
587 nir_foreach_instr_safe(instr
, block
) {
588 if (instr
->type
!= nir_instr_type_intrinsic
)
591 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(instr
);
592 if (load
->intrinsic
!= nir_intrinsic_load_view_index
&&
593 load
->intrinsic
!= nir_intrinsic_load_layer_id
)
597 layer
= find_layer_in_var(nir
);
599 b
.cursor
= nir_before_instr(instr
);
600 nir_ssa_def
*def
= nir_load_var(&b
, layer
);
601 nir_ssa_def_rewrite_uses(&load
->dest
.ssa
,
602 nir_src_for_ssa(def
));
604 nir_instr_remove(instr
);
613 radv_lower_fs_io(nir_shader
*nir
)
615 NIR_PASS_V(nir
, lower_view_index
);
616 nir_assign_io_var_locations(&nir
->inputs
, &nir
->num_inputs
,
617 MESA_SHADER_FRAGMENT
);
619 NIR_PASS_V(nir
, nir_lower_io
, nir_var_shader_in
, type_size_vec4
, 0);
621 /* This pass needs actual constants */
622 nir_opt_constant_folding(nir
);
624 NIR_PASS_V(nir
, nir_io_add_const_offset_to_base
, nir_var_shader_in
);
629 radv_alloc_shader_memory(struct radv_device
*device
,
630 struct radv_shader_variant
*shader
)
632 mtx_lock(&device
->shader_slab_mutex
);
633 list_for_each_entry(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
635 list_for_each_entry(struct radv_shader_variant
, s
, &slab
->shaders
, slab_list
) {
636 if (s
->bo_offset
- offset
>= shader
->code_size
) {
637 shader
->bo
= slab
->bo
;
638 shader
->bo_offset
= offset
;
639 list_addtail(&shader
->slab_list
, &s
->slab_list
);
640 mtx_unlock(&device
->shader_slab_mutex
);
641 return slab
->ptr
+ offset
;
643 offset
= align_u64(s
->bo_offset
+ s
->code_size
, 256);
645 if (slab
->size
- offset
>= shader
->code_size
) {
646 shader
->bo
= slab
->bo
;
647 shader
->bo_offset
= offset
;
648 list_addtail(&shader
->slab_list
, &slab
->shaders
);
649 mtx_unlock(&device
->shader_slab_mutex
);
650 return slab
->ptr
+ offset
;
654 mtx_unlock(&device
->shader_slab_mutex
);
655 struct radv_shader_slab
*slab
= calloc(1, sizeof(struct radv_shader_slab
));
657 slab
->size
= 256 * 1024;
658 slab
->bo
= device
->ws
->buffer_create(device
->ws
, slab
->size
, 256,
660 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
661 (device
->physical_device
->rad_info
.cpdma_prefetch_writes_memory
?
662 0 : RADEON_FLAG_READ_ONLY
),
663 RADV_BO_PRIORITY_SHADER
);
664 slab
->ptr
= (char*)device
->ws
->buffer_map(slab
->bo
);
665 list_inithead(&slab
->shaders
);
667 mtx_lock(&device
->shader_slab_mutex
);
668 list_add(&slab
->slabs
, &device
->shader_slabs
);
670 shader
->bo
= slab
->bo
;
671 shader
->bo_offset
= 0;
672 list_add(&shader
->slab_list
, &slab
->shaders
);
673 mtx_unlock(&device
->shader_slab_mutex
);
678 radv_destroy_shader_slabs(struct radv_device
*device
)
680 list_for_each_entry_safe(struct radv_shader_slab
, slab
, &device
->shader_slabs
, slabs
) {
681 device
->ws
->buffer_destroy(slab
->bo
);
684 mtx_destroy(&device
->shader_slab_mutex
);
687 /* For the UMR disassembler. */
688 #define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
689 #define DEBUGGER_NUM_MARKERS 5
692 radv_get_shader_binary_size(size_t code_size
)
694 return code_size
+ DEBUGGER_NUM_MARKERS
* 4;
697 static void radv_postprocess_config(const struct radv_physical_device
*pdevice
,
698 const struct ac_shader_config
*config_in
,
699 const struct radv_shader_info
*info
,
700 gl_shader_stage stage
,
701 struct ac_shader_config
*config_out
)
703 bool scratch_enabled
= config_in
->scratch_bytes_per_wave
> 0;
704 unsigned vgpr_comp_cnt
= 0;
705 unsigned num_input_vgprs
= info
->num_input_vgprs
;
707 if (stage
== MESA_SHADER_FRAGMENT
