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 instr
->intrinsic
!= nir_intrinsic_ssbo_atomic_comp_swap
) {
1697 /* XXX: The new raw/struct atomic intrinsics are buggy with
1698 * LLVM 8, see r358579.
1700 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1701 params
[arg_count
++] = ctx
->ac
.i32_0
; /* soffset */
1702 params
[arg_count
++] = ctx
->ac
.i32_0
; /* slc */
1704 ac_build_type_name_for_intr(return_type
, type
, sizeof(type
));
1705 snprintf(name
, sizeof(name
),
1706 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op
, type
);
1708 params
[arg_count
++] = ctx
->ac
.i32_0
; /* vindex */
1709 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1710 params
[arg_count
++] = ctx
->ac
.i1false
; /* slc */
1712 assert(return_type
== ctx
->ac
.i32
);
1713 snprintf(name
, sizeof(name
),
1714 "llvm.amdgcn.buffer.atomic.%s", op
);
1717 return ac_build_intrinsic(&ctx
->ac
, name
, return_type
, params
,
1721 static LLVMValueRef
visit_load_buffer(struct ac_nir_context
*ctx
,
1722 const nir_intrinsic_instr
*instr
)
1724 int elem_size_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
1725 int num_components
= instr
->num_components
;
1726 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
1727 unsigned cache_policy
= get_cache_policy(ctx
, access
, false, false);
1729 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
1730 LLVMValueRef rsrc
= ctx
->abi
->load_ssbo(ctx
->abi
,
1731 get_src(ctx
, instr
->src
[0]), false);
1732 LLVMValueRef vindex
= ctx
->ac
.i32_0
;
1734 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
1735 LLVMTypeRef def_elem_type
= num_components
> 1 ? LLVMGetElementType(def_type
) : def_type
;
1737 LLVMValueRef results
[4];
1738 for (int i
= 0; i
< num_components
;) {
1739 int num_elems
= num_components
- i
;
1740 if (elem_size_bytes
< 4 && nir_intrinsic_align(instr
) % 4 != 0)
1742 if (num_elems
* elem_size_bytes
> 16)
1743 num_elems
= 16 / elem_size_bytes
;
1744 int load_bytes
= num_elems
* elem_size_bytes
;
1746 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
, i
* elem_size_bytes
, false);
1750 if (load_bytes
== 1) {
1751 ret
= ac_build_tbuffer_load_byte(&ctx
->ac
,
1756 cache_policy
& ac_glc
);
1757 } else if (load_bytes
== 2) {
1758 ret
= ac_build_tbuffer_load_short(&ctx
->ac
,
1763 cache_policy
& ac_glc
);
1765 int num_channels
= util_next_power_of_two(load_bytes
) / 4;
1767 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_channels
,
1768 vindex
, offset
, immoffset
, 0,
1769 cache_policy
& ac_glc
, 0,
1773 LLVMTypeRef byte_vec
= LLVMVectorType(ctx
->ac
.i8
, ac_get_type_size(LLVMTypeOf(ret
)));
1774 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, byte_vec
, "");
1775 ret
= ac_trim_vector(&ctx
->ac
, ret
, load_bytes
);
1777 LLVMTypeRef ret_type
= LLVMVectorType(def_elem_type
, num_elems
);
1778 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ret_type
, "");
1780 for (unsigned j
= 0; j
< num_elems
; j
++) {
1781 results
[i
+ j
] = LLVMBuildExtractElement(ctx
->ac
.builder
, ret
, LLVMConstInt(ctx
->ac
.i32
, j
, false), "");
1786 return ac_build_gather_values(&ctx
->ac
, results
, num_components
);
1789 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
1790 const nir_intrinsic_instr
*instr
)
1793 LLVMValueRef rsrc
= get_src(ctx
, instr
->src
[0]);
1794 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
1795 int num_components
= instr
->num_components
;
1797 if (ctx
->abi
->load_ubo
)
1798 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
1800 if (instr
->dest
.ssa
.bit_size
== 64)
1801 num_components
*= 2;
1803 if (instr
->dest
.ssa
.bit_size
== 16 || instr
->dest
.ssa
.bit_size
== 8) {
1804 unsigned load_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
1805 LLVMValueRef results
[num_components
];
1806 for (unsigned i
= 0; i
< num_components
; ++i
) {
1807 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
,
1810 if (load_bytes
== 1) {
1811 results
[i
] = ac_build_tbuffer_load_byte(&ctx
->ac
,
1818 assert(load_bytes
== 2);
1819 results
[i
] = ac_build_tbuffer_load_short(&ctx
->ac
,
1827 ret
= ac_build_gather_values(&ctx
->ac
, results
, num_components
);
1829 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_components
, NULL
, offset
,
1830 NULL
, 0, false, false, true, true);
1832 ret
= ac_trim_vector(&ctx
->ac
, ret
, num_components
);
1835 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
1836 get_def_type(ctx
, &instr
->dest
.ssa
), "");
1840 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_instr
*instr
,
1841 bool vs_in
, unsigned *vertex_index_out
,
1842 LLVMValueRef
*vertex_index_ref
,
1843 unsigned *const_out
, LLVMValueRef
*indir_out
)
1845 nir_variable
*var
= nir_deref_instr_get_variable(instr
);
1846 nir_deref_path path
;
1847 unsigned idx_lvl
= 1;
1849 nir_deref_path_init(&path
, instr
, NULL
);
1851 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
1852 if (vertex_index_ref
) {
1853 *vertex_index_ref
= get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
);
1854 if (vertex_index_out
)
1855 *vertex_index_out
= 0;
1857 *vertex_index_out
= nir_src_as_uint(path
.path
[idx_lvl
]->arr
.index
);
1862 uint32_t const_offset
= 0;
1863 LLVMValueRef offset
= NULL
;
1865 if (var
->data
.compact
) {
1866 assert(instr
->deref_type
== nir_deref_type_array
);
1867 const_offset
= nir_src_as_uint(instr
->arr
.index
);
1871 for (; path
.path
[idx_lvl
]; ++idx_lvl
) {
1872 const struct glsl_type
*parent_type
= path
.path
[idx_lvl
- 1]->type
;
1873 if (path
.path
[idx_lvl
]->deref_type
== nir_deref_type_struct
) {
1874 unsigned index
= path
.path
[idx_lvl
]->strct
.index
;
1876 for (unsigned i
= 0; i
< index
; i
++) {
1877 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
1878 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
1880 } else if(path
.path
[idx_lvl
]->deref_type
== nir_deref_type_array
) {
1881 unsigned size
= glsl_count_attribute_slots(path
.path
[idx_lvl
]->type
, vs_in
);
1882 LLVMValueRef array_off
= LLVMBuildMul(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, size
, 0),
1883 get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
), "");
1885 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, array_off
, "");
1889 unreachable("Uhandled deref type in get_deref_instr_offset");
1893 nir_deref_path_finish(&path
);
1895 if (const_offset
&& offset
)
1896 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
1897 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
1900 *const_out
= const_offset
;
1901 *indir_out
= offset
;
1904 static LLVMValueRef
load_tess_varyings(struct ac_nir_context
*ctx
,
1905 nir_intrinsic_instr
*instr
,
1908 LLVMValueRef result
;
1909 LLVMValueRef vertex_index
= NULL
;
1910 LLVMValueRef indir_index
= NULL
;
1911 unsigned const_index
= 0;
1913 nir_variable
*var
= nir_deref_instr_get_variable(nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
));
1915 unsigned location
= var
->data
.location
;
1916 unsigned driver_location
= var
->data
.driver_location
;
1917 const bool is_patch
= var
->data
.patch
;
1918 const bool is_compact
= var
->data
.compact
;
1920 get_deref_offset(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
1921 false, NULL
, is_patch
? NULL
: &vertex_index
,
1922 &const_index
, &indir_index
);
1924 LLVMTypeRef dest_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
1926 LLVMTypeRef src_component_type
;
1927 if (LLVMGetTypeKind(dest_type
) == LLVMVectorTypeKind
)
1928 src_component_type
= LLVMGetElementType(dest_type
);
1930 src_component_type
= dest_type
;
1932 result
= ctx
->abi
->load_tess_varyings(ctx
->abi
, src_component_type
,
1933 vertex_index
, indir_index
,
1934 const_index
, location
, driver_location
,
1935 var
->data
.location_frac
,
1936 instr
->num_components
,
1937 is_patch
, is_compact
, load_inputs
);
1938 if (instr
->dest
.ssa
.bit_size
== 16) {
1939 result
= ac_to_integer(&ctx
->ac
, result
);
1940 result
= LLVMBuildTrunc(ctx
->ac
.builder
, result
, dest_type
, "");
1942 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, dest_type
, "");
1946 type_scalar_size_bytes(const struct glsl_type
*type
)
1948 assert(glsl_type_is_vector_or_scalar(type
) ||
1949 glsl_type_is_matrix(type
));
1950 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
1953 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
1954 nir_intrinsic_instr
*instr
)
1956 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
1957 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
1959 LLVMValueRef values
[8];
1961 int ve
= instr
->dest
.ssa
.num_components
;
1963 LLVMValueRef indir_index
;
1965 unsigned const_index
;
1966 unsigned stride
= 4;
1967 int mode
= deref
->mode
;
1970 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
1971 var
->data
.mode
== nir_var_shader_in
;
1972 idx
= var
->data
.driver_location
;
1973 comp
= var
->data
.location_frac
;
1974 mode
= var
->data
.mode
;
1976 get_deref_offset(ctx
, deref
, vs_in
, NULL
, NULL
,
1977 &const_index
, &indir_index
);
1979 if (var
->data
.compact
) {
1981 const_index
+= comp
;
1986 if (instr
->dest
.ssa
.bit_size
== 64 &&
1987 (deref
->mode
== nir_var_shader_in
||
1988 deref
->mode
== nir_var_shader_out
||
1989 deref
->mode
== nir_var_function_temp
))
1993 case nir_var_shader_in
:
1994 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
1995 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
1996 return load_tess_varyings(ctx
, instr
, true);
1999 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
2000 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
2001 LLVMValueRef indir_index
;
2002 unsigned const_index
, vertex_index
;
2003 get_deref_offset(ctx
, deref
, false, &vertex_index
, NULL
,
2004 &const_index
, &indir_index
);
2006 return ctx
->abi
->load_inputs(ctx
->abi
, var
->data
.location
,
2007 var
->data
.driver_location
,
2008 var
->data
.location_frac
,
2009 instr
->num_components
, vertex_index
, const_index
, type
);
2012 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2014 unsigned count
= glsl_count_attribute_slots(
2016 ctx
->stage
== MESA_SHADER_VERTEX
);
2018 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2019 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
2020 stride
, false, true);
2022 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2026 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* stride
];
2029 case nir_var_function_temp
:
2030 for (unsigned chan
= 0; chan
< ve
; chan
++) {
2032 unsigned count
= glsl_count_attribute_slots(
2035 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2036 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2037 stride
, true, true);
2039 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2043 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* stride
], "");
2047 case nir_var_mem_shared
: {
2048 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2049 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
2050 return LLVMBuildBitCast(ctx
->ac
.builder
, val
,
2051 get_def_type(ctx
, &instr
->dest
.ssa
),
2054 case nir_var_shader_out
:
2055 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2056 return load_tess_varyings(ctx
, instr
