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 <llvm/Config/llvm-config.h>
26 #include "ac_nir_to_llvm.h"
27 #include "ac_llvm_build.h"
28 #include "ac_llvm_util.h"
29 #include "ac_binary.h"
32 #include "nir/nir_deref.h"
33 #include "util/bitscan.h"
34 #include "util/u_math.h"
35 #include "ac_shader_abi.h"
36 #include "ac_shader_util.h"
38 struct ac_nir_context
{
39 struct ac_llvm_context ac
;
40 struct ac_shader_abi
*abi
;
41 const struct ac_shader_args
*args
;
43 gl_shader_stage stage
;
46 LLVMValueRef
*ssa_defs
;
49 LLVMValueRef constant_data
;
51 struct hash_table
*defs
;
52 struct hash_table
*phis
;
53 struct hash_table
*vars
;
54 struct hash_table
*verified_interp
;
56 LLVMValueRef main_function
;
57 LLVMBasicBlockRef continue_block
;
58 LLVMBasicBlockRef break_block
;
64 static LLVMValueRef
get_sampler_desc_index(struct ac_nir_context
*ctx
,
65 nir_deref_instr
*deref_instr
,
66 const nir_instr
*instr
,
69 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
70 nir_deref_instr
*deref_instr
,
71 enum ac_descriptor_type desc_type
,
72 const nir_instr
*instr
,
74 bool image
, bool write
);
77 build_store_values_extended(struct ac_llvm_context
*ac
,
80 unsigned value_stride
,
83 LLVMBuilderRef builder
= ac
->builder
;
86 for (i
= 0; i
< value_count
; i
++) {
87 LLVMValueRef ptr
= values
[i
* value_stride
];
88 LLVMValueRef index
= LLVMConstInt(ac
->i32
, i
, false);
89 LLVMValueRef value
= LLVMBuildExtractElement(builder
, vec
, index
, "");
90 LLVMBuildStore(builder
, value
, ptr
);
94 static LLVMTypeRef
get_def_type(struct ac_nir_context
*ctx
,
95 const nir_ssa_def
*def
)
97 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, def
->bit_size
);
98 if (def
->num_components
> 1) {
99 type
= LLVMVectorType(type
, def
->num_components
);
104 static LLVMValueRef
get_src(struct ac_nir_context
*nir
, nir_src src
)
107 return nir
->ssa_defs
[src
.ssa
->index
];
111 get_memory_ptr(struct ac_nir_context
*ctx
, nir_src src
, unsigned bit_size
)
113 LLVMValueRef ptr
= get_src(ctx
, src
);
114 ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ctx
->ac
.lds
, &ptr
, 1, "");
115 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
117 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, bit_size
);
119 return LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
120 LLVMPointerType(type
, addr_space
), "");
123 static LLVMBasicBlockRef
get_block(struct ac_nir_context
*nir
,
124 const struct nir_block
*b
)
126 struct hash_entry
*entry
= _mesa_hash_table_search(nir
->defs
, b
);
127 return (LLVMBasicBlockRef
)entry
->data
;
130 static LLVMValueRef
get_alu_src(struct ac_nir_context
*ctx
,
132 unsigned num_components
)
134 LLVMValueRef value
= get_src(ctx
, src
.src
);
135 bool need_swizzle
= false;
138 unsigned src_components
= ac_get_llvm_num_components(value
);
139 for (unsigned i
= 0; i
< num_components
; ++i
) {
140 assert(src
.swizzle
[i
] < src_components
);
141 if (src
.swizzle
[i
] != i
)
145 if (need_swizzle
|| num_components
!= src_components
) {
146 LLVMValueRef masks
[] = {
147 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[0], false),
148 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[1], false),
149 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[2], false),
150 LLVMConstInt(ctx
->ac
.i32
, src
.swizzle
[3], false)};
152 if (src_components
> 1 && num_components
== 1) {
153 value
= LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
155 } else if (src_components
== 1 && num_components
> 1) {
156 LLVMValueRef values
[] = {value
, value
, value
, value
};
157 value
= ac_build_gather_values(&ctx
->ac
, values
, num_components
);
159 LLVMValueRef swizzle
= LLVMConstVector(masks
, num_components
);
160 value
= LLVMBuildShuffleVector(ctx
->ac
.builder
, value
, value
,
169 static LLVMValueRef
emit_int_cmp(struct ac_llvm_context
*ctx
,
170 LLVMIntPredicate pred
, LLVMValueRef src0
,
173 LLVMTypeRef src0_type
= LLVMTypeOf(src0
);
174 LLVMTypeRef src1_type
= LLVMTypeOf(src1
);
176 if (LLVMGetTypeKind(src0_type
) == LLVMPointerTypeKind
&&
177 LLVMGetTypeKind(src1_type
) != LLVMPointerTypeKind
) {
178 src1
= LLVMBuildIntToPtr(ctx
->builder
, src1
, src0_type
, "");
179 } else if (LLVMGetTypeKind(src1_type
) == LLVMPointerTypeKind
&&
180 LLVMGetTypeKind(src0_type
) != LLVMPointerTypeKind
) {
181 src0
= LLVMBuildIntToPtr(ctx
->builder
, src0
, src1_type
, "");
184 LLVMValueRef result
= LLVMBuildICmp(ctx
->builder
, pred
, src0
, src1
, "");
185 return LLVMBuildSelect(ctx
->builder
, result
,
186 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
190 static LLVMValueRef
emit_float_cmp(struct ac_llvm_context
*ctx
,
191 LLVMRealPredicate pred
, LLVMValueRef src0
,
195 src0
= ac_to_float(ctx
, src0
);
196 src1
= ac_to_float(ctx
, src1
);
197 result
= LLVMBuildFCmp(ctx
->builder
, pred
, src0
, src1
, "");
198 return LLVMBuildSelect(ctx
->builder
, result
,
199 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false),
203 static LLVMValueRef
emit_intrin_1f_param(struct ac_llvm_context
*ctx
,
205 LLVMTypeRef result_type
,
208 char name
[64], type
[64];
209 LLVMValueRef params
[] = {
210 ac_to_float(ctx
, src0
),
213 ac_build_type_name_for_intr(LLVMTypeOf(params
[0]), type
, sizeof(type
));
214 ASSERTED
const int length
= snprintf(name
, sizeof(name
), "%s.%s", intrin
, type
);
215 assert(length
< sizeof(name
));
216 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 1, AC_FUNC_ATTR_READNONE
);
219 static LLVMValueRef
emit_intrin_2f_param(struct ac_llvm_context
*ctx
,
221 LLVMTypeRef result_type
,
222 LLVMValueRef src0
, LLVMValueRef src1
)
224 char name
[64], type
[64];
225 LLVMValueRef params
[] = {
226 ac_to_float(ctx
, src0
),
227 ac_to_float(ctx
, src1
),
230 ac_build_type_name_for_intr(LLVMTypeOf(params
[0]), type
, sizeof(type
));
231 ASSERTED
const int length
= snprintf(name
, sizeof(name
), "%s.%s", intrin
, type
);
232 assert(length
< sizeof(name
));
233 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 2, AC_FUNC_ATTR_READNONE
);
236 static LLVMValueRef
emit_intrin_3f_param(struct ac_llvm_context
*ctx
,
238 LLVMTypeRef result_type
,
239 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
241 char name
[64], type
[64];
242 LLVMValueRef params
[] = {
243 ac_to_float(ctx
, src0
),
244 ac_to_float(ctx
, src1
),
245 ac_to_float(ctx
, src2
),
248 ac_build_type_name_for_intr(LLVMTypeOf(params
[0]), type
, sizeof(type
));
249 ASSERTED
const int length
= snprintf(name
, sizeof(name
), "%s.%s", intrin
, type
);
250 assert(length
< sizeof(name
));
251 return ac_build_intrinsic(ctx
, name
, result_type
, params
, 3, AC_FUNC_ATTR_READNONE
);
254 static LLVMValueRef
emit_bcsel(struct ac_llvm_context
*ctx
,
255 LLVMValueRef src0
, LLVMValueRef src1
, LLVMValueRef src2
)
257 LLVMTypeRef src1_type
= LLVMTypeOf(src1
);
258 LLVMTypeRef src2_type
= LLVMTypeOf(src2
);
260 assert(LLVMGetTypeKind(LLVMTypeOf(src0
)) != LLVMVectorTypeKind
);
262 if (LLVMGetTypeKind(src1_type
) == LLVMPointerTypeKind
&&
263 LLVMGetTypeKind(src2_type
) != LLVMPointerTypeKind
) {
264 src2
= LLVMBuildIntToPtr(ctx
->builder
, src2
, src1_type
, "");
265 } else if (LLVMGetTypeKind(src2_type
) == LLVMPointerTypeKind
&&
266 LLVMGetTypeKind(src1_type
) != LLVMPointerTypeKind
) {
267 src1
= LLVMBuildIntToPtr(ctx
->builder
, src1
, src2_type
, "");
270 LLVMValueRef v
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
,
272 return LLVMBuildSelect(ctx
->builder
, v
,
273 ac_to_integer_or_pointer(ctx
, src1
),
274 ac_to_integer_or_pointer(ctx
, src2
), "");
277 static LLVMValueRef
emit_iabs(struct ac_llvm_context
*ctx
,
280 return ac_build_imax(ctx
, src0
, LLVMBuildNeg(ctx
->builder
, src0
, ""));
283 static LLVMValueRef
emit_uint_carry(struct ac_llvm_context
*ctx
,
285 LLVMValueRef src0
, LLVMValueRef src1
)
287 LLVMTypeRef ret_type
;
288 LLVMTypeRef types
[] = { ctx
->i32
, ctx
->i1
};
290 LLVMValueRef params
[] = { src0
, src1
};
291 ret_type
= LLVMStructTypeInContext(ctx
->context
, types
,
294 res
= ac_build_intrinsic(ctx
, intrin
, ret_type
,
295 params
, 2, AC_FUNC_ATTR_READNONE
);
297 res
= LLVMBuildExtractValue(ctx
->builder
, res
, 1, "");
298 res
= LLVMBuildZExt(ctx
->builder
, res
, ctx
->i32
, "");
302 static LLVMValueRef
emit_b2f(struct ac_llvm_context
*ctx
,
306 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
,
307 LLVMBuildBitCast(ctx
->builder
, LLVMConstReal(ctx
->f32
, 1.0), ctx
->i32
, ""),
309 result
= LLVMBuildBitCast(ctx
->builder
, result
, ctx
->f32
, "");
313 return LLVMBuildFPTrunc(ctx
->builder
, result
, ctx
->f16
, "");
317 return LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f64
, "");
319 unreachable("Unsupported bit size.");
323 static LLVMValueRef
emit_f2b(struct ac_llvm_context
*ctx
,
326 src0
= ac_to_float(ctx
, src0
);
327 LLVMValueRef zero
= LLVMConstNull(LLVMTypeOf(src0
));
328 return LLVMBuildSExt(ctx
->builder
,
329 LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
, src0
, zero
, ""),
333 static LLVMValueRef
emit_b2i(struct ac_llvm_context
*ctx
,
337 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, src0
, ctx
->i32_1
, "");
341 return LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i8
, "");
343 return LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i16
, "");
347 return LLVMBuildZExt(ctx
->builder
, result
, ctx
->i64
, "");
349 unreachable("Unsupported bit size.");
353 static LLVMValueRef
emit_i2b(struct ac_llvm_context
*ctx
,
356 LLVMValueRef zero
= LLVMConstNull(LLVMTypeOf(src0
));
357 return LLVMBuildSExt(ctx
->builder
,
358 LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src0
, zero
, ""),
362 static LLVMValueRef
emit_f2f16(struct ac_llvm_context
*ctx
,
366 LLVMValueRef cond
= NULL
;
368 src0
= ac_to_float(ctx
, src0
);
369 result
= LLVMBuildFPTrunc(ctx
->builder
, src0
, ctx
->f16
, "");
371 if (ctx
->chip_class
>= GFX8
) {
372 LLVMValueRef args
[2];
373 /* Check if the result is a denormal - and flush to 0 if so. */
375 args
[1] = LLVMConstInt(ctx
->i32
, N_SUBNORMAL
| P_SUBNORMAL
, false);
376 cond
= ac_build_intrinsic(ctx
, "llvm.amdgcn.class.f16", ctx
->i1
, args
, 2, AC_FUNC_ATTR_READNONE
);
379 /* need to convert back up to f32 */
380 result
= LLVMBuildFPExt(ctx
->builder
, result
, ctx
->f32
, "");
382 if (ctx
->chip_class
>= GFX8
)
383 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
386 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
387 * so compare the result and flush to 0 if it's smaller.
