2 * Copyright © 2016 Bas Nieuwenhuizen
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "ac_nir_to_llvm.h"
25 #include "ac_llvm_build.h"
26 #include "ac_llvm_util.h"
27 #include "ac_binary.h"
30 #include "../vulkan/radv_descriptor_set.h"
31 #include "util/bitscan.h"
32 #include <llvm-c/Transforms/Scalar.h>
33 #include "ac_shader_abi.h"
34 #include "ac_shader_info.h"
35 #include "ac_shader_util.h"
36 #include "ac_exp_param.h"
38 enum radeon_llvm_calling_convention
{
39 RADEON_LLVM_AMDGPU_VS
= 87,
40 RADEON_LLVM_AMDGPU_GS
= 88,
41 RADEON_LLVM_AMDGPU_PS
= 89,
42 RADEON_LLVM_AMDGPU_CS
= 90,
43 RADEON_LLVM_AMDGPU_HS
= 93,
46 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
47 #define RADEON_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
49 struct nir_to_llvm_context
;
51 struct ac_nir_context
{
52 struct ac_llvm_context ac
;
53 struct ac_shader_abi
*abi
;
55 gl_shader_stage stage
;
57 struct hash_table
*defs
;
58 struct hash_table
*phis
;
59 struct hash_table
*vars
;
61 LLVMValueRef main_function
;
62 LLVMBasicBlockRef continue_block
;
63 LLVMBasicBlockRef break_block
;
65 LLVMValueRef outputs
[RADEON_LLVM_MAX_OUTPUTS
* 4];
70 struct nir_to_llvm_context
*nctx
; /* TODO get rid of this */
73 struct nir_to_llvm_context
{
74 struct ac_llvm_context ac
;
75 const struct ac_nir_compiler_options
*options
;
76 struct ac_shader_variant_info
*shader_info
;
77 struct ac_shader_abi abi
;
78 struct ac_nir_context
*nir
;
80 unsigned max_workgroup_size
;
81 LLVMContextRef context
;
83 LLVMValueRef main_function
;
85 LLVMValueRef descriptor_sets
[AC_UD_MAX_SETS
];
86 LLVMValueRef ring_offsets
;
88 LLVMValueRef vertex_buffers
;
89 LLVMValueRef rel_auto_id
;
90 LLVMValueRef vs_prim_id
;
91 LLVMValueRef ls_out_layout
;
92 LLVMValueRef es2gs_offset
;
94 LLVMValueRef tcs_offchip_layout
;
95 LLVMValueRef tcs_out_offsets
;
96 LLVMValueRef tcs_out_layout
;
97 LLVMValueRef tcs_in_layout
;
99 LLVMValueRef merged_wave_info
;
100 LLVMValueRef tess_factor_offset
;
101 LLVMValueRef tes_rel_patch_id
;
105 LLVMValueRef gsvs_ring_stride
;
106 LLVMValueRef gsvs_num_entries
;
107 LLVMValueRef gs2vs_offset
;
108 LLVMValueRef gs_wave_id
;
109 LLVMValueRef gs_vtx_offset
[6];
111 LLVMValueRef esgs_ring
;
112 LLVMValueRef gsvs_ring
;
113 LLVMValueRef hs_ring_tess_offchip
;
114 LLVMValueRef hs_ring_tess_factor
;
116 LLVMValueRef sample_pos_offset
;
117 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
118 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
120 gl_shader_stage stage
;
122 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
125 uint64_t output_mask
;
126 uint8_t num_output_clips
;
127 uint8_t num_output_culls
;
129 bool is_gs_copy_shader
;
130 LLVMValueRef gs_next_vertex
;
131 unsigned gs_max_out_vertices
;
133 unsigned tes_primitive_mode
;
134 uint64_t tess_outputs_written
;
135 uint64_t tess_patch_outputs_written
;
137 uint32_t tcs_patch_outputs_read
;
138 uint64_t tcs_outputs_read
;
141 static inline struct nir_to_llvm_context
*
142 nir_to_llvm_context_from_abi(struct ac_shader_abi
*abi
)
144 struct nir_to_llvm_context
*ctx
= NULL
;
145 return container_of(abi
, ctx
, abi
);
148 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
149 const nir_deref_var
*deref
,
150 enum ac_descriptor_type desc_type
,
151 const nir_tex_instr
*instr
,
152 bool image
, bool write
);
154 static unsigned radeon_llvm_reg_index_soa(unsigned index
, unsigned chan
)
156 return (index
* 4) + chan
;
159 static unsigned shader_io_get_unique_index(gl_varying_slot slot
)
161 /* handle patch indices separate */
162 if (slot
== VARYING_SLOT_TESS_LEVEL_OUTER
)
164 if (slot
== VARYING_SLOT_TESS_LEVEL_INNER
)
166 if (slot
>= VARYING_SLOT_PATCH0
&& slot
<= VARYING_SLOT_TESS_MAX
)
167 return 2 + (slot
- VARYING_SLOT_PATCH0
);
169 if (slot
== VARYING_SLOT_POS
)
171 if (slot
== VARYING_SLOT_PSIZ
)
173 if (slot
== VARYING_SLOT_CLIP_DIST0
)
175 /* 3 is reserved for clip dist as well */
176 if (slot
>= VARYING_SLOT_VAR0
&& slot
<= VARYING_SLOT_VAR31
)
177 return 4 + (slot
- VARYING_SLOT_VAR0
);
178 unreachable("illegal slot in get unique index\n");
181 static void set_llvm_calling_convention(LLVMValueRef func
,
182 gl_shader_stage stage
)
184 enum radeon_llvm_calling_convention calling_conv
;
187 case MESA_SHADER_VERTEX
:
188 case MESA_SHADER_TESS_EVAL
:
189 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
191 case MESA_SHADER_GEOMETRY
:
192 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
194 case MESA_SHADER_TESS_CTRL
:
195 calling_conv
= HAVE_LLVM
>= 0x0500 ? RADEON_LLVM_AMDGPU_HS
: RADEON_LLVM_AMDGPU_VS
;
197 case MESA_SHADER_FRAGMENT
:
198 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
200 case MESA_SHADER_COMPUTE
:
201 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
204 unreachable("Unhandle shader type");
207 LLVMSetFunctionCallConv(func
, calling_conv
);
212 LLVMTypeRef types
[MAX_ARGS
];
213 LLVMValueRef
*assign
[MAX_ARGS
];
214 unsigned array_params_mask
;
217 uint8_t num_sgprs_used
;
218 uint8_t num_vgprs_used
;
221 enum ac_arg_regfile
{
227 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
228 LLVMValueRef
*param_ptr
)
230 assert(info
->count
< MAX_ARGS
);
232 info
->assign
[info
->count
] = param_ptr
;
233 info
->types
[info
->count
] = type
;
236 if (regfile
== ARG_SGPR
) {
237 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
240 assert(regfile
== ARG_VGPR
);
241 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
246 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
248 info
->array_params_mask
|= (1 << info
->count
);
249 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
252 static void assign_arguments(LLVMValueRef main_function
,
253 struct arg_info
*info
)
256 for (i
= 0; i
< info
->count
; i
++) {
258 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
263 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
264 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
265 unsigned num_return_elems
,
266 struct arg_info
*args
,
267 unsigned max_workgroup_size
,
270 LLVMTypeRef main_function_type
, ret_type
;
271 LLVMBasicBlockRef main_function_body
;
273 if (num_return_elems
)
274 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
275 num_return_elems
, true);
277 ret_type
= LLVMVoidTypeInContext(ctx
);
279 /* Setup the function */
281 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
282 LLVMValueRef main_function
=
283 LLVMAddFunction(module
, "main", main_function_type
);
285 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
286 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
288 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
289 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
290 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
292 if (args
->array_params_mask
& (1 << i
)) {
293 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
294 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
295 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
299 if (max_workgroup_size
) {
300 ac_llvm_add_target_dep_function_attr(main_function
,
301 "amdgpu-max-work-group-size",
305 /* These were copied from some LLVM test. */
306 LLVMAddTargetDependentFunctionAttr(main_function
,
307 "less-precise-fpmad",
309 LLVMAddTargetDependentFunctionAttr(main_function
,
312 LLVMAddTargetDependentFunctionAttr(main_function
,
315 LLVMAddTargetDependentFunctionAttr(main_function
,
318 LLVMAddTargetDependentFunctionAttr(main_function
,
319 "no-signed-zeros-fp-math",
322 return main_function
;
325 static LLVMValueRef
unpack_param(struct ac_llvm_context
*ctx
,
326 LLVMValueRef param
, unsigned rshift
,
329 LLVMValueRef value
= param
;
331 value
= LLVMBuildLShr(ctx
->builder
, value
,
332 LLVMConstInt(ctx
->i32
, rshift
, false), "");
334 if (rshift
+ bitwidth
< 32) {
335 unsigned mask
= (1 << bitwidth
) - 1;
336 value
= LLVMBuildAnd(ctx
->builder
, value
,
337 LLVMConstInt(ctx
->i32
, mask
, false), "");
342 static LLVMValueRef
get_rel_patch_id(struct nir_to_llvm_context
*ctx
)
344 switch (ctx
->stage
) {
345 case MESA_SHADER_TESS_CTRL
:
346 return unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
347 case MESA_SHADER_TESS_EVAL
:
348 return ctx
->tes_rel_patch_id
;
351 unreachable("Illegal stage");
355 /* Tessellation shaders pass outputs to the next shader using LDS.
357 * LS outputs = TCS inputs
358 * TCS outputs = TES inputs
361 * - TCS inputs for patch 0
362 * - TCS inputs for patch 1
363 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
365 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
366 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
367 * - TCS outputs for patch 1
368 * - Per-patch TCS outputs for patch 1
369 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
370 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
373 * All three shaders VS(LS), TCS, TES share the same LDS space.
376 get_tcs_in_patch_stride(struct nir_to_llvm_context
*ctx
)
378 if (ctx
->stage
== MESA_SHADER_VERTEX
)
379 return unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 0, 13);
380 else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
381 return unpack_param(&ctx
->ac
, ctx
->tcs_in_layout
, 0, 13);
389 get_tcs_out_patch_stride(struct nir_to_llvm_context
*ctx
)
391 return unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 0, 13);
395 get_tcs_out_patch0_offset(struct nir_to_llvm_context
*ctx
)
397 return LLVMBuildMul(ctx
->ac
.builder
,
398 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 0, 16),
399 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
403 get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context
*ctx
)
405 return LLVMBuildMul(ctx
->ac
.builder
,
406 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 16, 16),
407 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
411 get_tcs_in_current_patch_offset(struct nir_to_llvm_context
*ctx
)
413 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
414 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
416 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
420 get_tcs_out_current_patch_offset(struct nir_to_llvm_context
*ctx
)
422 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
423 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
424 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
426 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_offset
,
427 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
433 get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context
*ctx
)
435 LLVMValueRef patch0_patch_data_offset
=
436 get_tcs_out_patch0_patch_data_offset(ctx
);
437 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
438 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
440 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_patch_data_offset
,
441 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
447 set_loc(struct ac_userdata_info
*ud_info
, uint8_t *sgpr_idx
, uint8_t num_sgprs
,
448 uint32_t indirect_offset
)
450 ud_info
->sgpr_idx
= *sgpr_idx
;
451 ud_info
->num_sgprs
= num_sgprs
;
452 ud_info
->indirect
= indirect_offset
> 0;
453 ud_info
->indirect_offset
= indirect_offset
;
454 *sgpr_idx
+= num_sgprs
;
458 set_loc_shader(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
461 struct ac_userdata_info
*ud_info
=
462 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
465 set_loc(ud_info
, sgpr_idx
, num_sgprs
, 0);
469 set_loc_desc(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
470 uint32_t indirect_offset
)
472 struct ac_userdata_info
*ud_info
=
473 &ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[idx
];
476 set_loc(ud_info
, sgpr_idx
, 2, indirect_offset
);
479 struct user_sgpr_info
{
480 bool need_ring_offsets
;
482 bool indirect_all_descriptor_sets
;
485 static bool needs_view_index_sgpr(struct nir_to_llvm_context
*ctx
,
486 gl_shader_stage stage
)
489 case MESA_SHADER_VERTEX
:
490 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
491 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
494 case MESA_SHADER_TESS_EVAL
:
495 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
498 case MESA_SHADER_GEOMETRY
:
499 case MESA_SHADER_TESS_CTRL
:
500 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
510 count_vs_user_sgprs(struct nir_to_llvm_context
*ctx
)
514 count
+= ctx
->shader_info
->info
.vs
.has_vertex_buffers
? 2 : 0;
515 count
+= ctx
->shader_info
->info
.vs
.needs_draw_id
? 3 : 2;
520 static void allocate_user_sgprs(struct nir_to_llvm_context
*ctx
,
521 gl_shader_stage stage
,
522 bool has_previous_stage
,
523 gl_shader_stage previous_stage
,
524 bool needs_view_index
,
525 struct user_sgpr_info
*user_sgpr_info
)
527 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
529 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
530 if (stage
== MESA_SHADER_GEOMETRY
||
531 stage
== MESA_SHADER_VERTEX
||
532 stage
== MESA_SHADER_TESS_CTRL
||
533 stage
== MESA_SHADER_TESS_EVAL
||
534 ctx
->is_gs_copy_shader
)
535 user_sgpr_info
->need_ring_offsets
= true;
537 if (stage
== MESA_SHADER_FRAGMENT
&&
538 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
539 user_sgpr_info
->need_ring_offsets
= true;
541 /* 2 user sgprs will nearly always be allocated for scratch/rings */
542 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
543 user_sgpr_info
->sgpr_count
+= 2;
547 case MESA_SHADER_COMPUTE
:
548 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
549 user_sgpr_info
->sgpr_count
+= 3;
551 case MESA_SHADER_FRAGMENT
:
552 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
554 case MESA_SHADER_VERTEX
:
555 if (!ctx
->is_gs_copy_shader
)
556 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
557 if (ctx
->options
->key
.vs
.as_ls
)
558 user_sgpr_info
->sgpr_count
++;
560 case MESA_SHADER_TESS_CTRL
:
561 if (has_previous_stage
) {
562 if (previous_stage
== MESA_SHADER_VERTEX
)
563 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
564 user_sgpr_info
->sgpr_count
++;
566 user_sgpr_info
->sgpr_count
+= 4;
568 case MESA_SHADER_TESS_EVAL
:
569 user_sgpr_info
->sgpr_count
+= 1;
571 case MESA_SHADER_GEOMETRY
:
572 if (has_previous_stage
) {
573 if (previous_stage
== MESA_SHADER_VERTEX
) {
574 user_sgpr_info
->sgpr_count
+= count_vs_user_sgprs(ctx
);
576 user_sgpr_info
->sgpr_count
++;
579 user_sgpr_info
->sgpr_count
+= 2;
585 if (needs_view_index
)
586 user_sgpr_info
->sgpr_count
++;
588 if (ctx
->shader_info
->info
.loads_push_constants
)
589 user_sgpr_info
->sgpr_count
+= 2;
591 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
? 32 : 16;
592 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_info
->sgpr_count
;
594 if (remaining_sgprs
/ 2 < util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
)) {
595 user_sgpr_info
->sgpr_count
+= 2;
596 user_sgpr_info
->indirect_all_descriptor_sets
= true;
598 user_sgpr_info
->sgpr_count
+= util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
) * 2;
603 declare_global_input_sgprs(struct nir_to_llvm_context
*ctx
,
604 gl_shader_stage stage
,
605 bool has_previous_stage
,
606 gl_shader_stage previous_stage
,
607 const struct user_sgpr_info
*user_sgpr_info
,
608 struct arg_info
*args
,
609 LLVMValueRef
*desc_sets
)
611 LLVMTypeRef type
= ac_array_in_const_addr_space(ctx
->ac
.i8
);
612 unsigned num_sets
= ctx
->options
->layout
?
613 ctx
->options
->layout
->num_sets
: 0;
614 unsigned stage_mask
= 1 << stage
;
616 if (has_previous_stage
)
617 stage_mask
|= 1 << previous_stage
;
619 /* 1 for each descriptor set */
620 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
621 for (unsigned i
= 0; i
< num_sets
; ++i
) {
622 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
623 add_array_arg(args
, type
,
624 &ctx
->descriptor_sets
[i
]);
628 add_array_arg(args
, ac_array_in_const_addr_space(type
), desc_sets
);
631 if (ctx
->shader_info
->info
.loads_push_constants
) {
632 /* 1 for push constants and dynamic descriptors */
633 add_array_arg(args
, type
, &ctx
->abi
.push_constants
);
638 declare_vs_specific_input_sgprs(struct nir_to_llvm_context
*ctx
,
639 gl_shader_stage stage
,
640 bool has_previous_stage
,
641 gl_shader_stage previous_stage
,
642 struct arg_info
*args
)
644 if (!ctx
->is_gs_copy_shader
&&
645 (stage
== MESA_SHADER_VERTEX
||
646 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
647 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
648 add_arg(args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
649 &ctx
->vertex_buffers
);
651 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
652 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
653 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
654 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
660 declare_vs_input_vgprs(struct nir_to_llvm_context
*ctx
, struct arg_info
*args
)
662 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
663 if (!ctx
->is_gs_copy_shader
) {
664 if (ctx
->options
->key
.vs
.as_ls
) {
665 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
666 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
668 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
669 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
671 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
676 declare_tes_input_vgprs(struct nir_to_llvm_context
*ctx
, struct arg_info
*args
)
678 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
679 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
680 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
681 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
685 set_global_input_locs(struct nir_to_llvm_context
*ctx
, gl_shader_stage stage
,
686 bool has_previous_stage
, gl_shader_stage previous_stage
,
687 const struct user_sgpr_info
*user_sgpr_info
,
688 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
690 unsigned num_sets
= ctx
->options
->layout
?
