2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "radv_private.h"
29 #include "radv_shader.h"
30 #include "radv_shader_helper.h"
33 #include <llvm-c/Core.h>
34 #include <llvm-c/TargetMachine.h>
35 #include <llvm-c/Transforms/Scalar.h>
36 #include <llvm-c/Transforms/Utils.h>
40 #include "ac_binary.h"
41 #include "ac_llvm_util.h"
42 #include "ac_llvm_build.h"
43 #include "ac_shader_abi.h"
44 #include "ac_shader_util.h"
45 #include "ac_exp_param.h"
47 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
49 struct radv_shader_context
{
50 struct ac_llvm_context ac
;
51 const struct radv_nir_compiler_options
*options
;
52 struct radv_shader_variant_info
*shader_info
;
53 struct ac_shader_abi abi
;
55 unsigned max_workgroup_size
;
56 LLVMContextRef context
;
57 LLVMValueRef main_function
;
59 LLVMValueRef descriptor_sets
[RADV_UD_MAX_SETS
];
60 LLVMValueRef ring_offsets
;
62 LLVMValueRef vertex_buffers
;
63 LLVMValueRef rel_auto_id
;
64 LLVMValueRef vs_prim_id
;
65 LLVMValueRef es2gs_offset
;
68 LLVMValueRef merged_wave_info
;
69 LLVMValueRef tess_factor_offset
;
70 LLVMValueRef tes_rel_patch_id
;
74 LLVMValueRef gs2vs_offset
;
75 LLVMValueRef gs_wave_id
;
76 LLVMValueRef gs_vtx_offset
[6];
78 LLVMValueRef esgs_ring
;
79 LLVMValueRef gsvs_ring
[4];
80 LLVMValueRef hs_ring_tess_offchip
;
81 LLVMValueRef hs_ring_tess_factor
;
83 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
84 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
87 LLVMValueRef streamout_buffers
;
88 LLVMValueRef streamout_write_idx
;
89 LLVMValueRef streamout_config
;
90 LLVMValueRef streamout_offset
[4];
92 gl_shader_stage stage
;
94 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
99 bool is_gs_copy_shader
;
100 LLVMValueRef gs_next_vertex
[4];
101 unsigned gs_max_out_vertices
;
103 unsigned tes_primitive_mode
;
105 uint32_t tcs_patch_outputs_read
;
106 uint64_t tcs_outputs_read
;
107 uint32_t tcs_vertices_per_patch
;
108 uint32_t tcs_num_inputs
;
109 uint32_t tcs_num_patches
;
110 uint32_t max_gsvs_emit_size
;
111 uint32_t gsvs_vertex_size
;
114 enum radeon_llvm_calling_convention
{
115 RADEON_LLVM_AMDGPU_VS
= 87,
116 RADEON_LLVM_AMDGPU_GS
= 88,
117 RADEON_LLVM_AMDGPU_PS
= 89,
118 RADEON_LLVM_AMDGPU_CS
= 90,
119 RADEON_LLVM_AMDGPU_HS
= 93,
122 static inline struct radv_shader_context
*
123 radv_shader_context_from_abi(struct ac_shader_abi
*abi
)
125 struct radv_shader_context
*ctx
= NULL
;
126 return container_of(abi
, ctx
, abi
);
129 struct ac_build_if_state
131 struct radv_shader_context
*ctx
;
132 LLVMValueRef condition
;
133 LLVMBasicBlockRef entry_block
;
134 LLVMBasicBlockRef true_block
;
135 LLVMBasicBlockRef false_block
;
136 LLVMBasicBlockRef merge_block
;
139 static LLVMBasicBlockRef
140 ac_build_insert_new_block(struct radv_shader_context
*ctx
, const char *name
)
142 LLVMBasicBlockRef current_block
;
143 LLVMBasicBlockRef next_block
;
144 LLVMBasicBlockRef new_block
;
146 /* get current basic block */
147 current_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
149 /* chqeck if there's another block after this one */
150 next_block
= LLVMGetNextBasicBlock(current_block
);
152 /* insert the new block before the next block */
153 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
156 /* append new block after current block */
157 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
158 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
164 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
165 struct radv_shader_context
*ctx
,
166 LLVMValueRef condition
)
168 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
170 memset(ifthen
, 0, sizeof *ifthen
);
172 ifthen
->condition
= condition
;
173 ifthen
->entry_block
= block
;
175 /* create endif/merge basic block for the phi functions */
176 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
178 /* create/insert true_block before merge_block */
180 LLVMInsertBasicBlockInContext(ctx
->context
,
184 /* successive code goes into the true block */
185 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ifthen
->true_block
);
192 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
194 LLVMBuilderRef builder
= ifthen
->ctx
->ac
.builder
;
196 /* Insert branch to the merge block from current block */
197 LLVMBuildBr(builder
, ifthen
->merge_block
);
200 * Now patch in the various branch instructions.
203 /* Insert the conditional branch instruction at the end of entry_block */
204 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
205 if (ifthen
->false_block
) {
206 /* we have an else clause */
207 LLVMBuildCondBr(builder
, ifthen
->condition
,
208 ifthen
->true_block
, ifthen
->false_block
);
212 LLVMBuildCondBr(builder
, ifthen
->condition
,
213 ifthen
->true_block
, ifthen
->merge_block
);
216 /* Resume building code at end of the ifthen->merge_block */
217 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
221 static LLVMValueRef
get_rel_patch_id(struct radv_shader_context
*ctx
)
223 switch (ctx
->stage
) {
224 case MESA_SHADER_TESS_CTRL
:
225 return ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
226 case MESA_SHADER_TESS_EVAL
:
227 return ctx
->tes_rel_patch_id
;
230 unreachable("Illegal stage");
235 get_tcs_num_patches(struct radv_shader_context
*ctx
)
237 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
238 unsigned num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
239 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
240 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
241 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
242 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
243 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
244 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
245 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
246 unsigned num_patches
;
247 unsigned hardware_lds_size
;
249 /* Ensure that we only need one wave per SIMD so we don't need to check
250 * resource usage. Also ensures that the number of tcs in and out
251 * vertices per threadgroup are at most 256.
253 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
254 /* Make sure that the data fits in LDS. This assumes the shaders only
255 * use LDS for the inputs and outputs.
257 hardware_lds_size
= 32768;
259 /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
260 * threadgroup, even though there is more than 32 KiB LDS.
262 * Test: dEQP-VK.tessellation.shader_input_output.barrier
264 if (ctx
->options
->chip_class
>= CIK
&& ctx
->options
->family
!= CHIP_STONEY
)
265 hardware_lds_size
= 65536;
267 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
268 /* Make sure the output data fits in the offchip buffer */
269 num_patches
= MIN2(num_patches
, (ctx
->options
->tess_offchip_block_dw_size
* 4) / output_patch_size
);
270 /* Not necessary for correctness, but improves performance. The
271 * specific value is taken from the proprietary driver.
273 num_patches
= MIN2(num_patches
, 40);
275 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
276 if (ctx
->options
->chip_class
== SI
) {
277 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
278 num_patches
= MIN2(num_patches
, one_wave
);
284 calculate_tess_lds_size(struct radv_shader_context
*ctx
)
286 unsigned num_tcs_input_cp
= ctx
->options
->key
.tcs
.input_vertices
;
287 unsigned num_tcs_output_cp
;
288 unsigned num_tcs_outputs
, num_tcs_patch_outputs
;
289 unsigned input_vertex_size
, output_vertex_size
;
290 unsigned input_patch_size
, output_patch_size
;
291 unsigned pervertex_output_patch_size
;
292 unsigned output_patch0_offset
;
293 unsigned num_patches
;
296 num_tcs_output_cp
= ctx
->tcs_vertices_per_patch
;
297 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
298 num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
300 input_vertex_size
= ctx
->tcs_num_inputs
* 16;
301 output_vertex_size
= num_tcs_outputs
* 16;
303 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
305 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
306 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
308 num_patches
= ctx
->tcs_num_patches
;
309 output_patch0_offset
= input_patch_size
* num_patches
;
311 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
315 /* Tessellation shaders pass outputs to the next shader using LDS.
317 * LS outputs = TCS inputs
318 * TCS outputs = TES inputs
321 * - TCS inputs for patch 0
322 * - TCS inputs for patch 1
323 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
325 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
326 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
327 * - TCS outputs for patch 1
328 * - Per-patch TCS outputs for patch 1
329 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
330 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
333 * All three shaders VS(LS), TCS, TES share the same LDS space.
