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"
32 #include <llvm-c/Core.h>
33 #include <llvm-c/TargetMachine.h>
34 #include <llvm-c/Transforms/Scalar.h>
35 #if HAVE_LLVM >= 0x0700
36 #include <llvm-c/Transforms/Utils.h>
41 #include "ac_binary.h"
42 #include "ac_llvm_util.h"
43 #include "ac_llvm_build.h"
44 #include "ac_shader_abi.h"
45 #include "ac_shader_util.h"
46 #include "ac_exp_param.h"
48 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
50 struct radv_shader_context
{
51 struct ac_llvm_context ac
;
52 const struct radv_nir_compiler_options
*options
;
53 struct radv_shader_variant_info
*shader_info
;
54 struct ac_shader_abi abi
;
56 unsigned max_workgroup_size
;
57 LLVMContextRef context
;
58 LLVMValueRef main_function
;
60 LLVMValueRef descriptor_sets
[RADV_UD_MAX_SETS
];
61 LLVMValueRef ring_offsets
;
63 LLVMValueRef vertex_buffers
;
64 LLVMValueRef rel_auto_id
;
65 LLVMValueRef vs_prim_id
;
66 LLVMValueRef es2gs_offset
;
69 LLVMValueRef merged_wave_info
;
70 LLVMValueRef tess_factor_offset
;
71 LLVMValueRef tes_rel_patch_id
;
75 LLVMValueRef gs2vs_offset
;
76 LLVMValueRef gs_wave_id
;
77 LLVMValueRef gs_vtx_offset
[6];
79 LLVMValueRef esgs_ring
;
80 LLVMValueRef gsvs_ring
;
81 LLVMValueRef hs_ring_tess_offchip
;
82 LLVMValueRef hs_ring_tess_factor
;
84 LLVMValueRef sample_pos_offset
;
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];
94 uint8_t num_output_clips
;
95 uint8_t num_output_culls
;
97 bool is_gs_copy_shader
;
98 LLVMValueRef gs_next_vertex
;
99 unsigned gs_max_out_vertices
;
101 unsigned tes_primitive_mode
;
103 uint32_t tcs_patch_outputs_read
;
104 uint64_t tcs_outputs_read
;
105 uint32_t tcs_vertices_per_patch
;
106 uint32_t tcs_num_inputs
;
107 uint32_t tcs_num_patches
;
108 uint32_t max_gsvs_emit_size
;
109 uint32_t gsvs_vertex_size
;
112 enum radeon_llvm_calling_convention
{
113 RADEON_LLVM_AMDGPU_VS
= 87,
114 RADEON_LLVM_AMDGPU_GS
= 88,
115 RADEON_LLVM_AMDGPU_PS
= 89,
116 RADEON_LLVM_AMDGPU_CS
= 90,
117 RADEON_LLVM_AMDGPU_HS
= 93,
120 static inline struct radv_shader_context
*
121 radv_shader_context_from_abi(struct ac_shader_abi
*abi
)
123 struct radv_shader_context
*ctx
= NULL
;
124 return container_of(abi
, ctx
, abi
);
127 struct ac_build_if_state
129 struct radv_shader_context
*ctx
;
130 LLVMValueRef condition
;
131 LLVMBasicBlockRef entry_block
;
132 LLVMBasicBlockRef true_block
;
133 LLVMBasicBlockRef false_block
;
134 LLVMBasicBlockRef merge_block
;
137 static LLVMBasicBlockRef
138 ac_build_insert_new_block(struct radv_shader_context
*ctx
, const char *name
)
140 LLVMBasicBlockRef current_block
;
141 LLVMBasicBlockRef next_block
;
142 LLVMBasicBlockRef new_block
;
144 /* get current basic block */
145 current_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
147 /* chqeck if there's another block after this one */
148 next_block
= LLVMGetNextBasicBlock(current_block
);
150 /* insert the new block before the next block */
151 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
154 /* append new block after current block */
155 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
156 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
162 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
163 struct radv_shader_context
*ctx
,
164 LLVMValueRef condition
)
166 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
168 memset(ifthen
, 0, sizeof *ifthen
);
170 ifthen
->condition
= condition
;
171 ifthen
->entry_block
= block
;
173 /* create endif/merge basic block for the phi functions */
174 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
176 /* create/insert true_block before merge_block */
178 LLVMInsertBasicBlockInContext(ctx
->context
,
182 /* successive code goes into the true block */
183 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ifthen
->true_block
);
190 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
192 LLVMBuilderRef builder
= ifthen
->ctx
->ac
.builder
;
194 /* Insert branch to the merge block from current block */
195 LLVMBuildBr(builder
, ifthen
->merge_block
);
198 * Now patch in the various branch instructions.
201 /* Insert the conditional branch instruction at the end of entry_block */
202 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
203 if (ifthen
->false_block
) {
204 /* we have an else clause */
205 LLVMBuildCondBr(builder
, ifthen
->condition
,
206 ifthen
->true_block
, ifthen
->false_block
);
210 LLVMBuildCondBr(builder
, ifthen
->condition
,
211 ifthen
->true_block
, ifthen
->merge_block
);
214 /* Resume building code at end of the ifthen->merge_block */
215 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
219 static LLVMValueRef
get_rel_patch_id(struct radv_shader_context
*ctx
)
221 switch (ctx
->stage
) {
222 case MESA_SHADER_TESS_CTRL
:
223 return ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
224 case MESA_SHADER_TESS_EVAL
:
225 return ctx
->tes_rel_patch_id
;
228 unreachable("Illegal stage");
233 get_tcs_num_patches(struct radv_shader_context
*ctx
)
235 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
236 unsigned num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
237 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
238 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
239 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
240 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
241 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
242 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
243 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
244 unsigned num_patches
;
245 unsigned hardware_lds_size
;
247 /* Ensure that we only need one wave per SIMD so we don't need to check
248 * resource usage. Also ensures that the number of tcs in and out
249 * vertices per threadgroup are at most 256.
251 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
252 /* Make sure that the data fits in LDS. This assumes the shaders only
253 * use LDS for the inputs and outputs.
255 hardware_lds_size
= ctx
->options
->chip_class
>= CIK
? 65536 : 32768;
256 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
257 /* Make sure the output data fits in the offchip buffer */
258 num_patches
= MIN2(num_patches
, (ctx
->options
->tess_offchip_block_dw_size
* 4) / output_patch_size
);
259 /* Not necessary for correctness, but improves performance. The
260 * specific value is taken from the proprietary driver.
262 num_patches
= MIN2(num_patches
, 40);
264 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
265 if (ctx
->options
->chip_class
== SI
) {
266 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
267 num_patches
= MIN2(num_patches
, one_wave
);
273 calculate_tess_lds_size(struct radv_shader_context
*ctx
)
275 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
276 unsigned num_tcs_output_cp
;
277 unsigned num_tcs_outputs
, num_tcs_patch_outputs
;
278 unsigned input_vertex_size
, output_vertex_size
;
279 unsigned input_patch_size
, output_patch_size
;
280 unsigned pervertex_output_patch_size
;
281 unsigned output_patch0_offset
;
282 unsigned num_patches
;
285 num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
286 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
287 num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
289 input_vertex_size
= ctx
->tcs_num_inputs
* 16;
290 output_vertex_size
= num_tcs_outputs
* 16;
292 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
294 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
295 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
297 num_patches
= ctx
->tcs_num_patches
;
298 output_patch0_offset
= input_patch_size
* num_patches
;
300 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
304 /* Tessellation shaders pass outputs to the next shader using LDS.
306 * LS outputs = TCS inputs
307 * TCS outputs = TES inputs
310 * - TCS inputs for patch 0
311 * - TCS inputs for patch 1
312 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
314 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
315 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
316 * - TCS outputs for patch 1
317 * - Per-patch TCS outputs for patch 1
318 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
319 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
322 * All three shaders VS(LS), TCS, TES share the same LDS space.