389 LLVMValueRef temp
, cond2
;
390 temp
= emit_intrin_1f_param(ctx
, "llvm.fabs", ctx
->f32
, result
);
391 cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
,
392 LLVMBuildBitCast(ctx
->builder
, LLVMConstInt(ctx
->i32
, 0x38800000, false), ctx
->f32
, ""),
394 cond2
= LLVMBuildFCmp(ctx
->builder
, LLVMRealONE
,
395 temp
, ctx
->f32_0
, "");
396 cond
= LLVMBuildAnd(ctx
->builder
, cond
, cond2
, "");
397 result
= LLVMBuildSelect(ctx
->builder
, cond
, ctx
->f32_0
, result
, "");
402 static LLVMValueRef
emit_umul_high(struct ac_llvm_context
*ctx
,
403 LLVMValueRef src0
, LLVMValueRef src1
)
405 LLVMValueRef dst64
, result
;
406 src0
= LLVMBuildZExt(ctx
->builder
, src0
, ctx
->i64
, "");
407 src1
= LLVMBuildZExt(ctx
->builder
, src1
, ctx
->i64
, "");
409 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
410 dst64
= LLVMBuildLShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
411 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
415 static LLVMValueRef
emit_imul_high(struct ac_llvm_context
*ctx
,
416 LLVMValueRef src0
, LLVMValueRef src1
)
418 LLVMValueRef dst64
, result
;
419 src0
= LLVMBuildSExt(ctx
->builder
, src0
, ctx
->i64
, "");
420 src1
= LLVMBuildSExt(ctx
->builder
, src1
, ctx
->i64
, "");
422 dst64
= LLVMBuildMul(ctx
->builder
, src0
, src1
, "");
423 dst64
= LLVMBuildAShr(ctx
->builder
, dst64
, LLVMConstInt(ctx
->i64
, 32, false), "");
424 result
= LLVMBuildTrunc(ctx
->builder
, dst64
, ctx
->i32
, "");
428 static LLVMValueRef
emit_bfm(struct ac_llvm_context
*ctx
,
429 LLVMValueRef bits
, LLVMValueRef offset
)
431 /* mask = ((1 << bits) - 1) << offset */
432 return LLVMBuildShl(ctx
->builder
,
433 LLVMBuildSub(ctx
->builder
,
434 LLVMBuildShl(ctx
->builder
,
441 static LLVMValueRef
emit_bitfield_select(struct ac_llvm_context
*ctx
,
442 LLVMValueRef mask
, LLVMValueRef insert
,
446 * (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
447 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
449 return LLVMBuildXor(ctx
->builder
, base
,
450 LLVMBuildAnd(ctx
->builder
, mask
,
451 LLVMBuildXor(ctx
->builder
, insert
, base
, ""), ""), "");
454 static LLVMValueRef
emit_pack_2x16(struct ac_llvm_context
*ctx
,
456 LLVMValueRef (*pack
)(struct ac_llvm_context
*ctx
,
457 LLVMValueRef args
[2]))
459 LLVMValueRef comp
[2];
461 src0
= ac_to_float(ctx
, src0
);
462 comp
[0] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_0
, "");
463 comp
[1] = LLVMBuildExtractElement(ctx
->builder
, src0
, ctx
->i32_1
, "");
465 return LLVMBuildBitCast(ctx
->builder
, pack(ctx
, comp
), ctx
->i32
, "");
468 static LLVMValueRef
emit_unpack_half_2x16(struct ac_llvm_context
*ctx
,
471 LLVMValueRef const16
= LLVMConstInt(ctx
->i32
, 16, false);
472 LLVMValueRef temps
[2], val
;
475 for (i
= 0; i
< 2; i
++) {
476 val
= i
== 1 ? LLVMBuildLShr(ctx
->builder
, src0
, const16
, "") : src0
;
477 val
= LLVMBuildTrunc(ctx
->builder
, val
, ctx
->i16
, "");
478 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->f16
, "");
479 temps
[i
] = LLVMBuildFPExt(ctx
->builder
, val
, ctx
->f32
, "");
481 return ac_build_gather_values(ctx
, temps
, 2);
484 static LLVMValueRef
emit_ddxy(struct ac_nir_context
*ctx
,
492 if (op
== nir_op_fddx_fine
)
493 mask
= AC_TID_MASK_LEFT
;
494 else if (op
== nir_op_fddy_fine
)
495 mask
= AC_TID_MASK_TOP
;
497 mask
= AC_TID_MASK_TOP_LEFT
;
499 /* for DDX we want to next X pixel, DDY next Y pixel. */
500 if (op
== nir_op_fddx_fine
||
501 op
== nir_op_fddx_coarse
||
507 result
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, src0
);
511 struct waterfall_context
{
512 LLVMBasicBlockRef phi_bb
[2];
516 /* To deal with divergent descriptors we can create a loop that handles all
517 * lanes with the same descriptor on a given iteration (henceforth a
520 * These helper create the begin and end of the loop leaving the caller
521 * to implement the body.
524 * - ctx is the usal nir context
525 * - wctx is a temporary struct containing some loop info. Can be left uninitialized.
526 * - value is the possibly divergent value for which we built the loop
527 * - divergent is whether value is actually divergent. If false we just pass
530 static LLVMValueRef
enter_waterfall(struct ac_nir_context
*ctx
,
531 struct waterfall_context
*wctx
,
532 LLVMValueRef value
, bool divergent
)
534 /* If the app claims the value is divergent but it is constant we can
535 * end up with a dynamic index of NULL. */
539 wctx
->use_waterfall
= divergent
;
543 ac_build_bgnloop(&ctx
->ac
, 6000);
545 LLVMValueRef scalar_value
= ac_build_readlane(&ctx
->ac
, value
, NULL
);
547 LLVMValueRef active
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, value
,
548 scalar_value
, "uniform_active");
550 wctx
->phi_bb
[0] = LLVMGetInsertBlock(ctx
->ac
.builder
);
551 ac_build_ifcc(&ctx
->ac
, active
, 6001);
556 static LLVMValueRef
exit_waterfall(struct ac_nir_context
*ctx
,
557 struct waterfall_context
*wctx
,
560 LLVMValueRef ret
= NULL
;
561 LLVMValueRef phi_src
[2];
562 LLVMValueRef cc_phi_src
[2] = {
563 LLVMConstInt(ctx
->ac
.i32
, 0, false),
564 LLVMConstInt(ctx
->ac
.i32
, 0xffffffff, false),
567 if (!wctx
->use_waterfall
)
570 wctx
->phi_bb
[1] = LLVMGetInsertBlock(ctx
->ac
.builder
);
572 ac_build_endif(&ctx
->ac
, 6001);
575 phi_src
[0] = LLVMGetUndef(LLVMTypeOf(value
));
578 ret
= ac_build_phi(&ctx
->ac
, LLVMTypeOf(value
), 2, phi_src
, wctx
->phi_bb
);
582 * By using the optimization barrier on the exit decision, we decouple
583 * the operations from the break, and hence avoid LLVM hoisting the
584 * opteration into the break block.
586 LLVMValueRef cc
= ac_build_phi(&ctx
->ac
, ctx
->ac
.i32
, 2, cc_phi_src
, wctx
->phi_bb
);
587 ac_build_optimization_barrier(&ctx
->ac
, &cc
);
589 LLVMValueRef active
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, cc
, ctx
->ac
.i32_0
, "uniform_active2");
590 ac_build_ifcc(&ctx
->ac
, active
, 6002);
591 ac_build_break(&ctx
->ac
);
592 ac_build_endif(&ctx
->ac
, 6002);
594 ac_build_endloop(&ctx
->ac
, 6000);
598 static void visit_alu(struct ac_nir_context
*ctx
, const nir_alu_instr
*instr
)
600 LLVMValueRef src
[4], result
= NULL
;
601 unsigned num_components
= instr
->dest
.dest
.ssa
.num_components
;
602 unsigned src_components
;
603 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.dest
.ssa
);
604 bool saved_inexact
= false;
607 saved_inexact
= ac_disable_inexact_math(ctx
->ac
.builder
);
609 assert(nir_op_infos
[instr
->op
].num_inputs
<= ARRAY_SIZE(src
));
616 case nir_op_pack_half_2x16
:
617 case nir_op_pack_snorm_2x16
:
618 case nir_op_pack_unorm_2x16
:
621 case nir_op_unpack_half_2x16
:
624 case nir_op_cube_face_coord
:
625 case nir_op_cube_face_index
:
629 src_components
= num_components
;
632 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
633 src
[i
] = get_alu_src(ctx
, instr
->src
[i
], src_components
);
640 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
641 result
= LLVMBuildFNeg(ctx
->ac
.builder
, src
[0], "");
642 if (ctx
->ac
.float_mode
== AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO
) {
643 /* fneg will be optimized by backend compiler with sign
644 * bit removed via XOR. This is probably a LLVM bug.
