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 CONST_ADDR_SPACE 2
47 #define LOCAL_ADDR_SPACE 3
49 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
50 #define RADEON_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
52 struct nir_to_llvm_context
;
54 struct ac_nir_context
{
55 struct ac_llvm_context ac
;
56 struct ac_shader_abi
*abi
;
58 gl_shader_stage stage
;
60 struct hash_table
*defs
;
61 struct hash_table
*phis
;
62 struct hash_table
*vars
;
64 LLVMValueRef main_function
;
65 LLVMBasicBlockRef continue_block
;
66 LLVMBasicBlockRef break_block
;
68 LLVMValueRef outputs
[RADEON_LLVM_MAX_OUTPUTS
* 4];
73 struct nir_to_llvm_context
*nctx
; /* TODO get rid of this */
76 struct nir_to_llvm_context
{
77 struct ac_llvm_context ac
;
78 const struct ac_nir_compiler_options
*options
;
79 struct ac_shader_variant_info
*shader_info
;
80 struct ac_shader_abi abi
;
81 struct ac_nir_context
*nir
;
83 unsigned max_workgroup_size
;
84 LLVMContextRef context
;
86 LLVMBuilderRef builder
;
87 LLVMValueRef main_function
;
89 struct hash_table
*defs
;
90 struct hash_table
*phis
;
92 LLVMValueRef descriptor_sets
[AC_UD_MAX_SETS
];
93 LLVMValueRef ring_offsets
;
94 LLVMValueRef push_constants
;
95 LLVMValueRef view_index
;
96 LLVMValueRef num_work_groups
;
97 LLVMValueRef workgroup_ids
[3];
98 LLVMValueRef local_invocation_ids
;
101 LLVMValueRef vertex_buffers
;
102 LLVMValueRef rel_auto_id
;
103 LLVMValueRef vs_prim_id
;
104 LLVMValueRef ls_out_layout
;
105 LLVMValueRef es2gs_offset
;
107 LLVMValueRef tcs_offchip_layout
;
108 LLVMValueRef tcs_out_offsets
;
109 LLVMValueRef tcs_out_layout
;
110 LLVMValueRef tcs_in_layout
;
112 LLVMValueRef merged_wave_info
;
113 LLVMValueRef tess_factor_offset
;
114 LLVMValueRef tes_rel_patch_id
;
118 LLVMValueRef gsvs_ring_stride
;
119 LLVMValueRef gsvs_num_entries
;
120 LLVMValueRef gs2vs_offset
;
121 LLVMValueRef gs_wave_id
;
122 LLVMValueRef gs_vtx_offset
[6];
124 LLVMValueRef esgs_ring
;
125 LLVMValueRef gsvs_ring
;
126 LLVMValueRef hs_ring_tess_offchip
;
127 LLVMValueRef hs_ring_tess_factor
;
129 LLVMValueRef prim_mask
;
130 LLVMValueRef sample_pos_offset
;
131 LLVMValueRef persp_sample
, persp_center
, persp_centroid
;
132 LLVMValueRef linear_sample
, linear_center
, linear_centroid
;
134 gl_shader_stage stage
;
136 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
139 uint64_t output_mask
;
140 uint8_t num_output_clips
;
141 uint8_t num_output_culls
;
143 bool is_gs_copy_shader
;
144 LLVMValueRef gs_next_vertex
;
145 unsigned gs_max_out_vertices
;
147 unsigned tes_primitive_mode
;
148 uint64_t tess_outputs_written
;
149 uint64_t tess_patch_outputs_written
;
151 uint32_t tcs_patch_outputs_read
;
152 uint64_t tcs_outputs_read
;
155 static inline struct nir_to_llvm_context
*
156 nir_to_llvm_context_from_abi(struct ac_shader_abi
*abi
)
158 struct nir_to_llvm_context
*ctx
= NULL
;
159 return container_of(abi
, ctx
, abi
);
163 nir2llvmtype(struct ac_nir_context
*ctx
,
164 const struct glsl_type
*type
)
166 switch (glsl_get_base_type(glsl_without_array(type
))) {
170 case GLSL_TYPE_UINT64
:
171 case GLSL_TYPE_INT64
:
173 case GLSL_TYPE_DOUBLE
:
175 case GLSL_TYPE_FLOAT
:
178 assert(!"Unsupported type in nir2llvmtype()");
184 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
185 const nir_deref_var
*deref
,
186 enum ac_descriptor_type desc_type
,
187 const nir_tex_instr
*instr
,
188 bool image
, bool write
);
190 static unsigned radeon_llvm_reg_index_soa(unsigned index
, unsigned chan
)
192 return (index
* 4) + chan
;
195 static unsigned shader_io_get_unique_index(gl_varying_slot slot
)
197 /* handle patch indices separate */
198 if (slot
== VARYING_SLOT_TESS_LEVEL_OUTER
)
200 if (slot
== VARYING_SLOT_TESS_LEVEL_INNER
)
202 if (slot
>= VARYING_SLOT_PATCH0
&& slot
<= VARYING_SLOT_TESS_MAX
)
203 return 2 + (slot
- VARYING_SLOT_PATCH0
);
205 if (slot
== VARYING_SLOT_POS
)
207 if (slot
== VARYING_SLOT_PSIZ
)
209 if (slot
== VARYING_SLOT_CLIP_DIST0
)
211 /* 3 is reserved for clip dist as well */
212 if (slot
>= VARYING_SLOT_VAR0
&& slot
<= VARYING_SLOT_VAR31
)
213 return 4 + (slot
- VARYING_SLOT_VAR0
);
214 unreachable("illegal slot in get unique index\n");
217 static void set_llvm_calling_convention(LLVMValueRef func
,
218 gl_shader_stage stage
)
220 enum radeon_llvm_calling_convention calling_conv
;
223 case MESA_SHADER_VERTEX
:
224 case MESA_SHADER_TESS_EVAL
:
225 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
227 case MESA_SHADER_GEOMETRY
:
228 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
230 case MESA_SHADER_TESS_CTRL
:
231 calling_conv
= HAVE_LLVM
>= 0x0500 ? RADEON_LLVM_AMDGPU_HS
: RADEON_LLVM_AMDGPU_VS
;
233 case MESA_SHADER_FRAGMENT
:
234 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
236 case MESA_SHADER_COMPUTE
:
237 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
240 unreachable("Unhandle shader type");
243 LLVMSetFunctionCallConv(func
, calling_conv
);
248 LLVMTypeRef types
[MAX_ARGS
];
249 LLVMValueRef
*assign
[MAX_ARGS
];
250 unsigned array_params_mask
;
253 uint8_t num_sgprs_used
;
254 uint8_t num_vgprs_used
;
257 enum ac_arg_regfile
{
263 add_arg(struct arg_info
*info
, enum ac_arg_regfile regfile
, LLVMTypeRef type
,
264 LLVMValueRef
*param_ptr
)
266 assert(info
->count
< MAX_ARGS
);
268 info
->assign
[info
->count
] = param_ptr
;
269 info
->types
[info
->count
] = type
;
272 if (regfile
== ARG_SGPR
) {
273 info
->num_sgprs_used
+= ac_get_type_size(type
) / 4;
276 assert(regfile
== ARG_VGPR
);
277 info
->num_vgprs_used
+= ac_get_type_size(type
) / 4;
282 add_array_arg(struct arg_info
*info
, LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
284 info
->array_params_mask
|= (1 << info
->count
);
285 add_arg(info
, ARG_SGPR
, type
, param_ptr
);
288 static void assign_arguments(LLVMValueRef main_function
,
289 struct arg_info
*info
)
292 for (i
= 0; i
< info
->count
; i
++) {
294 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
299 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
300 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
301 unsigned num_return_elems
,
302 struct arg_info
*args
,
303 unsigned max_workgroup_size
,
306 LLVMTypeRef main_function_type
, ret_type
;
307 LLVMBasicBlockRef main_function_body
;
309 if (num_return_elems
)
310 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
311 num_return_elems
, true);
313 ret_type
= LLVMVoidTypeInContext(ctx
);
315 /* Setup the function */
317 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
318 LLVMValueRef main_function
=
319 LLVMAddFunction(module
, "main", main_function_type
);
321 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
322 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
324 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
325 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
326 if (args
->array_params_mask
& (1 << i
)) {
327 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
328 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_BYVAL
);
329 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
332 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
336 if (max_workgroup_size
) {
337 ac_llvm_add_target_dep_function_attr(main_function
,
338 "amdgpu-max-work-group-size",
342 /* These were copied from some LLVM test. */
343 LLVMAddTargetDependentFunctionAttr(main_function
,
344 "less-precise-fpmad",
346 LLVMAddTargetDependentFunctionAttr(main_function
,
349 LLVMAddTargetDependentFunctionAttr(main_function
,
352 LLVMAddTargetDependentFunctionAttr(main_function
,
356 return main_function
;
359 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
361 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
365 static int get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
367 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
368 type
= LLVMGetElementType(type
);
370 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
371 return LLVMGetIntTypeWidth(type
);
373 if (type
== ctx
->f16
)
375 if (type
== ctx
->f32
)
377 if (type
== ctx
->f64
)
380 unreachable("Unhandled type kind in get_elem_bits");
383 static LLVMValueRef
unpack_param(struct ac_llvm_context
*ctx
,
384 LLVMValueRef param
, unsigned rshift
,
387 LLVMValueRef value
= param
;
389 value
= LLVMBuildLShr(ctx
->builder
, value
,
390 LLVMConstInt(ctx
->i32
, rshift
, false), "");
392 if (rshift
+ bitwidth
< 32) {
393 unsigned mask
= (1 << bitwidth
) - 1;
394 value
= LLVMBuildAnd(ctx
->builder
, value
,
395 LLVMConstInt(ctx
->i32
, mask
, false), "");
400 static LLVMValueRef
get_rel_patch_id(struct nir_to_llvm_context
*ctx
)
402 switch (ctx
->stage
) {
403 case MESA_SHADER_TESS_CTRL
:
404 return unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
405 case MESA_SHADER_TESS_EVAL
:
406 return ctx
->tes_rel_patch_id
;
409 unreachable("Illegal stage");
413 /* Tessellation shaders pass outputs to the next shader using LDS.
415 * LS outputs = TCS inputs
416 * TCS outputs = TES inputs
419 * - TCS inputs for patch 0
420 * - TCS inputs for patch 1
421 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
423 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
424 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
425 * - TCS outputs for patch 1
426 * - Per-patch TCS outputs for patch 1
427 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
428 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
431 * All three shaders VS(LS), TCS, TES share the same LDS space.
434 get_tcs_in_patch_stride(struct nir_to_llvm_context
*ctx
)
436 if (ctx
->stage
== MESA_SHADER_VERTEX
)
437 return unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 0, 13);
438 else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
439 return unpack_param(&ctx
->ac
, ctx
->tcs_in_layout
, 0, 13);
447 get_tcs_out_patch_stride(struct nir_to_llvm_context
*ctx
)
449 return unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 0, 13);
453 get_tcs_out_patch0_offset(struct nir_to_llvm_context
*ctx
)
455 return LLVMBuildMul(ctx
->builder
,
456 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 0, 16),
457 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
461 get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context
*ctx
)
463 return LLVMBuildMul(ctx
->builder
,
464 unpack_param(&ctx
->ac
, ctx
->tcs_out_offsets
, 16, 16),
465 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
469 get_tcs_in_current_patch_offset(struct nir_to_llvm_context
*ctx
)
471 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
472 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
474 return LLVMBuildMul(ctx
->builder
, patch_stride
, rel_patch_id
, "");
478 get_tcs_out_current_patch_offset(struct nir_to_llvm_context
*ctx
)
480 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
481 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
482 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
484 return LLVMBuildAdd(ctx
->builder
, patch0_offset
,
485 LLVMBuildMul(ctx
->builder
, patch_stride
,
491 get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context
*ctx
)
493 LLVMValueRef patch0_patch_data_offset
=
494 get_tcs_out_patch0_patch_data_offset(ctx
);
495 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
496 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
498 return LLVMBuildAdd(ctx
->builder
, patch0_patch_data_offset
,
499 LLVMBuildMul(ctx
->builder
, patch_stride
,
505 set_loc(struct ac_userdata_info
*ud_info
, uint8_t *sgpr_idx
, uint8_t num_sgprs
,
506 uint32_t indirect_offset
)
508 ud_info
->sgpr_idx
= *sgpr_idx
;
509 ud_info
->num_sgprs
= num_sgprs
;
510 ud_info
->indirect
= indirect_offset
> 0;
511 ud_info
->indirect_offset
= indirect_offset
;
512 *sgpr_idx
+= num_sgprs
;
516 set_loc_shader(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
519 struct ac_userdata_info
*ud_info
=
520 &ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
];
523 set_loc(ud_info
, sgpr_idx
, num_sgprs
, 0);
527 set_loc_desc(struct nir_to_llvm_context
*ctx
, int idx
, uint8_t *sgpr_idx
,
528 uint32_t indirect_offset
)
530 struct ac_userdata_info
*ud_info
=
531 &ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[idx
];
534 set_loc(ud_info
, sgpr_idx
, 2, indirect_offset
);
537 struct user_sgpr_info
{
538 bool need_ring_offsets
;
540 bool indirect_all_descriptor_sets
;
543 static void allocate_user_sgprs(struct nir_to_llvm_context
*ctx
,
544 struct user_sgpr_info
*user_sgpr_info
)
546 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
548 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
549 if (ctx
->stage
== MESA_SHADER_GEOMETRY
||
550 ctx
->stage
== MESA_SHADER_VERTEX
||
551 ctx
->stage
== MESA_SHADER_TESS_CTRL
||
552 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
553 ctx
->is_gs_copy_shader
)
554 user_sgpr_info
->need_ring_offsets
= true;
556 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&&
557 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
558 user_sgpr_info
->need_ring_offsets
= true;
560 /* 2 user sgprs will nearly always be allocated for scratch/rings */
561 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
562 user_sgpr_info
->sgpr_count
+= 2;
565 switch (ctx
->stage
) {
566 case MESA_SHADER_COMPUTE
:
567 if (ctx
->shader_info
->info
.cs
.uses_grid_size
)
568 user_sgpr_info
->sgpr_count
+= 3;
570 case MESA_SHADER_FRAGMENT
:
571 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
573 case MESA_SHADER_VERTEX
:
574 if (!ctx
->is_gs_copy_shader
) {
575 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.vs
.has_vertex_buffers
? 2 : 0;
576 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
577 user_sgpr_info
->sgpr_count
+= 3;
579 user_sgpr_info
->sgpr_count
+= 2;
582 if (ctx
->options
->key
.vs
.as_ls
)
583 user_sgpr_info
->sgpr_count
++;
585 case MESA_SHADER_TESS_CTRL
:
586 user_sgpr_info
->sgpr_count
+= 4;
588 case MESA_SHADER_TESS_EVAL
:
589 user_sgpr_info
->sgpr_count
+= 1;
591 case MESA_SHADER_GEOMETRY
:
592 user_sgpr_info
->sgpr_count
+= 2;
598 if (ctx
->shader_info
->info
.needs_push_constants
)
599 user_sgpr_info
->sgpr_count
+= 2;
601 uint32_t remaining_sgprs
= 16 - 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
= const_array(ctx
->ac
.i8
, 1024 * 1024);
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
, const_array(type
, 32), desc_sets
);
639 if (ctx
->shader_info
->info
.needs_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
, const_array(ctx
->ac
.v4i32
, 16),
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 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->rel_auto_id
);
673 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->vs_prim_id
);
674 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.instance_id
);
679 declare_tes_input_vgprs(struct nir_to_llvm_context
*ctx
, struct arg_info
*args
)
681 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_u
);
682 add_arg(args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->tes_v
);
683 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->tes_rel_patch_id
);
684 add_arg(args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.tes_patch_id
);
688 set_global_input_locs(struct nir_to_llvm_context
*ctx
, gl_shader_stage stage
,
689 bool has_previous_stage
, gl_shader_stage previous_stage
,
690 const struct user_sgpr_info
*user_sgpr_info
,
691 LLVMValueRef desc_sets
, uint8_t *user_sgpr_idx
)
693 unsigned num_sets
= ctx
->options
->layout
?
694 ctx
->options
->layout
->num_sets
: 0;
695 unsigned stage_mask
= 1 << stage
;
697 if (has_previous_stage
)
698 stage_mask
|= 1 << previous_stage
;
700 if (!user_sgpr_info
->indirect_all_descriptor_sets
) {
701 for (unsigned i
= 0; i
< num_sets
; ++i
) {
702 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
703 set_loc_desc(ctx
, i
, user_sgpr_idx
, 0);
705 ctx
->descriptor_sets
[i
] = NULL
;
708 set_loc_shader(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
,
711 for (unsigned i
= 0; i
< num_sets
; ++i
) {
712 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& stage_mask
) {
713 set_loc_desc(ctx
, i
, user_sgpr_idx
, i
* 8);
714 ctx
->descriptor_sets
[i
] =
715 ac_build_load_to_sgpr(&ctx
->ac
,
717 LLVMConstInt(ctx
->ac
.i32
, i
, false));
720 ctx
->descriptor_sets
[i
] = NULL
;
722 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
725 if (ctx
->shader_info
->info
.needs_push_constants
) {
726 set_loc_shader(ctx
, AC_UD_PUSH_CONSTANTS
, user_sgpr_idx
, 2);
731 set_vs_specific_input_locs(struct nir_to_llvm_context
*ctx
,
732 gl_shader_stage stage
, bool has_previous_stage
,
733 gl_shader_stage previous_stage
,
734 uint8_t *user_sgpr_idx
)
736 if (!ctx
->is_gs_copy_shader
&&
737 (stage
== MESA_SHADER_VERTEX
||
738 (has_previous_stage
&& previous_stage
== MESA_SHADER_VERTEX
))) {
739 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
740 set_loc_shader(ctx
, AC_UD_VS_VERTEX_BUFFERS
,
745 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
748 set_loc_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
,
749 user_sgpr_idx
, vs_num
);
753 static void create_function(struct nir_to_llvm_context
*ctx
,
754 gl_shader_stage stage
,
755 bool has_previous_stage
,
756 gl_shader_stage previous_stage
)
758 uint8_t user_sgpr_idx
;
759 struct user_sgpr_info user_sgpr_info
;
760 struct arg_info args
= {};
761 LLVMValueRef desc_sets
;
763 allocate_user_sgprs(ctx
, &user_sgpr_info
);
765 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
766 add_arg(&args
, ARG_SGPR
, const_array(ctx
->ac
.v4i32
, 16),
771 case MESA_SHADER_COMPUTE
:
772 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
773 previous_stage
, &user_sgpr_info
,
776 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
777 add_arg(&args
, ARG_SGPR
, ctx
->ac
.v3i32
,
778 &ctx
->num_work_groups
);
781 for (int i
= 0; i
< 3; i
++) {
782 ctx
->workgroup_ids
[i
] = NULL
;
783 if (ctx
->shader_info
->info
.cs
.uses_block_id
[i
]) {
784 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
785 &ctx
->workgroup_ids
[i
]);
789 if (ctx
->shader_info
->info
.cs
.uses_local_invocation_idx
)
790 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->tg_size
);
791 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
,
792 &ctx
->local_invocation_ids
);
794 case MESA_SHADER_VERTEX
:
795 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
796 previous_stage
, &user_sgpr_info
,
798 declare_vs_specific_input_sgprs(ctx
, stage
, has_previous_stage
,
799 previous_stage
, &args
);
801 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.vs
.as_es
&& !ctx
->options
->key
.vs
.as_ls
&& ctx
->options
->key
.has_multiview_view_index
))
802 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->view_index
);
803 if (ctx
->options
->key
.vs
.as_es
)
804 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
806 else if (ctx
->options
->key
.vs
.as_ls
)
807 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
808 &ctx
->ls_out_layout
);
810 declare_vs_input_vgprs(ctx
, &args
);
812 case MESA_SHADER_TESS_CTRL
:
813 if (has_previous_stage
) {
814 // First 6 system regs
815 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
816 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
817 &ctx
->merged_wave_info
);
818 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
819 &ctx
->tess_factor_offset
);
821 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
822 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
823 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
825 declare_global_input_sgprs(ctx
, stage
,
828 &user_sgpr_info
, &args
,
830 declare_vs_specific_input_sgprs(ctx
, stage
,
832 previous_stage
, &args
);
834 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
835 &ctx
->ls_out_layout
);
837 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
838 &ctx
->tcs_offchip_layout
);
839 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
840 &ctx
->tcs_out_offsets
);
841 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
842 &ctx
->tcs_out_layout
);
843 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
844 &ctx
->tcs_in_layout
);
845 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
846 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
849 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
850 &ctx
->abi
.tcs_patch_id
);
851 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
852 &ctx
->abi
.tcs_rel_ids
);
854 declare_vs_input_vgprs(ctx
, &args
);
856 declare_global_input_sgprs(ctx
, stage
,
859 &user_sgpr_info
, &args
,
862 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
863 &ctx
->tcs_offchip_layout
);
864 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
865 &ctx
->tcs_out_offsets
);
866 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
867 &ctx
->tcs_out_layout
);
868 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
869 &ctx
->tcs_in_layout
);
870 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
871 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
874 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
875 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
876 &ctx
->tess_factor_offset
);
877 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
878 &ctx
->abi
.tcs_patch_id
);
879 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
880 &ctx
->abi
.tcs_rel_ids
);
883 case MESA_SHADER_TESS_EVAL
:
884 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
885 previous_stage
, &user_sgpr_info
,
888 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->tcs_offchip_layout
);
889 if (ctx
->shader_info
->info
.needs_multiview_view_index
|| (!ctx
->options
->key
.tes
.as_es
&& ctx
->options
->key
.has_multiview_view_index
))
890 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->view_index
);
892 if (ctx
->options
->key
.tes
.as_es
) {
893 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
894 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
895 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
898 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
);
899 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
901 declare_tes_input_vgprs(ctx
, &args
);
903 case MESA_SHADER_GEOMETRY
:
904 if (has_previous_stage
) {
905 // First 6 system regs
906 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
908 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
909 &ctx
->merged_wave_info
);
910 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->oc_lds
);
912 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // scratch offset
913 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
914 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, NULL
); // unknown
916 declare_global_input_sgprs(ctx
, stage
,
919 &user_sgpr_info
, &args
,
922 if (previous_stage
== MESA_SHADER_TESS_EVAL
) {
923 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
924 &ctx
->tcs_offchip_layout
);
926 declare_vs_specific_input_sgprs(ctx
, stage
,
932 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
933 &ctx
->gsvs_ring_stride
);
934 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
935 &ctx
->gsvs_num_entries
);
936 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
937 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
940 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
941 &ctx
->gs_vtx_offset
[0]);
942 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
943 &ctx
->gs_vtx_offset
[2]);
944 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
945 &ctx
->abi
.gs_prim_id
);
946 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
947 &ctx
->abi
.gs_invocation_id
);
948 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
949 &ctx
->gs_vtx_offset
[4]);
951 if (previous_stage
== MESA_SHADER_VERTEX
) {
952 declare_vs_input_vgprs(ctx
, &args
);
954 declare_tes_input_vgprs(ctx
, &args
);
957 declare_global_input_sgprs(ctx
, stage
,
960 &user_sgpr_info
, &args
,
963 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
964 &ctx
->gsvs_ring_stride
);
965 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
966 &ctx
->gsvs_num_entries
);
967 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
968 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
971 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs2vs_offset
);
972 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->gs_wave_id
);
973 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
974 &ctx
->gs_vtx_offset
[0]);
975 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
976 &ctx
->gs_vtx_offset
[1]);
977 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
978 &ctx
->abi
.gs_prim_id
);
979 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
980 &ctx
->gs_vtx_offset
[2]);
981 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
982 &ctx
->gs_vtx_offset
[3]);
983 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
984 &ctx
->gs_vtx_offset
[4]);
985 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
986 &ctx
->gs_vtx_offset
[5]);
987 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
,
988 &ctx
->abi
.gs_invocation_id
);
991 case MESA_SHADER_FRAGMENT
:
992 declare_global_input_sgprs(ctx
, stage
, has_previous_stage
,
993 previous_stage
, &user_sgpr_info
,
996 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
997 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
,
998 &ctx
->sample_pos_offset
);
1000 add_arg(&args
, ARG_SGPR
, ctx
->ac
.i32
, &ctx
->prim_mask
);
1001 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_sample
);
1002 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_center
);
1003 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->persp_centroid
);
1004 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v3i32
, NULL
); /* persp pull model */
1005 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_sample
);
1006 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_center
);
1007 add_arg(&args
, ARG_VGPR
, ctx
->ac
.v2i32
, &ctx
->linear_centroid
);
1008 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, NULL
); /* line stipple tex */
1009 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[0]);
1010 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[1]);
1011 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[2]);
1012 add_arg(&args
, ARG_VGPR
, ctx
->ac
.f32
, &ctx
->abi
.frag_pos
[3]);
1013 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.front_face
);
1014 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.ancillary
);
1015 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, &ctx
->abi
.sample_coverage
);
1016 add_arg(&args
, ARG_VGPR
, ctx
->ac
.i32
, NULL
); /* fixed pt */
1019 unreachable("Shader stage not implemented");
1022 ctx
->main_function
= create_llvm_function(
1023 ctx
->context
, ctx
->module
, ctx
->builder
, NULL
, 0, &args
,
1024 ctx
->max_workgroup_size
,
1025 ctx
->options
->unsafe_math
);
1026 set_llvm_calling_convention(ctx
->main_function
, stage
);
1029 ctx
->shader_info
->num_input_vgprs
= 0;
1030 ctx
->shader_info
->num_input_sgprs
= ctx
->options
->supports_spill
? 2 : 0;
1032 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
1034 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
1035 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
1037 assign_arguments(ctx
->main_function
, &args
);
1041 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
1042 set_loc_shader(ctx
, AC_UD_SCRATCH_RING_OFFSETS
,
1044 if (ctx
->options
->supports_spill
) {
1045 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
1046 LLVMPointerType(ctx
->ac
.i8
, CONST_ADDR_SPACE
),
1047 NULL
, 0, AC_FUNC_ATTR_READNONE
);
1048 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->builder
, ctx
->ring_offsets
,
1049 const_array(ctx
->ac
.v4i32
, 16), "");
1053 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1054 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1055 if (has_previous_stage
)
1058 set_global_input_locs(ctx
, stage
, has_previous_stage
, previous_stage
,
1059 &user_sgpr_info
, desc_sets
, &user_sgpr_idx
);
1062 case MESA_SHADER_COMPUTE
:
1063 if (ctx
->shader_info
->info
.cs
.uses_grid_size
) {
1064 set_loc_shader(ctx
, AC_UD_CS_GRID_SIZE
,
1068 case MESA_SHADER_VERTEX
:
1069 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1070 previous_stage
, &user_sgpr_idx
);
1071 if (ctx
->view_index
)
1072 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1073 if (ctx
->options
->key
.vs
.as_ls
) {
1074 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1077 if (ctx
->options
->key
