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
,
483 const struct radv_nir_compiler_options
*options
)
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 (options
->address32_hi
) {
515 ac_llvm_add_target_dep_function_attr(main_function
,
516 "amdgpu-32bit-address-high-bits",
517 options
->address32_hi
);
520 if (max_workgroup_size
) {
521 ac_llvm_add_target_dep_function_attr(main_function
,
522 "amdgpu-max-work-group-size",
525 if (options
->unsafe_math
) {
526 /* These were copied from some LLVM test. */
527 LLVMAddTargetDependentFunctionAttr(main_function
,
528 "less-precise-fpmad",
530 LLVMAddTargetDependentFunctionAttr(main_function
,
533 LLVMAddTargetDependentFunctionAttr(main_function
,
536 LLVMAddTargetDependentFunctionAttr(main_function
,
539 LLVMAddTargetDependentFunctionAttr(main_function
,
540 "no-signed-zeros-fp-math",
543 return main_function
;
548 set_loc(struct radv_userdata_info
*ud_info
, uint8_t *sgpr_idx
, uint8_t num_sgprs
,
549 uint32_t indirect_offset
)
551 ud_info
->sgpr_idx
= *sgpr_idx
;
552 ud_info
->num_sgprs
= num_sgprs
;
553 ud_info
->indirect
= indirect_offset
> 0;
554 ud_info
->indirect_offset
= indirect_offset
;
555 *sgpr_idx
+= num_sgprs
;
559 set_loc_shader(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
562 struct radv_userdata_info
*ud_info
=
563 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
566 set_loc(ud_info
, sgpr_idx
, num_sgprs
, 0);
570 set_loc_shader_ptr(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
572 bool use_32bit_pointers
= HAVE_32BIT_POINTERS
&&
573 idx
!= AC_UD_SCRATCH_RING_OFFSETS
;
575 set_loc_shader(ctx
, idx
, sgpr_idx
, use_32bit_pointers
? 1 : 2);
579 set_loc_desc(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
580 uint32_t indirect_offset
)
582 struct radv_userdata_info
*ud_info
=
583 &ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[idx
];
586 set_loc(ud_info
, sgpr_idx
, HAVE_32BIT_POINTERS
? 1 : 2, indirect_offset
);
589 struct user_sgpr_info
{
590 bool need_ring_offsets
;
591 bool indirect_all_descriptor_sets
;
594 static bool needs_view_index_sgpr(struct radv_shader_context
*ctx
,
595 gl_shader_stage stage
)
598 case MESA_SHADER_VERTEX
:
599 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
600 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
603 case MESA_SHADER_TESS_EVAL
:
604 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
607 case MESA_SHADER_GEOMETRY
:
608 case MESA_SHADER_TESS_CTRL
:
609 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
619 count_vs_user_sgprs(struct radv_shader_context
*ctx
)
623 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
)
624 count
+= HAVE_32BIT_POINTERS
? 1 : 2;
625 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
630 static void allocate_user_sgprs(struct radv_shader_context
*ctx
,
631 gl_shader_stage stage
,
632 bool has_previous_stage
,
633 gl_shader_stage previous_stage
,
634 bool needs_view_index
,
635 struct user_sgpr_info
*user_sgpr_info
)
637 uint8_t user_sgpr_count
= 0;
639 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
641 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
642 if (stage
== MESA_SHADER_GEOMETRY
||
643 stage
== MESA_SHADER_VERTEX
||
644 stage
== MESA_SHADER_TESS_CTRL
||
645 stage
== MESA_SHADER_TESS_EVAL
||
646 ctx
->is_gs_copy_shader
)
647 user_sgpr_info
->need_ring_offsets
= true;
649 if (stage
== MESA_SHADER_FRAGMENT
&&
650 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
651 user_sgpr_info
->need_ring_offsets
= true;
653 /* 2 user sgprs will nearly always be allocated for scratch/rings */
654 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
655 user_sgpr_count
+= 2;
659 case MESA_SHADER_COMPUTE
:
660 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
661 user_sgpr_count
+= 3;
663 case MESA_SHADER_FRAGMENT
:
664 user_sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
666 case MESA_SHADER_VERTEX
:
667 if (!ctx
->is_gs_copy_shader
)
668 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
670 case MESA_SHADER_TESS_CTRL
:
671 if (has_previous_stage
) {
672 if (previous_stage
== MESA_SHADER_VERTEX
)
673 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
676 case MESA_SHADER_TESS_EVAL
:
678 case MESA_SHADER_GEOMETRY
:
679 if (has_previous_stage
) {
680 if (previous_stage
== MESA_SHADER_VERTEX
) {
681 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
689 if (needs_view_index
)
692 if (ctx
->shader_info
->info
.loads_push_constants
)
693 user_sgpr_count
+= HAVE_32BIT_POINTERS
? 1 : 2;
695 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
? 32 : 16;
696 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_count
;
697 uint32_t num_desc_set
=
698 util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
);
700 if (remaining_sgprs
/ (HAVE_32BIT_POINTERS
? 1 : 2) < num_desc_set
) {
701 user_sgpr_info
->indirect_all_descriptor_sets
= true;
706 declare_global_input_sgprs(struct radv_shader_context
*ctx
,
707 gl_shader_stage stage
,
708 bool has_previous_stage
,
709 gl_shader_stage previous_stage
,
710 const struct user_sgpr_info
*user_sgpr_info
,
711 struct arg_info
*args
,
712 LLVMValueRef
*desc_sets
)
714 LLVMTypeRef type
= ac_array_in_const32_addr_space(ctx
->ac
.i8
);
715 unsigned num_sets
= ctx
->options
->layout
?
716 ctx
->options
->layout
->num_sets
: 0;
717 unsigned stage_mask
= 1 << stage
;
719 if (has_previous_stage
)
720 stage_mask
|= 1 << previous_stage
;
722 /* 1 for each descriptor set */
723 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
724 for (unsigned i
= 0; i
< num_sets
; ++i
) {
725 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
726 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
727 add_array_arg(args
, type
,
728 &ctx
->descriptor_sets
[i
]);
732 add_array_arg(args
, ac_array_in_const32_addr_space(type
), desc_sets
);
735 if (ctx
->shader_info
->info
.loads_push_constants
) {
736 /* 1 for push constants and dynamic descriptors */
737 add_array_arg(args
, type
, &ctx
->abi
.push_constants
);
742 declare_vs_specific_input_sgprs(struct radv_shader_context
*ctx
,
743 gl_shader_stage stage
,
744 bool has_previous_stage
,
745 gl_shader_stage previous_stage
,
746 struct arg_info
*args
)
748 if (!ctx
->is_gs_copy_shader
&&
749 (stage
== MESA_SHADER_VERTEX
||
750 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
751 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
752 add_arg(args
, ARG_SGPR
,
753 ac_array_in_const32_addr_space(ctx
->ac
.v4i32
),
754 &ctx
->vertex_buffers
);
756 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
757 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
758 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
759 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
765 declare_vs_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
767 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
768 if (!ctx
->is_gs_copy_shader
) {
769 if (ctx
->options
->key
.vs
.as_ls
) {
770 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
771 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
773 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
774 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
776 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
781 declare_tes_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
783 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
784 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
785 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
786 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
790 set_global_input_locs(struct radv_shader_context
*ctx
, gl_shader_stage stage
,
791 bool has_previous_stage
, gl_shader_stage previous_stage
,
792 const struct user_sgpr_info
*user_sgpr_info
,
793 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
795 unsigned num_sets
= ctx
->options
->layout
?