646 result
= ac_build_canonicalize(&ctx
->ac
, result
,
647 instr
->dest
.dest
.ssa
.bit_size
);
651 result
= LLVMBuildNeg(ctx
->ac
.builder
, src
[0], "");
654 result
= LLVMBuildNot(ctx
->ac
.builder
, src
[0], "");
657 result
= LLVMBuildAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
660 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
661 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
662 result
= LLVMBuildFAdd(ctx
->ac
.builder
, src
[0], src
[1], "");
665 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
666 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
667 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], src
[1], "");
670 result
= LLVMBuildSub(ctx
->ac
.builder
, src
[0], src
[1], "");
673 result
= LLVMBuildMul(ctx
->ac
.builder
, src
[0], src
[1], "");
676 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
679 result
= LLVMBuildURem(ctx
->ac
.builder
, src
[0], src
[1], "");
682 /* lower_fmod only lower 16-bit and 32-bit fmod */
683 assert(instr
->dest
.dest
.ssa
.bit_size
== 64);
684 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
685 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
686 result
= ac_build_fdiv(&ctx
->ac
, src
[0], src
[1]);
687 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
688 ac_to_float_type(&ctx
->ac
, def_type
), result
);
689 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[1] , result
, "");
690 result
= LLVMBuildFSub(ctx
->ac
.builder
, src
[0], result
, "");
693 result
= LLVMBuildSRem(ctx
->ac
.builder
, src
[0], src
[1], "");
696 result
= LLVMBuildSDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
699 result
= LLVMBuildUDiv(ctx
->ac
.builder
, src
[0], src
[1], "");
702 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
703 src
[1] = ac_to_float(&ctx
->ac
, src
[1]);
704 result
= LLVMBuildFMul(ctx
->ac
.builder
, src
[0], src
[1], "");
707 /* For doubles, we need precise division to pass GLCTS. */
708 if (ctx
->ac
.float_mode
== AC_FLOAT_MODE_DEFAULT_OPENGL
&&
709 ac_get_type_size(def_type
) == 8) {
710 result
= LLVMBuildFDiv(ctx
->ac
.builder
, ctx
->ac
.f64_1
,
711 ac_to_float(&ctx
->ac
, src
[0]), "");
713 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.amdgcn.rcp",
714 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
718 result
= LLVMBuildAnd(ctx
->ac
.builder
, src
[0], src
[1], "");
721 result
= LLVMBuildOr(ctx
->ac
.builder
, src
[0], src
[1], "");
724 result
= LLVMBuildXor(ctx
->ac
.builder
, src
[0], src
[1], "");
727 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
728 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
729 LLVMTypeOf(src
[0]), "");
730 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
731 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
732 LLVMTypeOf(src
[0]), "");
733 result
= LLVMBuildShl(ctx
->ac
.builder
, src
[0], src
[1], "");
736 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
737 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
738 LLVMTypeOf(src
[0]), "");
739 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
740 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
741 LLVMTypeOf(src
[0]), "");
742 result
= LLVMBuildAShr(ctx
->ac
.builder
, src
[0], src
[1], "");
745 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) < ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
746 src
[1] = LLVMBuildZExt(ctx
->ac
.builder
, src
[1],
747 LLVMTypeOf(src
[0]), "");
748 else if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[1])) > ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])))
749 src
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, src
[1],
750 LLVMTypeOf(src
[0]), "");
751 result
= LLVMBuildLShr(ctx
->ac
.builder
, src
[0], src
[1], "");
754 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSLT
, src
[0], src
[1]);
757 result
= emit_int_cmp(&ctx
->ac
, LLVMIntNE
, src
[0], src
[1]);
760 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, src
[0], src
[1]);
763 result
= emit_int_cmp(&ctx
->ac
, LLVMIntSGE
, src
[0], src
[1]);
766 result
= emit_int_cmp(&ctx
->ac
, LLVMIntULT
, src
[0], src
[1]);
769 result
= emit_int_cmp(&ctx
->ac
, LLVMIntUGE
, src
[0], src
[1]);
772 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOEQ
, src
[0], src
[1]);
775 result
= emit_float_cmp(&ctx
->ac
, LLVMRealUNE
, src
[0], src
[1]);
778 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOLT
, src
[0], src
[1]);
781 result
= emit_float_cmp(&ctx
->ac
, LLVMRealOGE
, src
[0], src
[1]);
784 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.fabs",
785 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
786 if (ctx
->ac
.float_mode
== AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO
) {
787 /* fabs will be optimized by backend compiler with sign
788 * bit removed via AND.
790 result
= ac_build_canonicalize(&ctx
->ac
, result
,
791 instr
->dest
.dest
.ssa
.bit_size
);
795 result
= emit_iabs(&ctx
->ac
, src
[0]);
798 result
= ac_build_imax(&ctx
->ac
, src
[0], src
[1]);
801 result
= ac_build_imin(&ctx
->ac
, src
[0], src
[1]);
804 result
= ac_build_umax(&ctx
->ac
, src
[0], src
[1]);
807 result
= ac_build_umin(&ctx
->ac
, src
[0], src
[1]);
810 result
= ac_build_isign(&ctx
->ac
, src
[0],
811 instr
->dest
.dest
.ssa
.bit_size
);
814 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
815 result
= ac_build_fsign(&ctx
->ac
, src
[0],
816 instr
->dest
.dest
.ssa
.bit_size
);
819 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.floor",
820 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
823 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.trunc",
824 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
827 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.ceil",
828 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
830 case nir_op_fround_even
:
831 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.rint",
832 ac_to_float_type(&ctx
->ac
, def_type
),src
[0]);
835 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
836 result
= ac_build_fract(&ctx
->ac
, src
[0],
837 instr
->dest
.dest
.ssa
.bit_size
);
840 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sin",
841 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
844 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.cos",
845 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
848 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.sqrt",
849 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
852 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.exp2",
853 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
856 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.log2",
857 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
860 result
= emit_intrin_1f_param(&ctx
->ac
, "llvm.amdgcn.rsq",
861 ac_to_float_type(&ctx
->ac
, def_type
), src
[0]);
863 case nir_op_frexp_exp
:
864 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
865 result
= ac_build_frexp_exp(&ctx
->ac
, src
[0],
866 ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])));
867 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) == 16)
868 result
= LLVMBuildSExt(ctx
->ac
.builder
, result
,
871 case nir_op_frexp_sig
:
872 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
873 result
= ac_build_frexp_mant(&ctx
->ac
, src
[0],
874 instr
->dest
.dest
.ssa
.bit_size
);
877 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.pow",
878 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
881 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.maxnum",
882 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
883 if (ctx
->ac
.chip_class
< GFX9
&&
884 instr
->dest
.dest
.ssa
.bit_size
== 32) {
885 /* Only pre-GFX9 chips do not flush denorms. */
886 result
= ac_build_canonicalize(&ctx
->ac
, result
,
887 instr
->dest
.dest
.ssa
.bit_size
);
891 result
= emit_intrin_2f_param(&ctx
->ac
, "llvm.minnum",
892 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1]);
893 if (ctx
->ac
.chip_class
< GFX9
&&
894 instr
->dest
.dest
.ssa
.bit_size
== 32) {
895 /* Only pre-GFX9 chips do not flush denorms. */
896 result
= ac_build_canonicalize(&ctx
->ac
, result
,
897 instr
->dest
.dest
.ssa
.bit_size
);
901 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
902 result
= emit_intrin_3f_param(&ctx
->ac
, ctx
->ac
.chip_class
>= GFX10
? "llvm.fma" : "llvm.fmuladd",
903 ac_to_float_type(&ctx
->ac
, def_type
), src
[0], src
[1], src
[2]);
906 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
907 if (ac_get_elem_bits(&ctx
->ac
, def_type
) == 32)
908 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f32", ctx
->ac
.f32
, src
, 2, AC_FUNC_ATTR_READNONE
);
909 else if (ac_get_elem_bits(&ctx
->ac
, def_type
) == 16)
910 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f16", ctx
->ac
.f16
, src
, 2, AC_FUNC_ATTR_READNONE
);
912 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ldexp.f64", ctx
->ac
.f64
, src
, 2, AC_FUNC_ATTR_READNONE
);
915 result
= emit_bfm(&ctx
->ac
, src
[0], src
[1]);
917 case nir_op_bitfield_select
:
918 result
= emit_bitfield_select(&ctx
->ac
, src
[0], src
[1], src
[2]);
921 result
= ac_build_bfe(&ctx
->ac
, src
[0], src
[1], src
[2], false);
924 result
= ac_build_bfe(&ctx
->ac
, src
[0], src
[1], src
[2], true);
926 case nir_op_bitfield_reverse
:
927 result
= ac_build_bitfield_reverse(&ctx
->ac
, src
[0]);
929 case nir_op_bit_count
:
930 result
= ac_build_bit_count(&ctx
->ac
, src
[0]);
935 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
936 src
[i
] = ac_to_integer(&ctx
->ac
, src
[i
]);
937 result
= ac_build_gather_values(&ctx
->ac
, src
, num_components
);
943 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
944 result
= LLVMBuildFPToSI(ctx
->ac
.builder
, src
[0], def_type
, "");
950 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
951 result
= LLVMBuildFPToUI(ctx
->ac
.builder
, src
[0], def_type
, "");
956 result
= LLVMBuildSIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
961 result
= LLVMBuildUIToFP(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
963 case nir_op_f2f16_rtz
:
966 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
968 /* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
969 * all f32->f16 conversions have to round towards zero, because both scalar
970 * and vec2 down-conversions have to round equally.