691 ctx
->options
->layout
->num_sets
: 0;
692 unsigned stage_mask
= 1 << stage
;
694 if (has_previous_stage
)
695 stage_mask
|= 1 << previous_stage
;
697 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
698 for (unsigned i
= 0; i
< num_sets
; ++i
) {
699 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
700 set_loc_desc(ctx
, i
, user_sgpr_idx
, 0);
702 ctx
->descriptor_sets
[i
] = NULL
;
705 set_loc_shader(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
708 for (unsigned i
= 0; i
< num_sets
; ++i
) {
709 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
710 set_loc_desc(ctx
, i
, user_sgpr_idx
, i
* 8);
711 ctx
->descriptor_sets
[i
] =
712 ac_build_load_to_sgpr(&ctx
->ac
,
714 LLVMConstInt(ctx
->ac
.i32
, i
, false));
717 ctx
->descriptor_sets
[i
] = NULL
;
719 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
722 if (ctx
->shader_info
->info
.loads_push_constants
) {
723 set_loc_shader(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
, 2);
728 set_vs_specific_input_locs(struct nir_to_llvm_context
*ctx
,
729 gl_shader_stage stage
, bool has_previous_stage
,
730 gl_shader_stage previous_stage
,
731 uint8_t *user_sgpr_idx
)
733 if (!ctx
->is_gs_copy_shader
&&
734 (stage
== MESA_SHADER_VERTEX
||
735 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
736 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
737 set_loc_shader(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
742 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
745 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
746 user_sgpr_idx
, vs_num
);
750 static void create_function(struct nir_to_llvm_context
*ctx
,
751 gl_shader_stage stage
,
752 bool has_previous_stage
,
753 gl_shader_stage previous_stage
)
755 uint8_t user_sgpr_idx
;
756 struct user_sgpr_info user_sgpr_info
;
757 struct arg_info args
= {};
758 LLVMValueRef desc_sets
;
759 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
760 allocate_user_sgprs(ctx
, stage
, has_previous_stage
,
761 previous_stage
, needs_view_index
, &user_sgpr_info
);
763 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
764 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
769 case MESA_SHADER_COMPUTE
:
770 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
771 previous_stage
, &user_sgpr_info
,
774 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
775 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
776 &ctx
->abi
.num_work_groups
);
779 for (int i
= 0; i
< 3; i
++) {
780 ctx
->abi
.workgroup_ids
[i
] = NULL
;
781 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
782 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
783 &ctx
->abi
.workgroup_ids
[i
]);
787 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
788 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.tg_size
);
789 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
790 &ctx
->abi
.local_invocation_ids
);
792 case MESA_SHADER_VERTEX
:
793 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
794 previous_stage
, &user_sgpr_info
,
796 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
797 previous_stage
, &args
);
799 if (needs_view_index
)
800 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
801 &ctx
->abi
.view_index
);
802 if (ctx
->options
->key
.vs
.as_es
)
803 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
805 else if (ctx
->options
->key
.vs
.as_ls
)
806 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
807 &ctx
->ls_out_layout
);
809 declare_vs_input_vgprs(ctx
, &args
);
811 case MESA_SHADER_TESS_CTRL
:
812 if (has_previous_stage
) {
813 // First 6 system regs
814 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
815 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
816 &ctx
->merged_wave_info
);
817 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
818 &ctx
->tess_factor_offset
);
820 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
821 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
822 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
824 declare_global_input_sgprs(ctx
, stage
,
827 &user_sgpr_info
, &args
,
829 declare_vs_specific_input_sgprs(ctx
, stage
,
831 previous_stage
, &args
);
833 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
834 &ctx
->ls_out_layout
);
836 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
837 &ctx
->tcs_offchip_layout
);
838 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
839 &ctx
->tcs_out_offsets
);
840 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
841 &ctx
->tcs_out_layout
);
842 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
843 &ctx
->tcs_in_layout
);
844 if (needs_view_index
)
845 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
846 &ctx
->abi
.view_index
);
848 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
849 &ctx
->abi
.tcs_patch_id
);
850 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
851 &ctx
->abi
.tcs_rel_ids
);
853 declare_vs_input_vgprs(ctx
, &args
);
855 declare_global_input_sgprs(ctx
, stage
,
858 &user_sgpr_info
, &args
,
861 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
862 &ctx
->tcs_offchip_layout
);
863 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
864 &ctx
->tcs_out_offsets
);
865 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
866 &ctx
->tcs_out_layout
);
867 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
868 &ctx
->tcs_in_layout
);
869 if (needs_view_index
)
870 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
871 &ctx
->abi
.view_index
);
873 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
874 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
875 &ctx
->tess_factor_offset
);
876 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
877 &ctx
->abi
.tcs_patch_id
);
878 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
879 &ctx
->abi
.tcs_rel_ids
);
882 case MESA_SHADER_TESS_EVAL
:
883 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
884 previous_stage
, &user_sgpr_info
,
887 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->tcs_offchip_layout
);
888 if (needs_view_index
)
889 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
890 &ctx
->abi
.view_index
);
892 if (ctx
->options
->key
.tes
.as_es
) {
893 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
894 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
895 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
898 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
899 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
901 declare_tes_input_vgprs(ctx
, &args
);
903 case MESA_SHADER_GEOMETRY
:
904 if (has_previous_stage
) {
905 // First 6 system regs
906 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
908 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
909 &ctx
->merged_wave_info
);
910 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
912 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
913 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
914 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
916 declare_global_input_sgprs(ctx
, stage
,
919 &user_sgpr_info
, &args
,
922 if (previous_stage
== MESA_SHADER_TESS_EVAL
) {
923 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
924 &ctx
->tcs_offchip_layout
);
926 declare_vs_specific_input_sgprs(ctx
, stage
,
932 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
933 &ctx
->gsvs_ring_stride
);
934 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
935 &ctx
->gsvs_num_entries
);
936 if (needs_view_index
)
937 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
938 &ctx
->abi
.view_index
);
940 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
941 &ctx
->gs_vtx_offset
[0]);
942 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
943 &ctx
->gs_vtx_offset
[2]);
944 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
945 &ctx
->abi
.gs_prim_id
);
946 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
947 &ctx
->abi
.gs_invocation_id
);
948 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
949 &ctx
->gs_vtx_offset
[4]);
951 if (previous_stage
== MESA_SHADER_VERTEX
) {
952 declare_vs_input_vgprs(ctx
, &args
);
954 declare_tes_input_vgprs(ctx
, &args
);
957 declare_global_input_sgprs(ctx
, stage
,
960 &user_sgpr_info
, &args
,
963 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
964 &ctx
->gsvs_ring_stride
);
965 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
966 &ctx
->gsvs_num_entries
);
967 if (needs_view_index
)
968 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
969 &ctx
->abi
.view_index
);
971 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
972 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
973 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
974 &ctx
->gs_vtx_offset
[0]);
975 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
976 &ctx
->gs_vtx_offset
[1]);
977 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
978 &ctx
->abi
.gs_prim_id
);
979 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
980 &ctx
->gs_vtx_offset
[2]);
981 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
982 &ctx
->gs_vtx_offset
[3]);
983 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
984 &ctx
->gs_vtx_offset
[4]);
985 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
986 &ctx
->gs_vtx_offset
[5]);
987 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
988 &ctx
->abi
.gs_invocation_id
);
991 case MESA_SHADER_FRAGMENT
:
992 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
993 previous_stage
, &user_sgpr_info
,
996 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
997 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
998 &ctx
->sample_pos_offset
);
1000 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1001 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1002 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1003 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1004 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1005 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1006 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1007 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1008 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1009 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1010 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1011 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1012 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1013 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1014 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1015 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1016 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1019 unreachable("Shader stage not implemented");
1022 ctx
->main_function
= create_llvm_function(
1023 ctx
->context
, ctx
->module
, ctx
->ac
.builder
, NULL
, 0, &args
,
1024 ctx
->max_workgroup_size
,
1025 ctx
->options
->unsafe_math
);
1026 set_llvm_calling_convention(ctx
->main_function
, stage
);
1029 ctx
->shader_info
->num_input_vgprs
= 0;
1030 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1032 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1034 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1035 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1037 assign_arguments(ctx
->main_function
, &args
);
1041 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1042 set_loc_shader(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1044 if (ctx
->options
->supports_spill
) {
1045 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1046 LLVMPointerType(ctx
->ac
.i8
, AC_CONST_ADDR_SPACE
),
1047 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1048 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
1049 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1053 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1054 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1055 if (has_previous_stage
)
1058 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1059 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1062 case MESA_SHADER_COMPUTE
:
1063 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1064 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1068 case MESA_SHADER_VERTEX
:
1069 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1070 previous_stage
, &user_sgpr_idx
);
1071 if (ctx
->abi
.view_index
)
1072 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1073 if (ctx
->options
->key
.vs
.as_ls
) {
1074 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1077 if (ctx
->options
->key
.vs
.as_ls
)
1078 ac_declare_lds_as_pointer(&ctx
->ac
);
1080 case MESA_SHADER_TESS_CTRL
:
1081 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1082 previous_stage
, &user_sgpr_idx
);
1083 if (has_previous_stage
)
1084 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1086 set_loc_shader(ctx
, AC_UD_TCS_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 4);
1087 if (ctx
->abi
.view_index
)
1088 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1089 ac_declare_lds_as_pointer(&ctx
->ac
);
1091 case MESA_SHADER_TESS_EVAL
:
1092 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 1);
1093 if (ctx
->abi
.view_index
)
1094 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1096 case MESA_SHADER_GEOMETRY
:
1097 if (has_previous_stage
) {
1098 if (previous_stage
== MESA_SHADER_VERTEX
)
1099 set_vs_specific_input_locs(ctx
, stage
,
1104 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
,
1107 set_loc_shader(ctx
, AC_UD_GS_VS_RING_STRIDE_ENTRIES
,
1109 if (ctx
->abi
.view_index
)
1110 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1111 if (has_previous_stage
)
1112 ac_declare_lds_as_pointer(&ctx
->ac
);
1114 case MESA_SHADER_FRAGMENT
:
1115 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
1116 set_loc_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
,
1121 unreachable("Shader stage not implemented");
1124 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1127 static LLVMValueRef
trim_vector(struct ac_llvm_context
*ctx
,
1128 LLVMValueRef value
, unsigned count
)
1130 unsigned num_components
= ac_get_llvm_num_components(value
);
1131 if (count
== num_components
)
1134 LLVMValueRef masks
[] = {
1135 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
1136 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false)};
1139 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
1142 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
1143 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
1147 build_store_values_extended(struct ac_llvm_context
*ac
,
1148 LLVMValueRef
*values
,
1149 unsigned value_count
,
1150 unsigned value_stride
,
1153 LLVMBuilderRef builder
= ac
->builder
;
1156 for (i
= 0; i
< value_count
; i
++) {
1157 LLVMValueRef ptr
= values
[i
* value_stride
];
1158 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
1159 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
1160 LLVMBuildStore(builder
, value
, ptr
);
1164 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
1165 const nir_ssa_def
*def
)
1167 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
1168 if (def
->num_components
> 1) {
1169 type
= LLVMVectorType(type
, def
->num_components
);
1174 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
1177 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, src
.ssa
);
1178 return (LLVMValueRef
)entry
->data
;
1182 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
1183 const struct nir_block
*b
)
1185 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
1186 return (LLVMBasicBlockRef
)entry
->data
;
1189 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
1191 unsigned num_components
)
1193 LLVMValueRef value
= get_src(ctx
, src
.src
);
1194 bool need_swizzle
= false;
1197 LLVMTypeRef type
= LLVMTypeOf(value
);
1198 unsigned src_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
1199 ? LLVMGetVectorSize(type
)
1202 for (unsigned i
= 0; i
< num_components
; ++i
) {
1203 assert(src
.swizzle
[i
] < src_components
);
1204 if (src
.swizzle
[i
] != i
)
1205 need_swizzle
= true;
1208 if (need_swizzle
|| num_components
!= src_components
) {
1209 LLVMValueRef masks
[] = {
1210 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
1211 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
1212 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
1213 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
1215 if (src_components
> 1 && num_components
== 1) {
1216 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1218 } else if (src_components
== 1 && num_components
> 1) {
1219 LLVMValueRef values
[] = {value
, value
, value
, value
};
1220 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
1222 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
1223 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
1227 assert(!src
.negate
);
1232 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
1233 LLVMIntPredicate pred
, LLVMValueRef src0
,
1236 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
1237 return LLVMBuildSelect(ctx
->builder
, result
,
1238 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1242 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
1243 LLVMRealPredicate pred
, LLVMValueRef src0
,
1246 LLVMValueRef result
;
1247 src0
= ac_to_float(ctx
, src0
);
1248 src1
= ac_to_float(ctx
, src1
);
1249 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
1250 return LLVMBuildSelect(ctx
->builder
, result
,
1251 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1255 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
1257 LLVMTypeRef result_type
,
1261 LLVMValueRef params
[] = {
1262 ac_to_float(ctx
, src0
),
1265 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1266 ac_get_elem_bits(ctx
, result_type
));
1267 assert(length
< sizeof(name
));
1268 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
1271 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
1273 LLVMTypeRef result_type
,
1274 LLVMValueRef src0
, LLVMValueRef src1
)
1277 LLVMValueRef params
[] = {
1278 ac_to_float(ctx
, src0
),
1279 ac_to_float(ctx
, src1
),
1282 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1283 ac_get_elem_bits(ctx
, result_type
));
1284 assert(length
< sizeof(name
));
1285 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
1288 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
1290 LLVMTypeRef result_type
,
1291 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1294 LLVMValueRef params
[] = {
1295 ac_to_float(ctx
, src0
),
1296 ac_to_float(ctx
, src1
),
1297 ac_to_float(ctx
, src2
),
1300 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1301 ac_get_elem_bits(ctx
, result_type
));
1302 assert(length
< sizeof(name
));
1303 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
1306 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
1307 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1309 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
1311 return LLVMBuildSelect(ctx
->builder
, v
, src1
, src2
, "");
1314 static LLVMValueRef
emit_minmax_int(struct ac_llvm_context
*ctx
,
1315 LLVMIntPredicate pred
,
1316 LLVMValueRef src0
, LLVMValueRef src1
)
1318 return LLVMBuildSelect(ctx
->builder
,
1319 LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, ""),
1324 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
1327 return emit_minmax_int(ctx
, LLVMIntSGT
, src0
,
1328 LLVMBuildNeg(ctx
->builder
, src0
, ""));
1331 static LLVMValueRef
emit_fsign(struct ac_llvm_context
*ctx
,
1335 LLVMValueRef cmp
, val
, zero
, one
;
1338 if (bitsize
== 32) {
1348 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
1349 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
1350 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
1351 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
1355 static LLVMValueRef
emit_isign(struct ac_llvm_context
*ctx
,
1356 LLVMValueRef src0
, unsigned bitsize
)
1358 LLVMValueRef cmp
, val
, zero
, one
;
1361 if (bitsize
== 32) {
1371 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
1372 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
1373 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
1374 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
1378 static LLVMValueRef
emit_ffract(struct ac_llvm_context
*ctx
,
1379 LLVMValueRef src0
, unsigned bitsize
)
1384 if (bitsize
== 32) {
1385 intr
= "llvm.floor.f32";
1388 intr
= "llvm.floor.f64";
1392 LLVMValueRef fsrc0
= ac_to_float(ctx
, src0
);
1393 LLVMValueRef params
[] = {
1396 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
1397 AC_FUNC_ATTR_READNONE
);
1398 return LLVMBuildFSub(ctx
->builder
, fsrc0
, floor
, "");
1401 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
1403 LLVMValueRef src0
, LLVMValueRef src1
)
1405 LLVMTypeRef ret_type
;
1406 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
1408 LLVMValueRef params
[] = { src0
, src1
};
1409 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
1412 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
1413 params
, 2, AC_FUNC_ATTR_READNONE
);
1415 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
1416 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
1420 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
1423 return LLVMBuildAnd(ctx
->builder
, src0
, LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""), "");
1426 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
1429 src0
= ac_to_float(ctx
, src0
);
1430 return LLVMBuildSExt(ctx
->builder
,
1431 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, ctx
->f32_0
, ""),
1435 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
1439 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
1444 return LLVMBuildZExt(ctx
->builder
, result
, ctx
->i64
, "");
1447 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
1450 return LLVMBuildSExt(ctx
->builder
,
1451 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, ctx
->i32_0
, ""),
1455 static LLVMValueRef
emit_f2f16(struct ac_llvm_context
*ctx
,
1458 LLVMValueRef result
;
1459 LLVMValueRef cond
= NULL
;
1461 src0
= ac_to_float(ctx
, src0
);
1462 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->f16
, "");
1464 if (ctx
->chip_class
>= VI
) {
1465 LLVMValueRef args
[2];
1466 /* Check if the result is a denormal - and flush to 0 if so. */
1468 args
[1] = LLVMConstInt(ctx
->i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
1469 cond
= ac_build_intrinsic(ctx
, "llvm.amdgcn.class.f16", ctx
->i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
1472 /* need to convert back up to f32 */
1473 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f32
, "");
1475 if (ctx
->chip_class
>= VI
)
1476 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
1479 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
1480 * so compare the result and flush to 0 if it's smaller.
1482 LLVMValueRef temp
, cond2
;
1483 temp
= emit_intrin_1f_param(ctx
, "llvm.fabs", ctx
->f32
, result
);
1484 cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUGT
,
1485 LLVMBuildBitCast(ctx
->builder
, LLVMConstInt(ctx
->i32
, 0x38800000, false), ctx
->f32
, ""),
1487 cond2
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
1488 temp
, ctx
->f32_0
, "");
1489 cond
= LLVMBuildAnd(ctx
->builder
, cond
, cond2
, "");
1490 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
1495 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
1496 LLVMValueRef src0
, LLVMValueRef src1
)
1498 LLVMValueRef dst64
, result
;
1499 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
1500 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
1502 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1503 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1504 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1508 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
1509 LLVMValueRef src0
, LLVMValueRef src1
)
1511 LLVMValueRef dst64
, result
;
1512 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
1513 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
1515 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1516 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1517 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1521 static LLVMValueRef
emit_bitfield_extract(struct ac_llvm_context
*ctx
,
1523 const LLVMValueRef srcs
[3])
1525 LLVMValueRef result
;
1526 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
1528 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
1529 result
= LLVMBuildSelect(ctx
->builder
, icond
, srcs
[0], result
, "");
1533 static LLVMValueRef
emit_bitfield_insert(struct ac_llvm_context
*ctx
,
1534 LLVMValueRef src0
, LLVMValueRef src1
,
1535 LLVMValueRef src2
, LLVMValueRef src3
)
1537 LLVMValueRef bfi_args
[3], result
;
1539 bfi_args
[0] = LLVMBuildShl(ctx
->builder
,
1540 LLVMBuildSub(ctx
->builder
,
1541 LLVMBuildShl(ctx
->builder
,
1546 bfi_args
[1] = LLVMBuildShl(ctx
->builder
, src1
, src2
, "");
1549 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, src3
, LLVMConstInt(ctx
->i32
, 32, false), "");
1552 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
1553 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
1555 result
= LLVMBuildXor(ctx
->builder
, bfi_args
[2],
1556 LLVMBuildAnd(ctx
->builder
, bfi_args
[0],
1557 LLVMBuildXor(ctx
->builder
, bfi_args
[1], bfi_args
[2], ""), ""), "");
1559 result
= LLVMBuildSelect(ctx
->builder
, icond
, src1
, result
, "");
1563 static LLVMValueRef
emit_pack_half_2x16(struct ac_llvm_context
*ctx
,
1566 LLVMValueRef comp
[2];
1568 src0
= ac_to_float(ctx
, src0
);
1569 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
1570 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
1572 return ac_build_cvt_pkrtz_f16(ctx
, comp
);
1575 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
1578 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
1579 LLVMValueRef temps
[2], result
, val
;
1582 for (i
= 0; i
< 2; i
++) {
1583 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
1584 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
1585 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
1586 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
1589 result
= LLVMBuildInsertElement(ctx
->builder
, LLVMGetUndef(ctx
->v2f32
), temps
[0],
1591 result
= LLVMBuildInsertElement(ctx
->builder
, result
, temps
[1],
1596 static LLVMValueRef
emit_ddxy(struct ac_nir_context
*ctx
,
1602 LLVMValueRef result
;
1604 if (op
== nir_op_fddx_fine
|| op
== nir_op_fddx
)
1605 mask
= AC_TID_MASK_LEFT
;
1606 else if (op
== nir_op_fddy_fine
|| op
== nir_op_fddy
)
1607 mask
= AC_TID_MASK_TOP
;
1609 mask
= AC_TID_MASK_TOP_LEFT
;
1611 /* for DDX we want to next X pixel, DDY next Y pixel. */
1612 if (op
== nir_op_fddx_fine
||
1613 op
== nir_op_fddx_coarse
||
1619 result
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, src0
);
1624 * this takes an I,J coordinate pair,
1625 * and works out the X and Y derivatives.
1626 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
1628 static LLVMValueRef
emit_ddxy_interp(
1629 struct ac_nir_context
*ctx
,
1630 LLVMValueRef interp_ij
)
1632 LLVMValueRef result
[4], a
;
1635 for (i
= 0; i
< 2; i
++) {
1636 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
1637 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1638 result
[i
] = emit_ddxy(ctx
, nir_op_fddx
, a
);
1639 result
[2+i
] = emit_ddxy(ctx
, nir_op_fddy
, a
);
1641 return ac_build_gather_values(&ctx
->ac
, result
, 4);
1644 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
1646 LLVMValueRef src
[4], result
= NULL
;
1647 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
1648 unsigned src_components
;
1649 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
1651 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
1652 switch (instr
->op
) {
1658 case nir_op_pack_half_2x16
:
1661 case nir_op_unpack_half_2x16
:
1665 src_components
= num_components
;
1668 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1669 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
1671 switch (instr
->op
) {
1677 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1678 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
1681 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
1684 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
1687 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1690 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1691 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1692 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1695 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1696 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1697 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1700 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1703 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1706 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1709 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
1712 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1713 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1714 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1715 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1716 ac_to_float_type(&ctx
->ac
, def_type
), result
);
1717 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
1718 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
1721 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1722 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1723 result
= LLVMBuildFRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1726 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1729 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1732 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1735 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1736 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1737 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1740 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1741 result
= ac_build_fdiv(&ctx
->ac
, instr
->dest
.dest
.ssa
.bit_size
== 32 ? ctx
->ac
.f32_1
: ctx
->ac
.f64_1
,
1745 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
1748 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
1751 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
1754 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0],
1755 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1756 LLVMTypeOf(src
[0]), ""),
1760 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0],
1761 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1762 LLVMTypeOf(src
[0]), ""),
1766 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0],
1767 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1768 LLVMTypeOf(src
[0]), ""),
1772 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1775 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
1778 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
1781 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
1784 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1787 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
1790 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUEQ
, src
[0], src
[1]);
1793 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
1796 result
= emit_float_cmp(&ctx
->ac
, LLVMRealULT
, src
[0], src
[1]);
1799 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUGE
, src
[0], src
[1]);
1802 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1803 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1806 result
= emit_iabs(&ctx
->ac
, src
[0]);
1809 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, src
[0], src
[1]);
1812 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1815 result
= emit_minmax_int(&ctx
->ac
, LLVMIntUGT
, src
[0], src
[1]);
1818 result
= emit_minmax_int(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1821 result
= emit_isign(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1824 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1825 result
= emit_fsign(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1828 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1829 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1832 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
1833 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1836 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
1837 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1839 case nir_op_fround_even
:
1840 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
1841 ac_to_float_type(&ctx
->ac
, def_type
),src
[0]);
1844 result
= emit_ffract(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1847 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
1848 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1851 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
1852 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1855 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1856 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1859 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
1860 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1863 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
1864 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1867 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1868 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1869 result
= ac_build_fdiv(&ctx
->ac
, instr
->dest
.dest
.ssa
.bit_size
== 32 ? ctx
->ac
.f32_1
: ctx
->ac
.f64_1
,
1873 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
1874 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1877 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1878 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1879 if (ctx
->ac
.chip_class
< GFX9
&&
1880 instr
->dest
.dest
.ssa
.bit_size
== 32) {
1881 /* Only pre-GFX9 chips do not flush denorms. */
1882 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1883 ac_to_float_type(&ctx
->ac
, def_type
),
1888 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1889 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1890 if (ctx
->ac
.chip_class
< GFX9
&&
1891 instr
->dest
.dest
.ssa
.bit_size
== 32) {
1892 /* Only pre-GFX9 chips do not flush denorms. */
1893 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1894 ac_to_float_type(&ctx
->ac
, def_type
),
1899 result
= emit_intrin_3f_param(&ctx
->ac
, "llvm.fmuladd",
1900 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
1902 case nir_op_ibitfield_extract
:
1903 result
= emit_bitfield_extract(&ctx
->ac
, true, src
);
1905 case nir_op_ubitfield_extract
:
1906 result
= emit_bitfield_extract(&ctx
->ac
, false, src
);
1908 case nir_op_bitfield_insert
:
1909 result
= emit_bitfield_insert(&ctx
->ac
, src
[0], src
[1], src
[2], src
[3]);
1911 case nir_op_bitfield_reverse
:
1912 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.bitreverse.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1914 case nir_op_bit_count
:
1915 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) == 32)
1916 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1918 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i64", ctx
->ac
.i64
, src
, 1, AC_FUNC_ATTR_READNONE
);
1919 result
= LLVMBuildTrunc(ctx
->ac
.builder
, result
, ctx
->ac
.i32
, "");
1925 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1926 src
[i
] = ac_to_integer(&ctx
->ac
, src
[i
]);
1927 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
1931 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1932 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
1936 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1937 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
1941 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1942 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1946 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1947 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1950 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1951 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1954 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1955 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1959 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1960 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
1961 result
= LLVMBuildZExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1963 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1967 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1968 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
1969 result
= LLVMBuildSExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1971 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1974 result
= emit_bcsel(&ctx
->ac
, src
[0], src
[1], src
[2]);
1976 case nir_op_find_lsb
:
1977 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1978 result
= ac_find_lsb(&ctx
->ac
, ctx
->ac
.i32
, src
[0]);
1980 case nir_op_ufind_msb
:
1981 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1982 result
= ac_build_umsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1984 case nir_op_ifind_msb
:
1985 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1986 result
= ac_build_imsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1988 case nir_op_uadd_carry
:
1989 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1990 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1991 result
= emit_uint_carry(&ctx
->ac
, "llvm.uadd.with.overflow.i32", src
[0], src
[1]);
1993 case nir_op_usub_borrow
:
1994 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1995 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1996 result
= emit_uint_carry(&ctx
->ac
, "llvm.usub.with.overflow.i32", src
[0], src
[1]);
1999 result
= emit_b2f(&ctx
->ac
, src
[0]);
2002 result
= emit_f2b(&ctx
->ac
, src
[0]);
2005 result
= emit_b2i(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
2008 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
2009 result
= emit_i2b(&ctx
->ac
, src
[0]);
2011 case nir_op_fquantize2f16
:
2012 result
= emit_f2f16(&ctx
->ac
, src
[0]);
2014 case nir_op_umul_high
:
2015 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
2016 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
2017 result
= emit_umul_high(&ctx
->ac
, src
[0], src
[1]);
2019 case nir_op_imul_high
:
2020 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
2021 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
2022 result
= emit_imul_high(&ctx
->ac
, src
[0], src
[1]);
2024 case nir_op_pack_half_2x16
:
2025 result
= emit_pack_half_2x16(&ctx
->ac
, src
[0]);
2027 case nir_op_unpack_half_2x16
:
2028 result
= emit_unpack_half_2x16(&ctx
->ac
, src
[0]);
2032 case nir_op_fddx_fine
:
2033 case nir_op_fddy_fine
:
2034 case nir_op_fddx_coarse
:
2035 case nir_op_fddy_coarse
:
2036 result
= emit_ddxy(ctx
, instr
->op
, src
[0]);
2039 case nir_op_unpack_64_2x32_split_x
: {
2040 assert(ac_get_llvm_num_components(src
[0]) == 1);
2041 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
2044 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
2049 case nir_op_unpack_64_2x32_split_y
: {
2050 assert(ac_get_llvm_num_components(src
[0]) == 1);
2051 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
2054 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
2059 case nir_op_pack_64_2x32_split
: {
2060 LLVMValueRef tmp
= LLVMGetUndef(ctx
->ac
.v2i32
);
2061 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
2062 src
[0], ctx
->ac
.i32_0
, "");
2063 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
2064 src
[1], ctx
->ac
.i32_1
, "");
2065 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i64
, "");
2070 fprintf(stderr
, "Unknown NIR alu instr: ");
2071 nir_print_instr(&instr
->instr
, stderr
);
2072 fprintf(stderr
, "\n");
2077 assert(instr
->dest
.dest
.is_ssa
);
2078 result
= ac_to_integer(&ctx
->ac
, result
);
2079 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.dest
.ssa
,
2084 static void visit_load_const(struct ac_nir_context
*ctx
,
2085 const nir_load_const_instr
*instr
)
2087 LLVMValueRef values
[4], value
= NULL
;
2088 LLVMTypeRef element_type
=
2089 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
2091 for (unsigned i
= 0; i
< instr
->def
.num_components
; ++i
) {
2092 switch (instr
->def
.bit_size
) {
2094 values
[i
] = LLVMConstInt(element_type
,
2095 instr
->value
.u32
[i
], false);
2098 values
[i
] = LLVMConstInt(element_type
,
2099 instr
->value
.u64
[i
], false);
2103 "unsupported nir load_const bit_size: %d\n",
2104 instr
->def
.bit_size
);
2108 if (instr
->def
.num_components
> 1) {
2109 value
= LLVMConstVector(values
, instr
->def
.num_components
);
2113 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, value
);
2116 static LLVMValueRef
cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
2119 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
2120 return LLVMBuildBitCast(ctx
->builder
, ptr
,
2121 LLVMPointerType(type
, addr_space
), "");
2125 get_buffer_size(struct ac_nir_context
*ctx
, LLVMValueRef descriptor
, bool in_elements
)
2128 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
2129 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
2132 if (ctx
->ac
.chip_class
== VI
&& in_elements
) {
2133 /* On VI, the descriptor contains the size in bytes,
2134 * but TXQ must return the size in elements.