796 ctx
->options
->layout
->num_sets
: 0;
797 unsigned stage_mask
= 1 << stage
;
799 if (has_previous_stage
)
800 stage_mask
|= 1 << previous_stage
;
802 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
803 for (unsigned i
= 0; i
< num_sets
; ++i
) {
804 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
805 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
806 set_loc_desc(ctx
, i
, user_sgpr_idx
, 0);
808 ctx
->descriptor_sets
[i
] = NULL
;
811 set_loc_shader_ptr(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
814 for (unsigned i
= 0; i
< num_sets
; ++i
) {
815 if ((ctx
->shader_info
->info
.desc_set_used_mask
& (1 << i
)) &&
816 ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
817 set_loc_desc(ctx
, i
, user_sgpr_idx
, i
* 8);
818 ctx
->descriptor_sets
[i
] =
819 ac_build_load_to_sgpr(&ctx
->ac
,
821 LLVMConstInt(ctx
->ac
.i32
, i
, false));
824 ctx
->descriptor_sets
[i
] = NULL
;
826 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
829 if (ctx
->shader_info
->info
.loads_push_constants
) {
830 set_loc_shader_ptr(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
);
835 set_vs_specific_input_locs(struct radv_shader_context
*ctx
,
836 gl_shader_stage stage
, bool has_previous_stage
,
837 gl_shader_stage previous_stage
,
838 uint8_t *user_sgpr_idx
)
840 if (!ctx
->is_gs_copy_shader
&&
841 (stage
== MESA_SHADER_VERTEX
||
842 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
843 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
844 set_loc_shader_ptr(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
849 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
852 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
853 user_sgpr_idx
, vs_num
);
857 static void set_llvm_calling_convention(LLVMValueRef func
,
858 gl_shader_stage stage
)
860 enum radeon_llvm_calling_convention calling_conv
;
863 case MESA_SHADER_VERTEX
:
864 case MESA_SHADER_TESS_EVAL
:
865 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
867 case MESA_SHADER_GEOMETRY
:
868 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
870 case MESA_SHADER_TESS_CTRL
:
871 calling_conv
= RADEON_LLVM_AMDGPU_HS
;
873 case MESA_SHADER_FRAGMENT
:
874 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
876 case MESA_SHADER_COMPUTE
:
877 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
880 unreachable("Unhandle shader type");
883 LLVMSetFunctionCallConv(func
, calling_conv
);
886 static void create_function(struct radv_shader_context
*ctx
,
887 gl_shader_stage stage
,
888 bool has_previous_stage
,
889 gl_shader_stage previous_stage
)
891 uint8_t user_sgpr_idx
;
892 struct user_sgpr_info user_sgpr_info
;
893 struct arg_info args
= {};
894 LLVMValueRef desc_sets
;
895 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
896 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
897 previous_stage
, needs_view_index
, &user_sgpr_info
);
899 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
900 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
905 case MESA_SHADER_COMPUTE
:
906 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
907 previous_stage
, &user_sgpr_info
,
910 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
911 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
912 &ctx
->abi
.num_work_groups
);
915 for (int i
= 0; i
< 3; i
++) {
916 ctx
->abi
.workgroup_ids
[i
] = NULL
;
917 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
918 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
919 &ctx
->abi
.workgroup_ids
[i
]);
923 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
924 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
925 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
926 &ctx
->abi
.local_invocation_ids
);
928 case MESA_SHADER_VERTEX
:
929 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
930 previous_stage
, &user_sgpr_info
,
932 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
933 previous_stage
, &args
);
935 if (needs_view_index
)
936 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
937 &ctx
->abi
.view_index
);
938 if (ctx
->options
->key
.vs
.as_es
)
939 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
942 declare_vs_input_vgprs(ctx
, &args
);
944 case MESA_SHADER_TESS_CTRL
:
945 if (has_previous_stage
) {
946 // First 6 system regs
947 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
948 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
949 &ctx
->merged_wave_info
);
950 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
951 &ctx
->tess_factor_offset
);
953 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
954 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
955 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
957 declare_global_input_sgprs(ctx
, stage
,
960 &user_sgpr_info
, &args
,
962 declare_vs_specific_input_sgprs(ctx
, stage
,
964 previous_stage
, &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_VGPR
, ctx
->ac
.i32
,
971 &ctx
->abi
.tcs_patch_id
);
972 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
973 &ctx
->abi
.tcs_rel_ids
);
975 declare_vs_input_vgprs(ctx
, &args
);
977 declare_global_input_sgprs(ctx
, stage
,
980 &user_sgpr_info
, &args
,
983 if (needs_view_index
)
984 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
985 &ctx
->abi
.view_index
);
987 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
988 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
989 &ctx
->tess_factor_offset
);
990 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
991 &ctx
->abi
.tcs_patch_id
);
992 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
993 &ctx
->abi
.tcs_rel_ids
);
996 case MESA_SHADER_TESS_EVAL
:
997 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
998 previous_stage
, &user_sgpr_info
,
1001 if (needs_view_index
)
1002 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1003 &ctx
->abi
.view_index
);
1005 if (ctx
->options
->key
.tes
.as_es
) {
1006 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1007 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1008 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1009 &ctx
->es2gs_offset
);
1011 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1012 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1014 declare_tes_input_vgprs(ctx
, &args
);
1016 case MESA_SHADER_GEOMETRY
:
1017 if (has_previous_stage
) {
1018 // First 6 system regs
1019 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1020 &ctx
->gs2vs_offset
);
1021 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1022 &ctx
->merged_wave_info
);
1023 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1025 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
1026 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1027 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1029 declare_global_input_sgprs(ctx
, stage
,
1032 &user_sgpr_info
, &args
,
1035 if (previous_stage
!= MESA_SHADER_TESS_EVAL
) {
1036 declare_vs_specific_input_sgprs(ctx
, stage
,
1042 if (needs_view_index
)
1043 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1044 &ctx
->abi
.view_index
);
1046 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1047 &ctx
->gs_vtx_offset
[0]);
1048 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1049 &ctx
->gs_vtx_offset
[2]);
1050 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1051 &ctx
->abi
.gs_prim_id
);
1052 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1053 &ctx
->abi
.gs_invocation_id
);
1054 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1055 &ctx
->gs_vtx_offset
[4]);
1057 if (previous_stage
== MESA_SHADER_VERTEX
) {
1058 declare_vs_input_vgprs(ctx
, &args
);
1060 declare_tes_input_vgprs(ctx
, &args
);
1063 declare_global_input_sgprs(ctx
, stage
,
1066 &user_sgpr_info
, &args
,
1069 if (needs_view_index
)
1070 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1071 &ctx
->abi
.view_index
);
1073 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
1074 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
1075 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1076 &ctx
->gs_vtx_offset
[0]);
1077 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1078 &ctx
->gs_vtx_offset
[1]);
1079 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1080 &ctx
->abi
.gs_prim_id
);
1081 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1082 &ctx
->gs_vtx_offset
[2]);
1083 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1084 &ctx
->gs_vtx_offset
[3]);
1085 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1086 &ctx
->gs_vtx_offset
[4]);
1087 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1088 &ctx
->gs_vtx_offset
[5]);
1089 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1090 &ctx
->abi
.gs_invocation_id
);
1093 case MESA_SHADER_FRAGMENT
:
1094 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
1095 previous_stage
, &user_sgpr_info
,
1098 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
1099 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1100 &ctx
->sample_pos_offset
);
1102 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1103 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1104 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1105 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1106 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1107 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1108 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1109 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1110 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1111 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1112 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1113 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1114 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1115 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1116 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1117 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1118 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1121 unreachable("Shader stage not implemented");
1124 ctx
->main_function
= create_llvm_function(
1125 ctx
->context
, ctx
->ac
.module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1126 ctx
->max_workgroup_size
, ctx
->options
);
1127 set_llvm_calling_convention(ctx
->main_function
, stage
);
1130 ctx
->shader_info
->num_input_vgprs
= 0;
1131 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1133 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1135 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1136 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1138 assign_arguments(ctx
->main_function
, &args
);
1142 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1143 set_loc_shader_ptr(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1145 if (ctx
->options
->supports_spill
) {
1146 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1147 LLVMPointerType(ctx
->ac
.i8
, AC_CONST_ADDR_SPACE
),
1148 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1149 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1150 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1154 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1155 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1156 if (has_previous_stage
)
1159 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1160 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1163 case MESA_SHADER_COMPUTE
:
1164 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1165 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1169 case MESA_SHADER_VERTEX
:
1170 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1171 previous_stage
, &user_sgpr_idx
);
1172 if (ctx
->abi
.view_index
)
1173 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1175 case MESA_SHADER_TESS_CTRL
:
1176 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1177 previous_stage
, &user_sgpr_idx
);
1178 if (ctx
->abi
.view_index
)
1179 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1181 case MESA_SHADER_TESS_EVAL
:
1182 if (ctx
->abi
.view_index
)
1183 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1185 case MESA_SHADER_GEOMETRY
:
1186 if (has_previous_stage
) {
1187 if (previous_stage
== MESA_SHADER_VERTEX
)
1188 set_vs_specific_input_locs(ctx
, stage
,
1193 if (ctx
->abi
.view_index
)
1194 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1196 case MESA_SHADER_FRAGMENT
:
1197 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
1198 set_loc_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
,
1203 unreachable("Shader stage not implemented");
1206 if (stage
== MESA_SHADER_TESS_CTRL
||
1207 (stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_ls
) ||
1208 /* GFX9 has the ESGS ring buffer in LDS. */
1209 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
1210 ac_declare_lds_as_pointer(&ctx
->ac
);
1213 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1218 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
1219 unsigned desc_set
, unsigned binding
)
1221 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1222 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
1223 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
1224 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
1225 unsigned base_offset
= layout
->binding
[binding
].offset
;
1226 LLVMValueRef offset
, stride
;
1228 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
1229 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
1230 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
1231 layout
->binding
[binding
].dynamic_offset_offset
;
1232 desc_ptr
= ctx
->abi
.push_constants
;
1233 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
1234 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1236 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
1238 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
1239 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, stride
, "");
1240 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
1242 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
1243 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
1244 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1250 /* The offchip buffer layout for TCS->TES is
1252 * - attribute 0 of patch 0 vertex 0
1253 * - attribute 0 of patch 0 vertex 1
1254 * - attribute 0 of patch 0 vertex 2
1256 * - attribute 0 of patch 1 vertex 0
1257 * - attribute 0 of patch 1 vertex 1
1259 * - attribute 1 of patch 0 vertex 0
1260 * - attribute 1 of patch 0 vertex 1
1262 * - per patch attribute 0 of patch 0
1263 * - per patch attribute 0 of patch 1
1266 * Note that every attribute has 4 components.