325 get_tcs_in_patch_stride(struct radv_shader_context
*ctx
)
327 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
328 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
329 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
331 input_patch_size
/= 4;
332 return LLVMConstInt(ctx
->ac
.i32
, input_patch_size
, false);
336 get_tcs_out_patch_stride(struct radv_shader_context
*ctx
)
338 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
339 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
340 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
341 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
342 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
343 output_patch_size
/= 4;
344 return LLVMConstInt(ctx
->ac
.i32
, output_patch_size
, false);
348 get_tcs_out_vertex_stride(struct radv_shader_context
*ctx
)
350 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
351 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
352 output_vertex_size
/= 4;
353 return LLVMConstInt(ctx
->ac
.i32
, output_vertex_size
, false);
357 get_tcs_out_patch0_offset(struct radv_shader_context
*ctx
)
359 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
360 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
361 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
362 uint32_t output_patch0_offset
= input_patch_size
;
363 unsigned num_patches
= ctx
->tcs_num_patches
;
365 output_patch0_offset
*= num_patches
;
366 output_patch0_offset
/= 4;
367 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
371 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context
*ctx
)
373 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
374 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
375 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
376 uint32_t output_patch0_offset
= input_patch_size
;
378 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
379 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
380 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
381 unsigned num_patches
= ctx
->tcs_num_patches
;
383 output_patch0_offset
*= num_patches
;
384 output_patch0_offset
+= pervertex_output_patch_size
;
385 output_patch0_offset
/= 4;
386 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
390 get_tcs_in_current_patch_offset(struct radv_shader_context
*ctx
)
392 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
393 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
395 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
399 get_tcs_out_current_patch_offset(struct radv_shader_context
*ctx
)
401 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
402 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
403 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
405 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_offset
,
406 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
412 get_tcs_out_current_patch_data_offset(struct radv_shader_context
*ctx
)
414 LLVMValueRef patch0_patch_data_offset
=
415 get_tcs_out_patch0_patch_data_offset(ctx
);
416 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
417 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
419 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_patch_data_offset
,
420 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
427 LLVMTypeRef types
[MAX_ARGS
];
428 LLVMValueRef
*assign
[MAX_ARGS
];
429 unsigned array_params_mask
;
432 uint8_t num_sgprs_used
;
433 uint8_t num_vgprs_used
;
436 enum ac_arg_regfile
{
442 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
443 LLVMValueRef
*param_ptr
)
445 assert(info
->count
< MAX_ARGS
);
447 info
->assign
[info
->count
] = param_ptr
;
448 info
->types
[info
->count
] = type
;
451 if (regfile
== ARG_SGPR
) {
452 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
455 assert(regfile
== ARG_VGPR
);
456 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
461 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
463 info
->array_params_mask
|= (1 << info
->count
);
464 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
467 static void assign_arguments(LLVMValueRef main_function
,
468 struct arg_info
*info
)
471 for (i
= 0; i
< info
->count
; i
++) {
473 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
478 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
479 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
480 unsigned num_return_elems
,
481 struct arg_info
*args
,
482 unsigned max_workgroup_size
,
485 LLVMTypeRef main_function_type
, ret_type
;
486 LLVMBasicBlockRef main_function_body
;
488 if (num_return_elems
)
489 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
490 num_return_elems
, true);
492 ret_type
= LLVMVoidTypeInContext(ctx
);
494 /* Setup the function */
496 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
497 LLVMValueRef main_function
=
498 LLVMAddFunction(module
, "main", main_function_type
);
500 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
501 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
503 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
504 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
505 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
507 if (args
->array_params_mask
& (1 << i
)) {
508 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
509 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
510 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
514 if (max_workgroup_size
) {
515 ac_llvm_add_target_dep_function_attr(main_function
,
516 "amdgpu-max-work-group-size",
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
, uint8_t num_sgprs
,
543 uint32_t indirect_offset
)
545 ud_info
->sgpr_idx
= *sgpr_idx
;
546 ud_info
->num_sgprs
= num_sgprs
;
547 ud_info
->indirect
= indirect_offset
> 0;
548 ud_info
->indirect_offset
= indirect_offset
;
549 *sgpr_idx
+= num_sgprs
;
553 set_loc_shader(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
556 struct radv_userdata_info
*ud_info
=
557 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
560 set_loc(ud_info
, sgpr_idx
, num_sgprs
, 0);
564 set_loc_desc(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
565 uint32_t indirect_offset
)
567 struct radv_userdata_info
*ud_info
=
568 &ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[idx
];
571 set_loc(ud_info
, sgpr_idx
, 2, indirect_offset
);
574 struct user_sgpr_info
{
575 bool need_ring_offsets
;
577 bool indirect_all_descriptor_sets
;
580 static bool needs_view_index_sgpr(struct radv_shader_context
*ctx
,
581 gl_shader_stage stage
)
584 case MESA_SHADER_VERTEX
:
585 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
586 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
589 case MESA_SHADER_TESS_EVAL
:
590 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
593 case MESA_SHADER_GEOMETRY
:
594 case MESA_SHADER_TESS_CTRL
:
595 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
605 count_vs_user_sgprs(struct radv_shader_context
*ctx
)
609 count
+= ctx
->shader_info
->info
.vs
.has_vertex_buffers
? 2 : 0;
610 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
615 static void allocate_user_sgprs(struct radv_shader_context
*ctx
,
616 gl_shader_stage stage
,
617 bool has_previous_stage
,
618 gl_shader_stage previous_stage
,
619 bool needs_view_index
,
620 struct user_sgpr_info
*user_sgpr_info
)
622 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
624 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
625 if (stage
== MESA_SHADER_GEOMETRY
||
626 stage
== MESA_SHADER_VERTEX
||
627 stage
== MESA_SHADER_TESS_CTRL
||
628 stage
== MESA_SHADER_TESS_EVAL
||
629 ctx
->is_gs_copy_shader
)
630 user_sgpr_info
->need_ring_offsets
= true;
632 if (stage
== MESA_SHADER_FRAGMENT
&&
633 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
634 user_sgpr_info
->need_ring_offsets
= true;
636 /* 2 user sgprs will nearly always be allocated for scratch/rings */
637 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
638 user_sgpr_info
->sgpr_count
+= 2;
642 case MESA_SHADER_COMPUTE
:
643 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
644 user_sgpr_info
->sgpr_count
+= 3;
646 case MESA_SHADER_FRAGMENT
:
647 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
649 case MESA_SHADER_VERTEX
:
650 if (!ctx
->is_gs_copy_shader
)
651 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
653 case MESA_SHADER_TESS_CTRL
:
654 if (has_previous_stage
) {
655 if (previous_stage
== MESA_SHADER_VERTEX
)
656 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
659 case MESA_SHADER_TESS_EVAL
:
661 case MESA_SHADER_GEOMETRY
:
662 if (has_previous_stage
) {
663 if (previous_stage
== MESA_SHADER_VERTEX
) {
664 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
672 if (needs_view_index
)
673 user_sgpr_info
->sgpr_count
++;
675 if (ctx
->shader_info
->info
.loads_push_constants
)
676 user_sgpr_info
->sgpr_count
+= 2;
678 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
? 32 : 16;
679 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_info
->sgpr_count
;
681 if (remaining_sgprs
/ 2 < util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
)) {
682 user_sgpr_info
->sgpr_count
+= 2;
683 user_sgpr_info
->indirect_all_descriptor_sets
= true;
685 user_sgpr_info
->sgpr_count
+= util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
) * 2;
690 declare_global_input_sgprs(struct radv_shader_context
*ctx
,
691 gl_shader_stage stage
,
692 bool has_previous_stage
,
693 gl_shader_stage previous_stage
,
694 const struct user_sgpr_info
*user_sgpr_info
,
695 struct arg_info
*args
,
696 LLVMValueRef
*desc_sets
)
698 LLVMTypeRef type
= ac_array_in_const_addr_space(ctx
->ac
.i8
);
699 unsigned num_sets
= ctx
->options
->layout
?
700 ctx
->options
->layout
->num_sets
: 0;
701 unsigned stage_mask
= 1 << stage
;
703 if (has_previous_stage
)
704 stage_mask
|= 1 << previous_stage
;
706 /* 1 for each descriptor set */
707 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
708 for (unsigned i
= 0; i
< num_sets
; ++i
) {
709 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
710 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
711 add_array_arg(args
, type
,
712 &ctx
->descriptor_sets
[i
]);
716 add_array_arg(args
, ac_array_in_const_addr_space(type
), desc_sets
);
719 if (ctx
->shader_info
->info
.loads_push_constants
) {
720 /* 1 for push constants and dynamic descriptors */
721 add_array_arg(args
, type
, &ctx
->abi
.push_constants
);
726 declare_vs_specific_input_sgprs(struct radv_shader_context
*ctx
,
727 gl_shader_stage stage
,
728 bool has_previous_stage
,
729 gl_shader_stage previous_stage
,
730 struct arg_info
*args
)
732 if (!ctx
->is_gs_copy_shader
&&
733 (stage
== MESA_SHADER_VERTEX
||
734 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
735 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
736 add_arg(args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
737 &ctx
->vertex_buffers
);
739 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
740 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
741 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
742 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
748 declare_vs_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
750 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
751 if (!ctx
->is_gs_copy_shader
) {
752 if (ctx
->options
->key
.vs
.as_ls
) {
753 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
754 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
756 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
757 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
759 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
764 declare_tes_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
766 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
767 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
768 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
769 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
773 set_global_input_locs(struct radv_shader_context
*ctx
, gl_shader_stage stage
,
774 bool has_previous_stage
, gl_shader_stage previous_stage
,
775 const struct user_sgpr_info
*user_sgpr_info
,
776 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
778 unsigned num_sets
= ctx
->options
->layout
?
