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 LLVMBuilderRef builder
;
84 LLVMValueRef main_function
;
86 struct hash_table
*defs
;
87 struct hash_table
*phis
;
89 LLVMValueRef descriptor_sets
[AC_UD_MAX_SETS
];
90 LLVMValueRef ring_offsets
;
91 LLVMValueRef push_constants
;
92 LLVMValueRef view_index
;
93 LLVMValueRef num_work_groups
;
96 LLVMValueRef vertex_buffers
;
97 LLVMValueRef rel_auto_id
;
98 LLVMValueRef vs_prim_id
;
99 LLVMValueRef ls_out_layout
;
100 LLVMValueRef es2gs_offset
;
102 LLVMValueRef tcs_offchip_layout
;
103 LLVMValueRef tcs_out_offsets
;
104 LLVMValueRef tcs_out_layout
;
105 LLVMValueRef tcs_in_layout
;
107 LLVMValueRef merged_wave_info
;
108 LLVMValueRef tess_factor_offset
;
109 LLVMValueRef tes_rel_patch_id
;
113 LLVMValueRef gsvs_ring_stride
;
114 LLVMValueRef gsvs_num_entries
;
115 LLVMValueRef gs2vs_offset
;
116 LLVMValueRef gs_wave_id
;
117 LLVMValueRef gs_vtx_offset
[6];
119 LLVMValueRef esgs_ring
;
120 LLVMValueRef gsvs_ring
;
121 LLVMValueRef hs_ring_tess_offchip
;
122 LLVMValueRef hs_ring_tess_factor
;
124 LLVMValueRef sample_pos_offset
;
125 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
126 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
128 gl_shader_stage stage
;
130 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
133 uint64_t output_mask
;
134 uint8_t num_output_clips
;
135 uint8_t num_output_culls
;
137 bool is_gs_copy_shader
;
138 LLVMValueRef gs_next_vertex
;
139 unsigned gs_max_out_vertices
;
141 unsigned tes_primitive_mode
;
142 uint64_t tess_outputs_written
;
143 uint64_t tess_patch_outputs_written
;
145 uint32_t tcs_patch_outputs_read
;
146 uint64_t tcs_outputs_read
;
149 static inline struct nir_to_llvm_context
*
150 nir_to_llvm_context_from_abi(struct ac_shader_abi
*abi
)
152 struct nir_to_llvm_context
*ctx
= NULL
;
153 return container_of(abi
, ctx
, abi
);
156 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
157 const nir_deref_var
*deref
,
158 enum ac_descriptor_type desc_type
,
159 const nir_tex_instr
*instr
,
160 bool image
, bool write
);
162 static unsigned radeon_llvm_reg_index_soa(unsigned index
, unsigned chan
)
164 return (index
* 4) + chan
;
167 static unsigned shader_io_get_unique_index(gl_varying_slot slot
)
169 /* handle patch indices separate */
170 if (slot
== VARYING_SLOT_TESS_LEVEL_OUTER
)
172 if (slot
== VARYING_SLOT_TESS_LEVEL_INNER
)
174 if (slot
>= VARYING_SLOT_PATCH0
&& slot
<= VARYING_SLOT_TESS_MAX
)
175 return 2 + (slot
- VARYING_SLOT_PATCH0
);
177 if (slot
== VARYING_SLOT_POS
)
179 if (slot
== VARYING_SLOT_PSIZ
)
181 if (slot
== VARYING_SLOT_CLIP_DIST0
)
183 /* 3 is reserved for clip dist as well */
184 if (slot
>= VARYING_SLOT_VAR0
&& slot
<= VARYING_SLOT_VAR31
)
185 return 4 + (slot
- VARYING_SLOT_VAR0
);
186 unreachable("illegal slot in get unique index\n");
189 static void set_llvm_calling_convention(LLVMValueRef func
,
190 gl_shader_stage stage
)
192 enum radeon_llvm_calling_convention calling_conv
;
195 case MESA_SHADER_VERTEX
:
196 case MESA_SHADER_TESS_EVAL
:
197 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
199 case MESA_SHADER_GEOMETRY
:
200 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
202 case MESA_SHADER_TESS_CTRL
:
203 calling_conv
= HAVE_LLVM
>= 0x0500 ? RADEON_LLVM_AMDGPU_HS
: RADEON_LLVM_AMDGPU_VS
;
205 case MESA_SHADER_FRAGMENT
:
206 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
208 case MESA_SHADER_COMPUTE
:
209 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
212 unreachable("Unhandle shader type");
215 LLVMSetFunctionCallConv(func
, calling_conv
);
220 LLVMTypeRef types
[MAX_ARGS
];
221 LLVMValueRef
*assign
[MAX_ARGS
];
222 unsigned array_params_mask
;
225 uint8_t num_sgprs_used
;
226 uint8_t num_vgprs_used
;
229 enum ac_arg_regfile
{
235 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
236 LLVMValueRef
*param_ptr
)
238 assert(info
->count
< MAX_ARGS
);
240 info
->assign
[info
->count
] = param_ptr
;
241 info
->types
[info
->count
] = type
;
244 if (regfile
== ARG_SGPR
) {
245 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
248 assert(regfile
== ARG_VGPR
);
249 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
254 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
256 info
->array_params_mask
|= (1 << info
->count
);
257 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
260 static void assign_arguments(LLVMValueRef main_function
,
261 struct arg_info
*info
)
264 for (i
= 0; i
< info
->count
; i
++) {
266 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
271 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
272 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
273 unsigned num_return_elems
,
274 struct arg_info
*args
,
275 unsigned max_workgroup_size
,
278 LLVMTypeRef main_function_type
, ret_type
;
279 LLVMBasicBlockRef main_function_body
;
281 if (num_return_elems
)
282 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
283 num_return_elems
, true);
285 ret_type
= LLVMVoidTypeInContext(ctx
);
287 /* Setup the function */
289 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
290 LLVMValueRef main_function
=
291 LLVMAddFunction(module
, "main", main_function_type
);
293 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
294 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
296 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
297 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
298 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
300 if (args
->array_params_mask
& (1 << i
)) {
301 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
302 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
303 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
307 if (max_workgroup_size
) {
308 ac_llvm_add_target_dep_function_attr(main_function
,
309 "amdgpu-max-work-group-size",
313 /* These were copied from some LLVM test. */
314 LLVMAddTargetDependentFunctionAttr(main_function
,
315 "less-precise-fpmad",
317 LLVMAddTargetDependentFunctionAttr(main_function
,
320 LLVMAddTargetDependentFunctionAttr(main_function
,
323 LLVMAddTargetDependentFunctionAttr(main_function
,
326 LLVMAddTargetDependentFunctionAttr(main_function
,
327 "no-signed-zeros-fp-math",
330 return main_function
;
333 static int get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
335 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
336 type
= LLVMGetElementType(type
);
338 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
339 return LLVMGetIntTypeWidth(type
);
341 if (type
== ctx
->f16
)
343 if (type
== ctx
->f32
)
345 if (type
== ctx
->f64
)
348 unreachable("Unhandled type kind in get_elem_bits");
351 static LLVMValueRef
unpack_param(struct ac_llvm_context
*ctx
,
352 LLVMValueRef param
, unsigned rshift
,
355 LLVMValueRef value
= param
;
357 value
= LLVMBuildLShr(ctx
->builder
, value
,
358 LLVMConstInt(ctx
->i32
, rshift
, false), "");
360 if (rshift
+ bitwidth
< 32) {
361 unsigned mask
= (1 << bitwidth
) - 1;
362 value
= LLVMBuildAnd(ctx
->builder
, value
,
363 LLVMConstInt(ctx
->i32
, mask
, false), "");
368 static LLVMValueRef
get_rel_patch_id(struct nir_to_llvm_context
*ctx
)
370 switch (ctx
->stage
) {
371 case MESA_SHADER_TESS_CTRL
:
372 return unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
373 case MESA_SHADER_TESS_EVAL
:
374 return ctx
->tes_rel_patch_id
;
377 unreachable("Illegal stage");
381 /* Tessellation shaders pass outputs to the next shader using LDS.
383 * LS outputs = TCS inputs
384 * TCS outputs = TES inputs
387 * - TCS inputs for patch 0
388 * - TCS inputs for patch 1
389 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
391 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
392 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
393 * - TCS outputs for patch 1
394 * - Per-patch TCS outputs for patch 1
395 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
396 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
399 * All three shaders VS(LS), TCS, TES share the same LDS space.
402 get_tcs_in_patch_stride(struct nir_to_llvm_context
*ctx
)
404 if (ctx
->stage
== MESA_SHADER_VERTEX
)
405 return unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 0, 13);
406 else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
407 return unpack_param(&ctx
->ac
, ctx
->tcs_in_layout
, 0, 13);
415 get_tcs_out_patch_stride(struct nir_to_llvm_context
*ctx
)
417 return unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 0, 13);
421 get_tcs_out_patch0_offset(struct nir_to_llvm_context
*ctx
)
423 return LLVMBuildMul(ctx
->builder
,
424 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 0, 16),
425 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
429 get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context
*ctx
)
431 return LLVMBuildMul(ctx
->builder
,
432 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 16, 16),
433 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
437 get_tcs_in_current_patch_offset(struct nir_to_llvm_context
*ctx
)
439 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
440 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
442 return LLVMBuildMul(ctx
->builder
, patch_stride
, rel_patch_id
, "");
446 get_tcs_out_current_patch_offset(struct nir_to_llvm_context
*ctx
)
448 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
449 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
450 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
452 return LLVMBuildAdd(ctx
->builder
, patch0_offset
,
453 LLVMBuildMul(ctx
->builder
, patch_stride
,
459 get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context
*ctx
)
461 LLVMValueRef patch0_patch_data_offset
=
462 get_tcs_out_patch0_patch_data_offset(ctx
);
463 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
464 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
466 return LLVMBuildAdd(ctx
->builder
, patch0_patch_data_offset
,
467 LLVMBuildMul(ctx
->builder
, patch_stride
,
473 set_loc(struct ac_userdata_info
*ud_info
, uint8_t *sgpr_idx
, uint8_t num_sgprs
,
474 uint32_t indirect_offset
)
476 ud_info
->sgpr_idx
= *sgpr_idx
;
477 ud_info
->num_sgprs
= num_sgprs
;
478 ud_info
->indirect
= indirect_offset
> 0;
479 ud_info
->indirect_offset
= indirect_offset
;
480 *sgpr_idx
+= num_sgprs
;
484 set_loc_shader(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
487 struct ac_userdata_info
*ud_info
=
488 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
491 set_loc(ud_info
, sgpr_idx
, num_sgprs
, 0);
495 set_loc_desc(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
496 uint32_t indirect_offset
)
498 struct ac_userdata_info
*ud_info
=
499 &ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[idx
];
502 set_loc(ud_info
, sgpr_idx
, 2, indirect_offset
);
505 struct user_sgpr_info
{
506 bool need_ring_offsets
;
508 bool indirect_all_descriptor_sets
;
511 static bool needs_view_index_sgpr(struct nir_to_llvm_context
*ctx
,
512 gl_shader_stage stage
)
515 case MESA_SHADER_VERTEX
:
516 if (ctx
->shader_info
->info
.needs_multiview_view_index
||
517 (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
520 case MESA_SHADER_TESS_EVAL
:
521 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
524 case MESA_SHADER_GEOMETRY
:
525 case MESA_SHADER_TESS_CTRL
:
526 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
535 static void allocate_user_sgprs(struct nir_to_llvm_context
*ctx
,
536 gl_shader_stage stage
,
537 bool needs_view_index
,
538 struct user_sgpr_info
*user_sgpr_info
)
540 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
542 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
543 if (stage
== MESA_SHADER_GEOMETRY
||
544 stage
== MESA_SHADER_VERTEX
||
545 stage
== MESA_SHADER_TESS_CTRL
||
546 stage
== MESA_SHADER_TESS_EVAL
||
547 ctx
->is_gs_copy_shader
)
548 user_sgpr_info
->need_ring_offsets
= true;
550 if (stage
== MESA_SHADER_FRAGMENT
&&
551 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
552 user_sgpr_info
->need_ring_offsets
= true;
554 /* 2 user sgprs will nearly always be allocated for scratch/rings */
555 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
556 user_sgpr_info
->sgpr_count
+= 2;
559 /* FIXME: fix the number of user sgprs for merged shaders on GFX9 */
561 case MESA_SHADER_COMPUTE
:
562 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
563 user_sgpr_info
->sgpr_count
+= 3;
565 case MESA_SHADER_FRAGMENT
:
566 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
568 case MESA_SHADER_VERTEX
:
569 if (!ctx
->is_gs_copy_shader
) {
570 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.vs
.has_vertex_buffers
? 2 : 0;
571 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
572 user_sgpr_info
->sgpr_count
+= 3;
574 user_sgpr_info
->sgpr_count
+= 2;
577 if (ctx
->options
->key
.vs
.as_ls
)
578 user_sgpr_info
->sgpr_count
++;
580 case MESA_SHADER_TESS_CTRL
:
581 user_sgpr_info
->sgpr_count
+= 4;
583 case MESA_SHADER_TESS_EVAL
:
584 user_sgpr_info
->sgpr_count
+= 1;
586 case MESA_SHADER_GEOMETRY
:
587 user_sgpr_info
->sgpr_count
+= 2;
593 if (needs_view_index
)
594 user_sgpr_info
->sgpr_count
++;
596 if (ctx
->shader_info
->info
.loads_push_constants
)
597 user_sgpr_info
->sgpr_count
+= 2;
599 uint32_t available_sgprs
= ctx
->options
->chip_class
>= GFX9
? 32 : 16;
600 uint32_t remaining_sgprs
= available_sgprs
- user_sgpr_info
->sgpr_count
;
602 if (remaining_sgprs
/ 2 < util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
)) {
603 user_sgpr_info
->sgpr_count
+= 2;
604 user_sgpr_info
->indirect_all_descriptor_sets
= true;
606 user_sgpr_info
->sgpr_count
+= util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
) * 2;
611 declare_global_input_sgprs(struct nir_to_llvm_context
*ctx
,
612 gl_shader_stage stage
,
613 bool has_previous_stage
,
614 gl_shader_stage previous_stage
,
615 const struct user_sgpr_info
*user_sgpr_info
,
616 struct arg_info
*args
,
617 LLVMValueRef
*desc_sets
)
619 LLVMTypeRef type
= ac_array_in_const_addr_space(ctx
->ac
.i8
);
620 unsigned num_sets
= ctx
->options
->layout
?
621 ctx
->options
->layout
->num_sets
: 0;
622 unsigned stage_mask
= 1 << stage
;
624 if (has_previous_stage
)
625 stage_mask
|= 1 << previous_stage
;
627 /* 1 for each descriptor set */
628 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
629 for (unsigned i
= 0; i
< num_sets
; ++i
) {
630 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
631 add_array_arg(args
, type
,
632 &ctx
->descriptor_sets
[i
]);
636 add_array_arg(args
, ac_array_in_const_addr_space(type
), desc_sets
);
639 if (ctx
->shader_info
->info
.loads_push_constants
) {
640 /* 1 for push constants and dynamic descriptors */
641 add_array_arg(args
, type
, &ctx
->push_constants
);
646 declare_vs_specific_input_sgprs(struct nir_to_llvm_context
*ctx
,
647 gl_shader_stage stage
,
648 bool has_previous_stage
,
649 gl_shader_stage previous_stage
,
650 struct arg_info
*args
)
652 if (!ctx
->is_gs_copy_shader
&&
653 (stage
== MESA_SHADER_VERTEX
||
654 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
655 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
656 add_arg(args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
657 &ctx
->vertex_buffers
);
659 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.base_vertex
);
660 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.start_instance
);
661 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
662 add_arg(args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.draw_id
);
668 declare_vs_input_vgprs(struct nir_to_llvm_context
*ctx
, struct arg_info
*args
)
670 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.vertex_id
);
671 if (!ctx
->is_gs_copy_shader
) {
672 if (ctx
->options
->key
.vs
.as_ls
) {
673 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
674 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
676 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
677 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
679 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* unused */
684 declare_tes_input_vgprs(struct nir_to_llvm_context
*ctx
, struct arg_info
*args
)
686 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
687 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
688 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
689 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
693 set_global_input_locs(struct nir_to_llvm_context
*ctx
, gl_shader_stage stage
,
694 bool has_previous_stage
, gl_shader_stage previous_stage
,
695 const struct user_sgpr_info
*user_sgpr_info
,
696 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
698 unsigned num_sets
= ctx
->options
->layout
?