336 get_tcs_in_patch_stride(struct radv_shader_context
*ctx
)
338 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
339 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
340 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
342 input_patch_size
/= 4;
343 return LLVMConstInt(ctx
->ac
.i32
, input_patch_size
, false);
347 get_tcs_out_patch_stride(struct radv_shader_context
*ctx
)
349 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
350 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.patch_outputs_written
);
351 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
352 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
353 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
354 output_patch_size
/= 4;
355 return LLVMConstInt(ctx
->ac
.i32
, output_patch_size
, false);
359 get_tcs_out_vertex_stride(struct radv_shader_context
*ctx
)
361 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
362 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
363 output_vertex_size
/= 4;
364 return LLVMConstInt(ctx
->ac
.i32
, output_vertex_size
, false);
368 get_tcs_out_patch0_offset(struct radv_shader_context
*ctx
)
370 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
371 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
372 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
373 uint32_t output_patch0_offset
= input_patch_size
;
374 unsigned num_patches
= ctx
->tcs_num_patches
;
376 output_patch0_offset
*= num_patches
;
377 output_patch0_offset
/= 4;
378 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
382 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context
*ctx
)
384 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
385 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
386 uint32_t input_patch_size
= ctx
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
387 uint32_t output_patch0_offset
= input_patch_size
;
389 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
390 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
391 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
392 unsigned num_patches
= ctx
->tcs_num_patches
;
394 output_patch0_offset
*= num_patches
;
395 output_patch0_offset
+= pervertex_output_patch_size
;
396 output_patch0_offset
/= 4;
397 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
401 get_tcs_in_current_patch_offset(struct radv_shader_context
*ctx
)
403 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
404 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
406 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
410 get_tcs_out_current_patch_offset(struct radv_shader_context
*ctx
)
412 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
413 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
414 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
416 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
421 get_tcs_out_current_patch_data_offset(struct radv_shader_context
*ctx
)
423 LLVMValueRef patch0_patch_data_offset
=
424 get_tcs_out_patch0_patch_data_offset(ctx
);
425 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
426 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
428 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
429 patch0_patch_data_offset
);
434 LLVMTypeRef types
[MAX_ARGS
];
435 LLVMValueRef
*assign
[MAX_ARGS
];
438 uint8_t num_sgprs_used
;
439 uint8_t num_vgprs_used
;
442 enum ac_arg_regfile
{
448 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
449 LLVMValueRef
*param_ptr
)
451 assert(info
->count
< MAX_ARGS
);
453 info
->assign
[info
->count
] = param_ptr
;
454 info
->types
[info
->count
] = type
;
457 if (regfile
== ARG_SGPR
) {
458 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
461 assert(regfile
== ARG_VGPR
);
462 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
466 static void assign_arguments(LLVMValueRef main_function
,
467 struct arg_info
*info
)
470 for (i
= 0; i
< info
->count
; i
++) {
472 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
477 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
478 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
479 unsigned num_return_elems
,
480 struct arg_info
*args
,
481 unsigned max_workgroup_size
,
482 const struct radv_nir_compiler_options
*options
)
484 LLVMTypeRef main_function_type
, ret_type
;
485 LLVMBasicBlockRef main_function_body
;
487 if (num_return_elems
)
488 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
489 num_return_elems
, true);
491 ret_type
= LLVMVoidTypeInContext(ctx
);
493 /* Setup the function */
495 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
496 LLVMValueRef main_function
=
497 LLVMAddFunction(module
, "main", main_function_type
);
499 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
500 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
502 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
503 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
504 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
506 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
508 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
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
,
551 ud_info
->sgpr_idx
= *sgpr_idx
;
552 ud_info
->num_sgprs
= num_sgprs
;
553 *sgpr_idx
+= num_sgprs
;
557 set_loc_shader(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
560 struct radv_userdata_info
*ud_info
=
561 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
564 set_loc(ud_info
, sgpr_idx
, num_sgprs
);
568 set_loc_shader_ptr(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
570 bool use_32bit_pointers
= idx
!= AC_UD_SCRATCH_RING_OFFSETS
;
572 set_loc_shader(ctx
, idx
, sgpr_idx
, use_32bit_pointers
? 1 : 2);
576 set_loc_desc(struct radv_shader_context
*ctx
, int idx
, uint8_t *sgpr_idx
)
578 struct radv_userdata_locations
*locs
=
579 &ctx
->shader_info
->user_sgprs_locs
;
580 struct radv_userdata_info
*ud_info
= &locs
->descriptor_sets
[idx
];
583 set_loc(ud_info
, sgpr_idx
, 1);
585 locs
->descriptor_sets_enabled
|= 1 << idx
;
588 struct user_sgpr_info
{
589 bool need_ring_offsets
;
590 bool indirect_all_descriptor_sets
;
593 static bool needs_view_index_sgpr(struct radv_shader_context
*ctx
,
594 gl_shader_stage stage
)
597 case MESA_SHADER_VERTEX
:
598 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
599 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
602 case MESA_SHADER_TESS_EVAL
:
603 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
606 case MESA_SHADER_GEOMETRY
:
607 case MESA_SHADER_TESS_CTRL
:
608 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
618 count_vs_user_sgprs(struct radv_shader_context
*ctx
)
622 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
)
624 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
629 static void allocate_user_sgprs(struct radv_shader_context
*ctx
,
630 gl_shader_stage stage
,
631 bool has_previous_stage
,
632 gl_shader_stage previous_stage
,
633 bool needs_view_index
,
634 struct user_sgpr_info
*user_sgpr_info
)
636 uint8_t user_sgpr_count
= 0;
638 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
640 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
641 if (stage
== MESA_SHADER_GEOMETRY
||
642 stage
== MESA_SHADER_VERTEX
||
643 stage
== MESA_SHADER_TESS_CTRL
||
644 stage
== MESA_SHADER_TESS_EVAL
||
645 ctx
->is_gs_copy_shader
)
646 user_sgpr_info
->need_ring_offsets
= true;
648 if (stage
== MESA_SHADER_FRAGMENT
&&
649 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
650 user_sgpr_info
->need_ring_offsets
= true;
652 /* 2 user sgprs will nearly always be allocated for scratch/rings */
653 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
654 user_sgpr_count
+= 2;
658 case MESA_SHADER_COMPUTE
:
659 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
660 user_sgpr_count
+= 3;
662 case MESA_SHADER_FRAGMENT
:
663 user_sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
665 case MESA_SHADER_VERTEX
:
666 if (!ctx
->is_gs_copy_shader
)
667 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
669 case MESA_SHADER_TESS_CTRL
:
670 if (has_previous_stage
) {
671 if (previous_stage
== MESA_SHADER_VERTEX
)
672 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
675 case MESA_SHADER_TESS_EVAL
:
677 case MESA_SHADER_GEOMETRY
:
678 if (has_previous_stage
) {
679 if (previous_stage
== MESA_SHADER_VERTEX
) {
680 user_sgpr_count
+= count_vs_user_sgprs(ctx
);
688 if (needs_view_index
)
691 if (ctx
->shader_info
->info
.loads_push_constants
)
694 if (ctx
->streamout_buffers
)
697 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
&& stage
!= MESA_SHADER_COMPUTE
? 32 : 16;
698 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_count
;
699 uint32_t num_desc_set
=
700 util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
);
702 if (remaining_sgprs
< num_desc_set
) {
703 user_sgpr_info
->indirect_all_descriptor_sets
= true;
708 declare_global_input_sgprs(struct radv_shader_context
*ctx
,
709 const struct user_sgpr_info
*user_sgpr_info
,
710 struct arg_info
*args
,
711 LLVMValueRef
*desc_sets
)
713 LLVMTypeRef type
= ac_array_in_const32_addr_space(ctx
->ac
.i8
);
715 /* 1 for each descriptor set */
716 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
717 uint32_t mask
= ctx
->shader_info
->info
.desc_set_used_mask
;
720 int i
= u_bit_scan(&mask
);
722 add_arg(args
, ARG_SGPR
, type
, &ctx
->descriptor_sets
[i
]);
725 add_arg(args
, ARG_SGPR
, ac_array_in_const32_addr_space(type
),
729 if (ctx
->shader_info
->info
.loads_push_constants
) {
730 /* 1 for push constants and dynamic descriptors */
731 add_arg(args
, ARG_SGPR
, type
, &ctx
->abi
.push_constants
);
734 if (ctx
->shader_info
->info
.so
.num_outputs
) {
735 add_arg(args
, ARG_SGPR
,
736 ac_array_in_const32_addr_space(ctx
->ac
.v4i32
),
737 &ctx
->streamout_buffers
);
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_streamout_sgprs(struct radv_shader_context
*ctx
, gl_shader_stage stage
,
782 struct arg_info
*args
)
786 /* Streamout SGPRs. */
787 if (ctx
->shader_info
->info
.so
.num_outputs
) {
788 assert(stage
== MESA_SHADER_VERTEX
||
789 stage
== MESA_SHADER_TESS_EVAL
);
791 if (stage
!= MESA_SHADER_TESS_EVAL
) {
792 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_config
);
794 args
->assign
[args
->count
- 1] = &ctx
->streamout_config
;
795 args
->types
[args
->count
- 1] = ctx
->ac
.i32
;
798 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_write_idx
);
801 /* A streamout buffer offset is loaded if the stride is non-zero. */
802 for (i
= 0; i
< 4; i
++) {
803 if (!ctx
->shader_info
->info
.so
.strides
[i
])
806 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->streamout_offset
[i
]);
811 declare_tes_input_vgprs(struct radv_shader_context
*ctx
, struct arg_info
*args
)
813 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
814 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
815 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
816 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
820 set_global_input_locs(struct radv_shader_context
*ctx
,
821 const struct user_sgpr_info
*user_sgpr_info
,
822 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
824 uint32_t mask
= ctx
->shader_info
->info
.desc_set_used_mask
;
826 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
828 int i
= u_bit_scan(&mask
);
830 set_loc_desc(ctx
, i
, user_sgpr_idx
);
833 set_loc_shader_ptr(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
837 int i
= u_bit_scan(&mask
);
839 ctx
->descriptor_sets
[i
] =
840 ac_build_load_to_sgpr(&ctx
->ac
, desc_sets
,
841 LLVMConstInt(ctx
->ac
.i32
, i
, false));
845 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
848 if (ctx
->shader_info
->info
.loads_push_constants
) {
849 set_loc_shader_ptr(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
);
852 if (ctx
->streamout_buffers
) {
853 set_loc_shader_ptr(ctx
, AC_UD_STREAMOUT_BUFFERS
,
859 set_vs_specific_input_locs(struct radv_shader_context
*ctx
,
860 gl_shader_stage stage
, bool has_previous_stage
,
861 gl_shader_stage previous_stage
,
862 uint8_t *user_sgpr_idx
)
864 if (!ctx
->is_gs_copy_shader
&&
865 (stage
== MESA_SHADER_VERTEX
||
866 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
867 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
868 set_loc_shader_ptr(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
873 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
876 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
877 user_sgpr_idx
, vs_num
);
881 static void set_llvm_calling_convention(LLVMValueRef func
,
882 gl_shader_stage stage
)
884 enum radeon_llvm_calling_convention calling_conv
;
887 case MESA_SHADER_VERTEX
:
888 case MESA_SHADER_TESS_EVAL
:
889 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
891 case MESA_SHADER_GEOMETRY
:
892 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
894 case MESA_SHADER_TESS_CTRL
:
895 calling_conv
= RADEON_LLVM_AMDGPU_HS
;
897 case MESA_SHADER_FRAGMENT
:
898 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
900 case MESA_SHADER_COMPUTE
:
901 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
904 unreachable("Unhandle shader type");
907 LLVMSetFunctionCallConv(func
, calling_conv
);
910 static void create_function(struct radv_shader_context
*ctx
,
911 gl_shader_stage stage
,
912 bool has_previous_stage
,
913 gl_shader_stage previous_stage
)
915 uint8_t user_sgpr_idx
;
916 struct user_sgpr_info user_sgpr_info
;
917 struct arg_info args
= {};
918 LLVMValueRef desc_sets
;
919 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
920 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
921 previous_stage
, needs_view_index
, &user_sgpr_info
);
923 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
924 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
929 case MESA_SHADER_COMPUTE
:
930 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
933 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
934 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
935 &ctx
->abi
.num_work_groups
);
938 for (int i
= 0; i
< 3; i
++) {
939 ctx
->abi
.workgroup_ids
[i
] = NULL
;
940 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
941 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
942 &ctx
->abi
.workgroup_ids
[i
]);
946 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
947 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
948 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
949 &ctx
->abi
.local_invocation_ids
);
951 case MESA_SHADER_VERTEX
:
952 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
955 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
956 previous_stage
, &args
);
958 if (needs_view_index
)
959 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
960 &ctx
->abi
.view_index
);
961 if (ctx
->options
->key
.vs
.as_es
) {
962 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
964 } else if (ctx
->options
->key
.vs
.as_ls
) {
965 /* no extra parameters */
967 declare_streamout_sgprs(ctx
, stage
, &args
);
970 declare_vs_input_vgprs(ctx
, &args
);
972 case MESA_SHADER_TESS_CTRL
:
973 if (has_previous_stage
) {
974 // First 6 system regs
975 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
976 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
977 &ctx
->merged_wave_info
);
978 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
979 &ctx
->tess_factor_offset
);
981 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
982 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
983 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
985 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
988 declare_vs_specific_input_sgprs(ctx
, stage
,
990 previous_stage
, &args
);
992 if (needs_view_index
)
993 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
994 &ctx
->abi
.view_index
);
996 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
997 &ctx
->abi
.tcs_patch_id
);
998 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
999 &ctx
->abi
.tcs_rel_ids
);
1001 declare_vs_input_vgprs(ctx
, &args
);
1003 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1006 if (needs_view_index
)
1007 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1008 &ctx
->abi
.view_index
);
1010 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1011 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1012 &ctx
->tess_factor_offset
);
1013 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1014 &ctx
->abi
.tcs_patch_id
);
1015 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1016 &ctx
->abi
.tcs_rel_ids
);
1019 case MESA_SHADER_TESS_EVAL
:
1020 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1023 if (needs_view_index
)
1024 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1025 &ctx
->abi
.view_index
);
1027 if (ctx
->options
->key
.tes
.as_es
) {
1028 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1029 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1030 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1031 &ctx
->es2gs_offset
);
1033 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
1034 declare_streamout_sgprs(ctx
, stage
, &args
);
1035 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1037 declare_tes_input_vgprs(ctx
, &args
);
1039 case MESA_SHADER_GEOMETRY
:
1040 if (has_previous_stage
) {
1041 // First 6 system regs
1042 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1043 &ctx
->gs2vs_offset
);
1044 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1045 &ctx
->merged_wave_info
);
1046 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
1048 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
1049 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1050 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
1052 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1055 if (previous_stage
!= MESA_SHADER_TESS_EVAL
) {
1056 declare_vs_specific_input_sgprs(ctx
, stage
,
1062 if (needs_view_index
)
1063 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1064 &ctx
->abi
.view_index
);
1066 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1067 &ctx
->gs_vtx_offset
[0]);
1068 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1069 &ctx
->gs_vtx_offset
[2]);
1070 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1071 &ctx
->abi
.gs_prim_id
);
1072 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1073 &ctx
->abi
.gs_invocation_id
);
1074 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1075 &ctx
->gs_vtx_offset
[4]);
1077 if (previous_stage
== MESA_SHADER_VERTEX
) {
1078 declare_vs_input_vgprs(ctx
, &args
);
1080 declare_tes_input_vgprs(ctx
, &args
);
1083 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1086 if (needs_view_index
)