972 if (ctx
->ac
.float_mode
== AC_FLOAT_MODE_DEFAULT_OPENGL
||
973 instr
->op
== nir_op_f2f16_rtz
) {
974 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
976 if (LLVMTypeOf(src
[0]) == ctx
->ac
.f64
)
977 src
[0] = LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ctx
->ac
.f32
, "");
979 /* Fast path conversion. This only works if NIR is vectorized
982 if (LLVMTypeOf(src
[0]) == ctx
->ac
.v2f32
) {
983 LLVMValueRef args
[] = {
984 ac_llvm_extract_elem(&ctx
->ac
, src
[0], 0),
985 ac_llvm_extract_elem(&ctx
->ac
, src
[0], 1),
987 result
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, args
);
991 assert(ac_get_llvm_num_components(src
[0]) == 1);
992 LLVMValueRef param
[2] = { src
[0], LLVMGetUndef(ctx
->ac
.f32
) };
993 result
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, param
);
994 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
996 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
997 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
999 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1002 case nir_op_f2f16_rtne
:
1005 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1006 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
1007 result
= LLVMBuildFPExt(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1009 result
= LLVMBuildFPTrunc(ctx
->ac
.builder
, src
[0], ac_to_float_type(&ctx
->ac
, def_type
), "");
1016 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
1017 result
= LLVMBuildZExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1019 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1026 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
[0])) < ac_get_elem_bits(&ctx
->ac
, def_type
))
1027 result
= LLVMBuildSExt(ctx
->ac
.builder
, src
[0], def_type
, "");
1029 result
= LLVMBuildTrunc(ctx
->ac
.builder
, src
[0], def_type
, "");
1031 case nir_op_b32csel
:
1032 result
= emit_bcsel(&ctx
->ac
, src
[0], src
[1], src
[2]);
1034 case nir_op_find_lsb
:
1035 result
= ac_find_lsb(&ctx
->ac
, ctx
->ac
.i32
, src
[0]);
1037 case nir_op_ufind_msb
:
1038 result
= ac_build_umsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1040 case nir_op_ifind_msb
:
1041 result
= ac_build_imsb(&ctx
->ac
, src
[0], ctx
->ac
.i32
);
1043 case nir_op_uadd_carry
:
1044 result
= emit_uint_carry(&ctx
->ac
, "llvm.uadd.with.overflow.i32", src
[0], src
[1]);
1046 case nir_op_usub_borrow
:
1047 result
= emit_uint_carry(&ctx
->ac
, "llvm.usub.with.overflow.i32", src
[0], src
[1]);
1052 result
= emit_b2f(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1055 result
= emit_f2b(&ctx
->ac
, src
[0]);
1061 result
= emit_b2i(&ctx
->ac
, src
[0], instr
->dest
.dest
.ssa
.bit_size
);
1064 result
= emit_i2b(&ctx
->ac
, src
[0]);
1066 case nir_op_fquantize2f16
:
1067 result
= emit_f2f16(&ctx
->ac
, src
[0]);
1069 case nir_op_umul_high
:
1070 result
= emit_umul_high(&ctx
->ac
, src
[0], src
[1]);
1072 case nir_op_imul_high
:
1073 result
= emit_imul_high(&ctx
->ac
, src
[0], src
[1]);
1075 case nir_op_pack_half_2x16
:
1076 result
= emit_pack_2x16(&ctx
->ac
, src
[0], ac_build_cvt_pkrtz_f16
);
1078 case nir_op_pack_snorm_2x16
:
1079 result
= emit_pack_2x16(&ctx
->ac
, src
[0], ac_build_cvt_pknorm_i16
);
1081 case nir_op_pack_unorm_2x16
:
1082 result
= emit_pack_2x16(&ctx
->ac
, src
[0], ac_build_cvt_pknorm_u16
);
1084 case nir_op_unpack_half_2x16
:
1085 result
= emit_unpack_half_2x16(&ctx
->ac
, src
[0]);
1089 case nir_op_fddx_fine
:
1090 case nir_op_fddy_fine
:
1091 case nir_op_fddx_coarse
:
1092 case nir_op_fddy_coarse
:
1093 result
= emit_ddxy(ctx
, instr
->op
, src
[0]);
1096 case nir_op_unpack_64_2x32_split_x
: {
1097 assert(ac_get_llvm_num_components(src
[0]) == 1);
1098 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1101 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1106 case nir_op_unpack_64_2x32_split_y
: {
1107 assert(ac_get_llvm_num_components(src
[0]) == 1);
1108 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1111 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1116 case nir_op_pack_64_2x32_split
: {
1117 LLVMValueRef tmp
= ac_build_gather_values(&ctx
->ac
, src
, 2);
1118 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i64
, "");
1122 case nir_op_pack_32_2x16_split
: {
1123 LLVMValueRef tmp
= ac_build_gather_values(&ctx
->ac
, src
, 2);
1124 result
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
1128 case nir_op_unpack_32_2x16_split_x
: {
1129 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1132 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1137 case nir_op_unpack_32_2x16_split_y
: {
1138 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, src
[0],
1141 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
,
1146 case nir_op_cube_face_coord
: {
1147 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1148 LLVMValueRef results
[2];
1150 for (unsigned chan
= 0; chan
< 3; chan
++)
1151 in
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src
[0], chan
);
1152 results
[0] = ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubesc",
1153 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1154 results
[1] = ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubetc",
1155 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1156 LLVMValueRef ma
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubema",
1157 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1158 results
[0] = ac_build_fdiv(&ctx
->ac
, results
[0], ma
);
1159 results
[1] = ac_build_fdiv(&ctx
->ac
, results
[1], ma
);
1160 LLVMValueRef offset
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
1161 results
[0] = LLVMBuildFAdd(ctx
->ac
.builder
, results
[0], offset
, "");
1162 results
[1] = LLVMBuildFAdd(ctx
->ac
.builder
, results
[1], offset
, "");
1163 result
= ac_build_gather_values(&ctx
->ac
, results
, 2);
1167 case nir_op_cube_face_index
: {
1168 src
[0] = ac_to_float(&ctx
->ac
, src
[0]);
1170 for (unsigned chan
= 0; chan
< 3; chan
++)
1171 in
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src
[0], chan
);
1172 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.cubeid",
1173 ctx
->ac
.f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
1178 fprintf(stderr
, "Unknown NIR alu instr: ");
1179 nir_print_instr(&instr
->instr
, stderr
);
1180 fprintf(stderr
, "\n");
1185 assert(instr
->dest
.dest
.is_ssa
);
1186 result
= ac_to_integer_or_pointer(&ctx
->ac
, result
);
1187 ctx
->ssa_defs
[instr
->dest
.dest
.ssa
.index
] = result
;
1191 ac_restore_inexact_math(ctx
->ac
.builder
, saved_inexact
);
1194 static void visit_load_const(struct ac_nir_context
*ctx
,
1195 const nir_load_const_instr
*instr
)
1197 LLVMValueRef values
[4], value
= NULL
;
1198 LLVMTypeRef element_type
=
1199 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
1201 for (unsigned i
= 0; i
< instr
->def
.num_components
; ++i
) {
1202 switch (instr
->def
.bit_size
) {
1204 values
[i
] = LLVMConstInt(element_type
,
1205 instr
->value
[i
].u8
, false);
1208 values
[i
] = LLVMConstInt(element_type
,
1209 instr
->value
[i
].u16
, false);
1212 values
[i
] = LLVMConstInt(element_type
,
1213 instr
->value
[i
].u32
, false);
1216 values
[i
] = LLVMConstInt(element_type
,
1217 instr
->value
[i
].u64
, false);
1221 "unsupported nir load_const bit_size: %d\n",
1222 instr
->def
.bit_size
);
1226 if (instr
->def
.num_components
> 1) {
1227 value
= LLVMConstVector(values
, instr
->def
.num_components
);
1231 ctx
->ssa_defs
[instr
->def
.index
] = value
;
1235 get_buffer_size(struct ac_nir_context
*ctx
, LLVMValueRef descriptor
, bool in_elements
)
1238 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1239 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1242 if (ctx
->ac
.chip_class
== GFX8
&& in_elements
) {
1243 /* On GFX8, the descriptor contains the size in bytes,
1244 * but TXQ must return the size in elements.
1245 * The stride is always non-zero for resources using TXQ.
1247 LLVMValueRef stride
=
1248 LLVMBuildExtractElement(ctx
->ac
.builder
, descriptor
,
1250 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
,
1251 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
1252 stride
= LLVMBuildAnd(ctx
->ac
.builder
, stride
,
1253 LLVMConstInt(ctx
->ac
.i32
, 0x3fff, false), "");
1255 size
= LLVMBuildUDiv(ctx
->ac
.builder
, size
, stride
, "");
1260 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1261 * incorrectly forces nearest filtering if the texture format is integer.
1262 * The only effect it has on Gather4, which always returns 4 texels for
1263 * bilinear filtering, is that the final coordinates are off by 0.5 of
1266 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1267 * or (0.5 / size) from the normalized coordinates.
1269 * However, cube textures with 8_8_8_8 data formats require a different
1270 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1271 * precision in 32-bit data formats, so it needs to be applied dynamically at
1272 * runtime. In this case, return an i1 value that indicates whether the
1273 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1275 static LLVMValueRef
lower_gather4_integer(struct ac_llvm_context
*ctx
,
1277 struct ac_image_args
*args
,
1278 const nir_tex_instr
*instr
)
1280 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1281 enum glsl_base_type stype
= glsl_get_sampler_result_type(type
);
1282 LLVMValueRef wa_8888
= NULL
;
1283 LLVMValueRef half_texel
[2];
1284 LLVMValueRef result
;
1286 assert(stype
== GLSL_TYPE_INT
|| stype
== GLSL_TYPE_UINT
);
1288 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1289 LLVMValueRef formats
;
1290 LLVMValueRef data_format
;
1291 LLVMValueRef wa_formats
;
1293 formats
= LLVMBuildExtractElement(ctx
->builder
, args
->resource
, ctx
->i32_1
, "");
1295 data_format
= LLVMBuildLShr(ctx
->builder
, formats
,
1296 LLVMConstInt(ctx
->i32
, 20, false), "");
1297 data_format
= LLVMBuildAnd(ctx
->builder
, data_format
,
1298 LLVMConstInt(ctx
->i32
, (1u << 6) - 1, false), "");
1299 wa_8888
= LLVMBuildICmp(
1300 ctx
->builder
, LLVMIntEQ
, data_format
,
1301 LLVMConstInt(ctx
->i32
, V_008F14_IMG_DATA_FORMAT_8_8_8_8
, false),
1304 uint32_t wa_num_format
=
1305 stype
== GLSL_TYPE_UINT
?
1306 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED
) :
1307 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED
);
1308 wa_formats
= LLVMBuildAnd(ctx
->builder
, formats
,
1309 LLVMConstInt(ctx
->i32
, C_008F14_NUM_FORMAT
, false),
1311 wa_formats
= LLVMBuildOr(ctx
->builder
, wa_formats
,
1312 LLVMConstInt(ctx
->i32
, wa_num_format
, false), "");
1314 formats
= LLVMBuildSelect(ctx
->builder
, wa_8888
, wa_formats
, formats
, "");
1315 args
->resource
= LLVMBuildInsertElement(
1316 ctx
->builder
, args
->resource
, formats
, ctx
->i32_1
, "");
1319 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
1321 half_texel
[0] = half_texel
[1] = LLVMConstReal(ctx
->f32
, -0.5);
1323 struct ac_image_args resinfo
= {};
1324 LLVMBasicBlockRef bbs
[2];
1326 LLVMValueRef unnorm
= NULL
;
1327 LLVMValueRef default_offset
= ctx
->f32_0
;
1328 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
&&
1330 /* In vulkan, whether the sampler uses unnormalized
1331 * coordinates or not is a dynamic property of the
1332 * sampler. Hence, to figure out whether or not we
1333 * need to divide by the texture size, we need to test
1334 * the sampler at runtime. This tests the bit set by
1335 * radv_init_sampler().
