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_exp_param.h"
37 enum radeon_llvm_calling_convention
{
38 RADEON_LLVM_AMDGPU_VS
= 87,
39 RADEON_LLVM_AMDGPU_GS
= 88,
40 RADEON_LLVM_AMDGPU_PS
= 89,
41 RADEON_LLVM_AMDGPU_CS
= 90,
44 #define CONST_ADDR_SPACE 2
45 #define LOCAL_ADDR_SPACE 3
47 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
48 #define RADEON_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
50 struct nir_to_llvm_context
;
52 struct ac_nir_context
{
53 struct ac_llvm_context ac
;
54 struct ac_shader_abi
*abi
;
56 gl_shader_stage stage
;
58 struct hash_table
*defs
;
59 struct hash_table
*phis
;
60 struct hash_table
*vars
;
62 LLVMValueRef main_function
;
63 LLVMBasicBlockRef continue_block
;
64 LLVMBasicBlockRef break_block
;
66 LLVMValueRef outputs
[RADEON_LLVM_MAX_OUTPUTS
* 4];
71 struct nir_to_llvm_context
*nctx
; /* TODO get rid of this */
74 struct nir_to_llvm_context
{
75 struct ac_llvm_context ac
;
76 const struct ac_nir_compiler_options
*options
;
77 struct ac_shader_variant_info
*shader_info
;
78 struct ac_shader_abi abi
;
79 struct ac_nir_context
*nir
;
81 unsigned max_workgroup_size
;
82 LLVMContextRef context
;
84 LLVMBuilderRef builder
;
85 LLVMValueRef main_function
;
87 struct hash_table
*defs
;
88 struct hash_table
*phis
;
90 LLVMValueRef descriptor_sets
[AC_UD_MAX_SETS
];
91 LLVMValueRef ring_offsets
;
92 LLVMValueRef push_constants
;
93 LLVMValueRef num_work_groups
;
94 LLVMValueRef workgroup_ids
;
95 LLVMValueRef local_invocation_ids
;
98 LLVMValueRef vertex_buffers
;
99 LLVMValueRef rel_auto_id
;
100 LLVMValueRef vs_prim_id
;
101 LLVMValueRef ls_out_layout
;
102 LLVMValueRef es2gs_offset
;
104 LLVMValueRef tcs_offchip_layout
;
105 LLVMValueRef tcs_out_offsets
;
106 LLVMValueRef tcs_out_layout
;
107 LLVMValueRef tcs_in_layout
;
109 LLVMValueRef tess_factor_offset
;
110 LLVMValueRef tcs_patch_id
;
111 LLVMValueRef tcs_rel_ids
;
112 LLVMValueRef tes_rel_patch_id
;
113 LLVMValueRef tes_patch_id
;
117 LLVMValueRef gsvs_ring_stride
;
118 LLVMValueRef gsvs_num_entries
;
119 LLVMValueRef gs2vs_offset
;
120 LLVMValueRef gs_wave_id
;
121 LLVMValueRef gs_vtx_offset
[6];
122 LLVMValueRef gs_prim_id
, gs_invocation_id
;
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
;
133 LLVMValueRef front_face
;
134 LLVMValueRef ancillary
;
135 LLVMValueRef sample_coverage
;
136 LLVMValueRef frag_pos
[4];
155 LLVMValueRef i1false
;
156 LLVMValueRef i32zero
;
158 LLVMValueRef f32zero
;
160 LLVMValueRef v4f32empty
;
162 unsigned uniform_md_kind
;
163 LLVMValueRef empty_md
;
164 gl_shader_stage stage
;
167 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
170 uint64_t output_mask
;
171 uint8_t num_output_clips
;
172 uint8_t num_output_culls
;
174 bool has_ds_bpermute
;
176 bool is_gs_copy_shader
;
177 LLVMValueRef gs_next_vertex
;
178 unsigned gs_max_out_vertices
;
180 unsigned tes_primitive_mode
;
181 uint64_t tess_outputs_written
;
182 uint64_t tess_patch_outputs_written
;
185 static inline struct nir_to_llvm_context
*
186 nir_to_llvm_context_from_abi(struct ac_shader_abi
*abi
)
188 struct nir_to_llvm_context
*ctx
= NULL
;
189 return container_of(abi
, ctx
, abi
);
192 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
193 const nir_deref_var
*deref
,
194 enum ac_descriptor_type desc_type
,
195 bool image
, bool write
);
197 static unsigned radeon_llvm_reg_index_soa(unsigned index
, unsigned chan
)
199 return (index
* 4) + chan
;
202 static unsigned shader_io_get_unique_index(gl_varying_slot slot
)
204 /* handle patch indices separate */
205 if (slot
== VARYING_SLOT_TESS_LEVEL_OUTER
)
207 if (slot
== VARYING_SLOT_TESS_LEVEL_INNER
)
209 if (slot
>= VARYING_SLOT_PATCH0
&& slot
<= VARYING_SLOT_TESS_MAX
)
210 return 2 + (slot
- VARYING_SLOT_PATCH0
);
212 if (slot
== VARYING_SLOT_POS
)
214 if (slot
== VARYING_SLOT_PSIZ
)
216 if (slot
== VARYING_SLOT_CLIP_DIST0
)
218 /* 3 is reserved for clip dist as well */
219 if (slot
>= VARYING_SLOT_VAR0
&& slot
<= VARYING_SLOT_VAR31
)
220 return 4 + (slot
- VARYING_SLOT_VAR0
);
221 unreachable("illegal slot in get unique index\n");
224 static unsigned llvm_get_type_size(LLVMTypeRef type
)
226 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
229 case LLVMIntegerTypeKind
:
230 return LLVMGetIntTypeWidth(type
) / 8;
231 case LLVMFloatTypeKind
:
233 case LLVMPointerTypeKind
:
235 case LLVMVectorTypeKind
:
236 return LLVMGetVectorSize(type
) *
237 llvm_get_type_size(LLVMGetElementType(type
));
244 static void set_llvm_calling_convention(LLVMValueRef func
,
245 gl_shader_stage stage
)
247 enum radeon_llvm_calling_convention calling_conv
;
250 case MESA_SHADER_VERTEX
:
251 case MESA_SHADER_TESS_CTRL
:
252 case MESA_SHADER_TESS_EVAL
:
253 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
255 case MESA_SHADER_GEOMETRY
:
256 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
258 case MESA_SHADER_FRAGMENT
:
259 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
261 case MESA_SHADER_COMPUTE
:
262 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
265 unreachable("Unhandle shader type");
268 LLVMSetFunctionCallConv(func
, calling_conv
);
273 LLVMTypeRef types
[MAX_ARGS
];
274 LLVMValueRef
*assign
[MAX_ARGS
];
275 unsigned array_params_mask
;
277 uint8_t user_sgpr_count
;
279 uint8_t num_user_sgprs_used
;
280 uint8_t num_sgprs_used
;
281 uint8_t num_vgprs_used
;
285 add_argument(struct arg_info
*info
,
286 LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
288 assert(info
->count
< MAX_ARGS
);
289 info
->assign
[info
->count
] = param_ptr
;
290 info
->types
[info
->count
] = type
;
295 add_sgpr_argument(struct arg_info
*info
,
296 LLVMTypeRef type
, LLVMValueRef
*param_ptr
)
298 add_argument(info
, type
, param_ptr
);
299 info
->num_sgprs_used
+= llvm_get_type_size(type
) / 4;
304 add_user_sgpr_argument(struct arg_info
*info
,
306 LLVMValueRef
*param_ptr
)
308 add_sgpr_argument(info
, type
, param_ptr
);
309 info
->num_user_sgprs_used
+= llvm_get_type_size(type
) / 4;
310 info
->user_sgpr_count
++;
314 add_vgpr_argument(struct arg_info
*info
,
316 LLVMValueRef
*param_ptr
)
318 add_argument(info
, type
, param_ptr
);
319 info
->num_vgprs_used
+= llvm_get_type_size(type
) / 4;
323 add_user_sgpr_array_argument(struct arg_info
*info
,
325 LLVMValueRef
*param_ptr
)
327 info
->array_params_mask
|= (1 << info
->count
);
328 add_user_sgpr_argument(info
, type
, param_ptr
);
331 static void assign_arguments(LLVMValueRef main_function
,
332 struct arg_info
*info
)
335 for (i
= 0; i
< info
->count
; i
++) {
337 *info
->assign
[i
] = LLVMGetParam(main_function
, i
);
342 create_llvm_function(LLVMContextRef ctx
, LLVMModuleRef module
,
343 LLVMBuilderRef builder
, LLVMTypeRef
*return_types
,
344 unsigned num_return_elems
,
345 struct arg_info
*args
,
346 unsigned max_workgroup_size
,
349 LLVMTypeRef main_function_type
, ret_type
;
350 LLVMBasicBlockRef main_function_body
;
352 if (num_return_elems
)
353 ret_type
= LLVMStructTypeInContext(ctx
, return_types
,
354 num_return_elems
, true);
356 ret_type
= LLVMVoidTypeInContext(ctx
);
358 /* Setup the function */
360 LLVMFunctionType(ret_type
, args
->types
, args
->count
, 0);
361 LLVMValueRef main_function
=
362 LLVMAddFunction(module
, "main", main_function_type
);
364 LLVMAppendBasicBlockInContext(ctx
, main_function
, "main_body");
365 LLVMPositionBuilderAtEnd(builder
, main_function_body
);
367 LLVMSetFunctionCallConv(main_function
, RADEON_LLVM_AMDGPU_CS
);
368 for (unsigned i
= 0; i
< args
->sgpr_count
; ++i
) {
369 if (args
->array_params_mask
& (1 << i
)) {
370 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
371 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_BYVAL
);
372 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
375 ac_add_function_attr(ctx
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
379 if (max_workgroup_size
) {
380 ac_llvm_add_target_dep_function_attr(main_function
,
381 "amdgpu-max-work-group-size",
385 /* These were copied from some LLVM test. */
386 LLVMAddTargetDependentFunctionAttr(main_function
,
387 "less-precise-fpmad",
389 LLVMAddTargetDependentFunctionAttr(main_function
,
392 LLVMAddTargetDependentFunctionAttr(main_function
,
395 LLVMAddTargetDependentFunctionAttr(main_function
,
399 return main_function
;
402 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
404 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
408 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
410 if (t
== ctx
->f16
|| t
== ctx
->i16
)
412 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
414 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
417 unreachable("Unhandled integer size");
420 static LLVMTypeRef
to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
422 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
423 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
424 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
425 LLVMGetVectorSize(t
));
427 return to_integer_type_scalar(ctx
, t
);
430 static LLVMValueRef
to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
432 LLVMTypeRef type
= LLVMTypeOf(v
);
433 return LLVMBuildBitCast(ctx
->builder
, v
, to_integer_type(ctx
, type
), "");
436 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
438 if (t
== ctx
->i16
|| t
== ctx
->f16
)
440 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
442 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
445 unreachable("Unhandled float size");
448 static LLVMTypeRef
to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
450 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
451 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
452 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
453 LLVMGetVectorSize(t
));
455 return to_float_type_scalar(ctx
, t
);
458 static LLVMValueRef
to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
460 LLVMTypeRef type
= LLVMTypeOf(v
);
461 return LLVMBuildBitCast(ctx
->builder
, v
, to_float_type(ctx
, type
), "");
464 static int get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
466 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
467 type
= LLVMGetElementType(type
);
469 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
470 return LLVMGetIntTypeWidth(type
);
472 if (type
== ctx
->f16
)
474 if (type
== ctx
->f32
)
476 if (type
== ctx
->f64
)
479 unreachable("Unhandled type kind in get_elem_bits");
482 static LLVMValueRef
unpack_param(struct nir_to_llvm_context
*ctx
,
483 LLVMValueRef param
, unsigned rshift
,
486 LLVMValueRef value
= param
;
488 value
= LLVMBuildLShr(ctx
->builder
, value
,
489 LLVMConstInt(ctx
->i32
, rshift
, false), "");
491 if (rshift
+ bitwidth
< 32) {
492 unsigned mask
= (1 << bitwidth
) - 1;
493 value
= LLVMBuildAnd(ctx
->builder
, value
,
494 LLVMConstInt(ctx
->i32
, mask
, false), "");
499 static LLVMValueRef
get_rel_patch_id(struct nir_to_llvm_context
*ctx
)
501 switch (ctx
->stage
) {
502 case MESA_SHADER_TESS_CTRL
:
503 return unpack_param(ctx
, ctx
->tcs_rel_ids
, 0, 8);
504 case MESA_SHADER_TESS_EVAL
:
505 return ctx
->tes_rel_patch_id
;
508 unreachable("Illegal stage");
512 /* Tessellation shaders pass outputs to the next shader using LDS.
514 * LS outputs = TCS inputs
515 * TCS outputs = TES inputs
518 * - TCS inputs for patch 0
519 * - TCS inputs for patch 1
520 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
522 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
523 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
524 * - TCS outputs for patch 1
525 * - Per-patch TCS outputs for patch 1
526 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
527 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
530 * All three shaders VS(LS), TCS, TES share the same LDS space.
533 get_tcs_in_patch_stride(struct nir_to_llvm_context
*ctx
)
535 if (ctx
->stage
== MESA_SHADER_VERTEX
)
536 return unpack_param(ctx
, ctx
->ls_out_layout
, 0, 13);
537 else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
538 return unpack_param(ctx
, ctx
->tcs_in_layout
, 0, 13);
546 get_tcs_out_patch_stride(struct nir_to_llvm_context
*ctx
)
548 return unpack_param(ctx
, ctx
->tcs_out_layout
, 0, 13);
552 get_tcs_out_patch0_offset(struct nir_to_llvm_context
*ctx
)
554 return LLVMBuildMul(ctx
->builder
,
555 unpack_param(ctx
, ctx
->tcs_out_offsets
, 0, 16),
556 LLVMConstInt(ctx
->i32
, 4, false), "");
560 get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context
*ctx
)
562 return LLVMBuildMul(ctx
->builder
,
563 unpack_param(ctx
, ctx
->tcs_out_offsets
, 16, 16),
564 LLVMConstInt(ctx
->i32
, 4, false), "");
568 get_tcs_in_current_patch_offset(struct nir_to_llvm_context
*ctx
)
570 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
571 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
573 return LLVMBuildMul(ctx
->builder
, patch_stride
, rel_patch_id
, "");
577 get_tcs_out_current_patch_offset(struct nir_to_llvm_context
*ctx
)
579 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
580 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
581 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
583 return LLVMBuildAdd(ctx
->builder
, patch0_offset
,
584 LLVMBuildMul(ctx
->builder
, patch_stride
,
590 get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context
*ctx
)
592 LLVMValueRef patch0_patch_data_offset
=
593 get_tcs_out_patch0_patch_data_offset(ctx
);
594 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
595 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
597 return LLVMBuildAdd(ctx
->builder
, patch0_patch_data_offset
,
598 LLVMBuildMul(ctx
->builder
, patch_stride
,
603 static void set_userdata_location(struct ac_userdata_info
*ud_info
, uint8_t *sgpr_idx
, uint8_t num_sgprs
)
605 ud_info
->sgpr_idx
= *sgpr_idx
;
606 ud_info
->num_sgprs
= num_sgprs
;
607 ud_info
->indirect
= false;
608 ud_info
->indirect_offset
= 0;
609 *sgpr_idx
+= num_sgprs
;
612 static void set_userdata_location_shader(struct nir_to_llvm_context
*ctx
,
613 int idx
, uint8_t *sgpr_idx
, uint8_t num_sgprs
)
615 set_userdata_location(&ctx
->shader_info
->user_sgprs_locs
.shader_data
[idx
], sgpr_idx
, num_sgprs
);
619 static void set_userdata_location_indirect(struct ac_userdata_info
*ud_info
, uint8_t sgpr_idx
, uint8_t num_sgprs
,
620 uint32_t indirect_offset
)
622 ud_info
->sgpr_idx
= sgpr_idx
;
623 ud_info
->num_sgprs
= num_sgprs
;
624 ud_info
->indirect
= true;
625 ud_info
->indirect_offset
= indirect_offset
;
628 static void declare_tess_lds(struct nir_to_llvm_context
*ctx
)
630 unsigned lds_size
= ctx
->options
->chip_class
>= CIK
? 65536 : 32768;
631 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32zero
,
632 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), LOCAL_ADDR_SPACE
),
636 struct user_sgpr_info
{
637 bool need_ring_offsets
;
639 bool indirect_all_descriptor_sets
;
642 static void allocate_user_sgprs(struct nir_to_llvm_context
*ctx
,
643 struct user_sgpr_info
*user_sgpr_info
)
645 memset(user_sgpr_info
, 0, sizeof(struct user_sgpr_info
));
647 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
648 if (ctx
->stage
== MESA_SHADER_GEOMETRY
||
649 ctx
->stage
== MESA_SHADER_VERTEX
||
650 ctx
->stage
== MESA_SHADER_TESS_CTRL
||
651 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
652 ctx
->is_gs_copy_shader
)
653 user_sgpr_info
->need_ring_offsets
= true;
655 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&&
656 ctx
->shader_info
->info
.ps
.needs_sample_positions
)
657 user_sgpr_info
->need_ring_offsets
= true;
659 /* 2 user sgprs will nearly always be allocated for scratch/rings */
660 if (ctx
->options
->supports_spill
|| user_sgpr_info
->need_ring_offsets
) {
661 user_sgpr_info
->sgpr_count
+= 2;
664 switch (ctx
->stage
) {
665 case MESA_SHADER_COMPUTE
:
666 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.cs
.grid_components_used
;
668 case MESA_SHADER_FRAGMENT
:
669 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.ps
.needs_sample_positions
;
671 case MESA_SHADER_VERTEX
:
672 if (!ctx
->is_gs_copy_shader
) {
673 user_sgpr_info
->sgpr_count
+= ctx
->shader_info
->info
.vs
.has_vertex_buffers
? 2 : 0;
674 if (ctx
->shader_info
->info
.vs
.needs_draw_id
) {
675 user_sgpr_info
->sgpr_count
+= 3;
677 user_sgpr_info
->sgpr_count
+= 2;
680 if (ctx
->options
->key
.vs
.as_ls
)
681 user_sgpr_info
->sgpr_count
++;
683 case MESA_SHADER_TESS_CTRL
:
684 user_sgpr_info
->sgpr_count
+= 4;
686 case MESA_SHADER_TESS_EVAL
:
687 user_sgpr_info
->sgpr_count
+= 1;
689 case MESA_SHADER_GEOMETRY
:
690 user_sgpr_info
->sgpr_count
+= 2;
696 if (ctx
->shader_info
->info
.needs_push_constants
)
697 user_sgpr_info
->sgpr_count
+= 2;
699 uint32_t remaining_sgprs
= 16 - user_sgpr_info
->sgpr_count
;
700 if (remaining_sgprs
/ 2 < util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
)) {
701 user_sgpr_info
->sgpr_count
+= 2;
702 user_sgpr_info
->indirect_all_descriptor_sets
= true;
704 user_sgpr_info
->sgpr_count
+= util_bitcount(ctx
->shader_info
->info
.desc_set_used_mask
) * 2;
708 static void create_function(struct nir_to_llvm_context
*ctx
)
710 unsigned num_sets
= ctx
->options
->layout
? ctx
->options
->layout
->num_sets
: 0;
711 uint8_t user_sgpr_idx
;
712 struct user_sgpr_info user_sgpr_info
;
713 struct arg_info args
= {};
714 LLVMValueRef desc_sets
;
716 allocate_user_sgprs(ctx
, &user_sgpr_info
);
717 if (user_sgpr_info
.need_ring_offsets
&& !ctx
->options
->supports_spill
) {
718 add_user_sgpr_argument(&args
, const_array(ctx
->v4i32
, 16), &ctx
->ring_offsets
); /* address of rings */
721 /* 1 for each descriptor set */
722 if (!user_sgpr_info
.indirect_all_descriptor_sets
) {
723 for (unsigned i
= 0; i
< num_sets
; ++i
) {
724 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& (1 << ctx
->stage
)) {
725 add_user_sgpr_array_argument(&args
, const_array(ctx
->i8
, 1024 * 1024), &ctx
->descriptor_sets
[i
]);
729 add_user_sgpr_array_argument(&args
, const_array(const_array(ctx
->i8
, 1024 * 1024), 32), &desc_sets
);
731 if (ctx
->shader_info
->info
.needs_push_constants
) {
732 /* 1 for push constants and dynamic descriptors */
733 add_user_sgpr_array_argument(&args
, const_array(ctx
->i8
, 1024 * 1024), &ctx
->push_constants
);
736 switch (ctx
->stage
) {
737 case MESA_SHADER_COMPUTE
:
738 if (ctx
->shader_info
->info
.cs
.grid_components_used
)
739 add_user_sgpr_argument(&args
, LLVMVectorType(ctx
->i32
, ctx
->shader_info
->info
.cs
.grid_components_used
), &ctx
->num_work_groups
); /* grid size */
740 add_sgpr_argument(&args
, LLVMVectorType(ctx
->i32
, 3), &ctx
->workgroup_ids
);
741 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->tg_size
);
742 add_vgpr_argument(&args
, LLVMVectorType(ctx
->i32
, 3), &ctx
->local_invocation_ids
);
744 case MESA_SHADER_VERTEX
:
745 if (!ctx
->is_gs_copy_shader
) {
746 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
)
747 add_user_sgpr_argument(&args
, const_array(ctx
->v4i32
, 16), &ctx
->vertex_buffers
); /* vertex buffers */
748 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->abi
.base_vertex
); // base vertex
749 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->abi
.start_instance
);// start instance
750 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
751 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->abi
.draw_id
); // draw id
753 if (ctx
->options
->key
.vs
.as_es
)
754 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->es2gs_offset
); // es2gs offset
755 else if (ctx
->options
->key
.vs
.as_ls
)
756 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->ls_out_layout
); // ls out layout
757 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->abi
.vertex_id
); // vertex id
758 if (!ctx
->is_gs_copy_shader
) {
759 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->rel_auto_id
); // rel auto id
760 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->vs_prim_id
); // vs prim id
761 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->abi
.instance_id
); // instance id
764 case MESA_SHADER_TESS_CTRL
:
765 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_offchip_layout
); // tcs offchip layout
766 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_out_offsets
); // tcs out offsets
767 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_out_layout
); // tcs out layout
768 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_in_layout
); // tcs in layout
769 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->oc_lds
); // param oc lds
770 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->tess_factor_offset
); // tess factor offset
771 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_patch_id
); // patch id
772 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_rel_ids
); // rel ids;
774 case MESA_SHADER_TESS_EVAL
:
775 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->tcs_offchip_layout
); // tcs offchip layout
776 if (ctx
->options
->key
.tes
.as_es
) {
777 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->oc_lds
); // OC LDS
778 add_sgpr_argument(&args
, ctx
->i32
, NULL
); //
779 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->es2gs_offset
); // es2gs offset
781 add_sgpr_argument(&args
, ctx
->i32
, NULL
); //
782 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->oc_lds
); // OC LDS
784 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->tes_u
); // tes_u
785 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->tes_v
); // tes_v
786 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->tes_rel_patch_id
); // tes rel patch id
787 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->tes_patch_id
); // tes patch id
789 case MESA_SHADER_GEOMETRY
:
790 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->gsvs_ring_stride
); // gsvs stride
791 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->gsvs_num_entries
); // gsvs num entires
792 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->gs2vs_offset
); // gs2vs offset
793 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->gs_wave_id
); // wave id
794 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[0]); // vtx0
795 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[1]); // vtx1
796 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_prim_id
); // prim id
797 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[2]);
798 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[3]);
799 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[4]);
800 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_vtx_offset
[5]);
801 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->gs_invocation_id
);
803 case MESA_SHADER_FRAGMENT
:
804 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
)
805 add_user_sgpr_argument(&args
, ctx
->i32
, &ctx
->sample_pos_offset
); /* sample position offset */
806 add_sgpr_argument(&args
, ctx
->i32
, &ctx
->prim_mask
); /* prim mask */
807 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->persp_sample
); /* persp sample */
808 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->persp_center
); /* persp center */
809 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->persp_centroid
); /* persp centroid */
810 add_vgpr_argument(&args
, ctx
->v3i32
, NULL
); /* persp pull model */
811 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->linear_sample
); /* linear sample */
812 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->linear_center
); /* linear center */
813 add_vgpr_argument(&args
, ctx
->v2i32
, &ctx
->linear_centroid
); /* linear centroid */
814 add_vgpr_argument(&args
, ctx
->f32
, NULL
); /* line stipple tex */
815 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->frag_pos
[0]); /* pos x float */
816 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->frag_pos
[1]); /* pos y float */
817 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->frag_pos
[2]); /* pos z float */
818 add_vgpr_argument(&args
, ctx
->f32
, &ctx
->frag_pos
[3]); /* pos w float */
819 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->front_face
); /* front face */
820 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->ancillary
); /* ancillary */
821 add_vgpr_argument(&args
, ctx
->i32
, &ctx
->sample_coverage
); /* sample coverage */
822 add_vgpr_argument(&args
, ctx
->i32
, NULL
); /* fixed pt */
825 unreachable("Shader stage not implemented");
828 ctx
->main_function
= create_llvm_function(
829 ctx
->context
, ctx
->module
, ctx
->builder
, NULL
, 0, &args
,
830 ctx
->max_workgroup_size
,
831 ctx
->options
->unsafe_math
);
832 set_llvm_calling_convention(ctx
->main_function
, ctx
->stage
);
835 ctx
->shader_info
->num_input_vgprs
= 0;
836 ctx
->shader_info
->num_input_sgprs
= ctx
->shader_info
->num_user_sgprs
=
837 ctx
->options
->supports_spill
? 2 : 0;
839 ctx
->shader_info
->num_user_sgprs
+= args
.num_user_sgprs_used
;
840 ctx
->shader_info
->num_input_sgprs
+= args
.num_sgprs_used
;
842 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
843 ctx
->shader_info
->num_input_vgprs
= args
.num_vgprs_used
;
845 assign_arguments(ctx
->main_function
, &args
);
849 if (ctx
->options
->supports_spill
|| user_sgpr_info
.need_ring_offsets
) {
850 set_userdata_location_shader(ctx
, AC_UD_SCRATCH_RING_OFFSETS
, &user_sgpr_idx
, 2);
851 if (ctx
->options
->supports_spill
) {
852 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
853 LLVMPointerType(ctx
->i8
, CONST_ADDR_SPACE
),
854 NULL
, 0, AC_FUNC_ATTR_READNONE
);
855 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->builder
, ctx
->ring_offsets
,
856 const_array(ctx
->v4i32
, 16), "");
860 if (!user_sgpr_info
.indirect_all_descriptor_sets
) {
861 for (unsigned i
= 0; i
< num_sets
; ++i
) {
862 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& (1 << ctx
->stage
)) {
863 set_userdata_location(&ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[i
], &user_sgpr_idx
, 2);
865 ctx
->descriptor_sets
[i
] = NULL
;
868 uint32_t desc_sgpr_idx
= user_sgpr_idx
;
869 set_userdata_location_shader(ctx
, AC_UD_INDIRECT_DESCRIPTOR_SETS
, &user_sgpr_idx
, 2);
871 for (unsigned i
= 0; i
< num_sets
; ++i
) {
872 if (ctx
->options
->layout
->set
[i
].layout
->shader_stages
& (1 << ctx
->stage
)) {
873 set_userdata_location_indirect(&ctx
->shader_info
->user_sgprs_locs
.descriptor_sets
[i
], desc_sgpr_idx
, 2, i
* 8);
874 ctx
->descriptor_sets
[i
] = ac_build_indexed_load_const(&ctx
->ac
, desc_sets
, LLVMConstInt(ctx
->i32
, i
, false));
877 ctx
->descriptor_sets
[i
] = NULL
;
879 ctx
->shader_info
->need_indirect_descriptor_sets
= true;
882 if (ctx
->shader_info
->info
.needs_push_constants
) {
883 set_userdata_location_shader(ctx
, AC_UD_PUSH_CONSTANTS
, &user_sgpr_idx
, 2);
886 switch (ctx
->stage
) {
887 case MESA_SHADER_COMPUTE
:
888 if (ctx
->shader_info
->info
.cs
.grid_components_used
) {
889 set_userdata_location_shader(ctx
, AC_UD_CS_GRID_SIZE
, &user_sgpr_idx
, ctx
->shader_info
->info
.cs
.grid_components_used
);
892 case MESA_SHADER_VERTEX
:
893 if (!ctx
->is_gs_copy_shader
) {
894 if (ctx
->shader_info
->info
.vs
.has_vertex_buffers
) {
895 set_userdata_location_shader(ctx
, AC_UD_VS_VERTEX_BUFFERS
, &user_sgpr_idx
, 2);
898 if (ctx
->shader_info
->info
.vs
.needs_draw_id
)
901 set_userdata_location_shader(ctx
, AC_UD_VS_BASE_VERTEX_START_INSTANCE
, &user_sgpr_idx
, vs_num
);
903 if (ctx
->options
->key
.vs
.as_ls
) {
904 set_userdata_location_shader(ctx
, AC_UD_VS_LS_TCS_IN_LAYOUT
, &user_sgpr_idx
, 1);
906 if (ctx
->options
->key
.vs
.as_ls
)
907 declare_tess_lds(ctx
);
909 case MESA_SHADER_TESS_CTRL
:
910 set_userdata_location_shader(ctx
, AC_UD_TCS_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 4);
911 declare_tess_lds(ctx
);
913 case MESA_SHADER_TESS_EVAL
:
914 set_userdata_location_shader(ctx
, AC_UD_TES_OFFCHIP_LAYOUT
, &user_sgpr_idx
, 1);
916 case MESA_SHADER_GEOMETRY
:
917 set_userdata_location_shader(ctx
, AC_UD_GS_VS_RING_STRIDE_ENTRIES
, &user_sgpr_idx
, 2);
919 case MESA_SHADER_FRAGMENT
:
920 if (ctx
->shader_info
->info
.ps
.needs_sample_positions
) {
921 set_userdata_location_shader(ctx
, AC_UD_PS_SAMPLE_POS_OFFSET
, &user_sgpr_idx
, 1);
925 unreachable("Shader stage not implemented");
929 static void setup_types(struct nir_to_llvm_context
*ctx
)
931 LLVMValueRef args
[4];
933 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
934 ctx
->i1
= LLVMIntTypeInContext(ctx
->context
, 1);
935 ctx
->i8
= LLVMIntTypeInContext(ctx
->context
, 8);
936 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
937 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
938 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
