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 #if HAVE_LLVM >= 0x0700
37 #include <llvm-c/Transforms/Utils.h>
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "ac_llvm_build.h"
45 #include "ac_shader_abi.h"
46 #include "ac_shader_util.h"
47 #include "ac_exp_param.h"
49 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
51 struct radv_shader_context
{
52 struct ac_llvm_context ac
;
53 const struct radv_nir_compiler_options
*options
;
54 struct radv_shader_variant_info
*shader_info
;
55 struct ac_shader_abi abi
;
57 unsigned max_workgroup_size
;
58 LLVMContextRef context
;
59 LLVMValueRef main_function
;
61 LLVMValueRef descriptor_sets
[RADV_UD_MAX_SETS
];
62 LLVMValueRef ring_offsets
;
64 LLVMValueRef vertex_buffers
;
65 LLVMValueRef rel_auto_id
;
66 LLVMValueRef vs_prim_id
;
67 LLVMValueRef es2gs_offset
;
70 LLVMValueRef merged_wave_info
;
71 LLVMValueRef tess_factor_offset
;
72 LLVMValueRef tes_rel_patch_id
;
76 LLVMValueRef gs2vs_offset
;
77 LLVMValueRef gs_wave_id
;
78 LLVMValueRef gs_vtx_offset
[6];
80 LLVMValueRef esgs_ring
;
81 LLVMValueRef gsvs_ring
[4];
82 LLVMValueRef hs_ring_tess_offchip
;
83 LLVMValueRef hs_ring_tess_factor
;
85 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
86 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
88 gl_shader_stage stage
;
90 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
95 bool is_gs_copy_shader
;
96 LLVMValueRef gs_next_vertex
[4];
97 unsigned gs_max_out_vertices
;
99 unsigned tes_primitive_mode
;
101 uint32_t tcs_patch_outputs_read
;
102 uint64_t tcs_outputs_read
;
103 uint32_t tcs_vertices_per_patch
;
104 uint32_t tcs_num_inputs
;
105 uint32_t tcs_num_patches
;
106 uint32_t max_gsvs_emit_size
;
107 uint32_t gsvs_vertex_size
;
110 enum radeon_llvm_calling_convention
{
111 RADEON_LLVM_AMDGPU_VS
= 87,
112 RADEON_LLVM_AMDGPU_GS
= 88,
113 RADEON_LLVM_AMDGPU_PS
= 89,
114 RADEON_LLVM_AMDGPU_CS
= 90,
115 RADEON_LLVM_AMDGPU_HS
= 93,
118 static inline struct radv_shader_context
*
119 radv_shader_context_from_abi(struct ac_shader_abi
*abi
)
121 struct radv_shader_context
*ctx
= NULL
;
122 return container_of(abi
, ctx
, abi
);
125 struct ac_build_if_state
127 struct radv_shader_context
*ctx
;
128 LLVMValueRef condition
;
129 LLVMBasicBlockRef entry_block
;
130 LLVMBasicBlockRef true_block
;
131 LLVMBasicBlockRef false_block
;
132 LLVMBasicBlockRef merge_block
;
135 static LLVMBasicBlockRef
136 ac_build_insert_new_block(struct radv_shader_context
*ctx
, const char *name
)
138 LLVMBasicBlockRef current_block
;
139 LLVMBasicBlockRef next_block
;
140 LLVMBasicBlockRef new_block
;
142 /* get current basic block */
143 current_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
145 /* chqeck if there's another block after this one */
146 next_block
= LLVMGetNextBasicBlock(current_block
);
148 /* insert the new block before the next block */
149 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
152 /* append new block after current block */
153 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
154 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
160 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
161 struct radv_shader_context
*ctx
,
162 LLVMValueRef condition
)
164 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
166 memset(ifthen
, 0, sizeof *ifthen
);
168 ifthen
->condition
= condition
;
169 ifthen
->entry_block
= block
;
171 /* create endif/merge basic block for the phi functions */
172 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
174 /* create/insert true_block before merge_block */
176 LLVMInsertBasicBlockInContext(ctx
->context
,
180 /* successive code goes into the true block */
181 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ifthen
->true_block
);
188 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
190 LLVMBuilderRef builder
= ifthen
->ctx
->ac
.builder
;
192 /* Insert branch to the merge block from current block */
193 LLVMBuildBr(builder
, ifthen
->merge_block
);
196 * Now patch in the various branch instructions.
199 /* Insert the conditional branch instruction at the end of entry_block */
200 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
201 if (ifthen
->false_block
) {
202 /* we have an else clause */
203 LLVMBuildCondBr(builder
, ifthen
->condition
,
204 ifthen
->true_block
, ifthen
->false_block
);
208 LLVMBuildCondBr(builder
, ifthen
->condition
,
209 ifthen
->true_block
, ifthen
->merge_block
);
212 /* Resume building code at end of the ifthen->merge_block */
213 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
217 static LLVMValueRef
get_rel_patch_id(struct radv_shader_context
*ctx
)
219 switch (ctx
->stage
) {
220 case MESA_SHADER_TESS_CTRL
:
221 return ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
222 case MESA_SHADER_TESS_EVAL
:
223 return ctx
->tes_rel_patch_id
;
226 unreachable("Illegal stage");
231 get_tcs_num_patches(struct radv_shader_context
*ctx
)
233 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
234 unsigned num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
235 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
236 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
237 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
238 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
239 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
240 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
241 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
242 unsigned num_patches
;
243 unsigned hardware_lds_size
;
245 /* Ensure that we only need one wave per SIMD so we don't need to check
246 * resource usage. Also ensures that the number of tcs in and out
247 * vertices per threadgroup are at most 256.
249 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
250 /* Make sure that the data fits in LDS. This assumes the shaders only
251 * use LDS for the inputs and outputs.
253 hardware_lds_size
= ctx
->options
->chip_class
>= CIK
? 65536 : 32768;
254 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
255 /* Make sure the output data fits in the offchip buffer */
256 num_patches
= MIN2(num_patches
, (ctx
->options
->tess_offchip_block_dw_size
* 4) / output_patch_size
);
257 /* Not necessary for correctness, but improves performance. The
258 * specific value is taken from the proprietary driver.
260 num_patches
= MIN2(num_patches
, 40);
262 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
263 if (ctx
->options
->chip_class
== SI
) {
264 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
265 num_patches
= MIN2(num_patches
, one_wave
);
271 calculate_tess_lds_size(struct radv_shader_context
*ctx
)
273 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
274 unsigned num_tcs_output_cp
;
275 unsigned num_tcs_outputs
, num_tcs_patch_outputs
;
276 unsigned input_vertex_size
, output_vertex_size
;
277 unsigned input_patch_size
, output_patch_size
;
278 unsigned pervertex_output_patch_size
;
279 unsigned output_patch0_offset
;
280 unsigned num_patches
;
283 num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
284 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
285 num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
287 input_vertex_size
= ctx
->tcs_num_inputs
* 16;
288 output_vertex_size
= num_tcs_outputs
* 16;
290 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
292 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
293 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
295 num_patches
= ctx
->tcs_num_patches
;
296 output_patch0_offset
= input_patch_size
* num_patches
;
298 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
302 /* Tessellation shaders pass outputs to the next shader using LDS.
304 * LS outputs = TCS inputs
305 * TCS outputs = TES inputs
308 * - TCS inputs for patch 0
309 * - TCS inputs for patch 1
310 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
312 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
313 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
314 * - TCS outputs for patch 1
315 * - Per-patch TCS outputs for patch 1
316 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
317 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
320 * All three shaders VS(LS), TCS, TES share the same LDS space.
323 get_tcs_in_patch_stride(struct radv_shader_context
*ctx
)
325 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
326 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
327 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
329 input_patch_size
/= 4;
330 return LLVMConstInt(ctx
->ac
.i32
, input_patch_size
, false);
334 get_tcs_out_patch_stride(struct radv_shader_context
*ctx
)
336 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
337 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
338 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
339 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
340 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
341 output_patch_size
/= 4;
342 return LLVMConstInt(ctx
->ac
.i32
, output_patch_size
, false);
346 get_tcs_out_vertex_stride(struct radv_shader_context
*ctx
)
348 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
349 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
350 output_vertex_size
/= 4;
351 return LLVMConstInt(ctx
->ac
.i32
, output_vertex_size
, false);
355 get_tcs_out_patch0_offset(struct radv_shader_context
*ctx
)
357 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
358 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
359 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
360 uint32_t output_patch0_offset
= input_patch_size
;
361 unsigned num_patches
= ctx
->tcs_num_patches
;
363 output_patch0_offset
*= num_patches
;
364 output_patch0_offset
/= 4;
365 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
369 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context
*ctx
)
371 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
372 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
373 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
374 uint32_t output_patch0_offset
= input_patch_size
;
376 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
377 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
378 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
379 unsigned num_patches
= ctx
->tcs_num_patches
;
381 output_patch0_offset
*= num_patches
;
382 output_patch0_offset
+= pervertex_output_patch_size
;
383 output_patch0_offset
/= 4;
384 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
388 get_tcs_in_current_patch_offset(struct radv_shader_context
*ctx
)
390 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