.vs
.as_ls
)
1078 ac_declare_lds_as_pointer(&ctx
->ac
);
1080 case MESA_SHADER_TESS_CTRL
:
1081 set_vs_specific_input_locs(ctx
, stage
, has_previous_stage
,
1082 previous_stage
, &user_sgpr_idx
);
1083 if (has_previous_stage
)
1084 set_loc_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
,
1086 set_loc_shader(ctx
, AC_UD_TCS_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 4);
1087 if (ctx
->view_index
)
1088 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1089 ac_declare_lds_as_pointer(&ctx
->ac
);
1091 case MESA_SHADER_TESS_EVAL
:
1092 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 1);
1093 if (ctx
->view_index
)
1094 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1096 case MESA_SHADER_GEOMETRY
:
1097 if (has_previous_stage
) {
1098 if (previous_stage
== MESA_SHADER_VERTEX
)
1099 set_vs_specific_input_locs(ctx
, stage
,
1104 set_loc_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
,
1107 set_loc_shader(ctx
, AC_UD_GS_VS_RING_STRIDE_ENTRIES
,
1109 if (ctx
->view_index
)
1110 set_loc_shader(ctx
, AC_UD_VIEW_INDEX
, &user_sgpr_idx
, 1);
1111 if (has_previous_stage
)
1112 ac_declare_lds_as_pointer(&ctx
->ac
);
1114 case MESA_SHADER_FRAGMENT
:
1115 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
1116 set_loc_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
,
1121 unreachable("Shader stage not implemented");
1124 ctx
->shader_info
->num_user_sgprs
= user_sgpr_idx
;
1127 static LLVMValueRef
trim_vector(struct ac_llvm_context
*ctx
,
1128 LLVMValueRef value
, unsigned count
)
1130 unsigned num_components
= ac_get_llvm_num_components(value
);
1131 if (count
== num_components
)
1134 LLVMValueRef masks
[] = {
1135 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
1136 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false)};
1139 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
1142 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
1143 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
1147 build_store_values_extended(struct ac_llvm_context
*ac
,
1148 LLVMValueRef
*values
,
1149 unsigned value_count
,
1150 unsigned value_stride
,
1153 LLVMBuilderRef builder
= ac
->builder
;
1156 for (i
= 0; i
< value_count
; i
++) {
1157 LLVMValueRef ptr
= values
[i
* value_stride
];
1158 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
1159 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
1160 LLVMBuildStore(builder
, value
, ptr
);
1164 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
1165 const nir_ssa_def
*def
)
1167 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
1168 if (def
->num_components
> 1) {
1169 type
= LLVMVectorType(type
, def
->num_components
);
1174 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
1177 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, src
.ssa
);
1178 return (LLVMValueRef
)entry
->data
;
1182 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
1183 const struct nir_block
*b
)
1185 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
1186 return (LLVMBasicBlockRef
)entry
->data
;
1189 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
1191 unsigned num_components
)
1193 LLVMValueRef value
= get_src(ctx
, src
.src
);
1194 bool need_swizzle
= false;
1197 LLVMTypeRef type
= LLVMTypeOf(value
);
1198 unsigned src_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
1199 ? LLVMGetVectorSize(type
)
1202 for (unsigned i
= 0; i
< num_components
; ++i
) {
1203 assert(src
.swizzle
[i
] < src_components
);
1204 if (src
.swizzle
[i
] != i
)
1205 need_swizzle
= true;
1208 if (need_swizzle
|| num_components
!= src_components
) {
1209 LLVMValueRef masks
[] = {
1210 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
1211 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
1212 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
1213 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
1215 if (src_components
> 1 && num_components
== 1) {
1216 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1218 } else if (src_components
== 1 && num_components
> 1) {
1219 LLVMValueRef values
[] = {value
, value
, value
, value
};
1220 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
1222 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
1223 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
1227 assert(!src
.negate
);
1232 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
1233 LLVMIntPredicate pred
, LLVMValueRef src0
,
1236 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
1237 return LLVMBuildSelect(ctx
->builder
, result
,
1238 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1242 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
1243 LLVMRealPredicate pred
, LLVMValueRef src0
,
1246 LLVMValueRef result
;
1247 src0
= ac_to_float(ctx
, src0
);
1248 src1
= ac_to_float(ctx
, src1
);
1249 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
1250 return LLVMBuildSelect(ctx
->builder
, result
,
1251 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1255 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
1257 LLVMTypeRef result_type
,
1261 LLVMValueRef params
[] = {
1262 ac_to_float(ctx
, src0
),
1265 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1266 get_elem_bits(ctx
, result_type
));
1267 assert(length
< sizeof(name
));
1268 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
1271 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
1273 LLVMTypeRef result_type
,
1274 LLVMValueRef src0
, LLVMValueRef src1
)
1277 LLVMValueRef params
[] = {
1278 ac_to_float(ctx
, src0
),
1279 ac_to_float(ctx
, src1
),
1282 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1283 get_elem_bits(ctx
, result_type
));
1284 assert(length
< sizeof(name
));
1285 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
1288 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
1290 LLVMTypeRef result_type
,
1291 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1294 LLVMValueRef params
[] = {
1295 ac_to_float(ctx
, src0
),
1296 ac_to_float(ctx
, src1
),
1297 ac_to_float(ctx
, src2
),
1300 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1301 get_elem_bits(ctx
, result_type
));
1302 assert(length
< sizeof(name
));
1303 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
1306 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
1307 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1309 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
1311 return LLVMBuildSelect(ctx
->builder
, v
, src1
, src2
, "");
1314 static LLVMValueRef
emit_minmax_int(struct ac_llvm_context
*ctx
,
1315 LLVMIntPredicate pred
,
1316 LLVMValueRef src0
, LLVMValueRef src1
)
1318 return LLVMBuildSelect(ctx
->builder
,
1319 LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, ""),
1324 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
1327 return emit_minmax_int(ctx
, LLVMIntSGT
, src0
,
1328 LLVMBuildNeg(ctx
->builder
, src0
, ""));
1331 static LLVMValueRef
emit_fsign(struct ac_llvm_context
*ctx
,
1334 LLVMValueRef cmp
, val
;
1336 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, ctx
->f32_0
, "");
1337 val
= LLVMBuildSelect(ctx
->builder
, cmp
, ctx
->f32_1
, src0
, "");
1338 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, ctx
->f32_0
, "");
1339 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(ctx
->f32
, -1.0), "");
1343 static LLVMValueRef
emit_isign(struct ac_llvm_context
*ctx
,
1346 LLVMValueRef cmp
, val
;
1348 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, ctx
->i32_0
, "");
1349 val
= LLVMBuildSelect(ctx
->builder
, cmp
, ctx
->i32_1
, src0
, "");
1350 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, ctx
->i32_0
, "");
1351 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(ctx
->i32
, -1, true), "");
1355 static LLVMValueRef
emit_ffract(struct ac_llvm_context
*ctx
,
1358 const char *intr
= "llvm.floor.f32";
1359 LLVMValueRef fsrc0
= ac_to_float(ctx
, src0
);
1360 LLVMValueRef params
[] = {
1363 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
,
1364 ctx
->f32
, params
, 1,
1365 AC_FUNC_ATTR_READNONE
);
1366 return LLVMBuildFSub(ctx
->builder
, fsrc0
, floor
, "");
1369 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
1371 LLVMValueRef src0
, LLVMValueRef src1
)
1373 LLVMTypeRef ret_type
;
1374 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
1376 LLVMValueRef params
[] = { src0
, src1
};
1377 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
1380 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
1381 params
, 2, AC_FUNC_ATTR_READNONE
);
1383 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
1384 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
1388 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
1391 return LLVMBuildAnd(ctx
->builder
, src0
, LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""), "");
1394 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
1397 src0
= ac_to_float(ctx
, src0
);
1398 return LLVMBuildSExt(ctx
->builder
,
1399 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, ctx
->f32_0
, ""),
1403 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
1406 return LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
1409 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
1412 return LLVMBuildSExt(ctx
->builder
,
1413 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, ctx
->i32_0
, ""),
1417 static LLVMValueRef
emit_f2f16(struct nir_to_llvm_context
*ctx
,
1420 LLVMValueRef result
;
1421 LLVMValueRef cond
= NULL
;
1423 src0
= ac_to_float(&ctx
->ac
, src0
);
1424 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->ac
.f16
, "");
1426 if (ctx
->options
->chip_class
>= VI
) {
1427 LLVMValueRef args
[2];
1428 /* Check if the result is a denormal - and flush to 0 if so. */
1430 args
[1] = LLVMConstInt(ctx
->ac
.i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
1431 cond
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.class.f16", ctx
->ac
.i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
1434 /* need to convert back up to f32 */
1435 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->ac
.f32
, "");
1437 if (ctx
->options
->chip_class
>= VI
)
1438 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->ac
.f32_0
, result
, "");
1441 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
1442 * so compare the result and flush to 0 if it's smaller.
1444 LLVMValueRef temp
, cond2
;
1445 temp
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1446 ctx
->ac
.f32
, result
);
1447 cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUGT
,
1448 LLVMBuildBitCast(ctx
->builder
, LLVMConstInt(ctx
->ac
.i32
, 0x38800000, false), ctx
->ac
.f32
, ""),
1450 cond2
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
1451 temp
, ctx
->ac
.f32_0
, "");
1452 cond
= LLVMBuildAnd(ctx
->builder
, cond
, cond2
, "");
1453 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->ac
.f32_0
, result
, "");
1458 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
1459 LLVMValueRef src0
, LLVMValueRef src1
)
1461 LLVMValueRef dst64
, result
;
1462 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
1463 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
1465 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1466 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1467 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1471 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
1472 LLVMValueRef src0
, LLVMValueRef src1
)
1474 LLVMValueRef dst64
, result
;
1475 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
1476 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
1478 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1479 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1480 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1484 static LLVMValueRef
emit_bitfield_extract(struct ac_llvm_context
*ctx
,
1486 const LLVMValueRef srcs
[3])
1488 LLVMValueRef result
;
1489 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
1491 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
1492 result
= LLVMBuildSelect(ctx
->builder
, icond
, srcs
[0], result
, "");
1496 static LLVMValueRef
emit_bitfield_insert(struct ac_llvm_context
*ctx
,
1497 LLVMValueRef src0
, LLVMValueRef src1
,
1498 LLVMValueRef src2
, LLVMValueRef src3
)
1500 LLVMValueRef bfi_args
[3], result
;
1502 bfi_args
[0] = LLVMBuildShl(ctx
->builder
,
1503 LLVMBuildSub(ctx
->builder
,
1504 LLVMBuildShl(ctx
->builder
,
1509 bfi_args
[1] = LLVMBuildShl(ctx
->builder
, src1
, src2
, "");
1512 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, src3
, LLVMConstInt(ctx
->i32
, 32, false), "");
1515 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
1516 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
1518 result
= LLVMBuildXor(ctx
->builder
, bfi_args
[2],
1519 LLVMBuildAnd(ctx
->builder
, bfi_args
[0],
1520 LLVMBuildXor(ctx
->builder
, bfi_args
[1], bfi_args
[2], ""), ""), "");
1522 result
= LLVMBuildSelect(ctx
->builder
, icond
, src1
, result
, "");
1526 static LLVMValueRef
emit_pack_half_2x16(struct ac_llvm_context
*ctx
,
1529 LLVMValueRef comp
[2];
1531 src0
= ac_to_float(ctx
, src0
);
1532 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
1533 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
1535 return ac_build_cvt_pkrtz_f16(ctx
, comp
);
1538 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
1541 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
1542 LLVMValueRef temps
[2], result
, val
;
1545 for (i
= 0; i
< 2; i
++) {
1546 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
1547 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
1548 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
1549 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
1552 result
= LLVMBuildInsertElement(ctx
->builder
, LLVMGetUndef(ctx
->v2f32
), temps
[0],
1554 result
= LLVMBuildInsertElement(ctx
->builder
, result
, temps
[1],
1559 static LLVMValueRef
emit_ddxy(struct ac_nir_context
*ctx
,
1565 LLVMValueRef result
;
1567 if (op
== nir_op_fddx_fine
|| op
== nir_op_fddx
)
1568 mask
= AC_TID_MASK_LEFT
;
1569 else if (op
== nir_op_fddy_fine
|| op
== nir_op_fddy
)
1570 mask
= AC_TID_MASK_TOP
;
1572 mask
= AC_TID_MASK_TOP_LEFT
;
1574 /* for DDX we want to next X pixel, DDY next Y pixel. */
1575 if (op
== nir_op_fddx_fine
||
1576 op
== nir_op_fddx_coarse
||
1582 result
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, src0
);
1587 * this takes an I,J coordinate pair,
1588 * and works out the X and Y derivatives.
1589 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
1591 static LLVMValueRef
emit_ddxy_interp(
1592 struct ac_nir_context
*ctx
,
1593 LLVMValueRef interp_ij
)
1595 LLVMValueRef result
[4], a
;
1598 for (i
= 0; i
< 2; i
++) {
1599 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
1600 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1601 result
[i
] = emit_ddxy(ctx
, nir_op_fddx
, a
);
1602 result
[2+i
] = emit_ddxy(ctx
, nir_op_fddy
, a
);
1604 return ac_build_gather_values(&ctx
->ac
, result
, 4);
1607 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
1609 LLVMValueRef src
[4], result
= NULL
;
1610 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
1611 unsigned src_components
;
1612 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
1614 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
1615 switch (instr
->op
) {
1621 case nir_op_pack_half_2x16
:
1624 case nir_op_unpack_half_2x16
:
1628 src_components
= num_components
;
1631 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1632 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
1634 switch (instr
->op
) {
1640 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1641 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
1644 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
1647 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
1650 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1653 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1654 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1655 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1658 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1659 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1660 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1663 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1666 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1669 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1672 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
1675 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1676 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1677 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1678 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1679 ac_to_float_type(&ctx
->ac
, def_type
), result
);
1680 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
1681 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
1684 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1685 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1686 result
= LLVMBuildFRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1689 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1692 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1695 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1698 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1699 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1700 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1703 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1704 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1705 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1708 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1709 result
= ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, src
[0]);
1712 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
1715 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
1718 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
1721 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0],
1722 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1723 LLVMTypeOf(src
[0]), ""),
1727 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0],
1728 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1729 LLVMTypeOf(src
[0]), ""),
1733 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0],
1734 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1735 LLVMTypeOf(src
[0]), ""),
1739 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1742 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
1745 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
1748 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
1751 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1754 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
1757 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUEQ
, src
[0], src
[1]);
1760 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
1763 result
= emit_float_cmp(&ctx
->ac
, LLVMRealULT
, src
[0], src
[1]);
1766 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUGE
, src
[0], src
[1]);
1769 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1770 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1773 result
= emit_iabs(&ctx
->ac
, src
[0]);
1776 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, src
[0], src
[1]);
1779 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1782 result
= emit_minmax_int(&ctx
->ac
, LLVMIntUGT
, src
[0], src
[1]);
1785 result
= emit_minmax_int(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1788 result
= emit_isign(&ctx
->ac
, src
[0]);
1791 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1792 result
= emit_fsign(&ctx
->ac
, src
[0]);
1795 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1796 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1799 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
1800 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1803 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
1804 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1806 case nir_op_fround_even
:
1807 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
1808 ac_to_float_type(&ctx
->ac
, def_type
),src
[0]);
1811 result
= emit_ffract(&ctx
->ac
, src
[0]);
1814 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
1815 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1818 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
1819 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1822 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1823 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1826 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
1827 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1830 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
1831 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1834 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1835 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
1836 result
= ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, result
);
1839 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
1840 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1843 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1844 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1845 if (instr
->dest
.dest
.ssa
.bit_size
== 32)
1846 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1847 ac_to_float_type(&ctx
->ac
, def_type
),
1851 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1852 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1853 if (instr
->dest
.dest
.ssa
.bit_size
== 32)
1854 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1855 ac_to_float_type(&ctx
->ac
, def_type
),
1859 result
= emit_intrin_3f_param(&ctx
->ac
, "llvm.fmuladd",
1860 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
1862 case nir_op_ibitfield_extract
:
1863 result
= emit_bitfield_extract(&ctx
->ac
, true, src
);
1865 case nir_op_ubitfield_extract
:
1866 result
= emit_bitfield_extract(&ctx
->ac
, false, src
);
1868 case nir_op_bitfield_insert
:
1869 result
= emit_bitfield_insert(&ctx
->ac
, src
[0], src
[1], src
[2], src
[3]);
1871 case nir_op_bitfield_reverse
:
1872 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.bitreverse.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1874 case nir_op_bit_count
:
1875 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1880 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1881 src
[i
] = ac_to_integer(&ctx
->ac
, src
[i
]);
1882 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
1886 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1887 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
1891 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1892 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
1896 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1897 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1901 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1902 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1905 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1906 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1909 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1913 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1914 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < get_elem_bits(&ctx
->ac
, def_type
))
1915 result
= LLVMBuildZExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1917 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1921 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1922 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < get_elem_bits(&ctx
->ac
, def_type
))
1923 result
= LLVMBuildSExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1925 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1928 result
= emit_bcsel(&ctx
->ac
, src
[0], src
[1], src
[2]);
1930 case nir_op_find_lsb
:
1931 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1932 result
= ac_find_lsb(&ctx
->ac
, ctx
->ac
.i32
, src
[0]);
1934 case nir_op_ufind_msb
:
1935 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1936 result
= ac_build_umsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1938 case nir_op_ifind_msb
:
1939 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1940 result
= ac_build_imsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1942 case nir_op_uadd_carry
:
1943 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1944 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1945 result
= emit_uint_carry(&ctx
->ac
, "llvm.uadd.with.overflow.i32", src
[0], src
[1]);
1947 case nir_op_usub_borrow
:
1948 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1949 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1950 result
= emit_uint_carry(&ctx
->ac
, "llvm.usub.with.overflow.i32", src
[0], src
[1]);
1953 result
= emit_b2f(&ctx
->ac
, src
[0]);
1956 result
= emit_f2b(&ctx
->ac
, src
[0]);
1959 result
= emit_b2i(&ctx
->ac
, src
[0]);
1962 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1963 result
= emit_i2b(&ctx
->ac
, src
[0]);
1965 case nir_op_fquantize2f16
:
1966 result
= emit_f2f16(ctx
->nctx
, src
[0]);
1968 case nir_op_umul_high
:
1969 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1970 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1971 result
= emit_umul_high(&ctx
->ac
, src
[0], src
[1]);
1973 case nir_op_imul_high
:
1974 src
[0] = ac_to_integer(&ctx
->ac
, src
[0]);
1975 src
[1] = ac_to_integer(&ctx
->ac
, src
[1]);
1976 result
= emit_imul_high(&ctx
->ac
, src
[0], src
[1]);
1978 case nir_op_pack_half_2x16
:
1979 result
= emit_pack_half_2x16(&ctx
->ac
, src
[0]);
1981 case nir_op_unpack_half_2x16
:
1982 result
= emit_unpack_half_2x16(&ctx
->ac
, src
[0]);
1986 case nir_op_fddx_fine
:
1987 case nir_op_fddy_fine
:
1988 case nir_op_fddx_coarse
:
1989 case nir_op_fddy_coarse
:
1990 result
= emit_ddxy(ctx
, instr
->op
, src
[0]);
1993 case nir_op_unpack_64_2x32_split_x
: {
1994 assert(instr
->src
[0].src
.ssa
->num_components
== 1);
1995 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1998 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
2003 case nir_op_unpack_64_2x32_split_y
: {
2004 assert(instr
->src
[0].src
.ssa
->num_components
== 1);
2005 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
2008 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
2013 case nir_op_pack_64_2x32_split
: {
2014 LLVMValueRef tmp
= LLVMGetUndef(ctx
->ac
.v2i32
);
2015 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
2016 src
[0], ctx
->ac
.i32_0
, "");
2017 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
2018 src
[1], ctx
->ac
.i32_1
, "");
2019 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i64
, "");
2024 fprintf(stderr
, "Unknown NIR alu instr: ");
2025 nir_print_instr(&instr
->instr
, stderr
);
2026 fprintf(stderr
, "\n");
2031 assert(instr
->dest
.dest
.is_ssa
);
2032 result
= ac_to_integer(&ctx
->ac
, result
);
2033 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.dest
.ssa
,
2038 static void visit_load_const(struct ac_nir_context
*ctx
,
2039 const nir_load_const_instr
*instr
)
2041 LLVMValueRef values
[4], value
= NULL
;
2042 LLVMTypeRef element_type
=
2043 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
2045 for (unsigned i
= 0; i
< instr
->def
.num_components
; ++i
) {
2046 switch (instr
->def
.bit_size
) {
2048 values
[i
] = LLVMConstInt(element_type
,
2049 instr
->value
.u32
[i
], false);
2052 values
[i
] = LLVMConstInt(element_type
,
2053 instr
->value
.u64
[i
], false);
2057 "unsupported nir load_const bit_size: %d\n",
2058 instr
->def
.bit_size
);
2062 if (instr
->def
.num_components
> 1) {
2063 value
= LLVMConstVector(values
, instr
->def
.num_components
);
2067 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, value
);
2070 static LLVMValueRef
cast_ptr(struct nir_to_llvm_context
*ctx
, LLVMValueRef ptr
,
2073 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
2074 return LLVMBuildBitCast(ctx
->builder
, ptr
,
2075 LLVMPointerType(type
, addr_space
), "");
2079 get_buffer_size(struct ac_nir_context
*ctx
, LLVMValueRef descriptor
, bool in_elements
)
2082 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
2083 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
2086 if (ctx
->ac
.chip_class
== VI
&& in_elements
) {
2087 /* On VI, the descriptor contains the size in bytes,
2088 * but TXQ must return the size in elements.
2089 * The stride is always non-zero for resources using TXQ.
