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 "nir/nir_deref.h"
31 #include "util/bitscan.h"
32 #include "util/u_math.h"
33 #include "ac_shader_abi.h"
34 #include "ac_shader_util.h"
36 struct ac_nir_context
{
37 struct ac_llvm_context ac
;
38 struct ac_shader_abi
*abi
;
40 gl_shader_stage stage
;
42 LLVMValueRef
*ssa_defs
;
44 struct hash_table
*defs
;
45 struct hash_table
*phis
;
46 struct hash_table
*vars
;
48 LLVMValueRef main_function
;
49 LLVMBasicBlockRef continue_block
;
50 LLVMBasicBlockRef break_block
;
56 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
57 nir_deref_instr
*deref_instr
,
58 enum ac_descriptor_type desc_type
,
59 const nir_instr
*instr
,
60 bool image
, bool write
);
63 build_store_values_extended(struct ac_llvm_context
*ac
,
66 unsigned value_stride
,
69 LLVMBuilderRef builder
= ac
->builder
;
72 for (i
= 0; i
< value_count
; i
++) {
73 LLVMValueRef ptr
= values
[i
* value_stride
];
74 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
75 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
76 LLVMBuildStore(builder
, value
, ptr
);
80 static enum ac_image_dim
81 get_ac_sampler_dim(const struct ac_llvm_context
*ctx
, enum glsl_sampler_dim dim
,
85 case GLSL_SAMPLER_DIM_1D
:
86 if (ctx
->chip_class
>= GFX9
)
87 return is_array
? ac_image_2darray
: ac_image_2d
;
88 return is_array
? ac_image_1darray
: ac_image_1d
;
89 case GLSL_SAMPLER_DIM_2D
:
90 case GLSL_SAMPLER_DIM_RECT
:
91 case GLSL_SAMPLER_DIM_EXTERNAL
:
92 return is_array
? ac_image_2darray
: ac_image_2d
;
93 case GLSL_SAMPLER_DIM_3D
:
95 case GLSL_SAMPLER_DIM_CUBE
:
97 case GLSL_SAMPLER_DIM_MS
:
98 return is_array
? ac_image_2darraymsaa
: ac_image_2dmsaa
;
99 case GLSL_SAMPLER_DIM_SUBPASS
:
100 return ac_image_2darray
;
101 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
102 return ac_image_2darraymsaa
;
104 unreachable("bad sampler dim");
108 static enum ac_image_dim
109 get_ac_image_dim(const struct ac_llvm_context
*ctx
, enum glsl_sampler_dim sdim
,
112 enum ac_image_dim dim
= get_ac_sampler_dim(ctx
, sdim
, is_array
);
114 if (dim
== ac_image_cube
||
115 (ctx
->chip_class
<= VI
&& dim
== ac_image_3d
))
116 dim
= ac_image_2darray
;
121 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
122 const nir_ssa_def
*def
)
124 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
125 if (def
->num_components
> 1) {
126 type
= LLVMVectorType(type
, def
->num_components
);
131 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
134 return nir
->ssa_defs
[src
.ssa
->index
];
138 get_memory_ptr(struct ac_nir_context
*ctx
, nir_src src
)
140 LLVMValueRef ptr
= get_src(ctx
, src
);
141 ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ctx
->ac
.lds
, &ptr
, 1, "");
142 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
144 return LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
145 LLVMPointerType(ctx
->ac
.i32
, addr_space
), "");
148 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
149 const struct nir_block
*b
)
151 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
152 return (LLVMBasicBlockRef
)entry
->data
;
155 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
157 unsigned num_components
)
159 LLVMValueRef value
= get_src(ctx
, src
.src
);
160 bool need_swizzle
= false;
163 unsigned src_components
= ac_get_llvm_num_components(value
);
164 for (unsigned i
= 0; i
< num_components
; ++i
) {
165 assert(src
.swizzle
[i
] < src_components
);
166 if (src
.swizzle
[i
] != i
)
170 if (need_swizzle
|| num_components
!= src_components
) {
171 LLVMValueRef masks
[] = {
172 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
173 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
174 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
175 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
177 if (src_components
> 1 && num_components
== 1) {
178 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
180 } else if (src_components
== 1 && num_components
> 1) {
181 LLVMValueRef values
[] = {value
, value
, value
, value
};
182 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
184 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
185 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
194 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
195 LLVMIntPredicate pred
, LLVMValueRef src0
,
198 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
199 return LLVMBuildSelect(ctx
->builder
, result
,
200 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
204 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
205 LLVMRealPredicate pred
, LLVMValueRef src0
,
209 src0
= ac_to_float(ctx
, src0
);
210 src1
= ac_to_float(ctx
, src1
);
211 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
212 return LLVMBuildSelect(ctx
->builder
, result
,
213 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
217 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
219 LLVMTypeRef result_type
,
223 LLVMValueRef params
[] = {
224 ac_to_float(ctx
, src0
),
227 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
228 ac_get_elem_bits(ctx
, result_type
));
229 assert(length
< sizeof(name
));
230 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
233 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
235 LLVMTypeRef result_type
,
236 LLVMValueRef src0
, LLVMValueRef src1
)
239 LLVMValueRef params
[] = {
240 ac_to_float(ctx
, src0
),
241 ac_to_float(ctx
, src1
),
244 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
245 ac_get_elem_bits(ctx
, result_type
));
246 assert(length
< sizeof(name
));
247 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
250 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
252 LLVMTypeRef result_type
,
253 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
256 LLVMValueRef params
[] = {
257 ac_to_float(ctx
, src0
),
258 ac_to_float(ctx
, src1
),
259 ac_to_float(ctx
, src2
),
262 MAYBE_UNUSED
const int length
= snprintf(name
, sizeof(name
), "%s.f%d", intrin
,
263 ac_get_elem_bits(ctx
, result_type
));
264 assert(length
< sizeof(name
));
265 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
268 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
269 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
271 assert(LLVMGetTypeKind(LLVMTypeOf(src0
)) != LLVMVectorTypeKind
);
273 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
275 return LLVMBuildSelect(ctx
->builder
, v
,
276 ac_to_integer_or_pointer(ctx
, src1
),
277 ac_to_integer_or_pointer(ctx
, src2
), "");
280 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
283 return ac_build_imax(ctx
, src0
, LLVMBuildNeg(ctx
->builder
, src0
, ""));
286 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
288 LLVMValueRef src0
, LLVMValueRef src1
)
290 LLVMTypeRef ret_type
;
291 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
293 LLVMValueRef params
[] = { src0
, src1
};
294 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
297 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
298 params
, 2, AC_FUNC_ATTR_READNONE
);
300 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
301 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
305 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
309 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
,
310 LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""),
312 result
= LLVMBuildBitCast(ctx
->builder
, result
, ctx
->f32
, "");
316 return LLVMBuildFPTrunc(ctx
->builder
, result
, ctx
->f16
, "");
320 return LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f64
, "");
322 unreachable("Unsupported bit size.");
326 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
329 src0
= ac_to_float(ctx
, src0
);
330 LLVMValueRef zero
= LLVMConstNull(LLVMTypeOf(src0
));
331 return LLVMBuildSExt(ctx
->builder
,
332 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, zero
, ""),
336 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
340 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
344 return LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i8
, "");
346 return LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i16
, "");
350 return LLVMBuildZExt(ctx
->builder
, result
, ctx
->i64
, "");
352 unreachable("Unsupported bit size.");
356 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
359 LLVMValueRef zero
= LLVMConstNull(LLVMTypeOf(src0
));
360 return LLVMBuildSExt(ctx
->builder
,
361 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, zero
, ""),
365 static LLVMValueRef
emit_f2f16(struct ac_llvm_context
*ctx
,
369 LLVMValueRef cond
= NULL
;
371 src0
= ac_to_float(ctx
, src0
);
372 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->f16
, "");
374 if (ctx
->chip_class
>= VI
) {
375 LLVMValueRef args
[2];
376 /* Check if the result is a denormal - and flush to 0 if so. */
378 args
[1] = LLVMConstInt(ctx
->i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
379 cond
= ac_build_intrinsic(ctx
, "llvm.amdgcn.class.f16", ctx
->i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
382 /* need to convert back up to f32 */
383 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f32
, "");
385 if (ctx
->chip_class
>= VI
)
386 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
389 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
390 * so compare the result and flush to 0 if it's smaller.
392 LLVMValueRef temp
, cond2
;
393 temp
= emit_intrin_1f_param(ctx
, "llvm.fabs", ctx
->f32
, result
);
394 cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUGT
,
395 LLVMBuildBitCast(ctx
->builder
, LLVMConstInt(ctx
->i32
, 0x38800000, false), ctx
->f32
, ""),
397 cond2
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
398 temp
, ctx
->f32_0
, "");
399 cond
= LLVMBuildAnd(ctx
->builder
, cond
, cond2
, "");
400 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
405 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
406 LLVMValueRef src0
, LLVMValueRef src1
)
408 LLVMValueRef dst64
, result
;
409 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
410 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
412 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
413 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
414 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
418 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
419 LLVMValueRef src0
, LLVMValueRef src1
)
421 LLVMValueRef dst64
, result
;
422 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
423 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
425 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
426 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
427 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
431 static LLVMValueRef
emit_bitfield_extract(struct ac_llvm_context
*ctx
,
433 const LLVMValueRef srcs
[3])
437 if (HAVE_LLVM
>= 0x0800) {
438 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
439 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
440 result
= LLVMBuildSelect(ctx
->builder
, icond
, srcs
[0], result
, "");
442 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
443 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
445 LLVMValueRef zero
= ctx
->i32_0
;
446 LLVMValueRef icond1
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], LLVMConstInt(ctx
->i32
, 32, false), "");
447 LLVMValueRef icond2
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, srcs
[2], zero
, "");
449 result
= ac_build_bfe(ctx
, srcs
[0], srcs
[1], srcs
[2], is_signed
);
450 result
= LLVMBuildSelect(ctx
->builder
, icond1
, srcs
[0], result
, "");
451 result
= LLVMBuildSelect(ctx
->builder
, icond2
, zero
, result
, "");
457 static LLVMValueRef
emit_bitfield_insert(struct ac_llvm_context
*ctx
,
458 LLVMValueRef src0
, LLVMValueRef src1
,
459 LLVMValueRef src2
, LLVMValueRef src3
)
461 LLVMValueRef bfi_args
[3], result
;
463 bfi_args
[0] = LLVMBuildShl(ctx
->builder
,
464 LLVMBuildSub(ctx
->builder
,
465 LLVMBuildShl(ctx
->builder
,
470 bfi_args
[1] = LLVMBuildShl(ctx
->builder
, src1
, src2
, "");
473 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, src3
, LLVMConstInt(ctx
->i32
, 32, false), "");
476 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
477 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
479 result
= LLVMBuildXor(ctx
->builder
, bfi_args
[2],
480 LLVMBuildAnd(ctx
->builder
, bfi_args
[0],
481 LLVMBuildXor(ctx
->builder
, bfi_args
[1], bfi_args
[2], ""), ""), "");
483 result
= LLVMBuildSelect(ctx
->builder
, icond
, src1
, result
, "");
487 static LLVMValueRef
emit_pack_half_2x16(struct ac_llvm_context
*ctx
,
490 LLVMValueRef comp
[2];
492 src0
= ac_to_float(ctx
, src0
);
493 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
494 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
496 return LLVMBuildBitCast(ctx
->builder
, ac_build_cvt_pkrtz_f16(ctx
, comp
),
500 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
503 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
504 LLVMValueRef temps
[2], val
;
507 for (i
= 0; i
< 2; i
++) {
508 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
509 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
510 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
511 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
513 return ac_build_gather_values(ctx
, temps
, 2);
516 static LLVMValueRef
emit_ddxy(struct ac_nir_context
*ctx
,
524 if (op
== nir_op_fddx_fine
)
525 mask
= AC_TID_MASK_LEFT
;
526 else if (op
== nir_op_fddy_fine
)
527 mask
= AC_TID_MASK_TOP
;
529 mask
= AC_TID_MASK_TOP_LEFT
;
531 /* for DDX we want to next X pixel, DDY next Y pixel. */
532 if (op
== nir_op_fddx_fine
||
533 op
== nir_op_fddx_coarse
||
539 result
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, src0
);
543 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
545 LLVMValueRef src
[4], result
= NULL
;
546 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
547 unsigned src_components
;
548 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
550 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
557 case nir_op_pack_half_2x16
:
560 case nir_op_unpack_half_2x16
:
563 case nir_op_cube_face_coord
:
564 case nir_op_cube_face_index
:
568 src_components
= num_components
;
571 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
572 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
580 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
581 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
584 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
587 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
590 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
593 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
594 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
595 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
598 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
599 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
600 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
603 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
606 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
609 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
612 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
615 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
616 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
617 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
618 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
619 ac_to_float_type(&ctx
->ac
, def_type
), result
);
620 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
621 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
624 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
625 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
626 result
= LLVMBuildFRem(ctx
->ac
.builder
, src
[0], src
[1], "");
629 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
632 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
635 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
638 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
639 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
640 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
643 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
644 result
= ac_build_fdiv(&ctx
->ac
, LLVMConstReal(LLVMTypeOf(src
[0]), 1.0), src
[0]);
647 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
650 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
653 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
656 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
657 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
658 LLVMTypeOf(src
[0]), "");
659 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
660 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
661 LLVMTypeOf(src
[0]), "");
662 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0], src
[1], "");
665 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
666 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
667 LLVMTypeOf(src
[0]), "");
668 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
669 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
670 LLVMTypeOf(src
[0]), "");
671 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0], src
[1], "");
674 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
675 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
676 LLVMTypeOf(src
[0]), "");
677 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
678 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
679 LLVMTypeOf(src
[0]), "");
680 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0], src
[1], "");
683 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
686 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
689 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
692 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
695 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
698 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
701 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOEQ
, src
[0], src
[1]);
704 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
707 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOLT
, src
[0], src
[1]);
710 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOGE
, src
[0], src
[1]);
713 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
714 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
717 result
= emit_iabs(&ctx
->ac
, src
[0]);
720 result
= ac_build_imax(&ctx
->ac
, src
[0], src
[1]);
723 result
= ac_build_imin(&ctx
->ac
, src
[0], src
[1]);
726 result
= ac_build_umax(&ctx
->ac
, src
[0], src
[1]);
729 result
= ac_build_umin(&ctx
->ac
, src
[0], src
[1]);
732 result
= ac_build_isign(&ctx
->ac
, src
[0],
733 instr
->dest
.dest
.ssa
.bit_size
);
736 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
737 result
= ac_build_fsign(&ctx
->ac
, src
[0],
738 instr
->dest
.dest
.ssa
.bit_size
);
741 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
742 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
745 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