1268 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
1270 uint32_t num_patches
= ctx
->tcs_num_patches
;
1271 uint32_t num_tcs_outputs
;
1272 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
1273 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
1275 num_tcs_outputs
= ctx
->options
->key
.tes
.tcs_num_outputs
;
1277 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
1278 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
1280 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
1283 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
1284 LLVMValueRef vertex_index
)
1286 LLVMValueRef param_stride
;
1288 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
* ctx
->tcs_num_patches
, false);
1290 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
1291 return param_stride
;
1294 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
1295 LLVMValueRef vertex_index
,
1296 LLVMValueRef param_index
)
1298 LLVMValueRef base_addr
;
1299 LLVMValueRef param_stride
, constant16
;
1300 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
1301 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
, false);
1302 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1303 param_stride
= calc_param_stride(ctx
, vertex_index
);
1305 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
1306 vertices_per_patch
, "");
1308 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1311 base_addr
= rel_patch_id
;
1314 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1315 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
1316 param_stride
, ""), "");
1318 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
1320 if (!vertex_index
) {
1321 LLVMValueRef patch_data_offset
= get_non_vertex_index_offset(ctx
);
1323 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1324 patch_data_offset
, "");
1329 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct radv_shader_context
*ctx
,
1331 unsigned const_index
,
1333 LLVMValueRef vertex_index
,
1334 LLVMValueRef indir_index
)
1336 LLVMValueRef param_index
;
1339 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
1342 if (const_index
&& !is_compact
)
1343 param
+= const_index
;
1344 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
1346 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
1350 get_dw_address(struct radv_shader_context
*ctx
,
1351 LLVMValueRef dw_addr
,
1353 unsigned const_index
,
1354 bool compact_const_index
,
1355 LLVMValueRef vertex_index
,
1356 LLVMValueRef stride
,
1357 LLVMValueRef indir_index
)
1362 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1363 LLVMBuildMul(ctx
->ac
.builder
,
1369 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1370 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
1371 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
1372 else if (const_index
&& !compact_const_index
)
1373 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1374 LLVMConstInt(ctx
->ac
.i32
, const_index
* 4, false), "");
1376 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1377 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
1379 if (const_index
&& compact_const_index
)
1380 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1381 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
1386 load_tcs_varyings(struct ac_shader_abi
*abi
,
1388 LLVMValueRef vertex_index
,
1389 LLVMValueRef indir_index
,
1390 unsigned const_index
,
1392 unsigned driver_location
,
1394 unsigned num_components
,
1399 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1400 LLVMValueRef dw_addr
, stride
;
1401 LLVMValueRef value
[4], result
;
1402 unsigned param
= shader_io_get_unique_index(location
);
1405 uint32_t input_vertex_size
= (ctx
->tcs_num_inputs
* 16) / 4;
1406 stride
= LLVMConstInt(ctx
->ac
.i32
, input_vertex_size
, false);
1407 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1410 stride
= get_tcs_out_vertex_stride(ctx
);
1411 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1413 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1418 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1421 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1422 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1423 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1426 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1431 store_tcs_output(struct ac_shader_abi
*abi
,
1432 const nir_variable
*var
,
1433 LLVMValueRef vertex_index
,
1434 LLVMValueRef param_index
,
1435 unsigned const_index
,
1439 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1440 const unsigned location
= var
->data
.location
;
1441 const unsigned component
= var
->data
.location_frac
;
1442 const bool is_patch
= var
->data
.patch
;
1443 const bool is_compact
= var
->data
.compact
;
1444 LLVMValueRef dw_addr
;
1445 LLVMValueRef stride
= NULL
;
1446 LLVMValueRef buf_addr
= NULL
;
1448 bool store_lds
= true;
1451 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
1454 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
1458 param
= shader_io_get_unique_index(location
);
1459 if (location
== VARYING_SLOT_CLIP_DIST0
&&
1460 is_compact
&& const_index
> 3) {
1466 stride
= get_tcs_out_vertex_stride(ctx
);
1467 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1469 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1472 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1474 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
1475 vertex_index
, param_index
);
1477 bool is_tess_factor
= false;
1478 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
1479 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
1480 is_tess_factor
= true;
1482 unsigned base
= is_compact
? const_index
: 0;
1483 for (unsigned chan
= 0; chan
< 8; chan
++) {
1484 if (!(writemask
& (1 << chan
)))
1486 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1488 if (store_lds
|| is_tess_factor
) {
1489 LLVMValueRef dw_addr_chan
=
1490 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1491 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
1492 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
1495 if (!is_tess_factor
&& writemask
!= 0xF)
1496 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
1497 buf_addr
, ctx
->oc_lds
,
1498 4 * (base
+ chan
), 1, 0, true, false);
1501 if (writemask
== 0xF) {
1502 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
1503 buf_addr
, ctx
->oc_lds
,
1504 (base
* 4), 1, 0, true, false);
1509 load_tes_input(struct ac_shader_abi
*abi
,
1511 LLVMValueRef vertex_index
,
1512 LLVMValueRef param_index
,
1513 unsigned const_index
,
1515 unsigned driver_location
,
1517 unsigned num_components
,
1522 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1523 LLVMValueRef buf_addr
;
1524 LLVMValueRef result
;
1525 unsigned param
= shader_io_get_unique_index(location
);
1527 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
1532 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
1533 is_compact
, vertex_index
, param_index
);
1535 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
1536 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
1538 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
1539 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
1540 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
1545 load_gs_input(struct ac_shader_abi
*abi
,
1547 unsigned driver_location
,
1549 unsigned num_components
,
1550 unsigned vertex_index
,
1551 unsigned const_index
,
1554 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1555 LLVMValueRef vtx_offset
;
1556 unsigned param
, vtx_offset_param
;
1557 LLVMValueRef value
[4], result
;
1559 vtx_offset_param
= vertex_index
;
1560 assert(vtx_offset_param
< 6);
1561 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
1562 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1564 param
= shader_io_get_unique_index(location
);
1566 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1567 if (ctx
->ac
.chip_class
>= GFX9
) {
1568 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1569 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1570 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
1571 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1573 LLVMValueRef soffset
=
1574 LLVMConstInt(ctx
->ac
.i32
,
1575 (param
* 4 + i
+ const_index
) * 256,
1578 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
1581 vtx_offset
, soffset
,
1582 0, 1, 0, true, false);
1584 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
],
1588 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1589 result
= ac_to_integer(&ctx
->ac
, result
);
1594 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
1596 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1597 ac_build_kill_if_false(&ctx
->ac
, visible
);
1600 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
1601 enum glsl_interp_mode interp
, unsigned location
)
1603 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1606 case INTERP_MODE_FLAT
:
1609 case INTERP_MODE_SMOOTH
:
1610 case INTERP_MODE_NONE
:
1611 if (location
== INTERP_CENTER
)
1612 return ctx
->persp_center
;
1613 else if (location
== INTERP_CENTROID
)
1614 return ctx
->persp_centroid
;
1615 else if (location
== INTERP_SAMPLE
)
1616 return ctx
->persp_sample
;
1618 case INTERP_MODE_NOPERSPECTIVE
:
1619 if (location
== INTERP_CENTER
)
1620 return ctx
->linear_center
;
1621 else if (location
== INTERP_CENTROID
)
1622 return ctx
->linear_centroid
;
1623 else if (location
== INTERP_SAMPLE
)
1624 return ctx
->linear_sample
;
1630 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
1631 LLVMValueRef sample_id
)
1633 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1635 LLVMValueRef result
;
1636 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
1638 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
1639 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
1641 sample_id
= LLVMBuildAdd(ctx
->ac
.builder
, sample_id
, ctx
->sample_pos_offset
, "");
1642 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
1648 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1650 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1651 uint8_t log2_ps_iter_samples
;
1653 if (ctx
->shader_info
->info
.ps
.force_persample
) {
1654 log2_ps_iter_samples
=
1655 util_logbase2(ctx
->options
->key
.fs
.num_samples
);
1657 log2_ps_iter_samples
= ctx
->options
->key
.fs
.log2_ps_iter_samples
;
1660 /* The bit pattern matches that used by fixed function fragment
1662 static const uint16_t ps_iter_masks
[] = {
1663 0xffff, /* not used */
1669 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
1671 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
1673 LLVMValueRef result
, sample_id
;
1674 sample_id
= ac_unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
1675 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
1676 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
1682 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
1684 LLVMValueRef gs_next_vertex
;
1685 LLVMValueRef can_emit
;
1687 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1689 assert(stream
== 0);
1691 /* Write vertex attribute values to GSVS ring */
1692 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
1693 ctx
->gs_next_vertex
,
1696 /* If this thread has already emitted the declared maximum number of
1697 * vertices, kill it: excessive vertex emissions are not supposed to
1698 * have any effect, and GS threads have no externally observable
1699 * effects other than emitting vertices.