1337 LLVMValueRef sampler0
=
1338 LLVMBuildExtractElement(ctx
->builder
, args
->sampler
, ctx
->i32_0
, "");
1339 sampler0
= LLVMBuildLShr(ctx
->builder
, sampler0
,
1340 LLVMConstInt(ctx
->i32
, 15, false), "");
1341 sampler0
= LLVMBuildAnd(ctx
->builder
, sampler0
, ctx
->i32_1
, "");
1342 unnorm
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, sampler0
, ctx
->i32_1
, "");
1343 default_offset
= LLVMConstReal(ctx
->f32
, -0.5);
1346 bbs
[0] = LLVMGetInsertBlock(ctx
->builder
);
1347 if (wa_8888
|| unnorm
) {
1348 assert(!(wa_8888
&& unnorm
));
1349 LLVMValueRef not_needed
= wa_8888
? wa_8888
: unnorm
;
1350 /* Skip the texture size query entirely if we don't need it. */
1351 ac_build_ifcc(ctx
, LLVMBuildNot(ctx
->builder
, not_needed
, ""), 2000);
1352 bbs
[1] = LLVMGetInsertBlock(ctx
->builder
);
1355 /* Query the texture size. */
1356 resinfo
.dim
= ac_get_sampler_dim(ctx
->chip_class
, instr
->sampler_dim
, instr
->is_array
);
1357 resinfo
.opcode
= ac_image_get_resinfo
;
1358 resinfo
.dmask
= 0xf;
1359 resinfo
.lod
= ctx
->i32_0
;
1360 resinfo
.resource
= args
->resource
;
1361 resinfo
.attributes
= AC_FUNC_ATTR_READNONE
;
1362 LLVMValueRef size
= ac_build_image_opcode(ctx
, &resinfo
);
1364 /* Compute -0.5 / size. */
1365 for (unsigned c
= 0; c
< 2; c
++) {
1367 LLVMBuildExtractElement(ctx
->builder
, size
,
1368 LLVMConstInt(ctx
->i32
, c
, 0), "");
1369 half_texel
[c
] = LLVMBuildUIToFP(ctx
->builder
, half_texel
[c
], ctx
->f32
, "");
1370 half_texel
[c
] = ac_build_fdiv(ctx
, ctx
->f32_1
, half_texel
[c
]);
1371 half_texel
[c
] = LLVMBuildFMul(ctx
->builder
, half_texel
[c
],
1372 LLVMConstReal(ctx
->f32
, -0.5), "");
1375 if (wa_8888
|| unnorm
) {
1376 ac_build_endif(ctx
, 2000);
1378 for (unsigned c
= 0; c
< 2; c
++) {
1379 LLVMValueRef values
[2] = { default_offset
, half_texel
[c
] };
1380 half_texel
[c
] = ac_build_phi(ctx
, ctx
->f32
, 2,
1386 for (unsigned c
= 0; c
< 2; c
++) {
1388 tmp
= LLVMBuildBitCast(ctx
->builder
, args
->coords
[c
], ctx
->f32
, "");
1389 args
->coords
[c
] = LLVMBuildFAdd(ctx
->builder
, tmp
, half_texel
[c
], "");
1392 args
->attributes
= AC_FUNC_ATTR_READNONE
;
1393 result
= ac_build_image_opcode(ctx
, args
);
1395 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1396 LLVMValueRef tmp
, tmp2
;
1398 /* if the cube workaround is in place, f2i the result. */
1399 for (unsigned c
= 0; c
< 4; c
++) {
1400 tmp
= LLVMBuildExtractElement(ctx
->builder
, result
, LLVMConstInt(ctx
->i32
, c
, false), "");
1401 if (stype
== GLSL_TYPE_UINT
)
1402 tmp2
= LLVMBuildFPToUI(ctx
->builder
, tmp
, ctx
->i32
, "");
1404 tmp2
= LLVMBuildFPToSI(ctx
->builder
, tmp
, ctx
->i32
, "");
1405 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->i32
, "");
1406 tmp2
= LLVMBuildBitCast(ctx
->builder
, tmp2
, ctx
->i32
, "");
1407 tmp
= LLVMBuildSelect(ctx
->builder
, wa_8888
, tmp2
, tmp
, "");
1408 tmp
= LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f32
, "");
1409 result
= LLVMBuildInsertElement(ctx
->builder
, result
, tmp
, LLVMConstInt(ctx
->i32
, c
, false), "");
1415 static nir_deref_instr
*get_tex_texture_deref(const nir_tex_instr
*instr
)
1417 nir_deref_instr
*texture_deref_instr
= NULL
;
1419 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1420 switch (instr
->src
[i
].src_type
) {
1421 case nir_tex_src_texture_deref
:
1422 texture_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
1428 return texture_deref_instr
;
1431 static LLVMValueRef
build_tex_intrinsic(struct ac_nir_context
*ctx
,
1432 const nir_tex_instr
*instr
,
1433 struct ac_image_args
*args
)
1435 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
1436 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
1438 assert(instr
->dest
.is_ssa
);
1439 return ac_build_buffer_load_format(&ctx
->ac
,
1443 util_last_bit(mask
),
1445 instr
->dest
.ssa
.bit_size
== 16);
1448 args
->opcode
= ac_image_sample
;
1450 switch (instr
->op
) {
1452 case nir_texop_txf_ms
:
1453 case nir_texop_samples_identical
:
1454 args
->opcode
= args
->level_zero
||
1455 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
?
1456 ac_image_load
: ac_image_load_mip
;
1457 args
->level_zero
= false;
1460 case nir_texop_query_levels
:
1461 args
->opcode
= ac_image_get_resinfo
;
1463 args
->lod
= ctx
->ac
.i32_0
;
1464 args
->level_zero
= false;
1467 if (ctx
->stage
!= MESA_SHADER_FRAGMENT
) {
1469 args
->level_zero
= true;
1473 args
->opcode
= ac_image_gather4
;
1474 if (!args
->lod
&& !args
->bias
)
1475 args
->level_zero
= true;
1478 args
->opcode
= ac_image_get_lod
;
1480 case nir_texop_fragment_fetch
:
1481 case nir_texop_fragment_mask_fetch
:
1482 args
->opcode
= ac_image_load
;
1483 args
->level_zero
= false;
1489 if (instr
->op
== nir_texop_tg4
&& ctx
->ac
.chip_class
<= GFX8
) {
1490 nir_deref_instr
*texture_deref_instr
= get_tex_texture_deref(instr
);
1491 nir_variable
*var
= nir_deref_instr_get_variable(texture_deref_instr
);
1492 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1493 enum glsl_base_type stype
= glsl_get_sampler_result_type(type
);
1494 if (stype
== GLSL_TYPE_UINT
|| stype
== GLSL_TYPE_INT
) {
1495 return lower_gather4_integer(&ctx
->ac
, var
, args
, instr
);
1499 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1500 if (instr
->op
== nir_texop_lod
&& ctx
->ac
.chip_class
== GFX9
) {
1501 if ((args
->dim
== ac_image_2darray
||
1502 args
->dim
== ac_image_2d
) && !args
->coords
[1]) {
1503 args
->coords
[1] = ctx
->ac
.i32_0
;
1507 args
->attributes
= AC_FUNC_ATTR_READNONE
;
1508 bool cs_derivs
= ctx
->stage
== MESA_SHADER_COMPUTE
&&
1509 ctx
->info
->cs
.derivative_group
!= DERIVATIVE_GROUP_NONE
;
1510 if (ctx
->stage
== MESA_SHADER_FRAGMENT
|| cs_derivs
) {
1511 /* Prevent texture instructions with implicit derivatives from being
1512 * sinked into branches. */
1513 switch (instr
->op
) {
1517 args
->attributes
|= AC_FUNC_ATTR_CONVERGENT
;
1524 return ac_build_image_opcode(&ctx
->ac
, args
);
1527 static LLVMValueRef
visit_vulkan_resource_reindex(struct ac_nir_context
*ctx
,
1528 nir_intrinsic_instr
*instr
)
1530 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
1531 LLVMValueRef index
= get_src(ctx
, instr
->src
[1]);
1533 LLVMValueRef result
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
1534 LLVMSetMetadata(result
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1538 static LLVMValueRef
visit_load_push_constant(struct ac_nir_context
*ctx
,
1539 nir_intrinsic_instr
*instr
)
1541 LLVMValueRef ptr
, addr
;
1542 LLVMValueRef src0
= get_src(ctx
, instr
->src
[0]);
1543 unsigned index
= nir_intrinsic_base(instr
);
1545 addr
= LLVMConstInt(ctx
->ac
.i32
, index
, 0);
1546 addr
= LLVMBuildAdd(ctx
->ac
.builder
, addr
, src0
, "");
1548 /* Load constant values from user SGPRS when possible, otherwise
1549 * fallback to the default path that loads directly from memory.
1551 if (LLVMIsConstant(src0
) &&
1552 instr
->dest
.ssa
.bit_size
== 32) {
1553 unsigned count
= instr
->dest
.ssa
.num_components
;
1554 unsigned offset
= index
;
1556 offset
+= LLVMConstIntGetZExtValue(src0
);
1559 offset
-= ctx
->args
->base_inline_push_consts
;
1561 unsigned num_inline_push_consts
= ctx
->args
->num_inline_push_consts
;
1562 if (offset
+ count
<= num_inline_push_consts
) {
1563 LLVMValueRef push_constants
[num_inline_push_consts
];
1564 for (unsigned i
= 0; i
< num_inline_push_consts
; i
++)
1565 push_constants
[i
] = ac_get_arg(&ctx
->ac
,
1566 ctx
->args
->inline_push_consts
[i
]);
1567 return ac_build_gather_values(&ctx
->ac
,
1568 push_constants
+ offset
,
1573 ptr
= LLVMBuildGEP(ctx
->ac
.builder
,
1574 ac_get_arg(&ctx
->ac
, ctx
->args
->push_constants
), &addr
, 1, "");
1576 if (instr
->dest
.ssa
.bit_size
== 8) {
1577 unsigned load_dwords
= instr
->dest
.ssa
.num_components
> 1 ? 2 : 1;
1578 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->ac
.i8
, 4 * load_dwords
);
1579 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, vec_type
);
1580 LLVMValueRef res
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1582 LLVMValueRef params
[3];
1583 if (load_dwords
> 1) {
1584 LLVMValueRef res_vec
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, ctx
->ac
.v2i32
, "");
1585 params
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, res_vec
, LLVMConstInt(ctx
->ac
.i32
, 1, false), "");
1586 params
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, res_vec
, LLVMConstInt(ctx
->ac
.i32
, 0, false), "");
1588 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, ctx
->ac
.i32
, "");
1589 params
[0] = ctx
->ac
.i32_0
;
1593 res
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.alignbyte", ctx
->ac
.i32
, params
, 3, 0);
1595 res
= LLVMBuildTrunc(ctx
->ac
.builder
, res
, LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.num_components
* 8), "");
1596 if (instr
->dest
.ssa
.num_components
> 1)
1597 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, LLVMVectorType(ctx
->ac
.i8
, instr
->dest
.ssa
.num_components
), "");
1599 } else if (instr
->dest
.ssa
.bit_size
== 16) {
1600 unsigned load_dwords
= instr
->dest
.ssa
.num_components
/ 2 + 1;
1601 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->ac
.i16
, 2 * load_dwords
);
1602 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, vec_type
);
1603 LLVMValueRef res
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1604 res
= LLVMBuildBitCast(ctx
->ac
.builder
, res
, vec_type
, "");
1605 LLVMValueRef cond
= LLVMBuildLShr(ctx
->ac
.builder
, addr
, ctx
->ac
.i32_1
, "");
1606 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
, ctx
->ac
.i1
, "");
1607 LLVMValueRef mask
[] = { LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
1608 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
1609 LLVMConstInt(ctx
->ac
.i32
, 4, false)};
1610 LLVMValueRef swizzle_aligned
= LLVMConstVector(&mask
[0], instr
->dest
.ssa
.num_components
);
1611 LLVMValueRef swizzle_unaligned
= LLVMConstVector(&mask
[1], instr
->dest
.ssa
.num_components
);
1612 LLVMValueRef shuffle_aligned
= LLVMBuildShuffleVector(ctx
->ac
.builder
, res
, res
, swizzle_aligned
, "");
1613 LLVMValueRef shuffle_unaligned
= LLVMBuildShuffleVector(ctx
->ac
.builder
, res
, res
, swizzle_unaligned
, "");
1614 res
= LLVMBuildSelect(ctx
->ac
.builder
, cond
, shuffle_unaligned
, shuffle_aligned
, "");
1615 return LLVMBuildBitCast(ctx
->ac
.builder
, res
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
1618 ptr
= ac_cast_ptr(&ctx
->ac
, ptr
, get_def_type(ctx
, &instr
->dest
.ssa
));
1620 return LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1623 static LLVMValueRef
visit_get_buffer_size(struct ac_nir_context
*ctx
,
1624 const nir_intrinsic_instr
*instr
)
1626 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
1628 return get_buffer_size(ctx
, ctx
->abi
->load_ssbo(ctx
->abi
, index
, false), false);
1631 static uint32_t widen_mask(uint32_t mask
, unsigned multiplier
)
1633 uint32_t new_mask
= 0;
1634 for(unsigned i
= 0; i
< 32 && (1u << i
) <= mask
; ++i
)
1635 if (mask
& (1u << i
))
1636 new_mask
|= ((1u << multiplier
) - 1u) << (i
* multiplier
);
1640 static LLVMValueRef
extract_vector_range(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
1641 unsigned start
, unsigned count
)
1643 LLVMValueRef mask
[] = {
1644 ctx
->i32_0
, ctx
->i32_1
,
1645 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false) };
1647 unsigned src_elements
= ac_get_llvm_num_components(src
);
1649 if (count
== src_elements
) {
1652 } else if (count
== 1) {
1653 assert(start
< src_elements
);
1654 return LLVMBuildExtractElement(ctx
->builder
, src
, mask
[start
], "");
1656 assert(start
+ count
<= src_elements
);
1658 LLVMValueRef swizzle
= LLVMConstVector(&mask
[start
], count
);
1659 return LLVMBuildShuffleVector(ctx
->builder
, src
, src
, swizzle
, "");
1663 static unsigned get_cache_policy(struct ac_nir_context
*ctx
,
1664 enum gl_access_qualifier access
,
1665 bool may_store_unaligned
,
1666 bool writeonly_memory
)
1668 unsigned cache_policy
= 0;
1670 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1671 * store opcodes not aligned to a dword are affected. The only way to
1672 * get unaligned stores is through shader images.