391 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
393 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
397 get_tcs_out_current_patch_offset(struct radv_shader_context
*ctx
)
399 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
400 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
401 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
403 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
408 get_tcs_out_current_patch_data_offset(struct radv_shader_context
*ctx
)
410 LLVMValueRef patch0_patch_data_offset
=
411 get_tcs_out_patch0_patch_data_offset(ctx
);
412 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
413 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
415 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
416 patch0_patch_data_offset
);
421 LLVMTypeRef types
[MAX_ARGS
];
422 LLVMValueRef
*assign
[MAX_ARGS
];
423 unsigned array_params_mask
;
426 uint8_t num_sgprs_used
;
427 uint8_t num_vgprs_used
;
430 enum ac_arg_regfile
{
436 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
437 LLVMValueRef
*param_ptr
)
439 assert(info
->count
< MAX_ARGS
);
441 info
->assign
[info
->count
] = param_ptr
;
442 info
->types
[info
->count
] = type
;
445 if (regfile
== ARG_SGPR
) {
446 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
449 assert(regfile
== ARG_VGPR
);
450 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
455 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
457 info
->array_params_mask
|= (1 << info
->count
);
458 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
461 static void assign_arguments(LLVMValueRef main_function
,
462 struct arg_info
*info
)
465 for (i
= 0; i
< info
->count
; i
++) {
467 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
472 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
473 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
474 unsigned num_return_elems
,
475 struct arg_info
*args
,
476 unsigned max_workgroup_size
,
477 const struct radv_nir_compiler_options
*options
)
479 LLVMTypeRef main_function_type
, ret_type
;
480 LLVMBasicBlockRef main_function_body
;
482 if (num_return_elems
)
483 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
484 num_return_elems
, true);
486 ret_type
= LLVMVoidTypeInContext(ctx
);
488 /* Setup the function */
490 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
491 LLVMValueRef main_function
=
492 LLVMAddFunction(module
, "main", main_function_type
);
494 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
495 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
497 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
498 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
499 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
501 if (args
->array_params_mask
& (1 << i
)) {
502 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
503 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
504 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
508 if (options
->address32_hi
) {
509 ac_llvm_add_target_dep_function_attr(main_function
,
510 "amdgpu-32bit-address-high-bits",
511 options
->address32_hi
);
514 if (max_workgroup_size
) {
515 ac_llvm_add_target_dep_function_attr(main_function
,
516 "amdgpu-max-work-group-size",
519 if (options
->unsafe_math
) {
520 /* These were copied from some LLVM test. */
521 LLVMAddTargetDependentFunctionAttr(main_function
,
522 "less-precise-fpmad",
524 LLVMAddTargetDependentFunctionAttr(main_function
,
527 LLVMAddTargetDependentFunctionAttr(main_function
,
530 LLVMAddTargetDependentFunctionAttr(main_function
,
533 LLVMAddTargetDependentFunctionAttr(main_function
,
534 "no-signed-zeros-fp-math",
537 return main_function
;
542 set_loc(struct radv_userdata_info
*ud_info
, uint8_t *sgpr_idx
,
543 uint8_t num_sgprs
, bool indirect
)
545 ud_info
->sgpr_idx
= *sgpr_idx
;
546 ud_info
->num_sgprs
= num_sgprs
;
547 ud_info
->indirect
= indirect
;
548 *sgpr_idx
+= num_sgprs
;
552 set_loc_shader(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
555 struct radv_userdata_info
*ud_info
=
556 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
559 set_loc(ud_info
, sgpr_idx
, num_sgprs
, false);
563 set_loc_shader_ptr(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
565 bool use_32bit_pointers
= HAVE_32BIT_POINTERS
&&
566 idx
!= AC_UD_SCRATCH_RING_OFFSETS
;
568 set_loc_shader(ctx
, idx
, sgpr_idx
, use_32bit_pointers
? 1 : 2);
572 set_loc_desc(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
575 struct radv_userdata_locations
*locs
=
576 &ctx
->shader_info
->user_sgprs_locs
;
577 struct radv_userdata_info
*ud_info
= &locs
->descriptor_sets
[idx
];
580 set_loc(ud_info
, sgpr_idx
, HAVE_32BIT_POINTERS
? 1 : 2, indirect
);
583 locs
->descriptor_sets_enabled
|= 1 << idx
;
586 struct user_sgpr_info
{
587 bool need_ring_offsets
;
588 bool indirect_all_descriptor_sets
;
591 static bool needs_view_index_sgpr(struct radv_shader_context
*ctx
,
592 gl_shader_stage stage
)
595 case MESA_SHADER_VERTEX
:
596 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
597 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
600 case MESA_SHADER_TESS_EVAL
:
601 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
604 case MESA_SHADER_GEOMETRY
:
605 case MESA_SHADER_TESS_CTRL
:
606 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
616 count_vs_user_sgprs(struct radv_shader_context
*ctx
)
620 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
)
621 count
+= HAVE_32BIT_POINTERS
? 1 : 2;
622 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
627 static void allocate_user_sgprs(struct radv_shader_context
*ctx
,
628 gl_shader_stage stage
,
629 bool has_previous_stage
,
630 gl_shader_stage previous_stage
,
631 bool needs_view_index
,
632 struct user_sgpr_info
*user_sgpr_info
)
634 uint8_t user_sgpr_count
= 0;
636 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
638 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
639 if (stage
== MESA_SHADER_GEOMETRY
||
640 stage
== MESA_SHADER_VERTEX
||
641 stage
== MESA_SHADER_TESS_CTRL
||
642 stage
== MESA_SHADER_TESS_EVAL
||
643 ctx
->is_gs_copy_shader
)
644 user_sgpr_info
->need_ring_offsets
= true;
646 if (stage
== MESA_SHADER_FRAGMENT
&&
647 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
648 user_sgpr_info
->need_ring_offsets
= true;
650 /* 2 user sgprs will nearly always be allocated for scratch/rings */
651 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
652 user_sgpr_count
+= 2;
656 case MESA_SHADER_COMPUTE
:
657 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
658 user_sgpr_count
+= 3;
660 case MESA_SHADER_FRAGMENT
:
661 user_sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
663 case MESA_SHADER_VERTEX
:
664 if (!ctx
->is_gs_copy_shader
)
665 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
667 case MESA_SHADER_TESS_CTRL
:
668 if (has_previous_stage
) {
669 if (previous_stage
== MESA_SHADER_VERTEX
)
670 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
673 case MESA_SHADER_TESS_EVAL
:
675 case MESA_SHADER_GEOMETRY
:
676 if (has_previous_stage
) {
677 if (previous_stage
== MESA_SHADER_VERTEX
) {
678 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
686 if (needs_view_index
)
689 if (ctx
->shader_info
->info
.loads_push_constants
)
690 user_sgpr_count
+= HAVE_32BIT_POINTERS
? 1 : 2;
692 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
&& stage
!= MESA_SHADER_COMPUTE
? 32 : 16;
693 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_count
;
694 uint32_t num_desc_set
=
695 util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
);
697 if (remaining_sgprs
/ (HAVE_32BIT_POINTERS
? 1 : 2) < num_desc_set
) {
698 user_sgpr_info
->indirect_all_descriptor_sets
= true;
703 declare_global_input_sgprs(struct radv_shader_context
*ctx
,
704 gl_shader_stage stage
,
705 bool has_previous_stage
,
706 gl_shader_stage previous_stage
,
707 const struct user_sgpr_info
*user_sgpr_info
,
708 struct arg_info
*args
,
709 LLVMValueRef
*desc_sets
)
711 LLVMTypeRef type
= ac_array_in_const32_addr_space(ctx
->ac
.i8
);
712 unsigned num_sets
= ctx
->options
->layout
?
713 ctx
->options
->layout
->num_sets
: 0;
714 unsigned stage_mask
= 1 << stage
;
716 if (has_previous_stage
)
717 stage_mask
|= 1 << previous_stage
;
719 /* 1 for each descriptor set */
720 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
721 for (unsigned i
= 0; i
< num_sets
; ++i
) {
722 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
723 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
724 add_array_arg(args
, type
,
725 &ctx
->descriptor_sets
[i
]);
729 add_array_arg(args
, ac_array_in_const32_addr_space(type
), desc_sets
);
732 if (ctx
->shader_info
->info
.loads_push_constants
) {
733 /* 1 for push constants and dynamic descriptors */
734 add_array_arg(args
, type
, &ctx
->abi
.push_constants
);
739 declare_vs_specific_input_sgprs(struct radv_shader_context
*ctx
,
740 gl_shader_stage stage
,
741 bool has_previous_stage
,
742 gl_shader_stage previous_stage
,
743 struct arg_info
*args
)
745 if (!ctx
->is_gs_copy_shader
&&
746 (stage
== MESA_SHADER_VERTEX
||
747 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
748 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
749 add_arg(args
, ARG_SGPR
,
750 ac_array_in_const32_addr_space(ctx
->ac
.v4i32
),
751 &ctx
->vertex_buffers
);
753 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
754 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
755 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
756 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
762 declare_vs_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
764 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
765 if (!ctx
->is_gs_copy_shader
) {
766 if (ctx
->options
->key
.vs
.as_ls
) {
767 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
768 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
770 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
771 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
773 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
778 declare_tes_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
780 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
781 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
782 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
783 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
787 set_global_input_locs(struct radv_shader_context
*ctx
, gl_shader_stage stage
,
788 bool has_previous_stage
, gl_shader_stage previous_stage
,
789 const struct user_sgpr_info
*user_sgpr_info
,
790 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
792 unsigned num_sets
= ctx
->options
->layout
?
