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 "radv_private.h"
29 #include "radv_shader.h"
30 #include "radv_shader_helper.h"
33 #include <llvm-c/Core.h>
34 #include <llvm-c/TargetMachine.h>
35 #include <llvm-c/Transforms/Scalar.h>
36 #include <llvm-c/Transforms/Utils.h>
40 #include "ac_binary.h"
41 #include "ac_llvm_util.h"
42 #include "ac_llvm_build.h"
43 #include "ac_shader_abi.h"
44 #include "ac_shader_util.h"
45 #include "ac_exp_param.h"
47 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
49 struct radv_shader_context
{
50 struct ac_llvm_context ac
;
51 const struct radv_nir_compiler_options
*options
;
52 struct radv_shader_variant_info
*shader_info
;
53 struct ac_shader_abi abi
;
55 unsigned max_workgroup_size
;
56 LLVMContextRef context
;
57 LLVMValueRef main_function
;
59 LLVMValueRef descriptor_sets
[RADV_UD_MAX_SETS
];
60 LLVMValueRef ring_offsets
;
62 LLVMValueRef vertex_buffers
;
63 LLVMValueRef rel_auto_id
;
64 LLVMValueRef vs_prim_id
;
65 LLVMValueRef es2gs_offset
;
68 LLVMValueRef merged_wave_info
;
69 LLVMValueRef tess_factor_offset
;
70 LLVMValueRef tes_rel_patch_id
;
74 LLVMValueRef gs2vs_offset
;
75 LLVMValueRef gs_wave_id
;
76 LLVMValueRef gs_vtx_offset
[6];
78 LLVMValueRef esgs_ring
;
79 LLVMValueRef gsvs_ring
[4];
80 LLVMValueRef hs_ring_tess_offchip
;
81 LLVMValueRef hs_ring_tess_factor
;
83 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
84 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
87 LLVMValueRef streamout_buffers
;
88 LLVMValueRef streamout_write_idx
;
89 LLVMValueRef streamout_config
;
90 LLVMValueRef streamout_offset
[4];
92 gl_shader_stage stage
;
94 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
99 bool is_gs_copy_shader
;
100 LLVMValueRef gs_next_vertex
[4];
101 unsigned gs_max_out_vertices
;
103 unsigned tes_primitive_mode
;
105 uint32_t tcs_patch_outputs_read
;
106 uint64_t tcs_outputs_read
;
107 uint32_t tcs_vertices_per_patch
;
108 uint32_t tcs_num_inputs
;
109 uint32_t tcs_num_patches
;
110 uint32_t max_gsvs_emit_size
;
111 uint32_t gsvs_vertex_size
;
114 enum radeon_llvm_calling_convention
{
115 RADEON_LLVM_AMDGPU_VS
= 87,
116 RADEON_LLVM_AMDGPU_GS
= 88,
117 RADEON_LLVM_AMDGPU_PS
= 89,
118 RADEON_LLVM_AMDGPU_CS
= 90,
119 RADEON_LLVM_AMDGPU_HS
= 93,
122 static inline struct radv_shader_context
*
123 radv_shader_context_from_abi(struct ac_shader_abi
*abi
)
125 struct radv_shader_context
*ctx
= NULL
;
126 return container_of(abi
, ctx
, abi
);
129 struct ac_build_if_state
131 struct radv_shader_context
*ctx
;
132 LLVMValueRef condition
;
133 LLVMBasicBlockRef entry_block
;
134 LLVMBasicBlockRef true_block
;
135 LLVMBasicBlockRef false_block
;
136 LLVMBasicBlockRef merge_block
;
139 static LLVMBasicBlockRef
140 ac_build_insert_new_block(struct radv_shader_context
*ctx
, const char *name
)
142 LLVMBasicBlockRef current_block
;
143 LLVMBasicBlockRef next_block
;
144 LLVMBasicBlockRef new_block
;
146 /* get current basic block */
147 current_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
149 /* chqeck if there's another block after this one */
150 next_block
= LLVMGetNextBasicBlock(current_block
);
152 /* insert the new block before the next block */
153 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
156 /* append new block after current block */
157 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
158 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
164 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
165 struct radv_shader_context
*ctx
,
166 LLVMValueRef condition
)
168 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
170 memset(ifthen
, 0, sizeof *ifthen
);
172 ifthen
->condition
= condition
;
173 ifthen
->entry_block
= block
;
175 /* create endif/merge basic block for the phi functions */
176 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
178 /* create/insert true_block before merge_block */
180 LLVMInsertBasicBlockInContext(ctx
->context
,
184 /* successive code goes into the true block */
185 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ifthen
->true_block
);
192 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
194 LLVMBuilderRef builder
= ifthen
->ctx
->ac
.builder
;
196 /* Insert branch to the merge block from current block */
197 LLVMBuildBr(builder
, ifthen
->merge_block
);
200 * Now patch in the various branch instructions.
203 /* Insert the conditional branch instruction at the end of entry_block */
204 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
205 if (ifthen
->false_block
) {
206 /* we have an else clause */
207 LLVMBuildCondBr(builder
, ifthen
->condition
,
208 ifthen
->true_block
, ifthen
->false_block
);
212 LLVMBuildCondBr(builder
, ifthen
->condition
,
213 ifthen
->true_block
, ifthen
->merge_block
);
216 /* Resume building code at end of the ifthen->merge_block */
217 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
221 static LLVMValueRef
get_rel_patch_id(struct radv_shader_context
*ctx
)
223 switch (ctx
->stage
) {
224 case MESA_SHADER_TESS_CTRL
:
225 return ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
226 case MESA_SHADER_TESS_EVAL
:
227 return ctx
->tes_rel_patch_id
;
230 unreachable("Illegal stage");
235 get_tcs_num_patches(struct radv_shader_context
*ctx
)
237 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
238 unsigned num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
239 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
240 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
241 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
242 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
243 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
244 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
245 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
246 unsigned num_patches
;
247 unsigned hardware_lds_size
;
249 /* Ensure that we only need one wave per SIMD so we don't need to check
250 * resource usage. Also ensures that the number of tcs in and out
251 * vertices per threadgroup are at most 256.
253 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
254 /* Make sure that the data fits in LDS. This assumes the shaders only
255 * use LDS for the inputs and outputs.
257 hardware_lds_size
= 32768;
259 /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
260 * threadgroup, even though there is more than 32 KiB LDS.
262 * Test: dEQP-VK.tessellation.shader_input_output.barrier
264 if (ctx
->options
->chip_class
>= CIK
&& ctx
->options
->family
!= CHIP_STONEY
)
265 hardware_lds_size
= 65536;
267 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
268 /* Make sure the output data fits in the offchip buffer */
269 num_patches
= MIN2(num_patches
, (ctx
->options
->tess_offchip_block_dw_size
* 4) / output_patch_size
);
270 /* Not necessary for correctness, but improves performance. The
271 * specific value is taken from the proprietary driver.
273 num_patches
= MIN2(num_patches
, 40);
275 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
276 if (ctx
->options
->chip_class
== SI
) {
277 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
278 num_patches
= MIN2(num_patches
, one_wave
);
284 calculate_tess_lds_size(struct radv_shader_context
*ctx
)
286 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
287 unsigned num_tcs_output_cp
;
288 unsigned num_tcs_outputs
, num_tcs_patch_outputs
;
289 unsigned input_vertex_size
, output_vertex_size
;
290 unsigned input_patch_size
, output_patch_size
;
291 unsigned pervertex_output_patch_size
;
292 unsigned output_patch0_offset
;
293 unsigned num_patches
;
296 num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
297 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
298 num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
300 input_vertex_size
= ctx
->tcs_num_inputs
* 16;
301 output_vertex_size
= num_tcs_outputs
* 16;
303 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
305 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
306 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
308 num_patches
= ctx
->tcs_num_patches
;
309 output_patch0_offset
= input_patch_size
* num_patches
;
311 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
315 /* Tessellation shaders pass outputs to the next shader using LDS.
317 * LS outputs = TCS inputs
318 * TCS outputs = TES inputs
321 * - TCS inputs for patch 0
322 * - TCS inputs for patch 1
323 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
325 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
326 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
327 * - TCS outputs for patch 1
328 * - Per-patch TCS outputs for patch 1
329 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
330 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
333 * All three shaders VS(LS), TCS, TES share the same LDS space.