1674 if (((may_store_unaligned
&& ctx
->ac
.chip_class
== GFX6
) ||
1675 /* If this is write-only, don't keep data in L1 to prevent
1676 * evicting L1 cache lines that may be needed by other
1680 access
& (ACCESS_COHERENT
| ACCESS_VOLATILE
))) {
1681 cache_policy
|= ac_glc
;
1684 if (access
& ACCESS_STREAM_CACHE_POLICY
)
1685 cache_policy
|= ac_slc
| ac_glc
;
1687 return cache_policy
;
1690 static LLVMValueRef
enter_waterfall_ssbo(struct ac_nir_context
*ctx
,
1691 struct waterfall_context
*wctx
,
1692 const nir_intrinsic_instr
*instr
,
1695 return enter_waterfall(ctx
, wctx
, get_src(ctx
, src
),
1696 nir_intrinsic_access(instr
) & ACCESS_NON_UNIFORM
);
1699 static void visit_store_ssbo(struct ac_nir_context
*ctx
,
1700 nir_intrinsic_instr
*instr
)
1702 if (ctx
->ac
.postponed_kill
) {
1703 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
1704 ctx
->ac
.postponed_kill
, "");
1705 ac_build_ifcc(&ctx
->ac
, cond
, 7000);
1708 LLVMValueRef src_data
= get_src(ctx
, instr
->src
[0]);
1709 int elem_size_bytes
= ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src_data
)) / 8;
1710 unsigned writemask
= nir_intrinsic_write_mask(instr
);
1711 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
1712 bool writeonly_memory
= access
& ACCESS_NON_READABLE
;
1713 unsigned cache_policy
= get_cache_policy(ctx
, access
, false, writeonly_memory
);
1715 struct waterfall_context wctx
;
1716 LLVMValueRef rsrc_base
= enter_waterfall_ssbo(ctx
, &wctx
, instr
, instr
->src
[1]);
1718 LLVMValueRef rsrc
= ctx
->abi
->load_ssbo(ctx
->abi
, rsrc_base
, true);
1719 LLVMValueRef base_data
= src_data
;
1720 base_data
= ac_trim_vector(&ctx
->ac
, base_data
, instr
->num_components
);
1721 LLVMValueRef base_offset
= get_src(ctx
, instr
->src
[2]);
1725 LLVMValueRef data
, offset
;
1726 LLVMTypeRef data_type
;
1728 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
1730 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1731 * writes into a 2-element and a 1-element write. */
1733 (elem_size_bytes
!= 4 || !ac_has_vec3_support(ctx
->ac
.chip_class
, false))) {
1734 writemask
|= 1 << (start
+ 2);
1737 int num_bytes
= count
* elem_size_bytes
; /* count in bytes */
1739 /* we can only store 4 DWords at the same time.
1740 * can only happen for 64 Bit vectors. */
1741 if (num_bytes
> 16) {
1742 writemask
|= ((1u << (count
- 2)) - 1u) << (start
+ 2);
1747 /* check alignment of 16 Bit stores */
1748 if (elem_size_bytes
== 2 && num_bytes
> 2 && (start
% 2) == 1) {
1749 writemask
|= ((1u << (count
- 1)) - 1u) << (start
+ 1);
1754 /* Due to alignment issues, split stores of 8-bit/16-bit
1757 if (ctx
->ac
.chip_class
== GFX6
&& count
> 1 && elem_size_bytes
< 4) {
1758 writemask
|= ((1u << (count
- 1)) - 1u) << (start
+ 1);
1760 num_bytes
= elem_size_bytes
;
1763 data
= extract_vector_range(&ctx
->ac
, base_data
, start
, count
);
1765 offset
= LLVMBuildAdd(ctx
->ac
.builder
, base_offset
,
1766 LLVMConstInt(ctx
->ac
.i32
, start
* elem_size_bytes
, false), "");
1768 if (num_bytes
== 1) {
1769 ac_build_tbuffer_store_byte(&ctx
->ac
, rsrc
, data
,
1770 offset
, ctx
->ac
.i32_0
,
1772 } else if (num_bytes
== 2) {
1773 ac_build_tbuffer_store_short(&ctx
->ac
, rsrc
, data
,
1774 offset
, ctx
->ac
.i32_0
,
1777 int num_channels
= num_bytes
/ 4;
1779 switch (num_bytes
) {
1780 case 16: /* v4f32 */
1781 data_type
= ctx
->ac
.v4f32
;
1783 case 12: /* v3f32 */
1784 data_type
= ctx
->ac
.v3f32
;
1787 data_type
= ctx
->ac
.v2f32
;
1790 data_type
= ctx
->ac
.f32
;
1793 unreachable("Malformed vector store.");
1795 data
= LLVMBuildBitCast(ctx
->ac
.builder
, data
, data_type
, "");
1797 ac_build_buffer_store_dword(&ctx
->ac
, rsrc
, data
,
1798 num_channels
, offset
,
1804 exit_waterfall(ctx
, &wctx
, NULL
);
1806 if (ctx
->ac
.postponed_kill
)
1807 ac_build_endif(&ctx
->ac
, 7000);
1810 static LLVMValueRef
emit_ssbo_comp_swap_64(struct ac_nir_context
*ctx
,
1811 LLVMValueRef descriptor
,
1812 LLVMValueRef offset
,
1813 LLVMValueRef compare
,
1814 LLVMValueRef exchange
)
1816 LLVMBasicBlockRef start_block
= NULL
, then_block
= NULL
;
1817 if (ctx
->abi
->robust_buffer_access
) {
1818 LLVMValueRef size
= ac_llvm_extract_elem(&ctx
->ac
, descriptor
, 2);
1820 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, offset
, size
, "");
1821 start_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
1823 ac_build_ifcc(&ctx
->ac
, cond
, -1);
1825 then_block
= LLVMGetInsertBlock(ctx
->ac
.builder
);
1828 LLVMValueRef ptr_parts
[2] = {
1829 ac_llvm_extract_elem(&ctx
->ac
, descriptor
, 0),
1830 LLVMBuildAnd(ctx
->ac
.builder
,
1831 ac_llvm_extract_elem(&ctx
->ac
, descriptor
, 1),
1832 LLVMConstInt(ctx
->ac
.i32
, 65535, 0), "")
1835 ptr_parts
[1] = LLVMBuildTrunc(ctx
->ac
.builder
, ptr_parts
[1], ctx
->ac
.i16
, "");
1836 ptr_parts
[1] = LLVMBuildSExt(ctx
->ac
.builder
, ptr_parts
[1], ctx
->ac
.i32
, "");
1838 offset
= LLVMBuildZExt(ctx
->ac
.builder
, offset
, ctx
->ac
.i64
, "");
1840 LLVMValueRef ptr
= ac_build_gather_values(&ctx
->ac
, ptr_parts
, 2);
1841 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
, ctx
->ac
.i64
, "");
1842 ptr
= LLVMBuildAdd(ctx
->ac
.builder
, ptr
, offset
, "");
1843 ptr
= LLVMBuildIntToPtr(ctx
->ac
.builder
, ptr
, LLVMPointerType(ctx
->ac
.i64
, AC_ADDR_SPACE_GLOBAL
), "");
1845 LLVMValueRef result
= ac_build_atomic_cmp_xchg(&ctx
->ac
, ptr
, compare
, exchange
, "singlethread-one-as");
1846 result
= LLVMBuildExtractValue(ctx
->ac
.builder
, result
, 0, "");
1848 if (ctx
->abi
->robust_buffer_access
) {
1849 ac_build_endif(&ctx
->ac
, -1);
1851 LLVMBasicBlockRef incoming_blocks
[2] = {
1856 LLVMValueRef incoming_values
[2] = {
1857 LLVMConstInt(ctx
->ac
.i64
, 0, 0),
1860 LLVMValueRef ret
= LLVMBuildPhi(ctx
->ac
.builder
, ctx
->ac
.i64
, "");
1861 LLVMAddIncoming(ret
, incoming_values
, incoming_blocks
, 2);
1868 static LLVMValueRef
visit_atomic_ssbo(struct ac_nir_context
*ctx
,
1869 nir_intrinsic_instr
*instr
)
1871 if (ctx
->ac
.postponed_kill
) {
1872 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
1873 ctx
->ac
.postponed_kill
, "");
1874 ac_build_ifcc(&ctx
->ac
, cond
, 7001);
1877 LLVMTypeRef return_type
= LLVMTypeOf(get_src(ctx
, instr
->src
[2]));
1879 char name
[64], type
[8];
1880 LLVMValueRef params
[6], descriptor
;
1881 LLVMValueRef result
;
1884 struct waterfall_context wctx
;
1885 LLVMValueRef rsrc_base
= enter_waterfall_ssbo(ctx
, &wctx
, instr
, instr
->src
[0]);
1887 switch (instr
->intrinsic
) {
1888 case nir_intrinsic_ssbo_atomic_add
:
1891 case nir_intrinsic_ssbo_atomic_imin
:
1894 case nir_intrinsic_ssbo_atomic_umin
:
1897 case nir_intrinsic_ssbo_atomic_imax
:
1900 case nir_intrinsic_ssbo_atomic_umax
:
1903 case nir_intrinsic_ssbo_atomic_and
:
1906 case nir_intrinsic_ssbo_atomic_or
:
1909 case nir_intrinsic_ssbo_atomic_xor
:
1912 case nir_intrinsic_ssbo_atomic_exchange
:
1915 case nir_intrinsic_ssbo_atomic_comp_swap
:
1922 descriptor
= ctx
->abi
->load_ssbo(ctx
->abi
,
1926 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
&&
1927 return_type
== ctx
->ac
.i64
) {
1928 result
= emit_ssbo_comp_swap_64(ctx
, descriptor
,
1929 get_src(ctx
, instr
->src
[1]),
1930 get_src(ctx
, instr
->src
[2]),
1931 get_src(ctx
, instr
->src
[3]));
1933 if (instr
->intrinsic
== nir_intrinsic_ssbo_atomic_comp_swap
) {
1934 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[3]), 0);
1936 params
[arg_count
++] = ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
1937 params
[arg_count
++] = descriptor
;
1939 if (LLVM_VERSION_MAJOR
>= 9) {
1940 /* XXX: The new raw/struct atomic intrinsics are buggy with
1941 * LLVM 8, see r358579.