) {
708 num_input_vgprs
= ac_get_fs_input_vgpr_cnt(config_in
, NULL
, NULL
);
711 unsigned num_vgprs
= MAX2(config_in
->num_vgprs
, num_input_vgprs
);
712 /* +3 for scratch wave offset and VCC */
713 unsigned num_sgprs
= MAX2(config_in
->num_sgprs
, info
->num_input_sgprs
+ 3);
714 unsigned num_shared_vgprs
= config_in
->num_shared_vgprs
;
715 /* shared VGPRs are introduced in Navi and are allocated in blocks of 8 (RDNA ref 3.6.5) */
716 assert((pdevice
->rad_info
.chip_class
>= GFX10
&& num_shared_vgprs
% 8 == 0)
717 || (pdevice
->rad_info
.chip_class
< GFX10
&& num_shared_vgprs
== 0));
718 unsigned num_shared_vgpr_blocks
= num_shared_vgprs
/ 8;
720 *config_out
= *config_in
;
721 config_out
->num_vgprs
= num_vgprs
;
722 config_out
->num_sgprs
= num_sgprs
;
723 config_out
->num_shared_vgprs
= num_shared_vgprs
;
725 config_out
->rsrc2
= S_00B12C_USER_SGPR(info
->num_user_sgprs
) |
726 S_00B12C_SCRATCH_EN(scratch_enabled
);
728 if (!pdevice
->use_ngg_streamout
) {
729 config_out
->rsrc2
|= S_00B12C_SO_BASE0_EN(!!info
->so
.strides
[0]) |
730 S_00B12C_SO_BASE1_EN(!!info
->so
.strides
[1]) |
731 S_00B12C_SO_BASE2_EN(!!info
->so
.strides
[2]) |
732 S_00B12C_SO_BASE3_EN(!!info
->so
.strides
[3]) |
733 S_00B12C_SO_EN(!!info
->so
.num_outputs
);
736 config_out
->rsrc1
= S_00B848_VGPRS((num_vgprs
- 1) /
737 (info
->wave_size
== 32 ? 8 : 4)) |
738 S_00B848_DX10_CLAMP(1) |
739 S_00B848_FLOAT_MODE(config_out
->float_mode
);
741 if (pdevice
->rad_info
.chip_class
>= GFX10
) {
742 config_out
->rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX10(info
->num_user_sgprs
>> 5);
744 config_out
->rsrc1
|= S_00B228_SGPRS((num_sgprs
- 1) / 8);
745 config_out
->rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX9(info
->num_user_sgprs
>> 5);
749 case MESA_SHADER_TESS_EVAL
:
751 config_out
->rsrc1
|= S_00B228_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
);
752 config_out
->rsrc2
|= S_00B22C_OC_LDS_EN(1);
753 } else if (info
->tes
.as_es
) {
754 assert(pdevice
->rad_info
.chip_class
<= GFX8
);
755 vgpr_comp_cnt
= info
->uses_prim_id
? 3 : 2;
757 config_out
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
759 bool enable_prim_id
= info
->tes
.export_prim_id
|| info
->uses_prim_id
;
760 vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
762 config_out
->rsrc1
|= S_00B128_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
);
763 config_out
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
765 config_out
->rsrc2
|= S_00B22C_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
767 case MESA_SHADER_TESS_CTRL
:
768 if (pdevice
->rad_info
.chip_class
>= GFX9
) {
769 /* We need at least 2 components for LS.
770 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
771 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
773 if (pdevice
->rad_info
.chip_class
>= GFX10
) {
774 vgpr_comp_cnt
= info
->vs
.needs_instance_id
? 3 : 1;
776 vgpr_comp_cnt
= info
->vs
.needs_instance_id
? 2 : 1;
779 config_out
->rsrc2
|= S_00B12C_OC_LDS_EN(1);
781 config_out
->rsrc1
|= S_00B428_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
) |
782 S_00B848_WGP_MODE(pdevice
->rad_info
.chip_class
>= GFX10
);
783 config_out
->rsrc2
|= S_00B42C_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
785 case MESA_SHADER_VERTEX
:
787 config_out
->rsrc1
|= S_00B228_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
);
788 } else if (info
->vs
.as_ls
) {
789 assert(pdevice
->rad_info
.chip_class
<= GFX8
);
790 /* We need at least 2 components for LS.