1087 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1088 &ctx
->abi
.view_index
);
1090 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
1091 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
1092 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1093 &ctx
->gs_vtx_offset
[0]);
1094 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1095 &ctx
->gs_vtx_offset
[1]);
1096 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1097 &ctx
->abi
.gs_prim_id
);
1098 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1099 &ctx
->gs_vtx_offset
[2]);
1100 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1101 &ctx
->gs_vtx_offset
[3]);
1102 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1103 &ctx
->gs_vtx_offset
[4]);
1104 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1105 &ctx
->gs_vtx_offset
[5]);
1106 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
1107 &ctx
->abi
.gs_invocation_id
);
1110 case MESA_SHADER_FRAGMENT
:
1111 declare_global_input_sgprs(ctx
, &user_sgpr_info
, &args
,
1114 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1115 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1116 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1117 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1118 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1119 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1120 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1121 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1122 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1123 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1124 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1125 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1126 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1127 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1128 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1129 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1130 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1133 unreachable("Shader stage not implemented");
1136 ctx
->main_function
= create_llvm_function(
1137 ctx
->context
, ctx
->ac
.module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1138 ctx
->max_workgroup_size
, ctx
->options
);
1139 set_llvm_calling_convention(ctx
->main_function
, stage
);
1142 ctx
->shader_info
->num_input_vgprs
= 0;
1143 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1145 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1147 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1148 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1150 assign_arguments(ctx
->main_function
, &args
);
1154 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1155 set_loc_shader_ptr(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1157 if (ctx
->options
->supports_spill
) {
1158 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1159 LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_CONST
),
1160 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1161 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1162 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1166 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1167 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1168 if (has_previous_stage
)
1171 set_global_input_locs(ctx
, &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1174 case MESA_SHADER_COMPUTE
:
1175 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1176 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1180 case MESA_SHADER_VERTEX
:
1181 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1182 previous_stage
, &user_sgpr_idx
);
1183 if (ctx
->abi
.view_index
)
1184 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1186 case MESA_SHADER_TESS_CTRL
:
1187 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1188 previous_stage
, &user_sgpr_idx
);
1189 if (ctx
->abi
.view_index
)
1190 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1192 case MESA_SHADER_TESS_EVAL
:
1193 if (ctx
->abi
.view_index
)
1194 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1196 case MESA_SHADER_GEOMETRY
:
1197 if (has_previous_stage
) {
1198 if (previous_stage
== MESA_SHADER_VERTEX
)
1199 set_vs_specific_input_locs(ctx
, stage
,
1204 if (ctx
->abi
.view_index
)
1205 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1207 case MESA_SHADER_FRAGMENT
:
1210 unreachable("Shader stage not implemented");
1213 if (stage
== MESA_SHADER_TESS_CTRL
||
1214 (stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_ls
) ||
1215 /* GFX9 has the ESGS ring buffer in LDS. */
1216 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
1217 ac_declare_lds_as_pointer(&ctx
->ac
);
1220 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1225 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
1226 unsigned desc_set
, unsigned binding
)
1228 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1229 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
1230 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
1231 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
1232 unsigned base_offset
= layout
->binding
[binding
].offset
;
1233 LLVMValueRef offset
, stride
;
1235 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
1236 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
1237 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
1238 layout
->binding
[binding
].dynamic_offset_offset
;
1239 desc_ptr
= ctx
->abi
.push_constants
;
1240 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
1241 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1243 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
1245 offset
= ac_build_imad(&ctx
->ac
, index
, stride
,
1246 LLVMConstInt(ctx
->ac
.i32
, base_offset
, false));
1248 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
1249 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
1250 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1256 /* The offchip buffer layout for TCS->TES is
1258 * - attribute 0 of patch 0 vertex 0
1259 * - attribute 0 of patch 0 vertex 1
1260 * - attribute 0 of patch 0 vertex 2
1262 * - attribute 0 of patch 1 vertex 0
1263 * - attribute 0 of patch 1 vertex 1
1265 * - attribute 1 of patch 0 vertex 0
1266 * - attribute 1 of patch 0 vertex 1
1268 * - per patch attribute 0 of patch 0
1269 * - per patch attribute 0 of patch 1
1272 * Note that every attribute has 4 components.
1274 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
1276 uint32_t num_patches
= ctx
->tcs_num_patches
;
1277 uint32_t num_tcs_outputs
;
1278 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
1279 num_tcs_outputs
= util_last_bit64(ctx
->shader_info
->info
.tcs
.outputs_written
);
1281 num_tcs_outputs
= ctx
->options
->key
.tes
.tcs_num_outputs
;
1283 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
1284 uint32_t pervertex_output_patch_size
= ctx
->tcs_vertices_per_patch
* output_vertex_size
;
1286 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
1289 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
1290 LLVMValueRef vertex_index
)
1292 LLVMValueRef param_stride
;
1294 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
* ctx
->tcs_num_patches
, false);
1296 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
1297 return param_stride
;
1300 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
1301 LLVMValueRef vertex_index
,
1302 LLVMValueRef param_index
)
1304 LLVMValueRef base_addr
;
1305 LLVMValueRef param_stride
, constant16
;
1306 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
1307 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_vertices_per_patch
, false);
1308 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
1309 param_stride
= calc_param_stride(ctx
, vertex_index
);
1311 base_addr
= ac_build_imad(&ctx
->ac
, rel_patch_id
,
1312 vertices_per_patch
, vertex_index
);
1314 base_addr
= rel_patch_id
;
1317 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1318 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
1319 param_stride
, ""), "");
1321 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
1323 if (!vertex_index
) {
1324 LLVMValueRef patch_data_offset
= get_non_vertex_index_offset(ctx
);
1326 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
1327 patch_data_offset
, "");
1332 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct radv_shader_context
*ctx
,
1334 unsigned const_index
,
1336 LLVMValueRef vertex_index
,
1337 LLVMValueRef indir_index
)
1339 LLVMValueRef param_index
;
1342 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
1345 if (const_index
&& !is_compact
)
1346 param
+= const_index
;
1347 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
1349 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
1353 get_dw_address(struct radv_shader_context
*ctx
,
1354 LLVMValueRef dw_addr
,
1356 unsigned const_index
,
1357 bool compact_const_index
,
1358 LLVMValueRef vertex_index
,
1359 LLVMValueRef stride
,
1360 LLVMValueRef indir_index
)
1365 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1366 LLVMBuildMul(ctx
->ac
.builder
,
1372 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1373 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
1374 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
1375 else if (const_index
&& !compact_const_index
)
1376 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1377 LLVMConstInt(ctx
->ac
.i32
, const_index
* 4, false), "");
1379 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1380 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
1382 if (const_index
&& compact_const_index
)
1383 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1384 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
1389 load_tcs_varyings(struct ac_shader_abi
*abi
,
1391 LLVMValueRef vertex_index
,
1392 LLVMValueRef indir_index
,
1393 unsigned const_index
,
1395 unsigned driver_location
,
1397 unsigned num_components
,
1402 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1403 LLVMValueRef dw_addr
, stride
;
1404 LLVMValueRef value
[4], result
;
1405 unsigned param
= shader_io_get_unique_index(location
);
1408 uint32_t input_vertex_size
= (ctx
->tcs_num_inputs
* 16) / 4;
1409 stride
= LLVMConstInt(ctx
->ac
.i32
, input_vertex_size
, false);
1410 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1413 stride
= get_tcs_out_vertex_stride(ctx
);
1414 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1416 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1421 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1424 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1425 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1426 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1429 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1434 store_tcs_output(struct ac_shader_abi
*abi
,
1435 const nir_variable
*var
,
1436 LLVMValueRef vertex_index
,
1437 LLVMValueRef param_index
,
1438 unsigned const_index
,
1442 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1443 const unsigned location
= var
->data
.location
;
1444 const unsigned component
= var
->data
.location_frac
;
1445 const bool is_patch
= var
->data
.patch
;
1446 const bool is_compact
= var
->data
.compact
;
1447 LLVMValueRef dw_addr
;
1448 LLVMValueRef stride
= NULL
;
1449 LLVMValueRef buf_addr
= NULL
;
1451 bool store_lds
= true;
1454 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
1457 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
1461 param
= shader_io_get_unique_index(location
);
1462 if (location
== VARYING_SLOT_CLIP_DIST0
&&
1463 is_compact
&& const_index
> 3) {
1469 stride
= get_tcs_out_vertex_stride(ctx
);
1470 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1472 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1475 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
1477 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
1478 vertex_index
, param_index
);
1480 bool is_tess_factor
= false;
1481 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
1482 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
1483 is_tess_factor
= true;
1485 unsigned base
= is_compact
? const_index
: 0;
1486 for (unsigned chan
= 0; chan
< 8; chan
++) {
1487 if (!(writemask
& (1 << chan
)))
1489 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1490 value
= ac_to_integer(&ctx
->ac
, value
);
1491 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
1493 if (store_lds
|| is_tess_factor
) {
1494 LLVMValueRef dw_addr_chan
=
1495 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1496 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
1497 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
1500 if (!is_tess_factor
&& writemask
!= 0xF)
1501 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
1502 buf_addr
, ctx
->oc_lds
,
1503 4 * (base
+ chan
), 1, 0, true, false);
1506 if (writemask
== 0xF) {
1507 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
1508 buf_addr
, ctx
->oc_lds
,
1509 (base
* 4), 1, 0, true, false);
1514 load_tes_input(struct ac_shader_abi
*abi
,
1516 LLVMValueRef vertex_index
,
1517 LLVMValueRef param_index
,
1518 unsigned const_index
,
1520 unsigned driver_location
,
1522 unsigned num_components
,
1527 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1528 LLVMValueRef buf_addr
;
1529 LLVMValueRef result
;
1530 unsigned param
= shader_io_get_unique_index(location
);
1532 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
1537 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
1538 is_compact
, vertex_index
, param_index
);
1540 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
1541 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
1543 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
1544 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
1545 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
1550 load_gs_input(struct ac_shader_abi
*abi
,
1552 unsigned driver_location
,
1554 unsigned num_components
,
1555 unsigned vertex_index
,
1556 unsigned const_index
,
1559 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1560 LLVMValueRef vtx_offset
;
1561 unsigned param
, vtx_offset_param
;
1562 LLVMValueRef value
[4], result
;
1564 vtx_offset_param
= vertex_index
;
1565 assert(vtx_offset_param
< 6);
1566 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
1567 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1569 param
= shader_io_get_unique_index(location
);
1571 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1572 if (ctx
->ac
.chip_class
>= GFX9
) {
1573 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1574 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1575 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
1576 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
1578 LLVMValueRef soffset
=
1579 LLVMConstInt(ctx
->ac
.i32
,
1580 (param
* 4 + i
+ const_index
) * 256,
1583 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
1586 vtx_offset
, soffset
,
1587 0, 1, 0, true, false);
1590 if (ac_get_type_size(type
) == 2) {
1591 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i32
, "");
1592 value
[i
] = LLVMBuildTrunc(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i16
, "");
1594 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], type
, "");
1596 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1597 result
= ac_to_integer(&ctx
->ac
, result
);
1602 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
1604 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1605 ac_build_kill_if_false(&ctx
->ac
, visible
);
1608 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
1609 enum glsl_interp_mode interp
, unsigned location
)
1611 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1614 case INTERP_MODE_FLAT
:
1617 case INTERP_MODE_SMOOTH
:
1618 case INTERP_MODE_NONE
:
1619 if (location
== INTERP_CENTER
)
1620 return ctx
->persp_center
;
1621 else if (location
== INTERP_CENTROID
)
1622 return ctx
->persp_centroid
;
1623 else if (location
== INTERP_SAMPLE
)
1624 return ctx
->persp_sample
;
1626 case INTERP_MODE_NOPERSPECTIVE
:
1627 if (location
== INTERP_CENTER
)
1628 return ctx
->linear_center
;
1629 else if (location
== INTERP_CENTROID
)
1630 return ctx
->linear_centroid
;
1631 else if (location
== INTERP_SAMPLE
)
1632 return ctx
->linear_sample
;
1639 radv_get_sample_pos_offset(uint32_t num_samples
)
1641 uint32_t sample_pos_offset
= 0;
1643 switch (num_samples
) {
1645 sample_pos_offset
= 1;
1648 sample_pos_offset
= 3;
1651 sample_pos_offset
= 7;
1656 return sample_pos_offset
;
1659 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
1660 LLVMValueRef sample_id
)
1662 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1664 LLVMValueRef result
;
1665 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
1667 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
1668 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
1670 uint32_t sample_pos_offset
=
1671 radv_get_sample_pos_offset(ctx
->options
->key
.fs
.num_samples
);
1674 LLVMBuildAdd(ctx
->ac
.builder
, sample_id
,
1675 LLVMConstInt(ctx
->ac
.i32
, sample_pos_offset
, false), "");
1676 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
1682 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1684 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1685 uint8_t log2_ps_iter_samples
;
1687 if (ctx
->shader_info
->info
.ps
.force_persample
) {
1688 log2_ps_iter_samples
=
1689 util_logbase2(ctx
->options
->key
.fs
.num_samples
);
1691 log2_ps_iter_samples
= ctx
->options
->key
.fs
.log2_ps_iter_samples
;
1694 /* The bit pattern matches that used by fixed function fragment
1696 static const uint16_t ps_iter_masks
[] = {
1697 0xffff, /* not used */
1703 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
1705 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
1707 LLVMValueRef result
, sample_id
;
1708 sample_id
= ac_unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
1709 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
1710 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
1716 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
1718 LLVMValueRef gs_next_vertex
;
1719 LLVMValueRef can_emit
;
1720 unsigned offset
= 0;
1721 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1723 /* Write vertex attribute values to GSVS ring */
1724 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
1725 ctx
->gs_next_vertex
[stream
],
1728 /* If this thread has already emitted the declared maximum number of
1729 * vertices, kill it: excessive vertex emissions are not supposed to
1730 * have any effect, and GS threads have no externally observable
1731 * effects other than emitting vertices.