2091 LLVMValueRef stride
=
2092 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
2094 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
,
2095 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
2096 stride
= LLVMBuildAnd(ctx
->ac
.builder
, stride
,
2097 LLVMConstInt(ctx
->ac
.i32
, 0x3fff, false), "");
2099 size
= LLVMBuildUDiv(ctx
->ac
.builder
, size
, stride
, "");
2105 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
2108 static void build_int_type_name(
2110 char *buf
, unsigned bufsize
)
2112 assert(bufsize
>= 6);
2114 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
2115 snprintf(buf
, bufsize
, "v%ui32",
2116 LLVMGetVectorSize(type
));
2121 static LLVMValueRef
radv_lower_gather4_integer(struct ac_llvm_context
*ctx
,
2122 struct ac_image_args
*args
,
2123 const nir_tex_instr
*instr
)
2125 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2126 LLVMValueRef coord
= args
->addr
;
2127 LLVMValueRef half_texel
[2];
2128 LLVMValueRef compare_cube_wa
= NULL
;
2129 LLVMValueRef result
;
2131 unsigned coord_vgpr_index
= (unsigned)args
->offset
+ (unsigned)args
->compare
;
2135 struct ac_image_args txq_args
= { 0 };
2137 txq_args
.da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
2138 txq_args
.opcode
= ac_image_get_resinfo
;
2139 txq_args
.dmask
= 0xf;
2140 txq_args
.addr
= ctx
->i32_0
;
2141 txq_args
.resource
= args
->resource
;
2142 LLVMValueRef size
= ac_build_image_opcode(ctx
, &txq_args
);
2144 for (c
= 0; c
< 2; c
++) {
2145 half_texel
[c
] = LLVMBuildExtractElement(ctx
->builder
, size
,
2146 LLVMConstInt(ctx
->i32
, c
, false), "");
2147 half_texel
[c
] = LLVMBuildUIToFP(ctx
->builder
, half_texel
[c
], ctx
->f32
, "");
2148 half_texel
[c
] = ac_build_fdiv(ctx
, ctx
->f32_1
, half_texel
[c
]);
2149 half_texel
[c
] = LLVMBuildFMul(ctx
->builder
, half_texel
[c
],
2150 LLVMConstReal(ctx
->f32
, -0.5), "");
2154 LLVMValueRef orig_coords
= args
->addr
;
2156 for (c
= 0; c
< 2; c
++) {
2158 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, coord_vgpr_index
+ c
, 0);
2159 tmp
= LLVMBuildExtractElement(ctx
->builder
, coord
, index
, "");
2160 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2161 tmp
= LLVMBuildFAdd(ctx
->builder
, tmp
, half_texel
[c
], "");
2162 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2163 coord
= LLVMBuildInsertElement(ctx
->builder
, coord
, tmp
, index
, "");
2168 * Apparantly cube has issue with integer types that the workaround doesn't solve,
2169 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
2170 * workaround by sampling using a scaled type and converting.
2171 * This is taken from amdgpu-pro shaders.
2173 /* NOTE this produces some ugly code compared to amdgpu-pro,
2174 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
2175 * and then reads them back. -pro generates two selects,
2176 * one s_cmp for the descriptor rewriting
2177 * one v_cmp for the coordinate and result changes.
2179 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2180 LLVMValueRef tmp
, tmp2
;
2182 /* workaround 8/8/8/8 uint/sint cube gather bug */
2183 /* first detect it then change to a scaled read and f2i */
2184 tmp
= LLVMBuildExtractElement(ctx
->builder
, args
->resource
, ctx
->i32_1
, "");
2187 /* extract the DATA_FORMAT */
2188 tmp
= ac_build_bfe(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 20, false),
2189 LLVMConstInt(ctx
->i32
, 6, false), false);
2191 /* is the DATA_FORMAT == 8_8_8_8 */
2192 compare_cube_wa
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tmp
, LLVMConstInt(ctx
->i32
, V_008F14_IMG_DATA_FORMAT_8_8_8_8
, false), "");
2194 if (stype
== GLSL_TYPE_UINT
)
2195 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
2196 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0x8000000, false),
2197 LLVMConstInt(ctx
->i32
, 0x10000000, false), "");
2199 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
2200 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0xc000000, false),
2201 LLVMConstInt(ctx
->i32
, 0x14000000, false), "");
2203 /* replace the NUM FORMAT in the descriptor */
2204 tmp2
= LLVMBuildAnd(ctx
->builder
, tmp2
, LLVMConstInt(ctx
->i32
, C_008F14_NUM_FORMAT_GFX6
, false), "");
2205 tmp2
= LLVMBuildOr(ctx
->builder
, tmp2
, tmp
, "");
2207 args
->resource
= LLVMBuildInsertElement(ctx
->builder
, args
->resource
, tmp2
, ctx
->i32_1
, "");
2209 /* don't modify the coordinates for this case */
2210 coord
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, orig_coords
, coord
, "");
2213 result
= ac_build_image_opcode(ctx
, args
);
2215 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2216 LLVMValueRef tmp
, tmp2
;
2218 /* if the cube workaround is in place, f2i the result. */
2219 for (c
= 0; c
< 4; c
++) {
2220 tmp
= LLVMBuildExtractElement(ctx
->builder
, result
, LLVMConstInt(ctx
->i32
, c
, false), "");
2221 if (stype
== GLSL_TYPE_UINT
)
2222 tmp2
= LLVMBuildFPToUI(ctx
->builder
, tmp
, ctx
->i32
, "");
2224 tmp2
= LLVMBuildFPToSI(ctx
->builder
, tmp
, ctx
->i32
, "");
2225 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2226 tmp2
= LLVMBuildBitCast(ctx
->builder
, tmp2
, ctx
->i32
, "");
2227 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, tmp2
, tmp
, "");
2228 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2229 result
= LLVMBuildInsertElement(ctx
->builder
, result
, tmp
, LLVMConstInt(ctx
->i32
, c
, false), "");
2235 static LLVMValueRef
build_tex_intrinsic(struct ac_nir_context
*ctx
,
2236 const nir_tex_instr
*instr
,
2238 struct ac_image_args
*args
)
2240 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
2241 return ac_build_buffer_load_format(&ctx
->ac
,
2248 args
->opcode
= ac_image_sample
;
2249 args
->compare
= instr
->is_shadow
;
2251 switch (instr
->op
) {
2253 case nir_texop_txf_ms
:
2254 case nir_texop_samples_identical
:
2255 args
->opcode
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
? ac_image_load
: ac_image_load_mip
;
2256 args
->compare
= false;
2257 args
->offset
= false;
2264 args
->level_zero
= true;
2269 case nir_texop_query_levels
:
2270 args
->opcode
= ac_image_get_resinfo
;
2273 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
2274 args
->level_zero
= true;
2280 args
->opcode
= ac_image_gather4
;
2281 args
->level_zero
= true;
2284 args
->opcode
= ac_image_get_lod
;
2285 args
->compare
= false;
2286 args
->offset
= false;
2292 if (instr
->op
== nir_texop_tg4
&& ctx
->ac
.chip_class
<= VI
) {
2293 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2294 if (stype
== GLSL_TYPE_UINT
|| stype
== GLSL_TYPE_INT
) {
2295 return radv_lower_gather4_integer(&ctx
->ac
, args
, instr
);
2298 return ac_build_image_opcode(&ctx
->ac
, args
);
2301 static LLVMValueRef
visit_vulkan_resource_index(struct nir_to_llvm_context
*ctx
,
2302 nir_intrinsic_instr
*instr
)
2304 LLVMValueRef index
= get_src(ctx
->nir
, instr
->src
[0]);
2305 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
2306 unsigned binding
= nir_intrinsic_binding(instr
);
2307 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
2308 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
2309 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
2310 unsigned base_offset
= layout
->binding
[binding
].offset
;
2311 LLVMValueRef offset
, stride
;
2313 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
2314 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
2315 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
2316 layout
->binding
[binding
].dynamic_offset_offset
;
2317 desc_ptr
= ctx
->push_constants
;
2318 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
2319 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2321 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
2323 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
2324 index
= LLVMBuildMul(ctx
->builder
, index
, stride
, "");
2325 offset
= LLVMBuildAdd(ctx
->builder
, offset
, index
, "");
2327 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
2328 desc_ptr
= cast_ptr(ctx
, desc_ptr
, ctx
->ac
.v4i32
);
2329 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2334 static LLVMValueRef
visit_vulkan_resource_reindex(struct nir_to_llvm_context
*ctx
,
2335 nir_intrinsic_instr
*instr
)
2337 LLVMValueRef ptr
= get_src(ctx
->nir
, instr
->src
[0]);
2338 LLVMValueRef index
= get_src(ctx
->nir
, instr
->src
[1]);
2340 LLVMValueRef result
= LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, "");
2341 LLVMSetMetadata(result
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
2345 static LLVMValueRef
visit_load_push_constant(struct nir_to_llvm_context
*ctx
,
2346 nir_intrinsic_instr
*instr
)
2348 LLVMValueRef ptr
, addr
;
2350 addr
= LLVMConstInt(ctx
->ac
.i32
, nir_intrinsic_base(instr
), 0);
2351 addr
= LLVMBuildAdd(ctx
->builder
, addr
, get_src(ctx
->nir
, instr
->src
[0]), "");
2353 ptr
= ac_build_gep0(&ctx
->ac
, ctx
->push_constants
, addr
);
2354 ptr
= cast_ptr(ctx
, ptr
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
));
2356 return LLVMBuildLoad(ctx
->builder
, ptr
, "");
2359 static LLVMValueRef
visit_get_buffer_size(struct ac_nir_context
*ctx
,
2360 const nir_intrinsic_instr
*instr
)
2362 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
2364 return get_buffer_size(ctx
, LLVMBuildLoad(ctx
->ac
.builder
, ptr
, ""), false);
2366 static void visit_store_ssbo(struct ac_nir_context
*ctx
,
2367 nir_intrinsic_instr
*instr
)
2369 const char *store_name
;
2370 LLVMValueRef src_data
= get_src(ctx
, instr
->src
[0]);
2371 LLVMTypeRef data_type
= ctx
->ac
.f32
;
2372 int elem_size_mult
= get_elem_bits(&ctx
->ac
, LLVMTypeOf(src_data
)) / 32;
2373 int components_32bit
= elem_size_mult
* instr
->num_components
;
2374 unsigned writemask
= nir_intrinsic_write_mask(instr
);
2375 LLVMValueRef base_data
, base_offset
;
2376 LLVMValueRef params
[6];
2378 params
[1] = ctx
->abi
->load_ssbo(ctx
->abi
,
2379 get_src(ctx
, instr
->src
[1]), true);
2380 params
[2] = ctx
->ac
.i32_0
; /* vindex */
2381 params
[4] = ctx
->ac
.i1false
; /* glc */
2382 params
[5] = ctx
->ac
.i1false
; /* slc */
2384 if (components_32bit
> 1)
2385 data_type
= LLVMVectorType(ctx
->ac
.f32
, components_32bit
);
2387 base_data
= ac_to_float(&ctx
->ac
, src_data
);
2388 base_data
= trim_vector(&ctx
->ac
, base_data
, instr
->num_components
);
2389 base_data
= LLVMBuildBitCast(ctx
->ac
.builder
, base_data
,
2391 base_offset
= get_src(ctx
, instr
->src
[2]); /* voffset */
2395 LLVMValueRef offset
;
2397 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
2399 /* Due to an LLVM limitation, split 3-element writes
2400 * into a 2-element and a 1-element write. */
2402 writemask
|= 1 << (start
+ 2);
2406 start
*= elem_size_mult
;
2407 count
*= elem_size_mult
;
2410 writemask
|= ((1u << (count
- 4)) - 1u) << (start
+ 4);
2415 store_name
= "llvm.amdgcn.buffer.store.v4f32";
2417 } else if (count
== 2) {
2418 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
2419 base_data
, LLVMConstInt(ctx
->ac
.i32
, start
, false), "");
2420 data
= LLVMBuildInsertElement(ctx
->ac
.builder
, LLVMGetUndef(ctx
->ac
.v2f32
), tmp
,
2423 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
2424 base_data
, LLVMConstInt(ctx
->ac
.i32
, start
+ 1, false), "");
2425 data
= LLVMBuildInsertElement(ctx
->ac
.builder
, data
, tmp
,
2427 store_name
= "llvm.amdgcn.buffer.store.v2f32";
2431 if (ac_get_llvm_num_components(base_data
) > 1)
2432 data
= LLVMBuildExtractElement(ctx
->ac
.builder
, base_data
,
2433 LLVMConstInt(ctx
->ac
.i32
, start
, false), "");
2436 store_name
= "llvm.amdgcn.buffer.store.f32";
2439 offset
= base_offset
;
2441 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, LLVMConstInt(ctx
->ac
.i32
, start
* 4, false), "");
2445 ac_build_intrinsic(&ctx
->ac
, store_name
,
2446 ctx
->ac
.voidt
, params
, 6, 0);
2450 static LLVMValueRef
visit_atomic_ssbo(struct ac_nir_context
*ctx
,
2451 const nir_intrinsic_instr
*instr
)
2454 LLVMValueRef params
[6];
2457 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
) {
2458 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[3]), 0);
2460 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
2461 params
[arg_count
++] = ctx
->abi
->load_ssbo(ctx
->abi
,
2462 get_src(ctx
, instr
->src
[0]),
2464 params
[arg_count
++] = ctx
->ac
.i32_0
; /* vindex */
2465 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
2466 params
[arg_count
++] = LLVMConstInt(ctx
->ac
.i1
, 0, false); /* slc */
2468 switch (instr
->intrinsic
) {
2469 case nir_intrinsic_ssbo_atomic_add
:
2470 name
= "llvm.amdgcn.buffer.atomic.add";
2472 case nir_intrinsic_ssbo_atomic_imin
:
2473 name
= "llvm.amdgcn.buffer.atomic.smin";
2475 case nir_intrinsic_ssbo_atomic_umin
:
2476 name
= "llvm.amdgcn.buffer.atomic.umin";
2478 case nir_intrinsic_ssbo_atomic_imax
:
2479 name
= "llvm.amdgcn.buffer.atomic.smax";
2481 case nir_intrinsic_ssbo_atomic_umax
:
2482 name
= "llvm.amdgcn.buffer.atomic.umax";
2484 case nir_intrinsic_ssbo_atomic_and
:
2485 name
= "llvm.amdgcn.buffer.atomic.and";
2487 case nir_intrinsic_ssbo_atomic_or
:
2488 name
= "llvm.amdgcn.buffer.atomic.or";
2490 case nir_intrinsic_ssbo_atomic_xor
:
2491 name
= "llvm.amdgcn.buffer.atomic.xor";
2493 case nir_intrinsic_ssbo_atomic_exchange
:
2494 name
= "llvm.amdgcn.buffer.atomic.swap";
2496 case nir_intrinsic_ssbo_atomic_comp_swap
:
2497 name
= "llvm.amdgcn.buffer.atomic.cmpswap";
2503 return ac_build_intrinsic(&ctx
->ac
, name
, ctx
->ac
.i32
, params
, arg_count
, 0);
2506 static LLVMValueRef
visit_load_buffer(struct ac_nir_context
*ctx
,
2507 const nir_intrinsic_instr
*instr
)
2509 LLVMValueRef results
[2];
2510 int load_components
;
2511 int num_components
= instr
->num_components
;
2512 if (instr
->dest
.ssa
.bit_size
== 64)
2513 num_components
*= 2;
2515 for (int i
= 0; i
< num_components
; i
+= load_components
) {
2516 load_components
= MIN2(num_components
- i
, 4);
2517 const char *load_name
;
2518 LLVMTypeRef data_type
= ctx
->ac
.f32
;
2519 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, i
* 4, false);
2520 offset
= LLVMBuildAdd(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]), offset
, "");
2522 if (load_components
== 3)
2523 data_type
= LLVMVectorType(ctx
->ac
.f32
, 4);
2524 else if (load_components
> 1)
2525 data_type
= LLVMVectorType(ctx
->ac
.f32
, load_components
);
2527 if (load_components
>= 3)
2528 load_name
= "llvm.amdgcn.buffer.load.v4f32";
2529 else if (load_components
== 2)
2530 load_name
= "llvm.amdgcn.buffer.load.v2f32";
2531 else if (load_components
== 1)
2532 load_name
= "llvm.amdgcn.buffer.load.f32";
2534 unreachable("unhandled number of components");
2536 LLVMValueRef params
[] = {
2537 ctx
->abi
->load_ssbo(ctx
->abi
,
2538 get_src(ctx
, instr
->src
[0]),
2546 results
[i
] = ac_build_intrinsic(&ctx
->ac
, load_name
, data_type
, params
, 5, 0);
2551 LLVMValueRef ret
= results
[0];
2552 if (num_components
> 4 || num_components
== 3) {
2553 LLVMValueRef masks
[] = {
2554 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
2555 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
2556 LLVMConstInt(ctx
->ac
.i32
, 4, false), LLVMConstInt(ctx
->ac
.i32
, 5, false),
2557 LLVMConstInt(ctx
->ac
.i32
, 6, false), LLVMConstInt(ctx
->ac
.i32
, 7, false)
2560 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
2561 ret
= LLVMBuildShuffleVector(ctx
->ac
.builder
, results
[0],
2562 results
[num_components
> 4 ? 1 : 0], swizzle
, "");
2565 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2566 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2569 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
2570 const nir_intrinsic_instr
*instr
)
2572 LLVMValueRef results
[8], ret
;
2573 LLVMValueRef rsrc
= get_src(ctx
, instr
->src
[0]);
2574 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
2575 int num_components
= instr
->num_components
;
2577 if (ctx
->abi
->load_ubo
)
2578 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
2580 if (instr
->dest
.ssa
.bit_size
== 64)
2581 num_components
*= 2;
2583 for (unsigned i
= 0; i
< num_components
; ++i
) {
2584 LLVMValueRef params
[] = {
2586 LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, 4 * i
, 0),
2589 results
[i
] = ac_build_intrinsic(&ctx
->ac
, "llvm.SI.load.const.v4i32", ctx
->ac
.f32
,
2591 AC_FUNC_ATTR_READNONE
|
2592 AC_FUNC_ATTR_LEGACY
);
2596 ret
= ac_build_gather_values(&ctx
->ac
, results
, num_components
);
2597 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2598 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2602 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_var
*deref
,
2603 bool vs_in
, unsigned *vertex_index_out
,
2604 LLVMValueRef
*vertex_index_ref
,
2605 unsigned *const_out
, LLVMValueRef
*indir_out
)
2607 unsigned const_offset
= 0;
2608 nir_deref
*tail
= &deref
->deref
;
2609 LLVMValueRef offset
= NULL
;
2611 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
2613 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2614 if (vertex_index_out
)
2615 *vertex_index_out
= deref_array
->base_offset
;
2617 if (vertex_index_ref
) {
2618 LLVMValueRef vtx
= LLVMConstInt(ctx
->ac
.i32
, deref_array
->base_offset
, false);
2619 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
2620 vtx
= LLVMBuildAdd(ctx
->ac
.builder
, vtx
, get_src(ctx
, deref_array
->indirect
), "");
2622 *vertex_index_ref
= vtx
;
2626 if (deref
->var
->data
.compact
) {
2627 assert(tail
->child
->deref_type
== nir_deref_type_array
);
2628 assert(glsl_type_is_scalar(glsl_without_array(deref
->var
->type
)));
2629 nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
2630 /* We always lower indirect dereferences for "compact" array vars. */
2631 assert(deref_array
->deref_array_type
== nir_deref_array_type_direct
);
2633 const_offset
= deref_array
->base_offset
;
2637 while (tail
->child
!= NULL
) {
2638 const struct glsl_type
*parent_type
= tail
->type
;
2641 if (tail
->deref_type
== nir_deref_type_array
) {
2642 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2643 LLVMValueRef index
, stride
, local_offset
;
2644 unsigned size
= glsl_count_attribute_slots(tail
->type
, vs_in
);
2646 const_offset
+= size
* deref_array
->base_offset
;
2647 if (deref_array
->deref_array_type
== nir_deref_array_type_direct
)
2650 assert(deref_array
->deref_array_type
== nir_deref_array_type_indirect
);
2651 index
= get_src(ctx
, deref_array
->indirect
);
2652 stride
= LLVMConstInt(ctx
->ac
.i32
, size
, 0);
2653 local_offset
= LLVMBuildMul(ctx
->ac
.builder
, stride
, index
, "");
2656 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, local_offset
, "");
2658 offset
= local_offset
;
2659 } else if (tail
->deref_type
== nir_deref_type_struct
) {
2660 nir_deref_struct
*deref_struct
= nir_deref_as_struct(tail
);
2662 for (unsigned i
= 0; i
< deref_struct
->index
; i
++) {
2663 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
2664 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
2667 unreachable("unsupported deref type");
2671 if (const_offset
&& offset
)
2672 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
2673 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
2676 *const_out
= const_offset
;
2677 *indir_out
= offset
;
2681 /* The offchip buffer layout for TCS->TES is
2683 * - attribute 0 of patch 0 vertex 0
2684 * - attribute 0 of patch 0 vertex 1
2685 * - attribute 0 of patch 0 vertex 2
2687 * - attribute 0 of patch 1 vertex 0
2688 * - attribute 0 of patch 1 vertex 1
2690 * - attribute 1 of patch 0 vertex 0
2691 * - attribute 1 of patch 0 vertex 1
2693 * - per patch attribute 0 of patch 0
2694 * - per patch attribute 0 of patch 1
2697 * Note that every attribute has 4 components.