791 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
792 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
794 vgpr_comp_cnt
= info
->vs
.needs_instance_id
? 2 : 1;
795 } else if (info
->vs
.as_es
) {
796 assert(pdevice
->rad_info
.chip_class
<= GFX8
);
797 /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
798 vgpr_comp_cnt
= info
->vs
.needs_instance_id
? 1 : 0;
800 /* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
801 * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
802 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
804 if (info
->vs
.needs_instance_id
&& pdevice
->rad_info
.chip_class
>= GFX10
) {
806 } else if (info
->vs
.export_prim_id
) {
808 } else if (info
->vs
.needs_instance_id
) {
814 config_out
->rsrc1
|= S_00B128_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
);
815 config_out
->rsrc2
|= S_00B12C_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
818 case MESA_SHADER_FRAGMENT
:
819 config_out
->rsrc1
|= S_00B028_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
);
820 config_out
->rsrc2
|= S_00B02C_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
822 case MESA_SHADER_GEOMETRY
:
823 config_out
->rsrc1
|= S_00B228_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
) |
824 S_00B848_WGP_MODE(pdevice
->rad_info
.chip_class
>= GFX10
);
825 config_out
->rsrc2
|= S_00B22C_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
827 case MESA_SHADER_COMPUTE
:
828 config_out
->rsrc1
|= S_00B848_MEM_ORDERED(pdevice
->rad_info
.chip_class
>= GFX10
) |
829 S_00B848_WGP_MODE(pdevice
->rad_info
.chip_class
>= GFX10
);
831 S_00B84C_TGID_X_EN(info
->cs
.uses_block_id
[0]) |
832 S_00B84C_TGID_Y_EN(info
->cs
.uses_block_id
[1]) |
833 S_00B84C_TGID_Z_EN(info
->cs
.uses_block_id
[2]) |
834 S_00B84C_TIDIG_COMP_CNT(info
->cs
.uses_thread_id
[2] ? 2 :
835 info
->cs
.uses_thread_id
[1] ? 1 : 0) |
836 S_00B84C_TG_SIZE_EN(info
->cs
.uses_local_invocation_idx
) |
837 S_00B84C_LDS_SIZE(config_in
->lds_size
);
838 config_out
->rsrc3
|= S_00B8A0_SHARED_VGPR_CNT(num_shared_vgpr_blocks
);
842 unreachable("unsupported shader type");
846 if (pdevice
->rad_info
.chip_class
>= GFX10
&& info
->is_ngg
&&
847 (stage
== MESA_SHADER_VERTEX
|| stage
== MESA_SHADER_TESS_EVAL
|| stage
== MESA_SHADER_GEOMETRY
)) {
848 unsigned gs_vgpr_comp_cnt
, es_vgpr_comp_cnt
;
849 gl_shader_stage es_stage
= stage
;
850 if (stage
== MESA_SHADER_GEOMETRY
)
851 es_stage
= info
->gs
.es_type
;
853 /* VGPR5-8: (VertexID, UserVGPR0, UserVGPR1, UserVGPR2 / InstanceID) */
854 if (es_stage
== MESA_SHADER_VERTEX
) {
855 es_vgpr_comp_cnt
= info
->vs
.needs_instance_id
? 3 : 0;
856 } else if (es_stage
== MESA_SHADER_TESS_EVAL
) {
857 bool enable_prim_id
= info
->tes
.export_prim_id
|| info
->uses_prim_id
;
858 es_vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
860 unreachable("Unexpected ES shader stage");
862 bool tes_triangles
= stage
== MESA_SHADER_TESS_EVAL
&&
863 info
->tes
.primitive_mode
>= 4; /* GL_TRIANGLES */
864 if (info
->uses_invocation_id
|| stage
== MESA_SHADER_VERTEX
) {
865 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
866 } else if (info
->uses_prim_id
) {
867 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
868 } else if (info
->gs
.vertices_in
>= 3 || tes_triangles
) {
869 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
871 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
874 config_out
->rsrc1
|= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
) |
875 S_00B228_WGP_MODE(1);
876 config_out
->rsrc2
|= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
877 S_00B22C_LDS_SIZE(config_in
->lds_size
) |
878 S_00B22C_OC_LDS_EN(es_stage
== MESA_SHADER_TESS_EVAL
);
879 } else if (pdevice
->rad_info
.chip_class
>= GFX9
&&
880 stage
== MESA_SHADER_GEOMETRY
) {
881 unsigned es_type
= info
->gs
.es_type
;
882 unsigned gs_vgpr_comp_cnt
, es_vgpr_comp_cnt
;
884 if (es_type
== MESA_SHADER_VERTEX
) {
885 /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
886 if (info
->vs
.needs_instance_id
) {
887 es_vgpr_comp_cnt
= pdevice
->rad_info
.chip_class
>= GFX10
? 3 : 1;
889 es_vgpr_comp_cnt
= 0;
891 } else if (es_type
== MESA_SHADER_TESS_EVAL
) {
892 es_vgpr_comp_cnt
= info
->uses_prim_id
? 3 : 2;
894 unreachable("invalid shader ES type");
897 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
898 * VGPR[0:4] are always loaded.