336 get_tcs_in_patch_stride(struct radv_shader_context
*ctx
)
338 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
339 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
340 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
342 input_patch_size
/= 4;
343 return LLVMConstInt(ctx
->ac
.i32
, input_patch_size
, false);
347 get_tcs_out_patch_stride(struct radv_shader_context
*ctx
)
349 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
350 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
351 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
352 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
353 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
354 output_patch_size
/= 4;
355 return LLVMConstInt(ctx
->ac
.i32
, output_patch_size
, false);
359 get_tcs_out_vertex_stride(struct radv_shader_context
*ctx
)
361 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
362 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
363 output_vertex_size
/= 4;
364 return LLVMConstInt(ctx
->ac
.i32
, output_vertex_size
, false);
368 get_tcs_out_patch0_offset(struct radv_shader_context
*ctx
)
370 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
371 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
372 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
373 uint32_t output_patch0_offset
= input_patch_size
;
374 unsigned num_patches
= ctx
->tcs_num_patches
;
376 output_patch0_offset
*= num_patches
;
377 output_patch0_offset
/= 4;
378 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
382 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context
*ctx
)
384 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
385 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
386 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
387 uint32_t output_patch0_offset
= input_patch_size
;
389 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
390 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
391 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
392 unsigned num_patches
= ctx
->tcs_num_patches
;
394 output_patch0_offset
*= num_patches
;
395 output_patch0_offset
+= pervertex_output_patch_size
;
396 output_patch0_offset
/= 4;
397 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
401 get_tcs_in_current_patch_offset(struct radv_shader_context
*ctx
)
403 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
404 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
406 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
410 get_tcs_out_current_patch_offset(struct radv_shader_context
*ctx
)
412 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
413 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
414 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
416 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
421 get_tcs_out_current_patch_data_offset(struct radv_shader_context
*ctx
)
423 LLVMValueRef patch0_patch_data_offset
=
424 get_tcs_out_patch0_patch_data_offset(ctx
);
425 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
426 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
428 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
429 patch0_patch_data_offset
);
434 LLVMTypeRef types
[MAX_ARGS
];
435 LLVMValueRef
*assign
[MAX_ARGS
];
436 unsigned array_params_mask
;
439 uint8_t num_sgprs_used
;
440 uint8_t num_vgprs_used
;
443 enum ac_arg_regfile
{
449 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
450 LLVMValueRef
*param_ptr
)
452 assert(info
->count
< MAX_ARGS
);
454 info
->assign
[info
->count
] = param_ptr
;
455 info
->types
[info
->count
] = type
;
458 if (regfile
== ARG_SGPR
) {
459 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
462 assert(regfile
== ARG_VGPR
);
463 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
468 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
470 info
->array_params_mask
|= (1 << info
->count
);
471 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
474 static void assign_arguments(LLVMValueRef main_function
,
475 struct arg_info
*info
)
478 for (i
= 0; i
< info
->count
; i
++) {
480 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
485 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
486 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
487 unsigned num_return_elems
,
488 struct arg_info
*args
,
489 unsigned max_workgroup_size
,
490 const struct radv_nir_compiler_options
*options
)
492 LLVMTypeRef main_function_type
, ret_type
;
493 LLVMBasicBlockRef main_function_body
;
495 if (num_return_elems
)
496 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
497 num_return_elems
, true);
499 ret_type
= LLVMVoidTypeInContext(ctx
);
501 /* Setup the function */
503 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
504 LLVMValueRef main_function
=
505 LLVMAddFunction(module
, "main", main_function_type
);
507 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
508 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
510 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
511 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
512 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
514 if (args
->array_params_mask
& (1 << i
)) {
515 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
516 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
517 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
521 if (options
->address32_hi
) {
522 ac_llvm_add_target_dep_function_attr(main_function
,
523 "amdgpu-32bit-address-high-bits",
524 options
->address32_hi
);
527 if (max_workgroup_size
) {
528 ac_llvm_add_target_dep_function_attr(main_function
,
529 "amdgpu-max-work-group-size",
532 if (options
->unsafe_math
) {
533 /* These were copied from some LLVM test. */
534 LLVMAddTargetDependentFunctionAttr(main_function
,
535 "less-precise-fpmad",
537 LLVMAddTargetDependentFunctionAttr(main_function
,
540 LLVMAddTargetDependentFunctionAttr(main_function
,
543 LLVMAddTargetDependentFunctionAttr(main_function
,
546 LLVMAddTargetDependentFunctionAttr(main_function
,
547 "no-signed-zeros-fp-math",
550 return main_function
;
555 set_loc(struct radv_userdata_info
*ud_info
, uint8_t *sgpr_idx
,
558 ud_info
->sgpr_idx
= *sgpr_idx
;
559 ud_info
->num_sgprs
= num_sgprs
;
560 *sgpr_idx
+= num_sgprs
;
564 set_loc_shader(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
567 struct radv_userdata_info
*ud_info
=
568 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
571 set_loc(ud_info
, sgpr_idx
, num_sgprs
);
575 set_loc_shader_ptr(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
577 bool use_32bit_pointers
= idx
!= AC_UD_SCRATCH_RING_OFFSETS
;
579 set_loc_shader(ctx
, idx
, sgpr_idx
, use_32bit_pointers
? 1 : 2);
583 set_loc_desc(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
585 struct radv_userdata_locations
*locs
=
586 &ctx
->shader_info
->user_sgprs_locs
;
587 struct radv_userdata_info
*ud_info
= &locs
->descriptor_sets
[idx
];
590 set_loc(ud_info
, sgpr_idx
, 1);
592 locs
->descriptor_sets_enabled
|= 1 << idx
;
595 struct user_sgpr_info
{
596 bool need_ring_offsets
;
597 bool indirect_all_descriptor_sets
;
600 static bool needs_view_index_sgpr(struct radv_shader_context
*ctx
,
601 gl_shader_stage stage
)
604 case MESA_SHADER_VERTEX
:
605 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
606 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
609 case MESA_SHADER_TESS_EVAL
:
610 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
613 case MESA_SHADER_GEOMETRY
:
614 case MESA_SHADER_TESS_CTRL
:
615 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
625 count_vs_user_sgprs(struct radv_shader_context
*ctx
)
629 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
)
631 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
636 static void allocate_user_sgprs(struct radv_shader_context
*ctx
,
637 gl_shader_stage stage
,
638 bool has_previous_stage
,
639 gl_shader_stage previous_stage
,
640 bool needs_view_index
,
641 struct user_sgpr_info
*user_sgpr_info
)
643 uint8_t user_sgpr_count
= 0;
645 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
647 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
648 if (stage
== MESA_SHADER_GEOMETRY
||
649 stage
== MESA_SHADER_VERTEX
||
650 stage
== MESA_SHADER_TESS_CTRL
||
651 stage
== MESA_SHADER_TESS_EVAL
||
652 ctx
->is_gs_copy_shader
)
653 user_sgpr_info
->need_ring_offsets
= true;
655 if (stage
== MESA_SHADER_FRAGMENT
&&
656 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
657 user_sgpr_info
->need_ring_offsets
= true;
659 /* 2 user sgprs will nearly always be allocated for scratch/rings */
660 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
661 user_sgpr_count
+= 2;
665 case MESA_SHADER_COMPUTE
:
666 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
667 user_sgpr_count
+= 3;
669 case MESA_SHADER_FRAGMENT
:
670 user_sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
672 case MESA_SHADER_VERTEX
:
673 if (!ctx
->is_gs_copy_shader
)
674 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
676 case MESA_SHADER_TESS_CTRL
:
677 if (has_previous_stage
) {
678 if (previous_stage
== MESA_SHADER_VERTEX
)
679 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
682 case MESA_SHADER_TESS_EVAL
:
684 case MESA_SHADER_GEOMETRY
:
685 if (has_previous_stage
) {
686 if (previous_stage
== MESA_SHADER_VERTEX
) {
687 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
695 if (needs_view_index
)
698 if (ctx
->shader_info
->info
.loads_push_constants
)
701 if (ctx
->streamout_buffers
)
704 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
&& stage
!= MESA_SHADER_COMPUTE
? 32 : 16;
705 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_count
;
706 uint32_t num_desc_set
=
707 util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
);
709 if (remaining_sgprs
< num_desc_set
) {
710 user_sgpr_info
->indirect_all_descriptor_sets
= true;
715 declare_global_input_sgprs(struct radv_shader_context
*ctx
,
716 const struct user_sgpr_info
*user_sgpr_info
,
717 struct arg_info
*args
,
718 LLVMValueRef
*desc_sets
)
720 LLVMTypeRef type
= ac_array_in_const32_addr_space(ctx
->ac
.i8
);
722 /* 1 for each descriptor set */
723 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
724 uint32_t mask
= ctx
->shader_info
->info
.desc_set_used_mask
;
727 int i
= u_bit_scan(&mask
);
729 add_array_arg(args
, type
, &ctx
->descriptor_sets
[i
]);
732 add_array_arg(args
, ac_array_in_const32_addr_space(type
), desc_sets
);
735 if (ctx
->shader_info
->info
.loads_push_constants
) {
736 /* 1 for push constants and dynamic descriptors */
737 add_array_arg(args
, type
, &ctx
->abi
.push_constants
);
740 if (ctx
->shader_info
->info
.so
.num_outputs
) {
741 add_arg(args
, ARG_SGPR
,
742 ac_array_in_const32_addr_space(ctx
->ac
.v4i32
),
743 &ctx
->streamout_buffers
);
748 declare_vs_specific_input_sgprs(struct radv_shader_context
*ctx
,
749 gl_shader_stage stage
,
750 bool has_previous_stage
,
751 gl_shader_stage previous_stage
,
752 struct arg_info
*args
)
754 if (!ctx
->is_gs_copy_shader
&&
755 (stage
== MESA_SHADER_VERTEX
||
756 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
757 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
758 add_arg(args
, ARG_SGPR
,
759 ac_array_in_const32_addr_space(ctx
->ac
.v4i32
),
760 &ctx
->vertex_buffers
);
762 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
763 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
764 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
765 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
771 declare_vs_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
773 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
774 if (!ctx
->is_gs_copy_shader
) {
775 if (ctx
->options
->key
.vs
.as_ls
) {
776 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
777 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
779 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
780 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
782 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
787 declare_streamout_sgprs(struct radv_shader_context
*ctx
, gl_shader_stage stage
,
788 struct arg_info
*args
)
792 /* Streamout SGPRs. */
793 if (ctx
->shader_info
->info
.so
.num_outputs
) {
794 assert(stage
== MESA_SHADER_VERTEX
||
795 stage
== MESA_SHADER_TESS_EVAL
);
797 if (stage
!= MESA_SHADER_TESS_EVAL
) {
798 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_config
);
800 args
->assign
[args
->count
- 1] = &ctx
->streamout_config
;
801 args
->types
[args
->count
- 1] = ctx
->ac
.i32
;
804 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_write_idx
);
807 /* A streamout buffer offset is loaded if the stride is non-zero. */
808 for (i
= 0; i
< 4; i
++) {
809 if (!ctx
->shader_info
->info
.so
.strides
[i
])
812 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_offset
[i
]);
817 declare_tes_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
819 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
820 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
821 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
822 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
826 set_global_input_locs(struct radv_shader_context
*ctx
,
827 const struct user_sgpr_info
*user_sgpr_info
,
828 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
830 uint32_t mask
= ctx
->shader_info
->info
.desc_set_used_mask
;
832 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
834 int i
= u_bit_scan(&mask
);
836 set_loc_desc(ctx
, i
, user_sgpr_idx
);
839 set_loc_shader_ptr(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
843 int i
= u_bit_scan(&mask
);
845 ctx
->descriptor_sets
[i
] =
846 ac_build_load_to_sgpr(&ctx
->ac
, desc_sets
,
847 LLVMConstInt(ctx
->ac
.i32
, i
, false));
851 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
854 if (ctx
->shader_info
->info
.loads_push_constants
) {
855 set_loc_shader_ptr(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
);
858 if (ctx
->streamout_buffers
) {
859 set_loc_shader_ptr(ctx
, AC_UD_STREAMOUT_BUFFERS
,
865 set_vs_specific_input_locs(struct radv_shader_context
*ctx
,
866 gl_shader_stage stage
, bool has_previous_stage
,
867 gl_shader_stage previous_stage
,
868 uint8_t *user_sgpr_idx
)
870 if (!ctx
->is_gs_copy_shader
&&
871 (stage
== MESA_SHADER_VERTEX
||
872 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
873 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
874 set_loc_shader_ptr(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
879 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
882 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
883 user_sgpr_idx
, vs_num
);
887 static void set_llvm_calling_convention(LLVMValueRef func
,
888 gl_shader_stage stage
)
890 enum radeon_llvm_calling_convention calling_conv
;
893 case MESA_SHADER_VERTEX
:
894 case MESA_SHADER_TESS_EVAL
:
895 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
897 case MESA_SHADER_GEOMETRY
:
898 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
900 case MESA_SHADER_TESS_CTRL
:
901 calling_conv
= RADEON_LLVM_AMDGPU_HS
;
903 case MESA_SHADER_FRAGMENT
:
904 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
906 case MESA_SHADER_COMPUTE
:
907 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
910 unreachable("Unhandle shader type");
913 LLVMSetFunctionCallConv(func
, calling_conv
);
916 static void create_function(struct radv_shader_context
*ctx
,
917 gl_shader_stage stage
,
918 bool has_previous_stage
,
919 gl_shader_stage previous_stage
)
921 uint8_t user_sgpr_idx
;
922 struct user_sgpr_info user_sgpr_info
;
923 struct arg_info args
= {};
924 LLVMValueRef desc_sets
;
925 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
926 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
927 previous_stage
, needs_view_index
, &user_sgpr_info
);
929 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
930 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
935 case MESA_SHADER_COMPUTE
:
936 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
939 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
940 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
941 &ctx
->abi
.num_work_groups
);
944 for (int i
= 0; i
< 3; i
++) {
945 ctx
->abi
.workgroup_ids
[i
] = NULL
;
946 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
947 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
948 &ctx
->abi
.workgroup_ids
[i
]);
952 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
953 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
954 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
955 &ctx
->abi
.local_invocation_ids
);
957 case MESA_SHADER_VERTEX
:
958 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
961 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
962 previous_stage
, &args
);
964 if (needs_view_index
)
965 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
966 &ctx
->abi
.view_index
);
967 if (ctx
->options
->key
.vs
.as_es
) {