939 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
940 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
941 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
942 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
943 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
944 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
945 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
946 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
947 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
949 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
950 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
951 ctx
->i32zero
= LLVMConstInt(ctx
->i32
, 0, false);
952 ctx
->i32one
= LLVMConstInt(ctx
->i32
, 1, false);
953 ctx
->f32zero
= LLVMConstReal(ctx
->f32
, 0.0);
954 ctx
->f32one
= LLVMConstReal(ctx
->f32
, 1.0);
956 args
[0] = ctx
->f32zero
;
957 args
[1] = ctx
->f32zero
;
958 args
[2] = ctx
->f32zero
;
959 args
[3] = ctx
->f32one
;
960 ctx
->v4f32empty
= LLVMConstVector(args
, 4);
962 ctx
->uniform_md_kind
=
963 LLVMGetMDKindIDInContext(ctx
->context
, "amdgpu.uniform", 14);
964 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
966 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
969 static int get_llvm_num_components(LLVMValueRef value
)
971 LLVMTypeRef type
= LLVMTypeOf(value
);
972 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
973 ? LLVMGetVectorSize(type
)
975 return num_components
;
978 static LLVMValueRef
llvm_extract_elem(struct ac_llvm_context
*ac
,
982 int count
= get_llvm_num_components(value
);
984 assert(index
< count
);
988 return LLVMBuildExtractElement(ac
->builder
, value
,
989 LLVMConstInt(ac
->i32
, index
, false), "");
992 static LLVMValueRef
trim_vector(struct ac_llvm_context
*ctx
,
993 LLVMValueRef value
, unsigned count
)
995 unsigned num_components
= get_llvm_num_components(value
);
996 if (count
== num_components
)
999 LLVMValueRef masks
[] = {
1000 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
1001 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false)};
1004 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
1007 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
1008 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
1012 build_store_values_extended(struct ac_llvm_context
*ac
,
1013 LLVMValueRef
*values
,
1014 unsigned value_count
,
1015 unsigned value_stride
,
1018 LLVMBuilderRef builder
= ac
->builder
;
1021 if (value_count
== 1) {
1022 LLVMBuildStore(builder
, vec
, values
[0]);
1026 for (i
= 0; i
< value_count
; i
++) {
1027 LLVMValueRef ptr
= values
[i
* value_stride
];
1028 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
1029 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
1030 LLVMBuildStore(builder
, value
, ptr
);
1034 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
1035 const nir_ssa_def
*def
)
1037 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
1038 if (def
->num_components
> 1) {
1039 type
= LLVMVectorType(type
, def
->num_components
);
1044 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
1047 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, src
.ssa
);
1048 return (LLVMValueRef
)entry
->data
;
1052 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
1053 const struct nir_block
*b
)
1055 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
1056 return (LLVMBasicBlockRef
)entry
->data
;
1059 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
1061 unsigned num_components
)
1063 LLVMValueRef value
= get_src(ctx
, src
.src
);
1064 bool need_swizzle
= false;
1067 LLVMTypeRef type
= LLVMTypeOf(value
);
1068 unsigned src_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
1069 ? LLVMGetVectorSize(type
)
1072 for (unsigned i
= 0; i
< num_components
; ++i
) {
1073 assert(src
.swizzle
[i
] < src_components
);
1074 if (src
.swizzle
[i
] != i
)
1075 need_swizzle
= true;
1078 if (need_swizzle
|| num_components
!= src_components
) {
1079 LLVMValueRef masks
[] = {
1080 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
1081 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
1082 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
1083 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
1085 if (src_components
> 1 && num_components
== 1) {
1086 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1088 } else if (src_components
== 1 && num_components
> 1) {
1089 LLVMValueRef values
[] = {value
, value
, value
, value
};
1090 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
1092 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
1093 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
1097 assert(!src
.negate
);
1102 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
1103 LLVMIntPredicate pred
, LLVMValueRef src0
,
1106 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
1107 return LLVMBuildSelect(ctx
->builder
, result
,
1108 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1109 LLVMConstInt(ctx
->i32
, 0, false), "");
1112 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
1113 LLVMRealPredicate pred
, LLVMValueRef src0
,
1116 LLVMValueRef result
;
1117 src0
= to_float(ctx
, src0
);
1118 src1
= to_float(ctx
, src1
);
1119 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
1120 return LLVMBuildSelect(ctx
->builder
, result
,
1121 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
1122 LLVMConstInt(ctx
->i32
, 0, false), "");
1125 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
1127 LLVMTypeRef result_type
,
1131 LLVMValueRef params
[] = {
1132 to_float(ctx
, src0
),
1135 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1136 get_elem_bits(ctx
, result_type
));
1137 assert(length
< sizeof(name
));
1138 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
1141 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
1143 LLVMTypeRef result_type
,
1144 LLVMValueRef src0
, LLVMValueRef src1
)
1147 LLVMValueRef params
[] = {
1148 to_float(ctx
, src0
),
1149 to_float(ctx
, src1
),
1152 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1153 get_elem_bits(ctx
, result_type
));
1154 assert(length
< sizeof(name
));
1155 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
1158 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
1160 LLVMTypeRef result_type
,
1161 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1164 LLVMValueRef params
[] = {
1165 to_float(ctx
, src0
),
1166 to_float(ctx
, src1
),
1167 to_float(ctx
, src2
),
1170 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
1171 get_elem_bits(ctx
, result_type
));
1172 assert(length
< sizeof(name
));
1173 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
1176 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
1177 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
1179 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
1181 return LLVMBuildSelect(ctx
->builder
, v
, src1
, src2
, "");
1184 static LLVMValueRef
emit_find_lsb(struct ac_llvm_context
*ctx
,
1187 LLVMValueRef params
[2] = {
1190 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
1191 * add special code to check for x=0. The reason is that
1192 * the LLVM behavior for x=0 is different from what we
1195 * The hardware already implements the correct behavior.
1197 LLVMConstInt(ctx
->i1
, 1, false),
1199 return ac_build_intrinsic(ctx
, "llvm.cttz.i32", ctx
->i32
, params
, 2, AC_FUNC_ATTR_READNONE
);
1202 static LLVMValueRef
emit_ifind_msb(struct ac_llvm_context
*ctx
,
1205 return ac_build_imsb(ctx
, src0
, ctx
->i32
);
1208 static LLVMValueRef
emit_ufind_msb(struct ac_llvm_context
*ctx
,
1211 return ac_build_umsb(ctx
, src0
, ctx
->i32
);
1214 static LLVMValueRef
emit_minmax_int(struct ac_llvm_context
*ctx
,
1215 LLVMIntPredicate pred
,
1216 LLVMValueRef src0
, LLVMValueRef src1
)
1218 return LLVMBuildSelect(ctx
->builder
,
1219 LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, ""),
1224 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
1227 return emit_minmax_int(ctx
, LLVMIntSGT
, src0
,
1228 LLVMBuildNeg(ctx
->builder
, src0
, ""));
1231 static LLVMValueRef
emit_fsign(struct ac_llvm_context
*ctx
,
1234 LLVMValueRef cmp
, val
;
1236 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, ctx
->f32_0
, "");
1237 val
= LLVMBuildSelect(ctx
->builder
, cmp
, ctx
->f32_1
, src0
, "");
1238 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, ctx
->f32_0
, "");
1239 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(ctx
->f32
, -1.0), "");
1243 static LLVMValueRef
emit_isign(struct ac_llvm_context
*ctx
,
1246 LLVMValueRef cmp
, val
;
1248 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, ctx
->i32_0
, "");
1249 val
= LLVMBuildSelect(ctx
->builder
, cmp
, ctx
->i32_1
, src0
, "");
1250 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, ctx
->i32_0
, "");
1251 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(ctx
->i32
, -1, true), "");
1255 static LLVMValueRef
emit_ffract(struct ac_llvm_context
*ctx
,
1258 const char *intr
= "llvm.floor.f32";
1259 LLVMValueRef fsrc0
= to_float(ctx
, src0
);
1260 LLVMValueRef params
[] = {
1263 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
,
1264 ctx
->f32
, params
, 1,
1265 AC_FUNC_ATTR_READNONE
);
1266 return LLVMBuildFSub(ctx
->builder
, fsrc0
, floor
, "");
1269 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
1271 LLVMValueRef src0
, LLVMValueRef src1
)
1273 LLVMTypeRef ret_type
;
1274 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
1276 LLVMValueRef params
[] = { src0
, src1
};
1277 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
1280 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
1281 params
, 2, AC_FUNC_ATTR_READNONE
);
1283 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
1284 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
1288 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
1291 return LLVMBuildAnd(ctx
->builder
, src0
, LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""), "");
1294 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
1297 src0
= to_float(ctx
, src0
);
1298 return LLVMBuildSExt(ctx
->builder
,
1299 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, ctx
->f32_0
, ""),
1303 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
1306 return LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
1309 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
1312 return LLVMBuildSExt(ctx
->builder
,
1313 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, ctx
->i32_0
, ""),
1317 static LLVMValueRef
emit_f2f16(struct nir_to_llvm_context
*ctx
,
1320 LLVMValueRef result
;
1323 src0
= to_float(&ctx
->ac
, src0
);
1324 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->f16
, "");
1326 /* TODO SI/CIK options here */
1327 if (ctx
->options
->chip_class
>= VI
) {
1328 LLVMValueRef args
[2];
1329 /* Check if the result is a denormal - and flush to 0 if so. */
1331 args
[1] = LLVMConstInt(ctx
->i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
1332 cond
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.class.f16", ctx
->i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
1335 /* need to convert back up to f32 */
1336 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f32
, "");
1338 if (ctx
->options
->chip_class
>= VI
)
1339 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32zero
, result
, "");
1344 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
1345 LLVMValueRef src0
, LLVMValueRef src1
)
1347 LLVMValueRef dst64
, result
;
1348 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
1349 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
1351 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1352 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1353 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1357 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
1358 LLVMValueRef src0
, LLVMValueRef src1
)
1360 LLVMValueRef dst64
, result
;
1361 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
1362 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
1364 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
1365 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
1366 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
1370 static LLVMValueRef
emit_bitfield_extract(struct ac_llvm_context
*ctx
,
1372 const LLVMValueRef srcs
[3])
1374 LLVMValueRef result
;
1375 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
1377 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
1378 result
= LLVMBuildSelect(ctx
->builder
, icond
, srcs
[0], result
, "");
1382 static LLVMValueRef
emit_bitfield_insert(struct ac_llvm_context
*ctx
,
1383 LLVMValueRef src0
, LLVMValueRef src1
,
1384 LLVMValueRef src2
, LLVMValueRef src3
)
1386 LLVMValueRef bfi_args
[3], result
;
1388 bfi_args
[0] = LLVMBuildShl(ctx
->builder
,
1389 LLVMBuildSub(ctx
->builder
,
1390 LLVMBuildShl(ctx
->builder
,
1395 bfi_args
[1] = LLVMBuildShl(ctx
->builder
, src1
, src2
, "");
1398 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, src3
, LLVMConstInt(ctx
->i32
, 32, false), "");
1401 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
1402 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
1404 result
= LLVMBuildXor(ctx
->builder
, bfi_args
[2],
1405 LLVMBuildAnd(ctx
->builder
, bfi_args
[0],
1406 LLVMBuildXor(ctx
->builder
, bfi_args
[1], bfi_args
[2], ""), ""), "");
1408 result
= LLVMBuildSelect(ctx
->builder
, icond
, src1
, result
, "");
1412 static LLVMValueRef
emit_pack_half_2x16(struct ac_llvm_context
*ctx
,
1415 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
1417 LLVMValueRef comp
[2];
1419 src0
= to_float(ctx
, src0
);
1420 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
1421 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
1422 for (i
= 0; i
< 2; i
++) {
1423 comp
[i
] = LLVMBuildFPTrunc(ctx
->builder
, comp
[i
], ctx
->f16
, "");
1424 comp
[i
] = LLVMBuildBitCast(ctx
->builder
, comp
[i
], ctx
->i16
, "");
1425 comp
[i
] = LLVMBuildZExt(ctx
->builder
, comp
[i
], ctx
->i32
, "");
1428 comp
[1] = LLVMBuildShl(ctx
->builder
, comp
[1], const16
, "");
1429 comp
[0] = LLVMBuildOr(ctx
->builder
, comp
[0], comp
[1], "");
1434 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
1437 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
1438 LLVMValueRef temps
[2], result
, val
;
1441 for (i
= 0; i
< 2; i
++) {
1442 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
1443 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
1444 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
1445 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
1448 LLVMTypeRef v2f32
= LLVMVectorType(ctx
->f32
, 2);
1449 result
= LLVMBuildInsertElement(ctx
->builder
, LLVMGetUndef(v2f32
), temps
[0],
1451 result
= LLVMBuildInsertElement(ctx
->builder
, result
, temps
[1],
1456 static LLVMValueRef
emit_ddxy(struct nir_to_llvm_context
*ctx
,
1462 LLVMValueRef result
;
1464 if (!ctx
->lds
&& !ctx
->has_ds_bpermute
)
1465 ctx
->lds
= LLVMAddGlobalInAddressSpace(ctx
->module
,
1466 LLVMArrayType(ctx
->i32
, 64),
1467 "ddxy_lds", LOCAL_ADDR_SPACE
);
1469 if (op
== nir_op_fddx_fine
|| op
== nir_op_fddx
)
1470 mask
= AC_TID_MASK_LEFT
;
1471 else if (op
== nir_op_fddy_fine
|| op
== nir_op_fddy
)
1472 mask
= AC_TID_MASK_TOP
;
1474 mask
= AC_TID_MASK_TOP_LEFT
;
1476 /* for DDX we want to next X pixel, DDY next Y pixel. */
1477 if (op
== nir_op_fddx_fine
||
1478 op
== nir_op_fddx_coarse
||
1484 result
= ac_build_ddxy(&ctx
->ac
, ctx
->has_ds_bpermute
,
1485 mask
, idx
, ctx
->lds
,
1491 * this takes an I,J coordinate pair,
1492 * and works out the X and Y derivatives.
1493 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
1495 static LLVMValueRef
emit_ddxy_interp(
1496 struct nir_to_llvm_context
*ctx
,
1497 LLVMValueRef interp_ij
)
1499 LLVMValueRef result
[4], a
;
1502 for (i
= 0; i
< 2; i
++) {
1503 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
1504 LLVMConstInt(ctx
->i32
, i
, false), "");
1505 result
[i
] = emit_ddxy(ctx
, nir_op_fddx
, a
);
1506 result
[2+i
] = emit_ddxy(ctx
, nir_op_fddy
, a
);
1508 return ac_build_gather_values(&ctx
->ac
, result
, 4);
1511 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
1513 LLVMValueRef src
[4], result
= NULL
;
1514 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
1515 unsigned src_components
;
1516 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
1518 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
1519 switch (instr
->op
) {
1525 case nir_op_pack_half_2x16
:
1528 case nir_op_unpack_half_2x16
:
1532 src_components
= num_components
;
1535 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1536 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
1538 switch (instr
->op
) {
1544 src
[0] = to_float(&ctx
->ac
, src
[0]);
1545 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
1548 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
1551 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
1554 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1557 src
[0] = to_float(&ctx
->ac
, src
[0]);
1558 src
[1] = to_float(&ctx
->ac
, src
[1]);
1559 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
1562 src
[0] = to_float(&ctx
->ac
, src
[0]);
1563 src
[1] = to_float(&ctx
->ac
, src
[1]);
1564 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1567 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
1570 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1573 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1576 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
1579 src
[0] = to_float(&ctx
->ac
, src
[0]);
1580 src
[1] = to_float(&ctx
->ac
, src
[1]);
1581 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1582 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1583 to_float_type(&ctx
->ac
, def_type
), result
);
1584 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
1585 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
1588 src
[0] = to_float(&ctx
->ac
, src
[0]);
1589 src
[1] = to_float(&ctx
->ac
, src
[1]);
1590 result
= LLVMBuildFRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1593 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
1596 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1599 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
1602 src
[0] = to_float(&ctx
->ac
, src
[0]);
1603 src
[1] = to_float(&ctx
->ac
, src
[1]);
1604 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
1607 src
[0] = to_float(&ctx
->ac
, src
[0]);
1608 src
[1] = to_float(&ctx
->ac
, src
[1]);
1609 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
1612 src
[0] = to_float(&ctx
->ac
, src
[0]);
1613 result
= ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, src
[0]);
1616 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
1619 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
1622 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
1625 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0],
1626 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1627 LLVMTypeOf(src
[0]), ""),
1631 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0],
1632 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1633 LLVMTypeOf(src
[0]), ""),
1637 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0],
1638 LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
1639 LLVMTypeOf(src
[0]), ""),
1643 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1646 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
1649 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
1652 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
1655 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1658 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
1661 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUEQ
, src
[0], src
[1]);
1664 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
1667 result
= emit_float_cmp(&ctx
->ac
, LLVMRealULT
, src
[0], src
[1]);
1670 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUGE
, src
[0], src
[1]);
1673 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
1674 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1677 result
= emit_iabs(&ctx
->ac
, src
[0]);
1680 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, src
[0], src
[1]);
1683 result
= emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
1686 result
= emit_minmax_int(&ctx
->ac
, LLVMIntUGT
, src
[0], src
[1]);
1689 result
= emit_minmax_int(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
1692 result
= emit_isign(&ctx
->ac
, src
[0]);
1695 src
[0] = to_float(&ctx
->ac
, src
[0]);
1696 result
= emit_fsign(&ctx
->ac
, src
[0]);
1699 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
1700 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1703 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
1704 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1707 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
1708 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1710 case nir_op_fround_even
:
1711 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
1712 to_float_type(&ctx
->ac
, def_type
),src
[0]);
1715 result
= emit_ffract(&ctx
->ac
, src
[0]);
1718 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
1719 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1722 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
1723 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1726 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1727 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1730 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
1731 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1734 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
1735 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1738 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
1739 to_float_type(&ctx
->ac
, def_type
), src
[0]);
1740 result
= ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, result
);
1743 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
1744 to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1747 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1748 to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1749 if (instr
->dest
.dest
.ssa
.bit_size
== 32)
1750 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1751 to_float_type(&ctx
->ac
, def_type
),
1755 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1756 to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1757 if (instr
->dest
.dest
.ssa
.bit_size
== 32)
1758 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
1759 to_float_type(&ctx
->ac
, def_type
),
1763 result
= emit_intrin_3f_param(&ctx
->ac
, "llvm.fma",
1764 to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
1766 case nir_op_ibitfield_extract
:
1767 result
= emit_bitfield_extract(&ctx
->ac
, true, src
);
1769 case nir_op_ubitfield_extract
:
1770 result
= emit_bitfield_extract(&ctx
->ac
, false, src
);
1772 case nir_op_bitfield_insert
:
1773 result
= emit_bitfield_insert(&ctx
->ac
, src
[0], src
[1], src
[2], src
[3]);
1775 case nir_op_bitfield_reverse
:
1776 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.bitreverse.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1778 case nir_op_bit_count
:
1779 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i32", ctx
->ac
.i32
, src
, 1, AC_FUNC_ATTR_READNONE
);
1784 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1785 src
[i
] = to_integer(&ctx
->ac
, src
[i
]);
1786 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
1790 src
[0] = to_float(&ctx
->ac
, src
[0]);
1791 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
1795 src
[0] = to_float(&ctx
->ac
, src
[0]);
1796 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
1800 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], to_float_type(&ctx
->ac
, def_type
), "");
1804 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], to_float_type(&ctx
->ac
, def_type
), "");
1807 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], to_float_type(&ctx
->ac
, def_type
), "");
1810 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], to_float_type(&ctx
->ac
, def_type
), "");
1814 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < get_elem_bits(&ctx
->ac
, def_type
))
1815 result
= LLVMBuildZExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1817 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1821 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < get_elem_bits(&ctx
->ac
, def_type
))
1822 result
= LLVMBuildSExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1824 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1827 result
= emit_bcsel(&ctx
->ac
, src
[0], src
[1], src
[2]);
1829 case nir_op_find_lsb
:
1830 result
= emit_find_lsb(&ctx
->ac
, src
[0]);
1832 case nir_op_ufind_msb
:
1833 result
= emit_ufind_msb(&ctx
->ac
, src
[0]);
1835 case nir_op_ifind_msb
:
1836 result
= emit_ifind_msb(&ctx
->ac
, src
[0]);
1838 case nir_op_uadd_carry
:
1839 result
= emit_uint_carry(&ctx
->ac
, "llvm.uadd.with.overflow.i32", src
[0], src
[1]);
1841 case nir_op_usub_borrow
:
1842 result
= emit_uint_carry(&ctx
->ac
, "llvm.usub.with.overflow.i32", src
[0], src
[1]);
1845 result
= emit_b2f(&ctx
->ac
, src
[0]);
1848 result
= emit_f2b(&ctx
->ac
, src
[0]);
1851 result
= emit_b2i(&ctx
->ac
, src
[0]);
1854 result
= emit_i2b(&ctx
->ac
, src
[0]);
1856 case nir_op_fquantize2f16
:
1857 result
= emit_f2f16(ctx
->nctx
, src
[0]);
1859 case nir_op_umul_high
:
1860 result
= emit_umul_high(&ctx
->ac
, src
[0], src
[1]);
1862 case nir_op_imul_high
:
1863 result
= emit_imul_high(&ctx
->ac
, src
[0], src
[1]);
1865 case nir_op_pack_half_2x16
:
1866 result
= emit_pack_half_2x16(&ctx
->ac
, src
[0]);
1868 case nir_op_unpack_half_2x16
:
1869 result
= emit_unpack_half_2x16(&ctx
->ac
, src
[0]);
1873 case nir_op_fddx_fine
:
1874 case nir_op_fddy_fine
:
1875 case nir_op_fddx_coarse
:
1876 case nir_op_fddy_coarse
:
1877 result
= emit_ddxy(ctx
->nctx
, instr
->op
, src
[0]);
1880 case nir_op_unpack_64_2x32_split_x
: {
1881 assert(instr
->src
[0].src
.ssa
->num_components
== 1);
1882 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1883 LLVMVectorType(ctx
->ac
.i32
, 2),
1885 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1890 case nir_op_unpack_64_2x32_split_y
: {
1891 assert(instr
->src
[0].src
.ssa
->num_components
== 1);
1892 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1893 LLVMVectorType(ctx
->ac
.i32
, 2),
1895 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1900 case nir_op_pack_64_2x32_split
: {
1901 LLVMValueRef tmp
= LLVMGetUndef(LLVMVectorType(ctx
->ac
.i32
, 2));
1902 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
1903 src
[0], ctx
->ac
.i32_0
, "");
1904 tmp
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp
,
1905 src
[1], ctx
->ac
.i32_1
, "");
1906 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i64
, "");
1911 fprintf(stderr
, "Unknown NIR alu instr: ");
1912 nir_print_instr(&instr
->instr
, stderr
);
1913 fprintf(stderr
, "\n");
1918 assert(instr
->dest
.dest
.is_ssa
);
1919 result
= to_integer(&ctx
->ac
, result
);
1920 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.dest
.ssa
,
1925 static void visit_load_const(struct ac_nir_context
*ctx
,
1926 const nir_load_const_instr
*instr
)
1928 LLVMValueRef values
[4], value
= NULL
;
1929 LLVMTypeRef element_type
=
1930 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
1932 for (unsigned i
= 0; i
< instr
->def
.num_components
; ++i
) {
1933 switch (instr
->def
.bit_size
) {
1935 values
[i
] = LLVMConstInt(element_type
,
1936 instr
->value
.u32
[i
], false);
1939 values
[i
] = LLVMConstInt(element_type
,
1940 instr
->value
.u64
[i
], false);
1944 "unsupported nir load_const bit_size: %d\n",
1945 instr
->def
.bit_size
);
1949 if (instr
->def
.num_components
> 1) {
1950 value
= LLVMConstVector(values
, instr
->def
.num_components
);
1954 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, value
);
1957 static LLVMValueRef
cast_ptr(struct nir_to_llvm_context
*ctx
, LLVMValueRef ptr
,
1960 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
1961 return LLVMBuildBitCast(ctx
->builder
, ptr
,
1962 LLVMPointerType(type
, addr_space
), "");
1966 get_buffer_size(struct ac_nir_context
*ctx
, LLVMValueRef descriptor
, bool in_elements
)
1969 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1970 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1973 if (ctx
->abi
->chip_class
>= VI
&& in_elements
) {
1974 /* On VI, the descriptor contains the size in bytes,
1975 * but TXQ must return the size in elements.