900 if (info
->uses_invocation_id
) {
901 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
902 } else if (info
->uses_prim_id
) {
903 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
904 } else if (info
->gs
.vertices_in
>= 3) {
905 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
907 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
910 config_out
->rsrc1
|= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
911 config_out
->rsrc2
|= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
912 S_00B22C_OC_LDS_EN(es_type
== MESA_SHADER_TESS_EVAL
);
913 } else if (pdevice
->rad_info
.chip_class
>= GFX9
&&
914 stage
== MESA_SHADER_TESS_CTRL
) {
915 config_out
->rsrc1
|= S_00B428_LS_VGPR_COMP_CNT(vgpr_comp_cnt
);
917 config_out
->rsrc1
|= S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
);
921 struct radv_shader_variant
*
922 radv_shader_variant_create(struct radv_device
*device
,
923 const struct radv_shader_binary
*binary
,
924 bool keep_shader_info
)
926 struct ac_shader_config config
= {0};
927 struct ac_rtld_binary rtld_binary
= {0};
928 struct radv_shader_variant
*variant
= calloc(1, sizeof(struct radv_shader_variant
));
932 variant
->ref_count
= 1;
934 if (binary
->type
== RADV_BINARY_TYPE_RTLD
) {
935 struct ac_rtld_symbol lds_symbols
[2];
936 unsigned num_lds_symbols
= 0;
937 const char *elf_data
= (const char *)((struct radv_shader_binary_rtld
*)binary
)->data
;
938 size_t elf_size
= ((struct radv_shader_binary_rtld
*)binary
)->elf_size
;
940 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
941 (binary
->stage
== MESA_SHADER_GEOMETRY
|| binary
->info
.is_ngg
) &&
942 !binary
->is_gs_copy_shader
) {
943 /* We add this symbol even on LLVM <= 8 to ensure that
944 * shader->config.lds_size is set correctly below.
946 struct ac_rtld_symbol
*sym
= &lds_symbols
[num_lds_symbols
++];
947 sym
->name
= "esgs_ring";
948 sym
->size
= binary
->info
.ngg_info
.esgs_ring_size
;
949 sym
->align
= 64 * 1024;
952 if (binary
->info
.is_ngg
&&
953 binary
->stage
== MESA_SHADER_GEOMETRY
) {
954 struct ac_rtld_symbol
*sym
= &lds_symbols
[num_lds_symbols
++];
955 sym
->name
= "ngg_emit";
956 sym
->size
= binary
->info
.ngg_info
.ngg_emit_size
* 4;
960 struct ac_rtld_open_info open_info
= {
961 .info
= &device
->physical_device
->rad_info
,
962 .shader_type
= binary
->stage
,
963 .wave_size
= binary
->info
.wave_size
,
965 .elf_ptrs
= &elf_data
,
966 .elf_sizes
= &elf_size
,
967 .num_shared_lds_symbols
= num_lds_symbols
,
968 .shared_lds_symbols
= lds_symbols
,
971 if (!ac_rtld_open(&rtld_binary
, open_info
)) {
976 if (!ac_rtld_read_config(&rtld_binary
, &config
)) {
977 ac_rtld_close(&rtld_binary
);
982 /* Enable 64-bit and 16-bit denormals, because there is no performance
985 * If denormals are enabled, all floating-point output modifiers are
988 * Don't enable denormals for 32-bit floats, because:
989 * - Floating-point output modifiers would be ignored by the hw.
990 * - Some opcodes don't support denormals, such as v_mad_f32. We would
991 * have to stop using those.
992 * - GFX6 & GFX7 would be very slow.