779 ctx
->options
->layout
->num_sets
: 0;
780 unsigned stage_mask
= 1 << stage
;
782 if (has_previous_stage
)
783 stage_mask
|= 1 << previous_stage
;
785 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
786 for (unsigned i
= 0; i
< num_sets
; ++i
) {
787 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
788 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
789 set_loc_desc(ctx
, i
, user_sgpr_idx
, 0);
791 ctx
->descriptor_sets
[i
] = NULL
;
794 set_loc_shader(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
797 for (unsigned i
= 0; i
< num_sets
; ++i
) {
798 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
799 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
800 set_loc_desc(ctx
, i
, user_sgpr_idx
, i
* 8);
801 ctx
->descriptor_sets
[i
] =
802 ac_build_load_to_sgpr(&ctx
->ac
,
804 LLVMConstInt(ctx
->ac
.i32
, i
, false));
807 ctx
->descriptor_sets
[i
] = NULL
;
809 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
812 if (ctx
->shader_info
->info
.loads_push_constants
) {
813 set_loc_shader(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
, 2);
818 set_vs_specific_input_locs(struct radv_shader_context
*ctx
,
819 gl_shader_stage stage
, bool has_previous_stage
,
820 gl_shader_stage previous_stage
,
821 uint8_t *user_sgpr_idx
)
823 if (!ctx
->is_gs_copy_shader
&&
824 (stage
== MESA_SHADER_VERTEX
||
825 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
826 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
827 set_loc_shader(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
832 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
835 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
836 user_sgpr_idx
, vs_num
);
840 static void set_llvm_calling_convention(LLVMValueRef func
,
841 gl_shader_stage stage
)
843 enum radeon_llvm_calling_convention calling_conv
;
846 case MESA_SHADER_VERTEX
:
847 case MESA_SHADER_TESS_EVAL
:
848 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
850 case MESA_SHADER_GEOMETRY
:
851 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
853 case MESA_SHADER_TESS_CTRL
:
854 calling_conv
= HAVE_LLVM
>= 0x0500 ? RADEON_LLVM_AMDGPU_HS
: RADEON_LLVM_AMDGPU_VS
;
856 case MESA_SHADER_FRAGMENT
:
857 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
859 case MESA_SHADER_COMPUTE
:
860 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
863 unreachable("Unhandle shader type");
866 LLVMSetFunctionCallConv(func
, calling_conv
);
869 static void create_function(struct radv_shader_context
*ctx
,
870 gl_shader_stage stage
,
871 bool has_previous_stage
,
872 gl_shader_stage previous_stage
)
874 uint8_t user_sgpr_idx
;
875 struct user_sgpr_info user_sgpr_info
;
876 struct arg_info args
= {};
877 LLVMValueRef desc_sets
;
878 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
879 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
880 previous_stage
, needs_view_index
, &user_sgpr_info
);
882 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
883 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
888 case MESA_SHADER_COMPUTE
:
889 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
890 previous_stage
, &user_sgpr_info
,
893 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
894 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
895 &ctx
->abi
.num_work_groups
);
898 for (int i
= 0; i
< 3; i
++) {
899 ctx
->abi
.workgroup_ids
[i
] = NULL
;
900 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
901 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
902 &ctx
->abi
.workgroup_ids
[i
]);
906 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
907 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
908 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
909 &ctx
->abi
.local_invocation_ids
);
911 case MESA_SHADER_VERTEX
:
912 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
913 previous_stage
, &user_sgpr_info
,
915 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
916 previous_stage
, &args
);
918 if (needs_view_index
)
919 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
920 &ctx
->abi
.view_index
);
921 if (ctx
->options
->key
.vs
.as_es
)
922 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
925 declare_vs_input_vgprs(ctx
, &args
);
927 case MESA_SHADER_TESS_CTRL
:
928 if (has_previous_stage
) {
929 // First 6 system regs
930 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
931 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
932 &ctx
->merged_wave_info
);
933 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
934 &ctx
->tess_factor_offset
);
936 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
937 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
938 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
940 declare_global_input_sgprs(ctx
, stage
,
943 &user_sgpr_info
, &args
,
945 declare_vs_specific_input_sgprs(ctx
, stage
,
947 previous_stage
, &args
);
949 if (needs_view_index
)
950 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
951 &ctx
->abi
.view_index
);
953 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
954 &ctx
->abi
.tcs_patch_id
);
955 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
956 &ctx
->abi
.tcs_rel_ids
);
958 declare_vs_input_vgprs(ctx
, &args
);
960 declare_global_input_sgprs(ctx
, stage
,
963 &user_sgpr_info
, &args
,
966 if (needs_view_index
)
967 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
968 &ctx
->abi
.view_index
);
970 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
971 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
972 &ctx
->tess_factor_offset
);
973 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
974 &ctx
->abi
.tcs_patch_id
);
975 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
976 &ctx
->abi
.tcs_rel_ids
);
979 case MESA_SHADER_TESS_EVAL
:
980 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
981 previous_stage
, &user_sgpr_info
,
984 if (needs_view_index
)
985 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
986 &ctx
->abi
.view_index
);
988 if (ctx
->options
->key
.tes
.as_es
) {
989 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
990 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
991 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
994 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
995 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
997 declare_tes_input_vgprs(ctx
, &args
);
999 case MESA_SHADER_GEOMETRY
:
1000 if (has_previous_stage
) {
1001 // First 6 system regs
1002 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1003 &ctx
->gs2vs_offset
);
1004 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1005 &ctx
->merged_wave_info
);
1006 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1008 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
1009 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1010 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1012 declare_global_input_sgprs(ctx
, stage
,
1015 &user_sgpr_info
, &args
,
1018 if (previous_stage
!= MESA_SHADER_TESS_EVAL
) {
1019 declare_vs_specific_input_sgprs(ctx
, stage
,
1025 if (needs_view_index
)
1026 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1027 &ctx
->abi
.view_index
);
1029 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1030 &ctx
->gs_vtx_offset
[0]);
1031 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1032 &ctx
->gs_vtx_offset
[2]);
1033 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1034 &ctx
->abi
.gs_prim_id
);
1035 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1036 &ctx
->abi
.gs_invocation_id
);
1037 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1038 &ctx
->gs_vtx_offset
[4]);
1040 if (previous_stage
== MESA_SHADER_VERTEX
) {
1041 declare_vs_input_vgprs(ctx
, &args
);
1043 declare_tes_input_vgprs(ctx
, &args
);
1046 declare_global_input_sgprs(ctx
, stage
,
1049 &user_sgpr_info
, &args
,
1052 if (needs_view_index
)
1053 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1054 &ctx
->abi
.view_index
);
1056 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
1057 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
1058 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1059 &ctx
->gs_vtx_offset
[0]);
1060 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1061 &ctx
->gs_vtx_offset
[1]);
1062 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1063 &ctx
->abi
.gs_prim_id
);
1064 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1065 &ctx
->gs_vtx_offset
[2]);
1066 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1067 &ctx
->gs_vtx_offset
[3]);
1068 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1069 &ctx
->gs_vtx_offset
[4]);
1070 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1071 &ctx
->gs_vtx_offset
[5]);
1072 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1073 &ctx
->abi
.gs_invocation_id
);
1076 case MESA_SHADER_FRAGMENT
:
1077 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
1078 previous_stage
, &user_sgpr_info
,
1081 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
1082 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1083 &ctx
->sample_pos_offset
);
1085 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1086 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1087 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1088 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1089 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1090 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1091 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1092 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1093 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1094 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1095 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1096 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1097 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1098 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1099 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1100 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1101 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1104 unreachable("Shader stage not implemented");
1107 ctx
->main_function
= create_llvm_function(
1108 ctx
->context
, ctx
->ac
.module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1109 ctx
->max_workgroup_size
,
1110 ctx
->options
->unsafe_math
);
1111 set_llvm_calling_convention(ctx
->main_function
, stage
);
1114 ctx
->shader_info
->num_input_vgprs
= 0;
1115 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1117 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1119 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1120 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1122 assign_arguments(ctx
->main_function
, &args
);
1126 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1127 set_loc_shader(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1129 if (ctx
->options
->supports_spill
) {
1130 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1131 LLVMPointerType(ctx
->ac
.i8
, AC_CONST_ADDR_SPACE
),
1132 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1133 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1134 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1138 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1139 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1140 if (has_previous_stage
)
1143 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1144 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1147 case MESA_SHADER_COMPUTE
:
1148 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1149 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1153 case MESA_SHADER_VERTEX
:
1154 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1155 previous_stage
, &user_sgpr_idx
);
1156 if (ctx
->abi
.view_index
)
1157 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1159 case MESA_SHADER_TESS_CTRL
:
1160 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1161 previous_stage
, &user_sgpr_idx
);
1162 if (ctx
->abi
.view_index
)
1163 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1165 case MESA_SHADER_TESS_EVAL
:
1166 if (ctx
->abi
.view_index
)
1167 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1169 case MESA_SHADER_GEOMETRY
:
1170 if (has_previous_stage
) {
1171 if (previous_stage
== MESA_SHADER_VERTEX
)
1172 set_vs_specific_input_locs(ctx
, stage
,
1177 if (ctx
->abi
.view_index
)
1178 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1180 case MESA_SHADER_FRAGMENT
:
1181 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
1182 set_loc_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
,
1187 unreachable("Shader stage not implemented");
1190 if (stage
== MESA_SHADER_TESS_CTRL
||
1191 (stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_ls
) ||
1192 /* GFX9 has the ESGS ring buffer in LDS. */
1193 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
1194 ac_declare_lds_as_pointer(&ctx
->ac
);
1197 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1202 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
1203 unsigned desc_set
, unsigned binding
)
1205 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1206 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
1207 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
1208 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
1209 unsigned base_offset
= layout
->binding
[binding
].offset
;
1210 LLVMValueRef offset
, stride
;
1212 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
1213 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
1214 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
1215 layout
->binding
[binding
].dynamic_offset_offset
;
1216 desc_ptr
= ctx
->abi
.push_constants
;
1217 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
1218 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1220 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
1222 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
1223 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, stride
, "");
1224 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
1226 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
1227 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
1228 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1234 /* The offchip buffer layout for TCS->TES is
1236 * - attribute 0 of patch 0 vertex 0
1237 * - attribute 0 of patch 0 vertex 1
1238 * - attribute 0 of patch 0 vertex 2
1240 * - attribute 0 of patch 1 vertex 0
1241 * - attribute 0 of patch 1 vertex 1
1243 * - attribute 1 of patch 0 vertex 0
1244 * - attribute 1 of patch 0 vertex 1
1246 * - per patch attribute 0 of patch 0
1247 * - per patch attribute 0 of patch 1
1250 * Note that every attribute has 4 components.