793 ctx
->options
->layout
->num_sets
: 0;
794 unsigned stage_mask
= 1 << stage
;
796 if (has_previous_stage
)
797 stage_mask
|= 1 << previous_stage
;
799 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
800 for (unsigned i
= 0; i
< num_sets
; ++i
) {
801 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
802 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
803 set_loc_desc(ctx
, i
, user_sgpr_idx
, false);
805 ctx
->descriptor_sets
[i
] = NULL
;
808 set_loc_shader_ptr(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
811 for (unsigned i
= 0; i
< num_sets
; ++i
) {
812 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
813 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
814 ctx
->descriptor_sets
[i
] =
815 ac_build_load_to_sgpr(&ctx
->ac
,
817 LLVMConstInt(ctx
->ac
.i32
, i
, false));
820 ctx
->descriptor_sets
[i
] = NULL
;
822 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
825 if (ctx
->shader_info
->info
.loads_push_constants
) {
826 set_loc_shader_ptr(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
);
831 set_vs_specific_input_locs(struct radv_shader_context
*ctx
,
832 gl_shader_stage stage
, bool has_previous_stage
,
833 gl_shader_stage previous_stage
,
834 uint8_t *user_sgpr_idx
)
836 if (!ctx
->is_gs_copy_shader
&&
837 (stage
== MESA_SHADER_VERTEX
||
838 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
839 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
840 set_loc_shader_ptr(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
845 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
848 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
849 user_sgpr_idx
, vs_num
);
853 static void set_llvm_calling_convention(LLVMValueRef func
,
854 gl_shader_stage stage
)
856 enum radeon_llvm_calling_convention calling_conv
;
859 case MESA_SHADER_VERTEX
:
860 case MESA_SHADER_TESS_EVAL
:
861 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
863 case MESA_SHADER_GEOMETRY
:
864 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
866 case MESA_SHADER_TESS_CTRL
:
867 calling_conv
= RADEON_LLVM_AMDGPU_HS
;
869 case MESA_SHADER_FRAGMENT
:
870 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
872 case MESA_SHADER_COMPUTE
:
873 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
876 unreachable("Unhandle shader type");
879 LLVMSetFunctionCallConv(func
, calling_conv
);
882 static void create_function(struct radv_shader_context
*ctx
,
883 gl_shader_stage stage
,
884 bool has_previous_stage
,
885 gl_shader_stage previous_stage
)
887 uint8_t user_sgpr_idx
;
888 struct user_sgpr_info user_sgpr_info
;
889 struct arg_info args
= {};
890 LLVMValueRef desc_sets
;
891 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
892 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
893 previous_stage
, needs_view_index
, &user_sgpr_info
);
895 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
896 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
901 case MESA_SHADER_COMPUTE
:
902 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
903 previous_stage
, &user_sgpr_info
,
906 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
907 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
908 &ctx
->abi
.num_work_groups
);
911 for (int i
= 0; i
< 3; i
++) {
912 ctx
->abi
.workgroup_ids
[i
] = NULL
;
913 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
914 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
915 &ctx
->abi
.workgroup_ids
[i
]);
919 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
920 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
921 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
922 &ctx
->abi
.local_invocation_ids
);
924 case MESA_SHADER_VERTEX
:
925 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
926 previous_stage
, &user_sgpr_info
,
928 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
929 previous_stage
, &args
);
931 if (needs_view_index
)
932 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
933 &ctx
->abi
.view_index
);
934 if (ctx
->options
->key
.vs
.as_es
)
935 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
938 declare_vs_input_vgprs(ctx
, &args
);
940 case MESA_SHADER_TESS_CTRL
:
941 if (has_previous_stage
) {
942 // First 6 system regs
943 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
944 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
945 &ctx
->merged_wave_info
);
946 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
947 &ctx
->tess_factor_offset
);
949 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
950 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
951 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
953 declare_global_input_sgprs(ctx
, stage
,
956 &user_sgpr_info
, &args
,
958 declare_vs_specific_input_sgprs(ctx
, stage
,
960 previous_stage
, &args
);
962 if (needs_view_index
)
963 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
964 &ctx
->abi
.view_index
);
966 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
967 &ctx
->abi
.tcs_patch_id
);
968 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
969 &ctx
->abi
.tcs_rel_ids
);
971 declare_vs_input_vgprs(ctx
, &args
);
973 declare_global_input_sgprs(ctx
, stage
,
976 &user_sgpr_info
, &args
,
979 if (needs_view_index
)
980 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
981 &ctx
->abi
.view_index
);
983 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
984 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
985 &ctx
->tess_factor_offset
);
986 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
987 &ctx
->abi
.tcs_patch_id
);
988 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
989 &ctx
->abi
.tcs_rel_ids
);
992 case MESA_SHADER_TESS_EVAL
:
993 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
994 previous_stage
, &user_sgpr_info
,
997 if (needs_view_index
)
998 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
999 &ctx
->abi
.view_index
);
1001 if (ctx
->options
->key
.tes
.as_es
) {
1002 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1003 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1004 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1005 &ctx
->es2gs_offset
);
1007 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1008 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1010 declare_tes_input_vgprs(ctx
, &args
);
1012 case MESA_SHADER_GEOMETRY
:
1013 if (has_previous_stage
) {
1014 // First 6 system regs
1015 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1016 &ctx
->gs2vs_offset
);
1017 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1018 &ctx
->merged_wave_info
);
1019 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1021 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
1022 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1023 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1025 declare_global_input_sgprs(ctx
, stage
,
1028 &user_sgpr_info
, &args
,
1031 if (previous_stage
!= MESA_SHADER_TESS_EVAL
) {
1032 declare_vs_specific_input_sgprs(ctx
, stage
,
1038 if (needs_view_index
)
1039 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1040 &ctx
->abi
.view_index
);
1042 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1043 &ctx
->gs_vtx_offset
[0]);
1044 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1045 &ctx
->gs_vtx_offset
[2]);
1046 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1047 &ctx
->abi
.gs_prim_id
);
1048 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1049 &ctx
->abi
.gs_invocation_id
);
1050 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1051 &ctx
->gs_vtx_offset
[4]);
1053 if (previous_stage
== MESA_SHADER_VERTEX
) {
1054 declare_vs_input_vgprs(ctx
, &args
);
1056 declare_tes_input_vgprs(ctx
, &args
);
1059 declare_global_input_sgprs(ctx
, stage
,
1062 &user_sgpr_info
, &args
,
1065 if (needs_view_index
)
1066 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1067 &ctx
->abi
.view_index
);
1069 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
1070 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
1071 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1072 &ctx
->gs_vtx_offset
[0]);
1073 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1074 &ctx
->gs_vtx_offset
[1]);
1075 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1076 &ctx
->abi
.gs_prim_id
);
1077 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1078 &ctx
->gs_vtx_offset
[2]);
1079 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1080 &ctx
->gs_vtx_offset
[3]);
1081 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1082 &ctx
->gs_vtx_offset
[4]);
1083 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1084 &ctx
->gs_vtx_offset
[5]);
1085 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1086 &ctx
->abi
.gs_invocation_id
);
1089 case MESA_SHADER_FRAGMENT
:
1090 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
1091 previous_stage
, &user_sgpr_info
,
1094 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1095 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1096 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1097 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1098 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1099 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1100 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1101 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1102 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1103 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1104 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1105 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1106 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1107 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1108 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1109 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1110 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1113 unreachable("Shader stage not implemented");
1116 ctx
->main_function
= create_llvm_function(
1117 ctx
->context
, ctx
->ac
.module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1118 ctx
->max_workgroup_size
, ctx
->options
);
1119 set_llvm_calling_convention(ctx
->main_function
, stage
);
1122 ctx
->shader_info
->num_input_vgprs
= 0;
1123 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1125 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1127 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1128 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1130 assign_arguments(ctx
->main_function
, &args
);
1134 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1135 set_loc_shader_ptr(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1137 if (ctx
->options
->supports_spill
) {
1138 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1139 LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_CONST
),
1140 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1141 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1142 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1146 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1147 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1148 if (has_previous_stage
)
1151 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1152 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1155 case MESA_SHADER_COMPUTE
:
1156 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1157 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1161 case MESA_SHADER_VERTEX
:
1162 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1163 previous_stage
, &user_sgpr_idx
);
1164 if (ctx
->abi
.view_index
)
1165 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1167 case MESA_SHADER_TESS_CTRL
:
1168 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1169 previous_stage
, &user_sgpr_idx
);
1170 if (ctx
->abi
.view_index
)
1171 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1173 case MESA_SHADER_TESS_EVAL
:
1174 if (ctx
->abi
.view_index
)
1175 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1177 case MESA_SHADER_GEOMETRY
:
1178 if (has_previous_stage
) {
1179 if (previous_stage
== MESA_SHADER_VERTEX
)
1180 set_vs_specific_input_locs(ctx
, stage
,
1185 if (ctx
->abi
.view_index
)
1186 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1188 case MESA_SHADER_FRAGMENT
:
1191 unreachable("Shader stage not implemented");
1194 if (stage
== MESA_SHADER_TESS_CTRL
||
1195 (stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_ls
) ||
1196 /* GFX9 has the ESGS ring buffer in LDS. */
1197 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
1198 ac_declare_lds_as_pointer(&ctx
->ac
);
1201 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1206 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
1207 unsigned desc_set
, unsigned binding
)
1209 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1210 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
1211 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
1212 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
1213 unsigned base_offset
= layout
->binding
[binding
].offset
;
1214 LLVMValueRef offset
, stride
;
1216 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
1217 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
1218 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
1219 layout
->binding
[binding
].dynamic_offset_offset
;
1220 desc_ptr
= ctx
->abi
.push_constants
;
1221 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
1222 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1224 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
1226 offset
= ac_build_imad(&ctx
->ac
, index
, stride
,
1227 LLVMConstInt(ctx
->ac
.i32
, base_offset
, false));
1229 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
1230 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
1231 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1237 /* The offchip buffer layout for TCS->TES is
1239 * - attribute 0 of patch 0 vertex 0
1240 * - attribute 0 of patch 0 vertex 1
1241 * - attribute 0 of patch 0 vertex 2
1243 * - attribute 0 of patch 1 vertex 0
1244 * - attribute 0 of patch 1 vertex 1
1246 * - attribute 1 of patch 0 vertex 0
1247 * - attribute 1 of patch 0 vertex 1
1249 * - per patch attribute 0 of patch 0
1250 * - per patch attribute 0 of patch 1
1253 * Note that every attribute has 4 components.
1255 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
1257 uint32_t num_patches
= ctx
->tcs_num_patches
;
1258 uint32_t num_tcs_outputs
;
1259 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
1260 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
1262 num_tcs_outputs
= ctx
->options
->key
.tes
.tcs_num_outputs
;
1264 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
1265 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
1267 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
1270 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
1271 LLVMValueRef vertex_index
)
1273 LLVMValueRef param_stride
;
1275 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
* ctx
->tcs_num_patches
, false);
1277 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
1278 return param_stride
;
1281 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
1282 LLVMValueRef vertex_index
,
1283 LLVMValueRef param_index
)
1285 LLVMValueRef base_addr
;
1286 LLVMValueRef param_stride
, constant16
;
1287 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
1288 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
, false);
1289 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1290 param_stride
= calc_param_stride(ctx
, vertex_index
);
1292 base_addr
= ac_build_imad(&ctx
->ac
, rel_patch_id
,
1293 vertices_per_patch
, vertex_index
);
1295 base_addr
= rel_patch_id
;
1298 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1299 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
1300 param_stride
, ""), "");
1302 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