994 config
.float_mode
|= V_00B028_FP_64_DENORMS
;
996 if (rtld_binary
.lds_size
> 0) {
997 unsigned alloc_granularity
= device
->physical_device
->rad_info
.chip_class
>= GFX7
? 512 : 256;
998 config
.lds_size
= align(rtld_binary
.lds_size
, alloc_granularity
) / alloc_granularity
;
1001 variant
->code_size
= rtld_binary
.rx_size
;
1002 variant
->exec_size
= rtld_binary
.exec_size
;
1004 assert(binary
->type
== RADV_BINARY_TYPE_LEGACY
);
1005 config
= ((struct radv_shader_binary_legacy
*)binary
)->config
;
1006 variant
->code_size
= radv_get_shader_binary_size(((struct radv_shader_binary_legacy
*)binary
)->code_size
);
1007 variant
->exec_size
= ((struct radv_shader_binary_legacy
*)binary
)->exec_size
;
1010 variant
->info
= binary
->info
;
1011 radv_postprocess_config(device
->physical_device
, &config
, &binary
->info
,
1012 binary
->stage
, &variant
->config
);
1014 if (radv_device_use_secure_compile(device
->instance
)) {
1015 if (binary
->type
== RADV_BINARY_TYPE_RTLD
)
1016 ac_rtld_close(&rtld_binary
);
1021 void *dest_ptr
= radv_alloc_shader_memory(device
, variant
);
1023 if (binary
->type
== RADV_BINARY_TYPE_RTLD
) {
1024 struct radv_shader_binary_rtld
* bin
= (struct radv_shader_binary_rtld
*)binary
;
1025 struct ac_rtld_upload_info info
= {
1026 .binary
= &rtld_binary
,
1027 .rx_va
= radv_buffer_get_va(variant
->bo
) + variant
->bo_offset
,
1031 if (!ac_rtld_upload(&info
)) {
1032 radv_shader_variant_destroy(device
, variant
);
1033 ac_rtld_close(&rtld_binary
);
1037 if (keep_shader_info
||
1038 (device
->instance
->debug_flags
& RADV_DEBUG_DUMP_SHADERS
)) {
1039 const char *disasm_data
;
1041 if (!ac_rtld_get_section_by_name(&rtld_binary
, ".AMDGPU.disasm", &disasm_data
, &disasm_size
)) {
1042 radv_shader_variant_destroy(device
, variant
);
1043 ac_rtld_close(&rtld_binary
);
1047 variant
->ir_string
= bin
->llvm_ir_size
? strdup((const char*)(bin
->data
+ bin
->elf_size
)) : NULL
;
1048 variant
->disasm_string
= malloc(disasm_size
+ 1);
1049 memcpy(variant
->disasm_string
, disasm_data
, disasm_size
);
1050 variant
->disasm_string
[disasm_size
] = 0;
1053 ac_rtld_close(&rtld_binary
);
1055 struct radv_shader_binary_legacy
* bin
= (struct radv_shader_binary_legacy
*)binary
;
1056 memcpy(dest_ptr
, bin
->data
, bin
->code_size
);
1058 /* Add end-of-code markers for the UMR disassembler. */
1059 uint32_t *ptr32
= (uint32_t *)dest_ptr
+ bin
->code_size
/ 4;
1060 for (unsigned i
= 0; i
< DEBUGGER_NUM_MARKERS
; i
++)
1061 ptr32
[i
] = DEBUGGER_END_OF_CODE_MARKER
;
1063 variant
->ir_string
= bin
->ir_size
? strdup((const char*)(bin
->data
+ bin
->code_size
)) : NULL
;
1064 variant
->disasm_string
= bin
->disasm_size
? strdup((const char*)(bin
->data
+ bin
->code_size
+ bin
->ir_size
)) : NULL
;
1070 radv_dump_nir_shaders(struct nir_shader
* const *shaders
,
1076 FILE *f
= open_memstream(&data
, &size
);
1078 for (int i
= 0; i
< shader_count
; ++i
)
1079 nir_print_shader(shaders
[i
], f
);
1083 ret
= malloc(size
+ 1);
1085 memcpy(ret
, data
, size
);
1092 static struct radv_shader_variant
*
1093 shader_variant_compile(struct radv_device
*device
,
1094 struct radv_shader_module
*module
,
1095 struct nir_shader
* const *shaders
,
1097 gl_shader_stage stage
,
1098 struct radv_shader_info
*info
,
1099 struct radv_nir_compiler_options
*options
,
1100 bool gs_copy_shader
,
1101 bool keep_shader_info
,
1103 struct radv_shader_binary
**binary_out
)
1105 enum radeon_family chip_family
= device
->physical_device
->rad_info
.family
;
1106 struct radv_shader_binary
*binary
= NULL
;
1108 options
->family
= chip_family
;
1109 options
->chip_class
= device
->physical_device
->rad_info
.chip_class
;
1110 options
->dump_shader
= radv_can_dump_shader(device
, module
, gs_copy_shader
);
1111 options
->dump_preoptir
= options
->dump_shader
&&
1112 device
->instance
->debug_flags
& RADV_DEBUG_PREOPTIR
;
1113 options