2135 * The stride is always non-zero for resources using TXQ.
2137 LLVMValueRef stride
=
2138 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
2140 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
,
2141 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
2142 stride
= LLVMBuildAnd(ctx
->ac
.builder
, stride
,
2143 LLVMConstInt(ctx
->ac
.i32
, 0x3fff, false), "");
2145 size
= LLVMBuildUDiv(ctx
->ac
.builder
, size
, stride
, "");
2151 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
2154 static void build_int_type_name(
2156 char *buf
, unsigned bufsize
)
2158 assert(bufsize
>= 6);
2160 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
2161 snprintf(buf
, bufsize
, "v%ui32",
2162 LLVMGetVectorSize(type
));
2167 static LLVMValueRef
radv_lower_gather4_integer(struct ac_llvm_context
*ctx
,
2168 struct ac_image_args
*args
,
2169 const nir_tex_instr
*instr
)
2171 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2172 LLVMValueRef coord
= args
->addr
;
2173 LLVMValueRef half_texel
[2];
2174 LLVMValueRef compare_cube_wa
= NULL
;
2175 LLVMValueRef result
;
2177 unsigned coord_vgpr_index
= (unsigned)args
->offset
+ (unsigned)args
->compare
;
2181 struct ac_image_args txq_args
= { 0 };
2183 txq_args
.da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
2184 txq_args
.opcode
= ac_image_get_resinfo
;
2185 txq_args
.dmask
= 0xf;
2186 txq_args
.addr
= ctx
->i32_0
;
2187 txq_args
.resource
= args
->resource
;
2188 LLVMValueRef size
= ac_build_image_opcode(ctx
, &txq_args
);
2190 for (c
= 0; c
< 2; c
++) {
2191 half_texel
[c
] = LLVMBuildExtractElement(ctx
->builder
, size
,
2192 LLVMConstInt(ctx
->i32
, c
, false), "");
2193 half_texel
[c
] = LLVMBuildUIToFP(ctx
->builder
, half_texel
[c
], ctx
->f32
, "");
2194 half_texel
[c
] = ac_build_fdiv(ctx
, ctx
->f32_1
, half_texel
[c
]);
2195 half_texel
[c
] = LLVMBuildFMul(ctx
->builder
, half_texel
[c
],
2196 LLVMConstReal(ctx
->f32
, -0.5), "");
2200 LLVMValueRef orig_coords
= args
->addr
;
2202 for (c
= 0; c
< 2; c
++) {
2204 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, coord_vgpr_index
+ c
, 0);
2205 tmp
= LLVMBuildExtractElement(ctx
->builder
, coord
, index
, "");
2206 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2207 tmp
= LLVMBuildFAdd(ctx
->builder
, tmp
, half_texel
[c
], "");
2208 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2209 coord
= LLVMBuildInsertElement(ctx
->builder
, coord
, tmp
, index
, "");
2214 * Apparantly cube has issue with integer types that the workaround doesn't solve,
2215 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
2216 * workaround by sampling using a scaled type and converting.
2217 * This is taken from amdgpu-pro shaders.
2219 /* NOTE this produces some ugly code compared to amdgpu-pro,
2220 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
2221 * and then reads them back. -pro generates two selects,
2222 * one s_cmp for the descriptor rewriting
2223 * one v_cmp for the coordinate and result changes.
2225 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2226 LLVMValueRef tmp
, tmp2
;
2228 /* workaround 8/8/8/8 uint/sint cube gather bug */
2229 /* first detect it then change to a scaled read and f2i */
2230 tmp
= LLVMBuildExtractElement(ctx
->builder
, args
->resource
, ctx
->i32_1
, "");
2233 /* extract the DATA_FORMAT */
2234 tmp
= ac_build_bfe(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 20, false),
2235 LLVMConstInt(ctx
->i32
, 6, false), false);
2237 /* is the DATA_FORMAT == 8_8_8_8 */
2238 compare_cube_wa
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tmp
, LLVMConstInt(ctx
->i32
, V_008F14_IMG_DATA_FORMAT_8_8_8_8
, false), "");
2240 if (stype
== GLSL_TYPE_UINT
)
2241 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
2242 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0x8000000, false),
2243 LLVMConstInt(ctx
->i32
, 0x10000000, false), "");
2245 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
2246 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0xc000000, false),
2247 LLVMConstInt(ctx
->i32
, 0x14000000, false), "");
2249 /* replace the NUM FORMAT in the descriptor */
2250 tmp2
= LLVMBuildAnd(ctx
->builder
, tmp2
, LLVMConstInt(ctx
->i32
, C_008F14_NUM_FORMAT_GFX6
, false), "");
2251 tmp2
= LLVMBuildOr(ctx
->builder
, tmp2
, tmp
, "");
2253 args
->resource
= LLVMBuildInsertElement(ctx
->builder
, args
->resource
, tmp2
, ctx
->i32_1
, "");
2255 /* don't modify the coordinates for this case */
2256 coord
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, orig_coords
, coord
, "");
2259 result
= ac_build_image_opcode(ctx
, args
);
2261 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2262 LLVMValueRef tmp
, tmp2
;
2264 /* if the cube workaround is in place, f2i the result. */
2265 for (c
= 0; c
< 4; c
++) {
2266 tmp
= LLVMBuildExtractElement(ctx
->builder
, result
, LLVMConstInt(ctx
->i32
, c
, false), "");
2267 if (stype
== GLSL_TYPE_UINT
)
2268 tmp2
= LLVMBuildFPToUI(ctx
->builder
, tmp
, ctx
->i32
, "");
2270 tmp2
= LLVMBuildFPToSI(ctx
->builder
, tmp
, ctx
->i32
, "");
2271 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2272 tmp2
= LLVMBuildBitCast(ctx
->builder
, tmp2
, ctx
->i32
, "");
2273 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, tmp2
, tmp
, "");
2274 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2275 result
= LLVMBuildInsertElement(ctx
->builder
, result
, tmp
, LLVMConstInt(ctx
->i32
, c
, false), "");
2281 static LLVMValueRef
build_tex_intrinsic(struct ac_nir_context
*ctx
,
2282 const nir_tex_instr
*instr
,
2284 struct ac_image_args
*args
)
2286 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
2287 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
2289 return ac_build_buffer_load_format(&ctx
->ac
,
2293 util_last_bit(mask
),
2297 args
->opcode
= ac_image_sample
;
2298 args
->compare
= instr
->is_shadow
;
2300 switch (instr
->op
) {
2302 case nir_texop_txf_ms
:
2303 case nir_texop_samples_identical
:
2304 args
->opcode
= lod_is_zero
||
2305 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
?
2306 ac_image_load
: ac_image_load_mip
;
2307 args
->compare
= false;
2308 args
->offset
= false;
2315 args
->level_zero
= true;
2320 case nir_texop_query_levels
:
2321 args
->opcode
= ac_image_get_resinfo
;
2324 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
2325 args
->level_zero
= true;
2331 args
->opcode
= ac_image_gather4
;
2332 args
->level_zero
= true;
2335 args
->opcode
= ac_image_get_lod
;
2336 args
->compare
= false;
2337 args
->offset
= false;
2343 if (instr
->op
== nir_texop_tg4
&& ctx
->ac
.chip_class
<= VI
) {
2344 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2345 if (stype
== GLSL_TYPE_UINT
|| stype
== GLSL_TYPE_INT
) {
2346 return radv_lower_gather4_integer(&ctx
->ac
, args
, instr
);
2349 return ac_build_image_opcode(&ctx
->ac
, args
);
2353 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
2354 unsigned desc_set
, unsigned binding
)
2356 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2357 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
2358 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
2359 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
2360 unsigned base_offset
= layout
->binding
[binding
].offset
;
2361 LLVMValueRef offset
, stride
;
2363 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
2364 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
2365 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
2366 layout
->binding
[binding
].dynamic_offset_offset
;
2367 desc_ptr
= ctx
->abi
.push_constants
;
2368 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
2369 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2371 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
2373 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
2374 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, stride
, "");
2375 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
2377 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
2378 desc_ptr
= cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
2379 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2384 static LLVMValueRef
visit_vulkan_resource_reindex(struct ac_nir_context
*ctx
,
2385 nir_intrinsic_instr
*instr
)
2387 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
2388 LLVMValueRef index
= get_src(ctx
, instr
->src
[1]);
2390 LLVMValueRef result
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
2391 LLVMSetMetadata(result
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2395 static LLVMValueRef
visit_load_push_constant(struct ac_nir_context
*ctx
,
2396 nir_intrinsic_instr
*instr
)
2398 LLVMValueRef ptr
, addr
;
2400 addr
= LLVMConstInt(ctx
->ac
.i32
, nir_intrinsic_base(instr
), 0);
2401 addr
= LLVMBuildAdd(ctx
->ac
.builder
, addr
,
2402 get_src(ctx
, instr
->src
[0]), "");
2404 ptr
= ac_build_gep0(&ctx
->ac
, ctx
->abi
->push_constants
, addr
);
2405 ptr
= cast_ptr(&ctx
->ac
, ptr
, get_def_type(ctx
, &instr
->dest
.ssa
));
2407 return LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
2410 static LLVMValueRef
visit_get_buffer_size(struct ac_nir_context
*ctx
,
2411 const nir_intrinsic_instr
*instr
)
2413 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
2415 return get_buffer_size(ctx
, ctx
->abi
->load_ssbo(ctx
->abi
, index
, false), false);
2418 static uint32_t widen_mask(uint32_t mask
, unsigned multiplier
)
2420 uint32_t new_mask
= 0;
2421 for(unsigned i
= 0; i
< 32 && (1u << i
) <= mask
; ++i
)
2422 if (mask
& (1u << i
))
2423 new_mask
|= ((1u << multiplier
) - 1u) << (i
* multiplier
);
2427 static LLVMValueRef
extract_vector_range(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
2428 unsigned start
, unsigned count
)
2430 LLVMTypeRef type
= LLVMTypeOf(src
);
2432 if (LLVMGetTypeKind(type
) != LLVMVectorTypeKind
) {
2438 unsigned src_elements
= LLVMGetVectorSize(type
);
2439 assert(start
< src_elements
);
2440 assert(start
+ count
<= src_elements
);
2442 if (start
== 0 && count
== src_elements
)
2446 return LLVMBuildExtractElement(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, start
, false), "");
2449 LLVMValueRef indices
[8];
2450 for (unsigned i
= 0; i
< count
; ++i
)
2451 indices
[i
] = LLVMConstInt(ctx
->i32
, start
+ i
, false);
2453 LLVMValueRef swizzle
= LLVMConstVector(indices
, count
);
2454 return LLVMBuildShuffleVector(ctx
->builder
, src
, src
, swizzle
, "");
2457 static void visit_store_ssbo(struct ac_nir_context
*ctx
,
2458 nir_intrinsic_instr
*instr
)
2460 const char *store_name
;
2461 LLVMValueRef src_data
= get_src(ctx
, instr
->src
[0]);
2462 LLVMTypeRef data_type
= ctx
->ac
.f32
;
2463 int elem_size_mult
= ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src_data
)) / 32;
2464 int components_32bit
= elem_size_mult
* instr
->num_components
;
2465 unsigned writemask
= nir_intrinsic_write_mask(instr
);
2466 LLVMValueRef base_data
, base_offset
;
2467 LLVMValueRef params
[6];
2469 params
[1] = ctx
->abi
->load_ssbo(ctx
->abi
,
2470 get_src(ctx
, instr
->src
[1]), true);
2471 params
[2] = ctx
->ac
.i32_0
; /* vindex */
2472 params
[4] = ctx
->ac
.i1false
; /* glc */
2473 params
[5] = ctx
->ac
.i1false
; /* slc */
2475 if (components_32bit
> 1)
2476 data_type
= LLVMVectorType(ctx
->ac
.f32
, components_32bit
);
2478 writemask
= widen_mask(writemask
, elem_size_mult
);
2480 base_data
= ac_to_float(&ctx
->ac
, src_data
);
2481 base_data
= trim_vector(&ctx
->ac
, base_data
, instr
->num_components
);
2482 base_data
= LLVMBuildBitCast(ctx
->ac
.builder
, base_data
,
2484 base_offset
= get_src(ctx
, instr
->src
[2]); /* voffset */
2488 LLVMValueRef offset
;
2490 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
2492 /* Due to an LLVM limitation, split 3-element writes
2493 * into a 2-element and a 1-element write. */
2495 writemask
|= 1 << (start
+ 2);
2500 writemask
|= ((1u << (count
- 4)) - 1u) << (start
+ 4);
2505 store_name
= "llvm.amdgcn.buffer.store.v4f32";
2506 } else if (count
== 2) {
2507 store_name
= "llvm.amdgcn.buffer.store.v2f32";
2511 store_name
= "llvm.amdgcn.buffer.store.f32";
2513 data
= extract_vector_range(&ctx
->ac
, base_data
, start
, count
);
2515 offset
= base_offset
;
2517 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, LLVMConstInt(ctx
->ac
.i32
, start
* 4, false), "");
2521 ac_build_intrinsic(&ctx
->ac
, store_name
,
2522 ctx
->ac
.voidt
, params
, 6, 0);
2526 static LLVMValueRef
visit_atomic_ssbo(struct ac_nir_context
*ctx
,
2527 const nir_intrinsic_instr
*instr
)
2530 LLVMValueRef params
[6];
2533 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
) {
2534 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[3]), 0);
2536 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
2537 params
[arg_count
++] = ctx
->abi
->load_ssbo(ctx
->abi
,
2538 get_src(ctx
, instr
->src
[0]),
2540 params
[arg_count
++] = ctx
->ac
.i32_0
; /* vindex */
2541 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
2542 params
[arg_count
++] = LLVMConstInt(ctx
->ac
.i1
, 0, false); /* slc */
2544 switch (instr
->intrinsic
) {
2545 case nir_intrinsic_ssbo_atomic_add
:
2546 name
= "llvm.amdgcn.buffer.atomic.add";
2548 case nir_intrinsic_ssbo_atomic_imin
:
2549 name
= "llvm.amdgcn.buffer.atomic.smin";
2551 case nir_intrinsic_ssbo_atomic_umin
:
2552 name
= "llvm.amdgcn.buffer.atomic.umin";
2554 case nir_intrinsic_ssbo_atomic_imax
:
2555 name
= "llvm.amdgcn.buffer.atomic.smax";
2557 case nir_intrinsic_ssbo_atomic_umax
:
2558 name
= "llvm.amdgcn.buffer.atomic.umax";
2560 case nir_intrinsic_ssbo_atomic_and
:
2561 name
= "llvm.amdgcn.buffer.atomic.and";
2563 case nir_intrinsic_ssbo_atomic_or
:
2564 name
= "llvm.amdgcn.buffer.atomic.or";
2566 case nir_intrinsic_ssbo_atomic_xor
:
2567 name
= "llvm.amdgcn.buffer.atomic.xor";
2569 case nir_intrinsic_ssbo_atomic_exchange
:
2570 name
= "llvm.amdgcn.buffer.atomic.swap";
2572 case nir_intrinsic_ssbo_atomic_comp_swap
:
2573 name
= "llvm.amdgcn.buffer.atomic.cmpswap";
2579 return ac_build_intrinsic(&ctx
->ac
, name
, ctx
->ac
.i32
, params
, arg_count
, 0);
2582 static LLVMValueRef
visit_load_buffer(struct ac_nir_context
*ctx
,
2583 const nir_intrinsic_instr
*instr
)
2585 LLVMValueRef results
[2];
2586 int load_components
;
2587 int num_components
= instr
->num_components
;
2588 if (instr
->dest
.ssa
.bit_size
== 64)
2589 num_components
*= 2;
2591 for (int i
= 0; i
< num_components
; i
+= load_components
) {
2592 load_components
= MIN2(num_components
- i
, 4);
2593 const char *load_name
;
2594 LLVMTypeRef data_type
= ctx
->ac
.f32
;
2595 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, i
* 4, false);
2596 offset
= LLVMBuildAdd(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]), offset
, "");
2598 if (load_components
== 3)
2599 data_type
= LLVMVectorType(ctx
->ac
.f32
, 4);
2600 else if (load_components
> 1)
2601 data_type
= LLVMVectorType(ctx
->ac
.f32
, load_components
);
2603 if (load_components
>= 3)
2604 load_name
= "llvm.amdgcn.buffer.load.v4f32";
2605 else if (load_components
== 2)
2606 load_name
= "llvm.amdgcn.buffer.load.v2f32";
2607 else if (load_components
== 1)
2608 load_name
= "llvm.amdgcn.buffer.load.f32";
2610 unreachable("unhandled number of components");
2612 LLVMValueRef params
[] = {
2613 ctx
->abi
->load_ssbo(ctx
->abi
,
2614 get_src(ctx
, instr
->src
[0]),
2622 results
[i
> 0 ? 1 : 0] = ac_build_intrinsic(&ctx
->ac
, load_name
, data_type
, params
, 5, 0);
2626 LLVMValueRef ret
= results
[0];
2627 if (num_components
> 4 || num_components
== 3) {
2628 LLVMValueRef masks
[] = {
2629 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
2630 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
2631 LLVMConstInt(ctx
->ac
.i32
, 4, false), LLVMConstInt(ctx
->ac
.i32
, 5, false),
2632 LLVMConstInt(ctx
->ac
.i32
, 6, false), LLVMConstInt(ctx
->ac
.i32
, 7, false)
2635 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
2636 ret
= LLVMBuildShuffleVector(ctx
->ac
.builder
, results
[0],
2637 results
[num_components
> 4 ? 1 : 0], swizzle
, "");
2640 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2641 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2644 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
2645 const nir_intrinsic_instr
*instr
)
2648 LLVMValueRef rsrc
= get_src(ctx
, instr
->src
[0]);
2649 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
2650 int num_components
= instr
->num_components
;
2652 if (ctx
->abi
->load_ubo
)
2653 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
2655 if (instr
->dest
.ssa
.bit_size
== 64)
2656 num_components
*= 2;
2658 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_components
, NULL
, offset
,
2659 NULL
, 0, false, false, true, true);
2660 ret
= trim_vector(&ctx
->ac
, ret
, num_components
);
2661 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2662 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2666 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_var
*deref
,
2667 bool vs_in
, unsigned *vertex_index_out
,
2668 LLVMValueRef
*vertex_index_ref
,
2669 unsigned *const_out
, LLVMValueRef
*indir_out
)
2671 unsigned const_offset
= 0;
2672 nir_deref
*tail
= &deref
->deref
;
2673 LLVMValueRef offset
= NULL
;
2675 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
2677 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2678 if (vertex_index_out
)
2679 *vertex_index_out
= deref_array
->base_offset
;
2681 if (vertex_index_ref
) {
2682 LLVMValueRef vtx
= LLVMConstInt(ctx
->ac
.i32
, deref_array
->base_offset
, false);
2683 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
2684 vtx
= LLVMBuildAdd(ctx
->ac
.builder
, vtx
, get_src(ctx
, deref_array
->indirect
), "");
2686 *vertex_index_ref
= vtx
;
2690 if (deref
->var
->data
.compact
) {
2691 assert(tail
->child
->deref_type
== nir_deref_type_array
);
2692 assert(glsl_type_is_scalar(glsl_without_array(deref
->var
->type
)));
2693 nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
2694 /* We always lower indirect dereferences for "compact" array vars. */
2695 assert(deref_array
->deref_array_type
== nir_deref_array_type_direct
);
2697 const_offset
= deref_array
->base_offset
;
2701 while (tail
->child
!= NULL
) {
2702 const struct glsl_type
*parent_type
= tail
->type
;
2705 if (tail
->deref_type
== nir_deref_type_array
) {
2706 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2707 LLVMValueRef index
, stride
, local_offset
;
2708 unsigned size
= glsl_count_attribute_slots(tail
->type
, vs_in
);
2710 const_offset
+= size
* deref_array
->base_offset
;
2711 if (deref_array
->deref_array_type
== nir_deref_array_type_direct
)
2714 assert(deref_array
->deref_array_type
== nir_deref_array_type_indirect
);
2715 index
= get_src(ctx
, deref_array
->indirect
);
2716 stride
= LLVMConstInt(ctx
->ac
.i32
, size
, 0);
2717 local_offset
= LLVMBuildMul(ctx
->ac
.builder
, stride
, index
, "");
2720 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, local_offset
, "");
2722 offset
= local_offset
;
2723 } else if (tail
->deref_type
== nir_deref_type_struct
) {
2724 nir_deref_struct
*deref_struct
= nir_deref_as_struct(tail
);
2726 for (unsigned i
= 0; i
< deref_struct
->index
; i
++) {
2727 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
2728 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
2731 unreachable("unsupported deref type");
2735 if (const_offset
&& offset
)
2736 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
2737 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
2740 *const_out
= const_offset
;
2741 *indir_out
= offset
;
2745 /* The offchip buffer layout for TCS->TES is
2747 * - attribute 0 of patch 0 vertex 0
2748 * - attribute 0 of patch 0 vertex 1
2749 * - attribute 0 of patch 0 vertex 2
2751 * - attribute 0 of patch 1 vertex 0
2752 * - attribute 0 of patch 1 vertex 1
2754 * - attribute 1 of patch 0 vertex 0
2755 * - attribute 1 of patch 0 vertex 1
2757 * - per patch attribute 0 of patch 0
2758 * - per patch attribute 0 of patch 1
2761 * Note that every attribute has 4 components.