699 ctx
->options
->layout
->num_sets
: 0;
700 unsigned stage_mask
= 1 << stage
;
702 if (has_previous_stage
)
703 stage_mask
|= 1 << previous_stage
;
705 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
706 for (unsigned i
= 0; i
< num_sets
; ++i
) {
707 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
708 set_loc_desc(ctx
, i
, user_sgpr_idx
, 0);
710 ctx
->descriptor_sets
[i
] = NULL
;
713 set_loc_shader(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
716 for (unsigned i
= 0; i
< num_sets
; ++i
) {
717 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
718 set_loc_desc(ctx
, i
, user_sgpr_idx
, i
* 8);
719 ctx
->descriptor_sets
[i
] =
720 ac_build_load_to_sgpr(&ctx
->ac
,
722 LLVMConstInt(ctx
->ac
.i32
, i
, false));
725 ctx
->descriptor_sets
[i
] = NULL
;
727 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
730 if (ctx
->shader_info
->info
.loads_push_constants
) {
731 set_loc_shader(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
, 2);
736 set_vs_specific_input_locs(struct nir_to_llvm_context
*ctx
,
737 gl_shader_stage stage
, bool has_previous_stage
,
738 gl_shader_stage previous_stage
,
739 uint8_t *user_sgpr_idx
)
741 if (!ctx
->is_gs_copy_shader
&&
742 (stage
== MESA_SHADER_VERTEX
||
743 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
744 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
745 set_loc_shader(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
750 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
753 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
754 user_sgpr_idx
, vs_num
);
758 static void create_function(struct nir_to_llvm_context
*ctx
,
759 gl_shader_stage stage
,
760 bool has_previous_stage
,
761 gl_shader_stage previous_stage
)
763 uint8_t user_sgpr_idx
;
764 struct user_sgpr_info user_sgpr_info
;
765 struct arg_info args
= {};
766 LLVMValueRef desc_sets
;
767 bool needs_view_index
= needs_view_index_sgpr(ctx
, stage
);
768 allocate_user_sgprs(ctx
, stage
, needs_view_index
, &user_sgpr_info
);
770 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
771 add_arg(&args
, ARG_SGPR
, ac_array_in_const_addr_space(ctx
->ac
.v4i32
),
776 case MESA_SHADER_COMPUTE
:
777 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
778 previous_stage
, &user_sgpr_info
,
781 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
782 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
783 &ctx
->num_work_groups
);
786 for (int i
= 0; i
< 3; i
++) {
787 ctx
->abi
.workgroup_ids
[i
] = NULL
;
788 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
789 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
790 &ctx
->abi
.workgroup_ids
[i
]);
794 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
795 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->tg_size
);
796 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
797 &ctx
->abi
.local_invocation_ids
);
799 case MESA_SHADER_VERTEX
:
800 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
801 previous_stage
, &user_sgpr_info
,
803 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
804 previous_stage
, &args
);
806 if (needs_view_index
)
807 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->view_index
);
808 if (ctx
->options
->key
.vs
.as_es
)
809 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
811 else if (ctx
->options
->key
.vs
.as_ls
)
812 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
813 &ctx
->ls_out_layout
);
815 declare_vs_input_vgprs(ctx
, &args
);
817 case MESA_SHADER_TESS_CTRL
:
818 if (has_previous_stage
) {
819 // First 6 system regs
820 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
821 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
822 &ctx
->merged_wave_info
);
823 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
824 &ctx
->tess_factor_offset
);
826 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
827 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
828 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
830 declare_global_input_sgprs(ctx
, stage
,
833 &user_sgpr_info
, &args
,
835 declare_vs_specific_input_sgprs(ctx
, stage
,
837 previous_stage
, &args
);
839 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
840 &ctx
->ls_out_layout
);
842 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
843 &ctx
->tcs_offchip_layout
);
844 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
845 &ctx
->tcs_out_offsets
);
846 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
847 &ctx
->tcs_out_layout
);
848 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
849 &ctx
->tcs_in_layout
);
850 if (needs_view_index
)
851 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
854 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
855 &ctx
->abi
.tcs_patch_id
);
856 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
857 &ctx
->abi
.tcs_rel_ids
);
859 declare_vs_input_vgprs(ctx
, &args
);
861 declare_global_input_sgprs(ctx
, stage
,
864 &user_sgpr_info
, &args
,
867 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
868 &ctx
->tcs_offchip_layout
);
869 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
870 &ctx
->tcs_out_offsets
);
871 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
872 &ctx
->tcs_out_layout
);
873 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
874 &ctx
->tcs_in_layout
);
875 if (needs_view_index
)
876 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
879 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
880 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
881 &ctx
->tess_factor_offset
);
882 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
883 &ctx
->abi
.tcs_patch_id
);
884 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
885 &ctx
->abi
.tcs_rel_ids
);
888 case MESA_SHADER_TESS_EVAL
:
889 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
890 previous_stage
, &user_sgpr_info
,
893 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->tcs_offchip_layout
);
894 if (needs_view_index
)
895 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->view_index
);
897 if (ctx
->options
->key
.tes
.as_es
) {
898 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
899 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
900 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
903 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
904 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
906 declare_tes_input_vgprs(ctx
, &args
);
908 case MESA_SHADER_GEOMETRY
:
909 if (has_previous_stage
) {
910 // First 6 system regs
911 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
913 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
914 &ctx
->merged_wave_info
);
915 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
917 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
918 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
919 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
921 declare_global_input_sgprs(ctx
, stage
,
924 &user_sgpr_info
, &args
,
927 if (previous_stage
== MESA_SHADER_TESS_EVAL
) {
928 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
929 &ctx
->tcs_offchip_layout
);
931 declare_vs_specific_input_sgprs(ctx
, stage
,
937 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
938 &ctx
->gsvs_ring_stride
);
939 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
940 &ctx
->gsvs_num_entries
);
941 if (needs_view_index
)
942 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
945 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
946 &ctx
->gs_vtx_offset
[0]);
947 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
948 &ctx
->gs_vtx_offset
[2]);
949 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
950 &ctx
->abi
.gs_prim_id
);
951 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
952 &ctx
->abi
.gs_invocation_id
);
953 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
954 &ctx
->gs_vtx_offset
[4]);
956 if (previous_stage
== MESA_SHADER_VERTEX
) {
957 declare_vs_input_vgprs(ctx
, &args
);
959 declare_tes_input_vgprs(ctx
, &args
);
962 declare_global_input_sgprs(ctx
, stage
,
965 &user_sgpr_info
, &args
,
968 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
969 &ctx
->gsvs_ring_stride
);
970 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
971 &ctx
->gsvs_num_entries
);
972 if (needs_view_index
)
973 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
976 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
977 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
978 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
979 &ctx
->gs_vtx_offset
[0]);
980 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
981 &ctx
->gs_vtx_offset
[1]);
982 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
983 &ctx
->abi
.gs_prim_id
);
984 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
985 &ctx
->gs_vtx_offset
[2]);
986 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
987 &ctx
->gs_vtx_offset
[3]);
988 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
989 &ctx
->gs_vtx_offset
[4]);
990 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
991 &ctx
->gs_vtx_offset
[5]);
992 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
993 &ctx
->abi
.gs_invocation_id
);
996 case MESA_SHADER_FRAGMENT
:
997 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
998 previous_stage
, &user_sgpr_info
,
1001 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
1002 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
1003 &ctx
->sample_pos_offset
);
1005 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->abi
.prim_mask
);
1006 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1007 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1008 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1009 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1010 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1011 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1012 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1013 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1014 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1015 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1016 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1017 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1018 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1019 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1020 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1021 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1024 unreachable("Shader stage not implemented");
1027 ctx
->main_function
= create_llvm_function(
1028 ctx
->context
, ctx
->module
, ctx
->builder
, NULL
, 0, &args
,
1029 ctx
->max_workgroup_size
,
1030 ctx
->options
->unsafe_math
);
1031 set_llvm_calling_convention(ctx
->main_function
, stage
);
1034 ctx
->shader_info
->num_input_vgprs
= 0;
1035 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1037 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1039 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1040 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1042 assign_arguments(ctx
->main_function
, &args
);
1046 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1047 set_loc_shader(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1049 if (ctx
->options
->supports_spill
) {
1050 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1051 LLVMPointerType(ctx
->ac
.i8
, AC_CONST_ADDR_SPACE
),
1052 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1053 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->builder
, ctx
->ring_offsets
,
1054 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
1058 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1059 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1060 if (has_previous_stage
)
1063 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1064 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1067 case MESA_SHADER_COMPUTE
:
1068 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1069 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1073 case MESA_SHADER_VERTEX
:
1074 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1075 previous_stage
, &user_sgpr_idx
);
1076 if (ctx
->view_index
)
1077 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1078 if (ctx
->options
->key
.vs
.as_ls
) {
1079 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1082 if (ctx
->options
->key
.vs
.as_ls
)
1083 ac_declare_lds_as_pointer(&ctx
->ac
);
1085 case MESA_SHADER_TESS_CTRL
:
1086 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1087 previous_stage
, &user_sgpr_idx
);
1088 if (has_previous_stage
)
1089 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1091 set_loc_shader(ctx
, AC_UD_TCS_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 4);
1092 if (ctx
->view_index
)
1093 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1094 ac_declare_lds_as_pointer(&ctx
->ac
);
1096 case MESA_SHADER_TESS_EVAL
:
1097 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 1);
1098 if (ctx
->view_index
)
1099 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1101 case MESA_SHADER_GEOMETRY
:
1102 if (has_previous_stage
) {
1103 if (previous_stage
== MESA_SHADER_VERTEX
)
1104 set_vs_specific_input_locs(ctx
, stage
,
1109 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
,
1112 set_loc_shader(ctx
, AC_UD_GS_VS_RING_STRIDE_ENTRIES
,
1114 if (ctx
->view_index
)
1115 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1116 if (has_previous_stage
)
1117 ac_declare_lds_as_pointer(&ctx
->ac
);
1119 case MESA_SHADER_FRAGMENT
:
1120 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
1121 set_loc_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
,
1126 unreachable("Shader stage not implemented");
1129 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1132 static LLVMValueRef
trim_vector(struct ac_llvm_context
*ctx
,
1133 LLVMValueRef value
, unsigned count
)
1135 unsigned num_components
= ac_get_llvm_num_components(value
);
1136 if (count
== num_components
)
1139 LLVMValueRef masks
[] = {
1140 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
1141 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false)};
1144 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
1147 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
1148 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
1152 build_store_values_extended(struct ac_llvm_context
*ac
,
1153 LLVMValueRef
*values
,
1154 unsigned value_count
,
1155 unsigned value_stride
,
1158 LLVMBuilderRef builder
= ac
->builder
;
1161 for (i
= 0; i
< value_count
; i
++) {
1162 LLVMValueRef ptr
= values
[i
* value_stride
];
1163 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
1164 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
1165 LLVMBuildStore(builder
, value
, ptr
);
1169 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
1170 const nir_ssa_def
*def
)
1172 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
1173 if (def
->num_components
> 1) {
1174 type
= LLVMVectorType(type
, def
->num_components
);
1179 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
1182 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, src
.ssa
);
1183 return (LLVMValueRef
)entry
->data
;
1187 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
1188 const struct nir_block
*b
)
1190 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
1191 return (LLVMBasicBlockRef
)entry
->data
;
1194 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
1196 unsigned num_components
)
1198 LLVMValueRef value
= get_src(ctx
, src
.src
);
1199 bool need_swizzle
= false;
1202 LLVMTypeRef type
= LLVMTypeOf(value
);
1203 unsigned src_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
1204 ? LLVMGetVectorSize(type
)
1207 for (unsigned i
= 0; i
< num_components
; ++i
) {
1208 assert(src
.swizzle
[i
] < src_components
);
1209 if (src
.swizzle
[i
] != i
)
1210 need_swizzle
= true;
1213 if (need_swizzle
|| num_components
!= src_components
) {
1214 LLVMValueRef masks
[] = {
1215 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
1216 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
1217 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
1218 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
1220 if (src_components
> 1 && num_components
== 1) {
1221 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1223 } else if (src_components
== 1 && num_components
> 1) {
1224 LLVMValueRef values
[] = {value
, value
, value
, value
};
1225 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
1227 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
1228 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
1232 assert(!src
.negate
);
1237 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
1238 LLVMIntPredicate pred
, LLVMValueRef src0
,
1241 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
1242 return LLVMBuildSelect(ctx
->builder
, result
,
1243 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1247 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
1248 LLVMRealPredicate pred
, LLVMValueRef src0
,
1251 LLVMValueRef result
;
1252 src0
= ac_to_float(ctx
, src0
);
1253 src1
= ac_to_float(ctx
, src1
);
1254 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
1255 return LLVMBuildSelect(ctx
->builder
, result
,
1256 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1260 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
1262 LLVMTypeRef result_type
,
1266 LLVMValueRef params
[] = {
1267 ac_to_float(ctx
, src0
),
1270 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1271 get_elem_bits(ctx
, result_type
));
1272 assert(length
< sizeof(name
));
1273 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
1276 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
1278 LLVMTypeRef result_type
,
1279 LLVMValueRef src0
, LLVMValueRef src1
)
1282 LLVMValueRef params
[] = {
1283 ac_to_float(ctx
, src0
),
1284 ac_to_float(ctx
, src1
),
1287 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1288 get_elem_bits(ctx
, result_type
));
1289 assert(length
< sizeof(name
));
1290 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
1293 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
1295 LLVMTypeRef result_type
,
1296 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1299 LLVMValueRef params
[] = {
1300 ac_to_float(ctx
, src0
),
1301 ac_to_float(ctx
, src1
),
1302 ac_to_float(ctx
, src2
),
1305 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1306 get_elem_bits(ctx
, result_type
));
1307 assert(length
< sizeof(name
));
1308 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
1311 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
1312 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1314 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
1316 return LLVMBuildSelect(ctx
->builder
, v
, src1
, src2
, "");
1319 static LLVMValueRef
emit_minmax_int(struct ac_llvm_context
*ctx
,
1320 LLVMIntPredicate pred
,
1321 LLVMValueRef src0
, LLVMValueRef src1
)
1323 return LLVMBuildSelect(ctx
->builder
,
1324 LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, ""),
1329 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
1332 return emit_minmax_int(ctx
, LLVMIntSGT
, src0
,
1333 LLVMBuildNeg(ctx
->builder
, src0
, ""));
1336 static LLVMValueRef
emit_fsign(struct ac_llvm_context
*ctx
,
1340 LLVMValueRef cmp
, val
, zero
, one
;
1343 if (bitsize
== 32) {
1353 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
1354 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
1355 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
1356 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
1360 static LLVMValueRef
emit_isign(struct ac_llvm_context
*ctx
,
1361 LLVMValueRef src0
, unsigned bitsize
)
1363 LLVMValueRef cmp
, val
, zero
, one
;
1366 if (bitsize
== 32) {
1376 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
1377 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
1378 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
1379 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
1383 static LLVMValueRef
emit_ffract(struct ac_llvm_context
*ctx
,
1384 LLVMValueRef src0
, unsigned bitsize
)
1389 if (bitsize
== 32) {
1390 intr
= "llvm.floor.f32";
1393 intr
= "llvm.floor.f64";
1397 LLVMValueRef fsrc0
= ac_to_float(ctx
, src0
);
1398 LLVMValueRef params
[] = {
1401 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
1402 AC_FUNC_ATTR_READNONE
);
1403 return LLVMBuildFSub(ctx
->builder
, fsrc0
, floor
, "");
1406 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
1408 LLVMValueRef src0
, LLVMValueRef src1
)
1410 LLVMTypeRef ret_type
;
1411 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
1413 LLVMValueRef params
[] = { src0
, src1
};
1414 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
1417 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
1418 params
, 2, AC_FUNC_ATTR_READNONE
);
1420 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
1421 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
1425 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
1428 return LLVMBuildAnd(ctx
->builder
, src0
, LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""), "");
1431 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
1434 src0
= ac_to_float(ctx
, src0
);
1435 return LLVMBuildSExt(ctx
->builder
,
1436 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, ctx
->f32_0
, ""),
1440 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
1444 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
1449 return LLVMBuildZExt(ctx
->builder
, result
, ctx
->i64
, "");
1452 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
1455 return LLVMBuildSExt(ctx
->builder
,
1456 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, ctx
->i32_0
, ""),
1460 static LLVMValueRef
emit_f2f16(struct nir_to_llvm_context
*ctx
,
1463 LLVMValueRef result
;
1464 LLVMValueRef cond
= NULL
;
1466 src0
= ac_to_float(&ctx
->ac
, src0
);
1467 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->ac
.f16
, "");
1469 if (ctx
->options
->chip_class
>= VI
) {
1470 LLVMValueRef args
[2];
1471 /* Check if the result is a denormal - and flush to 0 if so. */
1473 args
[1] = LLVMConstInt(ctx
->ac
.i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
1474 cond
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.class.f16", ctx
->ac
.i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
1477 /* need to convert back up to f32 */
1478 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->ac
.f32
, "");
1480 if (ctx
->options
->chip_class
>= VI
)
1481 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->ac
.f32_0
, result
, "");
1484 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
1485 * so compare the result and flush to 0 if it's smaller.