, false);
2059 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2061 unsigned count
= glsl_count_attribute_slots(
2064 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2065 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2066 stride
, true, true);
2068 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2072 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
2073 ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
],
2078 case nir_var_mem_global
: {
2079 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2080 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2081 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2082 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2084 LLVMTypeRef result_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
2085 if (stride
!= natural_stride
) {
2086 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMGetElementType(result_type
),
2087 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2088 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2090 for (unsigned i
= 0; i
< instr
->dest
.ssa
.num_components
; ++i
) {
2091 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, i
* stride
/ natural_stride
, 0);
2092 values
[i
] = LLVMBuildLoad(ctx
->ac
.builder
,
2093 ac_build_gep_ptr(&ctx
->ac
, address
, offset
), "");
2095 return ac_build_gather_values(&ctx
->ac
, values
, instr
->dest
.ssa
.num_components
);
2097 LLVMTypeRef ptr_type
= LLVMPointerType(result_type
,
2098 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2099 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2100 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
2105 unreachable("unhandle variable mode");
2107 ret
= ac_build_varying_gather_values(&ctx
->ac
, values
, ve
, comp
);
2108 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
2112 visit_store_var(struct ac_nir_context
*ctx
,
2113 nir_intrinsic_instr
*instr
)
2115 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2116 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2118 LLVMValueRef temp_ptr
, value
;
2121 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[1]));
2122 int writemask
= instr
->const_index
[0];
2123 LLVMValueRef indir_index
;
2124 unsigned const_index
;
2127 get_deref_offset(ctx
, deref
, false,
2128 NULL
, NULL
, &const_index
, &indir_index
);
2129 idx
= var
->data
.driver_location
;
2130 comp
= var
->data
.location_frac
;
2132 if (var
->data
.compact
) {
2133 const_index
+= comp
;
2138 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64 &&
2139 (deref
->mode
== nir_var_shader_out
||
2140 deref
->mode
== nir_var_function_temp
)) {
2142 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2143 LLVMVectorType(ctx
->ac
.f32
, ac_get_llvm_num_components(src
) * 2),
2146 writemask
= widen_mask(writemask
, 2);
2149 writemask
= writemask
<< comp
;
2151 switch (deref
->mode
) {
2152 case nir_var_shader_out
:
2154 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2155 LLVMValueRef vertex_index
= NULL
;
2156 LLVMValueRef indir_index
= NULL
;
2157 unsigned const_index
= 0;
2158 const bool is_patch
= var
->data
.patch
;
2160 get_deref_offset(ctx
, deref
, false, NULL
,
2161 is_patch
? NULL
: &vertex_index
,
2162 &const_index
, &indir_index
);
2164 ctx
->abi
->store_tcs_outputs(ctx
->abi
, var
,
2165 vertex_index
, indir_index
,
2166 const_index
, src
, writemask
);
2170 for (unsigned chan
= 0; chan
< 8; chan
++) {
2172 if (!(writemask
& (1 << chan
)))
2175 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- comp
);
2177 if (var
->data
.compact
)
2180 unsigned count
= glsl_count_attribute_slots(
2183 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2184 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2185 stride
, true, true);
2187 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2188 value
, indir_index
, "");
2189 build_store_values_extended(&ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
,
2190 count
, stride
, tmp_vec
);
2193 temp_ptr
= ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
];
2195 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2199 case nir_var_function_temp
:
2200 for (unsigned chan
= 0; chan
< 8; chan
++) {
2201 if (!(writemask
& (1 << chan
)))
2204 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
2206 unsigned count
= glsl_count_attribute_slots(
2209 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2210 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2213 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2214 value
, indir_index
, "");
2215 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
2218 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
2220 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2225 case nir_var_mem_global
:
2226 case nir_var_mem_shared
: {
2227 int writemask
= instr
->const_index
[0];
2228 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2229 LLVMValueRef val
= get_src(ctx
, instr
->src
[1]);
2231 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2232 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2233 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2235 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2236 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2237 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2239 if (writemask
== (1u << ac_get_llvm_num_components(val
)) - 1 &&
2240 stride
== natural_stride
) {
2241 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2242 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2243 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2245 val
= LLVMBuildBitCast(ctx
->ac
.builder
, val
,
2246 LLVMGetElementType(LLVMTypeOf(address
)), "");
2247 LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
2249 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMGetElementType(LLVMTypeOf(val
)),
2250 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2251 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2252 for (unsigned chan
= 0; chan
< 4; chan
++) {
2253 if (!(writemask
& (1 << chan
)))
2256 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, chan
* stride
/ natural_stride
, 0);
2258 LLVMValueRef ptr
= ac_build_gep_ptr(&ctx
->ac
, address
, offset
);
2259 LLVMValueRef src
= ac_llvm_extract_elem(&ctx
->ac
, val
,
2261 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2262 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
2263 LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
2274 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
2277 case GLSL_SAMPLER_DIM_BUF
:
2279 case GLSL_SAMPLER_DIM_1D
:
2280 return array
? 2 : 1;
2281 case GLSL_SAMPLER_DIM_2D
:
2282 return array
? 3 : 2;
2283 case GLSL_SAMPLER_DIM_MS
:
2284 return array
? 4 : 3;
2285 case GLSL_SAMPLER_DIM_3D
:
2286 case GLSL_SAMPLER_DIM_CUBE
:
2288 case GLSL_SAMPLER_DIM_RECT
:
2289 case GLSL_SAMPLER_DIM_SUBPASS
:
2291 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
2299 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
2300 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
2301 LLVMValueRef coord_z
,
2302 LLVMValueRef sample_index
,
2303 LLVMValueRef fmask_desc_ptr
)
2305 unsigned sample_chan
= coord_z
? 3 : 2;
2306 LLVMValueRef addr
[4] = {coord_x
, coord_y
, coord_z
};
2307 addr
[sample_chan
] = sample_index
;
2309 ac_apply_fmask_to_sample(ctx
, fmask_desc_ptr
, addr
, coord_z
!= NULL
);
2310 return addr
[sample_chan
];
2313 static nir_deref_instr
*get_image_deref(const nir_intrinsic_instr
*instr
)
2315 assert(instr
->src
[0].is_ssa
);
2316 return nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2319 static LLVMValueRef
get_image_descriptor(struct ac_nir_context
*ctx
,
2320 const nir_intrinsic_instr
*instr
,
2321 enum ac_descriptor_type desc_type
,
2324 nir_deref_instr
*deref_instr
=
2325 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_deref
?
2326 nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
) : NULL
;
2328 return get_sampler_desc(ctx
, deref_instr
, desc_type
, &instr
->instr
, true, write
);
2331 static void get_image_coords(struct ac_nir_context
*ctx
,
2332 const nir_intrinsic_instr
*instr
,
2333 struct ac_image_args
*args
,
2334 enum glsl_sampler_dim dim
,
2337 LLVMValueRef src0
= get_src(ctx
, instr
->src
[1]);
2338 LLVMValueRef masks
[] = {
2339 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
2340 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
2342 LLVMValueRef sample_index
= ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
2345 bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
2346 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2347 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
2348 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2349 bool gfx9_1d
= ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
;
2350 count
= image_type_to_components_count(dim
, is_array
);
2352 if (is_ms
&& (instr
->intrinsic
== nir_intrinsic_image_deref_load
||
2353 instr
->intrinsic
== nir_intrinsic_bindless_image_load
)) {
2354 LLVMValueRef fmask_load_address
[3];
2357 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2358 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
2360 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
2362 fmask_load_address
[2] = NULL
;
2364 for (chan
= 0; chan
< 2; ++chan
)
2365 fmask_load_address
[chan
] =
2366 LLVMBuildAdd(ctx
->ac
.builder
, fmask_load_address
[chan
],
2367 LLVMBuildFPToUI(ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
2368 ctx
->ac
.i32
, ""), "");
2369 fmask_load_address
[2] = ac_to_integer(&ctx
->ac
, ctx
->abi
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
2371 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
2372 fmask_load_address
[0],
2373 fmask_load_address
[1],
2374 fmask_load_address
[2],
2376 get_sampler_desc(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
2377 AC_DESC_FMASK
, &instr
->instr
, false, false));
2379 if (count
== 1 && !gfx9_1d
) {
2380 if (instr
->src
[1].ssa
->num_components
)
2381 args
->coords
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2383 args
->coords
[0] = src0
;
2388 for (chan
= 0; chan
< count
; ++chan
) {
2389 args
->coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src0
, chan
);
2392 for (chan
= 0; chan
< 2; ++chan
) {
2393 args
->coords
[chan
] = LLVMBuildAdd(
2394 ctx
->ac
.builder
, args
->coords
[chan
],
2396 ctx
->ac
.builder
, ctx
->abi
->frag_pos
[chan
],
2397 ctx
->ac
.i32
, ""), "");
2399 args
->coords
[2] = ac_to_integer(&ctx
->ac
,
2400 ctx
->abi
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)]);
2406 args
->coords
[2] = args
->coords
[1];
2407 args
->coords
[1] = ctx
->ac
.i32_0
;
2409 args
->coords
[1] = ctx
->ac
.i32_0
;
2414 args
->coords
[count
] = sample_index
;
2420 static LLVMValueRef
get_image_buffer_descriptor(struct ac_nir_context
*ctx
,
2421 const nir_intrinsic_instr
*instr
, bool write
)
2423 LLVMValueRef rsrc
= get_image_descriptor(ctx
, instr
, AC_DESC_BUFFER
, write
);
2424 if (ctx
->abi
->gfx9_stride_size_workaround
) {
2425 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2426 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 1, 0), "");