1976 * The stride is always non-zero for resources using TXQ.
1978 LLVMValueRef stride
=
1979 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1980 LLVMConstInt(ctx
->ac
.i32
, 1, false), "");
1981 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
,
1982 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
1983 stride
= LLVMBuildAnd(ctx
->ac
.builder
, stride
,
1984 LLVMConstInt(ctx
->ac
.i32
, 0x3fff, false), "");
1986 size
= LLVMBuildUDiv(ctx
->ac
.builder
, size
, stride
, "");
1992 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
1995 static void build_int_type_name(
1997 char *buf
, unsigned bufsize
)
1999 assert(bufsize
>= 6);
2001 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
2002 snprintf(buf
, bufsize
, "v%ui32",
2003 LLVMGetVectorSize(type
));
2008 static LLVMValueRef
radv_lower_gather4_integer(struct ac_llvm_context
*ctx
,
2009 struct ac_image_args
*args
,
2010 const nir_tex_instr
*instr
)
2012 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2013 LLVMValueRef coord
= args
->addr
;
2014 LLVMValueRef half_texel
[2];
2015 LLVMValueRef compare_cube_wa
;
2016 LLVMValueRef result
;
2018 unsigned coord_vgpr_index
= (unsigned)args
->offset
+ (unsigned)args
->compare
;
2022 struct ac_image_args txq_args
= { 0 };
2024 txq_args
.da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
2025 txq_args
.opcode
= ac_image_get_resinfo
;
2026 txq_args
.dmask
= 0xf;
2027 txq_args
.addr
= ctx
->i32_0
;
2028 txq_args
.resource
= args
->resource
;
2029 LLVMValueRef size
= ac_build_image_opcode(ctx
, &txq_args
);
2031 for (c
= 0; c
< 2; c
++) {
2032 half_texel
[c
] = LLVMBuildExtractElement(ctx
->builder
, size
,
2033 LLVMConstInt(ctx
->i32
, c
, false), "");
2034 half_texel
[c
] = LLVMBuildUIToFP(ctx
->builder
, half_texel
[c
], ctx
->f32
, "");
2035 half_texel
[c
] = ac_build_fdiv(ctx
, ctx
->f32_1
, half_texel
[c
]);
2036 half_texel
[c
] = LLVMBuildFMul(ctx
->builder
, half_texel
[c
],
2037 LLVMConstReal(ctx
->f32
, -0.5), "");
2041 LLVMValueRef orig_coords
= args
->addr
;
2043 for (c
= 0; c
< 2; c
++) {
2045 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, coord_vgpr_index
+ c
, 0);
2046 tmp
= LLVMBuildExtractElement(ctx
->builder
, coord
, index
, "");
2047 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2048 tmp
= LLVMBuildFAdd(ctx
->builder
, tmp
, half_texel
[c
], "");
2049 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2050 coord
= LLVMBuildInsertElement(ctx
->builder
, coord
, tmp
, index
, "");
2055 * Apparantly cube has issue with integer types that the workaround doesn't solve,
2056 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
2057 * workaround by sampling using a scaled type and converting.
2058 * This is taken from amdgpu-pro shaders.
2060 /* NOTE this produces some ugly code compared to amdgpu-pro,
2061 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
2062 * and then reads them back. -pro generates two selects,
2063 * one s_cmp for the descriptor rewriting
2064 * one v_cmp for the coordinate and result changes.
2066 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2067 LLVMValueRef tmp
, tmp2
;
2069 /* workaround 8/8/8/8 uint/sint cube gather bug */
2070 /* first detect it then change to a scaled read and f2i */
2071 tmp
= LLVMBuildExtractElement(ctx
->builder
, args
->resource
, ctx
->i32_1
, "");
2074 /* extract the DATA_FORMAT */
2075 tmp
= ac_build_bfe(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 20, false),
2076 LLVMConstInt(ctx
->i32
, 6, false), false);
2078 /* is the DATA_FORMAT == 8_8_8_8 */
2079 compare_cube_wa
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tmp
, LLVMConstInt(ctx
->i32
, V_008F14_IMG_DATA_FORMAT_8_8_8_8
, false), "");
2081 if (stype
== GLSL_TYPE_UINT
)
2082 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
2083 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0x8000000, false),
2084 LLVMConstInt(ctx
->i32
, 0x10000000, false), "");
2086 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
2087 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0xc000000, false),
2088 LLVMConstInt(ctx
->i32
, 0x14000000, false), "");
2090 /* replace the NUM FORMAT in the descriptor */
2091 tmp2
= LLVMBuildAnd(ctx
->builder
, tmp2
, LLVMConstInt(ctx
->i32
, C_008F14_NUM_FORMAT_GFX6
, false), "");
2092 tmp2
= LLVMBuildOr(ctx
->builder
, tmp2
, tmp
, "");
2094 args
->resource
= LLVMBuildInsertElement(ctx
->builder
, args
->resource
, tmp2
, ctx
->i32_1
, "");
2096 /* don't modify the coordinates for this case */
2097 coord
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, orig_coords
, coord
, "");
2100 result
= ac_build_image_opcode(ctx
, args
);
2102 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2103 LLVMValueRef tmp
, tmp2
;
2105 /* if the cube workaround is in place, f2i the result. */
2106 for (c
= 0; c
< 4; c
++) {
2107 tmp
= LLVMBuildExtractElement(ctx
->builder
, result
, LLVMConstInt(ctx
->i32
, c
, false), "");
2108 if (stype
== GLSL_TYPE_UINT
)
2109 tmp2
= LLVMBuildFPToUI(ctx
->builder
, tmp
, ctx
->i32
, "");
2111 tmp2
= LLVMBuildFPToSI(ctx
->builder
, tmp
, ctx
->i32
, "");
2112 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
2113 tmp2
= LLVMBuildBitCast(ctx
->builder
, tmp2
, ctx
->i32
, "");
2114 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, tmp2
, tmp
, "");
2115 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
2116 result
= LLVMBuildInsertElement(ctx
->builder
, result
, tmp
, LLVMConstInt(ctx
->i32
, c
, false), "");
2122 static LLVMValueRef
build_tex_intrinsic(struct ac_nir_context
*ctx
,
2123 const nir_tex_instr
*instr
,
2125 struct ac_image_args
*args
)
2127 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
2128 return ac_build_buffer_load_format(&ctx
->ac
,
2131 LLVMConstInt(ctx
->ac
.i32
, 0, false),
2135 args
->opcode
= ac_image_sample
;
2136 args
->compare
= instr
->is_shadow
;
2138 switch (instr
->op
) {
2140 case nir_texop_txf_ms
:
2141 case nir_texop_samples_identical
:
2142 args
->opcode
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
? ac_image_load
: ac_image_load_mip
;
2143 args
->compare
= false;
2144 args
->offset
= false;
2151 args
->level_zero
= true;
2156 case nir_texop_query_levels
:
2157 args
->opcode
= ac_image_get_resinfo
;
2160 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
)
2161 args
->level_zero
= true;
2167 args
->opcode
= ac_image_gather4
;
2168 args
->level_zero
= true;
2171 args
->opcode
= ac_image_get_lod
;
2172 args
->compare
= false;
2173 args
->offset
= false;
2179 if (instr
->op
== nir_texop_tg4
) {
2180 enum glsl_base_type stype
= glsl_get_sampler_result_type(instr
->texture
->var
->type
);
2181 if (stype
== GLSL_TYPE_UINT
|| stype
== GLSL_TYPE_INT
) {
2182 return radv_lower_gather4_integer(&ctx
->ac
, args
, instr
);
2185 return ac_build_image_opcode(&ctx
->ac
, args
);
2188 static LLVMValueRef
visit_vulkan_resource_index(struct nir_to_llvm_context
*ctx
,
2189 nir_intrinsic_instr
*instr
)
2191 LLVMValueRef index
= get_src(ctx
->nir
, instr
->src
[0]);
2192 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
2193 unsigned binding
= nir_intrinsic_binding(instr
);
2194 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
2195 struct radv_pipeline_layout
*pipeline_layout
= ctx
->options
->layout
;
2196 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
2197 unsigned base_offset
= layout
->binding
[binding
].offset
;
2198 LLVMValueRef offset
, stride
;
2200 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
2201 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
2202 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
2203 layout
->binding
[binding
].dynamic_offset_offset
;
2204 desc_ptr
= ctx
->push_constants
;
2205 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
2206 stride
= LLVMConstInt(ctx
->i32
, 16, false);
2208 stride
= LLVMConstInt(ctx
->i32
, layout
->binding
[binding
].size
, false);
2210 offset
= LLVMConstInt(ctx
->i32
, base_offset
, false);
2211 index
= LLVMBuildMul(ctx
->builder
, index
, stride
, "");
2212 offset
= LLVMBuildAdd(ctx
->builder
, offset
, index
, "");
2214 desc_ptr
= ac_build_gep0(&ctx
->ac
, desc_ptr
, offset
);
2215 desc_ptr
= cast_ptr(ctx
, desc_ptr
, ctx
->v4i32
);
2216 LLVMSetMetadata(desc_ptr
, ctx
->uniform_md_kind
, ctx
->empty_md
);
2218 return LLVMBuildLoad(ctx
->builder
, desc_ptr
, "");
2221 static LLVMValueRef
visit_load_push_constant(struct nir_to_llvm_context
*ctx
,
2222 nir_intrinsic_instr
*instr
)
2224 LLVMValueRef ptr
, addr
;
2226 addr
= LLVMConstInt(ctx
->i32
, nir_intrinsic_base(instr
), 0);
2227 addr
= LLVMBuildAdd(ctx
->builder
, addr
, get_src(ctx
->nir
, instr
->src
[0]), "");
2229 ptr
= ac_build_gep0(&ctx
->ac
, ctx
->push_constants
, addr
);
2230 ptr
= cast_ptr(ctx
, ptr
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
));
2232 return LLVMBuildLoad(ctx
->builder
, ptr
, "");
2235 static LLVMValueRef
visit_get_buffer_size(struct ac_nir_context
*ctx
,
2236 const nir_intrinsic_instr
*instr
)
2238 LLVMValueRef desc
= get_src(ctx
, instr
->src
[0]);
2240 return get_buffer_size(ctx
, desc
, false);
2242 static void visit_store_ssbo(struct nir_to_llvm_context
*ctx
,
2243 nir_intrinsic_instr
*instr
)
2245 const char *store_name
;
2246 LLVMValueRef src_data
= get_src(ctx
->nir
, instr
->src
[0]);
2247 LLVMTypeRef data_type
= ctx
->f32
;
2248 int elem_size_mult
= get_elem_bits(&ctx
->ac
, LLVMTypeOf(src_data
)) / 32;
2249 int components_32bit
= elem_size_mult
* instr
->num_components
;
2250 unsigned writemask
= nir_intrinsic_write_mask(instr
);
2251 LLVMValueRef base_data
, base_offset
;
2252 LLVMValueRef params
[6];
2254 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2255 ctx
->shader_info
->fs
.writes_memory
= true;
2257 params
[1] = get_src(ctx
->nir
, instr
->src
[1]);
2258 params
[2] = LLVMConstInt(ctx
->i32
, 0, false); /* vindex */
2259 params
[4] = ctx
->i1false
; /* glc */
2260 params
[5] = ctx
->i1false
; /* slc */
2262 if (components_32bit
> 1)
2263 data_type
= LLVMVectorType(ctx
->f32
, components_32bit
);
2265 base_data
= to_float(&ctx
->ac
, src_data
);
2266 base_data
= trim_vector(&ctx
->ac
, base_data
, instr
->num_components
);
2267 base_data
= LLVMBuildBitCast(ctx
->builder
, base_data
,
2269 base_offset
= get_src(ctx
->nir
, instr
->src
[2]); /* voffset */
2273 LLVMValueRef offset
;
2275 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
2277 /* Due to an LLVM limitation, split 3-element writes
2278 * into a 2-element and a 1-element write. */
2280 writemask
|= 1 << (start
+ 2);
2284 start
*= elem_size_mult
;
2285 count
*= elem_size_mult
;
2288 writemask
|= ((1u << (count
- 4)) - 1u) << (start
+ 4);
2293 store_name
= "llvm.amdgcn.buffer.store.v4f32";
2295 } else if (count
== 2) {
2296 tmp
= LLVMBuildExtractElement(ctx
->builder
,
2297 base_data
, LLVMConstInt(ctx
->i32
, start
, false), "");
2298 data
= LLVMBuildInsertElement(ctx
->builder
, LLVMGetUndef(ctx
->v2f32
), tmp
,
2301 tmp
= LLVMBuildExtractElement(ctx
->builder
,
2302 base_data
, LLVMConstInt(ctx
->i32
, start
+ 1, false), "");
2303 data
= LLVMBuildInsertElement(ctx
->builder
, data
, tmp
,
2305 store_name
= "llvm.amdgcn.buffer.store.v2f32";
2309 if (get_llvm_num_components(base_data
) > 1)
2310 data
= LLVMBuildExtractElement(ctx
->builder
, base_data
,
2311 LLVMConstInt(ctx
->i32
, start
, false), "");
2314 store_name
= "llvm.amdgcn.buffer.store.f32";
2317 offset
= base_offset
;
2319 offset
= LLVMBuildAdd(ctx
->builder
, offset
, LLVMConstInt(ctx
->i32
, start
* 4, false), "");
2323 ac_build_intrinsic(&ctx
->ac
, store_name
,
2324 ctx
->voidt
, params
, 6, 0);
2328 static LLVMValueRef
visit_atomic_ssbo(struct nir_to_llvm_context
*ctx
,
2329 const nir_intrinsic_instr
*instr
)
2332 LLVMValueRef params
[6];
2334 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2335 ctx
->shader_info
->fs
.writes_memory
= true;
2337 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
) {
2338 params
[arg_count
++] = llvm_extract_elem(&ctx
->ac
, get_src(ctx
->nir
, instr
->src
[3]), 0);
2340 params
[arg_count
++] = llvm_extract_elem(&ctx
->ac
, get_src(ctx
->nir
, instr
->src
[2]), 0);
2341 params
[arg_count
++] = get_src(ctx
->nir
, instr
->src
[0]);
2342 params
[arg_count
++] = LLVMConstInt(ctx
->i32
, 0, false); /* vindex */
2343 params
[arg_count
++] = get_src(ctx
->nir
, instr
->src
[1]); /* voffset */
2344 params
[arg_count
++] = ctx
->i1false
; /* slc */
2346 switch (instr
->intrinsic
) {
2347 case nir_intrinsic_ssbo_atomic_add
:
2348 name
= "llvm.amdgcn.buffer.atomic.add";
2350 case nir_intrinsic_ssbo_atomic_imin
:
2351 name
= "llvm.amdgcn.buffer.atomic.smin";
2353 case nir_intrinsic_ssbo_atomic_umin
:
2354 name
= "llvm.amdgcn.buffer.atomic.umin";
2356 case nir_intrinsic_ssbo_atomic_imax
:
2357 name
= "llvm.amdgcn.buffer.atomic.smax";
2359 case nir_intrinsic_ssbo_atomic_umax
:
2360 name
= "llvm.amdgcn.buffer.atomic.umax";
2362 case nir_intrinsic_ssbo_atomic_and
:
2363 name
= "llvm.amdgcn.buffer.atomic.and";
2365 case nir_intrinsic_ssbo_atomic_or
:
2366 name
= "llvm.amdgcn.buffer.atomic.or";
2368 case nir_intrinsic_ssbo_atomic_xor
:
2369 name
= "llvm.amdgcn.buffer.atomic.xor";
2371 case nir_intrinsic_ssbo_atomic_exchange
:
2372 name
= "llvm.amdgcn.buffer.atomic.swap";
2374 case nir_intrinsic_ssbo_atomic_comp_swap
:
2375 name
= "llvm.amdgcn.buffer.atomic.cmpswap";
2381 return ac_build_intrinsic(&ctx
->ac
, name
, ctx
->i32
, params
, arg_count
, 0);
2384 static LLVMValueRef
visit_load_buffer(struct nir_to_llvm_context
*ctx
,
2385 const nir_intrinsic_instr
*instr
)
2387 LLVMValueRef results
[2];
2388 int load_components
;
2389 int num_components
= instr
->num_components
;
2390 if (instr
->dest
.ssa
.bit_size
== 64)
2391 num_components
*= 2;
2393 for (int i
= 0; i
< num_components
; i
+= load_components
) {
2394 load_components
= MIN2(num_components
- i
, 4);
2395 const char *load_name
;
2396 LLVMTypeRef data_type
= ctx
->f32
;
2397 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, i
* 4, false);
2398 offset
= LLVMBuildAdd(ctx
->builder
, get_src(ctx
->nir
, instr
->src
[1]), offset
, "");
2400 if (load_components
== 3)
2401 data_type
= LLVMVectorType(ctx
->f32
, 4);
2402 else if (load_components
> 1)
2403 data_type
= LLVMVectorType(ctx
->f32
, load_components
);
2405 if (load_components
>= 3)
2406 load_name
= "llvm.amdgcn.buffer.load.v4f32";
2407 else if (load_components
== 2)
2408 load_name
= "llvm.amdgcn.buffer.load.v2f32";
2409 else if (load_components
== 1)
2410 load_name
= "llvm.amdgcn.buffer.load.f32";
2412 unreachable("unhandled number of components");
2414 LLVMValueRef params
[] = {
2415 get_src(ctx
->nir
, instr
->src
[0]),
2416 LLVMConstInt(ctx
->i32
, 0, false),
2422 results
[i
] = ac_build_intrinsic(&ctx
->ac
, load_name
, data_type
, params
, 5, 0);
2426 LLVMValueRef ret
= results
[0];
2427 if (num_components
> 4 || num_components
== 3) {
2428 LLVMValueRef masks
[] = {
2429 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
2430 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false),
2431 LLVMConstInt(ctx
->i32
, 4, false), LLVMConstInt(ctx
->i32
, 5, false),
2432 LLVMConstInt(ctx
->i32
, 6, false), LLVMConstInt(ctx
->i32
, 7, false)
2435 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
2436 ret
= LLVMBuildShuffleVector(ctx
->builder
, results
[0],
2437 results
[num_components
> 4 ? 1 : 0], swizzle
, "");
2440 return LLVMBuildBitCast(ctx
->builder
, ret
,
2441 get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
2444 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
2445 const nir_intrinsic_instr
*instr
)
2447 LLVMValueRef results
[8], ret
;
2448 LLVMValueRef rsrc
= get_src(ctx
, instr
->src
[0]);
2449 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
2450 int num_components
= instr
->num_components
;
2452 if (ctx
->abi
->load_ubo
)
2453 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
2455 if (instr
->dest
.ssa
.bit_size
== 64)
2456 num_components
*= 2;
2458 for (unsigned i
= 0; i
< num_components
; ++i
) {
2459 LLVMValueRef params
[] = {
2461 LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, 4 * i
, 0),
2464 results
[i
] = ac_build_intrinsic(&ctx
->ac
, "llvm.SI.load.const.v4i32", ctx
->ac
.f32
,
2466 AC_FUNC_ATTR_READNONE
|
2467 AC_FUNC_ATTR_LEGACY
);
2471 ret
= ac_build_gather_values(&ctx
->ac
, results
, instr
->num_components
);
2472 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2473 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2477 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_var
*deref
,
2478 bool vs_in
, unsigned *vertex_index_out
,
2479 LLVMValueRef
*vertex_index_ref
,
2480 unsigned *const_out
, LLVMValueRef
*indir_out
)
2482 unsigned const_offset
= 0;
2483 nir_deref
*tail
= &deref
->deref
;
2484 LLVMValueRef offset
= NULL
;
2486 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
2488 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2489 if (vertex_index_out
)
2490 *vertex_index_out
= deref_array
->base_offset
;
2492 if (vertex_index_ref
) {
2493 LLVMValueRef vtx
= LLVMConstInt(ctx
->ac
.i32
, deref_array
->base_offset
, false);
2494 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
2495 vtx
= LLVMBuildAdd(ctx
->ac
.builder
, vtx
, get_src(ctx
, deref_array
->indirect
), "");
2497 *vertex_index_ref
= vtx
;
2501 if (deref
->var
->data
.compact
) {
2502 assert(tail
->child
->deref_type
== nir_deref_type_array
);
2503 assert(glsl_type_is_scalar(glsl_without_array(deref
->var
->type
)));
2504 nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
2505 /* We always lower indirect dereferences for "compact" array vars. */
2506 assert(deref_array
->deref_array_type
== nir_deref_array_type_direct
);
2508 const_offset
= deref_array
->base_offset
;
2512 while (tail
->child
!= NULL
) {
2513 const struct glsl_type
*parent_type
= tail
->type
;
2516 if (tail
->deref_type
== nir_deref_type_array
) {
2517 nir_deref_array
*deref_array
= nir_deref_as_array(tail
);
2518 LLVMValueRef index
, stride
, local_offset
;
2519 unsigned size
= glsl_count_attribute_slots(tail
->type
, vs_in
);
2521 const_offset
+= size
* deref_array
->base_offset
;
2522 if (deref_array
->deref_array_type
== nir_deref_array_type_direct
)
2525 assert(deref_array
->deref_array_type
== nir_deref_array_type_indirect
);
2526 index
= get_src(ctx
, deref_array
->indirect
);
2527 stride
= LLVMConstInt(ctx
->ac
.i32
, size
, 0);
2528 local_offset
= LLVMBuildMul(ctx
->ac
.builder
, stride
, index
, "");
2531 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, local_offset
, "");
2533 offset
= local_offset
;
2534 } else if (tail
->deref_type
== nir_deref_type_struct
) {
2535 nir_deref_struct
*deref_struct
= nir_deref_as_struct(tail
);
2537 for (unsigned i
= 0; i
< deref_struct
->index
; i
++) {
2538 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
2539 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
2542 unreachable("unsupported deref type");
2546 if (const_offset
&& offset
)
2547 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
2548 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
2551 *const_out
= const_offset
;
2552 *indir_out
= offset
;
2556 lds_load(struct nir_to_llvm_context
*ctx
,
2557 LLVMValueRef dw_addr
)
2560 value
= ac_build_indexed_load(&ctx
->ac
, ctx
->lds
, dw_addr
, false);
2565 lds_store(struct nir_to_llvm_context
*ctx
,
2566 LLVMValueRef dw_addr
, LLVMValueRef value
)
2568 value
= LLVMBuildBitCast(ctx
->builder
, value
, ctx
->i32
, "");
2569 ac_build_indexed_store(&ctx
->ac
, ctx
->lds
,
2573 /* The offchip buffer layout for TCS->TES is
2575 * - attribute 0 of patch 0 vertex 0
2576 * - attribute 0 of patch 0 vertex 1
2577 * - attribute 0 of patch 0 vertex 2
2579 * - attribute 0 of patch 1 vertex 0
2580 * - attribute 0 of patch 1 vertex 1
2582 * - attribute 1 of patch 0 vertex 0
2583 * - attribute 1 of patch 0 vertex 1
2585 * - per patch attribute 0 of patch 0
2586 * - per patch attribute 0 of patch 1
2589 * Note that every attribute has 4 components.