2699 static LLVMValueRef
get_tcs_tes_buffer_address(struct nir_to_llvm_context
*ctx
,
2700 LLVMValueRef vertex_index
,
2701 LLVMValueRef param_index
)
2703 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
2704 LLVMValueRef param_stride
, constant16
;
2705 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
2707 vertices_per_patch
= unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 9, 6);
2708 num_patches
= unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 0, 9);
2709 total_vertices
= LLVMBuildMul(ctx
->builder
, vertices_per_patch
,
2712 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
2714 base_addr
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
2715 vertices_per_patch
, "");
2717 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2720 param_stride
= total_vertices
;
2722 base_addr
= rel_patch_id
;
2723 param_stride
= num_patches
;
2726 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2727 LLVMBuildMul(ctx
->builder
, param_index
,
2728 param_stride
, ""), "");
2730 base_addr
= LLVMBuildMul(ctx
->builder
, base_addr
, constant16
, "");
2732 if (!vertex_index
) {
2733 LLVMValueRef patch_data_offset
=
2734 unpack_param(&ctx
->ac
, ctx
->tcs_offchip_layout
, 16, 16);
2736 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2737 patch_data_offset
, "");
2742 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct nir_to_llvm_context
*ctx
,
2744 unsigned const_index
,
2746 LLVMValueRef vertex_index
,
2747 LLVMValueRef indir_index
)
2749 LLVMValueRef param_index
;
2752 param_index
= LLVMBuildAdd(ctx
->builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
2755 if (const_index
&& !is_compact
)
2756 param
+= const_index
;
2757 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
2759 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
2763 mark_tess_output(struct nir_to_llvm_context
*ctx
,
2764 bool is_patch
, uint32_t param
)
2768 ctx
->tess_patch_outputs_written
|= (1ull << param
);
2770 ctx
->tess_outputs_written
|= (1ull << param
);
2774 get_dw_address(struct nir_to_llvm_context
*ctx
,
2775 LLVMValueRef dw_addr
,
2777 unsigned const_index
,
2778 bool compact_const_index
,
2779 LLVMValueRef vertex_index
,
2780 LLVMValueRef stride
,
2781 LLVMValueRef indir_index
)
2786 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2787 LLVMBuildMul(ctx
->builder
,
2793 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2794 LLVMBuildMul(ctx
->builder
, indir_index
,
2795 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
2796 else if (const_index
&& !compact_const_index
)
2797 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2798 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
2800 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2801 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
2803 if (const_index
&& compact_const_index
)
2804 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2805 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
2810 load_tcs_input(struct ac_shader_abi
*abi
,
2811 LLVMValueRef vertex_index
,
2812 LLVMValueRef indir_index
,
2813 unsigned const_index
,
2815 unsigned driver_location
,
2817 unsigned num_components
,
2821 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2822 LLVMValueRef dw_addr
, stride
;
2823 LLVMValueRef value
[4], result
;
2824 unsigned param
= shader_io_get_unique_index(location
);
2826 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_in_layout
, 13, 8);
2827 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
2828 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2831 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
2832 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
2833 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2836 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
2841 load_tcs_output(struct nir_to_llvm_context
*ctx
,
2842 nir_intrinsic_instr
*instr
)
2844 LLVMValueRef dw_addr
;
2845 LLVMValueRef stride
= NULL
;
2846 LLVMValueRef value
[4], result
;
2847 LLVMValueRef vertex_index
= NULL
;
2848 LLVMValueRef indir_index
= NULL
;
2849 unsigned const_index
= 0;
2851 const bool per_vertex
= nir_is_per_vertex_io(instr
->variables
[0]->var
, ctx
->stage
);
2852 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
2853 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2854 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2855 false, NULL
, per_vertex
? &vertex_index
: NULL
,
2856 &const_index
, &indir_index
);
2858 if (!instr
->variables
[0]->var
->data
.patch
) {
2859 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 13, 8);
2860 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2862 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2865 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2868 unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
2869 for (unsigned i
= comp
; i
< instr
->num_components
+ comp
; i
++) {
2870 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
2871 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2874 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, instr
->num_components
, comp
);
2875 result
= LLVMBuildBitCast(ctx
->builder
, result
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
2880 store_tcs_output(struct ac_shader_abi
*abi
,
2881 LLVMValueRef vertex_index
,
2882 LLVMValueRef param_index
,
2883 unsigned const_index
,
2885 unsigned driver_location
,
2892 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2893 LLVMValueRef dw_addr
;
2894 LLVMValueRef stride
= NULL
;
2895 LLVMValueRef buf_addr
= NULL
;
2897 bool store_lds
= true;
2900 if (!(ctx
->tcs_patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
2903 if (!(ctx
->tcs_outputs_read
& (1ULL << location
)))
2907 param
= shader_io_get_unique_index(location
);
2908 if (location
== VARYING_SLOT_CLIP_DIST0
&&
2909 is_compact
&& const_index
> 3) {
2915 stride
= unpack_param(&ctx
->ac
, ctx
->tcs_out_layout
, 13, 8);
2916 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2918 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2921 mark_tess_output(ctx
, is_patch
, param
);
2923 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2925 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
2926 vertex_index
, param_index
);
2928 bool is_tess_factor
= false;
2929 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
2930 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
2931 is_tess_factor
= true;
2933 unsigned base
= is_compact
? const_index
: 0;
2934 for (unsigned chan
= 0; chan
< 8; chan
++) {
2935 if (!(writemask
& (1 << chan
)))
2937 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
2939 if (store_lds
|| is_tess_factor
)
2940 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
2942 if (!is_tess_factor
&& writemask
!= 0xF)
2943 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
2944 buf_addr
, ctx
->oc_lds
,
2945 4 * (base
+ chan
), 1, 0, true, false);
2947 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2951 if (writemask
== 0xF) {
2952 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
2953 buf_addr
, ctx
->oc_lds
,
2954 (base
* 4), 1, 0, true, false);
2959 load_tes_input(struct ac_shader_abi
*abi
,
2960 LLVMValueRef vertex_index
,
2961 LLVMValueRef param_index
,
2962 unsigned const_index
,
2964 unsigned driver_location
,
2966 unsigned num_components
,
2970 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
2971 LLVMValueRef buf_addr
;
2972 LLVMValueRef result
;
2973 unsigned param
= shader_io_get_unique_index(location
);
2975 if (location
== VARYING_SLOT_CLIP_DIST0
&& is_compact
&& const_index
> 3) {
2980 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
2981 is_compact
, vertex_index
, param_index
);
2983 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
2984 buf_addr
= LLVMBuildAdd(ctx
->builder
, buf_addr
, comp_offset
, "");
2986 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
2987 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
2988 result
= trim_vector(&ctx
->ac
, result
, num_components
);
2993 load_gs_input(struct ac_shader_abi
*abi
,
2995 unsigned driver_location
,
2997 unsigned num_components
,
2998 unsigned vertex_index
,
2999 unsigned const_index
,
3002 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
3003 LLVMValueRef vtx_offset
;
3004 LLVMValueRef args
[9];
3005 unsigned param
, vtx_offset_param
;
3006 LLVMValueRef value
[4], result
;
3008 vtx_offset_param
= vertex_index
;
3009 assert(vtx_offset_param
< 6);
3010 vtx_offset
= LLVMBuildMul(ctx
->builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
3011 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3013 param
= shader_io_get_unique_index(location
);
3015 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
3016 if (ctx
->ac
.chip_class
>= GFX9
) {
3017 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
3018 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
3019 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
3020 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
3022 args
[0] = ctx
->esgs_ring
;
3023 args
[1] = vtx_offset
;
3024 args
[2] = LLVMConstInt(ctx
->ac
.i32
, (param
* 4 + i
+ const_index
) * 256, false);
3025 args
[3] = ctx
->ac
.i32_0
;
3026 args
[4] = ctx
->ac
.i32_1
; /* OFFEN */
3027 args
[5] = ctx
->ac
.i32_0
; /* IDXEN */
3028 args
[6] = ctx
->ac
.i32_1
; /* GLC */
3029 args
[7] = ctx
->ac
.i32_0
; /* SLC */
3030 args
[8] = ctx
->ac
.i32_0
; /* TFE */
3032 value
[i
] = ac_build_intrinsic(&ctx
->ac
, "llvm.SI.buffer.load.dword.i32.i32",
3033 ctx
->ac
.i32
, args
, 9,
3034 AC_FUNC_ATTR_READONLY
|
3035 AC_FUNC_ATTR_LEGACY
);
3038 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
3044 build_gep_for_deref(struct ac_nir_context
*ctx
,
3045 nir_deref_var
*deref
)
3047 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, deref
->var
);
3048 assert(entry
->data
);
3049 LLVMValueRef val
= entry
->data
;
3050 nir_deref
*tail
= deref
->deref
.child
;
3051 while (tail
!= NULL
) {
3052 LLVMValueRef offset
;
3053 switch (tail
->deref_type
) {
3054 case nir_deref_type_array
: {
3055 nir_deref_array
*array
= nir_deref_as_array(tail
);
3056 offset
= LLVMConstInt(ctx
->ac
.i32
, array
->base_offset
, 0);
3057 if (array
->deref_array_type
==
3058 nir_deref_array_type_indirect
) {
3059 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
3066 case nir_deref_type_struct
: {
3067 nir_deref_struct
*deref_struct
=
3068 nir_deref_as_struct(tail
);
3069 offset
= LLVMConstInt(ctx
->ac
.i32
,
3070 deref_struct
->index
, 0);
3074 unreachable("bad deref type");
3076 val
= ac_build_gep0(&ctx
->ac
, val
, offset
);
3082 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
3083 nir_intrinsic_instr
*instr
)
3085 LLVMValueRef values
[8];
3086 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
3087 int ve
= instr
->dest
.ssa
.num_components
;
3088 unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3089 LLVMValueRef indir_index
;
3091 unsigned const_index
;
3092 unsigned stride
= instr
->variables
[0]->var
->data
.compact
? 1 : 4;
3093 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
3094 instr
->variables
[0]->var
->data
.mode
== nir_var_shader_in
;
3095 get_deref_offset(ctx
, instr
->variables
[0], vs_in
, NULL
, NULL
,
3096 &const_index
, &indir_index
);
3098 if (instr
->dest
.ssa
.bit_size
== 64)
3101 switch (instr
->variables
[0]->var
->data
.mode
) {
3102 case nir_var_shader_in
:
3103 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3104 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3105 LLVMValueRef result
;
3106 LLVMValueRef vertex_index
= NULL
;
3107 LLVMValueRef indir_index
= NULL
;
3108 unsigned const_index
= 0;
3109 unsigned location
= instr
->variables
[0]->var
->data
.location
;
3110 unsigned driver_location
= instr
->variables
[0]->var
->data
.driver_location
;
3111 const bool is_patch
= instr
->variables
[0]->var
->data
.patch
;
3112 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
3114 get_deref_offset(ctx
, instr
->variables
[0],
3115 false, NULL
, is_patch
? NULL
: &vertex_index
,
3116 &const_index
, &indir_index
);
3118 result
= ctx
->abi
->load_tess_inputs(ctx
->abi
, vertex_index
, indir_index
,
3119 const_index
, location
, driver_location
,
3120 instr
->variables
[0]->var
->data
.location_frac
,
3121 instr
->num_components
,
3122 is_patch
, is_compact
);
3123 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3126 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3127 LLVMValueRef indir_index
;
3128 unsigned const_index
, vertex_index
;
3129 get_deref_offset(ctx
, instr
->variables
[0],
3130 false, &vertex_index
, NULL
,
3131 &const_index
, &indir_index
);
3132 return ctx
->abi
->load_inputs(ctx
->abi
, instr
->variables
[0]->var
->data
.location
,
3133 instr
->variables
[0]->var
->data
.driver_location
,
3134 instr
->variables
[0]->var
->data
.location_frac
, ve
,
3135 vertex_index
, const_index
,
3136 nir2llvmtype(ctx
, instr
->variables
[0]->var
->type
));
3139 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
3141 unsigned count
= glsl_count_attribute_slots(
3142 instr
->variables
[0]->var
->type
,
3143 ctx
->stage
== MESA_SHADER_VERTEX
);
3145 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3146 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
3147 stride
, false, true);
3149 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3153 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* stride
];
3157 for (unsigned chan
= 0; chan
< ve
; chan
++) {
3159 unsigned count
= glsl_count_attribute_slots(
3160 instr
->variables
[0]->var
->type
, false);
3162 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3163 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3164 stride
, true, true);
3166 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3170 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* stride
], "");
3174 case nir_var_shared
: {
3175 LLVMValueRef address
= build_gep_for_deref(ctx
,
3176 instr
->variables
[0]);
3177 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
3178 return LLVMBuildBitCast(ctx
->ac
.builder
, val
,
3179 get_def_type(ctx
, &instr
->dest
.ssa
),
3182 case nir_var_shader_out
:
3183 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
3184 return load_tcs_output(ctx
->nctx
, instr
);
3186 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
3188 unsigned count
= glsl_count_attribute_slots(
3189 instr
->variables
[0]->var
->type
, false);
3191 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3192 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3193 stride
, true, true);
3195 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3199 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
3200 ctx
->outputs
[idx
+ chan
+ const_index
* stride
],
3206 unreachable("unhandle variable mode");
3208 ret
= ac_build_varying_gather_values(&ctx
->ac
, values
, ve
, comp
);
3209 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3213 visit_store_var(struct ac_nir_context
*ctx
,
3214 nir_intrinsic_instr
*instr
)
3216 LLVMValueRef temp_ptr
, value
;
3217 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
3218 unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3219 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3220 int writemask
= instr
->const_index
[0] << comp
;
3221 LLVMValueRef indir_index
;
3222 unsigned const_index
;
3223 get_deref_offset(ctx
, instr
->variables
[0], false,
3224 NULL
, NULL
, &const_index
, &indir_index
);
3226 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64) {
3227 int old_writemask
= writemask
;
3229 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
3230 LLVMVectorType(ctx
->ac
.f32
, ac_get_llvm_num_components(src
) * 2),
3234 for (unsigned chan
= 0; chan
< 4; chan
++) {
3235 if (old_writemask
& (1 << chan
))
3236 writemask
|= 3u << (2 * chan
);
3240 switch (instr
->variables
[0]->var
->data
.mode
) {
3241 case nir_var_shader_out
:
3243 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3244 LLVMValueRef vertex_index
= NULL
;
3245 LLVMValueRef indir_index
= NULL
;
3246 unsigned const_index
= 0;
3247 const unsigned location
= instr
->variables
[0]->var
->data
.location
;
3248 const unsigned driver_location
= instr
->variables
[0]->var
->data
.driver_location
;
3249 const unsigned comp
= instr
->variables
[0]->var
->data
.location_frac
;
3250 const bool is_patch
= instr
->variables
[0]->var
->data
.patch
;
3251 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
3253 get_deref_offset(ctx
, instr
->variables
[0],
3254 false, NULL
, is_patch
? NULL
: &vertex_index
,
3255 &const_index
, &indir_index
);
3257 ctx
->abi
->store_tcs_outputs(ctx
->abi
, vertex_index
, indir_index
,
3258 const_index
, location
, driver_location
,
3259 src
, comp
, is_patch
, is_compact
, writemask
);
3263 for (unsigned chan
= 0; chan
< 8; chan
++) {
3265 if (!(writemask
& (1 << chan
)))
3268 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- comp
);
3270 if (instr
->variables
[0]->var
->data
.compact
)
3273 unsigned count
= glsl_count_attribute_slots(
3274 instr
->variables
[0]->var
->type
, false);
3276 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3277 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3278 stride
, true, true);
3280 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3281 value
, indir_index
, "");
3282 build_store_values_extended(&ctx
->ac
, ctx
->outputs
+ idx
+ chan
,
3283 count
, stride
, tmp_vec
);
3286 temp_ptr
= ctx
->outputs
[idx
+ chan
+ const_index
* stride
];
3288 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3293 for (unsigned chan
= 0; chan
< 8; chan
++) {
3294 if (!(writemask
& (1 << chan
)))
3297 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
3299 unsigned count
= glsl_count_attribute_slots(
3300 instr
->variables
[0]->var
->type
, false);
3302 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3303 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3306 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3307 value
, indir_index
, "");
3308 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
3311 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
3313 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3317 case nir_var_shared
: {
3318 int writemask
= instr
->const_index
[0];
3319 LLVMValueRef address
= build_gep_for_deref(ctx
,
3320 instr
->variables
[0]);
3321 LLVMValueRef val
= get_src(ctx
, instr
->src
[0]);
3322 unsigned components
=
3323 glsl_get_vector_elements(
3324 nir_deref_tail(&instr
->variables
[0]->deref
)->type
);
3325 if (writemask
== (1 << components
) - 1) {
3326 val
= LLVMBuildBitCast(
3327 ctx
->ac
.builder
, val
,
3328 LLVMGetElementType(LLVMTypeOf(address
)), "");
3329 LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
3331 for (unsigned chan
= 0; chan
< 4; chan
++) {
3332 if (!(writemask
& (1 << chan
)))
3335 LLVMBuildStructGEP(ctx
->ac
.builder
,
3337 LLVMValueRef src
= ac_llvm_extract_elem(&ctx
->ac
, val
,
3339 src
= LLVMBuildBitCast(
3340 ctx
->ac
.builder
, src
,
3341 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
3342 LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
3352 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
3355 case GLSL_SAMPLER_DIM_BUF
:
3357 case GLSL_SAMPLER_DIM_1D
:
3358 return array
? 2 : 1;
3359 case GLSL_SAMPLER_DIM_2D
:
3360 return array
? 3 : 2;
3361 case GLSL_SAMPLER_DIM_MS
:
3362 return array
? 4 : 3;
3363 case GLSL_SAMPLER_DIM_3D
:
3364 case GLSL_SAMPLER_DIM_CUBE
:
3366 case GLSL_SAMPLER_DIM_RECT
:
3367 case GLSL_SAMPLER_DIM_SUBPASS
:
3369 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
3379 /* Adjust the sample index according to FMASK.
3381 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3382 * which is the identity mapping. Each nibble says which physical sample
3383 * should be fetched to get that sample.
3385 * For example, 0x11111100 means there are only 2 samples stored and
3386 * the second sample covers 3/4 of the pixel. When reading samples 0
3387 * and 1, return physical sample 0 (determined by the first two 0s
3388 * in FMASK), otherwise return physical sample 1.
3390 * The sample index should be adjusted as follows:
3391 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
3393 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
3394 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
3395 LLVMValueRef coord_z
,
3396 LLVMValueRef sample_index
,
3397 LLVMValueRef fmask_desc_ptr
)
3399 LLVMValueRef fmask_load_address
[4];
3402 fmask_load_address
[0] = coord_x
;
3403 fmask_load_address
[1] = coord_y
;
3405 fmask_load_address
[2] = coord_z
;
3406 fmask_load_address
[3] = LLVMGetUndef(ctx
->i32
);
3409 struct ac_image_args args
= {0};
3411 args
.opcode
= ac_image_load
;
3412 args
.da
= coord_z
? true : false;
3413 args
.resource
= fmask_desc_ptr
;
3415 args
.addr
= ac_build_gather_values(ctx
, fmask_load_address
, coord_z
? 4 : 2);
3417 res
= ac_build_image_opcode(ctx
, &args
);
3419 res
= ac_to_integer(ctx
, res
);
3420 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, false);
3421 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xf, false);
3423 LLVMValueRef fmask
= LLVMBuildExtractElement(ctx
->builder
,
3427 LLVMValueRef sample_index4
=
3428 LLVMBuildMul(ctx
->builder
, sample_index
, four
, "");
3429 LLVMValueRef shifted_fmask
=
3430 LLVMBuildLShr(ctx
->builder
, fmask
, sample_index4
, "");
3431 LLVMValueRef final_sample
=
3432 LLVMBuildAnd(ctx
->builder
, shifted_fmask
, F
, "");
3434 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3435 * resource descriptor is 0 (invalid),
3437 LLVMValueRef fmask_desc
=
3438 LLVMBuildBitCast(ctx
->builder
, fmask_desc_ptr
,
3441 LLVMValueRef fmask_word1
=
3442 LLVMBuildExtractElement(ctx
->builder
, fmask_desc
,
3445 LLVMValueRef word1_is_nonzero
=
3446 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3447 fmask_word1
, ctx
->i32_0
, "");
3449 /* Replace the MSAA sample index. */
3451 LLVMBuildSelect(ctx
->builder
, word1_is_nonzero
,
3452 final_sample
, sample_index
, "");
3453 return sample_index
;
3456 static LLVMValueRef
get_image_coords(struct ac_nir_context
*ctx
,
3457 const nir_intrinsic_instr
*instr
)
3459 const struct glsl_type
*type
= instr
->variables
[0]->var
->type
;
3460 if(instr
->variables
[0]->deref
.child
)
3461 type
= instr
->variables
[0]->deref
.child
->type
;
3463 LLVMValueRef src0
= get_src(ctx
, instr
->src
[0]);
3464 LLVMValueRef coords
[4];
3465 LLVMValueRef masks
[] = {
3466 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
3467 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
3470 LLVMValueRef sample_index
= ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[1]), 0);
3473 enum glsl_sampler_dim dim
= glsl_get_sampler_dim(type
);
3474 bool is_array
= glsl_sampler_type_is_array(type
);
3475 bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3476 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3477 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
3478 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3479 bool gfx9_1d
= ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
;
3480 count
= image_type_to_components_count(dim
, is_array
);
3483 LLVMValueRef fmask_load_address
[3];
3486 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3487 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
3489 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
3491 fmask_load_address
[2] = NULL
;
3493 for (chan
= 0; chan
< 2; ++chan
)
3494 fmask_load_address
[chan
] =
3495 LLVMBuildAdd(ctx
->ac
.builder
, fmask_load_address
[chan
],
3496 LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
3497 ctx
->ac
.i32
, ""), "");
3498 fmask_load_address
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
3500 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
3501 fmask_load_address
[0],
3502 fmask_load_address
[1],
3503 fmask_load_address
[2],
3505 get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_FMASK
, NULL
, true, false));
3507 if (count
== 1 && !gfx9_1d
) {
3508 if (instr
->src
[0].ssa
->num_components
)
3509 res
= LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3516 for (chan
= 0; chan
< count
; ++chan
) {
3517 coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src0
, chan
);
3520 for (chan
= 0; chan
< 2; ++chan
)
3521 coords
[chan
] = LLVMBuildAdd(ctx
->ac
.builder
, coords
[chan
], LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
3522 ctx
->ac
.i32
, ""), "");
3523 coords
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
3529 coords
[2] = coords
[1];
3530 coords
[1] = ctx
->ac
.i32_0
;
3532 coords
[1] = ctx
->ac
.i32_0
;
3537 coords
[count
] = sample_index
;
3542 coords
[3] = LLVMGetUndef(ctx
->ac
.i32
);
3545 res
= ac_build_gather_values(&ctx
->ac
, coords
, count
);
3550 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
3551 const nir_intrinsic_instr
*instr
)
3553 LLVMValueRef params
[7];
3555 char intrinsic_name
[64];
3556 const nir_variable
*var
= instr
->variables
[0]->var
;
3557 const struct glsl_type
*type
= var
->type
;
3559 if(instr
->variables
[0]->deref
.child
)
3560 type
= instr
->variables
[0]->deref
.child
->type
;
3562 type
= glsl_without_array(type
);
3563 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3564 params
[0] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, NULL
, true, false);
3565 params
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3566 ctx
->ac
.i32_0
, ""); /* vindex */
3567 params
[2] = ctx
->ac
.i32_0
; /* voffset */
3568 params
[3] = ctx
->ac
.i1false
; /* glc */
3569 params
[4] = ctx
->ac
.i1false
; /* slc */
3570 res
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.buffer.load.format.v4f32", ctx
->ac
.v4f32
,
3573 res
= trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
3574 res
= ac_to_integer(&ctx
->ac
, res
);
3576 bool is_da
= glsl_sampler_type_is_array(type
) ||
3577 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
||
3578 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_SUBPASS
||
3579 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_SUBPASS_MS
;
3580 LLVMValueRef da
= is_da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
;
3581 LLVMValueRef glc
= ctx
->ac
.i1false
;
3582 LLVMValueRef slc
= ctx
->ac
.i1false
;
3584 params
[0] = get_image_coords(ctx
, instr
);
3585 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, false);
3586 params
[2] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3587 if (HAVE_LLVM
<= 0x0309) {
3588 params
[3] = ctx
->ac
.i1false
; /* r128 */
3593 LLVMValueRef lwe
= ctx
->ac
.i1false
;
3600 ac_get_image_intr_name("llvm.amdgcn.image.load",
3601 ctx
->ac
.v4f32
, /* vdata */
3602 LLVMTypeOf(params
[0]), /* coords */
3603 LLVMTypeOf(params
[1]), /* rsrc */
3604 intrinsic_name
, sizeof(intrinsic_name
));
3606 res
= ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.v4f32
,
3607 params
, 7, AC_FUNC_ATTR_READONLY
);
3609 return ac_to_integer(&ctx
->ac
, res
);
3612 static void visit_image_store(struct ac_nir_context
*ctx
,
3613 nir_intrinsic_instr
*instr
)
3615 LLVMValueRef params
[8];
3616 char intrinsic_name
[64];
3617 const nir_variable
*var
= instr
->variables
[0]->var
;
3618 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3619 LLVMValueRef glc
= ctx
->ac
.i1false
;
3620 bool force_glc
= ctx
->ac
.chip_class
== SI
;
3622 glc
= ctx
->ac
.i1true
;
3624 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3625 params
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2])); /* data */
3626 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, NULL
, true, true);
3627 params
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3628 ctx
->ac
.i32_0
, ""); /* vindex */
3629 params
[3] = ctx
->ac
.i32_0
; /* voffset */
3630 params
[4] = glc
; /* glc */
3631 params
[5] = ctx
->ac
.i1false
; /* slc */
3632 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.buffer.store.format.v4f32", ctx
->ac
.voidt
,
3635 bool is_da
= glsl_sampler_type_is_array(type
) ||
3636 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3637 LLVMValueRef da
= is_da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
;
3638 LLVMValueRef slc
= ctx
->ac
.i1false
;
3640 params
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2]));
3641 params
[1] = get_image_coords(ctx
, instr
); /* coords */
3642 params
[2] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, true);
3643 params
[3] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3644 if (HAVE_LLVM
<= 0x0309) {
3645 params
[4] = ctx
->ac
.i1false
; /* r128 */
3650 LLVMValueRef lwe
= ctx
->ac
.i1false
;
3657 ac_get_image_intr_name("llvm.amdgcn.image.store",
3658 LLVMTypeOf(params
[0]), /* vdata */
3659 LLVMTypeOf(params
[1]), /* coords */
3660 LLVMTypeOf(params
[2]), /* rsrc */
3661 intrinsic_name
, sizeof(intrinsic_name
));
3663 ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.voidt
,
3669 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
3670 const nir_intrinsic_instr
*instr
)
3672 LLVMValueRef params
[7];
3673 int param_count
= 0;
3674 const nir_variable
*var
= instr
->variables
[0]->var
;
3676 const char *atomic_name
;
3677 char intrinsic_name
[41];
3678 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3679 MAYBE_UNUSED
int length
;
3681 bool is_unsigned
= glsl_get_sampler_result_type(type
) == GLSL_TYPE_UINT
;
3683 switch (instr
->intrinsic
) {
3684 case nir_intrinsic_image_atomic_add
:
3685 atomic_name
= "add";
3687 case nir_intrinsic_image_atomic_min
:
3688 atomic_name
= is_unsigned
? "umin" : "smin";
3690 case nir_intrinsic_image_atomic_max
:
3691 atomic_name
= is_unsigned
? "umax" : "smax";
3693 case nir_intrinsic_image_atomic_and
:
3694 atomic_name
= "and";
3696 case nir_intrinsic_image_atomic_or
:
3699 case nir_intrinsic_image_atomic_xor
:
3700 atomic_name
= "xor";
3702 case nir_intrinsic_image_atomic_exchange
:
3703 atomic_name
= "swap";
3705 case nir_intrinsic_image_atomic_comp_swap
:
3706 atomic_name
= "cmpswap";
3712 if (instr
->intrinsic
== nir_intrinsic_image_atomic_comp_swap
)
3713 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
3714 params
[param_count
++] = get_src(ctx
, instr
->src
[2]);
3716 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3717 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
,
3719 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3720 ctx
->ac
.i32_0
, ""); /* vindex */
3721 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
3722 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
3724 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3725 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
3727 char coords_type
[8];
3729 bool da
= glsl_sampler_type_is_array(type
) ||
3730 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3732 LLVMValueRef coords
= params
[param_count
++] = get_image_coords(ctx
, instr
);
3733 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
,
3735 params
[param_count
++] = ctx
->ac
.i1false
; /* r128 */
3736 params
[param_count
++] = da
? ctx
->ac
.i1true
: ctx
->ac
.i1false
; /* da */
3737 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
3739 build_int_type_name(LLVMTypeOf(coords
),
3740 coords_type
, sizeof(coords_type
));
3742 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3743 "llvm.amdgcn.image.atomic.%s.%s", atomic_name
, coords_type
);
3746 assert(length
< sizeof(intrinsic_name
));
3747 return ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
, params
, param_count
, 0);
3750 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
3751 const nir_intrinsic_instr
*instr
)
3754 const nir_variable
*var
= instr
->variables
[0]->var
;
3755 const struct glsl_type
*type
= instr
->variables
[0]->var
->type
;
3756 bool da
= glsl_sampler_type_is_array(var
->type
) ||
3757 glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_CUBE
;
3758 if(instr
->variables
[0]->deref
.child
)
3759 type
= instr
->variables
[0]->deref
.child
->type
;
3761 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
)
3762 return get_buffer_size(ctx
,
3763 get_sampler_desc(ctx
, instr
->variables
[0],
3764 AC_DESC_BUFFER
, NULL
, true, false), true);
3766 struct ac_image_args args
= { 0 };
3770 args
.resource
= get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, NULL
, true, false);
3771 args
.opcode
= ac_image_get_resinfo
;
3772 args
.addr
= ctx
->ac
.i32_0
;
3774 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
3776 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3778 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
&&
3779 glsl_sampler_type_is_array(type
)) {
3780 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3781 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3782 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3783 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
3785 if (ctx
->ac
.chip_class
>= GFX9
&&
3786 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_1D
&&
3787 glsl_sampler_type_is_array(type
)) {
3788 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3789 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, layers
,
3796 #define NOOP_WAITCNT 0xf7f
3797 #define LGKM_CNT 0x07f
3798 #define VM_CNT 0xf70
3800 static void emit_membar(struct nir_to_llvm_context
*ctx
,
3801 const nir_intrinsic_instr
*instr
)
3803 unsigned waitcnt
= NOOP_WAITCNT
;
3805 switch (instr
->intrinsic
) {
3806 case nir_intrinsic_memory_barrier
:
3807 case nir_intrinsic_group_memory_barrier
:
3808 waitcnt
&= VM_CNT
& LGKM_CNT
;
3810 case nir_intrinsic_memory_barrier_atomic_counter
:
3811 case nir_intrinsic_memory_barrier_buffer
:
3812 case nir_intrinsic_memory_barrier_image
:
3815 case nir_intrinsic_memory_barrier_shared
:
3816 waitcnt
&= LGKM_CNT
;
3821 if (waitcnt
!= NOOP_WAITCNT
)
3822 ac_build_waitcnt(&ctx
->ac
, waitcnt
);
3825 static void emit_barrier(struct nir_to_llvm_context
*ctx
)
3827 /* SI only (thanks to a hw bug workaround):
3828 * The real barrier instruction isn’t needed, because an entire patch
3829 * always fits into a single wave.