1487 LLVMValueRef temp
, cond2
;
1488 temp
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1489 ctx
->ac
.f32
, result
);
1490 cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUGT
,
1491 LLVMBuildBitCast(ctx
->builder
, LLVMConstInt(ctx
->ac
.i32
, 0x38800000, false), ctx
->ac
.f32
, ""),
1493 cond2
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
1494 temp
, ctx
->ac
.f32_0
, "");
1495 cond
= LLVMBuildAnd(ctx
->builder
, cond
, cond2
, "");
1496 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->ac
.f32_0
, result
, "");
1501 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
1502 LLVMValueRef src0
, LLVMValueRef src1
)
1504 LLVMValueRef dst64
, result
;
1505 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
1506 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
1508 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1509 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1510 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1514 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
1515 LLVMValueRef src0
, LLVMValueRef src1
)
1517 LLVMValueRef dst64
, result
;
1518 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
1519 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
1521 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1522 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1523 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1527 static LLVMValueRef
emit_bitfield_extract(struct ac_llvm_context
*ctx
,
1529 const LLVMValueRef srcs
[3])
1531 LLVMValueRef result
;
1532 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
1534 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
1535 result
= LLVMBuildSelect(ctx
->builder
, icond
, srcs
[0], result
, "");
1539 static LLVMValueRef
emit_bitfield_insert(struct ac_llvm_context
*ctx
,
1540 LLVMValueRef src0
, LLVMValueRef src1
,
1541 LLVMValueRef src2
, LLVMValueRef src3
)
1543 LLVMValueRef bfi_args
[3], result
;
1545 bfi_args
[0] = LLVMBuildShl(ctx
->builder
,
1546 LLVMBuildSub(ctx
->builder
,
1547 LLVMBuildShl(ctx
->builder
,
1552 bfi_args
[1] = LLVMBuildShl(ctx
->builder
, src1
, src2
, "");
1555 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, src3
, LLVMConstInt(ctx
->i32
, 32, false), "");
1558 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
1559 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
1561 result
= LLVMBuildXor(ctx
->builder
, bfi_args
[2],
1562 LLVMBuildAnd(ctx
->builder
, bfi_args
[0],
1563 LLVMBuildXor(ctx
->builder
, bfi_args
[1], bfi_args
[2], ""), ""), "");
1565 result
= LLVMBuildSelect(ctx
->builder
, icond
, src1
, result
, "");
1569 static LLVMValueRef
emit_pack_half_2x16(struct ac_llvm_context
*ctx
,
1572 LLVMValueRef comp
[2];
1574 src0
= ac_to_float(ctx
, src0
);
1575 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
1576 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
1578 return ac_build_cvt_pkrtz_f16(ctx
, comp
);
1581 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
1584 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
1585 LLVMValueRef temps
[2], result
, val
;
1588 for (i
= 0; i
< 2; i
++) {
1589 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
1590 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
1591 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
1592 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
1595 result
= LLVMBuildInsertElement(ctx
->builder
, LLVMGetUndef(ctx
->v2f32
), temps
[0],
1597 result
= LLVMBuildInsertElement(ctx
->builder
, result
, temps
[1],
1602 static LLVMValueRef
emit_ddxy(struct ac_nir_context
*ctx
,
1608 LLVMValueRef result
;
1610 if (op
== nir_op_fddx_fine
|| op
== nir_op_fddx
)
1611 mask
= AC_TID_MASK_LEFT
;
1612 else if (op
== nir_op_fddy_fine
|| op
== nir_op_fddy
)
1613 mask
= AC_TID_MASK_TOP
;
1615 mask
= AC_TID_MASK_TOP_LEFT
;
1617 /* for DDX we want to next X pixel, DDY next Y pixel. */
1618 if (op
== nir_op_fddx_fine
||
1619 op
== nir_op_fddx_coarse
||
1625 result
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, src0
);
1630 * this takes an I,J coordinate pair,
1631 * and works out the X and Y derivatives.
1632 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
1634 static LLVMValueRef
emit_ddxy_interp(
1635 struct ac_nir_context
*ctx
,
1636 LLVMValueRef interp_ij
)
1638 LLVMValueRef result
[4], a
;
1641 for (i
= 0; i
< 2; i
++) {
1642 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
1643 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1644 result
[i
] = emit_ddxy(ctx
, nir_op_fddx
, a
);
1645 result
[2+i
] = emit_ddxy(ctx
, nir_op_fddy
, a
);
1647 return ac_build_gather_values(&ctx
->ac
, result
, 4);
1650 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
1652 LLVMValueRef src
[4], result
= NULL
;
1653 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
1654 unsigned src_components
;
1655 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
1657 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
1658 switch (instr
->op
) {
1664 case nir_op_pack_half_2x16
:
1667 case nir_op_unpack_half_2x16
:
1671 src_components
= num_components
;
1674 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1675 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
1677 switch (instr
->op
) {
1683 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1684 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
1687 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
1690 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
1693 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1696 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1697 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1698 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1701 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1702 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1703 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1706 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1709 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1712 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1715 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
1718 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1719 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1720 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1721 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1722 ac_to_float_type(&ctx
->ac
, def_type
), result
);
1723 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
1724 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
1727 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1728 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1729 result
= LLVMBuildFRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1732 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1735 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1738 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1741 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1742 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1743 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1746 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1747 result
= ac_build_fdiv(&ctx
->ac
, instr
->dest
.dest
.ssa
.bit_size
== 32 ? ctx
->ac
.f32_1
: ctx
->ac
.f64_1
,
1751 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
1754 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
1757 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
1760 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0],
1761 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1762 LLVMTypeOf(src
[0]), ""),
1766 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0],
1767 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1768 LLVMTypeOf(src
[0]), ""),
1772 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0],
1773 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1774 LLVMTypeOf(src
[0]), ""),
1778 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1781 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
1784 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
1787 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
1790 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1793 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
1796 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUEQ
, src
[0], src
[1]);
1799 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
1802 result
= emit_float_cmp(&ctx
->ac
, LLVMRealULT
, src
[0], src
[1]);
1805 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUGE
, src
[0], src
[1]);
1808 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1809 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1812 result
= emit_iabs(&ctx
->ac
, src
[0]);
1815 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, src
[0], src
[1]);
1818 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1821 result
= emit_minmax_int(&ctx
->ac
, LLVMIntUGT
, src
[0], src
[1]);
1824 result
= emit_minmax_int(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1827 result
= emit_isign(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1830 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1831 result
= emit_fsign(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1834 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1835 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1838 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
1839 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1842 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
1843 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1845 case nir_op_fround_even
:
1846 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
1847 ac_to_float_type(&ctx
->ac
, def_type
),src
[0]);
1850 result
= emit_ffract(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1853 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
1854 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1857 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
1858 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1861 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1862 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1865 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
1866 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1869 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
1870 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1873 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1874 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1875 result
= ac_build_fdiv(&ctx
->ac
, instr
->dest
.dest
.ssa
.bit_size
== 32 ? ctx
->ac
.f32_1
: ctx
->ac
.f64_1
,
1879 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
1880 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1883 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1884 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1885 if (ctx
->ac
.chip_class
< GFX9
&&
1886 instr
->dest
.dest
.ssa
.bit_size
== 32) {
1887 /* Only pre-GFX9 chips do not flush denorms. */
1888 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1889 ac_to_float_type(&ctx
->ac
, def_type
),
1894 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1895 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1896 if (ctx
->ac
.chip_class
< GFX9
&&
1897 instr
->dest
.dest
.ssa
.bit_size
== 32) {
1898 /* Only pre-GFX9 chips do not flush denorms. */
1899 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1900 ac_to_float_type(&ctx
->ac
, def_type
),
1905 result
= emit_intrin_3f_param(&ctx
->ac
, "llvm.fmuladd",
1906 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
1908 case nir_op_ibitfield_extract
:
1909 result
= emit_bitfield_extract(&ctx
->ac
, true, src
);
1911 case nir_op_ubitfield_extract
:
1912 result
= emit_bitfield_extract(&ctx
->ac
, false, src
);
1914 case nir_op_bitfield_insert
:
1915 result
= emit_bitfield_insert(&ctx
->ac
, src
[0], src
[1], src
[2], src
[3]);
1917 case nir_op_bitfield_reverse
:
1918 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.bitreverse.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1920 case nir_op_bit_count
:
1921 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1926 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1927 src
[i
] = ac_to_integer(&ctx
->ac
, src
[i
]);
1928 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
1932 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1933 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
1937 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1938 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
1942 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1943 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1947 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1948 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1951 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1952 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
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 (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < 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 (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < 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
->nctx
, 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(instr
->src
[0].src
.ssa
->num_components
== 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(instr
->src
[0].src
.ssa
->num_components
== 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 nir_to_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
);
2352 static LLVMValueRef
visit_vulkan_resource_index(struct nir_to_llvm_context
*ctx
,
2353 nir_intrinsic_instr
*instr
)
2355 LLVMValueRef index
= get_src(ctx
->nir
, instr
->src
[0]);
2356 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
2357 unsigned binding
= nir_intrinsic_binding(instr
);
2358 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
2359 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
2360 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
2361 unsigned base_offset
= layout
->binding
[binding
].offset
;
2362 LLVMValueRef offset
, stride
;
2364 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
2365 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
2366 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
2367 layout
->binding
[binding
].dynamic_offset_offset
;
2368 desc_ptr
= ctx
->push_constants
;
2369 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
2370 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2372 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
2374 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
2375 index
= LLVMBuildMul(ctx
->builder
, index
, stride
, "");
2376 offset
= LLVMBuildAdd(ctx
->builder
, offset
, index
, "");
2378 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
2379 desc_ptr
= cast_ptr(ctx
, desc_ptr
, ctx
->ac
.v4i32
);
2380 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2385 static LLVMValueRef
visit_vulkan_resource_reindex(struct nir_to_llvm_context
*ctx
,
2386 nir_intrinsic_instr
*instr
)
2388 LLVMValueRef ptr
= get_src(ctx
->nir
, instr
->src
[0]);
2389 LLVMValueRef index
= get_src(ctx
->nir
, instr
->src
[1]);
2391 LLVMValueRef result
= LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, "");
2392 LLVMSetMetadata(result
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2396 static LLVMValueRef
visit_load_push_constant(struct nir_to_llvm_context
*ctx
,
2397 nir_intrinsic_instr
*instr
)
2399 LLVMValueRef ptr
, addr
;
2401 addr
= LLVMConstInt(ctx
->ac
.i32
, nir_intrinsic_base(instr
), 0);
2402 addr
= LLVMBuildAdd(ctx
->builder
, addr
, get_src(ctx
->nir
, instr
->src
[0]), "");
2404 ptr
= ac_build_gep0(&ctx
->ac
, ctx
->push_constants
, addr
);
2405 ptr
= cast_ptr(ctx
, ptr
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
));
2407 return LLVMBuildLoad(ctx
->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
= 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
->builder
, vertices_per_patch
,
2776 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2778 base_addr
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
2779 vertices_per_patch
, "");
2781 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2784 param_stride
= total_vertices
;
2786 base_addr
= rel_patch_id
;
2787 param_stride
= num_patches
;
2790 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2791 LLVMBuildMul(ctx
->builder
, param_index
,
2792 param_stride
, ""), "");
2794 base_addr
= LLVMBuildMul(ctx
->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
->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
->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
->builder
, dw_addr
,
2851 LLVMBuildMul(ctx
->builder
,
2857 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2858 LLVMBuildMul(ctx
->builder
, indir_index
,
2859 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
2860 else if (const_index
&& !compact_const_index
)
2861 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2862 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
2864 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2865 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
2867 if (const_index
&& compact_const_index
)
2868 dw_addr
= LLVMBuildAdd(ctx
->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
->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
->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
->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
->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
->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 (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 nir_to_llvm_context
*ctx
,
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(&ctx
->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 nir_to_llvm_context
*ctx
)
3904 LLVMValueRef result
;
3905 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
3906 result
= LLVMBuildAnd(ctx
->builder
, ctx
->tg_size
,
3907 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
3909 return LLVMBuildAdd(ctx
->builder
, result
, thread_id
, "");
3912 static LLVMValueRef
visit_var_atomic(struct nir_to_llvm_context
*ctx
,
3913 const nir_intrinsic_instr
*instr
)
3915 LLVMValueRef ptr
, result
;
3916 LLVMValueRef src
= get_src(ctx
->nir
, instr
->src
[0]);
3917 ptr
= build_gep_for_deref(ctx
->nir
, instr
->variables
[0]);
3919 if (instr
->intrinsic
== nir_intrinsic_var_atomic_comp_swap
) {
3920 LLVMValueRef src1
= get_src(ctx
->nir
, instr
->src
[1]);
3921 result
= LLVMBuildAtomicCmpXchg(ctx
->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
->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
->builder
, ptr
,
4006 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
4008 sample_id
= LLVMBuildAdd(ctx
->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_nir_context
*ctx
)
4025 uint8_t log2_ps_iter_samples
= ctx
->nctx
->shader_info
->info
.ps
.force_persample
? ctx
->nctx
->options
->key
.fs
.log2_num_samples
: ctx
->nctx
->options
->key
.fs
.log2_ps_iter_samples
;
4027 /* The bit pattern matches that used by fixed function fragment
4029 static const uint16_t ps_iter_masks
[] = {
4030 0xffff, /* not used */
4036 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
4038 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
4040 LLVMValueRef result
, sample_id
;
4041 sample_id
= unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
4042 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
4043 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
, ctx
->abi
->sample_coverage
, "");
4047 static LLVMValueRef
visit_interp(struct ac_nir_context
*ctx
,
4048 const nir_intrinsic_instr
*instr
)
4050 LLVMValueRef result
[4];
4051 LLVMValueRef interp_param
, attr_number
;
4054 LLVMValueRef src_c0
= NULL
;
4055 LLVMValueRef src_c1
= NULL
;
4056 LLVMValueRef src0
= NULL
;
4057 int input_index
= instr
->variables
[0]->var
->data
.location
- VARYING_SLOT_VAR0
;
4058 switch (instr
->intrinsic
) {
4059 case nir_intrinsic_interp_var_at_centroid
:
4060 location
= INTERP_CENTROID
;
4062 case nir_intrinsic_interp_var_at_sample
:
4063 case nir_intrinsic_interp_var_at_offset
:
4064 location
= INTERP_CENTER
;
4065 src0
= get_src(ctx
, instr
->src
[0]);
4071 if (instr
->intrinsic
== nir_intrinsic_interp_var_at_offset
) {
4072 src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_0
, ""));
4073 src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_1
, ""));
4074 } else if (instr
->intrinsic
== nir_intrinsic_interp_var_at_sample
) {
4075 LLVMValueRef sample_position
;
4076 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
4078 /* fetch sample ID */
4079 sample_position
= ctx
->abi
->load_sample_position(ctx
->abi
, src0
);
4081 src_c0
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_0
, "");
4082 src_c0
= LLVMBuildFSub(ctx
->ac
.builder
, src_c0
, halfval
, "");
4083 src_c1
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_1
, "");
4084 src_c1
= LLVMBuildFSub(ctx
->ac
.builder
, src_c1
, halfval
, "");
4086 interp_param
= ctx
->abi
->lookup_interp_param(ctx
->abi
, instr
->variables
[0]->var
->data
.interpolation
, location
);
4087 attr_number
= LLVMConstInt(ctx
->ac
.i32
, input_index
, false);
4089 if (location
== INTERP_CENTER
) {
4090 LLVMValueRef ij_out
[2];
4091 LLVMValueRef ddxy_out
= emit_ddxy_interp(ctx
, interp_param
);
4094 * take the I then J parameters, and the DDX/Y for it, and
4095 * calculate the IJ inputs for the interpolator.
4096 * temp1 = ddx * offset/sample.x + I;
4097 * interp_param.I = ddy * offset/sample.y + temp1;
4098 * temp1 = ddx * offset/sample.x + J;
4099 * interp_param.J = ddy * offset/sample.y + temp1;
4101 for (unsigned i
= 0; i
< 2; i
++) {
4102 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
4103 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
4104 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4105 ddxy_out
, ix_ll
, "");
4106 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4107 ddxy_out
, iy_ll
, "");
4108 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4109 interp_param
, ix_ll
, "");
4110 LLVMValueRef temp1
, temp2
;
4112 interp_el
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_el
,
4115 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, src_c0
, "");
4116 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
4118 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, src_c1
, "");
4119 temp2
= LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
4121 ij_out
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
,
4122 temp2
, ctx
->ac
.i32
, "");
4124 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
4128 for (chan
= 0; chan
< 4; chan
++) {
4129 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
4132 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
,
4133 interp_param
, ctx
->ac
.v2f32
, "");
4134 LLVMValueRef i
= LLVMBuildExtractElement(
4135 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_0
, "");
4136 LLVMValueRef j
= LLVMBuildExtractElement(
4137 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_1
, "");
4139 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
4140 llvm_chan
, attr_number
,
4141 ctx
->abi
->prim_mask
, i
, j
);
4143 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
4144 LLVMConstInt(ctx
->ac
.i32
, 2, false),
4145 llvm_chan
, attr_number
,
4146 ctx
->abi
->prim_mask
);
4149 return ac_build_varying_gather_values(&ctx
->ac
, result
, instr
->num_components
,
4150 instr
->variables
[0]->var
->data
.location_frac
);
4154 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
4156 LLVMValueRef gs_next_vertex
;
4157 LLVMValueRef can_emit
;
4159 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4161 assert(stream
== 0);
4163 /* Write vertex attribute values to GSVS ring */
4164 gs_next_vertex
= LLVMBuildLoad(ctx
->builder
,
4165 ctx
->gs_next_vertex
,
4168 /* If this thread has already emitted the declared maximum number of
4169 * vertices, kill it: excessive vertex emissions are not supposed to
4170 * have any effect, and GS threads have no externally observable
4171 * effects other than emitting vertices.