2763 static LLVMValueRef
get_tcs_tes_buffer_address(struct nir_to_llvm_context
*ctx
,
2764 LLVMValueRef vertex_index
,
2765 LLVMValueRef param_index
)
2767 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
2768 LLVMValueRef param_stride
, constant16
;
2769 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
2771 vertices_per_patch
= unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 9, 6);
2772 num_patches
= unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 0, 9);
2773 total_vertices
= LLVMBuildMul(ctx
->ac
.builder
, vertices_per_patch
,
2776 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2778 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
2779 vertices_per_patch
, "");
2781 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
2784 param_stride
= total_vertices
;
2786 base_addr
= rel_patch_id
;
2787 param_stride
= num_patches
;
2790 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
2791 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
2792 param_stride
, ""), "");
2794 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
2796 if (!vertex_index
) {
2797 LLVMValueRef patch_data_offset
=
2798 unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 16, 16);
2800 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
2801 patch_data_offset
, "");
2806 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct nir_to_llvm_context
*ctx
,
2808 unsigned const_index
,
2810 LLVMValueRef vertex_index
,
2811 LLVMValueRef indir_index
)
2813 LLVMValueRef param_index
;
2816 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
2819 if (const_index
&& !is_compact
)
2820 param
+= const_index
;
2821 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
2823 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
2827 mark_tess_output(struct nir_to_llvm_context
*ctx
,
2828 bool is_patch
, uint32_t param
)
2832 ctx
->tess_patch_outputs_written
|= (1ull << param
);
2834 ctx
->tess_outputs_written
|= (1ull << param
);
2838 get_dw_address(struct nir_to_llvm_context
*ctx
,
2839 LLVMValueRef dw_addr
,
2841 unsigned const_index
,
2842 bool compact_const_index
,
2843 LLVMValueRef vertex_index
,
2844 LLVMValueRef stride
,
2845 LLVMValueRef indir_index
)
2850 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2851 LLVMBuildMul(ctx
->ac
.builder
,
2857 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2858 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
2859 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
2860 else if (const_index
&& !compact_const_index
)
2861 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2862 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
2864 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2865 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
2867 if (const_index
&& compact_const_index
)
2868 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2869 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
2874 load_tcs_varyings(struct ac_shader_abi
*abi
,
2875 LLVMValueRef vertex_index
,
2876 LLVMValueRef indir_index
,
2877 unsigned const_index
,
2879 unsigned driver_location
,
2881 unsigned num_components
,
2886 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2887 LLVMValueRef dw_addr
, stride
;
2888 LLVMValueRef value
[4], result
;
2889 unsigned param
= shader_io_get_unique_index(location
);
2892 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_in_layout
, 13, 8);
2893 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
2896 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 13, 8);
2897 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2899 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2904 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2907 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
2908 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
2909 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2912 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
2917 store_tcs_output(struct ac_shader_abi
*abi
,
2918 LLVMValueRef vertex_index
,
2919 LLVMValueRef param_index
,
2920 unsigned const_index
,
2922 unsigned driver_location
,
2929 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2930 LLVMValueRef dw_addr
;
2931 LLVMValueRef stride
= NULL
;
2932 LLVMValueRef buf_addr
= NULL
;
2934 bool store_lds
= true;
2937 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
2940 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
2944 param
= shader_io_get_unique_index(location
);
2945 if (location
== VARYING_SLOT_CLIP_DIST0
&&
2946 is_compact
&& const_index
> 3) {
2952 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 13, 8);
2953 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2955 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2958 mark_tess_output(ctx
, is_patch
, param
);
2960 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2962 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
2963 vertex_index
, param_index
);
2965 bool is_tess_factor
= false;
2966 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
2967 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
2968 is_tess_factor
= true;
2970 unsigned base
= is_compact
? const_index
: 0;
2971 for (unsigned chan
= 0; chan
< 8; chan
++) {
2972 if (!(writemask
& (1 << chan
)))
2974 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
2976 if (store_lds
|| is_tess_factor
) {
2977 LLVMValueRef dw_addr_chan
=
2978 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2979 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
2980 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
2983 if (!is_tess_factor
&& writemask
!= 0xF)
2984 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
2985 buf_addr
, ctx
->oc_lds
,
2986 4 * (base
+ chan
), 1, 0, true, false);
2989 if (writemask
== 0xF) {
2990 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
2991 buf_addr
, ctx
->oc_lds
,
2992 (base
* 4), 1, 0, true, false);
2997 load_tes_input(struct ac_shader_abi
*abi
,
2998 LLVMValueRef vertex_index
,
2999 LLVMValueRef param_index
,
3000 unsigned const_index
,
3002 unsigned driver_location
,
3004 unsigned num_components
,
3009 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
3010 LLVMValueRef buf_addr
;
3011 LLVMValueRef result
;
3012 unsigned param
= shader_io_get_unique_index(location
);
3014 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
3019 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
3020 is_compact
, vertex_index
, param_index
);
3022 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
3023 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
3025 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
3026 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
3027 result
= trim_vector(&ctx
->ac
, result
, num_components
);
3032 load_gs_input(struct ac_shader_abi
*abi
,
3034 unsigned driver_location
,
3036 unsigned num_components
,
3037 unsigned vertex_index
,
3038 unsigned const_index
,
3041 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
3042 LLVMValueRef vtx_offset
;
3043 unsigned param
, vtx_offset_param
;
3044 LLVMValueRef value
[4], result
;
3046 vtx_offset_param
= vertex_index
;
3047 assert(vtx_offset_param
< 6);
3048 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
3049 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3051 param
= shader_io_get_unique_index(location
);
3053 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
3054 if (ctx
->ac
.chip_class
>= GFX9
) {
3055 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
3056 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
3057 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
3058 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
3060 LLVMValueRef soffset
=
3061 LLVMConstInt(ctx
->ac
.i32
,
3062 (param
* 4 + i
+ const_index
) * 256,
3065 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
3068 vtx_offset
, soffset
,
3069 0, 1, 0, true, false);
3071 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
],
3075 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
3076 result
= ac_to_integer(&ctx
->ac
, result
);
3081 build_gep_for_deref(struct ac_nir_context
*ctx
,
3082 nir_deref_var
*deref
)
3084 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, deref
->var
);
3085 assert(entry
->data
);
3086 LLVMValueRef val
= entry
->data
;
3087 nir_deref
*tail
= deref
->deref
.child
;
3088 while (tail
!= NULL
) {
3089 LLVMValueRef offset
;
3090 switch (tail
->deref_type
) {
3091 case nir_deref_type_array
: {
3092 nir_deref_array
*array
= nir_deref_as_array(tail
);
3093 offset
= LLVMConstInt(ctx
->ac
.i32
, array
->base_offset
, 0);
3094 if (array
->deref_array_type
==
3095 nir_deref_array_type_indirect
) {
3096 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
3103 case nir_deref_type_struct
: {
3104 nir_deref_struct
*deref_struct
=
3105 nir_deref_as_struct(tail
);
3106 offset
= LLVMConstInt(ctx
->ac
.i32
,
3107 deref_struct
->index
, 0);
3111 unreachable("bad deref type");
3113 val
= ac_build_gep0(&ctx
->ac
, val
, offset
);
3119 static LLVMValueRef
load_tess_varyings(struct ac_nir_context
*ctx
,
3120 nir_intrinsic_instr
*instr
,
3123 LLVMValueRef result
;
3124 LLVMValueRef vertex_index
= NULL
;
3125 LLVMValueRef indir_index
= NULL
;
3126 unsigned const_index
= 0;
3127 unsigned location
= instr
->variables
[0]->var
->data
.location
;
3128 unsigned driver_location
= instr
->variables
[0]->var
->data
.driver_location
;
3129 const bool is_patch
= instr
->variables
[0]->var
->data
.patch
;
3130 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
3132 get_deref_offset(ctx
, instr
->variables
[0],
3133 false, NULL
, is_patch
? NULL
: &vertex_index
,
3134 &const_index
, &indir_index
);
3136 result
= ctx
->abi
->load_tess_varyings(ctx
->abi
, vertex_index
, indir_index
,
3137 const_index
, location
, driver_location
,
3138 instr
->variables
[0]->var
->data
.location_frac
,
3139 instr
->num_components
,
3140 is_patch
, is_compact
, load_inputs
);
3141 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3144 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
3145 nir_intrinsic_instr
*instr
)
3147 LLVMValueRef values
[8];
3148 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
3149 int ve
= instr
->dest
.ssa
.num_components
;
3150 unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3151 LLVMValueRef indir_index
;
3153 unsigned const_index
;
3154 unsigned stride
= instr
->variables
[0]->var
->data
.compact
? 1 : 4;
3155 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
3156 instr
->variables
[0]->var
->data
.mode
== nir_var_shader_in
;
3157 get_deref_offset(ctx
, instr
->variables
[0], vs_in
, NULL
, NULL
,
3158 &const_index
, &indir_index
);
3160 if (instr
->dest
.ssa
.bit_size
== 64)
3163 switch (instr
->variables
[0]->var
->data
.mode
) {
3164 case nir_var_shader_in
:
3165 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3166 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3167 return load_tess_varyings(ctx
, instr
, true);
3170 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3171 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
3172 LLVMValueRef indir_index
;
3173 unsigned const_index
, vertex_index
;
3174 get_deref_offset(ctx
, instr
->variables
[0],
3175 false, &vertex_index
, NULL
,
3176 &const_index
, &indir_index
);
3178 return ctx
->abi
->load_inputs(ctx
->abi
, instr
->variables
[0]->var
->data
.location
,
3179 instr
->variables
[0]->var
->data
.driver_location
,
3180 instr
->variables
[0]->var
->data
.location_frac
, ve
,
3181 vertex_index
, const_index
, type
);
3184 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
3186 unsigned count
= glsl_count_attribute_slots(
3187 instr
->variables
[0]->var
->type
,
3188 ctx
->stage
== MESA_SHADER_VERTEX
);
3190 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3191 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
3192 stride
, false, true);
3194 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3198 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* stride
];
3202 for (unsigned chan
= 0; chan
< ve
; chan
++) {
3204 unsigned count
= glsl_count_attribute_slots(
3205 instr
->variables
[0]->var
->type
, false);
3207 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3208 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3209 stride
, true, true);
3211 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3215 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* stride
], "");
3219 case nir_var_shared
: {
3220 LLVMValueRef address
= build_gep_for_deref(ctx
,
3221 instr
->variables
[0]);
3222 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
3223 return LLVMBuildBitCast(ctx
->ac
.builder
, val
,
3224 get_def_type(ctx
, &instr
->dest
.ssa
),
3227 case nir_var_shader_out
:
3228 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3229 return load_tess_varyings(ctx
, instr
, false);
3232 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
3234 unsigned count
= glsl_count_attribute_slots(
3235 instr
->variables
[0]->var
->type
, false);
3237 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3238 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3239 stride
, true, true);
3241 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3245 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
3246 ctx
->outputs
[idx
+ chan
+ const_index
* stride
],
3252 unreachable("unhandle variable mode");
3254 ret
= ac_build_varying_gather_values(&ctx
->ac
, values
, ve
, comp
);
3255 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3259 visit_store_var(struct ac_nir_context
*ctx
,
3260 nir_intrinsic_instr
*instr
)
3262 LLVMValueRef temp_ptr
, value
;
3263 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
3264 unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3265 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3266 int writemask
= instr
->const_index
[0] << comp
;
3267 LLVMValueRef indir_index
;
3268 unsigned const_index
;
3269 get_deref_offset(ctx
, instr
->variables
[0], false,
3270 NULL
, NULL
, &const_index
, &indir_index
);
3272 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64) {
3274 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
3275 LLVMVectorType(ctx
->ac
.f32
, ac_get_llvm_num_components(src
) * 2),
3278 writemask
= widen_mask(writemask
, 2);
3281 switch (instr
->variables
[0]->var
->data
.mode
) {
3282 case nir_var_shader_out
:
3284 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3285 LLVMValueRef vertex_index
= NULL
;
3286 LLVMValueRef indir_index
= NULL
;
3287 unsigned const_index
= 0;
3288 const unsigned location
= instr
->variables
[0]->var
->data
.location
;
3289 const unsigned driver_location
= instr
->variables
[0]->var
->data
.driver_location
;
3290 const unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3291 const bool is_patch
= instr
->variables
[0]->var
->data
.patch
;
3292 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
3294 get_deref_offset(ctx
, instr
->variables
[0],
3295 false, NULL
, is_patch
? NULL
: &vertex_index
,
3296 &const_index
, &indir_index
);
3298 ctx
->abi
->store_tcs_outputs(ctx
->abi
, vertex_index
, indir_index
,
3299 const_index
, location
, driver_location
,
3300 src
, comp
, is_patch
, is_compact
, writemask
);
3304 for (unsigned chan
= 0; chan
< 8; chan
++) {
3306 if (!(writemask
& (1 << chan
)))
3309 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- comp
);
3311 if (instr
->variables
[0]->var
->data
.compact
)
3314 unsigned count
= glsl_count_attribute_slots(
3315 instr
->variables
[0]->var
->type
, false);
3317 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3318 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3319 stride
, true, true);
3321 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3322 value
, indir_index
, "");
3323 build_store_values_extended(&ctx
->ac
, ctx
->outputs
+ idx
+ chan
,
3324 count
, stride
, tmp_vec
);
3327 temp_ptr
= ctx
->outputs
[idx
+ chan
+ const_index
* stride
];
3329 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3334 for (unsigned chan
= 0; chan
< 8; chan
++) {
3335 if (!(writemask
& (1 << chan
)))
3338 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
3340 unsigned count
= glsl_count_attribute_slots(
3341 instr
->variables
[0]->var
->type
, false);
3343 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3344 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3347 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3348 value
, indir_index
, "");
3349 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
3352 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
3354 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3358 case nir_var_shared
: {
3359 int writemask
= instr
->const_index
[0];
3360 LLVMValueRef address
= build_gep_for_deref(ctx
,
3361 instr
->variables
[0]);
3362 LLVMValueRef val
= get_src(ctx
, instr
->src
[0]);
3363 unsigned components
=
3364 glsl_get_vector_elements(
3365 nir_deref_tail(&instr
->variables
[0]->deref
)->type
);
3366 if (writemask
== (1 << components
) - 1) {
3367 val
= LLVMBuildBitCast(
3368 ctx
->ac
.builder
, val
,
3369 LLVMGetElementType(LLVMTypeOf(address
)), "");
3370 LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
3372 for (unsigned chan
= 0; chan
< 4; chan
++) {
3373 if (!(writemask
& (1 << chan
)))
3376 LLVMBuildStructGEP(ctx
->ac
.builder
,
3378 LLVMValueRef src
= ac_llvm_extract_elem(&ctx
->ac
, val
,
3380 src
= LLVMBuildBitCast(
3381 ctx
->ac
.builder
, src
,
3382 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
3383 LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
3393 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
3396 case GLSL_SAMPLER_DIM_BUF
:
3398 case GLSL_SAMPLER_DIM_1D
:
3399 return array
? 2 : 1;
3400 case GLSL_SAMPLER_DIM_2D
:
3401 return array
? 3 : 2;
3402 case GLSL_SAMPLER_DIM_MS
:
3403 return array
? 4 : 3;
3404 case GLSL_SAMPLER_DIM_3D
:
3405 case GLSL_SAMPLER_DIM_CUBE
:
3407 case GLSL_SAMPLER_DIM_RECT
:
3408 case GLSL_SAMPLER_DIM_SUBPASS
:
3410 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
3420 /* Adjust the sample index according to FMASK.
3422 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3423 * which is the identity mapping. Each nibble says which physical sample
3424 * should be fetched to get that sample.
3426 * For example, 0x11111100 means there are only 2 samples stored and
3427 * the second sample covers 3/4 of the pixel. When reading samples 0
3428 * and 1, return physical sample 0 (determined by the first two 0s
3429 * in FMASK), otherwise return physical sample 1.
3431 * The sample index should be adjusted as follows:
3432 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
3434 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
3435 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
3436 LLVMValueRef coord_z
,
3437 LLVMValueRef sample_index
,
3438 LLVMValueRef fmask_desc_ptr
)
3440 LLVMValueRef fmask_load_address
[4];
3443 fmask_load_address
[0] = coord_x
;
3444 fmask_load_address
[1] = coord_y
;
3446 fmask_load_address
[2] = coord_z
;
3447 fmask_load_address
[3] = LLVMGetUndef(ctx
->i32
);
3450 struct ac_image_args args
= {0};
3452 args
.opcode
= ac_image_load
;
3453 args
.da
= coord_z
? true : false;
3454 args
.resource
= fmask_desc_ptr
;
3456 args
.addr
= ac_build_gather_values(ctx
, fmask_load_address
, coord_z
? 4 : 2);
3458 res
= ac_build_image_opcode(ctx
, &args
);
3460 res
= ac_to_integer(ctx
, res
);
3461 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, false);
3462 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xf, false);
3464 LLVMValueRef fmask
= LLVMBuildExtractElement(ctx
->builder
,
3468 LLVMValueRef sample_index4
=
3469 LLVMBuildMul(ctx
->builder
, sample_index
, four
, "");
3470 LLVMValueRef shifted_fmask
=
3471 LLVMBuildLShr(ctx
->builder
, fmask
, sample_index4
, "");
3472 LLVMValueRef final_sample
=
3473 LLVMBuildAnd(ctx
->builder
, shifted_fmask
, F
, "");
3475 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3476 * resource descriptor is 0 (invalid),
3478 LLVMValueRef fmask_desc
=
3479 LLVMBuildBitCast(ctx
->builder
, fmask_desc_ptr
,
3482 LLVMValueRef fmask_word1
=
3483 LLVMBuildExtractElement(ctx
->builder
, fmask_desc
,
3486 LLVMValueRef word1_is_nonzero
=
3487 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3488 fmask_word1
, ctx
->i32_0
, "");
3490 /* Replace the MSAA sample index. */
3492 LLVMBuildSelect(ctx
->builder
, word1_is_nonzero
,
3493 final_sample
, sample_index
, "");
3494 return sample_index
;
3497 static LLVMValueRef
get_image_coords(struct ac_nir_context
*ctx
,
3498 const nir_intrinsic_instr
*instr
)
3500 const struct glsl_type
*type
= glsl_without_array(instr
->variables
[0]->var
->type
);
3502 LLVMValueRef src0
= get_src(ctx
, instr
->src
[0]);
3503 LLVMValueRef coords
[4];
3504 LLVMValueRef masks
[] = {
3505 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
3506 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
3509 LLVMValueRef sample_index
= ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[1]), 0);
3512 enum glsl_sampler_dim dim
= glsl_get_sampler_dim(type
);
3513 bool is_array
= glsl_sampler_type_is_array(type
);
3514 bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3515 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3516 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
3517 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3518 bool gfx9_1d
= ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
;
3519 count
= image_type_to_components_count(dim
, is_array
);
3522 LLVMValueRef fmask_load_address
[3];
3525 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3526 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
3528 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
3530 fmask_load_address
[2] = NULL
;
3532 for (chan
= 0; chan
< 2; ++chan
)
3533 fmask_load_address
[chan
] =
3534 LLVMBuildAdd(ctx
->ac
.builder
, fmask_load_address
[chan
],
3535 LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
3536 ctx
->ac
.i32
, ""), "");
3537 fmask_load_address
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
3539 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
3540 fmask_load_address
[0],
3541 fmask_load_address
[1],
3542 fmask_load_address
[2],
3544 get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_FMASK
, NULL
, true, false));
3546 if (count
== 1 && !gfx9_1d
) {
3547 if (instr
->src
[0].ssa
->num_components
)
3548 res
= LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3555 for (chan
= 0; chan
< count
; ++chan
) {
3556 coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src0
, chan
);
3559 for (chan
= 0; chan
< 2; ++chan
)
3560 coords
[chan
] = LLVMBuildAdd(ctx
->ac
.builder
, coords
[chan
], LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
3561 ctx
->ac
.i32
, ""), "");
3562 coords
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
3568 coords
[2] = coords
[1];
3569 coords
[1] = ctx
->ac
.i32_0
;
3571 coords
[1] = ctx
->ac
.i32_0
;
3576 coords
[count
] = sample_index
;
3581 coords
[3] = LLVMGetUndef(ctx
->ac
.i32
);
3584 res
= ac_build_gather_values(&ctx
->ac
, coords
, count
);
3589 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
3590 const nir_intrinsic_instr
*instr
)
3592 LLVMValueRef params
[7];
3594 char intrinsic_name
[64];
3595 const nir_variable
*var
= instr
->variables
[0]->var
;
3596 const struct glsl_type
*type
= var
->type
;
3598 if(instr
->variables
[0]->deref
.child
)
3599 type
= instr
->variables
[0]->deref
.child
->type
;
3601 type
= glsl_without_array(type
);
3603 const enum glsl_sampler_dim dim
= glsl_get_sampler_dim(type
);
3604 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
3605 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
3606 unsigned num_channels
= util_last_bit(mask
);
3607 LLVMValueRef rsrc
, vindex
;
3609 rsrc
= get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, NULL
, true, false);
3610 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3613 /* TODO: set "glc" and "can_speculate" when OpenGL needs it. */
3614 res
= ac_build_buffer_load_format(&ctx
->ac
, rsrc
, vindex
,
3615 ctx
->ac
.i32_0
, num_channels
,
3617 res
= ac_build_expand_to_vec4(&ctx
->ac
, res
, num_channels
);
3619 res
= trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
3620 res
= ac_to_integer(&ctx
->ac
, res
);
3622 bool is_da
= glsl_sampler_type_is_array(type
) ||
3623 dim
== GLSL_SAMPLER_DIM_CUBE
||
3624 dim
== GLSL_SAMPLER_DIM_3D
||
3625 dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3626 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
;
3627 LLVMValueRef da
= is_da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
;
3628 LLVMValueRef glc
= ctx
->ac
.i1false
;
3629 LLVMValueRef slc
= ctx
->ac
.i1false
;
3631 params
[0] = get_image_coords(ctx
, instr
);
3632 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, false);
3633 params
[2] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3636 params
[5] = ctx
->ac
.i1false
;
3639 ac_get_image_intr_name("llvm.amdgcn.image.load",
3640 ctx
->ac
.v4f32
, /* vdata */
3641 LLVMTypeOf(params
[0]), /* coords */
3642 LLVMTypeOf(params
[1]), /* rsrc */
3643 intrinsic_name
, sizeof(intrinsic_name
));
3645 res
= ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.v4f32
,
3646 params
, 7, AC_FUNC_ATTR_READONLY
);
3648 return ac_to_integer(&ctx
->ac
, res
);
3651 static void visit_image_store(struct ac_nir_context
*ctx
,
3652 nir_intrinsic_instr
*instr
)
3654 LLVMValueRef params
[8];
3655 char intrinsic_name
[64];
3656 const nir_variable
*var
= instr
->variables
[0]->var
;
3657 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3658 const enum glsl_sampler_dim dim
= glsl_get_sampler_dim(type
);
3659 LLVMValueRef glc
= ctx
->ac
.i1false
;
3660 bool force_glc
= ctx
->ac
.chip_class
== SI
;
3662 glc
= ctx
->ac
.i1true
;
3664 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
3665 params
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2])); /* data */
3666 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, NULL
, true, true);
3667 params
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3668 ctx
->ac
.i32_0
, ""); /* vindex */
3669 params
[3] = ctx
->ac
.i32_0
; /* voffset */
3670 params
[4] = glc
; /* glc */
3671 params
[5] = ctx
->ac
.i1false
; /* slc */
3672 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.buffer.store.format.v4f32", ctx
->ac
.voidt
,
3675 bool is_da
= glsl_sampler_type_is_array(type
) ||
3676 dim
== GLSL_SAMPLER_DIM_CUBE
||
3677 dim
== GLSL_SAMPLER_DIM_3D
;
3678 LLVMValueRef da
= is_da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
;
3679 LLVMValueRef slc
= ctx
->ac
.i1false
;
3681 params
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2]));
3682 params
[1] = get_image_coords(ctx
, instr
); /* coords */
3683 params
[2] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, true);
3684 params
[3] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3687 params
[6] = ctx
->ac
.i1false
;
3690 ac_get_image_intr_name("llvm.amdgcn.image.store",
3691 LLVMTypeOf(params
[0]), /* vdata */
3692 LLVMTypeOf(params
[1]), /* coords */
3693 LLVMTypeOf(params
[2]), /* rsrc */
3694 intrinsic_name
, sizeof(intrinsic_name
));
3696 ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.voidt
,
3702 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
3703 const nir_intrinsic_instr
*instr
)
3705 LLVMValueRef params
[7];
3706 int param_count
= 0;
3707 const nir_variable
*var
= instr
->variables
[0]->var
;
3709 const char *atomic_name
;
3710 char intrinsic_name
[41];
3711 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3712 MAYBE_UNUSED
int length
;
3714 bool is_unsigned
= glsl_get_sampler_result_type(type
) == GLSL_TYPE_UINT
;
3716 switch (instr
->intrinsic
) {
3717 case nir_intrinsic_image_atomic_add
:
3718 atomic_name
= "add";
3720 case nir_intrinsic_image_atomic_min
:
3721 atomic_name
= is_unsigned
? "umin" : "smin";
3723 case nir_intrinsic_image_atomic_max
:
3724 atomic_name
= is_unsigned
? "umax" : "smax";
3726 case nir_intrinsic_image_atomic_and
:
3727 atomic_name
= "and";
3729 case nir_intrinsic_image_atomic_or
:
3732 case nir_intrinsic_image_atomic_xor
:
3733 atomic_name
= "xor";
3735 case nir_intrinsic_image_atomic_exchange
:
3736 atomic_name
= "swap";
3738 case nir_intrinsic_image_atomic_comp_swap
:
3739 atomic_name
= "cmpswap";
3745 if (instr
->intrinsic
== nir_intrinsic_image_atomic_comp_swap
)
3746 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
3747 params
[param_count
++] = get_src(ctx
, instr
->src
[2]);
3749 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3750 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
,
3752 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3753 ctx
->ac
.i32_0
, ""); /* vindex */
3754 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
3755 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
3757 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3758 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
3760 char coords_type
[8];
3762 bool da
= glsl_sampler_type_is_array(type
) ||
3763 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3765 LLVMValueRef coords
= params
[param_count
++] = get_image_coords(ctx
, instr
);
3766 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
,
3768 params
[param_count
++] = ctx
->ac
.i1false
; /* r128 */
3769 params
[param_count
++] = da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
; /* da */
3770 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
3772 build_int_type_name(LLVMTypeOf(coords
),
3773 coords_type
, sizeof(coords_type
));
3775 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3776 "llvm.amdgcn.image.atomic.%s.%s", atomic_name
, coords_type
);
3779 assert(length
< sizeof(intrinsic_name
));
3780 return ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
, params
, param_count
, 0);
3783 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
3784 const nir_intrinsic_instr
*instr
)
3787 const nir_variable
*var
= instr
->variables
[0]->var
;
3788 const struct glsl_type
*type
= instr
->variables
[0]->var
->type
;
3789 bool da
= glsl_sampler_type_is_array(var
->type
) ||
3790 glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_CUBE
||
3791 glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_3D
;
3792 if(instr
->variables
[0]->deref
.child
)
3793 type
= instr
->variables
[0]->deref
.child
->type
;
3795 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
)
3796 return get_buffer_size(ctx
,
3797 get_sampler_desc(ctx
, instr
->variables
[0],
3798 AC_DESC_BUFFER
, NULL
, true, false), true);
3800 struct ac_image_args args
= { 0 };
3804 args
.resource
= get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, false);
3805 args
.opcode
= ac_image_get_resinfo
;
3806 args
.addr
= ctx
->ac
.i32_0
;
3808 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
3810 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3812 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
&&
3813 glsl_sampler_type_is_array(type
)) {
3814 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3815 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3816 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3817 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
3819 if (ctx
->ac
.chip_class
>= GFX9
&&
3820 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_1D
&&
3821 glsl_sampler_type_is_array(type
)) {
3822 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3823 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, layers
,
3830 #define NOOP_WAITCNT 0xf7f
3831 #define LGKM_CNT 0x07f
3832 #define VM_CNT 0xf70
3834 static void emit_membar(struct ac_llvm_context
*ac
,
3835 const nir_intrinsic_instr
*instr
)
3837 unsigned waitcnt
= NOOP_WAITCNT
;
3839 switch (instr
->intrinsic
) {
3840 case nir_intrinsic_memory_barrier
:
3841 case nir_intrinsic_group_memory_barrier
:
3842 waitcnt
&= VM_CNT
& LGKM_CNT
;
3844 case nir_intrinsic_memory_barrier_atomic_counter
:
3845 case nir_intrinsic_memory_barrier_buffer
:
3846 case nir_intrinsic_memory_barrier_image
:
3849 case nir_intrinsic_memory_barrier_shared
:
3850 waitcnt
&= LGKM_CNT
;
3855 if (waitcnt
!= NOOP_WAITCNT
)
3856 ac_build_waitcnt(ac
, waitcnt
);
3859 static void emit_barrier(struct ac_llvm_context
*ac
, gl_shader_stage stage
)
3861 /* SI only (thanks to a hw bug workaround):
3862 * The real barrier instruction isn’t needed, because an entire patch
3863 * always fits into a single wave.