968 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
970 } else if (ctx
->options
->key
.vs
.as_ls
) {
971 /* no extra parameters */
973 declare_streamout_sgprs(ctx
, stage
, &args
);
976 declare_vs_input_vgprs(ctx
, &args
);
978 case MESA_SHADER_TESS_CTRL
:
979 if (has_previous_stage
) {
980 // First 6 system regs
981 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
982 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
983 &ctx
->merged_wave_info
);
984 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
985 &ctx
->tess_factor_offset
);
987 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
988 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
989 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
991 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
994 declare_vs_specific_input_sgprs(ctx
, stage
,
996 previous_stage
, &args
);
998 if (needs_view_index
)
999 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1000 &ctx
->abi
.view_index
);
1002 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1003 &ctx
->abi
.tcs_patch_id
);
1004 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1005 &ctx
->abi
.tcs_rel_ids
);
1007 declare_vs_input_vgprs(ctx
, &args
);
1009 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1012 if (needs_view_index
)
1013 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1014 &ctx
->abi
.view_index
);
1016 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1017 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1018 &ctx
->tess_factor_offset
);
1019 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1020 &ctx
->abi
.tcs_patch_id
);
1021 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1022 &ctx
->abi
.tcs_rel_ids
);
1025 case MESA_SHADER_TESS_EVAL
:
1026 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1029 if (needs_view_index
)
1030 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1031 &ctx
->abi
.view_index
);
1033 if (ctx
->options
->key
.tes
.as_es
) {
1034 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1035 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1036 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1037 &ctx
->es2gs_offset
);
1039 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1040 declare_streamout_sgprs(ctx
, stage
, &args
);
1041 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1043 declare_tes_input_vgprs(ctx
, &args
);
1045 case MESA_SHADER_GEOMETRY
:
1046 if (has_previous_stage
) {
1047 // First 6 system regs
1048 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1049 &ctx
->gs2vs_offset
);
1050 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1051 &ctx
->merged_wave_info
);
1052 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1054 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
1055 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1056 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1058 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1061 if (previous_stage
!= MESA_SHADER_TESS_EVAL
) {
1062 declare_vs_specific_input_sgprs(ctx
, stage
,
1068 if (needs_view_index
)
1069 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1070 &ctx
->abi
.view_index
);
1072 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1073 &ctx
->gs_vtx_offset
[0]);
1074 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1075 &ctx
->gs_vtx_offset
[2]);
1076 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1077 &ctx
->abi
.gs_prim_id
);
1078 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1079 &ctx
->abi
.gs_invocation_id
);
1080 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1081 &ctx
->gs_vtx_offset
[4]);
1083 if (previous_stage
== MESA_SHADER_VERTEX
) {
1084 declare_vs_input_vgprs(ctx
, &args
);
1086 declare_tes_input_vgprs(ctx
, &args
);
1089 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1092 if (needs_view_index
)
1093 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1094 &ctx
->abi
.view_index
);
1096 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
1097 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
1098 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1099 &ctx
->gs_vtx_offset
[0]);
1100 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1101 &ctx
->gs_vtx_offset
[1]);
1102 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1103 &ctx
->abi
.gs_prim_id
);
1104 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1105 &ctx
->gs_vtx_offset
[2]);
1106 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1107 &ctx
->gs_vtx_offset
[3]);
1108 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1109 &ctx
->gs_vtx_offset
[4]);
1110 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1111 &ctx
->gs_vtx_offset
[5]);
1112 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1113 &ctx
->abi
.gs_invocation_id
);
1116 case MESA_SHADER_FRAGMENT
:
1117 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1120 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1121 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1122 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1123 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1124 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1125 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1126 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1127 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1128 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1129 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1130 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1131 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1132 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1133 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1134 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1135 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1136 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1139 unreachable("Shader stage not implemented");
1142 ctx
->main_function
= create_llvm_function(
1143 ctx
->context
, ctx
->ac
.module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1144 ctx
->max_workgroup_size
, ctx
->options
);
1145 set_llvm_calling_convention(ctx
->main_function
, stage
);
1148 ctx
->shader_info
->num_input_vgprs
= 0;
1149 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1151 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1153 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1154 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1156 assign_arguments(ctx
->main_function
, &args
);
1160 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1161 set_loc_shader_ptr(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1163 if (ctx
->options
->supports_spill
) {
1164 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1165 LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_CONST
),
1166 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1167 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1168 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1172 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1173 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1174 if (has_previous_stage
)
1177 set_global_input_locs(ctx
, &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1180 case MESA_SHADER_COMPUTE
:
1181 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1182 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1186 case MESA_SHADER_VERTEX
:
1187 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1188 previous_stage
, &user_sgpr_idx
);
1189 if (ctx
->abi
.view_index
)
1190 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1192 case MESA_SHADER_TESS_CTRL
:
1193 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1194 previous_stage
, &user_sgpr_idx
);
1195 if (ctx
->abi
.view_index
)
1196 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1198 case MESA_SHADER_TESS_EVAL
:
1199 if (ctx
->abi
.view_index
)
1200 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1202 case MESA_SHADER_GEOMETRY
:
1203 if (has_previous_stage
) {
1204 if (previous_stage
== MESA_SHADER_VERTEX
)
1205 set_vs_specific_input_locs(ctx
, stage
,
1210 if (ctx
->abi
.view_index
)
1211 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1213 case MESA_SHADER_FRAGMENT
:
1216 unreachable("Shader stage not implemented");
1219 if (stage
== MESA_SHADER_TESS_CTRL
||
1220 (stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_ls
) ||
1221 /* GFX9 has the ESGS ring buffer in LDS. */
1222 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
1223 ac_declare_lds_as_pointer(&ctx
->ac
);
1226 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1231 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
1232 unsigned desc_set
, unsigned binding
)
1234 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1235 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
1236 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
1237 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
1238 unsigned base_offset
= layout
->binding
[binding
].offset
;
1239 LLVMValueRef offset
, stride
;
1241 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
1242 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
1243 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
1244 layout
->binding
[binding
].dynamic_offset_offset
;
1245 desc_ptr
= ctx
->abi
.push_constants
;
1246 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
1247 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1249 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
1251 offset
= ac_build_imad(&ctx
->ac
, index
, stride
,
1252 LLVMConstInt(ctx
->ac
.i32
, base_offset
, false));
1254 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
1255 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
1256 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1262 /* The offchip buffer layout for TCS->TES is
1264 * - attribute 0 of patch 0 vertex 0
1265 * - attribute 0 of patch 0 vertex 1
1266 * - attribute 0 of patch 0 vertex 2
1268 * - attribute 0 of patch 1 vertex 0
1269 * - attribute 0 of patch 1 vertex 1
1271 * - attribute 1 of patch 0 vertex 0
1272 * - attribute 1 of patch 0 vertex 1
1274 * - per patch attribute 0 of patch 0
1275 * - per patch attribute 0 of patch 1
1278 * Note that every attribute has 4 components.
1280 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
1282 uint32_t num_patches
= ctx
->tcs_num_patches
;
1283 uint32_t num_tcs_outputs
;
1284 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
1285 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
1287 num_tcs_outputs
= ctx
->options
->key
.tes
.tcs_num_outputs
;
1289 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
1290 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
1292 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
1295 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
1296 LLVMValueRef vertex_index
)
1298 LLVMValueRef param_stride
;
1300 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
* ctx
->tcs_num_patches
, false);
1302 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
1303 return param_stride
;
1306 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
1307 LLVMValueRef vertex_index
,
1308 LLVMValueRef param_index
)
1310 LLVMValueRef base_addr
;
1311 LLVMValueRef param_stride
, constant16
;
1312 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
1313 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
, false);
1314 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1315 param_stride
= calc_param_stride(ctx
, vertex_index
);
1317 base_addr
= ac_build_imad(&ctx
->ac
, rel_patch_id
,
1318 vertices_per_patch
, vertex_index
);
1320 base_addr
= rel_patch_id
;
1323 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1324 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
1325 param_stride
, ""), "");
1327 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
1329 if (!vertex_index
) {
1330 LLVMValueRef patch_data_offset
= get_non_vertex_index_offset(ctx
);
1332 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1333 patch_data_offset
, "");
1338 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct radv_shader_context
*ctx
,
1340 unsigned const_index
,
1342 LLVMValueRef vertex_index
,
1343 LLVMValueRef indir_index
)
1345 LLVMValueRef param_index
;
1348 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
1351 if (const_index
&& !is_compact
)
1352 param
+= const_index
;
1353 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
1355 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
1359 get_dw_address(struct radv_shader_context
*ctx
,
1360 LLVMValueRef dw_addr
,
1362 unsigned const_index
,
1363 bool compact_const_index
,
1364 LLVMValueRef vertex_index
,
1365 LLVMValueRef stride
,
1366 LLVMValueRef indir_index
)
1371 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1372 LLVMBuildMul(ctx
->ac
.builder
,
1378 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1379 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
1380 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
1381 else if (const_index
&& !compact_const_index
)
1382 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1383 LLVMConstInt(ctx
->ac
.i32
, const_index
* 4, false), "");
1385 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1386 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
1388 if (const_index
&& compact_const_index
)
1389 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1390 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
1395 load_tcs_varyings(struct ac_shader_abi
*abi
,
1397 LLVMValueRef vertex_index
,
1398 LLVMValueRef indir_index
,
1399 unsigned const_index
,
1401 unsigned driver_location
,
1403 unsigned num_components
,
1408 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1409 LLVMValueRef dw_addr
, stride
;
1410 LLVMValueRef value
[4], result
;
1411 unsigned param
= shader_io_get_unique_index(location
);
1414 uint32_t input_vertex_size
= (ctx
->tcs_num_inputs
* 16) / 4;
1415 stride
= LLVMConstInt(ctx
->ac
.i32
, input_vertex_size
, false);
1416 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1419 stride
= get_tcs_out_vertex_stride(ctx
);
1420 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1422 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1427 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1430 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1431 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1432 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1435 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1440 store_tcs_output(struct ac_shader_abi
*abi
,
1441 const nir_variable
*var
,
1442 LLVMValueRef vertex_index
,
1443 LLVMValueRef param_index
,
1444 unsigned const_index
,
1448 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1449 const unsigned location
= var
->data
.location
;
1450 const unsigned component
= var
->data
.location_frac
;
1451 const bool is_patch
= var
->data
.patch
;
1452 const bool is_compact
= var
->data
.compact
;
1453 LLVMValueRef dw_addr
;
1454 LLVMValueRef stride
= NULL
;
1455 LLVMValueRef buf_addr
= NULL
;
1457 bool store_lds
= true;
1460 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
1463 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
1467 param
= shader_io_get_unique_index(location
);
1468 if (location
== VARYING_SLOT_CLIP_DIST0
&&
1469 is_compact
&& const_index
> 3) {
1475 stride
= get_tcs_out_vertex_stride(ctx
);
1476 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1478 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1481 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1483 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
1484 vertex_index
, param_index
);
1486 bool is_tess_factor
= false;
1487 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
1488 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
1489 is_tess_factor
= true;
1491 unsigned base
= is_compact
? const_index
: 0;
1492 for (unsigned chan
= 0; chan
< 8; chan
++) {
1493 if (!(writemask
& (1 << chan
)))
1495 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1496 value
= ac_to_integer(&ctx
->ac
, value
);
1497 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
1499 if (store_lds
|| is_tess_factor
) {
1500 LLVMValueRef dw_addr_chan
=
1501 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1502 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
1503 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
1506 if (!is_tess_factor
&& writemask
!= 0xF)
1507 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
1508 buf_addr
, ctx
->oc_lds
,
1509 4 * (base
+ chan
), 1, 0, true, false);
1512 if (writemask
== 0xF) {
1513 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
1514 buf_addr
, ctx
->oc_lds
,
1515 (base
* 4), 1, 0, true, false);
1520 load_tes_input(struct ac_shader_abi
*abi
,
1522 LLVMValueRef vertex_index
,
1523 LLVMValueRef param_index
,
1524 unsigned const_index
,
1526 unsigned driver_location
,
1528 unsigned num_components