2427 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
, LLVMConstInt(ctx
->ac
.i32
, 16, 0), "");
2429 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->ac
.builder
,
2430 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntUGT
, elem_count
, stride
, ""),
2431 elem_count
, stride
, "");
2433 rsrc
= LLVMBuildInsertElement(ctx
->ac
.builder
, rsrc
, new_elem_count
,
2434 LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2439 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
2440 const nir_intrinsic_instr
*instr
,
2445 enum glsl_sampler_dim dim
;
2446 enum gl_access_qualifier access
;
2449 dim
= nir_intrinsic_image_dim(instr
);
2450 access
= nir_intrinsic_access(instr
);
2451 is_array
= nir_intrinsic_image_array(instr
);
2453 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2454 const struct glsl_type
*type
= image_deref
->type
;
2455 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2456 dim
= glsl_get_sampler_dim(type
);
2457 access
= var
->data
.image
.access
;
2458 is_array
= glsl_sampler_type_is_array(type
);
2461 struct ac_image_args args
= {};
2463 args
.cache_policy
= get_cache_policy(ctx
, access
, false, false);
2465 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2466 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
2467 unsigned num_channels
= util_last_bit(mask
);
2468 LLVMValueRef rsrc
, vindex
;
2470 rsrc
= get_image_buffer_descriptor(ctx
, instr
, false);
2471 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
2474 /* TODO: set "can_speculate" when OpenGL needs it. */
2475 res
= ac_build_buffer_load_format(&ctx
->ac
, rsrc
, vindex
,
2476 ctx
->ac
.i32_0
, num_channels
,
2477 !!(args
.cache_policy
& ac_glc
),
2479 res
= ac_build_expand_to_vec4(&ctx
->ac
, res
, num_channels
);
2481 res
= ac_trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
2482 res
= ac_to_integer(&ctx
->ac
, res
);
2484 args
.opcode
= ac_image_load
;
2485 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2486 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2487 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2489 args
.attributes
= AC_FUNC_ATTR_READONLY
;
2491 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
2496 static void visit_image_store(struct ac_nir_context
*ctx
,
2497 nir_intrinsic_instr
*instr
,
2502 enum glsl_sampler_dim dim
;
2503 enum gl_access_qualifier access
;
2506 dim
= nir_intrinsic_image_dim(instr
);
2507 access
= nir_intrinsic_access(instr
);
2508 is_array
= nir_intrinsic_image_array(instr
);
2510 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2511 const struct glsl_type
*type
= image_deref
->type
;
2512 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2513 dim
= glsl_get_sampler_dim(type
);
2514 access
= var
->data
.image
.access
;
2515 is_array
= glsl_sampler_type_is_array(type
);
2518 bool writeonly_memory
= access
& ACCESS_NON_READABLE
;
2519 struct ac_image_args args
= {};
2521 args
.cache_policy
= get_cache_policy(ctx
, access
, true, writeonly_memory
);
2523 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2524 LLVMValueRef rsrc
= get_image_buffer_descriptor(ctx
, instr
, true);
2525 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2526 unsigned src_channels
= ac_get_llvm_num_components(src
);
2527 LLVMValueRef vindex
;
2529 if (src_channels
== 3)
2530 src
= ac_build_expand_to_vec4(&ctx
->ac
, src
, 3);
2532 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
,
2533 get_src(ctx
, instr
->src
[1]),
2536 ac_build_buffer_store_format(&ctx
->ac
, rsrc
, src
, vindex
,
2537 ctx
->ac
.i32_0
, src_channels
,
2538 args
.cache_policy
& ac_glc
,
2541 args
.opcode
= ac_image_store
;
2542 args
.data
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2543 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2544 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, true);
2545 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2548 ac_build_image_opcode(&ctx
->ac
, &args
);
2553 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
2554 const nir_intrinsic_instr
*instr
,
2557 LLVMValueRef params
[7];
2558 int param_count
= 0;
2560 bool cmpswap
= instr
->intrinsic
== nir_intrinsic_image_deref_atomic_comp_swap
||
2561 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_comp_swap
;
2562 const char *atomic_name
;
2563 char intrinsic_name
[64];
2564 enum ac_atomic_op atomic_subop
;
2565 MAYBE_UNUSED
int length
;
2567 enum glsl_sampler_dim dim
;
2568 bool is_unsigned
= false;
2571 if (instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_min
||
2572 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_max
) {
2573 const GLenum format
= nir_intrinsic_format(instr
);
2574 assert(format
== GL_R32UI
|| format
== GL_R32I
);
2575 is_unsigned
= format
== GL_R32UI
;
2577 dim
= nir_intrinsic_image_dim(instr
);
2578 is_array
= nir_intrinsic_image_array(instr
);
2580 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2581 is_unsigned
= glsl_get_sampler_result_type(type
) == GLSL_TYPE_UINT
;
2582 dim
= glsl_get_sampler_dim(type
);
2583 is_array
= glsl_sampler_type_is_array(type
);
2586 switch (instr
->intrinsic
) {
2587 case nir_intrinsic_bindless_image_atomic_add
:
2588 case nir_intrinsic_image_deref_atomic_add
:
2589 atomic_name
= "add";
2590 atomic_subop
= ac_atomic_add
;
2592 case nir_intrinsic_bindless_image_atomic_min
:
2593 case nir_intrinsic_image_deref_atomic_min
:
2594 atomic_name
= is_unsigned
? "umin" : "smin";
2595 atomic_subop
= is_unsigned
? ac_atomic_umin
: ac_atomic_smin
;
2597 case nir_intrinsic_bindless_image_atomic_max
:
2598 case nir_intrinsic_image_deref_atomic_max
:
2599 atomic_name
= is_unsigned
? "umax" : "smax";
2600 atomic_subop
= is_unsigned
? ac_atomic_umax
: ac_atomic_smax
;
2602 case nir_intrinsic_bindless_image_atomic_and
:
2603 case nir_intrinsic_image_deref_atomic_and
:
2604 atomic_name
= "and";
2605 atomic_subop
= ac_atomic_and
;
2607 case nir_intrinsic_bindless_image_atomic_or
:
2608 case nir_intrinsic_image_deref_atomic_or
:
2610 atomic_subop
= ac_atomic_or
;
2612 case nir_intrinsic_bindless_image_atomic_xor
:
2613 case nir_intrinsic_image_deref_atomic_xor
:
2614 atomic_name
= "xor";
2615 atomic_subop
= ac_atomic_xor
;
2617 case nir_intrinsic_bindless_image_atomic_exchange
:
2618 case nir_intrinsic_image_deref_atomic_exchange
:
2619 atomic_name
= "swap";
2620 atomic_subop
= ac_atomic_swap
;
2622 case nir_intrinsic_bindless_image_atomic_comp_swap
:
2623 case nir_intrinsic_image_deref_atomic_comp_swap
:
2624 atomic_name
= "cmpswap";
2625 atomic_subop
= 0; /* not used */
2632 params
[param_count
++] = get_src(ctx
, instr
->src
[4]);
2633 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
2635 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2636 params
[param_count
++] = get_image_buffer_descriptor(ctx
, instr
, true);
2637 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
2638 ctx
->ac
.i32_0
, ""); /* vindex */
2639 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
2640 if (HAVE_LLVM
>= 0x900) {
2641 /* XXX: The new raw/struct atomic intrinsics are buggy
2642 * with LLVM 8, see r358579.
2644 params
[param_count
++] = ctx
->ac
.i32_0
; /* soffset */
2645 params
[param_count
++] = ctx
->ac
.i32_0
; /* slc */
2647 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
2648 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name
);
2650 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
2652 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
2653 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
2656 assert(length
< sizeof(intrinsic_name
));
2657 return ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
,
2658 params
, param_count
, 0);
2660 struct ac_image_args args
= {};
2661 args
.opcode
= cmpswap
? ac_image_atomic_cmpswap
: ac_image_atomic
;
2662 args
.atomic
= atomic_subop
;
2663 args
.data
[0] = params
[0];
2665 args
.data
[1] = params
[1];
2666 get_image_coords(ctx
, instr
, &args
, dim
, is_array
);
2667 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, true);
2668 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2670 return ac_build_image_opcode(&ctx
->ac
, &args
);
2674 static LLVMValueRef
visit_image_samples(struct ac_nir_context
*ctx
,
2675 const nir_intrinsic_instr
*instr
,
2678 enum glsl_sampler_dim dim
;
2681 dim
= nir_intrinsic_image_dim(instr
);
2682 is_array
= nir_intrinsic_image_array(instr
);
2684 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2685 dim
= glsl_get_sampler_dim(type
);
2686 is_array
= glsl_sampler_type_is_array(type
);
2689 struct ac_image_args args
= { 0 };
2690 args
.dim
= get_ac_sampler_dim(&ctx
->ac
, dim
, is_array
);
2692 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2693 args
.opcode
= ac_image_get_resinfo
;
2694 args
.lod
= ctx
->ac
.i32_0
;
2695 args
.attributes
= AC_FUNC_ATTR_READNONE
;
2697 return ac_build_image_opcode(&ctx
->ac
, &args
);
2700 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
2701 const nir_intrinsic_instr
*instr
,
2706 enum glsl_sampler_dim dim
;
2709 dim
= nir_intrinsic_image_dim(instr
);
2710 is_array
= nir_intrinsic_image_array(instr
);
2712 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2713 dim
= glsl_get_sampler_dim(type
);
2714 is_array
= glsl_sampler_type_is_array(type
);
2717 if (dim
== GLSL_SAMPLER_DIM_BUF
)
2718 return get_buffer_size(ctx
, get_image_descriptor(ctx
, instr
, AC_DESC_BUFFER
, false), true);
2720 struct ac_image_args args
= { 0 };
2722 args
.dim
= get_ac_image_dim(&ctx
->ac
, dim
, is_array
);
2724 args
.resource
= get_image_descriptor(ctx
, instr
, AC_DESC_IMAGE
, false);
2725 args
.opcode
= ac_image_get_resinfo
;
2726 args
.lod
= ctx
->ac
.i32_0
;
2727 args
.attributes
= AC_FUNC_ATTR_READNONE
;
2729 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
2731 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
2733 if (dim
== GLSL_SAMPLER_DIM_CUBE
&& is_array
) {
2734 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
2735 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
2736 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
2737 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
2739 if (ctx
->ac
.chip_class
>= GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
&& is_array
) {
2740 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
2741 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, layers
,
2748 static void emit_membar(struct ac_llvm_context
*ac
,
2749 const nir_intrinsic_instr
*instr
)
2751 unsigned waitcnt
= NOOP_WAITCNT
;
2753 switch (instr
->intrinsic
) {
2754 case nir_intrinsic_memory_barrier
:
2755 case nir_intrinsic_group_memory_barrier
:
2756 waitcnt
&= VM_CNT
& LGKM_CNT
;
2758 case nir_intrinsic_memory_barrier_atomic_counter
:
2759 case nir_intrinsic_memory_barrier_buffer
:
2760 case nir_intrinsic_memory_barrier_image
:
2763 case nir_intrinsic_memory_barrier_shared
:
2764 waitcnt
&= LGKM_CNT
;
2769 if (waitcnt
!= NOOP_WAITCNT
)
2770 ac_build_waitcnt(ac
, waitcnt
);
2773 void ac_emit_barrier(struct ac_llvm_context
*ac
, gl_shader_stage stage
)
2775 /* SI only (thanks to a hw bug workaround):
2776 * The real barrier instruction isn’t needed, because an entire patch
2777 * always fits into a single wave.