3831 if (ctx
->options
->chip_class
== SI
&&
3832 ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3833 ac_build_waitcnt(&ctx
->ac
, LGKM_CNT
& VM_CNT
);
3836 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.s.barrier",
3837 ctx
->ac
.voidt
, NULL
, 0, AC_FUNC_ATTR_CONVERGENT
);
3840 static void emit_discard_if(struct ac_nir_context
*ctx
,
3841 const nir_intrinsic_instr
*instr
)
3845 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3846 get_src(ctx
, instr
->src
[0]),
3848 ac_build_kill_if_false(&ctx
->ac
, cond
);
3852 visit_load_local_invocation_index(struct nir_to_llvm_context
*ctx
)
3854 LLVMValueRef result
;
3855 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
3856 result
= LLVMBuildAnd(ctx
->builder
, ctx
->tg_size
,
3857 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
3859 return LLVMBuildAdd(ctx
->builder
, result
, thread_id
, "");
3862 static LLVMValueRef
visit_var_atomic(struct nir_to_llvm_context
*ctx
,
3863 const nir_intrinsic_instr
*instr
)
3865 LLVMValueRef ptr
, result
;
3866 LLVMValueRef src
= get_src(ctx
->nir
, instr
->src
[0]);
3867 ptr
= build_gep_for_deref(ctx
->nir
, instr
->variables
[0]);
3869 if (instr
->intrinsic
== nir_intrinsic_var_atomic_comp_swap
) {
3870 LLVMValueRef src1
= get_src(ctx
->nir
, instr
->src
[1]);
3871 result
= LLVMBuildAtomicCmpXchg(ctx
->builder
,
3873 LLVMAtomicOrderingSequentiallyConsistent
,
3874 LLVMAtomicOrderingSequentiallyConsistent
,
3877 LLVMAtomicRMWBinOp op
;
3878 switch (instr
->intrinsic
) {
3879 case nir_intrinsic_var_atomic_add
:
3880 op
= LLVMAtomicRMWBinOpAdd
;
3882 case nir_intrinsic_var_atomic_umin
:
3883 op
= LLVMAtomicRMWBinOpUMin
;
3885 case nir_intrinsic_var_atomic_umax
:
3886 op
= LLVMAtomicRMWBinOpUMax
;
3888 case nir_intrinsic_var_atomic_imin
:
3889 op
= LLVMAtomicRMWBinOpMin
;
3891 case nir_intrinsic_var_atomic_imax
:
3892 op
= LLVMAtomicRMWBinOpMax
;
3894 case nir_intrinsic_var_atomic_and
:
3895 op
= LLVMAtomicRMWBinOpAnd
;
3897 case nir_intrinsic_var_atomic_or
:
3898 op
= LLVMAtomicRMWBinOpOr
;
3900 case nir_intrinsic_var_atomic_xor
:
3901 op
= LLVMAtomicRMWBinOpXor
;
3903 case nir_intrinsic_var_atomic_exchange
:
3904 op
= LLVMAtomicRMWBinOpXchg
;
3910 result
= LLVMBuildAtomicRMW(ctx
->builder
, op
, ptr
, ac_to_integer(&ctx
->ac
, src
),
3911 LLVMAtomicOrderingSequentiallyConsistent
,
3917 #define INTERP_CENTER 0
3918 #define INTERP_CENTROID 1
3919 #define INTERP_SAMPLE 2
3921 static LLVMValueRef
lookup_interp_param(struct nir_to_llvm_context
*ctx
,
3922 enum glsl_interp_mode interp
, unsigned location
)
3925 case INTERP_MODE_FLAT
:
3928 case INTERP_MODE_SMOOTH
:
3929 case INTERP_MODE_NONE
:
3930 if (location
== INTERP_CENTER
)
3931 return ctx
->persp_center
;
3932 else if (location
== INTERP_CENTROID
)
3933 return ctx
->persp_centroid
;
3934 else if (location
== INTERP_SAMPLE
)
3935 return ctx
->persp_sample
;
3937 case INTERP_MODE_NOPERSPECTIVE
:
3938 if (location
== INTERP_CENTER
)
3939 return ctx
->linear_center
;
3940 else if (location
== INTERP_CENTROID
)
3941 return ctx
->linear_centroid
;
3942 else if (location
== INTERP_SAMPLE
)
3943 return ctx
->linear_sample
;
3949 static LLVMValueRef
load_sample_position(struct nir_to_llvm_context
*ctx
,
3950 LLVMValueRef sample_id
)
3952 LLVMValueRef result
;
3953 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false));
3955 ptr
= LLVMBuildBitCast(ctx
->builder
, ptr
,
3956 const_array(ctx
->ac
.v2f32
, 64), "");
3958 sample_id
= LLVMBuildAdd(ctx
->builder
, sample_id
, ctx
->sample_pos_offset
, "");
3959 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
3964 static LLVMValueRef
load_sample_pos(struct ac_nir_context
*ctx
)
3966 LLVMValueRef values
[2];
3968 values
[0] = emit_ffract(&ctx
->ac
, ctx
->abi
->frag_pos
[0]);
3969 values
[1] = emit_ffract(&ctx
->ac
, ctx
->abi
->frag_pos
[1]);
3970 return ac_build_gather_values(&ctx
->ac
, values
, 2);
3973 static LLVMValueRef
visit_interp(struct nir_to_llvm_context
*ctx
,
3974 const nir_intrinsic_instr
*instr
)
3976 LLVMValueRef result
[4];
3977 LLVMValueRef interp_param
, attr_number
;
3980 LLVMValueRef src_c0
= NULL
;
3981 LLVMValueRef src_c1
= NULL
;
3982 LLVMValueRef src0
= NULL
;
3983 int input_index
= instr
->variables
[0]->var
->data
.location
- VARYING_SLOT_VAR0
;
3984 switch (instr
->intrinsic
) {
3985 case nir_intrinsic_interp_var_at_centroid
:
3986 location
= INTERP_CENTROID
;
3988 case nir_intrinsic_interp_var_at_sample
:
3989 case nir_intrinsic_interp_var_at_offset
:
3990 location
= INTERP_CENTER
;
3991 src0
= get_src(ctx
->nir
, instr
->src
[0]);
3997 if (instr
->intrinsic
== nir_intrinsic_interp_var_at_offset
) {
3998 src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->ac
.i32_0
, ""));
3999 src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->ac
.i32_1
, ""));
4000 } else if (instr
->intrinsic
== nir_intrinsic_interp_var_at_sample
) {
4001 LLVMValueRef sample_position
;
4002 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
4004 /* fetch sample ID */
4005 sample_position
= load_sample_position(ctx
, src0
);
4007 src_c0
= LLVMBuildExtractElement(ctx
->builder
, sample_position
, ctx
->ac
.i32_0
, "");
4008 src_c0
= LLVMBuildFSub(ctx
->builder
, src_c0
, halfval
, "");
4009 src_c1
= LLVMBuildExtractElement(ctx
->builder
, sample_position
, ctx
->ac
.i32_1
, "");
4010 src_c1
= LLVMBuildFSub(ctx
->builder
, src_c1
, halfval
, "");
4012 interp_param
= lookup_interp_param(ctx
, instr
->variables
[0]->var
->data
.interpolation
, location
);
4013 attr_number
= LLVMConstInt(ctx
->ac
.i32
, input_index
, false);
4015 if (location
== INTERP_CENTER
) {
4016 LLVMValueRef ij_out
[2];
4017 LLVMValueRef ddxy_out
= emit_ddxy_interp(ctx
->nir
, interp_param
);
4020 * take the I then J parameters, and the DDX/Y for it, and
4021 * calculate the IJ inputs for the interpolator.
4022 * temp1 = ddx * offset/sample.x + I;
4023 * interp_param.I = ddy * offset/sample.y + temp1;
4024 * temp1 = ddx * offset/sample.x + J;
4025 * interp_param.J = ddy * offset/sample.y + temp1;
4027 for (unsigned i
= 0; i
< 2; i
++) {
4028 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
4029 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
4030 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->builder
,
4031 ddxy_out
, ix_ll
, "");
4032 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->builder
,
4033 ddxy_out
, iy_ll
, "");
4034 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->builder
,
4035 interp_param
, ix_ll
, "");
4036 LLVMValueRef temp1
, temp2
;
4038 interp_el
= LLVMBuildBitCast(ctx
->builder
, interp_el
,
4041 temp1
= LLVMBuildFMul(ctx
->builder
, ddx_el
, src_c0
, "");
4042 temp1
= LLVMBuildFAdd(ctx
->builder
, temp1
, interp_el
, "");
4044 temp2
= LLVMBuildFMul(ctx
->builder
, ddy_el
, src_c1
, "");
4045 temp2
= LLVMBuildFAdd(ctx
->builder
, temp2
, temp1
, "");
4047 ij_out
[i
] = LLVMBuildBitCast(ctx
->builder
,
4048 temp2
, ctx
->ac
.i32
, "");
4050 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
4054 for (chan
= 0; chan
< 4; chan
++) {
4055 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
4058 interp_param
= LLVMBuildBitCast(ctx
->builder
,
4059 interp_param
, ctx
->ac
.v2f32
, "");
4060 LLVMValueRef i
= LLVMBuildExtractElement(
4061 ctx
->builder
, interp_param
, ctx
->ac
.i32_0
, "");
4062 LLVMValueRef j
= LLVMBuildExtractElement(
4063 ctx
->builder
, interp_param
, ctx
->ac
.i32_1
, "");
4065 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
4066 llvm_chan
, attr_number
,
4067 ctx
->prim_mask
, i
, j
);
4069 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
4070 LLVMConstInt(ctx
->ac
.i32
, 2, false),
4071 llvm_chan
, attr_number
,
4075 return ac_build_varying_gather_values(&ctx
->ac
, result
, instr
->num_components
,
4076 instr
->variables
[0]->var
->data
.location_frac
);
4080 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
4082 LLVMValueRef gs_next_vertex
;
4083 LLVMValueRef can_emit
;
4085 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4087 /* Write vertex attribute values to GSVS ring */
4088 gs_next_vertex
= LLVMBuildLoad(ctx
->builder
,
4089 ctx
->gs_next_vertex
,
4092 /* If this thread has already emitted the declared maximum number of
4093 * vertices, kill it: excessive vertex emissions are not supposed to
4094 * have any effect, and GS threads have no externally observable
4095 * effects other than emitting vertices.
4097 can_emit
= LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, gs_next_vertex
,
4098 LLVMConstInt(ctx
->ac
.i32
, ctx
->gs_max_out_vertices
, false), "");
4099 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4101 /* loop num outputs */
4103 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
4104 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
4109 if (!(ctx
->output_mask
& (1ull << i
)))
4112 if (i
== VARYING_SLOT_CLIP_DIST0
) {
4113 /* pack clip and cull into a single set of slots */
4114 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
4118 for (unsigned j
= 0; j
< length
; j
++) {
4119 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
,
4121 LLVMValueRef voffset
= LLVMConstInt(ctx
->ac
.i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
4122 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, gs_next_vertex
, "");
4123 voffset
= LLVMBuildMul(ctx
->builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
4125 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->ac
.i32
, "");
4127 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
4129 voffset
, ctx
->gs2vs_offset
, 0,
4135 gs_next_vertex
= LLVMBuildAdd(ctx
->builder
, gs_next_vertex
,
4137 LLVMBuildStore(ctx
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
4139 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
4143 visit_end_primitive(struct nir_to_llvm_context
*ctx
,
4144 const nir_intrinsic_instr
*instr
)
4146 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
4150 visit_load_tess_coord(struct nir_to_llvm_context
*ctx
,
4151 const nir_intrinsic_instr
*instr
)
4153 LLVMValueRef coord
[4] = {
4160 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
4161 coord
[2] = LLVMBuildFSub(ctx
->builder
, ctx
->ac
.f32_1
,
4162 LLVMBuildFAdd(ctx
->builder
, coord
[0], coord
[1], ""), "");
4164 LLVMValueRef result
= ac_build_gather_values(&ctx
->ac
, coord
, instr
->num_components
);
4165 return LLVMBuildBitCast(ctx
->builder
, result
,
4166 get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
4169 static void visit_intrinsic(struct ac_nir_context
*ctx
,
4170 nir_intrinsic_instr
*instr
)
4172 LLVMValueRef result
= NULL
;
4174 switch (instr
->intrinsic
) {
4175 case nir_intrinsic_load_work_group_id
: {
4176 LLVMValueRef values
[3];
4178 for (int i
= 0; i
< 3; i
++) {
4179 values
[i
] = ctx
->nctx
->workgroup_ids
[i
] ?
4180 ctx
->nctx
->workgroup_ids
[i
] : ctx
->ac
.i32_0
;
4183 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
4186 case nir_intrinsic_load_base_vertex
: {
4187 result
= ctx
->abi
->base_vertex
;
4190 case nir_intrinsic_load_vertex_id_zero_base
: {
4191 result
= ctx
->abi
->vertex_id
;
4194 case nir_intrinsic_load_local_invocation_id
: {
4195 result
= ctx
->nctx
->local_invocation_ids
;
4198 case nir_intrinsic_load_base_instance
:
4199 result
= ctx
->abi
->start_instance
;
4201 case nir_intrinsic_load_draw_id
:
4202 result
= ctx
->abi
->draw_id
;
4204 case nir_intrinsic_load_view_index
:
4205 result
= ctx
->nctx
->view_index
? ctx
->nctx
->view_index
: ctx
->ac
.i32_0
;
4207 case nir_intrinsic_load_invocation_id
:
4208 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
4209 result
= unpack_param(&ctx
->ac
, ctx
->abi
->tcs_rel_ids
, 8, 5);
4211 result
= ctx
->abi
->gs_invocation_id
;
4213 case nir_intrinsic_load_primitive_id
:
4214 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
4215 result
= ctx
->abi
->gs_prim_id
;
4216 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
4217 result
= ctx
->abi
->tcs_patch_id
;
4218 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
4219 result
= ctx
->abi
->tes_patch_id
;
4221 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
4223 case nir_intrinsic_load_sample_id
:
4224 result
= unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
4226 case nir_intrinsic_load_sample_pos
:
4227 result
= load_sample_pos(ctx
);
4229 case nir_intrinsic_load_sample_mask_in
:
4230 result
= ctx
->abi
->sample_coverage
;
4232 case nir_intrinsic_load_frag_coord
: {
4233 LLVMValueRef values
[4] = {
4234 ctx
->abi
->frag_pos
[0],
4235 ctx
->abi
->frag_pos
[1],
4236 ctx
->abi
->frag_pos
[2],
4237 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, ctx
->abi
->frag_pos
[3])
4239 result
= ac_build_gather_values(&ctx
->ac
, values
, 4);
4242 case nir_intrinsic_load_front_face
:
4243 result
= ctx
->abi
->front_face
;
4245 case nir_intrinsic_load_instance_id
:
4246 result
= ctx
->abi
->instance_id
;
4248 case nir_intrinsic_load_num_work_groups
:
4249 result
= ctx
->nctx
->num_work_groups
;
4251 case nir_intrinsic_load_local_invocation_index
:
4252 result
= visit_load_local_invocation_index(ctx
->nctx
);
4254 case nir_intrinsic_load_push_constant
:
4255 result
= visit_load_push_constant(ctx
->nctx
, instr
);
4257 case nir_intrinsic_vulkan_resource_index
:
4258 result
= visit_vulkan_resource_index(ctx
->nctx
, instr
);
4260 case nir_intrinsic_vulkan_resource_reindex
:
4261 result
= visit_vulkan_resource_reindex(ctx
->nctx
, instr
);
4263 case nir_intrinsic_store_ssbo
:
4264 visit_store_ssbo(ctx
, instr
);
4266 case nir_intrinsic_load_ssbo
:
4267 result
= visit_load_buffer(ctx
, instr
);
4269 case nir_intrinsic_ssbo_atomic_add
:
4270 case nir_intrinsic_ssbo_atomic_imin
:
4271 case nir_intrinsic_ssbo_atomic_umin
:
4272 case nir_intrinsic_ssbo_atomic_imax
:
4273 case nir_intrinsic_ssbo_atomic_umax
:
4274 case nir_intrinsic_ssbo_atomic_and
:
4275 case nir_intrinsic_ssbo_atomic_or
:
4276 case nir_intrinsic_ssbo_atomic_xor
:
4277 case nir_intrinsic_ssbo_atomic_exchange
:
4278 case nir_intrinsic_ssbo_atomic_comp_swap
:
4279 result
= visit_atomic_ssbo(ctx
, instr
);
4281 case nir_intrinsic_load_ubo
:
4282 result
= visit_load_ubo_buffer(ctx
, instr
);
4284 case nir_intrinsic_get_buffer_size
:
4285 result
= visit_get_buffer_size(ctx
, instr
);
4287 case nir_intrinsic_load_var
:
4288 result
= visit_load_var(ctx
, instr
);
4290 case nir_intrinsic_store_var
:
4291 visit_store_var(ctx
, instr
);
4293 case nir_intrinsic_image_load
:
4294 result
= visit_image_load(ctx
, instr
);
4296 case nir_intrinsic_image_store
:
4297 visit_image_store(ctx
, instr
);
4299 case nir_intrinsic_image_atomic_add
:
4300 case nir_intrinsic_image_atomic_min
:
4301 case nir_intrinsic_image_atomic_max
:
4302 case nir_intrinsic_image_atomic_and
:
4303 case nir_intrinsic_image_atomic_or
:
4304 case nir_intrinsic_image_atomic_xor
:
4305 case nir_intrinsic_image_atomic_exchange
:
4306 case nir_intrinsic_image_atomic_comp_swap
:
4307 result
= visit_image_atomic(ctx
, instr
);
4309 case nir_intrinsic_image_size
:
4310 result
= visit_image_size(ctx
, instr
);
4312 case nir_intrinsic_discard
:
4313 ac_build_intrinsic(&ctx
->ac
, "llvm.AMDGPU.kilp",
4314 LLVMVoidTypeInContext(ctx
->ac
.context
),
4315 NULL
, 0, AC_FUNC_ATTR_LEGACY
);
4317 case nir_intrinsic_discard_if
:
4318 emit_discard_if(ctx
, instr
);
4320 case nir_intrinsic_memory_barrier
:
4321 case nir_intrinsic_group_memory_barrier
:
4322 case nir_intrinsic_memory_barrier_atomic_counter
:
4323 case nir_intrinsic_memory_barrier_buffer
:
4324 case nir_intrinsic_memory_barrier_image
:
4325 case nir_intrinsic_memory_barrier_shared
:
4326 emit_membar(ctx
->nctx
, instr
);
4328 case nir_intrinsic_barrier
:
4329 emit_barrier(ctx
->nctx
);
4331 case nir_intrinsic_var_atomic_add
:
4332 case nir_intrinsic_var_atomic_imin
:
4333 case nir_intrinsic_var_atomic_umin
:
4334 case nir_intrinsic_var_atomic_imax
:
4335 case nir_intrinsic_var_atomic_umax
:
4336 case nir_intrinsic_var_atomic_and
:
4337 case nir_intrinsic_var_atomic_or
:
4338 case nir_intrinsic_var_atomic_xor
:
4339 case nir_intrinsic_var_atomic_exchange
:
4340 case nir_intrinsic_var_atomic_comp_swap
:
4341 result
= visit_var_atomic(ctx
->nctx
, instr
);
4343 case nir_intrinsic_interp_var_at_centroid
:
4344 case nir_intrinsic_interp_var_at_sample
:
4345 case nir_intrinsic_interp_var_at_offset
:
4346 result
= visit_interp(ctx
->nctx
, instr
);
4348 case nir_intrinsic_emit_vertex
:
4349 assert(instr
->const_index
[0] == 0);
4350 ctx
->abi
->emit_vertex(ctx
->abi
, 0, ctx
->outputs
);
4352 case nir_intrinsic_end_primitive
:
4353 visit_end_primitive(ctx
->nctx
, instr
);
4355 case nir_intrinsic_load_tess_coord
:
4356 result
= visit_load_tess_coord(ctx
->nctx
, instr
);
4358 case nir_intrinsic_load_patch_vertices_in
:
4359 result
= LLVMConstInt(ctx
->ac
.i32
, ctx
->nctx
->options
->key
.tcs
.input_vertices
, false);
4362 fprintf(stderr
, "Unknown intrinsic: ");
4363 nir_print_instr(&instr
->instr
, stderr
);
4364 fprintf(stderr
, "\n");
4368 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4372 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
4373 LLVMValueRef buffer_ptr
, bool write
)
4375 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4377 if (write
&& ctx
->stage
== MESA_SHADER_FRAGMENT
)
4378 ctx
->shader_info
->fs
.writes_memory
= true;
4380 return LLVMBuildLoad(ctx
->builder
, buffer_ptr
, "");
4383 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
4385 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4387 return LLVMBuildLoad(ctx
->builder
, buffer_ptr
, "");
4390 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
4391 unsigned descriptor_set
,
4392 unsigned base_index
,
4393 unsigned constant_index
,
4395 enum ac_descriptor_type desc_type
,
4396 bool image
, bool write
)
4398 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4399 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
4400 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
4401 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
4402 unsigned offset
= binding
->offset
;
4403 unsigned stride
= binding
->size
;
4405 LLVMBuilderRef builder
= ctx
->builder
;
4408 assert(base_index
< layout
->binding_count
);
4410 if (write
&& ctx
->stage
== MESA_SHADER_FRAGMENT
)
4411 ctx
->shader_info
->fs
.writes_memory
= true;
4413 switch (desc_type
) {
4415 type
= ctx
->ac
.v8i32
;
4419 type
= ctx
->ac
.v8i32
;
4423 case AC_DESC_SAMPLER
:
4424 type
= ctx
->ac
.v4i32
;
4425 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
4430 case AC_DESC_BUFFER
:
4431 type
= ctx
->ac
.v4i32
;
4435 unreachable("invalid desc_type\n");
4438 offset
+= constant_index
* stride
;
4440 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
4441 (!index
|| binding
->immutable_samplers_equal
)) {
4442 if (binding
->immutable_samplers_equal
)
4445 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
4447 LLVMValueRef constants
[] = {
4448 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
4449 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
4450 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
4451 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
4453 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
4456 assert(stride
% type_size
== 0);
4459 index
= ctx
->ac
.i32_0
;
4461 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
4463 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
4464 list
= LLVMBuildPointerCast(builder
, list
, const_array(type
, 0), "");
4466 return ac_build_load_to_sgpr(&ctx
->ac
, list
, index
);
4469 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
4470 const nir_deref_var
*deref
,
4471 enum ac_descriptor_type desc_type
,
4472 const nir_tex_instr
*tex_instr
,
4473 bool image
, bool write
)
4475 LLVMValueRef index
= NULL
;
4476 unsigned constant_index
= 0;
4477 unsigned descriptor_set
;
4478 unsigned base_index
;
4481 assert(tex_instr
&& !image
);
4483 base_index
= tex_instr
->sampler_index
;
4485 const nir_deref
*tail
= &deref
->deref
;
4486 while (tail
->child
) {
4487 const nir_deref_array
*child
= nir_deref_as_array(tail
->child
);
4488 unsigned array_size
= glsl_get_aoa_size(tail
->child
->type
);
4493 assert(child
->deref_array_type
!= nir_deref_array_type_wildcard
);
4495 if (child
->deref_array_type
== nir_deref_array_type_indirect
) {
4496 LLVMValueRef indirect
= get_src(ctx
, child
->indirect
);
4498 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
4499 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
4504 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
4507 constant_index
+= child
->base_offset
* array_size
;
4509 tail
= &child
->deref
;
4511 descriptor_set
= deref
->var
->data
.descriptor_set
;
4512 base_index
= deref
->var
->data
.binding
;
4515 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
4518 constant_index
, index
,
4519 desc_type
, image
, write
);
4522 static void set_tex_fetch_args(struct ac_llvm_context
*ctx
,
4523 struct ac_image_args
*args
,
4524 const nir_tex_instr
*instr
,
4526 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4527 LLVMValueRef
*param
, unsigned count
,
4530 unsigned is_rect
= 0;
4531 bool da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
4533 if (op
== nir_texop_lod
)
4535 /* Pad to power of two vector */
4536 while (count
< util_next_power_of_two(count
))
4537 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4540 args
->addr
= ac_build_gather_values(ctx
, param
, count
);
4542 args
->addr
= param
[0];
4544 args
->resource
= res_ptr
;
4545 args
->sampler
= samp_ptr
;
4547 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
&& op
== nir_texop_txf
) {
4548 args
->addr
= param
[0];
4552 args
->dmask
= dmask
;
4553 args
->unorm
= is_rect
;
4557 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4560 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4561 * filtering manually. The driver sets img7 to a mask clearing
4562 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4563 * s_and_b32 samp0, samp0, img7
4566 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4568 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
4569 LLVMValueRef res
, LLVMValueRef samp
)
4571 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4572 LLVMValueRef img7
, samp0
;
4574 if (ctx
->ac
.chip_class
>= VI
)
4577 img7
= LLVMBuildExtractElement(builder
, res
,
4578 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
4579 samp0
= LLVMBuildExtractElement(builder
, samp
,
4580 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4581 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4582 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4583 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4586 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
4587 nir_tex_instr
*instr
,
4588 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
4589 LLVMValueRef
*fmask_ptr
)
4591 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
4592 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_BUFFER
, instr
, false, false);
4594 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_IMAGE
, instr
, false, false);
4597 *samp_ptr
= get_sampler_desc(ctx
, instr
->sampler
, AC_DESC_SAMPLER
, instr
, false, false);
4599 *samp_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_SAMPLER
, instr
, false, false);