,
1533 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1534 LLVMValueRef buf_addr
;
1535 LLVMValueRef result
;
1536 unsigned param
= shader_io_get_unique_index(location
);
1538 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
1543 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
1544 is_compact
, vertex_index
, param_index
);
1546 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
1547 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
1549 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
1550 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
1551 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
1556 load_gs_input(struct ac_shader_abi
*abi
,
1558 unsigned driver_location
,
1560 unsigned num_components
,
1561 unsigned vertex_index
,
1562 unsigned const_index
,
1565 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1566 LLVMValueRef vtx_offset
;
1567 unsigned param
, vtx_offset_param
;
1568 LLVMValueRef value
[4], result
;
1570 vtx_offset_param
= vertex_index
;
1571 assert(vtx_offset_param
< 6);
1572 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
1573 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1575 param
= shader_io_get_unique_index(location
);
1577 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1578 if (ctx
->ac
.chip_class
>= GFX9
) {
1579 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1580 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1581 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
1582 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1584 LLVMValueRef soffset
=
1585 LLVMConstInt(ctx
->ac
.i32
,
1586 (param
* 4 + i
+ const_index
) * 256,
1589 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
1592 vtx_offset
, soffset
,
1593 0, 1, 0, true, false);
1596 if (ac_get_type_size(type
) == 2) {
1597 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i32
, "");
1598 value
[i
] = LLVMBuildTrunc(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i16
, "");
1600 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], type
, "");
1602 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1603 result
= ac_to_integer(&ctx
->ac
, result
);
1608 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
1610 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1611 ac_build_kill_if_false(&ctx
->ac
, visible
);
1614 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
1615 enum glsl_interp_mode interp
, unsigned location
)
1617 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1620 case INTERP_MODE_FLAT
:
1623 case INTERP_MODE_SMOOTH
:
1624 case INTERP_MODE_NONE
:
1625 if (location
== INTERP_CENTER
)
1626 return ctx
->persp_center
;
1627 else if (location
== INTERP_CENTROID
)
1628 return ctx
->persp_centroid
;
1629 else if (location
== INTERP_SAMPLE
)
1630 return ctx
->persp_sample
;
1632 case INTERP_MODE_NOPERSPECTIVE
:
1633 if (location
== INTERP_CENTER
)
1634 return ctx
->linear_center
;
1635 else if (location
== INTERP_CENTROID
)
1636 return ctx
->linear_centroid
;
1637 else if (location
== INTERP_SAMPLE
)
1638 return ctx
->linear_sample
;
1645 radv_get_sample_pos_offset(uint32_t num_samples
)
1647 uint32_t sample_pos_offset
= 0;
1649 switch (num_samples
) {
1651 sample_pos_offset
= 1;
1654 sample_pos_offset
= 3;
1657 sample_pos_offset
= 7;
1662 return sample_pos_offset
;
1665 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
1666 LLVMValueRef sample_id
)
1668 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1670 LLVMValueRef result
;
1671 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
1673 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
1674 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
1676 uint32_t sample_pos_offset
=
1677 radv_get_sample_pos_offset(ctx
->options
->key
.fs
.num_samples
);
1680 LLVMBuildAdd(ctx
->ac
.builder
, sample_id
,
1681 LLVMConstInt(ctx
->ac
.i32
, sample_pos_offset
, false), "");
1682 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
1688 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1690 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1691 uint8_t log2_ps_iter_samples
;
1693 if (ctx
->shader_info
->info
.ps
.force_persample
) {
1694 log2_ps_iter_samples
=
1695 util_logbase2(ctx
->options
->key
.fs
.num_samples
);
1697 log2_ps_iter_samples
= ctx
->options
->key
.fs
.log2_ps_iter_samples
;
1700 /* The bit pattern matches that used by fixed function fragment
1702 static const uint16_t ps_iter_masks
[] = {
1703 0xffff, /* not used */
1709 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
1711 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
1713 LLVMValueRef result
, sample_id
;
1714 sample_id
= ac_unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
1715 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
1716 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
1722 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
1724 LLVMValueRef gs_next_vertex
;
1725 LLVMValueRef can_emit
;
1726 unsigned offset
= 0;
1727 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1729 /* Write vertex attribute values to GSVS ring */
1730 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
1731 ctx
->gs_next_vertex
[stream
],
1734 /* If this thread has already emitted the declared maximum number of
1735 * vertices, kill it: excessive vertex emissions are not supposed to
1736 * have any effect, and GS threads have no externally observable
1737 * effects other than emitting vertices.
1739 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
1740 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
1741 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1743 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1744 unsigned output_usage_mask
=
1745 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
1746 uint8_t output_stream
=
1747 ctx
->shader_info
->info
.gs
.output_streams
[i
];
1748 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1749 int length
= util_last_bit(output_usage_mask
);
1751 if (!(ctx
->output_mask
& (1ull << i
)) ||
1752 output_stream
!= stream
)
1755 for (unsigned j
= 0; j
< length
; j
++) {
1756 if (!(output_usage_mask
& (1 << j
)))
1759 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1761 LLVMValueRef voffset
=
1762 LLVMConstInt(ctx
->ac
.i32
, offset
*
1763 ctx
->gs_max_out_vertices
, false);
1767 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1768 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1770 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1771 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1773 ac_build_buffer_store_dword(&ctx
->ac
,
1774 ctx
->gsvs_ring
[stream
],
1776 voffset
, ctx
->gs2vs_offset
, 0,
1781 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1783 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
1785 ac_build_sendmsg(&ctx
->ac
,
1786 AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
1791 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1793 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1794 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1798 load_tess_coord(struct ac_shader_abi
*abi
)
1800 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1802 LLVMValueRef coord
[4] = {
1809 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
1810 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1811 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1813 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1817 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1819 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1820 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
1824 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1826 return abi
->base_vertex
;
1829 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1830 LLVMValueRef buffer_ptr
, bool write
)
1832 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1833 LLVMValueRef result
;
1835 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1837 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1838 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1843 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1845 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1846 LLVMValueRef result
;
1848 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1850 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1851 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1856 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1857 unsigned descriptor_set
,
1858 unsigned base_index
,
1859 unsigned constant_index
,
1861 enum ac_descriptor_type desc_type
,
1862 bool image
, bool write
,
1865 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1866 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1867 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
1868 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1869 unsigned offset
= binding
->offset
;
1870 unsigned stride
= binding
->size
;
1872 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1875 assert(base_index
< layout
->binding_count
);
1877 switch (desc_type
) {
1879 type
= ctx
->ac
.v8i32
;
1883 type
= ctx
->ac
.v8i32
;
1887 case AC_DESC_SAMPLER
:
1888 type
= ctx
->ac
.v4i32
;
1889 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
1894 case AC_DESC_BUFFER
:
1895 type
= ctx
->ac
.v4i32
;
1899 unreachable("invalid desc_type\n");
1902 offset
+= constant_index
* stride
;
1904 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1905 (!index
|| binding
->immutable_samplers_equal
)) {
1906 if (binding
->immutable_samplers_equal
)
1909 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1911 LLVMValueRef constants
[] = {
1912 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1913 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1914 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1915 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1917 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1920 assert(stride
% type_size
== 0);
1923 index
= ctx
->ac
.i32_0
;
1925 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1927 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
1928 list
= LLVMBuildPointerCast(builder
, list
,
1929 ac_array_in_const32_addr_space(type
), "");
1931 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
1934 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1935 * so we may need to fix it up. */
1937 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1938 unsigned adjustment
,
1941 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1944 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1946 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1947 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1949 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1951 /* For the integer-like cases, do a natural sign extension.
1953 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1954 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1957 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1958 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1959 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1960 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1962 /* Convert back to the right type. */
1963 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1965 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1966 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1967 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1968 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1969 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1970 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1977 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1978 struct nir_variable
*variable
)
1980 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
1981 LLVMValueRef t_offset
;
1982 LLVMValueRef t_list
;
1984 LLVMValueRef buffer_index
;
1985 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1986 uint8_t input_usage_mask
=
1987 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
1988 unsigned num_channels
= util_last_bit(input_usage_mask
);
1990 variable
->data
.driver_location
= variable
->data
.location
* 4;
1992 enum glsl_base_type type
= glsl_get_base_type(variable
->type
);
1993 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1994 LLVMValueRef output
[4];
1995 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1997 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1998 uint32_t divisor
= ctx
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
2001 buffer_index
= ctx
->abi
.instance_id
;
2004 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
2005 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
2008 if (ctx
->options
->key
.vs
.as_ls
) {
2009 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
2010 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2012 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
2013 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2016 buffer_index
= ctx
->ac
.i32_0
;
2019 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.start_instance
, buffer_index
, "");
2021 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
2022 ctx
->abi
.base_vertex
, "");
2023 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_index
, false);
2025 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
2027 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
2030 num_channels
, false, true);
2032 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
2034 for (unsigned chan
= 0; chan
< 4; chan
++) {
2035 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2036 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
2037 if (type
== GLSL_TYPE_FLOAT16
) {
2038 output
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f32
, "");
2039 output
[chan
] = LLVMBuildFPTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f16
, "");
2043 unsigned alpha_adjust
= (ctx
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
2044 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
2046 for (unsigned chan
= 0; chan
< 4; chan
++) {
2047 output
[chan
] = ac_to_integer(&ctx
->ac
, output
[chan
]);
2048 if (type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
)
2049 output
[chan
] = LLVMBuildTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.i16
, "");
2051 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] = output
[chan
];
2056 static void interp_fs_input(struct radv_shader_context
*ctx
,
2058 LLVMValueRef interp_param
,
2059 LLVMValueRef prim_mask
,
2060 LLVMValueRef result
[4])
2062 LLVMValueRef attr_number
;
2065 bool interp
= !LLVMIsUndef(interp_param
);
2067 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
2069 /* fs.constant returns the param from the middle vertex, so it's not
2070 * really useful for flat shading. It's meant to be used for custom
2071 * interpolation (but the intrinsic can't fetch from the other two
2074 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
2075 * to do the right thing. The only reason we use fs.constant is that
2076 * fs.interp cannot be used on integers, because they can be equal
2080 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
2083 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2085 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2089 for (chan
= 0; chan
< 4; chan
++) {
2090 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2093 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
2098 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
2099 LLVMConstInt(ctx
->ac
.i32
, 2, false),
2103 result
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, result
[chan
], ctx
->ac
.i32
, "");
2104 result
[chan
] = LLVMBuildTruncOrBitCast(ctx
->ac
.builder
, result
[chan
], LLVMTypeOf(interp_param
), "");
2110 handle_fs_input_decl(struct radv_shader_context
*ctx
,
2111 struct nir_variable
*variable
)
2113 int idx
= variable
->data
.location
;
2114 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2115 LLVMValueRef interp
= NULL
;
2118 variable
->data
.driver_location
= idx
* 4;
2119 mask
= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
2121 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
2122 unsigned interp_type
;
2123 if (variable
->data
.sample
)
2124 interp_type
= INTERP_SAMPLE
;
2125 else if (variable
->data
.centroid
)
2126 interp_type
= INTERP_CENTROID
;
2128 interp_type
= INTERP_CENTER
;
2130 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
2132 bool is_16bit
= glsl_type_is_16bit(glsl_without_array(variable
->type
));
2133 LLVMTypeRef type
= is_16bit
? ctx
->ac
.i16
: ctx
->ac
.i32
;
2135 interp
= LLVMGetUndef(type
);
2137 for (unsigned i
= 0; i
< attrib_count
; ++i
)
2138 ctx
->inputs
[ac_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
2140 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2141 /* Do not account for the number of components inside the array
2142 * of clip/cull distances because this might wrongly set other
2143 * bits like primitive ID or layer.