1943 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1944 params
[arg_count
++] = ctx
->ac
.i32_0
; /* soffset */
1945 params
[arg_count
++] = ctx
->ac
.i32_0
; /* slc */
1947 ac_build_type_name_for_intr(return_type
, type
, sizeof(type
));
1948 snprintf(name
, sizeof(name
),
1949 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op
, type
);
1951 params
[arg_count
++] = ctx
->ac
.i32_0
; /* vindex */
1952 params
[arg_count
++] = get_src(ctx
, instr
->src
[1]); /* voffset */
1953 params
[arg_count
++] = ctx
->ac
.i1false
; /* slc */
1955 assert(return_type
== ctx
->ac
.i32
);
1956 snprintf(name
, sizeof(name
),
1957 "llvm.amdgcn.buffer.atomic.%s", op
);
1960 result
= ac_build_intrinsic(&ctx
->ac
, name
, return_type
, params
,
1964 result
= exit_waterfall(ctx
, &wctx
, result
);
1965 if (ctx
->ac
.postponed_kill
)
1966 ac_build_endif(&ctx
->ac
, 7001);
1970 static LLVMValueRef
visit_load_buffer(struct ac_nir_context
*ctx
,
1971 nir_intrinsic_instr
*instr
)
1973 struct waterfall_context wctx
;
1974 LLVMValueRef rsrc_base
= enter_waterfall_ssbo(ctx
, &wctx
, instr
, instr
->src
[0]);
1976 int elem_size_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
1977 int num_components
= instr
->num_components
;
1978 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
1979 unsigned cache_policy
= get_cache_policy(ctx
, access
, false, false);
1981 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
1982 LLVMValueRef rsrc
= ctx
->abi
->load_ssbo(ctx
->abi
, rsrc_base
, false);
1983 LLVMValueRef vindex
= ctx
->ac
.i32_0
;
1985 LLVMTypeRef def_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
1986 LLVMTypeRef def_elem_type
= num_components
> 1 ? LLVMGetElementType(def_type
) : def_type
;
1988 LLVMValueRef results
[4];
1989 for (int i
= 0; i
< num_components
;) {
1990 int num_elems
= num_components
- i
;
1991 if (elem_size_bytes
< 4 && nir_intrinsic_align(instr
) % 4 != 0)
1993 if (num_elems
* elem_size_bytes
> 16)
1994 num_elems
= 16 / elem_size_bytes
;
1995 int load_bytes
= num_elems
* elem_size_bytes
;
1997 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
, i
* elem_size_bytes
, false);
2001 if (load_bytes
== 1) {
2002 ret
= ac_build_tbuffer_load_byte(&ctx
->ac
,
2008 } else if (load_bytes
== 2) {
2009 ret
= ac_build_tbuffer_load_short(&ctx
->ac
,
2016 int num_channels
= util_next_power_of_two(load_bytes
) / 4;
2017 bool can_speculate
= access
& ACCESS_CAN_REORDER
;
2019 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_channels
,
2020 vindex
, offset
, immoffset
, 0,
2021 cache_policy
, can_speculate
, false);
2024 LLVMTypeRef byte_vec
= LLVMVectorType(ctx
->ac
.i8
, ac_get_type_size(LLVMTypeOf(ret
)));
2025 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, byte_vec
, "");
2026 ret
= ac_trim_vector(&ctx
->ac
, ret
, load_bytes
);
2028 LLVMTypeRef ret_type
= LLVMVectorType(def_elem_type
, num_elems
);
2029 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ret_type
, "");
2031 for (unsigned j
= 0; j
< num_elems
; j
++) {
2032 results
[i
+ j
] = LLVMBuildExtractElement(ctx
->ac
.builder
, ret
, LLVMConstInt(ctx
->ac
.i32
, j
, false), "");
2037 LLVMValueRef ret
= ac_build_gather_values(&ctx
->ac
, results
, num_components
);
2038 return exit_waterfall(ctx
, &wctx
, ret
);
2041 static LLVMValueRef
enter_waterfall_ubo(struct ac_nir_context
*ctx
,
2042 struct waterfall_context
*wctx
,
2043 const nir_intrinsic_instr
*instr
)
2045 return enter_waterfall(ctx
, wctx
, get_src(ctx
, instr
->src
[0]),
2046 nir_intrinsic_access(instr
) & ACCESS_NON_UNIFORM
);
2049 static LLVMValueRef
visit_load_ubo_buffer(struct ac_nir_context
*ctx
,
2050 nir_intrinsic_instr
*instr
)
2052 struct waterfall_context wctx
;
2053 LLVMValueRef rsrc_base
= enter_waterfall_ubo(ctx
, &wctx
, instr
);
2056 LLVMValueRef rsrc
= rsrc_base
;
2057 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
2058 int num_components
= instr
->num_components
;
2060 if (ctx
->abi
->load_ubo
)
2061 rsrc
= ctx
->abi
->load_ubo(ctx
->abi
, rsrc
);
2063 if (instr
->dest
.ssa
.bit_size
== 64)
2064 num_components
*= 2;
2066 if (instr
->dest
.ssa
.bit_size
== 16 || instr
->dest
.ssa
.bit_size
== 8) {
2067 unsigned load_bytes
= instr
->dest
.ssa
.bit_size
/ 8;
2068 LLVMValueRef results
[num_components
];
2069 for (unsigned i
= 0; i
< num_components
; ++i
) {
2070 LLVMValueRef immoffset
= LLVMConstInt(ctx
->ac
.i32
,
2073 if (load_bytes
== 1) {
2074 results
[i
] = ac_build_tbuffer_load_byte(&ctx
->ac
,
2081 assert(load_bytes
== 2);
2082 results
[i
] = ac_build_tbuffer_load_short(&ctx
->ac
,
2090 ret
= ac_build_gather_values(&ctx
->ac
, results
, num_components
);
2092 ret
= ac_build_buffer_load(&ctx
->ac
, rsrc
, num_components
, NULL
, offset
,
2093 NULL
, 0, 0, true, true);
2095 ret
= ac_trim_vector(&ctx
->ac
, ret
, num_components
);
2098 ret
= LLVMBuildBitCast(ctx
->ac
.builder
, ret
,
2099 get_def_type(ctx
, &instr
->dest
.ssa
), "");
2101 return exit_waterfall(ctx
, &wctx
, ret
);
2105 get_deref_offset(struct ac_nir_context
*ctx
, nir_deref_instr
*instr
,
2106 bool vs_in
, unsigned *vertex_index_out
,
2107 LLVMValueRef
*vertex_index_ref
,
2108 unsigned *const_out
, LLVMValueRef
*indir_out
)
2110 nir_variable
*var
= nir_deref_instr_get_variable(instr
);
2111 nir_deref_path path
;
2112 unsigned idx_lvl
= 1;
2114 nir_deref_path_init(&path
, instr
, NULL
);
2116 if (vertex_index_out
!= NULL
|| vertex_index_ref
!= NULL
) {
2117 if (vertex_index_ref
) {
2118 *vertex_index_ref
= get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
);
2119 if (vertex_index_out
)
2120 *vertex_index_out
= 0;
2122 *vertex_index_out
= nir_src_as_uint(path
.path
[idx_lvl
]->arr
.index
);
2127 uint32_t const_offset
= 0;
2128 LLVMValueRef offset
= NULL
;
2130 if (var
->data
.compact
) {
2131 assert(instr
->deref_type
== nir_deref_type_array
);
2132 const_offset
= nir_src_as_uint(instr
->arr
.index
);
2136 for (; path
.path
[idx_lvl
]; ++idx_lvl
) {
2137 const struct glsl_type
*parent_type
= path
.path
[idx_lvl
- 1]->type
;
2138 if (path
.path
[idx_lvl
]->deref_type
== nir_deref_type_struct
) {
2139 unsigned index
= path
.path
[idx_lvl
]->strct
.index
;
2141 for (unsigned i
= 0; i
< index
; i
++) {
2142 const struct glsl_type
*ft
= glsl_get_struct_field(parent_type
, i
);
2143 const_offset
+= glsl_count_attribute_slots(ft
, vs_in
);
2145 } else if(path
.path
[idx_lvl
]->deref_type
== nir_deref_type_array
) {
2146 unsigned size
= glsl_count_attribute_slots(path
.path
[idx_lvl
]->type
, vs_in
);
2147 if (nir_src_is_const(path
.path
[idx_lvl
]->arr
.index
)) {
2148 const_offset
+= size
*
2149 nir_src_as_uint(path
.path
[idx_lvl
]->arr
.index
);
2151 LLVMValueRef array_off
= LLVMBuildMul(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, size
, 0),
2152 get_src(ctx
, path
.path
[idx_lvl
]->arr
.index
), "");
2154 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, array_off
, "");
2159 unreachable("Uhandled deref type in get_deref_instr_offset");
2163 nir_deref_path_finish(&path
);
2165 if (const_offset
&& offset
)
2166 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
2167 LLVMConstInt(ctx
->ac
.i32
, const_offset
, 0),
2170 *const_out
= const_offset
;
2171 *indir_out
= offset
;
2174 static LLVMValueRef
load_tess_varyings(struct ac_nir_context
*ctx
,
2175 nir_intrinsic_instr
*instr
,
2178 LLVMValueRef result
;
2179 LLVMValueRef vertex_index
= NULL
;
2180 LLVMValueRef indir_index
= NULL
;
2181 unsigned const_index
= 0;
2183 nir_variable
*var
= nir_deref_instr_get_variable(nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
));
2185 unsigned location
= var
->data
.location
;
2186 unsigned driver_location
= var
->data
.driver_location
;
2187 const bool is_patch
= var
->data
.patch
||
2188 var
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
2189 var
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
;
2190 const bool is_compact
= var
->data
.compact
;
2192 get_deref_offset(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
2193 false, NULL
, is_patch
? NULL
: &vertex_index
,
2194 &const_index
, &indir_index
);
2196 LLVMTypeRef dest_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
2198 LLVMTypeRef src_component_type
;
2199 if (LLVMGetTypeKind(dest_type
) == LLVMVectorTypeKind
)
2200 src_component_type
= LLVMGetElementType(dest_type
);
2202 src_component_type
= dest_type
;
2204 result
= ctx
->abi
->load_tess_varyings(ctx
->abi
, src_component_type
,
2205 vertex_index
, indir_index
,
2206 const_index
, location
, driver_location
,
2207 var
->data
.location_frac
,
2208 instr
->num_components
,
2209 is_patch
, is_compact
, load_inputs
);
2210 if (instr
->dest
.ssa
.bit_size
== 16) {
2211 result
= ac_to_integer(&ctx
->ac
, result
);
2212 result
= LLVMBuildTrunc(ctx
->ac
.builder
, result
, dest_type
, "");
2214 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, dest_type
, "");
2218 type_scalar_size_bytes(const struct glsl_type
*type
)
2220 assert(glsl_type_is_vector_or_scalar(type
) ||
2221 glsl_type_is_matrix(type
));
2222 return glsl_type_is_boolean(type
) ? 4 : glsl_get_bit_size(type
) / 8;
2225 static LLVMValueRef
visit_load_var(struct ac_nir_context
*ctx
,
2226 nir_intrinsic_instr
*instr
)
2228 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2229 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2231 LLVMValueRef values
[8];
2233 int ve
= instr
->dest
.ssa
.num_components
;
2235 LLVMValueRef indir_index
;
2237 unsigned const_index
;
2238 unsigned stride
= 4;
2239 int mode
= deref
->mode
;
2242 bool vs_in
= ctx
->stage
== MESA_SHADER_VERTEX
&&
2243 var
->data
.mode
== nir_var_shader_in
;
2244 idx
= var
->data
.driver_location
;
2245 comp