4600 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
4601 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4603 if (fmask_ptr
&& !instr
->sampler
&& (instr
->op
== nir_texop_txf_ms
||
4604 instr
->op
== nir_texop_samples_identical
))
4605 *fmask_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_FMASK
, instr
, false, false);
4608 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
4611 coord
= ac_to_float(ctx
, coord
);
4612 coord
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &coord
, 1, 0);
4613 coord
= ac_to_integer(ctx
, coord
);
4617 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
4619 LLVMValueRef result
= NULL
;
4620 struct ac_image_args args
= { 0 };
4621 unsigned dmask
= 0xf;
4622 LLVMValueRef address
[16];
4623 LLVMValueRef coords
[5];
4624 LLVMValueRef coord
= NULL
, lod
= NULL
, comparator
= NULL
;
4625 LLVMValueRef bias
= NULL
, offsets
= NULL
;
4626 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
, sample_index
= NULL
;
4627 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
4628 LLVMValueRef derivs
[6];
4629 unsigned chan
, count
= 0;
4630 unsigned const_src
= 0, num_deriv_comp
= 0;
4631 bool lod_is_zero
= false;
4633 tex_fetch_ptrs(ctx
, instr
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4635 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4636 switch (instr
->src
[i
].src_type
) {
4637 case nir_tex_src_coord
:
4638 coord
= get_src(ctx
, instr
->src
[i
].src
);
4640 case nir_tex_src_projector
:
4642 case nir_tex_src_comparator
:
4643 comparator
= get_src(ctx
, instr
->src
[i
].src
);
4645 case nir_tex_src_offset
:
4646 offsets
= get_src(ctx
, instr
->src
[i
].src
);
4649 case nir_tex_src_bias
:
4650 bias
= get_src(ctx
, instr
->src
[i
].src
);
4652 case nir_tex_src_lod
: {
4653 nir_const_value
*val
= nir_src_as_const_value(instr
->src
[i
].src
);
4655 if (val
&& val
->i32
[0] == 0)
4657 lod
= get_src(ctx
, instr
->src
[i
].src
);
4660 case nir_tex_src_ms_index
:
4661 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
4663 case nir_tex_src_ms_mcs
:
4665 case nir_tex_src_ddx
:
4666 ddx
= get_src(ctx
, instr
->src
[i
].src
);
4667 num_deriv_comp
= instr
->src
[i
].src
.ssa
->num_components
;
4669 case nir_tex_src_ddy
:
4670 ddy
= get_src(ctx
, instr
->src
[i
].src
);
4672 case nir_tex_src_texture_offset
:
4673 case nir_tex_src_sampler_offset
:
4674 case nir_tex_src_plane
:
4680 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
4681 result
= get_buffer_size(ctx
, res_ptr
, true);
4685 if (instr
->op
== nir_texop_texture_samples
) {
4686 LLVMValueRef res
, samples
, is_msaa
;
4687 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res_ptr
, ctx
->ac
.v8i32
, "");
4688 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4689 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4690 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4691 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
4692 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
4693 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4694 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
4695 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4697 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4698 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
4699 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
4700 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
4701 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
4703 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
4710 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4711 coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
4713 if (offsets
&& instr
->op
!= nir_texop_txf
) {
4714 LLVMValueRef offset
[3], pack
;
4715 for (chan
= 0; chan
< 3; ++chan
)
4716 offset
[chan
] = ctx
->ac
.i32_0
;
4719 for (chan
= 0; chan
< ac_get_llvm_num_components(offsets
); chan
++) {
4720 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, offsets
, chan
);
4721 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
4722 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
4724 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
4725 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
4727 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
4728 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
4729 address
[count
++] = pack
;
4732 /* pack LOD bias value */
4733 if (instr
->op
== nir_texop_txb
&& bias
) {
4734 address
[count
++] = bias
;
4737 /* Pack depth comparison value */
4738 if (instr
->is_shadow
&& comparator
) {
4739 LLVMValueRef z
= ac_to_float(&ctx
->ac
,
4740 ac_llvm_extract_elem(&ctx
->ac
, comparator
, 0));
4742 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
4743 * so the depth comparison value isn't clamped for Z16 and
4744 * Z24 anymore. Do it manually here.
4746 * It's unnecessary if the original texture format was
4747 * Z32_FLOAT, but we don't know that here.
4749 if (ctx
->ac
.chip_class
== VI
&& ctx
->abi
->clamp_shadow_reference
)
4750 z
= ac_build_clamp(&ctx
->ac
, z
);
4752 address
[count
++] = z
;
4755 /* pack derivatives */
4757 int num_src_deriv_channels
, num_dest_deriv_channels
;
4758 switch (instr
->sampler_dim
) {
4759 case GLSL_SAMPLER_DIM_3D
:
4760 case GLSL_SAMPLER_DIM_CUBE
:
4762 num_src_deriv_channels
= 3;
4763 num_dest_deriv_channels
= 3;
4765 case GLSL_SAMPLER_DIM_2D
:
4767 num_src_deriv_channels
= 2;
4768 num_dest_deriv_channels
= 2;
4771 case GLSL_SAMPLER_DIM_1D
:
4772 num_src_deriv_channels
= 1;
4773 if (ctx
->ac
.chip_class
>= GFX9
) {
4774 num_dest_deriv_channels
= 2;
4777 num_dest_deriv_channels
= 1;
4783 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
4784 derivs
[i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
4785 derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
, ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
4787 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
4788 derivs
[i
] = ctx
->ac
.f32_0
;
4789 derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
4793 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& coord
) {
4794 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4795 coords
[chan
] = ac_to_float(&ctx
->ac
, coords
[chan
]);
4796 if (instr
->coord_components
== 3)
4797 coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
4798 ac_prepare_cube_coords(&ctx
->ac
,
4799 instr
->op
== nir_texop_txd
, instr
->is_array
,
4800 instr
->op
== nir_texop_lod
, coords
, derivs
);
4806 for (unsigned i
= 0; i
< num_deriv_comp
* 2; i
++)
4807 address
[count
++] = derivs
[i
];
4810 /* Pack texture coordinates */
4812 address
[count
++] = coords
[0];
4813 if (instr
->coord_components
> 1) {
4814 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&& instr
->is_array
&& instr
->op
!= nir_texop_txf
) {
4815 coords
[1] = apply_round_slice(&ctx
->ac
, coords
[1]);
4817 address
[count
++] = coords
[1];
4819 if (instr
->coord_components
> 2) {
4820 /* This seems like a bit of a hack - but it passes Vulkan CTS with it */
4821 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_3D
&&
4822 instr
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
&&
4823 instr
->op
!= nir_texop_txf
) {
4824 coords
[2] = apply_round_slice(&ctx
->ac
, coords
[2]);
4826 address
[count
++] = coords
[2];
4829 if (ctx
->ac
.chip_class
>= GFX9
) {
4830 LLVMValueRef filler
;
4831 if (instr
->op
== nir_texop_txf
)
4832 filler
= ctx
->ac
.i32_0
;
4834 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
4836 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
) {
4837 /* No nir_texop_lod, because it does not take a slice
4838 * even with array textures. */
4839 if (instr
->is_array
&& instr
->op
!= nir_texop_lod
) {
4840 address
[count
] = address
[count
- 1];
4841 address
[count
- 1] = filler
;
4844 address
[count
++] = filler
;
4850 if (lod
&& ((instr
->op
== nir_texop_txl
&& !lod_is_zero
) ||
4851 instr
->op
== nir_texop_txf
)) {
4852 address
[count
++] = lod
;
4853 } else if (instr
->op
== nir_texop_txf_ms
&& sample_index
) {
4854 address
[count
++] = sample_index
;
4855 } else if(instr
->op
== nir_texop_txs
) {
4858 address
[count
++] = lod
;
4860 address
[count
++] = ctx
->ac
.i32_0
;
4863 for (chan
= 0; chan
< count
; chan
++) {
4864 address
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
,
4865 address
[chan
], ctx
->ac
.i32
, "");
4868 if (instr
->op
== nir_texop_samples_identical
) {
4869 LLVMValueRef txf_address
[4];
4870 struct ac_image_args txf_args
= { 0 };
4871 unsigned txf_count
= count
;
4872 memcpy(txf_address
, address
, sizeof(txf_address
));
4874 if (!instr
->is_array
)
4875 txf_address
[2] = ctx
->ac
.i32_0
;
4876 txf_address
[3] = ctx
->ac
.i32_0
;
4878 set_tex_fetch_args(&ctx
->ac
, &txf_args
, instr
, nir_texop_txf
,
4880 txf_address
, txf_count
, 0xf);
4882 result
= build_tex_intrinsic(ctx
, instr
, false, &txf_args
);
4884 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4885 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
4889 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
4890 instr
->op
!= nir_texop_txs
) {
4891 unsigned sample_chan
= instr
->is_array
? 3 : 2;
4892 address
[sample_chan
] = adjust_sample_index_using_fmask(&ctx
->ac
,
4895 instr
->is_array
? address
[2] : NULL
,
4896 address
[sample_chan
],
4900 if (offsets
&& instr
->op
== nir_texop_txf
) {
4901 nir_const_value
*const_offset
=
4902 nir_src_as_const_value(instr
->src
[const_src
].src
);
4903 int num_offsets
= instr
->src
[const_src
].src
.ssa
->num_components
;
4904 assert(const_offset
);
4905 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
4906 if (num_offsets
> 2)
4907 address
[2] = LLVMBuildAdd(ctx
->ac
.builder
,
4908 address
[2], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[2], false), "");
4909 if (num_offsets
> 1)
4910 address
[1] = LLVMBuildAdd(ctx
->ac
.builder
,
4911 address
[1], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[1], false), "");
4912 address
[0] = LLVMBuildAdd(ctx
->ac
.builder
,
4913 address
[0], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[0], false), "");
4917 /* TODO TG4 support */
4918 if (instr
->op
== nir_texop_tg4
) {
4919 if (instr
->is_shadow
)
4922 dmask
= 1 << instr
->component
;
4924 set_tex_fetch_args(&ctx
->ac
, &args
, instr
, instr
->op
,
4925 res_ptr
, samp_ptr
, address
, count
, dmask
);
4927 result
= build_tex_intrinsic(ctx
, instr
, lod_is_zero
, &args
);
4929 if (instr
->op
== nir_texop_query_levels
)
4930 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4931 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
4932 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
4933 instr
->op
!= nir_texop_tg4
)
4934 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4935 else if (instr
->op
== nir_texop_txs
&&
4936 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
4938 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
4939 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
4940 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
4941 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
4942 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
4943 } else if (ctx
->ac
.chip_class
>= GFX9
&&
4944 instr
->op
== nir_texop_txs
&&
4945 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
4947 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
4948 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
4949 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
4951 } else if (instr
->dest
.ssa
.num_components
!= 4)
4952 result
= trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
4956 assert(instr
->dest
.is_ssa
);
4957 result
= ac_to_integer(&ctx
->ac
, result
);
4958 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4963 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
4965 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
4966 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
4968 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4969 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
4972 static void visit_post_phi(struct ac_nir_context
*ctx
,
4973 nir_phi_instr
*instr
,
4974 LLVMValueRef llvm_phi
)
4976 nir_foreach_phi_src(src
, instr
) {
4977 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
4978 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
4980 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
4984 static void phi_post_pass(struct ac_nir_context
*ctx
)
4986 struct hash_entry
*entry
;
4987 hash_table_foreach(ctx
->phis
, entry
) {
4988 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
4989 (LLVMValueRef
)entry
->data
);
4994 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
4995 const nir_ssa_undef_instr
*instr
)
4997 unsigned num_components
= instr
->def
.num_components
;
5000 if (num_components
== 1)
5001 undef
= LLVMGetUndef(ctx
->ac
.i32
);
5003 undef
= LLVMGetUndef(LLVMVectorType(ctx
->ac
.i32
, num_components
));
5005 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, undef
);
5008 static void visit_jump(struct ac_nir_context
*ctx
,
5009 const nir_jump_instr
*instr
)
5011 switch (instr
->type
) {
5012 case nir_jump_break
:
5013 LLVMBuildBr(ctx
->ac
.builder
, ctx
->break_block
);
5014 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5016 case nir_jump_continue
:
5017 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5018 LLVMClearInsertionPosition(ctx
->ac
.builder
);
5021 fprintf(stderr
, "Unknown NIR jump instr: ");
5022 nir_print_instr(&instr
->instr
, stderr
);
5023 fprintf(stderr
, "\n");
5028 static void visit_cf_list(struct ac_nir_context
*ctx
,
5029 struct exec_list
*list
);
5031 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
5033 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
5034 nir_foreach_instr(instr
, block
)
5036 switch (instr
->type
) {
5037 case nir_instr_type_alu
:
5038 visit_alu(ctx
, nir_instr_as_alu(instr
));
5040 case nir_instr_type_load_const
:
5041 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
5043 case nir_instr_type_intrinsic
:
5044 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
5046 case nir_instr_type_tex
:
5047 visit_tex(ctx
, nir_instr_as_tex(instr
));
5049 case nir_instr_type_phi
:
5050 visit_phi(ctx
, nir_instr_as_phi(instr
));
5052 case nir_instr_type_ssa_undef
:
5053 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
5055 case nir_instr_type_jump
:
5056 visit_jump(ctx
, nir_instr_as_jump(instr
));
5059 fprintf(stderr
, "Unknown NIR instr type: ");
5060 nir_print_instr(instr
, stderr
);
5061 fprintf(stderr
, "\n");
5066 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
5069 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
5071 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
5073 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5074 LLVMBasicBlockRef merge_block
=
5075 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5076 LLVMBasicBlockRef if_block
=
5077 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5078 LLVMBasicBlockRef else_block
= merge_block
;
5079 if (!exec_list_is_empty(&if_stmt
->else_list
))
5080 else_block
= LLVMAppendBasicBlockInContext(
5081 ctx
->ac
.context
, fn
, "");
5083 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, value
,
5085 LLVMBuildCondBr(ctx
->ac
.builder
, cond
, if_block
, else_block
);
5087 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, if_block
);
5088 visit_cf_list(ctx
, &if_stmt
->then_list
);
5089 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5090 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5092 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
5093 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, else_block
);
5094 visit_cf_list(ctx
, &if_stmt
->else_list
);
5095 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5096 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
5099 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, merge_block
);
5102 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
5104 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
5105 LLVMBasicBlockRef continue_parent
= ctx
->continue_block
;
5106 LLVMBasicBlockRef break_parent
= ctx
->break_block
;
5108 ctx
->continue_block
=
5109 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5111 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
5113 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5114 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->continue_block
);
5115 visit_cf_list(ctx
, &loop
->body
);
5117 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
5118 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
5119 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->break_block
);
5121 ctx
->continue_block
= continue_parent
;
5122 ctx
->break_block
= break_parent
;
5125 static void visit_cf_list(struct ac_nir_context
*ctx
,
5126 struct exec_list
*list
)
5128 foreach_list_typed(nir_cf_node
, node
, node
, list
)
5130 switch (node
->type
) {
5131 case nir_cf_node_block
:
5132 visit_block(ctx
, nir_cf_node_as_block(node
));
5135 case nir_cf_node_if
:
5136 visit_if(ctx
, nir_cf_node_as_if(node
));
5139 case nir_cf_node_loop
:
5140 visit_loop(ctx
, nir_cf_node_as_loop(node
));
5150 handle_vs_input_decl(struct nir_to_llvm_context
*ctx
,
5151 struct nir_variable
*variable
)
5153 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
5154 LLVMValueRef t_offset
;
5155 LLVMValueRef t_list
;
5157 LLVMValueRef buffer_index
;
5158 int index
= variable
->data
.location
- VERT_ATTRIB_GENERIC0
;
5159 int idx
= variable
->data
.location
;
5160 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
5162 variable
->data
.driver_location
= idx
* 4;
5164 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << index
)) {
5165 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.instance_id
,
5166 ctx
->abi
.start_instance
, "");
5167 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(3,
5168 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5170 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.vertex_id
,
5171 ctx
->abi
.base_vertex
, "");
5173 for (unsigned i
= 0; i
< attrib_count
; ++i
, ++idx
) {
5174 t_offset
= LLVMConstInt(ctx
->ac
.i32
, index
+ i
, false);
5176 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
5178 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
5183 for (unsigned chan
= 0; chan
< 4; chan
++) {
5184 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5185 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
, chan
)] =
5186 ac_to_integer(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
,
5187 input
, llvm_chan
, ""));
5192 static void interp_fs_input(struct nir_to_llvm_context
*ctx
,
5194 LLVMValueRef interp_param
,
5195 LLVMValueRef prim_mask
,
5196 LLVMValueRef result
[4])
5198 LLVMValueRef attr_number
;
5201 bool interp
= interp_param
!= NULL
;
5203 attr_number
= LLVMConstInt(ctx
->ac
.i32
, attr
, false);
5205 /* fs.constant returns the param from the middle vertex, so it's not
5206 * really useful for flat shading. It's meant to be used for custom
5207 * interpolation (but the intrinsic can't fetch from the other two
5210 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
5211 * to do the right thing. The only reason we use fs.constant is that
5212 * fs.interp cannot be used on integers, because they can be equal
5216 interp_param
= LLVMBuildBitCast(ctx
->builder
, interp_param
,
5219 i
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
5221 j
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
5225 for (chan
= 0; chan
< 4; chan
++) {
5226 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
5229 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
5234 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
5235 LLVMConstInt(ctx
->ac
.i32
, 2, false),
5244 handle_fs_input_decl(struct nir_to_llvm_context
*ctx
,
5245 struct nir_variable
*variable
)
5247 int idx
= variable
->data
.location
;
5248 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5249 LLVMValueRef interp
;
5251 variable
->data
.driver_location
= idx
* 4;
5252 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
5254 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
5255 unsigned interp_type
;
5256 if (variable
->data
.sample
) {
5257 interp_type
= INTERP_SAMPLE
;
5258 ctx
->shader_info
->info
.ps
.force_persample
= true;
5259 } else if (variable
->data
.centroid
)
5260 interp_type
= INTERP_CENTROID
;
5262 interp_type
= INTERP_CENTER
;
5264 interp
= lookup_interp_param(ctx
, variable
->data
.interpolation
, interp_type
);
5268 for (unsigned i
= 0; i
< attrib_count
; ++i
)
5269 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
5274 handle_vs_inputs(struct nir_to_llvm_context
*ctx
,
5275 struct nir_shader
*nir
) {
5276 nir_foreach_variable(variable
, &nir
->inputs
)
5277 handle_vs_input_decl(ctx
, variable
);
5281 prepare_interp_optimize(struct nir_to_llvm_context
*ctx
,
5282 struct nir_shader
*nir
)
5284 if (!ctx
->options
->key
.fs
.multisample
)
5287 bool uses_center
= false;
5288 bool uses_centroid
= false;
5289 nir_foreach_variable(variable
, &nir
->inputs
) {
5290 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
5291 variable
->data
.sample
)
5294 if (variable
->data
.centroid
)
5295 uses_centroid
= true;
5300 if (uses_center
&& uses_centroid
) {
5301 LLVMValueRef sel
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, ctx
->prim_mask
, ctx
->ac
.i32_0
, "");
5302 ctx
->persp_centroid
= LLVMBuildSelect(ctx
->builder
, sel
, ctx
->persp_center
, ctx
->persp_centroid
, "");
5303 ctx
->linear_centroid
= LLVMBuildSelect(ctx
->builder
, sel
, ctx
->linear_center
, ctx
->linear_centroid
, "");
5308 handle_fs_inputs(struct nir_to_llvm_context
*ctx
,
5309 struct nir_shader
*nir
)
5311 prepare_interp_optimize(ctx
, nir
);
5313 nir_foreach_variable(variable
, &nir
->inputs
)
5314 handle_fs_input_decl(ctx
, variable
);
5318 if (ctx
->shader_info
->info
.ps
.uses_input_attachments
||
5319 ctx
->shader_info
->info
.needs_multiview_view_index
)
5320 ctx
->input_mask
|= 1ull << VARYING_SLOT_LAYER
;
5322 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
5323 LLVMValueRef interp_param
;
5324 LLVMValueRef
*inputs
= ctx
->inputs
+radeon_llvm_reg_index_soa(i
, 0);
5326 if (!(ctx
->input_mask
& (1ull << i
)))
5329 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
5330 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
5331 interp_param
= *inputs
;
5332 interp_fs_input(ctx
, index
, interp_param
, ctx
->prim_mask
,
5336 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
5338 } else if (i
== VARYING_SLOT_POS
) {
5339 for(int i
= 0; i
< 3; ++i
)
5340 inputs
[i
] = ctx
->abi
.frag_pos
[i
];
5342 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
5343 ctx
->abi
.frag_pos
[3]);
5346 ctx
->shader_info
->fs
.num_interp
= index
;
5347 if (ctx
->input_mask
& (1 << VARYING_SLOT_PNTC
))
5348 ctx
->shader_info
->fs
.has_pcoord
= true;
5349 if (ctx
->input_mask
& (1 << VARYING_SLOT_PRIMITIVE_ID
))
5350 ctx
->shader_info
->fs
.prim_id_input
= true;
5351 if (ctx
->input_mask
& (1 << VARYING_SLOT_LAYER
))
5352 ctx
->shader_info
->fs
.layer_input
= true;
5353 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
5355 if (ctx
->shader_info
->info
.needs_multiview_view_index
)
5356 ctx
->view_index
= ctx
->inputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5360 ac_build_alloca(struct ac_llvm_context