2591 static LLVMValueRef
get_tcs_tes_buffer_address(struct nir_to_llvm_context
*ctx
,
2592 LLVMValueRef vertex_index
,
2593 LLVMValueRef param_index
)
2595 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
2596 LLVMValueRef param_stride
, constant16
;
2597 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
2599 vertices_per_patch
= unpack_param(ctx
, ctx
->tcs_offchip_layout
, 9, 6);
2600 num_patches
= unpack_param(ctx
, ctx
->tcs_offchip_layout
, 0, 9);
2601 total_vertices
= LLVMBuildMul(ctx
->builder
, vertices_per_patch
,
2604 constant16
= LLVMConstInt(ctx
->i32
, 16, false);
2606 base_addr
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
2607 vertices_per_patch
, "");
2609 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2612 param_stride
= total_vertices
;
2614 base_addr
= rel_patch_id
;
2615 param_stride
= num_patches
;
2618 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2619 LLVMBuildMul(ctx
->builder
, param_index
,
2620 param_stride
, ""), "");
2622 base_addr
= LLVMBuildMul(ctx
->builder
, base_addr
, constant16
, "");
2624 if (!vertex_index
) {
2625 LLVMValueRef patch_data_offset
=
2626 unpack_param(ctx
, ctx
->tcs_offchip_layout
, 16, 16);
2628 base_addr
= LLVMBuildAdd(ctx
->builder
, base_addr
,
2629 patch_data_offset
, "");
2634 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct nir_to_llvm_context
*ctx
,
2636 unsigned const_index
,
2638 LLVMValueRef vertex_index
,
2639 LLVMValueRef indir_index
)
2641 LLVMValueRef param_index
;
2644 param_index
= LLVMBuildAdd(ctx
->builder
, LLVMConstInt(ctx
->i32
, param
, false),
2647 if (const_index
&& !is_compact
)
2648 param
+= const_index
;
2649 param_index
= LLVMConstInt(ctx
->i32
, param
, false);
2651 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
2655 mark_tess_output(struct nir_to_llvm_context
*ctx
,
2656 bool is_patch
, uint32_t param
)
2660 ctx
->tess_patch_outputs_written
|= (1ull << param
);
2662 ctx
->tess_outputs_written
|= (1ull << param
);
2666 get_dw_address(struct nir_to_llvm_context
*ctx
,
2667 LLVMValueRef dw_addr
,
2669 unsigned const_index
,
2670 bool compact_const_index
,
2671 LLVMValueRef vertex_index
,
2672 LLVMValueRef stride
,
2673 LLVMValueRef indir_index
)
2678 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2679 LLVMBuildMul(ctx
->builder
,
2685 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2686 LLVMBuildMul(ctx
->builder
, indir_index
,
2687 LLVMConstInt(ctx
->i32
, 4, false), ""), "");
2688 else if (const_index
&& !compact_const_index
)
2689 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2690 LLVMConstInt(ctx
->i32
, const_index
, false), "");
2692 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2693 LLVMConstInt(ctx
->i32
, param
* 4, false), "");
2695 if (const_index
&& compact_const_index
)
2696 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2697 LLVMConstInt(ctx
->i32
, const_index
, false), "");
2702 load_tcs_input(struct nir_to_llvm_context
*ctx
,
2703 nir_intrinsic_instr
*instr
)
2705 LLVMValueRef dw_addr
, stride
;
2706 unsigned const_index
;
2707 LLVMValueRef vertex_index
;
2708 LLVMValueRef indir_index
;
2710 LLVMValueRef value
[4], result
;
2711 const bool per_vertex
= nir_is_per_vertex_io(instr
->variables
[0]->var
, ctx
->stage
);
2712 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
2713 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2714 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2715 false, NULL
, per_vertex
? &vertex_index
: NULL
,
2716 &const_index
, &indir_index
);
2718 stride
= unpack_param(ctx
, ctx
->tcs_in_layout
, 13, 8);
2719 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
2720 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2723 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2724 value
[i
] = lds_load(ctx
, dw_addr
);
2725 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2728 result
= ac_build_gather_values(&ctx
->ac
, value
, instr
->num_components
);
2729 result
= LLVMBuildBitCast(ctx
->builder
, result
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
2734 load_tcs_output(struct nir_to_llvm_context
*ctx
,
2735 nir_intrinsic_instr
*instr
)
2737 LLVMValueRef dw_addr
, stride
;
2738 LLVMValueRef value
[4], result
;
2739 LLVMValueRef vertex_index
= NULL
;
2740 LLVMValueRef indir_index
= NULL
;
2741 unsigned const_index
= 0;
2743 const bool per_vertex
= nir_is_per_vertex_io(instr
->variables
[0]->var
, ctx
->stage
);
2744 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
2745 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2746 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2747 false, NULL
, per_vertex
? &vertex_index
: NULL
,
2748 &const_index
, &indir_index
);
2750 if (!instr
->variables
[0]->var
->data
.patch
) {
2751 stride
= unpack_param(ctx
, ctx
->tcs_out_layout
, 13, 8);
2752 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2754 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2757 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2760 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2761 value
[i
] = lds_load(ctx
, dw_addr
);
2762 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2765 result
= ac_build_gather_values(&ctx
->ac
, value
, instr
->num_components
);
2766 result
= LLVMBuildBitCast(ctx
->builder
, result
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
2771 store_tcs_output(struct nir_to_llvm_context
*ctx
,
2772 nir_intrinsic_instr
*instr
,
2776 LLVMValueRef stride
, dw_addr
;
2777 LLVMValueRef buf_addr
= NULL
;
2778 LLVMValueRef vertex_index
= NULL
;
2779 LLVMValueRef indir_index
= NULL
;
2780 unsigned const_index
= 0;
2782 const bool per_vertex
= nir_is_per_vertex_io(instr
->variables
[0]->var
, ctx
->stage
);
2783 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
2785 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2786 false, NULL
, per_vertex
? &vertex_index
: NULL
,
2787 &const_index
, &indir_index
);
2789 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2790 if (instr
->variables
[0]->var
->data
.location
== VARYING_SLOT_CLIP_DIST0
&&
2791 is_compact
&& const_index
> 3) {
2796 if (!instr
->variables
[0]->var
->data
.patch
) {
2797 stride
= unpack_param(ctx
, ctx
->tcs_out_layout
, 13, 8);
2798 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
2800 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
2803 mark_tess_output(ctx
, instr
->variables
[0]->var
->data
.patch
, param
);
2805 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
2807 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
2808 vertex_index
, indir_index
);
2810 unsigned base
= is_compact
? const_index
: 0;
2811 for (unsigned chan
= 0; chan
< 8; chan
++) {
2812 bool is_tess_factor
= false;
2813 if (!(writemask
& (1 << chan
)))
2815 LLVMValueRef value
= llvm_extract_elem(&ctx
->ac
, src
, chan
);
2817 lds_store(ctx
, dw_addr
, value
);
2819 if (instr
->variables
[0]->var
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
2820 instr
->variables
[0]->var
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
2821 is_tess_factor
= true;
2823 if (!is_tess_factor
&& writemask
!= 0xF)
2824 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
2825 buf_addr
, ctx
->oc_lds
,
2826 4 * (base
+ chan
), 1, 0, true, false);
2828 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
,
2832 if (writemask
== 0xF) {
2833 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
2834 buf_addr
, ctx
->oc_lds
,
2835 (base
* 4), 1, 0, true, false);
2840 load_tes_input(struct nir_to_llvm_context
*ctx
,
2841 const nir_intrinsic_instr
*instr
)
2843 LLVMValueRef buf_addr
;
2844 LLVMValueRef result
;
2845 LLVMValueRef vertex_index
= NULL
;
2846 LLVMValueRef indir_index
= NULL
;
2847 unsigned const_index
= 0;
2849 const bool per_vertex
= nir_is_per_vertex_io(instr
->variables
[0]->var
, ctx
->stage
);
2850 const bool is_compact
= instr
->variables
[0]->var
->data
.compact
;
2852 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2853 false, NULL
, per_vertex
? &vertex_index
: NULL
,
2854 &const_index
, &indir_index
);
2855 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2856 if (instr
->variables
[0]->var
->data
.location
== VARYING_SLOT_CLIP_DIST0
&&
2857 is_compact
&& const_index
> 3) {
2861 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
2862 is_compact
, vertex_index
, indir_index
);
2864 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, instr
->num_components
, NULL
,
2865 buf_addr
, ctx
->oc_lds
, is_compact
? (4 * const_index
) : 0, 1, 0, true, false);
2866 result
= trim_vector(&ctx
->ac
, result
, instr
->num_components
);
2867 result
= LLVMBuildBitCast(ctx
->builder
, result
, get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
2872 load_gs_input(struct nir_to_llvm_context
*ctx
,
2873 nir_intrinsic_instr
*instr
)
2875 LLVMValueRef indir_index
, vtx_offset
;
2876 unsigned const_index
;
2877 LLVMValueRef args
[9];
2878 unsigned param
, vtx_offset_param
;
2879 LLVMValueRef value
[4], result
;
2880 unsigned vertex_index
;
2881 get_deref_offset(ctx
->nir
, instr
->variables
[0],
2882 false, &vertex_index
, NULL
,
2883 &const_index
, &indir_index
);
2884 vtx_offset_param
= vertex_index
;
2885 assert(vtx_offset_param
< 6);
2886 vtx_offset
= LLVMBuildMul(ctx
->builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
2887 LLVMConstInt(ctx
->i32
, 4, false), "");
2889 param
= shader_io_get_unique_index(instr
->variables
[0]->var
->data
.location
);
2890 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2892 args
[0] = ctx
->esgs_ring
;
2893 args
[1] = vtx_offset
;
2894 args
[2] = LLVMConstInt(ctx
->i32
, (param
* 4 + i
+ const_index
) * 256, false);
2895 args
[3] = ctx
->i32zero
;
2896 args
[4] = ctx
->i32one
; /* OFFEN */
2897 args
[5] = ctx
->i32zero
; /* IDXEN */
2898 args
[6] = ctx
->i32one
; /* GLC */
2899 args
[7] = ctx
->i32zero
; /* SLC */
2900 args
[8] = ctx
->i32zero
; /* TFE */
2902 value
[i
] = ac_build_intrinsic(&ctx
->ac
, "llvm.SI.buffer.load.dword.i32.i32",
2904 AC_FUNC_ATTR_READONLY
|
2905 AC_FUNC_ATTR_LEGACY
);
2907 result
= ac_build_gather_values(&ctx
->ac
, value
, instr
->num_components
);
2913 build_gep_for_deref(struct ac_nir_context
*ctx
,
2914 nir_deref_var
*deref
)
2916 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, deref
->var
);
2917 assert(entry
->data
);
2918 LLVMValueRef val
= entry
->data
;
2919 nir_deref
*tail
= deref
->deref
.child
;
2920 while (tail
!= NULL
) {
2921 LLVMValueRef offset
;
2922 switch (tail
->deref_type
) {
2923 case nir_deref_type_array
: {
2924 nir_deref_array
*array
= nir_deref_as_array(tail
);
2925 offset
= LLVMConstInt(ctx
->ac
.i32
, array
->base_offset
, 0);
2926 if (array
->deref_array_type
==
2927 nir_deref_array_type_indirect
) {
2928 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
2935 case nir_deref_type_struct
: {
2936 nir_deref_struct
*deref_struct
=
2937 nir_deref_as_struct(tail
);
2938 offset
= LLVMConstInt(ctx
->ac
.i32
,
2939 deref_struct
->index
, 0);
2943 unreachable("bad deref type");
2945 val
= ac_build_gep0(&ctx
->ac
, val
, offset
);
2951 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
2952 nir_intrinsic_instr
*instr
)
2954 LLVMValueRef values
[8];
2955 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
2956 int ve
= instr
->dest
.ssa
.num_components
;
2957 LLVMValueRef indir_index
;
2959 unsigned const_index
;
2960 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
2961 instr
->variables
[0]->var
->data
.mode
== nir_var_shader_in
;
2962 get_deref_offset(ctx
, instr
->variables
[0], vs_in
, NULL
, NULL
,
2963 &const_index
, &indir_index
);
2965 if (instr
->dest
.ssa
.bit_size
== 64)
2968 switch (instr
->variables
[0]->var
->data
.mode
) {
2969 case nir_var_shader_in
:
2970 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
2971 return load_tcs_input(ctx
->nctx
, instr
);
2972 if (ctx
->stage
== MESA_SHADER_TESS_EVAL
)
2973 return load_tes_input(ctx
->nctx
, instr
);
2974 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
2975 return load_gs_input(ctx
->nctx
, instr
);
2977 for (unsigned chan
= 0; chan
< ve
; chan
++) {
2979 unsigned count
= glsl_count_attribute_slots(
2980 instr
->variables
[0]->var
->type
,
2981 ctx
->stage
== MESA_SHADER_VERTEX
);
2983 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2984 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
2987 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2991 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* 4];
2995 for (unsigned chan
= 0; chan
< ve
; chan
++) {
2997 unsigned count
= glsl_count_attribute_slots(
2998 instr
->variables
[0]->var
->type
, false);
3000 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3001 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3004 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3008 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* 4], "");
3012 case nir_var_shared
: {
3013 LLVMValueRef address
= build_gep_for_deref(ctx
,
3014 instr
->variables
[0]);
3015 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
3016 return LLVMBuildBitCast(ctx
->ac
.builder
, val
,
3017 get_def_type(ctx
, &instr
->dest
.ssa
),
3020 case nir_var_shader_out
:
3021 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
3022 return load_tcs_output(ctx
->nctx
, instr
);
3023 for (unsigned chan
= 0; chan
< ve
; chan
++) {
3025 unsigned count
= glsl_count_attribute_slots(
3026 instr
->variables
[0]->var
->type
, false);
3028 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3029 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3032 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3036 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
3037 ctx
->outputs
[idx
+ chan
+ const_index
* 4],
3043 unreachable("unhandle variable mode");
3045 ret
= ac_build_gather_values(&ctx
->ac
, values
, ve
);
3046 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3050 visit_store_var(struct ac_nir_context
*ctx
,
3051 nir_intrinsic_instr
*instr
)
3053 LLVMValueRef temp_ptr
, value
;
3054 int idx
= instr
->variables
[0]->var
->data
.driver_location
;
3055 LLVMValueRef src
= to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3056 int writemask
= instr
->const_index
[0];
3057 LLVMValueRef indir_index
;
3058 unsigned const_index
;
3059 get_deref_offset(ctx
, instr
->variables
[0], false,
3060 NULL
, NULL
, &const_index
, &indir_index
);
3062 if (get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64) {
3063 int old_writemask
= writemask
;
3065 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
3066 LLVMVectorType(ctx
->ac
.f32
, get_llvm_num_components(src
) * 2),
3070 for (unsigned chan
= 0; chan
< 4; chan
++) {
3071 if (old_writemask
& (1 << chan
))
3072 writemask
|= 3u << (2 * chan
);
3076 switch (instr
->variables
[0]->var
->data
.mode
) {
3077 case nir_var_shader_out
:
3079 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3080 store_tcs_output(ctx
->nctx
, instr
, src
, writemask
);
3084 for (unsigned chan
= 0; chan
< 8; chan
++) {
3086 if (!(writemask
& (1 << chan
)))
3089 value
= llvm_extract_elem(&ctx
->ac
, src
, chan
);
3091 if (instr
->variables
[0]->var
->data
.compact
)
3094 unsigned count
= glsl_count_attribute_slots(
3095 instr
->variables
[0]->var
->type
, false);
3097 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3098 &ctx
->ac
, ctx
->outputs
+ idx
+ chan
, count
,
3101 if (get_llvm_num_components(tmp_vec
) > 1) {
3102 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3103 value
, indir_index
, "");
3106 build_store_values_extended(&ctx
->ac
, ctx
->outputs
+ idx
+ chan
,
3107 count
, stride
, tmp_vec
);
3110 temp_ptr
= ctx
->outputs
[idx
+ chan
+ const_index
* stride
];
3112 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3117 for (unsigned chan
= 0; chan
< 8; chan
++) {
3118 if (!(writemask
& (1 << chan
)))
3121 value
= llvm_extract_elem(&ctx
->ac
, src
, chan
);
3123 unsigned count
= glsl_count_attribute_slots(
3124 instr
->variables
[0]->var
->type
, false);
3126 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
3127 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
3130 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
3131 value
, indir_index
, "");
3132 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
3135 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
3137 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
3141 case nir_var_shared
: {
3142 int writemask
= instr
->const_index
[0];
3143 LLVMValueRef address
= build_gep_for_deref(ctx
,
3144 instr
->variables
[0]);
3145 LLVMValueRef val
= get_src(ctx
, instr
->src
[0]);
3146 unsigned components
=
3147 glsl_get_vector_elements(
3148 nir_deref_tail(&instr
->variables
[0]->deref
)->type
);
3149 if (writemask
== (1 << components
) - 1) {
3150 val
= LLVMBuildBitCast(
3151 ctx
->ac
.builder
, val
,
3152 LLVMGetElementType(LLVMTypeOf(address
)), "");
3153 LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
3155 for (unsigned chan
= 0; chan
< 4; chan
++) {
3156 if (!(writemask
& (1 << chan
)))
3159 LLVMBuildStructGEP(ctx
->ac
.builder
,
3161 LLVMValueRef src
= llvm_extract_elem(&ctx
->ac
, val
,
3163 src
= LLVMBuildBitCast(
3164 ctx
->ac
.builder
, src
,
3165 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
3166 LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
3176 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
3179 case GLSL_SAMPLER_DIM_BUF
:
3181 case GLSL_SAMPLER_DIM_1D
:
3182 return array
? 2 : 1;
3183 case GLSL_SAMPLER_DIM_2D
:
3184 return array
? 3 : 2;
3185 case GLSL_SAMPLER_DIM_MS
:
3186 return array
? 4 : 3;
3187 case GLSL_SAMPLER_DIM_3D
:
3188 case GLSL_SAMPLER_DIM_CUBE
:
3190 case GLSL_SAMPLER_DIM_RECT
:
3191 case GLSL_SAMPLER_DIM_SUBPASS
:
3193 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
3203 /* Adjust the sample index according to FMASK.
3205 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3206 * which is the identity mapping. Each nibble says which physical sample
3207 * should be fetched to get that sample.
3209 * For example, 0x11111100 means there are only 2 samples stored and
3210 * the second sample covers 3/4 of the pixel. When reading samples 0
3211 * and 1, return physical sample 0 (determined by the first two 0s
3212 * in FMASK), otherwise return physical sample 1.
3214 * The sample index should be adjusted as follows:
3215 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
3217 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
3218 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
3219 LLVMValueRef coord_z
,
3220 LLVMValueRef sample_index
,
3221 LLVMValueRef fmask_desc_ptr
)
3223 LLVMValueRef fmask_load_address
[4];
3226 fmask_load_address
[0] = coord_x
;
3227 fmask_load_address
[1] = coord_y
;
3229 fmask_load_address
[2] = coord_z
;
3230 fmask_load_address
[3] = LLVMGetUndef(ctx
->i32
);
3233 struct ac_image_args args
= {0};
3235 args
.opcode
= ac_image_load
;
3236 args
.da
= coord_z
? true : false;
3237 args
.resource
= fmask_desc_ptr
;
3239 args
.addr
= ac_build_gather_values(ctx
, fmask_load_address
, coord_z
? 4 : 2);
3241 res
= ac_build_image_opcode(ctx
, &args
);
3243 res
= to_integer(ctx
, res
);
3244 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, false);
3245 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xf, false);
3247 LLVMValueRef fmask
= LLVMBuildExtractElement(ctx
->builder
,
3251 LLVMValueRef sample_index4
=
3252 LLVMBuildMul(ctx
->builder
, sample_index
, four
, "");
3253 LLVMValueRef shifted_fmask
=
3254 LLVMBuildLShr(ctx
->builder
, fmask
, sample_index4
, "");
3255 LLVMValueRef final_sample
=
3256 LLVMBuildAnd(ctx
->builder
, shifted_fmask
, F
, "");
3258 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3259 * resource descriptor is 0 (invalid),
3261 LLVMValueRef fmask_desc
=
3262 LLVMBuildBitCast(ctx
->builder
, fmask_desc_ptr
,
3265 LLVMValueRef fmask_word1
=
3266 LLVMBuildExtractElement(ctx
->builder
, fmask_desc
,
3269 LLVMValueRef word1_is_nonzero
=
3270 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3271 fmask_word1
, ctx
->i32_0
, "");
3273 /* Replace the MSAA sample index. */
3275 LLVMBuildSelect(ctx
->builder
, word1_is_nonzero
,
3276 final_sample
, sample_index
, "");
3277 return sample_index
;
3280 static LLVMValueRef
get_image_coords(struct ac_nir_context
*ctx
,
3281 const nir_intrinsic_instr
*instr
)
3283 const struct glsl_type
*type
= instr
->variables
[0]->var
->type
;
3284 if(instr
->variables
[0]->deref
.child
)
3285 type
= instr
->variables
[0]->deref
.child
->type
;
3287 LLVMValueRef src0
= get_src(ctx
, instr
->src
[0]);
3288 LLVMValueRef coords
[4];
3289 LLVMValueRef masks
[] = {
3290 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
3291 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
3294 LLVMValueRef sample_index
= llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[1]), 0);
3297 enum glsl_sampler_dim dim
= glsl_get_sampler_dim(type
);
3298 bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3299 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3300 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
3301 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
3303 count
= image_type_to_components_count(dim
,
3304 glsl_sampler_type_is_array(type
));
3307 LLVMValueRef fmask_load_address
[3];
3310 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3311 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
3312 if (glsl_sampler_type_is_array(type
))
3313 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
3315 fmask_load_address
[2] = NULL
;
3317 for (chan
= 0; chan
< 2; ++chan
)
3318 fmask_load_address
[chan
] = LLVMBuildAdd(ctx
->ac
.builder
, fmask_load_address
[chan
], LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->nctx
->frag_pos
[chan
], ctx
->ac
.i32
, ""), "");
3320 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
3321 fmask_load_address
[0],
3322 fmask_load_address
[1],
3323 fmask_load_address
[2],
3325 get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_FMASK
, true, false));
3328 if (instr
->src
[0].ssa
->num_components
)
3329 res
= LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
3336 for (chan
= 0; chan
< count
; ++chan
) {
3337 coords
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[chan
], "");
3341 for (chan
= 0; chan
< count
; ++chan
)
3342 coords
[chan
] = LLVMBuildAdd(ctx
->ac
.builder
, coords
[chan
], LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->nctx
->frag_pos
[chan
], ctx
->ac
.i32
, ""), "");
3345 coords
[count
] = sample_index
;
3350 coords
[3] = LLVMGetUndef(ctx
->ac
.i32
);
3353 res
= ac_build_gather_values(&ctx
->ac
, coords
, count
);
3358 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
3359 const nir_intrinsic_instr
*instr
)
3361 LLVMValueRef params
[7];
3363 char intrinsic_name
[64];
3364 const nir_variable
*var
= instr
->variables
[0]->var
;
3365 const struct glsl_type
*type
= var
->type
;
3366 LLVMValueRef i1false
= LLVMConstInt(ctx
->ac
.i1
, 0, false);
3367 LLVMValueRef i1true
= LLVMConstInt(ctx
->ac
.i1
, 1, false);
3369 if(instr
->variables
[0]->deref
.child
)
3370 type
= instr
->variables
[0]->deref
.child
->type
;
3372 type
= glsl_without_array(type
);
3373 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3374 params
[0] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, true, false);
3375 params
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3376 ctx
->ac
.i32_0
, ""); /* vindex */
3377 params
[2] = ctx
->ac
.i32_0
; /* voffset */
3378 params
[3] = i1false
; /* glc */
3379 params
[4] = i1false
; /* slc */
3380 res
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.buffer.load.format.v4f32", ctx
->ac
.v4f32
,
3383 res
= trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
3384 res
= to_integer(&ctx
->ac
, res
);
3386 bool is_da
= glsl_sampler_type_is_array(type
) ||
3387 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3388 LLVMValueRef da
= is_da
? i1true
: i1false
;
3389 LLVMValueRef glc
= i1false
;
3390 LLVMValueRef slc
= i1false
;
3392 params
[0] = get_image_coords(ctx
, instr
);
3393 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, true, false);
3394 params
[2] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3395 if (HAVE_LLVM
<= 0x0309) {
3396 params
[3] = i1false
; /* r128 */
3401 LLVMValueRef lwe
= i1false
;
3408 ac_get_image_intr_name("llvm.amdgcn.image.load",
3409 ctx
->ac
.v4f32
, /* vdata */
3410 LLVMTypeOf(params
[0]), /* coords */
3411 LLVMTypeOf(params
[1]), /* rsrc */
3412 intrinsic_name
, sizeof(intrinsic_name
));
3414 res
= ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.v4f32
,
3415 params
, 7, AC_FUNC_ATTR_READONLY
);
3417 return to_integer(&ctx
->ac
, res
);
3420 static void visit_image_store(struct ac_nir_context
*ctx
,
3421 nir_intrinsic_instr
*instr
)
3423 LLVMValueRef params
[8];
3424 char intrinsic_name
[64];
3425 const nir_variable
*var
= instr
->variables
[0]->var
;
3426 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3427 LLVMValueRef i1false
= LLVMConstInt(ctx
->ac
.i1
, 0, false);
3428 LLVMValueRef i1true
= LLVMConstInt(ctx
->ac
.i1
, 1, false);
3429 LLVMValueRef glc
= i1false
;
3430 bool force_glc
= ctx
->abi
->chip_class
== SI
;
3434 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3435 params
[0] = to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2])); /* data */
3436 params
[1] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
, true, true);
3437 params
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3438 ctx
->ac
.i32_0
, ""); /* vindex */
3439 params
[3] = ctx
->ac
.i32_0
; /* voffset */
3440 params
[4] = glc
; /* glc */
3441 params
[5] = i1false
; /* slc */
3442 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.buffer.store.format.v4f32", ctx
->ac
.voidt
,
3445 bool is_da
= glsl_sampler_type_is_array(type
) ||
3446 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3447 LLVMValueRef da
= is_da
? i1true
: i1false
;
3448 LLVMValueRef slc
= i1false
;
3450 params
[0] = to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[2]));
3451 params
[1] = get_image_coords(ctx
, instr
); /* coords */
3452 params
[2] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, true, true);
3453 params
[3] = LLVMConstInt(ctx
->ac
.i32
, 15, false); /* dmask */
3454 if (HAVE_LLVM
<= 0x0309) {
3455 params
[4] = i1false
; /* r128 */
3460 LLVMValueRef lwe
= i1false
;
3467 ac_get_image_intr_name("llvm.amdgcn.image.store",
3468 LLVMTypeOf(params
[0]), /* vdata */
3469 LLVMTypeOf(params
[1]), /* coords */
3470 LLVMTypeOf(params
[2]), /* rsrc */
3471 intrinsic_name
, sizeof(intrinsic_name
));
3473 ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.voidt
,
3479 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
3480 const nir_intrinsic_instr
*instr
)
3482 LLVMValueRef params
[6];
3483 int param_count
= 0;
3484 const nir_variable
*var
= instr
->variables
[0]->var
;
3486 const char *atomic_name
;
3487 char intrinsic_name
[41];
3488 const struct glsl_type
*type
= glsl_without_array(var
->type
);
3489 LLVMValueRef i1false
= LLVMConstInt(ctx
->ac
.i1
, 0, false);
3490 LLVMValueRef i1true
= LLVMConstInt(ctx
->ac
.i1
, 1, false);
3491 MAYBE_UNUSED
int length
;
3493 switch (instr
->intrinsic
) {
3494 case nir_intrinsic_image_atomic_add
:
3495 atomic_name
= "add";
3497 case nir_intrinsic_image_atomic_min
:
3498 atomic_name
= "smin";
3500 case nir_intrinsic_image_atomic_max
:
3501 atomic_name
= "smax";
3503 case nir_intrinsic_image_atomic_and
:
3504 atomic_name
= "and";
3506 case nir_intrinsic_image_atomic_or
:
3509 case nir_intrinsic_image_atomic_xor
:
3510 atomic_name
= "xor";
3512 case nir_intrinsic_image_atomic_exchange
:
3513 atomic_name
= "swap";
3515 case nir_intrinsic_image_atomic_comp_swap
:
3516 atomic_name
= "cmpswap";
3522 if (instr
->intrinsic
== nir_intrinsic_image_atomic_comp_swap
)
3523 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
3524 params
[param_count
++] = get_src(ctx
, instr
->src
[2]);
3526 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
) {
3527 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_BUFFER
,
3529 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[0]),
3530 ctx
->ac
.i32_0
, ""); /* vindex */
3531 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
3532 params
[param_count
++] = i1false
; /* slc */
3534 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3535 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
3537 char coords_type
[8];
3539 bool da
= glsl_sampler_type_is_array(type
) ||
3540 glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
;
3542 LLVMValueRef coords
= params
[param_count
++] = get_image_coords(ctx
, instr
);
3543 params
[param_count
++] = get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
,
3545 params
[param_count
++] = i1false
; /* r128 */
3546 params
[param_count
++] = da
? i1true
: i1false
; /* da */
3547 params
[param_count
++] = i1false
; /* slc */
3549 build_int_type_name(LLVMTypeOf(coords
),
3550 coords_type
, sizeof(coords_type
));
3552 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3553 "llvm.amdgcn.image.atomic.%s.%s", atomic_name
, coords_type
);
3556 assert(length
< sizeof(intrinsic_name
));
3557 return ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
, params
, param_count
, 0);
3560 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
3561 const nir_intrinsic_instr
*instr
)
3564 const nir_variable
*var
= instr
->variables
[0]->var
;
3565 const struct glsl_type
*type
= instr
->variables
[0]->var
->type
;
3566 bool da
= glsl_sampler_type_is_array(var
->type
) ||
3567 glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_CUBE
;
3568 if(instr
->variables
[0]->deref
.child
)
3569 type
= instr
->variables
[0]->deref
.child
->type
;
3571 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_BUF
)
3572 return get_buffer_size(ctx
,
3573 get_sampler_desc(ctx
, instr
->variables
[0],
3574 AC_DESC_BUFFER
, true, false), true);
3576 struct ac_image_args args
= { 0 };
3580 args
.resource
= get_sampler_desc(ctx
, instr
->variables
[0], AC_DESC_IMAGE
, true, false);
3581 args
.opcode
= ac_image_get_resinfo
;
3582 args
.addr
= ctx
->ac
.i32_0
;
3584 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
3586 if (glsl_get_sampler_dim(type
) == GLSL_SAMPLER_DIM_CUBE
&&
3587 glsl_sampler_type_is_array(type
)) {
3588 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3589 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3590 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3591 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3592 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
3597 #define NOOP_WAITCNT 0xf7f
3598 #define LGKM_CNT 0x07f
3599 #define VM_CNT 0xf70
3601 static void emit_waitcnt(struct nir_to_llvm_context
*ctx
,
3604 LLVMValueRef args
[1] = {
3605 LLVMConstInt(ctx
->i32
, simm16
, false),
3607 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.s.waitcnt",
3608 ctx
->voidt
, args
, 1, 0);
3611 static void emit_barrier(struct nir_to_llvm_context
*ctx
)
3613 /* SI only (thanks to a hw bug workaround):
3614 * The real barrier instruction isn’t needed, because an entire patch
3615 * always fits into a single wave.