1701 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
1702 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
1703 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1705 /* loop num outputs */
1707 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1708 unsigned output_usage_mask
=
1709 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
1710 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1715 if (!(ctx
->output_mask
& (1ull << i
)))
1718 if (i
== VARYING_SLOT_CLIP_DIST0
) {
1719 /* pack clip and cull into a single set of slots */
1720 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
1723 output_usage_mask
= (1 << length
) - 1;
1726 for (unsigned j
= 0; j
< length
; j
++) {
1727 if (!(output_usage_mask
& (1 << j
)))
1730 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1732 LLVMValueRef voffset
= LLVMConstInt(ctx
->ac
.i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
1733 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1734 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1736 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1738 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
1740 voffset
, ctx
->gs2vs_offset
, 0,
1746 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1748 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
1750 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
1754 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1756 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1757 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1761 load_tess_coord(struct ac_shader_abi
*abi
)
1763 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1765 LLVMValueRef coord
[4] = {
1772 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
1773 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1774 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1776 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1780 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1782 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1783 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
1787 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1789 return abi
->base_vertex
;
1792 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1793 LLVMValueRef buffer_ptr
, bool write
)
1795 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1796 LLVMValueRef result
;
1798 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1800 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1801 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1806 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1808 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1809 LLVMValueRef result
;
1811 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1813 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1814 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1819 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1820 unsigned descriptor_set
,
1821 unsigned base_index
,
1822 unsigned constant_index
,
1824 enum ac_descriptor_type desc_type
,
1825 bool image
, bool write
,
1828 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1829 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1830 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
1831 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1832 unsigned offset
= binding
->offset
;
1833 unsigned stride
= binding
->size
;
1835 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1838 assert(base_index
< layout
->binding_count
);
1840 switch (desc_type
) {
1842 type
= ctx
->ac
.v8i32
;
1846 type
= ctx
->ac
.v8i32
;
1850 case AC_DESC_SAMPLER
:
1851 type
= ctx
->ac
.v4i32
;
1852 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
1857 case AC_DESC_BUFFER
:
1858 type
= ctx
->ac
.v4i32
;
1862 unreachable("invalid desc_type\n");
1865 offset
+= constant_index
* stride
;
1867 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1868 (!index
|| binding
->immutable_samplers_equal
)) {
1869 if (binding
->immutable_samplers_equal
)
1872 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1874 LLVMValueRef constants
[] = {
1875 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1876 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1877 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1878 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1880 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1883 assert(stride
% type_size
== 0);
1886 index
= ctx
->ac
.i32_0
;
1888 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1890 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
1891 list
= LLVMBuildPointerCast(builder
, list
,
1892 ac_array_in_const32_addr_space(type
), "");
1894 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
1897 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1898 * so we may need to fix it up. */
1900 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1901 unsigned adjustment
,
1904 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1907 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1909 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1910 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1912 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1914 /* For the integer-like cases, do a natural sign extension.
1916 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1917 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1920 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1921 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1922 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1923 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1925 /* Convert back to the right type. */
1926 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1928 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1929 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1930 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1931 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1932 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1933 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1940 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1941 struct nir_variable
*variable
)
1943 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
1944 LLVMValueRef t_offset
;
1945 LLVMValueRef t_list
;
1947 LLVMValueRef buffer_index
;
1948 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1949 uint8_t input_usage_mask
=
1950 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
1951 unsigned num_channels
= util_last_bit(input_usage_mask
);
1953 variable
->data
.driver_location
= variable
->data
.location
* 4;
1955 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1956 LLVMValueRef output
[4];
1957 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1959 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1960 uint32_t divisor
= ctx
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
1963 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.instance_id
,
1964 ctx
->abi
.start_instance
, "");
1967 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
1968 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
1971 if (ctx
->options
->key
.vs
.as_ls
) {
1972 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1973 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1975 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
1976 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
1979 buffer_index
= ctx
->ac
.i32_0
;
1982 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
1983 ctx
->abi
.base_vertex
, "");
1984 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_index
, false);
1986 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
1988 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
1991 num_channels
, false, true);
1993 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
1995 for (unsigned chan
= 0; chan
< 4; chan
++) {
1996 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
1997 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
2000 unsigned alpha_adjust
= (ctx
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
2001 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
2003 for (unsigned chan
= 0; chan
< 4; chan
++) {
2004 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] =
2005 ac_to_integer(&ctx
->ac
, output
[chan
]);
2010 static void interp_fs_input(struct radv_shader_context
*ctx
,
2012 LLVMValueRef interp_param
,
2013 LLVMValueRef prim_mask
,
2014 LLVMValueRef result
[4])
2016 LLVMValueRef attr_number
;
2019 bool interp
= interp_param
!= NULL
;
2021 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
2023 /* fs.constant returns the param from the middle vertex, so it's not
2024 * really useful for flat shading. It's meant to be used for custom
2025 * interpolation (but the intrinsic can't fetch from the other two
2028 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
2029 * to do the right thing. The only reason we use fs.constant is that
2030 * fs.interp cannot be used on integers, because they can be equal
2034 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
2037 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2039 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2043 for (chan
= 0; chan
< 4; chan
++) {
2044 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2047 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
2052 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
2053 LLVMConstInt(ctx
->ac
.i32
, 2, false),
2062 handle_fs_input_decl(struct radv_shader_context
*ctx
,
2063 struct nir_variable
*variable
)
2065 int idx
= variable
->data
.location
;
2066 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2067 LLVMValueRef interp
;
2069 variable
->data
.driver_location
= idx
* 4;
2070 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
2072 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
2073 unsigned interp_type
;
2074 if (variable
->data
.sample
)
2075 interp_type
= INTERP_SAMPLE
;
2076 else if (variable
->data
.centroid
)
2077 interp_type
= INTERP_CENTROID
;
2079 interp_type
= INTERP_CENTER
;
2081 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
2085 for (unsigned i
= 0; i
< attrib_count
; ++i
)
2086 ctx
->inputs
[ac_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
2091 handle_vs_inputs(struct radv_shader_context
*ctx
,
2092 struct nir_shader
*nir
) {
2093 nir_foreach_variable(variable
, &nir
->inputs
)
2094 handle_vs_input_decl(ctx
, variable
);
2098 prepare_interp_optimize(struct radv_shader_context
*ctx
,
2099 struct nir_shader
*nir
)
2101 bool uses_center
= false;
2102 bool uses_centroid
= false;
2103 nir_foreach_variable(variable
, &nir
->inputs
) {
2104 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
2105 variable
->data
.sample
)
2108 if (variable
->data
.centroid
)
2109 uses_centroid
= true;
2114 if (uses_center
&& uses_centroid
) {
2115 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
2116 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
2117 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
2122 handle_fs_inputs(struct radv_shader_context
*ctx
,
2123 struct nir_shader
*nir
)
2125 prepare_interp_optimize(ctx
, nir
);
2127 nir_foreach_variable(variable
, &nir
->inputs
)
2128 handle_fs_input_decl(ctx
, variable
);
2132 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
2133 ctx
->shader_info
->info