1733 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
1734 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
1735 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1737 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1738 unsigned output_usage_mask
=
1739 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
1740 uint8_t output_stream
=
1741 ctx
->shader_info
->info
.gs
.output_streams
[i
];
1742 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1743 int length
= util_last_bit(output_usage_mask
);
1745 if (!(ctx
->output_mask
& (1ull << i
)) ||
1746 output_stream
!= stream
)
1749 for (unsigned j
= 0; j
< length
; j
++) {
1750 if (!(output_usage_mask
& (1 << j
)))
1753 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1755 LLVMValueRef voffset
=
1756 LLVMConstInt(ctx
->ac
.i32
, offset
*
1757 ctx
->gs_max_out_vertices
, false);
1761 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1762 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1764 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1765 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1767 ac_build_buffer_store_dword(&ctx
->ac
,
1768 ctx
->gsvs_ring
[stream
],
1770 voffset
, ctx
->gs2vs_offset
, 0,
1775 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1777 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
1779 ac_build_sendmsg(&ctx
->ac
,
1780 AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
1785 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1787 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1788 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1792 load_tess_coord(struct ac_shader_abi
*abi
)
1794 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1796 LLVMValueRef coord
[4] = {
1803 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
1804 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1805 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1807 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1811 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1813 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1814 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
1818 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1820 return abi
->base_vertex
;
1823 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1824 LLVMValueRef buffer_ptr
, bool write
)
1826 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1827 LLVMValueRef result
;
1829 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1831 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1832 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1837 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1839 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1840 LLVMValueRef result
;
1842 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1844 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1845 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1850 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1851 unsigned descriptor_set
,
1852 unsigned base_index
,
1853 unsigned constant_index
,
1855 enum ac_descriptor_type desc_type
,
1856 bool image
, bool write
,
1859 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1860 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1861 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
1862 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1863 unsigned offset
= binding
->offset
;
1864 unsigned stride
= binding
->size
;
1866 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1869 assert(base_index
< layout
->binding_count
);
1871 switch (desc_type
) {
1873 type
= ctx
->ac
.v8i32
;
1877 type
= ctx
->ac
.v8i32
;
1881 case AC_DESC_SAMPLER
:
1882 type
= ctx
->ac
.v4i32
;
1883 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
1888 case AC_DESC_BUFFER
:
1889 type
= ctx
->ac
.v4i32
;
1893 unreachable("invalid desc_type\n");
1896 offset
+= constant_index
* stride
;
1898 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1899 (!index
|| binding
->immutable_samplers_equal
)) {
1900 if (binding
->immutable_samplers_equal
)
1903 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1905 LLVMValueRef constants
[] = {
1906 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1907 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1908 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1909 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1911 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1914 assert(stride
% type_size
== 0);
1917 index
= ctx
->ac
.i32_0
;
1919 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1921 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
1922 list
= LLVMBuildPointerCast(builder
, list
,
1923 ac_array_in_const32_addr_space(type
), "");
1925 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
1928 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1929 * so we may need to fix it up. */
1931 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1932 unsigned adjustment
,
1935 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1938 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1940 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1941 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1943 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1945 /* For the integer-like cases, do a natural sign extension.
1947 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1948 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1951 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1952 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1953 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1954 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1956 /* Convert back to the right type. */
1957 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1959 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1960 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1961 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1962 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1963 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1964 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1971 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1972 struct nir_variable
*variable
)
1974 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
1975 LLVMValueRef t_offset
;
1976 LLVMValueRef t_list
;
1978 LLVMValueRef buffer_index
;
1979 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1980 uint8_t input_usage_mask
=
1981 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
1982 unsigned num_channels
= util_last_bit(input_usage_mask
);
1984 variable
->data
.driver_location
= variable
->data
.location
* 4;
1986 enum glsl_base_type type
= glsl_get_base_type(variable
->type
);
1987 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1988 LLVMValueRef output
[4];
1989 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1991 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1992 uint32_t divisor
= ctx
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
1995 buffer_index
= ctx
->abi
.instance_id
;
1998 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
1999 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
2002 if (ctx
->options
->key
.vs
.as_ls
) {
2003 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
2004 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2006 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
2007 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2010 buffer_index
= ctx
->ac
.i32_0
;
2013 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.start_instance
, buffer_index
, "");
2015 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
2016 ctx
->abi
.base_vertex
, "");
2017 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_index
, false);
2019 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
2021 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
2024 num_channels
, false, true);
2026 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
2028 for (unsigned chan
= 0; chan
< 4; chan
++) {
2029 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2030 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
2031 if (type
== GLSL_TYPE_FLOAT16
) {
2032 output
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f32
, "");
2033 output
[chan
] = LLVMBuildFPTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f16
, "");
2037 unsigned alpha_adjust
= (ctx
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
2038 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
2040 for (unsigned chan
= 0; chan
< 4; chan
++) {
2041 output
[chan
] = ac_to_integer(&ctx
->ac
, output
[chan
]);
2042 if (type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
)
2043 output
[chan
] = LLVMBuildTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.i16
, "");
2045 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] = output
[chan
];
2050 static void interp_fs_input(struct radv_shader_context
*ctx
,
2052 LLVMValueRef interp_param
,
2053 LLVMValueRef prim_mask
,
2054 LLVMValueRef result
[4])
2056 LLVMValueRef attr_number
;
2059 bool interp
= !LLVMIsUndef(interp_param
);
2061 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
2063 /* fs.constant returns the param from the middle vertex, so it's not
2064 * really useful for flat shading. It's meant to be used for custom
2065 * interpolation (but the intrinsic can't fetch from the other two
2068 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
2069 * to do the right thing. The only reason we use fs.constant is that
2070 * fs.interp cannot be used on integers, because they can be equal
2074 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
2077 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2079 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
2083 for (chan
= 0; chan
< 4; chan
++) {
2084 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
2087 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
2092 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
2093 LLVMConstInt(ctx
->ac
.i32
, 2, false),
2097 result
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, result
[chan
], ctx
->ac
.i32
, "");
2098 result
[chan
] = LLVMBuildTruncOrBitCast(ctx
->ac
.builder
, result
[chan
], LLVMTypeOf(interp_param
), "");
2104 handle_fs_input_decl(struct radv_shader_context
*ctx
,
2105 struct nir_variable
*variable
)
2107 int idx
= variable
->data
.location
;
2108 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2109 LLVMValueRef interp
= NULL
;
2112 variable
->data
.driver_location
= idx
* 4;
2113 mask
= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
2115 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
2116 unsigned interp_type
;
2117 if (variable
->data
.sample
)
2118 interp_type
= INTERP_SAMPLE
;
2119 else if (variable
->data
.centroid
)
2120 interp_type
= INTERP_CENTROID
;
2122 interp_type
= INTERP_CENTER
;
2124 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
2126 bool is_16bit
= glsl_type_is_16bit(glsl_without_array(variable
->type
));
2127 LLVMTypeRef type
= is_16bit
? ctx
->ac
.i16
: ctx
->ac
.i32
;
2129 interp
= LLVMGetUndef(type
);
2131 for (unsigned i
= 0; i
< attrib_count
; ++i
)
2132 ctx
->inputs
[ac_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
2134 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2135 /* Do not account for the number of components inside the array
2136 * of clip/cull distances because this might wrongly set other
2137 * bits like primitive ID or layer.