3617 if (ctx
->options
->chip_class
== SI
&&
3618 ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3619 emit_waitcnt(ctx
, LGKM_CNT
& VM_CNT
);
3622 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.s.barrier",
3623 ctx
->voidt
, NULL
, 0, AC_FUNC_ATTR_CONVERGENT
);
3626 static void emit_discard_if(struct nir_to_llvm_context
*ctx
,
3627 const nir_intrinsic_instr
*instr
)
3631 cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3632 get_src(ctx
->nir
, instr
->src
[0]),
3635 cond
= LLVMBuildSelect(ctx
->builder
, cond
,
3636 LLVMConstReal(ctx
->f32
, -1.0f
),
3638 ac_build_kill(&ctx
->ac
, cond
);
3642 visit_load_local_invocation_index(struct nir_to_llvm_context
*ctx
)
3644 LLVMValueRef result
;
3645 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
3646 result
= LLVMBuildAnd(ctx
->builder
, ctx
->tg_size
,
3647 LLVMConstInt(ctx
->i32
, 0xfc0, false), "");
3649 return LLVMBuildAdd(ctx
->builder
, result
, thread_id
, "");
3652 static LLVMValueRef
visit_var_atomic(struct nir_to_llvm_context
*ctx
,
3653 const nir_intrinsic_instr
*instr
)
3655 LLVMValueRef ptr
, result
;
3656 LLVMValueRef src
= get_src(ctx
->nir
, instr
->src
[0]);
3657 ptr
= build_gep_for_deref(ctx
->nir
, instr
->variables
[0]);
3659 if (instr
->intrinsic
== nir_intrinsic_var_atomic_comp_swap
) {
3660 LLVMValueRef src1
= get_src(ctx
->nir
, instr
->src
[1]);
3661 result
= LLVMBuildAtomicCmpXchg(ctx
->builder
,
3663 LLVMAtomicOrderingSequentiallyConsistent
,
3664 LLVMAtomicOrderingSequentiallyConsistent
,
3667 LLVMAtomicRMWBinOp op
;
3668 switch (instr
->intrinsic
) {
3669 case nir_intrinsic_var_atomic_add
:
3670 op
= LLVMAtomicRMWBinOpAdd
;
3672 case nir_intrinsic_var_atomic_umin
:
3673 op
= LLVMAtomicRMWBinOpUMin
;
3675 case nir_intrinsic_var_atomic_umax
:
3676 op
= LLVMAtomicRMWBinOpUMax
;
3678 case nir_intrinsic_var_atomic_imin
:
3679 op
= LLVMAtomicRMWBinOpMin
;
3681 case nir_intrinsic_var_atomic_imax
:
3682 op
= LLVMAtomicRMWBinOpMax
;
3684 case nir_intrinsic_var_atomic_and
:
3685 op
= LLVMAtomicRMWBinOpAnd
;
3687 case nir_intrinsic_var_atomic_or
:
3688 op
= LLVMAtomicRMWBinOpOr
;
3690 case nir_intrinsic_var_atomic_xor
:
3691 op
= LLVMAtomicRMWBinOpXor
;
3693 case nir_intrinsic_var_atomic_exchange
:
3694 op
= LLVMAtomicRMWBinOpXchg
;
3700 result
= LLVMBuildAtomicRMW(ctx
->builder
, op
, ptr
, to_integer(&ctx
->ac
, src
),
3701 LLVMAtomicOrderingSequentiallyConsistent
,
3707 #define INTERP_CENTER 0
3708 #define INTERP_CENTROID 1
3709 #define INTERP_SAMPLE 2
3711 static LLVMValueRef
lookup_interp_param(struct nir_to_llvm_context
*ctx
,
3712 enum glsl_interp_mode interp
, unsigned location
)
3715 case INTERP_MODE_FLAT
:
3718 case INTERP_MODE_SMOOTH
:
3719 case INTERP_MODE_NONE
:
3720 if (location
== INTERP_CENTER
)
3721 return ctx
->persp_center
;
3722 else if (location
== INTERP_CENTROID
)
3723 return ctx
->persp_centroid
;
3724 else if (location
== INTERP_SAMPLE
)
3725 return ctx
->persp_sample
;
3727 case INTERP_MODE_NOPERSPECTIVE
:
3728 if (location
== INTERP_CENTER
)
3729 return ctx
->linear_center
;
3730 else if (location
== INTERP_CENTROID
)
3731 return ctx
->linear_centroid
;
3732 else if (location
== INTERP_SAMPLE
)
3733 return ctx
->linear_sample
;
3739 static LLVMValueRef
load_sample_position(struct nir_to_llvm_context
*ctx
,
3740 LLVMValueRef sample_id
)
3742 LLVMValueRef result
;
3743 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_PS_SAMPLE_POSITIONS
, false));
3745 ptr
= LLVMBuildBitCast(ctx
->builder
, ptr
,
3746 const_array(ctx
->v2f32
, 64), "");
3748 sample_id
= LLVMBuildAdd(ctx
->builder
, sample_id
, ctx
->sample_pos_offset
, "");
3749 result
= ac_build_indexed_load(&ctx
->ac
, ptr
, sample_id
, false);
3754 static LLVMValueRef
load_sample_pos(struct nir_to_llvm_context
*ctx
)
3756 LLVMValueRef values
[2];
3758 values
[0] = emit_ffract(&ctx
->ac
, ctx
->frag_pos
[0]);
3759 values
[1] = emit_ffract(&ctx
->ac
, ctx
->frag_pos
[1]);
3760 return ac_build_gather_values(&ctx
->ac
, values
, 2);
3763 static LLVMValueRef
visit_interp(struct nir_to_llvm_context
*ctx
,
3764 const nir_intrinsic_instr
*instr
)
3766 LLVMValueRef result
[2];
3767 LLVMValueRef interp_param
, attr_number
;
3770 LLVMValueRef src_c0
, src_c1
;
3772 int input_index
= instr
->variables
[0]->var
->data
.location
- VARYING_SLOT_VAR0
;
3773 switch (instr
->intrinsic
) {
3774 case nir_intrinsic_interp_var_at_centroid
:
3775 location
= INTERP_CENTROID
;
3777 case nir_intrinsic_interp_var_at_sample
:
3778 case nir_intrinsic_interp_var_at_offset
:
3779 location
= INTERP_CENTER
;
3780 src0
= get_src(ctx
->nir
, instr
->src
[0]);
3786 if (instr
->intrinsic
== nir_intrinsic_interp_var_at_offset
) {
3787 src_c0
= to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32zero
, ""));
3788 src_c1
= to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32one
, ""));
3789 } else if (instr
->intrinsic
== nir_intrinsic_interp_var_at_sample
) {
3790 LLVMValueRef sample_position
;
3791 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
3793 /* fetch sample ID */
3794 sample_position
= load_sample_position(ctx
, src0
);
3796 src_c0
= LLVMBuildExtractElement(ctx
->builder
, sample_position
, ctx
->i32zero
, "");
3797 src_c0
= LLVMBuildFSub(ctx
->builder
, src_c0
, halfval
, "");
3798 src_c1
= LLVMBuildExtractElement(ctx
->builder
, sample_position
, ctx
->i32one
, "");
3799 src_c1
= LLVMBuildFSub(ctx
->builder
, src_c1
, halfval
, "");
3801 interp_param
= lookup_interp_param(ctx
, instr
->variables
[0]->var
->data
.interpolation
, location
);
3802 attr_number
= LLVMConstInt(ctx
->i32
, input_index
, false);
3804 if (location
== INTERP_SAMPLE
|| location
== INTERP_CENTER
) {
3805 LLVMValueRef ij_out
[2];
3806 LLVMValueRef ddxy_out
= emit_ddxy_interp(ctx
, interp_param
);
3809 * take the I then J parameters, and the DDX/Y for it, and
3810 * calculate the IJ inputs for the interpolator.
3811 * temp1 = ddx * offset/sample.x + I;
3812 * interp_param.I = ddy * offset/sample.y + temp1;
3813 * temp1 = ddx * offset/sample.x + J;
3814 * interp_param.J = ddy * offset/sample.y + temp1;
3816 for (unsigned i
= 0; i
< 2; i
++) {
3817 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, false);
3818 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, false);
3819 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->builder
,
3820 ddxy_out
, ix_ll
, "");
3821 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->builder
,
3822 ddxy_out
, iy_ll
, "");
3823 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->builder
,
3824 interp_param
, ix_ll
, "");
3825 LLVMValueRef temp1
, temp2
;
3827 interp_el
= LLVMBuildBitCast(ctx
->builder
, interp_el
,
3830 temp1
= LLVMBuildFMul(ctx
->builder
, ddx_el
, src_c0
, "");
3831 temp1
= LLVMBuildFAdd(ctx
->builder
, temp1
, interp_el
, "");
3833 temp2
= LLVMBuildFMul(ctx
->builder
, ddy_el
, src_c1
, "");
3834 temp2
= LLVMBuildFAdd(ctx
->builder
, temp2
, temp1
, "");
3836 ij_out
[i
] = LLVMBuildBitCast(ctx
->builder
,
3837 temp2
, ctx
->i32
, "");
3839 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
3843 for (chan
= 0; chan
< 2; chan
++) {
3844 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->i32
, chan
, false);
3847 interp_param
= LLVMBuildBitCast(ctx
->builder
,
3848 interp_param
, LLVMVectorType(ctx
->f32
, 2), "");
3849 LLVMValueRef i
= LLVMBuildExtractElement(
3850 ctx
->builder
, interp_param
, ctx
->i32zero
, "");
3851 LLVMValueRef j
= LLVMBuildExtractElement(
3852 ctx
->builder
, interp_param
, ctx
->i32one
, "");
3854 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
3855 llvm_chan
, attr_number
,
3856 ctx
->prim_mask
, i
, j
);
3858 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
3859 LLVMConstInt(ctx
->i32
, 2, false),
3860 llvm_chan
, attr_number
,
3864 return ac_build_gather_values(&ctx
->ac
, result
, 2);
3868 visit_emit_vertex(struct nir_to_llvm_context
*ctx
,
3869 const nir_intrinsic_instr
*instr
)
3871 LLVMValueRef gs_next_vertex
;
3872 LLVMValueRef can_emit
, kill
;
3875 assert(instr
->const_index
[0] == 0);
3876 /* Write vertex attribute values to GSVS ring */
3877 gs_next_vertex
= LLVMBuildLoad(ctx
->builder
,
3878 ctx
->gs_next_vertex
,
3881 /* If this thread has already emitted the declared maximum number of
3882 * vertices, kill it: excessive vertex emissions are not supposed to
3883 * have any effect, and GS threads have no externally observable
3884 * effects other than emitting vertices.
3886 can_emit
= LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, gs_next_vertex
,
3887 LLVMConstInt(ctx
->i32
, ctx
->gs_max_out_vertices
, false), "");
3889 kill
= LLVMBuildSelect(ctx
->builder
, can_emit
,
3890 LLVMConstReal(ctx
->f32
, 1.0f
),
3891 LLVMConstReal(ctx
->f32
, -1.0f
), "");
3892 ac_build_kill(&ctx
->ac
, kill
);
3894 /* loop num outputs */
3896 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
3897 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
3902 if (!(ctx
->output_mask
& (1ull << i
)))
3905 if (i
== VARYING_SLOT_CLIP_DIST0
) {
3906 /* pack clip and cull into a single set of slots */
3907 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
3911 for (unsigned j
= 0; j
< length
; j
++) {
3912 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
,
3914 LLVMValueRef voffset
= LLVMConstInt(ctx
->i32
, (slot
* 4 + j
) * ctx
->gs_max_out_vertices
, false);
3915 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, gs_next_vertex
, "");
3916 voffset
= LLVMBuildMul(ctx
->builder
, voffset
, LLVMConstInt(ctx
->i32
, 4, false), "");
3918 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->i32
, "");
3920 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->gsvs_ring
,
3922 voffset
, ctx
->gs2vs_offset
, 0,
3928 gs_next_vertex
= LLVMBuildAdd(ctx
->builder
, gs_next_vertex
,
3930 LLVMBuildStore(ctx
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
);
3932 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
3936 visit_end_primitive(struct nir_to_llvm_context
*ctx
,
3937 const nir_intrinsic_instr
*instr
)
3939 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (0 << 8), ctx
->gs_wave_id
);
3943 visit_load_tess_coord(struct nir_to_llvm_context
*ctx
,
3944 const nir_intrinsic_instr
*instr
)
3946 LLVMValueRef coord
[4] = {
3953 if (ctx
->tes_primitive_mode
== GL_TRIANGLES
)
3954 coord
[2] = LLVMBuildFSub(ctx
->builder
, ctx
->f32one
,
3955 LLVMBuildFAdd(ctx
->builder
, coord
[0], coord
[1], ""), "");
3957 LLVMValueRef result
= ac_build_gather_values(&ctx
->ac
, coord
, instr
->num_components
);
3958 return LLVMBuildBitCast(ctx
->builder
, result
,
3959 get_def_type(ctx
->nir
, &instr
->dest
.ssa
), "");
3962 static void visit_intrinsic(struct ac_nir_context
*ctx
,
3963 nir_intrinsic_instr
*instr
)
3965 LLVMValueRef result
= NULL
;
3967 switch (instr
->intrinsic
) {
3968 case nir_intrinsic_load_work_group_id
: {
3969 result
= ctx
->nctx
->workgroup_ids
;
3972 case nir_intrinsic_load_base_vertex
: {
3973 result
= ctx
->abi
->base_vertex
;
3976 case nir_intrinsic_load_vertex_id_zero_base
: {
3977 result
= ctx
->abi
->vertex_id
;
3980 case nir_intrinsic_load_local_invocation_id
: {
3981 result
= ctx
->nctx
->local_invocation_ids
;
3984 case nir_intrinsic_load_base_instance
:
3985 result
= ctx
->abi
->start_instance
;
3987 case nir_intrinsic_load_draw_id
:
3988 result
= ctx
->abi
->draw_id
;
3990 case nir_intrinsic_load_invocation_id
:
3991 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
3992 result
= unpack_param(ctx
->nctx
, ctx
->nctx
->tcs_rel_ids
, 8, 5);
3994 result
= ctx
->nctx
->gs_invocation_id
;
3996 case nir_intrinsic_load_primitive_id
:
3997 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3998 ctx
->nctx
->shader_info
->gs
.uses_prim_id
= true;
3999 result
= ctx
->nctx
->gs_prim_id
;
4000 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
4001 ctx
->nctx
->shader_info
->tcs
.uses_prim_id
= true;
4002 result
= ctx
->nctx
->tcs_patch_id
;
4003 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
4004 ctx
->nctx
->shader_info
->tcs
.uses_prim_id
= true;
4005 result
= ctx
->nctx
->tes_patch_id
;
4007 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
4009 case nir_intrinsic_load_sample_id
:
4010 ctx
->nctx
->shader_info
->fs
.force_persample
= true;
4011 result
= unpack_param(ctx
->nctx
, ctx
->nctx
->ancillary
, 8, 4);
4013 case nir_intrinsic_load_sample_pos
:
4014 ctx
->nctx
->shader_info
->fs
.force_persample
= true;
4015 result
= load_sample_pos(ctx
->nctx
);
4017 case nir_intrinsic_load_sample_mask_in
:
4018 result
= ctx
->nctx
->sample_coverage
;
4020 case nir_intrinsic_load_front_face
:
4021 result
= ctx
->nctx
->front_face
;
4023 case nir_intrinsic_load_instance_id
:
4024 result
= ctx
->abi
->instance_id
;
4026 case nir_intrinsic_load_num_work_groups
:
4027 result
= ctx
->nctx
->num_work_groups
;
4029 case nir_intrinsic_load_local_invocation_index
:
4030 result
= visit_load_local_invocation_index(ctx
->nctx
);
4032 case nir_intrinsic_load_push_constant
:
4033 result
= visit_load_push_constant(ctx
->nctx
, instr
);
4035 case nir_intrinsic_vulkan_resource_index
:
4036 result
= visit_vulkan_resource_index(ctx
->nctx
, instr
);
4038 case nir_intrinsic_store_ssbo
:
4039 visit_store_ssbo(ctx
->nctx
, instr
);
4041 case nir_intrinsic_load_ssbo
:
4042 result
= visit_load_buffer(ctx
->nctx
, instr
);
4044 case nir_intrinsic_ssbo_atomic_add
:
4045 case nir_intrinsic_ssbo_atomic_imin
:
4046 case nir_intrinsic_ssbo_atomic_umin
:
4047 case nir_intrinsic_ssbo_atomic_imax
:
4048 case nir_intrinsic_ssbo_atomic_umax
:
4049 case nir_intrinsic_ssbo_atomic_and
:
4050 case nir_intrinsic_ssbo_atomic_or
:
4051 case nir_intrinsic_ssbo_atomic_xor
:
4052 case nir_intrinsic_ssbo_atomic_exchange
:
4053 case nir_intrinsic_ssbo_atomic_comp_swap
:
4054 result
= visit_atomic_ssbo(ctx
->nctx
, instr
);
4056 case nir_intrinsic_load_ubo
:
4057 result
= visit_load_ubo_buffer(ctx
, instr
);
4059 case nir_intrinsic_get_buffer_size
:
4060 result
= visit_get_buffer_size(ctx
, instr
);
4062 case nir_intrinsic_load_var
:
4063 result
= visit_load_var(ctx
, instr
);
4065 case nir_intrinsic_store_var
:
4066 visit_store_var(ctx
, instr
);
4068 case nir_intrinsic_image_load
:
4069 result
= visit_image_load(ctx
, instr
);
4071 case nir_intrinsic_image_store
:
4072 visit_image_store(ctx
, instr
);
4074 case nir_intrinsic_image_atomic_add
:
4075 case nir_intrinsic_image_atomic_min
:
4076 case nir_intrinsic_image_atomic_max
:
4077 case nir_intrinsic_image_atomic_and
:
4078 case nir_intrinsic_image_atomic_or
:
4079 case nir_intrinsic_image_atomic_xor
:
4080 case nir_intrinsic_image_atomic_exchange
:
4081 case nir_intrinsic_image_atomic_comp_swap
:
4082 result
= visit_image_atomic(ctx
, instr
);
4084 case nir_intrinsic_image_size
:
4085 result
= visit_image_size(ctx
, instr
);
4087 case nir_intrinsic_discard
:
4088 ac_build_intrinsic(&ctx
->ac
, "llvm.AMDGPU.kilp",
4089 LLVMVoidTypeInContext(ctx
->ac
.context
),
4090 NULL
, 0, AC_FUNC_ATTR_LEGACY
);
4092 case nir_intrinsic_discard_if
:
4093 emit_discard_if(ctx
->nctx
, instr
);
4095 case nir_intrinsic_memory_barrier
:
4096 emit_waitcnt(ctx
->nctx
, VM_CNT
);
4098 case nir_intrinsic_barrier
:
4099 emit_barrier(ctx
->nctx
);
4101 case nir_intrinsic_var_atomic_add
:
4102 case nir_intrinsic_var_atomic_imin
:
4103 case nir_intrinsic_var_atomic_umin
:
4104 case nir_intrinsic_var_atomic_imax
:
4105 case nir_intrinsic_var_atomic_umax
:
4106 case nir_intrinsic_var_atomic_and
:
4107 case nir_intrinsic_var_atomic_or
:
4108 case nir_intrinsic_var_atomic_xor
:
4109 case nir_intrinsic_var_atomic_exchange
:
4110 case nir_intrinsic_var_atomic_comp_swap
:
4111 result
= visit_var_atomic(ctx
->nctx
, instr
);
4113 case nir_intrinsic_interp_var_at_centroid
:
4114 case nir_intrinsic_interp_var_at_sample
:
4115 case nir_intrinsic_interp_var_at_offset
:
4116 result
= visit_interp(ctx
->nctx
, instr
);
4118 case nir_intrinsic_emit_vertex
:
4119 visit_emit_vertex(ctx
->nctx
, instr
);
4121 case nir_intrinsic_end_primitive
:
4122 visit_end_primitive(ctx
->nctx
, instr
);
4124 case nir_intrinsic_load_tess_coord
:
4125 result
= visit_load_tess_coord(ctx
->nctx
, instr
);
4127 case nir_intrinsic_load_patch_vertices_in
:
4128 result
= LLVMConstInt(ctx
->ac
.i32
, ctx
->nctx
->options
->key
.tcs
.input_vertices
, false);
4131 fprintf(stderr
, "Unknown intrinsic: ");
4132 nir_print_instr(&instr
->instr
, stderr
);
4133 fprintf(stderr
, "\n");
4137 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4141 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
4142 unsigned descriptor_set
,
4143 unsigned base_index
,
4144 unsigned constant_index
,
4146 enum ac_descriptor_type desc_type
,
4147 bool image
, bool write
)
4149 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
4150 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
4151 struct radv_descriptor_set_layout
*layout
= ctx
->options
->layout
->set
[descriptor_set
].layout
;
4152 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
4153 unsigned offset
= binding
->offset
;
4154 unsigned stride
= binding
->size
;
4156 LLVMBuilderRef builder
= ctx
->builder
;
4159 assert(base_index
< layout
->binding_count
);
4161 if (write
&& ctx
->stage
== MESA_SHADER_FRAGMENT
)
4162 ctx
->shader_info
->fs
.writes_memory
= true;
4164 switch (desc_type
) {
4174 case AC_DESC_SAMPLER
:
4176 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
4181 case AC_DESC_BUFFER
:
4186 unreachable("invalid desc_type\n");
4189 offset
+= constant_index
* stride
;
4191 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
4192 (!index
|| binding
->immutable_samplers_equal
)) {
4193 if (binding
->immutable_samplers_equal
)
4196 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
4198 LLVMValueRef constants
[] = {
4199 LLVMConstInt(ctx
->i32
, samplers
[constant_index
* 4 + 0], 0),
4200 LLVMConstInt(ctx
->i32
, samplers
[constant_index
* 4 + 1], 0),
4201 LLVMConstInt(ctx
->i32
, samplers
[constant_index
* 4 + 2], 0),
4202 LLVMConstInt(ctx
->i32
, samplers
[constant_index
* 4 + 3], 0),
4204 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
4207 assert(stride
% type_size
== 0);
4210 index
= ctx
->i32zero
;
4212 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, stride
/ type_size
, 0), "");
4214 list
= ac_build_gep0(&ctx
->ac
, list
, LLVMConstInt(ctx
->i32
, offset
, 0));
4215 list
= LLVMBuildPointerCast(builder
, list
, const_array(type
, 0), "");
4217 return ac_build_indexed_load_const(&ctx
->ac
, list
, index
);
4220 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
4221 const nir_deref_var
*deref
,
4222 enum ac_descriptor_type desc_type
,
4223 bool image
, bool write
)
4225 LLVMValueRef index
= NULL
;
4226 unsigned constant_index
= 0;
4227 const nir_deref
*tail
= &deref
->deref
;
4229 while (tail
->child
) {
4230 const nir_deref_array
*child
= nir_deref_as_array(tail
->child
);
4231 unsigned array_size
= glsl_get_aoa_size(tail
->child
->type
);
4236 assert(child
->deref_array_type
!= nir_deref_array_type_wildcard
);
4238 if (child
->deref_array_type
== nir_deref_array_type_indirect
) {
4239 LLVMValueRef indirect
= get_src(ctx
, child
->indirect
);
4241 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
4242 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
4247 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
4250 constant_index
+= child
->base_offset
* array_size
;
4252 tail
= &child
->deref
;
4255 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
4256 deref
->var
->data
.descriptor_set
,
4257 deref
->var
->data
.binding
,
4258 constant_index
, index
,
4259 desc_type
, image
, write
);
4262 static void set_tex_fetch_args(struct ac_llvm_context
*ctx
,
4263 struct ac_image_args
*args
,
4264 const nir_tex_instr
*instr
,
4266 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4267 LLVMValueRef
*param
, unsigned count
,
4270 unsigned is_rect
= 0;
4271 bool da
= instr
->is_array
|| instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
;
4273 if (op
== nir_texop_lod
)
4275 /* Pad to power of two vector */
4276 while (count
< util_next_power_of_two(count
))
4277 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4280 args
->addr
= ac_build_gather_values(ctx
, param
, count
);
4282 args
->addr
= param
[0];
4284 args
->resource
= res_ptr
;
4285 args
->sampler
= samp_ptr
;
4287 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
&& op
== nir_texop_txf
) {
4288 args
->addr
= param
[0];
4292 args
->dmask
= dmask
;
4293 args
->unorm
= is_rect
;
4297 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4300 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4301 * filtering manually. The driver sets img7 to a mask clearing
4302 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4303 * s_and_b32 samp0, samp0, img7
4306 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4308 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
4309 LLVMValueRef res
, LLVMValueRef samp
)
4311 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4312 LLVMValueRef img7
, samp0
;
4314 if (ctx
->abi
->chip_class
>= VI
)
4317 img7
= LLVMBuildExtractElement(builder