4173 can_emit
= LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, gs_next_vertex
,
4174 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
4175 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4177 /* loop num outputs */
4179 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
4180 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
4185 if (!(ctx
->output_mask
& (1ull << i
)))
4188 if (i
== VARYING_SLOT_CLIP_DIST0
) {
4189 /* pack clip and cull into a single set of slots */
4190 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
4194 for (unsigned j
= 0; j
< length
; j
++) {
4195 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
,
4197 LLVMValueRef voffset
= LLVMConstInt(ctx
->ac
.i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
4198 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, gs_next_vertex
, "");
4199 voffset
= LLVMBuildMul(ctx
->builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
4201 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->ac
.i32
, "");
4203 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
4205 voffset
, ctx
->gs2vs_offset
, 0,
4211 gs_next_vertex
= LLVMBuildAdd(ctx
->builder
, gs_next_vertex
,
4213 LLVMBuildStore(ctx
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
4215 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
4219 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
4221 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4222 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
4226 load_tess_coord(struct ac_shader_abi
*abi
, LLVMTypeRef type
,
4227 unsigned num_components
)
4229 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4231 LLVMValueRef coord
[4] = {
4238 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
4239 coord
[2] = LLVMBuildFSub(ctx
->builder
, ctx
->ac
.f32_1
,
4240 LLVMBuildFAdd(ctx
->builder
, coord
[0], coord
[1], ""), "");
4242 LLVMValueRef result
= ac_build_gather_values(&ctx
->ac
, coord
, num_components
);
4243 return LLVMBuildBitCast(ctx
->builder
, result
, type
, "");
4247 load_patch_vertices_in(struct ac_shader_abi
*abi
)
4249 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4250 return LLVMConstInt(ctx
->ac
.i32
, ctx
->options
->key
.tcs
.input_vertices
, false);
4253 static void visit_intrinsic(struct ac_nir_context
*ctx
,
4254 nir_intrinsic_instr
*instr
)
4256 LLVMValueRef result
= NULL
;
4258 switch (instr
->intrinsic
) {
4259 case nir_intrinsic_ballot
:
4260 result
= ac_build_ballot(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4262 case nir_intrinsic_read_invocation
:
4263 case nir_intrinsic_read_first_invocation
: {
4264 LLVMValueRef args
[2];
4267 args
[0] = get_src(ctx
, instr
->src
[0]);
4270 const char *intr_name
;
4271 if (instr
->intrinsic
== nir_intrinsic_read_invocation
) {
4273 intr_name
= "llvm.amdgcn.readlane";
4276 args
[1] = get_src(ctx
, instr
->src
[1]);
4279 intr_name
= "llvm.amdgcn.readfirstlane";
4282 /* We currently have no other way to prevent LLVM from lifting the icmp
4283 * calls to a dominating basic block.
4285 ac_build_optimization_barrier(&ctx
->ac
, &args
[0]);
4287 result
= ac_build_intrinsic(&ctx
->ac
, intr_name
,
4288 ctx
->ac
.i32
, args
, num_args
,
4289 AC_FUNC_ATTR_READNONE
|
4290 AC_FUNC_ATTR_CONVERGENT
);
4293 case nir_intrinsic_load_subgroup_invocation
:
4294 result
= ac_get_thread_id(&ctx
->ac
);
4296 case nir_intrinsic_load_work_group_id
: {
4297 LLVMValueRef values
[3];
4299 for (int i
= 0; i
< 3; i
++) {
4300 values
[i
] = ctx
->abi
->workgroup_ids
[i
] ?
4301 ctx
->abi
->workgroup_ids
[i
] : ctx
->ac
.i32_0
;
4304 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
4307 case nir_intrinsic_load_base_vertex
: {
4308 result
= ctx
->abi
->base_vertex
;
4311 case nir_intrinsic_load_vertex_id_zero_base
: {
4312 result
= ctx
->abi
->vertex_id
;
4315 case nir_intrinsic_load_local_invocation_id
: {
4316 result
= ctx
->abi
->local_invocation_ids
;
4319 case nir_intrinsic_load_base_instance
:
4320 result
= ctx
->abi
->start_instance
;
4322 case nir_intrinsic_load_draw_id
:
4323 result
= ctx
->abi
->draw_id
;
4325 case nir_intrinsic_load_view_index
:
4326 result
= ctx
->nctx
->view_index
? ctx
->nctx
->view_index
: ctx
->ac
.i32_0
;
4328 case nir_intrinsic_load_invocation_id
:
4329 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
4330 result
= unpack_param(&ctx
->ac
, ctx
->abi
->tcs_rel_ids
, 8, 5);
4332 result
= ctx
->abi
->gs_invocation_id
;
4334 case nir_intrinsic_load_primitive_id
:
4335 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
4336 result
= ctx
->abi
->gs_prim_id
;
4337 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
4338 result
= ctx
->abi
->tcs_patch_id
;
4339 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
4340 result
= ctx
->abi
->tes_patch_id
;
4342 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
4344 case nir_intrinsic_load_sample_id
:
4345 result
= unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
4347 case nir_intrinsic_load_sample_pos
:
4348 result
= load_sample_pos(ctx
);
4350 case nir_intrinsic_load_sample_mask_in
:
4352 result
= load_sample_mask_in(ctx
);
4354 result
= ctx
->abi
->sample_coverage
;
4356 case nir_intrinsic_load_frag_coord
: {
4357 LLVMValueRef values
[4] = {
4358 ctx
->abi
->frag_pos
[0],
4359 ctx
->abi
->frag_pos
[1],
4360 ctx
->abi
->frag_pos
[2],
4361 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, ctx
->abi
->frag_pos
[3])
4363 result
= ac_build_gather_values(&ctx
->ac
, values
, 4);
4366 case nir_intrinsic_load_front_face
:
4367 result
= ctx
->abi
->front_face
;
4369 case nir_intrinsic_load_helper_invocation
:
4370 result
= visit_load_helper_invocation(ctx
);
4372 case nir_intrinsic_load_instance_id
:
4373 result
= ctx
->abi
->instance_id
;
4375 case nir_intrinsic_load_num_work_groups
:
4376 result
= ctx
->nctx
->num_work_groups
;
4378 case nir_intrinsic_load_local_invocation_index
:
4379 result
= visit_load_local_invocation_index(ctx
->nctx
);
4381 case nir_intrinsic_load_push_constant
:
4382 result
= visit_load_push_constant(ctx
->nctx
, instr
);
4384 case nir_intrinsic_vulkan_resource_index
:
4385 result
= visit_vulkan_resource_index(ctx
->nctx
, instr
);
4387 case nir_intrinsic_vulkan_resource_reindex
:
4388 result
= visit_vulkan_resource_reindex(ctx
->nctx
, instr
);
4390 case nir_intrinsic_store_ssbo
:
4391 visit_store_ssbo(ctx
, instr
);
4393 case nir_intrinsic_load_ssbo
:
4394 result
= visit_load_buffer(ctx
, instr
);
4396 case nir_intrinsic_ssbo_atomic_add
:
4397 case nir_intrinsic_ssbo_atomic_imin
:
4398 case nir_intrinsic_ssbo_atomic_umin
:
4399 case nir_intrinsic_ssbo_atomic_imax
:
4400 case nir_intrinsic_ssbo_atomic_umax
:
4401 case nir_intrinsic_ssbo_atomic_and
:
4402 case nir_intrinsic_ssbo_atomic_or
:
4403 case nir_intrinsic_ssbo_atomic_xor
:
4404 case nir_intrinsic_ssbo_atomic_exchange
:
4405 case nir_intrinsic_ssbo_atomic_comp_swap
:
4406 result
= visit_atomic_ssbo(ctx
, instr
);
4408 case nir_intrinsic_load_ubo
:
4409 result
= visit_load_ubo_buffer(ctx
, instr
);
4411 case nir_intrinsic_get_buffer_size
:
4412 result
= visit_get_buffer_size(ctx
, instr
);
4414 case nir_intrinsic_load_var
:
4415 result
= visit_load_var(ctx
, instr
);
4417 case nir_intrinsic_store_var
:
4418 visit_store_var(ctx
, instr
);
4420 case nir_intrinsic_image_load
:
4421 result
= visit_image_load(ctx
, instr
);
4423 case nir_intrinsic_image_store
:
4424 visit_image_store(ctx
, instr
);
4426 case nir_intrinsic_image_atomic_add
:
4427 case nir_intrinsic_image_atomic_min
:
4428 case nir_intrinsic_image_atomic_max
:
4429 case nir_intrinsic_image_atomic_and
:
4430 case nir_intrinsic_image_atomic_or
:
4431 case nir_intrinsic_image_atomic_xor
:
4432 case nir_intrinsic_image_atomic_exchange
:
4433 case nir_intrinsic_image_atomic_comp_swap
:
4434 result
= visit_image_atomic(ctx
, instr
);
4436 case nir_intrinsic_image_size
:
4437 result
= visit_image_size(ctx
, instr
);
4439 case nir_intrinsic_discard
:
4440 case nir_intrinsic_discard_if
:
4441 emit_discard(ctx
, instr
);
4443 case nir_intrinsic_memory_barrier
:
4444 case nir_intrinsic_group_memory_barrier
:
4445 case nir_intrinsic_memory_barrier_atomic_counter
:
4446 case nir_intrinsic_memory_barrier_buffer
:
4447 case nir_intrinsic_memory_barrier_image
:
4448 case nir_intrinsic_memory_barrier_shared
:
4449 emit_membar(ctx
->nctx
, instr
);
4451 case nir_intrinsic_barrier
:
4452 emit_barrier(&ctx
->ac
, ctx
->stage
);
4454 case nir_intrinsic_var_atomic_add
:
4455 case nir_intrinsic_var_atomic_imin
:
4456 case nir_intrinsic_var_atomic_umin
:
4457 case nir_intrinsic_var_atomic_imax
:
4458 case nir_intrinsic_var_atomic_umax
:
4459 case nir_intrinsic_var_atomic_and
:
4460 case nir_intrinsic_var_atomic_or
:
4461 case nir_intrinsic_var_atomic_xor
:
4462 case nir_intrinsic_var_atomic_exchange
:
4463 case nir_intrinsic_var_atomic_comp_swap
:
4464 result
= visit_var_atomic(ctx
->nctx
, instr
);
4466 case nir_intrinsic_interp_var_at_centroid
:
4467 case nir_intrinsic_interp_var_at_sample
:
4468 case nir_intrinsic_interp_var_at_offset
:
4469 result
= visit_interp(ctx
, instr
);
4471 case nir_intrinsic_emit_vertex
:
4472 ctx
->abi
->emit_vertex(ctx
->abi
, nir_intrinsic_stream_id(instr
), ctx
->outputs
);
4474 case nir_intrinsic_end_primitive
:
4475 ctx
->abi
->emit_primitive(ctx
->abi
, nir_intrinsic_stream_id(instr
));
4477 case nir_intrinsic_load_tess_coord
: {
4478 LLVMTypeRef type
= ctx
->nctx
?
4479 get_def_type(ctx
->nctx
->nir
, &instr
->dest
.ssa
) :
4481 result
= ctx
->abi
->load_tess_coord(ctx
->abi
, type
, instr
->num_components
);
4484 case nir_intrinsic_load_tess_level_outer
:
4485 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
);
4487 case nir_intrinsic_load_tess_level_inner
:
4488 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
);
4490 case nir_intrinsic_load_patch_vertices_in
:
4491 result
= ctx
->abi
->load_patch_vertices_in(ctx
->abi
);
4493 case nir_intrinsic_vote_all
: {
4494 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4495 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4498 case nir_intrinsic_vote_any
: {
4499 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4500 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4503 case nir_intrinsic_vote_eq
: {
4504 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4505 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4509 fprintf(stderr
, "Unknown intrinsic: ");
4510 nir_print_instr(&instr
->instr
, stderr
);
4511 fprintf(stderr
, "\n");
4515 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4519 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
4520 LLVMValueRef buffer_ptr
, bool write
)
4522 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4523 LLVMValueRef result
;
4525 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
4527 result
= LLVMBuildLoad(ctx
->builder
, buffer_ptr
, "");
4528 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
4533 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
4535 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4536 LLVMValueRef result
;
4538 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
4540 result
= LLVMBuildLoad(ctx
->builder
, buffer_ptr
, "");
4541 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
4546 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
4547 unsigned descriptor_set
,
4548 unsigned base_index
,
4549 unsigned constant_index
,
4551 enum ac_descriptor_type desc_type
,
4552 bool image
, bool write
)
4554 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4555 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
4556 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
4557 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
4558 unsigned offset
= binding
->offset
;
4559 unsigned stride
= binding
->size
;
4561 LLVMBuilderRef builder
= ctx
->builder
;
4564 assert(base_index
< layout
->binding_count
);
4566 switch (desc_type
) {
4568 type
= ctx
->ac
.v8i32
;
4572 type
= ctx
->ac
.v8i32
;
4576 case AC_DESC_SAMPLER
:
4577 type
= ctx
->ac
.v4i32
;
4578 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
4583 case AC_DESC_BUFFER
:
4584 type
= ctx
->ac
.v4i32
;
4588 unreachable("invalid desc_type\n");
4591 offset
+= constant_index
* stride
;
4593 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
4594 (!index
|| binding
->immutable_samplers_equal
)) {
4595 if (binding
->immutable_samplers_equal
)
4598 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
4600 LLVMValueRef constants
[] = {
4601 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
4602 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
4603 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
4604 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
4606 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
4609 assert(stride
% type_size
== 0);
4612 index
= ctx
->ac
.i32_0
;
4614 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
4616 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
4617 list
= LLVMBuildPointerCast(builder
, list
, ac_array_in_const_addr_space(type
), "");
4619 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
4622 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
4623 const nir_deref_var
*deref
,
4624 enum ac_descriptor_type desc_type
,
4625 const nir_tex_instr
*tex_instr
,
4626 bool image
, bool write
)
4628 LLVMValueRef index
= NULL
;
4629 unsigned constant_index
= 0;
4630 unsigned descriptor_set
;
4631 unsigned base_index
;
4634 assert(tex_instr
&& !image
);
4636 base_index
= tex_instr
->sampler_index
;
4638 const nir_deref
*tail
= &deref
->deref
;
4639 while (tail
->child
) {
4640 const nir_deref_array
*child
= nir_deref_as_array(tail
->child
);
4641 unsigned array_size
= glsl_get_aoa_size(tail
->child
->type
);
4646 assert(child
->deref_array_type
!= nir_deref_array_type_wildcard
);
4648 if (child
->deref_array_type
== nir_deref_array_type_indirect
) {
4649 LLVMValueRef indirect
= get_src(ctx
, child
->indirect
);
4651 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
4652 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
4657 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
4660 constant_index
+= child
->base_offset
* array_size
;
4662 tail
= &child
->deref
;
4664 descriptor_set
= deref
->var
->data
.descriptor_set
;
4665 base_index
= deref
->var
->data
.binding
;
4668 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
4671 constant_index
, index
,
4672 desc_type
, image
, write
);
4675 static void set_tex_fetch_args(struct ac_llvm_context
*ctx
,
4676 struct ac_image_args
*args
,
4677 const nir_tex_instr
*instr
,
4679 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4680 LLVMValueRef
*param
, unsigned count
,
4683 unsigned is_rect
= 0;
4684 bool da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
4686 if (op
== nir_texop_lod
)
4688 /* Pad to power of two vector */
4689 while (count
< util_next_power_of_two(count
))
4690 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4693 args
->addr
= ac_build_gather_values(ctx
, param
, count
);
4695 args
->addr
= param
[0];
4697 args
->resource
= res_ptr
;
4698 args
->sampler
= samp_ptr
;
4700 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
&& op
== nir_texop_txf
) {
4701 args
->addr
= param
[0];
4705 args
->dmask
= dmask
;
4706 args
->unorm
= is_rect
;
4710 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4713 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4714 * filtering manually. The driver sets img7 to a mask clearing
4715 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4716 * s_and_b32 samp0, samp0, img7
4719 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4721 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
4722 LLVMValueRef res
, LLVMValueRef samp
)
4724 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4725 LLVMValueRef img7
, samp0
;
4727 if (ctx
->ac
.chip_class
>= VI
)
4730 img7
= LLVMBuildExtractElement(builder
, res
,
4731 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
4732 samp0
= LLVMBuildExtractElement(builder
, samp
,
4733 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4734 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4735 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4736 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4739 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
4740 nir_tex_instr
*instr
,
4741 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
4742 LLVMValueRef
*fmask_ptr
)
4744 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
4745 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_BUFFER
, instr
, false, false);
4747 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_IMAGE
, instr
, false, false);
4750 *samp_ptr
= get_sampler_desc(ctx
, instr
->sampler
, AC_DESC_SAMPLER
, instr
, false, false);
4752 *samp_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_SAMPLER
, instr
, false, false);
4753 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
4754 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4756 if (fmask_ptr
&& !instr
->sampler
&& (instr
->op
== nir_texop_txf_ms
||
4757 instr
->op
== nir_texop_samples_identical
))
4758 *fmask_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_FMASK
, instr
, false, false);
4761 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
4764 coord
= ac_to_float(ctx
, coord
);
4765 coord
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &coord
, 1, 0);
4766 coord
= ac_to_integer(ctx
, coord
);
4770 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
4772 LLVMValueRef result
= NULL
;
4773 struct ac_image_args args
= { 0 };
4774 unsigned dmask
= 0xf;
4775 LLVMValueRef address
[16];
4776 LLVMValueRef coords
[5];
4777 LLVMValueRef coord
= NULL
, lod
= NULL
, comparator
= NULL
;
4778 LLVMValueRef bias
= NULL
, offsets
= NULL
;
4779 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
, sample_index
= NULL
;
4780 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
4781 LLVMValueRef derivs
[6];
4782 unsigned chan
, count
= 0;
4783 unsigned const_src
= 0, num_deriv_comp
= 0;
4784 bool lod_is_zero
= false;
4786 tex_fetch_ptrs(ctx
, instr
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4788 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4789 switch (instr
->src
[i
].src_type
) {
4790 case nir_tex_src_coord
:
4791 coord
= get_src(ctx
, instr
->src
[i
].src
);
4793 case nir_tex_src_projector
:
4795 case nir_tex_src_comparator
:
4796 comparator
= get_src(ctx
, instr
->src
[i
].src
);
4798 case nir_tex_src_offset
:
4799 offsets
= get_src(ctx
, instr
->src
[i
].src
);
4802 case nir_tex_src_bias
:
4803 bias
= get_src(ctx
, instr
->src
[i
].src
);
4805 case nir_tex_src_lod
: {
4806 nir_const_value
*val
= nir_src_as_const_value(instr
->src
[i
].src
);
4808 if (val
&& val
->i32
[0] == 0)
4810 lod
= get_src(ctx
, instr
->src
[i
].src
);
4813 case nir_tex_src_ms_index
:
4814 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
4816 case nir_tex_src_ms_mcs
:
4818 case nir_tex_src_ddx
:
4819 ddx
= get_src(ctx
, instr
->src
[i
].src
);
4820 num_deriv_comp
= instr
->src
[i
].src
.ssa
->num_components
;
4822 case nir_tex_src_ddy
:
4823 ddy
= get_src(ctx
, instr
->src
[i
].src
);
4825 case nir_tex_src_texture_offset
:
4826 case nir_tex_src_sampler_offset
:
4827 case nir_tex_src_plane
:
4833 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
4834 result
= get_buffer_size(ctx
, res_ptr
, true);
4838 if (instr
->op
== nir_texop_texture_samples
) {
4839 LLVMValueRef res
, samples
, is_msaa
;
4840 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res_ptr
, ctx
->ac
.v8i32
, "");
4841 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4842 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4843 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4844 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
4845 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
4846 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4847 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
4848 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4850 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4851 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
4852 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
4853 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
4854 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
4856 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
4863 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4864 coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
4866 if (offsets
&& instr
->op
!= nir_texop_txf
) {
4867 LLVMValueRef offset
[3], pack
;
4868 for (chan
= 0; chan
< 3; ++chan
)
4869 offset
[chan
] = ctx
->ac
.i32_0
;
4872 for (chan
= 0; chan
< ac_get_llvm_num_components(offsets
); chan
++) {
4873 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, offsets
, chan
);
4874 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
4875 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
4877 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
4878 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
4880 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
4881 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
4882 address
[count
++] = pack
;
4885 /* pack LOD bias value */
4886 if (instr
->op
== nir_texop_txb
&& bias
) {
4887 address
[count
++] = bias
;
4890 /* Pack depth comparison value */
4891 if (instr
->is_shadow
&& comparator
) {
4892 LLVMValueRef z
= ac_to_float(&ctx
->ac
,
4893 ac_llvm_extract_elem(&ctx
->ac
, comparator
, 0));
4895 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
4896 * so the depth comparison value isn't clamped for Z16 and
4897 * Z24 anymore. Do it manually here.