746 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
749 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
750 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
752 case nir_op_fround_even
:
753 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
754 ac_to_float_type(&ctx
->ac
, def_type
),src
[0]);
757 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
758 result
= ac_build_fract(&ctx
->ac
, src
[0],
759 instr
->dest
.dest
.ssa
.bit_size
);
762 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
763 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
766 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
767 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
770 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
771 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
774 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
775 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
778 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
779 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
782 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
783 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
784 result
= ac_build_fdiv(&ctx
->ac
, LLVMConstReal(LLVMTypeOf(result
), 1.0), result
);
786 case nir_op_frexp_exp
:
787 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
788 result
= ac_build_frexp_exp(&ctx
->ac
, src
[0],
789 ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])));
790 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) == 16)
791 result
= LLVMBuildSExt(ctx
->ac
.builder
, result
,
794 case nir_op_frexp_sig
:
795 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
796 result
= ac_build_frexp_mant(&ctx
->ac
, src
[0],
797 instr
->dest
.dest
.ssa
.bit_size
);
800 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
801 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
804 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
805 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
806 if (ctx
->ac
.chip_class
< GFX9
&&
807 instr
->dest
.dest
.ssa
.bit_size
== 32) {
808 /* Only pre-GFX9 chips do not flush denorms. */
809 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
810 ac_to_float_type(&ctx
->ac
, def_type
),
815 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
816 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
817 if (ctx
->ac
.chip_class
< GFX9
&&
818 instr
->dest
.dest
.ssa
.bit_size
== 32) {
819 /* Only pre-GFX9 chips do not flush denorms. */
820 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.canonicalize",
821 ac_to_float_type(&ctx
->ac
, def_type
),
826 result
= emit_intrin_3f_param(&ctx
->ac
, "llvm.fmuladd",
827 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
830 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
831 if (ac_get_elem_bits(&ctx
->ac
, def_type
) == 32)
832 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f32", ctx
->ac
.f32
, src
, 2, AC_FUNC_ATTR_READNONE
);
833 else if (ac_get_elem_bits(&ctx
->ac
, def_type
) == 16)
834 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f16", ctx
->ac
.f16
, src
, 2, AC_FUNC_ATTR_READNONE
);
836 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f64", ctx
->ac
.f64
, src
, 2, AC_FUNC_ATTR_READNONE
);
838 case nir_op_ibitfield_extract
:
839 result
= emit_bitfield_extract(&ctx
->ac
, true, src
);
841 case nir_op_ubitfield_extract
:
842 result
= emit_bitfield_extract(&ctx
->ac
, false, src
);
844 case nir_op_bitfield_insert
:
845 result
= emit_bitfield_insert(&ctx
->ac
, src
[0], src
[1], src
[2], src
[3]);
847 case nir_op_bitfield_reverse
:
848 result
= ac_build_bitfield_reverse(&ctx
->ac
, src
[0]);
850 case nir_op_bit_count
:
851 result
= ac_build_bit_count(&ctx
->ac
, src
[0]);
856 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
857 src
[i
] = ac_to_integer(&ctx
->ac
, src
[i
]);
858 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
864 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
865 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
871 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
872 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
877 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
882 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
884 case nir_op_f2f16_rtz
:
885 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
886 if (LLVMTypeOf(src
[0]) == ctx
->ac
.f64
)
887 src
[0] = LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ctx
->ac
.f32
, "");
888 LLVMValueRef param
[2] = { src
[0], ctx
->ac
.f32_0
};
889 result
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, param
);
890 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
892 case nir_op_f2f16_rtne
:
896 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
897 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
898 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
900 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
906 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
907 result
= LLVMBuildZExt(ctx
->ac
.builder
, src
[0], def_type
, "");
909 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
915 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
916 result
= LLVMBuildSExt(ctx
->ac
.builder
, src
[0], def_type
, "");
918 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
921 result
= emit_bcsel(&ctx
->ac
, src
[0], src
[1], src
[2]);
923 case nir_op_find_lsb
:
924 result
= ac_find_lsb(&ctx
->ac
, ctx
->ac
.i32
, src
[0]);
926 case nir_op_ufind_msb
:
927 result
= ac_build_umsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
929 case nir_op_ifind_msb
:
930 result
= ac_build_imsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
932 case nir_op_uadd_carry
:
933 result
= emit_uint_carry(&ctx
->ac
, "llvm.uadd.with.overflow.i32", src
[0], src
[1]);
935 case nir_op_usub_borrow
:
936 result
= emit_uint_carry(&ctx
->ac
, "llvm.usub.with.overflow.i32", src
[0], src
[1]);
941 result
= emit_b2f(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
944 result
= emit_f2b(&ctx
->ac
, src
[0]);
950 result
= emit_b2i(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
953 result
= emit_i2b(&ctx
->ac
, src
[0]);
955 case nir_op_fquantize2f16
:
956 result
= emit_f2f16(&ctx
->ac
, src
[0]);
958 case nir_op_umul_high
:
959 result
= emit_umul_high(&ctx
->ac
, src
[0], src
[1]);
961 case nir_op_imul_high
:
962 result
= emit_imul_high(&ctx
->ac
, src
[0], src
[1]);
964 case nir_op_pack_half_2x16
:
965 result
= emit_pack_half_2x16(&ctx
->ac
, src
[0]);
967 case nir_op_unpack_half_2x16
:
968 result
= emit_unpack_half_2x16(&ctx
->ac
, src
[0]);
972 case nir_op_fddx_fine
:
973 case nir_op_fddy_fine
:
974 case nir_op_fddx_coarse
:
975 case nir_op_fddy_coarse
:
976 result
= emit_ddxy(ctx
, instr
->op
, src
[0]);
979 case nir_op_unpack_64_2x32_split_x
: {
980 assert(ac_get_llvm_num_components(src
[0]) == 1);
981 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
984 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
989 case nir_op_unpack_64_2x32_split_y
: {
990 assert(ac_get_llvm_num_components(src
[0]) == 1);
991 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
994 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
999 case nir_op_pack_64_2x32_split
: {
1000 LLVMValueRef tmp
= ac_build_gather_values(&ctx
->ac
, src
, 2);
1001 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i64
, "");
1005 case nir_op_pack_32_2x16_split
: {
1006 LLVMValueRef tmp
= ac_build_gather_values(&ctx
->ac
, src
, 2);
1007 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
1011 case nir_op_unpack_32_2x16_split_x
: {
1012 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1015 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1020 case nir_op_unpack_32_2x16_split_y
: {
1021 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1024 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1029 case nir_op_cube_face_coord
: {
1030 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1031 LLVMValueRef results
[2];
1033 for (unsigned chan
= 0; chan
< 3; chan
++)
1034 in
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src
[0], chan
);
1035 results
[0] = ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubesc",
1036 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1037 results
[1] = ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubetc",
1038 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1039 LLVMValueRef ma
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubema",
1040 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1041 results
[0] = ac_build_fdiv(&ctx
->ac
, results
[0], ma
);
1042 results
[1] = ac_build_fdiv(&ctx
->ac
, results
[1], ma
);
1043 LLVMValueRef offset
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
1044 results
[0] = LLVMBuildFAdd(ctx
->ac
.builder
, results
[0], offset
, "");
1045 results
[1] = LLVMBuildFAdd(ctx
->ac
.builder
, results
[1], offset
, "");
1046 result
= ac_build_gather_values(&ctx
->ac
, results
, 2);
1050 case nir_op_cube_face_index
: {
1051 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1053 for (unsigned chan
= 0; chan
< 3; chan
++)
1054 in
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src
[0], chan
);
1055 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubeid",
1056 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1061 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1062 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1063 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
1064 ac_to_float_type(&ctx
->ac
, def_type
), result
, src
[2]);
1067 result
= ac_build_umin(&ctx
->ac
, src
[0], src
[1]);
1068 result
= ac_build_umin(&ctx
->ac
, result
, src
[2]);
1071 result
= ac_build_imin(&ctx
->ac
, src
[0], src
[1]);
1072 result
= ac_build_imin(&ctx
->ac
, result
, src
[2]);
1075 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1076 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
1077 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
1078 ac_to_float_type(&ctx
->ac
, def_type
), result
, src
[2]);
1081 result
= ac_build_umax(&ctx
->ac
, src
[0], src
[1]);
1082 result
= ac_build_umax(&ctx
->ac
, result
, src
[2]);
1085 result
= ac_build_imax(&ctx
->ac
, src
[0], src
[1]);
1086 result
= ac_build_imax(&ctx
->ac
, result
, src
[2]);
1088 case nir_op_fmed3
: {
1089 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1090 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
1091 src
[2] = ac_to_float(&ctx
->ac
, src
[2]);
1092 result
= ac_build_fmed3(&ctx
->ac
, src
[0], src
[1], src
[2],
1093 instr
->dest
.dest
.ssa
.bit_size
);
1096 case nir_op_imed3
: {
1097 LLVMValueRef tmp1
= ac_build_imin(&ctx
->ac
, src
[0], src
[1]);
1098 LLVMValueRef tmp2
= ac_build_imax(&ctx
->ac
, src
[0], src
[1]);
1099 tmp2
= ac_build_imin(&ctx
->ac
, tmp2
, src
[2]);
1100 result
= ac_build_imax(&ctx
->ac
, tmp1
, tmp2
);
1103 case nir_op_umed3
: {
1104 LLVMValueRef tmp1
= ac_build_umin(&ctx
->ac
, src
[0], src
[1]);
1105 LLVMValueRef tmp2
= ac_build_umax(&ctx
->ac
, src
[0], src
[1]);
1106 tmp2
= ac_build_umin(&ctx
->ac
, tmp2
, src
[2]);
1107 result
= ac_build_umax(&ctx
->ac
, tmp1
, tmp2
);
1112 fprintf(stderr
, "Unknown NIR alu instr: ");
1113 nir_print_instr(&instr
->instr
, stderr
);
1114 fprintf(stderr
, "\n");
1119 assert(instr
->dest
.dest
.is_ssa
);
1120 result
= ac_to_integer_or_pointer(&ctx
->ac
, result
);
1121 ctx
->ssa_defs
[instr
->dest
.dest
.ssa
.index
] = result
;
1125 static void visit_load_const(struct ac_nir_context
*ctx
,
1126 const nir_load_const_instr
*instr
)
1128 LLVMValueRef values
[4], value
= NULL
;
1129 LLVMTypeRef element_type
=
1130 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
1132 for (unsigned i
= 0; i
< instr
->def
.num_components
; ++i
) {
1133 switch (instr
->def
.bit_size
) {
1135 values
[i
] = LLVMConstInt(element_type
,
1136 instr
->value
[i
].u8
, false);
1139 values
[i
] = LLVMConstInt(element_type
,
1140 instr
->value
[i
].u16
, false);
1143 values
[i
] = LLVMConstInt(element_type
,
1144 instr
->value
[i
].u32
, false);
1147 values
[i
] = LLVMConstInt(element_type
,
1148 instr
->value
[i
].u64
, false);
1152 "unsupported nir load_const bit_size: %d\n",
1153 instr
->def
.bit_size
);
1157 if (instr
->def
.num_components
> 1) {
1158 value
= LLVMConstVector(values
, instr
->def
.num_components
);
1162 ctx
->ssa_defs
[instr
->def
.index
] = value
;
1166 get_buffer_size(struct ac_nir_context
*ctx
, LLVMValueRef descriptor
, bool in_elements
)
1169 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1170 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1173 if (ctx
->ac
.chip_class
== VI
&& in_elements
) {
1174 /* On VI, the descriptor contains the size in bytes,
1175 * but TXQ must return the size in elements.
1176 * The stride is always non-zero for resources using TXQ.
1178 LLVMValueRef stride
=
1179 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1181 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
,
1182 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
1183 stride
= LLVMBuildAnd(ctx
->ac
.builder
, stride
,
1184 LLVMConstInt(ctx
->ac
.i32
, 0x3fff, false), "");
1186 size
= LLVMBuildUDiv(ctx
->ac
.builder
, size
, stride
, "");
1191 static LLVMValueRef
lower_gather4_integer(struct ac_llvm_context
*ctx
,
1193 struct ac_image_args
*args
,
1194 const nir_tex_instr
*instr
)
1196 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1197 enum glsl_base_type stype
= glsl_get_sampler_result_type(type
);
1198 LLVMValueRef half_texel
[2];
1199 LLVMValueRef compare_cube_wa
= NULL
;
1200 LLVMValueRef result
;
1204 struct ac_image_args txq_args
= { 0 };
1206 txq_args
.dim
= get_ac_sampler_dim(ctx
, instr
->sampler_dim
, instr
->is_array
);
1207 txq_args
.opcode
= ac_image_get_resinfo
;
1208 txq_args
.dmask
= 0xf;
1209 txq_args
.lod
= ctx
->i32_0
;
1210 txq_args
.resource
= args
->resource
;
1211 txq_args
.attributes
= AC_FUNC_ATTR_READNONE
;
1212 LLVMValueRef size
= ac_build_image_opcode(ctx
, &txq_args
);
1214 for (unsigned c
= 0; c
< 2; c
++) {
1215 half_texel
[c
] = LLVMBuildExtractElement(ctx
->builder
, size
,
1216 LLVMConstInt(ctx
->i32
, c
, false), "");
1217 half_texel
[c
] = LLVMBuildUIToFP(ctx
->builder
, half_texel
[c
], ctx
->f32
, "");
1218 half_texel
[c
] = ac_build_fdiv(ctx
, ctx
->f32_1
, half_texel
[c
]);
1219 half_texel
[c
] = LLVMBuildFMul(ctx
->builder
, half_texel
[c
],
1220 LLVMConstReal(ctx
->f32
, -0.5), "");
1224 LLVMValueRef orig_coords
[2] = { args
->coords
[0], args
->coords
[1] };
1226 for (unsigned c
= 0; c
< 2; c
++) {
1228 tmp
= LLVMBuildBitCast(ctx
->builder
, args
->coords
[c
], ctx
->f32
, "");
1229 args
->coords
[c
] = LLVMBuildFAdd(ctx
->builder
, tmp
, half_texel
[c
], "");
1233 * Apparantly cube has issue with integer types that the workaround doesn't solve,
1234 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
1235 * workaround by sampling using a scaled type and converting.
1236 * This is taken from amdgpu-pro shaders.
1238 /* NOTE this produces some ugly code compared to amdgpu-pro,
1239 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
1240 * and then reads them back. -pro generates two selects,
1241 * one s_cmp for the descriptor rewriting
1242 * one v_cmp for the coordinate and result changes.
1244 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1245 LLVMValueRef tmp
, tmp2
;
1247 /* workaround 8/8/8/8 uint/sint cube gather bug */
1248 /* first detect it then change to a scaled read and f2i */
1249 tmp
= LLVMBuildExtractElement(ctx
->builder
, args
->resource
, ctx
->i32_1
, "");
1252 /* extract the DATA_FORMAT */
1253 tmp
= ac_build_bfe(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 20, false),
1254 LLVMConstInt(ctx
->i32
, 6, false), false);
1256 /* is the DATA_FORMAT == 8_8_8_8 */
1257 compare_cube_wa
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tmp
, LLVMConstInt(ctx
->i32
, V_008F14_IMG_DATA_FORMAT_8_8_8_8
, false), "");
1259 if (stype
== GLSL_TYPE_UINT
)
1260 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
1261 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0x8000000, false),
1262 LLVMConstInt(ctx
->i32
, 0x10000000, false), "");
1264 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
1265 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, LLVMConstInt(ctx
->i32
, 0xc000000, false),
1266 LLVMConstInt(ctx
->i32
, 0x14000000, false), "");
1268 /* replace the NUM FORMAT in the descriptor */
1269 tmp2
= LLVMBuildAnd(ctx
->builder
, tmp2
, LLVMConstInt(ctx
->i32
, C_008F14_NUM_FORMAT_GFX6
, false), "");
1270 tmp2
= LLVMBuildOr(ctx
->builder
, tmp2
, tmp
, "");
1272 args
->resource
= LLVMBuildInsertElement(ctx
->builder
, args
->resource
, tmp2
, ctx
->i32_1
, "");
1274 /* don't modify the coordinates for this case */
1275 for (unsigned c
= 0; c
< 2; ++c
)
1276 args
->coords
[c
] = LLVMBuildSelect(
1277 ctx
->builder
, compare_cube_wa
,
1278 orig_coords
[c
], args
->coords
[c
], "");
1281 args
->attributes
= AC_FUNC_ATTR_READNONE
;
1282 result
= ac_build_image_opcode(ctx
, args
);
1284 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1285 LLVMValueRef tmp
, tmp2
;
1287 /* if the cube workaround is in place, f2i the result. */
1288 for (unsigned c
= 0; c
< 4; c
++) {
1289 tmp
= LLVMBuildExtractElement(ctx
->builder
, result
, LLVMConstInt(ctx
->i32
, c
, false), "");
1290 if (stype
== GLSL_TYPE_UINT
)
1291 tmp2
= LLVMBuildFPToUI(ctx
->builder
, tmp
, ctx
->i32
, "");
1293 tmp2
= LLVMBuildFPToSI(ctx
->builder
, tmp
, ctx
->i32
, "");
1294 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
1295 tmp2
= LLVMBuildBitCast(ctx
->builder
, tmp2
, ctx
->i32
, "");
1296 tmp
= LLVMBuildSelect(ctx
->builder
, compare_cube_wa
, tmp2
, tmp
, "");
1297 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
1298 result
= LLVMBuildInsertElement(ctx
->builder
, result
, tmp
, LLVMConstInt(ctx
->i32
, c
, false), "");
1304 static nir_deref_instr
*get_tex_texture_deref(const nir_tex_instr
*instr
)
1306 nir_deref_instr
*texture_deref_instr
= NULL
;
1308 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1309 switch (instr
->src
[i
].src_type
) {
1310 case nir_tex_src_texture_deref
:
1311 texture_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
1317 return texture_deref_instr
;
1320 static LLVMValueRef
build_tex_intrinsic(struct ac_nir_context
*ctx
,
1321 const nir_tex_instr
*instr
,
1322 struct ac_image_args
*args
)
1324 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
1325 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
1327 if (ctx
->abi
->gfx9_stride_size_workaround
) {
1328 return ac_build_buffer_load_format_gfx9_safe(&ctx
->ac
,
1332 util_last_bit(mask
),
1335 return ac_build_buffer_load_format(&ctx
->ac
,
1339 util_last_bit(mask
),
1344 args
->opcode
= ac_image_sample
;
1346 switch (instr
->op
) {
1348 case nir_texop_txf_ms
:
1349 case nir_texop_samples_identical
:
1350 args
->opcode
= args
->level_zero
||
1351 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
?