*ac
,
5364 LLVMBuilderRef builder
= ac
->builder
;
5365 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
5366 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
5367 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
5368 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
5369 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
5373 LLVMPositionBuilderBefore(first_builder
, first_instr
);
5375 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
5378 res
= LLVMBuildAlloca(first_builder
, type
, name
);
5379 LLVMBuildStore(builder
, LLVMConstNull(type
), res
);
5381 LLVMDisposeBuilder(first_builder
);
5386 static LLVMValueRef
si_build_alloca_undef(struct ac_llvm_context
*ac
,
5390 LLVMValueRef ptr
= ac_build_alloca(ac
, type
, name
);
5391 LLVMBuildStore(ac
->builder
, LLVMGetUndef(type
), ptr
);
5396 scan_shader_output_decl(struct nir_to_llvm_context
*ctx
,
5397 struct nir_variable
*variable
,
5398 struct nir_shader
*shader
,
5399 gl_shader_stage stage
)
5401 int idx
= variable
->data
.location
+ variable
->data
.index
;
5402 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5403 uint64_t mask_attribs
;
5405 variable
->data
.driver_location
= idx
* 4;
5407 /* tess ctrl has it's own load/store paths for outputs */
5408 if (stage
== MESA_SHADER_TESS_CTRL
)
5411 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
5412 if (stage
== MESA_SHADER_VERTEX
||
5413 stage
== MESA_SHADER_TESS_EVAL
||
5414 stage
== MESA_SHADER_GEOMETRY
) {
5415 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5416 int length
= shader
->info
.clip_distance_array_size
+
5417 shader
->info
.cull_distance_array_size
;
5418 if (stage
== MESA_SHADER_VERTEX
) {
5419 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5420 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5422 if (stage
== MESA_SHADER_TESS_EVAL
) {
5423 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << shader
->info
.clip_distance_array_size
) - 1;
5424 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << shader
->info
.cull_distance_array_size
) - 1;
5431 mask_attribs
= 1ull << idx
;
5435 ctx
->output_mask
|= mask_attribs
;
5439 handle_shader_output_decl(struct ac_nir_context
*ctx
,
5440 struct nir_shader
*nir
,
5441 struct nir_variable
*variable
)
5443 unsigned output_loc
= variable
->data
.driver_location
/ 4;
5444 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5446 /* tess ctrl has it's own load/store paths for outputs */
5447 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
5450 if (ctx
->stage
== MESA_SHADER_VERTEX
||
5451 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
5452 ctx
->stage
== MESA_SHADER_GEOMETRY
) {
5453 int idx
= variable
->data
.location
+ variable
->data
.index
;
5454 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5455 int length
= nir
->info
.clip_distance_array_size
+
5456 nir
->info
.cull_distance_array_size
;
5465 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
5466 for (unsigned chan
= 0; chan
< 4; chan
++) {
5467 ctx
->outputs
[radeon_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
5468 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5474 glsl_base_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5475 enum glsl_base_type type
)
5479 case GLSL_TYPE_UINT
:
5480 case GLSL_TYPE_BOOL
:
5481 case GLSL_TYPE_SUBROUTINE
:
5483 case GLSL_TYPE_FLOAT
: /* TODO handle mediump */
5485 case GLSL_TYPE_INT64
:
5486 case GLSL_TYPE_UINT64
:
5488 case GLSL_TYPE_DOUBLE
:
5491 unreachable("unknown GLSL type");
5496 glsl_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5497 const struct glsl_type
*type
)
5499 if (glsl_type_is_scalar(type
)) {
5500 return glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
));
5503 if (glsl_type_is_vector(type
)) {
5504 return LLVMVectorType(
5505 glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
)),
5506 glsl_get_vector_elements(type
));
5509 if (glsl_type_is_matrix(type
)) {
5510 return LLVMArrayType(
5511 glsl_to_llvm_type(ctx
, glsl_get_column_type(type
)),
5512 glsl_get_matrix_columns(type
));
5515 if (glsl_type_is_array(type
)) {
5516 return LLVMArrayType(
5517 glsl_to_llvm_type(ctx
, glsl_get_array_element(type
)),
5518 glsl_get_length(type
));
5521 assert(glsl_type_is_struct(type
));
5523 LLVMTypeRef member_types
[glsl_get_length(type
)];
5525 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
5527 glsl_to_llvm_type(ctx
,
5528 glsl_get_struct_field(type
, i
));
5531 return LLVMStructTypeInContext(ctx
->context
, member_types
,
5532 glsl_get_length(type
), false);
5536 setup_locals(struct ac_nir_context
*ctx
,
5537 struct nir_function
*func
)
5540 ctx
->num_locals
= 0;
5541 nir_foreach_variable(variable
, &func
->impl
->locals
) {
5542 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5543 variable
->data
.driver_location
= ctx
->num_locals
* 4;
5544 ctx
->num_locals
+= attrib_count
;
5546 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
5550 for (i
= 0; i
< ctx
->num_locals
; i
++) {
5551 for (j
= 0; j
< 4; j
++) {
5552 ctx
->locals
[i
* 4 + j
] =
5553 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
5559 setup_shared(struct ac_nir_context
*ctx
,
5560 struct nir_shader
*nir
)
5562 nir_foreach_variable(variable
, &nir
->shared
) {
5563 LLVMValueRef shared
=
5564 LLVMAddGlobalInAddressSpace(
5565 ctx
->ac
.module
, glsl_to_llvm_type(ctx
->nctx
, variable
->type
),
5566 variable
->name
? variable
->name
: "",
5568 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
5573 emit_float_saturate(struct ac_llvm_context
*ctx
, LLVMValueRef v
, float lo
, float hi
)
5575 v
= ac_to_float(ctx
, v
);
5576 v
= emit_intrin_2f_param(ctx
, "llvm.maxnum", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, lo
));
5577 return emit_intrin_2f_param(ctx
, "llvm.minnum", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, hi
));
5581 static LLVMValueRef
emit_pack_int16(struct nir_to_llvm_context
*ctx
,
5582 LLVMValueRef src0
, LLVMValueRef src1
)
5584 LLVMValueRef const16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
5585 LLVMValueRef comp
[2];
5587 comp
[0] = LLVMBuildAnd(ctx
->builder
, src0
, LLVMConstInt(ctx
->ac
.i32
, 65535, 0), "");
5588 comp
[1] = LLVMBuildAnd(ctx
->builder
, src1
, LLVMConstInt(ctx
->ac
.i32
, 65535, 0), "");
5589 comp
[1] = LLVMBuildShl(ctx
->builder
, comp
[1], const16
, "");
5590 return LLVMBuildOr(ctx
->builder
, comp
[0], comp
[1], "");
5593 /* Initialize arguments for the shader export intrinsic */
5595 si_llvm_init_export_args(struct nir_to_llvm_context
*ctx
,
5596 LLVMValueRef
*values
,
5598 struct ac_export_args
*args
)
5600 /* Default is 0xf. Adjusted below depending on the format. */
5601 args
->enabled_channels
= 0xf;
5603 /* Specify whether the EXEC mask represents the valid mask */
5604 args
->valid_mask
= 0;
5606 /* Specify whether this is the last export */
5609 /* Specify the target we are exporting */
5610 args
->target
= target
;
5612 args
->compr
= false;
5613 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
5614 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
5615 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
5616 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
5621 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
5622 LLVMValueRef val
[4];
5623 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
5624 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
5625 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
5626 bool is_int10
= (ctx
->options
->key
.fs
.is_int10
>> index
) & 1;
5628 switch(col_format
) {
5629 case V_028714_SPI_SHADER_ZERO
:
5630 args
->enabled_channels
= 0; /* writemask */
5631 args
->target
= V_008DFC_SQ_EXP_NULL
;
5634 case V_028714_SPI_SHADER_32_R
:
5635 args
->enabled_channels
= 1;
5636 args
->out
[0] = values
[0];
5639 case V_028714_SPI_SHADER_32_GR
:
5640 args
->enabled_channels
= 0x3;
5641 args
->out
[0] = values
[0];
5642 args
->out
[1] = values
[1];
5645 case V_028714_SPI_SHADER_32_AR
:
5646 args
->enabled_channels
= 0x9;
5647 args
->out
[0] = values
[0];
5648 args
->out
[3] = values
[3];
5651 case V_028714_SPI_SHADER_FP16_ABGR
:
5654 for (unsigned chan
= 0; chan
< 2; chan
++) {
5655 LLVMValueRef pack_args
[2] = {
5657 values
[2 * chan
+ 1]
5659 LLVMValueRef packed
;
5661 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
5662 args
->out
[chan
] = packed
;
5666 case V_028714_SPI_SHADER_UNORM16_ABGR
:
5667 for (unsigned chan
= 0; chan
< 4; chan
++) {
5668 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
5669 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5670 LLVMConstReal(ctx
->ac
.f32
, 65535), "");
5671 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5672 LLVMConstReal(ctx
->ac
.f32
, 0.5), "");
5673 val
[chan
] = LLVMBuildFPToUI(ctx
->builder
, val
[chan
],
5678 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5679 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5682 case V_028714_SPI_SHADER_SNORM16_ABGR
:
5683 for (unsigned chan
= 0; chan
< 4; chan
++) {
5684 val
[chan
] = emit_float_saturate(&ctx
->ac
, values
[chan
], -1, 1);
5685 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5686 LLVMConstReal(ctx
->ac
.f32
, 32767), "");
5688 /* If positive, add 0.5, else add -0.5. */
5689 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5690 LLVMBuildSelect(ctx
->builder
,
5691 LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
,
5692 val
[chan
], ctx
->ac
.f32_0
, ""),
5693 LLVMConstReal(ctx
->ac
.f32
, 0.5),
5694 LLVMConstReal(ctx
->ac
.f32
, -0.5), ""), "");
5695 val
[chan
] = LLVMBuildFPToSI(ctx
->builder
, val
[chan
], ctx
->ac
.i32
, "");
5699 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5700 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5703 case V_028714_SPI_SHADER_UINT16_ABGR
: {
5704 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5705 is_int8
? 255 : is_int10
? 1023 : 65535, 0);
5706 LLVMValueRef max_alpha
= !is_int10
? max_rgb
: LLVMConstInt(ctx
->ac
.i32
, 3, 0);
5708 for (unsigned chan
= 0; chan
< 4; chan
++) {
5709 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
5710 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntULT
, val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
5714 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5715 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5719 case V_028714_SPI_SHADER_SINT16_ABGR
: {
5720 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5721 is_int8
? 127 : is_int10
? 511 : 32767, 0);
5722 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->ac
.i32
,
5723 is_int8
? -128 : is_int10
? -512 : -32768, 0);
5724 LLVMValueRef max_alpha
= !is_int10
? max_rgb
: ctx
->ac
.i32_1
;
5725 LLVMValueRef min_alpha
= !is_int10
? min_rgb
: LLVMConstInt(ctx
->ac
.i32
, -2, 0);
5728 for (unsigned chan
= 0; chan
< 4; chan
++) {
5729 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
5730 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
5731 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, val
[chan
], chan
== 3 ? min_alpha
: min_rgb
);
5735 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5736 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5741 case V_028714_SPI_SHADER_32_ABGR
:
5742 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5746 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5748 for (unsigned i
= 0; i
< 4; ++i
)
5749 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
5753 handle_vs_outputs_post(struct nir_to_llvm_context
*ctx
,
5754 bool export_prim_id
,
5755 struct ac_vs_output_info
*outinfo
)
5757 uint32_t param_count
= 0;
5759 unsigned pos_idx
, num_pos_exports
= 0;
5760 struct ac_export_args args
, pos_args
[4] = {};
5761 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
5764 if (ctx
->options
->key
.has_multiview_view_index
) {
5765 LLVMValueRef
* tmp_out
= &ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
5767 for(unsigned i
= 0; i
< 4; ++i
)
5768 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
5769 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5772 LLVMBuildStore(ctx
->builder
, ac_to_float(&ctx
->ac
, ctx
->view_index
), *tmp_out
);
5773 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
5776 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
5777 sizeof(outinfo
->vs_output_param_offset
));
5779 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
5780 LLVMValueRef slots
[8];
5783 if (outinfo
->cull_dist_mask
)
5784 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
5786 i
= VARYING_SLOT_CLIP_DIST0
;
5787 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
5788 slots
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5789 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5791 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
5792 slots
[i
] = LLVMGetUndef(ctx
->ac
.f32
);
5794 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
5795 target
= V_008DFC_SQ_EXP_POS
+ 3;
5796 si_llvm_init_export_args(ctx
, &slots
[4], target
, &args
);
5797 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5798 &args
, sizeof(args
));
5801 target
= V_008DFC_SQ_EXP_POS
+ 2;
5802 si_llvm_init_export_args(ctx
, &slots
[0], target
, &args
);
5803 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5804 &args
, sizeof(args
));
5808 LLVMValueRef pos_values
[4] = {ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_0
, ctx
->ac
.f32_1
};
5809 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
5810 for (unsigned j
= 0; j
< 4; j
++)
5811 pos_values
[j
] = LLVMBuildLoad(ctx
->builder
,
5812 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_POS
, j
)], "");
5814 si_llvm_init_export_args(ctx
, pos_values
, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
5816 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
5817 outinfo
->writes_pointsize
= true;
5818 psize_value
= LLVMBuildLoad(ctx
->builder
,
5819 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_PSIZ
, 0)], "");
5822 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
5823 outinfo
->writes_layer
= true;
5824 layer_value
= LLVMBuildLoad(ctx
->builder
,
5825 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)], "");
5828 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
5829 outinfo
->writes_viewport_index
= true;
5830 viewport_index_value
= LLVMBuildLoad(ctx
->builder
,
5831 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_VIEWPORT
, 0)], "");
5834 if (outinfo
->writes_pointsize
||
5835 outinfo
->writes_layer
||
5836 outinfo
->writes_viewport_index
) {
5837 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
5838 (outinfo
->writes_layer
== true ? 4 : 0));
5839 pos_args
[1].valid_mask
= 0;
5840 pos_args
[1].done
= 0;
5841 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
5842 pos_args
[1].compr
= 0;
5843 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
5844 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
5845 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
5846 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
5848 if (outinfo
->writes_pointsize
== true)
5849 pos_args
[1].out
[0] = psize_value
;
5850 if (outinfo
->writes_layer
== true)
5851 pos_args
[1].out
[2] = layer_value
;
5852 if (outinfo
->writes_viewport_index
== true) {
5853 if (ctx
->options
->chip_class
>= GFX9
) {
5854 /* GFX9 has the layer in out.z[10:0] and the viewport
5855 * index in out.z[19:16].
5857 LLVMValueRef v
= viewport_index_value
;
5858 v
= ac_to_integer(&ctx
->ac
, v
);
5859 v
= LLVMBuildShl(ctx
->builder
, v
,
5860 LLVMConstInt(ctx
->ac
.i32
, 16, false),
5862 v
= LLVMBuildOr(ctx
->builder
, v
,
5863 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
5865 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
5866 pos_args
[1].enabled_channels
|= 1 << 2;
5868 pos_args
[1].out
[3] = viewport_index_value
;
5869 pos_args
[1].enabled_channels
|= 1 << 3;
5873 for (i
= 0; i
< 4; i
++) {
5874 if (pos_args
[i
].out
[0])
5879 for (i
= 0; i
< 4; i
++) {
5880 if (!pos_args
[i
].out
[0])
5883 /* Specify the target we are exporting */
5884 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
5885 if (pos_idx
== num_pos_exports
)
5886 pos_args
[i
].done
= 1;
5887 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
5890 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5891 LLVMValueRef values
[4];
5892 if (!(ctx
->output_mask
& (1ull << i
)))
5895 for (unsigned j
= 0; j
< 4; j
++)
5896 values
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5897 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5899 if (i
== VARYING_SLOT_LAYER
) {
5900 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5901 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
;
5903 } else if (i
== VARYING_SLOT_PRIMITIVE_ID
) {
5904 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5905 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
5907 } else if (i
>= VARYING_SLOT_VAR0
) {
5908 outinfo
->export_mask
|= 1u << (i
- VARYING_SLOT_VAR0
);
5909 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5910 outinfo
->vs_output_param_offset
[i
] = param_count
;
5915 si_llvm_init_export_args(ctx
, values
, target
, &args
);
5917 if (target
>= V_008DFC_SQ_EXP_POS
&&
5918 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
5919 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5920 &args
, sizeof(args
));
5922 ac_build_export(&ctx
->ac
, &args
);
5926 if (export_prim_id
) {
5927 LLVMValueRef values
[4];
5928 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5929 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
5932 values
[0] = ctx
->vs_prim_id
;
5933 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
5934 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5935 for (unsigned j
= 1; j
< 4; j
++)
5936 values
[j
] = ctx
->ac
.f32_0
;
5937 si_llvm_init_export_args(ctx
, values
, target
, &args
);
5938 ac_build_export(&ctx
->ac
, &args
);
5939 outinfo
->export_prim_id
= true;
5942 outinfo
->pos_exports
= num_pos_exports
;
5943 outinfo
->param_exports
= param_count
;
5947 handle_es_outputs_post(struct nir_to_llvm_context
*ctx
,
5948 struct ac_es_output_info
*outinfo
)
5951 uint64_t max_output_written
= 0;
5952 LLVMValueRef lds_base
= NULL
;
5954 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5958 if (!(ctx
->output_mask
& (1ull << i
)))
5961 if (i
== VARYING_SLOT_CLIP_DIST0
)
5962 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
5964 param_index
= shader_io_get_unique_index(i
);
5966 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
5969 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
5971 if (ctx
->ac
.chip_class
>= GFX9
) {
5972 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
5973 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
5974 LLVMValueRef wave_idx
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
5975 LLVMConstInt(ctx
->ac
.i32
, 24, false),
5976 LLVMConstInt(ctx
->ac
.i32
, 4, false), false);
5977 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
5978 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
5979 LLVMConstInt(ctx
->ac
.i32
, 64, false), ""), "");
5980 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
5981 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
5984 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5985 LLVMValueRef dw_addr
;
5986 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
5990 if (!(ctx
->output_mask
& (1ull << i
)))
5993 if (i
== VARYING_SLOT_CLIP_DIST0
)
5994 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
5996 param_index
= shader_io_get_unique_index(i
);
5999 dw_addr
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6000 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
6003 for (j
= 0; j
< length
; j
++) {
6004 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], "");
6005 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->ac
.i32
, "");
6007 if (ctx
->ac
.chip_class
>= GFX9
) {
6008 ac_lds_store(&ctx
->ac
, dw_addr
,
6009 LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], ""));
6010 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6012 ac_build_buffer_store_dword(&ctx
->ac
,
6015 NULL
, ctx
->es2gs_offset
,
6016 (4 * param_index
+ j
) * 4,
6024 handle_ls_outputs_post(struct nir_to_llvm_context
*ctx
)
6026 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
6027 LLVMValueRef vertex_dw_stride
= unpack_param(&ctx
->ac
, ctx
->ls_out_layout
, 13, 8);
6028 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->builder
, vertex_id
,
6029 vertex_dw_stride
, "");
6031 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6032 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
6035 if (!(ctx
->output_mask
& (1ull << i
)))
6038 if (i
== VARYING_SLOT_CLIP_DIST0
)
6039 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6040 int param
= shader_io_get_unique_index(i
);
6041 mark_tess_output(ctx
, false, param
);
6043 mark_tess_output(ctx
, false, param
+ 1);
6044 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->builder
, base_dw_addr
,
6045 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
6047 for (unsigned j
= 0; j
< length
; j
++) {
6048 ac_lds_store(&ctx
->ac
, dw_addr
,
6049 LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], ""));
6050 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
, ctx
->ac
.i32_1
, "");
6055 struct ac_build_if_state
6057 struct nir_to_llvm_context
*ctx
;
6058 LLVMValueRef condition
;
6059 LLVMBasicBlockRef entry_block
;
6060 LLVMBasicBlockRef true_block
;
6061 LLVMBasicBlockRef false_block
;
6062 LLVMBasicBlockRef merge_block
;
6065 static LLVMBasicBlockRef
6066 ac_build_insert_new_block(struct nir_to_llvm_context
*ctx
, const char *name
)
6068 LLVMBasicBlockRef current_block
;
6069 LLVMBasicBlockRef next_block
;
6070 LLVMBasicBlockRef new_block
;
6072 /* get current basic block */
6073 current_block
= LLVMGetInsertBlock(ctx
->builder
);
6075 /* chqeck if there's another block after this one */
6076 next_block
= LLVMGetNextBasicBlock(current_block
);
6078 /* insert the new block before the next block */
6079 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
6082 /* append new block after current block */
6083 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
6084 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
6090 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
6091 struct nir_to_llvm_context
*ctx
,
6092 LLVMValueRef condition
)
6094 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->builder
);
6096 memset(ifthen
, 0, sizeof *ifthen
);
6098 ifthen
->condition
= condition
;
6099 ifthen
->entry_block
= block
;
6101 /* create endif/merge basic block for the phi functions */
6102 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
6104 /* create/insert true_block before merge_block */
6105 ifthen
->true_block
=
6106 LLVMInsertBasicBlockInContext(ctx
->context
,
6107 ifthen
->merge_block
,
6110 /* successive code goes into the true block */
6111 LLVMPositionBuilderAtEnd(ctx
->builder
, ifthen
->true_block
);
6115 * End a conditional.
6118 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
6120 LLVMBuilderRef builder
= ifthen
->ctx
->builder
;
6122 /* Insert branch to the merge block from current block */
6123 LLVMBuildBr(builder
, ifthen
->merge_block
);
6126 * Now patch in the various branch instructions.