4899 * It's unnecessary if the original texture format was
4900 * Z32_FLOAT, but we don't know that here.
4902 if (ctx
->ac
.chip_class
== VI
&& ctx
->abi
->clamp_shadow_reference
)
4903 z
= ac_build_clamp(&ctx
->ac
, z
);
4905 address
[count
++] = z
;
4908 /* pack derivatives */
4910 int num_src_deriv_channels
, num_dest_deriv_channels
;
4911 switch (instr
->sampler_dim
) {
4912 case GLSL_SAMPLER_DIM_3D
:
4913 case GLSL_SAMPLER_DIM_CUBE
:
4915 num_src_deriv_channels
= 3;
4916 num_dest_deriv_channels
= 3;
4918 case GLSL_SAMPLER_DIM_2D
:
4920 num_src_deriv_channels
= 2;
4921 num_dest_deriv_channels
= 2;
4924 case GLSL_SAMPLER_DIM_1D
:
4925 num_src_deriv_channels
= 1;
4926 if (ctx
->ac
.chip_class
>= GFX9
) {
4927 num_dest_deriv_channels
= 2;
4930 num_dest_deriv_channels
= 1;
4936 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
4937 derivs
[i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
4938 derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
4940 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
4941 derivs
[i
] = ctx
->ac
.f32_0
;
4942 derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
4946 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& coord
) {
4947 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4948 coords
[chan
] = ac_to_float(&ctx
->ac
, coords
[chan
]);
4949 if (instr
->coord_components
== 3)
4950 coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
4951 ac_prepare_cube_coords(&ctx
->ac
,
4952 instr
->op
== nir_texop_txd
, instr
->is_array
,
4953 instr
->op
== nir_texop_lod
, coords
, derivs
);
4959 for (unsigned i
= 0; i
< num_deriv_comp
* 2; i
++)
4960 address
[count
++] = derivs
[i
];
4963 /* Pack texture coordinates */
4965 address
[count
++] = coords
[0];
4966 if (instr
->coord_components
> 1) {
4967 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&& instr
->is_array
&& instr
->op
!= nir_texop_txf
) {
4968 coords
[1] = apply_round_slice(&ctx
->ac
, coords
[1]);
4970 address
[count
++] = coords
[1];
4972 if (instr
->coord_components
> 2) {
4973 /* This seems like a bit of a hack - but it passes Vulkan CTS with it */
4974 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_3D
&&
4975 instr
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
&&
4976 instr
->op
!= nir_texop_txf
) {
4977 coords
[2] = apply_round_slice(&ctx
->ac
, coords
[2]);
4979 address
[count
++] = coords
[2];
4982 if (ctx
->ac
.chip_class
>= GFX9
) {
4983 LLVMValueRef filler
;
4984 if (instr
->op
== nir_texop_txf
)
4985 filler
= ctx
->ac
.i32_0
;
4987 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
4989 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
) {
4990 /* No nir_texop_lod, because it does not take a slice
4991 * even with array textures. */
4992 if (instr
->is_array
&& instr
->op
!= nir_texop_lod
) {
4993 address
[count
] = address
[count
- 1];
4994 address
[count
- 1] = filler
;
4997 address
[count
++] = filler
;
5003 if (lod
&& ((instr
->op
== nir_texop_txl
&& !lod_is_zero
) ||
5004 instr
->op
== nir_texop_txf
)) {
5005 address
[count
++] = lod
;
5006 } else if (instr
->op
== nir_texop_txf_ms
&& sample_index
) {
5007 address
[count
++] = sample_index
;
5008 } else if(instr
->op
== nir_texop_txs
) {
5011 address
[count
++] = lod
;
5013 address
[count
++] = ctx
->ac
.i32_0
;
5016 for (chan
= 0; chan
< count
; chan
++) {
5017 address
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
,
5018 address
[chan
], ctx
->ac
.i32
, "");
5021 if (instr
->op
== nir_texop_samples_identical
) {
5022 LLVMValueRef txf_address
[4];
5023 struct ac_image_args txf_args
= { 0 };
5024 unsigned txf_count
= count
;
5025 memcpy(txf_address
, address
, sizeof(txf_address
));
5027 if (!instr
->is_array
)
5028 txf_address
[2] = ctx
->ac
.i32_0
;
5029 txf_address
[3] = ctx
->ac
.i32_0
;
5031 set_tex_fetch_args(&ctx
->ac
, &txf_args
, instr
, nir_texop_txf
,
5033 txf_address
, txf_count
, 0xf);
5035 result
= build_tex_intrinsic(ctx
, instr
, false, &txf_args
);
5037 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
5038 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
5042 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
5043 instr
->op
!= nir_texop_txs
) {
5044 unsigned sample_chan
= instr
->is_array
? 3 : 2;
5045 address
[sample_chan
] = adjust_sample_index_using_fmask(&ctx
->ac
,
5048 instr
->is_array
? address
[2] : NULL
,
5049 address
[sample_chan
],
5053 if (offsets
&& instr
->op
== nir_texop_txf
) {
5054 nir_const_value
*const_offset
=
5055 nir_src_as_const_value(instr
->src
[const_src
].src
);
5056 int num_offsets
= instr
->src
[const_src
].src
.ssa
->num_components
;
5057 assert(const_offset
);
5058 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
5059 if (num_offsets
> 2)
5060 address
[2] = LLVMBuildAdd(ctx
->ac
.builder
,
5061 address
[2], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[2], false), "");
5062 if (num_offsets
> 1)
5063 address
[1] = LLVMBuildAdd(ctx
->ac
.builder
,
5064 address
[1], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[1], false), "");
5065 address
[0] = LLVMBuildAdd(ctx
->ac
.builder
,
5066 address
[0], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[0], false), "");
5070 /* TODO TG4 support */
5071 if (instr
->op
== nir_texop_tg4
) {
5072 if (instr
->is_shadow
)
5075 dmask
= 1 << instr
->component
;
5077 set_tex_fetch_args(&ctx
->ac
, &args
, instr
, instr
->op
,
5078 res_ptr
, samp_ptr
, address
, count
, dmask
);
5080 result
= build_tex_intrinsic(ctx
, instr
, lod_is_zero
, &args
);
5082 if (instr
->op
== nir_texop_query_levels
)
5083 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
5084 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
5085 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
5086 instr
->op
!= nir_texop_tg4
)
5087 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
5088 else if (instr
->op
== nir_texop_txs
&&
5089 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
5091 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
5092 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
5093 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
5094 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
5095 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
5096 } else if (ctx
->ac
.chip_class
>= GFX9
&&
5097 instr
->op
== nir_texop_txs
&&
5098 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
5100 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
5101 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
5102 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
5104 } else if (instr
->dest
.ssa
.num_components
!= 4)
5105 result
= trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
5109 assert(instr
->dest
.is_ssa
);
5110 result
= ac_to_integer(&ctx
->ac
, result
);
5111 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
5116 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
5118 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
5119 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
5121 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
5122 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
5125 static void visit_post_phi(struct ac_nir_context
*ctx
,
5126 nir_phi_instr
*instr
,
5127 LLVMValueRef llvm_phi
)
5129 nir_foreach_phi_src(src
, instr
) {
5130 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
5131 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
5133 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
5137 static void phi_post_pass(struct ac_nir_context
*ctx
)
5139 struct hash_entry
*entry
;
5140 hash_table_foreach(ctx
->phis
, entry
) {
5141 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
5142 (LLVMValueRef
)entry
->data
);
5147 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
5148 const nir_ssa_undef_instr
*instr
)
5150 unsigned num_components
= instr
->def
.num_components
;
5151 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
5154 if (num_components
== 1)
5155 undef
= LLVMGetUndef(type
);
5157 undef
= LLVMGetUndef(LLVMVectorType(type
, num_components
));
5159 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, undef
);
5162 static void visit_jump(struct ac_nir_context
*ctx
,
5163 const nir_jump_instr
*instr
)
5165 switch (instr
->type
) {
5166 case nir_jump_break
:
5167 LLVMBuildBr(ctx
->ac
.builder
, ctx
->break_block
);
5168 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5170 case nir_jump_continue
:
5171 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5172 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5175 fprintf(stderr
, "Unknown NIR jump instr: ");
5176 nir_print_instr(&instr
->instr
, stderr
);
5177 fprintf(stderr
, "\n");
5182 static void visit_cf_list(struct ac_nir_context
*ctx
,
5183 struct exec_list
*list
);
5185 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
5187 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
5188 nir_foreach_instr(instr
, block
)
5190 switch (instr
->type
) {
5191 case nir_instr_type_alu
:
5192 visit_alu(ctx
, nir_instr_as_alu(instr
));
5194 case nir_instr_type_load_const
:
5195 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
5197 case nir_instr_type_intrinsic
:
5198 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
5200 case nir_instr_type_tex
:
5201 visit_tex(ctx
, nir_instr_as_tex(instr
));
5203 case nir_instr_type_phi
:
5204 visit_phi(ctx
, nir_instr_as_phi(instr
));
5206 case nir_instr_type_ssa_undef
:
5207 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
5209 case nir_instr_type_jump
:
5210 visit_jump(ctx
, nir_instr_as_jump(instr
));
5213 fprintf(stderr
, "Unknown NIR instr type: ");
5214 nir_print_instr(instr
, stderr
);
5215 fprintf(stderr
, "\n");
5220 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
5223 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
5225 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
5227 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5228 LLVMBasicBlockRef merge_block
=
5229 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5230 LLVMBasicBlockRef if_block
=
5231 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5232 LLVMBasicBlockRef else_block
= merge_block
;
5233 if (!exec_list_is_empty(&if_stmt
->else_list
))
5234 else_block
= LLVMAppendBasicBlockInContext(
5235 ctx
->ac
.context
, fn
, "");
5237 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, value
,
5239 LLVMBuildCondBr(ctx
->ac
.builder
, cond
, if_block
, else_block
);
5241 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, if_block
);
5242 visit_cf_list(ctx
, &if_stmt
->then_list
);
5243 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5244 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5246 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
5247 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, else_block
);
5248 visit_cf_list(ctx
, &if_stmt
->else_list
);
5249 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5250 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5253 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, merge_block
);
5256 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
5258 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5259 LLVMBasicBlockRef continue_parent
= ctx
->continue_block
;
5260 LLVMBasicBlockRef break_parent
= ctx
->break_block
;
5262 ctx
->continue_block
=
5263 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5265 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5267 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5268 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->continue_block
);
5269 visit_cf_list(ctx
, &loop
->body
);
5271 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5272 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5273 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->break_block
);
5275 ctx
->continue_block
= continue_parent
;
5276 ctx
->break_block
= break_parent
;
5279 static void visit_cf_list(struct ac_nir_context
*ctx
,
5280 struct exec_list
*list
)
5282 foreach_list_typed(nir_cf_node
, node
, node
, list
)
5284 switch (node
->type
) {
5285 case nir_cf_node_block
:
5286 visit_block(ctx
, nir_cf_node_as_block(node
));
5289 case nir_cf_node_if
:
5290 visit_if(ctx
, nir_cf_node_as_if(node
));
5293 case nir_cf_node_loop
:
5294 visit_loop(ctx
, nir_cf_node_as_loop(node
));
5304 handle_vs_input_decl(struct nir_to_llvm_context
*ctx
,
5305 struct nir_variable
*variable
)
5307 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
5308 LLVMValueRef t_offset
;
5309 LLVMValueRef t_list
;
5311 LLVMValueRef buffer_index
;
5312 int index
= variable
->data
.location
- VERT_ATTRIB_GENERIC0
;
5313 int idx
= variable
->data
.location
;
5314 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
5316 variable
->data
.driver_location
= idx
* 4;
5318 for (unsigned i
= 0; i
< attrib_count
; ++i
, ++idx
) {
5319 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << (index
+ i
))) {
5320 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.instance_id
,
5321 ctx
->abi
.start_instance
, "");
5322 if (ctx
->options
->key
.vs
.as_ls
) {
5323 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
5324 MAX2(2, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5326 ctx
->shader_info
->vs
.vgpr_comp_cnt
=
5327 MAX2(1, ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5330 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.vertex_id
,
5331 ctx
->abi
.base_vertex
, "");
5332 t_offset
= LLVMConstInt(ctx
->ac
.i32
, index
+ i
, false);
5334 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
5336 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
5341 for (unsigned chan
= 0; chan
< 4; chan
++) {
5342 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5343 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
, chan
)] =
5344 ac_to_integer(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
,
5345 input
, llvm_chan
, ""));
5350 static void interp_fs_input(struct nir_to_llvm_context
*ctx
,
5352 LLVMValueRef interp_param
,
5353 LLVMValueRef prim_mask
,
5354 LLVMValueRef result
[4])
5356 LLVMValueRef attr_number
;
5359 bool interp
= interp_param
!= NULL
;
5361 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
5363 /* fs.constant returns the param from the middle vertex, so it's not
5364 * really useful for flat shading. It's meant to be used for custom
5365 * interpolation (but the intrinsic can't fetch from the other two
5368 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
5369 * to do the right thing. The only reason we use fs.constant is that
5370 * fs.interp cannot be used on integers, because they can be equal
5374 interp_param
= LLVMBuildBitCast(ctx
->builder
, interp_param
,
5377 i
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
5379 j
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
5383 for (chan
= 0; chan
< 4; chan
++) {
5384 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5387 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
5392 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
5393 LLVMConstInt(ctx
->ac
.i32
, 2, false),
5402 handle_fs_input_decl(struct nir_to_llvm_context
*ctx
,
5403 struct nir_variable
*variable
)
5405 int idx
= variable
->data
.location
;
5406 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5407 LLVMValueRef interp
;
5409 variable
->data
.driver_location
= idx
* 4;
5410 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
5412 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
5413 unsigned interp_type
;
5414 if (variable
->data
.sample
) {
5415 interp_type
= INTERP_SAMPLE
;
5416 ctx
->shader_info
->info
.ps
.force_persample
= true;
5417 } else if (variable
->data
.centroid
)
5418 interp_type
= INTERP_CENTROID
;
5420 interp_type
= INTERP_CENTER
;
5422 interp
= lookup_interp_param(&ctx
->abi
, variable
->data
.interpolation
, interp_type
);
5426 for (unsigned i
= 0; i
< attrib_count
; ++i
)
5427 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
5432 handle_vs_inputs(struct nir_to_llvm_context
*ctx
,
5433 struct nir_shader
*nir
) {
5434 nir_foreach_variable(variable
, &nir
->inputs
)
5435 handle_vs_input_decl(ctx
, variable
);
5439 prepare_interp_optimize(struct nir_to_llvm_context
*ctx
,
5440 struct nir_shader
*nir
)
5442 if (!ctx
->options
->key
.fs
.multisample
)
5445 bool uses_center
= false;
5446 bool uses_centroid
= false;
5447 nir_foreach_variable(variable
, &nir
->inputs
) {
5448 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
5449 variable
->data
.sample
)
5452 if (variable
->data
.centroid
)
5453 uses_centroid
= true;
5458 if (uses_center
&& uses_centroid
) {
5459 LLVMValueRef sel
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, ctx
->abi
.prim_mask
, ctx
->ac
.i32_0
, "");
5460 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
5461 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
5466 handle_fs_inputs(struct nir_to_llvm_context
*ctx
,
5467 struct nir_shader