2145 mask
= 1ull << VARYING_SLOT_CLIP_DIST0
;
2148 ctx
->input_mask
|= mask
;
2152 handle_vs_inputs(struct radv_shader_context
*ctx
,
2153 struct nir_shader
*nir
) {
2154 nir_foreach_variable(variable
, &nir
->inputs
)
2155 handle_vs_input_decl(ctx
, variable
);
2159 prepare_interp_optimize(struct radv_shader_context
*ctx
,
2160 struct nir_shader
*nir
)
2162 bool uses_center
= false;
2163 bool uses_centroid
= false;
2164 nir_foreach_variable(variable
, &nir
->inputs
) {
2165 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
2166 variable
->data
.sample
)
2169 if (variable
->data
.centroid
)
2170 uses_centroid
= true;
2175 if (uses_center
&& uses_centroid
) {
2176 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
2177 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
2178 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
2183 handle_fs_inputs(struct radv_shader_context
*ctx
,
2184 struct nir_shader
*nir
)
2186 prepare_interp_optimize(ctx
, nir
);
2188 nir_foreach_variable(variable
, &nir
->inputs
)
2189 handle_fs_input_decl(ctx
, variable
);
2193 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
2194 ctx
->shader_info
->info
.needs_multiview_view_index
) {
2195 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
2196 ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)] = LLVMGetUndef(ctx
->ac
.i32
);
2199 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
2200 LLVMValueRef interp_param
;
2201 LLVMValueRef
*inputs
= ctx
->inputs
+ac_llvm_reg_index_soa(i
, 0);
2203 if (!(ctx
->input_mask
& (1ull << i
)))
2206 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
2207 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
2208 interp_param
= *inputs
;
2209 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
2212 if (LLVMIsUndef(interp_param
))
2213 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
2214 if (i
>= VARYING_SLOT_VAR0
)
2215 ctx
->abi
.fs_input_attr_indices
[i
- VARYING_SLOT_VAR0
] = index
;
2217 } else if (i
== VARYING_SLOT_CLIP_DIST0
) {
2218 int length
= ctx
->shader_info
->info
.ps
.num_input_clips_culls
;
2220 for (unsigned j
= 0; j
< length
; j
+= 4) {
2221 inputs
= ctx
->inputs
+ ac_llvm_reg_index_soa(i
, j
);
2223 interp_param
= *inputs
;
2224 interp_fs_input(ctx
, index
, interp_param
,
2225 ctx
->abi
.prim_mask
, inputs
);
2228 } else if (i
== VARYING_SLOT_POS
) {
2229 for(int i
= 0; i
< 3; ++i
)
2230 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
2232 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2233 ctx
->abi
.frag_pos
[3]);
2236 ctx
->shader_info
->fs
.num_interp
= index
;
2237 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
2239 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
2240 ctx
->abi
.view_index
= ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2244 scan_shader_output_decl(struct radv_shader_context
*ctx
,
2245 struct nir_variable
*variable
,
2246 struct nir_shader
*shader
,
2247 gl_shader_stage stage
)
2249 int idx
= variable
->data
.location
+ variable
->data
.index
;
2250 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2251 uint64_t mask_attribs
;
2253 variable
->data
.driver_location
= idx
* 4;
2255 /* tess ctrl has it's own load/store paths for outputs */
2256 if (stage
== MESA_SHADER_TESS_CTRL
)
2259 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
2260 if (stage
== MESA_SHADER_VERTEX
||
2261 stage
== MESA_SHADER_TESS_EVAL
||
2262 stage
== MESA_SHADER_GEOMETRY
) {
2263 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2264 if (stage
== MESA_SHADER_VERTEX
) {
2265 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2266 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2267 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2269 if (stage
== MESA_SHADER_TESS_EVAL
) {
2270 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2271 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2272 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2275 mask_attribs
= 1ull << idx
;
2279 ctx
->output_mask
|= mask_attribs
;
2283 /* Initialize arguments for the shader export intrinsic */
2285 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
2286 LLVMValueRef
*values
,
2287 unsigned enabled_channels
,
2289 struct ac_export_args
*args
)
2291 /* Specify the channels that are enabled. */
2292 args
->enabled_channels
= enabled_channels
;
2294 /* Specify whether the EXEC mask represents the valid mask */
2295 args
->valid_mask
= 0;
2297 /* Specify whether this is the last export */
2300 /* Specify the target we are exporting */
2301 args
->target
= target
;
2303 args
->compr
= false;
2304 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
2305 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
2306 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
2307 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
2312 bool is_16bit
= ac_get_type_size(LLVMTypeOf(values
[0])) == 2;
2313 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2314 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
2315 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
2316 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
2317 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
2320 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2321 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2322 unsigned bits
, bool hi
) = NULL
;
2324 switch(col_format
) {
2325 case V_028714_SPI_SHADER_ZERO
:
2326 args
->enabled_channels
= 0; /* writemask */
2327 args
->target
= V_008DFC_SQ_EXP_NULL
;
2330 case V_028714_SPI_SHADER_32_R
:
2331 args
->enabled_channels
= 1;
2332 args
->out
[0] = values
[0];
2335 case V_028714_SPI_SHADER_32_GR
:
2336 args
->enabled_channels
= 0x3;
2337 args
->out
[0] = values
[0];
2338 args
->out
[1] = values
[1];
2341 case V_028714_SPI_SHADER_32_AR
:
2342 args
->enabled_channels
= 0x9;
2343 args
->out
[0] = values
[0];
2344 args
->out
[3] = values
[3];
2347 case V_028714_SPI_SHADER_FP16_ABGR
:
2348 args
->enabled_channels
= 0x5;
2349 packf
= ac_build_cvt_pkrtz_f16
;
2351 for (unsigned chan
= 0; chan
< 4; chan
++)
2352 values
[chan
] = LLVMBuildFPExt(ctx
->ac
.builder
,
2358 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2359 args
->enabled_channels
= 0x5;
2360 packf
= ac_build_cvt_pknorm_u16
;
2363 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2364 args
->enabled_channels
= 0x5;
2365 packf
= ac_build_cvt_pknorm_i16
;
2368 case V_028714_SPI_SHADER_UINT16_ABGR
:
2369 args
->enabled_channels
= 0x5;
2370 packi
= ac_build_cvt_pk_u16
;
2372 for (unsigned chan
= 0; chan
< 4; chan
++)
2373 values
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
,
2379 case V_028714_SPI_SHADER_SINT16_ABGR
:
2380 args
->enabled_channels
= 0x5;
2381 packi
= ac_build_cvt_pk_i16
;
2383 for (unsigned chan
= 0; chan
< 4; chan
++)
2384 values
[chan
] = LLVMBuildSExt(ctx
->ac
.builder
,
2391 case V_028714_SPI_SHADER_32_ABGR
:
2392 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2396 /* Pack f16 or norm_i16/u16. */
2398 for (chan
= 0; chan
< 2; chan
++) {
2399 LLVMValueRef pack_args
[2] = {
2401 values
[2 * chan
+ 1]
2403 LLVMValueRef packed
;
2405 packed
= packf(&ctx
->ac
, pack_args
);
2406 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2408 args
->compr
= 1; /* COMPR flag */
2413 for (chan
= 0; chan
< 2; chan
++) {
2414 LLVMValueRef pack_args
[2] = {
2415 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2416 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2418 LLVMValueRef packed
;
2420 packed
= packi(&ctx
->ac
, pack_args
,
2421 is_int8
? 8 : is_int10
? 10 : 16,
2423 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2425 args
->compr
= 1; /* COMPR flag */
2431 for (unsigned chan
= 0; chan
< 4; chan
++) {
2432 values
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i16
, "");
2433 args
->out
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i32
, "");
2436 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2438 for (unsigned i
= 0; i
< 4; ++i
) {
2439 if (!(args
->enabled_channels
& (1 << i
)))
2442 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
2447 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
2448 LLVMValueRef
*values
, unsigned enabled_channels
)
2450 struct ac_export_args args
;
2452 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
2453 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2454 ac_build_export(&ctx
->ac
, &args
);
2458 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
2460 LLVMValueRef output
=
2461 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
2463 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
2467 radv_emit_stream_output(struct radv_shader_context
*ctx
,
2468 LLVMValueRef
const *so_buffers
,
2469 LLVMValueRef
const *so_write_offsets
,
2470 const struct radv_stream_output
*output
)
2472 unsigned num_comps
= util_bitcount(output
->component_mask
);
2473 unsigned loc
= output
->location
;
2474 unsigned buf
= output
->buffer
;
2475 unsigned offset
= output
->offset
;
2477 LLVMValueRef out
[4];
2479 assert(num_comps
&& num_comps
<= 4);
2480 if (!num_comps
|| num_comps
> 4)
2483 /* Get the first component. */
2484 start
= ffs(output
->component_mask
) - 1;
2486 /* Load the output as int. */
2487 for (int i
= 0; i
< num_comps
; i
++) {
2488 out
[i
] = ac_to_integer(&ctx
->ac
,
2489 radv_load_output(ctx
, loc
, start
+ i
));
2492 /* Pack the output. */
2493 LLVMValueRef vdata
= NULL
;
2495 switch (num_comps
) {
2496 case 1: /* as i32 */
2499 case 2: /* as v2i32 */
2500 case 3: /* as v4i32 (aligned to 4) */
2501 out
[3] = LLVMGetUndef(ctx
->ac
.i32
);
2503 case 4: /* as v4i32 */
2504 vdata
= ac_build_gather_values(&ctx
->ac
, out
,
2505 util_next_power_of_two(num_comps
));
2509 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf
],
2510 vdata
, num_comps
, so_write_offsets
[buf
],
2511 ctx
->ac
.i32_0
, offset
,
2516 radv_emit_streamout(struct radv_shader_context
*ctx
, unsigned stream
)
2518 struct ac_build_if_state if_ctx
;
2521 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2522 assert(ctx
->streamout_config
);
2523 LLVMValueRef so_vtx_count
=
2524 ac_build_bfe(&ctx
->ac
, ctx
->streamout_config
,
2525 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2526 LLVMConstInt(ctx
->ac
.i32
, 7, false), false);
2528 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2530 /* can_emit = tid < so_vtx_count; */
2531 LLVMValueRef can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
2532 tid
, so_vtx_count
, "");
2534 /* Emit the streamout code conditionally. This actually avoids
2535 * out-of-bounds buffer access. The hw tells us via the SGPR
2536 * (so_vtx_count) which threads are allowed to emit streamout data.