1252 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
1254 uint32_t num_patches
= ctx
->tcs_num_patches
;
1255 uint32_t num_tcs_outputs
;
1256 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
1257 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
1259 num_tcs_outputs
= ctx
->options
->key
.tes
.tcs_num_outputs
;
1261 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
1262 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
1264 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
1267 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
1268 LLVMValueRef vertex_index
)
1270 LLVMValueRef param_stride
;
1272 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
* ctx
->tcs_num_patches
, false);
1274 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
1275 return param_stride
;
1278 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
1279 LLVMValueRef vertex_index
,
1280 LLVMValueRef param_index
)
1282 LLVMValueRef base_addr
;
1283 LLVMValueRef param_stride
, constant16
;
1284 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
1285 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
, false);
1286 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1287 param_stride
= calc_param_stride(ctx
, vertex_index
);
1289 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
1290 vertices_per_patch
, "");
1292 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
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
);
1472 if (store_lds
|| is_tess_factor
) {
1473 LLVMValueRef dw_addr_chan
=
1474 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1475 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
1476 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
1479 if (!is_tess_factor
&& writemask
!= 0xF)
1480 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
1481 buf_addr
, ctx
->oc_lds
,
1482 4 * (base
+ chan
), 1, 0, true, false);
1485 if (writemask
== 0xF) {
1486 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
1487 buf_addr
, ctx
->oc_lds
,
1488 (base
* 4), 1, 0, true, false);
1493 load_tes_input(struct ac_shader_abi
*abi
,
1495 LLVMValueRef vertex_index
,
1496 LLVMValueRef param_index
,
1497 unsigned const_index
,
1499 unsigned driver_location
,
1501 unsigned num_components
,
1506 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1507 LLVMValueRef buf_addr
;
1508 LLVMValueRef result
;
1509 unsigned param
= shader_io_get_unique_index(location
);
1511 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
1516 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
1517 is_compact
, vertex_index
, param_index
);
1519 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
1520 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
1522 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
1523 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
1524 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
1529 load_gs_input(struct ac_shader_abi
*abi
,
1531 unsigned driver_location
,
1533 unsigned num_components
,
1534 unsigned vertex_index
,
1535 unsigned const_index
,
1538 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1539 LLVMValueRef vtx_offset
;
1540 unsigned param
, vtx_offset_param
;
1541 LLVMValueRef value
[4], result
;
1543 vtx_offset_param
= vertex_index
;
1544 assert(vtx_offset_param
< 6);
1545 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
1546 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1548 param
= shader_io_get_unique_index(location
);
1550 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1551 if (ctx
->ac
.chip_class
>= GFX9
) {
1552 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1553 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1554 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
1555 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1557 LLVMValueRef soffset
=
1558 LLVMConstInt(ctx
->ac
.i32
,
1559 (param
* 4 + i
+ const_index
) * 256,
1562 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
1565 vtx_offset
, soffset
,
1566 0, 1, 0, true, false);
1568 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
],
1572 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1573 result
= ac_to_integer(&ctx
->ac
, result
);
1578 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
1580 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1581 ac_build_kill_if_false(&ctx
->ac
, visible
);
1584 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
1585 enum glsl_interp_mode interp
, unsigned location
)
1587 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1590 case INTERP_MODE_FLAT
:
1593 case INTERP_MODE_SMOOTH
:
1594 case INTERP_MODE_NONE
:
1595 if (location
== INTERP_CENTER
)
1596 return ctx
->persp_center
;
1597 else if (location
== INTERP_CENTROID
)
1598 return ctx
->persp_centroid
;
1599 else if (location
== INTERP_SAMPLE
)
1600 return ctx
->persp_sample
;
1602 case INTERP_MODE_NOPERSPECTIVE
:
1603 if (location
== INTERP_CENTER
)
1604 return ctx
->linear_center
;
1605 else if (location
== INTERP_CENTROID
)
1606 return ctx
->linear_centroid
;
1607 else if (location
== INTERP_SAMPLE
)
1608 return ctx
->linear_sample
;
1614 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
1615 LLVMValueRef sample_id
)
1617 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1619 LLVMValueRef result
;
1620 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
1622 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
1623 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
1625 sample_id
= LLVMBuildAdd(ctx
->ac
.builder
, sample_id
, ctx
->sample_pos_offset
, "");
1626 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
1632 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1634 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1635 uint8_t log2_ps_iter_samples
= ctx
->shader_info
->info
.ps
.force_persample
?
1636 ctx
->options
->key
.fs
.log2_num_samples
:
1637 ctx
->options
->key
.fs
.log2_ps_iter_samples
;
1639 /* The bit pattern matches that used by fixed function fragment
1641 static const uint16_t ps_iter_masks
[] = {
1642 0xffff, /* not used */
1648 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
1650 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
1652 LLVMValueRef result
, sample_id
;
1653 sample_id
= ac_unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
1654 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
1655 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
1661 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
1663 LLVMValueRef gs_next_vertex
;
1664 LLVMValueRef can_emit
;
1666 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1668 assert(stream
== 0);
1670 /* Write vertex attribute values to GSVS ring */
1671 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
1672 ctx
->gs_next_vertex
,
1675 /* If this thread has already emitted the declared maximum number of
1676 * vertices, kill it: excessive vertex emissions are not supposed to
1677 * have any effect, and GS threads have no externally observable
1678 * effects other than emitting vertices.
1680 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
1681 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
1682 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1684 /* loop num outputs */
1686 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1687 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1692 if (!(ctx
->output_mask
& (1ull << i
)))
1695 if (i
== VARYING_SLOT_CLIP_DIST0
) {
1696 /* pack clip and cull into a single set of slots */
1697 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
1701 for (unsigned j
= 0; j
< length
; j
++) {
1702 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1704 LLVMValueRef voffset
= LLVMConstInt(ctx
->ac
.i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
1705 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1706 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1708 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1710 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
1712 voffset
, ctx
->gs2vs_offset
, 0,
1718 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1720 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
1722 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
1726 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1728 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1729 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1733 load_tess_coord(struct ac_shader_abi
*abi
)
1735 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1737 LLVMValueRef coord
[4] = {
1744 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
1745 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1746 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1748 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1752 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1754 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1755 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
1759 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1761 return abi
->base_vertex
;
1764 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1765 LLVMValueRef buffer_ptr
, bool write
)
1767 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1768 LLVMValueRef result
;
1770 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1772 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1773 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1778 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1780 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1781 LLVMValueRef result
;
1783 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1785 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1786 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1791 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1792 unsigned descriptor_set
,
1793 unsigned base_index
,
1794 unsigned constant_index
,
1796 enum ac_descriptor_type desc_type
,
1797 bool image
, bool write
,
1800 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1801 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1802 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
1803 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1804 unsigned offset
= binding
->offset
;
1805 unsigned stride
= binding
->size
;
1807 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1810 assert(base_index
< layout
->binding_count
);
1812 switch (desc_type
) {
1814 type
= ctx
->ac
.v8i32
;
1818 type
= ctx
->ac
.v8i32
;
1822 case AC_DESC_SAMPLER
:
1823 type
= ctx
->ac
.v4i32
;
1824 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
1829 case AC_DESC_BUFFER
:
1830 type
= ctx
->ac
.v4i32
;
1834 unreachable("invalid desc_type\n");
1837 offset
+= constant_index
* stride
;
1839 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1840 (!index
|| binding
->immutable_samplers_equal
)) {
1841 if (binding
->immutable_samplers_equal
)
1844 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1846 LLVMValueRef constants
[] = {
1847 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1848 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1849 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1850 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1852 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1855 assert(stride
% type_size
== 0);
1858 index
= ctx
->ac
.i32_0
;
1860 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1862 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
1863 list
= LLVMBuildPointerCast(builder
, list
, ac_array_in_const_addr_space(type
), "");
1865 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
1868 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1869 * so we may need to fix it up. */
1871 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1872 unsigned adjustment
,
1875 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1878 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1880 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1881 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1883 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1885 /* For the integer-like cases, do a natural sign extension.