*nir
)
5469 prepare_interp_optimize(ctx
, nir
);
5471 nir_foreach_variable(variable
, &nir
->inputs
)
5472 handle_fs_input_decl(ctx
, variable
);
5476 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
5477 ctx
->shader_info
->info
.needs_multiview_view_index
)
5478 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
5480 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
5481 LLVMValueRef interp_param
;
5482 LLVMValueRef
*inputs
= ctx
->inputs
+radeon_llvm_reg_index_soa(i
, 0);
5484 if (!(ctx
->input_mask
& (1ull << i
)))
5487 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
5488 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
5489 interp_param
= *inputs
;
5490 interp_fs_input(ctx
, index
, interp_param
, ctx
->abi
.prim_mask
,
5494 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
5496 } else if (i
== VARYING_SLOT_POS
) {
5497 for(int i
= 0; i
< 3; ++i
)
5498 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
5500 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
5501 ctx
->abi
.frag_pos
[3]);
5504 ctx
->shader_info
->fs
.num_interp
= index
;
5505 if (ctx
->input_mask
& (1 << VARYING_SLOT_PNTC
))
5506 ctx
->shader_info
->fs
.has_pcoord
= true;
5507 if (ctx
->input_mask
& (1 << VARYING_SLOT_PRIMITIVE_ID
))
5508 ctx
->shader_info
->fs
.prim_id_input
= true;
5509 if (ctx
->input_mask
& (1 << VARYING_SLOT_LAYER
))
5510 ctx
->shader_info
->fs
.layer_input
= true;
5511 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
5513 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
5514 ctx
->view_index
= ctx
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5518 ac_build_alloca(struct ac_llvm_context
*ac
,
5522 LLVMBuilderRef builder
= ac
->builder
;
5523 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
5524 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
5525 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
5526 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
5527 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
5531 LLVMPositionBuilderBefore(first_builder
, first_instr
);
5533 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
5536 res
= LLVMBuildAlloca(first_builder
, type
, name
);
5537 LLVMBuildStore(builder
, LLVMConstNull(type
), res
);
5539 LLVMDisposeBuilder(first_builder
);
5544 static LLVMValueRef
si_build_alloca_undef(struct ac_llvm_context
*ac
,
5548 LLVMValueRef ptr
= ac_build_alloca(ac
, type
, name
);
5549 LLVMBuildStore(ac
->builder
, LLVMGetUndef(type
), ptr
);
5554 scan_shader_output_decl(struct nir_to_llvm_context
*ctx
,
5555 struct nir_variable
*variable
,
5556 struct nir_shader
*shader
,
5557 gl_shader_stage stage
)
5559 int idx
= variable
->data
.location
+ variable
->data
.index
;
5560 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5561 uint64_t mask_attribs
;
5563 variable
->data
.driver_location
= idx
* 4;
5565 /* tess ctrl has it's own load/store paths for outputs */
5566 if (stage
== MESA_SHADER_TESS_CTRL
)
5569 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
5570 if (stage
== MESA_SHADER_VERTEX
||
5571 stage
== MESA_SHADER_TESS_EVAL
||
5572 stage
== MESA_SHADER_GEOMETRY
) {
5573 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5574 int length
= shader
->info
.clip_distance_array_size
+
5575 shader
->info
.cull_distance_array_size
;
5576 if (stage
== MESA_SHADER_VERTEX
) {
5577 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5578 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5580 if (stage
== MESA_SHADER_TESS_EVAL
) {
5581 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5582 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5589 mask_attribs
= 1ull << idx
;
5593 ctx
->output_mask
|= mask_attribs
;
5597 handle_shader_output_decl(struct ac_nir_context
*ctx
,
5598 struct nir_shader
*nir
,
5599 struct nir_variable
*variable
)
5601 unsigned output_loc
= variable
->data
.driver_location
/ 4;
5602 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5604 /* tess ctrl has it's own load/store paths for outputs */
5605 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
5608 if (ctx
->stage
== MESA_SHADER_VERTEX
||
5609 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
5610 ctx
->stage
== MESA_SHADER_GEOMETRY
) {
5611 int idx
= variable
->data
.location
+ variable
->data
.index
;
5612 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5613 int length
= nir
->info
.clip_distance_array_size
+
5614 nir
->info
.cull_distance_array_size
;
5623 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
5624 for (unsigned chan
= 0; chan
< 4; chan
++) {
5625 ctx
->outputs
[radeon_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
5626 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5632 glsl_base_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5633 enum glsl_base_type type
)
5637 case GLSL_TYPE_UINT
:
5638 case GLSL_TYPE_BOOL
:
5639 case GLSL_TYPE_SUBROUTINE
:
5641 case GLSL_TYPE_FLOAT
: /* TODO handle mediump */
5643 case GLSL_TYPE_INT64
:
5644 case GLSL_TYPE_UINT64
:
5646 case GLSL_TYPE_DOUBLE
:
5649 unreachable("unknown GLSL type");
5654 glsl_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5655 const struct glsl_type
*type
)
5657 if (glsl_type_is_scalar(type
)) {
5658 return glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
));
5661 if (glsl_type_is_vector(type
)) {
5662 return LLVMVectorType(
5663 glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
)),
5664 glsl_get_vector_elements(type
));
5667 if (glsl_type_is_matrix(type
)) {
5668 return LLVMArrayType(
5669 glsl_to_llvm_type(ctx
, glsl_get_column_type(type
)),
5670 glsl_get_matrix_columns(type
));
5673 if (glsl_type_is_array(type
)) {
5674 return LLVMArrayType(
5675 glsl_to_llvm_type(ctx
, glsl_get_array_element(type
)),
5676 glsl_get_length(type
));
5679 assert(glsl_type_is_struct(type
));
5681 LLVMTypeRef member_types
[glsl_get_length(type
)];
5683 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
5685 glsl_to_llvm_type(ctx
,
5686 glsl_get_struct_field(type
, i
));
5689 return LLVMStructTypeInContext(ctx
->context
, member_types
,
5690 glsl_get_length(type
), false);
5694 setup_locals(struct ac_nir_context
*ctx
,
5695 struct nir_function
*func
)
5698 ctx
->num_locals
= 0;
5699 nir_foreach_variable(variable
, &func
->impl
->locals
) {
5700 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5701 variable
->data
.driver_location
= ctx
->num_locals
* 4;
5702 variable
->data
.location_frac
= 0;
5703 ctx
->num_locals
+= attrib_count
;
5705 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
5709 for (i
= 0; i
< ctx
->num_locals
; i
++) {
5710 for (j
= 0; j
< 4; j
++) {
5711 ctx
->locals
[i
* 4 + j
] =
5712 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
5718 setup_shared(struct ac_nir_context
*ctx
,
5719 struct nir_shader
*nir
)
5721 nir_foreach_variable(variable
, &nir
->shared
) {
5722 LLVMValueRef shared
=
5723 LLVMAddGlobalInAddressSpace(
5724 ctx
->ac
.module
, glsl_to_llvm_type(ctx
->nctx
, variable
->type
),
5725 variable
->name
? variable
->name
: "",
5726 AC_LOCAL_ADDR_SPACE
);
5727 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
5732 emit_float_saturate(struct ac_llvm_context
*ctx
, LLVMValueRef v
, float lo
, float hi
)
5734 v
= ac_to_float(ctx
, v
);
5735 v
= emit_intrin_2f_param(ctx
, "llvm.maxnum", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, lo
));
5736 return emit_intrin_2f_param(ctx
, "llvm.minnum", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, hi
));
5740 static LLVMValueRef
emit_pack_int16(struct nir_to_llvm_context
*ctx
,
5741 LLVMValueRef src0
, LLVMValueRef src1
)
5743 LLVMValueRef const16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
5744 LLVMValueRef comp
[2];
5746 comp
[0] = LLVMBuildAnd(ctx
->builder
, src0
, LLVMConstInt(ctx
->ac
.i32
, 65535, 0), "");
5747 comp
[1] = LLVMBuildAnd(ctx
->builder
, src1
, LLVMConstInt(ctx
->ac
.i32
, 65535, 0), "");
5748 comp
[1] = LLVMBuildShl(ctx
->builder
, comp
[1], const16
, "");
5749 return LLVMBuildOr(ctx
->builder
, comp
[0], comp
[1], "");
5752 /* Initialize arguments for the shader export intrinsic */
5754 si_llvm_init_export_args(struct nir_to_llvm_context
*ctx
,
5755 LLVMValueRef
*values
,
5757 struct ac_export_args
*args
)
5759 /* Default is 0xf. Adjusted below depending on the format. */
5760 args
->enabled_channels
= 0xf;
5762 /* Specify whether the EXEC mask represents the valid mask */
5763 args
->valid_mask
= 0;
5765 /* Specify whether this is the last export */
5768 /* Specify the target we are exporting */
5769 args
->target
= target
;
5771 args
->compr
= false;
5772 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
5773 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
5774 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
5775 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
5780 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
5781 LLVMValueRef val
[4];
5782 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
5783 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
5784 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
5785 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
5787 switch(col_format
) {
5788 case V_028714_SPI_SHADER_ZERO
:
5789 args
->enabled_channels
= 0; /* writemask */
5790 args
->target
= V_008DFC_SQ_EXP_NULL
;
5793 case V_028714_SPI_SHADER_32_R
:
5794 args
->enabled_channels
= 1;
5795 args
->out
[0] = values
[0];
5798 case V_028714_SPI_SHADER_32_GR
:
5799 args
->enabled_channels
= 0x3;
5800 args
->out
[0] = values
[0];
5801 args
->out
[1] = values
[1];
5804 case V_028714_SPI_SHADER_32_AR
:
5805 args
->enabled_channels
= 0x9;
5806 args
->out
[0] = values
[0];
5807 args
->out
[3] = values
[3];
5810 case V_028714_SPI_SHADER_FP16_ABGR
:
5813 for (unsigned chan
= 0; chan
< 2; chan
++) {
5814 LLVMValueRef pack_args
[2] = {
5816 values
[2 * chan
+ 1]
5818 LLVMValueRef packed
;
5820 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
5821 args
->out
[chan
] = packed
;
5825 case V_028714_SPI_SHADER_UNORM16_ABGR
:
5826 for (unsigned chan
= 0; chan
< 4; chan
++) {
5827 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
5828 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5829 LLVMConstReal(ctx
->ac
.f32
, 65535), "");
5830 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5831 LLVMConstReal(ctx
->ac
.f32
, 0.5), "");
5832 val
[chan
] = LLVMBuildFPToUI(ctx
->builder
, val
[chan
],
5837 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5838 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5841 case V_028714_SPI_SHADER_SNORM16_ABGR
:
5842 for (unsigned chan
= 0; chan
< 4; chan
++) {
5843 val
[chan
] = emit_float_saturate(&ctx
->ac
, values
[chan
], -1, 1);
5844 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5845 LLVMConstReal(ctx
->ac
.f32
, 32767), "");
5847 /* If positive, add 0.5, else add -0.5. */
5848 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5849 LLVMBuildSelect(ctx
->builder
,
5850 LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
,
5851 val
[chan
], ctx
->ac
.f32_0
, ""),
5852 LLVMConstReal(ctx
->ac
.f32
, 0.5),
5853 LLVMConstReal(ctx
->ac
.f32
, -0.5), ""), "");
5854 val
[chan
] = LLVMBuildFPToSI(ctx
->builder
, val
[chan
], ctx
->ac
.i32
, "");
5858 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5859 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5862 case V_028714_SPI_SHADER_UINT16_ABGR
: {
5863 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5864 is_int8
? 255 : is_int10
? 1023 : 65535, 0);
5865 LLVMValueRef max_alpha
= !is_int10
? max_rgb
: LLVMConstInt(ctx
->ac
.i32
, 3, 0);
5867 for (unsigned chan
= 0; chan
< 4; chan
++) {
5868 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
5869 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntULT
, val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
5873 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5874 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5878 case V_028714_SPI_SHADER_SINT16_ABGR
: {
5879 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5880 is_int8
? 127 : is_int10
? 511 : 32767, 0);
5881 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5882 is_int8
? -128 : is_int10
? -512 : -32768, 0);
5883 LLVMValueRef max_alpha
= !is_int10
? max_rgb
: ctx
->ac
.i32_1
;
5884 LLVMValueRef min_alpha
= !is_int10
? min_rgb
: LLVMConstInt(ctx
->ac
.i32
, -2, 0);
5887 for (unsigned chan
= 0; chan
< 4; chan
++) {
5888 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
5889 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
5890 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, val
[chan
], chan
== 3 ? min_alpha
: min_rgb
);
5894 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5895 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5900 case V_028714_SPI_SHADER_32_ABGR
:
5901 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5905 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5907 for (unsigned i
= 0; i
< 4; ++i
)
5908 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
5912 handle_vs_outputs_post(struct nir_to_llvm_context
*ctx
,
5913 bool export_prim_id
,
5914 struct ac_vs_output_info
*outinfo
)
5916 uint32_t param_count
= 0;
5918 unsigned pos_idx
, num_pos_exports
= 0;
5919 struct ac_export_args args
, pos_args
[4] = {};
5920 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
5923 if (ctx
->options
->key
.has_multiview_view_index
) {
5924 LLVMValueRef
* tmp_out
= &ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5926 for(unsigned i
= 0; i
< 4; ++i
)
5927 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
5928 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5931 LLVMBuildStore(ctx
->builder
, ac_to_float(&ctx
->ac
, ctx
->view_index
), *tmp_out
);
5932 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
5935 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
5936 sizeof(outinfo
->vs_output_param_offset
));
5938 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
5939 LLVMValueRef slots
[8];
5942 if (outinfo
->cull_dist_mask
)
5943 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
5945 i
= VARYING_SLOT_CLIP_DIST0
;
5946 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
5947 slots
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5948 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5950 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
5951 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
5953 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
5954 target
= V_008DFC_SQ_EXP_POS
+ 3;
5955 si_llvm_init_export_args(ctx
, &slots
[4], target
, &args
);
5956 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5957 &args
, sizeof(args
));
5960 target
= V_008DFC_SQ_EXP_POS
+ 2;
5961 si_llvm_init_export_args(ctx
, &slots
[0], target
, &args
);
5962 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5963 &args
, sizeof(args
));
5967 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
5968 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
5969 for (unsigned j
= 0; j
< 4; j
++)
5970 pos_values
[j
] = LLVMBuildLoad(ctx
->builder
,
5971 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_POS
, j
)], "");
5973 si_llvm_init_export_args(ctx
, pos_values
, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
5975 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
5976 outinfo
->writes_pointsize
= true;
5977 psize_value
= LLVMBuildLoad(ctx
->builder
,
5978 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_PSIZ
, 0)], "");
5981 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
5982 outinfo
->writes_layer
= true;
5983 layer_value
= LLVMBuildLoad(ctx
->builder
,
5984 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)], "");
5987 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
5988 outinfo
->writes_viewport_index
= true;
5989 viewport_index_value
= LLVMBuildLoad(ctx
->builder
,
5990 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_VIEWPORT
, 0)], "");
5993 if (outinfo
->writes_pointsize
||
5994 outinfo
->writes_layer
||
5995 outinfo
->writes_viewport_index
) {
5996 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
5997 (outinfo
->writes_layer
== true ? 4 : 0));
5998 pos_args
[1].valid_mask
= 0;
5999 pos_args
[1].done
= 0;
6000 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
6001 pos_args
[1].compr
= 0;
6002 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
6003 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
6004 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
6005 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
6007 if (outinfo
->writes_pointsize
== true)
6008 pos_args
[1].out
[0] = psize_value
;
6009 if (outinfo
->writes_layer
== true)
6010 pos_args
[1].out
[2] = layer_value
;
6011 if (outinfo
->writes_viewport_index
== true) {
6012 if (ctx
->options
->chip_class
>= GFX9
) {
6013 /* GFX9 has the layer in out.z[10:0] and the viewport
6014 * index in out.z[19:16].