.needs_multiview_view_index
)
2134 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
2136 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
2137 LLVMValueRef interp_param
;
2138 LLVMValueRef
*inputs
= ctx
->inputs
+ac_llvm_reg_index_soa(i
, 0);
2140 if (!(ctx
->input_mask
& (1ull << i
)))
2143 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
2144 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
2145 interp_param
= *inputs
;
2146 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
2150 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
2152 } else if (i
== VARYING_SLOT_POS
) {
2153 for(int i
= 0; i
< 3; ++i
)
2154 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
2156 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2157 ctx
->abi
.frag_pos
[3]);
2160 ctx
->shader_info
->fs
.num_interp
= index
;
2161 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
2163 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
2164 ctx
->abi
.view_index
= ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2168 scan_shader_output_decl(struct radv_shader_context
*ctx
,
2169 struct nir_variable
*variable
,
2170 struct nir_shader
*shader
,
2171 gl_shader_stage stage
)
2173 int idx
= variable
->data
.location
+ variable
->data
.index
;
2174 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2175 uint64_t mask_attribs
;
2177 variable
->data
.driver_location
= idx
* 4;
2179 /* tess ctrl has it's own load/store paths for outputs */
2180 if (stage
== MESA_SHADER_TESS_CTRL
)
2183 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
2184 if (stage
== MESA_SHADER_VERTEX
||
2185 stage
== MESA_SHADER_TESS_EVAL
||
2186 stage
== MESA_SHADER_GEOMETRY
) {
2187 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2188 int length
= shader
->info
.clip_distance_array_size
+
2189 shader
->info
.cull_distance_array_size
;
2190 if (stage
== MESA_SHADER_VERTEX
) {
2191 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2192 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2194 if (stage
== MESA_SHADER_TESS_EVAL
) {
2195 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2196 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2203 mask_attribs
= 1ull << idx
;
2207 ctx
->output_mask
|= mask_attribs
;
2211 /* Initialize arguments for the shader export intrinsic */
2213 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
2214 LLVMValueRef
*values
,
2215 unsigned enabled_channels
,
2217 struct ac_export_args
*args
)
2219 /* Specify the channels that are enabled. */
2220 args
->enabled_channels
= enabled_channels
;
2222 /* Specify whether the EXEC mask represents the valid mask */
2223 args
->valid_mask
= 0;
2225 /* Specify whether this is the last export */
2228 /* Specify the target we are exporting */
2229 args
->target
= target
;
2231 args
->compr
= false;
2232 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
2233 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
2234 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
2235 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
2237 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
2238 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
2239 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
2240 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
2241 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
2244 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2245 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2246 unsigned bits
, bool hi
) = NULL
;
2248 switch(col_format
) {
2249 case V_028714_SPI_SHADER_ZERO
:
2250 args
->enabled_channels
= 0; /* writemask */
2251 args
->target
= V_008DFC_SQ_EXP_NULL
;
2254 case V_028714_SPI_SHADER_32_R
:
2255 args
->enabled_channels
= 1;
2256 args
->out
[0] = values
[0];
2259 case V_028714_SPI_SHADER_32_GR
:
2260 args
->enabled_channels
= 0x3;
2261 args
->out
[0] = values
[0];
2262 args
->out
[1] = values
[1];
2265 case V_028714_SPI_SHADER_32_AR
:
2266 args
->enabled_channels
= 0x9;
2267 args
->out
[0] = values
[0];
2268 args
->out
[3] = values
[3];
2271 case V_028714_SPI_SHADER_FP16_ABGR
:
2272 args
->enabled_channels
= 0x5;
2273 packf
= ac_build_cvt_pkrtz_f16
;
2276 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2277 args
->enabled_channels
= 0x5;
2278 packf
= ac_build_cvt_pknorm_u16
;
2281 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2282 args
->enabled_channels
= 0x5;
2283 packf
= ac_build_cvt_pknorm_i16
;
2286 case V_028714_SPI_SHADER_UINT16_ABGR
:
2287 args
->enabled_channels
= 0x5;
2288 packi
= ac_build_cvt_pk_u16
;
2291 case V_028714_SPI_SHADER_SINT16_ABGR
:
2292 args
->enabled_channels
= 0x5;
2293 packi
= ac_build_cvt_pk_i16
;
2297 case V_028714_SPI_SHADER_32_ABGR
:
2298 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2302 /* Pack f16 or norm_i16/u16. */
2304 for (chan
= 0; chan
< 2; chan
++) {
2305 LLVMValueRef pack_args
[2] = {
2307 values
[2 * chan
+ 1]
2309 LLVMValueRef packed
;
2311 packed
= packf(&ctx
->ac
, pack_args
);
2312 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2314 args
->compr
= 1; /* COMPR flag */
2319 for (chan
= 0; chan
< 2; chan
++) {
2320 LLVMValueRef pack_args
[2] = {
2321 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2322 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2324 LLVMValueRef packed
;
2326 packed
= packi(&ctx
->ac
, pack_args
,
2327 is_int8
? 8 : is_int10
? 10 : 16,
2329 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2331 args
->compr
= 1; /* COMPR flag */
2336 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2338 for (unsigned i
= 0; i
< 4; ++i
) {
2339 if (!(args
->enabled_channels
& (1 << i
)))
2342 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
2347 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
2348 LLVMValueRef
*values
, unsigned enabled_channels
)
2350 struct ac_export_args args
;
2352 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
2353 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2354 ac_build_export(&ctx
->ac
, &args
);
2358 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
2360 LLVMValueRef output
=
2361 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
2363 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
2367 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2368 bool export_prim_id
, bool export_layer_id
,
2369 struct radv_vs_output_info
*outinfo
)
2371 uint32_t param_count
= 0;
2373 unsigned pos_idx
, num_pos_exports
= 0;
2374 struct ac_export_args args
, pos_args
[4] = {};
2375 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2378 if (ctx
->options
->key
.has_multiview_view_index
) {
2379 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2381 for(unsigned i
= 0; i
< 4; ++i
)
2382 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2383 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2386 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
2387 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2390 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2391 sizeof(outinfo
->vs_output_param_offset
));
2393 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
2394 LLVMValueRef slots
[8];
2397 if (outinfo
->cull_dist_mask
)
2398 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
2400 i
= VARYING_SLOT_CLIP_DIST0
;
2401 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
2402 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2404 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
2405 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
2407 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
2408 target
= V_008DFC_SQ_EXP_POS
+ 3;
2409 si_llvm_init_export_args(ctx
, &slots
[4], 0xf, target
, &args
);
2410 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2411 &args
, sizeof(args
));
2414 target
= V_008DFC_SQ_EXP_POS
+ 2;
2415 si_llvm_init_export_args(ctx
, &slots
[0], 0xf, target
, &args
);
2416 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2417 &args
, sizeof(args
));
2421 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
2422 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
2423 for (unsigned j
= 0; j
< 4; j
++)
2424 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
2426 si_llvm_init_export_args(ctx
, pos_values
, 0xf, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2428 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
2429 outinfo
->writes_pointsize
= true;
2430 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
2433 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
2434 outinfo
->writes_layer
= true;
2435 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
2438 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
2439 outinfo
->writes_viewport_index
= true;
2440 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
2443 if (outinfo
->writes_pointsize
||
2444 outinfo
->writes_layer
||
2445 outinfo
->writes_viewport_index
) {
2446 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
2447 (outinfo
->writes_layer
== true ? 4 : 0));
2448 pos_args
[1].valid_mask
= 0;
2449 pos_args
[1].done
= 0;
2450 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2451 pos_args
[1].compr
= 0;
2452 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2453 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2454 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2455 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2457 if (outinfo
->writes_pointsize
== true)
2458 pos_args
[1].out
[0] = psize_value
;
2459 if (outinfo
->writes_layer
== true)
2460 pos_args
[1].out
[2] = layer_value
;
2461 if (outinfo
->writes_viewport_index
== true) {
2462 if (ctx
->options
->chip_class
>= GFX9
) {
2463 /* GFX9 has the layer in out.z[10:0] and the viewport
2464 * index in out.z[19:16].
2466 LLVMValueRef v
= viewport_index_value
;
2467 v
= ac_to_integer(&ctx
->ac
, v
);
2468 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2469 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2471 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2472 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2474 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2475 pos_args
[1].enabled_channels
|= 1 << 2;
2477 pos_args
[1].out
[3] = viewport_index_value
;
2478 pos_args
[1].enabled_channels
|= 1 << 3;
2482 for (i
= 0; i
< 4; i
++) {
2483 if (pos_args
[i
].out
[0])
2488 for (i
= 0; i
< 4; i
++) {
2489 if (!pos_args
[i
].out
[0])
2492 /* Specify the target we are exporting */
2493 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2494 if (pos_idx
== num_pos_exports
)
2495 pos_args
[i
].done
= 1;
2496 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2499 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2500 LLVMValueRef values
[4];
2501 if (!(ctx
->output_mask
& (1ull << i
)))
2504 if (i
!= VARYING_SLOT_LAYER
&&
2505 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
2506 i
< VARYING_SLOT_VAR0
)
2509 for (unsigned j
= 0; j
< 4; j
++)
2510 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2512 unsigned output_usage_mask
;
2514 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2515 !ctx
->is_gs_copy_shader
) {
2517 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2518 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2520 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2522 assert(ctx
->is_gs_copy_shader
);
2524 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