2538 ac_nir_build_if(&if_ctx
, ctx
, can_emit
);
2540 /* The buffer offset is computed as follows:
2541 * ByteOffset = streamout_offset[buffer_id]*4 +
2542 * (streamout_write_index + thread_id)*stride[buffer_id] +
2545 LLVMValueRef so_write_index
= ctx
->streamout_write_idx
;
2547 /* Compute (streamout_write_index + thread_id). */
2549 LLVMBuildAdd(ctx
->ac
.builder
, so_write_index
, tid
, "");
2551 /* Load the descriptor and compute the write offset for each
2554 LLVMValueRef so_write_offset
[4] = {};
2555 LLVMValueRef so_buffers
[4] = {};
2556 LLVMValueRef buf_ptr
= ctx
->streamout_buffers
;
2558 for (i
= 0; i
< 4; i
++) {
2559 uint16_t stride
= ctx
->shader_info
->info
.so
.strides
[i
];
2564 LLVMValueRef offset
=
2565 LLVMConstInt(ctx
->ac
.i32
, i
, false);
2567 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
,
2570 LLVMValueRef so_offset
= ctx
->streamout_offset
[i
];
2572 so_offset
= LLVMBuildMul(ctx
->ac
.builder
, so_offset
,
2573 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
2575 so_write_offset
[i
] =
2576 ac_build_imad(&ctx
->ac
, so_write_index
,
2577 LLVMConstInt(ctx
->ac
.i32
,
2582 /* Write streamout data. */
2583 for (i
= 0; i
< ctx
->shader_info
->info
.so
.num_outputs
; i
++) {
2584 struct radv_stream_output
*output
=
2585 &ctx
->shader_info
->info
.so
.outputs
[i
];
2587 if (stream
!= output
->stream
)
2590 radv_emit_stream_output(ctx
, so_buffers
,
2591 so_write_offset
, output
);
2594 ac_nir_build_endif(&if_ctx
);
2598 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2599 bool export_prim_id
, bool export_layer_id
,
2600 struct radv_vs_output_info
*outinfo
)
2602 uint32_t param_count
= 0;
2604 unsigned pos_idx
, num_pos_exports
= 0;
2605 struct ac_export_args args
, pos_args
[4] = {};
2606 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2609 if (ctx
->options
->key
.has_multiview_view_index
) {
2610 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2612 for(unsigned i
= 0; i
< 4; ++i
)
2613 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2614 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2617 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
2618 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2621 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2622 sizeof(outinfo
->vs_output_param_offset
));
2624 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
2625 unsigned output_usage_mask
, length
;
2626 LLVMValueRef slots
[8];
2629 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2630 !ctx
->is_gs_copy_shader
) {
2632 ctx
->shader_info
->info
.vs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2633 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2635 ctx
->shader_info
->info
.tes
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2637 assert(ctx
->is_gs_copy_shader
);
2639 ctx
->shader_info
->info
.gs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2642 length
= util_last_bit(output_usage_mask
);
2644 i
= VARYING_SLOT_CLIP_DIST0
;
2645 for (j
= 0; j
< length
; j
++)
2646 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2648 for (i
= length
; i
< 8; i
++)
2649 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
2652 target
= V_008DFC_SQ_EXP_POS
+ 3;
2653 si_llvm_init_export_args(ctx
, &slots
[4], 0xf, target
, &args
);
2654 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2655 &args
, sizeof(args
));
2658 target
= V_008DFC_SQ_EXP_POS
+ 2;
2659 si_llvm_init_export_args(ctx
, &slots
[0], 0xf, target
, &args
);
2660 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2661 &args
, sizeof(args
));
2663 /* Export the clip/cull distances values to the next stage. */
2664 radv_export_param(ctx
, param_count
, &slots
[0], 0xf);
2665 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST0
] = param_count
++;
2667 radv_export_param(ctx
, param_count
, &slots
[4], 0xf);
2668 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST1
] = param_count
++;
2672 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
2673 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
2674 for (unsigned j
= 0; j
< 4; j
++)
2675 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
2677 si_llvm_init_export_args(ctx
, pos_values
, 0xf, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2679 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
2680 outinfo
->writes_pointsize
= true;
2681 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
2684 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
2685 outinfo
->writes_layer
= true;
2686 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
2689 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
2690 outinfo
->writes_viewport_index
= true;
2691 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
2694 if (ctx
->shader_info
->info
.so
.num_outputs
&&
2695 !ctx
->is_gs_copy_shader
) {
2696 /* The GS copy shader emission already emits streamout. */
2697 radv_emit_streamout(ctx
, 0);
2700 if (outinfo
->writes_pointsize
||
2701 outinfo
->writes_layer
||
2702 outinfo
->writes_viewport_index
) {
2703 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
2704 (outinfo
->writes_layer
== true ? 4 : 0));
2705 pos_args
[1].valid_mask
= 0;
2706 pos_args
[1].done
= 0;
2707 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2708 pos_args
[1].compr
= 0;
2709 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2710 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2711 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2712 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2714 if (outinfo
->writes_pointsize
== true)
2715 pos_args
[1].out
[0] = psize_value
;
2716 if (outinfo
->writes_layer
== true)
2717 pos_args
[1].out
[2] = layer_value
;
2718 if (outinfo
->writes_viewport_index
== true) {
2719 if (ctx
->options
->chip_class
>= GFX9
) {
2720 /* GFX9 has the layer in out.z[10:0] and the viewport
2721 * index in out.z[19:16].
2723 LLVMValueRef v
= viewport_index_value
;
2724 v
= ac_to_integer(&ctx
->ac
, v
);
2725 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2726 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2728 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2729 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2731 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2732 pos_args
[1].enabled_channels
|= 1 << 2;
2734 pos_args
[1].out
[3] = viewport_index_value
;
2735 pos_args
[1].enabled_channels
|= 1 << 3;
2739 for (i
= 0; i
< 4; i
++) {
2740 if (pos_args
[i
].out
[0])
2745 for (i
= 0; i
< 4; i
++) {
2746 if (!pos_args
[i
].out
[0])
2749 /* Specify the target we are exporting */
2750 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2751 if (pos_idx
== num_pos_exports
)
2752 pos_args
[i
].done
= 1;
2753 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2756 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2757 LLVMValueRef values
[4];
2758 if (!(ctx
->output_mask
& (1ull << i
)))
2761 if (i
!= VARYING_SLOT_LAYER
&&
2762 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
2763 i
< VARYING_SLOT_VAR0
)
2766 for (unsigned j
= 0; j
< 4; j
++)
2767 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2769 unsigned output_usage_mask
;
2771 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2772 !ctx
->is_gs_copy_shader
) {
2774 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2775 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2777 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2779 assert(ctx
->is_gs_copy_shader
);
2781 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
2784 radv_export_param(ctx
, param_count
, values
, output_usage_mask
);
2786 outinfo
->vs_output_param_offset
[i
] = param_count
++;
2789 if (export_prim_id
) {
2790 LLVMValueRef values
[4];
2792 values
[0] = ctx
->vs_prim_id
;
2793 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
2794 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2795 for (unsigned j
= 1; j
< 4; j
++)
2796 values
[j
] = ctx
->ac
.f32_0
;
2798 radv_export_param(ctx
, param_count
, values
, 0x1);
2800 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
2801 outinfo
->export_prim_id
= true;
2804 if (export_layer_id
&& layer_value
) {
2805 LLVMValueRef values
[4];
2807 values
[0] = layer_value
;
2808 for (unsigned j
= 1; j
< 4; j
++)
2809 values
[j
] = ctx
->ac
.f32_0
;
2811 radv_export_param(ctx
, param_count
, values
, 0x1);
2813 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
++;
2816 outinfo
->pos_exports
= num_pos_exports
;
2817 outinfo
->param_exports
= param_count
;
2821 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2822 struct radv_es_output_info
*outinfo
)
2825 uint64_t max_output_written
= 0;
2826 LLVMValueRef lds_base
= NULL
;
2828 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2829 unsigned output_usage_mask
;
2833 if (!(ctx
->output_mask
& (1ull << i
)))
2836 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2838 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2840 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2842 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2845 if (i
== VARYING_SLOT_CLIP_DIST0
)
2846 length
= util_last_bit(output_usage_mask
);
2848 param_index
= shader_io_get_unique_index(i
);
2850 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
2853 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
2855 if (ctx
->ac
.chip_class
>= GFX9
) {
2856 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2857 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2858 LLVMValueRef wave_idx
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 24, 4);
2859 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2860 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2861 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
2862 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2863 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2866 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2867 LLVMValueRef dw_addr
= NULL
;
2868 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2869 unsigned output_usage_mask
;
2873 if (!(ctx
->output_mask
& (1ull << i
)))
2876 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2878 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2880 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2882 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2885 if (i
== VARYING_SLOT_CLIP_DIST0
)
2886 length
= util_last_bit(output_usage_mask
);
2888 param_index
= shader_io_get_unique_index(i
);
2891 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2892 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2896 for (j
= 0; j
< length
; j
++) {
2897 if (!(output_usage_mask
& (1 << j
)))
2900 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2901 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
2902 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2904 if (ctx
->ac
.chip_class
>= GFX9
) {
2905 LLVMValueRef dw_addr_offset
=
2906 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2907 LLVMConstInt(ctx
->ac
.i32
,
2910 ac_lds_store(&ctx
->ac
, dw_addr_offset
, out_val
);
2912 ac_build_buffer_store_dword(&ctx
->ac
,
2915 NULL
, ctx
->es2gs_offset
,
2916 (4 * param_index
+ j
) * 4,
2924 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2926 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2927 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->shader_info
->info
.vs
.ls_outputs_written
);
2928 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2929 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2930 vertex_dw_stride
, "");
2932 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2933 unsigned output_usage_mask
=
2934 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2935 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2938 if (!(ctx
->output_mask
& (1ull << i
)))
2941 if (i
== VARYING_SLOT_CLIP_DIST0
)
2942 length
= util_last_bit(output_usage_mask
);
2944 int param
= shader_io_get_unique_index(i
);
2945 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2946 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2948 for (unsigned j
= 0; j
< length
; j
++) {
2949 LLVMValueRef value
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2950 value
= ac_to_integer(&ctx
->ac
, value
);
2951 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
2952 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
2953 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2959 write_tess_factors(struct radv_shader_context
*ctx
)
2961 unsigned stride
, outer_comps
, inner_comps
;
2962 struct ac_build_if_state if_ctx
, inner_if_ctx
;
2963 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2964 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
2965 unsigned tess_inner_index
= 0, tess_outer_index
;
2966 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
2967 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2969 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
2971 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
2991 ac_nir_build_if(&if_ctx
, ctx
,
2992 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2993 invocation_id
, ctx
->ac
.i32_0
, ""));
2995 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
2998 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2999 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3000 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
3003 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
3004 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3005 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
3007 for (i
= 0; i
< 4; i
++) {
3008 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3009 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3013 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
3014 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
3015 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3017 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
3019 for (i
= 0; i
< outer_comps
; i
++) {
3021 ac_lds_load(&ctx
->ac
, lds_outer
);
3022 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3025 for (i
= 0; i
< inner_comps
; i
++) {
3026 inner
[i
] = out
[outer_comps
+i
] =
3027 ac_lds_load(&ctx
->ac
, lds_inner
);