1887 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1888 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1891 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1892 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1893 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1894 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1896 /* Convert back to the right type. */
1897 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1899 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1900 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1901 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1902 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1903 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1904 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1911 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1912 struct nir_variable
*variable
)
1914 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
1915 LLVMValueRef t_offset
;
1916 LLVMValueRef t_list
;
1918 LLVMValueRef buffer_index
;
1919 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1920 uint8_t input_usage_mask
=
1921 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
1922 unsigned num_channels
= util_last_bit(input_usage_mask
);
1924 variable
->data
.driver_location
= variable
->data
.location
* 4;
1926 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1927 LLVMValueRef output
[4];
1928 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1930 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1931 uint32_t divisor
= ctx
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
1934 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.instance_id
,
1935 ctx
->abi
.start_instance
, "");
1938 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
1939 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
1942 if (ctx
->options
->key
.vs
.as_ls
) {
1943 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1944 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1946 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1947 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1950 buffer_index
= ctx
->ac
.i32_0
;
1953 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
1954 ctx
->abi
.base_vertex
, "");
1955 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_index
, false);
1957 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
1959 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
1962 num_channels
, false, true);
1964 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
1966 for (unsigned chan
= 0; chan
< 4; chan
++) {
1967 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
1968 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
1971 unsigned alpha_adjust
= (ctx
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
1972 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
1974 for (unsigned chan
= 0; chan
< 4; chan
++) {
1975 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] =
1976 ac_to_integer(&ctx
->ac
, output
[chan
]);
1981 static void interp_fs_input(struct radv_shader_context
*ctx
,
1983 LLVMValueRef interp_param
,
1984 LLVMValueRef prim_mask
,
1985 LLVMValueRef result
[4])
1987 LLVMValueRef attr_number
;
1990 bool interp
= interp_param
!= NULL
;
1992 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
1994 /* fs.constant returns the param from the middle vertex, so it's not
1995 * really useful for flat shading. It's meant to be used for custom
1996 * interpolation (but the intrinsic can't fetch from the other two
1999 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
2000 * to do the right thing. The only reason we use fs.constant is that
2001 * fs.interp cannot be used on integers, because they can be equal
2005 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
2008 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2010 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2014 for (chan
= 0; chan
< 4; chan
++) {
2015 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2018 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
2023 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
2024 LLVMConstInt(ctx
->ac
.i32
, 2, false),
2033 handle_fs_input_decl(struct radv_shader_context
*ctx
,
2034 struct nir_variable
*variable
)
2036 int idx
= variable
->data
.location
;
2037 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2038 LLVMValueRef interp
;
2040 variable
->data
.driver_location
= idx
* 4;
2041 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
2043 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
2044 unsigned interp_type
;
2045 if (variable
->data
.sample
)
2046 interp_type
= INTERP_SAMPLE
;
2047 else if (variable
->data
.centroid
)
2048 interp_type
= INTERP_CENTROID
;
2050 interp_type
= INTERP_CENTER
;
2052 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
2056 for (unsigned i
= 0; i
< attrib_count
; ++i
)
2057 ctx
->inputs
[ac_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
2062 handle_vs_inputs(struct radv_shader_context
*ctx
,
2063 struct nir_shader
*nir
) {
2064 nir_foreach_variable(variable
, &nir
->inputs
)
2065 handle_vs_input_decl(ctx
, variable
);
2069 prepare_interp_optimize(struct radv_shader_context
*ctx
,
2070 struct nir_shader
*nir
)
2072 if (!ctx
->options
->key
.fs
.multisample
)
2075 bool uses_center
= false;
2076 bool uses_centroid
= false;
2077 nir_foreach_variable(variable
, &nir
->inputs
) {
2078 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
2079 variable
->data
.sample
)
2082 if (variable
->data
.centroid
)
2083 uses_centroid
= true;
2088 if (uses_center
&& uses_centroid
) {
2089 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
2090 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
2091 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
2096 handle_fs_inputs(struct radv_shader_context
*ctx
,
2097 struct nir_shader
*nir
)
2099 prepare_interp_optimize(ctx
, nir
);
2101 nir_foreach_variable(variable
, &nir
->inputs
)
2102 handle_fs_input_decl(ctx
, variable
);
2106 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
2107 ctx
->shader_info
->info
.needs_multiview_view_index
)
2108 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
2110 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
2111 LLVMValueRef interp_param
;
2112 LLVMValueRef
*inputs
= ctx
->inputs
+ac_llvm_reg_index_soa(i
, 0);
2114 if (!(ctx
->input_mask
& (1ull << i
)))
2117 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
2118 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
2119 interp_param
= *inputs
;
2120 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
2124 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
2126 } else if (i
== VARYING_SLOT_POS
) {
2127 for(int i
= 0; i
< 3; ++i
)
2128 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
2130 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2131 ctx
->abi
.frag_pos
[3]);
2134 ctx
->shader_info
->fs
.num_interp
= index
;
2135 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
2137 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
2138 ctx
->abi
.view_index
= ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2142 scan_shader_output_decl(struct radv_shader_context
*ctx
,
2143 struct nir_variable
*variable
,
2144 struct nir_shader
*shader
,
2145 gl_shader_stage stage
)
2147 int idx
= variable
->data
.location
+ variable
->data
.index
;
2148 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2149 uint64_t mask_attribs
;
2151 variable
->data
.driver_location
= idx
* 4;
2153 /* tess ctrl has it's own load/store paths for outputs */
2154 if (stage
== MESA_SHADER_TESS_CTRL
)
2157 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
2158 if (stage
== MESA_SHADER_VERTEX
||
2159 stage
== MESA_SHADER_TESS_EVAL
||
2160 stage
== MESA_SHADER_GEOMETRY
) {
2161 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2162 int length
= shader
->info
.clip_distance_array_size
+
2163 shader
->info
.cull_distance_array_size
;
2164 if (stage
== MESA_SHADER_VERTEX
) {
2165 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2166 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2168 if (stage
== MESA_SHADER_TESS_EVAL
) {
2169 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2170 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2177 mask_attribs
= 1ull << idx
;
2181 ctx
->output_mask
|= mask_attribs
;
2185 /* Initialize arguments for the shader export intrinsic */
2187 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
2188 LLVMValueRef
*values
,
2189 unsigned enabled_channels
,
2191 struct ac_export_args
*args
)
2193 /* Specify the channels that are enabled. */
2194 args
->enabled_channels
= enabled_channels
;
2196 /* Specify whether the EXEC mask represents the valid mask */
2197 args
->valid_mask
= 0;
2199 /* Specify whether this is the last export */
2202 /* Specify the target we are exporting */
2203 args
->target
= target
;
2205 args
->compr
= false;
2206 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
2207 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
2208 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
2209 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
2211 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
2212 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
2213 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
2214 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
2215 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
2218 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2219 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2220 unsigned bits
, bool hi
) = NULL
;
2222 switch(col_format
) {
2223 case V_028714_SPI_SHADER_ZERO
:
2224 args
->enabled_channels
= 0; /* writemask */
2225 args
->target
= V_008DFC_SQ_EXP_NULL
;
2228 case V_028714_SPI_SHADER_32_R
:
2229 args
->enabled_channels
= 1;
2230 args
->out
[0] = values
[0];
2233 case V_028714_SPI_SHADER_32_GR
:
2234 args
->enabled_channels
= 0x3;
2235 args
->out
[0] = values
[0];
2236 args
->out
[1] = values
[1];
2239 case V_028714_SPI_SHADER_32_AR
:
2240 args
->enabled_channels
= 0x9;
2241 args
->out
[0] = values
[0];
2242 args
->out
[3] = values
[3];
2245 case V_028714_SPI_SHADER_FP16_ABGR
:
2246 args
->enabled_channels
= 0x5;
2247 packf
= ac_build_cvt_pkrtz_f16
;
2250 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2251 args
->enabled_channels
= 0x5;
2252 packf
= ac_build_cvt_pknorm_u16
;
2255 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2256 args
->enabled_channels
= 0x5;
2257 packf
= ac_build_cvt_pknorm_i16
;
2260 case V_028714_SPI_SHADER_UINT16_ABGR
:
2261 args
->enabled_channels
= 0x5;
2262 packi
= ac_build_cvt_pk_u16
;
2265 case V_028714_SPI_SHADER_SINT16_ABGR
:
2266 args
->enabled_channels
= 0x5;
2267 packi
= ac_build_cvt_pk_i16
;
2271 case V_028714_SPI_SHADER_32_ABGR
:
2272 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2276 /* Pack f16 or norm_i16/u16. */
2278 for (chan
= 0; chan
< 2; chan
++) {
2279 LLVMValueRef pack_args
[2] = {
2281 values
[2 * chan
+ 1]
2283 LLVMValueRef packed
;
2285 packed
= packf(&ctx
->ac
, pack_args
);
2286 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2288 args
->compr
= 1; /* COMPR flag */
2293 for (chan
= 0; chan
< 2; chan
++) {
2294 LLVMValueRef pack_args
[2] = {
2295 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2296 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2298 LLVMValueRef packed
;
2300 packed
= packi(&ctx
->ac
, pack_args
,
2301 is_int8
? 8 : is_int10
? 10 : 16,
2303 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2305 args
->compr
= 1; /* COMPR flag */
2310 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2312 for (unsigned i
= 0; i
< 4; ++i
) {
2313 if (!(args
->enabled_channels
& (1 << i
)))
2316 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
2321 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
2322 LLVMValueRef
*values
, unsigned enabled_channels
)
2324 struct ac_export_args args
;
2326 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
2327 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2328 ac_build_export(&ctx
->ac
, &args
);
2332 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
2334 LLVMValueRef output
=
2335 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
2337 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
2341 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2342 bool export_prim_id
, bool export_layer_id
,
2343 struct radv_vs_output_info
*outinfo
)
2345 uint32_t param_count
= 0;
2347 unsigned pos_idx
, num_pos_exports
= 0;
2348 struct ac_export_args args
, pos_args
[4] = {};
2349 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2352 if (ctx
->options
->key
.has_multiview_view_index
) {
2353 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2355 for(unsigned i
= 0; i
< 4; ++i
)
2356 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2357 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2360 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
2361 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2364 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2365 sizeof(outinfo
->vs_output_param_offset
));
2367 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
2368 LLVMValueRef slots
[8];
2371 if (outinfo
->cull_dist_mask
)
2372 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
2374 i
= VARYING_SLOT_CLIP_DIST0
;
2375 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
2376 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2378 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
2379 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
2381 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
2382 target
= V_008DFC_SQ_EXP_POS
+ 3;
2383 si_llvm_init_export_args(ctx
, &slots
[4], 0xf, target
, &args
);
2384 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2385 &args
, sizeof(args
));
2388 target
= V_008DFC_SQ_EXP_POS
+ 2;
2389 si_llvm_init_export_args(ctx
, &slots
[0], 0xf, target
, &args
);
2390 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2391 &args
, sizeof(args
));
2395 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
2396 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
2397 for (unsigned j
= 0; j
< 4; j
++)
2398 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
2400 si_llvm_init_export_args(ctx
, pos_values
, 0xf, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2402 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
2403 outinfo
->writes_pointsize
= true;
2404 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
2407 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
2408 outinfo
->writes_layer
= true;
2409 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
2412 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
2413 outinfo
->writes_viewport_index
= true;
2414 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
2417 if (outinfo
->writes_pointsize
||
2418 outinfo
->writes_layer
||
2419 outinfo
->writes_viewport_index
) {
2420 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
2421 (outinfo
->writes_layer
== true ? 4 : 0));
2422 pos_args
[1].valid_mask
= 0;
2423 pos_args
[1].done
= 0;
2424 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2425 pos_args
[1].compr
= 0;
2426 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2427 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2428 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2429 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2431 if (outinfo
->writes_pointsize
== true)
2432 pos_args
[1].out
[0] = psize_value
;
2433 if (outinfo
->writes_layer
== true)
2434 pos_args
[1].out
[2] = layer_value
;
2435 if (outinfo
->writes_viewport_index
== true) {
2436 if (ctx
->options
->chip_class
>= GFX9
) {
2437 /* GFX9 has the layer in out.z[10:0] and the viewport
2438 * index in out.z[19:16].