2527 radv_export_param(ctx
, param_count
, values
, output_usage_mask
);
2529 outinfo
->vs_output_param_offset
[i
] = param_count
++;
2532 if (export_prim_id
) {
2533 LLVMValueRef values
[4];
2535 values
[0] = ctx
->vs_prim_id
;
2536 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
2537 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2538 for (unsigned j
= 1; j
< 4; j
++)
2539 values
[j
] = ctx
->ac
.f32_0
;
2541 radv_export_param(ctx
, param_count
, values
, 0x1);
2543 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
2544 outinfo
->export_prim_id
= true;
2547 if (export_layer_id
&& layer_value
) {
2548 LLVMValueRef values
[4];
2550 values
[0] = layer_value
;
2551 for (unsigned j
= 1; j
< 4; j
++)
2552 values
[j
] = ctx
->ac
.f32_0
;
2554 radv_export_param(ctx
, param_count
, values
, 0x1);
2556 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
++;
2559 outinfo
->pos_exports
= num_pos_exports
;
2560 outinfo
->param_exports
= param_count
;
2564 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2565 struct radv_es_output_info
*outinfo
)
2568 uint64_t max_output_written
= 0;
2569 LLVMValueRef lds_base
= NULL
;
2571 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2575 if (!(ctx
->output_mask
& (1ull << i
)))
2578 if (i
== VARYING_SLOT_CLIP_DIST0
)
2579 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2581 param_index
= shader_io_get_unique_index(i
);
2583 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
2586 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
2588 if (ctx
->ac
.chip_class
>= GFX9
) {
2589 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2590 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2591 LLVMValueRef wave_idx
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
2592 LLVMConstInt(ctx
->ac
.i32
, 24, false),
2593 LLVMConstInt(ctx
->ac
.i32
, 4, false), false);
2594 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2595 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2596 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
2597 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2598 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2601 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2602 LLVMValueRef dw_addr
= NULL
;
2603 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2604 unsigned output_usage_mask
;
2608 if (!(ctx
->output_mask
& (1ull << i
)))
2611 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2613 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2615 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2617 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2620 if (i
== VARYING_SLOT_CLIP_DIST0
) {
2621 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2622 output_usage_mask
= (1 << length
) - 1;
2625 param_index
= shader_io_get_unique_index(i
);
2628 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2629 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2633 for (j
= 0; j
< length
; j
++) {
2634 if (!(output_usage_mask
& (1 << j
)))
2637 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2638 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2640 if (ctx
->ac
.chip_class
>= GFX9
) {
2641 LLVMValueRef dw_addr_offset
=
2642 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2643 LLVMConstInt(ctx
->ac
.i32
,
2646 ac_lds_store(&ctx
->ac
, dw_addr_offset
,
2647 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
2649 ac_build_buffer_store_dword(&ctx
->ac
,
2652 NULL
, ctx
->es2gs_offset
,
2653 (4 * param_index
+ j
) * 4,
2661 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2663 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2664 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->shader_info
->info
.vs
.ls_outputs_written
);
2665 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2666 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2667 vertex_dw_stride
, "");
2669 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2670 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2673 if (!(ctx
->output_mask
& (1ull << i
)))
2676 if (i
== VARYING_SLOT_CLIP_DIST0
)
2677 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
2678 int param
= shader_io_get_unique_index(i
);
2679 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2680 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2682 for (unsigned j
= 0; j
< length
; j
++) {
2683 ac_lds_store(&ctx
->ac
, dw_addr
,
2684 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
2685 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2691 write_tess_factors(struct radv_shader_context
*ctx
)
2693 unsigned stride
, outer_comps
, inner_comps
;
2694 struct ac_build_if_state if_ctx
, inner_if_ctx
;
2695 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2696 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
2697 unsigned tess_inner_index
= 0, tess_outer_index
;
2698 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
2699 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2701 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
2703 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
2723 ac_nir_build_if(&if_ctx
, ctx
,
2724 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2725 invocation_id
, ctx
->ac
.i32_0
, ""));
2727 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
2730 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2731 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2732 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
2735 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2736 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2737 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
2739 for (i
= 0; i
< 4; i
++) {
2740 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2741 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
2745 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
2746 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
2747 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2749 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
2751 for (i
= 0; i
< outer_comps
; i
++) {
2753 ac_lds_load(&ctx
->ac
, lds_outer
);
2754 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
2757 for (i
= 0; i
< inner_comps
; i
++) {
2758 inner
[i
] = out
[outer_comps
+i
] =
2759 ac_lds_load(&ctx
->ac
, lds_inner
);
2760 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
2765 /* Convert the outputs to vectors for stores. */
2766 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
2770 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
2773 buffer
= ctx
->hs_ring_tess_factor
;
2774 tf_base
= ctx
->tess_factor_offset
;
2775 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
2776 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
2777 unsigned tf_offset
= 0;
2779 if (ctx
->options
->chip_class
<= VI
) {
2780 ac_nir_build_if(&inner_if_ctx
, ctx
,
2781 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2782 rel_patch_id
, ctx
->ac
.i32_0
, ""));
2784 /* Store the dynamic HS control word. */
2785 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
2786 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
2787 1, ctx
->ac
.i32_0
, tf_base
,
2788 0, 1, 0, true, false);
2791 ac_nir_build_endif(&inner_if_ctx
);
2794 /* Store the tessellation factors. */
2795 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
2796 MIN2(stride
, 4), byteoffset
, tf_base
,
2797 tf_offset
, 1, 0, true, false);
2799 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
2800 stride
- 4, byteoffset
, tf_base
,
2801 16 + tf_offset
, 1, 0, true, false);
2803 //store to offchip for TES to read - only if TES reads them
2804 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
2805 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
2806 LLVMValueRef tf_inner_offset
;
2807 unsigned param_outer
, param_inner
;
2809 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2810 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2811 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
2813 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
2814 util_next_power_of_two(outer_comps
));
2816 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
2817 outer_comps
, tf_outer_offset
,
2818 ctx
->oc_lds
, 0, 1, 0, true, false);
2820 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2821 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
2822 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
2824 inner_vec
= inner_comps
== 1 ? inner
[0] :
2825 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
2826 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
2827 inner_comps
, tf_inner_offset
,
2828 ctx
->oc_lds
, 0, 1, 0, true, false);
2831 ac_nir_build_endif(&if_ctx
);
2835 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
2837 write_tess_factors(ctx
);
2841 si_export_mrt_color(struct radv_shader_context
*ctx
,
2842 LLVMValueRef
*color
, unsigned index
,
2843 struct ac_export_args
*args
)
2846 si_llvm_init_export_args(ctx
, color
, 0xf,
2847 V_008DFC_SQ_EXP_MRT
+ index
, args
);
2848 if (!args
->enabled_channels
)
2849 return false; /* unnecessary NULL export */
2855 radv_export_mrt_z(struct radv_shader_context
*ctx
,
2856 LLVMValueRef depth
, LLVMValueRef stencil
,
2857 LLVMValueRef samplemask
)
2859 struct ac_export_args args
;
2861 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
2863 ac_build_export(&ctx
->ac
, &args
);
2867 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
2870 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2871 struct ac_export_args color_args
[8];
2873 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2874 LLVMValueRef values
[4];
2876 if (!(ctx
->output_mask
& (1ull << i
)))
2879 if (i
< FRAG_RESULT_DATA0
)
2882 for (unsigned j
= 0; j
< 4; j
++)
2883 values
[j
] = ac_to_float(&ctx
->ac
,
2884 radv_load_output(ctx
, i
, j
));
2886 bool ret
= si_export_mrt_color(ctx
, values
,
2887 i
- FRAG_RESULT_DATA0
,
2888 &color_args
[index
]);
2893 /* Process depth, stencil, samplemask. */
2894 if (ctx
->shader_info
->info
.ps
.writes_z
) {
2895 depth
= ac_to_float(&ctx
->ac
,
2896 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
2898 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
2899 stencil
= ac_to_float(&ctx
->ac
,
2900 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
2902 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
2903 samplemask
= ac_to_float(&ctx
->ac
,
2904 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
2907 /* Set the DONE bit on last non-null color export only if Z isn't
2911 !ctx
->shader_info
->info
.ps
.writes_z
&&
2912 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
2913 !ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
2914 unsigned last
= index
- 1;
2916 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
2917 color_args
[last
].done
= 1; /* DONE bit */
2920 /* Export PS outputs. */
2921 for (unsigned i
= 0; i
< index
; i
++)
2922 ac_build_export(&ctx
->ac
, &color_args
[i
]);
2924 if (depth
|| stencil
|| samplemask
)
2925 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
2927 ac_build_export_null(&ctx
->ac
);
2931 emit_gs_epilogue(struct radv_shader_context
*ctx
)
2933 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
2937 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
2938 LLVMValueRef
*addrs
)
2940 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