1304 if (!vertex_index
) {
1305 LLVMValueRef patch_data_offset
= get_non_vertex_index_offset(ctx
);
1307 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1308 patch_data_offset
, "");
1313 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct radv_shader_context
*ctx
,
1315 unsigned const_index
,
1317 LLVMValueRef vertex_index
,
1318 LLVMValueRef indir_index
)
1320 LLVMValueRef param_index
;
1323 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
1326 if (const_index
&& !is_compact
)
1327 param
+= const_index
;
1328 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
1330 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
1334 get_dw_address(struct radv_shader_context
*ctx
,
1335 LLVMValueRef dw_addr
,
1337 unsigned const_index
,
1338 bool compact_const_index
,
1339 LLVMValueRef vertex_index
,
1340 LLVMValueRef stride
,
1341 LLVMValueRef indir_index
)
1346 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1347 LLVMBuildMul(ctx
->ac
.builder
,
1353 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1354 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
1355 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
1356 else if (const_index
&& !compact_const_index
)
1357 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1358 LLVMConstInt(ctx
->ac
.i32
, const_index
* 4, false), "");
1360 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1361 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
1363 if (const_index
&& compact_const_index
)
1364 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1365 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
1370 load_tcs_varyings(struct ac_shader_abi
*abi
,
1372 LLVMValueRef vertex_index
,
1373 LLVMValueRef indir_index
,
1374 unsigned const_index
,
1376 unsigned driver_location
,
1378 unsigned num_components
,
1383 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1384 LLVMValueRef dw_addr
, stride
;
1385 LLVMValueRef value
[4], result
;
1386 unsigned param
= shader_io_get_unique_index(location
);
1389 uint32_t input_vertex_size
= (ctx
->tcs_num_inputs
* 16) / 4;
1390 stride
= LLVMConstInt(ctx
->ac
.i32
, input_vertex_size
, false);
1391 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1394 stride
= get_tcs_out_vertex_stride(ctx
);
1395 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1397 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1402 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1405 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1406 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1407 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1410 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1415 store_tcs_output(struct ac_shader_abi
*abi
,
1416 const nir_variable
*var
,
1417 LLVMValueRef vertex_index
,
1418 LLVMValueRef param_index
,
1419 unsigned const_index
,
1423 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1424 const unsigned location
= var
->data
.location
;
1425 const unsigned component
= var
->data
.location_frac
;
1426 const bool is_patch
= var
->data
.patch
;
1427 const bool is_compact
= var
->data
.compact
;
1428 LLVMValueRef dw_addr
;
1429 LLVMValueRef stride
= NULL
;
1430 LLVMValueRef buf_addr
= NULL
;
1432 bool store_lds
= true;
1435 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
1438 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
1442 param
= shader_io_get_unique_index(location
);
1443 if (location
== VARYING_SLOT_CLIP_DIST0
&&
1444 is_compact
&& const_index
> 3) {
1450 stride
= get_tcs_out_vertex_stride(ctx
);
1451 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1453 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1456 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1458 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
1459 vertex_index
, param_index
);
1461 bool is_tess_factor
= false;
1462 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
1463 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
1464 is_tess_factor
= true;
1466 unsigned base
= is_compact
? const_index
: 0;
1467 for (unsigned chan
= 0; chan
< 8; chan
++) {
1468 if (!(writemask
& (1 << chan
)))
1470 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1471 value
= ac_to_integer(&ctx
->ac
, value
);
1472 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
1474 if (store_lds
|| is_tess_factor
) {
1475 LLVMValueRef dw_addr_chan
=
1476 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1477 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
1478 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
1481 if (!is_tess_factor
&& writemask
!= 0xF)
1482 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
1483 buf_addr
, ctx
->oc_lds
,
1484 4 * (base
+ chan
), 1, 0, true, false);
1487 if (writemask
== 0xF) {
1488 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
1489 buf_addr
, ctx
->oc_lds
,
1490 (base
* 4), 1, 0, true, false);
1495 load_tes_input(struct ac_shader_abi
*abi
,
1497 LLVMValueRef vertex_index
,
1498 LLVMValueRef param_index
,
1499 unsigned const_index
,
1501 unsigned driver_location
,
1503 unsigned num_components
,
1508 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1509 LLVMValueRef buf_addr
;
1510 LLVMValueRef result
;
1511 unsigned param
= shader_io_get_unique_index(location
);
1513 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
1518 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
1519 is_compact
, vertex_index
, param_index
);
1521 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
1522 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
1524 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
1525 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
1526 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
1531 load_gs_input(struct ac_shader_abi
*abi
,
1533 unsigned driver_location
,
1535 unsigned num_components
,
1536 unsigned vertex_index
,
1537 unsigned const_index
,
1540 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1541 LLVMValueRef vtx_offset
;
1542 unsigned param
, vtx_offset_param
;
1543 LLVMValueRef value
[4], result
;
1545 vtx_offset_param
= vertex_index
;
1546 assert(vtx_offset_param
< 6);
1547 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
1548 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1550 param
= shader_io_get_unique_index(location
);
1552 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1553 if (ctx
->ac
.chip_class
>= GFX9
) {
1554 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1555 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1556 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
1557 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1559 LLVMValueRef soffset
=
1560 LLVMConstInt(ctx
->ac
.i32
,
1561 (param
* 4 + i
+ const_index
) * 256,
1564 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
1567 vtx_offset
, soffset
,
1568 0, 1, 0, true, false);
1571 if (ac_get_type_size(type
) == 2) {
1572 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i32
, "");
1573 value
[i
] = LLVMBuildTrunc(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i16
, "");
1575 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], type
, "");
1577 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1578 result
= ac_to_integer(&ctx
->ac
, result
);
1583 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
1585 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1586 ac_build_kill_if_false(&ctx
->ac
, visible
);
1589 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
1590 enum glsl_interp_mode interp
, unsigned location
)
1592 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1595 case INTERP_MODE_FLAT
:
1598 case INTERP_MODE_SMOOTH
:
1599 case INTERP_MODE_NONE
:
1600 if (location
== INTERP_CENTER
)
1601 return ctx
->persp_center
;
1602 else if (location
== INTERP_CENTROID
)
1603 return ctx
->persp_centroid
;
1604 else if (location
== INTERP_SAMPLE
)
1605 return ctx
->persp_sample
;
1607 case INTERP_MODE_NOPERSPECTIVE
:
1608 if (location
== INTERP_CENTER
)
1609 return ctx
->linear_center
;
1610 else if (location
== INTERP_CENTROID
)
1611 return ctx
->linear_centroid
;
1612 else if (location
== INTERP_SAMPLE
)
1613 return ctx
->linear_sample
;
1620 radv_get_sample_pos_offset(uint32_t num_samples
)
1622 uint32_t sample_pos_offset
= 0;
1624 switch (num_samples
) {
1626 sample_pos_offset
= 1;
1629 sample_pos_offset
= 3;
1632 sample_pos_offset
= 7;
1635 sample_pos_offset
= 15;
1640 return sample_pos_offset
;
1643 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
1644 LLVMValueRef sample_id
)
1646 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1648 LLVMValueRef result
;
1649 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
1651 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
1652 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
1654 uint32_t sample_pos_offset
=
1655 radv_get_sample_pos_offset(ctx
->options
->key
.fs
.num_samples
);
1658 LLVMBuildAdd(ctx
->ac
.builder
, sample_id
,
1659 LLVMConstInt(ctx
->ac
.i32
, sample_pos_offset
, false), "");
1660 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
1666 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1668 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1669 uint8_t log2_ps_iter_samples
;
1671 if (ctx
->shader_info
->info
.ps
.force_persample
) {
1672 log2_ps_iter_samples
=
1673 util_logbase2(ctx
->options
->key
.fs
.num_samples
);
1675 log2_ps_iter_samples
= ctx
->options
->key
.fs
.log2_ps_iter_samples
;
1678 /* The bit pattern matches that used by fixed function fragment
1680 static const uint16_t ps_iter_masks
[] = {
1681 0xffff, /* not used */
1687 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
1689 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
1691 LLVMValueRef result
, sample_id
;
1692 sample_id
= ac_unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
1693 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
1694 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
1700 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
1702 LLVMValueRef gs_next_vertex
;
1703 LLVMValueRef can_emit
;
1704 unsigned offset
= 0;
1705 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1707 /* Write vertex attribute values to GSVS ring */
1708 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
1709 ctx
->gs_next_vertex
[stream
],
1712 /* If this thread has already emitted the declared maximum number of
1713 * vertices, kill it: excessive vertex emissions are not supposed to
1714 * have any effect, and GS threads have no externally observable
1715 * effects other than emitting vertices.
1717 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
1718 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
1719 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1721 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1722 unsigned output_usage_mask
=
1723 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
1724 uint8_t output_stream
=
1725 ctx
->shader_info
->info
.gs
.output_streams
[i
];
1726 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1727 int length
= util_last_bit(output_usage_mask
);
1729 if (!(ctx
->output_mask
& (1ull << i
)) ||
1730 output_stream
!= stream
)
1733 for (unsigned j
= 0; j
< length
; j
++) {
1734 if (!(output_usage_mask
& (1 << j
)))
1737 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1739 LLVMValueRef voffset
=
1740 LLVMConstInt(ctx
->ac
.i32
, offset
*
1741 ctx
->gs_max_out_vertices
, false);
1745 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1746 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1748 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1749 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1751 ac_build_buffer_store_dword(&ctx
->ac
,
1752 ctx
->gsvs_ring
[stream
],
1754 voffset
, ctx
->gs2vs_offset
, 0,
1759 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1761 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
1763 ac_build_sendmsg(&ctx
->ac
,
1764 AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
1769 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1771 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1772 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1776 load_tess_coord(struct ac_shader_abi
*abi
)
1778 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1780 LLVMValueRef coord
[4] = {
1787 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
1788 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1789 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1791 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1795 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1797 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1798 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
1802 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1804 return abi
->base_vertex
;
1807 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1808 LLVMValueRef buffer_ptr
, bool write
)
1810 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1811 LLVMValueRef result
;
1813 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1815 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1816 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1821 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1823 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1824 LLVMValueRef result
;
1826 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1828 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1829 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1834 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1835 unsigned descriptor_set
,
1836 unsigned base_index
,
1837 unsigned constant_index
,
1839 enum ac_descriptor_type desc_type
,
1840 bool image
, bool write
,
1843 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1844 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1845 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
1846 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1847 unsigned offset
= binding
->offset
;
1848 unsigned stride
= binding
->size
;
1850 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1853 assert(base_index
< layout
->binding_count
);
1855 switch (desc_type
) {
1857 type
= ctx
->ac
.v8i32
;
1861 type
= ctx
->ac
.v8i32
;
1865 case AC_DESC_SAMPLER
:
1866 type
= ctx
->ac
.v4i32
;
1867 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
1872 case AC_DESC_BUFFER
:
1873 type
= ctx
->ac
.v4i32
;
1877 unreachable("invalid desc_type\n");
1880 offset
+= constant_index
* stride
;
1882 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1883 (!index
|| binding
->immutable_samplers_equal
)) {
1884 if (binding
->immutable_samplers_equal
)
1887 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1889 LLVMValueRef constants
[] = {
1890 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1891 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1892 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1893 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1895 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1898 assert(stride
% type_size
== 0);
1901 index
= ctx
->ac
.i32_0
;
1903 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1905 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
1906 list
= LLVMBuildPointerCast(builder
, list
,
1907 ac_array_in_const32_addr_space(type
), "");
1909 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
1912 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1913 * so we may need to fix it up. */
1915 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1916 unsigned adjustment
,
1919 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1922 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1924 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1925 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1927 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1929 /* For the integer-like cases, do a natural sign extension.