2139 mask
= 1ull << VARYING_SLOT_CLIP_DIST0
;
2142 ctx
->input_mask
|= mask
;
2146 handle_vs_inputs(struct radv_shader_context
*ctx
,
2147 struct nir_shader
*nir
) {
2148 nir_foreach_variable(variable
, &nir
->inputs
)
2149 handle_vs_input_decl(ctx
, variable
);
2153 prepare_interp_optimize(struct radv_shader_context
*ctx
,
2154 struct nir_shader
*nir
)
2156 bool uses_center
= false;
2157 bool uses_centroid
= false;
2158 nir_foreach_variable(variable
, &nir
->inputs
) {
2159 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
2160 variable
->data
.sample
)
2163 if (variable
->data
.centroid
)
2164 uses_centroid
= true;
2169 if (uses_center
&& uses_centroid
) {
2170 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
2171 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
2172 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
2177 handle_fs_inputs(struct radv_shader_context
*ctx
,
2178 struct nir_shader
*nir
)
2180 prepare_interp_optimize(ctx
, nir
);
2182 nir_foreach_variable(variable
, &nir
->inputs
)
2183 handle_fs_input_decl(ctx
, variable
);
2187 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
2188 ctx
->shader_info
->info
.needs_multiview_view_index
) {
2189 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
2190 ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)] = LLVMGetUndef(ctx
->ac
.i32
);
2193 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
2194 LLVMValueRef interp_param
;
2195 LLVMValueRef
*inputs
= ctx
->inputs
+ac_llvm_reg_index_soa(i
, 0);
2197 if (!(ctx
->input_mask
& (1ull << i
)))
2200 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
2201 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
2202 interp_param
= *inputs
;
2203 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
2206 if (LLVMIsUndef(interp_param
))
2207 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
2208 if (i
>= VARYING_SLOT_VAR0
)
2209 ctx
->abi
.fs_input_attr_indices
[i
- VARYING_SLOT_VAR0
] = index
;
2211 } else if (i
== VARYING_SLOT_CLIP_DIST0
) {
2212 int length
= ctx
->shader_info
->info
.ps
.num_input_clips_culls
;
2214 for (unsigned j
= 0; j
< length
; j
+= 4) {
2215 inputs
= ctx
->inputs
+ ac_llvm_reg_index_soa(i
, j
);
2217 interp_param
= *inputs
;
2218 interp_fs_input(ctx
, index
, interp_param
,
2219 ctx
->abi
.prim_mask
, inputs
);
2222 } else if (i
== VARYING_SLOT_POS
) {
2223 for(int i
= 0; i
< 3; ++i
)
2224 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
2226 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2227 ctx
->abi
.frag_pos
[3]);
2230 ctx
->shader_info
->fs
.num_interp
= index
;
2231 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
2233 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
2234 ctx
->abi
.view_index
= ctx
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2238 scan_shader_output_decl(struct radv_shader_context
*ctx
,
2239 struct nir_variable
*variable
,
2240 struct nir_shader
*shader
,
2241 gl_shader_stage stage
)
2243 int idx
= variable
->data
.location
+ variable
->data
.index
;
2244 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
2245 uint64_t mask_attribs
;
2247 variable
->data
.driver_location
= idx
* 4;
2249 /* tess ctrl has it's own load/store paths for outputs */
2250 if (stage
== MESA_SHADER_TESS_CTRL
)
2253 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
2254 if (stage
== MESA_SHADER_VERTEX
||
2255 stage
== MESA_SHADER_TESS_EVAL
||
2256 stage
== MESA_SHADER_GEOMETRY
) {
2257 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
2258 if (stage
== MESA_SHADER_VERTEX
) {
2259 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2260 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2261 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2263 if (stage
== MESA_SHADER_TESS_EVAL
) {
2264 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
2265 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
2266 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
<<= shader
->info
.clip_distance_array_size
;
2269 mask_attribs
= 1ull << idx
;
2273 ctx
->output_mask
|= mask_attribs
;
2277 /* Initialize arguments for the shader export intrinsic */
2279 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
2280 LLVMValueRef
*values
,
2281 unsigned enabled_channels
,
2283 struct ac_export_args
*args
)
2285 /* Specify the channels that are enabled. */
2286 args
->enabled_channels
= enabled_channels
;
2288 /* Specify whether the EXEC mask represents the valid mask */
2289 args
->valid_mask
= 0;
2291 /* Specify whether this is the last export */
2294 /* Specify the target we are exporting */
2295 args
->target
= target
;
2297 args
->compr
= false;
2298 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
2299 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
2300 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
2301 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
2306 bool is_16bit
= ac_get_type_size(LLVMTypeOf(values
[0])) == 2;
2307 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2308 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
2309 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
2310 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
2311 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
2314 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2315 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2316 unsigned bits
, bool hi
) = NULL
;
2318 switch(col_format
) {
2319 case V_028714_SPI_SHADER_ZERO
:
2320 args
->enabled_channels
= 0; /* writemask */
2321 args
->target
= V_008DFC_SQ_EXP_NULL
;
2324 case V_028714_SPI_SHADER_32_R
:
2325 args
->enabled_channels
= 1;
2326 args
->out
[0] = values
[0];
2329 case V_028714_SPI_SHADER_32_GR
:
2330 args
->enabled_channels
= 0x3;
2331 args
->out
[0] = values
[0];
2332 args
->out
[1] = values
[1];
2335 case V_028714_SPI_SHADER_32_AR
:
2336 args
->enabled_channels
= 0x9;
2337 args
->out
[0] = values
[0];
2338 args
->out
[3] = values
[3];
2341 case V_028714_SPI_SHADER_FP16_ABGR
:
2342 args
->enabled_channels
= 0x5;
2343 packf
= ac_build_cvt_pkrtz_f16
;
2345 for (unsigned chan
= 0; chan
< 4; chan
++)
2346 values
[chan
] = LLVMBuildFPExt(ctx
->ac
.builder
,
2352 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2353 args
->enabled_channels
= 0x5;
2354 packf
= ac_build_cvt_pknorm_u16
;
2357 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2358 args
->enabled_channels
= 0x5;
2359 packf
= ac_build_cvt_pknorm_i16
;
2362 case V_028714_SPI_SHADER_UINT16_ABGR
:
2363 args
->enabled_channels
= 0x5;
2364 packi
= ac_build_cvt_pk_u16
;
2366 for (unsigned chan
= 0; chan
< 4; chan
++)
2367 values
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
,
2373 case V_028714_SPI_SHADER_SINT16_ABGR
:
2374 args
->enabled_channels
= 0x5;
2375 packi
= ac_build_cvt_pk_i16
;
2377 for (unsigned chan
= 0; chan
< 4; chan
++)
2378 values
[chan
] = LLVMBuildSExt(ctx
->ac
.builder
,
2385 case V_028714_SPI_SHADER_32_ABGR
:
2386 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2390 /* Pack f16 or norm_i16/u16. */
2392 for (chan
= 0; chan
< 2; chan
++) {
2393 LLVMValueRef pack_args
[2] = {
2395 values
[2 * chan
+ 1]
2397 LLVMValueRef packed
;
2399 packed
= packf(&ctx
->ac
, pack_args
);
2400 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2402 args
->compr
= 1; /* COMPR flag */
2407 for (chan
= 0; chan
< 2; chan
++) {
2408 LLVMValueRef pack_args
[2] = {
2409 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2410 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2412 LLVMValueRef packed
;
2414 packed
= packi(&ctx
->ac
, pack_args
,
2415 is_int8
? 8 : is_int10
? 10 : 16,
2417 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2419 args
->compr
= 1; /* COMPR flag */
2425 for (unsigned chan
= 0; chan
< 4; chan
++) {
2426 values
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i16
, "");
2427 args
->out
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i32
, "");
2430 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2432 for (unsigned i
= 0; i
< 4; ++i
) {
2433 if (!(args
->enabled_channels
& (1 << i
)))
2436 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
2441 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
2442 LLVMValueRef
*values
, unsigned enabled_channels
)
2444 struct ac_export_args args
;
2446 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
2447 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2448 ac_build_export(&ctx
->ac
, &args
);
2452 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
2454 LLVMValueRef output
=
2455 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
2457 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
2461 radv_emit_stream_output(struct radv_shader_context
*ctx
,
2462 LLVMValueRef
const *so_buffers
,
2463 LLVMValueRef
const *so_write_offsets
,
2464 const struct radv_stream_output
*output
)
2466 unsigned num_comps
= util_bitcount(output
->component_mask
);
2467 unsigned loc
= output
->location
;
2468 unsigned buf
= output
->buffer
;
2469 unsigned offset
= output
->offset
;
2471 LLVMValueRef out
[4];
2473 assert(num_comps
&& num_comps
<= 4);
2474 if (!num_comps
|| num_comps
> 4)
2477 /* Get the first component. */
2478 start
= ffs(output
->component_mask
) - 1;
2480 /* Load the output as int. */
2481 for (int i
= 0; i
< num_comps
; i
++) {
2482 out
[i
] = ac_to_integer(&ctx
->ac
,
2483 radv_load_output(ctx
, loc
, start
+ i
));
2486 /* Pack the output. */
2487 LLVMValueRef vdata
= NULL
;
2489 switch (num_comps
) {
2490 case 1: /* as i32 */
2493 case 2: /* as v2i32 */
2494 case 3: /* as v4i32 (aligned to 4) */
2495 out
[3] = LLVMGetUndef(ctx
->ac
.i32
);
2497 case 4: /* as v4i32 */
2498 vdata
= ac_build_gather_values(&ctx
->ac
, out
,
2499 util_next_power_of_two(num_comps
));
2503 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf
],
2504 vdata
, num_comps
, so_write_offsets
[buf
],
2505 ctx
->ac
.i32_0
, offset
,
2510 radv_emit_streamout(struct radv_shader_context
*ctx
, unsigned stream
)
2512 struct ac_build_if_state if_ctx
;
2515 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2516 assert(ctx
->streamout_config
);
2517 LLVMValueRef so_vtx_count
=
2518 ac_build_bfe(&ctx
->ac
, ctx
->streamout_config
,
2519 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2520 LLVMConstInt(ctx
->ac
.i32
, 7, false), false);
2522 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2524 /* can_emit = tid < so_vtx_count; */
2525 LLVMValueRef can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
2526 tid
, so_vtx_count
, "");
2528 /* Emit the streamout code conditionally. This actually avoids
2529 * out-of-bounds buffer access. The hw tells us via the SGPR
2530 * (so_vtx_count) which threads are allowed to emit streamout data.