2779 if (ac
->chip_class
== SI
&& stage
== MESA_SHADER_TESS_CTRL
) {
2780 ac_build_waitcnt(ac
, LGKM_CNT
& VM_CNT
);
2783 ac_build_s_barrier(ac
);
2786 static void emit_discard(struct ac_nir_context
*ctx
,
2787 const nir_intrinsic_instr
*instr
)
2791 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
2792 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2793 get_src(ctx
, instr
->src
[0]),
2796 assert(instr
->intrinsic
== nir_intrinsic_discard
);
2797 cond
= ctx
->ac
.i1false
;
2800 ctx
->abi
->emit_kill(ctx
->abi
, cond
);
2804 visit_load_local_invocation_index(struct ac_nir_context
*ctx
)
2806 LLVMValueRef result
;
2807 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
2808 result
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2809 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
2811 return LLVMBuildAdd(ctx
->ac
.builder
, result
, thread_id
, "");
2815 visit_load_subgroup_id(struct ac_nir_context
*ctx
)
2817 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
2818 LLVMValueRef result
;
2819 result
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2820 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
2821 return LLVMBuildLShr(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 6, false), "");
2823 return LLVMConstInt(ctx
->ac
.i32
, 0, false);
2828 visit_load_num_subgroups(struct ac_nir_context
*ctx
)
2830 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
2831 return LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
->tg_size
,
2832 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
2834 return LLVMConstInt(ctx
->ac
.i32
, 1, false);
2839 visit_first_invocation(struct ac_nir_context
*ctx
)
2841 LLVMValueRef active_set
= ac_build_ballot(&ctx
->ac
, ctx
->ac
.i32_1
);
2843 /* The second argument is whether cttz(0) should be defined, but we do not care. */
2844 LLVMValueRef args
[] = {active_set
, ctx
->ac
.i1false
};
2845 LLVMValueRef result
= ac_build_intrinsic(&ctx
->ac
,
2847 ctx
->ac
.i64
, args
, 2,
2848 AC_FUNC_ATTR_NOUNWIND
|
2849 AC_FUNC_ATTR_READNONE
);
2851 return LLVMBuildTrunc(ctx
->ac
.builder
, result
, ctx
->ac
.i32
, "");
2855 visit_load_shared(struct ac_nir_context
*ctx
,
2856 const nir_intrinsic_instr
*instr
)
2858 LLVMValueRef values
[4], derived_ptr
, index
, ret
;
2860 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0]);
2862 for (int chan
= 0; chan
< instr
->num_components
; chan
++) {
2863 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
2864 derived_ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
2865 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, derived_ptr
, "");
2868 ret
= ac_build_gather_values(&ctx
->ac
, values
, instr
->num_components
);
2869 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
2873 visit_store_shared(struct ac_nir_context
*ctx
,
2874 const nir_intrinsic_instr
*instr
)
2876 LLVMValueRef derived_ptr
, data
,index
;
2877 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2879 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[1]);
2880 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
2882 int writemask
= nir_intrinsic_write_mask(instr
);
2883 for (int chan
= 0; chan
< 4; chan
++) {
2884 if (!(writemask
& (1 << chan
))) {
2887 data
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
2888 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
2889 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
2890 LLVMBuildStore(builder
, data
, derived_ptr
);
2894 static LLVMValueRef
visit_var_atomic(struct ac_nir_context
*ctx
,
2895 const nir_intrinsic_instr
*instr
,
2896 LLVMValueRef ptr
, int src_idx
)
2898 LLVMValueRef result
;
2899 LLVMValueRef src
= get_src(ctx
, instr
->src
[src_idx
]);
2901 if (instr
->intrinsic
== nir_intrinsic_shared_atomic_comp_swap
||
2902 instr
->intrinsic
== nir_intrinsic_deref_atomic_comp_swap
) {
2903 LLVMValueRef src1
= get_src(ctx
, instr
->src
[src_idx
+ 1]);
2904 result
= LLVMBuildAtomicCmpXchg(ctx
->ac
.builder
,
2906 LLVMAtomicOrderingSequentiallyConsistent
,
2907 LLVMAtomicOrderingSequentiallyConsistent
,
2909 result
= LLVMBuildExtractValue(ctx
->ac
.builder
, result
, 0, "");
2911 LLVMAtomicRMWBinOp op
;
2912 switch (instr
->intrinsic
) {
2913 case nir_intrinsic_shared_atomic_add
:
2914 case nir_intrinsic_deref_atomic_add
:
2915 op
= LLVMAtomicRMWBinOpAdd
;
2917 case nir_intrinsic_shared_atomic_umin
:
2918 case nir_intrinsic_deref_atomic_umin
:
2919 op
= LLVMAtomicRMWBinOpUMin
;
2921 case nir_intrinsic_shared_atomic_umax
:
2922 case nir_intrinsic_deref_atomic_umax
:
2923 op
= LLVMAtomicRMWBinOpUMax
;
2925 case nir_intrinsic_shared_atomic_imin
:
2926 case nir_intrinsic_deref_atomic_imin
:
2927 op
= LLVMAtomicRMWBinOpMin
;
2929 case nir_intrinsic_shared_atomic_imax
:
2930 case nir_intrinsic_deref_atomic_imax
:
2931 op
= LLVMAtomicRMWBinOpMax
;
2933 case nir_intrinsic_shared_atomic_and
:
2934 case nir_intrinsic_deref_atomic_and
:
2935 op
= LLVMAtomicRMWBinOpAnd
;
2937 case nir_intrinsic_shared_atomic_or
:
2938 case nir_intrinsic_deref_atomic_or
:
2939 op
= LLVMAtomicRMWBinOpOr
;
2941 case nir_intrinsic_shared_atomic_xor
:
2942 case nir_intrinsic_deref_atomic_xor
:
2943 op
= LLVMAtomicRMWBinOpXor
;
2945 case nir_intrinsic_shared_atomic_exchange
:
2946 case nir_intrinsic_deref_atomic_exchange
:
2947 op
= LLVMAtomicRMWBinOpXchg
;
2953 result
= LLVMBuildAtomicRMW(ctx
->ac
.builder
, op
, ptr
, ac_to_integer(&ctx
->ac
, src
),
2954 LLVMAtomicOrderingSequentiallyConsistent
,
2960 static LLVMValueRef
load_sample_pos(struct ac_nir_context
*ctx
)
2962 LLVMValueRef values
[2];
2963 LLVMValueRef pos
[2];
2965 pos
[0] = ac_to_float(&ctx
->ac
, ctx
->abi
->frag_pos
[0]);
2966 pos
[1] = ac_to_float(&ctx
->ac
, ctx
->abi
->frag_pos
[1]);
2968 values
[0] = ac_build_fract(&ctx
->ac
, pos
[0], 32);
2969 values
[1] = ac_build_fract(&ctx
->ac
, pos
[1], 32);
2970 return ac_build_gather_values(&ctx
->ac
, values
, 2);
2973 static LLVMValueRef
visit_interp(struct ac_nir_context
*ctx
,
2974 const nir_intrinsic_instr
*instr
)
2976 LLVMValueRef result
[4];
2977 LLVMValueRef interp_param
;
2980 LLVMValueRef src_c0
= NULL
;
2981 LLVMValueRef src_c1
= NULL
;
2982 LLVMValueRef src0
= NULL
;
2984 nir_deref_instr
*deref_instr
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2985 nir_variable
*var
= nir_deref_instr_get_variable(deref_instr
);
2986 int input_base
= ctx
->abi
->fs_input_attr_indices
[var
->data
.location
- VARYING_SLOT_VAR0
];
2987 switch (instr
->intrinsic
) {
2988 case nir_intrinsic_interp_deref_at_centroid
:
2989 location
= INTERP_CENTROID
;
2991 case nir_intrinsic_interp_deref_at_sample
:
2992 case nir_intrinsic_interp_deref_at_offset
:
2993 location
= INTERP_CENTER
;
2994 src0
= get_src(ctx
, instr
->src
[1]);
3000 if (instr
->intrinsic
== nir_intrinsic_interp_deref_at_offset
) {
3001 src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_0
, ""));
3002 src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, ctx
->ac
.i32_1
, ""));
3003 } else if (instr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
) {
3004 LLVMValueRef sample_position
;
3005 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
3007 /* fetch sample ID */
3008 sample_position
= ctx
->abi
->load_sample_position(ctx
->abi
, src0
);
3010 src_c0
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_0
, "");
3011 src_c0
= LLVMBuildFSub(ctx
->ac
.builder
, src_c0
, halfval
, "");
3012 src_c1
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
, ctx
->ac
.i32_1
, "");
3013 src_c1
= LLVMBuildFSub(ctx
->ac
.builder
, src_c1
, halfval
, "");
3015 interp_param
= ctx
->abi
->lookup_interp_param(ctx
->abi
, var
->data
.interpolation
, location
);
3017 if (location
== INTERP_CENTER
) {
3018 LLVMValueRef ij_out
[2];
3019 LLVMValueRef ddxy_out
= ac_build_ddxy_interp(&ctx
->ac
, interp_param
);
3022 * take the I then J parameters, and the DDX/Y for it, and
3023 * calculate the IJ inputs for the interpolator.
3024 * temp1 = ddx * offset/sample.x + I;
3025 * interp_param.I = ddy * offset/sample.y + temp1;
3026 * temp1 = ddx * offset/sample.x + J;
3027 * interp_param.J = ddy * offset/sample.y + temp1;
3029 for (unsigned i
= 0; i
< 2; i
++) {
3030 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
3031 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
3032 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3033 ddxy_out
, ix_ll
, "");
3034 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3035 ddxy_out
, iy_ll
, "");
3036 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3037 interp_param
, ix_ll
, "");
3038 LLVMValueRef temp1
, temp2
;
3040 interp_el
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_el
,
3043 temp1
= ac_build_fmad(&ctx
->ac
, ddx_el
, src_c0
, interp_el
);
3044 temp2
= ac_build_fmad(&ctx
->ac
, ddy_el
, src_c1
, temp1
);
3046 ij_out
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
,
3047 temp2
, ctx
->ac
.i32
, "");
3049 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
3053 LLVMValueRef attrib_idx
= ctx
->ac
.i32_0
;
3054 while(deref_instr
->deref_type
!= nir_deref_type_var
) {
3055 if (deref_instr
->deref_type
== nir_deref_type_array
) {
3056 unsigned array_size
= glsl_count_attribute_slots(deref_instr
->type
, false);
3058 LLVMValueRef offset
;
3059 if (nir_src_is_const(deref_instr
->arr
.index
)) {
3060 offset
= LLVMConstInt(ctx
->ac
.i32
, array_size
* nir_src_as_uint(deref_instr
->arr
.index
), false);
3062 LLVMValueRef indirect
= get_src(ctx
, deref_instr
->arr
.index
);
3064 offset
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
3065 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
3068 attrib_idx
= LLVMBuildAdd(ctx
->ac
.builder
, attrib_idx
, offset
, "");
3069 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3070 } else if (deref_instr
->deref_type
== nir_deref_type_struct
) {
3071 LLVMValueRef offset
;
3072 unsigned sidx
= deref_instr
->strct
.index
;
3073 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3074 offset
= LLVMConstInt(ctx
->ac
.i32
, glsl_get_struct_location_offset(deref_instr
->type
, sidx
), false);
3075 attrib_idx
= LLVMBuildAdd(ctx
->ac
.builder
, attrib_idx
, offset
, "");
3077 unreachable("Unsupported deref type");
3082 unsigned attrib_size
= glsl_count_attribute_slots(var
->type
, false);
3083 for (chan
= 0; chan
< 4; chan
++) {
3084 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->ac
.f32
, attrib_size
));
3085 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
3087 for (unsigned idx
= 0; idx
< attrib_size
; ++idx
) {
3088 LLVMValueRef v
, attr_number
;
3090 attr_number
= LLVMConstInt(ctx
->ac
.i32
, input_base
+ idx
, false);
3092 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
,
3093 interp_param
, ctx
->ac
.v2f32
, "");
3094 LLVMValueRef i
= LLVMBuildExtractElement(
3095 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_0
, "");
3096 LLVMValueRef j
= LLVMBuildExtractElement(
3097 ctx
->ac
.builder
, interp_param
, ctx
->ac
.i32_1
, "");
3099 v
= ac_build_fs_interp(&ctx
->ac
, llvm_chan
, attr_number
,
3100 ctx
->abi
->prim_mask
, i
, j
);
3102 v
= ac_build_fs_interp_mov(&ctx
->ac
, LLVMConstInt(ctx
->ac
.i32
, 2, false),
3103 llvm_chan
, attr_number
, ctx
->abi
->prim_mask
);
3106 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
, gather
, v
,
3107 LLVMConstInt(ctx
->ac
.i32
, idx
, false), "");
3110 result
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, gather
, attrib_idx
, "");
3113 return ac_build_varying_gather_values(&ctx
->ac
, result
, instr
->num_components
,
3114 var
->data
.location_frac
);
3117 static void visit_intrinsic(struct ac_nir_context
*ctx
,
3118 nir_intrinsic_instr
*instr
)
3120 LLVMValueRef result
= NULL
;
3122 switch (instr
->intrinsic
) {
3123 case nir_intrinsic_ballot
:
3124 result
= ac_build_ballot(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3126 case nir_intrinsic_read_invocation
:
3127 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3128 get_src(ctx
, instr
->src
[1]));
3130 case nir_intrinsic_read_first_invocation
:
3131 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), NULL
);
3133 case nir_intrinsic_load_subgroup_invocation
:
3134 result
= ac_get_thread_id(&ctx
->ac
);
3136 case nir_intrinsic_load_work_group_id
: {
3137 LLVMValueRef values
[3];
3139 for (int i
= 0; i
< 3; i
++) {
3140 values
[i
] = ctx
->abi
->workgroup_ids
[i
] ?