, res
,
4318 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
4319 samp0
= LLVMBuildExtractElement(builder
, samp
,
4320 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4321 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4322 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4323 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4326 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
4327 nir_tex_instr
*instr
,
4328 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
4329 LLVMValueRef
*fmask_ptr
)
4331 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
4332 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_BUFFER
, false, false);
4334 *res_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_IMAGE
, false, false);
4337 *samp_ptr
= get_sampler_desc(ctx
, instr
->sampler
, AC_DESC_SAMPLER
, false, false);
4339 *samp_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_SAMPLER
, false, false);
4340 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
4341 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4343 if (fmask_ptr
&& !instr
->sampler
&& (instr
->op
== nir_texop_txf_ms
||
4344 instr
->op
== nir_texop_samples_identical
))
4345 *fmask_ptr
= get_sampler_desc(ctx
, instr
->texture
, AC_DESC_FMASK
, false, false);
4348 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
4351 coord
= to_float(ctx
, coord
);
4352 coord
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &coord
, 1, 0);
4353 coord
= to_integer(ctx
, coord
);
4357 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
4359 LLVMValueRef result
= NULL
;
4360 struct ac_image_args args
= { 0 };
4361 unsigned dmask
= 0xf;
4362 LLVMValueRef address
[16];
4363 LLVMValueRef coords
[5];
4364 LLVMValueRef coord
= NULL
, lod
= NULL
, comparator
= NULL
;
4365 LLVMValueRef bias
= NULL
, offsets
= NULL
;
4366 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
, sample_index
= NULL
;
4367 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
4368 LLVMValueRef derivs
[6];
4369 unsigned chan
, count
= 0;
4370 unsigned const_src
= 0, num_deriv_comp
= 0;
4371 bool lod_is_zero
= false;
4373 tex_fetch_ptrs(ctx
, instr
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4375 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4376 switch (instr
->src
[i
].src_type
) {
4377 case nir_tex_src_coord
:
4378 coord
= get_src(ctx
, instr
->src
[i
].src
);
4380 case nir_tex_src_projector
:
4382 case nir_tex_src_comparator
:
4383 comparator
= get_src(ctx
, instr
->src
[i
].src
);
4385 case nir_tex_src_offset
:
4386 offsets
= get_src(ctx
, instr
->src
[i
].src
);
4389 case nir_tex_src_bias
:
4390 bias
= get_src(ctx
, instr
->src
[i
].src
);
4392 case nir_tex_src_lod
: {
4393 nir_const_value
*val
= nir_src_as_const_value(instr
->src
[i
].src
);
4395 if (val
&& val
->i32
[0] == 0)
4397 lod
= get_src(ctx
, instr
->src
[i
].src
);
4400 case nir_tex_src_ms_index
:
4401 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
4403 case nir_tex_src_ms_mcs
:
4405 case nir_tex_src_ddx
:
4406 ddx
= get_src(ctx
, instr
->src
[i
].src
);
4407 num_deriv_comp
= instr
->src
[i
].src
.ssa
->num_components
;
4409 case nir_tex_src_ddy
:
4410 ddy
= get_src(ctx
, instr
->src
[i
].src
);
4412 case nir_tex_src_texture_offset
:
4413 case nir_tex_src_sampler_offset
:
4414 case nir_tex_src_plane
:
4420 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
4421 result
= get_buffer_size(ctx
, res_ptr
, true);
4425 if (instr
->op
== nir_texop_texture_samples
) {
4426 LLVMValueRef res
, samples
, is_msaa
;
4427 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res_ptr
, ctx
->ac
.v8i32
, "");
4428 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4429 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4430 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4431 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
4432 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
4433 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4434 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
4435 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4437 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4438 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
4439 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
4440 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
4441 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
4443 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
4450 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4451 coords
[chan
] = llvm_extract_elem(&ctx
->ac
, coord
, chan
);
4453 if (offsets
&& instr
->op
!= nir_texop_txf
) {
4454 LLVMValueRef offset
[3], pack
;
4455 for (chan
= 0; chan
< 3; ++chan
)
4456 offset
[chan
] = ctx
->ac
.i32_0
;
4459 for (chan
= 0; chan
< get_llvm_num_components(offsets
); chan
++) {
4460 offset
[chan
] = llvm_extract_elem(&ctx
->ac
, offsets
, chan
);
4461 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
4462 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
4464 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
4465 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
4467 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
4468 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
4469 address
[count
++] = pack
;
4472 /* pack LOD bias value */
4473 if (instr
->op
== nir_texop_txb
&& bias
) {
4474 address
[count
++] = bias
;
4477 /* Pack depth comparison value */
4478 if (instr
->is_shadow
&& comparator
) {
4479 address
[count
++] = llvm_extract_elem(&ctx
->ac
, comparator
, 0);
4482 /* pack derivatives */
4484 switch (instr
->sampler_dim
) {
4485 case GLSL_SAMPLER_DIM_3D
:
4486 case GLSL_SAMPLER_DIM_CUBE
:
4489 case GLSL_SAMPLER_DIM_2D
:
4493 case GLSL_SAMPLER_DIM_1D
:
4498 for (unsigned i
= 0; i
< num_deriv_comp
; i
++) {
4499 derivs
[i
] = to_float(&ctx
->ac
, llvm_extract_elem(&ctx
->ac
, ddx
, i
));
4500 derivs
[num_deriv_comp
+ i
] = to_float(&ctx
->ac
, llvm_extract_elem(&ctx
->ac
, ddy
, i
));
4504 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& coord
) {
4505 if (instr
->is_array
&& instr
->op
!= nir_texop_lod
)
4506 coords
[3] = apply_round_slice(&ctx
->ac
, coords
[3]);
4507 for (chan
= 0; chan
< instr
->coord_components
; chan
++)
4508 coords
[chan
] = to_float(&ctx
->ac
, coords
[chan
]);
4509 if (instr
->coord_components
== 3)
4510 coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
4511 ac_prepare_cube_coords(&ctx
->ac
,
4512 instr
->op
== nir_texop_txd
, instr
->is_array
,
4519 for (unsigned i
= 0; i
< num_deriv_comp
* 2; i
++)
4520 address
[count
++] = derivs
[i
];
4523 /* Pack texture coordinates */
4525 address
[count
++] = coords
[0];
4526 if (instr
->coord_components
> 1) {
4527 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&& instr
->is_array
&& instr
->op
!= nir_texop_txf
) {
4528 coords
[1] = apply_round_slice(&ctx
->ac
, coords
[1]);
4530 address
[count
++] = coords
[1];
4532 if (instr
->coord_components
> 2) {
4533 /* This seems like a bit of a hack - but it passes Vulkan CTS with it */
4534 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_3D
&&
4535 instr
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
&&
4536 instr
->op
!= nir_texop_txf
) {
4537 coords
[2] = apply_round_slice(&ctx
->ac
, coords
[2]);
4539 address
[count
++] = coords
[2];
4544 if (lod
&& ((instr
->op
== nir_texop_txl
&& !lod_is_zero
) ||
4545 instr
->op
== nir_texop_txf
)) {
4546 address
[count
++] = lod
;
4547 } else if (instr
->op
== nir_texop_txf_ms
&& sample_index
) {
4548 address
[count
++] = sample_index
;
4549 } else if(instr
->op
== nir_texop_txs
) {
4552 address
[count
++] = lod
;
4554 address
[count
++] = ctx
->ac
.i32_0
;
4557 for (chan
= 0; chan
< count
; chan
++) {
4558 address
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
,
4559 address
[chan
], ctx
->ac
.i32
, "");
4562 if (instr
->op
== nir_texop_samples_identical
) {
4563 LLVMValueRef txf_address
[4];
4564 struct ac_image_args txf_args
= { 0 };
4565 unsigned txf_count
= count
;
4566 memcpy(txf_address
, address
, sizeof(txf_address
));
4568 if (!instr
->is_array
)
4569 txf_address
[2] = ctx
->ac
.i32_0
;
4570 txf_address
[3] = ctx
->ac
.i32_0
;
4572 set_tex_fetch_args(&ctx
->ac
, &txf_args
, instr
, nir_texop_txf
,
4574 txf_address
, txf_count
, 0xf);
4576 result
= build_tex_intrinsic(ctx
, instr
, false, &txf_args
);
4578 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4579 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
4583 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
4584 instr
->op
!= nir_texop_txs
) {
4585 unsigned sample_chan
= instr
->is_array
? 3 : 2;
4586 address
[sample_chan
] = adjust_sample_index_using_fmask(&ctx
->ac
,
4589 instr
->is_array
? address
[2] : NULL
,
4590 address
[sample_chan
],
4594 if (offsets
&& instr
->op
== nir_texop_txf
) {
4595 nir_const_value
*const_offset
=
4596 nir_src_as_const_value(instr
->src
[const_src
].src
);
4597 int num_offsets
= instr
->src
[const_src
].src
.ssa
->num_components
;
4598 assert(const_offset
);
4599 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
4600 if (num_offsets
> 2)
4601 address
[2] = LLVMBuildAdd(ctx
->ac
.builder
,
4602 address
[2], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[2], false), "");
4603 if (num_offsets
> 1)
4604 address
[1] = LLVMBuildAdd(ctx
->ac
.builder
,
4605 address
[1], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[1], false), "");
4606 address
[0] = LLVMBuildAdd(ctx
->ac
.builder
,
4607 address
[0], LLVMConstInt(ctx
->ac
.i32
, const_offset
->i32
[0], false), "");
4611 /* TODO TG4 support */
4612 if (instr
->op
== nir_texop_tg4
) {
4613 if (instr
->is_shadow
)
4616 dmask
= 1 << instr
->component
;
4618 set_tex_fetch_args(&ctx
->ac
, &args
, instr
, instr
->op
,
4619 res_ptr
, samp_ptr
, address
, count
, dmask
);
4621 result
= build_tex_intrinsic(ctx
, instr
, lod_is_zero
, &args
);
4623 if (instr
->op
== nir_texop_query_levels
)
4624 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4625 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
4626 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
4627 instr
->op
!= nir_texop_tg4
)
4628 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4629 else if (instr
->op
== nir_texop_txs
&&
4630 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
4632 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
4633 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
4634 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
4635 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
4636 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
4637 } else if (instr
->dest
.ssa
.num_components
!= 4)
4638 result
= trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
4642 assert(instr
->dest
.is_ssa
);
4643 result
= to_integer(&ctx
->ac
, result
);
4644 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4649 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
4651 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
4652 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
4654 _mesa_hash_table_insert(ctx
->defs
, &instr
->dest
.ssa
, result
);
4655 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
4658 static void visit_post_phi(struct ac_nir_context
*ctx
,
4659 nir_phi_instr
*instr
,
4660 LLVMValueRef llvm_phi
)
4662 nir_foreach_phi_src(src
, instr
) {
4663 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
4664 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
4666 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
4670 static void phi_post_pass(struct ac_nir_context
*ctx
)
4672 struct hash_entry
*entry
;
4673 hash_table_foreach(ctx
->phis
, entry
) {
4674 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
4675 (LLVMValueRef
)entry
->data
);
4680 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
4681 const nir_ssa_undef_instr
*instr
)
4683 unsigned num_components
= instr
->def
.num_components
;
4686 if (num_components
== 1)
4687 undef
= LLVMGetUndef(ctx
->ac
.i32
);
4689 undef
= LLVMGetUndef(LLVMVectorType(ctx
->ac
.i32
, num_components
));
4691 _mesa_hash_table_insert(ctx
->defs
, &instr
->def
, undef
);
4694 static void visit_jump(struct ac_nir_context
*ctx
,
4695 const nir_jump_instr
*instr
)
4697 switch (instr
->type
) {
4698 case nir_jump_break
:
4699 LLVMBuildBr(ctx
->ac
.builder
, ctx
->break_block
);
4700 LLVMClearInsertionPosition(ctx
->ac
.builder
);
4702 case nir_jump_continue
:
4703 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
4704 LLVMClearInsertionPosition(ctx
->ac
.builder
);
4707 fprintf(stderr
, "Unknown NIR jump instr: ");
4708 nir_print_instr(&instr
->instr
, stderr
);
4709 fprintf(stderr
, "\n");
4714 static void visit_cf_list(struct ac_nir_context
*ctx
,
4715 struct exec_list
*list
);
4717 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
4719 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
4720 nir_foreach_instr(instr
, block
)
4722 switch (instr
->type
) {
4723 case nir_instr_type_alu
:
4724 visit_alu(ctx
, nir_instr_as_alu(instr
));
4726 case nir_instr_type_load_const
:
4727 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
4729 case nir_instr_type_intrinsic
:
4730 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
4732 case nir_instr_type_tex
:
4733 visit_tex(ctx
, nir_instr_as_tex(instr
));
4735 case nir_instr_type_phi
:
4736 visit_phi(ctx
, nir_instr_as_phi(instr
));
4738 case nir_instr_type_ssa_undef
:
4739 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
4741 case nir_instr_type_jump
:
4742 visit_jump(ctx
, nir_instr_as_jump(instr
));
4745 fprintf(stderr
, "Unknown NIR instr type: ");
4746 nir_print_instr(instr
, stderr
);
4747 fprintf(stderr
, "\n");
4752 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
4755 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
4757 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
4759 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
4760 LLVMBasicBlockRef merge_block
=
4761 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
4762 LLVMBasicBlockRef if_block
=
4763 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
4764 LLVMBasicBlockRef else_block
= merge_block
;
4765 if (!exec_list_is_empty(&if_stmt
->else_list
))
4766 else_block
= LLVMAppendBasicBlockInContext(
4767 ctx
->ac
.context
, fn
, "");
4769 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, value
,
4770 LLVMConstInt(ctx
->ac
.i32
, 0, false), "");
4771 LLVMBuildCondBr(ctx
->ac
.builder
, cond
, if_block
, else_block
);
4773 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, if_block
);
4774 visit_cf_list(ctx
, &if_stmt
->then_list
);
4775 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
4776 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
4778 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
4779 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, else_block
);
4780 visit_cf_list(ctx
, &if_stmt
->else_list
);
4781 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
4782 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
4785 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, merge_block
);
4788 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
4790 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
4791 LLVMBasicBlockRef continue_parent
= ctx
->continue_block
;
4792 LLVMBasicBlockRef break_parent
= ctx
->break_block
;
4794 ctx
->continue_block
=
4795 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
4797 LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
4799 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
4800 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->continue_block
);
4801 visit_cf_list(ctx
, &loop
->body
);
4803 if (LLVMGetInsertBlock(ctx
->ac
.builder
))
4804 LLVMBuildBr(ctx
->ac
.builder
, ctx
->continue_block
);
4805 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, ctx
->break_block
);
4807 ctx
->continue_block
= continue_parent
;
4808 ctx
->break_block
= break_parent
;
4811 static void visit_cf_list(struct ac_nir_context
*ctx
,
4812 struct exec_list
*list
)
4814 foreach_list_typed(nir_cf_node
, node
, node
, list
)
4816 switch (node
->type
) {
4817 case nir_cf_node_block
:
4818 visit_block(ctx
, nir_cf_node_as_block(node
));
4821 case nir_cf_node_if
:
4822 visit_if(ctx
, nir_cf_node_as_if(node
));
4825 case nir_cf_node_loop
:
4826 visit_loop(ctx
, nir_cf_node_as_loop(node
));
4836 handle_vs_input_decl(struct nir_to_llvm_context
*ctx
,
4837 struct nir_variable
*variable
)
4839 LLVMValueRef t_list_ptr
= ctx
->vertex_buffers
;
4840 LLVMValueRef t_offset
;
4841 LLVMValueRef t_list
;
4843 LLVMValueRef buffer_index
;
4844 int index
= variable
->data
.location
- VERT_ATTRIB_GENERIC0
;
4845 int idx
= variable
->data
.location
;
4846 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
4848 variable
->data
.driver_location
= idx
* 4;
4850 if (ctx
->options
->key
.vs
.instance_rate_inputs
& (1u << index
)) {
4851 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.instance_id
,
4852 ctx
->abi
.start_instance
, "");
4853 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(3,
4854 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
4856 buffer_index
= LLVMBuildAdd(ctx
->builder
, ctx
->abi
.vertex_id
,
4857 ctx
->abi
.base_vertex
, "");
4859 for (unsigned i
= 0; i
< attrib_count
; ++i
, ++idx
) {
4860 t_offset
= LLVMConstInt(ctx
->i32
, index
+ i
, false);
4862 t_list
= ac_build_indexed_load_const(&ctx
->ac
, t_list_ptr
, t_offset
);
4864 input
= ac_build_buffer_load_format(&ctx
->ac
, t_list
,
4866 LLVMConstInt(ctx
->i32
, 0, false),
4869 for (unsigned chan
= 0; chan
< 4; chan
++) {
4870 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->i32
, chan
, false);
4871 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
, chan
)] =
4872 to_integer(&ctx
->ac
, LLVMBuildExtractElement(ctx
->builder
,
4873 input
, llvm_chan
, ""));
4878 static void interp_fs_input(struct nir_to_llvm_context
*ctx
,
4880 LLVMValueRef interp_param
,
4881 LLVMValueRef prim_mask
,
4882 LLVMValueRef result
[4])
4884 LLVMValueRef attr_number
;
4887 bool interp
= interp_param
!= NULL
;
4889 attr_number
= LLVMConstInt(ctx
->i32
, attr
, false);
4891 /* fs.constant returns the param from the middle vertex, so it's not
4892 * really useful for flat shading. It's meant to be used for custom
4893 * interpolation (but the intrinsic can't fetch from the other two
4896 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
4897 * to do the right thing. The only reason we use fs.constant is that
4898 * fs.interp cannot be used on integers, because they can be equal
4902 interp_param
= LLVMBuildBitCast(ctx
->builder
, interp_param
,
4903 LLVMVectorType(ctx
->f32
, 2), "");
4905 i
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
4907 j
= LLVMBuildExtractElement(ctx
->builder
, interp_param
,
4911 for (chan
= 0; chan
< 4; chan
++) {
4912 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->i32
, chan
, false);
4915 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
4920 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
4921 LLVMConstInt(ctx
->i32
, 2, false),
4930 handle_fs_input_decl(struct nir_to_llvm_context
*ctx
,
4931 struct nir_variable
*variable
)
4933 int idx
= variable
->data
.location
;
4934 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
4935 LLVMValueRef interp
;
4937 variable
->data
.driver_location
= idx
* 4;
4938 ctx
->input_mask
|= ((1ull << attrib_count
) - 1) << variable
->data
.location
;
4940 if (glsl_get_base_type(glsl_without_array(variable
->type
)) == GLSL_TYPE_FLOAT
) {
4941 unsigned interp_type
;
4942 if (variable
->data
.sample
) {
4943 interp_type
= INTERP_SAMPLE
;
4944 ctx
->shader_info
->fs
.force_persample
= true;
4945 } else if (variable
->data
.centroid
)
4946 interp_type
= INTERP_CENTROID
;
4948 interp_type
= INTERP_CENTER
;
4950 interp
= lookup_interp_param(ctx
, variable
->data
.interpolation
, interp_type
);
4954 for (unsigned i
= 0; i
< attrib_count
; ++i
)
4955 ctx
->inputs
[radeon_llvm_reg_index_soa(idx
+ i
, 0)] = interp
;
4960 handle_shader_input_decl(struct nir_to_llvm_context
*ctx
,
4961 struct nir_variable
*variable
)
4963 switch (ctx
->stage
) {
4964 case MESA_SHADER_VERTEX
:
4965 handle_vs_input_decl(ctx
, variable
);
4967 case MESA_SHADER_FRAGMENT
:
4968 handle_fs_input_decl(ctx
, variable
);
4977 handle_fs_inputs_pre(struct nir_to_llvm_context
*ctx
,
4978 struct nir_shader
*nir
)
4981 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_INPUTS
; ++i
) {
4982 LLVMValueRef interp_param
;
4983 LLVMValueRef
*inputs
= ctx
->inputs
+radeon_llvm_reg_index_soa(i
, 0);
4985 if (!(ctx
->input_mask
& (1ull << i
)))
4988 if (i
>= VARYING_SLOT_VAR0
|| i
== VARYING_SLOT_PNTC
||
4989 i
== VARYING_SLOT_PRIMITIVE_ID
|| i
== VARYING_SLOT_LAYER
) {
4990 interp_param
= *inputs
;
4991 interp_fs_input(ctx
, index
, interp_param
, ctx
->prim_mask
,
4995 ctx
->shader_info
->fs
.flat_shaded_mask
|= 1u << index
;
4997 } else if (i
== VARYING_SLOT_POS
) {
4998 for(int i
= 0; i
< 3; ++i
)
4999 inputs
[i
] = ctx
->frag_pos
[i
];
5001 inputs