1931 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1932 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1935 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1936 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1937 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1938 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1940 /* Convert back to the right type. */
1941 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1943 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1944 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1945 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1946 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1947 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1948 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1955 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1956 struct nir_variable
*variable
)
1958 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
1959 LLVMValueRef t_offset
;
1960 LLVMValueRef t_list
;
1962 LLVMValueRef buffer_index
;
1963 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1964 uint8_t input_usage_mask
=
1965 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
1966 unsigned num_channels
= util_last_bit(input_usage_mask
);
1968 variable
->data
.driver_location
= variable
->data
.location
* 4;
1970 enum glsl_base_type type
= glsl_get_base_type(variable
->type
);
1971 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1972 LLVMValueRef output
[4];
1973 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1975 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1976 uint32_t divisor
= ctx
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
1979 buffer_index
= ctx
->abi
.instance_id
;
1982 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
1983 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
1986 if (ctx
->options
->key
.vs
.as_ls
) {
1987 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1988 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1990 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1991 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1994 buffer_index
= ctx
->ac
.i32_0
;
1997 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.start_instance
, buffer_index
, "");
1999 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
2000 ctx
->abi
.base_vertex
, "");
2001 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_index
, false);
2003 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
2005 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
2008 num_channels
, false, true);
2010 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
2012 for (unsigned chan
= 0; chan
< 4; chan
++) {
2013 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2014 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
2015 if (type
== GLSL_TYPE_FLOAT16
) {
2016 output
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f32
, "");
2017 output
[chan
] = LLVMBuildFPTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f16
, "");
2021 unsigned alpha_adjust
= (ctx
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
2022 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
2024 for (unsigned chan
= 0; chan
< 4; chan
++) {
2025 output
[chan
] = ac_to_integer(&ctx
->ac
, output
[chan
]);
2026 if (type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
)
2027 output
[chan
] = LLVMBuildTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.i16
, "");
2029 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] = output
[chan
];
2034 static void interp_fs_input(struct radv_shader_context
*ctx
,
2036 LLVMValueRef interp_param
,
2037 LLVMValueRef prim_mask
,
2038 LLVMValueRef result
[4])
2040 LLVMValueRef attr_number
;
2043 bool interp
= !LLVMIsUndef(interp_param
);
2045 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
2047 /* fs.constant returns the param from the middle vertex, so it's not
2048 * really useful for flat shading. It's meant to be used for custom
2049 * interpolation (but the intrinsic can't fetch from the other two
2052 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
2053 * to do the right thing. The only reason we use fs.constant is that
2054 * fs.interp cannot be used on integers, because they can be equal
2058 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
2061 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2063 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2067 for (chan
= 0; chan
< 4; chan
++) {
2068 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2071 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
2076 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
2077 LLVMConstInt(ctx
->ac
.i32
, 2, false),
2081 result
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, result
[chan
], ctx
->ac
.i32
, "");
2082 result
[chan
] = LLVMBuildTruncOrBitCast(ctx
->ac
.builder
, result
[chan
], LLVMTypeOf(interp_param
), "");
2088 handle_fs_input_decl(struct radv_shader_context
*ctx
,
2089 struct nir_variable
*variable
)
2091 int idx
= variable
->data
.location
;
2092 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2093 LLVMValueRef interp
= NULL
;
2096 variable
->data
.driver_location
= idx
* 4;
2097 mask
= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
2099 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
2100 unsigned interp_type
;
2101 if (variable
->data
.sample
)
2102 interp_type
= INTERP_SAMPLE
;
2103 else if (variable
->data
.centroid
)
2104 interp_type
= INTERP_CENTROID
;
2106 interp_type
= INTERP_CENTER
;
2108 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
2110 bool is_16bit
= glsl_type_is_16bit(variable
->type
);
2111 LLVMTypeRef type
= is_16bit
? ctx
->ac
.i16
: ctx
->ac
.i32
;
2113 interp
= LLVMGetUndef(type
);
2115 for (unsigned i
= 0; i
< attrib_count
; ++i
)
2116 ctx
->inputs
[ac_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
2118 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2119 /* Do not account for the number of components inside the array
2120 * of clip/cull distances because this might wrongly set other
2121 * bits like primitive ID or layer.
2123 mask
= 1ull << VARYING_SLOT_CLIP_DIST0
;
2126 ctx
->input_mask
|= mask
;
2130 handle_vs_inputs(struct radv_shader_context
*ctx
,
2131 struct nir_shader
*nir
) {
2132 nir_foreach_variable(variable
, &nir
->inputs
)
2133 handle_vs_input_decl(ctx
, variable
);
2137 prepare_interp_optimize(struct radv_shader_context
*ctx
,
2138 struct nir_shader
*nir
)
2140 bool uses_center
= false;
2141 bool uses_centroid
= false;
2142 nir_foreach_variable(variable
, &nir
->inputs
) {
2143 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
2144 variable
->data
.sample
)
2147 if (variable
->data
.centroid
)
2148 uses_centroid
= true;
2153 if (uses_center
&& uses_centroid
) {
2154 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
2155 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
2156 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
2161 handle_fs_inputs(struct radv_shader_context
*ctx
,
2162 struct nir_shader
*nir
)
2164 prepare_interp_optimize(ctx
, nir
);
2166 nir_foreach_variable(variable
, &nir
->inputs
)
2167 handle_fs_input_decl(ctx
, variable
);
2171 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
2172 ctx
->shader_info
->info
.needs_multiview_view_index
) {
2173 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
2174 ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)] = LLVMGetUndef(ctx
->ac
.i32
);
2177 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
2178 LLVMValueRef interp_param
;
2179 LLVMValueRef
*inputs
= ctx
->inputs
+ac_llvm_reg_index_soa(i
, 0);
2181 if (!(ctx
->input_mask
& (1ull << i
)))
2184 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
2185 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
2186 interp_param
= *inputs
;
2187 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
2190 if (LLVMIsUndef(interp_param
))
2191 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
2193 } else if (i
== VARYING_SLOT_CLIP_DIST0
) {
2194 int length
= ctx
->shader_info
->info
.ps
.num_input_clips_culls
;
2196 for (unsigned j
= 0; j
< length
; j
+= 4) {
2197 inputs
= ctx
->inputs
+ ac_llvm_reg_index_soa(i
, j
);
2199 interp_param
= *inputs
;
2200 interp_fs_input(ctx
, index
, interp_param
,
2201 ctx
->abi
.prim_mask
, inputs
);
2204 } else if (i
== VARYING_SLOT_POS
) {
2205 for(int i
= 0; i
< 3; ++i
)
2206 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
2208 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2209 ctx
->abi
.frag_pos
[3]);
2212 ctx
->shader_info
->fs
.num_interp
= index
;
2213 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
2215 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
2216 ctx
->abi
.view_index
= ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2220 scan_shader_output_decl(struct radv_shader_context
*ctx
,
2221 struct nir_variable
*variable
,
2222 struct nir_shader
*shader
,
2223 gl_shader_stage stage
)
2225 int idx
= variable
->data
.location
+ variable
->data
.index
;
2226 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2227 uint64_t mask_attribs
;
2229 variable
->data
.driver_location
= idx
* 4;
2231 /* tess ctrl has it's own load/store paths for outputs */
2232 if (stage
== MESA_SHADER_TESS_CTRL
)
2235 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
2236 if (stage
== MESA_SHADER_VERTEX
||
2237 stage
== MESA_SHADER_TESS_EVAL
||
2238 stage
== MESA_SHADER_GEOMETRY
) {
2239 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2240 if (stage
== MESA_SHADER_VERTEX
) {
2241 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2242 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2243 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2245 if (stage
== MESA_SHADER_TESS_EVAL
) {
2246 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2247 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2248 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2251 mask_attribs
= 1ull << idx
;
2255 ctx
->output_mask
|= mask_attribs
;
2259 /* Initialize arguments for the shader export intrinsic */
2261 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
2262 LLVMValueRef
*values
,
2263 unsigned enabled_channels
,
2265 struct ac_export_args
*args
)
2267 /* Specify the channels that are enabled. */
2268 args
->enabled_channels
= enabled_channels
;
2270 /* Specify whether the EXEC mask represents the valid mask */
2271 args
->valid_mask
= 0;
2273 /* Specify whether this is the last export */
2276 /* Specify the target we are exporting */
2277 args
->target
= target
;
2279 args
->compr
= false;
2280 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
2281 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
2282 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
2283 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
2288 bool is_16bit
= ac_get_type_size(LLVMTypeOf(values
[0])) == 2;
2289 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2290 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
2291 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
2292 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
2293 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
2296 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2297 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2298 unsigned bits
, bool hi
) = NULL
;
2300 switch(col_format
) {
2301 case V_028714_SPI_SHADER_ZERO
:
2302 args
->enabled_channels
= 0; /* writemask */
2303 args
->target
= V_008DFC_SQ_EXP_NULL
;
2306 case V_028714_SPI_SHADER_32_R
:
2307 args
->enabled_channels
= 1;
2308 args
->out
[0] = values
[0];
2311 case V_028714_SPI_SHADER_32_GR
:
2312 args
->enabled_channels
= 0x3;
2313 args
->out
[0] = values
[0];
2314 args
->out
[1] = values
[1];
2317 case V_028714_SPI_SHADER_32_AR
:
2318 args
->enabled_channels
= 0x9;
2319 args
->out
[0] = values
[0];
2320 args
->out
[3] = values
[3];
2323 case V_028714_SPI_SHADER_FP16_ABGR
:
2324 args
->enabled_channels
= 0x5;
2325 packf
= ac_build_cvt_pkrtz_f16
;
2327 for (unsigned chan
= 0; chan
< 4; chan
++)
2328 values
[chan
] = LLVMBuildFPExt(ctx
->ac
.builder
,
2334 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2335 args
->enabled_channels
= 0x5;
2336 packf
= ac_build_cvt_pknorm_u16
;
2339 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2340 args
->enabled_channels
= 0x5;
2341 packf
= ac_build_cvt_pknorm_i16
;
2344 case V_028714_SPI_SHADER_UINT16_ABGR
:
2345 args
->enabled_channels
= 0x5;
2346 packi
= ac_build_cvt_pk_u16
;
2348 for (unsigned chan
= 0; chan
< 4; chan
++)
2349 values
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
,
2355 case V_028714_SPI_SHADER_SINT16_ABGR
:
2356 args
->enabled_channels
= 0x5;
2357 packi
= ac_build_cvt_pk_i16
;
2359 for (unsigned chan
= 0; chan
< 4; chan
++)
2360 values
[chan
] = LLVMBuildSExt(ctx
->ac
.builder
,
2367 case V_028714_SPI_SHADER_32_ABGR
:
2368 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2372 /* Pack f16 or norm_i16/u16. */
2374 for (chan
= 0; chan
< 2; chan
++) {
2375 LLVMValueRef pack_args
[2] = {
2377 values
[2 * chan
+ 1]
2379 LLVMValueRef packed
;
2381 packed
= packf(&ctx
->ac
, pack_args
);
2382 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2384 args
->compr
= 1; /* COMPR flag */
2389 for (chan
= 0; chan
< 2; chan
++) {
2390 LLVMValueRef pack_args
[2] = {
2391 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2392 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2394 LLVMValueRef packed
;
2396 packed
= packi(&ctx
->ac
, pack_args
,
2397 is_int8
? 8 : is_int10
? 10 : 16,
2399 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2401 args
->compr
= 1; /* COMPR flag */
2407 for (unsigned chan
= 0; chan
< 4; chan
++) {
2408 values
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i16
, "");
2409 args
->out
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i32
, "");
2412 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2414 for (unsigned i
= 0; i
< 4; ++i
) {
2415 if (!(args
->enabled_channels
& (1 << i
)))
2418 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
2423 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
2424 LLVMValueRef
*values
, unsigned enabled_channels
)
2426 struct ac_export_args args
;
2428 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
2429 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2430 ac_build_export(&ctx
->ac
, &args
);
2434 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
2436 LLVMValueRef output
=
2437 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
2439 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
2443 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2444 bool export_prim_id
, bool export_layer_id
,
2445 struct radv_vs_output_info
*outinfo
)
2447 uint32_t param_count
= 0;
2449 unsigned pos_idx
, num_pos_exports
= 0;
2450 struct ac_export_args args
, pos_args
[4] = {};
2451 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2454 if (ctx
->options
->key
.has_multiview_view_index
) {
2455 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2457 for(unsigned i
= 0; i
< 4; ++i
)
2458 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2459 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2462 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
2463 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2466 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2467 sizeof(outinfo
->vs_output_param_offset
));
2469 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
2470 unsigned output_usage_mask
, length
;
2471 LLVMValueRef slots
[8];
2474 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2475 !ctx
->is_gs_copy_shader
) {
2477 ctx
->shader_info
->info
.vs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2478 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2480 ctx
->shader_info
->info
.tes
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2482 assert(ctx
->is_gs_copy_shader
);
2484 ctx
->shader_info
->info
.gs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2487 length
= util_last_bit(output_usage_mask
);
2489 i
= VARYING_SLOT_CLIP_DIST0
;
2490 for (j
= 0; j
< length
; j
++)
2491 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2493 for (i
= length
; i
< 8; i
++)
2494 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
2497 target
= V_008DFC_SQ_EXP_POS
+ 3;
2498 si_llvm_init_export_args(ctx
, &slots
[4], 0xf, target
, &args
);
2499 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2500 &args
, sizeof(args
));
2503 target
= V_008DFC_SQ_EXP_POS
+ 2;
2504 si_llvm_init_export_args(ctx
, &slots
[0], 0xf, target
, &args
);
2505 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2506 &args
, sizeof(args
));
2508 /* Export the clip/cull distances values to the next stage. */
2509 radv_export_param(ctx
, param_count
, &slots
[0], 0xf);
2510 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST0
] = param_count
++;
2512 radv_export_param(ctx
, param_count
, &slots
[4], 0xf);
2513 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST1
] = param_count
++;
2517 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
2518 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
2519 for (unsigned j
= 0; j
< 4; j
++)
2520 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
2522 si_llvm_init_export_args(ctx
, pos_values
, 0xf, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2524 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
2525 outinfo
->writes_pointsize
= true;
2526 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
2529 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
2530 outinfo
->writes_layer
= true;
2531 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
2534 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
2535 outinfo
->writes_viewport_index
= true;
2536 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
2539 if (outinfo
->writes_pointsize
||
2540 outinfo
->writes_layer
||
2541 outinfo
->writes_viewport_index
) {
2542 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
2543 (outinfo
->writes_layer
== true ? 4 : 0));
2544 pos_args
[1].valid_mask
= 0;
2545 pos_args
[1].done
= 0;
2546 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2547 pos_args
[1].compr
= 0;
2548 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2549 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2550 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2551 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2553 if (outinfo
->writes_pointsize
== true)
2554 pos_args
[1].out
[0] = psize_value
;
2555 if (outinfo
->writes_layer
== true)
2556 pos_args
[1].out
[2] = layer_value
;
2557 if (outinfo
->writes_viewport_index
== true) {
2558 if (ctx
->options
->chip_class
>= GFX9
) {
2559 /* GFX9 has the layer in out.z[10:0] and the viewport
2560 * index in out.z[19:16].