3028 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
3033 /* Convert the outputs to vectors for stores. */
3034 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
3038 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
3041 buffer
= ctx
->hs_ring_tess_factor
;
3042 tf_base
= ctx
->tess_factor_offset
;
3043 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3044 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
3045 unsigned tf_offset
= 0;
3047 if (ctx
->options
->chip_class
<= VI
) {
3048 ac_nir_build_if(&inner_if_ctx
, ctx
,
3049 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3050 rel_patch_id
, ctx
->ac
.i32_0
, ""));
3052 /* Store the dynamic HS control word. */
3053 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3054 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
3055 1, ctx
->ac
.i32_0
, tf_base
,
3056 0, 1, 0, true, false);
3059 ac_nir_build_endif(&inner_if_ctx
);
3062 /* Store the tessellation factors. */
3063 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3064 MIN2(stride
, 4), byteoffset
, tf_base
,
3065 tf_offset
, 1, 0, true, false);
3067 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3068 stride
- 4, byteoffset
, tf_base
,
3069 16 + tf_offset
, 1, 0, true, false);
3071 //store to offchip for TES to read - only if TES reads them
3072 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
3073 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
3074 LLVMValueRef tf_inner_offset
;
3075 unsigned param_outer
, param_inner
;
3077 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
3078 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3079 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
3081 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
3082 util_next_power_of_two(outer_comps
));
3084 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
3085 outer_comps
, tf_outer_offset
,
3086 ctx
->oc_lds
, 0, 1, 0, true, false);
3088 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
3089 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3090 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
3092 inner_vec
= inner_comps
== 1 ? inner
[0] :
3093 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
3094 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
3095 inner_comps
, tf_inner_offset
,
3096 ctx
->oc_lds
, 0, 1, 0, true, false);
3099 ac_nir_build_endif(&if_ctx
);
3103 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
3105 write_tess_factors(ctx
);
3109 si_export_mrt_color(struct radv_shader_context
*ctx
,
3110 LLVMValueRef
*color
, unsigned index
,
3111 struct ac_export_args
*args
)
3114 si_llvm_init_export_args(ctx
, color
, 0xf,
3115 V_008DFC_SQ_EXP_MRT
+ index
, args
);
3116 if (!args
->enabled_channels
)
3117 return false; /* unnecessary NULL export */
3123 radv_export_mrt_z(struct radv_shader_context
*ctx
,
3124 LLVMValueRef depth
, LLVMValueRef stencil
,
3125 LLVMValueRef samplemask
)
3127 struct ac_export_args args
;
3129 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3131 ac_build_export(&ctx
->ac
, &args
);
3135 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
3138 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3139 struct ac_export_args color_args
[8];
3141 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3142 LLVMValueRef values
[4];
3144 if (!(ctx
->output_mask
& (1ull << i
)))
3147 if (i
< FRAG_RESULT_DATA0
)
3150 for (unsigned j
= 0; j
< 4; j
++)
3151 values
[j
] = ac_to_float(&ctx
->ac
,
3152 radv_load_output(ctx
, i
, j
));
3154 bool ret
= si_export_mrt_color(ctx
, values
,
3155 i
- FRAG_RESULT_DATA0
,
3156 &color_args
[index
]);
3161 /* Process depth, stencil, samplemask. */
3162 if (ctx
->shader_info
->info
.ps
.writes_z
) {
3163 depth
= ac_to_float(&ctx
->ac
,
3164 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
3166 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
3167 stencil
= ac_to_float(&ctx
->ac
,
3168 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
3170 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3171 samplemask
= ac_to_float(&ctx
->ac
,
3172 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
3175 /* Set the DONE bit on last non-null color export only if Z isn't
3179 !ctx
->shader_info
->info
.ps
.writes_z
&&
3180 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
3181 !ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3182 unsigned last
= index
- 1;
3184 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
3185 color_args
[last
].done
= 1; /* DONE bit */
3188 /* Export PS outputs. */
3189 for (unsigned i
= 0; i
< index
; i
++)
3190 ac_build_export(&ctx
->ac
, &color_args
[i
]);
3192 if (depth
|| stencil
|| samplemask
)
3193 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
3195 ac_build_export_null(&ctx
->ac
);
3199 emit_gs_epilogue(struct radv_shader_context
*ctx
)
3201 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
3205 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
3206 LLVMValueRef
*addrs
)
3208 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
3210 switch (ctx
->stage
) {
3211 case MESA_SHADER_VERTEX
:
3212 if (ctx
->options
->key
.vs
.as_ls
)
3213 handle_ls_outputs_post(ctx
);
3214 else if (ctx
->options
->key
.vs
.as_es
)
3215 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
3217 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
3218 ctx
->options
->key
.vs
.export_layer_id
,
3219 &ctx
->shader_info
->vs
.outinfo
);
3221 case MESA_SHADER_FRAGMENT
:
3222 handle_fs_outputs_post(ctx
);
3224 case MESA_SHADER_GEOMETRY
:
3225 emit_gs_epilogue(ctx
);
3227 case MESA_SHADER_TESS_CTRL
:
3228 handle_tcs_outputs_post(ctx
);
3230 case MESA_SHADER_TESS_EVAL
:
3231 if (ctx
->options
->key
.tes
.as_es
)
3232 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
3234 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
3235 ctx
->options
->key
.tes
.export_layer_id
,
3236 &ctx
->shader_info
->tes
.outinfo
);
3243 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
,
3244 LLVMPassManagerRef passmgr
,
3245 const struct radv_nir_compiler_options
*options
)
3247 LLVMRunPassManager(passmgr
, ctx
->ac
.module
);
3248 LLVMDisposeBuilder(ctx
->ac
.builder
);
3250 ac_llvm_context_dispose(&ctx
->ac
);
3254 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
3256 struct radv_vs_output_info
*outinfo
;
3258 switch (ctx
->stage
) {
3259 case MESA_SHADER_FRAGMENT
:
3260 case MESA_SHADER_COMPUTE
:
3261 case MESA_SHADER_TESS_CTRL
:
3262 case MESA_SHADER_GEOMETRY
:
3264 case MESA_SHADER_VERTEX
:
3265 if (ctx
->options
->key
.vs
.as_ls
||
3266 ctx
->options
->key
.vs
.as_es
)
3268 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
3270 case MESA_SHADER_TESS_EVAL
:
3271 if (ctx
->options
->key
.vs
.as_es
)
3273 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
3276 unreachable("Unhandled shader type");
3279 ac_optimize_vs_outputs(&ctx
->ac
,
3281 outinfo
->vs_output_param_offset
,
3283 &outinfo
->param_exports
);
3287 ac_setup_rings(struct radv_shader_context
*ctx
)
3289 if (ctx
->options
->chip_class
<= VI
&&
3290 (ctx
->stage
== MESA_SHADER_GEOMETRY
||
3291 ctx
->options
->key
.vs
.as_es
|| ctx
->options
->key
.tes
.as_es
)) {
3292 unsigned ring
= ctx
->stage
== MESA_SHADER_GEOMETRY
? RING_ESGS_GS
3294 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, ring
, false);
3296 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
,
3301 if (ctx
->is_gs_copy_shader
) {
3303 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3304 LLVMConstInt(ctx
->ac
.i32
,
3305 RING_GSVS_VS
, false));
3308 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3309 /* The conceptual layout of the GSVS ring is
3310 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3311 * but the real memory layout is swizzled across
3313 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3315 * Override the buffer descriptor accordingly.
3317 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->ac
.i64
, 2);
3318 uint64_t stream_offset
= 0;
3319 unsigned num_records
= 64;
3320 LLVMValueRef base_ring
;
3323 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3324 LLVMConstInt(ctx
->ac
.i32
,
3325 RING_GSVS_GS
, false));
3327 for (unsigned stream
= 0; stream
< 4; stream
++) {
3328 unsigned num_components
, stride
;
3329 LLVMValueRef ring
, tmp
;
3332 ctx
->shader_info
->info
.gs
.num_stream_output_components
[stream
];
3334 if (!num_components
)
3337 stride
= 4 * num_components
* ctx
->gs_max_out_vertices
;
3339 /* Limit on the stride field for <= CIK. */
3340 assert(stride
< (1 << 14));
3342 ring
= LLVMBuildBitCast(ctx
->ac
.builder
,
3343 base_ring
, v2i64
, "");
3344 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3345 ring
, ctx
->ac
.i32_0
, "");
3346 tmp
= LLVMBuildAdd(ctx
->ac
.builder
, tmp
,
3347 LLVMConstInt(ctx
->ac
.i64
,
3348 stream_offset
, 0), "");
3349 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
,
3350 ring
, tmp
, ctx
->ac
.i32_0
, "");
3352 stream_offset
+= stride
* 64;
3354 ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ring
,
3357 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ring
,
3359 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
,
3360 LLVMConstInt(ctx
->ac
.i32
,
3361 S_008F04_STRIDE(stride
), false), "");
3362 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
, tmp
,
3365 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
,
3366 LLVMConstInt(ctx
->ac
.i32
,
3367 num_records
, false),
3368 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3370 ctx
->gsvs_ring
[stream
] = ring
;
3374 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3375 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3376 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3377 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3382 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
3383 const struct nir_shader
*nir
)
3385 switch (nir
->info
.stage
) {
3386 case MESA_SHADER_TESS_CTRL
:
3387 return chip_class
>= CIK
? 128 : 64;
3388 case MESA_SHADER_GEOMETRY
:
3389 return chip_class
>= GFX9
? 128 : 64;
3390 case MESA_SHADER_COMPUTE
:
3396 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
3397 nir
->info
.cs
.local_size
[1] *
3398 nir
->info
.cs
.local_size
[2];
3399 return max_workgroup_size
;
3402 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3403 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3405 LLVMValueRef count
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 8, 8);
3406 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3408 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
3409 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
3410 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
3413 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
)
3415 for(int i
= 5; i
>= 0; --i
) {
3416 ctx
->gs_vtx_offset
[i
] = ac_unpack_param(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
3420 ctx
->gs_wave_id
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 16, 8);
3425 LLVMModuleRef
ac_translate_nir_to_llvm(struct ac_llvm_compiler
*ac_llvm
,
3426 struct nir_shader
*const *shaders
,
3428 struct radv_shader_variant_info
*shader_info
,
3429 const struct radv_nir_compiler_options
*options
)
3431 struct radv_shader_context ctx
= {0};
3433 ctx
.options
= options
;
3434 ctx
.shader_info
= shader_info
;
3436 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3437 ctx
.context
= ctx
.ac
.context
;
3438 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3440 enum ac_float_mode float_mode
=
3441 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3442 AC_FLOAT_MODE_DEFAULT
;
3444 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3446 memset(shader_info
, 0, sizeof(*shader_info
));
3448 for(int i
= 0; i
< shader_count
; ++i
)
3449 radv_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
3451 for (i
= 0; i
< RADV_UD_MAX_SETS
; i
++)
3452 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
3453 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
3454 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
3456 ctx
.max_workgroup_size
= 0;
3457 for (int i
= 0; i
< shader_count
; ++i
) {
3458 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
3459 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
3463 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
3464 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
3466 ctx
.abi
.inputs
= &ctx
.inputs
[0];
3467 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
3468 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
3469 ctx
.abi
.load_ubo
= radv_load_ubo
;
3470 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
3471 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
3472 ctx
.abi
.load_resource
= radv_load_resource
;
3473 ctx
.abi
.clamp_shadow_reference
= false;
3474 ctx
.abi
.gfx9_stride_size_workaround
= ctx
.ac
.chip_class
== GFX9
&& HAVE_LLVM
< 0x800;
3476 if (shader_count
>= 2)
3477 ac_init_exec_full_mask(&ctx
.ac
);
3479 if (ctx
.ac
.chip_class
== GFX9
&&
3480 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
3481 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
3483 for(int i
= 0; i
< shader_count
; ++i
) {
3484 ctx
.stage
= shaders
[i
]->info
.stage
;
3485 ctx
.output_mask
= 0;
3487 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3488 for (int i
= 0; i
< 4; i
++) {
3489 ctx
.gs_next_vertex
[i
] =
3490 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
3492 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
3493 ctx
.abi
.load_inputs
= load_gs_input
;
3494 ctx
.abi
.emit_primitive
= visit_end_primitive
;
3495 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3496 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
3497 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
3498 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
3499 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3500 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
3501 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3502 if (shader_count
== 1)
3503 ctx
.tcs_num_inputs
= ctx
.options
->key
.tcs
.num_inputs
;
3505 ctx
.tcs_num_inputs
= util_last_bit64(shader_info
->info
.vs
.ls_outputs_written
);
3506 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