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->f32one
, ctx
->frag_pos
[3]);
5004 ctx
->shader_info
->fs
.num_interp
= index
;
5005 if (ctx
->input_mask
& (1 << VARYING_SLOT_PNTC
))
5006 ctx
->shader_info
->fs
.has_pcoord
= true;
5007 if (ctx
->input_mask
& (1 << VARYING_SLOT_PRIMITIVE_ID
))
5008 ctx
->shader_info
->fs
.prim_id_input
= true;
5009 if (ctx
->input_mask
& (1 << VARYING_SLOT_LAYER
))
5010 ctx
->shader_info
->fs
.layer_input
= true;
5011 ctx
->shader_info
->fs
.input_mask
= ctx
->input_mask
>> VARYING_SLOT_VAR0
;
5015 ac_build_alloca(struct ac_llvm_context
*ac
,
5019 LLVMBuilderRef builder
= ac
->builder
;
5020 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
5021 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
5022 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
5023 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
5024 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
5028 LLVMPositionBuilderBefore(first_builder
, first_instr
);
5030 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
5033 res
= LLVMBuildAlloca(first_builder
, type
, name
);
5034 LLVMBuildStore(builder
, LLVMConstNull(type
), res
);
5036 LLVMDisposeBuilder(first_builder
);
5041 static LLVMValueRef
si_build_alloca_undef(struct ac_llvm_context
*ac
,
5045 LLVMValueRef ptr
= ac_build_alloca(ac
, type
, name
);
5046 LLVMBuildStore(ac
->builder
, LLVMGetUndef(type
), ptr
);
5051 scan_shader_output_decl(struct nir_to_llvm_context
*ctx
,
5052 struct nir_variable
*variable
)
5054 int idx
= variable
->data
.location
+ variable
->data
.index
;
5055 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5056 uint64_t mask_attribs
;
5058 variable
->data
.driver_location
= idx
* 4;
5060 /* tess ctrl has it's own load/store paths for outputs */
5061 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
5064 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
5065 if (ctx
->stage
== MESA_SHADER_VERTEX
||
5066 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
5067 ctx
->stage
== MESA_SHADER_GEOMETRY
) {
5068 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5069 int length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
5070 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
5071 ctx
->shader_info
->vs
.outinfo
.clip_dist_mask
= (1 << ctx
->num_output_clips
) - 1;
5072 ctx
->shader_info
->vs
.outinfo
.cull_dist_mask
= (1 << ctx
->num_output_culls
) - 1;
5074 if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
5075 ctx
->shader_info
->tes
.outinfo
.clip_dist_mask
= (1 << ctx
->num_output_clips
) - 1;
5076 ctx
->shader_info
->tes
.outinfo
.cull_dist_mask
= (1 << ctx
->num_output_culls
) - 1;
5083 mask_attribs
= 1ull << idx
;
5087 ctx
->output_mask
|= mask_attribs
;
5091 handle_shader_output_decl(struct ac_nir_context
*ctx
,
5092 struct nir_shader
*nir
,
5093 struct nir_variable
*variable
)
5095 unsigned output_loc
= variable
->data
.driver_location
/ 4;
5096 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5098 /* tess ctrl has it's own load/store paths for outputs */
5099 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
5102 if (ctx
->stage
== MESA_SHADER_VERTEX
||
5103 ctx
->stage
== MESA_SHADER_TESS_EVAL
||
5104 ctx
->stage
== MESA_SHADER_GEOMETRY
) {
5105 int idx
= variable
->data
.location
+ variable
->data
.index
;
5106 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5107 int length
= nir
->info
.clip_distance_array_size
+
5108 nir
->info
.cull_distance_array_size
;
5117 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
5118 for (unsigned chan
= 0; chan
< 4; chan
++) {
5119 ctx
->outputs
[radeon_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
5120 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
5126 glsl_base_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5127 enum glsl_base_type type
)
5131 case GLSL_TYPE_UINT
:
5132 case GLSL_TYPE_BOOL
:
5133 case GLSL_TYPE_SUBROUTINE
:
5135 case GLSL_TYPE_FLOAT
: /* TODO handle mediump */
5137 case GLSL_TYPE_INT64
:
5138 case GLSL_TYPE_UINT64
:
5140 case GLSL_TYPE_DOUBLE
:
5143 unreachable("unknown GLSL type");
5148 glsl_to_llvm_type(struct nir_to_llvm_context
*ctx
,
5149 const struct glsl_type
*type
)
5151 if (glsl_type_is_scalar(type
)) {
5152 return glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
));
5155 if (glsl_type_is_vector(type
)) {
5156 return LLVMVectorType(
5157 glsl_base_to_llvm_type(ctx
, glsl_get_base_type(type
)),
5158 glsl_get_vector_elements(type
));
5161 if (glsl_type_is_matrix(type
)) {
5162 return LLVMArrayType(
5163 glsl_to_llvm_type(ctx
, glsl_get_column_type(type
)),
5164 glsl_get_matrix_columns(type
));
5167 if (glsl_type_is_array(type
)) {
5168 return LLVMArrayType(
5169 glsl_to_llvm_type(ctx
, glsl_get_array_element(type
)),
5170 glsl_get_length(type
));
5173 assert(glsl_type_is_struct(type
));
5175 LLVMTypeRef member_types
[glsl_get_length(type
)];
5177 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
5179 glsl_to_llvm_type(ctx
,
5180 glsl_get_struct_field(type
, i
));
5183 return LLVMStructTypeInContext(ctx
->context
, member_types
,
5184 glsl_get_length(type
), false);
5188 setup_locals(struct ac_nir_context
*ctx
,
5189 struct nir_function
*func
)
5192 ctx
->num_locals
= 0;
5193 nir_foreach_variable(variable
, &func
->impl
->locals
) {
5194 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5195 variable
->data
.driver_location
= ctx
->num_locals
* 4;
5196 ctx
->num_locals
+= attrib_count
;
5198 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
5202 for (i
= 0; i
< ctx
->num_locals
; i
++) {
5203 for (j
= 0; j
< 4; j
++) {
5204 ctx
->locals
[i
* 4 + j
] =
5205 si_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
5211 setup_shared(struct ac_nir_context
*ctx
,
5212 struct nir_shader
*nir
)
5214 nir_foreach_variable(variable
, &nir
->shared
) {
5215 LLVMValueRef shared
=
5216 LLVMAddGlobalInAddressSpace(
5217 ctx
->ac
.module
, glsl_to_llvm_type(ctx
->nctx
, variable
->type
),
5218 variable
->name
? variable
->name
: "",
5220 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
5225 emit_float_saturate(struct ac_llvm_context
*ctx
, LLVMValueRef v
, float lo
, float hi
)
5227 v
= to_float(ctx
, v
);
5228 v
= emit_intrin_2f_param(ctx
, "llvm.maxnum.f32", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, lo
));
5229 return emit_intrin_2f_param(ctx
, "llvm.minnum.f32", ctx
->f32
, v
, LLVMConstReal(ctx
->f32
, hi
));
5233 static LLVMValueRef
emit_pack_int16(struct nir_to_llvm_context
*ctx
,
5234 LLVMValueRef src0
, LLVMValueRef src1
)
5236 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
5237 LLVMValueRef comp
[2];
5239 comp
[0] = LLVMBuildAnd(ctx
->builder
, src0
, LLVMConstInt(ctx
-> i32
, 65535, 0), "");
5240 comp
[1] = LLVMBuildAnd(ctx
->builder
, src1
, LLVMConstInt(ctx
-> i32
, 65535, 0), "");
5241 comp
[1] = LLVMBuildShl(ctx
->builder
, comp
[1], const16
, "");
5242 return LLVMBuildOr(ctx
->builder
, comp
[0], comp
[1], "");
5245 /* Initialize arguments for the shader export intrinsic */
5247 si_llvm_init_export_args(struct nir_to_llvm_context
*ctx
,
5248 LLVMValueRef
*values
,
5250 struct ac_export_args
*args
)
5252 /* Default is 0xf. Adjusted below depending on the format. */
5253 args
->enabled_channels
= 0xf;
5255 /* Specify whether the EXEC mask represents the valid mask */
5256 args
->valid_mask
= 0;
5258 /* Specify whether this is the last export */
5261 /* Specify the target we are exporting */
5262 args
->target
= target
;
5264 args
->compr
= false;
5265 args
->out
[0] = LLVMGetUndef(ctx
->f32
);
5266 args
->out
[1] = LLVMGetUndef(ctx
->f32
);
5267 args
->out
[2] = LLVMGetUndef(ctx
->f32
);
5268 args
->out
[3] = LLVMGetUndef(ctx
->f32
);
5273 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& target
>= V_008DFC_SQ_EXP_MRT
) {
5274 LLVMValueRef val
[4];
5275 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
5276 unsigned col_format
= (ctx
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
5277 bool is_int8
= (ctx
->options
->key
.fs
.is_int8
>> index
) & 1;
5279 switch(col_format
) {
5280 case V_028714_SPI_SHADER_ZERO
:
5281 args
->enabled_channels
= 0; /* writemask */
5282 args
->target
= V_008DFC_SQ_EXP_NULL
;
5285 case V_028714_SPI_SHADER_32_R
:
5286 args
->enabled_channels
= 1;
5287 args
->out
[0] = values
[0];
5290 case V_028714_SPI_SHADER_32_GR
:
5291 args
->enabled_channels
= 0x3;
5292 args
->out
[0] = values
[0];
5293 args
->out
[1] = values
[1];
5296 case V_028714_SPI_SHADER_32_AR
:
5297 args
->enabled_channels
= 0x9;
5298 args
->out
[0] = values
[0];
5299 args
->out
[3] = values
[3];
5302 case V_028714_SPI_SHADER_FP16_ABGR
:
5305 for (unsigned chan
= 0; chan
< 2; chan
++) {
5306 LLVMValueRef pack_args
[2] = {
5308 values
[2 * chan
+ 1]
5310 LLVMValueRef packed
;
5312 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
5313 args
->out
[chan
] = packed
;
5317 case V_028714_SPI_SHADER_UNORM16_ABGR
:
5318 for (unsigned chan
= 0; chan
< 4; chan
++) {
5319 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
5320 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5321 LLVMConstReal(ctx
->f32
, 65535), "");
5322 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5323 LLVMConstReal(ctx
->f32
, 0.5), "");
5324 val
[chan
] = LLVMBuildFPToUI(ctx
->builder
, val
[chan
],
5329 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5330 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5333 case V_028714_SPI_SHADER_SNORM16_ABGR
:
5334 for (unsigned chan
= 0; chan
< 4; chan
++) {
5335 val
[chan
] = emit_float_saturate(&ctx
->ac
, values
[chan
], -1, 1);
5336 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
5337 LLVMConstReal(ctx
->f32
, 32767), "");
5339 /* If positive, add 0.5, else add -0.5. */
5340 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
5341 LLVMBuildSelect(ctx
->builder
,
5342 LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
,
5343 val
[chan
], ctx
->f32zero
, ""),
5344 LLVMConstReal(ctx
->f32
, 0.5),
5345 LLVMConstReal(ctx
->f32
, -0.5), ""), "");
5346 val
[chan
] = LLVMBuildFPToSI(ctx
->builder
, val
[chan
], ctx
->i32
, "");
5350 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5351 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5354 case V_028714_SPI_SHADER_UINT16_ABGR
: {
5355 LLVMValueRef max
= LLVMConstInt(ctx
->i32
, is_int8
? 255 : 65535, 0);
5357 for (unsigned chan
= 0; chan
< 4; chan
++) {
5358 val
[chan
] = to_integer(&ctx
->ac
, values
[chan
]);
5359 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntULT
, val
[chan
], max
);
5363 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5364 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5368 case V_028714_SPI_SHADER_SINT16_ABGR
: {
5369 LLVMValueRef max
= LLVMConstInt(ctx
->i32
, is_int8
? 127 : 32767, 0);
5370 LLVMValueRef min
= LLVMConstInt(ctx
->i32
, is_int8
? -128 : -32768, 0);
5373 for (unsigned chan
= 0; chan
< 4; chan
++) {
5374 val
[chan
] = to_integer(&ctx
->ac
, values
[chan
]);
5375 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSLT
, val
[chan
], max
);
5376 val
[chan
] = emit_minmax_int(&ctx
->ac
, LLVMIntSGT
, val
[chan
], min
);
5380 args
->out
[0] = emit_pack_int16(ctx
, val
[0], val
[1]);
5381 args
->out
[1] = emit_pack_int16(ctx
, val
[2], val
[3]);
5386 case V_028714_SPI_SHADER_32_ABGR
:
5387 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5391 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
5393 for (unsigned i
= 0; i
< 4; ++i
)
5394 args
->out
[i
] = to_float(&ctx
->ac
, args
->out
[i
]);
5398 handle_vs_outputs_post(struct nir_to_llvm_context
*ctx
,
5399 bool export_prim_id
,
5400 struct ac_vs_output_info
*outinfo
)
5402 uint32_t param_count
= 0;
5404 unsigned pos_idx
, num_pos_exports
= 0;
5405 struct ac_export_args args
, pos_args
[4] = {};
5406 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
5409 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
5410 sizeof(outinfo
->vs_output_param_offset
));
5412 if (ctx
->output_mask
& (1ull << VARYING_SLOT_CLIP_DIST0
)) {
5413 LLVMValueRef slots
[8];
5416 if (outinfo
->cull_dist_mask
)
5417 outinfo
->cull_dist_mask
<<= ctx
->num_output_clips
;
5419 i
= VARYING_SLOT_CLIP_DIST0
;
5420 for (j
= 0; j
< ctx
->num_output_clips
+ ctx
->num_output_culls
; j
++)
5421 slots
[j
] = to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5422 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5424 for (i
= ctx
->num_output_clips
+ ctx
->num_output_culls
; i
< 8; i
++)
5425 slots
[i
] = LLVMGetUndef(ctx
->f32
);
5427 if (ctx
->num_output_clips
+ ctx
->num_output_culls
> 4) {
5428 target
= V_008DFC_SQ_EXP_POS
+ 3;
5429 si_llvm_init_export_args(ctx
, &slots
[4], target
, &args
);
5430 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5431 &args
, sizeof(args
));
5434 target
= V_008DFC_SQ_EXP_POS
+ 2;
5435 si_llvm_init_export_args(ctx
, &slots
[0], target
, &args
);
5436 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5437 &args
, sizeof(args
));
5441 LLVMValueRef pos_values
[4] = {ctx
->f32zero
, ctx
->f32zero
, ctx
->f32zero
, ctx
->f32one
};
5442 if (ctx
->output_mask
& (1ull << VARYING_SLOT_POS
)) {
5443 for (unsigned j
= 0; j
< 4; j
++)
5444 pos_values
[j
] = LLVMBuildLoad(ctx
->builder
,
5445 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_POS
, j
)], "");
5447 si_llvm_init_export_args(ctx
, pos_values
, V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
5449 if (ctx
->output_mask
& (1ull << VARYING_SLOT_PSIZ
)) {
5450 outinfo
->writes_pointsize
= true;
5451 psize_value
= LLVMBuildLoad(ctx
->builder
,
5452 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_PSIZ
, 0)], "");
5455 if (ctx
->output_mask
& (1ull << VARYING_SLOT_LAYER
)) {
5456 outinfo
->writes_layer
= true;
5457 layer_value
= LLVMBuildLoad(ctx
->builder
,
5458 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)], "");
5461 if (ctx
->output_mask
& (1ull << VARYING_SLOT_VIEWPORT
)) {
5462 outinfo
->writes_viewport_index
= true;
5463 viewport_index_value
= LLVMBuildLoad(ctx
->builder
,
5464 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(VARYING_SLOT_VIEWPORT
, 0)], "");
5467 uint32_t mask
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
5468 (outinfo
->writes_layer
== true ? 4 : 0) |
5469 (outinfo
->writes_viewport_index
== true ? 8 : 0));
5471 pos_args
[1].enabled_channels
= mask
;
5472 pos_args
[1].valid_mask
= 0;
5473 pos_args
[1].done
= 0;
5474 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
5475 pos_args
[1].compr
= 0;
5476 pos_args
[1].out
[0] = ctx
->f32zero
; /* X */
5477 pos_args
[1].out
[1] = ctx
->f32zero
; /* Y */
5478 pos_args
[1].out
[2] = ctx
->f32zero
; /* Z */
5479 pos_args
[1].out
[3] = ctx
->f32zero
; /* W */
5481 if (outinfo
->writes_pointsize
== true)
5482 pos_args
[1].out
[0] = psize_value
;
5483 if (outinfo
->writes_layer
== true)
5484 pos_args
[1].out
[2] = layer_value
;
5485 if (outinfo
->writes_viewport_index
== true)
5486 pos_args
[1].out
[3] = viewport_index_value
;
5488 for (i
= 0; i
< 4; i
++) {
5489 if (pos_args
[i
].out
[0])
5494 for (i
= 0; i
< 4; i
++) {
5495 if (!pos_args
[i
].out
[0])
5498 /* Specify the target we are exporting */
5499 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
5500 if (pos_idx
== num_pos_exports
)
5501 pos_args
[i
].done
= 1;
5502 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
5505 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5506 LLVMValueRef values
[4];
5507 if (!(ctx
->output_mask
& (1ull << i
)))
5510 for (unsigned j
= 0; j
< 4; j
++)
5511 values
[j
] = to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5512 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5514 if (i
== VARYING_SLOT_LAYER
) {
5515 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5516 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = param_count
;
5518 } else if (i
== VARYING_SLOT_PRIMITIVE_ID
) {
5519 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5520 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
5522 } else if (i
>= VARYING_SLOT_VAR0
) {
5523 outinfo
->export_mask
|= 1u << (i
- VARYING_SLOT_VAR0
);
5524 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5525 outinfo
->vs_output_param_offset
[i
] = param_count
;
5530 si_llvm_init_export_args(ctx
, values
, target
, &args
);
5532 if (target
>= V_008DFC_SQ_EXP_POS
&&
5533 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
5534 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
5535 &args
, sizeof(args
));
5537 ac_build_export(&ctx
->ac
, &args
);
5541 if (export_prim_id
) {
5542 LLVMValueRef values
[4];
5543 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
5544 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
;
5547 values
[0] = ctx
->vs_prim_id
;
5548 ctx
->shader_info
->vs
.vgpr_comp_cnt
= MAX2(2,
5549 ctx
->shader_info
->vs
.vgpr_comp_cnt
);
5550 for (unsigned j
= 1; j
< 4; j
++)
5551 values
[j
] = ctx
->f32zero
;
5552 si_llvm_init_export_args(ctx
, values
, target
, &args
);
5553 ac_build_export(&ctx
->ac
, &args
);
5554 outinfo
->export_prim_id
= true;
5557 outinfo
->pos_exports
= num_pos_exports
;
5558 outinfo
->param_exports
= param_count
;
5562 handle_es_outputs_post(struct nir_to_llvm_context
*ctx
,
5563 struct ac_es_output_info
*outinfo
)
5566 uint64_t max_output_written
= 0;
5567 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5568 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
5572 if (!(ctx
->output_mask
& (1ull << i
)))
5575 if (i
== VARYING_SLOT_CLIP_DIST0
)
5576 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
5578 param_index
= shader_io_get_unique_index(i
);
5580 max_output_written
= MAX2(param_index
+ (length
> 4), max_output_written
);
5582 for (j
= 0; j
< length
; j
++) {
5583 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], "");
5584 out_val
= LLVMBuildBitCast(ctx
->builder
, out_val
, ctx
->i32
, "");
5586 ac_build_buffer_store_dword(&ctx
->ac
,
5589 NULL
, ctx
->es2gs_offset
,
5590 (4 * param_index
+ j
) * 4,
5594 outinfo
->esgs_itemsize
= (max_output_written
+ 1) * 16;
5598 handle_ls_outputs_post(struct nir_to_llvm_context
*ctx
)
5600 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
5601 LLVMValueRef vertex_dw_stride
= unpack_param(ctx
, ctx
->ls_out_layout
, 13, 8);
5602 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->builder
, vertex_id
,
5603 vertex_dw_stride
, "");
5605 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5606 LLVMValueRef
*out_ptr
= &ctx
->nir
->outputs
[i
* 4];
5609 if (!(ctx
->output_mask
& (1ull << i
)))
5612 if (i
== VARYING_SLOT_CLIP_DIST0
)
5613 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
5614 int param
= shader_io_get_unique_index(i
);
5615 mark_tess_output(ctx
, false, param
);
5617 mark_tess_output(ctx
, false, param
+ 1);
5618 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->builder
, base_dw_addr
,
5619 LLVMConstInt(ctx
->i32
, param
* 4, false),
5621 for (unsigned j
= 0; j
< length
; j
++) {
5622 lds_store(ctx
, dw_addr
,
5623 LLVMBuildLoad(ctx
->builder
, out_ptr
[j
], ""));
5624 dw_addr
= LLVMBuildAdd(ctx
->builder
, dw_addr
, ctx
->i32one
, "");
5629 struct ac_build_if_state
5631 struct nir_to_llvm_context
*ctx
;
5632 LLVMValueRef condition
;
5633 LLVMBasicBlockRef entry_block
;
5634 LLVMBasicBlockRef true_block
;
5635 LLVMBasicBlockRef false_block
;
5636 LLVMBasicBlockRef merge_block
;
5639 static LLVMBasicBlockRef
5640 ac_build_insert_new_block(struct nir_to_llvm_context
*ctx
, const char *name
)
5642 LLVMBasicBlockRef current_block
;
5643 LLVMBasicBlockRef next_block
;
5644 LLVMBasicBlockRef new_block
;
5646 /* get current basic block */
5647 current_block
= LLVMGetInsertBlock(ctx
->builder
);
5649 /* chqeck if there's another block after this one */
5650 next_block
= LLVMGetNextBasicBlock(current_block
);
5652 /* insert the new block before the next block */
5653 new_block
= LLVMInsertBasicBlockInContext(ctx
->context
, next_block
, name
);
5656 /* append new block after current block */
5657 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
5658 new_block
= LLVMAppendBasicBlockInContext(ctx
->context
, function
, name
);
5664 ac_nir_build_if(struct ac_build_if_state
*ifthen
,
5665 struct nir_to_llvm_context
*ctx
,
5666 LLVMValueRef condition
)
5668 LLVMBasicBlockRef block
= LLVMGetInsertBlock(ctx
->builder
);
5670 memset(ifthen
, 0, sizeof *ifthen
);
5672 ifthen
->condition
= condition
;
5673 ifthen
->entry_block
= block
;
5675 /* create endif/merge basic block for the phi functions */
5676 ifthen
->merge_block
= ac_build_insert_new_block(ctx
, "endif-block");
5678 /* create/insert true_block before merge_block */
5679 ifthen
->true_block
=
5680 LLVMInsertBasicBlockInContext(ctx
->context
,
5681 ifthen
->merge_block
,
5684 /* successive code goes into the true block */
5685 LLVMPositionBuilderAtEnd(ctx
->builder
, ifthen
->true_block
);
5689 * End a conditional.
5692 ac_nir_build_endif(struct ac_build_if_state
*ifthen
)
5694 LLVMBuilderRef builder
= ifthen
->ctx
->builder
;
5696 /* Insert branch to the merge block from current block */
5697 LLVMBuildBr(builder
, ifthen
->merge_block
);
5700 * Now patch in the various branch instructions.