3141 ctx
->abi
->workgroup_ids
[i
] : ctx
->ac
.i32_0
;
3144 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
3147 case nir_intrinsic_load_base_vertex
:
3148 case nir_intrinsic_load_first_vertex
:
3149 result
= ctx
->abi
->load_base_vertex(ctx
->abi
);
3151 case nir_intrinsic_load_local_group_size
:
3152 result
= ctx
->abi
->load_local_group_size(ctx
->abi
);
3154 case nir_intrinsic_load_vertex_id
:
3155 result
= LLVMBuildAdd(ctx
->ac
.builder
, ctx
->abi
->vertex_id
,
3156 ctx
->abi
->base_vertex
, "");
3158 case nir_intrinsic_load_vertex_id_zero_base
: {
3159 result
= ctx
->abi
->vertex_id
;
3162 case nir_intrinsic_load_local_invocation_id
: {
3163 result
= ctx
->abi
->local_invocation_ids
;
3166 case nir_intrinsic_load_base_instance
:
3167 result
= ctx
->abi
->start_instance
;
3169 case nir_intrinsic_load_draw_id
:
3170 result
= ctx
->abi
->draw_id
;
3172 case nir_intrinsic_load_view_index
:
3173 result
= ctx
->abi
->view_index
;
3175 case nir_intrinsic_load_invocation_id
:
3176 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
3177 result
= ac_unpack_param(&ctx
->ac
, ctx
->abi
->tcs_rel_ids
, 8, 5);
3179 result
= ctx
->abi
->gs_invocation_id
;
3181 case nir_intrinsic_load_primitive_id
:
3182 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3183 result
= ctx
->abi
->gs_prim_id
;
3184 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3185 result
= ctx
->abi
->tcs_patch_id
;
3186 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3187 result
= ctx
->abi
->tes_patch_id
;
3189 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
3191 case nir_intrinsic_load_sample_id
:
3192 result
= ac_unpack_param(&ctx
->ac
, ctx
->abi
->ancillary
, 8, 4);
3194 case nir_intrinsic_load_sample_pos
:
3195 result
= load_sample_pos(ctx
);
3197 case nir_intrinsic_load_sample_mask_in
:
3198 result
= ctx
->abi
->load_sample_mask_in(ctx
->abi
);
3200 case nir_intrinsic_load_frag_coord
: {
3201 LLVMValueRef values
[4] = {
3202 ctx
->abi
->frag_pos
[0],
3203 ctx
->abi
->frag_pos
[1],
3204 ctx
->abi
->frag_pos
[2],
3205 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, ctx
->abi
->frag_pos
[3])
3207 result
= ac_to_integer(&ctx
->ac
,
3208 ac_build_gather_values(&ctx
->ac
, values
, 4));
3211 case nir_intrinsic_load_front_face
:
3212 result
= ctx
->abi
->front_face
;
3214 case nir_intrinsic_load_helper_invocation
:
3215 result
= ac_build_load_helper_invocation(&ctx
->ac
);
3217 case nir_intrinsic_load_instance_id
:
3218 result
= ctx
->abi
->instance_id
;
3220 case nir_intrinsic_load_num_work_groups
:
3221 result
= ctx
->abi
->num_work_groups
;
3223 case nir_intrinsic_load_local_invocation_index
:
3224 result
= visit_load_local_invocation_index(ctx
);
3226 case nir_intrinsic_load_subgroup_id
:
3227 result
= visit_load_subgroup_id(ctx
);
3229 case nir_intrinsic_load_num_subgroups
:
3230 result
= visit_load_num_subgroups(ctx
);
3232 case nir_intrinsic_first_invocation
:
3233 result
= visit_first_invocation(ctx
);
3235 case nir_intrinsic_load_push_constant
:
3236 result
= visit_load_push_constant(ctx
, instr
);
3238 case nir_intrinsic_vulkan_resource_index
: {
3239 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
3240 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
3241 unsigned binding
= nir_intrinsic_binding(instr
);
3243 result
= ctx
->abi
->load_resource(ctx
->abi
, index
, desc_set
,
3247 case nir_intrinsic_vulkan_resource_reindex
:
3248 result
= visit_vulkan_resource_reindex(ctx
, instr
);
3250 case nir_intrinsic_store_ssbo
:
3251 visit_store_ssbo(ctx
, instr
);
3253 case nir_intrinsic_load_ssbo
:
3254 result
= visit_load_buffer(ctx
, instr
);
3256 case nir_intrinsic_ssbo_atomic_add
:
3257 case nir_intrinsic_ssbo_atomic_imin
:
3258 case nir_intrinsic_ssbo_atomic_umin
:
3259 case nir_intrinsic_ssbo_atomic_imax
:
3260 case nir_intrinsic_ssbo_atomic_umax
:
3261 case nir_intrinsic_ssbo_atomic_and
:
3262 case nir_intrinsic_ssbo_atomic_or
:
3263 case nir_intrinsic_ssbo_atomic_xor
:
3264 case nir_intrinsic_ssbo_atomic_exchange
:
3265 case nir_intrinsic_ssbo_atomic_comp_swap
:
3266 result
= visit_atomic_ssbo(ctx
, instr
);
3268 case nir_intrinsic_load_ubo
:
3269 result
= visit_load_ubo_buffer(ctx
, instr
);
3271 case nir_intrinsic_get_buffer_size
:
3272 result
= visit_get_buffer_size(ctx
, instr
);
3274 case nir_intrinsic_load_deref
:
3275 result
= visit_load_var(ctx
, instr
);
3277 case nir_intrinsic_store_deref
:
3278 visit_store_var(ctx
, instr
);
3280 case nir_intrinsic_load_shared
:
3281 result
= visit_load_shared(ctx
, instr
);
3283 case nir_intrinsic_store_shared
:
3284 visit_store_shared(ctx
, instr
);
3286 case nir_intrinsic_bindless_image_samples
:
3287 result
= visit_image_samples(ctx
, instr
, true);
3289 case nir_intrinsic_image_deref_samples
:
3290 result
= visit_image_samples(ctx
, instr
, false);
3292 case nir_intrinsic_bindless_image_load
:
3293 result
= visit_image_load(ctx
, instr
, true);
3295 case nir_intrinsic_image_deref_load
:
3296 result
= visit_image_load(ctx
, instr
, false);
3298 case nir_intrinsic_bindless_image_store
:
3299 visit_image_store(ctx
, instr
, true);
3301 case nir_intrinsic_image_deref_store
:
3302 visit_image_store(ctx
, instr
, false);
3304 case nir_intrinsic_bindless_image_atomic_add
:
3305 case nir_intrinsic_bindless_image_atomic_min
:
3306 case nir_intrinsic_bindless_image_atomic_max
:
3307 case nir_intrinsic_bindless_image_atomic_and
:
3308 case nir_intrinsic_bindless_image_atomic_or
:
3309 case nir_intrinsic_bindless_image_atomic_xor
:
3310 case nir_intrinsic_bindless_image_atomic_exchange
:
3311 case nir_intrinsic_bindless_image_atomic_comp_swap
:
3312 result
= visit_image_atomic(ctx
, instr
, true);
3314 case nir_intrinsic_image_deref_atomic_add
:
3315 case nir_intrinsic_image_deref_atomic_min
:
3316 case nir_intrinsic_image_deref_atomic_max
:
3317 case nir_intrinsic_image_deref_atomic_and
:
3318 case nir_intrinsic_image_deref_atomic_or
:
3319 case nir_intrinsic_image_deref_atomic_xor
:
3320 case nir_intrinsic_image_deref_atomic_exchange
:
3321 case nir_intrinsic_image_deref_atomic_comp_swap
:
3322 result
= visit_image_atomic(ctx
, instr
, false);
3324 case nir_intrinsic_bindless_image_size
:
3325 result
= visit_image_size(ctx
, instr
, true);
3327 case nir_intrinsic_image_deref_size
:
3328 result
= visit_image_size(ctx
, instr
, false);
3330 case nir_intrinsic_shader_clock
:
3331 result
= ac_build_shader_clock(&ctx
->ac
);
3333 case nir_intrinsic_discard
:
3334 case nir_intrinsic_discard_if
:
3335 emit_discard(ctx
, instr
);
3337 case nir_intrinsic_memory_barrier
:
3338 case nir_intrinsic_group_memory_barrier
:
3339 case nir_intrinsic_memory_barrier_atomic_counter
:
3340 case nir_intrinsic_memory_barrier_buffer
:
3341 case nir_intrinsic_memory_barrier_image
:
3342 case nir_intrinsic_memory_barrier_shared
:
3343 emit_membar(&ctx
->ac
, instr
);
3345 case nir_intrinsic_barrier
:
3346 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
3348 case nir_intrinsic_shared_atomic_add
:
3349 case nir_intrinsic_shared_atomic_imin
:
3350 case nir_intrinsic_shared_atomic_umin
:
3351 case nir_intrinsic_shared_atomic_imax
:
3352 case nir_intrinsic_shared_atomic_umax
:
3353 case nir_intrinsic_shared_atomic_and
:
3354 case nir_intrinsic_shared_atomic_or
:
3355 case nir_intrinsic_shared_atomic_xor
:
3356 case nir_intrinsic_shared_atomic_exchange
:
3357 case nir_intrinsic_shared_atomic_comp_swap
: {
3358 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0]);
3359 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3362 case nir_intrinsic_deref_atomic_add
:
3363 case nir_intrinsic_deref_atomic_imin
:
3364 case nir_intrinsic_deref_atomic_umin
:
3365 case nir_intrinsic_deref_atomic_imax
:
3366 case nir_intrinsic_deref_atomic_umax
:
3367 case nir_intrinsic_deref_atomic_and
:
3368 case nir_intrinsic_deref_atomic_or
:
3369 case nir_intrinsic_deref_atomic_xor
:
3370 case nir_intrinsic_deref_atomic_exchange
:
3371 case nir_intrinsic_deref_atomic_comp_swap
: {
3372 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
3373 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3376 case nir_intrinsic_interp_deref_at_centroid
:
3377 case nir_intrinsic_interp_deref_at_sample
:
3378 case nir_intrinsic_interp_deref_at_offset
:
3379 result
= visit_interp(ctx
, instr
);
3381 case nir_intrinsic_emit_vertex
:
3382 ctx
->abi
->emit_vertex(ctx
->abi
, nir_intrinsic_stream_id(instr
), ctx
->abi
->outputs
);
3384 case nir_intrinsic_end_primitive
:
3385 ctx
->abi
->emit_primitive(ctx
->abi
, nir_intrinsic_stream_id(instr
));
3387 case nir_intrinsic_load_tess_coord
:
3388 result
= ctx
->abi
->load_tess_coord(ctx
->abi
);
3390 case nir_intrinsic_load_tess_level_outer
:
3391 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
);
3393 case nir_intrinsic_load_tess_level_inner
:
3394 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
);
3396 case nir_intrinsic_load_patch_vertices_in
:
3397 result
= ctx
->abi
->load_patch_vertices_in(ctx
->abi
);
3399 case nir_intrinsic_vote_all
: {
3400 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3401 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
3404 case nir_intrinsic_vote_any
: {
3405 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3406 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
3409 case nir_intrinsic_shuffle
:
3410 result
= ac_build_shuffle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3411 get_src(ctx
, instr
->src
[1]));
3413 case nir_intrinsic_reduce
:
3414 result
= ac_build_reduce(&ctx
->ac
,
3415 get_src(ctx
, instr
->src
[0]),
3416 instr
->const_index
[0],
3417 instr
->const_index
[1]);
3419 case nir_intrinsic_inclusive_scan
:
3420 result
= ac_build_inclusive_scan(&ctx
->ac
,
3421 get_src(ctx
, instr
->src
[0]),
3422 instr
->const_index
[0]);
3424 case nir_intrinsic_exclusive_scan
:
3425 result
= ac_build_exclusive_scan(&ctx
->ac
,
3426 get_src(ctx
, instr
->src
[0]),
3427 instr
->const_index
[0]);
3429 case nir_intrinsic_quad_broadcast
: {
3430 unsigned lane
= nir_src_as_uint(instr
->src
[1]);
3431 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3432 lane
, lane
, lane
, lane
);
3435 case nir_intrinsic_quad_swap_horizontal