3865 if (ac
->chip_class
== SI
&& stage
== MESA_SHADER_TESS_CTRL
) {
3866 ac_build_waitcnt(ac
, LGKM_CNT
& VM_CNT
);
3869 ac_build_intrinsic(ac
, "llvm.amdgcn.s.barrier",
3870 ac
->voidt
, NULL
, 0, AC_FUNC_ATTR_CONVERGENT
);
3873 static void emit_discard(struct ac_nir_context
*ctx
,
3874 const nir_intrinsic_instr
*instr
)
3878 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
3879 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3880 get_src(ctx
, instr
->src
[0]),
3883 assert(instr
->intrinsic
== nir_intrinsic_discard
);
3884 cond
= LLVMConstInt(ctx
->ac
.i1
, false, 0);
3887 ac_build_kill_if_false(&ctx
->ac
, cond
);
3891 visit_load_helper_invocation(struct ac_nir_context
*ctx
)
3893 LLVMValueRef result
= ac_build_intrinsic(&ctx
->ac
,
3894 "llvm.amdgcn.ps.live",
3895 ctx
->ac
.i1
, NULL
, 0,
3896 AC_FUNC_ATTR_READNONE
);
3897 result
= LLVMBuildNot(ctx
->ac
.builder
, result
, "");
3898 return LLVMBuildSExt(ctx
->ac
.builder
, result
, ctx
->ac
.i32
, "");
3902 visit_load_local_invocation_index(struct ac_nir_context
*ctx
)
3904 LLVMValueRef result
;
3905 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
3906 result
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
3907 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
3909 return LLVMBuildAdd(ctx
->ac
.builder
, result
, thread_id
, "");
3912 static LLVMValueRef
visit_var_atomic(struct ac_nir_context
*ctx
,
3913 const nir_intrinsic_instr
*instr
)
3915 LLVMValueRef ptr
, result
;
3916 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
3917 ptr
= build_gep_for_deref(ctx
, instr
->variables
[0]);
3919 if (instr
->intrinsic
== nir_intrinsic_var_atomic_comp_swap
) {
3920 LLVMValueRef src1
= get_src(ctx
, instr
->src
[1]);
3921 result
= LLVMBuildAtomicCmpXchg(ctx
->ac
.builder
,
3923 LLVMAtomicOrderingSequentiallyConsistent
,
3924 LLVMAtomicOrderingSequentiallyConsistent
,
3927 LLVMAtomicRMWBinOp op
;
3928 switch (instr
->intrinsic
) {
3929 case nir_intrinsic_var_atomic_add
:
3930 op
= LLVMAtomicRMWBinOpAdd
;
3932 case nir_intrinsic_var_atomic_umin
:
3933 op
= LLVMAtomicRMWBinOpUMin
;
3935 case nir_intrinsic_var_atomic_umax
:
3936 op
= LLVMAtomicRMWBinOpUMax
;
3938 case nir_intrinsic_var_atomic_imin
:
3939 op
= LLVMAtomicRMWBinOpMin
;
3941 case nir_intrinsic_var_atomic_imax
:
3942 op
= LLVMAtomicRMWBinOpMax
;
3944 case nir_intrinsic_var_atomic_and
:
3945 op
= LLVMAtomicRMWBinOpAnd
;
3947 case nir_intrinsic_var_atomic_or
:
3948 op
= LLVMAtomicRMWBinOpOr
;
3950 case nir_intrinsic_var_atomic_xor
:
3951 op
= LLVMAtomicRMWBinOpXor
;
3953 case nir_intrinsic_var_atomic_exchange
:
3954 op
= LLVMAtomicRMWBinOpXchg
;
3960 result
= LLVMBuildAtomicRMW(ctx
->ac
.builder
, op
, ptr
, ac_to_integer(&ctx
->ac
, src
),
3961 LLVMAtomicOrderingSequentiallyConsistent
,
3967 static LLVMValueRef
lookup_interp_param(struct ac_shader_abi
*abi
,
3968 enum glsl_interp_mode interp
, unsigned location
)
3970 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
3973 case INTERP_MODE_FLAT
:
3976 case INTERP_MODE_SMOOTH
:
3977 case INTERP_MODE_NONE
:
3978 if (location
== INTERP_CENTER
)
3979 return ctx
->persp_center
;
3980 else if (location
== INTERP_CENTROID
)
3981 return ctx
->persp_centroid
;
3982 else if (location
== INTERP_SAMPLE
)
3983 return ctx
->persp_sample
;
3985 case INTERP_MODE_NOPERSPECTIVE
:
3986 if (location
== INTERP_CENTER
)
3987 return ctx
->linear_center
;
3988 else if (location
== INTERP_CENTROID
)
3989 return ctx
->linear_centroid
;
3990 else if (location
== INTERP_SAMPLE
)
3991 return ctx
->linear_sample
;
3997 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
3998 LLVMValueRef sample_id
)
4000 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4002 LLVMValueRef result
;
4003 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
4005 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
4006 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
4008 sample_id
= LLVMBuildAdd(ctx
->ac
.builder
, sample_id
, ctx
->sample_pos_offset
, "");
4009 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
4014 static LLVMValueRef
load_sample_pos(struct ac_nir_context
*ctx
)
4016 LLVMValueRef values
[2];
4018 values
[0] = emit_ffract(&ctx
->ac
, ctx
->abi
->frag_pos
[0], 32);
4019 values
[1] = emit_ffract(&ctx
->ac
, ctx
->abi
->frag_pos
[1], 32);
4020 return ac_build_gather_values(&ctx
->ac
, values
, 2);
4023 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
4025 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4026 uint8_t log2_ps_iter_samples
= ctx
->shader_info
->info
.ps
.force_persample
?
4027 ctx
->options
->key
.fs
.log2_num_samples
:
4028 ctx
->options
->key
.fs
.log2_ps_iter_samples
;
4030 /* The bit pattern matches that used by fixed function fragment
4032 static const uint16_t ps_iter_masks
[] = {
4033 0xffff, /* not used */
4039 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
4041 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
4043 LLVMValueRef result
, sample_id
;
4044 sample_id
= unpack_param(&ctx
->ac
, abi
->ancillary
, 8, 4);
4045 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
4046 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, abi
->sample_coverage
, "");
4050 static LLVMValueRef
visit_interp(struct ac_nir_context
*ctx
,
4051 const nir_intrinsic_instr
*instr
)
4053 LLVMValueRef result
[4];
4054 LLVMValueRef interp_param
, attr_number
;
4057 LLVMValueRef src_c0
= NULL
;
4058 LLVMValueRef src_c1
= NULL
;
4059 LLVMValueRef src0
= NULL
;
4060 int input_index
= instr
->variables
[0]->var
->data
.location
- VARYING_SLOT_VAR0
;
4061 switch (instr
->intrinsic
) {
4062 case nir_intrinsic_interp_var_at_centroid
:
4063 location
= INTERP_CENTROID
;
4065 case nir_intrinsic_interp_var_at_sample
:
4066 case nir_intrinsic_interp_var_at_offset
:
4067 location
= INTERP_CENTER
;
4068 src0
= get_src(ctx
, instr
->src
[0]);
4074 if (instr
->intrinsic
== nir_intrinsic_interp_var_at_offset
) {
4075 src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_0
, ""));
4076 src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_1
, ""));
4077 } else if (instr
->intrinsic
== nir_intrinsic_interp_var_at_sample
) {
4078 LLVMValueRef sample_position
;
4079 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
4081 /* fetch sample ID */
4082 sample_position
= ctx
->abi
->load_sample_position(ctx
->abi
, src0
);
4084 src_c0
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_0
, "");
4085 src_c0
= LLVMBuildFSub(ctx
->ac
.builder
, src_c0
, halfval
, "");
4086 src_c1
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_1
, "");
4087 src_c1
= LLVMBuildFSub(ctx
->ac
.builder
, src_c1
, halfval
, "");
4089 interp_param
= ctx
->abi
->lookup_interp_param(ctx
->abi
, instr
->variables
[0]->var
->data
.interpolation
, location
);
4090 attr_number
= LLVMConstInt(ctx
->ac
.i32
, input_index
, false);
4092 if (location
== INTERP_CENTER
) {
4093 LLVMValueRef ij_out
[2];
4094 LLVMValueRef ddxy_out
= emit_ddxy_interp(ctx
, interp_param
);
4097 * take the I then J parameters, and the DDX/Y for it, and
4098 * calculate the IJ inputs for the interpolator.
4099 * temp1 = ddx * offset/sample.x + I;
4100 * interp_param.I = ddy * offset/sample.y + temp1;
4101 * temp1 = ddx * offset/sample.x + J;
4102 * interp_param.J = ddy * offset/sample.y + temp1;
4104 for (unsigned i
= 0; i
< 2; i
++) {
4105 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
4106 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
4107 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4108 ddxy_out
, ix_ll
, "");
4109 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4110 ddxy_out
, iy_ll
, "");
4111 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4112 interp_param
, ix_ll
, "");
4113 LLVMValueRef temp1
, temp2
;
4115 interp_el
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_el
,
4118 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, src_c0
, "");
4119 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
4121 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, src_c1
, "");
4122 temp2
= LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
4124 ij_out
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
,
4125 temp2
, ctx
->ac
.i32
, "");
4127 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
4131 for (chan
= 0; chan
< 4; chan
++) {
4132 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
4135 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
,
4136 interp_param
, ctx
->ac
.v2f32
, "");
4137 LLVMValueRef i
= LLVMBuildExtractElement(
4138 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_0
, "");
4139 LLVMValueRef j
= LLVMBuildExtractElement(
4140 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_1
, "");
4142 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
4143 llvm_chan
, attr_number
,
4144 ctx
->abi
->prim_mask
, i
, j
);
4146 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
4147 LLVMConstInt(ctx
->ac
.i32
, 2, false),
4148 llvm_chan
, attr_number
,
4149 ctx
->abi
->prim_mask
);
4152 return ac_build_varying_gather_values(&ctx
->ac
, result
, instr
->num_components
,
4153 instr
->variables
[0]->var
->data
.location_frac
);
4157 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
4159 LLVMValueRef gs_next_vertex
;
4160 LLVMValueRef can_emit
;
4162 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4164 assert(stream
== 0);
4166 /* Write vertex attribute values to GSVS ring */
4167 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4168 ctx
->gs_next_vertex
,
4171 /* If this thread has already emitted the declared maximum number of
4172 * vertices, kill it: excessive vertex emissions are not supposed to
4173 * have any effect, and GS threads have no externally observable
4174 * effects other than emitting vertices.
4176 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4177 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
4178 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4180 /* loop num outputs */
4182 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
4183 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
4188 if (!(ctx
->output_mask
& (1ull << i
)))
4191 if (i
== VARYING_SLOT_CLIP_DIST0
) {
4192 /* pack clip and cull into a single set of slots */
4193 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
4197 for (unsigned j
= 0; j
< length
; j
++) {
4198 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
4200 LLVMValueRef voffset
= LLVMConstInt(ctx
->ac
.i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
4201 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
4202 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
4204 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
4206 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
4208 voffset
, ctx
->gs2vs_offset
, 0,
4214 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
4216 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
4218 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
4222 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
4224 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4225 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
4229 load_tess_coord(struct ac_shader_abi
*abi
)
4231 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4233 LLVMValueRef coord
[4] = {
4240 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
4241 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
4242 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
4244 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
4248 load_patch_vertices_in(struct ac_shader_abi
*abi
)
4250 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4251 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
4254 static void visit_intrinsic(struct ac_nir_context
*ctx
,
4255 nir_intrinsic_instr
*instr
)
4257 LLVMValueRef result
= NULL
;
4259 switch (instr
->intrinsic
) {
4260 case nir_intrinsic_ballot
:
4261 result
= ac_build_ballot(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4263 case nir_intrinsic_read_invocation
:
4264 case nir_intrinsic_read_first_invocation
: {
4265 LLVMValueRef args
[2];
4268 args
[0] = get_src(ctx
, instr
->src
[0]);
4271 const char *intr_name
;
4272 if (instr
->intrinsic
== nir_intrinsic_read_invocation
) {
4274 intr_name
= "llvm.amdgcn.readlane";
4277 args
[1] = get_src(ctx
, instr
->src
[1]);
4280 intr_name
= "llvm.amdgcn.readfirstlane";
4283 /* We currently have no other way to prevent LLVM from lifting the icmp
4284 * calls to a dominating basic block.
4286 ac_build_optimization_barrier(&ctx
->ac
, &args
[0]);
4288 result
= ac_build_intrinsic(&ctx
->ac
, intr_name
,
4289 ctx
->ac
.i32
, args
, num_args
,
4290 AC_FUNC_ATTR_READNONE
|
4291 AC_FUNC_ATTR_CONVERGENT
);
4294 case nir_intrinsic_load_subgroup_invocation
:
4295 result
= ac_get_thread_id(&ctx
->ac
);
4297 case nir_intrinsic_load_work_group_id
: {
4298 LLVMValueRef values
[3];
4300 for (int i
= 0; i
< 3; i
++) {
4301 values
[i
] = ctx
->abi
->workgroup_ids
[i
] ?