5703 /* Insert the conditional branch instruction at the end of entry_block */
5704 LLVMPositionBuilderAtEnd(builder
, ifthen
->entry_block
);
5705 if (ifthen
->false_block
) {
5706 /* we have an else clause */
5707 LLVMBuildCondBr(builder
, ifthen
->condition
,
5708 ifthen
->true_block
, ifthen
->false_block
);
5711 /* no else clause */
5712 LLVMBuildCondBr(builder
, ifthen
->condition
,
5713 ifthen
->true_block
, ifthen
->merge_block
);
5716 /* Resume building code at end of the ifthen->merge_block */
5717 LLVMPositionBuilderAtEnd(builder
, ifthen
->merge_block
);
5721 write_tess_factors(struct nir_to_llvm_context
*ctx
)
5723 unsigned stride
, outer_comps
, inner_comps
;
5724 struct ac_build_if_state if_ctx
, inner_if_ctx
;
5725 LLVMValueRef invocation_id
= unpack_param(ctx
, ctx
->tcs_rel_ids
, 8, 5);
5726 LLVMValueRef rel_patch_id
= unpack_param(ctx
, ctx
->tcs_rel_ids
, 0, 8);
5727 unsigned tess_inner_index
, tess_outer_index
;
5728 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
5729 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
5733 switch (ctx
->options
->key
.tcs
.primitive_mode
) {
5753 ac_nir_build_if(&if_ctx
, ctx
,
5754 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
5755 invocation_id
, ctx
->i32zero
, ""));
5757 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
5758 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
5760 mark_tess_output(ctx
, true, tess_inner_index
);
5761 mark_tess_output(ctx
, true, tess_outer_index
);
5762 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
5763 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_base
,
5764 LLVMConstInt(ctx
->i32
, tess_inner_index
* 4, false), "");
5765 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_base
,
5766 LLVMConstInt(ctx
->i32
, tess_outer_index
* 4, false), "");
5768 for (i
= 0; i
< 4; i
++) {
5769 inner
[i
] = LLVMGetUndef(ctx
->i32
);
5770 outer
[i
] = LLVMGetUndef(ctx
->i32
);
5774 if (ctx
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
5775 outer
[0] = out
[1] = lds_load(ctx
, lds_outer
);
5776 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
5777 LLVMConstInt(ctx
->i32
, 1, false), "");
5778 outer
[1] = out
[0] = lds_load(ctx
, lds_outer
);
5780 for (i
= 0; i
< outer_comps
; i
++) {
5782 lds_load(ctx
, lds_outer
);
5783 lds_outer
= LLVMBuildAdd(ctx
->builder
, lds_outer
,
5784 LLVMConstInt(ctx
->i32
, 1, false), "");
5786 for (i
= 0; i
< inner_comps
; i
++) {
5787 inner
[i
] = out
[outer_comps
+i
] =
5788 lds_load(ctx
, lds_inner
);
5789 lds_inner
= LLVMBuildAdd(ctx
->builder
, lds_inner
,
5790 LLVMConstInt(ctx
->i32
, 1, false), "");
5794 /* Convert the outputs to vectors for stores. */
5795 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
5799 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
5802 buffer
= ctx
->hs_ring_tess_factor
;
5803 tf_base
= ctx
->tess_factor_offset
;
5804 byteoffset
= LLVMBuildMul(ctx
->builder
, rel_patch_id
,
5805 LLVMConstInt(ctx
->i32
, 4 * stride
, false), "");
5807 ac_nir_build_if(&inner_if_ctx
, ctx
,
5808 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
5809 rel_patch_id
, ctx
->i32zero
, ""));
5811 /* Store the dynamic HS control word. */
5812 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
5813 LLVMConstInt(ctx
->i32
, 0x80000000, false),
5814 1, ctx
->i32zero
, tf_base
,
5815 0, 1, 0, true, false);
5816 ac_nir_build_endif(&inner_if_ctx
);
5818 /* Store the tessellation factors. */
5819 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
5820 MIN2(stride
, 4), byteoffset
, tf_base
,
5821 4, 1, 0, true, false);
5823 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
5824 stride
- 4, byteoffset
, tf_base
,
5825 20, 1, 0, true, false);
5827 //TODO store to offchip for TES to read - only if TES reads them
5829 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
5830 LLVMValueRef tf_inner_offset
;
5831 unsigned param_outer
, param_inner
;
5833 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
5834 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
5835 LLVMConstInt(ctx
->i32
, param_outer
, 0));
5837 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
5838 util_next_power_of_two(outer_comps
));
5840 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
5841 outer_comps
, tf_outer_offset
,
5842 ctx
->oc_lds
, 0, 1, 0, true, false);
5844 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
5845 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
5846 LLVMConstInt(ctx
->i32
, param_inner
, 0));
5848 inner_vec
= inner_comps
== 1 ? inner
[0] :
5849 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
5850 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
5851 inner_comps
, tf_inner_offset
,
5852 ctx
->oc_lds
, 0, 1, 0, true, false);
5855 ac_nir_build_endif(&if_ctx
);
5859 handle_tcs_outputs_post(struct nir_to_llvm_context
*ctx
)
5861 write_tess_factors(ctx
);
5865 si_export_mrt_color(struct nir_to_llvm_context
*ctx
,
5866 LLVMValueRef
*color
, unsigned param
, bool is_last
,
5867 struct ac_export_args
*args
)
5870 si_llvm_init_export_args(ctx
, color
, param
,
5874 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
5875 args
->done
= 1; /* DONE bit */
5876 } else if (!args
->enabled_channels
)
5877 return false; /* unnecessary NULL export */
5883 si_export_mrt_z(struct nir_to_llvm_context
*ctx
,
5884 LLVMValueRef depth
, LLVMValueRef stencil
,
5885 LLVMValueRef samplemask
)
5887 struct ac_export_args args
;
5889 args
.enabled_channels
= 0;
5890 args
.valid_mask
= 1;
5892 args
.target
= V_008DFC_SQ_EXP_MRTZ
;
5895 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R, depth */
5896 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G, stencil test val[0:7], stencil op val[8:15] */
5897 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B, sample mask */
5898 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A, alpha to mask */
5901 args
.out
[0] = depth
;
5902 args
.enabled_channels
|= 0x1;
5906 args
.out
[1] = stencil
;
5907 args
.enabled_channels
|= 0x2;
5911 args
.out
[2] = samplemask
;
5912 args
.enabled_channels
|= 0x4;
5915 /* SI (except OLAND and HAINAN) has a bug that it only looks
5916 * at the X writemask component. */
5917 if (ctx
->options
->chip_class
== SI
&&
5918 ctx
->options
->family
!= CHIP_OLAND
&&
5919 ctx
->options
->family
!= CHIP_HAINAN
)
5920 args
.enabled_channels
|= 0x1;
5922 ac_build_export(&ctx
->ac
, &args
);
5926 handle_fs_outputs_post(struct nir_to_llvm_context
*ctx
)
5929 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
5930 struct ac_export_args color_args
[8];
5932 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
5933 LLVMValueRef values
[4];
5935 if (!(ctx
->output_mask
& (1ull << i
)))
5938 if (i
== FRAG_RESULT_DEPTH
) {
5939 ctx
->shader_info
->fs
.writes_z
= true;
5940 depth
= to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5941 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
5942 } else if (i
== FRAG_RESULT_STENCIL
) {
5943 ctx
->shader_info
->fs
.writes_stencil
= true;
5944 stencil
= to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5945 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
5946 } else if (i
== FRAG_RESULT_SAMPLE_MASK
) {
5947 ctx
->shader_info
->fs
.writes_sample_mask
= true;
5948 samplemask
= to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5949 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, 0)], ""));
5952 for (unsigned j
= 0; j
< 4; j
++)
5953 values
[j
] = to_float(&ctx
->ac
, LLVMBuildLoad(ctx
->builder
,
5954 ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)], ""));
5956 if (!ctx
->shader_info
->fs
.writes_z
&& !ctx
->shader_info
->fs
.writes_stencil
&& !ctx
->shader_info
->fs
.writes_sample_mask
)
5957 last
= ctx
->output_mask
<= ((1ull << (i
+ 1)) - 1);
5959 bool ret
= si_export_mrt_color(ctx
, values
, V_008DFC_SQ_EXP_MRT
+ (i
- FRAG_RESULT_DATA0
), last
, &color_args
[index
]);
5965 for (unsigned i
= 0; i
< index
; i
++)
5966 ac_build_export(&ctx
->ac
, &color_args
[i
]);
5967 if (depth
|| stencil
|| samplemask
)
5968 si_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
5970 si_export_mrt_color(ctx
, NULL
, V_008DFC_SQ_EXP_NULL
, true, &color_args
[0]);
5971 ac_build_export(&ctx
->ac
, &color_args
[0]);
5974 ctx
->shader_info
->fs
.output_mask
= index
? ((1ull << index
) - 1) : 0;
5978 emit_gs_epilogue(struct nir_to_llvm_context
*ctx
)
5980 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
5984 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
5985 LLVMValueRef
*addrs
)
5987 struct nir_to_llvm_context
*ctx
= nir_to_llvm_context_from_abi(abi
);
5989 switch (ctx
->stage
) {
5990 case MESA_SHADER_VERTEX
:
5991 if (ctx
->options
->key
.vs
.as_ls
)
5992 handle_ls_outputs_post(ctx
);
5993 else if (ctx
->options
->key
.vs
.as_es
)
5994 handle_es_outputs_post(ctx
, &ctx
->shader_info
->vs
.es_info
);
5996 handle_vs_outputs_post(ctx
, ctx
->options
->key
.vs
.export_prim_id
,
5997 &ctx
->shader_info
->vs
.outinfo
);
5999 case MESA_SHADER_FRAGMENT
:
6000 handle_fs_outputs_post(ctx
);
6002 case MESA_SHADER_GEOMETRY
:
6003 emit_gs_epilogue(ctx
);
6005 case MESA_SHADER_TESS_CTRL
:
6006 handle_tcs_outputs_post(ctx
);
6008 case MESA_SHADER_TESS_EVAL
:
6009 if (ctx
->options
->key
.tes
.as_es
)
6010 handle_es_outputs_post(ctx
, &ctx
->shader_info
->tes
.es_info
);
6012 handle_vs_outputs_post(ctx
, ctx
->options
->key
.tes
.export_prim_id
,
6013 &ctx
->shader_info
->tes
.outinfo
);
6020 static void ac_llvm_finalize_module(struct nir_to_llvm_context
* ctx
)
6022 LLVMPassManagerRef passmgr
;
6023 /* Create the pass manager */
6024 passmgr
= LLVMCreateFunctionPassManagerForModule(
6027 /* This pass should eliminate all the load and store instructions */
6028 LLVMAddPromoteMemoryToRegisterPass(passmgr
);
6030 /* Add some optimization passes */
6031 LLVMAddScalarReplAggregatesPass(passmgr
);
6032 LLVMAddLICMPass(passmgr
);
6033 LLVMAddAggressiveDCEPass(passmgr
);
6034 LLVMAddCFGSimplificationPass(passmgr
);
6035 LLVMAddInstructionCombiningPass(passmgr
);
6038 LLVMInitializeFunctionPassManager(passmgr
);
6039 LLVMRunFunctionPassManager(passmgr
, ctx
->main_function
);
6040 LLVMFinalizeFunctionPassManager(passmgr
);
6042 LLVMDisposeBuilder(ctx
->builder
);
6043 LLVMDisposePassManager(passmgr
);
6047 ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context
*ctx
)
6049 struct ac_vs_output_info
*outinfo
;
6051 switch (ctx
->stage
) {
6052 case MESA_SHADER_FRAGMENT
:
6053 case MESA_SHADER_COMPUTE
:
6054 case MESA_SHADER_TESS_CTRL
:
6055 case MESA_SHADER_GEOMETRY
:
6057 case MESA_SHADER_VERTEX
:
6058 if (ctx
->options
->key
.vs
.as_ls
||
6059 ctx
->options
->key
.vs
.as_es
)
6061 outinfo
= &ctx
->shader_info
->vs
.outinfo
;
6063 case MESA_SHADER_TESS_EVAL
:
6064 if (ctx
->options
->key
.vs
.as_es
)
6066 outinfo
= &ctx
->shader_info
->tes
.outinfo
;
6069 unreachable("Unhandled shader type");
6072 ac_optimize_vs_outputs(&ctx
->ac
,
6074 outinfo
->vs_output_param_offset
,
6076 &outinfo
->param_exports
);
6080 ac_setup_rings(struct nir_to_llvm_context
*ctx
)
6082 if ((ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->options
->key
.vs
.as_es
) ||
6083 (ctx
->stage
== MESA_SHADER_TESS_EVAL
&& ctx
->options
->key
.tes
.as_es
)) {
6084 ctx
->esgs_ring
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_ESGS_VS
, false));
6087 if (ctx
->is_gs_copy_shader
) {
6088 ctx
->gsvs_ring
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_GSVS_VS
, false));
6090 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
6092 ctx
->esgs_ring
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_ESGS_GS
, false));
6093 ctx
->gsvs_ring
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_GSVS_GS
, false));
6095 ctx
->gsvs_ring
= LLVMBuildBitCast(ctx
->builder
, ctx
->gsvs_ring
, ctx
->v4i32
, "");
6097 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->gsvs_num_entries
, LLVMConstInt(ctx
->i32
, 2, false), "");
6098 tmp
= LLVMBuildExtractElement(ctx
->builder
, ctx
->gsvs_ring
, ctx
->i32one
, "");
6099 tmp
= LLVMBuildOr(ctx
->builder
, tmp
, ctx
->gsvs_ring_stride
, "");
6100 ctx
->gsvs_ring
= LLVMBuildInsertElement(ctx
->builder
, ctx
->gsvs_ring
, tmp
, ctx
->i32one
, "");
6103 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
6104 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
6105 ctx
->hs_ring_tess_offchip
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_HS_TESS_OFFCHIP
, false));
6106 ctx
->hs_ring_tess_factor
= ac_build_indexed_load_const(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->i32
, RING_HS_TESS_FACTOR
, false));
6111 ac_nir_get_max_workgroup_size(enum chip_class chip_class
,
6112 const struct nir_shader
*nir
)
6114 switch (nir
->stage
) {
6115 case MESA_SHADER_TESS_CTRL
:
6116 return chip_class
>= CIK
? 128 : 64;
6117 case MESA_SHADER_GEOMETRY
:
6119 case MESA_SHADER_COMPUTE
:
6125 unsigned max_workgroup_size
= nir
->info
.cs
.local_size
[0] *
6126 nir
->info
.cs
.local_size
[1] *
6127 nir
->info
.cs
.local_size
[2];
6128 return max_workgroup_size
;
6131 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
6132 struct nir_shader
*nir
, struct nir_to_llvm_context
*nctx
)
6134 struct ac_nir_context ctx
= {};
6135 struct nir_function
*func
;
6144 ctx
.stage
= nir
->stage
;
6146 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
6148 nir_foreach_variable(variable
, &nir
->outputs
)
6149 handle_shader_output_decl(&ctx
, nir
, variable
);
6151 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6152 _mesa_key_pointer_equal
);
6153 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6154 _mesa_key_pointer_equal
);
6155 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
6156 _mesa_key_pointer_equal
);
6158 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
6160 setup_locals(&ctx
, func
);
6162 if (nir
->stage
== MESA_SHADER_COMPUTE
)
6163 setup_shared(&ctx
, nir
);
6165 visit_cf_list(&ctx
, &func
->impl
->body
);
6166 phi_post_pass(&ctx
);
6168 ctx
.abi
->emit_outputs(ctx
.abi
, RADEON_LLVM_MAX_OUTPUTS
,
6172 ralloc_free(ctx
.defs
);
6173 ralloc_free(ctx
.phis
);
6174 ralloc_free(ctx
.vars
);
6181 LLVMModuleRef
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm
,
6182 struct nir_shader
*nir
,
6183 struct ac_shader_variant_info
*shader_info
,
6184 const struct ac_nir_compiler_options
*options
)
6186 struct nir_to_llvm_context ctx
= {0};
6188 ctx
.options
= options
;
6189 ctx
.shader_info
= shader_info
;
6190 ctx
.context
= LLVMContextCreate();
6191 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
6193 ac_llvm_context_init(&ctx
.ac
, ctx
.context
);
6194 ctx
.ac
.module
= ctx
.module
;
6196 ctx
.has_ds_bpermute
= ctx
.options
->chip_class
>= VI
;
6198 memset(shader_info
, 0, sizeof(*shader_info
));
6200 ac_nir_shader_info_pass(nir
, options
, &shader_info
->info
);
6202 LLVMSetTarget(ctx
.module
, options
->supports_spill
? "amdgcn-mesa-mesa3d" : "amdgcn--");
6204 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
6205 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
6206 LLVMSetDataLayout(ctx
.module
, data_layout_str
);
6207 LLVMDisposeTargetData(data_layout
);
6208 LLVMDisposeMessage(data_layout_str
);
6212 ctx
.builder
= LLVMCreateBuilderInContext(ctx
.context
);
6213 ctx
.ac
.builder
= ctx
.builder
;
6214 ctx
.stage
= nir
->stage
;
6215 ctx
.max_workgroup_size
= ac_nir_get_max_workgroup_size(ctx
.options
->chip_class
, nir
);
6217 for (i
= 0; i
< AC_UD_MAX_SETS
; i
++)
6218 shader_info
->user_sgprs_locs
.descriptor_sets
[i
].sgpr_idx
= -1;
6219 for (i
= 0; i
< AC_UD_MAX_UD
; i
++)
6220 shader_info
->user_sgprs_locs
.shader_data
[i
].sgpr_idx
= -1;
6222 create_function(&ctx
);
6224 if (nir
->stage
== MESA_SHADER_GEOMETRY
) {
6225 ctx
.gs_next_vertex
= ac_build_alloca(&ctx
.ac
, ctx
.i32
, "gs_next_vertex");
6227 ctx
.gs_max_out_vertices
= nir
->info
.gs
.vertices_out
;
6228 } else if (nir
->stage
== MESA_SHADER_TESS_EVAL
) {
6229 ctx
.tes_primitive_mode
= nir
->info
.tess
.primitive_mode
;
6230 } else if (nir
->stage
== MESA_SHADER_VERTEX
) {
6231 if (shader_info
->info
.vs
.needs_instance_id
) {
6232 ctx
.shader_info
->vs
.vgpr_comp_cnt
=
6233 MAX2(3, ctx
.shader_info
->vs
.vgpr_comp_cnt
);
6235 } else if (nir
->stage
== MESA_SHADER_FRAGMENT
) {
6236 shader_info
->fs
.can_discard
= nir
->info
.fs
.uses_discard
;
6239 ac_setup_rings(&ctx
);
6241 ctx
.num_output_clips
= nir
->info
.clip_distance_array_size
;
6242 ctx
.num_output_culls
= nir
->info
.cull_distance_array_size
;
6244 nir_foreach_variable(variable
, &nir
->inputs
)
6245 handle_shader_input_decl(&ctx
, variable
);
6247 if (nir
->stage
== MESA_SHADER_FRAGMENT
)
6248 handle_fs_inputs_pre(&ctx
, nir
);
6250 ctx
.abi
.chip_class
= options
->chip_class
;
6251 ctx
.abi
.inputs
= &ctx
.inputs
[0];
6252 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
6253 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
6255 nir_foreach_variable(variable
, &nir
->outputs
)
6256 scan_shader_output_decl(&ctx
, variable
);
6258 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, nir
, &ctx
);
6260 LLVMBuildRetVoid(ctx
.builder
);
6262 ac_llvm_finalize_module(&ctx
);
6264 ac_nir_eliminate_const_vs_outputs(&ctx
);
6266 if (nir
->stage
== MESA_SHADER_GEOMETRY
) {
6267 unsigned addclip
= ctx
.num_output_clips
+ ctx
.num_output_culls
> 4;
6268 shader_info
->gs
.gsvs_vertex_size
= (util_bitcount64(ctx
.output_mask
) + addclip
) * 16;
6269 shader_info
->gs
.max_gsvs_emit_size
= shader_info
->gs
.gsvs_vertex_size
*
6270 nir
->info
.gs
.vertices_out
;
6271 } else if (nir
->stage
== MESA_SHADER_TESS_CTRL
) {
6272 shader_info
->tcs
.outputs_written
= ctx
.tess_outputs_written
;
6273 shader_info
->tcs
.patch_outputs_written
= ctx
.tess_patch_outputs_written
;
6274 } else if (nir
->stage
== MESA_SHADER_VERTEX
&& ctx
.options
->key
.vs
.as_ls
) {
6275 shader_info
->vs
.outputs_written
= ctx
.tess_outputs_written
;
6281 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
6283 unsigned *retval
= (unsigned *)context
;
6284 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
6285 char *description
= LLVMGetDiagInfoDescription(di
);
6287 if (severity
== LLVMDSError
) {
6289 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
6293 LLVMDisposeMessage(description
);
6296 static unsigned ac_llvm_compile(LLVMModuleRef M
,
6297 struct ac_shader_binary
*binary
,
6298 LLVMTargetMachineRef tm
)
6300 unsigned retval
= 0;
6302 LLVMContextRef llvm_ctx
;
6303 LLVMMemoryBufferRef out_buffer
;
6304 unsigned buffer_size
;
6305 const char *buffer_data
;
6308 /* Setup Diagnostic Handler*/
6309 llvm_ctx
= LLVMGetModuleContext(M
);
6311 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
6315 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
,
6318 /* Process Errors/Warnings */
6320 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
6326 /* Extract Shader Code*/
6327 buffer_size
= LLVMGetBufferSize(out_buffer
);
6328 buffer_data
= LLVMGetBufferStart(out_buffer
);
6330 ac_elf_read(buffer_data
, buffer_size
, binary
);
6333 LLVMDisposeMemoryBuffer(out_buffer
);
6339 static void ac_compile_llvm_module(LLVMTargetMachineRef tm
,
6340 LLVMModuleRef llvm_module
,
6341 struct ac_shader_binary
*binary
,
6342 struct ac_shader_config
*config
,
6343 struct ac_shader_variant_info
*shader_info
,
6344 gl_shader_stage stage
,
6345 bool dump_shader
, bool supports_spill
)
6348 ac_dump_module(llvm_module
);
6350 memset(binary
, 0, sizeof(*binary
));
6351 int v
= ac_llvm_compile(llvm_module
, binary
, tm
);
6353 fprintf(stderr
, "compile failed\n");
6357 fprintf(stderr
, "disasm:\n%s\n", binary
->disasm_string
);
6359 ac_shader_binary_read_config(binary
, config
, 0, supports_spill
);
6361 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
6362 LLVMDisposeModule(llvm_module
);
6363 LLVMContextDispose(ctx
);
6365 if (stage
== MESA_SHADER_FRAGMENT
) {
6366 shader_info
->num_input_vgprs
= 0;
6367 if (G_0286CC_PERSP_SAMPLE_ENA(config
->spi_ps_input_addr
))
6368 shader_info
->num_input_vgprs
+= 2;
6369 if (G_0286CC_PERSP_CENTER_ENA(config
->spi_ps_input_addr
))
6370 shader_info
->num_input_vgprs
+= 2;
6371 if (G_0286CC_PERSP_CENTROID_ENA(config
->spi_ps_input_addr
))
6372 shader_info
->num_input_vgprs
+= 2;
6373 if (G_0286CC_PERSP_PULL_MODEL_ENA(config
->spi_ps_input_addr
))
6374 shader_info
->num_input_vgprs
+= 3;
6375 if (G_0286CC_LINEAR_SAMPLE_ENA(config
->spi_ps_input_addr
))
6376 shader_info
->num_input_vgprs
+= 2;
6377 if (G_0286CC_LINEAR_CENTER_ENA(config
->spi_ps_input_addr
))
6378 shader_info
->num_input_vgprs
+= 2;
6379 if (G_0286CC_LINEAR_CENTROID_ENA(config
->spi_ps_input_addr
))
6380 shader_info
->num_input_vgprs
+= 2;
6381 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config
->spi_ps_input_addr
))
6382 shader_info
->num_input_vgprs
+= 1;
6383 if (G_0286CC_POS_X_FLOAT_ENA(config
->spi_ps_input_addr
))
6384 shader_info
->num_input_vgprs
+= 1;
6385 if (G_0286CC_POS_Y_FLOAT_ENA(config
->spi_ps_input_addr
))
6386 shader_info
->num_input_vgprs
+= 1;
6387 if (G_0286CC_POS_Z_FLOAT_ENA(config
->spi_ps_input_addr
))
6388 shader_info
->num_input_vgprs
+= 1;
6389 if (G_0286CC_POS_W_FLOAT_ENA(config
->spi_ps_input_addr
))
6390 shader_info
->num_input_vgprs
+= 1;
6391 if (G_0286CC_FRONT_FACE_ENA(config
->spi_ps_input_addr
))
6392 shader_info
->num_input_vgprs
+= 1;
6393 if (G_0286CC_ANCILLARY_ENA(config
->spi_ps_input_addr
))
6394 shader_info
->num_input_vgprs
+= 1;
6395 if (G_0286CC_SAMPLE_COVERAGE_ENA(config
->spi_ps_input_addr
))
6396 shader_info
->num_input_vgprs
+= 1;
6397 if (G_0286CC_POS_FIXED_PT_ENA(config
->spi_ps_input_addr
))
6398 shader_info
->num_input_vgprs
+= 1;
6400 config
->num_vgprs
= MAX2(config
->num_vgprs
, shader_info
->num_input_vgprs
);
6402 /* +3 for scratch wave offset and VCC */
6403 config
->num_sgprs
= MAX2(config
->num_sgprs
,
6404 shader_info
->num_input_sgprs
+ 3);
6407 void ac_compile_nir_shader(LLVMTargetMachineRef tm
,
6408 struct ac_shader_binary
*binary
,
6409 struct ac_shader_config
*config
,
6410 struct ac_shader_variant_info
*shader_info
,
6411 struct nir_shader
*nir
,
6412 const struct ac_nir_compiler_options
*options
,
6416 LLVMModuleRef llvm_module
= ac_translate_nir_to_llvm(tm
, nir
, shader_info
,
6419 ac_compile_llvm_module(tm
, llvm_module
, binary
, config
, shader_info
, nir
->stage
, dump_shader
, options
->supports_spill
);
6420 switch (nir
->stage
) {
6421 case MESA_SHADER_COMPUTE
:
6422 for (int i
= 0; i
< 3; ++i
)
6423 shader_info
->cs
.block_size
[i
] = nir
->info
.cs
.local_size
[i
];
6425 case MESA_SHADER_FRAGMENT
:
6426 shader_info
->fs
.early_fragment_test
= nir
->info
.fs
.early_fragment_tests
;
6428 case MESA_SHADER_GEOMETRY
:
6429 shader_info
->gs
.vertices_in
= nir
->info
.gs
.vertices_in
;
6430 shader_info
->gs
.vertices_out
= nir
->info
.gs
.vertices_out
;
6431 shader_info
->gs
.output_prim
= nir
->info
.gs
.output_primitive
;
6432 shader_info
->gs
.invocations
= nir
->info
.gs
.invocations
;
6434 case MESA_SHADER_TESS_EVAL
:
6435 shader_info
->tes
.primitive_mode
= nir
->info
.tess
.primitive_mode
;
6436 shader_info
->tes
.spacing
= nir
->info
.tess
.spacing
;
6437 shader_info
->tes
.ccw
= nir
->info
.tess
.ccw
;
6438 shader_info
->tes
.point_mode
= nir
->info
.tess
.point_mode
;
6439 shader_info
->tes
.as_es
= options
->key
.tes
.as_es
;
6441 case MESA_SHADER_TESS_CTRL
:
6442 shader_info
->tcs
.tcs_vertices_out
= nir
->info
.tess
.tcs_vertices_out
;
6444 case MESA_SHADER_VERTEX
:
6445 shader_info
->vs
.as_es
= options
->key
.vs
.as_es
;
6446 shader_info
->vs
.as_ls
= options
->key
.vs
.as_ls
;
6447 /* in LS mode we need at least 1, invocation id needs 3, handled elsewhere */
6448 if (options
->key
.vs
.as_ls
)
6449 shader_info
->vs
.vgpr_comp_cnt
= MAX2(1, shader_info
->vs
.vgpr_comp_cnt
);
6457 ac_gs_copy_shader_emit(struct nir_to_llvm_context
*ctx
)
6459 LLVMValueRef args
[9];
6460 args
[0] = ctx
->gsvs_ring
;
6461 args
[1] = LLVMBuildMul(ctx
->builder
, ctx
->abi
.vertex_id
, LLVMConstInt(ctx
->i32
, 4, false), "");
6462 args
[3] = ctx
->i32zero
;
6463 args
[4] = ctx
->i32one
; /* OFFEN */
6464 args
[5] = ctx
->i32zero
; /* IDXEN */
6465 args
[6] = ctx
->i32one
; /* GLC */
6466 args
[7] = ctx
->i32one
; /* SLC */
6467 args
[8] = ctx
->i32zero
; /* TFE */
6471 for (unsigned i
= 0; i
< RADEON_LLVM_MAX_OUTPUTS
; ++i
) {
6475 if (!(ctx
->output_mask
& (1ull << i
)))
6478 if (i
== VARYING_SLOT_CLIP_DIST0
) {
6479 /* unpack clip and cull from a single set of slots */
6480 length
= ctx
->num_output_clips
+ ctx
->num_output_culls
;
6485 for (unsigned j
= 0; j
< length
; j
++) {
6487 args
[2] = LLVMConstInt(ctx
->i32
,
6489 ctx
->gs_max_out_vertices
* 16 * 4, false);
6491 value
= ac_build_intrinsic(&ctx
->ac
,
6492 "llvm.SI.buffer.load.dword.i32.i32",
6494 AC_FUNC_ATTR_READONLY
|
6495 AC_FUNC_ATTR_LEGACY
);
6497 LLVMBuildStore(ctx
->builder
,
6498 to_float(&ctx
->ac
, value
), ctx
->nir
->outputs
[radeon_llvm_reg_index_soa(i
, j
)]);
6502 handle_vs_outputs_post(ctx
, false, &ctx
->shader_info
->vs
.outinfo
);
6505 void ac_create_gs_copy_shader(LLVMTargetMachineRef tm
,
6506 struct nir_shader
*geom_shader
,
6507 struct ac_shader_binary
*binary
,
6508 struct ac_shader_config
*config
,
6509 struct ac_shader_variant_info
*shader_info
,
6510 const struct ac_nir_compiler_options
*options
,
6513 struct nir_to_llvm_context ctx
= {0};
6514 ctx
.context
= LLVMContextCreate();
6515 ctx
.module
= LLVMModuleCreateWithNameInContext("shader", ctx
.context
);
6516 ctx
.options
= options
;
6517 ctx
.shader_info
= shader_info
;
6519 ac_llvm_context_init(&ctx
.ac
, ctx
.context
);
6520 ctx
.ac
.module
= ctx
.module
;
6522 ctx
.is_gs_copy_shader
= true;
6523 LLVMSetTarget(ctx
.module
, "amdgcn--");
6526 ctx
.builder
= LLVMCreateBuilderInContext(ctx
.context
);
6527 ctx
.ac
.builder
= ctx
.builder
;
6528 ctx
.stage
= MESA_SHADER_VERTEX
;
6530 create_function(&ctx
);
6532 ctx
.gs_max_out_vertices
= geom_shader
->info
.gs
.vertices_out
;
6533 ac_setup_rings(&ctx
);
6535 ctx
.num_output_clips
= geom_shader
->info
.clip_distance_array_size
;
6536 ctx
.num_output_culls
= geom_shader
->info
.cull_distance_array_size
;
6538 struct ac_nir_context nir_ctx
= {};
6539 nir_ctx
.ac
= ctx
.ac
;
6540 nir_ctx
.abi
= &ctx
.abi
;
6542 nir_ctx
.nctx
= &ctx
;
6545 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
6546 scan_shader_output_decl(&ctx
, variable
);
6547 handle_shader_output_decl(&nir_ctx
, geom_shader
, variable
);
6550 ac_gs_copy_shader_emit(&ctx
);
6554 LLVMBuildRetVoid(ctx
.builder
);
6556 ac_llvm_finalize_module(&ctx
);
6558 ac_compile_llvm_module(tm
, ctx
.module
, binary
, config
, shader_info
,
6560 dump_shader
, options
->supports_spill
);