2440 LLVMValueRef v
= viewport_index_value
;
2441 v
= ac_to_integer(&ctx
->ac
, v
);
2442 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2443 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2445 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2446 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2448 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2449 pos_args
[1].enabled_channels
|= 1 << 2;
2451 pos_args
[1].out
[3] = viewport_index_value
;
2452 pos_args
[1].enabled_channels
|= 1 << 3;
2456 for (i
= 0; i
< 4; i
++) {
2457 if (pos_args
[i
].out
[0])
2462 for (i
= 0; i
< 4; i
++) {
2463 if (!pos_args
[i
].out
[0])
2466 /* Specify the target we are exporting */
2467 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2468 if (pos_idx
== num_pos_exports
)
2469 pos_args
[i
].done
= 1;
2470 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2473 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2474 LLVMValueRef values
[4];
2475 if (!(ctx
->output_mask
& (1ull << i
)))
2478 if (i
!= VARYING_SLOT_LAYER
&&
2479 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
2480 i
< VARYING_SLOT_VAR0
)
2483 for (unsigned j
= 0; j
< 4; j
++)
2484 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2486 unsigned output_usage_mask
;
2488 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2489 !ctx
->is_gs_copy_shader
) {
2491 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2492 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2494 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2496 /* Enable all channels for the GS copy shader because
2497 * we don't know the output usage mask currently.
2499 output_usage_mask
= 0xf;
2502 radv_export_param(ctx
, param_count
, values
, output_usage_mask
);
2504 outinfo
->vs_output_param_offset
[i
] = param_count
++;
2507 if (export_prim_id
) {
2508 LLVMValueRef values
[4];
2510 values
[0] = ctx
->vs_prim_id
;
2511 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
2512 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2513 for (unsigned j
= 1; j
< 4; j
++)
2514 values
[j
] = ctx
->ac
.f32_0
;
2516 radv_export_param(ctx
, param_count
, values
, 0x1);
2518 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
2519 outinfo
->export_prim_id
= true;
2522 if (export_layer_id
&& layer_value
) {
2523 LLVMValueRef values
[4];
2525 values
[0] = layer_value
;
2526 for (unsigned j
= 1; j
< 4; j
++)
2527 values
[j
] = ctx
->ac
.f32_0
;
2529 radv_export_param(ctx
, param_count
, values
, 0x1);
2531 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
++;
2534 outinfo
->pos_exports
= num_pos_exports
;
2535 outinfo
->param_exports
= param_count
;
2539 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2540 struct radv_es_output_info
*outinfo
)
2543 uint64_t max_output_written
= 0;
2544 LLVMValueRef lds_base
= NULL
;
2546 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2550 if (!(ctx
->output_mask
& (1ull << i
)))
2553 if (i
== VARYING_SLOT_CLIP_DIST0
)
2554 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2556 param_index
= shader_io_get_unique_index(i
);
2558 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
2561 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
2563 if (ctx
->ac
.chip_class
>= GFX9
) {
2564 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2565 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2566 LLVMValueRef wave_idx
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
2567 LLVMConstInt(ctx
->ac
.i32
, 24, false),
2568 LLVMConstInt(ctx
->ac
.i32
, 4, false), false);
2569 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2570 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2571 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
2572 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2573 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2576 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2577 LLVMValueRef dw_addr
= NULL
;
2578 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2582 if (!(ctx
->output_mask
& (1ull << i
)))
2585 if (i
== VARYING_SLOT_CLIP_DIST0
)
2586 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2588 param_index
= shader_io_get_unique_index(i
);
2591 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2592 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2595 for (j
= 0; j
< length
; j
++) {
2596 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2597 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2599 if (ctx
->ac
.chip_class
>= GFX9
) {
2600 ac_lds_store(&ctx
->ac
, dw_addr
,
2601 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
2602 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2604 ac_build_buffer_store_dword(&ctx
->ac
,
2607 NULL
, ctx
->es2gs_offset
,
2608 (4 * param_index
+ j
) * 4,
2616 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2618 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2619 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->shader_info
->info
.vs
.ls_outputs_written
);
2620 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2621 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2622 vertex_dw_stride
, "");
2624 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2625 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2628 if (!(ctx
->output_mask
& (1ull << i
)))
2631 if (i
== VARYING_SLOT_CLIP_DIST0
)
2632 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2633 int param
= shader_io_get_unique_index(i
);
2634 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2635 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2637 for (unsigned j
= 0; j
< length
; j
++) {
2638 ac_lds_store(&ctx
->ac
, dw_addr
,
2639 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
2640 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2646 write_tess_factors(struct radv_shader_context
*ctx
)
2648 unsigned stride
, outer_comps
, inner_comps
;
2649 struct ac_build_if_state if_ctx
, inner_if_ctx
;
2650 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2651 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
2652 unsigned tess_inner_index
= 0, tess_outer_index
;
2653 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
2654 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2656 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
2658 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
2678 ac_nir_build_if(&if_ctx
, ctx
,
2679 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2680 invocation_id
, ctx
->ac
.i32_0
, ""));
2682 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
2685 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2686 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2687 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
2690 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2691 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2692 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
2694 for (i
= 0; i
< 4; i
++) {
2695 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2696 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2700 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
2701 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
2702 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2704 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
2706 for (i
= 0; i
< outer_comps
; i
++) {
2708 ac_lds_load(&ctx
->ac
, lds_outer
);
2709 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2712 for (i
= 0; i
< inner_comps
; i
++) {
2713 inner
[i
] = out
[outer_comps
+i
] =
2714 ac_lds_load(&ctx
->ac
, lds_inner
);
2715 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
2720 /* Convert the outputs to vectors for stores. */
2721 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
2725 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
2728 buffer
= ctx
->hs_ring_tess_factor
;
2729 tf_base
= ctx
->tess_factor_offset
;
2730 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
2731 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
2732 unsigned tf_offset
= 0;
2734 if (ctx
->options
->chip_class
<= VI
) {
2735 ac_nir_build_if(&inner_if_ctx
, ctx
,
2736 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2737 rel_patch_id
, ctx
->ac
.i32_0
, ""));
2739 /* Store the dynamic HS control word. */
2740 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
2741 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
2742 1, ctx
->ac
.i32_0
, tf_base
,
2743 0, 1, 0, true, false);
2746 ac_nir_build_endif(&inner_if_ctx
);
2749 /* Store the tessellation factors. */
2750 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
2751 MIN2(stride
, 4), byteoffset
, tf_base
,
2752 tf_offset
, 1, 0, true, false);
2754 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
2755 stride
- 4, byteoffset
, tf_base
,
2756 16 + tf_offset
, 1, 0, true, false);
2758 //store to offchip for TES to read - only if TES reads them
2759 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
2760 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
2761 LLVMValueRef tf_inner_offset
;
2762 unsigned param_outer
, param_inner
;
2764 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2765 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2766 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
2768 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
2769 util_next_power_of_two(outer_comps
));
2771 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
2772 outer_comps
, tf_outer_offset
,
2773 ctx
->oc_lds
, 0, 1, 0, true, false);
2775 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2776 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2777 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
2779 inner_vec
= inner_comps
== 1 ? inner
[0] :
2780 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
2781 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
2782 inner_comps
, tf_inner_offset
,
2783 ctx
->oc_lds
, 0, 1, 0, true, false);
2786 ac_nir_build_endif(&if_ctx
);
2790 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
2792 write_tess_factors(ctx
);
2796 si_export_mrt_color(struct radv_shader_context
*ctx
,
2797 LLVMValueRef
*color
, unsigned index
,
2798 struct ac_export_args
*args
)
2801 si_llvm_init_export_args(ctx
, color
, 0xf,
2802 V_008DFC_SQ_EXP_MRT
+ index
, args
);
2803 if (!args
->enabled_channels
)
2804 return false; /* unnecessary NULL export */
2810 radv_export_mrt_z(struct radv_shader_context
*ctx
,
2811 LLVMValueRef depth
, LLVMValueRef stencil
,
2812 LLVMValueRef samplemask
)
2814 struct ac_export_args args
;
2816 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
2818 ac_build_export(&ctx
->ac
, &args
);
2822 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
2825 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2826 struct ac_export_args color_args
[8];
2828 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2829 LLVMValueRef values
[4];
2831 if (!(ctx
->output_mask
& (1ull << i
)))
2834 if (i
< FRAG_RESULT_DATA0
)
2837 for (unsigned j
= 0; j
< 4; j
++)
2838 values
[j
] = ac_to_float(&ctx
->ac
,
2839 radv_load_output(ctx
, i
, j
));
2841 bool ret
= si_export_mrt_color(ctx
, values
,
2842 i
- FRAG_RESULT_DATA0
,
2843 &color_args
[index
]);
2848 /* Process depth, stencil, samplemask. */
2849 if (ctx
->shader_info
->info
.ps
.writes_z
) {
2850 depth
= ac_to_float(&ctx
->ac
,
2851 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
2853 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
2854 stencil
= ac_to_float(&ctx
->ac
,
2855 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
2857 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
2858 samplemask
= ac_to_float(&ctx
->ac
,
2859 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