->record_ir
= keep_shader_info
;
1114 options
->check_ir
= device
->instance
->debug_flags
& RADV_DEBUG_CHECKIR
;
1115 options
->tess_offchip_block_dw_size
= device
->tess_offchip_block_dw_size
;
1116 options
->address32_hi
= device
->physical_device
->rad_info
.address32_hi
;
1117 options
->has_ls_vgpr_init_bug
= device
->physical_device
->rad_info
.has_ls_vgpr_init_bug
;
1118 options
->use_ngg_streamout
= device
->physical_device
->use_ngg_streamout
;
1120 struct radv_shader_args args
= {};
1121 args
.options
= options
;
1122 args
.shader_info
= info
;
1123 args
.is_gs_copy_shader
= gs_copy_shader
;
1124 radv_declare_shader_args(&args
,
1125 gs_copy_shader
? MESA_SHADER_VERTEX
1126 : shaders
[shader_count
- 1]->info
.stage
,
1128 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
1129 : MESA_SHADER_VERTEX
);
1131 if (!use_aco
|| options
->dump_shader
|| options
->record_ir
)
1132 ac_init_llvm_once();
1135 aco_compile_shader(shader_count
, shaders
, &binary
, &args
);
1136 binary
->info
= *info
;
1138 enum ac_target_machine_options tm_options
= 0;
1139 struct ac_llvm_compiler ac_llvm
;
1140 bool thread_compiler
;
1142 tm_options
|= AC_TM_SUPPORTS_SPILL
;
1143 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
1144 tm_options
|= AC_TM_SISCHED
;
1145 if (options
->check_ir
)
1146 tm_options
|= AC_TM_CHECK_IR
;
1147 if (device
->instance
->debug_flags
& RADV_DEBUG_NO_LOAD_STORE_OPT
)
1148 tm_options
|= AC_TM_NO_LOAD_STORE_OPT
;
1150 thread_compiler
= !(device
->instance
->debug_flags
& RADV_DEBUG_NOTHREADLLVM
);
1151 radv_init_llvm_compiler(&ac_llvm
,
1153 chip_family
, tm_options
,
1156 if (gs_copy_shader
) {
1157 assert(shader_count
== 1);
1158 radv_compile_gs_copy_shader(&ac_llvm
, *shaders
, &binary
,
1161 radv_compile_nir_shader(&ac_llvm
, &binary
, &args
,
1162 shaders
, shader_count
);
1165 binary
->info
= *info
;
1166 radv_destroy_llvm_compiler(&ac_llvm
, thread_compiler
);
1169 struct radv_shader_variant
*variant
= radv_shader_variant_create(device
, binary
,
1175 variant
->aco_used
= use_aco
;
1177 if (options
->dump_shader
) {
1178 fprintf(stderr
, "disasm:\n%s\n", variant
->disasm_string
);
1182 if (keep_shader_info
) {
1183 variant
->nir_string
= radv_dump_nir_shaders(shaders
, shader_count
);
1184 if (!gs_copy_shader
&& !module
->nir
) {
1185 variant
->spirv
= malloc(module
->size
);
1186 if (!variant
->spirv
) {
1192 memcpy(variant
->spirv
, module
->data
, module
->size
);
1193 variant
->spirv_size
= module
->size
;
1198 *binary_out
= binary
;
1205 struct radv_shader_variant
*
1206 radv_shader_variant_compile(struct radv_device
*device
,
1207 struct radv_shader_module
*module
,
1208 struct nir_shader
*const *shaders
,
1210 struct radv_pipeline_layout
*layout
,
1211 const struct radv_shader_variant_key
*key
,
1212 struct radv_shader_info
*info
,
1213 bool keep_shader_info
,
1215 struct radv_shader_binary
**binary_out
)
1217 struct radv_nir_compiler_options options
= {0};
1219 options
.layout
= layout
;
1223 options
.explicit_scratch_args
= use_aco
;
1224 options
.robust_buffer_access
= device
->robust_buffer_access
;
1226 return shader_variant_compile(device
, module
, shaders
, shader_count
, shaders
[shader_count
- 1]->info
.stage
, info
,
1227 &options
, false, keep_shader_info
, use_aco
, binary_out
);
1230 struct radv_shader_variant
*
1231 radv_create_gs_copy_shader(struct radv_device
*device
,
1232 struct nir_shader
*shader
,
1233 struct radv_shader_info
*info
,
1234 struct radv_shader_binary
**binary_out
,
1235 bool keep_shader_info
,
1238 struct radv_nir_compiler_options options
= {0};
1240 options
.key
.has_multiview_view_index
= multiview
;
1242 return shader_variant_compile(device
, NULL
, &shader
, 1, MESA_SHADER_VERTEX
,
1243 info
, &options
, true, keep_shader_info
, false, binary_out
);
1247 radv_shader_variant_destroy(struct radv_device
*device
,