2532 ac_nir_build_if(&if_ctx
, ctx
, can_emit
);
2534 /* The buffer offset is computed as follows:
2535 * ByteOffset = streamout_offset[buffer_id]*4 +
2536 * (streamout_write_index + thread_id)*stride[buffer_id] +
2539 LLVMValueRef so_write_index
= ctx
->streamout_write_idx
;
2541 /* Compute (streamout_write_index + thread_id). */
2543 LLVMBuildAdd(ctx
->ac
.builder
, so_write_index
, tid
, "");
2545 /* Load the descriptor and compute the write offset for each
2548 LLVMValueRef so_write_offset
[4] = {};
2549 LLVMValueRef so_buffers
[4] = {};
2550 LLVMValueRef buf_ptr
= ctx
->streamout_buffers
;
2552 for (i
= 0; i
< 4; i
++) {
2553 uint16_t stride
= ctx
->shader_info
->info
.so
.strides
[i
];
2558 LLVMValueRef offset
=
2559 LLVMConstInt(ctx
->ac
.i32
, i
, false);
2561 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
,
2564 LLVMValueRef so_offset
= ctx
->streamout_offset
[i
];
2566 so_offset
= LLVMBuildMul(ctx
->ac
.builder
, so_offset
,
2567 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
2569 so_write_offset
[i
] =
2570 ac_build_imad(&ctx
->ac
, so_write_index
,
2571 LLVMConstInt(ctx
->ac
.i32
,
2576 /* Write streamout data. */
2577 for (i
= 0; i
< ctx
->shader_info
->info
.so
.num_outputs
; i
++) {
2578 struct radv_stream_output
*output
=
2579 &ctx
->shader_info
->info
.so
.outputs
[i
];
2581 if (stream
!= output
->stream
)
2584 radv_emit_stream_output(ctx
, so_buffers
,
2585 so_write_offset
, output
);
2588 ac_nir_build_endif(&if_ctx
);
2592 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2593 bool export_prim_id
, bool export_layer_id
,
2594 struct radv_vs_output_info
*outinfo
)
2596 uint32_t param_count
= 0;
2598 unsigned pos_idx
, num_pos_exports
= 0;
2599 struct ac_export_args args
, pos_args
[4] = {};
2600 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2603 if (ctx
->options
->key
.has_multiview_view_index
) {
2604 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2606 for(unsigned i
= 0; i
< 4; ++i
)
2607 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2608 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2611 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
2612 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2615 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2616 sizeof(outinfo
->vs_output_param_offset
));
2618 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
2619 unsigned output_usage_mask
, length
;
2620 LLVMValueRef slots
[8];
2623 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2624 !ctx
->is_gs_copy_shader
) {
2626 ctx
->shader_info
->info
.vs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2627 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2629 ctx
->shader_info
->info
.tes
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2631 assert(ctx
->is_gs_copy_shader
);
2633 ctx
->shader_info
->info
.gs
.output_usage_mask
[VARYING_SLOT_CLIP_DIST0
];
2636 length
= util_last_bit(output_usage_mask
);
2638 i
= VARYING_SLOT_CLIP_DIST0
;
2639 for (j
= 0; j
< length
; j
++)
2640 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2642 for (i
= length
; i
< 8; i
++)
2643 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
2646 target
= V_008DFC_SQ_EXP_POS
+ 3;
2647 si_llvm_init_export_args(ctx
, &slots
[4], 0xf, target
, &args
);
2648 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2649 &args
, sizeof(args
));
2652 target
= V_008DFC_SQ_EXP_POS
+ 2;
2653 si_llvm_init_export_args(ctx
, &slots
[0], 0xf, target
, &args
);
2654 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2655 &args
, sizeof(args
));
2657 /* Export the clip/cull distances values to the next stage. */
2658 radv_export_param(ctx
, param_count
, &slots
[0], 0xf);
2659 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST0
] = param_count
++;
2661 radv_export_param(ctx
, param_count
, &slots
[4], 0xf);
2662 outinfo
->vs_output_param_offset
[VARYING_SLOT_CLIP_DIST1
] = param_count
++;
2666 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
2667 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
2668 for (unsigned j
= 0; j
< 4; j
++)
2669 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
2671 si_llvm_init_export_args(ctx
, pos_values
, 0xf, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2673 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
2674 outinfo
->writes_pointsize
= true;
2675 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
2678 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
2679 outinfo
->writes_layer
= true;
2680 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
2683 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
2684 outinfo
->writes_viewport_index
= true;
2685 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
2688 if (ctx
->shader_info
->info
.so
.num_outputs
&&
2689 !ctx
->is_gs_copy_shader
) {
2690 /* The GS copy shader emission already emits streamout. */
2691 radv_emit_streamout(ctx
, 0);
2694 if (outinfo
->writes_pointsize
||
2695 outinfo
->writes_layer
||
2696 outinfo
->writes_viewport_index
) {
2697 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
2698 (outinfo
->writes_layer
== true ? 4 : 0));
2699 pos_args
[1].valid_mask
= 0;
2700 pos_args
[1].done
= 0;
2701 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2702 pos_args
[1].compr
= 0;
2703 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2704 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2705 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2706 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2708 if (outinfo
->writes_pointsize
== true)
2709 pos_args
[1].out
[0] = psize_value
;
2710 if (outinfo
->writes_layer
== true)
2711 pos_args
[1].out
[2] = layer_value
;
2712 if (outinfo
->writes_viewport_index
== true) {
2713 if (ctx
->options
->chip_class
>= GFX9
) {
2714 /* GFX9 has the layer in out.z[10:0] and the viewport
2715 * index in out.z[19:16].
2717 LLVMValueRef v
= viewport_index_value
;
2718 v
= ac_to_integer(&ctx
->ac
, v
);
2719 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2720 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2722 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2723 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2725 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2726 pos_args
[1].enabled_channels
|= 1 << 2;
2728 pos_args
[1].out
[3] = viewport_index_value
;
2729 pos_args
[1].enabled_channels
|= 1 << 3;
2733 for (i
= 0; i
< 4; i
++) {
2734 if (pos_args
[i
].out
[0])
2739 for (i
= 0; i
< 4; i
++) {
2740 if (!pos_args
[i
].out
[0])
2743 /* Specify the target we are exporting */
2744 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2745 if (pos_idx
== num_pos_exports
)
2746 pos_args
[i
].done
= 1;
2747 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2750 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2751 LLVMValueRef values
[4];
2752 if (!(ctx
->output_mask
& (1ull << i
)))
2755 if (i
!= VARYING_SLOT_LAYER
&&
2756 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
2757 i
< VARYING_SLOT_VAR0
)
2760 for (unsigned j
= 0; j
< 4; j
++)
2761 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2763 unsigned output_usage_mask
;
2765 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2766 !ctx
->is_gs_copy_shader
) {
2768 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2769 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2771 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2773 assert(ctx
->is_gs_copy_shader
);
2775 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
2778 radv_export_param(ctx
, param_count
, values
, output_usage_mask
);
2780 outinfo
->vs_output_param_offset
[i
] = param_count
++;
2783 if (export_prim_id
) {
2784 LLVMValueRef values
[4];
2786 values
[0] = ctx
->vs_prim_id
;
2787 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
2788 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
2789 for (unsigned j
= 1; j
< 4; j
++)
2790 values
[j
] = ctx
->ac
.f32_0
;
2792 radv_export_param(ctx
, param_count
, values
, 0x1);
2794 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
2795 outinfo
->export_prim_id
= true;
2798 if (export_layer_id
&& layer_value
) {
2799 LLVMValueRef values
[4];
2801 values
[0] = layer_value
;
2802 for (unsigned j
= 1; j
< 4; j
++)
2803 values
[j
] = ctx
->ac
.f32_0
;
2805 radv_export_param(ctx
, param_count
, values
, 0x1);
2807 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
++;
2810 outinfo
->pos_exports
= num_pos_exports
;
2811 outinfo
->param_exports
= param_count
;
2815 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2816 struct radv_es_output_info
*outinfo
)
2819 uint64_t max_output_written
= 0;
2820 LLVMValueRef lds_base
= NULL
;
2822 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2823 unsigned output_usage_mask
;
2827 if (!(ctx
->output_mask
& (1ull << i
)))
2830 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2832 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2834 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2836 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2839 if (i
== VARYING_SLOT_CLIP_DIST0
)
2840 length
= util_last_bit(output_usage_mask
);
2842 param_index
= shader_io_get_unique_index(i
);
2844 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
2847 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
2849 if (ctx
->ac
.chip_class
>= GFX9
) {
2850 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2851 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2852 LLVMValueRef wave_idx
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 24, 4);
2853 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2854 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2855 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
2856 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2857 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2860 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2861 LLVMValueRef dw_addr
= NULL
;
2862 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2863 unsigned output_usage_mask
;
2867 if (!(ctx
->output_mask
& (1ull << i
)))
2870 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2872 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2874 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2876 ctx
->shader_info
->info
.tes
.output_usage_mask
[i
];
2879 if (i
== VARYING_SLOT_CLIP_DIST0
)
2880 length
= util_last_bit(output_usage_mask
);
2882 param_index
= shader_io_get_unique_index(i
);
2885 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2886 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2890 for (j
= 0; j
< length
; j
++) {
2891 if (!(output_usage_mask
& (1 << j
)))
2894 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2895 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
2896 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2898 if (ctx
->ac
.chip_class
>= GFX9
) {
2899 LLVMValueRef dw_addr_offset
=
2900 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2901 LLVMConstInt(ctx
->ac
.i32
,
2904 ac_lds_store(&ctx
->ac
, dw_addr_offset
, out_val
);
2906 ac_build_buffer_store_dword(&ctx
->ac
,
2909 NULL
, ctx
->es2gs_offset
,
2910 (4 * param_index
+ j
) * 4,
2918 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2920 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2921 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->shader_info
->info
.vs
.ls_outputs_written
);
2922 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2923 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2924 vertex_dw_stride
, "");
2926 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2927 unsigned output_usage_mask
=
2928 ctx
->shader_info
->info
.vs
.output_usage_mask
[i
];
2929 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2932 if (!(ctx
->output_mask
& (1ull << i
)))
2935 if (i
== VARYING_SLOT_CLIP_DIST0
)
2936 length
= util_last_bit(output_usage_mask
);
2938 int param
= shader_io_get_unique_index(i
);
2939 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2940 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2942 for (unsigned j
= 0; j
< length
; j
++) {
2943 LLVMValueRef value
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2944 value
= ac_to_integer(&ctx
->ac
, value
);
2945 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
2946 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
2947 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2953 write_tess_factors(struct radv_shader_context
*ctx
)
2955 unsigned stride
, outer_comps
, inner_comps
;
2956 struct ac_build_if_state if_ctx
, inner_if_ctx
;
2957 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2958 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
2959 unsigned tess_inner_index
= 0, tess_outer_index
;
2960 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
2961 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2963 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
2965 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
2985 ac_nir_build_if(&if_ctx
, ctx
,
2986 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2987 invocation_id
, ctx
->ac
.i32_0
, ""));
2989 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
2992 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
2993 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2994 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
2997 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
2998 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2999 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
3001 for (i
= 0; i
< 4; i
++) {
3002 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3003 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3007 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
3008 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
3009 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3011 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
3013 for (i
= 0; i
< outer_comps
; i
++) {
3015 ac_lds_load(&ctx
->ac
, lds_outer
);
3016 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3019 for (i
= 0; i
< inner_comps
; i
++) {
3020 inner
[i
] = out
[outer_comps
+i
] =
3021 ac_lds_load(&ctx
->ac
, lds_inner
);
3022 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