4302 ctx
->abi
->workgroup_ids
[i
] : ctx
->ac
.i32_0
;
4305 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
4308 case nir_intrinsic_load_base_vertex
: {
4309 result
= ctx
->abi
->base_vertex
;
4312 case nir_intrinsic_load_local_group_size
:
4313 result
= ctx
->abi
->load_local_group_size(ctx
->abi
);
4315 case nir_intrinsic_load_vertex_id_zero_base
: {
4316 result
= ctx
->abi
->vertex_id
;
4319 case nir_intrinsic_load_local_invocation_id
: {
4320 result
= ctx
->abi
->local_invocation_ids
;
4323 case nir_intrinsic_load_base_instance
:
4324 result
= ctx
->abi
->start_instance
;
4326 case nir_intrinsic_load_draw_id
:
4327 result
= ctx
->abi
->draw_id
;
4329 case nir_intrinsic_load_view_index
:
4330 result
= ctx
->abi
->view_index
;
4332 case nir_intrinsic_load_invocation_id
:
4333 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
4334 result
= unpack_param(&ctx
->ac
, ctx
->abi
->tcs_rel_ids
, 8, 5);
4336 result
= ctx
->abi
->gs_invocation_id
;
4338 case nir_intrinsic_load_primitive_id
:
4339 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
4340 result
= ctx
->abi
->gs_prim_id
;
4341 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
4342 result
= ctx
->abi
->tcs_patch_id
;
4343 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
4344 result
= ctx
->abi
->tes_patch_id
;
4346 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
4348 case nir_intrinsic_load_sample_id
:
4349 result
= unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
4351 case nir_intrinsic_load_sample_pos
:
4352 result
= load_sample_pos(ctx
);
4354 case nir_intrinsic_load_sample_mask_in
:
4355 result
= ctx
->abi
->load_sample_mask_in(ctx
->abi
);
4357 case nir_intrinsic_load_frag_coord
: {
4358 LLVMValueRef values
[4] = {
4359 ctx
->abi
->frag_pos
[0],
4360 ctx
->abi
->frag_pos
[1],
4361 ctx
->abi
->frag_pos
[2],
4362 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, ctx
->abi
->frag_pos
[3])
4364 result
= ac_build_gather_values(&ctx
->ac
, values
, 4);
4367 case nir_intrinsic_load_front_face
:
4368 result
= ctx
->abi
->front_face
;
4370 case nir_intrinsic_load_helper_invocation
:
4371 result
= visit_load_helper_invocation(ctx
);
4373 case nir_intrinsic_load_instance_id
:
4374 result
= ctx
->abi
->instance_id
;
4376 case nir_intrinsic_load_num_work_groups
:
4377 result
= ctx
->abi
->num_work_groups
;
4379 case nir_intrinsic_load_local_invocation_index
:
4380 result
= visit_load_local_invocation_index(ctx
);
4382 case nir_intrinsic_load_push_constant
:
4383 result
= visit_load_push_constant(ctx
, instr
);
4385 case nir_intrinsic_vulkan_resource_index
: {
4386 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
4387 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
4388 unsigned binding
= nir_intrinsic_binding(instr
);
4390 result
= ctx
->abi
->load_resource(ctx
->abi
, index
, desc_set
,
4394 case nir_intrinsic_vulkan_resource_reindex
:
4395 result
= visit_vulkan_resource_reindex(ctx
, instr
);
4397 case nir_intrinsic_store_ssbo
:
4398 visit_store_ssbo(ctx
, instr
);
4400 case nir_intrinsic_load_ssbo
:
4401 result
= visit_load_buffer(ctx
, instr
);
4403 case nir_intrinsic_ssbo_atomic_add
:
4404 case nir_intrinsic_ssbo_atomic_imin
:
4405 case nir_intrinsic_ssbo_atomic_umin
:
4406 case nir_intrinsic_ssbo_atomic_imax
:
4407 case nir_intrinsic_ssbo_atomic_umax
:
4408 case nir_intrinsic_ssbo_atomic_and
:
4409 case nir_intrinsic_ssbo_atomic_or
:
4410 case nir_intrinsic_ssbo_atomic_xor
:
4411 case nir_intrinsic_ssbo_atomic_exchange
:
4412 case nir_intrinsic_ssbo_atomic_comp_swap
:
4413 result
= visit_atomic_ssbo(ctx
, instr
);
4415 case nir_intrinsic_load_ubo
:
4416 result
= visit_load_ubo_buffer(ctx
, instr
);
4418 case nir_intrinsic_get_buffer_size
:
4419 result
= visit_get_buffer_size(ctx
, instr
);
4421 case nir_intrinsic_load_var
:
4422 result
= visit_load_var(ctx
, instr
);
4424 case nir_intrinsic_store_var
:
4425 visit_store_var(ctx
, instr
);
4427 case nir_intrinsic_image_load
:
4428 result
= visit_image_load(ctx
, instr
);
4430 case nir_intrinsic_image_store
:
4431 visit_image_store(ctx
, instr
);
4433 case nir_intrinsic_image_atomic_add
:
4434 case nir_intrinsic_image_atomic_min
:
4435 case nir_intrinsic_image_atomic_max
:
4436 case nir_intrinsic_image_atomic_and
:
4437 case nir_intrinsic_image_atomic_or
:
4438 case nir_intrinsic_image_atomic_xor
:
4439 case nir_intrinsic_image_atomic_exchange
:
4440 case nir_intrinsic_image_atomic_comp_swap
:
4441 result
= visit_image_atomic(ctx
, instr
);
4443 case nir_intrinsic_image_size
:
4444 result
= visit_image_size(ctx
, instr
);
4446 case nir_intrinsic_shader_clock
:
4447 result
= ac_build_shader_clock(&ctx
->ac
);
4449 case nir_intrinsic_discard
:
4450 case nir_intrinsic_discard_if
:
4451 emit_discard(ctx
, instr
);
4453 case nir_intrinsic_memory_barrier
:
4454 case nir_intrinsic_group_memory_barrier
:
4455 case nir_intrinsic_memory_barrier_atomic_counter
:
4456 case nir_intrinsic_memory_barrier_buffer
:
4457 case nir_intrinsic_memory_barrier_image
:
4458 case nir_intrinsic_memory_barrier_shared
:
4459 emit_membar(&ctx
->ac
, instr
);
4461 case nir_intrinsic_barrier
:
4462 emit_barrier(&ctx
->ac
, ctx
->stage
);
4464 case nir_intrinsic_var_atomic_add
:
4465 case nir_intrinsic_var_atomic_imin
:
4466 case nir_intrinsic_var_atomic_umin
:
4467 case nir_intrinsic_var_atomic_imax
:
4468 case nir_intrinsic_var_atomic_umax
:
4469 case nir_intrinsic_var_atomic_and
:
4470 case nir_intrinsic_var_atomic_or
:
4471 case nir_intrinsic_var_atomic_xor
:
4472 case nir_intrinsic_var_atomic_exchange
:
4473 case nir_intrinsic_var_atomic_comp_swap
:
4474 result
= visit_var_atomic(ctx
, instr
);
4476 case nir_intrinsic_interp_var_at_centroid
:
4477 case nir_intrinsic_interp_var_at_sample
:
4478 case nir_intrinsic_interp_var_at_offset
:
4479 result
= visit_interp(ctx
, instr
);
4481 case nir_intrinsic_emit_vertex
:
4482 ctx
->abi
->emit_vertex(ctx
->abi
, nir_intrinsic_stream_id(instr
), ctx
->outputs
);
4484 case nir_intrinsic_end_primitive
:
4485 ctx
->abi
->emit_primitive(ctx
->abi
, nir_intrinsic_stream_id(instr
));
4487 case nir_intrinsic_load_tess_coord
:
4488 result
= ctx
->abi
->load_tess_coord(ctx
->abi
);
4490 case nir_intrinsic_load_tess_level_outer
:
4491 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
);
4493 case nir_intrinsic_load_tess_level_inner
:
4494 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
);
4496 case nir_intrinsic_load_patch_vertices_in
:
4497 result
= ctx
->abi
->load_patch_vertices_in(ctx
->abi
);
4499 case nir_intrinsic_vote_all
: {
4500 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4501 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4504 case nir_intrinsic_vote_any
: {
4505 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4506 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4509 case nir_intrinsic_vote_eq
: {
4510 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4511 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4515 fprintf(stderr
, "Unknown intrinsic: ");
4516 nir_print_instr(&instr
->instr
, stderr
);
4517 fprintf(stderr
, "\n");
4521 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4525 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
4526 LLVMValueRef buffer_ptr
, bool write
)
4528 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4529 LLVMValueRef result
;
4531 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
4533 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
4534 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
4539 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
4541 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4542 LLVMValueRef result
;
4544 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
4546 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
4547 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
4552 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
4553 unsigned descriptor_set
,
4554 unsigned base_index
,
4555 unsigned constant_index
,
4557 enum ac_descriptor_type desc_type
,
4558 bool image
, bool write
)
4560 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4561 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
4562 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
4563 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
4564 unsigned offset
= binding
->offset
;
4565 unsigned stride
= binding
->size
;
4567 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4570 assert(base_index
< layout
->binding_count
);
4572 switch (desc_type
) {
4574 type
= ctx
->ac
.v8i32
;
4578 type
= ctx
->ac
.v8i32
;
4582 case AC_DESC_SAMPLER
:
4583 type
= ctx
->ac
.v4i32
;
4584 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
4589 case AC_DESC_BUFFER
:
4590 type
= ctx
->ac
.v4i32
;
4594 unreachable("invalid desc_type\n");
4597 offset
+= constant_index
* stride
;
4599 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
4600 (!index
|| binding
->immutable_samplers_equal
)) {
4601 if (binding
->immutable_samplers_equal
)
4604 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
4606 LLVMValueRef constants
[] = {
4607 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
4608 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
4609 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
4610 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
4612 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
4615 assert(stride
% type_size
== 0);
4618 index
= ctx
->ac
.i32_0
;
4620 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
4622 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
4623 list
= LLVMBuildPointerCast(builder
, list
, ac_array_in_const_addr_space(type
), "");
4625 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
4628 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
4629 const nir_deref_var
*deref
,
4630 enum ac_descriptor_type desc_type
,
4631 const nir_tex_instr
*tex_instr
,
4632 bool image
, bool write
)
4634 LLVMValueRef index
= NULL
;
4635 unsigned constant_index
= 0;
4636 unsigned descriptor_set
;
4637 unsigned base_index
;
4640 assert(tex_instr
&& !image
);
4642 base_index
= tex_instr
->sampler_index
;
4644 const nir_deref
*tail
= &deref
->deref
;
4645 while (tail
->child
) {
4646 const nir_deref_array
*child
= nir_deref_as_array(tail
->child
);
4647 unsigned array_size
= glsl_get_aoa_size(tail
->child
->type
);
4652 assert(child
->deref_array_type
!= nir_deref_array_type_wildcard
);
4654 if (child
->deref_array_type
== nir_deref_array_type_indirect
) {
4655 LLVMValueRef indirect
= get_src(ctx
, child
->indirect
);
4657 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
4658 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
4663 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
4666 constant_index
+= child
->base_offset
* array_size
;
4668 tail
= &child
->deref
;
4670 descriptor_set
= deref
->var
->data
.descriptor_set
;
4671 base_index
= deref
->var
->data
.binding
;
4674 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
4677 constant_index
, index
,
4678 desc_type
, image
, write
);
4681 static void set_tex_fetch_args(struct ac_llvm_context
*ctx
,
4682 struct ac_image_args
*args
,
4683 const nir_tex_instr
*instr
,
4685 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4686 LLVMValueRef
*param
, unsigned count
,
4689 unsigned is_rect
= 0;
4690 bool da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
4692 if (op
== nir_texop_lod
)
4694 /* Pad to power of two vector */
4695 while (count
< util_next_power_of_two(count
))
4696 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4699 args
->addr
= ac_build_gather_values(ctx
, param
, count
);
4701 args
->addr
= param
[0];
4703 args
->resource
= res_ptr
;
4704 args
->sampler
= samp_ptr
;
4706 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
&& op
== nir_texop_txf
) {
4707 args
->addr
= param
[0];
4711 args
->dmask
= dmask
;
4712 args
->unorm
= is_rect
;
4716 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4719 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4720 * filtering manually. The driver sets img7 to a mask clearing
4721 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4722 * s_and_b32 samp0, samp0, img7
4725 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4727 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
4728 LLVMValueRef res
, LLVMValueRef samp
)
4730 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4731 LLVMValueRef img7
, samp0
;
4733 if (ctx
->ac
.chip_class
>= VI
)
4736 img7
= LLVMBuildExtractElement(builder
, res
,
4737 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
4738 samp0
= LLVMBuildExtractElement(builder
, samp
,
4739 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4740 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4741 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4742 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4745 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
4746 nir_tex_instr
*instr
,
4747 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
4748 LLVMValueRef
*fmask_ptr
)
4750 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
4751 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_BUFFER
, instr
, false, false);
4753 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_IMAGE
, instr
, false, false);
4756 *samp_ptr
= get_sampler_desc(ctx
, instr
->sampler
, AC_DESC_SAMPLER
, instr
, false, false);
4758 *samp_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_SAMPLER
, instr
, false, false);
4759 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
4760 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4762 if (fmask_ptr
&& !instr
->sampler
&& (instr
->op
== nir_texop_txf_ms
||
4763 instr
->op
== nir_texop_samples_identical
))
4764 *fmask_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_FMASK
, instr
, false, false);
4767 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
4770 coord
= ac_to_float(ctx
, coord
);
4771 coord
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &coord
, 1, 0);
4772 coord
= ac_to_integer(ctx
, coord
);
4776 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
4778 LLVMValueRef result
= NULL
;
4779 struct ac_image_args args
= { 0 };
4780 unsigned dmask
= 0xf;
4781 LLVMValueRef address
[16];
4782 LLVMValueRef coords
[5];
4783 LLVMValueRef coord
= NULL
, lod
= NULL
, comparator
= NULL
;
4784 LLVMValueRef bias
= NULL
, offsets
= NULL
;
4785 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
, sample_index
= NULL
;
4786 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
4787 LLVMValueRef derivs
[6];
4788 unsigned chan
, count
= 0;
4789 unsigned const_src
= 0, num_deriv_comp
= 0;
4790 bool lod_is_zero
= false;
4792 tex_fetch_ptrs(ctx
, instr
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4794 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4795 switch (instr
->src
[i
].src_type
) {
4796 case nir_tex_src_coord
:
4797 coord
= get_src(ctx
, instr
->src
[i
].src
);
4799 case nir_tex_src_projector
:
4801 case nir_tex_src_comparator
:
4802 comparator
= get_src(ctx
, instr
->src
[i
].src
);
4804 case nir_tex_src_offset
:
4805 offsets
= get_src(ctx
, instr
->src
[i
].src
);
4808 case nir_tex_src_bias
:
4809 bias
= get_src(ctx
, instr
->src
[i
].src
);
4811 case nir_tex_src_lod
: {
4812 nir_const_value
*val
= nir_src_as_const_value(instr
->src
[i
].src
);
4814 if (val
&& val
->i32
[0] == 0)
4816 lod
= get_src(ctx
, instr
->src
[i
].src
);
4819 case nir_tex_src_ms_index
:
4820 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
4822 case nir_tex_src_ms_mcs
:
4824 case nir_tex_src_ddx
:
4825 ddx
= get_src(ctx
, instr
->src
[i
].src
);
4826 num_deriv_comp
= instr
->src
[i
].src
.ssa
->num_components
;
4828 case nir_tex_src_ddy
:
4829 ddy
= get_src(ctx
, instr
->src
[i
].src
);
4831 case nir_tex_src_texture_offset
:
4832 case nir_tex_src_sampler_offset
:
4833 case nir_tex_src_plane
:
4839 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
4840 result
= get_buffer_size(ctx
, res_ptr
, true);
4844 if (instr
->op
== nir_texop_texture_samples
) {
4845 LLVMValueRef res
, samples
, is_msaa
;
4846 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res_ptr
, ctx
->ac
.v8i32
, "");
4847 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4848 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4849 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4850 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
4851 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
4852 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4853 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
4854 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4856 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4857 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
4858 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
4859 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
4860 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
4862 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
4869 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4870 coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
4872 if (offsets
&& instr
->op
!= nir_texop_txf
) {
4873 LLVMValueRef offset
[3], pack
;
4874 for (chan
= 0; chan
< 3; ++chan
)
4875 offset
[chan
] = ctx
->ac
.i32_0
;
4878 for (chan
= 0; chan
< ac_get_llvm_num_components(offsets
); chan
++) {
4879 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, offsets
, chan
);
4880 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
4881 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
4883 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
4884 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
4886 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
4887 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
4888 address
[count
++] = pack
;
4891 /* pack LOD bias value */
4892 if (instr
->op
== nir_texop_txb
&& bias
) {
4893 address
[count
++] = bias
;
4896 /* Pack depth comparison value */
4897 if (instr
->is_shadow
&& comparator
) {
4898 LLVMValueRef z
= ac_to_float(&ctx
->ac
,
4899 ac_llvm_extract_elem(&ctx
->ac
, comparator
, 0));
4901 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
4902 * so the depth comparison value isn't clamped for Z16 and
4903 * Z24 anymore. Do it manually here.
4905 * It's unnecessary if the original texture format was
4906 * Z32_FLOAT, but we don't know that here.
4908 if (ctx
->ac
.chip_class
== VI
&& ctx
->abi
->clamp_shadow_reference
)
4909 z
= ac_build_clamp(&ctx
->ac
, z
);
4911 address
[count
++] = z
;
4914 /* pack derivatives */
4916 int num_src_deriv_channels
, num_dest_deriv_channels
;
4917 switch (instr
->sampler_dim
) {
4918 case GLSL_SAMPLER_DIM_3D
:
4919 case GLSL_SAMPLER_DIM_CUBE
:
4921 num_src_deriv_channels
= 3;
4922 num_dest_deriv_channels
= 3;
4924 case GLSL_SAMPLER_DIM_2D
:
4926 num_src_deriv_channels
= 2;
4927 num_dest_deriv_channels
= 2;
4930 case GLSL_SAMPLER_DIM_1D
:
4931 num_src_deriv_channels
= 1;
4932 if (ctx
->ac
.chip_class
>= GFX9
) {
4933 num_dest_deriv_channels
= 2;
4936 num_dest_deriv_channels
= 1;
4942 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
4943 derivs
[i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
4944 derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
4946 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
4947 derivs
[i
] = ctx
->ac
.f32_0
;
4948 derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
4952 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& coord
) {
4953 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4954 coords
[chan
] = ac_to_float(&ctx
->ac
, coords
[chan
]);
4955 if (instr
->coord_components
== 3)
4956 coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
4957 ac_prepare_cube_coords(&ctx
->ac
,
4958 instr
->op
== nir_texop_txd
, instr
->is_array
,
4959 instr
->op
== nir_texop_lod
, coords
, derivs
);
4965 for (unsigned i
= 0; i
< num_deriv_comp
* 2; i
++)
4966 address
[count
++] = derivs
[i
];
4969 /* Pack texture coordinates */
4971 address
[count
++] = coords
[0];
4972 if (instr
->coord_components
> 1) {
4973 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&& instr
->is_array
&& instr
->op
!= nir_texop_txf
) {
4974 coords
[1] = apply_round_slice(&ctx
->ac
, coords
[1]);
4976 address
[count
++] = coords
[1];
4978 if (instr
->coord_components
> 2) {
4979 /* This seems like a bit of a hack - but it passes Vulkan CTS with it */
4980 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_3D
&&
4981 instr
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
&&
4982 instr
->op
!= nir_texop_txf
) {
4983 coords
[2] = apply_round_slice(&ctx
->ac
, coords
[2]);
4985 address
[count
++] = coords
[2];
4988 if (ctx
->ac
.chip_class
>= GFX9
) {
4989 LLVMValueRef filler
;
4990 if (instr
->op
== nir_texop_txf
)
4991 filler
= ctx
->ac
.i32_0
;
4993 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
4995 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
) {
4996 /* No nir_texop_lod, because it does not take a slice
4997 * even with array textures. */
4998 if (instr
->is_array
&& instr
->op
!= nir_texop_lod
) {
4999 address
[count
] = address
[count
- 1];
5000 address
[count
- 1] = filler
;
5003 address
[count
++] = filler
;
5009 if (lod
&& ((instr
->op
== nir_texop_txl
&& !lod_is_zero
) ||
5010 instr
->op
== nir_texop_txf
)) {
5011 address
[count
++] = lod
;
5012 } else if (instr
->op
== nir_texop_txf_ms
&& sample_index
) {
5013 address
[count
++] = sample_index
;
5014 } else if(instr
->op
== nir_texop_txs
) {
5017 address
[count
++] = lod
;
5019 address
[count
++] = ctx
->ac
.i32_0
;
5022 for (chan
= 0; chan
< count
; chan
++) {
5023 address
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
,
5024 address
[chan
], ctx
->ac
.i32
, "");
5027 if (instr
->op
== nir_texop_samples_identical
) {
5028 LLVMValueRef txf_address
[4];
5029 struct ac_image_args txf_args
= { 0 };
5030 unsigned txf_count
= count
;
5031 memcpy(txf_address
, address
, sizeof(txf_address
));
5033 if (!instr
->is_array
)
5034 txf_address
[2] = ctx
->ac
.i32_0
;
5035 txf_address
[3] = ctx
->ac
.i32_0
;
5037 set_tex_fetch_args(&ctx
->ac
, &txf_args
, instr
, nir_texop_txf
,
5039 txf_address
, txf_count
, 0xf);
5041 result
= build_tex_intrinsic(ctx
, instr
, false, &txf_args
);
5043 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
5044 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
5048 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
5049 instr
->op
!= nir_texop_txs
) {
5050 unsigned sample_chan
= instr
->is_array
? 3 : 2;
5051 address
[sample_chan
] = adjust_sample_index_using_fmask(&ctx
->ac
,
5054 instr
->is_array
? address
[2] : NULL
,
5055 address
[sample_chan
],
5059 if (offsets
&& instr
->op
== nir_texop_txf
) {
5060 nir_const_value
*const_offset
=
5061 nir_src_as_const_value(instr
->src
[const_src
].src
);
5062 int num_offsets
= instr
->src
[const_src
].src
.ssa
->num_components
;
5063 assert(const_offset
);
5064 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
5065 if (num_offsets
> 2)
5066 address
[2] = LLVMBuildAdd(ctx
->ac
.builder
,
5067 address
[2], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[2], false), "");
5068 if (num_offsets
> 1)
5069 address
[1] = LLVMBuildAdd(ctx
->ac
.builder
,
5070 address
[1], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[1], false), "");
5071 address
[0] = LLVMBuildAdd(ctx
->ac
.builder
,
5072 address
[0], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[0], false), "");
5076 /* TODO TG4 support */
5077 if (instr
->op
== nir_texop_tg4
) {
5078 if (instr
->is_shadow
)
5081 dmask
= 1 << instr
->component
;
5083 set_tex_fetch_args(&ctx
->ac
, &args
, instr
, instr
->op
,
5084 res_ptr
, samp_ptr
, address
, count
, dmask
);
5086 result
= build_tex_intrinsic(ctx
, instr
, lod_is_zero
, &args
);
5088 if (instr
->op
== nir_texop_query_levels
)
5089 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
5090 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
5091 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
5092 instr
->op
!= nir_texop_tg4
)
5093 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
5094 else if (instr
->op
== nir_texop_txs
&&
5095 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
5097 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
5098 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
5099 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
5100 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
5101 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
5102 } else if (ctx
->ac
.chip_class
>= GFX9
&&
5103 instr
->op
== nir_texop_txs
&&
5104 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
5106 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
5107 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
5108 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
5110 } else if (instr
->dest
.ssa
.num_components
!= 4)
5111 result
= trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
5115 assert(instr
->dest
.is_ssa
);
5116 result
= ac_to_integer(&ctx
->ac
, result
);
5117 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
5122 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
5124 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
5125 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
5127 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
5128 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
5131 static void visit_post_phi(struct ac_nir_context
*ctx
,
5132 nir_phi_instr
*instr
,
5133 LLVMValueRef llvm_phi
)
5135 nir_foreach_phi_src(src
, instr
) {
5136 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
5137 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
5139 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
5143 static void phi_post_pass(struct ac_nir_context
*ctx
)
5145 struct hash_entry
*entry
;
5146 hash_table_foreach(ctx
->phis
, entry
) {
5147 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
5148 (LLVMValueRef
)entry
->data
);
5153 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
5154 const nir_ssa_undef_instr
*instr
)
5156 unsigned num_components
= instr
->def
.num_components
;
5157 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
5160 if (num_components
== 1)
5161 undef
= LLVMGetUndef(type
);
5163 undef
= LLVMGetUndef(LLVMVectorType(type
, num_components
));
5165 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, undef
);
5168 static void visit_jump(struct ac_nir_context
*ctx
,
5169 const nir_jump_instr
*instr
)
5171 switch (instr
->type
) {
5172 case nir_jump_break
:
5173 LLVMBuildBr(ctx
->ac
.builder
, ctx
->break_block
);
5174 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5176 case nir_jump_continue
:
5177 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5178 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5181 fprintf(stderr
, "Unknown NIR jump instr: ");
5182 nir_print_instr(&instr
->instr
, stderr
);
5183 fprintf(stderr
, "\n");
5188 static void visit_cf_list(struct ac_nir_context
*ctx
,
5189 struct exec_list
*list
);
5191 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
5193 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
5194 nir_foreach_instr(instr
, block
)
5196 switch (instr
->type
) {
5197 case nir_instr_type_alu
:
5198 visit_alu(ctx
, nir_instr_as_alu(instr
));
5200 case nir_instr_type_load_const
:
5201 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
5203 case nir_instr_type_intrinsic
:
5204 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
5206 case nir_instr_type_tex
:
5207 visit_tex(ctx
, nir_instr_as_tex(instr
));
5209 case nir_instr_type_phi
:
5210 visit_phi(ctx
, nir_instr_as_phi(instr
));
5212 case nir_instr_type_ssa_undef
:
5213 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
5215 case nir_instr_type_jump
:
5216 visit_jump(ctx
, nir_instr_as_jump(instr
));
5219 fprintf(stderr
, "Unknown NIR instr type: ");
5220 nir_print_instr(instr
, stderr
);
5221 fprintf(stderr
, "\n");
5226 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
5229 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
5231 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
5233 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5234 LLVMBasicBlockRef merge_block
=
5235 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5236 LLVMBasicBlockRef if_block
=
5237 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5238 LLVMBasicBlockRef else_block