3507 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3508 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
3509 ctx
.abi
.load_tess_varyings
= load_tes_input
;
3510 ctx
.abi
.load_tess_coord
= load_tess_coord
;
3511 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3512 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3513 ctx
.tcs_num_patches
= ctx
.options
->key
.tes
.num_patches
;
3514 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
3515 if (shader_info
->info
.vs
.needs_instance_id
) {
3516 if (ctx
.options
->key
.vs
.as_ls
) {
3517 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3518 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3520 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3521 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3524 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
3525 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
3526 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
3527 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
3528 ctx
.abi
.load_sample_position
= load_sample_position
;
3529 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
3530 ctx
.abi
.emit_kill
= radv_emit_kill
;
3534 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
3536 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
3537 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
3539 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3540 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
3541 shaders
[i
]->info
.cull_distance_array_size
> 4;
3542 ctx
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
3543 ctx
.max_gsvs_emit_size
= ctx
.gsvs_vertex_size
*
3544 shaders
[i
]->info
.gs
.vertices_out
;
3547 ac_setup_rings(&ctx
);
3549 LLVMBasicBlockRef merge_block
;
3550 if (shader_count
>= 2) {
3551 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
3552 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3553 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3555 LLVMValueRef count
= ac_unpack_param(&ctx
.ac
, ctx
.merged_wave_info
, 8 * i
, 8);
3556 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
3557 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
3558 thread_id
, count
, "");
3559 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
3561 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
3564 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
3565 handle_fs_inputs(&ctx
, shaders
[i
]);
3566 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
3567 handle_vs_inputs(&ctx
, shaders
[i
]);
3568 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
3569 prepare_gs_input_vgprs(&ctx
);
3571 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
]);
3573 if (shader_count
>= 2) {
3574 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
3575 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
3578 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3579 shader_info
->gs
.gsvs_vertex_size
= ctx
.gsvs_vertex_size
;
3580 shader_info
->gs
.max_gsvs_emit_size
= ctx
.max_gsvs_emit_size
;
3581 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3582 shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
3583 shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
3587 LLVMBuildRetVoid(ctx
.ac
.builder
);
3589 if (options
->dump_preoptir
)
3590 ac_dump_module(ctx
.ac
.module
);
3592 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3594 if (shader_count
== 1)
3595 ac_nir_eliminate_const_vs_outputs(&ctx
);
3597 if (options
->dump_shader
) {
3598 ctx
.shader_info
->private_mem_vgprs
=
3599 ac_count_scratch_private_memory(ctx
.main_function
);
3602 return ctx
.ac
.module
;
3605 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
3607 unsigned *retval
= (unsigned *)context
;
3608 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
3609 char *description
= LLVMGetDiagInfoDescription(di
);
3611 if (severity
== LLVMDSError
) {
3613 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
3617 LLVMDisposeMessage(description
);
3620 static unsigned ac_llvm_compile(LLVMModuleRef M
,
3621 struct ac_shader_binary
*binary
,
3622 struct ac_llvm_compiler
*ac_llvm
)
3624 unsigned retval
= 0;
3625 LLVMContextRef llvm_ctx
;
3627 /* Setup Diagnostic Handler*/
3628 llvm_ctx
= LLVMGetModuleContext(M
);
3630 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
3634 if (!radv_compile_to_binary(ac_llvm
, M
, binary
))
3639 static void ac_compile_llvm_module(struct ac_llvm_compiler
*ac_llvm
,
3640 LLVMModuleRef llvm_module
,
3641 struct ac_shader_binary
*binary
,
3642 struct ac_shader_config
*config
,
3643 struct radv_shader_variant_info
*shader_info
,
3644 gl_shader_stage stage
,
3645 const struct radv_nir_compiler_options
*options
)
3647 if (options
->dump_shader
)
3648 ac_dump_module(llvm_module
);
3650 memset(binary
, 0, sizeof(*binary
));
3652 if (options
->record_llvm_ir
) {
3653 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
3654 binary
->llvm_ir_string
= strdup(llvm_ir
);
3655 LLVMDisposeMessage(llvm_ir
);
3658 int v
= ac_llvm_compile(llvm_module
, binary
, ac_llvm
);
3660 fprintf(stderr
, "compile failed\n");
3663 if (options
->dump_shader
)
3664 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
3666 ac_shader_binary_read_config(binary
, config
, 0, options
->supports_spill
);
3668 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
3669 LLVMDisposeModule(llvm_module
);
3670 LLVMContextDispose(ctx
);
3672 if (stage
== MESA_SHADER_FRAGMENT
) {
3673 shader_info
->num_input_vgprs
= 0;
3674 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
3675 shader_info
->num_input_vgprs
+= 2;
3676 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
3677 shader_info
->num_input_vgprs
+= 2;
3678 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
3679 shader_info
->num_input_vgprs
+= 2;
3680 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
3681 shader_info
->num_input_vgprs
+= 3;
3682 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
3683 shader_info
->num_input_vgprs
+= 2;
3684 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
3685 shader_info
->num_input_vgprs
+= 2;
3686 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
3687 shader_info
->num_input_vgprs
+= 2;
3688 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
3689 shader_info
->num_input_vgprs
+= 1;
3690 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
3691 shader_info
->num_input_vgprs
+= 1;
3692 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
3693 shader_info
->num_input_vgprs
+= 1;
3694 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
3695 shader_info
->num_input_vgprs
+= 1;
3696 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
3697 shader_info
->num_input_vgprs
+= 1;
3698 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
3699 shader_info
->num_input_vgprs
+= 1;
3700 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
3701 shader_info
->num_input_vgprs
+= 1;
3702 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
3703 shader_info
->num_input_vgprs
+= 1;
3704 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
3705 shader_info
->num_input_vgprs
+= 1;
3707 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
3709 /* +3 for scratch wave offset and VCC */
3710 config
->num_sgprs
= MAX2(config
->num_sgprs
,
3711 shader_info
->num_input_sgprs
+ 3);
3713 /* Enable 64-bit and 16-bit denormals, because there is no performance
3716 * If denormals are enabled, all floating-point output modifiers are
3719 * Don't enable denormals for 32-bit floats, because:
3720 * - Floating-point output modifiers would be ignored by the hw.
3721 * - Some opcodes don't support denormals, such as v_mad_f32. We would
3722 * have to stop using those.
3723 * - SI & CI would be very slow.
3725 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
3729 ac_fill_shader_info(struct radv_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct radv_nir_compiler_options
*options
)
3731 switch (nir
->info
.stage
) {
3732 case MESA_SHADER_COMPUTE
:
3733 for (int i
= 0; i
< 3; ++i
)
3734 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
3736 case MESA_SHADER_FRAGMENT
:
3737 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
3739 case MESA_SHADER_GEOMETRY
:
3740 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
3741 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
3742 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
3743 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
3745 case MESA_SHADER_TESS_EVAL
:
3746 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
3747 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
3748 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
3749 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
3750 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
3752 case MESA_SHADER_TESS_CTRL
:
3753 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
3755 case MESA_SHADER_VERTEX
:
3756 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
3757 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
3758 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
3759 if (options
->key
.vs
.as_ls
)
3760 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
3768 radv_compile_nir_shader(struct ac_llvm_compiler
*ac_llvm
,
3769 struct ac_shader_binary
*binary
,
3770 struct ac_shader_config
*config
,
3771 struct radv_shader_variant_info
*shader_info
,
3772 struct nir_shader
*const *nir
,
3774 const struct radv_nir_compiler_options
*options
)
3777 LLVMModuleRef llvm_module
;
3779 llvm_module
= ac_translate_nir_to_llvm(ac_llvm
, nir
, nir_count
, shader_info
,
3782 ac_compile_llvm_module(ac_llvm
, llvm_module
, binary
, config
, shader_info
,
3783 nir
[0]->info
.stage
, options
);
3785 for (int i
= 0; i
< nir_count
; ++i
)
3786 ac_fill_shader_info(shader_info
, nir
[i
], options
);
3788 /* Determine the ES type (VS or TES) for the GS on GFX9. */
3789 if (options
->chip_class
== GFX9
) {
3790 if (nir_count
== 2 &&
3791 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3792 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
3798 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
3800 LLVMValueRef vtx_offset
=
3801 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
3802 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3803 LLVMValueRef stream_id
;
3805 /* Fetch the vertex stream ID. */
3806 if (ctx
->shader_info
->info
.so
.num_outputs
) {
3808 ac_unpack_param(&ctx
->ac
, ctx
->streamout_config
, 24, 2);
3810 stream_id
= ctx
->ac
.i32_0
;
3813 LLVMBasicBlockRef end_bb
;
3814 LLVMValueRef switch_inst
;
3816 end_bb
= LLVMAppendBasicBlockInContext(ctx
->ac
.context
,
3817 ctx
->main_function
, "end");
3818 switch_inst
= LLVMBuildSwitch(ctx
->ac
.builder
, stream_id
, end_bb
, 4);
3820 for (unsigned stream
= 0; stream
< 4; stream
++) {
3821 unsigned num_components
=
3822 ctx
->shader_info
->info
.gs
.num_stream_output_components
[stream
];
3823 LLVMBasicBlockRef bb
;
3826 if (!num_components
)
3829 if (stream
> 0 && !ctx
->shader_info
->info
.so
.num_outputs
)
3832 bb
= LLVMInsertBasicBlockInContext(ctx
->ac
.context
, end_bb
, "out");
3833 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
->ac
.i32
, stream
, 0), bb
);
3834 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, bb
);
3837 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3838 unsigned output_usage_mask
=
3839 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
3840 unsigned output_stream
=
3841 ctx
->shader_info
->info
.gs
.output_streams
[i
];
3842 int length
= util_last_bit(output_usage_mask
);
3844 if (!(ctx
->output_mask
& (1ull << i
)) ||
3845 output_stream
!= stream
)
3848 for (unsigned j
= 0; j
< length
; j
++) {
3849 LLVMValueRef value
, soffset
;
3851 if (!(output_usage_mask
& (1 << j
)))
3854 soffset
= LLVMConstInt(ctx
->ac
.i32
,
3856 ctx
->gs_max_out_vertices
* 16 * 4, false);
3860 value
= ac_build_buffer_load(&ctx
->ac
,
3863 vtx_offset
, soffset
,
3864 0, 1, 1, true, false);
3866 LLVMTypeRef type
= LLVMGetAllocatedType(ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3867 if (ac_get_type_size(type
) == 2) {
3868 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
3869 value
= LLVMBuildTrunc(ctx
->ac
.builder
, value
, ctx
->ac
.i16
, "");
3872 LLVMBuildStore(ctx
->ac
.builder
,
3873 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3877 if (ctx
->shader_info
->info
.so
.num_outputs
)
3878 radv_emit_streamout(ctx
, stream
);
3881 handle_vs_outputs_post(ctx
, false, false,
3882 &ctx
->shader_info
->vs
.outinfo
);
3885 LLVMBuildBr(ctx
->ac
.builder
, end_bb
);
3888 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, end_bb
);
3892 radv_compile_gs_copy_shader(struct ac_llvm_compiler
*ac_llvm
,
3893 struct nir_shader
*geom_shader
,
3894 struct ac_shader_binary
*binary
,
3895 struct ac_shader_config
*config
,
3896 struct radv_shader_variant_info
*shader_info
,
3897 const struct radv_nir_compiler_options
*options
)
3899 struct radv_shader_context ctx
= {0};
3900 ctx
.options
= options
;
3901 ctx
.shader_info
= shader_info
;
3903 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3904 ctx
.context
= ctx
.ac
.context
;
3905 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3907 ctx
.is_gs_copy_shader
= true;
3909 enum ac_float_mode float_mode
=
3910 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3911 AC_FLOAT_MODE_DEFAULT
;
3913 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3914 ctx
.stage
= MESA_SHADER_VERTEX
;
3916 radv_nir_shader_info_pass(geom_shader
, options
, &shader_info
->info
);
3918 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
3920 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
3921 ac_setup_rings(&ctx
);
3923 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
3924 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
3925 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
3926 variable
, MESA_SHADER_VERTEX
);
3929 ac_gs_copy_shader_emit(&ctx
);
3931 LLVMBuildRetVoid(ctx
.ac
.builder
);
3933 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3935 ac_compile_llvm_module(ac_llvm
, ctx
.ac
.module
, binary
, config
, shader_info
,
3936 MESA_SHADER_VERTEX
, options
);