2862 /* Set the DONE bit on last non-null color export only if Z isn't
2866 !ctx
->shader_info
->info
.ps
.writes_z
&&
2867 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
2868 !ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
2869 unsigned last
= index
- 1;
2871 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
2872 color_args
[last
].done
= 1; /* DONE bit */
2875 /* Export PS outputs. */
2876 for (unsigned i
= 0; i
< index
; i
++)
2877 ac_build_export(&ctx
->ac
, &color_args
[i
]);
2879 if (depth
|| stencil
|| samplemask
)
2880 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
2882 ac_build_export_null(&ctx
->ac
);
2886 emit_gs_epilogue(struct radv_shader_context
*ctx
)
2888 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
2892 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
2893 LLVMValueRef
*addrs
)
2895 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
2897 switch (ctx
->stage
) {
2898 case MESA_SHADER_VERTEX
:
2899 if (ctx
->options
->key
.vs
.as_ls
)
2900 handle_ls_outputs_post(ctx
);
2901 else if (ctx
->options
->key
.vs
.as_es
)
2902 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
2904 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
2905 ctx
->options
->key
.vs
.export_layer_id
,
2906 &ctx
->shader_info
->vs
.outinfo
);
2908 case MESA_SHADER_FRAGMENT
:
2909 handle_fs_outputs_post(ctx
);
2911 case MESA_SHADER_GEOMETRY
:
2912 emit_gs_epilogue(ctx
);
2914 case MESA_SHADER_TESS_CTRL
:
2915 handle_tcs_outputs_post(ctx
);
2917 case MESA_SHADER_TESS_EVAL
:
2918 if (ctx
->options
->key
.tes
.as_es
)
2919 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
2921 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
2922 ctx
->options
->key
.tes
.export_layer_id
,
2923 &ctx
->shader_info
->tes
.outinfo
);
2930 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
)
2932 LLVMPassManagerRef passmgr
;
2933 /* Create the pass manager */
2934 passmgr
= LLVMCreateFunctionPassManagerForModule(
2937 /* This pass should eliminate all the load and store instructions */
2938 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
2940 /* Add some optimization passes */
2941 LLVMAddScalarReplAggregatesPass(passmgr
);
2942 LLVMAddLICMPass(passmgr
);
2943 LLVMAddAggressiveDCEPass(passmgr
);
2944 LLVMAddCFGSimplificationPass(passmgr
);
2945 LLVMAddInstructionCombiningPass(passmgr
);
2948 LLVMInitializeFunctionPassManager(passmgr
);
2949 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
2950 LLVMFinalizeFunctionPassManager(passmgr
);
2952 LLVMDisposeBuilder(ctx
->ac
.builder
);
2953 LLVMDisposePassManager(passmgr
);
2955 ac_llvm_context_dispose(&ctx
->ac
);
2959 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
2961 struct radv_vs_output_info
*outinfo
;
2963 switch (ctx
->stage
) {
2964 case MESA_SHADER_FRAGMENT
:
2965 case MESA_SHADER_COMPUTE
:
2966 case MESA_SHADER_TESS_CTRL
:
2967 case MESA_SHADER_GEOMETRY
:
2969 case MESA_SHADER_VERTEX
:
2970 if (ctx
->options
->key
.vs
.as_ls
||
2971 ctx
->options
->key
.vs
.as_es
)
2973 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
2975 case MESA_SHADER_TESS_EVAL
:
2976 if (ctx
->options
->key
.vs
.as_es
)
2978 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
2981 unreachable("Unhandled shader type");
2984 ac_optimize_vs_outputs(&ctx
->ac
,
2986 outinfo
->vs_output_param_offset
,
2988 &outinfo
->param_exports
);
2992 ac_setup_rings(struct radv_shader_context
*ctx
)
2994 if ((ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_es
) ||
2995 (ctx
->stage
== MESA_SHADER_TESS_EVAL
&& ctx
->options
->key
.tes
.as_es
)) {
2996 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_VS
, false));
2999 if (ctx
->is_gs_copy_shader
) {
3000 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_VS
, false));
3002 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3004 uint32_t num_entries
= 64;
3005 LLVMValueRef gsvs_ring_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->max_gsvs_emit_size
, false);
3006 LLVMValueRef gsvs_ring_desc
= LLVMConstInt(ctx
->ac
.i32
, ctx
->max_gsvs_emit_size
<< 16, false);
3007 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_GS
, false));
3008 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_GS
, false));
3010 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.v4i32
, "");
3012 tmp
= LLVMConstInt(ctx
->ac
.i32
, num_entries
, false);
3013 if (ctx
->options
->chip_class
>= VI
)
3014 tmp
= LLVMBuildMul(ctx
->ac
.builder
, gsvs_ring_stride
, tmp
, "");
3015 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3016 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.i32_1
, "");
3017 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
, gsvs_ring_desc
, "");
3018 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, tmp
, ctx
->ac
.i32_1
, "");
3021 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3022 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3023 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3024 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3029 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
3030 const struct nir_shader
*nir
)
3032 switch (nir
->info
.stage
) {
3033 case MESA_SHADER_TESS_CTRL
:
3034 return chip_class
>= CIK
? 128 : 64;
3035 case MESA_SHADER_GEOMETRY
:
3036 return chip_class
>= GFX9
? 128 : 64;
3037 case MESA_SHADER_COMPUTE
:
3043 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
3044 nir
->info
.cs
.local_size
[1] *
3045 nir
->info
.cs
.local_size
[2];
3046 return max_workgroup_size
;
3049 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3050 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3052 LLVMValueRef count
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
3053 LLVMConstInt(ctx
->ac
.i32
, 8, false),
3054 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
3055 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3057 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
3058 ctx
->vs_prim_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.vertex_id
, ctx
->vs_prim_id
, "");
3059 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
3060 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
3063 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
)
3065 for(int i
= 5; i
>= 0; --i
) {
3066 ctx
->gs_vtx_offset
[i
] = ac_build_bfe(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
3067 LLVMConstInt(ctx
->ac
.i32
, (i
& 1) * 16, false),
3068 LLVMConstInt(ctx
->ac
.i32
, 16, false), false);
3071 ctx
->gs_wave_id
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
3072 LLVMConstInt(ctx
->ac
.i32
, 16, false),
3073 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
3078 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
3079 struct nir_shader
*const *shaders
,
3081 struct radv_shader_variant_info
*shader_info
,
3082 const struct radv_nir_compiler_options
*options
)
3084 struct radv_shader_context ctx
= {0};
3086 ctx
.options
= options
;
3087 ctx
.shader_info
= shader_info
;
3088 ctx
.context
= LLVMContextCreate();
3090 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
3092 ctx
.ac
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
3093 LLVMSetTarget(ctx
.ac
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
3095 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
3096 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
3097 LLVMSetDataLayout(ctx
.ac
.module
, data_layout_str
);
3098 LLVMDisposeTargetData(data_layout
);
3099 LLVMDisposeMessage(data_layout_str
);
3101 enum ac_float_mode float_mode
=
3102 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3103 AC_FLOAT_MODE_DEFAULT
;
3105 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3107 memset(shader_info
, 0, sizeof(*shader_info
));
3109 for(int i
= 0; i
< shader_count
; ++i
)
3110 radv_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
3112 for (i
= 0; i
< RADV_UD_MAX_SETS
; i
++)
3113 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
3114 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
3115 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
3117 ctx
.max_workgroup_size
= 0;
3118 for (int i
= 0; i
< shader_count
; ++i
) {
3119 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
3120 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
3124 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
3125 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
3127 ctx
.abi
.inputs
= &ctx
.inputs
[0];
3128 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
3129 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
3130 ctx
.abi
.load_ubo
= radv_load_ubo
;
3131 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
3132 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
3133 ctx
.abi
.load_resource
= radv_load_resource
;
3134 ctx
.abi
.clamp_shadow_reference
= false;
3135 ctx
.abi
.gfx9_stride_size_workaround
= ctx
.ac
.chip_class
== GFX9
;
3137 if (shader_count
>= 2)
3138 ac_init_exec_full_mask(&ctx
.ac
);
3140 if (ctx
.ac
.chip_class
== GFX9
&&
3141 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
3142 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
3144 for(int i
= 0; i
< shader_count
; ++i
) {
3145 ctx
.stage
= shaders
[i
]->info
.stage
;
3146 ctx
.output_mask
= 0;
3147 ctx
.num_output_clips
= shaders
[i
]->info
.clip_distance_array_size
;
3148 ctx
.num_output_culls
= shaders
[i
]->info
.cull_distance_array_size
;
3150 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3151 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "gs_next_vertex");
3152 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
3153 ctx
.abi
.load_inputs
= load_gs_input
;
3154 ctx
.abi
.emit_primitive
= visit_end_primitive
;
3155 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3156 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
3157 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
3158 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
3159 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3160 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
3161 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3162 if (shader_count
== 1)
3163 ctx
.tcs_num_inputs
= ctx
.options
->key
.tcs
.num_inputs
;
3165 ctx
.tcs_num_inputs
= util_last_bit64(shader_info
->info
.vs
.ls_outputs_written
);
3166 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
3167 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3168 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
3169 ctx
.abi
.load_tess_varyings
= load_tes_input
;
3170 ctx
.abi
.load_tess_coord
= load_tess_coord
;
3171 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3172 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3173 ctx
.tcs_num_patches
= ctx
.options
->key
.tes
.num_patches
;
3174 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
3175 if (shader_info
->info
.vs
.needs_instance_id
) {
3176 if (ctx
.options
->key
.vs
.as_ls
) {
3177 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3178 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3180 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3181 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3184 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
3185 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
3186 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
3187 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
3188 ctx
.abi
.load_sample_position
= load_sample_position
;
3189 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
3190 ctx
.abi
.emit_kill
= radv_emit_kill
;
3194 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
3196 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
3197 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