:
3436 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 1, 0, 3 ,2);
3438 case nir_intrinsic_quad_swap_vertical
:
3439 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 2, 3, 0 ,1);
3441 case nir_intrinsic_quad_swap_diagonal
:
3442 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 3, 2, 1 ,0);
3445 fprintf(stderr
, "Unknown intrinsic: ");
3446 nir_print_instr(&instr
->instr
, stderr
);
3447 fprintf(stderr
, "\n");
3451 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3455 static LLVMValueRef
get_bindless_index_from_uniform(struct ac_nir_context
*ctx
,
3456 unsigned base_index
,
3457 unsigned constant_index
,
3458 LLVMValueRef dynamic_index
)
3460 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, base_index
* 4, 0);
3461 LLVMValueRef index
= LLVMBuildAdd(ctx
->ac
.builder
, dynamic_index
,
3462 LLVMConstInt(ctx
->ac
.i32
, constant_index
, 0), "");
3464 /* Bindless uniforms are 64bit so multiple index by 8 */
3465 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i32
, 8, 0), "");
3466 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
3468 LLVMValueRef ubo_index
= ctx
->abi
->load_ubo(ctx
->abi
, ctx
->ac
.i32_0
);
3470 LLVMValueRef ret
= ac_build_buffer_load(&ctx
->ac
, ubo_index
, 1, NULL
, offset
,
3471 NULL
, 0, false, false, true, true);
3473 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ctx
->ac
.i32
, "");
3476 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
3477 nir_deref_instr
*deref_instr
,
3478 enum ac_descriptor_type desc_type
,
3479 const nir_instr
*instr
,
3480 bool image
, bool write
)
3482 LLVMValueRef index
= NULL
;
3483 unsigned constant_index
= 0;
3484 unsigned descriptor_set
;
3485 unsigned base_index
;
3486 bool bindless
= false;
3491 nir_intrinsic_instr
*img_instr
= nir_instr_as_intrinsic(instr
);
3494 index
= get_src(ctx
, img_instr
->src
[0]);
3496 nir_tex_instr
*tex_instr
= nir_instr_as_tex(instr
);
3497 int sampSrcIdx
= nir_tex_instr_src_index(tex_instr
,
3498 nir_tex_src_sampler_handle
);
3499 if (sampSrcIdx
!= -1) {
3502 index
= get_src(ctx
, tex_instr
->src
[sampSrcIdx
].src
);
3504 assert(tex_instr
&& !image
);
3505 base_index
= tex_instr
->sampler_index
;
3509 while(deref_instr
->deref_type
!= nir_deref_type_var
) {
3510 if (deref_instr
->deref_type
== nir_deref_type_array
) {
3511 unsigned array_size
= glsl_get_aoa_size(deref_instr
->type
);
3515 if (nir_src_is_const(deref_instr
->arr
.index
)) {
3516 constant_index
+= array_size
* nir_src_as_uint(deref_instr
->arr
.index
);
3518 LLVMValueRef indirect
= get_src(ctx
, deref_instr
->arr
.index
);
3520 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
3521 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
3526 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
3529 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3530 } else if (deref_instr
->deref_type
== nir_deref_type_struct
) {
3531 unsigned sidx
= deref_instr
->strct
.index
;
3532 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
3533 constant_index
+= glsl_get_struct_location_offset(deref_instr
->type
, sidx
);
3535 unreachable("Unsupported deref type");
3538 descriptor_set
= deref_instr
->var
->data
.descriptor_set
;
3540 if (deref_instr
->var
->data
.bindless
) {
3541 /* For now just assert on unhandled variable types */
3542 assert(deref_instr
->var
->data
.mode
== nir_var_uniform
);
3544 base_index
= deref_instr
->var
->data
.driver_location
;
3547 index
= index
? index
: ctx
->ac
.i32_0
;
3548 index
= get_bindless_index_from_uniform(ctx
, base_index
,
3549 constant_index
, index
);
3551 base_index
= deref_instr
->var
->data
.binding
;
3554 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
3557 constant_index
, index
,
3558 desc_type
, image
, write
, bindless
);
3561 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3564 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3565 * filtering manually. The driver sets img7 to a mask clearing
3566 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3567 * s_and_b32 samp0, samp0, img7
3570 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3572 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
3573 LLVMValueRef res
, LLVMValueRef samp
)
3575 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3576 LLVMValueRef img7
, samp0
;
3578 if (ctx
->ac
.chip_class
>= VI
)
3581 img7
= LLVMBuildExtractElement(builder
, res
,
3582 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
3583 samp0
= LLVMBuildExtractElement(builder
, samp
,
3584 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
3585 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
3586 return LLVMBuildInsertElement(builder
, samp
, samp0
,
3587 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
3590 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
3591 nir_tex_instr
*instr
,
3592 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
3593 LLVMValueRef
*fmask_ptr
)
3595 nir_deref_instr
*texture_deref_instr
= NULL
;
3596 nir_deref_instr
*sampler_deref_instr
= NULL
;
3598 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3599 switch (instr
->src
[i
].src_type
) {
3600 case nir_tex_src_texture_deref
:
3601 texture_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
3603 case nir_tex_src_sampler_deref
:
3604 sampler_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
3611 if (!sampler_deref_instr
)
3612 sampler_deref_instr
= texture_deref_instr
;
3614 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
3615 *res_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_BUFFER
, &instr
->instr
, false, false);
3617 *res_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_IMAGE
, &instr
->instr
, false, false);
3619 *samp_ptr
= get_sampler_desc(ctx
, sampler_deref_instr
, AC_DESC_SAMPLER
, &instr
->instr
, false, false);
3620 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
3621 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
3623 if (fmask_ptr
&& (instr
->op
== nir_texop_txf_ms
||
3624 instr
->op
== nir_texop_samples_identical
))
3625 *fmask_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_FMASK
, &instr
->instr
, false, false);
3628 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
3631 coord
= ac_to_float(ctx
, coord
);
3632 coord
= ac_build_round(ctx
, coord
);
3633 coord
= ac_to_integer(ctx
, coord
);
3637 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
3639 LLVMValueRef result
= NULL
;
3640 struct ac_image_args args
= { 0 };
3641 LLVMValueRef fmask_ptr
= NULL
, sample_index
= NULL
;
3642 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
3643 unsigned offset_src
= 0;
3645 tex_fetch_ptrs(ctx
, instr
, &args
.resource
, &args
.sampler
, &fmask_ptr
);
3647 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3648 switch (instr
->src
[i
].src_type
) {
3649 case nir_tex_src_coord
: {
3650 LLVMValueRef coord
= get_src(ctx
, instr
->src
[i
].src
);
3651 for (unsigned chan
= 0; chan
< instr
->coord_components
; ++chan
)
3652 args
.coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
3655 case nir_tex_src_projector
:
3657 case nir_tex_src_comparator
:
3658 if (instr
->is_shadow
)
3659 args
.compare
= get_src(ctx
, instr
->src
[i
].src
);
3661 case nir_tex_src_offset
:
3662 args
.offset
= get_src(ctx
, instr
->src
[i
].src
);
3665 case nir_tex_src_bias
:
3666 if (instr
->op
== nir_texop_txb
)
3667 args
.bias
= get_src(ctx
, instr
->src
[i
].src
);
3669 case nir_tex_src_lod
: {
3670 if (nir_src_is_const(instr
->src
[i
].src
) && nir_src_as_uint(instr
->src
[i
].src
) == 0)
3671 args
.level_zero
= true;
3673 args
.lod
= get_src(ctx
, instr
->src
[i
].src
);
3676 case nir_tex_src_ms_index
:
3677 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
3679 case nir_tex_src_ms_mcs
:
3681 case nir_tex_src_ddx
:
3682 ddx
= get_src(ctx
, instr
->src
[i
].src
);
3684 case nir_tex_src_ddy
:
3685 ddy
= get_src(ctx
, instr
->src
[i
].src
);
3687 case nir_tex_src_texture_offset
:
3688 case nir_tex_src_sampler_offset
:
3689 case nir_tex_src_plane
:
3695 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
3696 result
= get_buffer_size(ctx
, args
.resource
, true);
3700 if (instr
->op
== nir_texop_texture_samples
) {
3701 LLVMValueRef res
, samples
, is_msaa
;
3702 res
= LLVMBuildBitCast(ctx
->ac
.builder
, args
.resource
, ctx
->ac
.v8i32
, "");
3703 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
3704 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
3705 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
3706 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
3707 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
3708 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
3709 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
3710 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
3712 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
3713 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
3714 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
3715 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
3716 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
3718 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
3724 if (args
.offset
&& instr
->op
!= nir_texop_txf
) {
3725 LLVMValueRef offset
[3], pack
;
3726 for (unsigned chan
= 0; chan
< 3; ++chan
)
3727 offset
[chan
] = ctx
->ac
.i32_0
;
3729 unsigned num_components
= ac_get_llvm_num_components(args
.offset
);
3730 for (unsigned chan
= 0; chan
< num_components
; chan
++) {
3731 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, args
.offset
, chan
);
3732 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
3733 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
3735 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
3736 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
3738 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
3739 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
3743 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
3744 * so the depth comparison value isn't clamped for Z16 and
3745 * Z24 anymore. Do it manually here.
3747 * It's unnecessary if the original texture format was
3748 * Z32_FLOAT, but we don't know that here.