1352 ac_image_load
: ac_image_load_mip
;
1353 args
->level_zero
= false;
1356 case nir_texop_query_levels
:
1357 args
->opcode
= ac_image_get_resinfo
;
1359 args
->lod
= ctx
->ac
.i32_0
;
1360 args
->level_zero
= false;
1363 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
) {
1365 args
->level_zero
= true;
1369 args
->opcode
= ac_image_gather4
;
1370 args
->level_zero
= true;
1373 args
->opcode
= ac_image_get_lod
;
1379 if (instr
->op
== nir_texop_tg4
&& ctx
->ac
.chip_class
<= VI
) {
1380 nir_deref_instr
*texture_deref_instr
= get_tex_texture_deref(instr
);
1381 nir_variable
*var
= nir_deref_instr_get_variable(texture_deref_instr
);
1382 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1383 enum glsl_base_type stype
= glsl_get_sampler_result_type(type
);
1384 if (stype
== GLSL_TYPE_UINT
|| stype
== GLSL_TYPE_INT
) {
1385 return lower_gather4_integer(&ctx
->ac
, var
, args
, instr
);
1389 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1390 if (instr
->op
== nir_texop_lod
&& ctx
->ac
.chip_class
>= GFX9
) {
1391 if ((args
->dim
== ac_image_2darray
||
1392 args
->dim
== ac_image_2d
) && !args
->coords
[1]) {
1393 args
->coords
[1] = ctx
->ac
.i32_0
;
1397 args
->attributes
= AC_FUNC_ATTR_READNONE
;
1398 return ac_build_image_opcode(&ctx
->ac
, args
);
1401 static LLVMValueRef
visit_vulkan_resource_reindex(struct ac_nir_context
*ctx
,
1402 nir_intrinsic_instr
*instr
)
1404 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
1405 LLVMValueRef index
= get_src(ctx
, instr
->src
[1]);
1407 LLVMValueRef result
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
1408 LLVMSetMetadata(result
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1412 static LLVMValueRef
visit_load_push_constant(struct ac_nir_context
*ctx
,
1413 nir_intrinsic_instr
*instr
)
1415 LLVMValueRef ptr
, addr
;
1416 LLVMValueRef src0
= get_src(ctx
, instr
->src
[0]);
1417 unsigned index
= nir_intrinsic_base(instr
);
1419 addr
= LLVMConstInt(ctx
->ac
.i32
, index
, 0);
1420 addr
= LLVMBuildAdd(ctx
->ac
.builder
, addr
, src0
, "");
1422 /* Load constant values from user SGPRS when possible, otherwise
1423 * fallback to the default path that loads directly from memory.
1425 if (LLVMIsConstant(src0
) &&
1426 instr
->dest
.ssa
.bit_size
== 32) {
1427 unsigned count
= instr
->dest
.ssa
.num_components
;
1428 unsigned offset
= index
;
1430 offset
+= LLVMConstIntGetZExtValue(src0
);
1433 offset
-= ctx
->abi
->base_inline_push_consts
;
1435 if (offset
+ count
<= ctx
->abi
->num_inline_push_consts
) {
1436 return ac_build_gather_values(&ctx
->ac
,
1437 ctx
->abi
->inline_push_consts
+ offset
,
1442 ptr
= ac_build_gep0(&ctx
->ac
, ctx
->abi
->push_constants
, addr
);
1444 if (instr
->dest
.ssa
.bit_size
== 8) {
1445 unsigned load_dwords
= instr
->dest
.ssa
.num_components
> 1 ? 2 : 1;
1446 LLVMTypeRef vec_type
= LLVMVectorType(LLVMInt8TypeInContext(ctx
->ac
.context
), 4 * load_dwords
);
1447 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, vec_type
);
1448 LLVMValueRef res
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1450 LLVMValueRef params
[3];
1451 if (load_dwords
> 1) {
1452 LLVMValueRef res_vec
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, LLVMVectorType(ctx
->ac
.i32
, 2), "");
1453 params
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, res_vec
, LLVMConstInt(ctx
->ac
.i32
, 1, false), "");
1454 params
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, res_vec
, LLVMConstInt(ctx
->ac
.i32
, 0, false), "");
1456 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, ctx
->ac
.i32
, "");
1457 params
[0] = ctx
->ac
.i32_0
;
1461 res
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.alignbyte", ctx
->ac
.i32
, params
, 3, 0);
1463 res
= LLVMBuildTrunc(ctx
->ac
.builder
, res
, LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.num_components
* 8), "");
1464 if (instr
->dest
.ssa
.num_components
> 1)
1465 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, LLVMVectorType(LLVMInt8TypeInContext(ctx
->ac
.context
), instr
->dest
.ssa
.num_components
), "");
1467 } else if (instr
->dest
.ssa
.bit_size
== 16) {
1468 unsigned load_dwords
= instr
->dest
.ssa
.num_components
/ 2 + 1;
1469 LLVMTypeRef vec_type
= LLVMVectorType(LLVMInt16TypeInContext(ctx
->ac
.context
), 2 * load_dwords
);
1470 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, vec_type
);
1471 LLVMValueRef res
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1472 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, vec_type
, "");
1473 LLVMValueRef cond
= LLVMBuildLShr(ctx
->ac
.builder
, addr
, ctx
->ac
.i32_1
, "");
1474 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
, ctx
->ac
.i1
, "");
1475 LLVMValueRef mask
[] = { LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
1476 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
1477 LLVMConstInt(ctx
->ac
.i32
, 4, false)};
1478 LLVMValueRef swizzle_aligned
= LLVMConstVector(&mask
[0], instr
->dest
.ssa
.num_components
);
1479 LLVMValueRef swizzle_unaligned
= LLVMConstVector(&mask
[1], instr
->dest
.ssa
.num_components
);
1480 LLVMValueRef shuffle_aligned
= LLVMBuildShuffleVector(ctx
->ac
.builder
, res
, res
, swizzle_aligned
, "");
1481 LLVMValueRef shuffle_unaligned
= LLVMBuildShuffleVector(ctx
->ac
.builder
, res
, res
, swizzle_unaligned
, "");
1482 res
= LLVMBuildSelect(ctx
->ac
.builder
, cond
, shuffle_unaligned
, shuffle_aligned
, "");
1483 return LLVMBuildBitCast(ctx
->ac
.builder
, res
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
1486 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, get_def_type(ctx
, &instr
->dest
.ssa
));
1488 return LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1491 static LLVMValueRef
visit_get_buffer_size(struct ac_nir_context
*ctx
,
1492 const nir_intrinsic_instr
*instr
)
1494 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
1496 return get_buffer_size(ctx
, ctx
->abi
->load_ssbo(ctx
->abi
, index
, false), false);
1499 static uint32_t widen_mask(uint32_t mask
, unsigned multiplier
)
1501 uint32_t new_mask
= 0;
1502 for(unsigned i
= 0; i
< 32 && (1u << i
) <= mask
; ++i
)
1503 if (mask
& (1u << i
))
1504 new_mask
|= ((1u << multiplier
) - 1u) << (i
* multiplier
);
1508 static LLVMValueRef
extract_vector_range(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
1509 unsigned start
, unsigned count
)
1511 LLVMValueRef mask
[] = {
1512 ctx
->i32_0
, ctx
->i32_1
,
1513 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false) };
1515 unsigned src_elements
= ac_get_llvm_num_components(src
);
1517 if (count
== src_elements
) {
1520 } else if (count
== 1) {
1521 assert(start
< src_elements
);
1522 return LLVMBuildExtractElement(ctx
->builder
, src
, mask
[start
], "");
1524 assert(start
+ count
<= src_elements
);
1526 LLVMValueRef swizzle
= LLVMConstVector(&mask
[start
], count
);
1527 return LLVMBuildShuffleVector(ctx
->builder
, src
, src
, swizzle
, "");
1531 static unsigned get_cache_policy(struct ac_nir_context
*ctx
,
1532 enum gl_access_qualifier access
,
1533 bool may_store_unaligned
,
1534 bool writeonly_memory
)
1536 unsigned cache_policy
= 0;
1538 /* SI has a TC L1 bug causing corruption of 8bit/16bit stores. All
1539 * store opcodes not aligned to a dword are affected. The only way to
1540 * get unaligned stores is through shader images.
1542 if (((may_store_unaligned
&& ctx
->ac
.chip_class
== SI
) ||
1543 /* If this is write-only, don't keep data in L1 to prevent
1544 * evicting L1 cache lines that may be needed by other
1548 access
& (ACCESS_COHERENT
| ACCESS_VOLATILE
))) {
1549 cache_policy
|= ac_glc
;
1552 return cache_policy
;
1555 static void visit_store_ssbo(struct ac_nir_context
*ctx
,
1556 nir_intrinsic_instr
*instr
)
1558 LLVMValueRef src_data
= get_src(ctx
, instr
->src
[0]);
1559 int elem_size_bytes
= ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src_data
)) / 8;
1560 unsigned writemask
= nir_intrinsic_write_mask(instr
);
1561 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
1562 bool writeonly_memory
= access
& ACCESS_NON_READABLE
;
1563 unsigned cache_policy
= get_cache_policy(ctx
, access
, false, writeonly_memory
);
1565 LLVMValueRef rsrc
= ctx
->abi
->load_ssbo(ctx
->abi
,
1566 get_src(ctx
, instr
->src
[1]), true);
1567 LLVMValueRef base_data
= src_data
;
1568 base_data
= ac_trim_vector(&ctx
->ac
, base_data
, instr
->num_components
);
1569 LLVMValueRef base_offset
= get_src(ctx
, instr
->src
[2]);
1573 LLVMValueRef data
, offset
;
1574 LLVMTypeRef data_type
;
1576 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
1578 /* Due to an LLVM limitation, split 3-element writes
1579 * into a 2-element and a 1-element write. */
1581 writemask
|= 1 << (start
+ 2);
1584 int num_bytes
= count
* elem_size_bytes
; /* count in bytes */
1586 /* we can only store 4 DWords at the same time.
1587 * can only happen for 64 Bit vectors. */
1588 if (num_bytes
> 16) {
1589 writemask
|= ((1u << (count
- 2)) - 1u) << (start
+ 2);
1594 /* check alignment of 16 Bit stores */
1595 if (elem_size_bytes
== 2 && num_bytes
> 2 && (start
% 2) == 1) {
1596 writemask
|= ((1u << (count
- 1)) - 1u) << (start
+ 1);
1600 data
= extract_vector_range(&ctx
->ac
, base_data
, start
, count
);
1602 offset
= LLVMBuildAdd(ctx
->ac
.builder
, base_offset
,
1603 LLVMConstInt(ctx
->ac
.i32
, start
* elem_size_bytes
, false), "");
1605 if (num_bytes
== 1) {
1606 ac_build_tbuffer_store_byte(&ctx
->ac
, rsrc
, data
,
1607 offset
, ctx
->ac
.i32_0
,
1608 cache_policy
& ac_glc
,
1610 } else if (num_bytes
== 2) {
1611 ac_build_tbuffer_store_short(&ctx
->ac
, rsrc
, data
,
1612 offset
, ctx
->ac
.i32_0
,
1613 cache_policy
& ac_glc
,
1616 int num_channels
= num_bytes
/ 4;
1618 switch (num_bytes
) {
1619 case 16: /* v4f32 */
1620 data_type
= ctx
->ac
.v4f32
;
1623 data_type
= ctx
->ac
.v2f32
;
1626 data_type
= ctx
->ac
.f32
;
1629 unreachable("Malformed vector store.");
1631 data
= LLVMBuildBitCast(ctx
->ac
.builder
, data
, data_type
, "");
1633 ac_build_buffer_store_dword(&ctx
->ac
, rsrc
, data
,
1634 num_channels
, offset
,
1636 cache_policy
& ac_glc
,
1637 false, writeonly_memory
,
1643 static LLVMValueRef
visit_atomic_ssbo(struct ac_nir_context
*ctx
,
1644 const nir_intrinsic_instr
*instr
)
1646 LLVMTypeRef return_type
= LLVMTypeOf(get_src(ctx
, instr
->src
[2]));
1648 char name
[64], type
[8];
1649 LLVMValueRef params
[6];
1652 switch (instr
->intrinsic
) {
1653 case nir_intrinsic_ssbo_atomic_add
:
1656 case nir_intrinsic_ssbo_atomic_imin
:
1659 case nir_intrinsic_ssbo_atomic_umin
:
1662 case nir_intrinsic_ssbo_atomic_imax
:
1665 case nir_intrinsic_ssbo_atomic_umax
:
1668 case nir_intrinsic_ssbo_atomic_and
:
1671 case nir_intrinsic_ssbo_atomic_or
:
1674 case nir_intrinsic_ssbo_atomic_xor
:
1677 case nir_intrinsic_ssbo_atomic_exchange
:
1680 case nir_intrinsic_ssbo_atomic_comp_swap
:
1687 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
) {
1688 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[3]), 0);
1690 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
1691 params
[arg_count
++] = ctx
->abi
->load_ssbo(ctx
->abi
,
1692 get_src(ctx
, instr
->src
[0]),
1695 if (HAVE_LLVM
>= 0x900) {
1696 /* XXX: The new raw/struct atomic intrinsics are buggy with
1697 * LLVM 8, see r358579.