1248 struct radv_shader_variant
*variant
)
1250 if (!p_atomic_dec_zero(&variant
->ref_count
))
1253 mtx_lock(&device
->shader_slab_mutex
);
1254 list_del(&variant
->slab_list
);
1255 mtx_unlock(&device
->shader_slab_mutex
);
1257 free(variant
->spirv
);
1258 free(variant
->nir_string
);
1259 free(variant
->disasm_string
);
1260 free(variant
->ir_string
);
1265 radv_get_shader_name(struct radv_shader_info
*info
,
1266 gl_shader_stage stage
)
1269 case MESA_SHADER_VERTEX
:
1271 return "Vertex Shader as LS";
1272 else if (info
->vs
.as_es
)
1273 return "Vertex Shader as ES";
1274 else if (info
->is_ngg
)
1275 return "Vertex Shader as ESGS";
1277 return "Vertex Shader as VS";
1278 case MESA_SHADER_TESS_CTRL
:
1279 return "Tessellation Control Shader";
1280 case MESA_SHADER_TESS_EVAL
:
1281 if (info
->tes
.as_es
)
1282 return "Tessellation Evaluation Shader as ES";
1283 else if (info
->is_ngg
)
1284 return "Tessellation Evaluation Shader as ESGS";
1286 return "Tessellation Evaluation Shader as VS";
1287 case MESA_SHADER_GEOMETRY
:
1288 return "Geometry Shader";
1289 case MESA_SHADER_FRAGMENT
:
1290 return "Pixel Shader";
1291 case MESA_SHADER_COMPUTE
:
1292 return "Compute Shader";
1294 return "Unknown shader";
1299 radv_get_max_workgroup_size(enum chip_class chip_class
,
1300 gl_shader_stage stage
,
1301 const unsigned *sizes
)
1304 case MESA_SHADER_TESS_CTRL
:
1305 return chip_class
>= GFX7
? 128 : 64;
1306 case MESA_SHADER_GEOMETRY
:
1307 return chip_class
>= GFX9
? 128 : 64;
1308 case MESA_SHADER_COMPUTE
:
1314 unsigned max_workgroup_size
= sizes
[0] * sizes
[1] * sizes
[2];
1315 return max_workgroup_size
;
1319 radv_get_max_waves(struct radv_device
*device
,
1320 struct radv_shader_variant
*variant
,
1321 gl_shader_stage stage
)
1323 enum chip_class chip_class
= device
->physical_device
->rad_info
.chip_class
;
1324 unsigned lds_increment
= chip_class
>= GFX7
? 512 : 256;
1325 uint8_t wave_size
= variant
->info
.wave_size
;
1326 struct ac_shader_config
*conf
= &variant
->config
;
1327 unsigned max_simd_waves
;
1328 unsigned lds_per_wave
= 0;
1330 max_simd_waves
= device
->physical_device
->rad_info
.max_wave64_per_simd
;
1332 if (stage
== MESA_SHADER_FRAGMENT
) {
1333 lds_per_wave
= conf
->lds_size
* lds_increment
+
1334 align(variant
->info
.ps
.num_interp
* 48,
1336 } else if (stage
== MESA_SHADER_COMPUTE
) {
1337 unsigned max_workgroup_size
=
1338 radv_get_max_workgroup_size(chip_class
, stage
, variant
->info
.cs
.block_size
);
1339 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
1340 DIV_ROUND_UP(max_workgroup_size
, wave_size
);
1343 if (conf
->num_sgprs
) {
1344 unsigned sgprs
= align(conf
->num_sgprs
, chip_class
>= GFX8
? 16 : 8);
1346 MIN2(max_simd_waves
,
1347 device
->physical_device
->rad_info
.num_physical_sgprs_per_simd
/
1351 if (conf
->num_vgprs
) {
1352 unsigned vgprs
= align(conf
->num_vgprs
, wave_size
== 32 ? 8 : 4);
1354 MIN2(max_simd_waves
,
1355 RADV_NUM_PHYSICAL_VGPRS
/ vgprs
);
1358 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
1362 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
1364 return max_simd_waves
;
1368 generate_shader_stats(struct radv_device
*device
,
1369 struct radv_shader_variant
*variant
,
1370 gl_shader_stage stage
,
1371 struct _mesa_string_buffer
*buf
)
1373 struct ac_shader_config
*conf
= &variant
->config
;
1374 unsigned max_simd_waves
= radv_get_max_waves(device
, variant
, stage
);
1376 if (stage
== MESA_SHADER_FRAGMENT
) {
1377 _mesa_string_buffer_printf(buf
, "*** SHADER CONFIG ***\n"
1378 "SPI_PS_INPUT_ADDR = 0x%04x\n"
1379 "SPI_PS_INPUT_ENA = 0x%04x\n",
1380 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
1383 _mesa_string_buffer_printf(buf
, "*** SHADER STATS ***\n"
1386 "Spilled SGPRs: %d\n"
1387 "Spilled VGPRs: %d\n"
1388 "PrivMem VGPRS: %d\n"
1389 "Code Size: %d bytes\n"
1391 "Scratch: %d bytes per wave\n"
1393 "********************\n\n\n",