6016 LLVMValueRef v
= viewport_index_value
;
6017 v
= ac_to_integer(&ctx
->ac
, v
);
6018 v
= LLVMBuildShl(ctx
->builder
, v
,
6019 LLVMConstInt(ctx
->ac
.i32
, 16, false),
6021 v
= LLVMBuildOr(ctx
->builder
, v
,
6022 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
6024 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
6025 pos_args
[1].enabled_channels
|= 1 << 2;
6027 pos_args
[1].out
[3] = viewport_index_value
;
6028 pos_args
[1].enabled_channels
|= 1 << 3;
6032 for (i
= 0; i
< 4; i
++) {
6033 if (pos_args
[i
].out
[0])
6038 for (i
= 0; i
< 4; i
++) {
6039 if (!pos_args
[i
].out
[0])
6042 /* Specify the target we are exporting */
6043 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
6044 if (pos_idx
== num_pos_exports
)
6045 pos_args
[i
].done
= 1;
6046 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
6049 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6050 LLVMValueRef values
[4];
6051 if (!(ctx
->output_mask
& (1ull << i
)))
6054 for (unsigned j
= 0; j
< 4; j
++)
6055 values
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6056 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
6058 if (i
== VARYING_SLOT_LAYER
) {
6059 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
6060 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
;
6062 } else if (i
== VARYING_SLOT_PRIMITIVE_ID
) {
6063 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
6064 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
6066 } else if (i
>= VARYING_SLOT_VAR0
) {
6067 outinfo
->export_mask
|= 1u << (i
- VARYING_SLOT_VAR0
);
6068 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
6069 outinfo
->vs_output_param_offset
[i
] = param_count
;
6074 si_llvm_init_export_args(ctx
, values
, target
, &args
);
6076 if (target
>= V_008DFC_SQ_EXP_POS
&&
6077 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
6078 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
6079 &args
, sizeof(args
));
6081 ac_build_export(&ctx
->ac
, &args
);
6085 if (export_prim_id
) {
6086 LLVMValueRef values
[4];
6087 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
6088 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
6091 values
[0] = ctx
->vs_prim_id
;
6092 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
6093 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
6094 for (unsigned j
= 1; j
< 4; j
++)
6095 values
[j
] = ctx
->ac
.f32_0
;
6096 si_llvm_init_export_args(ctx
, values
, target
, &args
);
6097 ac_build_export(&ctx
->ac
, &args
);
6098 outinfo
->export_prim_id
= true;
6101 outinfo
->pos_exports
= num_pos_exports
;
6102 outinfo
->param_exports
= param_count
;
6106 handle_es_outputs_post(struct nir_to_llvm_context
*ctx
,
6107 struct ac_es_output_info
*outinfo
)
6110 uint64_t max_output_written
= 0;
6111 LLVMValueRef lds_base
= NULL
;
6113 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6117 if (!(ctx
->output_mask
& (1ull << i
)))
6120 if (i
== VARYING_SLOT_CLIP_DIST0
)
6121 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6123 param_index
= shader_io_get_unique_index(i
);
6125 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
6128 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
6130 if (ctx
->ac
.chip_class
>= GFX9
) {
6131 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
6132 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
6133 LLVMValueRef wave_idx
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6134 LLVMConstInt(ctx
->ac
.i32
, 24, false),
6135 LLVMConstInt(ctx
->ac
.i32
, 4, false), false);
6136 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
6137 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
6138 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
6139 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
6140 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
6143 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6144 LLVMValueRef dw_addr
;
6145 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
6149 if (!(ctx
->output_mask
& (1ull << i
)))
6152 if (i
== VARYING_SLOT_CLIP_DIST0
)
6153 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6155 param_index
= shader_io_get_unique_index(i
);
6158 dw_addr
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6159 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
6162 for (j
= 0; j
< length
; j
++) {
6163 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], "");
6164 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->ac
.i32
, "");
6166 if (ctx
->ac
.chip_class
>= GFX9
) {
6167 ac_lds_store(&ctx
->ac
, dw_addr
,
6168 LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], ""));
6169 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6171 ac_build_buffer_store_dword(&ctx
->ac
,
6174 NULL
, ctx
->es2gs_offset
,
6175 (4 * param_index
+ j
) * 4,
6183 handle_ls_outputs_post(struct nir_to_llvm_context
*ctx
)
6185 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
6186 LLVMValueRef vertex_dw_stride
= unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 13, 8);
6187 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->builder
, vertex_id
,
6188 vertex_dw_stride
, "");
6190 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6191 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
6194 if (!(ctx
->output_mask
& (1ull << i
)))
6197 if (i
== VARYING_SLOT_CLIP_DIST0
)
6198 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6199 int param
= shader_io_get_unique_index(i
);
6200 mark_tess_output(ctx
, false, param
);
6202 mark_tess_output(ctx
, false, param
+ 1);
6203 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->builder
, base_dw_addr
,
6204 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
6206 for (unsigned j
= 0; j
< length
; j
++) {
6207 ac_lds_store(&ctx
->ac
, dw_addr
,
6208 LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], ""));
6209 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6214 struct ac_build_if_state
6216 struct nir_to_llvm_context
*ctx
;
6217 LLVMValueRef condition
;
6218 LLVMBasicBlockRef entry_block
;
6219 LLVMBasicBlockRef true_block
;
6220 LLVMBasicBlockRef false_block
;
6221 LLVMBasicBlockRef merge_block
;
6224 static LLVMBasicBlockRef
6225 ac_build_insert_new_block(struct nir_to_llvm_context
*ctx
, const char *name
)
6227 LLVMBasicBlockRef current_block
;
6228 LLVMBasicBlockRef next_block
;
6229 LLVMBasicBlockRef new_block
;
6231 /* get current basic block */
6232 current_block
= LLVMGetInsertBlock(ctx
->builder
);
6234 /* chqeck if there's another block after this one */
6235 next_block
= LLVMGetNextBasicBlock(current_block
);
6237 /* insert the new block before the next block */
6238 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
6241 /* append new block after current block */
6242 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
6243 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
6249 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
6250 struct nir_to_llvm_context
*ctx
,
6251 LLVMValueRef condition
)
6253 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->builder
);
6255 memset(ifthen
, 0, sizeof *ifthen
);
6257 ifthen
->condition
= condition
;
6258 ifthen
->entry_block
= block
;
6260 /* create endif/merge basic block for the phi functions */
6261 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
6263 /* create/insert true_block before merge_block */
6264 ifthen
->true_block
=
6265 LLVMInsertBasicBlockInContext(ctx
->context
,
6266 ifthen
->merge_block
,
6269 /* successive code goes into the true block */
6270 LLVMPositionBuilderAtEnd(ctx
->builder
, ifthen
->true_block
);
6274 * End a conditional.
6277 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
6279 LLVMBuilderRef builder
= ifthen
->ctx
->builder
;
6281 /* Insert branch to the merge block from current block */
6282 LLVMBuildBr(builder
, ifthen
->merge_block
);
6285 * Now patch in the various branch instructions.
6288 /* Insert the conditional branch instruction at the end of entry_block */
6289 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
6290 if (ifthen
->false_block
) {
6291 /* we have an else clause */
6292 LLVMBuildCondBr(builder
, ifthen
->condition
,
6293 ifthen
->true_block
, ifthen
->false_block
);
6296 /* no else clause */
6297 LLVMBuildCondBr(builder
, ifthen
->condition
,
6298 ifthen
->true_block
, ifthen
->merge_block
);
6301 /* Resume building code at end of the ifthen->merge_block */
6302 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
6306 write_tess_factors(struct nir_to_llvm_context
*ctx
)
6308 unsigned stride
, outer_comps
, inner_comps
;
6309 struct ac_build_if_state if_ctx
, inner_if_ctx
;
6310 LLVMValueRef invocation_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
6311 LLVMValueRef rel_patch_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
6312 unsigned tess_inner_index
, tess_outer_index
;
6313 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
6314 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
6316 emit_barrier(&ctx
->ac
, ctx
->stage
);
6318 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
6338 ac_nir_build_if(&if_ctx
, ctx
,
6339 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
6340 invocation_id
, ctx
->ac
.i32_0
, ""));
6342 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6343 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6345 mark_tess_output(ctx
, true, tess_inner_index
);
6346 mark_tess_output(ctx
, true, tess_outer_index
);
6347 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
6348 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6349 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
6350 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6351 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
6353 for (i
= 0; i
< 4; i
++) {
6354 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6355 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6359 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
6360 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
6361 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
6363 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
6365 for (i
= 0; i
< outer_comps
; i
++) {
6367 ac_lds_load(&ctx
->ac
, lds_outer
);
6368 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
6371 for (i
= 0; i
< inner_comps
; i
++) {
6372 inner
[i
] = out
[outer_comps
+i
] =
6373 ac_lds_load(&ctx
->ac
, lds_inner
);
6374 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_inner
,
6379 /* Convert the outputs to vectors for stores. */
6380 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
6384 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
6387 buffer
= ctx
->hs_ring_tess_factor
;
6388 tf_base
= ctx
->tess_factor_offset
;
6389 byteoffset
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
6390 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
6391 unsigned tf_offset
= 0;
6393 if (ctx
->options
->chip_class
<= VI
) {
6394 ac_nir_build_if(&inner_if_ctx
, ctx
,
6395 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
6396 rel_patch_id
, ctx
->ac
.i32_0
, ""));
6398 /* Store the dynamic HS control word. */
6399 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
6400 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
6401 1, ctx
->ac
.i32_0
, tf_base
,
6402 0, 1, 0, true, false);
6405 ac_nir_build_endif(&inner_if_ctx
);
6408 /* Store the tessellation factors. */
6409 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
6410 MIN2(stride
, 4), byteoffset
, tf_base
,
6411 tf_offset
, 1, 0, true, false);
6413 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
6414 stride
- 4, byteoffset
, tf_base
,
6415 16 + tf_offset
, 1, 0, true, false);
6417 //store to offchip for TES to read - only if TES reads them
6418 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
6419 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
6420 LLVMValueRef tf_inner_offset
;
6421 unsigned param_outer
, param_inner
;
6423 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6424 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6425 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
6427 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
6428 util_next_power_of_two(outer_comps
));
6430 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
6431 outer_comps
, tf_outer_offset
,
6432 ctx
->oc_lds
, 0, 1, 0, true, false);
6434 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6435 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6436 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
6438 inner_vec
= inner_comps
== 1 ? inner
[0] :
6439 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
6440 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
6441 inner_comps
, tf_inner_offset
,
6442 ctx
->oc_lds
, 0, 1, 0, true, false);
6445 ac_nir_build_endif(&if_ctx
);
6449 handle_tcs_outputs_post(struct nir_to_llvm_context
*ctx
)
6451 write_tess_factors(ctx
);
6455 si_export_mrt_color(struct nir_to_llvm_context
*ctx
,
6456 LLVMValueRef
*color
, unsigned param
, bool is_last
,
6457 struct ac_export_args
*args
)
6460 si_llvm_init_export_args(ctx
, color
, param
,
6464 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
6465 args
->done
= 1; /* DONE bit */
6466 } else if (!args
->enabled_channels
)
6467 return false; /* unnecessary NULL export */
6473 radv_export_mrt_z(struct nir_to_llvm_context
*ctx
,
6474 LLVMValueRef depth
, LLVMValueRef stencil
,
6475 LLVMValueRef samplemask
)
6477 struct ac_export_args args
;
6479 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
6481 ac_build_export(&ctx
->ac
, &args
);
6485 handle_fs_outputs_post(struct nir_to_llvm_context
*ctx
)
6488 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
6489 struct ac_export_args color_args
[8];
6491 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6492 LLVMValueRef values
[4];
6494 if (!(ctx
->output_mask
& (1ull << i
)))
6497 if (i
== FRAG_RESULT_DEPTH
) {
6498 ctx
->shader_info
->fs
.writes_z
= true;
6499 depth
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6500 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6501 } else if (i
== FRAG_RESULT_STENCIL
) {
6502 ctx
->shader_info
->fs
.writes_stencil
= true;
6503 stencil
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6504 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6505 } else if (i
== FRAG_RESULT_SAMPLE_MASK
) {
6506 ctx
->shader_info
->fs
.writes_sample_mask
= true;
6507 samplemask
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6508 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6511 for (unsigned j
= 0; j
< 4; j
++)
6512 values
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6513 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
6515 if (!ctx
->shader_info
->fs
.writes_z
&& !ctx
->shader_info
->fs
.writes_stencil
&& !ctx
->shader_info
->fs
.writes_sample_mask
)
6516 last
= ctx
->output_mask
<= ((1ull << (i
+ 1)) - 1);
6518 bool ret
= si_export_mrt_color(ctx
, values
, V_008DFC_SQ_EXP_MRT
+ (i
- FRAG_RESULT_DATA0
), last
, &color_args
[index
]);
6524 for (unsigned i
= 0; i
< index
; i
++)
6525 ac_build_export(&ctx
->ac
, &color_args
[i
]);
6526 if (depth
|| stencil
|| samplemask
)
6527 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
6529 si_export_mrt_color(ctx
, NULL
, V_008DFC_SQ_EXP_NULL
, true, &color_args
[0]);
6530 ac_build_export(&ctx
->ac
, &color_args
[0]);
6535 emit_gs_epilogue(struct nir_to_llvm_context
*ctx
)
6537 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
6541 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
6542 LLVMValueRef
*addrs
)
6544 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
6546 switch (ctx
->stage
) {
6547 case MESA_SHADER_VERTEX
:
6548 if (ctx
->options
->key
.vs
.as_ls
)
6549 handle_ls_outputs_post(ctx
);
6550 else if (ctx
->options
->key
.vs
.as_es
)
6551 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
6553 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
6554 &ctx
->shader_info
->vs
.outinfo
);
6556 case MESA_SHADER_FRAGMENT
:
6557 handle_fs_outputs_post(ctx
);
6559 case MESA_SHADER_GEOMETRY
:
6560 emit_gs_epilogue(ctx
);
6562 case MESA_SHADER_TESS_CTRL
:
6563 handle_tcs_outputs_post(ctx
);
6565 case MESA_SHADER_TESS_EVAL
:
6566 if (ctx
->options
->key
.tes
.as_es
)
6567 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
6569 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
6570 &ctx
->shader_info
->tes
.outinfo
);
6577 static void ac_llvm_finalize_module(struct nir_to_llvm_context
* ctx
)
6579 LLVMPassManagerRef passmgr
;
6580 /* Create the pass manager */
6581 passmgr
= LLVMCreateFunctionPassManagerForModule(
6584 /* This pass should eliminate all the load and store instructions */
6585 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
6587 /* Add some optimization passes */
6588 LLVMAddScalarReplAggregatesPass(passmgr
);
6589 LLVMAddLICMPass(passmgr
);
6590 LLVMAddAggressiveDCEPass(passmgr
);
6591 LLVMAddCFGSimplificationPass(passmgr
);
6592 LLVMAddInstructionCombiningPass(passmgr
);
6595 LLVMInitializeFunctionPassManager(passmgr
);
6596 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
6597 LLVMFinalizeFunctionPassManager(passmgr
);
6599 LLVMDisposeBuilder(ctx
->builder
);
6600 LLVMDisposePassManager(passmgr
);
6604 ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context
*ctx
)
6606 struct ac_vs_output_info
*outinfo
;
6608 switch (ctx
->stage
) {
6609 case MESA_SHADER_FRAGMENT
:
6610 case MESA_SHADER_COMPUTE
:
6611 case MESA_SHADER_TESS_CTRL
:
6612 case MESA_SHADER_GEOMETRY
:
6614 case MESA_SHADER_VERTEX
:
6615 if (ctx
->options
->key
.vs
.as_ls
||
6616 ctx
->options
->key
.vs
.as_es
)
6618 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
6620 case MESA_SHADER_TESS_EVAL
:
6621 if (ctx
->options
->key
.vs
.as_es
)
6623 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
6626 unreachable("Unhandled shader type");
6629 ac_optimize_vs_outputs(&ctx
->ac
,
6631 outinfo
->vs_output_param_offset
,
6633 &outinfo
->param_exports
);
6637 ac_setup_rings(struct nir_to_llvm_context
*ctx
)
6639 if ((ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_es
) ||
6640 (ctx
->stage
== MESA_SHADER_TESS_EVAL
&& ctx
->options
->key
.tes
.as_es
)) {
6641 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_VS
, false));
6644 if (ctx
->is_gs_copy_shader
) {
6645 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_VS
, false));
6647 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
6649 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_GS
, false));
6650 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_GS
, false));
6652 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->builder
, ctx
->gsvs_ring
, ctx
->ac
.v4i32
, "");
6654 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->gsvs_num_entries
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
6655 tmp
= LLVMBuildExtractElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->ac
.i32_1
, "");
6656 tmp
= LLVMBuildOr(ctx
->builder
, tmp
, ctx
->gsvs_ring_stride
, "");