1699 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1700 params
[arg_count
++] = ctx
->ac
.i32_0
; /* soffset */
1701 params
[arg_count
++] = ctx
->ac
.i32_0
; /* slc */
1703 ac_build_type_name_for_intr(return_type
, type
, sizeof(type
));
1704 snprintf(name
, sizeof(name
),
1705 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op
, type
);
1707 params
[arg_count
++] = ctx
->ac
.i32_0
; /* vindex */
1708 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1709 params
[arg_count
++] = ctx
->ac
.i1false
; /* slc */
1711 assert(return_type
== ctx
->ac
.i32
);
1712 snprintf(name
, sizeof(name
),
1713 "llvm.amdgcn.buffer.atomic.%s", op
);
1716 return ac_build_intrinsic(&ctx
->ac
, name
, return_type
, params
,
1720 static LLVMValueRef
visit_load_buffer(struct ac_nir_context
*ctx
,
1721 const nir_intrinsic_instr
*instr
)
1723 int elem_size_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
1724 int num_components
= instr
->num_components
;
1725 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
1726 unsigned cache_policy
= get_cache_policy(ctx
, access
, false, false);
1728 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
1729 LLVMValueRef rsrc
= ctx
->abi
->load_ssbo(ctx
->abi
,
1730 get_src(ctx
, instr
->src
[0]), false);
1731 LLVMValueRef vindex
= ctx
->ac
.i32_0
;
1733 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
1734 LLVMTypeRef def_elem_type
= num_components
> 1 ? LLVMGetElementType(def_type
) : def_type
;
1736 LLVMValueRef results
[4];
1737 for (int i
= 0; i
< num_components
;) {
1738 int num_elems
= num_components
- i
;
1739 if (elem_size_bytes
< 4 && nir_intrinsic_align(instr
) % 4 != 0)
1741 if (num_elems
* elem_size_bytes
> 16)
1742 num_elems
= 16 / elem_size_bytes
;
1743 int load_bytes
= num_elems
* elem_size_bytes
;
1745 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
, i
* elem_size_bytes
, false);
1749 if (load_bytes
== 1) {
1750 ret
= ac_build_tbuffer_load_byte(&ctx
->ac
,
1755 cache_policy
& ac_glc
);
1756 } else if (load_bytes
== 2) {
1757 ret
= ac_build_tbuffer_load_short(&ctx
->ac
,
1762 cache_policy
& ac_glc
);
1764 int num_channels
= util_next_power_of_two(load_bytes
) / 4;
1766 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_channels
,
1767 vindex
, offset
, immoffset
, 0,
1768 cache_policy
& ac_glc
, 0,
1772 LLVMTypeRef byte_vec
= LLVMVectorType(ctx
->ac
.i8
, ac_get_type_size(LLVMTypeOf(ret
)));
1773 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, byte_vec
, "");
1774 ret
= ac_trim_vector(&ctx
->ac
, ret
, load_bytes
);
1776 LLVMTypeRef ret_type
= LLVMVectorType(def_elem_type
, num_elems
);
1777 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ret_type
, "");
1779 for (unsigned j
= 0; j
< num_elems
; j
++) {
1780 results
[i
+ j
] = LLVMBuildExtractElement(ctx
->ac
.builder
, ret
, LLVMConstInt(ctx
->ac
.i32
, j
, false), "");
1785 return ac_build_gather_values(&ctx
->ac
, results
, num_components
);
1788 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
1789 const nir_intrinsic_instr
*instr
)
1792 LLVMValueRef rsrc
= get_src(ctx
, instr
->src
[0]);
1793 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
1794 int num_components
= instr
->num_components
;
1796 if (ctx
->abi
->load_ubo
)
1797 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
1799 if (instr
->dest
.ssa
.bit_size
== 64)
1800 num_components
*= 2;
1802 if (instr
->dest
.ssa
.bit_size
== 16 || instr
->dest
.ssa
.bit_size
== 8) {
1803 unsigned load_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
1804 LLVMValueRef results
[num_components
];
1805 for (unsigned i
= 0; i
< num_components
; ++i
) {
1806 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
,
1809 if (load_bytes
== 1) {
1810 results
[i
] = ac_build_tbuffer_load_byte(&ctx
->ac
,
1817 assert(load_bytes
== 2);
1818 results
[i
] = ac_build_tbuffer_load_short(&ctx
->ac
,
1826 ret
= ac_build_gather_values(&ctx
->ac
, results
, num_components
);
1828 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_components
, NULL
, offset
,
1829 NULL
, 0, false, false, true, true);
1831 ret
= ac_trim_vector(&ctx
->ac
, ret
, num_components
);
1834 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
1835 get_def_type(ctx
, &instr
->dest
.ssa
), "");
1839 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_instr
*instr
,
1840 bool vs_in
, unsigned *vertex_index_out
,
1841 LLVMValueRef
*vertex_index_ref
,
1842 unsigned *const_out
, LLVMValueRef
*indir_out
)
1844 nir_variable
*var
= nir_deref_instr_get_variable(instr
);
1845 nir_deref_path path
;
1846 unsigned idx_lvl
= 1;
1848 nir_deref_path_init(&path
, instr
, NULL
);
1850 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
1851 if (vertex_index_ref
) {
1852 *vertex_index_ref
= get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
);
1853 if (vertex_index_out
)
1854 *vertex_index_out
= 0;
1856 *vertex_index_out
= nir_src_as_uint(path
.path
[idx_lvl
]->arr
.index
);
1861 uint32_t const_offset
= 0;
1862 LLVMValueRef offset
= NULL
;
1864 if (var
->data
.compact
) {
1865 assert(instr
->deref_type
== nir_deref_type_array
);
1866 const_offset
= nir_src_as_uint(instr
->arr
.index
);
1870 for (; path
.path
[idx_lvl
]; ++idx_lvl
) {
1871 const struct glsl_type
*parent_type
= path
.path
[idx_lvl
- 1]->type
;
1872 if (path
.path
[idx_lvl
]->deref_type
== nir_deref_type_struct
) {
1873 unsigned index
= path
.path
[idx_lvl
]->strct
.index
;
1875 for (unsigned i
= 0; i
< index
; i
++) {
1876 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
1877 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
1879 } else if(path
.path
[idx_lvl
]->deref_type
== nir_deref_type_array
) {
1880 unsigned size
= glsl_count_attribute_slots(path
.path
[idx_lvl
]->type
, vs_in
);
1881 LLVMValueRef array_off
= LLVMBuildMul(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, size
, 0),
1882 get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
), "");
1884 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, array_off
, "");
1888 unreachable("Uhandled deref type in get_deref_instr_offset");
1892 nir_deref_path_finish(&path
);
1894 if (const_offset
&& offset
)
1895 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
1896 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
1899 *const_out
= const_offset
;
1900 *indir_out
= offset
;
1903 static LLVMValueRef
load_tess_varyings(struct ac_nir_context
*ctx
,
1904 nir_intrinsic_instr
*instr
,
1907 LLVMValueRef result
;
1908 LLVMValueRef vertex_index
= NULL
;
1909 LLVMValueRef indir_index
= NULL
;
1910 unsigned const_index
= 0;
1912 nir_variable
*var
= nir_deref_instr_get_variable(nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
));
1914 unsigned location
= var
->data
.location
;
1915 unsigned driver_location
= var
->data
.driver_location
;
1916 const bool is_patch
= var
->data
.patch
;
1917 const bool is_compact
= var
->data
.compact
;
1919 get_deref_offset(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
1920 false, NULL
, is_patch
? NULL
: &vertex_index
,
1921 &const_index
, &indir_index
);
1923 LLVMTypeRef dest_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
1925 LLVMTypeRef src_component_type
;
1926 if (LLVMGetTypeKind(dest_type
) == LLVMVectorTypeKind
)
1927 src_component_type
= LLVMGetElementType(dest_type
);
1929 src_component_type
= dest_type
;
1931 result
= ctx
->abi
->load_tess_varyings(ctx
->abi
, src_component_type
,
1932 vertex_index
, indir_index
,
1933 const_index
, location
, driver_location
,
1934 var
->data
.location_frac
,
1935 instr
->num_components
,
1936 is_patch
, is_compact
, load_inputs
);
1937 if (instr
->dest
.ssa
.bit_size
== 16) {
1938 result
= ac_to_integer(&ctx
->ac
, result
);
1939 result
= LLVMBuildTrunc(ctx
->ac
.builder
, result
, dest_type
, "");
1941 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, dest_type
, "");
1945 type_scalar_size_bytes(const struct glsl_type
*type
)
1947 assert(glsl_type_is_vector_or_scalar(type
) ||
1948 glsl_type_is_matrix(type
));
1949 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
1952 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
1953 nir_intrinsic_instr
*instr
)
1955 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
1956 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
1958 LLVMValueRef values
[8];
1960 int ve
= instr
->dest
.ssa
.num_components
;
1962 LLVMValueRef indir_index
;
1964 unsigned const_index
;
1965 unsigned stride
= 4;
1966 int mode
= deref
->mode
;
1969 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
1970 var
->data
.mode
== nir_var_shader_in
;
1971 idx
= var
->data
.driver_location
;
1972 comp
= var
->data
.location_frac
;
1973 mode
= var
->data
.mode
;
1975 get_deref_offset(ctx
, deref
, vs_in
, NULL
, NULL
,
1976 &const_index
, &indir_index
);
1978 if (var
->data
.compact
) {
1980 const_index
+= comp
;
1985 if (instr
->dest
.ssa
.bit_size
== 64 &&
1986 (deref
->mode
== nir_var_shader_in
||
1987 deref
->mode
== nir_var_shader_out
||
1988 deref
->mode
== nir_var_function_temp
))
1992 case nir_var_shader_in
:
1993 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
1994 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
1995 return load_tess_varyings(ctx
, instr
, true);
1998 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
1999 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
2000 LLVMValueRef indir_index
;
2001 unsigned const_index
, vertex_index
;
2002 get_deref_offset(ctx
, deref
, false, &vertex_index
, NULL
,
2003 &const_index
, &indir_index
);
2005 return ctx
->abi
->load_inputs(ctx
->abi
, var
->data
.location
,
2006 var
->data
.driver_location
,
2007 var
->data
.location_frac
,
2008 instr
->num_components
, vertex_index
, const_index
, type
);
2011 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2013 unsigned count
= glsl_count_attribute_slots(
2015 ctx
->stage
== MESA_SHADER_VERTEX
);
2017 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2018 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
2019 stride
, false, true);
2021 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2025 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* stride
];
2028 case nir_var_function_temp
:
2029 for (unsigned chan
= 0; chan
< ve
; chan
++) {
2031 unsigned count
= glsl_count_attribute_slots(
2034 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2035 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2036 stride
, true, true);
2038 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2042 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* stride
], "");
2046 case nir_var_mem_shared
: {
2047 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2048 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
2049 return LLVMBuildBitCast(ctx
->ac
.builder
, val
,
2050 get_def_type(ctx
, &instr
->dest
.ssa
),
2053 case nir_var_shader_out
:
2054 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2055 return load_tess_varyings(ctx
, instr
, false);
2058 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2060 unsigned count
= glsl_count_attribute_slots(
2063 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2064 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2065 stride
, true, true);
2067 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2071 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
2072 ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
],
2077 case nir_var_mem_global
: {
2078 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2079 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2080 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2081 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2083 LLVMTypeRef result_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
2084 if (stride
!= natural_stride
) {
2085 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMGetElementType(result_type
),
2086 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2087 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2089 for (unsigned i
= 0; i
< instr
->dest
.ssa
.num_components
; ++i
) {
2090 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, i
* stride
/ natural_stride
, 0);
2091 values
[i
] = LLVMBuildLoad(ctx
->ac
.builder
,
2092 ac_build_gep_ptr(&ctx
->ac
, address
, offset
), "");
2094 return ac_build_gather_values(&ctx
->ac
, values
, instr
->dest
.ssa
.num_components
);
2096 LLVMTypeRef ptr_type
= LLVMPointerType(result_type
,
2097 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2098 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2099 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
2104 unreachable("unhandle variable mode");
2106 ret
= ac_build_varying_gather_values(&ctx
->ac
, values
, ve
, comp
);
2107 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
2111 visit_store_var(struct ac_nir_context
*ctx
,
2112 nir_intrinsic_instr
*instr
)
2114 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2115 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2117 LLVMValueRef temp_ptr
, value
;
2120 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[1]));
2121 int writemask
= instr
->const_index
[0];
2122 LLVMValueRef indir_index
;
2123 unsigned const_index
;
2126 get_deref_offset(ctx
, deref
, false,
2127 NULL
, NULL
, &const_index
, &indir_index
);
2128 idx
= var
->data
.driver_location
;
2129 comp
= var
->data
.location_frac
;
2131 if (var
->data
.compact
) {
2132 const_index
+= comp
;
2137 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64 &&
2138 (deref
->mode
== nir_var_shader_out
||
2139 deref
->mode
== nir_var_function_temp
)) {
2141 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2142 LLVMVectorType(ctx
->ac
.f32
, ac_get_llvm_num_components(src
) * 2),
2145 writemask
= widen_mask(writemask
, 2);
2148 writemask
= writemask
<< comp
;
2150 switch (deref
->mode
) {
2151 case nir_var_shader_out
:
2153 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2154 LLVMValueRef vertex_index
= NULL
;
2155 LLVMValueRef indir_index
= NULL
;
2156 unsigned const_index
= 0;
2157 const bool is_patch
= var
->data
.patch
;
2159 get_deref_offset(ctx
, deref
, false, NULL
,
2160 is_patch
? NULL
: &vertex_index
,
2161 &const_index
, &indir_index
);
2163 ctx
->abi
->store_tcs_outputs(ctx
->abi
, var
,
2164 vertex_index
, indir_index
,
2165 const_index
, src
, writemask
);
2169 for (unsigned chan
= 0; chan
< 8; chan
++) {
2171 if (!(writemask
& (1 << chan
)))
2174 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- comp
);
2176 if (var
->data
.compact
)
2179 unsigned count
= glsl_count_attribute_slots(
2182 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2183 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2184 stride
, true, true);
2186 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2187 value
, indir_index
, "");
2188 build_store_values_extended(&ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
,
2189 count
, stride
, tmp_vec
);
2192 temp_ptr
= ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
];
2194 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2198 case nir_var_function_temp
:
2199 for (unsigned chan
= 0; chan
< 8; chan
++) {
2200 if (!(writemask
& (1 << chan
)))
2203 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
2205 unsigned count
= glsl_count_attribute_slots(
2208 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2209 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2212 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2213 value
, indir_index
, "");
2214 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
2217 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
2219 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2224 case nir_var_mem_global
:
2225 case nir_var_mem_shared
: {
2226 int writemask
= instr
->const_index
[0];
2227 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2228 LLVMValueRef val
= get_src(ctx
, instr
->src
[1]);
2230 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2231 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2232 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2234 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2235 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2236 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2238 if (writemask
== (1u << ac_get_llvm_num_components(val
)) - 1 &&
2239 stride
== natural_stride
) {
2240 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2241 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2242 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2244 val
= LLVMBuildBitCast(ctx
->ac
.builder
, val
,
2245 LLVMGetElementType(LLVMTypeOf(address
)), "");
2246 LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
2248 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMGetElementType(LLVMTypeOf(val
)),
2249 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2250 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2251 for (unsigned chan
= 0; chan
< 4; chan
++) {
2252 if (!(writemask
& (1 << chan
)))
2255 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, chan
* stride
/ natural_stride
, 0);
2257 LLVMValueRef ptr
= ac_build_gep_ptr(&ctx
->ac
, address
, offset
);
2258 LLVMValueRef src
= ac_llvm_extract_elem(&ctx
->ac
, val
,
2260 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2261 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
2262 LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
2273 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
2276 case GLSL_SAMPLER_DIM_BUF
:
2278 case GLSL_SAMPLER_DIM_1D
:
2279 return array
? 2 : 1;
2280 case GLSL_SAMPLER_DIM_2D
:
2281 return array
? 3 : 2;
2282 case GLSL_SAMPLER_DIM_MS
:
2283 return array
? 4 : 3;
2284 case GLSL_SAMPLER_DIM_3D
:
2285 case GLSL_SAMPLER_DIM_CUBE
:
2287 case GLSL_SAMPLER_DIM_RECT
:
2288 case GLSL_SAMPLER_DIM_SUBPASS
:
2290 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
2298 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
2299 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
2300 LLVMValueRef coord_z
,
2301 LLVMValueRef sample_index
,
2302 LLVMValueRef fmask_desc_ptr
)
2304 unsigned sample_chan
= coord_z
? 3 : 2;
2305 LLVMValueRef addr
[4] = {coord_x
, coord_y
, coord_z
};
2306 addr
[sample_chan
] = sample_index
;
2308 ac_apply_fmask_to_sample(ctx
, fmask_desc_ptr
, addr
, coord_z
!= NULL
);
2309 return addr
[sample_chan
];
2312 static nir_deref_instr
*get_image_deref(const nir_intrinsic_instr
*instr
)
2314 assert(instr
->src
[0].is_ssa
);
2315 return nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2318 static LLVMValueRef
get_image_descriptor(struct ac_nir_context
*ctx
,
2319 const nir_intrinsic_instr
*instr
,
2320 enum ac_descriptor_type desc_type
,
2323 nir_deref_instr
*deref_instr
=
2324 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_deref
?