3750 if (args
.compare
&& ctx
->ac
.chip_class
>= VI
&& ctx
->abi
->clamp_shadow_reference
)
3751 args
.compare
= ac_build_clamp(&ctx
->ac
, ac_to_float(&ctx
->ac
, args
.compare
));
3753 /* pack derivatives */
3755 int num_src_deriv_channels
, num_dest_deriv_channels
;
3756 switch (instr
->sampler_dim
) {
3757 case GLSL_SAMPLER_DIM_3D
:
3758 case GLSL_SAMPLER_DIM_CUBE
:
3759 num_src_deriv_channels
= 3;
3760 num_dest_deriv_channels
= 3;
3762 case GLSL_SAMPLER_DIM_2D
:
3764 num_src_deriv_channels
= 2;
3765 num_dest_deriv_channels
= 2;
3767 case GLSL_SAMPLER_DIM_1D
:
3768 num_src_deriv_channels
= 1;
3769 if (ctx
->ac
.chip_class
>= GFX9
) {
3770 num_dest_deriv_channels
= 2;
3772 num_dest_deriv_channels
= 1;
3777 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
3778 args
.derivs
[i
] = ac_to_float(&ctx
->ac
,
3779 ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
3780 args
.derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
,
3781 ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
3783 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
3784 args
.derivs
[i
] = ctx
->ac
.f32_0
;
3785 args
.derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
3789 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& args
.coords
[0]) {
3790 for (unsigned chan
= 0; chan
< instr
->coord_components
; chan
++)
3791 args
.coords
[chan
] = ac_to_float(&ctx
->ac
, args
.coords
[chan
]);
3792 if (instr
->coord_components
== 3)
3793 args
.coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
3794 ac_prepare_cube_coords(&ctx
->ac
,
3795 instr
->op
== nir_texop_txd
, instr
->is_array
,
3796 instr
->op
== nir_texop_lod
, args
.coords
, args
.derivs
);
3799 /* Texture coordinates fixups */
3800 if (instr
->coord_components
> 1 &&
3801 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3803 instr
->op
!= nir_texop_txf
) {
3804 args
.coords
[1] = apply_round_slice(&ctx
->ac
, args
.coords
[1]);
3807 if (instr
->coord_components
> 2 &&
3808 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
||
3809 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
||
3810 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS
||
3811 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
) &&
3813 instr
->op
!= nir_texop_txf
&& instr
->op
!= nir_texop_txf_ms
) {
3814 args
.coords
[2] = apply_round_slice(&ctx
->ac
, args
.coords
[2]);
3817 if (ctx
->ac
.chip_class
>= GFX9
&&
3818 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3819 instr
->op
!= nir_texop_lod
) {
3820 LLVMValueRef filler
;
3821 if (instr
->op
== nir_texop_txf
)
3822 filler
= ctx
->ac
.i32_0
;
3824 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
3826 if (instr
->is_array
)
3827 args
.coords
[2] = args
.coords
[1];
3828 args
.coords
[1] = filler
;
3831 /* Pack sample index */
3832 if (instr
->op
== nir_texop_txf_ms
&& sample_index
)
3833 args
.coords
[instr
->coord_components
] = sample_index
;
3835 if (instr
->op
== nir_texop_samples_identical
) {
3836 struct ac_image_args txf_args
= { 0 };
3837 memcpy(txf_args
.coords
, args
.coords
, sizeof(txf_args
.coords
));
3839 txf_args
.dmask
= 0xf;
3840 txf_args
.resource
= fmask_ptr
;
3841 txf_args
.dim
= instr
->is_array
? ac_image_2darray
: ac_image_2d
;
3842 result
= build_tex_intrinsic(ctx
, instr
, &txf_args
);
3844 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
3845 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
3849 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
&&
3850 instr
->op
!= nir_texop_txs
) {
3851 unsigned sample_chan
= instr
->is_array
? 3 : 2;
3852 args
.coords
[sample_chan
] = adjust_sample_index_using_fmask(
3853 &ctx
->ac
, args
.coords
[0], args
.coords
[1],
3854 instr
->is_array
? args
.coords
[2] : NULL
,
3855 args
.coords
[sample_chan
], fmask_ptr
);
3858 if (args
.offset
&& instr
->op
== nir_texop_txf
) {
3859 int num_offsets
= instr
->src
[offset_src
].src
.ssa
->num_components
;
3860 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
3861 for (unsigned i
= 0; i
< num_offsets
; ++i
) {
3862 args
.coords
[i
] = LLVMBuildAdd(
3863 ctx
->ac
.builder
, args
.coords
[i
],
3864 LLVMConstInt(ctx
->ac
.i32
, nir_src_comp_as_uint(instr
->src
[offset_src
].src
, i
), false), "");
3869 /* TODO TG4 support */
3871 if (instr
->op
== nir_texop_tg4
) {
3872 if (instr
->is_shadow
)
3875 args
.dmask
= 1 << instr
->component
;
3878 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_BUF
)
3879 args
.dim
= get_ac_sampler_dim(&ctx
->ac
, instr
->sampler_dim
, instr
->is_array
);
3880 result
= build_tex_intrinsic(ctx
, instr
, &args
);
3882 if (instr
->op
== nir_texop_query_levels
)
3883 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
3884 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
3885 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
3886 instr
->op
!= nir_texop_tg4
)
3887 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
3888 else if (instr
->op
== nir_texop_txs
&&
3889 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
3891 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3892 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3893 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
3894 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3895 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
3896 } else if (ctx
->ac
.chip_class
>= GFX9
&&
3897 instr
->op
== nir_texop_txs
&&
3898 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
3900 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3901 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
3902 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
3904 } else if (instr
->dest
.ssa
.num_components
!= 4)
3905 result
= ac_trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
3909 assert(instr
->dest
.is_ssa
);
3910 result
= ac_to_integer(&ctx
->ac
, result
);
3911 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3916 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
3918 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
3919 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
3921 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
3922 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
3925 static void visit_post_phi(struct ac_nir_context
*ctx
,
3926 nir_phi_instr
*instr
,
3927 LLVMValueRef llvm_phi
)
3929 nir_foreach_phi_src(src
, instr
) {
3930 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
3931 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
3933 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
3937 static void phi_post_pass(struct ac_nir_context
*ctx
)
3939 hash_table_foreach(ctx
->phis
, entry
) {
3940 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
3941 (LLVMValueRef
)entry
->data
);
3946 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
3947 const nir_ssa_undef_instr
*instr
)
3949 unsigned num_components
= instr
->def
.num_components
;
3950 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
3953 if (num_components
== 1)
3954 undef
= LLVMGetUndef(type
);
3956 undef
= LLVMGetUndef(LLVMVectorType(type
, num_components
));
3958 ctx
->ssa_defs
[instr
->def
.index
] = undef
;
3961 static void visit_jump(struct ac_llvm_context
*ctx
,
3962 const nir_jump_instr
*instr
)
3964 switch (instr
->type
) {
3965 case nir_jump_break
:
3966 ac_build_break(ctx
);
3968 case nir_jump_continue
:
3969 ac_build_continue(ctx
);
3972 fprintf(stderr
, "Unknown NIR jump instr: ");
3973 nir_print_instr(&instr
->instr
, stderr
);
3974 fprintf(stderr
, "\n");
3980 glsl_base_to_llvm_type(struct ac_llvm_context
*ac
,
3981 enum glsl_base_type type
)
3985 case GLSL_TYPE_UINT
:
3986 case GLSL_TYPE_BOOL
:
3987 case GLSL_TYPE_SUBROUTINE
:
3989 case GLSL_TYPE_INT8
:
3990 case GLSL_TYPE_UINT8
:
3992 case GLSL_TYPE_INT16
:
3993 case GLSL_TYPE_UINT16
:
3995 case GLSL_TYPE_FLOAT
:
3997 case GLSL_TYPE_FLOAT16
:
3999 case GLSL_TYPE_INT64
:
4000 case GLSL_TYPE_UINT64
:
4002 case GLSL_TYPE_DOUBLE
:
4005 unreachable("unknown GLSL type");
4010 glsl_to_llvm_type(struct ac_llvm_context
*ac
,
4011 const struct glsl_type
*type
)
4013 if (glsl_type_is_scalar(type
)) {
4014 return glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
));
4017 if (glsl_type_is_vector(type
)) {
4018 return LLVMVectorType(
4019 glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
)),
4020 glsl_get_vector_elements(type
));
4023 if (glsl_type_is_matrix(type
)) {
4024 return LLVMArrayType(
4025 glsl_to_llvm_type(ac
, glsl_get_column_type(type
)),
4026 glsl_get_matrix_columns(type
));
4029 if (glsl_type_is_array(type
)) {
4030 return LLVMArrayType(
4031 glsl_to_llvm_type(ac
, glsl_get_array_element(type
)),
4032 glsl_get_length(type
));
4035 assert(glsl_type_is_struct_or_ifc(type
));
4037 LLVMTypeRef member_types
[glsl_get_length(type
)];
4039 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
4041 glsl_to_llvm_type(ac
,
4042 glsl_get_struct_field(type
, i
));
4045 return LLVMStructTypeInContext(ac
->context
, member_types
,
4046 glsl_get_length(type
), false);
4049 static void visit_deref(struct ac_nir_context
*ctx
,
4050 nir_deref_instr
*instr
)
4052 if (instr
->mode
!= nir_var_mem_shared
&&
4053 instr
->mode
!= nir_var_mem_global
)
4056 LLVMValueRef result
= NULL
;
4057 switch(instr
->deref_type
) {
4058 case nir_deref_type_var
: {
4059 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, instr
->var
);
4060 result
= entry
->data
;
4063 case nir_deref_type_struct
:
4064 if (instr
->mode
== nir_var_mem_global
) {
4065 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
4066 uint64_t offset
= glsl_get_struct_field_offset(parent
->type
,
4067 instr
->strct
.index
);
4068 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4069 LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
4071 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4072 LLVMConstInt(ctx
->ac
.i32
, instr
->strct
.index
, 0));
4075 case nir_deref_type_array
:
4076 if (instr
->mode
== nir_var_mem_global
) {
4077 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
4078 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
4080 if ((glsl_type_is_matrix(parent
->type
) &&
4081 glsl_matrix_type_is_row_major(parent
->type
)) ||
4082 (glsl_type_is_vector(parent
->type
) && stride
== 0))
4083 stride
= type_scalar_size_bytes(parent
->type
);
4086 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
4087 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
4088 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
4090 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
4092 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
4094 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4095 get_src(ctx
, instr
->arr
.index
));
4098 case nir_deref_type_ptr_as_array
:
4099 if (instr
->mode
== nir_var_mem_global
) {
4100 unsigned stride
= nir_deref_instr_ptr_as_array_stride(instr
);
4102 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
4103 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
4104 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
4106 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
4108 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
4110 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
4111 get_src(ctx
, instr
->arr
.index
));
4114 case nir_deref_type_cast
: {
4115 result
= get_src(ctx
, instr
->parent
);
4117 /* We can't use the structs from LLVM because the shader
4118 * specifies its own offsets. */
4119 LLVMTypeRef pointee_type
= ctx
->ac
.i8
;
4120 if (instr
->mode
== nir_var_mem_shared
)
4121 pointee_type
= glsl_to_llvm_type(&ctx
->ac
, instr
->type
);
4123 unsigned address_space
;
4125 switch(instr
->mode
) {
4126 case nir_var_mem_shared
:
4127 address_space
= AC_ADDR_SPACE_LDS
;
4129 case nir_var_mem_global
:
4130 address_space
= AC_ADDR_SPACE_GLOBAL
;
4133 unreachable("Unhandled address space");
4136 LLVMTypeRef type
= LLVMPointerType(pointee_type
, address_space
);
4138 if (LLVMTypeOf(result
) != type
) {
4139 if (LLVMGetTypeKind(LLVMTypeOf(result
)) == LLVMVectorTypeKind
) {
4140 result
= LLVMBuildBitCast(ctx
->ac
.builder
, result
,
4143 result
= LLVMBuildIntToPtr(ctx
->ac
.builder
, result
,
4150 unreachable("Unhandled deref_instr deref type");