6129 /* Insert the conditional branch instruction at the end of entry_block */
6130 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
6131 if (ifthen
->false_block
) {
6132 /* we have an else clause */
6133 LLVMBuildCondBr(builder
, ifthen
->condition
,
6134 ifthen
->true_block
, ifthen
->false_block
);
6137 /* no else clause */
6138 LLVMBuildCondBr(builder
, ifthen
->condition
,
6139 ifthen
->true_block
, ifthen
->merge_block
);
6142 /* Resume building code at end of the ifthen->merge_block */
6143 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
6147 write_tess_factors(struct nir_to_llvm_context
*ctx
)
6149 unsigned stride
, outer_comps
, inner_comps
;
6150 struct ac_build_if_state if_ctx
, inner_if_ctx
;
6151 LLVMValueRef invocation_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 8, 5);
6152 LLVMValueRef rel_patch_id
= unpack_param(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
, 0, 8);
6153 unsigned tess_inner_index
, tess_outer_index
;
6154 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
6155 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
6159 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
6179 ac_nir_build_if(&if_ctx
, ctx
,
6180 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
6181 invocation_id
, ctx
->ac
.i32_0
, ""));
6183 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6184 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6186 mark_tess_output(ctx
, true, tess_inner_index
);
6187 mark_tess_output(ctx
, true, tess_outer_index
);
6188 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
6189 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6190 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
6191 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_base
,
6192 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
6194 for (i
= 0; i
< 4; i
++) {
6195 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6196 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
6200 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
6201 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
6202 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
6204 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
6206 for (i
= 0; i
< outer_comps
; i
++) {
6208 ac_lds_load(&ctx
->ac
, lds_outer
);
6209 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
6212 for (i
= 0; i
< inner_comps
; i
++) {
6213 inner
[i
] = out
[outer_comps
+i
] =
6214 ac_lds_load(&ctx
->ac
, lds_inner
);
6215 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_inner
,
6220 /* Convert the outputs to vectors for stores. */
6221 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
6225 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
6228 buffer
= ctx
->hs_ring_tess_factor
;
6229 tf_base
= ctx
->tess_factor_offset
;
6230 byteoffset
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
6231 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
6232 unsigned tf_offset
= 0;
6234 if (ctx
->options
->chip_class
<= VI
) {
6235 ac_nir_build_if(&inner_if_ctx
, ctx
,
6236 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
6237 rel_patch_id
, ctx
->ac
.i32_0
, ""));
6239 /* Store the dynamic HS control word. */
6240 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
6241 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
6242 1, ctx
->ac
.i32_0
, tf_base
,
6243 0, 1, 0, true, false);
6246 ac_nir_build_endif(&inner_if_ctx
);
6249 /* Store the tessellation factors. */
6250 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
6251 MIN2(stride
, 4), byteoffset
, tf_base
,
6252 tf_offset
, 1, 0, true, false);
6254 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
6255 stride
- 4, byteoffset
, tf_base
,
6256 16 + tf_offset
, 1, 0, true, false);
6258 //store to offchip for TES to read - only if TES reads them
6259 if (ctx
->options
->key
.tcs
.tes_reads_tess_factors
) {
6260 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
6261 LLVMValueRef tf_inner_offset
;
6262 unsigned param_outer
, param_inner
;
6264 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
6265 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6266 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
6268 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
6269 util_next_power_of_two(outer_comps
));
6271 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
6272 outer_comps
, tf_outer_offset
,
6273 ctx
->oc_lds
, 0, 1, 0, true, false);
6275 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
6276 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
6277 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
6279 inner_vec
= inner_comps
== 1 ? inner
[0] :
6280 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
6281 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
6282 inner_comps
, tf_inner_offset
,
6283 ctx
->oc_lds
, 0, 1, 0, true, false);
6286 ac_nir_build_endif(&if_ctx
);
6290 handle_tcs_outputs_post(struct nir_to_llvm_context
*ctx
)
6292 write_tess_factors(ctx
);
6296 si_export_mrt_color(struct nir_to_llvm_context
*ctx
,
6297 LLVMValueRef
*color
, unsigned param
, bool is_last
,
6298 struct ac_export_args
*args
)
6301 si_llvm_init_export_args(ctx
, color
, param
,
6305 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
6306 args
->done
= 1; /* DONE bit */
6307 } else if (!args
->enabled_channels
)
6308 return false; /* unnecessary NULL export */
6314 radv_export_mrt_z(struct nir_to_llvm_context
*ctx
,
6315 LLVMValueRef depth
, LLVMValueRef stencil
,
6316 LLVMValueRef samplemask
)
6318 struct ac_export_args args
;
6320 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
6322 ac_build_export(&ctx
->ac
, &args
);
6326 handle_fs_outputs_post(struct nir_to_llvm_context
*ctx
)
6329 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
6330 struct ac_export_args color_args
[8];
6332 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6333 LLVMValueRef values
[4];
6335 if (!(ctx
->output_mask
& (1ull << i
)))
6338 if (i
== FRAG_RESULT_DEPTH
) {
6339 ctx
->shader_info
->fs
.writes_z
= true;
6340 depth
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6341 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6342 } else if (i
== FRAG_RESULT_STENCIL
) {
6343 ctx
->shader_info
->fs
.writes_stencil
= true;
6344 stencil
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6345 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6346 } else if (i
== FRAG_RESULT_SAMPLE_MASK
) {
6347 ctx
->shader_info
->fs
.writes_sample_mask
= true;
6348 samplemask
= ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6349 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
6352 for (unsigned j
= 0; j
< 4; j
++)
6353 values
[j
] = ac_to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
6354 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
6356 if (!ctx
->shader_info
->fs
.writes_z
&& !ctx
->shader_info
->fs
.writes_stencil
&& !ctx
->shader_info
->fs
.writes_sample_mask
)
6357 last
= ctx
->output_mask
<= ((1ull << (i
+ 1)) - 1);
6359 bool ret
= si_export_mrt_color(ctx
, values
, V_008DFC_SQ_EXP_MRT
+ (i
- FRAG_RESULT_DATA0
), last
, &color_args
[index
]);
6365 for (unsigned i
= 0; i
< index
; i
++)
6366 ac_build_export(&ctx
->ac
, &color_args
[i
]);
6367 if (depth
|| stencil
|| samplemask
)
6368 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
6370 si_export_mrt_color(ctx
, NULL
, V_008DFC_SQ_EXP_NULL
, true, &color_args
[0]);
6371 ac_build_export(&ctx
->ac
, &color_args
[0]);
6374 ctx
->shader_info
->fs
.output_mask
= index
? ((1ull << index
) - 1) : 0;
6378 emit_gs_epilogue(struct nir_to_llvm_context
*ctx
)
6380 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
6384 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
6385 LLVMValueRef
*addrs
)
6387 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
6389 switch (ctx
->stage
) {
6390 case MESA_SHADER_VERTEX
:
6391 if (ctx
->options
->key
.vs
.as_ls
)
6392 handle_ls_outputs_post(ctx
);
6393 else if (ctx
->options
->key
.vs
.as_es
)
6394 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
6396 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
6397 &ctx
->shader_info
->vs
.outinfo
);
6399 case MESA_SHADER_FRAGMENT
:
6400 handle_fs_outputs_post(ctx
);
6402 case MESA_SHADER_GEOMETRY
:
6403 emit_gs_epilogue(ctx
);
6405 case MESA_SHADER_TESS_CTRL
:
6406 handle_tcs_outputs_post(ctx
);
6408 case MESA_SHADER_TESS_EVAL
:
6409 if (ctx
->options
->key
.tes
.as_es
)
6410 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
6412 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
6413 &ctx
->shader_info
->tes
.outinfo
);
6420 static void ac_llvm_finalize_module(struct nir_to_llvm_context
* ctx
)
6422 LLVMPassManagerRef passmgr
;
6423 /* Create the pass manager */
6424 passmgr
= LLVMCreateFunctionPassManagerForModule(
6427 /* This pass should eliminate all the load and store instructions */
6428 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
6430 /* Add some optimization passes */
6431 LLVMAddScalarReplAggregatesPass(passmgr
);
6432 LLVMAddLICMPass(passmgr
);
6433 LLVMAddAggressiveDCEPass(passmgr
);
6434 LLVMAddCFGSimplificationPass(passmgr
);
6435 LLVMAddInstructionCombiningPass(passmgr
);
6438 LLVMInitializeFunctionPassManager(passmgr
);
6439 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
6440 LLVMFinalizeFunctionPassManager(passmgr
);
6442 LLVMDisposeBuilder(ctx
->builder
);
6443 LLVMDisposePassManager(passmgr
);
6447 ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context
*ctx
)
6449 struct ac_vs_output_info
*outinfo
;
6451 switch (ctx
->stage
) {
6452 case MESA_SHADER_FRAGMENT
:
6453 case MESA_SHADER_COMPUTE
:
6454 case MESA_SHADER_TESS_CTRL
:
6455 case MESA_SHADER_GEOMETRY
:
6457 case MESA_SHADER_VERTEX
:
6458 if (ctx
->options
->key
.vs
.as_ls
||
6459 ctx
->options
->key
.vs
.as_es
)
6461 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
6463 case MESA_SHADER_TESS_EVAL
:
6464 if (ctx
->options
->key
.vs
.as_es
)
6466 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
6469 unreachable("Unhandled shader type");
6472 ac_optimize_vs_outputs(&ctx
->ac
,
6474 outinfo
->vs_output_param_offset
,
6476 &outinfo
->param_exports
);
6480 ac_setup_rings(struct nir_to_llvm_context
*ctx
)
6482 if ((ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_es
) ||
6483 (ctx
->stage
== MESA_SHADER_TESS_EVAL
&& ctx
->options
->key
.tes
.as_es
)) {
6484 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_VS
, false));
6487 if (ctx
->is_gs_copy_shader
) {
6488 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_VS
, false));
6490 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
6492 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_ESGS_GS
, false));
6493 ctx
->gsvs_ring
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_GSVS_GS
, false));
6495 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->builder
, ctx
->gsvs_ring
, ctx
->ac
.v4i32
, "");
6497 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->gsvs_num_entries
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
6498 tmp
= LLVMBuildExtractElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->ac
.i32_1
, "");
6499 tmp
= LLVMBuildOr(ctx
->builder
, tmp
, ctx
->gsvs_ring_stride
, "");
6500 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, tmp
, ctx
->ac
.i32_1
, "");
6503 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
6504 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
6505 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
6506 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
6511 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
6512 const struct nir_shader
*nir
)
6514 switch (nir
->info
.stage
) {
6515 case MESA_SHADER_TESS_CTRL
:
6516 return chip_class
>= CIK
? 128 : 64;
6517 case MESA_SHADER_GEOMETRY
:
6518 return chip_class
>= GFX9
? 128 : 64;
6519 case MESA_SHADER_COMPUTE
:
6525 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
6526 nir
->info
.cs
.local_size
[1] *
6527 nir
->info
.cs
.local_size
[2];
6528 return max_workgroup_size
;
6531 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
6532 static void ac_nir_fixup_ls_hs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6534 LLVMValueRef count
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6535 LLVMConstInt(ctx
->ac
.i32
, 8, false),
6536 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6537 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
6539 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->rel_auto_id
, ctx
->abi
.instance_id
, "");
6540 ctx
->vs_prim_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.vertex_id
, ctx
->vs_prim_id
, "");
6541 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_rel_ids
, ctx
->rel_auto_id
, "");
6542 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
, ctx
->abi
.tcs_patch_id
, ctx
->abi
.vertex_id
, "");
6545 static void prepare_gs_input_vgprs(struct nir_to_llvm_context
*ctx
)
6547 for(int i
= 5; i
>= 0; --i
) {
6548 ctx
->gs_vtx_offset
[i
] = ac_build_bfe(&ctx
->ac
, ctx
->gs_vtx_offset
[i
& ~1],
6549 LLVMConstInt(ctx
->ac
.i32
, (i
& 1) * 16, false),
6550 LLVMConstInt(ctx
->ac
.i32
, 16, false), false);
6553 ctx
->gs_wave_id
= ac_build_bfe(&ctx
->ac
, ctx
->merged_wave_info
,
6554 LLVMConstInt(ctx
->ac
.i32
, 16, false),
6555 LLVMConstInt(ctx
->ac
.i32
, 8, false), false);
6558 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
6559 struct nir_shader
*nir
, struct nir_to_llvm_context
*nctx
)
6561 struct ac_nir_context ctx
= {};
6562 struct nir_function
*func
;
6571 ctx
.stage
= nir
->info
.stage
;
6573 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6575 nir_foreach_variable(variable
, &nir
->outputs
)
6576 handle_shader_output_decl(&ctx
, nir
, variable
);
6578 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6579 _mesa_key_pointer_equal
);
6580 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6581 _mesa_key_pointer_equal
);
6582 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6583 _mesa_key_pointer_equal
);
6585 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
6587 setup_locals(&ctx
, func
);
6589 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
)
6590 setup_shared(&ctx
, nir
);
6592 visit_cf_list(&ctx
, &func
->impl
->body
);
6593 phi_post_pass(&ctx
);
6595 ctx
.abi
->emit_outputs(ctx
.abi
, RADEON_LLVM_MAX_OUTPUTS
,
6599 ralloc_free(ctx
.defs
);
6600 ralloc_free(ctx
.phis
);
6601 ralloc_free(ctx
.vars
);
6608 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
6609 struct nir_shader
*const *shaders
,
6611 struct ac_shader_variant_info
*shader_info
,
6612 const struct ac_nir_compiler_options
*options
)
6614 struct nir_to_llvm_context ctx
= {0};
6616 ctx
.options
= options
;
6617 ctx
.shader_info
= shader_info
;
6618 ctx
.context
= LLVMContextCreate();
6619 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
6621 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
6623 ctx
.ac
.module
= ctx
.module
;
6624 LLVMSetTarget(ctx
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
6626 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
6627 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
6628 LLVMSetDataLayout(ctx
.module
, data_layout_str
);
6629 LLVMDisposeTargetData(data_layout
);
6630 LLVMDisposeMessage(data_layout_str
);
6632 ctx
.builder
= LLVMCreateBuilderInContext(ctx
.context
);
6633 ctx
.ac
.builder
= ctx
.builder
;
6635 memset(shader_info
, 0, sizeof(*shader_info
));
6637 for(int i
= 0; i
< shader_count
; ++i
)
6638 ac_nir_shader_info_pass(shaders
[i
], options
, &shader_info
->info
);
6640 for (i
= 0; i
< AC_UD_MAX_SETS
; i
++)
6641 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
6642 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
6643 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
6645 ctx
.max_workgroup_size
= 0;
6646 for (int i
= 0; i
< shader_count
; ++i
) {
6647 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
6648 ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
,
6652 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2,
6653 shader_count
>= 2 ? shaders
[shader_count
- 2]->info
.stage
: MESA_SHADER_VERTEX
);
6655 ctx
.abi
.inputs
= &ctx
.inputs
[0];
6656 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
6657 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
6658 ctx
.abi
.load_ubo
= radv_load_ubo
;
6659 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
6660 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
6661 ctx
.abi
.clamp_shadow_reference
= false;
6663 if (shader_count
>= 2)
6664 ac_init_exec_full_mask(&ctx
.ac
);
6666 if (ctx
.ac
.chip_class
== GFX9
&&
6667 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
6668 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
6670 for(int i
= 0; i
< shader_count
; ++i
) {
6671 ctx
.stage
= shaders
[i
]->info
.stage
;
6672 ctx
.output_mask
= 0;
6673 ctx
.tess_outputs_written
= 0;
6674 ctx
.num_output_clips
= shaders
[i
]->info
.clip_distance_array_size
;
6675 ctx
.num_output_culls
= shaders
[i
]->info
.cull_distance_array_size
;
6677 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6678 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "gs_next_vertex");
6679 ctx
.gs_max_out_vertices
= shaders
[i
]->info
.gs
.vertices_out
;
6680 ctx
.abi
.load_inputs
= load_gs_input
;
6681 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6682 ctx
.tcs_outputs_read
= shaders
[i
]->info
.outputs_read
;
6683 ctx
.tcs_patch_outputs_read
= shaders
[i
]->info
.patch_outputs_read
;
6684 ctx
.abi
.load_tess_inputs
= load_tcs_input
;
6685 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
6686 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
6687 ctx
.tes_primitive_mode
= shaders
[i
]->info
.tess
.primitive_mode
;
6688 ctx
.abi
.load_tess_inputs
= load_tes_input
;
6689 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
6690 if (shader_info
->info
.vs
.needs_instance_id
) {
6691 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6692 MAX2(3, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6694 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
6695 shader_info
->fs
.can_discard
= shaders
[i
]->info
.fs
.uses_discard
;
6701 ac_setup_rings(&ctx
);
6703 LLVMBasicBlockRef merge_block
;
6704 if (shader_count
>= 2) {
6705 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6706 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6707 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
6709 LLVMValueRef count
= ac_build_bfe(&ctx
.ac
, ctx
.merged_wave_info
,
6710 LLVMConstInt(ctx
.ac
.i32
, 8 * i
, false),
6711 LLVMConstInt(ctx
.ac
.i32
, 8, false), false);
6712 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
6713 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
6714 thread_id
, count
, "");
6715 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
6717 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
6720 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
6721 handle_fs_inputs(&ctx
, shaders
[i
]);
6722 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
6723 handle_vs_inputs(&ctx
, shaders
[i
]);
6724 else if(shader_count
>= 2 && shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
6725 prepare_gs_input_vgprs(&ctx
);
6727 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
6728 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
6730 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, shaders
[i
], &ctx
);
6732 if (shader_count
>= 2) {
6733 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
6734 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
6737 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
6738 unsigned addclip
= shaders
[i
]->info
.clip_distance_array_size
+
6739 shaders
[i
]->info
.cull_distance_array_size
> 4;
6740 shader_info
->gs
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
6741 shader_info
->gs
.max_gsvs_emit_size
= shader_info
->gs
.gsvs_vertex_size
*
6742 shaders
[i
]->info
.gs
.vertices_out
;
6743 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
6744 shader_info
->tcs
.outputs_written
= ctx
.tess_outputs_written
;
6745 shader_info
->tcs
.patch_outputs_written
= ctx
.tess_patch_outputs_written
;
6746 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
&& ctx
.options
->key
.vs
.as_ls
) {
6747 shader_info
->vs
.outputs_written
= ctx
.tess_outputs_written
;
6751 LLVMBuildRetVoid(ctx
.builder
);
6753 ac_llvm_finalize_module(&ctx
);
6755 if (shader_count
== 1)
6756 ac_nir_eliminate_const_vs_outputs(&ctx
);
6761 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
6763 unsigned *retval
= (unsigned *)context
;
6764 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
6765 char *description
= LLVMGetDiagInfoDescription(di
);
6767 if (severity
== LLVMDSError
) {
6769 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
6773 LLVMDisposeMessage(description
);
6776 static unsigned ac_llvm_compile(LLVMModuleRef M
,
6777 struct ac_shader_binary
*binary
,
6778 LLVMTargetMachineRef tm
)
6780 unsigned retval
= 0;
6782 LLVMContextRef llvm_ctx
;
6783 LLVMMemoryBufferRef out_buffer
;
6784 unsigned buffer_size
;
6785 const char *buffer_data
;
6788 /* Setup Diagnostic Handler*/
6789 llvm_ctx
= LLVMGetModuleContext(M
);
6791 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
6795 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
6798 /* Process Errors/Warnings */
6800 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
6806 /* Extract Shader Code*/
6807 buffer_size
= LLVMGetBufferSize(out_buffer
);
6808 buffer_data
= LLVMGetBufferStart(out_buffer
);
6810 ac_elf_read(buffer_data
, buffer_size
, binary
);
6813 LLVMDisposeMemoryBuffer(out_buffer
);
6819 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
6820 LLVMModuleRef llvm_module
,
6821 struct ac_shader_binary
*binary
,
6822 struct ac_shader_config
*config
,
6823 struct ac_shader_variant_info
*shader_info
,
6824 gl_shader_stage stage
,
6825 bool dump_shader
, bool supports_spill
)
6828 ac_dump_module(llvm_module
);
6830 memset(binary
, 0, sizeof(*binary
));
6831 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
6833 fprintf(stderr
, "compile failed\n");
6837 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
6839 ac_shader_binary_read_config(binary
, config
, 0, supports_spill
);
6841 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
6842 LLVMDisposeModule(llvm_module
);
6843 LLVMContextDispose(ctx
);
6845 if (stage
== MESA_SHADER_FRAGMENT
) {
6846 shader_info
->num_input_vgprs
= 0;
6847 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
6848 shader_info
->num_input_vgprs
+= 2;
6849 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
6850 shader_info
->num_input_vgprs
+= 2;
6851 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
6852 shader_info
->num_input_vgprs
+= 2;
6853 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
6854 shader_info
->num_input_vgprs
+= 3;
6855 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
6856 shader_info
->num_input_vgprs
+= 2;
6857 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
6858 shader_info
->num_input_vgprs
+= 2;
6859 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
6860 shader_info
->num_input_vgprs
+= 2;
6861 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
6862 shader_info
->num_input_vgprs
+= 1;
6863 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
6864 shader_info
->num_input_vgprs
+= 1;
6865 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
6866 shader_info
->num_input_vgprs
+= 1;
6867 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
6868 shader_info
->num_input_vgprs
+= 1;
6869 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
6870 shader_info
->num_input_vgprs
+= 1;
6871 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
6872 shader_info
->num_input_vgprs
+= 1;
6873 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
6874 shader_info
->num_input_vgprs
+= 1;
6875 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
6876 shader_info
->num_input_vgprs
+= 1;
6877 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
6878 shader_info
->num_input_vgprs
+= 1;
6880 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
6882 /* +3 for scratch wave offset and VCC */
6883 config
->num_sgprs
= MAX2(config
->num_sgprs
,
6884 shader_info
->num_input_sgprs
+ 3);
6888 ac_fill_shader_info(struct ac_shader_variant_info
*shader_info
, struct nir_shader
*nir
, const struct ac_nir_compiler_options
*options
)
6890 switch (nir
->info
.stage
) {
6891 case MESA_SHADER_COMPUTE
:
6892 for (int i
= 0; i
< 3; ++i
)
6893 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
6895 case MESA_SHADER_FRAGMENT
:
6896 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
6898 case MESA_SHADER_GEOMETRY
:
6899 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
6900 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
6901 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
6902 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
6904 case MESA_SHADER_TESS_EVAL
:
6905 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
6906 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
6907 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
6908 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
6909 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
6911 case MESA_SHADER_TESS_CTRL
:
6912 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
6914 case MESA_SHADER_VERTEX
:
6915 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
6916 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
6917 /* in LS mode we need at least 1, invocation id needs 3, handled elsewhere */
6918 if (options
->key
.vs
.as_ls
)
6919 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
6926 void ac_compile_nir_shader(LLVMTargetMachineRef tm
,
6927 struct ac_shader_binary
*binary
,
6928 struct ac_shader_config
*config
,
6929 struct ac_shader_variant_info
*shader_info
,
6930 struct nir_shader
*const *nir
,
6932 const struct ac_nir_compiler_options
*options
,
6936 LLVMModuleRef llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, nir_count
, shader_info
,
6939 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
, nir
[0]->info
.stage
, dump_shader
, options
->supports_spill
);
6940 for (int i
= 0; i
< nir_count
; ++i
)
6941 ac_fill_shader_info(shader_info
, nir
[i
], options
);
6945 ac_gs_copy_shader_emit(struct nir_to_llvm_context
*ctx
)
6947 LLVMValueRef args
[9];
6948 args
[0] = ctx
->gsvs_ring
;
6949 args
[1] = LLVMBuildMul(ctx
->builder
, ctx
->abi
.vertex_id
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
6950 args
[3] = ctx
->ac
.i32_0
;
6951 args
[4] = ctx
->ac
.i32_1
; /* OFFEN */
6952 args
[5] = ctx
->ac
.i32_0
; /* IDXEN */
6953 args
[6] = ctx
->ac
.i32_1
; /* GLC */
6954 args
[7] = ctx
->ac
.i32_1
; /* SLC */
6955 args
[8] = ctx
->ac
.i32_0
; /* TFE */
6959 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6963 if (!(ctx
->output_mask
& (1ull << i
)))
6966 if (i
== VARYING_SLOT_CLIP_DIST0
) {
6967 /* unpack clip and cull from a single set of slots */
6968 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6973 for (unsigned j
= 0; j
< length
; j
++) {
6975 args
[2] = LLVMConstInt(ctx
->ac
.i32
,
6977 ctx
->gs_max_out_vertices
* 16 * 4, false);
6979 value
= ac_build_intrinsic(&ctx
->ac
,
6980 "llvm.SI.buffer.load.dword.i32.i32",
6981 ctx
->ac
.i32
, args
, 9,
6982 AC_FUNC_ATTR_READONLY
|
6983 AC_FUNC_ATTR_LEGACY
);
6985 LLVMBuildStore(ctx
->builder
,
6986 ac_to_float(&ctx
->ac
, value
), ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)]);
6990 handle_vs_outputs_post(ctx
, false, &ctx
->shader_info
->vs
.outinfo
);
6993 void ac_create_gs_copy_shader(LLVMTargetMachineRef tm
,
6994 struct nir_shader
*geom_shader
,
6995 struct ac_shader_binary
*binary
,
6996 struct ac_shader_config
*config
,
6997 struct ac_shader_variant_info
*shader_info
,
6998 const struct ac_nir_compiler_options
*options
,
7001 struct nir_to_llvm_context ctx
= {0};
7002 ctx
.context
= LLVMContextCreate();
7003 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
7004 ctx
.options
= options
;
7005 ctx
.shader_info
= shader_info
;
7007 ac_llvm_context_init(&ctx
.ac
, ctx
.context
, options
->chip_class
,
7009 ctx
.ac
.module
= ctx
.module
;
7011 ctx
.is_gs_copy_shader
= true;
7012 LLVMSetTarget(ctx
.module
, "amdgcn--");
7014 ctx
.builder
= LLVMCreateBuilderInContext(ctx
.context
);
7015 ctx
.ac
.builder
= ctx
.builder
;
7016 ctx
.stage
= MESA_SHADER_VERTEX
;
7018 create_function(&ctx
, MESA_SHADER_VERTEX
, false, MESA_SHADER_VERTEX
);
7020 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
7021 ac_setup_rings(&ctx
);
7023 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
7024 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
7026 struct ac_nir_context nir_ctx
= {};
7027 nir_ctx
.ac
= ctx
.ac
;
7028 nir_ctx
.abi
= &ctx
.abi
;
7030 nir_ctx
.nctx
= &ctx
;
7033 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
7034 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
7035 handle_shader_output_decl(&nir_ctx
, geom_shader
, variable
);
7038 ac_gs_copy_shader_emit(&ctx
);
7042 LLVMBuildRetVoid(ctx
.builder
);
7044 ac_llvm_finalize_module(&ctx
);
7046 ac_compile_llvm_module(tm
, ctx
.module
, binary
, config
, shader_info
,
7048 dump_shader
, options
->supports_spill
);