2325 nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
) : NULL
;
2327 return get_sampler_desc(ctx
, deref_instr
, desc_type
, &instr
->instr
, true, write
);
2330 static void get_image_coords(struct ac_nir_context
*ctx
,
2331 const nir_intrinsic_instr
*instr
,
2332 struct ac_image_args
*args
,
2333 enum glsl_sampler_dim dim
,
2336 LLVMValueRef src0
= get_src(ctx
, instr
->src
[1]);
2337 LLVMValueRef masks
[] = {
2338 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
2339 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
2341 LLVMValueRef sample_index
= ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
2344 bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
2345 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2346 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
2347 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2348 bool gfx9_1d
= ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
;
2349 count
= image_type_to_components_count(dim
, is_array
);
2351 if (is_ms
&& (instr
->intrinsic
== nir_intrinsic_image_deref_load
||
2352 instr
->intrinsic
== nir_intrinsic_bindless_image_load
)) {
2353 LLVMValueRef fmask_load_address
[3];
2356 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2357 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
2359 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
2361 fmask_load_address
[2] = NULL
;
2363 for (chan
= 0; chan
< 2; ++chan
)
2364 fmask_load_address
[chan
] =
2365 LLVMBuildAdd(ctx
->ac
.builder
, fmask_load_address
[chan
],
2366 LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
2367 ctx
->ac
.i32
, ""), "");
2368 fmask_load_address
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
2370 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
2371 fmask_load_address
[0],
2372 fmask_load_address
[1],
2373 fmask_load_address
[2],
2375 get_sampler_desc(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
2376 AC_DESC_FMASK
, &instr
->instr
, false, false));
2378 if (count
== 1 && !gfx9_1d
) {
2379 if (instr
->src
[1].ssa
->num_components
)
2380 args
->coords
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2382 args
->coords
[0] = src0
;
2387 for (chan
= 0; chan
< count
; ++chan
) {
2388 args
->coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src0
, chan
);
2391 for (chan
= 0; chan
< 2; ++chan
) {
2392 args
->coords
[chan
] = LLVMBuildAdd(
2393 ctx
->ac
.builder
, args
->coords
[chan
],
2395 ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
2396 ctx
->ac
.i32
, ""), "");
2398 args
->coords
[2] = ac_to_integer(&ctx
->ac
,
2399 ctx
->abi
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
2405 args
->coords
[2] = args
->coords
[1];
2406 args
->coords
[1] = ctx
->ac
.i32_0
;
2408 args
->coords
[1] = ctx
->ac
.i32_0
;
2413 args
->coords
[count
] = sample_index
;
2419 static LLVMValueRef
get_image_buffer_descriptor(struct ac_nir_context
*ctx
,
2420 const nir_intrinsic_instr
*instr
, bool write
)
2422 LLVMValueRef rsrc
= get_image_descriptor(ctx
, instr
, AC_DESC_BUFFER
, write
);
2423 if (ctx
->abi
->gfx9_stride_size_workaround
) {
2424 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2425 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 1, 0), "");
2426 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
, LLVMConstInt(ctx
->ac
.i32
, 16, 0), "");
2428 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->ac
.builder
,
2429 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntUGT
, elem_count
, stride
, ""),
2430 elem_count
, stride
, "");
2432 rsrc
= LLVMBuildInsertElement(ctx
->ac
.builder
, rsrc
, new_elem_count
,
2433 LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2438 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
2439 const nir_intrinsic_instr
*instr
,
2444 enum glsl_sampler_dim dim
;
2445 enum gl_access_qualifier access
;
2448 dim
= nir_intrinsic_image_dim(instr
);
2449 access
= nir_intrinsic_access(instr
);
2450 is_array
= nir_intrinsic_image_array(instr
);
2452 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2453 const struct glsl_type
*type
= image_deref
->type
;
2454 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2455 dim
= glsl_get_sampler_dim(type
);
2456 access
= var
->data
.image
.access
;
2457 is_array
= glsl_sampler_type_is_array(type
);
2460 struct ac_image_args args
= {};
2462 args
.cache_policy
= get_cache_policy(ctx
, access
, false, false);
2464 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2465 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
2466 unsigned num_channels
= util_last_bit(mask
);
2467 LLVMValueRef rsrc
, vindex
;
2469 rsrc
= get_image_buffer_descriptor(ctx
, instr
, false);
2470 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
2473 /* TODO: set "can_speculate" when OpenGL needs it. */
2474 res
= ac_build_buffer_load_format(&ctx
->ac
, rsrc
, vindex
,
2475 ctx
->ac
.i32_0
, num_channels
,
2476 !!(args
.cache_policy
& ac_glc
),
2478 res
= ac_build_expand_to_vec4(&ctx
->ac
, res
, num_channels
);
2480 res
= ac_trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
2481 res
= ac_to_integer(&ctx
->ac
, res
);
2483 args
.opcode
= ac_image_load
;
2484 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2485 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2486 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2488 args
.attributes
= AC_FUNC_ATTR_READONLY
;
2490 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
2495 static void visit_image_store(struct ac_nir_context
*ctx
,
2496 nir_intrinsic_instr
*instr
,
2501 enum glsl_sampler_dim dim
;
2502 enum gl_access_qualifier access
;
2505 dim
= nir_intrinsic_image_dim(instr
);
2506 access
= nir_intrinsic_access(instr
);
2507 is_array
= nir_intrinsic_image_array(instr
);
2509 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2510 const struct glsl_type
*type
= image_deref
->type
;
2511 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2512 dim
= glsl_get_sampler_dim(type
);
2513 access
= var
->data
.image
.access
;
2514 is_array
= glsl_sampler_type_is_array(type
);
2517 bool writeonly_memory
= access
& ACCESS_NON_READABLE
;
2518 struct ac_image_args args
= {};
2520 args
.cache_policy
= get_cache_policy(ctx
, access
, true, writeonly_memory
);
2522 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2523 LLVMValueRef rsrc
= get_image_buffer_descriptor(ctx
, instr
, true);
2524 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2525 unsigned src_channels
= ac_get_llvm_num_components(src
);
2526 LLVMValueRef vindex
;
2528 if (src_channels
== 3)
2529 src
= ac_build_expand_to_vec4(&ctx
->ac
, src
, 3);
2531 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
,
2532 get_src(ctx
, instr
->src
[1]),
2535 ac_build_buffer_store_format(&ctx
->ac
, rsrc
, src
, vindex
,
2536 ctx
->ac
.i32_0
, src_channels
,
2537 args
.cache_policy
& ac_glc
,
2540 args
.opcode
= ac_image_store
;
2541 args
.data
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2542 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2543 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, true);
2544 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2547 ac_build_image_opcode(&ctx
->ac
, &args
);
2552 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
2553 const nir_intrinsic_instr
*instr
,
2556 LLVMValueRef params
[7];
2557 int param_count
= 0;
2559 bool cmpswap
= instr
->intrinsic
== nir_intrinsic_image_deref_atomic_comp_swap
||
2560 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_comp_swap
;
2561 const char *atomic_name
;
2562 char intrinsic_name
[64];
2563 enum ac_atomic_op atomic_subop
;
2564 MAYBE_UNUSED
int length
;
2566 enum glsl_sampler_dim dim
;
2567 bool is_unsigned
= false;
2570 if (instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_min
||
2571 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_max
) {
2572 const GLenum format
= nir_intrinsic_format(instr
);
2573 assert(format
== GL_R32UI
|| format
== GL_R32I
);
2574 is_unsigned
= format
== GL_R32UI
;
2576 dim
= nir_intrinsic_image_dim(instr
);
2577 is_array
= nir_intrinsic_image_array(instr
);
2579 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2580 is_unsigned
= glsl_get_sampler_result_type(type
) == GLSL_TYPE_UINT
;
2581 dim
= glsl_get_sampler_dim(type
);
2582 is_array
= glsl_sampler_type_is_array(type
);
2585 switch (instr
->intrinsic
) {
2586 case nir_intrinsic_bindless_image_atomic_add
:
2587 case nir_intrinsic_image_deref_atomic_add
:
2588 atomic_name
= "add";
2589 atomic_subop
= ac_atomic_add
;
2591 case nir_intrinsic_bindless_image_atomic_min
:
2592 case nir_intrinsic_image_deref_atomic_min
:
2593 atomic_name
= is_unsigned
? "umin" : "smin";
2594 atomic_subop
= is_unsigned
? ac_atomic_umin
: ac_atomic_smin
;
2596 case nir_intrinsic_bindless_image_atomic_max
:
2597 case nir_intrinsic_image_deref_atomic_max
:
2598 atomic_name
= is_unsigned
? "umax" : "smax";
2599 atomic_subop
= is_unsigned
? ac_atomic_umax
: ac_atomic_smax
;
2601 case nir_intrinsic_bindless_image_atomic_and
:
2602 case nir_intrinsic_image_deref_atomic_and
:
2603 atomic_name
= "and";
2604 atomic_subop
= ac_atomic_and
;
2606 case nir_intrinsic_bindless_image_atomic_or
:
2607 case nir_intrinsic_image_deref_atomic_or
:
2609 atomic_subop
= ac_atomic_or
;
2611 case nir_intrinsic_bindless_image_atomic_xor
:
2612 case nir_intrinsic_image_deref_atomic_xor
:
2613 atomic_name
= "xor";
2614 atomic_subop
= ac_atomic_xor
;
2616 case nir_intrinsic_bindless_image_atomic_exchange
:
2617 case nir_intrinsic_image_deref_atomic_exchange
:
2618 atomic_name
= "swap";
2619 atomic_subop
= ac_atomic_swap
;
2621 case nir_intrinsic_bindless_image_atomic_comp_swap
:
2622 case nir_intrinsic_image_deref_atomic_comp_swap
:
2623 atomic_name
= "cmpswap";
2624 atomic_subop
= 0; /* not used */
2631 params
[param_count
++] = get_src(ctx
, instr
->src
[4]);
2632 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
2634 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2635 params
[param_count
++] = get_image_buffer_descriptor(ctx
, instr
, true);
2636 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
2637 ctx
->ac
.i32_0
, ""); /* vindex */
2638 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
2639 if (HAVE_LLVM
>= 0x900) {
2640 /* XXX: The new raw/struct atomic intrinsics are buggy
2641 * with LLVM 8, see r358579.
2643 params
[param_count
++] = ctx
->ac
.i32_0
; /* soffset */
2644 params
[param_count
++] = ctx
->ac
.i32_0
; /* slc */
2646 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
2647 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name
);
2649 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
2651 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
2652 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
2655 assert(length
< sizeof(intrinsic_name
));
2656 return ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
,
2657 params
, param_count
, 0);
2659 struct ac_image_args args
= {};
2660 args
.opcode
= cmpswap
? ac_image_atomic_cmpswap
: ac_image_atomic
;
2661 args
.atomic
= atomic_subop
;
2662 args
.data
[0] = params
[0];
2664 args
.data
[1] = params
[1];
2665 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2666 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, true);
2667 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2669 return ac_build_image_opcode(&ctx
->ac
, &args
);
2673 static LLVMValueRef
visit_image_samples(struct ac_nir_context
*ctx
,
2674 const nir_intrinsic_instr
*instr
,
2677 enum glsl_sampler_dim dim
;
2680 dim
= nir_intrinsic_image_dim(instr
);
2681 is_array
= nir_intrinsic_image_array(instr
);
2683 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2684 dim
= glsl_get_sampler_dim(type
);
2685 is_array
= glsl_sampler_type_is_array(type
);
2688 struct ac_image_args args
= { 0 };
2689 args
.dim
= get_ac_sampler_dim(&ctx
->ac
, dim
, is_array
);
2691 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2692 args
.opcode
= ac_image_get_resinfo
;
2693 args
.lod
= ctx
->ac
.i32_0
;
2694 args
.attributes
= AC_FUNC_ATTR_READNONE
;
2696 return ac_build_image_opcode(&ctx
->ac
, &args
);
2699 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
2700 const nir_intrinsic_instr
*instr
,
2705 enum glsl_sampler_dim dim
;
2708 dim
= nir_intrinsic_image_dim(instr
);
2709 is_array
= nir_intrinsic_image_array(instr
);
2711 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2712 dim
= glsl_get_sampler_dim(type
);
2713 is_array
= glsl_sampler_type_is_array(type
);
2716 if (dim
== GLSL_SAMPLER_DIM_BUF
)
2717 return get_buffer_size(ctx
, get_image_descriptor(ctx
, instr
, AC_DESC_BUFFER
, false), true);
2719 struct ac_image_args args
= { 0 };
2721 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2723 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2724 args
.opcode
= ac_image_get_resinfo
;
2725 args
.lod
= ctx
->ac
.i32_0
;
2726 args
.attributes
= AC_FUNC_ATTR_READNONE
;
2728 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
2730 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
2732 if (dim
== GLSL_SAMPLER_DIM_CUBE
&& is_array
) {
2733 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
2734 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
2735 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
2736 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
2738 if (ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
&& is_array
) {
2739 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
2740 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, layers
,
2747 static void emit_membar(struct ac_llvm_context
*ac
,
2748 const nir_intrinsic_instr
*instr
)
2750 unsigned waitcnt
= NOOP_WAITCNT
;
2752 switch (instr
->intrinsic
) {
2753 case nir_intrinsic_memory_barrier
:
2754 case nir_intrinsic_group_memory_barrier
:
2755 waitcnt
&= VM_CNT
& LGKM_CNT
;
2757 case nir_intrinsic_memory_barrier_atomic_counter
:
2758 case nir_intrinsic_memory_barrier_buffer
:
2759 case nir_intrinsic_memory_barrier_image
:
2762 case nir_intrinsic_memory_barrier_shared
:
2763 waitcnt
&= LGKM_CNT
;
2768 if (waitcnt
!= NOOP_WAITCNT
)
2769 ac_build_waitcnt(ac
, waitcnt
);
2772 void ac_emit_barrier(struct ac_llvm_context
*ac
, gl_shader_stage stage
)
2774 /* SI only (thanks to a hw bug workaround):
2775 * The real barrier instruction isn’t needed, because an entire patch
2776 * always fits into a single wave.
2778 if (ac
->chip_class
== SI
&& stage
== MESA_SHADER_TESS_CTRL
) {
2779 ac_build_waitcnt(ac
, LGKM_CNT
& VM_CNT
);
2782 ac_build_s_barrier(ac
);
2785 static void emit_discard(struct ac_nir_context
*ctx
,
2786 const nir_intrinsic_instr
*instr
)
2790 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
2791 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2792 get_src(ctx
, instr
->src
[0]),
2795 assert(instr
->intrinsic
== nir_intrinsic_discard
);
2796 cond
= ctx
->ac
.i1false
;
2799 ctx
->abi
->emit_kill(ctx
->abi
, cond
);
2803 visit_load_local_invocation_index(struct ac_nir_context
*ctx
)
2805 LLVMValueRef result
;
2806 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
2807 result
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2808 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
2810 return LLVMBuildAdd(ctx
->ac
.builder
, result
, thread_id
, "");
2814 visit_load_subgroup_id(struct ac_nir_context
*ctx
)
2816 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
2817 LLVMValueRef result
;
2818 result
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2819 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
2820 return LLVMBuildLShr(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 6, false), "");
2822 return LLVMConstInt(ctx
->ac
.i32
, 0, false);
2827 visit_load_num_subgroups(struct ac_nir_context
*ctx
)
2829 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
2830 return LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2831 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
2833 return LLVMConstInt(ctx
->ac
.i32
, 1, false);
2838 visit_first_invocation(struct ac_nir_context
*ctx
)
2840 LLVMValueRef active_set
= ac_build_ballot(&ctx
->ac
, ctx
->ac
.i32_1
);
2842 /* The second argument is whether cttz(0) should be defined, but we do not care. */
2843 LLVMValueRef args
[] = {active_set
, ctx
->ac
.i1false
};
2844 LLVMValueRef result
= ac_build_intrinsic(&ctx
->ac
,
2846 ctx
->ac
.i64
, args
, 2,
2847 AC_FUNC_ATTR_NOUNWIND
|
2848 AC_FUNC_ATTR_READNONE
);
2850 return LLVMBuildTrunc(ctx
->ac
.builder
, result
, ctx
->ac
.i32
, "");
2854 visit_load_shared(struct ac_nir_context
*ctx
,
2855 const nir_intrinsic_instr
*instr
)
2857 LLVMValueRef values
[4], derived_ptr
, index
, ret
;
2859 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0]);
2861 for (int chan
= 0; chan
< instr
->num_components
; chan
++) {
2862 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
2863 derived_ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
2864 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, derived_ptr
, "");
2867 ret
= ac_build_gather_values(&ctx
->ac
, values
, instr
->num_components
);
2868 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
2872 visit_store_shared(struct ac_nir_context
*ctx
,
2873 const nir_intrinsic_instr
*instr
)
2875 LLVMValueRef derived_ptr
, data
,index
;
2876 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2878 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[1]);
2879 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
2881 int writemask
= nir_intrinsic_write_mask(instr
);
2882 for (int chan
= 0; chan
< 4; chan
++) {
2883 if (!(writemask
& (1 << chan
))) {
2886 data
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
2887 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
2888 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
2889 LLVMBuildStore(builder
, data
, derived_ptr
);
2893 static LLVMValueRef
visit_var_atomic(struct ac_nir_context
*ctx
,
2894 const nir_intrinsic_instr
*instr
,
2895 LLVMValueRef ptr
, int src_idx
)
2897 LLVMValueRef result
;
2898 LLVMValueRef src
= get_src(ctx
, instr
->src
[src_idx
]);
2900 if (instr
->intrinsic
== nir_intrinsic_shared_atomic_comp_swap
||
2901 instr
->intrinsic
== nir_intrinsic_deref_atomic_comp_swap
) {
2902 LLVMValueRef src1
= get_src(ctx
, instr
->src
[src_idx
+ 1]);
2903 result
= LLVMBuildAtomicCmpXchg(ctx
->ac
.builder
,
2905 LLVMAtomicOrderingSequentiallyConsistent
,
2906 LLVMAtomicOrderingSequentiallyConsistent
,
2908 result
= LLVMBuildExtractValue(ctx
->ac
.builder
, result
, 0, "");
2910 LLVMAtomicRMWBinOp op
;
2911 switch (instr
->intrinsic
) {
2912 case nir_intrinsic_shared_atomic_add
:
2913 case nir_intrinsic_deref_atomic_add
:
2914 op
= LLVMAtomicRMWBinOpAdd
;
2916 case nir_intrinsic_shared_atomic_umin
:
2917 case nir_intrinsic_deref_atomic_umin
:
2918 op
= LLVMAtomicRMWBinOpUMin
;
2920 case nir_intrinsic_shared_atomic_umax
:
2921 case nir_intrinsic_deref_atomic_umax
:
2922 op
= LLVMAtomicRMWBinOpUMax
;
2924 case nir_intrinsic_shared_atomic_imin
:
2925 case nir_intrinsic_deref_atomic_imin
:
2926 op
= LLVMAtomicRMWBinOpMin
;
2928 case nir_intrinsic_shared_atomic_imax
:
2929 case nir_intrinsic_deref_atomic_imax
:
2930 op
= LLVMAtomicRMWBinOpMax
;
2932 case nir_intrinsic_shared_atomic_and
:
2933 case nir_intrinsic_deref_atomic_and
:
2934 op
= LLVMAtomicRMWBinOpAnd
;
2936 case nir_intrinsic_shared_atomic_or
:
2937 case nir_intrinsic_deref_atomic_or
:
2938 op
= LLVMAtomicRMWBinOpOr
;
2940 case nir_intrinsic_shared_atomic_xor
:
2941 case nir_intrinsic_deref_atomic_xor
:
2942 op
= LLVMAtomicRMWBinOpXor
;
2944 case nir_intrinsic_shared_atomic_exchange
:
2945 case nir_intrinsic_deref_atomic_exchange
:
2946 op
= LLVMAtomicRMWBinOpXchg
;
2952 result
= LLVMBuildAtomicRMW(ctx
->ac
.builder
, op
, ptr
, ac_to_integer(&ctx
->ac
, src
),
2953 LLVMAtomicOrderingSequentiallyConsistent
,
2959 static LLVMValueRef
load_sample_pos(struct ac_nir_context
*ctx
)
2961 LLVMValueRef values
[2];
2962 LLVMValueRef pos
[2];
2964 pos
[0] = ac_to_float(&ctx
->ac
, ctx
->abi
->frag_pos
[0]);
2965 pos
[1] = ac_to_float(&ctx
->ac
, ctx
->abi
->frag_pos
[1]);
2967 values
[0] = ac_build_fract(&ctx
->ac
, pos
[0], 32);
2968 values
[1] = ac_build_fract(&ctx
->ac
, pos
[1], 32);
2969 return ac_build_gather_values(&ctx
->ac
, values
, 2);
2972 static LLVMValueRef
visit_interp(struct ac_nir_context
*ctx
,
2973 const nir_intrinsic_instr
*instr
)
2975 LLVMValueRef result
[4];
2976 LLVMValueRef interp_param
;
2979 LLVMValueRef src_c0
= NULL
;
2980 LLVMValueRef src_c1
= NULL
;
2981 LLVMValueRef src0
= NULL
;
2983 nir_deref_instr
*deref_instr
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2984 nir_variable
*var
= nir_deref_instr_get_variable(deref_instr
);
2985 int input_base
= ctx
->abi
->fs_input_attr_indices
[var
->data
.location
- VARYING_SLOT_VAR0
];
2986 switch (instr
->intrinsic
) {
2987 case nir_intrinsic_interp_deref_at_centroid
:
2988 location
= INTERP_CENTROID
;
2990 case nir_intrinsic_interp_deref_at_sample
:
2991 case nir_intrinsic_interp_deref_at_offset
:
2992 location
= INTERP_CENTER
;
2993 src0
= get_src(ctx
, instr
->src
[1]);
2999 if (instr
->intrinsic
== nir_intrinsic_interp_deref_at_offset
) {
3000 src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_0
, ""));
3001 src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_1
, ""));
3002 } else if (instr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
) {
3003 LLVMValueRef sample_position
;
3004 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
3006 /* fetch sample ID */
3007 sample_position
= ctx
->abi
->load_sample_position(ctx
->abi
, src0
);
3009 src_c0
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_0
, "");
3010 src_c0
= LLVMBuildFSub(ctx
->ac
.builder
, src_c0
, halfval
, "");
3011 src_c1
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_1
, "");
3012 src_c1
= LLVMBuildFSub(ctx
->ac
.builder
, src_c1
, halfval
, "");
3014 interp_param
= ctx
->abi
->lookup_interp_param(ctx
->abi
, var
->data
.interpolation
, location
);
3016 if (location
== INTERP_CENTER
) {
3017 LLVMValueRef ij_out
[2];
3018 LLVMValueRef ddxy_out
= ac_build_ddxy_interp(&ctx
->ac
, interp_param
);
3021 * take the I then J parameters, and the DDX/Y for it, and
3022 * calculate the IJ inputs for the interpolator.