2562 LLVMValueRef v
= viewport_index_value
;
2563 v
= ac_to_integer(&ctx
->ac
, v
);
2564 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2565 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2567 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2568 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2570 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2571 pos_args
[1].enabled_channels
|= 1 << 2;
2573 pos_args
[1].out
[3] = viewport_index_value
;
2574 pos_args
[1].enabled_channels
|= 1 << 3;
2578 for (i
= 0; i
< 4; i
++) {
2579 if (pos_args
[i
].out
[0])
2584 for (i
= 0; i
< 4; i
++) {
2585 if (!pos_args
[i
].out
[0])
2588 /* Specify the target we are exporting */
2589 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2590 if (pos_idx
== num_pos_exports
)
2591 pos_args
[i
].done
= 1;
2592 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2595 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2596 LLVMValueRef values
[4];
2597 if (!(ctx
->output_mask
& (1ull << i
)))
2600 if (i
!= VARYING_SLOT_LAYER
&&
2601 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
2602 i
< VARYING_SLOT_VAR0
)
2605 for (unsigned j
= 0; j
< 4; j
++)
2606 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2608 unsigned output_usage_mask
;
2610 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2611 !ctx
->is_gs_copy_shader
) {
2613 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2614 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2616 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2618 assert(ctx
->is_gs_copy_shader
);
2620 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
2623 radv_export_param(ctx
, param_count
, values
, output_usage_mask
);
2625 outinfo
->vs_output_param_offset
[i
] = param_count
++;
2628 if (export_prim_id
) {
2629 LLVMValueRef values
[4];
2631 values
[0] = ctx
->vs_prim_id
;
2632 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
2633 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2634 for (unsigned j
= 1; j
< 4; j
++)
2635 values
[j
] = ctx
->ac
.f32_0
;
2637 radv_export_param(ctx
, param_count
, values
, 0x1);
2639 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
2640 outinfo
->export_prim_id
= true;
2643 if (export_layer_id
&& layer_value
) {
2644 LLVMValueRef values
[4];
2646 values
[0] = layer_value
;
2647 for (unsigned j
= 1; j
< 4; j
++)
2648 values
[j
] = ctx
->ac
.f32_0
;
2650 radv_export_param(ctx
, param_count
, values
, 0x1);
2652 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
++;
2655 outinfo
->pos_exports
= num_pos_exports
;
2656 outinfo
->param_exports
= param_count
;
2660 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2661 struct radv_es_output_info
*outinfo
)
2664 uint64_t max_output_written
= 0;
2665 LLVMValueRef lds_base
= NULL
;
2667 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2668 unsigned output_usage_mask
;
2672 if (!(ctx
->output_mask
& (1ull << i
)))
2675 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2677 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2679 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2681 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2684 if (i
== VARYING_SLOT_CLIP_DIST0
)
2685 length
= util_last_bit(output_usage_mask
);
2687 param_index
= shader_io_get_unique_index(i
);
2689 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
2692 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
2694 if (ctx
->ac
.chip_class
>= GFX9
) {
2695 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2696 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2697 LLVMValueRef wave_idx
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 24, 4);
2698 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2699 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2700 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
2701 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2702 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2705 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2706 LLVMValueRef dw_addr
= NULL
;
2707 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2708 unsigned output_usage_mask
;
2712 if (!(ctx
->output_mask
& (1ull << i
)))
2715 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2717 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2719 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2721 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2724 if (i
== VARYING_SLOT_CLIP_DIST0
)
2725 length
= util_last_bit(output_usage_mask
);
2727 param_index
= shader_io_get_unique_index(i
);
2730 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2731 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2735 for (j
= 0; j
< length
; j
++) {
2736 if (!(output_usage_mask
& (1 << j
)))
2739 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2740 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
2741 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2743 if (ctx
->ac
.chip_class
>= GFX9
) {
2744 LLVMValueRef dw_addr_offset
=
2745 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2746 LLVMConstInt(ctx
->ac
.i32
,
2749 ac_lds_store(&ctx
->ac
, dw_addr_offset
, out_val
);
2751 ac_build_buffer_store_dword(&ctx
->ac
,
2754 NULL
, ctx
->es2gs_offset
,
2755 (4 * param_index
+ j
) * 4,
2763 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2765 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2766 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->shader_info
->info
.vs
.ls_outputs_written
);
2767 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2768 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2769 vertex_dw_stride
, "");
2771 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2772 unsigned output_usage_mask
=
2773 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2774 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2777 if (!(ctx
->output_mask
& (1ull << i
)))
2780 if (i
== VARYING_SLOT_CLIP_DIST0
)
2781 length
= util_last_bit(output_usage_mask
);
2783 int param
= shader_io_get_unique_index(i
);
2784 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2785 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2787 for (unsigned j
= 0; j
< length
; j
++) {
2788 LLVMValueRef value
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2789 value
= ac_to_integer(&ctx
->ac
, value
);
2790 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
2791 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
2792 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2798 write_tess_factors(struct radv_shader_context
*ctx
)
2800 unsigned stride
, outer_comps
, inner_comps
;
2801 struct ac_build_if_state if_ctx
, inner_if_ctx
;
2802 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2803 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
2804 unsigned tess_inner_index
= 0, tess_outer_index
;
2805 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
2806 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2808 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
2810 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
2830 ac_nir_build_if(&if_ctx
, ctx
,
2831 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2832 invocation_id
, ctx
->ac
.i32_0
, ""));
2834 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
2837 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2838 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2839 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
2842 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2843 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2844 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
2846 for (i
= 0; i
< 4; i
++) {
2847 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2848 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2852 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
2853 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
2854 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2856 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
2858 for (i
= 0; i
< outer_comps
; i
++) {
2860 ac_lds_load(&ctx
->ac
, lds_outer
);
2861 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2864 for (i
= 0; i
< inner_comps
; i
++) {
2865 inner
[i
] = out
[outer_comps
+i
] =
2866 ac_lds_load(&ctx
->ac
, lds_inner
);
2867 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
2872 /* Convert the outputs to vectors for stores. */
2873 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
2877 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
2880 buffer
= ctx
->hs_ring_tess_factor
;
2881 tf_base
= ctx
->tess_factor_offset
;
2882 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
2883 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
2884 unsigned tf_offset
= 0;
2886 if (ctx
->options
->chip_class
<= VI
) {
2887 ac_nir_build_if(&inner_if_ctx
, ctx
,
2888 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2889 rel_patch_id
, ctx
->ac
.i32_0
, ""));
2891 /* Store the dynamic HS control word. */
2892 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
2893 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
2894 1, ctx
->ac
.i32_0
, tf_base
,
2895 0, 1, 0, true, false);
2898 ac_nir_build_endif(&inner_if_ctx
);
2901 /* Store the tessellation factors. */
2902 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
2903 MIN2(stride
, 4), byteoffset
, tf_base
,
2904 tf_offset
, 1, 0, true, false);
2906 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
2907 stride
- 4, byteoffset
, tf_base
,
2908 16 + tf_offset
, 1, 0, true, false);
2910 //store to offchip for TES to read - only if TES reads them
2911 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
2912 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
2913 LLVMValueRef tf_inner_offset
;
2914 unsigned param_outer
, param_inner
;
2916 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2917 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2918 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
2920 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
2921 util_next_power_of_two(outer_comps
));
2923 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
2924 outer_comps
, tf_outer_offset
,
2925 ctx
->oc_lds
, 0, 1, 0, true, false);
2927 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2928 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2929 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
2931 inner_vec
= inner_comps
== 1 ? inner
[0] :
2932 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
2933 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
2934 inner_comps
, tf_inner_offset
,
2935 ctx
->oc_lds
, 0, 1, 0, true, false);
2938 ac_nir_build_endif(&if_ctx
);
2942 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
2944 write_tess_factors(ctx
);
2948 si_export_mrt_color(struct radv_shader_context
*ctx
,
2949 LLVMValueRef
*color
, unsigned index
,
2950 struct ac_export_args
*args
)
2953 si_llvm_init_export_args(ctx
, color
, 0xf,
2954 V_008DFC_SQ_EXP_MRT
+ index
, args
);
2955 if (!args
->enabled_channels
)
2956 return false; /* unnecessary NULL export */
2962 radv_export_mrt_z(struct radv_shader_context
*ctx
,
2963 LLVMValueRef depth
, LLVMValueRef stencil
,
2964 LLVMValueRef samplemask
)
2966 struct ac_export_args args
;
2968 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
2970 ac_build_export(&ctx
->ac
, &args
);
2974 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
2977 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2978 struct ac_export_args color_args
[8];
2980 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2981 LLVMValueRef values
[4];
2983 if (!(ctx
->output_mask
& (1ull << i
)))
2986 if (i
< FRAG_RESULT_DATA0
)
2989 for (unsigned j
= 0; j
< 4; j
++)
2990 values
[j
] = ac_to_float(&ctx
->ac
,