3027 /* Convert the outputs to vectors for stores. */
3028 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
3032 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
3035 buffer
= ctx
->hs_ring_tess_factor
;
3036 tf_base
= ctx
->tess_factor_offset
;
3037 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3038 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
3039 unsigned tf_offset
= 0;
3041 if (ctx
->options
->chip_class
<= VI
) {
3042 ac_nir_build_if(&inner_if_ctx
, ctx
,
3043 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3044 rel_patch_id
, ctx
->ac
.i32_0
, ""));
3046 /* Store the dynamic HS control word. */
3047 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3048 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
3049 1, ctx
->ac
.i32_0
, tf_base
,
3050 0, 1, 0, true, false);
3053 ac_nir_build_endif(&inner_if_ctx
);
3056 /* Store the tessellation factors. */
3057 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3058 MIN2(stride
, 4), byteoffset
, tf_base
,
3059 tf_offset
, 1, 0, true, false);
3061 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3062 stride
- 4, byteoffset
, tf_base
,
3063 16 + tf_offset
, 1, 0, true, false);
3065 //store to offchip for TES to read - only if TES reads them
3066 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
3067 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
3068 LLVMValueRef tf_inner_offset
;
3069 unsigned param_outer
, param_inner
;
3071 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
3072 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3073 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
3075 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
3076 util_next_power_of_two(outer_comps
));
3078 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
3079 outer_comps
, tf_outer_offset
,
3080 ctx
->oc_lds
, 0, 1, 0, true, false);
3082 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
3083 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3084 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
3086 inner_vec
= inner_comps
== 1 ? inner
[0] :
3087 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
3088 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
3089 inner_comps
, tf_inner_offset
,
3090 ctx
->oc_lds
, 0, 1, 0, true, false);
3093 ac_nir_build_endif(&if_ctx
);
3097 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
3099 write_tess_factors(ctx
);
3103 si_export_mrt_color(struct radv_shader_context
*ctx
,
3104 LLVMValueRef
*color
, unsigned index
,
3105 struct ac_export_args
*args
)
3108 si_llvm_init_export_args(ctx
, color
, 0xf,
3109 V_008DFC_SQ_EXP_MRT
+ index
, args
);
3110 if (!args
->enabled_channels
)
3111 return false; /* unnecessary NULL export */
3117 radv_export_mrt_z(struct radv_shader_context
*ctx
,
3118 LLVMValueRef depth
, LLVMValueRef stencil
,
3119 LLVMValueRef samplemask
)
3121 struct ac_export_args args
;
3123 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3125 ac_build_export(&ctx
->ac
, &args
);
3129 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
3132 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3133 struct ac_export_args color_args
[8];
3135 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3136 LLVMValueRef values
[4];
3138 if (!(ctx
->output_mask
& (1ull << i
)))
3141 if (i
< FRAG_RESULT_DATA0
)
3144 for (unsigned j
= 0; j
< 4; j
++)
3145 values
[j
] = ac_to_float(&ctx
->ac
,
3146 radv_load_output(ctx
, i
, j
));
3148 bool ret
= si_export_mrt_color(ctx
, values
,
3149 i
- FRAG_RESULT_DATA0
,
3150 &color_args
[index
]);
3155 /* Process depth, stencil, samplemask. */
3156 if (ctx
->shader_info
->info
.ps
.writes_z
) {
3157 depth
= ac_to_float(&ctx
->ac
,
3158 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
3160 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
3161 stencil
= ac_to_float(&ctx
->ac
,
3162 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
3164 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3165 samplemask
= ac_to_float(&ctx
->ac
,
3166 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
3169 /* Set the DONE bit on last non-null color export only if Z isn't
3173 !ctx
->shader_info
->info
.ps
.writes_z
&&
3174 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
3175 !ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
3176 unsigned last
= index
- 1;
3178 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
3179 color_args
[last
].done
= 1; /* DONE bit */
3182 /* Export PS outputs. */
3183 for (unsigned i
= 0; i
< index
; i
++)
3184 ac_build_export(&ctx
->ac
, &color_args
[i
]);
3186 if (depth
|| stencil
|| samplemask
)
3187 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
3189 ac_build_export_null(&ctx
->ac
);
3193 emit_gs_epilogue(struct radv_shader_context
*ctx
)
3195 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
3199 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
3200 LLVMValueRef
*addrs
)
3202 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
3204 switch (ctx
->stage
) {
3205 case MESA_SHADER_VERTEX
:
3206 if (ctx
->options
->key
.vs
.as_ls
)
3207 handle_ls_outputs_post(ctx
);
3208 else if (ctx
->options
->key
.vs
.as_es
)
3209 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
3211 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
3212 ctx
->options
->key
.vs
.export_layer_id
,
3213 &ctx
->shader_info
->vs
.outinfo
);
3215 case MESA_SHADER_FRAGMENT
:
3216 handle_fs_outputs_post(ctx
);
3218 case MESA_SHADER_GEOMETRY
:
3219 emit_gs_epilogue(ctx
);
3221 case MESA_SHADER_TESS_CTRL
:
3222 handle_tcs_outputs_post(ctx
);
3224 case MESA_SHADER_TESS_EVAL
:
3225 if (ctx
->options
->key
.tes
.as_es
)
3226 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
3228 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
3229 ctx
->options
->key
.tes
.export_layer_id
,
3230 &ctx
->shader_info
->tes
.outinfo
);
3237 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
,
3238 LLVMPassManagerRef passmgr
,
3239 const struct radv_nir_compiler_options
*options
)
3241 LLVMRunPassManager(passmgr
, ctx
->ac
.module
);
3242 LLVMDisposeBuilder(ctx
->ac
.builder
);
3244 ac_llvm_context_dispose(&ctx
->ac
);
3248 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
3250 struct radv_vs_output_info
*outinfo
;
3252 switch (ctx
->stage
) {
3253 case MESA_SHADER_FRAGMENT
:
3254 case MESA_SHADER_COMPUTE
:
3255 case MESA_SHADER_TESS_CTRL
:
3256 case MESA_SHADER_GEOMETRY
:
3258 case MESA_SHADER_VERTEX
:
3259 if (ctx
->options
->key
.vs
.as_ls
||
3260 ctx
->options
->key
.vs
.as_es
)
3262 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
3264 case MESA_SHADER_TESS_EVAL
:
3265 if (ctx
->options
->key
.vs
.as_es
)
3267 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
3270 unreachable("Unhandled shader type");
3273 ac_optimize_vs_outputs(&ctx
->ac
,
3275 outinfo
->vs_output_param_offset
,
3277 &outinfo
->param_exports
);
3281 ac_setup_rings(struct radv_shader_context
*ctx
)
3283 if (ctx
->options
->chip_class
<= VI
&&
3284 (ctx
->stage
== MESA_SHADER_GEOMETRY
||
3285 ctx
->options
->key
.vs
.as_es
|| ctx
->options
->key
.tes
.as_es
)) {
3286 unsigned ring
= ctx
->stage
== MESA_SHADER_GEOMETRY
? RING_ESGS_GS
3288 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, ring
, false);
3290 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
,
3295 if (ctx
->is_gs_copy_shader
) {
3297 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3298 LLVMConstInt(ctx
->ac
.i32
,
3299 RING_GSVS_VS
, false));
3302 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3303 /* The conceptual layout of the GSVS ring is
3304 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3305 * but the real memory layout is swizzled across
3307 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3309 * Override the buffer descriptor accordingly.
3311 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->ac
.i64
, 2);
3312 uint64_t stream_offset
= 0;
3313 unsigned num_records
= 64;
3314 LLVMValueRef base_ring
;
3317 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3318 LLVMConstInt(ctx
->ac
.i32
,
3319 RING_GSVS_GS
, false));
3321 for (unsigned stream
= 0; stream
< 4; stream
++) {
3322 unsigned num_components
, stride
;
3323 LLVMValueRef ring
, tmp
;
3326 ctx
->shader_info
->info
.gs
.num_stream_output_components
[stream
];
3328 if (!num_components
)
3331 stride
= 4 * num_components
* ctx
->gs_max_out_vertices
;
3333 /* Limit on the stride field for <= CIK. */
3334 assert(stride
< (1 << 14));
3336 ring
= LLVMBuildBitCast(ctx
->ac
.builder
,
3337 base_ring
, v2i64
, "");
3338 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3339 ring
, ctx
->ac
.i32_0
, "");
3340 tmp
= LLVMBuildAdd(ctx
->ac
.builder
, tmp
,
3341 LLVMConstInt(ctx
->ac
.i64
,
3342 stream_offset
, 0), "");
3343 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
,
3344 ring
, tmp
, ctx
->ac
.i32_0
, "");
3346 stream_offset
+= stride
* 64;
3348 ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ring
,
3351 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ring
,
3353 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
,
3354 LLVMConstInt(ctx
->ac
.i32
,
3355 S_008F04_STRIDE(stride
), false), "");
3356 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
, tmp
,
3359 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
,
3360 LLVMConstInt(ctx
->ac
.i32
,
3361 num_records
, false),
3362 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3364 ctx
->gsvs_ring
[stream
] = ring
;
3368 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3369 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3370 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3371 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3376 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
3377 const struct nir_shader
*nir
)
3379 switch (nir
->info
.stage
) {
3380 case MESA_SHADER_TESS_CTRL
:
3381 return chip_class
>= CIK
? 128 : 64;
3382 case MESA_SHADER_GEOMETRY
:
3383 return chip_class
>= GFX9
? 128 : 64;
3384 case MESA_SHADER_COMPUTE
:
3390 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
3391 nir
->info
.cs
.local_size
[1] *
3392 nir
->info
.cs
.local_size
[2];
3393 return max_workgroup_size
;
3396 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3397 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3399 LLVMValueRef count
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 8, 8);
3400 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3402 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
3403 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
3404 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
3407 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
)
3409 for(int i
= 5; i
>= 0; --i
) {
3410 ctx
->gs_vtx_offset
[i
] = ac_unpack_param(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
3414 ctx
->gs_wave_id
= ac_unpack_param(&ctx
->ac
, ctx
->merged_wave_info
, 16, 8);
3419 LLVMModuleRef
ac_translate_nir_to_llvm(struct ac_llvm_compiler
*ac_llvm
,
3420 struct nir_shader
*const *shaders
,
3422 struct radv_shader_variant_info
*shader_info
,
3423 const struct radv_nir_compiler_options
*options
)
3425 struct radv_shader_context ctx
= {0};
3427 ctx
.options
= options
;
3428 ctx
.shader_info
= shader_info
;
3430 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3431 ctx
.context
= ctx
.ac
.context
;
3432 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3434 enum ac_float_mode float_mode
=
3435 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3436 AC_FLOAT_MODE_DEFAULT
;
3438 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3440 memset(shader_info
, 0, sizeof(*shader_info
));
3442 for(int i
= 0; i
< shader_count
; ++i
)
3443 radv_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
3445 for (i
= 0; i
< RADV_UD_MAX_SETS
; i
++)
3446 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
3447 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
3448 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
3450 ctx
.max_workgroup_size
= 0;
3451 for (int i
= 0; i
< shader_count
; ++i
) {
3452 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
3453 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
3457 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
3458 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
3460 ctx
.abi
.inputs
= &ctx
.inputs
[0];
3461 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
3462 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
3463 ctx
.abi
.load_ubo
= radv_load_ubo
;
3464 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
3465 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
3466 ctx
.abi
.load_resource
= radv_load_resource
;
3467 ctx
.abi
.clamp_shadow_reference
= false;
3468 ctx
.abi
.gfx9_stride_size_workaround
= ctx
.ac
.chip_class
== GFX9
&& HAVE_LLVM
< 0x800;
3470 if (shader_count
>= 2)
3471 ac_init_exec_full_mask(&ctx
.ac
);
3473 if (ctx
.ac
.chip_class
== GFX9
&&
3474 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
3475 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
3477 for(int i
= 0; i
< shader_count
; ++i
) {
3478 ctx
.stage
= shaders
[i
]->info
.stage
;
3479 ctx
.output_mask
= 0;
3481 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3482 for (int i
= 0; i
< 4; i
++) {
3483 ctx
.gs_next_vertex
[i
] =
3484 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
3486 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
3487 ctx
.abi
.load_inputs
= load_gs_input
;
3488 ctx
.abi
.emit_primitive
= visit_end_primitive
;
3489 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3490 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
3491 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
3492 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
3493 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3494 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
3495 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3496 if (shader_count
== 1)
3497 ctx
.tcs_num_inputs
= ctx
.options
->key
.tcs
.num_inputs
;
3499 ctx
.tcs_num_inputs
= util_last_bit64(shader_info
->info
.vs
.ls_outputs_written
);
3500 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