3023 * temp1 = ddx * offset/sample.x + I;
3024 * interp_param.I = ddy * offset/sample.y + temp1;
3025 * temp1 = ddx * offset/sample.x + J;
3026 * interp_param.J = ddy * offset/sample.y + temp1;
3028 for (unsigned i
= 0; i
< 2; i
++) {
3029 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
3030 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
3031 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3032 ddxy_out
, ix_ll
, "");
3033 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3034 ddxy_out
, iy_ll
, "");
3035 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3036 interp_param
, ix_ll
, "");
3037 LLVMValueRef temp1
, temp2
;
3039 interp_el
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_el
,
3042 temp1
= ac_build_fmad(&ctx
->ac
, ddx_el
, src_c0
, interp_el
);
3043 temp2
= ac_build_fmad(&ctx
->ac
, ddy_el
, src_c1
, temp1
);
3045 ij_out
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
,
3046 temp2
, ctx
->ac
.i32
, "");
3048 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
3052 LLVMValueRef attrib_idx
= ctx
->ac
.i32_0
;
3053 while(deref_instr
->deref_type
!= nir_deref_type_var
) {
3054 if (deref_instr
->deref_type
== nir_deref_type_array
) {
3055 unsigned array_size
= glsl_count_attribute_slots(deref_instr
->type
, false);
3057 LLVMValueRef offset
;
3058 if (nir_src_is_const(deref_instr
->arr
.index
)) {
3059 offset
= LLVMConstInt(ctx
->ac
.i32
, array_size
* nir_src_as_uint(deref_instr
->arr
.index
), false);
3061 LLVMValueRef indirect
= get_src(ctx
, deref_instr
->arr
.index
);
3063 offset
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
3064 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
3067 attrib_idx
= LLVMBuildAdd(ctx
->ac
.builder
, attrib_idx
, offset
, "");
3068 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3069 } else if (deref_instr
->deref_type
== nir_deref_type_struct
) {
3070 LLVMValueRef offset
;
3071 unsigned sidx
= deref_instr
->strct
.index
;
3072 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3073 offset
= LLVMConstInt(ctx
->ac
.i32
, glsl_get_struct_location_offset(deref_instr
->type
, sidx
), false);
3074 attrib_idx
= LLVMBuildAdd(ctx
->ac
.builder
, attrib_idx
, offset
, "");
3076 unreachable("Unsupported deref type");
3081 unsigned attrib_size
= glsl_count_attribute_slots(var
->type
, false);
3082 for (chan
= 0; chan
< 4; chan
++) {
3083 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->ac
.f32
, attrib_size
));
3084 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
3086 for (unsigned idx
= 0; idx
< attrib_size
; ++idx
) {
3087 LLVMValueRef v
, attr_number
;
3089 attr_number
= LLVMConstInt(ctx
->ac
.i32
, input_base
+ idx
, false);
3091 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
,
3092 interp_param
, ctx
->ac
.v2f32
, "");
3093 LLVMValueRef i
= LLVMBuildExtractElement(
3094 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_0
, "");
3095 LLVMValueRef j
= LLVMBuildExtractElement(
3096 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_1
, "");
3098 v
= ac_build_fs_interp(&ctx
->ac
, llvm_chan
, attr_number
,
3099 ctx
->abi
->prim_mask
, i
, j
);
3101 v
= ac_build_fs_interp_mov(&ctx
->ac
, LLVMConstInt(ctx
->ac
.i32
, 2, false),
3102 llvm_chan
, attr_number
, ctx
->abi
->prim_mask
);
3105 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
, gather
, v
,
3106 LLVMConstInt(ctx
->ac
.i32
, idx
, false), "");
3109 result
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, gather
, attrib_idx
, "");
3112 return ac_build_varying_gather_values(&ctx
->ac
, result
, instr
->num_components
,
3113 var
->data
.location_frac
);
3116 static void visit_intrinsic(struct ac_nir_context
*ctx
,
3117 nir_intrinsic_instr
*instr
)
3119 LLVMValueRef result
= NULL
;
3121 switch (instr
->intrinsic
) {
3122 case nir_intrinsic_ballot
:
3123 result
= ac_build_ballot(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3125 case nir_intrinsic_read_invocation
:
3126 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3127 get_src(ctx
, instr
->src
[1]));
3129 case nir_intrinsic_read_first_invocation
:
3130 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), NULL
);
3132 case nir_intrinsic_load_subgroup_invocation
:
3133 result
= ac_get_thread_id(&ctx
->ac
);
3135 case nir_intrinsic_load_work_group_id
: {
3136 LLVMValueRef values
[3];
3138 for (int i
= 0; i
< 3; i
++) {
3139 values
[i
] = ctx
->abi
->workgroup_ids
[i
] ?
3140 ctx
->abi
->workgroup_ids
[i
] : ctx
->ac
.i32_0
;
3143 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
3146 case nir_intrinsic_load_base_vertex
:
3147 case nir_intrinsic_load_first_vertex
:
3148 result
= ctx
->abi
->load_base_vertex(ctx
->abi
);
3150 case nir_intrinsic_load_local_group_size
:
3151 result
= ctx
->abi
->load_local_group_size(ctx
->abi
);
3153 case nir_intrinsic_load_vertex_id
:
3154 result
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
->vertex_id
,
3155 ctx
->abi
->base_vertex
, "");
3157 case nir_intrinsic_load_vertex_id_zero_base
: {
3158 result
= ctx
->abi
->vertex_id
;
3161 case nir_intrinsic_load_local_invocation_id
: {
3162 result
= ctx
->abi
->local_invocation_ids
;
3165 case nir_intrinsic_load_base_instance
:
3166 result
= ctx
->abi
->start_instance
;
3168 case nir_intrinsic_load_draw_id
:
3169 result
= ctx
->abi
->draw_id
;
3171 case nir_intrinsic_load_view_index
:
3172 result
= ctx
->abi
->view_index
;
3174 case nir_intrinsic_load_invocation_id
:
3175 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
3176 result
= ac_unpack_param(&ctx
->ac
, ctx
->abi
->tcs_rel_ids
, 8, 5);
3178 result
= ctx
->abi
->gs_invocation_id
;
3180 case nir_intrinsic_load_primitive_id
:
3181 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3182 result
= ctx
->abi
->gs_prim_id
;
3183 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3184 result
= ctx
->abi
->tcs_patch_id
;
3185 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3186 result
= ctx
->abi
->tes_patch_id
;
3188 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
3190 case nir_intrinsic_load_sample_id
:
3191 result
= ac_unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
3193 case nir_intrinsic_load_sample_pos
:
3194 result
= load_sample_pos(ctx
);
3196 case nir_intrinsic_load_sample_mask_in
:
3197 result
= ctx
->abi
->load_sample_mask_in(ctx
->abi
);
3199 case nir_intrinsic_load_frag_coord
: {
3200 LLVMValueRef values
[4] = {
3201 ctx
->abi
->frag_pos
[0],
3202 ctx
->abi
->frag_pos
[1],
3203 ctx
->abi
->frag_pos
[2],
3204 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, ctx
->abi
->frag_pos
[3])
3206 result
= ac_to_integer(&ctx
->ac
,
3207 ac_build_gather_values(&ctx
->ac
, values
, 4));
3210 case nir_intrinsic_load_front_face
:
3211 result
= ctx
->abi
->front_face
;
3213 case nir_intrinsic_load_helper_invocation
:
3214 result
= ac_build_load_helper_invocation(&ctx
->ac
);
3216 case nir_intrinsic_load_instance_id
:
3217 result
= ctx
->abi
->instance_id
;
3219 case nir_intrinsic_load_num_work_groups
:
3220 result
= ctx
->abi
->num_work_groups
;
3222 case nir_intrinsic_load_local_invocation_index
:
3223 result
= visit_load_local_invocation_index(ctx
);
3225 case nir_intrinsic_load_subgroup_id
:
3226 result
= visit_load_subgroup_id(ctx
);
3228 case nir_intrinsic_load_num_subgroups
:
3229 result
= visit_load_num_subgroups(ctx
);
3231 case nir_intrinsic_first_invocation
:
3232 result
= visit_first_invocation(ctx
);
3234 case nir_intrinsic_load_push_constant
:
3235 result
= visit_load_push_constant(ctx
, instr
);
3237 case nir_intrinsic_vulkan_resource_index
: {
3238 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
3239 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
3240 unsigned binding
= nir_intrinsic_binding(instr
);
3242 result
= ctx
->abi
->load_resource(ctx
->abi
, index
, desc_set
,
3246 case nir_intrinsic_vulkan_resource_reindex
:
3247 result
= visit_vulkan_resource_reindex(ctx
, instr
);
3249 case nir_intrinsic_store_ssbo
:
3250 visit_store_ssbo(ctx
, instr
);
3252 case nir_intrinsic_load_ssbo
:
3253 result
= visit_load_buffer(ctx
, instr
);
3255 case nir_intrinsic_ssbo_atomic_add
:
3256 case nir_intrinsic_ssbo_atomic_imin
:
3257 case nir_intrinsic_ssbo_atomic_umin
:
3258 case nir_intrinsic_ssbo_atomic_imax
:
3259 case nir_intrinsic_ssbo_atomic_umax
:
3260 case nir_intrinsic_ssbo_atomic_and
:
3261 case nir_intrinsic_ssbo_atomic_or
:
3262 case nir_intrinsic_ssbo_atomic_xor
:
3263 case nir_intrinsic_ssbo_atomic_exchange
:
3264 case nir_intrinsic_ssbo_atomic_comp_swap
:
3265 result
= visit_atomic_ssbo(ctx
, instr
);
3267 case nir_intrinsic_load_ubo
:
3268 result
= visit_load_ubo_buffer(ctx
, instr
);
3270 case nir_intrinsic_get_buffer_size
:
3271 result
= visit_get_buffer_size(ctx
, instr
);
3273 case nir_intrinsic_load_deref
:
3274 result
= visit_load_var(ctx
, instr
);
3276 case nir_intrinsic_store_deref
:
3277 visit_store_var(ctx
, instr
);
3279 case nir_intrinsic_load_shared
:
3280 result
= visit_load_shared(ctx
, instr
);
3282 case nir_intrinsic_store_shared
:
3283 visit_store_shared(ctx
, instr
);
3285 case nir_intrinsic_bindless_image_samples
:
3286 result
= visit_image_samples(ctx
, instr
, true);
3288 case nir_intrinsic_image_deref_samples
:
3289 result
= visit_image_samples(ctx
, instr
, false);
3291 case nir_intrinsic_bindless_image_load
:
3292 result
= visit_image_load(ctx
, instr
, true);
3294 case nir_intrinsic_image_deref_load
:
3295 result
= visit_image_load(ctx
, instr
, false);
3297 case nir_intrinsic_bindless_image_store
:
3298 visit_image_store(ctx
, instr
, true);
3300 case nir_intrinsic_image_deref_store
:
3301 visit_image_store(ctx
, instr
, false);
3303 case nir_intrinsic_bindless_image_atomic_add
:
3304 case nir_intrinsic_bindless_image_atomic_min
:
3305 case nir_intrinsic_bindless_image_atomic_max
:
3306 case nir_intrinsic_bindless_image_atomic_and
:
3307 case nir_intrinsic_bindless_image_atomic_or
:
3308 case nir_intrinsic_bindless_image_atomic_xor
:
3309 case nir_intrinsic_bindless_image_atomic_exchange
:
3310 case nir_intrinsic_bindless_image_atomic_comp_swap
:
3311 result
= visit_image_atomic(ctx
, instr
, true);
3313 case nir_intrinsic_image_deref_atomic_add
:
3314 case nir_intrinsic_image_deref_atomic_min
:
3315 case nir_intrinsic_image_deref_atomic_max
:
3316 case nir_intrinsic_image_deref_atomic_and
:
3317 case nir_intrinsic_image_deref_atomic_or
:
3318 case nir_intrinsic_image_deref_atomic_xor
:
3319 case nir_intrinsic_image_deref_atomic_exchange
:
3320 case nir_intrinsic_image_deref_atomic_comp_swap
:
3321 result
= visit_image_atomic(ctx
, instr
, false);
3323 case nir_intrinsic_bindless_image_size
:
3324 result
= visit_image_size(ctx
, instr
, true);
3326 case nir_intrinsic_image_deref_size
:
3327 result
= visit_image_size(ctx
, instr
, false);
3329 case nir_intrinsic_shader_clock
:
3330 result
= ac_build_shader_clock(&ctx
->ac
);
3332 case nir_intrinsic_discard
:
3333 case nir_intrinsic_discard_if
:
3334 emit_discard(ctx
, instr
);
3336 case nir_intrinsic_memory_barrier
:
3337 case nir_intrinsic_group_memory_barrier
:
3338 case nir_intrinsic_memory_barrier_atomic_counter
:
3339 case nir_intrinsic_memory_barrier_buffer
:
3340 case nir_intrinsic_memory_barrier_image
:
3341 case nir_intrinsic_memory_barrier_shared
:
3342 emit_membar(&ctx
->ac
, instr
);
3344 case nir_intrinsic_barrier
:
3345 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
3347 case nir_intrinsic_shared_atomic_add
:
3348 case nir_intrinsic_shared_atomic_imin
:
3349 case nir_intrinsic_shared_atomic_umin
:
3350 case nir_intrinsic_shared_atomic_imax
:
3351 case nir_intrinsic_shared_atomic_umax
:
3352 case nir_intrinsic_shared_atomic_and
:
3353 case nir_intrinsic_shared_atomic_or
:
3354 case nir_intrinsic_shared_atomic_xor
:
3355 case nir_intrinsic_shared_atomic_exchange
:
3356 case nir_intrinsic_shared_atomic_comp_swap
: {
3357 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0]);
3358 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3361 case nir_intrinsic_deref_atomic_add
:
3362 case nir_intrinsic_deref_atomic_imin
:
3363 case nir_intrinsic_deref_atomic_umin
:
3364 case nir_intrinsic_deref_atomic_imax
:
3365 case nir_intrinsic_deref_atomic_umax
:
3366 case nir_intrinsic_deref_atomic_and
:
3367 case nir_intrinsic_deref_atomic_or
:
3368 case nir_intrinsic_deref_atomic_xor
:
3369 case nir_intrinsic_deref_atomic_exchange
:
3370 case nir_intrinsic_deref_atomic_comp_swap
: {
3371 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
3372 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3375 case nir_intrinsic_interp_deref_at_centroid
:
3376 case nir_intrinsic_interp_deref_at_sample
:
3377 case nir_intrinsic_interp_deref_at_offset
:
3378 result
= visit_interp(ctx
, instr
);
3380 case nir_intrinsic_emit_vertex
:
3381 ctx
->abi
->emit_vertex(ctx
->abi
, nir_intrinsic_stream_id(instr
), ctx
->abi
->outputs
);
3383 case nir_intrinsic_end_primitive
:
3384 ctx
->abi
->emit_primitive(ctx
->abi
, nir_intrinsic_stream_id(instr
));
3386 case nir_intrinsic_load_tess_coord
:
3387 result
= ctx
->abi
->load_tess_coord(ctx
->abi
);
3389 case nir_intrinsic_load_tess_level_outer
:
3390 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
);
3392 case nir_intrinsic_load_tess_level_inner
:
3393 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
);
3395 case nir_intrinsic_load_patch_vertices_in
:
3396 result
= ctx
->abi
->load_patch_vertices_in(ctx
->abi
);
3398 case nir_intrinsic_vote_all
: {
3399 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3400 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
3403 case nir_intrinsic_vote_any
: {
3404 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3405 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
3408 case nir_intrinsic_shuffle
:
3409 result
= ac_build_shuffle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3410 get_src(ctx
, instr
->src
[1]));
3412 case nir_intrinsic_reduce
:
3413 result
= ac_build_reduce(&ctx
->ac
,
3414 get_src(ctx
, instr
->src
[0]),
3415 instr
->const_index
[0],
3416 instr
->const_index
[1]);
3418 case nir_intrinsic_inclusive_scan
:
3419 result
= ac_build_inclusive_scan(&ctx
->ac
,
3420 get_src(ctx
, instr
->src
[0]),
3421 instr
->const_index
[0]);
3423 case nir_intrinsic_exclusive_scan
:
3424 result
= ac_build_exclusive_scan(&ctx
->ac
,
3425 get_src(ctx
, instr
->src
[0]),
3426 instr
->const_index
[0]);
3428 case nir_intrinsic_quad_broadcast
: {
3429 unsigned lane
= nir_src_as_uint(instr
->src
[1]);
3430 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3431 lane
, lane
, lane
, lane
);
3434 case nir_intrinsic_quad_swap_horizontal
:
3435 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 1, 0, 3 ,2);
3437 case nir_intrinsic_quad_swap_vertical
:
3438 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 2, 3, 0 ,1);
3440 case nir_intrinsic_quad_swap_diagonal
:
3441 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 3, 2, 1 ,0);
3444 fprintf(stderr
, "Unknown intrinsic: ");
3445 nir_print_instr(&instr
->instr
, stderr
);
3446 fprintf(stderr
, "\n");
3450 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3454 static LLVMValueRef
get_bindless_index_from_uniform(struct ac_nir_context
*ctx
,
3455 unsigned base_index
,
3456 unsigned constant_index
,
3457 LLVMValueRef dynamic_index
)
3459 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, base_index
* 4, 0);
3460 LLVMValueRef index
= LLVMBuildAdd(ctx
->ac
.builder
, dynamic_index
,
3461 LLVMConstInt(ctx
->ac
.i32
, constant_index
, 0), "");
3463 /* Bindless uniforms are 64bit so multiple index by 8 */
3464 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i32
, 8, 0), "");
3465 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
3467 LLVMValueRef ubo_index
= ctx
->abi
->load_ubo(ctx
->abi
, ctx
->ac
.i32_0
);
3469 LLVMValueRef ret
= ac_build_buffer_load(&ctx
->ac
, ubo_index
, 1, NULL
, offset
,
3470 NULL
, 0, false, false, true, true);
3472 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ctx
->ac
.i32
, "");
3475 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
3476 nir_deref_instr
*deref_instr
,
3477 enum ac_descriptor_type desc_type
,
3478 const nir_instr
*instr
,
3479 bool image
, bool write
)
3481 LLVMValueRef index
= NULL
;
3482 unsigned constant_index
= 0;
3483 unsigned descriptor_set
;
3484 unsigned base_index
;
3485 bool bindless
= false;
3490 nir_intrinsic_instr