2991 radv_load_output(ctx
, i
, j
));
2993 bool ret
= si_export_mrt_color(ctx
, values
,
2994 i
- FRAG_RESULT_DATA0
,
2995 &color_args
[index
]);
3000 /* Process depth, stencil, samplemask. */
3001 if (ctx
->shader_info
->info
.ps
.writes_z
) {
3002 depth
= ac_to_float(&ctx
->ac
,
3003 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
3005 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
3006 stencil
= ac_to_float(&ctx
->ac
,
3007 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
3009 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3010 samplemask
= ac_to_float(&ctx
->ac
,
3011 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
3014 /* Set the DONE bit on last non-null color export only if Z isn't
3018 !ctx
->shader_info
->info
.ps
.writes_z
&&
3019 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
3020 !ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3021 unsigned last
= index
- 1;
3023 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
3024 color_args
[last
].done
= 1; /* DONE bit */
3027 /* Export PS outputs. */
3028 for (unsigned i
= 0; i
< index
; i
++)
3029 ac_build_export(&ctx
->ac
, &color_args
[i
]);
3031 if (depth
|| stencil
|| samplemask
)
3032 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
3034 ac_build_export_null(&ctx
->ac
);
3038 emit_gs_epilogue(struct radv_shader_context
*ctx
)
3040 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
3044 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
3045 LLVMValueRef
*addrs
)
3047 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
3049 switch (ctx
->stage
) {
3050 case MESA_SHADER_VERTEX
:
3051 if (ctx
->options
->key
.vs
.as_ls
)
3052 handle_ls_outputs_post(ctx
);
3053 else if (ctx
->options
->key
.vs
.as_es
)
3054 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
3056 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
3057 ctx
->options
->key
.vs
.export_layer_id
,
3058 &ctx
->shader_info
->vs
.outinfo
);
3060 case MESA_SHADER_FRAGMENT
:
3061 handle_fs_outputs_post(ctx
);
3063 case MESA_SHADER_GEOMETRY
:
3064 emit_gs_epilogue(ctx
);
3066 case MESA_SHADER_TESS_CTRL
:
3067 handle_tcs_outputs_post(ctx
);
3069 case MESA_SHADER_TESS_EVAL
:
3070 if (ctx
->options
->key
.tes
.as_es
)
3071 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
3073 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
3074 ctx
->options
->key
.tes
.export_layer_id
,
3075 &ctx
->shader_info
->tes
.outinfo
);
3082 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
,
3083 LLVMPassManagerRef passmgr
,
3084 const struct radv_nir_compiler_options
*options
)
3086 LLVMRunPassManager(passmgr
, ctx
->ac
.module
);
3087 LLVMDisposeBuilder(ctx
->ac
.builder
);
3089 ac_llvm_context_dispose(&ctx
->ac
);
3093 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
3095 struct radv_vs_output_info
*outinfo
;
3097 switch (ctx
->stage
) {
3098 case MESA_SHADER_FRAGMENT
:
3099 case MESA_SHADER_COMPUTE
:
3100 case MESA_SHADER_TESS_CTRL
:
3101 case MESA_SHADER_GEOMETRY
:
3103 case MESA_SHADER_VERTEX
:
3104 if (ctx
->options
->key
.vs
.as_ls
||
3105 ctx
->options
->key
.vs
.as_es
)
3107 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
3109 case MESA_SHADER_TESS_EVAL
:
3110 if (ctx
->options
->key
.vs
.as_es
)
3112 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
3115 unreachable("Unhandled shader type");
3118 ac_optimize_vs_outputs(&ctx
->ac
,
3120 outinfo
->vs_output_param_offset
,
3122 &outinfo
->param_exports
);
3126 ac_setup_rings(struct radv_shader_context
*ctx
)
3128 if (ctx
->options
->chip_class
<= VI
&&
3129 (ctx
->stage
== MESA_SHADER_GEOMETRY
||
3130 ctx
->options
->key
.vs
.as_es
|| ctx
->options
->key
.tes
.as_es
)) {
3131 unsigned ring
= ctx
->stage
== MESA_SHADER_GEOMETRY
? RING_ESGS_GS
3133 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, ring
, false);
3135 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
,
3140 if (ctx
->is_gs_copy_shader
) {
3142 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3143 LLVMConstInt(ctx
->ac
.i32
,
3144 RING_GSVS_VS
, false));
3147 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3148 /* The conceptual layout of the GSVS ring is
3149 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3150 * but the real memory layout is swizzled across
3152 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3154 * Override the buffer descriptor accordingly.
3156 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->ac
.i64
, 2);
3157 uint64_t stream_offset
= 0;
3158 unsigned num_records
= 64;
3159 LLVMValueRef base_ring
;
3162 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3163 LLVMConstInt(ctx
->ac
.i32
,
3164 RING_GSVS_GS
, false));
3166 for (unsigned stream
= 0; stream
< 4; stream
++) {
3167 unsigned num_components
, stride
;
3168 LLVMValueRef ring
, tmp
;
3171 ctx
->shader_info
->info
.gs
.num_stream_output_components
[stream
];
3173 if (!num_components
)
3176 stride
= 4 * num_components
* ctx
->gs_max_out_vertices
;
3178 /* Limit on the stride field for <= CIK. */
3179 assert(stride
< (1 << 14));
3181 ring
= LLVMBuildBitCast(ctx
->ac
.builder
,
3182 base_ring
, v2i64
, "");
3183 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3184 ring
, ctx
->ac
.i32_0
, "");
3185 tmp
= LLVMBuildAdd(ctx
->ac
.builder
, tmp
,
3186 LLVMConstInt(ctx
->ac
.i64
,
3187 stream_offset
, 0), "");
3188 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
,
3189 ring
, tmp
, ctx
->ac
.i32_0
, "");
3191 stream_offset
+= stride
* 64;
3193 ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ring
,
3196 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ring
,
3198 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
,
3199 LLVMConstInt(ctx
->ac
.i32
,
3200 S_008F04_STRIDE(stride
), false), "");
3201 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
, tmp
,
3204 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
,
3205 LLVMConstInt(ctx
->ac
.i32
,
3206 num_records
, false),
3207 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3209 ctx
->gsvs_ring
[stream
] = ring
;
3213 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3214 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3215 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3216 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3221 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
3222 const struct nir_shader
*nir
)
3224 switch (nir
->info
.stage
) {
3225 case MESA_SHADER_TESS_CTRL
:
3226 return chip_class
>= CIK
? 128 : 64;
3227 case MESA_SHADER_GEOMETRY
:
3228 return chip_class
>= GFX9
? 128 : 64;
3229 case MESA_SHADER_COMPUTE
:
3235 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
3236 nir
->info
.cs
.local_size
[1] *
3237 nir
->info
.cs
.local_size
[2];
3238 return max_workgroup_size
;
3241 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3242 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3244 LLVMValueRef count
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 8, 8);
3245 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3247 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
3248 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
3249 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
3252 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
)
3254 for(int i
= 5; i
>= 0; --i
) {
3255 ctx
->gs_vtx_offset
[i
] = ac_unpack_param(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
3259 ctx
->gs_wave_id
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 16, 8);
3264 LLVMModuleRef
ac_translate_nir_to_llvm(struct ac_llvm_compiler
*ac_llvm
,
3265 struct nir_shader
*const *shaders
,
3267 struct radv_shader_variant_info
*shader_info
,
3268 const struct radv_nir_compiler_options
*options
)
3270 struct radv_shader_context ctx
= {0};
3272 ctx
.options
= options
;
3273 ctx
.shader_info
= shader_info
;
3275 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3276 ctx
.context
= ctx
.ac
.context
;
3277 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3279 enum ac_float_mode float_mode
=
3280 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3281 AC_FLOAT_MODE_DEFAULT
;
3283 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3285 memset(shader_info
, 0, sizeof(*shader_info
));
3287 for(int i
= 0; i
< shader_count
; ++i
)
3288 radv_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
3290 for (i
= 0; i
< RADV_UD_MAX_SETS
; i
++)
3291 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
3292 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
3293 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
3295 ctx
.max_workgroup_size
= 0;
3296 for (int i
= 0; i
< shader_count
; ++i
) {
3297 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
3298 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
3302 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
3303 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
3305 ctx
.abi
.inputs
= &ctx
.inputs
[0];
3306 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
3307 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
3308 ctx
.abi
.load_ubo
= radv_load_ubo
;
3309 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
3310 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
3311 ctx
.abi
.load_resource
= radv_load_resource
;
3312 ctx
.abi
.clamp_shadow_reference
= false;
3313 ctx
.abi
.gfx9_stride_size_workaround
= ctx
.ac
.chip_class
== GFX9
;
3315 if (shader_count
>= 2)
3316 ac_init_exec_full_mask(&ctx
.ac
);
3318 if (ctx
.ac
.chip_class
== GFX9
&&
3319 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
3320 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
3322 for(int i
= 0; i
< shader_count
; ++i
) {
3323 ctx
.stage
= shaders
[i
]->info
.stage
;
3324 ctx
.output_mask
= 0;
3326 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3327 for (int i
= 0; i
< 4; i
++) {
3328 ctx
.gs_next_vertex
[i
] =
3329 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
3331 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
3332 ctx
.abi
.load_inputs
= load_gs_input
;
3333 ctx
.abi
.emit_primitive
= visit_end_primitive
;
3334 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3335 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
3336 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
3337 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
3338 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3339 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
3340 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3341 if (shader_count
== 1)
3342 ctx
.tcs_num_inputs
= ctx
.options
->key
.tcs
.num_inputs
;
3344 ctx
.tcs_num_inputs
= util_last_bit64(shader_info
->info
.vs
.ls_outputs_written
);
3345 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
3346 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3347 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
3348 ctx
.abi
.load_tess_varyings
= load_tes_input
;
3349 ctx
.abi
.load_tess_coord
= load_tess_coord
;
3350 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3351 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3352 ctx
.tcs_num_patches
= ctx
.options
->key
.tes
.num_patches
;
3353 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
3354 if (shader_info
->info
.vs
.needs_instance_id
) {
3355 if (ctx
.options
->key
.vs
.as_ls
) {
3356 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3357 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3359 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3360 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3363 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
3364 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
3365 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
3366 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
3367 ctx
.abi
.load_sample_position