3199 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3200 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
3201 shaders
[i
]->info
.cull_distance_array_size
> 4;
3202 ctx
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
3203 ctx
.max_gsvs_emit_size
= ctx
.gsvs_vertex_size
*
3204 shaders
[i
]->info
.gs
.vertices_out
;
3207 ac_setup_rings(&ctx
);
3209 LLVMBasicBlockRef merge_block
;
3210 if (shader_count
>= 2) {
3211 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
3212 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3213 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3215 LLVMValueRef count
= ac_build_bfe(&ctx
.ac
, ctx
.merged_wave_info
,
3216 LLVMConstInt(ctx
.ac
.i32
, 8 * i
, false),
3217 LLVMConstInt(ctx
.ac
.i32
, 8, false), false);
3218 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
3219 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
3220 thread_id
, count
, "");
3221 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
3223 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
3226 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
3227 handle_fs_inputs(&ctx
, shaders
[i
]);
3228 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
3229 handle_vs_inputs(&ctx
, shaders
[i
]);
3230 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
3231 prepare_gs_input_vgprs(&ctx
);
3233 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
]);
3235 if (shader_count
>= 2) {
3236 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
3237 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
3240 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3241 shader_info
->gs
.gsvs_vertex_size
= ctx
.gsvs_vertex_size
;
3242 shader_info
->gs
.max_gsvs_emit_size
= ctx
.max_gsvs_emit_size
;
3243 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3244 shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
3245 shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
3249 LLVMBuildRetVoid(ctx
.ac
.builder
);
3251 if (options
->dump_preoptir
)
3252 ac_dump_module(ctx
.ac
.module
);
3254 ac_llvm_finalize_module(&ctx
);
3256 if (shader_count
== 1)
3257 ac_nir_eliminate_const_vs_outputs(&ctx
);
3259 if (options
->dump_shader
) {
3260 ctx
.shader_info
->private_mem_vgprs
=
3261 ac_count_scratch_private_memory(ctx
.main_function
);
3264 return ctx
.ac
.module
;
3267 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
3269 unsigned *retval
= (unsigned *)context
;
3270 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
3271 char *description
= LLVMGetDiagInfoDescription(di
);
3273 if (severity
== LLVMDSError
) {
3275 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
3279 LLVMDisposeMessage(description
);
3282 static unsigned ac_llvm_compile(LLVMModuleRef M
,
3283 struct ac_shader_binary
*binary
,
3284 LLVMTargetMachineRef tm
)
3286 unsigned retval
= 0;
3288 LLVMContextRef llvm_ctx
;
3289 LLVMMemoryBufferRef out_buffer
;
3290 unsigned buffer_size
;
3291 const char *buffer_data
;
3294 /* Setup Diagnostic Handler*/
3295 llvm_ctx
= LLVMGetModuleContext(M
);
3297 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
3301 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
3304 /* Process Errors/Warnings */
3306 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
3312 /* Extract Shader Code*/
3313 buffer_size
= LLVMGetBufferSize(out_buffer
);
3314 buffer_data
= LLVMGetBufferStart(out_buffer
);
3316 ac_elf_read(buffer_data
, buffer_size
, binary
);
3319 LLVMDisposeMemoryBuffer(out_buffer
);
3325 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
3326 LLVMModuleRef llvm_module
,
3327 struct ac_shader_binary
*binary
,
3328 struct ac_shader_config
*config
,
3329 struct radv_shader_variant_info
*shader_info
,
3330 gl_shader_stage stage
,
3331 const struct radv_nir_compiler_options
*options
)
3333 if (options
->dump_shader
)
3334 ac_dump_module(llvm_module
);
3336 memset(binary
, 0, sizeof(*binary
));
3338 if (options
->record_llvm_ir
) {
3339 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
3340 binary
->llvm_ir_string
= strdup(llvm_ir
);
3341 LLVMDisposeMessage(llvm_ir
);
3344 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
3346 fprintf(stderr
, "compile failed\n");
3349 if (options
->dump_shader
)
3350 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
3352 ac_shader_binary_read_config(binary
, config
, 0, options
->supports_spill
);
3354 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
3355 LLVMDisposeModule(llvm_module
);
3356 LLVMContextDispose(ctx
);
3358 if (stage
== MESA_SHADER_FRAGMENT
) {
3359 shader_info
->num_input_vgprs
= 0;
3360 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
3361 shader_info
->num_input_vgprs
+= 2;
3362 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
3363 shader_info
->num_input_vgprs
+= 2;
3364 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
3365 shader_info
->num_input_vgprs
+= 2;
3366 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
3367 shader_info
->num_input_vgprs
+= 3;
3368 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
3369 shader_info
->num_input_vgprs
+= 2;
3370 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
3371 shader_info
->num_input_vgprs
+= 2;
3372 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
3373 shader_info
->num_input_vgprs
+= 2;
3374 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
3375 shader_info
->num_input_vgprs
+= 1;
3376 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
3377 shader_info
->num_input_vgprs
+= 1;
3378 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
3379 shader_info
->num_input_vgprs
+= 1;
3380 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
3381 shader_info
->num_input_vgprs
+= 1;
3382 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
3383 shader_info
->num_input_vgprs
+= 1;
3384 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
3385 shader_info
->num_input_vgprs
+= 1;
3386 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
3387 shader_info
->num_input_vgprs
+= 1;
3388 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
3389 shader_info
->num_input_vgprs
+= 1;
3390 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
3391 shader_info
->num_input_vgprs
+= 1;
3393 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
3395 /* +3 for scratch wave offset and VCC */
3396 config
->num_sgprs
= MAX2(config
->num_sgprs
,
3397 shader_info
->num_input_sgprs
+ 3);
3399 /* Enable 64-bit and 16-bit denormals, because there is no performance
3402 * If denormals are enabled, all floating-point output modifiers are
3405 * Don't enable denormals for 32-bit floats, because:
3406 * - Floating-point output modifiers would be ignored by the hw.
3407 * - Some opcodes don't support denormals, such as v_mad_f32. We would
3408 * have to stop using those.
3409 * - SI & CI would be very slow.
3411 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
3415 ac_fill_shader_info(struct radv_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct radv_nir_compiler_options
*options
)
3417 switch (nir
->info
.stage
) {
3418 case MESA_SHADER_COMPUTE
:
3419 for (int i
= 0; i
< 3; ++i
)
3420 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
3422 case MESA_SHADER_FRAGMENT
:
3423 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
3425 case MESA_SHADER_GEOMETRY
:
3426 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
3427 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
3428 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
3429 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
3431 case MESA_SHADER_TESS_EVAL
:
3432 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
3433 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
3434 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
3435 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
3436 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
3438 case MESA_SHADER_TESS_CTRL
:
3439 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
3441 case MESA_SHADER_VERTEX
:
3442 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
3443 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
3444 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
3445 if (options
->key
.vs
.as_ls
)
3446 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
3454 radv_compile_nir_shader(LLVMTargetMachineRef tm
,
3455 struct ac_shader_binary
*binary
,
3456 struct ac_shader_config
*config
,
3457 struct radv_shader_variant_info
*shader_info
,
3458 struct nir_shader
*const *nir
,
3460 const struct radv_nir_compiler_options
*options
)
3463 LLVMModuleRef llvm_module
;
3465 llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, nir_count
, shader_info
,
3468 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
,
3469 nir
[0]->info
.stage
, options
);
3471 for (int i
= 0; i
< nir_count
; ++i
)
3472 ac_fill_shader_info(shader_info
, nir
[i
], options
);
3474 /* Determine the ES type (VS or TES) for the GS on GFX9. */
3475 if (options
->chip_class
== GFX9
) {
3476 if (nir_count
== 2 &&
3477 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3478 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
3484 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
3486 LLVMValueRef vtx_offset
=
3487 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
3488 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3491 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3495 if (!(ctx
->output_mask
& (1ull << i
)))
3498 if (i
== VARYING_SLOT_CLIP_DIST0
) {
3499 /* unpack clip and cull from a single set of slots */
3500 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
3505 for (unsigned j
= 0; j
< length
; j
++) {
3506 LLVMValueRef value
, soffset
;
3508 soffset
= LLVMConstInt(ctx
->ac
.i32
,
3510 ctx
->gs_max_out_vertices
* 16 * 4, false);
3512 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->gsvs_ring
,
3514 vtx_offset
, soffset
,
3515 0, 1, 1, true, false);
3517 LLVMBuildStore(ctx
->ac
.builder
,
3518 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3522 handle_vs_outputs_post(ctx
, false, false, &ctx
->shader_info
->vs
.outinfo
);
3526 radv_compile_gs_copy_shader(LLVMTargetMachineRef tm
,
3527 struct nir_shader
*geom_shader
,
3528 struct ac_shader_binary
*binary
,
3529 struct ac_shader_config
*config
,
3530 struct radv_shader_variant_info
*shader_info
,
3531 const struct radv_nir_compiler_options
*options
)
3533 struct radv_shader_context ctx
= {0};
3534 ctx
.context
= LLVMContextCreate();
3535 ctx
.options
= options
;
3536 ctx
.shader_info
= shader_info
;
3538 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
3540 ctx
.ac
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
3542 ctx
.is_gs_copy_shader
= true;
3543 LLVMSetTarget(ctx
.ac
.module
, "amdgcn--");
3545 enum ac_float_mode float_mode
=
3546 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3547 AC_FLOAT_MODE_DEFAULT
;
3549 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3550 ctx
.stage
= MESA_SHADER_VERTEX
;
3552 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
3554 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
3555 ac_setup_rings(&ctx
);
3557 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
3558 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
3560 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
3561 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
3562 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
3563 variable
, MESA_SHADER_VERTEX
);
3566 ac_gs_copy_shader_emit(&ctx
);
3568 LLVMBuildRetVoid(ctx
.ac
.builder
);
3570 ac_llvm_finalize_module(&ctx
);
3572 ac_compile_llvm_module(tm
, ctx
.ac
.module
, binary
, config
, shader_info
,
3573 MESA_SHADER_VERTEX
, options
);