1394 conf
->num_sgprs
, conf
->num_vgprs
,
1395 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
1396 variant
->info
.private_mem_vgprs
, variant
->exec_size
,
1397 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
1402 radv_shader_dump_stats(struct radv_device
*device
,
1403 struct radv_shader_variant
*variant
,
1404 gl_shader_stage stage
,
1407 struct _mesa_string_buffer
*buf
= _mesa_string_buffer_create(NULL
, 256);
1409 generate_shader_stats(device
, variant
, stage
, buf
);
1411 fprintf(file
, "\n%s:\n", radv_get_shader_name(&variant
->info
, stage
));
1412 fprintf(file
, "%s", buf
->buf
);
1414 _mesa_string_buffer_destroy(buf
);
1418 radv_GetShaderInfoAMD(VkDevice _device
,
1419 VkPipeline _pipeline
,
1420 VkShaderStageFlagBits shaderStage
,
1421 VkShaderInfoTypeAMD infoType
,
1425 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1426 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
1427 gl_shader_stage stage
= vk_to_mesa_shader_stage(shaderStage
);
1428 struct radv_shader_variant
*variant
= pipeline
->shaders
[stage
];
1429 struct _mesa_string_buffer
*buf
;
1430 VkResult result
= VK_SUCCESS
;
1432 /* Spec doesn't indicate what to do if the stage is invalid, so just
1433 * return no info for this. */
1435 return vk_error(device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1438 case VK_SHADER_INFO_TYPE_STATISTICS_AMD
:
1440 *pInfoSize
= sizeof(VkShaderStatisticsInfoAMD
);
1442 unsigned lds_multiplier
= device
->physical_device
->rad_info
.chip_class
>= GFX7
? 512 : 256;
1443 struct ac_shader_config
*conf
= &variant
->config
;
1445 VkShaderStatisticsInfoAMD statistics
= {};
1446 statistics
.shaderStageMask
= shaderStage
;
1447 statistics
.numPhysicalVgprs
= RADV_NUM_PHYSICAL_VGPRS
;
1448 statistics
.numPhysicalSgprs
= device
->physical_device
->rad_info
.num_physical_sgprs_per_simd
;
1449 statistics
.numAvailableSgprs
= statistics
.numPhysicalSgprs
;
1451 if (stage
== MESA_SHADER_COMPUTE
) {
1452 unsigned *local_size
= variant
->info
.cs
.block_size
;
1453 unsigned workgroup_size
= local_size
[0] * local_size
[1] * local_size
[2];
1455 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
/
1456 ceil((double)workgroup_size
/ statistics
.numPhysicalVgprs
);
1458 statistics
.computeWorkGroupSize
[0] = local_size
[0];
1459 statistics
.computeWorkGroupSize
[1] = local_size
[1];
1460 statistics
.computeWorkGroupSize
[2] = local_size
[2];
1462 statistics
.numAvailableVgprs
= statistics
.numPhysicalVgprs
;
1465 statistics
.resourceUsage
.numUsedVgprs
= conf
->num_vgprs
;
1466 statistics
.resourceUsage
.numUsedSgprs
= conf
->num_sgprs
;
1467 statistics
.resourceUsage
.ldsSizePerLocalWorkGroup
= 32768;
1468 statistics
.resourceUsage
.ldsUsageSizeInBytes
= conf
->lds_size
* lds_multiplier
;
1469 statistics
.resourceUsage
.scratchMemUsageInBytes
= conf
->scratch_bytes_per_wave
;
1471 size_t size
= *pInfoSize
;
1472 *pInfoSize
= sizeof(statistics
);
1474 memcpy(pInfo
, &statistics
, MIN2(size
, *pInfoSize
));
1476 if (size
< *pInfoSize
)
1477 result
= VK_INCOMPLETE
;
1481 case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD
:
1482 buf
= _mesa_string_buffer_create(NULL
, 1024);
1484 _mesa_string_buffer_printf(buf
, "%s:\n", radv_get_shader_name(&variant
->info
, stage
));
1485 _mesa_string_buffer_printf(buf
, "%s\n\n", variant
->ir_string
);
1486 _mesa_string_buffer_printf(buf
, "%s\n\n", variant
->disasm_string
);
1487 generate_shader_stats(device
, variant
, stage
, buf
);
1489 /* Need to include the null terminator. */
1490 size_t length
= buf
->length
+ 1;
1493 *pInfoSize
= length
;
1495 size_t size
= *pInfoSize
;
1496 *pInfoSize
= length
;
1498 memcpy(pInfo
, buf
->buf
, MIN2(size
, length
));
1501 result
= VK_INCOMPLETE
;
1504 _mesa_string_buffer_destroy(buf
);
1507 /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
1508 result
= VK_ERROR_FEATURE_NOT_PRESENT
;