= load_sample_position
;
3368 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
3369 ctx
.abi
.emit_kill
= radv_emit_kill
;
3373 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
3375 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
3376 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
3378 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3379 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
3380 shaders
[i
]->info
.cull_distance_array_size
> 4;
3381 ctx
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
3382 ctx
.max_gsvs_emit_size
= ctx
.gsvs_vertex_size
*
3383 shaders
[i
]->info
.gs
.vertices_out
;
3386 ac_setup_rings(&ctx
);
3388 LLVMBasicBlockRef merge_block
;
3389 if (shader_count
>= 2) {
3390 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
3391 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3392 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3394 LLVMValueRef count
= ac_unpack_param(&ctx
.ac
, ctx
.merged_wave_info
, 8 * i
, 8);
3395 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
3396 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
3397 thread_id
, count
, "");
3398 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
3400 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
3403 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
3404 handle_fs_inputs(&ctx
, shaders
[i
]);
3405 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
3406 handle_vs_inputs(&ctx
, shaders
[i
]);
3407 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
3408 prepare_gs_input_vgprs(&ctx
);
3410 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
]);
3412 if (shader_count
>= 2) {
3413 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
3414 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
3417 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3418 shader_info
->gs
.gsvs_vertex_size
= ctx
.gsvs_vertex_size
;
3419 shader_info
->gs
.max_gsvs_emit_size
= ctx
.max_gsvs_emit_size
;
3420 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3421 shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
3422 shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
3426 LLVMBuildRetVoid(ctx
.ac
.builder
);
3428 if (options
->dump_preoptir
)
3429 ac_dump_module(ctx
.ac
.module
);
3431 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3433 if (shader_count
== 1)
3434 ac_nir_eliminate_const_vs_outputs(&ctx
);
3436 if (options
->dump_shader
) {
3437 ctx
.shader_info
->private_mem_vgprs
=
3438 ac_count_scratch_private_memory(ctx
.main_function
);
3441 return ctx
.ac
.module
;
3444 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
3446 unsigned *retval
= (unsigned *)context
;
3447 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
3448 char *description
= LLVMGetDiagInfoDescription(di
);
3450 if (severity
== LLVMDSError
) {
3452 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
3456 LLVMDisposeMessage(description
);
3459 static unsigned ac_llvm_compile(LLVMModuleRef M
,
3460 struct ac_shader_binary
*binary
,
3461 struct ac_llvm_compiler
*ac_llvm
)
3463 unsigned retval
= 0;
3464 LLVMContextRef llvm_ctx
;
3466 /* Setup Diagnostic Handler*/
3467 llvm_ctx
= LLVMGetModuleContext(M
);
3469 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
3473 if (!radv_compile_to_binary(ac_llvm
, M
, binary
))
3478 static void ac_compile_llvm_module(struct ac_llvm_compiler
*ac_llvm
,
3479 LLVMModuleRef llvm_module
,
3480 struct ac_shader_binary
*binary
,
3481 struct ac_shader_config
*config
,
3482 struct radv_shader_variant_info
*shader_info
,
3483 gl_shader_stage stage
,
3484 const struct radv_nir_compiler_options
*options
)
3486 if (options
->dump_shader
)
3487 ac_dump_module(llvm_module
);
3489 memset(binary
, 0, sizeof(*binary
));
3491 if (options
->record_llvm_ir
) {
3492 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
3493 binary
->llvm_ir_string
= strdup(llvm_ir
);
3494 LLVMDisposeMessage(llvm_ir
);
3497 int v
= ac_llvm_compile(llvm_module
, binary
, ac_llvm
);
3499 fprintf(stderr
, "compile failed\n");
3502 if (options
->dump_shader
)
3503 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
3505 ac_shader_binary_read_config(binary
, config
, 0, options
->supports_spill
);
3507 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
3508 LLVMDisposeModule(llvm_module
);
3509 LLVMContextDispose(ctx
);
3511 if (stage
== MESA_SHADER_FRAGMENT
) {
3512 shader_info
->num_input_vgprs
= 0;
3513 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
3514 shader_info
->num_input_vgprs
+= 2;
3515 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
3516 shader_info
->num_input_vgprs
+= 2;
3517 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
3518 shader_info
->num_input_vgprs
+= 2;
3519 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
3520 shader_info
->num_input_vgprs
+= 3;
3521 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
3522 shader_info
->num_input_vgprs
+= 2;
3523 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
3524 shader_info
->num_input_vgprs
+= 2;
3525 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
3526 shader_info
->num_input_vgprs
+= 2;
3527 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
3528 shader_info
->num_input_vgprs
+= 1;
3529 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
3530 shader_info
->num_input_vgprs
+= 1;
3531 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
3532 shader_info
->num_input_vgprs
+= 1;
3533 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
3534 shader_info
->num_input_vgprs
+= 1;
3535 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
3536 shader_info
->num_input_vgprs
+= 1;
3537 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
3538 shader_info
->num_input_vgprs
+= 1;
3539 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
3540 shader_info
->num_input_vgprs
+= 1;
3541 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
3542 shader_info
->num_input_vgprs
+= 1;
3543 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
3544 shader_info
->num_input_vgprs
+= 1;
3546 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
3548 /* +3 for scratch wave offset and VCC */
3549 config
->num_sgprs
= MAX2(config
->num_sgprs
,
3550 shader_info
->num_input_sgprs
+ 3);
3552 /* Enable 64-bit and 16-bit denormals, because there is no performance
3555 * If denormals are enabled, all floating-point output modifiers are
3558 * Don't enable denormals for 32-bit floats, because:
3559 * - Floating-point output modifiers would be ignored by the hw.
3560 * - Some opcodes don't support denormals, such as v_mad_f32. We would
3561 * have to stop using those.
3562 * - SI & CI would be very slow.
3564 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
3568 ac_fill_shader_info(struct radv_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct radv_nir_compiler_options
*options
)
3570 switch (nir
->info
.stage
) {
3571 case MESA_SHADER_COMPUTE
:
3572 for (int i
= 0; i
< 3; ++i
)
3573 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
3575 case MESA_SHADER_FRAGMENT
:
3576 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
3578 case MESA_SHADER_GEOMETRY
:
3579 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
3580 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
3581 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
3582 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
3584 case MESA_SHADER_TESS_EVAL
:
3585 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
3586 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
3587 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
3588 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
3589 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
3591 case MESA_SHADER_TESS_CTRL
:
3592 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
3594 case MESA_SHADER_VERTEX
:
3595 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
3596 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
3597 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
3598 if (options
->key
.vs
.as_ls
)
3599 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
3607 radv_compile_nir_shader(struct ac_llvm_compiler
*ac_llvm
,
3608 struct ac_shader_binary
*binary
,
3609 struct ac_shader_config
*config
,
3610 struct radv_shader_variant_info
*shader_info
,
3611 struct nir_shader
*const *nir
,
3613 const struct radv_nir_compiler_options
*options
)
3616 LLVMModuleRef llvm_module
;
3618 llvm_module
= ac_translate_nir_to_llvm(ac_llvm
, nir
, nir_count
, shader_info
,
3621 ac_compile_llvm_module(ac_llvm
, llvm_module
, binary
, config
, shader_info
,
3622 nir
[0]->info
.stage
, options
);
3624 for (int i
= 0; i
< nir_count
; ++i
)
3625 ac_fill_shader_info(shader_info
, nir
[i
], options
);
3627 /* Determine the ES type (VS or TES) for the GS on GFX9. */
3628 if (options
->chip_class
== GFX9
) {
3629 if (nir_count
== 2 &&
3630 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3631 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
3637 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
3639 LLVMValueRef vtx_offset
=
3640 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
3641 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3642 unsigned offset
= 0;
3644 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3645 unsigned output_usage_mask
=
3646 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
3647 int length
= util_last_bit(output_usage_mask
);
3649 if (!(ctx
->output_mask
& (1ull << i
)))
3652 for (unsigned j
= 0; j
< length
; j
++) {
3653 LLVMValueRef value
, soffset
;
3655 if (!(output_usage_mask
& (1 << j
)))
3658 soffset
= LLVMConstInt(ctx
->ac
.i32
,
3660 ctx
->gs_max_out_vertices
* 16 * 4, false);
3664 value
= ac_build_buffer_load(&ctx
->ac
,
3667 vtx_offset
, soffset
,
3668 0, 1, 1, true, false);
3670 LLVMTypeRef type
= LLVMGetAllocatedType(ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3671 if (ac_get_type_size(type
) == 2) {
3672 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
3673 value
= LLVMBuildTrunc(ctx
->ac
.builder
, value
, ctx
->ac
.i16
, "");
3676 LLVMBuildStore(ctx
->ac
.builder
,
3677 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3680 handle_vs_outputs_post(ctx
, false, false, &ctx
->shader_info
->vs
.outinfo
);
3684 radv_compile_gs_copy_shader(struct ac_llvm_compiler
*ac_llvm
,
3685 struct nir_shader
*geom_shader
,
3686 struct ac_shader_binary
*binary
,
3687 struct ac_shader_config
*config
,
3688 struct radv_shader_variant_info
*shader_info
,
3689 const struct radv_nir_compiler_options
*options
)
3691 struct radv_shader_context ctx
= {0};
3692 ctx
.options
= options
;
3693 ctx
.shader_info
= shader_info
;
3695 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3696 ctx
.context
= ctx
.ac
.context
;
3697 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3699 ctx
.is_gs_copy_shader
= true;
3701 enum ac_float_mode float_mode
=
3702 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3703 AC_FLOAT_MODE_DEFAULT
;
3705 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3706 ctx
.stage
= MESA_SHADER_VERTEX
;
3708 radv_nir_shader_info_pass(geom_shader
, options
, &shader_info
->info
);
3710 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
3712 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
3713 ac_setup_rings(&ctx
);
3715 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
3716 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
3717 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
3718 variable
, MESA_SHADER_VERTEX
);
3721 ac_gs_copy_shader_emit(&ctx
);
3723 LLVMBuildRetVoid(ctx
.ac
.builder
);
3725 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3727 ac_compile_llvm_module(ac_llvm
, ctx
.ac
.module
, binary
, config
, shader_info
,
3728 MESA_SHADER_VERTEX
, options
);