6657 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, tmp
, ctx
->ac
.i32_1
, "");
6660 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
6661 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
6662 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
6663 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
6668 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
6669 const struct nir_shader
*nir
)
6671 switch (nir
->info
.stage
) {
6672 case MESA_SHADER_TESS_CTRL
:
6673 return chip_class
>= CIK
? 128 : 64;
6674 case MESA_SHADER_GEOMETRY
:
6675 return chip_class
>= GFX9
? 128 : 64;
6676 case MESA_SHADER_COMPUTE
:
6682 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
6683 nir
->info
.cs
.local_size
[1] *
6684 nir
->info
.cs
.local_size
[2];
6685 return max_workgroup_size
;
6688 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
6689 static void ac_nir_fixup_ls_hs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6691 LLVMValueRef count
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6692 LLVMConstInt(ctx
->ac
.i32
, 8, false),
6693 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6694 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
6696 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
6697 ctx
->vs_prim_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.vertex_id
, ctx
->vs_prim_id
, "");
6698 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
6699 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
6702 static void prepare_gs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6704 for(int i
= 5; i
>= 0; --i
) {
6705 ctx
->gs_vtx_offset
[i
] = ac_build_bfe(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
6706 LLVMConstInt(ctx
->ac
.i32
, (i
& 1) * 16, false),
6707 LLVMConstInt(ctx
->ac
.i32
, 16, false), false);
6710 ctx
->gs_wave_id
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6711 LLVMConstInt(ctx
->ac
.i32
, 16, false),
6712 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6715 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
6716 struct nir_shader
*nir
, struct nir_to_llvm_context
*nctx
)
6718 struct ac_nir_context ctx
= {};
6719 struct nir_function
*func
;
6728 ctx
.stage
= nir
->info
.stage
;
6730 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6732 nir_foreach_variable(variable
, &nir
->outputs
)
6733 handle_shader_output_decl(&ctx
, nir
, variable
);
6735 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6736 _mesa_key_pointer_equal
);
6737 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6738 _mesa_key_pointer_equal
);
6739 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6740 _mesa_key_pointer_equal
);
6742 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
6744 setup_locals(&ctx
, func
);
6746 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
)
6747 setup_shared(&ctx
, nir
);
6749 visit_cf_list(&ctx
, &func
->impl
->body
);
6750 phi_post_pass(&ctx
);
6752 ctx
.abi
->emit_outputs(ctx
.abi
, RADEON_LLVM_MAX_OUTPUTS
,
6756 ralloc_free(ctx
.defs
);
6757 ralloc_free(ctx
.phis
);
6758 ralloc_free(ctx
.vars
);
6765 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
6766 struct nir_shader
*const *shaders
,
6768 struct ac_shader_variant_info
*shader_info
,
6769 const struct ac_nir_compiler_options
*options
)
6771 struct nir_to_llvm_context ctx
= {0};
6773 ctx
.options
= options
;
6774 ctx
.shader_info
= shader_info
;
6775 ctx
.context
= LLVMContextCreate();
6776 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
6778 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
6780 ctx
.ac
.module
= ctx
.module
;
6781 LLVMSetTarget(ctx
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
6783 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
6784 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
6785 LLVMSetDataLayout(ctx
.module
, data_layout_str
);
6786 LLVMDisposeTargetData(data_layout
);
6787 LLVMDisposeMessage(data_layout_str
);
6789 enum ac_float_mode float_mode
=
6790 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
6791 AC_FLOAT_MODE_DEFAULT
;
6793 ctx
.builder
= ac_create_builder(ctx
.context
, float_mode
);
6794 ctx
.ac
.builder
= ctx
.builder
;
6796 memset(shader_info
, 0, sizeof(*shader_info
));
6798 for(int i
= 0; i
< shader_count
; ++i
)
6799 ac_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
6801 for (i
= 0; i
< AC_UD_MAX_SETS
; i
++)
6802 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
6803 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
6804 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
6806 ctx
.max_workgroup_size
= 0;
6807 for (int i
= 0; i
< shader_count
; ++i
) {
6808 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
6809 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
6813 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
6814 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
6816 ctx
.abi
.inputs
= &ctx
.inputs
[0];
6817 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
6818 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
6819 ctx
.abi
.load_ubo
= radv_load_ubo
;
6820 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
6821 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
6822 ctx
.abi
.clamp_shadow_reference
= false;
6824 if (shader_count
>= 2)
6825 ac_init_exec_full_mask(&ctx
.ac
);
6827 if (ctx
.ac
.chip_class
== GFX9
&&
6828 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
6829 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
6831 for(int i
= 0; i
< shader_count
; ++i
) {
6832 ctx
.stage
= shaders
[i
]->info
.stage
;
6833 ctx
.output_mask
= 0;
6834 ctx
.tess_outputs_written
= 0;
6835 ctx
.num_output_clips
= shaders
[i
]->info
.clip_distance_array_size
;
6836 ctx
.num_output_culls
= shaders
[i
]->info
.cull_distance_array_size
;
6838 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6839 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "gs_next_vertex");
6840 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
6841 ctx
.abi
.load_inputs
= load_gs_input
;
6842 ctx
.abi
.emit_primitive
= visit_end_primitive
;
6843 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6844 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
6845 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
6846 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
6847 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
6848 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
6849 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
6850 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
6851 ctx
.abi
.load_tess_varyings
= load_tes_input
;
6852 ctx
.abi
.load_tess_coord
= load_tess_coord
;
6853 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
6854 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
6855 if (shader_info
->info
.vs
.needs_instance_id
) {
6856 if (ctx
.options
->key
.vs
.as_ls
) {
6857 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6858 MAX2(2, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6860 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6861 MAX2(1, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6864 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
6865 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
6866 ctx
.abi
.lookup_interp_param
= lookup_interp_param
;
6867 ctx
.abi
.load_sample_position
= load_sample_position
;
6871 emit_barrier(&ctx
.ac
, ctx
.stage
);
6873 ac_setup_rings(&ctx
);
6875 LLVMBasicBlockRef merge_block
;
6876 if (shader_count
>= 2) {
6877 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6878 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6879 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6881 LLVMValueRef count
= ac_build_bfe(&ctx
.ac
, ctx
.merged_wave_info
,
6882 LLVMConstInt(ctx
.ac
.i32
, 8 * i
, false),
6883 LLVMConstInt(ctx
.ac
.i32
, 8, false), false);
6884 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
6885 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
6886 thread_id
, count
, "");
6887 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
6889 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
6892 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
6893 handle_fs_inputs(&ctx
, shaders
[i
]);
6894 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
6895 handle_vs_inputs(&ctx
, shaders
[i
]);
6896 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
6897 prepare_gs_input_vgprs(&ctx
);
6899 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
6900 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
6902 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
], &ctx
);
6904 if (shader_count
>= 2) {
6905 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
6906 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
6909 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6910 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
6911 shaders
[i
]->info
.cull_distance_array_size
> 4;
6912 shader_info
->gs
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
6913 shader_info
->gs
.max_gsvs_emit_size
= shader_info
->gs
.gsvs_vertex_size
*
6914 shaders
[i
]->info
.gs
.vertices_out
;
6915 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6916 shader_info
->tcs
.outputs_written
= ctx
.tess_outputs_written
;
6917 shader_info
->tcs
.patch_outputs_written
= ctx
.tess_patch_outputs_written
;
6918 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
&& ctx
.options
->key
.vs
.as_ls
) {
6919 shader_info
->vs
.outputs_written
= ctx
.tess_outputs_written
;
6923 LLVMBuildRetVoid(ctx
.builder
);
6925 if (options
->dump_preoptir
)
6926 ac_dump_module(ctx
.module
);
6928 ac_llvm_finalize_module(&ctx
);
6930 if (shader_count
== 1)
6931 ac_nir_eliminate_const_vs_outputs(&ctx
);
6936 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
6938 unsigned *retval
= (unsigned *)context
;
6939 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
6940 char *description
= LLVMGetDiagInfoDescription(di
);
6942 if (severity
== LLVMDSError
) {
6944 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
6948 LLVMDisposeMessage(description
);
6951 static unsigned ac_llvm_compile(LLVMModuleRef M
,
6952 struct ac_shader_binary
*binary
,
6953 LLVMTargetMachineRef tm
)
6955 unsigned retval
= 0;
6957 LLVMContextRef llvm_ctx
;
6958 LLVMMemoryBufferRef out_buffer
;
6959 unsigned buffer_size
;
6960 const char *buffer_data
;
6963 /* Setup Diagnostic Handler*/
6964 llvm_ctx
= LLVMGetModuleContext(M
);
6966 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
6970 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
6973 /* Process Errors/Warnings */
6975 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
6981 /* Extract Shader Code*/
6982 buffer_size
= LLVMGetBufferSize(out_buffer
);
6983 buffer_data
= LLVMGetBufferStart(out_buffer
);
6985 ac_elf_read(buffer_data
, buffer_size
, binary
);
6988 LLVMDisposeMemoryBuffer(out_buffer
);
6994 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
6995 LLVMModuleRef llvm_module
,
6996 struct ac_shader_binary
*binary
,
6997 struct ac_shader_config
*config
,
6998 struct ac_shader_variant_info
*shader_info
,
6999 gl_shader_stage stage
,
7000 bool dump_shader
, bool supports_spill
)
7003 ac_dump_module(llvm_module
);
7005 memset(binary
, 0, sizeof(*binary
));
7006 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
7008 fprintf(stderr
, "compile failed\n");
7012 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
7014 ac_shader_binary_read_config(binary
, config
, 0, supports_spill
);
7016 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
7017 LLVMDisposeModule(llvm_module
);
7018 LLVMContextDispose(ctx
);
7020 if (stage
== MESA_SHADER_FRAGMENT
) {
7021 shader_info
->num_input_vgprs
= 0;
7022 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
7023 shader_info
->num_input_vgprs
+= 2;
7024 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
7025 shader_info
->num_input_vgprs
+= 2;
7026 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
7027 shader_info
->num_input_vgprs
+= 2;
7028 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
7029 shader_info
->num_input_vgprs
+= 3;
7030 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
7031 shader_info
->num_input_vgprs
+= 2;
7032 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
7033 shader_info
->num_input_vgprs
+= 2;
7034 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
7035 shader_info
->num_input_vgprs
+= 2;
7036 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
7037 shader_info
->num_input_vgprs
+= 1;
7038 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
7039 shader_info
->num_input_vgprs
+= 1;
7040 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
7041 shader_info
->num_input_vgprs
+= 1;
7042 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
7043 shader_info
->num_input_vgprs
+= 1;
7044 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
7045 shader_info
->num_input_vgprs
+= 1;
7046 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
7047 shader_info
->num_input_vgprs
+= 1;
7048 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
7049 shader_info
->num_input_vgprs
+= 1;
7050 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
7051 shader_info
->num_input_vgprs
+= 1;
7052 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
7053 shader_info
->num_input_vgprs
+= 1;
7055 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
7057 /* +3 for scratch wave offset and VCC */
7058 config
->num_sgprs
= MAX2(config
->num_sgprs
,
7059 shader_info
->num_input_sgprs
+ 3);
7061 /* Enable 64-bit and 16-bit denormals, because there is no performance
7064 * If denormals are enabled, all floating-point output modifiers are
7067 * Don't enable denormals for 32-bit floats, because:
7068 * - Floating-point output modifiers would be ignored by the hw.
7069 * - Some opcodes don't support denormals, such as v_mad_f32. We would
7070 * have to stop using those.
7071 * - SI & CI would be very slow.
7073 config
->float_mode
|= V_00B028_FP_64_DENORMS
;
7077 ac_fill_shader_info(struct ac_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct ac_nir_compiler_options
*options
)
7079 switch (nir
->info
.stage
) {
7080 case MESA_SHADER_COMPUTE
:
7081 for (int i
= 0; i
< 3; ++i
)
7082 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
7084 case MESA_SHADER_FRAGMENT
:
7085 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
7087 case MESA_SHADER_GEOMETRY
:
7088 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
7089 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
7090 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
7091 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
7093 case MESA_SHADER_TESS_EVAL
:
7094 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
7095 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
7096 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
7097 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
7098 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
7100 case MESA_SHADER_TESS_CTRL
:
7101 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
7103 case MESA_SHADER_VERTEX
:
7104 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
7105 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
7106 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
7107 if (options
->key
.vs
.as_ls
)
7108 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
7115 void ac_compile_nir_shader(LLVMTargetMachineRef tm
,
7116 struct ac_shader_binary
*binary
,
7117 struct ac_shader_config
*config
,
7118 struct ac_shader_variant_info
*shader_info
,
7119 struct nir_shader
*const *nir
,
7121 const struct ac_nir_compiler_options
*options
,
7125 LLVMModuleRef llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, nir_count
, shader_info
,
7128 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
, nir
[0]->info
.stage
, dump_shader
, options
->supports_spill
);
7129 for (int i
= 0; i
< nir_count
; ++i
)
7130 ac_fill_shader_info(shader_info
, nir
[i
], options
);
7132 /* Determine the ES type (VS or TES) for the GS on GFX9. */
7133 if (options
->chip_class
== GFX9
) {
7134 if (nir_count
== 2 &&
7135 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
7136 shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
7142 ac_gs_copy_shader_emit(struct nir_to_llvm_context
*ctx
)
7144 LLVMValueRef vtx_offset
=
7145 LLVMBuildMul(ctx
->builder
, ctx
->abi
.vertex_id
,
7146 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
7149 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
7153 if (!(ctx
->output_mask
& (1ull << i
)))
7156 if (i
== VARYING_SLOT_CLIP_DIST0
) {
7157 /* unpack clip and cull from a single set of slots */
7158 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
7163 for (unsigned j
= 0; j
< length
; j
++) {
7164 LLVMValueRef value
, soffset
;
7166 soffset
= LLVMConstInt(ctx
->ac
.i32
,
7168 ctx
->gs_max_out_vertices
* 16 * 4, false);
7170 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->gsvs_ring
,
7172 vtx_offset
, soffset
,
7173 0, 1, 1, true, false);
7175 LLVMBuildStore(ctx
->builder
,
7176 ac_to_float(&ctx
->ac
, value
), ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)]);
7180 handle_vs_outputs_post(ctx
, false, &ctx
->shader_info
->vs
.outinfo
);
7183 void ac_create_gs_copy_shader(LLVMTargetMachineRef tm
,
7184 struct nir_shader
*geom_shader
,
7185 struct ac_shader_binary
*binary
,
7186 struct ac_shader_config
*config
,
7187 struct ac_shader_variant_info
*shader_info
,
7188 const struct ac_nir_compiler_options
*options
,
7191 struct nir_to_llvm_context ctx
= {0};
7192 ctx
.context
= LLVMContextCreate();
7193 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
7194 ctx
.options
= options
;
7195 ctx
.shader_info
= shader_info
;
7197 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
7199 ctx
.ac
.module
= ctx
.module
;
7201 ctx
.is_gs_copy_shader
= true;
7202 LLVMSetTarget(ctx
.module
, "amdgcn--");
7204 enum ac_float_mode float_mode
=
7205 options
->unsafe_math
? AC_FLOAT_MODE_UNSAFE_FP_MATH
:
7206 AC_FLOAT_MODE_DEFAULT
;
7208 ctx
.builder
= ac_create_builder(ctx
.context
, float_mode
);
7209 ctx
.ac
.builder
= ctx
.builder
;
7210 ctx
.stage
= MESA_SHADER_VERTEX
;
7212 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
7214 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
7215 ac_setup_rings(&ctx
);
7217 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
7218 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
7220 struct ac_nir_context nir_ctx
= {};
7221 nir_ctx
.ac
= ctx
.ac
;
7222 nir_ctx
.abi
= &ctx
.abi
;
7224 nir_ctx
.nctx
= &ctx
;
7227 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
7228 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
7229 handle_shader_output_decl(&nir_ctx
, geom_shader
, variable
);
7232 ac_gs_copy_shader_emit(&ctx
);
7236 LLVMBuildRetVoid(ctx
.builder
);
7238 ac_llvm_finalize_module(&ctx
);
7240 ac_compile_llvm_module(tm
, ctx
.module
, binary
, config
, shader_info
,
7242 dump_shader
, options
->supports_spill
);