3501 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3502 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
3503 ctx
.abi
.load_tess_varyings
= load_tes_input
;
3504 ctx
.abi
.load_tess_coord
= load_tess_coord
;
3505 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
3506 ctx
.tcs_vertices_per_patch
= shaders
[i
]->info
.tess
.tcs_vertices_out
;
3507 ctx
.tcs_num_patches
= ctx
.options
->key
.tes
.num_patches
;
3508 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
3509 if (shader_info
->info
.vs
.needs_instance_id
) {
3510 if (ctx
.options
->key
.vs
.as_ls
) {
3511 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3512 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3514 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
3515 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
3518 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
3519 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
3520 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
3521 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
3522 ctx
.abi
.load_sample_position
= load_sample_position
;
3523 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
3524 ctx
.abi
.emit_kill
= radv_emit_kill
;
3528 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
3530 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
3531 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
3533 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3534 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
3535 shaders
[i
]->info
.cull_distance_array_size
> 4;
3536 ctx
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
3537 ctx
.max_gsvs_emit_size
= ctx
.gsvs_vertex_size
*
3538 shaders
[i
]->info
.gs
.vertices_out
;
3541 ac_setup_rings(&ctx
);
3543 LLVMBasicBlockRef merge_block
;
3544 if (shader_count
>= 2) {
3545 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
3546 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3547 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
3549 LLVMValueRef count
= ac_unpack_param(&ctx
.ac
, ctx
.merged_wave_info
, 8 * i
, 8);
3550 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
3551 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
3552 thread_id
, count
, "");
3553 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
3555 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
3558 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
3559 handle_fs_inputs(&ctx
, shaders
[i
]);
3560 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
3561 handle_vs_inputs(&ctx
, shaders
[i
]);
3562 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
3563 prepare_gs_input_vgprs(&ctx
);
3565 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
]);
3567 if (shader_count
>= 2) {
3568 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
3569 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
3572 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3573 shader_info
->gs
.gsvs_vertex_size
= ctx
.gsvs_vertex_size
;
3574 shader_info
->gs
.max_gsvs_emit_size
= ctx
.max_gsvs_emit_size
;
3575 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3576 shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
3577 shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
3581 LLVMBuildRetVoid(ctx
.ac
.builder
);
3583 if (options
->dump_preoptir
)
3584 ac_dump_module(ctx
.ac
.module
);
3586 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3588 if (shader_count
== 1)
3589 ac_nir_eliminate_const_vs_outputs(&ctx
);
3591 if (options
->dump_shader
) {
3592 ctx
.shader_info
->private_mem_vgprs
=
3593 ac_count_scratch_private_memory(ctx
.main_function
);
3596 return ctx
.ac
.module
;
3599 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
3601 unsigned *retval
= (unsigned *)context
;
3602 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
3603 char *description
= LLVMGetDiagInfoDescription(di
);
3605 if (severity
== LLVMDSError
) {
3607 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
3611 LLVMDisposeMessage(description
);
3614 static unsigned ac_llvm_compile(LLVMModuleRef M
,
3615 struct ac_shader_binary
*binary
,
3616 struct ac_llvm_compiler
*ac_llvm
)
3618 unsigned retval
= 0;
3619 LLVMContextRef llvm_ctx
;
3621 /* Setup Diagnostic Handler*/
3622 llvm_ctx
= LLVMGetModuleContext(M
);
3624 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
3628 if (!radv_compile_to_binary(ac_llvm
, M
, binary
))
3633 static void ac_compile_llvm_module(struct ac_llvm_compiler
*ac_llvm
,
3634 LLVMModuleRef llvm_module
,
3635 struct ac_shader_binary
*binary
,
3636 struct ac_shader_config
*config
,
3637 struct radv_shader_variant_info
*shader_info
,
3638 gl_shader_stage stage
,
3639 const struct radv_nir_compiler_options
*options
)
3641 if (options
->dump_shader
)
3642 ac_dump_module(llvm_module
);
3644 memset(binary
, 0, sizeof(*binary
));
3646 if (options
->record_llvm_ir
) {
3647 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
3648 binary
->llvm_ir_string
= strdup(llvm_ir
);
3649 LLVMDisposeMessage(llvm_ir
);
3652 int v
= ac_llvm_compile(llvm_module
, binary
, ac_llvm
);
3654 fprintf(stderr
, "compile failed\n");
3657 if (options
->dump_shader
)
3658 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
3660 ac_shader_binary_read_config(binary
, config
, 0, options
->supports_spill
);
3662 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
3663 LLVMDisposeModule(llvm_module
);
3664 LLVMContextDispose(ctx
);
3666 if (stage
== MESA_SHADER_FRAGMENT
) {
3667 shader_info
->num_input_vgprs
= 0;
3668 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
3669 shader_info
->num_input_vgprs
+= 2;
3670 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
3671 shader_info
->num_input_vgprs
+= 2;
3672 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
3673 shader_info
->num_input_vgprs
+= 2;
3674 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
3675 shader_info
->num_input_vgprs
+= 3;
3676 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
3677 shader_info
->num_input_vgprs
+= 2;
3678 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
3679 shader_info
->num_input_vgprs
+= 2;
3680 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
3681 shader_info
->num_input_vgprs
+= 2;
3682 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
3683 shader_info
->num_input_vgprs
+= 1;
3684 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
3685 shader_info
->num_input_vgprs
+= 1;
3686 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
3687 shader_info
->num_input_vgprs
+= 1;
3688 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
3689 shader_info
->num_input_vgprs
+= 1;
3690 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
3691 shader_info
->num_input_vgprs
+= 1;
3692 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
3693 shader_info
->num_input_vgprs
+= 1;
3694 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
3695 shader_info
->num_input_vgprs
+= 1;
3696 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
3697 shader_info
->num_input_vgprs
+= 1;
3698 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
3699 shader_info
->num_input_vgprs
+= 1;
3701 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
3703 /* +3 for scratch wave offset and VCC */
3704 config
->num_sgprs
= MAX2(config
->num_sgprs
,
3705 shader_info
->num_input_sgprs
+ 3);
3707 /* Enable 64-bit and 16-bit denormals, because there is no performance
3710 * If denormals are enabled, all floating-point output modifiers are
3713 * Don't enable denormals for 32-bit floats, because:
3714 * - Floating-point output modifiers would be ignored by the hw.
3715 * - Some opcodes don't support denormals, such as v_mad_f32. We would
3716 * have to stop using those.
3717 * - SI & CI would be very slow.
3719 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
3723 ac_fill_shader_info(struct radv_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct radv_nir_compiler_options
*options
)
3725 switch (nir
->info
.stage
) {
3726 case MESA_SHADER_COMPUTE
:
3727 for (int i
= 0; i
< 3; ++i
)
3728 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
3730 case MESA_SHADER_FRAGMENT
:
3731 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
3733 case MESA_SHADER_GEOMETRY
:
3734 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
3735 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
3736 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
3737 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
3739 case MESA_SHADER_TESS_EVAL
:
3740 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
3741 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
3742 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
3743 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
3744 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
3746 case MESA_SHADER_TESS_CTRL
:
3747 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
3749 case MESA_SHADER_VERTEX
:
3750 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
3751 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
3752 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
3753 if (options
->key
.vs
.as_ls
)
3754 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
3762 radv_compile_nir_shader(struct ac_llvm_compiler
*ac_llvm
,
3763 struct ac_shader_binary
*binary
,
3764 struct ac_shader_config
*config
,
3765 struct radv_shader_variant_info
*shader_info
,
3766 struct nir_shader
*const *nir
,
3768 const struct radv_nir_compiler_options
*options
)
3771 LLVMModuleRef llvm_module
;
3773 llvm_module
= ac_translate_nir_to_llvm(ac_llvm
, nir
, nir_count
, shader_info
,
3776 ac_compile_llvm_module(ac_llvm
, llvm_module
, binary
, config
, shader_info
,
3777 nir
[0]->info
.stage
, options
);
3779 for (int i
= 0; i
< nir_count
; ++i
)
3780 ac_fill_shader_info(shader_info
, nir
[i
], options
);
3782 /* Determine the ES type (VS or TES) for the GS on GFX9. */
3783 if (options
->chip_class
== GFX9
) {
3784 if (nir_count
== 2 &&
3785 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
3786 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
3792 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
3794 LLVMValueRef vtx_offset
=
3795 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
3796 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3797 LLVMValueRef stream_id
;
3799 /* Fetch the vertex stream ID. */
3800 if (ctx
->shader_info
->info
.so
.num_outputs
) {
3802 ac_unpack_param(&ctx
->ac
, ctx
->streamout_config
, 24, 2);
3804 stream_id
= ctx
->ac
.i32_0
;
3807 LLVMBasicBlockRef end_bb
;
3808 LLVMValueRef switch_inst
;
3810 end_bb
= LLVMAppendBasicBlockInContext(ctx
->ac
.context
,
3811 ctx
->main_function
, "end");
3812 switch_inst
= LLVMBuildSwitch(ctx
->ac
.builder
, stream_id
, end_bb
, 4);
3814 for (unsigned stream
= 0; stream
< 4; stream
++) {
3815 unsigned num_components
=
3816 ctx
->shader_info
->info
.gs
.num_stream_output_components
[stream
];
3817 LLVMBasicBlockRef bb
;
3820 if (!num_components
)
3823 if (stream
> 0 && !ctx
->shader_info
->info
.so
.num_outputs
)
3826 bb
= LLVMInsertBasicBlockInContext(ctx
->ac
.context
, end_bb
, "out");
3827 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
->ac
.i32
, stream
, 0), bb
);
3828 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, bb
);
3831 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3832 unsigned output_usage_mask
=
3833 ctx
->shader_info
->info
.gs
.output_usage_mask
[i
];
3834 unsigned output_stream
=
3835 ctx
->shader_info
->info
.gs
.output_streams
[i
];
3836 int length
= util_last_bit(output_usage_mask
);
3838 if (!(ctx
->output_mask
& (1ull << i
)) ||
3839 output_stream
!= stream
)
3842 for (unsigned j
= 0; j
< length
; j
++) {
3843 LLVMValueRef value
, soffset
;
3845 if (!(output_usage_mask
& (1 << j
)))
3848 soffset
= LLVMConstInt(ctx
->ac
.i32
,
3850 ctx
->gs_max_out_vertices
* 16 * 4, false);
3854 value
= ac_build_buffer_load(&ctx
->ac
,
3857 vtx_offset
, soffset
,
3858 0, 1, 1, true, false);
3860 LLVMTypeRef type
= LLVMGetAllocatedType(ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3861 if (ac_get_type_size(type
) == 2) {
3862 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
3863 value
= LLVMBuildTrunc(ctx
->ac
.builder
, value
, ctx
->ac
.i16
, "");
3866 LLVMBuildStore(ctx
->ac
.builder
,
3867 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3871 if (ctx
->shader_info
->info
.so
.num_outputs
)
3872 radv_emit_streamout(ctx
, stream
);
3875 handle_vs_outputs_post(ctx
, false, false,
3876 &ctx
->shader_info
->vs
.outinfo
);
3879 LLVMBuildBr(ctx
->ac
.builder
, end_bb
);
3882 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, end_bb
);
3886 radv_compile_gs_copy_shader(struct ac_llvm_compiler
*ac_llvm
,
3887 struct nir_shader
*geom_shader
,
3888 struct ac_shader_binary
*binary
,
3889 struct ac_shader_config
*config
,
3890 struct radv_shader_variant_info
*shader_info
,
3891 const struct radv_nir_compiler_options
*options
)
3893 struct radv_shader_context ctx
= {0};
3894 ctx
.options
= options
;
3895 ctx
.shader_info
= shader_info
;
3897 ac_llvm_context_init(&ctx
.ac
, options
->chip_class
, options
->family
);
3898 ctx
.context
= ctx
.ac
.context
;
3899 ctx
.ac
.module
= ac_create_module(ac_llvm
->tm
, ctx
.context
);
3901 ctx
.is_gs_copy_shader
= true;
3903 enum ac_float_mode float_mode
=
3904 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
3905 AC_FLOAT_MODE_DEFAULT
;
3907 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
3908 ctx
.stage
= MESA_SHADER_VERTEX
;
3910 radv_nir_shader_info_pass(geom_shader
, options
, &shader_info
->info
);
3912 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
3914 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
3915 ac_setup_rings(&ctx
);
3917 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
3918 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
3919 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
3920 variable
, MESA_SHADER_VERTEX
);
3923 ac_gs_copy_shader_emit(&ctx
);
3925 LLVMBuildRetVoid(ctx
.ac
.builder
);
3927 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, options
);
3929 ac_compile_llvm_module(ac_llvm
, ctx
.ac
.module
, binary
, config
, shader_info
,
3930 MESA_SHADER_VERTEX
, options
);