4153 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
4156 static void visit_cf_list(struct ac_nir_context
*ctx
,
4157 struct exec_list
*list
);
4159 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
4161 LLVMBasicBlockRef llvm_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
4162 nir_foreach_instr(instr
, block
)
4164 switch (instr
->type
) {
4165 case nir_instr_type_alu
:
4166 visit_alu(ctx
, nir_instr_as_alu(instr
));
4168 case nir_instr_type_load_const
:
4169 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
4171 case nir_instr_type_intrinsic
:
4172 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
4174 case nir_instr_type_tex
:
4175 visit_tex(ctx
, nir_instr_as_tex(instr
));
4177 case nir_instr_type_phi
:
4178 visit_phi(ctx
, nir_instr_as_phi(instr
));
4180 case nir_instr_type_ssa_undef
:
4181 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
4183 case nir_instr_type_jump
:
4184 visit_jump(&ctx
->ac
, nir_instr_as_jump(instr
));
4186 case nir_instr_type_deref
:
4187 visit_deref(ctx
, nir_instr_as_deref(instr
));
4190 fprintf(stderr
, "Unknown NIR instr type: ");
4191 nir_print_instr(instr
, stderr
);
4192 fprintf(stderr
, "\n");
4197 _mesa_hash_table_insert(ctx
->defs
, block
, llvm_block
);
4200 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
4202 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
4204 nir_block
*then_block
=
4205 (nir_block
*) exec_list_get_head(&if_stmt
->then_list
);
4207 ac_build_uif(&ctx
->ac
, value
, then_block
->index
);
4209 visit_cf_list(ctx
, &if_stmt
->then_list
);
4211 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
4212 nir_block
*else_block
=
4213 (nir_block
*) exec_list_get_head(&if_stmt
->else_list
);
4215 ac_build_else(&ctx
->ac
, else_block
->index
);
4216 visit_cf_list(ctx
, &if_stmt
->else_list
);
4219 ac_build_endif(&ctx
->ac
, then_block
->index
);
4222 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
4224 nir_block
*first_loop_block
=
4225 (nir_block
*) exec_list_get_head(&loop
->body
);
4227 ac_build_bgnloop(&ctx
->ac
, first_loop_block
->index
);
4229 visit_cf_list(ctx
, &loop
->body
);
4231 ac_build_endloop(&ctx
->ac
, first_loop_block
->index
);
4234 static void visit_cf_list(struct ac_nir_context
*ctx
,
4235 struct exec_list
*list
)
4237 foreach_list_typed(nir_cf_node
, node
, node
, list
)
4239 switch (node
->type
) {
4240 case nir_cf_node_block
:
4241 visit_block(ctx
, nir_cf_node_as_block(node
));
4244 case nir_cf_node_if
:
4245 visit_if(ctx
, nir_cf_node_as_if(node
));
4248 case nir_cf_node_loop
:
4249 visit_loop(ctx
, nir_cf_node_as_loop(node
));
4259 ac_handle_shader_output_decl(struct ac_llvm_context
*ctx
,
4260 struct ac_shader_abi
*abi
,
4261 struct nir_shader
*nir
,
4262 struct nir_variable
*variable
,
4263 gl_shader_stage stage
)
4265 unsigned output_loc
= variable
->data
.driver_location
/ 4;
4266 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
4268 /* tess ctrl has it's own load/store paths for outputs */
4269 if (stage
== MESA_SHADER_TESS_CTRL
)
4272 if (stage
== MESA_SHADER_VERTEX
||
4273 stage
== MESA_SHADER_TESS_EVAL
||
4274 stage
== MESA_SHADER_GEOMETRY
) {
4275 int idx
= variable
->data
.location
+ variable
->data
.index
;
4276 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
4277 int length
= nir
->info
.clip_distance_array_size
+
4278 nir
->info
.cull_distance_array_size
;
4287 bool is_16bit
= glsl_type_is_16bit(glsl_without_array(variable
->type
));
4288 LLVMTypeRef type
= is_16bit
? ctx
->f16
: ctx
->f32
;
4289 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
4290 for (unsigned chan
= 0; chan
< 4; chan
++) {
4291 abi
->outputs
[ac_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
4292 ac_build_alloca_undef(ctx
, type
, "");
4298 setup_locals(struct ac_nir_context
*ctx
,
4299 struct nir_function
*func
)
4302 ctx
->num_locals
= 0;
4303 nir_foreach_variable(variable
, &func
->impl
->locals
) {
4304 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
4305 variable
->data
.driver_location
= ctx
->num_locals
* 4;
4306 variable
->data
.location_frac
= 0;
4307 ctx
->num_locals
+= attrib_count
;
4309 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
4313 for (i
= 0; i
< ctx
->num_locals
; i
++) {
4314 for (j
= 0; j
< 4; j
++) {
4315 ctx
->locals
[i
* 4 + j
] =
4316 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
4322 setup_shared(struct ac_nir_context
*ctx
,
4323 struct nir_shader
*nir
)
4325 nir_foreach_variable(variable
, &nir
->shared
) {
4326 LLVMValueRef shared
=
4327 LLVMAddGlobalInAddressSpace(
4328 ctx
->ac
.module
, glsl_to_llvm_type(&ctx
->ac
, variable
->type
),
4329 variable
->name
? variable
->name
: "",
4331 _mesa_hash_table_insert(ctx
->vars
, variable
, shared
);
4335 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
4336 struct nir_shader
*nir
)
4338 struct ac_nir_context ctx
= {};
4339 struct nir_function
*func
;
4344 ctx
.stage
= nir
->info
.stage
;
4346 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
4348 nir_foreach_variable(variable
, &nir
->outputs
)
4349 ac_handle_shader_output_decl(&ctx
.ac
, ctx
.abi
, nir
, variable
,
4352 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4353 _mesa_key_pointer_equal
);
4354 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4355 _mesa_key_pointer_equal
);
4356 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
4357 _mesa_key_pointer_equal
);
4359 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
4361 nir_index_ssa_defs(func
->impl
);
4362 ctx
.ssa_defs
= calloc(func
->impl
->ssa_alloc
, sizeof(LLVMValueRef
));
4364 setup_locals(&ctx
, func
);
4366 if (gl_shader_stage_is_compute(nir
->info
.stage
))
4367 setup_shared(&ctx
, nir
);
4369 visit_cf_list(&ctx
, &func
->impl
->body
);
4370 phi_post_pass(&ctx
);
4372 if (!gl_shader_stage_is_compute(nir
->info
.stage
))
4373 ctx
.abi
->emit_outputs(ctx
.abi
, AC_LLVM_MAX_OUTPUTS
,
4378 ralloc_free(ctx
.defs
);
4379 ralloc_free(ctx
.phis
);
4380 ralloc_free(ctx
.vars
);
4384 ac_lower_indirect_derefs(struct nir_shader
*nir
, enum chip_class chip_class
)
4386 /* While it would be nice not to have this flag, we are constrained
4387 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
4390 bool llvm_has_working_vgpr_indexing
= chip_class
<= VI
;
4392 /* TODO: Indirect indexing of GS inputs is unimplemented.
4394 * TCS and TES load inputs directly from LDS or offchip memory, so
4395 * indirect indexing is trivial.
4397 nir_variable_mode indirect_mask
= 0;
4398 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
4399 (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
&&
4400 nir
->info
.stage
!= MESA_SHADER_TESS_EVAL
&&
4401 !llvm_has_working_vgpr_indexing
)) {
4402 indirect_mask
|= nir_var_shader_in
;
4404 if (!llvm_has_working_vgpr_indexing
&&
4405 nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
4406 indirect_mask
|= nir_var_shader_out
;
4408 /* TODO: We shouldn't need to do this, however LLVM isn't currently
4409 * smart enough to handle indirects without causing excess spilling
4410 * causing the gpu to hang.
4412 * See the following thread for more details of the problem:
4413 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
4415 indirect_mask
|= nir_var_function_temp
;
4417 nir_lower_indirect_derefs(nir
, indirect_mask
);
4421 get_inst_tessfactor_writemask(nir_intrinsic_instr
*intrin
)
4423 if (intrin
->intrinsic
!= nir_intrinsic_store_deref
)
4427 nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[0]));
4429 if (var
->data
.mode
!= nir_var_shader_out
)
4432 unsigned writemask
= 0;
4433 const int location
= var
->data
.location
;
4434 unsigned first_component
= var
->data
.location_frac
;
4435 unsigned num_comps
= intrin
->dest
.ssa
.num_components
;
4437 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
)
4438 writemask
= ((1 << (num_comps
+ 1)) - 1) << first_component
;
4439 else if (location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
4440 writemask
= (((1 << (num_comps
+ 1)) - 1) << first_component
) << 4;
4446 scan_tess_ctrl(nir_cf_node
*cf_node
, unsigned *upper_block_tf_writemask
,
4447 unsigned *cond_block_tf_writemask
,
4448 bool *tessfactors_are_def_in_all_invocs
, bool is_nested_cf
)
4450 switch (cf_node
->type
) {
4451 case nir_cf_node_block
: {
4452 nir_block
*block
= nir_cf_node_as_block(cf_node
);
4453 nir_foreach_instr(instr
, block
) {
4454 if (instr
->type
!= nir_instr_type_intrinsic
)
4457 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
4458 if (intrin
->intrinsic
== nir_intrinsic_barrier
) {
4460 /* If we find a barrier in nested control flow put this in the
4461 * too hard basket. In GLSL this is not possible but it is in
4465 *tessfactors_are_def_in_all_invocs
= false;
4469 /* The following case must be prevented:
4470 * gl_TessLevelInner = ...;
4472 * if (gl_InvocationID == 1)
4473 * gl_TessLevelInner = ...;
4475 * If you consider disjoint code segments separated by barriers, each
4476 * such segment that writes tess factor channels should write the same
4477 * channels in all codepaths within that segment.
4479 if (upper_block_tf_writemask
|| cond_block_tf_writemask
) {
4480 /* Accumulate the result: */
4481 *tessfactors_are_def_in_all_invocs
&=
4482 !(*cond_block_tf_writemask
& ~(*upper_block_tf_writemask
));
4484 /* Analyze the next code segment from scratch. */
4485 *upper_block_tf_writemask
= 0;
4486 *cond_block_tf_writemask
= 0;
4489 *upper_block_tf_writemask
|= get_inst_tessfactor_writemask(intrin
);
4494 case nir_cf_node_if
: {
4495 unsigned then_tessfactor_writemask
= 0;
4496 unsigned else_tessfactor_writemask
= 0;
4498 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
4499 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->then_list
) {
4500 scan_tess_ctrl(nested_node
, &then_tessfactor_writemask
,
4501 cond_block_tf_writemask
,
4502 tessfactors_are_def_in_all_invocs
, true);
4505 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->else_list
) {
4506 scan_tess_ctrl(nested_node
, &else_tessfactor_writemask
,
4507 cond_block_tf_writemask
,
4508 tessfactors_are_def_in_all_invocs
, true);
4511 if (then_tessfactor_writemask
|| else_tessfactor_writemask
) {
4512 /* If both statements write the same tess factor channels,
4513 * we can say that the upper block writes them too.
4515 *upper_block_tf_writemask
|= then_tessfactor_writemask
&
4516 else_tessfactor_writemask
;
4517 *cond_block_tf_writemask
|= then_tessfactor_writemask
|
4518 else_tessfactor_writemask
;
4523 case nir_cf_node_loop
: {
4524 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
4525 foreach_list_typed(nir_cf_node
, nested_node
, node
, &loop
->body
) {
4526 scan_tess_ctrl(nested_node
, cond_block_tf_writemask
,
4527 cond_block_tf_writemask
,
4528 tessfactors_are_def_in_all_invocs
, true);
4534 unreachable("unknown cf node type");
4539 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader
*nir
)
4541 assert(nir
->info
.stage
== MESA_SHADER_TESS_CTRL
);
4543 /* The pass works as follows:
4544 * If all codepaths write tess factors, we can say that all
4545 * invocations define tess factors.
4547 * Each tess factor channel is tracked separately.
4549 unsigned main_block_tf_writemask
= 0; /* if main block writes tess factors */
4550 unsigned cond_block_tf_writemask
= 0; /* if cond block writes tess factors */
4552 /* Initial value = true. Here the pass will accumulate results from
4553 * multiple segments surrounded by barriers. If tess factors aren't
4554 * written at all, it's a shader bug and we don't care if this will be
4557 bool tessfactors_are_def_in_all_invocs
= true;
4559 nir_foreach_function(function
, nir
) {
4560 if (function
->impl
) {
4561 foreach_list_typed(nir_cf_node
, node
, node
, &function
->impl
->body
) {
4562 scan_tess_ctrl(node
, &main_block_tf_writemask
,
4563 &cond_block_tf_writemask
,
4564 &tessfactors_are_def_in_all_invocs
,
4570 /* Accumulate the result for the last code segment separated by a
4573 if (main_block_tf_writemask
|| cond_block_tf_writemask
) {
4574 tessfactors_are_def_in_all_invocs
&=
4575 !(cond_block_tf_writemask
& ~main_block_tf_writemask
);
4578 return tessfactors_are_def_in_all_invocs
;