2942 switch (ctx
->stage
) {
2943 case MESA_SHADER_VERTEX
:
2944 if (ctx
->options
->key
.vs
.as_ls
)
2945 handle_ls_outputs_post(ctx
);
2946 else if (ctx
->options
->key
.vs
.as_es
)
2947 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
2949 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
2950 ctx
->options
->key
.vs
.export_layer_id
,
2951 &ctx
->shader_info
->vs
.outinfo
);
2953 case MESA_SHADER_FRAGMENT
:
2954 handle_fs_outputs_post(ctx
);
2956 case MESA_SHADER_GEOMETRY
:
2957 emit_gs_epilogue(ctx
);
2959 case MESA_SHADER_TESS_CTRL
:
2960 handle_tcs_outputs_post(ctx
);
2962 case MESA_SHADER_TESS_EVAL
:
2963 if (ctx
->options
->key
.tes
.as_es
)
2964 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
2966 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
2967 ctx
->options
->key
.tes
.export_layer_id
,
2968 &ctx
->shader_info
->tes
.outinfo
);
2975 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
,
2976 const struct radv_nir_compiler_options
*options
)
2978 LLVMPassManagerRef passmgr
;
2979 /* Create the pass manager */
2980 passmgr
= LLVMCreateFunctionPassManagerForModule(
2983 if (options
->check_ir
)
2984 LLVMAddVerifierPass(passmgr
);
2986 /* This pass should eliminate all the load and store instructions */
2987 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
2989 /* Add some optimization passes */
2990 LLVMAddScalarReplAggregatesPass(passmgr
);
2991 LLVMAddLICMPass(passmgr
);
2992 LLVMAddAggressiveDCEPass(passmgr
);
2993 LLVMAddCFGSimplificationPass(passmgr
);
2994 /* This is recommended by the instruction combining pass. */
2995 LLVMAddEarlyCSEMemSSAPass(passmgr
);
2996 LLVMAddInstructionCombiningPass(passmgr
);
2999 LLVMInitializeFunctionPassManager(passmgr
);
3000 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
3001 LLVMFinalizeFunctionPassManager(passmgr
);
3003 LLVMDisposeBuilder(ctx
->ac
.builder
);
3004 LLVMDisposePassManager(passmgr
);
3006 ac_llvm_context_dispose(&ctx
->ac
);
3010 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
3012 struct radv_vs_output_info
*outinfo
;
3014 switch (ctx
->stage
) {
3015 case MESA_SHADER_FRAGMENT
:
3016 case MESA_SHADER_COMPUTE
:
3017 case MESA_SHADER_TESS_CTRL
:
3018 case MESA_SHADER_GEOMETRY
:
3020 case MESA_SHADER_VERTEX
:
3021 if (ctx
->options
->key
.vs
.as_ls
||
3022 ctx
->options
->key
.vs
.as_es
)
3024 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
3026 case MESA_SHADER_TESS_EVAL
:
3027 if (ctx
->options
->key
.vs
.as_es
)
3029 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
3032 unreachable("Unhandled shader type");
3035 ac_optimize_vs_outputs(&ctx
->ac
,
3037 outinfo
->vs_output_param_offset
,
3039 &outinfo
->param_exports
);
3043 ac_setup_rings(struct radv_shader_context
*ctx
)
3045 if (ctx
->options
->chip_class
<= VI
&&
3046 (ctx
->stage
== MESA_SHADER_GEOMETRY
||
3047 ctx
->options
->key
.vs
.as_es
|| ctx
->options
->key
.tes
.as_es
)) {
3048 unsigned ring
= ctx
->stage
== MESA_SHADER_GEOMETRY
? RING_ESGS_GS
3050 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, ring
, false);
3052 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
,
3057 if (ctx
->is_gs_copy_shader
) {
3058 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_VS
, false));
3060 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3062 uint32_t num_entries
= 64;
3063 LLVMValueRef gsvs_ring_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->max_gsvs_emit_size
, false);
3064 LLVMValueRef gsvs_ring_desc
= LLVMConstInt(ctx
->ac
.i32
, ctx
->max_gsvs_emit_size
<< 16, false);
3065 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_GS
, false));
3067 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.v4i32
, "");
3069 tmp
= LLVMConstInt(ctx
->ac
.i32
, num_entries
, false);
3070 if (ctx
->options
->chip_class
>= VI
)
3071 tmp
= LLVMBuildMul(ctx
->ac
.builder
, gsvs_ring_stride
, tmp
, "");
3072 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3073 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.i32_1
, "");
3074 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
, gsvs_ring_desc
, "");
3075 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, tmp
, ctx
->ac
.i32_1
, "");
3078 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3079 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3080 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3081 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3086 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
3087 const struct nir_shader
*nir
)
3089 switch (nir
->info
.stage
) {
3090 case MESA_SHADER_TESS_CTRL
:
3091 return chip_class
>= CIK
? 128 : 64;
3092 case MESA_SHADER_GEOMETRY
:
3093 return chip_class
>= GFX9
? 128 : 64;
3094 case MESA_SHADER_COMPUTE
:
3100 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
3101 nir
->info
.cs
.local_size
[1] *
3102 nir
->info
.cs
.local_size
[2];
3103 return max_workgroup_size
;
3106 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3107 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3109 LLVMValueRef count
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
3110 LLVMConstInt(ctx
->ac
.i32
, 8, false),
3111 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
3112 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3114 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
3115 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
3116 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
3119 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
)
3121 for(int i
= 5; i
>= 0; --i
) {
3122 ctx
->gs_vtx_offset
[i
] = ac_build_bfe(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
3123 LLVMConstInt(ctx
->ac
.i32
, (i
& 1) * 16, false),
3124 LLVMConstInt(ctx
->ac
.i32
, 16, false), false);
3127 ctx
->gs_wave_id
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
3128 LLVMConstInt(ctx
->ac
.i32
, 16, false),
3129 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
3134 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
3135 struct nir_shader
*const *shaders
,
3137 struct radv_shader_variant_info
*shader_info
,
3138 const struct radv_nir_compiler_options
*options
)
3140 struct radv_shader_context ctx
= {0};
3142 ctx
.options
= options
;
3143 ctx
.shader_info
= shader_info
;
3144 ctx
.context
= LLVMContextCreate();
3146 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
3148 ctx
.ac
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
3149 LLVMSetTarget(ctx
.ac
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
3151 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
3152 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
3153 LLVMSetDataLayout(ctx
.ac
.module
, data_layout_str
);
3154 LLVMDisposeTargetData(data_layout
);
3155 LLVMDisposeMessage(data_layout_str
);
3157 enum ac_float_mode float_mode
=
3158 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3159 AC_FLOAT_MODE_DEFAULT
;
3161 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3163 memset(shader_info
, 0, sizeof(*shader_info
));
3165 for(int i
= 0; i
< shader_count
; ++i
)
3166 radv_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
3168 for (i
= 0; i
< RADV_UD_MAX_SETS
; i
++)
3169 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
3170 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
3171 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
3173 ctx
.max_workgroup_size
= 0;
3174 for (int i
= 0; i
< shader_count
; ++i
) {
3175 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
3176 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
3180 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
3181 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
3183 ctx
.abi
.inputs
= &ctx
.inputs
[0];
3184 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
3185 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
3186 ctx
.abi
.load_ubo
= radv_load_ubo
;
3187 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
3188 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
3189 ctx
.abi
.load_resource
= radv_load_resource
;
3190 ctx
.abi
.clamp_shadow_reference
= false;
3191 ctx
.abi
.gfx9_stride_size_workaround
= ctx
.ac
.chip_class
== GFX9
;
3193 if (shader_count
>= 2)
3194 ac_init_exec_full_mask(&ctx
.ac
);
3196 if (ctx
.ac
.chip_class
== GFX9
&&
3197 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
3198 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
3200 for(int i
= 0; i
< shader_count
; ++i
) {
3201 ctx
.stage
= shaders
[i
]->info
.stage
;
3202 ctx
.output_mask
= 0;
3203 ctx
.num_output_clips
= shaders
[i
]->info
.clip_distance_array_size
;
3204 ctx
.num_output_culls
= shaders
[i
]->info
.cull_distance_array_size
;
3206 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3207 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "gs_next_vertex");
3208 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
3209 ctx
.abi
.load_inputs
= load_gs_input
;
3210 ctx
.abi
.emit_primitive
= visit_end_primitive
;
3211 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3212 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
3213 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
3214 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
3215 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3216 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
3217 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3218 if (shader_count
== 1)
3219 ctx
.tcs_num_inputs
= ctx
.options
->key
.tcs
.num_inputs
;
3221 ctx
.tcs_num_inputs
= util_last_bit64(shader_info
->info
.vs
.ls_outputs_written
);
3222 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
3223 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3224 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
3225 ctx
.abi
.load_tess_varyings
= load_tes_input
;
3226 ctx
.abi
.load_tess_coord
= load_tess_coord
;
3227 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3228 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3229 ctx
.tcs_num_patches
= ctx
.options
->key
.tes
.num_patches
;
3230 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
3231 if (shader_info
->info
.vs
.needs_instance_id
) {
3232 if (ctx
.options
->key
.vs
.as_ls
) {
3233 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3234 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3236 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3237 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3240 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
3241 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
3242 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
3243 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
3244 ctx
.abi
.load_sample_position
= load_sample_position
;
3245 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
3246 ctx
.abi
.emit_kill
= radv_emit_kill
;
3250 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
3252 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
3253 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
3255 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3256 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
3257 shaders
[i
]->info
.cull_distance_array_size
> 4;
3258 ctx
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
3259 ctx
.max_gsvs_emit_size
= ctx
.gsvs_vertex_size
*
3260 shaders
[i
]->info
.gs
.vertices_out
;
3263 ac_setup_rings(&ctx
);
3265 LLVMBasicBlockRef merge_block
;