*img_instr
= nir_instr_as_intrinsic(instr
);
3493 index
= get_src(ctx
, img_instr
->src
[0]);
3495 nir_tex_instr
*tex_instr
= nir_instr_as_tex(instr
);
3496 int sampSrcIdx
= nir_tex_instr_src_index(tex_instr
,
3497 nir_tex_src_sampler_handle
);
3498 if (sampSrcIdx
!= -1) {
3501 index
= get_src(ctx
, tex_instr
->src
[sampSrcIdx
].src
);
3503 assert(tex_instr
&& !image
);
3504 base_index
= tex_instr
->sampler_index
;
3508 while(deref_instr
->deref_type
!= nir_deref_type_var
) {
3509 if (deref_instr
->deref_type
== nir_deref_type_array
) {
3510 unsigned array_size
= glsl_get_aoa_size(deref_instr
->type
);
3514 if (nir_src_is_const(deref_instr
->arr
.index
)) {
3515 constant_index
+= array_size
* nir_src_as_uint(deref_instr
->arr
.index
);
3517 LLVMValueRef indirect
= get_src(ctx
, deref_instr
->arr
.index
);
3519 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
3520 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
3525 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
3528 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3529 } else if (deref_instr
->deref_type
== nir_deref_type_struct
) {
3530 unsigned sidx
= deref_instr
->strct
.index
;
3531 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3532 constant_index
+= glsl_get_struct_location_offset(deref_instr
->type
, sidx
);
3534 unreachable("Unsupported deref type");
3537 descriptor_set
= deref_instr
->var
->data
.descriptor_set
;
3539 if (deref_instr
->var
->data
.bindless
) {
3540 /* For now just assert on unhandled variable types */
3541 assert(deref_instr
->var
->data
.mode
== nir_var_uniform
);
3543 base_index
= deref_instr
->var
->data
.driver_location
;
3546 index
= index
? index
: ctx
->ac
.i32_0
;
3547 index
= get_bindless_index_from_uniform(ctx
, base_index
,
3548 constant_index
, index
);
3550 base_index
= deref_instr
->var
->data
.binding
;
3553 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
3556 constant_index
, index
,
3557 desc_type
, image
, write
, bindless
);
3560 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3563 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3564 * filtering manually. The driver sets img7 to a mask clearing
3565 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3566 * s_and_b32 samp0, samp0, img7
3569 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3571 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
3572 LLVMValueRef res
, LLVMValueRef samp
)
3574 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3575 LLVMValueRef img7
, samp0
;
3577 if (ctx
->ac
.chip_class
>= VI
)
3580 img7
= LLVMBuildExtractElement(builder
, res
,
3581 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
3582 samp0
= LLVMBuildExtractElement(builder
, samp
,
3583 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
3584 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
3585 return LLVMBuildInsertElement(builder
, samp
, samp0
,
3586 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
3589 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
3590 nir_tex_instr
*instr
,
3591 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
3592 LLVMValueRef
*fmask_ptr
)
3594 nir_deref_instr
*texture_deref_instr
= NULL
;
3595 nir_deref_instr
*sampler_deref_instr
= NULL
;
3597 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3598 switch (instr
->src
[i
].src_type
) {
3599 case nir_tex_src_texture_deref
:
3600 texture_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
3602 case nir_tex_src_sampler_deref
:
3603 sampler_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
3610 if (!sampler_deref_instr
)
3611 sampler_deref_instr
= texture_deref_instr
;
3613 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
3614 *res_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_BUFFER
, &instr
->instr
, false, false);
3616 *res_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_IMAGE
, &instr
->instr
, false, false);
3618 *samp_ptr
= get_sampler_desc(ctx
, sampler_deref_instr
, AC_DESC_SAMPLER
, &instr
->instr
, false, false);
3619 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
3620 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
3622 if (fmask_ptr
&& (instr
->op
== nir_texop_txf_ms
||
3623 instr
->op
== nir_texop_samples_identical
))
3624 *fmask_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_FMASK
, &instr
->instr
, false, false);
3627 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
3630 coord
= ac_to_float(ctx
, coord
);
3631 coord
= ac_build_round(ctx
, coord
);
3632 coord
= ac_to_integer(ctx
, coord
);
3636 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
3638 LLVMValueRef result
= NULL
;
3639 struct ac_image_args args
= { 0 };
3640 LLVMValueRef fmask_ptr
= NULL
, sample_index
= NULL
;
3641 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
3642 unsigned offset_src
= 0;
3644 tex_fetch_ptrs(ctx
, instr
, &args
.resource
, &args
.sampler
, &fmask_ptr
);
3646 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3647 switch (instr
->src
[i
].src_type
) {
3648 case nir_tex_src_coord
: {
3649 LLVMValueRef coord
= get_src(ctx
, instr
->src
[i
].src
);
3650 for (unsigned chan
= 0; chan
< instr
->coord_components
; ++chan
)
3651 args
.coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
3654 case nir_tex_src_projector
:
3656 case nir_tex_src_comparator
:
3657 if (instr
->is_shadow
)
3658 args
.compare
= get_src(ctx
, instr
->src
[i
].src
);
3660 case nir_tex_src_offset
:
3661 args
.offset
= get_src(ctx
, instr
->src
[i
].src
);
3664 case nir_tex_src_bias
:
3665 if (instr
->op
== nir_texop_txb
)
3666 args
.bias
= get_src(ctx
, instr
->src
[i
].src
);
3668 case nir_tex_src_lod
: {
3669 if (nir_src_is_const(instr
->src
[i
].src
) && nir_src_as_uint(instr
->src
[i
].src
) == 0)
3670 args
.level_zero
= true;
3672 args
.lod
= get_src(ctx
, instr
->src
[i
].src
);
3675 case nir_tex_src_ms_index
:
3676 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
3678 case nir_tex_src_ms_mcs
:
3680 case nir_tex_src_ddx
:
3681 ddx
= get_src(ctx
, instr
->src
[i
].src
);
3683 case nir_tex_src_ddy
:
3684 ddy
= get_src(ctx
, instr
->src
[i
].src
);
3686 case nir_tex_src_texture_offset
:
3687 case nir_tex_src_sampler_offset
:
3688 case nir_tex_src_plane
:
3694 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
3695 result
= get_buffer_size(ctx
, args
.resource
, true);
3699 if (instr
->op
== nir_texop_texture_samples
) {
3700 LLVMValueRef res
, samples
, is_msaa
;
3701 res
= LLVMBuildBitCast(ctx
->ac
.builder
, args
.resource
, ctx
->ac
.v8i32
, "");
3702 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
3703 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
3704 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
3705 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
3706 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
3707 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
3708 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
3709 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
3711 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
3712 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
3713 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
3714 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
3715 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
3717 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
3723 if (args
.offset
&& instr
->op
!= nir_texop_txf
) {
3724 LLVMValueRef offset
[3], pack
;
3725 for (unsigned chan
= 0; chan
< 3; ++chan
)
3726 offset
[chan
] = ctx
->ac
.i32_0
;
3728 unsigned num_components
= ac_get_llvm_num_components(args
.offset
);
3729 for (unsigned chan
= 0; chan
< num_components
; chan
++) {
3730 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, args
.offset
, chan
);
3731 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
3732 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
3734 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
3735 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
3737 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
3738 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
3742 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
3743 * so the depth comparison value isn't clamped for Z16 and
3744 * Z24 anymore. Do it manually here.
3746 * It's unnecessary if the original texture format was
3747 * Z32_FLOAT, but we don't know that here.
3749 if (args
.compare
&& ctx
->ac
.chip_class
>= VI
&& ctx
->abi
->clamp_shadow_reference
)
3750 args
.compare
= ac_build_clamp(&ctx
->ac
, ac_to_float(&ctx
->ac
, args
.compare
));
3752 /* pack derivatives */
3754 int num_src_deriv_channels
, num_dest_deriv_channels
;
3755 switch (instr
->sampler_dim
) {
3756 case GLSL_SAMPLER_DIM_3D
:
3757 case GLSL_SAMPLER_DIM_CUBE
:
3758 num_src_deriv_channels
= 3;
3759 num_dest_deriv_channels
= 3;
3761 case GLSL_SAMPLER_DIM_2D
:
3763 num_src_deriv_channels
= 2;
3764 num_dest_deriv_channels
= 2;
3766 case GLSL_SAMPLER_DIM_1D
:
3767 num_src_deriv_channels
= 1;
3768 if (ctx
->ac
.chip_class
>= GFX9
) {
3769 num_dest_deriv_channels
= 2;
3771 num_dest_deriv_channels
= 1;
3776 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
3777 args
.derivs
[i
] = ac_to_float(&ctx
->ac
,
3778 ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
3779 args
.derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
,
3780 ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
3782 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
3783 args
.derivs
[i
] = ctx
->ac
.f32_0
;
3784 args
.derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
3788 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& args
.coords
[0]) {
3789 for (unsigned chan
= 0; chan
< instr
->coord_components
; chan
++)
3790 args
.coords
[chan
] = ac_to_float(&ctx
->ac
, args
.coords
[chan
]);
3791 if (instr
->coord_components
== 3)
3792 args
.coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
3793 ac_prepare_cube_coords(&ctx
->ac
,
3794 instr
->op
== nir_texop_txd
, instr
->is_array
,
3795 instr
->op
== nir_texop_lod
, args
.coords
, args
.derivs
);
3798 /* Texture coordinates fixups */
3799 if (instr
->coord_components
> 1 &&
3800 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3802 instr
->op
!= nir_texop_txf
) {
3803 args
.coords
[1] = apply_round_slice(&ctx
->ac
, args
.coords
[1]);
3806 if (instr
->coord_components
> 2 &&
3807 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
||
3808 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
||
3809 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3810 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
) &&
3812 instr
->op
!= nir_texop_txf
&& instr
->op
!= nir_texop_txf_ms
) {
3813 args
.coords
[2] = apply_round_slice(&ctx
->ac
, args
.coords
[2]);
3816 if (ctx
->ac
.chip_class
>= GFX9
&&
3817 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3818 instr
->op
!= nir_texop_lod
) {
3819 LLVMValueRef filler
;
3820 if (instr
->op
== nir_texop_txf
)
3821 filler
= ctx
->ac
.i32_0
;
3823 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
3825 if (instr
->is_array
)
3826 args
.coords
[2] = args
.coords
[1];
3827 args
.coords
[1] = filler
;
3830 /* Pack sample index */
3831 if (instr
->op
== nir_texop_txf_ms
&& sample_index
)
3832 args
.coords
[instr
->coord_components
] = sample_index
;
3834 if (instr
->op
== nir_texop_samples_identical
) {
3835 struct ac_image_args txf_args
= { 0 };
3836 memcpy(txf_args
.coords
, args
.coords
, sizeof(txf_args
.coords
));
3838 txf_args
.dmask
= 0xf;
3839 txf_args
.resource
= fmask_ptr
;
3840 txf_args
.dim
= instr
->is_array
? ac_image_2darray
: ac_image_2d
;
3841 result
= build_tex_intrinsic(ctx
, instr
, &txf_args
);
3843 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
3844 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
3848 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
3849 instr
->op
!= nir_texop_txs
) {
3850 unsigned sample_chan
= instr
->is_array
? 3 : 2;
3851 args
.coords
[sample_chan
] = adjust_sample_index_using_fmask(
3852 &ctx
->ac
, args
.coords
[0], args
.coords
[1],
3853 instr
->is_array
? args
.coords
[2] : NULL
,
3854 args
.coords
[sample_chan
], fmask_ptr
);
3857 if (args
.offset
&& instr
->op
== nir_texop_txf
) {
3858 int num_offsets
= instr
->src
[offset_src
].src
.ssa
->num_components
;
3859 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
3860 for (unsigned i
= 0; i
< num_offsets
; ++i
) {
3861 args
.coords
[i
] = LLVMBuildAdd(
3862 ctx
->ac
.builder
, args
.coords
[i
],
3863 LLVMConstInt(ctx
->ac
.i32
, nir_src_comp_as_uint(instr
->src
[offset_src
].src
, i
), false), "");
3868 /* TODO TG4 support */
3870 if (instr
->op
== nir_texop_tg4
) {
3871 if (instr
->is_shadow
)
3874 args
.dmask
= 1 << instr
->component
;
3877 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_BUF
)
3878 args
.dim
= get_ac_sampler_dim(&ctx
->ac
, instr
->sampler_dim
, instr
->is_array
);
3879 result
= build_tex_intrinsic(ctx
, instr
, &args
);
3881 if (instr
->op
== nir_texop_query_levels
)
3882 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
3883 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
3884 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
3885 instr
->op
!= nir_texop_tg4
)
3886 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
3887 else if (instr
->op
== nir_texop_txs
&&
3888 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
3890 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3891 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3892 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
3893 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3894 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
3895 } else if (ctx
->ac
.chip_class
>= GFX9
&&
3896 instr
->op
== nir_texop_txs
&&
3897 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3899 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3900 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
3901 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
3903 } else if (instr
->dest
.ssa
.num_components
!= 4)
3904 result
= ac_trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
3908 assert(instr
->dest
.is_ssa
);
3909 result
= ac_to_integer(&ctx
->ac
, result
);
3910 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3915 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
3917 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
3918 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
3920 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3921 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
3924 static void visit_post_phi(struct ac_nir_context
*ctx
,
3925 nir_phi_instr
*instr
,
3926 LLVMValueRef llvm_phi
)
3928 nir_foreach_phi_src(src
, instr
) {
3929 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
3930 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
3932 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
3936 static void phi_post_pass(struct ac_nir_context
*ctx
)
3938 hash_table_foreach(ctx
->phis
, entry
) {
3939 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
3940 (LLVMValueRef
)entry
->data
);
3945 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
3946 const nir_ssa_undef_instr
*instr
)
3948 unsigned num_components
= instr
->def
.num_components
;
3949 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
3952 if (num_components
== 1)
3953 undef
= LLVMGetUndef(type
);
3955 undef
= LLVMGetUndef(LLVMVectorType(type
, num_components
));
3957 ctx
->ssa_defs
[instr
->def
.index
] = undef
;
3960 static void visit_jump(struct ac_llvm_context
*ctx
,
3961 const nir_jump_instr
*instr
)
3963 switch (instr
->type
) {
3964 case nir_jump_break
:
3965 ac_build_break(ctx
);
3967 case nir_jump_continue
:
3968 ac_build_continue(ctx
);
3971 fprintf(stderr
, "Unknown NIR jump instr: ");
3972 nir_print_instr(&instr
->instr
, stderr
);
3973 fprintf(stderr
, "\n");
3979 glsl_base_to_llvm_type(struct ac_llvm_context
*ac
,
3980 enum glsl_base_type type
)
3984 case GLSL_TYPE_UINT
:
3985 case GLSL_TYPE_BOOL
:
3986 case GLSL_TYPE_SUBROUTINE
:
3988 case GLSL_TYPE_INT8
:
3989 case GLSL_TYPE_UINT8
:
3991 case GLSL_TYPE_INT16
:
3992 case GLSL_TYPE_UINT16
:
3994 case GLSL_TYPE_FLOAT
:
3996 case GLSL_TYPE_FLOAT16
:
3998 case GLSL_TYPE_INT64
:
3999 case GLSL_TYPE_UINT64
:
4001 case GLSL_TYPE_DOUBLE
:
4004 unreachable("unknown GLSL type");
4009 glsl_to_llvm_type(struct ac_llvm_context
*ac
,
4010 const struct glsl_type
*type
)
4012 if (glsl_type_is_scalar(type
)) {
4013 return glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
));
4016 if (glsl_type_is_vector(type
)) {
4017 return LLVMVectorType(
4018 glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
)),
4019 glsl_get_vector_elements(type
));
4022 if (glsl_type_is_matrix(type
)) {
4023 return LLVMArrayType(
4024 glsl_to_llvm_type(ac
, glsl_get_column_type(type
)),
4025 glsl_get_matrix_columns(type
));
4028 if (glsl_type_is_array(type
)) {
4029 return LLVMArrayType(
4030 glsl_to_llvm_type(ac
, glsl_get_array_element(type
)),
4031 glsl_get_length(type
));
4034 assert(glsl_type_is_struct_or_ifc(type
));
4036 LLVMTypeRef member_types
[glsl_get_length(type
)];
4038 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
4040 glsl_to_llvm_type(ac
,
4041 glsl_get_struct_field(type
, i
));
4044 return LLVMStructTypeInContext(ac
->context
, member_types
,
4045 glsl_get_length(type
), false);
4048 static void visit_deref(struct ac_nir_context
*ctx
,
4049 nir_deref_instr
*instr
)
4051 if (instr
->mode
!= nir_var_mem_shared
&&
4052 instr
->mode
!= nir_var_mem_global
)
4055 LLVMValueRef result
= NULL
;
4056 switch(instr
->deref_type
) {
4057 case nir_deref_type_var
: {
4058 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, instr
->var
);
4059 result
= entry
->data
;
4062 case nir_deref_type_struct
:
4063 if (instr
->mode
== nir_var_mem_global
) {
4064 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
4065 uint64_t offset
= glsl_get_struct_field_offset(parent
->type
,
4066 instr
->strct
.index
);
4067 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4068 LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
4070 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4071 LLVMConstInt(ctx
->ac
.i32
, instr
->strct
.index
, 0));
4074 case nir_deref_type_array
:
4075 if (instr
->mode
== nir_var_mem_global
) {
4076 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
4077 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
4079 if ((glsl_type_is_matrix(parent
->type
) &&
4080 glsl_matrix_type_is_row_major(parent
->type
)) ||
4081 (glsl_type_is_vector(parent
->type
) && stride
== 0))
4082 stride
= type_scalar_size_bytes(parent
->type
);
4085 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
4086 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
4087 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
4089 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
4091 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
4093 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4094 get_src(ctx
, instr
->arr
.index
));
4097 case nir_deref_type_ptr_as_array
:
4098 if (instr
->mode
== nir_var_mem_global
) {
4099 unsigned stride
= nir_deref_instr_ptr_as_array_stride(instr
);
4101 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
4102 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
4103 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
4105 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
4107 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
4109 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4110 get_src(ctx
, instr
->arr
.index
));
4113 case nir_deref_type_cast
: {
4114 result
= get_src(ctx
, instr
->parent
);
4116 /* We can't use the structs from LLVM because the shader
4117 * specifies its own offsets. */
4118 LLVMTypeRef pointee_type
= ctx
->ac
.i8
;
4119 if (instr
->mode
== nir_var_mem_shared
)
4120 pointee_type
= glsl_to_llvm_type(&ctx
->ac
, instr
->type
);
4122 unsigned address_space
;
4124 switch(instr
->mode
) {
4125 case nir_var_mem_shared
:
4126 address_space
= AC_ADDR_SPACE_LDS
;
4128 case nir_var_mem_global
:
4129 address_space
= AC_ADDR_SPACE_GLOBAL
;
4132 unreachable("Unhandled address space");
4135 LLVMTypeRef type
= LLVMPointerType(pointee_type
, address_space
);
4137 if (LLVMTypeOf(result
) != type
) {
4138 if (LLVMGetTypeKind(LLVMTypeOf(result
)) == LLVMVectorTypeKind
) {
4139 result
= LLVMBuildBitCast(ctx
->ac
.builder
, result
,
4142 result
= LLVMBuildIntToPtr(ctx
->ac
.builder
, result
,
4149 unreachable("Unhandled deref_instr deref type");
4152 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
4155 static void visit_cf_list(struct ac_nir_context
*ctx
,
4156 struct exec_list
*list
);
4158 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
4160 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
4161 nir_foreach_instr(instr
, block
)
4163 switch (instr
->type
) {
4164 case nir_instr_type_alu
:
4165 visit_alu(ctx
, nir_instr_as_alu(instr
));
4167 case nir_instr_type_load_const
:
4168 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
4170 case nir_instr_type_intrinsic
:
4171 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
4173 case nir_instr_type_tex
:
4174 visit_tex(ctx
, nir_instr_as_tex(instr
));
4176 case nir_instr_type_phi
:
4177 visit_phi(ctx
, nir_instr_as_phi(instr
));
4179 case nir_instr_type_ssa_undef
:
4180 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
4182 case nir_instr_type_jump
:
4183 visit_jump(&ctx
->ac
, nir_instr_as_jump(instr
));
4185 case nir_instr_type_deref
:
4186 visit_deref(ctx
, nir_instr_as_deref(instr
));
4189 fprintf(stderr
, "Unknown NIR instr type: ");
4190 nir_print_instr(instr
, stderr
);
4191 fprintf(stderr
, "\n");
4196 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
4199 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
4201 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
4203 nir_block
*then_block
=
4204 (nir_block
*) exec_list_get_head(&if_stmt
->then_list
);
4206 ac_build_uif(&ctx
->ac
, value
, then_block
->index
);
4208 visit_cf_list(ctx
, &if_stmt
->then_list
);
4210 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
4211 nir_block
*else_block
=
4212 (nir_block
*) exec_list_get_head(&if_stmt
->else_list
);
4214 ac_build_else(&ctx
->ac
, else_block
->index
);
4215 visit_cf_list(ctx
, &if_stmt
->else_list
);
4218 ac_build_endif(&ctx
->ac
, then_block
->index
);
4221 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
4223 nir_block
*first_loop_block
=
4224 (nir_block
*) exec_list_get_head(&loop
->body
);
4226 ac_build_bgnloop(&ctx
->ac
, first_loop_block
->index
);
4228 visit_cf_list(ctx
, &loop
->body
);
4230 ac_build_endloop(&ctx
->ac
, first_loop_block
->index
);
4233 static void visit_cf_list(struct ac_nir_context
*ctx
,
4234 struct exec_list
*list
)
4236 foreach_list_typed(nir_cf_node
, node
, node
, list
)
4238 switch (node
->type
) {
4239 case nir_cf_node_block
:
4240 visit_block(ctx
, nir_cf_node_as_block(node
));
4243 case nir_cf_node_if
:
4244 visit_if(ctx
, nir_cf_node_as_if(node
));
4247 case nir_cf_node_loop
:
4248 visit_loop(ctx
, nir_cf_node_as_loop(node
));
4258 ac_handle_shader_output_decl(struct ac_llvm_context
*ctx
,
4259 struct ac_shader_abi
*abi
,
4260 struct nir_shader
*nir
,
4261 struct nir_variable
*variable
,
4262 gl_shader_stage stage
)
4264 unsigned output_loc
= variable
->data
.driver_location
/ 4;
4265 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
4267 /* tess ctrl has it's own load/store paths for outputs */
4268 if (stage
== MESA_SHADER_TESS_CTRL
)
4271 if (stage
== MESA_SHADER_VERTEX
||
4272 stage
== MESA_SHADER_TESS_EVAL
||
4273 stage
== MESA_SHADER_GEOMETRY
) {
4274 int idx
= variable
->data
.location
+ variable
->data
.index
;
4275 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
4276 int length
= nir
->info
.clip_distance_array_size
+
4277 nir
->info
.cull_distance_array_size
;
4286 bool is_16bit
= glsl_type_is_16bit(glsl_without_array(variable
->type
));
4287 LLVMTypeRef type
= is_16bit
? ctx
->f16
: ctx
->f32
;
4288 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
4289 for (unsigned chan
= 0; chan
< 4; chan
++) {
4290 abi
->outputs
[ac_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
4291 ac_build_alloca_undef(ctx
, type
, "");
4297 setup_locals(struct ac_nir_context
*ctx
,
4298 struct nir_function
*func
)
4301 ctx
->num_locals
= 0;
4302 nir_foreach_variable(variable
, &func
->impl
->locals
) {
4303 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
4304 variable
->data
.driver_location
= ctx
->num_locals
* 4;
4305 variable
->data
.location_frac
= 0;
4306 ctx
->num_locals
+= attrib_count
;
4308 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
4312 for (i
= 0; i
< ctx
->num_locals
; i
++) {
4313 for (j
= 0; j
< 4; j
++) {
4314 ctx
->locals
[i
* 4 + j
] =
4315 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
4321 setup_shared(struct ac_nir_context
*ctx
,
4322 struct nir_shader
*nir
)
4324 nir_foreach_variable(variable
, &nir
->shared
) {
4325 LLVMValueRef shared
=
4326 LLVMAddGlobalInAddressSpace(
4327 ctx
->ac
.module
, glsl_to_llvm_type(&ctx
->ac
, variable
->type
),
4328 variable
->name
? variable
->name
: "",
4330 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
4334 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
4335 struct nir_shader
*nir
)
4337 struct ac_nir_context ctx
= {};
4338 struct nir_function
*func
;
4343 ctx
.stage
= nir
->info
.stage
;
4345 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
4347 nir_foreach_variable(variable
, &nir
->outputs
)
4348 ac_handle_shader_output_decl(&ctx
.ac
, ctx
.abi
, nir
, variable
,
4351 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4352 _mesa_key_pointer_equal
);
4353 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4354 _mesa_key_pointer_equal
);
4355 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4356 _mesa_key_pointer_equal
);
4358 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
4360 nir_index_ssa_defs(func
->impl
);
4361 ctx
.ssa_defs
= calloc(func
->impl
->ssa_alloc
, sizeof(LLVMValueRef
));
4363 setup_locals(&ctx
, func
);
4365 if (gl_shader_stage_is_compute(nir
->info
.stage
))
4366 setup_shared(&ctx
, nir
);
4368 visit_cf_list(&ctx
, &func
->impl
->body
);
4369 phi_post_pass(&ctx
);
4371 if (!gl_shader_stage_is_compute(nir
->info
.stage
))
4372 ctx
.abi
->emit_outputs(ctx
.abi
, AC_LLVM_MAX_OUTPUTS
,
4377 ralloc_free(ctx
.defs
);
4378 ralloc_free(ctx
.phis
);
4379 ralloc_free(ctx
.vars
);
4383 ac_lower_indirect_derefs(struct nir_shader
*nir
, enum chip_class chip_class
)
4385 /* While it would be nice not to have this flag, we are constrained
4386 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
4389 bool llvm_has_working_vgpr_indexing
= chip_class
<= VI
;
4391 /* TODO: Indirect indexing of GS inputs is unimplemented.
4393 * TCS and TES load inputs directly from LDS or offchip memory, so
4394 * indirect indexing is trivial.
4396 nir_variable_mode indirect_mask
= 0;
4397 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
4398 (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
&&
4399 nir
->info
.stage
!= MESA_SHADER_TESS_EVAL
&&
4400 !llvm_has_working_vgpr_indexing
)) {
4401 indirect_mask
|= nir_var_shader_in
;
4403 if (!llvm_has_working_vgpr_indexing
&&
4404 nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
4405 indirect_mask
|= nir_var_shader_out
;
4407 /* TODO: We shouldn't need to do this, however LLVM isn't currently
4408 * smart enough to handle indirects without causing excess spilling
4409 * causing the gpu to hang.
4411 * See the following thread for more details of the problem:
4412 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
4414 indirect_mask
|= nir_var_function_temp
;
4416 nir_lower_indirect_derefs(nir
, indirect_mask
);
4420 get_inst_tessfactor_writemask(nir_intrinsic_instr
*intrin
)
4422 if (intrin
->intrinsic
!= nir_intrinsic_store_deref
)
4426 nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[0]));
4428 if (var
->data
.mode
!= nir_var_shader_out
)
4431 unsigned writemask
= 0;
4432 const int location
= var
->data
.location
;
4433 unsigned first_component
= var
->data
.location_frac
;
4434 unsigned num_comps
= intrin
->dest
.ssa
.num_components
;
4436 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
)
4437 writemask
= ((1 << (num_comps
+ 1)) - 1) << first_component
;
4438 else if (location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
4439 writemask
= (((1 << (num_comps
+ 1)) - 1) << first_component
) << 4;
4445 scan_tess_ctrl(nir_cf_node
*cf_node
, unsigned *upper_block_tf_writemask
,
4446 unsigned *cond_block_tf_writemask
,
4447 bool *tessfactors_are_def_in_all_invocs
, bool is_nested_cf
)
4449 switch (cf_node
->type
) {
4450 case nir_cf_node_block
: {
4451 nir_block
*block
= nir_cf_node_as_block(cf_node
);
4452 nir_foreach_instr(instr
, block
) {
4453 if (instr
->type
!= nir_instr_type_intrinsic
)
4456 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
4457 if (intrin
->intrinsic
== nir_intrinsic_barrier
) {
4459 /* If we find a barrier in nested control flow put this in the
4460 * too hard basket. In GLSL this is not possible but it is in
4464 *tessfactors_are_def_in_all_invocs
= false;
4468 /* The following case must be prevented:
4469 * gl_TessLevelInner = ...;
4471 * if (gl_InvocationID == 1)
4472 * gl_TessLevelInner = ...;
4474 * If you consider disjoint code segments separated by barriers, each
4475 * such segment that writes tess factor channels should write the same
4476 * channels in all codepaths within that segment.
4478 if (upper_block_tf_writemask
|| cond_block_tf_writemask
) {
4479 /* Accumulate the result: */
4480 *tessfactors_are_def_in_all_invocs
&=
4481 !(*cond_block_tf_writemask
& ~(*upper_block_tf_writemask
));
4483 /* Analyze the next code segment from scratch. */
4484 *upper_block_tf_writemask
= 0;
4485 *cond_block_tf_writemask
= 0;
4488 *upper_block_tf_writemask
|= get_inst_tessfactor_writemask(intrin
);
4493 case nir_cf_node_if
: {
4494 unsigned then_tessfactor_writemask
= 0;
4495 unsigned else_tessfactor_writemask
= 0;
4497 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
4498 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->then_list
) {
4499 scan_tess_ctrl(nested_node
, &then_tessfactor_writemask
,
4500 cond_block_tf_writemask
,
4501 tessfactors_are_def_in_all_invocs
, true);
4504 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->else_list
) {
4505 scan_tess_ctrl(nested_node
, &else_tessfactor_writemask
,
4506 cond_block_tf_writemask
,
4507 tessfactors_are_def_in_all_invocs
, true);
4510 if (then_tessfactor_writemask
|| else_tessfactor_writemask
) {
4511 /* If both statements write the same tess factor channels,
4512 * we can say that the upper block writes them too.
4514 *upper_block_tf_writemask
|= then_tessfactor_writemask
&
4515 else_tessfactor_writemask
;
4516 *cond_block_tf_writemask
|= then_tessfactor_writemask
|
4517 else_tessfactor_writemask
;
4522 case nir_cf_node_loop
: {
4523 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
4524 foreach_list_typed(nir_cf_node
, nested_node
, node
, &loop
->body
) {
4525 scan_tess_ctrl(nested_node
, cond_block_tf_writemask
,
4526 cond_block_tf_writemask
,
4527 tessfactors_are_def_in_all_invocs
, true);
4533 unreachable("unknown cf node type");
4538 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader
*nir
)
4540 assert(nir
->info
.stage
== MESA_SHADER_TESS_CTRL
);
4542 /* The pass works as follows:
4543 * If all codepaths write tess factors, we can say that all
4544 * invocations define tess factors.
4546 * Each tess factor channel is tracked separately.
4548 unsigned main_block_tf_writemask
= 0; /* if main block writes tess factors */
4549 unsigned cond_block_tf_writemask
= 0; /* if cond block writes tess factors */
4551 /* Initial value = true. Here the pass will accumulate results from
4552 * multiple segments surrounded by barriers. If tess factors aren't
4553 * written at all, it's a shader bug and we don't care if this will be
4556 bool tessfactors_are_def_in_all_invocs
= true;
4558 nir_foreach_function(function
, nir
) {
4559 if (function
->impl
) {
4560 foreach_list_typed(nir_cf_node
, node
, node
, &function
->impl
->body
) {
4561 scan_tess_ctrl(node
, &main_block_tf_writemask
,
4562 &cond_block_tf_writemask
,
4563 &tessfactors_are_def_in_all_invocs
,
4569 /* Accumulate the result for the last code segment separated by a
4572 if (main_block_tf_writemask
|| cond_block_tf_writemask
) {
4573 tessfactors_are_def_in_all_invocs
&=
4574 !(cond_block_tf_writemask
& ~main_block_tf_writemask
);
4577 return tessfactors_are_def_in_all_invocs
;