= merge_block
;
5239 if (!exec_list_is_empty(&if_stmt
->else_list
))
5240 else_block
= LLVMAppendBasicBlockInContext(
5241 ctx
->ac
.context
, fn
, "");
5243 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, value
,
5245 LLVMBuildCondBr(ctx
->ac
.builder
, cond
, if_block
, else_block
);
5247 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, if_block
);
5248 visit_cf_list(ctx
, &if_stmt
->then_list
);
5249 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5250 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5252 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
5253 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, else_block
);
5254 visit_cf_list(ctx
, &if_stmt
->else_list
);
5255 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5256 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5259 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, merge_block
);
5262 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
5264 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5265 LLVMBasicBlockRef continue_parent
= ctx
->continue_block
;
5266 LLVMBasicBlockRef break_parent
= ctx
->break_block
;
5268 ctx
->continue_block
=
5269 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5271 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5273 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5274 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->continue_block
);
5275 visit_cf_list(ctx
, &loop
->body
);
5277 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5278 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5279 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->break_block
);
5281 ctx
->continue_block
= continue_parent
;
5282 ctx
->break_block
= break_parent
;
5285 static void visit_cf_list(struct ac_nir_context
*ctx
,
5286 struct exec_list
*list
)
5288 foreach_list_typed(nir_cf_node
, node
, node
, list
)
5290 switch (node
->type
) {
5291 case nir_cf_node_block
:
5292 visit_block(ctx
, nir_cf_node_as_block(node
));
5295 case nir_cf_node_if
:
5296 visit_if(ctx
, nir_cf_node_as_if(node
));
5299 case nir_cf_node_loop
:
5300 visit_loop(ctx
, nir_cf_node_as_loop(node
));
5310 handle_vs_input_decl(struct nir_to_llvm_context
*ctx
,
5311 struct nir_variable
*variable
)
5313 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
5314 LLVMValueRef t_offset
;
5315 LLVMValueRef t_list
;
5317 LLVMValueRef buffer_index
;
5318 int index
= variable
->data
.location
- VERT_ATTRIB_GENERIC0
;
5319 int idx
= variable
->data
.location
;
5320 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
5321 uint8_t input_usage_mask
=
5322 ctx
->shader_info
->info
.vs
.input_usage_mask
[variable
->data
.location
];
5323 unsigned num_channels
= util_last_bit(input_usage_mask
);
5325 variable
->data
.driver_location
= idx
* 4;
5327 for (unsigned i
= 0; i
< attrib_count
; ++i
, ++idx
) {
5328 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << (index
+ i
))) {
5329 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.instance_id
,
5330 ctx
->abi
.start_instance
, "");
5331 if (ctx
->options
->key
.vs
.as_ls
) {
5332 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
5333 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5335 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
5336 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5339 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
5340 ctx
->abi
.base_vertex
, "");
5341 t_offset
= LLVMConstInt(ctx
->ac
.i32
, index
+ i
, false);
5343 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
5345 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
5348 num_channels
, false, true);
5350 input
= ac_build_expand_to_vec4(&ctx
->ac
, input
, num_channels
);
5352 for (unsigned chan
= 0; chan
< 4; chan
++) {
5353 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5354 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
, chan
)] =
5355 ac_to_integer(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
,
5356 input
, llvm_chan
, ""));
5361 static void interp_fs_input(struct nir_to_llvm_context
*ctx
,
5363 LLVMValueRef interp_param
,
5364 LLVMValueRef prim_mask
,
5365 LLVMValueRef result
[4])
5367 LLVMValueRef attr_number
;
5370 bool interp
= interp_param
!= NULL
;
5372 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
5374 /* fs.constant returns the param from the middle vertex, so it's not
5375 * really useful for flat shading. It's meant to be used for custom
5376 * interpolation (but the intrinsic can't fetch from the other two
5379 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
5380 * to do the right thing. The only reason we use fs.constant is that
5381 * fs.interp cannot be used on integers, because they can be equal
5385 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
5388 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
5390 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
5394 for (chan
= 0; chan
< 4; chan
++) {
5395 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5398 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
5403 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
5404 LLVMConstInt(ctx
->ac
.i32
, 2, false),
5413 handle_fs_input_decl(struct nir_to_llvm_context
*ctx
,
5414 struct nir_variable
*variable
)
5416 int idx
= variable
->data
.location
;
5417 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5418 LLVMValueRef interp
;
5420 variable
->data
.driver_location
= idx
* 4;
5421 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
5423 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
5424 unsigned interp_type
;
5425 if (variable
->data
.sample
) {
5426 interp_type
= INTERP_SAMPLE
;
5427 ctx
->shader_info
->info
.ps
.force_persample
= true;
5428 } else if (variable
->data
.centroid
)
5429 interp_type
= INTERP_CENTROID
;
5431 interp_type
= INTERP_CENTER
;
5433 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
5437 for (unsigned i
= 0; i
< attrib_count
; ++i
)
5438 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
5443 handle_vs_inputs(struct nir_to_llvm_context
*ctx
,
5444 struct nir_shader
*nir
) {
5445 nir_foreach_variable(variable
, &nir
->inputs
)
5446 handle_vs_input_decl(ctx
, variable
);
5450 prepare_interp_optimize(struct nir_to_llvm_context
*ctx
,
5451 struct nir_shader
*nir
)
5453 if (!ctx
->options
->key
.fs
.multisample
)
5456 bool uses_center
= false;
5457 bool uses_centroid
= false;
5458 nir_foreach_variable(variable
, &nir
->inputs
) {
5459 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
5460 variable
->data
.sample
)
5463 if (variable
->data
.centroid
)
5464 uses_centroid
= true;
5469 if (uses_center
&& uses_centroid
) {
5470 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
5471 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
5472 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->ac
.builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
5477 handle_fs_inputs(struct nir_to_llvm_context
*ctx
,
5478 struct nir_shader
*nir
)
5480 prepare_interp_optimize(ctx
, nir
);
5482 nir_foreach_variable(variable
, &nir
->inputs
)
5483 handle_fs_input_decl(ctx
, variable
);
5487 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
5488 ctx
->shader_info
->info
.needs_multiview_view_index
)
5489 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
5491 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
5492 LLVMValueRef interp_param
;
5493 LLVMValueRef
*inputs
= ctx
->inputs
+radeon_llvm_reg_index_soa(i
, 0);
5495 if (!(ctx
->input_mask
& (1ull << i
)))
5498 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
5499 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
5500 interp_param
= *inputs
;
5501 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
5505 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
5507 } else if (i
== VARYING_SLOT_POS
) {
5508 for(int i
= 0; i
< 3; ++i
)
5509 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
5511 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
5512 ctx
->abi
.frag_pos
[3]);
5515 ctx
->shader_info
->fs
.num_interp
= index
;
5516 if (ctx
->input_mask
& (1 << VARYING_SLOT_PNTC
))
5517 ctx
->shader_info
->fs
.has_pcoord
= true;
5518 if (ctx
->input_mask
& (1 << VARYING_SLOT_PRIMITIVE_ID
))
5519 ctx
->shader_info
->fs
.prim_id_input
= true;
5520 if (ctx
->input_mask
& (1 << VARYING_SLOT_LAYER
))
5521 ctx
->shader_info
->fs
.layer_input
= true;
5522 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
5524 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
5525 ctx
->abi
.view_index
= ctx
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5529 ac_build_alloca(struct ac_llvm_context
*ac
,
5533 LLVMBuilderRef builder
= ac
->builder
;
5534 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
5535 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
5536 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
5537 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
5538 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
5542 LLVMPositionBuilderBefore(first_builder
, first_instr
);
5544 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
5547 res
= LLVMBuildAlloca(first_builder
, type
, name
);
5548 LLVMBuildStore(builder
, LLVMConstNull(type
), res
);
5550 LLVMDisposeBuilder(first_builder
);
5555 static LLVMValueRef
si_build_alloca_undef(struct ac_llvm_context
*ac
,
5559 LLVMValueRef ptr
= ac_build_alloca(ac
, type
, name
);
5560 LLVMBuildStore(ac
->builder
, LLVMGetUndef(type
), ptr
);
5565 scan_shader_output_decl(struct nir_to_llvm_context
*ctx
,
5566 struct nir_variable
*variable
,
5567 struct nir_shader
*shader
,
5568 gl_shader_stage stage
)
5570 int idx
= variable
->data
.location
+ variable
->data
.index
;
5571 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5572 uint64_t mask_attribs
;
5574 variable
->data
.driver_location
= idx
* 4;
5576 /* tess ctrl has it's own load/store paths for outputs */
5577 if (stage
== MESA_SHADER_TESS_CTRL
)
5580 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
5581 if (stage
== MESA_SHADER_VERTEX
||
5582 stage
== MESA_SHADER_TESS_EVAL
||
5583 stage
== MESA_SHADER_GEOMETRY
) {
5584 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5585 int length
= shader
->info
.clip_distance_array_size
+
5586 shader
->info
.cull_distance_array_size
;
5587 if (stage
== MESA_SHADER_VERTEX
) {
5588 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5589 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5591 if (stage
== MESA_SHADER_TESS_EVAL
) {
5592 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5593 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5600 mask_attribs
= 1ull << idx
;
5604 ctx
->output_mask
|= mask_attribs
;
5608 handle_shader_output_decl(struct ac_nir_context
*ctx
,
5609 struct nir_shader
*nir
,
5610 struct nir_variable
*variable
)
5612 unsigned output_loc
= variable
->data
.driver_location
/ 4;
5613 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5615 /* tess ctrl has it's own load/store paths for outputs */
5616 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
5619 if (ctx
->stage
== MESA_SHADER_VERTEX
||
5620 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
5621 ctx
->stage
== MESA_SHADER_GEOMETRY
) {
5622 int idx
= variable
->data
.location
+ variable
->data
.index
;
5623 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5624 int length
= nir
->info
.clip_distance_array_size
+
5625 nir
->info
.cull_distance_array_size
;
5634 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
5635 for (unsigned chan
= 0; chan
< 4; chan
++) {
5636 ctx
->outputs
[radeon_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
5637 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5643 glsl_base_to_llvm_type(struct ac_llvm_context
*ac
,
5644 enum glsl_base_type type
)
5648 case GLSL_TYPE_UINT
:
5649 case GLSL_TYPE_BOOL
:
5650 case GLSL_TYPE_SUBROUTINE
:
5652 case GLSL_TYPE_FLOAT
: /* TODO handle mediump */
5654 case GLSL_TYPE_INT64
:
5655 case GLSL_TYPE_UINT64
:
5657 case GLSL_TYPE_DOUBLE
:
5660 unreachable("unknown GLSL type");
5665 glsl_to_llvm_type(struct ac_llvm_context
*ac
,
5666 const struct glsl_type
*type
)
5668 if (glsl_type_is_scalar(type
)) {
5669 return glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
));
5672 if (glsl_type_is_vector(type
)) {
5673 return LLVMVectorType(
5674 glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
)),
5675 glsl_get_vector_elements(type
));
5678 if (glsl_type_is_matrix(type
)) {
5679 return LLVMArrayType(
5680 glsl_to_llvm_type(ac
, glsl_get_column_type(type
)),
5681 glsl_get_matrix_columns(type
));
5684 if (glsl_type_is_array(type
)) {
5685 return LLVMArrayType(
5686 glsl_to_llvm_type(ac
, glsl_get_array_element(type
)),
5687 glsl_get_length(type
));
5690 assert(glsl_type_is_struct(type
));
5692 LLVMTypeRef member_types
[glsl_get_length(type
)];
5694 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
5696 glsl_to_llvm_type(ac
,
5697 glsl_get_struct_field(type
, i
));
5700 return LLVMStructTypeInContext(ac
->context
, member_types
,
5701 glsl_get_length(type
), false);
5705 setup_locals(struct ac_nir_context
*ctx
,
5706 struct nir_function
*func
)
5709 ctx
->num_locals
= 0;
5710 nir_foreach_variable(variable
, &func
->impl
->locals
) {
5711 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5712 variable
->data
.driver_location
= ctx
->num_locals
* 4;
5713 variable
->data
.location_frac
= 0;
5714 ctx
->num_locals
+= attrib_count
;
5716 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
5720 for (i
= 0; i
< ctx
->num_locals
; i
++) {
5721 for (j
= 0; j
< 4; j
++) {
5722 ctx
->locals
[i
* 4 + j
] =
5723 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
5729 setup_shared(struct ac_nir_context
*ctx
,
5730 struct nir_shader
*nir
)
5732 nir_foreach_variable(variable
, &nir
->shared
) {
5733 LLVMValueRef shared
=
5734 LLVMAddGlobalInAddressSpace(
5735 ctx
->ac
.module
, glsl_to_llvm_type(&ctx
->ac
, variable
->type
),
5736 variable
->name
? variable
->name
: "",
5737 AC_LOCAL_ADDR_SPACE
);
5738 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
5742 /* Initialize arguments for the shader export intrinsic */
5744 si_llvm_init_export_args(struct nir_to_llvm_context
*ctx
,
5745 LLVMValueRef
*values
,
5747 struct ac_export_args
*args
)
5749 /* Default is 0xf. Adjusted below depending on the format. */
5750 args
->enabled_channels
= 0xf;
5752 /* Specify whether the EXEC mask represents the valid mask */
5753 args
->valid_mask
= 0;
5755 /* Specify whether this is the last export */
5758 /* Specify the target we are exporting */
5759 args
->target
= target
;
5761 args
->compr
= false;
5762 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
5763 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
5764 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
5765 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
5767 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
5768 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
5769 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
5770 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
5771 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
5774 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
5775 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
5776 unsigned bits
, bool hi
) = NULL
;
5778 switch(col_format
) {
5779 case V_028714_SPI_SHADER_ZERO
:
5780 args
->enabled_channels
= 0; /* writemask */
5781 args
->target
= V_008DFC_SQ_EXP_NULL
;
5784 case V_028714_SPI_SHADER_32_R
:
5785 args
->enabled_channels
= 1;
5786 args
->out
[0] = values
[0];
5789 case V_028714_SPI_SHADER_32_GR
:
5790 args
->enabled_channels
= 0x3;
5791 args
->out
[0] = values
[0];
5792 args
->out
[1] = values
[1];
5795 case V_028714_SPI_SHADER_32_AR
:
5796 args
->enabled_channels
= 0x9;
5797 args
->out
[0] = values
[0];
5798 args
->out
[3] = values
[3];
5801 case V_028714_SPI_SHADER_FP16_ABGR
:
5802 packf
= ac_build_cvt_pkrtz_f16
;
5805 case V_028714_SPI_SHADER_UNORM16_ABGR
:
5806 packf
= ac_build_cvt_pknorm_u16
;
5809 case V_028714_SPI_SHADER_SNORM16_ABGR
:
5810 packf
= ac_build_cvt_pknorm_i16
;
5813 case V_028714_SPI_SHADER_UINT16_ABGR
:
5814 packi
= ac_build_cvt_pk_u16
;
5817 case V_028714_SPI_SHADER_SINT16_ABGR
:
5818 packi
= ac_build_cvt_pk_i16
;
5822 case V_028714_SPI_SHADER_32_ABGR
:
5823 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5827 /* Pack f16 or norm_i16/u16. */
5829 for (chan
= 0; chan
< 2; chan
++) {
5830 LLVMValueRef pack_args
[2] = {
5832 values
[2 * chan
+ 1]
5834 LLVMValueRef packed
;
5836 packed
= packf(&ctx
->ac
, pack_args
);
5837 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
5839 args
->compr
= 1; /* COMPR flag */
5844 for (chan
= 0; chan
< 2; chan
++) {
5845 LLVMValueRef pack_args
[2] = {
5846 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
5847 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
5849 LLVMValueRef packed
;
5851 packed
= packi(&ctx
->ac
, pack_args
,
5852 is_int8
? 8 : is_int10
? 10 : 16,
5854 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
5856 args
->compr
= 1; /* COMPR flag */
5861 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5863 for (unsigned i
= 0; i
< 4; ++i
)
5864 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
5868 radv_export_param(struct nir_to_llvm_context
*ctx
, unsigned index
,
5869 LLVMValueRef
*values
)
5871 struct ac_export_args args
;
5873 si_llvm_init_export_args(ctx
, values
,
5874 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
5875 ac_build_export(&ctx
->ac
, &args
);
5879 radv_load_output(struct nir_to_llvm_context
*ctx
, unsigned index
, unsigned chan
)
5881 LLVMValueRef output
=
5882 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(index
, chan
)];
5884 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
5888 handle_vs_outputs_post(struct nir_to_llvm_context
*ctx
,
5889 bool export_prim_id
,
5890 struct ac_vs_output_info
*outinfo
)
5892 uint32_t param_count
= 0;
5894 unsigned pos_idx
, num_pos_exports
= 0;
5895 struct ac_export_args args
, pos_args
[4] = {};
5896 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
5899 if (ctx
->options
->key
.has_multiview_view_index
) {
5900 LLVMValueRef
* tmp_out
= &ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5902 for(unsigned i
= 0; i
< 4; ++i
)
5903 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
5904 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5907 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, ctx
->abi
.view_index
), *tmp_out
);
5908 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
5911 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
5912 sizeof(outinfo
->vs_output_param_offset
));
5914 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
5915 LLVMValueRef slots
[8];
5918 if (outinfo
->cull_dist_mask
)
5919 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
5921 i
= VARYING_SLOT_CLIP_DIST0
;
5922 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
5923 slots
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
5925 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
5926 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
5928 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
5929 target
= V_008DFC_SQ_EXP_POS
+ 3;
5930 si_llvm_init_export_args(ctx
, &slots
[4], target
, &args
);
5931 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5932 &args
, sizeof(args
));
5935 target
= V_008DFC_SQ_EXP_POS
+ 2;
5936 si_llvm_init_export_args(ctx
, &slots
[0], target
, &args
);
5937 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5938 &args
, sizeof(args
));
5942 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
5943 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
5944 for (unsigned j
= 0; j
< 4; j
++)
5945 pos_values
[j
] = radv_load_output(ctx
, VARYING_SLOT_POS
, j
);
5947 si_llvm_init_export_args(ctx
, pos_values
, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
5949 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
5950 outinfo
->writes_pointsize
= true;
5951 psize_value
= radv_load_output(ctx
, VARYING_SLOT_PSIZ
, 0);
5954 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
5955 outinfo
->writes_layer
= true;
5956 layer_value
= radv_load_output(ctx
, VARYING_SLOT_LAYER
, 0);
5959 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
5960 outinfo
->writes_viewport_index
= true;
5961 viewport_index_value
= radv_load_output(ctx
, VARYING_SLOT_VIEWPORT
, 0);
5964 if (outinfo
->writes_pointsize
||
5965 outinfo
->writes_layer
||
5966 outinfo
->writes_viewport_index
) {
5967 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
5968 (outinfo
->writes_layer
== true ? 4 : 0));
5969 pos_args
[1].valid_mask
= 0;
5970 pos_args
[1].done
= 0;
5971 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
5972 pos_args
[1].compr
= 0;
5973 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
5974 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
5975 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
5976 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
5978 if (outinfo
->writes_pointsize
== true)
5979 pos_args
[1].out
[0] = psize_value
;
5980 if (outinfo
->writes_layer
== true)
5981 pos_args
[1].out
[2] = layer_value
;
5982 if (outinfo
->writes_viewport_index
== true) {
5983 if (ctx
->options
->chip_class
>= GFX9
) {
5984 /* GFX9 has the layer in out.z[10:0] and the viewport
5985 * index in out.z[19:16].
5987 LLVMValueRef v
= viewport_index_value
;
5988 v
= ac_to_integer(&ctx
->ac
, v
);
5989 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
5990 LLVMConstInt(ctx
->ac
.i32
, 16, false),
5992 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
5993 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
5995 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
5996 pos_args
[1].enabled_channels
|= 1 << 2;
5998 pos_args
[1].out
[3] = viewport_index_value
;
5999 pos_args
[1].enabled_channels
|= 1 << 3;
6003 for (i
= 0; i
< 4; i
++) {
6004 if (pos_args
[i
].out
[0])
6009 for (i
= 0; i
< 4; i
++) {
6010 if (!pos_args
[i
].out
[0])
6013 /* Specify the target we are exporting */
6014 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
6015 if (pos_idx
== num_pos_exports
)
6016 pos_args
[i
].done
= 1;
6017 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
6020 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6021 LLVMValueRef values
[4];
6022 if (!(ctx
->output_mask
& (1ull << i
)))
6025 if (i
!= VARYING_SLOT_LAYER
&&
6026 i
!= VARYING_SLOT_PRIMITIVE_ID
&&
6027 i
< VARYING_SLOT_VAR0
)
6030 for (unsigned j
= 0; j
< 4; j
++)
6031 values
[j
] = ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
6033 radv_export_param(ctx
, param_count
, values
);
6035 outinfo
->vs_output_param_offset
[i
] = param_count
++;
6038 if (export_prim_id
) {
6039 LLVMValueRef values
[4];
6041 values
[0] = ctx
->vs_prim_id
;
6042 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
6043 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
6044 for (unsigned j
= 1; j
< 4; j
++)
6045 values
[j
] = ctx
->ac
.f32_0
;
6047 radv_export_param(ctx
, param_count
, values
);
6049 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
6050 outinfo
->export_prim_id
= true;
6053 outinfo
->pos_exports
= num_pos_exports
;
6054 outinfo
->param_exports
= param_count
;
6058 handle_es_outputs_post(struct nir_to_llvm_context
*ctx
,
6059 struct ac_es_output_info
*outinfo
)
6062 uint64_t max_output_written
= 0;
6063 LLVMValueRef lds_base
= NULL
;
6065 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6069 if (!(ctx
->output_mask
& (1ull << i
)))
6072 if (i
== VARYING_SLOT_CLIP_DIST0
)
6073 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6075 param_index
= shader_io_get_unique_index(i
);
6077 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
6080 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
6082 if (ctx
->ac
.chip_class
>= GFX9
) {
6083 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
6084 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
6085 LLVMValueRef wave_idx
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6086 LLVMConstInt(ctx
->ac
.i32
, 24, false),
6087 LLVMConstInt(ctx
->ac
.i32
, 4, false), false);
6088 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
6089 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
6090 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
6091 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
6092 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
6095 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6096 LLVMValueRef dw_addr
;
6097 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
6101 if (!(ctx
->output_mask
& (1ull << i
)))
6104 if (i
== VARYING_SLOT_CLIP_DIST0
)
6105 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6107 param_index
= shader_io_get_unique_index(i
);
6110 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
6111 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
6114 for (j
= 0; j
< length
; j
++) {
6115 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
6116 out_val
= LLVMBuildBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
6118 if (ctx
->ac
.chip_class
>= GFX9
) {
6119 ac_lds_store(&ctx
->ac
, dw_addr
,
6120 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
6121 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6123 ac_build_buffer_store_dword(&ctx
->ac
,
6126 NULL
, ctx
->es2gs_offset
,
6127 (4 * param_index
+ j
) * 4,
6135 handle_ls_outputs_post(struct nir_to_llvm_context
*ctx
)
6137 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
6138 LLVMValueRef vertex_dw_stride
= unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 13, 8);
6139 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
6140 vertex_dw_stride
, "");
6142 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6143 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
6146 if (!(ctx
->output_mask
& (1ull << i
)))
6149 if (i
== VARYING_SLOT_CLIP_DIST0
)
6150 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6151 int param
= shader_io_get_unique_index(i
);
6152 mark_tess_output(ctx
, false, param
);
6154 mark_tess_output(ctx
, false, param
+ 1);
6155 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
6156 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
6158 for (unsigned j
= 0; j
< length
; j
++) {
6159 ac_lds_store(&ctx
->ac
, dw_addr
,
6160 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], ""));
6161 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6166 struct ac_build_if_state
6168 struct nir_to_llvm_context
*ctx
;
6169 LLVMValueRef condition
;
6170 LLVMBasicBlockRef entry_block
;
6171 LLVMBasicBlockRef true_block
;
6172 LLVMBasicBlockRef false_block
;
6173 LLVMBasicBlockRef merge_block
;
6176 static LLVMBasicBlockRef
6177 ac_build_insert_new_block(struct nir_to_llvm_context
*ctx
, const char *name
)
6179 LLVMBasicBlockRef current_block
;
6180 LLVMBasicBlockRef next_block
;
6181 LLVMBasicBlockRef new_block
;
6183 /* get current basic block */
6184 current_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
6186 /* chqeck if there's another block after this one */
6187 next_block
= LLVMGetNextBasicBlock(current_block
);
6189 /* insert the new block before the next block */
6190 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
6193 /* append new block after current block */
6194 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
6195 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
6201 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
6202 struct nir_to_llvm_context
*ctx
,
6203 LLVMValueRef condition
)
6205 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
6207 memset(ifthen
, 0, sizeof *ifthen
);
6209 ifthen
->condition
= condition
;
6210 ifthen
->entry_block
= block
;
6212 /* create endif/merge basic block for the phi functions */
6213 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
6215 /* create/insert true_block before merge_block */
6216 ifthen
->true_block
=
6217 LLVMInsertBasicBlockInContext(ctx
->context
,
6218 ifthen
->merge_block
,
6221 /* successive code goes into the true block */
6222 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ifthen
->true_block
);
6226 * End a conditional.
6229 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
6231 LLVMBuilderRef builder
= ifthen
->ctx
->ac
.builder
;
6233 /* Insert branch to the merge block from current block */
6234 LLVMBuildBr(builder
, ifthen
->merge_block
);
6237 * Now patch in the various branch instructions.