3266 if (shader_count
>= 2) {
3267 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
3268 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3269 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3271 LLVMValueRef count
= ac_build_bfe(&ctx
.ac
, ctx
.merged_wave_info
,
3272 LLVMConstInt(ctx
.ac
.i32
, 8 * i
, false),
3273 LLVMConstInt(ctx
.ac
.i32
, 8, false), false);
3274 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
3275 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
3276 thread_id
, count
, "");
3277 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
3279 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
3282 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
3283 handle_fs_inputs(&ctx
, shaders
[i
]);
3284 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
3285 handle_vs_inputs(&ctx
, shaders
[i
]);
3286 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
3287 prepare_gs_input_vgprs(&ctx
);
3289 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
]);
3291 if (shader_count
>= 2) {
3292 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
3293 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
3296 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3297 shader_info
->gs
.gsvs_vertex_size
= ctx
.gsvs_vertex_size
;
3298 shader_info
->gs
.max_gsvs_emit_size
= ctx
.max_gsvs_emit_size
;
3299 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3300 shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
3301 shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
3305 LLVMBuildRetVoid(ctx
.ac
.builder
);
3307 if (options
->dump_preoptir
)
3308 ac_dump_module(ctx
.ac
.module
);
3310 ac_llvm_finalize_module(&ctx
, options
);
3312 if (shader_count
== 1)
3313 ac_nir_eliminate_const_vs_outputs(&ctx
);
3315 if (options
->dump_shader
) {
3316 ctx
.shader_info
->private_mem_vgprs
=
3317 ac_count_scratch_private_memory(ctx
.main_function
);
3320 return ctx
.ac
.module
;
3323 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
3325 unsigned *retval
= (unsigned *)context
;
3326 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
3327 char *description
= LLVMGetDiagInfoDescription(di
);
3329 if (severity
== LLVMDSError
) {
3331 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
3335 LLVMDisposeMessage(description
);
3338 static unsigned ac_llvm_compile(LLVMModuleRef M
,
3339 struct ac_shader_binary
*binary
,
3340 LLVMTargetMachineRef tm
)
3342 unsigned retval
= 0;
3344 LLVMContextRef llvm_ctx
;
3345 LLVMMemoryBufferRef out_buffer
;
3346 unsigned buffer_size
;
3347 const char *buffer_data
;
3350 /* Setup Diagnostic Handler*/
3351 llvm_ctx
= LLVMGetModuleContext(M
);
3353 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
3357 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
3360 /* Process Errors/Warnings */
3362 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
3368 /* Extract Shader Code*/
3369 buffer_size
= LLVMGetBufferSize(out_buffer
);
3370 buffer_data
= LLVMGetBufferStart(out_buffer
);
3372 ac_elf_read(buffer_data
, buffer_size
, binary
);
3375 LLVMDisposeMemoryBuffer(out_buffer
);
3381 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
3382 LLVMModuleRef llvm_module
,
3383 struct ac_shader_binary
*binary
,
3384 struct ac_shader_config
*config
,
3385 struct radv_shader_variant_info
*shader_info
,
3386 gl_shader_stage stage
,
3387 const struct radv_nir_compiler_options
*options
)
3389 if (options
->dump_shader
)
3390 ac_dump_module(llvm_module
);
3392 memset(binary
, 0, sizeof(*binary
));
3394 if (options
->record_llvm_ir
) {
3395 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
3396 binary
->llvm_ir_string
= strdup(llvm_ir
);
3397 LLVMDisposeMessage(llvm_ir
);
3400 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
3402 fprintf(stderr
, "compile failed\n");
3405 if (options
->dump_shader
)
3406 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
3408 ac_shader_binary_read_config(binary
, config
, 0, options
->supports_spill
);
3410 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
3411 LLVMDisposeModule(llvm_module
);
3412 LLVMContextDispose(ctx
);
3414 if (stage
== MESA_SHADER_FRAGMENT
) {
3415 shader_info
->num_input_vgprs
= 0;
3416 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
3417 shader_info
->num_input_vgprs
+= 2;
3418 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
3419 shader_info
->num_input_vgprs
+= 2;
3420 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
3421 shader_info
->num_input_vgprs
+= 2;
3422 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
3423 shader_info
->num_input_vgprs
+= 3;
3424 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
3425 shader_info
->num_input_vgprs
+= 2;
3426 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
3427 shader_info
->num_input_vgprs
+= 2;
3428 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
3429 shader_info
->num_input_vgprs
+= 2;
3430 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
3431 shader_info
->num_input_vgprs
+= 1;
3432 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
3433 shader_info
->num_input_vgprs
+= 1;
3434 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
3435 shader_info
->num_input_vgprs
+= 1;
3436 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
3437 shader_info
->num_input_vgprs
+= 1;
3438 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
3439 shader_info
->num_input_vgprs
+= 1;
3440 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
3441 shader_info
->num_input_vgprs
+= 1;
3442 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
3443 shader_info
->num_input_vgprs
+= 1;
3444 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
3445 shader_info
->num_input_vgprs
+= 1;
3446 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
3447 shader_info
->num_input_vgprs
+= 1;
3449 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
3451 /* +3 for scratch wave offset and VCC */
3452 config
->num_sgprs
= MAX2(config
->num_sgprs
,
3453 shader_info
->num_input_sgprs
+ 3);
3455 /* Enable 64-bit and 16-bit denormals, because there is no performance
3458 * If denormals are enabled, all floating-point output modifiers are
3461 * Don't enable denormals for 32-bit floats, because:
3462 * - Floating-point output modifiers would be ignored by the hw.
3463 * - Some opcodes don't support denormals, such as v_mad_f32. We would
3464 * have to stop using those.
3465 * - SI & CI would be very slow.
3467 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
3471 ac_fill_shader_info(struct radv_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct radv_nir_compiler_options
*options
)
3473 switch (nir
->info
.stage
) {
3474 case MESA_SHADER_COMPUTE
:
3475 for (int i
= 0; i
< 3; ++i
)
3476 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
3478 case MESA_SHADER_FRAGMENT
:
3479 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
3481 case MESA_SHADER_GEOMETRY
:
3482 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
3483 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
3484 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
3485 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
3487 case MESA_SHADER_TESS_EVAL
:
3488 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
3489 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
3490 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
3491 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
3492 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
3494 case MESA_SHADER_TESS_CTRL
:
3495 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
3497 case MESA_SHADER_VERTEX
:
3498 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
3499 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
3500 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
3501 if (options
->key
.vs
.as_ls
)
3502 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
3510 radv_compile_nir_shader(LLVMTargetMachineRef tm
,
3511 struct ac_shader_binary
*binary
,
3512 struct ac_shader_config
*config
,
3513 struct radv_shader_variant_info
*shader_info
,
3514 struct nir_shader
*const *nir
,
3516 const struct radv_nir_compiler_options
*options
)
3519 LLVMModuleRef llvm_module
;
3521 llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, nir_count
, shader_info
,
3524 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
,
3525 nir
[0]->info
.stage
, options
);
3527 for (int i
= 0; i
< nir_count
; ++i
)
3528 ac_fill_shader_info(shader_info
, nir
[i
], options
);
3530 /* Determine the ES type (VS or TES) for the GS on GFX9. */
3531 if (options
->chip_class
== GFX9
) {
3532 if (nir_count
== 2 &&
3533 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3534 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
3540 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
3542 LLVMValueRef vtx_offset
=
3543 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
3544 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3547 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3551 if (!(ctx
->output_mask
& (1ull << i
)))
3554 if (i
== VARYING_SLOT_CLIP_DIST0
) {
3555 /* unpack clip and cull from a single set of slots */
3556 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
3561 for (unsigned j
= 0; j
< length
; j
++) {
3562 LLVMValueRef value
, soffset
;
3564 soffset
= LLVMConstInt(ctx
->ac
.i32
,
3566 ctx
->gs_max_out_vertices
* 16 * 4, false);
3568 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->gsvs_ring
,
3570 vtx_offset
, soffset
,
3571 0, 1, 1, true, false);
3573 LLVMBuildStore(ctx
->ac
.builder
,
3574 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3578 handle_vs_outputs_post(ctx
, false, false, &ctx
->shader_info
->vs
.outinfo
);
3582 radv_compile_gs_copy_shader(LLVMTargetMachineRef tm
,
3583 struct nir_shader
*geom_shader
,
3584 struct ac_shader_binary
*binary
,
3585 struct ac_shader_config
*config
,
3586 struct radv_shader_variant_info
*shader_info
,
3587 const struct radv_nir_compiler_options
*options
)
3589 struct radv_shader_context ctx
= {0};
3590 ctx
.context
= LLVMContextCreate();
3591 ctx
.options
= options
;
3592 ctx
.shader_info
= shader_info
;
3594 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
3596 ctx
.ac
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
3598 ctx
.is_gs_copy_shader
= true;
3599 LLVMSetTarget(ctx
.ac
.module
, "amdgcn--");
3601 enum ac_float_mode float_mode
=
3602 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3603 AC_FLOAT_MODE_DEFAULT
;
3605 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3606 ctx
.stage
= MESA_SHADER_VERTEX
;
3608 radv_nir_shader_info_pass(geom_shader
, options
, &shader_info
->info
);
3610 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
3612 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
3613 ac_setup_rings(&ctx
);
3615 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
3616 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
3618 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
3619 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
3620 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
3621 variable
, MESA_SHADER_VERTEX
);
3624 ac_gs_copy_shader_emit(&ctx
);
3626 LLVMBuildRetVoid(ctx
.ac
.builder
);
3628 ac_llvm_finalize_module(&ctx
, options
);
3630 ac_compile_llvm_module(tm
, ctx
.ac
.module
, binary
, config
, shader_info
,
3631 MESA_SHADER_VERTEX
, options
);