6240 /* Insert the conditional branch instruction at the end of entry_block */
6241 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
6242 if (ifthen
->false_block
) {
6243 /* we have an else clause */
6244 LLVMBuildCondBr(builder
, ifthen
->condition
,
6245 ifthen
->true_block
, ifthen
->false_block
);
6248 /* no else clause */
6249 LLVMBuildCondBr(builder
, ifthen
->condition
,
6250 ifthen
->true_block
, ifthen
->merge_block
);
6253 /* Resume building code at end of the ifthen->merge_block */
6254 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
6258 write_tess_factors(struct nir_to_llvm_context
*ctx
)
6260 unsigned stride
, outer_comps
, inner_comps
;
6261 struct ac_build_if_state if_ctx
, inner_if_ctx
;
6262 LLVMValueRef invocation_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
6263 LLVMValueRef rel_patch_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
6264 unsigned tess_inner_index
, tess_outer_index
;
6265 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
6266 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
6268 emit_barrier(&ctx
->ac
, ctx
->stage
);
6270 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
6290 ac_nir_build_if(&if_ctx
, ctx
,
6291 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
6292 invocation_id
, ctx
->ac
.i32_0
, ""));
6294 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6295 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6297 mark_tess_output(ctx
, true, tess_inner_index
);
6298 mark_tess_output(ctx
, true, tess_outer_index
);
6299 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
6300 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
6301 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
6302 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
6303 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
6305 for (i
= 0; i
< 4; i
++) {
6306 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6307 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6311 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
6312 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
6313 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
6315 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
6317 for (i
= 0; i
< outer_comps
; i
++) {
6319 ac_lds_load(&ctx
->ac
, lds_outer
);
6320 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
6323 for (i
= 0; i
< inner_comps
; i
++) {
6324 inner
[i
] = out
[outer_comps
+i
] =
6325 ac_lds_load(&ctx
->ac
, lds_inner
);
6326 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
6331 /* Convert the outputs to vectors for stores. */
6332 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
6336 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
6339 buffer
= ctx
->hs_ring_tess_factor
;
6340 tf_base
= ctx
->tess_factor_offset
;
6341 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
6342 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
6343 unsigned tf_offset
= 0;
6345 if (ctx
->options
->chip_class
<= VI
) {
6346 ac_nir_build_if(&inner_if_ctx
, ctx
,
6347 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
6348 rel_patch_id
, ctx
->ac
.i32_0
, ""));
6350 /* Store the dynamic HS control word. */
6351 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
6352 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
6353 1, ctx
->ac
.i32_0
, tf_base
,
6354 0, 1, 0, true, false);
6357 ac_nir_build_endif(&inner_if_ctx
);
6360 /* Store the tessellation factors. */
6361 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
6362 MIN2(stride
, 4), byteoffset
, tf_base
,
6363 tf_offset
, 1, 0, true, false);
6365 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
6366 stride
- 4, byteoffset
, tf_base
,
6367 16 + tf_offset
, 1, 0, true, false);
6369 //store to offchip for TES to read - only if TES reads them
6370 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
6371 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
6372 LLVMValueRef tf_inner_offset
;
6373 unsigned param_outer
, param_inner
;
6375 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6376 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6377 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
6379 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
6380 util_next_power_of_two(outer_comps
));
6382 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
6383 outer_comps
, tf_outer_offset
,
6384 ctx
->oc_lds
, 0, 1, 0, true, false);
6386 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6387 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6388 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
6390 inner_vec
= inner_comps
== 1 ? inner
[0] :
6391 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
6392 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
6393 inner_comps
, tf_inner_offset
,
6394 ctx
->oc_lds
, 0, 1, 0, true, false);
6397 ac_nir_build_endif(&if_ctx
);
6401 handle_tcs_outputs_post(struct nir_to_llvm_context
*ctx
)
6403 write_tess_factors(ctx
);
6407 si_export_mrt_color(struct nir_to_llvm_context
*ctx
,
6408 LLVMValueRef
*color
, unsigned index
, bool is_last
,
6409 struct ac_export_args
*args
)
6412 si_llvm_init_export_args(ctx
, color
,
6413 V_008DFC_SQ_EXP_MRT
+ index
, args
);
6416 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
6417 args
->done
= 1; /* DONE bit */
6418 } else if (!args
->enabled_channels
)
6419 return false; /* unnecessary NULL export */
6425 radv_export_mrt_z(struct nir_to_llvm_context
*ctx
,
6426 LLVMValueRef depth
, LLVMValueRef stencil
,
6427 LLVMValueRef samplemask
)
6429 struct ac_export_args args
;
6431 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
6433 ac_build_export(&ctx
->ac
, &args
);
6437 handle_fs_outputs_post(struct nir_to_llvm_context
*ctx
)
6440 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
6441 struct ac_export_args color_args
[8];
6443 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6444 LLVMValueRef values
[4];
6447 if (!(ctx
->output_mask
& (1ull << i
)))
6450 if (i
< FRAG_RESULT_DATA0
)
6453 for (unsigned j
= 0; j
< 4; j
++)
6454 values
[j
] = ac_to_float(&ctx
->ac
,
6455 radv_load_output(ctx
, i
, j
));
6457 if (!ctx
->shader_info
->info
.ps
.writes_z
&&
6458 !ctx
->shader_info
->info
.ps
.writes_stencil
&&
6459 !ctx
->shader_info
->info
.ps
.writes_sample_mask
)
6460 last
= ctx
->output_mask
<= ((1ull << (i
+ 1)) - 1);
6462 bool ret
= si_export_mrt_color(ctx
, values
,
6463 i
- FRAG_RESULT_DATA0
,
6464 last
, &color_args
[index
]);
6469 /* Process depth, stencil, samplemask. */
6470 if (ctx
->shader_info
->info
.ps
.writes_z
) {
6471 depth
= ac_to_float(&ctx
->ac
,
6472 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
6474 if (ctx
->shader_info
->info
.ps
.writes_stencil
) {
6475 stencil
= ac_to_float(&ctx
->ac
,
6476 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
6478 if (ctx
->shader_info
->info
.ps
.writes_sample_mask
) {
6479 samplemask
= ac_to_float(&ctx
->ac
,
6480 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
6483 /* Export PS outputs. */
6484 for (unsigned i
= 0; i
< index
; i
++)
6485 ac_build_export(&ctx
->ac
, &color_args
[i
]);
6487 if (depth
|| stencil
|| samplemask
)
6488 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
6490 ac_build_export_null(&ctx
->ac
);
6494 emit_gs_epilogue(struct nir_to_llvm_context
*ctx
)
6496 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
6500 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
6501 LLVMValueRef
*addrs
)
6503 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
6505 switch (ctx
->stage
) {
6506 case MESA_SHADER_VERTEX
:
6507 if (ctx
->options
->key
.vs
.as_ls
)
6508 handle_ls_outputs_post(ctx
);
6509 else if (ctx
->options
->key
.vs
.as_es
)
6510 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
6512 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
6513 &ctx
->shader_info
->vs
.outinfo
);
6515 case MESA_SHADER_FRAGMENT
:
6516 handle_fs_outputs_post(ctx
);
6518 case MESA_SHADER_GEOMETRY
:
6519 emit_gs_epilogue(ctx
);
6521 case MESA_SHADER_TESS_CTRL
:
6522 handle_tcs_outputs_post(ctx
);
6524 case MESA_SHADER_TESS_EVAL
:
6525 if (ctx
->options
->key
.tes
.as_es
)
6526 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
6528 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
6529 &ctx
->shader_info
->tes
.outinfo
);
6536 static void ac_llvm_finalize_module(struct nir_to_llvm_context
* ctx
)
6538 LLVMPassManagerRef passmgr
;
6539 /* Create the pass manager */
6540 passmgr
= LLVMCreateFunctionPassManagerForModule(
6543 /* This pass should eliminate all the load and store instructions */
6544 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
6546 /* Add some optimization passes */
6547 LLVMAddScalarReplAggregatesPass(passmgr
);
6548 LLVMAddLICMPass(passmgr
);
6549 LLVMAddAggressiveDCEPass(passmgr
);
6550 LLVMAddCFGSimplificationPass(passmgr
);
6551 LLVMAddInstructionCombiningPass(passmgr
);
6554 LLVMInitializeFunctionPassManager(passmgr
);
6555 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
6556 LLVMFinalizeFunctionPassManager(passmgr
);
6558 LLVMDisposeBuilder(ctx
->ac
.builder
);
6559 LLVMDisposePassManager(passmgr
);
6563 ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context
*ctx
)
6565 struct ac_vs_output_info
*outinfo
;
6567 switch (ctx
->stage
) {
6568 case MESA_SHADER_FRAGMENT
:
6569 case MESA_SHADER_COMPUTE
:
6570 case MESA_SHADER_TESS_CTRL
:
6571 case MESA_SHADER_GEOMETRY
:
6573 case MESA_SHADER_VERTEX
:
6574 if (ctx
->options
->key
.vs
.as_ls
||
6575 ctx
->options
->key
.vs
.as_es
)
6577 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
6579 case MESA_SHADER_TESS_EVAL
:
6580 if (ctx
->options
->key
.vs
.as_es
)
6582 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
6585 unreachable("Unhandled shader type");
6588 ac_optimize_vs_outputs(&ctx
->ac
,
6590 outinfo
->vs_output_param_offset
,
6592 &outinfo
->param_exports
);
6596 ac_setup_rings(struct nir_to_llvm_context
*ctx
)
6598 if ((ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_es
) ||
6599 (ctx
->stage
== MESA_SHADER_TESS_EVAL
&& ctx
->options
->key
.tes
.as_es
)) {
6600 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_VS
, false));
6603 if (ctx
->is_gs_copy_shader
) {
6604 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_VS
, false));
6606 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
6608 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_GS
, false));
6609 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_GS
, false));
6611 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.v4i32
, "");
6613 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->gsvs_num_entries
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
6614 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, ctx
->ac
.i32_1
, "");
6615 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
, ctx
->gsvs_ring_stride
, "");
6616 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ctx
->gsvs_ring
, tmp
, ctx
->ac
.i32_1
, "");
6619 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
6620 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
6621 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
6622 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
6627 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
6628 const struct nir_shader
*nir
)
6630 switch (nir
->info
.stage
) {
6631 case MESA_SHADER_TESS_CTRL
:
6632 return chip_class
>= CIK
? 128 : 64;
6633 case MESA_SHADER_GEOMETRY
:
6634 return chip_class
>= GFX9
? 128 : 64;
6635 case MESA_SHADER_COMPUTE
:
6641 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
6642 nir
->info
.cs
.local_size
[1] *
6643 nir
->info
.cs
.local_size
[2];
6644 return max_workgroup_size
;
6647 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
6648 static void ac_nir_fixup_ls_hs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6650 LLVMValueRef count
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6651 LLVMConstInt(ctx
->ac
.i32
, 8, false),
6652 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6653 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
6655 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
6656 ctx
->vs_prim_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.vertex_id
, ctx
->vs_prim_id
, "");
6657 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
6658 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
6661 static void prepare_gs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6663 for(int i
= 5; i
>= 0; --i
) {
6664 ctx
->gs_vtx_offset
[i
] = ac_build_bfe(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
6665 LLVMConstInt(ctx
->ac
.i32
, (i
& 1) * 16, false),
6666 LLVMConstInt(ctx
->ac
.i32
, 16, false), false);
6669 ctx
->gs_wave_id
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6670 LLVMConstInt(ctx
->ac
.i32
, 16, false),
6671 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6674 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
6675 struct nir_shader
*nir
, struct nir_to_llvm_context
*nctx
)
6677 struct ac_nir_context ctx
= {};
6678 struct nir_function
*func
;
6687 ctx
.stage
= nir
->info
.stage
;
6689 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6691 nir_foreach_variable(variable
, &nir
->outputs
)
6692 handle_shader_output_decl(&ctx
, nir
, variable
);
6694 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6695 _mesa_key_pointer_equal
);
6696 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6697 _mesa_key_pointer_equal
);
6698 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6699 _mesa_key_pointer_equal
);
6701 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
6703 setup_locals(&ctx
, func
);
6705 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
)
6706 setup_shared(&ctx
, nir
);
6708 visit_cf_list(&ctx
, &func
->impl
->body
);
6709 phi_post_pass(&ctx
);
6711 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
6712 ctx
.abi
->emit_outputs(ctx
.abi
, RADEON_LLVM_MAX_OUTPUTS
,
6716 ralloc_free(ctx
.defs
);
6717 ralloc_free(ctx
.phis
);
6718 ralloc_free(ctx
.vars
);
6725 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
6726 struct nir_shader
*const *shaders
,
6728 struct ac_shader_variant_info
*shader_info
,
6729 const struct ac_nir_compiler_options
*options
)
6731 struct nir_to_llvm_context ctx
= {0};
6733 ctx
.options
= options
;
6734 ctx
.shader_info
= shader_info
;
6735 ctx
.context
= LLVMContextCreate();
6736 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
6738 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
6740 ctx
.ac
.module
= ctx
.module
;
6741 LLVMSetTarget(ctx
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
6743 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
6744 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
6745 LLVMSetDataLayout(ctx
.module
, data_layout_str
);
6746 LLVMDisposeTargetData(data_layout
);
6747 LLVMDisposeMessage(data_layout_str
);
6749 enum ac_float_mode float_mode
=
6750 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
6751 AC_FLOAT_MODE_DEFAULT
;
6753 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
6755 memset(shader_info
, 0, sizeof(*shader_info
));
6757 for(int i
= 0; i
< shader_count
; ++i
)
6758 ac_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
6760 for (i
= 0; i
< AC_UD_MAX_SETS
; i
++)
6761 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
6762 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
6763 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
6765 ctx
.max_workgroup_size
= 0;
6766 for (int i
= 0; i
< shader_count
; ++i
) {
6767 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
6768 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
6772 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
6773 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
6775 ctx
.abi
.inputs
= &ctx
.inputs
[0];
6776 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
6777 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
6778 ctx
.abi
.load_ubo
= radv_load_ubo
;
6779 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
6780 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
6781 ctx
.abi
.load_resource
= radv_load_resource
;
6782 ctx
.abi
.clamp_shadow_reference
= false;
6784 if (shader_count
>= 2)
6785 ac_init_exec_full_mask(&ctx
.ac
);
6787 if (ctx
.ac
.chip_class
== GFX9
&&
6788 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
6789 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
6791 for(int i
= 0; i
< shader_count
; ++i
) {
6792 ctx
.stage
= shaders
[i
]->info
.stage
;
6793 ctx
.output_mask
= 0;
6794 ctx
.tess_outputs_written
= 0;
6795 ctx
.num_output_clips
= shaders
[i
]->info
.clip_distance_array_size
;
6796 ctx
.num_output_culls
= shaders
[i
]->info
.cull_distance_array_size
;
6798 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6799 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "gs_next_vertex");
6800 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
6801 ctx
.abi
.load_inputs
= load_gs_input
;
6802 ctx
.abi
.emit_primitive
= visit_end_primitive
;
6803 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6804 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
6805 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
6806 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
6807 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
6808 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
6809 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
6810 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
6811 ctx
.abi
.load_tess_varyings
= load_tes_input
;
6812 ctx
.abi
.load_tess_coord
= load_tess_coord
;
6813 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
6814 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
6815 if (shader_info
->info
.vs
.needs_instance_id
) {
6816 if (ctx
.options
->key
.vs
.as_ls
) {
6817 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6818 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6820 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6821 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6824 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
6825 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
6826 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
6827 ctx
.abi
.load_sample_position
= load_sample_position
;
6828 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
6832 emit_barrier(&ctx
.ac
, ctx
.stage
);
6834 ac_setup_rings(&ctx
);
6836 LLVMBasicBlockRef merge_block
;
6837 if (shader_count
>= 2) {
6838 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6839 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6840 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6842 LLVMValueRef count
= ac_build_bfe(&ctx
.ac
, ctx
.merged_wave_info
,
6843 LLVMConstInt(ctx
.ac
.i32
, 8 * i
, false),
6844 LLVMConstInt(ctx
.ac
.i32
, 8, false), false);
6845 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
6846 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
6847 thread_id
, count
, "");
6848 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
6850 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
6853 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
6854 handle_fs_inputs(&ctx
, shaders
[i
]);
6855 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
6856 handle_vs_inputs(&ctx
, shaders
[i
]);
6857 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
6858 prepare_gs_input_vgprs(&ctx
);
6860 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
6861 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
6863 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
], &ctx
);
6865 if (shader_count
>= 2) {
6866 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
6867 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
6870 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6871 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
6872 shaders
[i
]->info
.cull_distance_array_size
> 4;
6873 shader_info
->gs
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
6874 shader_info
->gs
.max_gsvs_emit_size
= shader_info
->gs
.gsvs_vertex_size
*
6875 shaders
[i
]->info
.gs
.vertices_out
;
6876 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6877 shader_info
->tcs
.outputs_written
= ctx
.tess_outputs_written
;
6878 shader_info
->tcs
.patch_outputs_written
= ctx
.tess_patch_outputs_written
;
6879 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
&& ctx
.options
->key
.vs
.as_ls
) {
6880 shader_info
->vs
.outputs_written
= ctx
.tess_outputs_written
;
6884 LLVMBuildRetVoid(ctx
.ac
.builder
);
6886 if (options
->dump_preoptir
)
6887 ac_dump_module(ctx
.module
);
6889 ac_llvm_finalize_module(&ctx
);
6891 if (shader_count
== 1)
6892 ac_nir_eliminate_const_vs_outputs(&ctx
);
6897 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
6899 unsigned *retval
= (unsigned *)context
;
6900 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
6901 char *description
= LLVMGetDiagInfoDescription(di
);
6903 if (severity
== LLVMDSError
) {
6905 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
6909 LLVMDisposeMessage(description
);
6912 static unsigned ac_llvm_compile(LLVMModuleRef M
,
6913 struct ac_shader_binary
*binary
,
6914 LLVMTargetMachineRef tm
)
6916 unsigned retval
= 0;
6918 LLVMContextRef llvm_ctx
;
6919 LLVMMemoryBufferRef out_buffer
;
6920 unsigned buffer_size
;
6921 const char *buffer_data
;
6924 /* Setup Diagnostic Handler*/
6925 llvm_ctx
= LLVMGetModuleContext(M
);
6927 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
6931 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
6934 /* Process Errors/Warnings */
6936 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
6942 /* Extract Shader Code*/
6943 buffer_size
= LLVMGetBufferSize(out_buffer
);
6944 buffer_data
= LLVMGetBufferStart(out_buffer
);
6946 ac_elf_read(buffer_data
, buffer_size
, binary
);
6949 LLVMDisposeMemoryBuffer(out_buffer
);
6955 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
6956 LLVMModuleRef llvm_module
,
6957 struct ac_shader_binary
*binary
,
6958 struct ac_shader_config
*config
,
6959 struct ac_shader_variant_info
*shader_info
,
6960 gl_shader_stage stage
,
6961 bool dump_shader
, bool supports_spill
)
6964 ac_dump_module(llvm_module
);
6966 memset(binary
, 0, sizeof(*binary
));
6967 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
6969 fprintf(stderr
, "compile failed\n");
6973 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
6975 ac_shader_binary_read_config(binary
, config
, 0, supports_spill
);
6977 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
6978 LLVMDisposeModule(llvm_module
);
6979 LLVMContextDispose(ctx
);
6981 if (stage
== MESA_SHADER_FRAGMENT
) {
6982 shader_info
->num_input_vgprs
= 0;
6983 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
6984 shader_info
->num_input_vgprs
+= 2;
6985 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
6986 shader_info
->num_input_vgprs
+= 2;
6987 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
6988 shader_info
->num_input_vgprs
+= 2;
6989 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
6990 shader_info
->num_input_vgprs
+= 3;
6991 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
6992 shader_info
->num_input_vgprs
+= 2;
6993 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
6994 shader_info
->num_input_vgprs
+= 2;
6995 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
6996 shader_info
->num_input_vgprs
+= 2;
6997 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
6998 shader_info
->num_input_vgprs
+= 1;
6999 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
7000 shader_info
->num_input_vgprs
+= 1;
7001 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
7002 shader_info
->num_input_vgprs
+= 1;
7003 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
7004 shader_info
->num_input_vgprs
+= 1;
7005 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
7006 shader_info
->num_input_vgprs
+= 1;
7007 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
7008 shader_info
->num_input_vgprs
+= 1;
7009 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
7010 shader_info
->num_input_vgprs
+= 1;
7011 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
7012 shader_info
->num_input_vgprs
+= 1;
7013 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
7014 shader_info
->num_input_vgprs
+= 1;
7016 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
7018 /* +3 for scratch wave offset and VCC */
7019 config
->num_sgprs
= MAX2(config
->num_sgprs
,
7020 shader_info
->num_input_sgprs
+ 3);
7022 /* Enable 64-bit and 16-bit denormals, because there is no performance
7025 * If denormals are enabled, all floating-point output modifiers are
7028 * Don't enable denormals for 32-bit floats, because:
7029 * - Floating-point output modifiers would be ignored by the hw.
7030 * - Some opcodes don't support denormals, such as v_mad_f32. We would
7031 * have to stop using those.
7032 * - SI & CI would be very slow.
7034 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
7038 ac_fill_shader_info(struct ac_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct ac_nir_compiler_options
*options
)
7040 switch (nir
->info
.stage
) {
7041 case MESA_SHADER_COMPUTE
:
7042 for (int i
= 0; i
< 3; ++i
)
7043 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
7045 case MESA_SHADER_FRAGMENT
:
7046 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
7048 case MESA_SHADER_GEOMETRY
:
7049 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
7050 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
7051 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
7052 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
7054 case MESA_SHADER_TESS_EVAL
:
7055 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
7056 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
7057 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
7058 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
7059 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
7061 case MESA_SHADER_TESS_CTRL
:
7062 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
7064 case MESA_SHADER_VERTEX
:
7065 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
7066 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
7067 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
7068 if (options
->key
.vs
.as_ls
)
7069 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
7076 void ac_compile_nir_shader(LLVMTargetMachineRef tm
,
7077 struct ac_shader_binary
*binary
,
7078 struct ac_shader_config
*config
,
7079 struct ac_shader_variant_info
*shader_info
,
7080 struct nir_shader
*const *nir
,
7082 const struct ac_nir_compiler_options
*options
,
7086 LLVMModuleRef llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, nir_count
, shader_info
,
7089 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
, nir
[0]->info
.stage
, dump_shader
, options
->supports_spill
);
7090 for (int i
= 0; i
< nir_count
; ++i
)
7091 ac_fill_shader_info(shader_info
, nir
[i
], options
);
7093 /* Determine the ES type (VS or TES) for the GS on GFX9. */
7094 if (options
->chip_class
== GFX9
) {
7095 if (nir_count
== 2 &&
7096 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
7097 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
7103 ac_gs_copy_shader_emit(struct nir_to_llvm_context
*ctx
)
7105 LLVMValueRef vtx_offset
=
7106 LLVMBuildMul(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
7107 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
7110 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
7114 if (!(ctx
->output_mask
& (1ull << i
)))
7117 if (i
== VARYING_SLOT_CLIP_DIST0
) {
7118 /* unpack clip and cull from a single set of slots */
7119 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
7124 for (unsigned j
= 0; j
< length
; j
++) {
7125 LLVMValueRef value
, soffset
;
7127 soffset
= LLVMConstInt(ctx
->ac
.i32
,
7129 ctx
->gs_max_out_vertices
* 16 * 4, false);
7131 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->gsvs_ring
,
7133 vtx_offset
, soffset
,
7134 0, 1, 1, true, false);
7136 LLVMBuildStore(ctx
->ac
.builder
,
7137 ac_to_float(&ctx
->ac
, value
), ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)]);
7141 handle_vs_outputs_post(ctx
, false, &ctx
->shader_info
->vs
.outinfo
);
7144 void ac_create_gs_copy_shader(LLVMTargetMachineRef tm
,
7145 struct nir_shader
*geom_shader
,
7146 struct ac_shader_binary
*binary
,
7147 struct ac_shader_config
*config
,
7148 struct ac_shader_variant_info
*shader_info
,
7149 const struct ac_nir_compiler_options
*options
,
7152 struct nir_to_llvm_context ctx
= {0};
7153 ctx
.context
= LLVMContextCreate();
7154 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
7155 ctx
.options
= options
;
7156 ctx
.shader_info
= shader_info
;
7158 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
7160 ctx
.ac
.module
= ctx
.module
;
7162 ctx
.is_gs_copy_shader
= true;
7163 LLVMSetTarget(ctx
.module
, "amdgcn--");
7165 enum ac_float_mode float_mode
=
7166 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
7167 AC_FLOAT_MODE_DEFAULT
;
7169 ctx
.ac
.builder
= ac_create_builder(ctx
.context
, float_mode
);
7170 ctx
.stage
= MESA_SHADER_VERTEX
;
7172 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
7174 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
7175 ac_setup_rings(&ctx
);
7177 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
7178 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
7180 struct ac_nir_context nir_ctx
= {};
7181 nir_ctx
.ac
= ctx
.ac
;
7182 nir_ctx
.abi
= &ctx
.abi
;
7184 nir_ctx
.nctx
= &ctx
;
7187 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
7188 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
7189 handle_shader_output_decl(&nir_ctx
, geom_shader
, variable
);
7192 ac_gs_copy_shader_emit(&ctx
);
7196 LLVMBuildRetVoid(ctx
.ac
.builder
);
7198 ac_llvm_finalize_module(&ctx
);
7200 ac_compile_llvm_module(tm
, ctx
.module
, binary
, config
, shader_info
,
7202 dump_shader
, options
->supports_spill
);