= var
->data
.location_frac
;
2246 mode
= var
->data
.mode
;
2248 get_deref_offset(ctx
, deref
, vs_in
, NULL
, NULL
,
2249 &const_index
, &indir_index
);
2251 if (var
->data
.compact
) {
2253 const_index
+= comp
;
2258 if (instr
->dest
.ssa
.bit_size
== 64 &&
2259 (deref
->mode
== nir_var_shader_in
||
2260 deref
->mode
== nir_var_shader_out
||
2261 deref
->mode
== nir_var_function_temp
))
2265 case nir_var_shader_in
:
2266 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
2267 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2268 return load_tess_varyings(ctx
, instr
, true);
2271 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
2272 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
2273 LLVMValueRef indir_index
;
2274 unsigned const_index
, vertex_index
;
2275 get_deref_offset(ctx
, deref
, false, &vertex_index
, NULL
,
2276 &const_index
, &indir_index
);
2277 assert(indir_index
== NULL
);
2279 return ctx
->abi
->load_inputs(ctx
->abi
, var
->data
.location
,
2280 var
->data
.driver_location
,
2281 var
->data
.location_frac
,
2282 instr
->num_components
, vertex_index
, const_index
, type
);
2285 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2287 unsigned count
= glsl_count_attribute_slots(
2289 ctx
->stage
== MESA_SHADER_VERTEX
);
2291 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2292 &ctx
->ac
, ctx
->abi
->inputs
+ idx
+ chan
, count
,
2293 stride
, false, true);
2295 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2299 values
[chan
] = ctx
->abi
->inputs
[idx
+ chan
+ const_index
* stride
];
2302 case nir_var_function_temp
:
2303 for (unsigned chan
= 0; chan
< ve
; chan
++) {
2305 unsigned count
= glsl_count_attribute_slots(
2308 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2309 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2310 stride
, true, true);
2312 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2316 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, ctx
->locals
[idx
+ chan
+ const_index
* stride
], "");
2320 case nir_var_shader_out
:
2321 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2322 return load_tess_varyings(ctx
, instr
, false);
2325 if (ctx
->stage
== MESA_SHADER_FRAGMENT
&&
2326 var
->data
.fb_fetch_output
&&
2327 ctx
->abi
->emit_fbfetch
)
2328 return ctx
->abi
->emit_fbfetch(ctx
->abi
);
2330 for (unsigned chan
= comp
; chan
< ve
+ comp
; chan
++) {
2332 unsigned count
= glsl_count_attribute_slots(
2335 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2336 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2337 stride
, true, true);
2339 values
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
2343 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
,
2344 ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
],
2349 case nir_var_mem_global
: {
2350 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2351 LLVMTypeRef result_type
= get_def_type(ctx
, &instr
->dest
.ssa
);
2352 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2353 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2354 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2355 int elem_size_bytes
= ac_get_elem_bits(&ctx
->ac
, result_type
) / 8;
2356 bool split_loads
= ctx
->ac
.chip_class
== GFX6
&& elem_size_bytes
< 4;
2358 if (stride
!= natural_stride
|| split_loads
) {
2359 if (LLVMGetTypeKind(result_type
) == LLVMVectorTypeKind
)
2360 result_type
= LLVMGetElementType(result_type
);
2362 LLVMTypeRef ptr_type
= LLVMPointerType(result_type
,
2363 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2364 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2366 for (unsigned i
= 0; i
< instr
->dest
.ssa
.num_components
; ++i
) {
2367 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, i
* stride
/ natural_stride
, 0);
2368 values
[i
] = LLVMBuildLoad(ctx
->ac
.builder
,
2369 ac_build_gep_ptr(&ctx
->ac
, address
, offset
), "");
2371 if (nir_intrinsic_access(instr
) & (ACCESS_COHERENT
| ACCESS_VOLATILE
))
2372 LLVMSetOrdering(values
[i
], LLVMAtomicOrderingMonotonic
);
2374 return ac_build_gather_values(&ctx
->ac
, values
, instr
->dest
.ssa
.num_components
);
2376 LLVMTypeRef ptr_type
= LLVMPointerType(result_type
,
2377 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2378 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2379 LLVMValueRef val
= LLVMBuildLoad(ctx
->ac
.builder
, address
, "");
2381 if (nir_intrinsic_access(instr
) & (ACCESS_COHERENT
| ACCESS_VOLATILE
))
2382 LLVMSetOrdering(val
, LLVMAtomicOrderingMonotonic
);
2387 unreachable("unhandle variable mode");
2389 ret
= ac_build_varying_gather_values(&ctx
->ac
, values
, ve
, comp
);
2390 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
2394 visit_store_var(struct ac_nir_context
*ctx
,
2395 nir_intrinsic_instr
*instr
)
2397 if (ctx
->ac
.postponed_kill
) {
2398 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
2399 ctx
->ac
.postponed_kill
, "");
2400 ac_build_ifcc(&ctx
->ac
, cond
, 7002);
2403 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2404 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2406 LLVMValueRef temp_ptr
, value
;
2409 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[1]));
2410 int writemask
= instr
->const_index
[0];
2411 LLVMValueRef indir_index
;
2412 unsigned const_index
;
2415 get_deref_offset(ctx
, deref
, false,
2416 NULL
, NULL
, &const_index
, &indir_index
);
2417 idx
= var
->data
.driver_location
;
2418 comp
= var
->data
.location_frac
;
2420 if (var
->data
.compact
) {
2421 const_index
+= comp
;
2426 if (ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(src
)) == 64 &&
2427 (deref
->mode
== nir_var_shader_out
||
2428 deref
->mode
== nir_var_function_temp
)) {
2430 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2431 LLVMVectorType(ctx
->ac
.f32
, ac_get_llvm_num_components(src
) * 2),
2434 writemask
= widen_mask(writemask
, 2);
2437 writemask
= writemask
<< comp
;
2439 switch (deref
->mode
) {
2440 case nir_var_shader_out
:
2442 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
2443 LLVMValueRef vertex_index
= NULL
;
2444 LLVMValueRef indir_index
= NULL
;
2445 unsigned const_index
= 0;
2446 const bool is_patch
= var
->data
.patch
||
2447 var
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
2448 var
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
;
2450 get_deref_offset(ctx
, deref
, false, NULL
,
2451 is_patch
? NULL
: &vertex_index
,
2452 &const_index
, &indir_index
);
2454 ctx
->abi
->store_tcs_outputs(ctx
->abi
, var
,
2455 vertex_index
, indir_index
,
2456 const_index
, src
, writemask
);
2460 for (unsigned chan
= 0; chan
< 8; chan
++) {
2462 if (!(writemask
& (1 << chan
)))
2465 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- comp
);
2467 if (var
->data
.compact
)
2470 unsigned count
= glsl_count_attribute_slots(
2473 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2474 &ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
, count
,
2475 stride
, true, true);
2477 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2478 value
, indir_index
, "");
2479 build_store_values_extended(&ctx
->ac
, ctx
->abi
->outputs
+ idx
+ chan
,
2480 count
, stride
, tmp_vec
);
2483 temp_ptr
= ctx
->abi
->outputs
[idx
+ chan
+ const_index
* stride
];
2485 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2489 case nir_var_function_temp
:
2490 for (unsigned chan
= 0; chan
< 8; chan
++) {
2491 if (!(writemask
& (1 << chan
)))
2494 value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
2496 unsigned count
= glsl_count_attribute_slots(
2499 LLVMValueRef tmp_vec
= ac_build_gather_values_extended(
2500 &ctx
->ac
, ctx
->locals
+ idx
+ chan
, count
,
2503 tmp_vec
= LLVMBuildInsertElement(ctx
->ac
.builder
, tmp_vec
,
2504 value
, indir_index
, "");
2505 build_store_values_extended(&ctx
->ac
, ctx
->locals
+ idx
+ chan
,
2508 temp_ptr
= ctx
->locals
[idx
+ chan
+ const_index
* 4];
2510 LLVMBuildStore(ctx
->ac
.builder
, value
, temp_ptr
);
2515 case nir_var_mem_global
: {
2516 int writemask
= instr
->const_index
[0];
2517 LLVMValueRef address
= get_src(ctx
, instr
->src
[0]);
2518 LLVMValueRef val
= get_src(ctx
, instr
->src
[1]);
2520 unsigned explicit_stride
= glsl_get_explicit_stride(deref
->type
);
2521 unsigned natural_stride
= type_scalar_size_bytes(deref
->type
);
2522 unsigned stride
= explicit_stride
? explicit_stride
: natural_stride
;
2523 int elem_size_bytes
= ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(val
)) / 8;
2524 bool split_stores
= ctx
->ac
.chip_class
== GFX6
&& elem_size_bytes
< 4;
2526 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2527 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2528 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2530 if (writemask
== (1u << ac_get_llvm_num_components(val
)) - 1 &&
2531 stride
== natural_stride
&& !split_stores
) {
2532 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(val
),
2533 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2534 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2536 val
= LLVMBuildBitCast(ctx
->ac
.builder
, val
,
2537 LLVMGetElementType(LLVMTypeOf(address
)), "");
2538 LLVMValueRef store
= LLVMBuildStore(ctx
->ac
.builder
, val
, address
);
2540 if (nir_intrinsic_access(instr
) & (ACCESS_COHERENT
| ACCESS_VOLATILE
))
2541 LLVMSetOrdering(store
, LLVMAtomicOrderingMonotonic
);
2543 LLVMTypeRef val_type
= LLVMTypeOf(val
);
2544 if (LLVMGetTypeKind(LLVMTypeOf(val
)) == LLVMVectorTypeKind
)
2545 val_type
= LLVMGetElementType(val_type
);
2547 LLVMTypeRef ptr_type
= LLVMPointerType(val_type
,
2548 LLVMGetPointerAddressSpace(LLVMTypeOf(address
)));
2549 address
= LLVMBuildBitCast(ctx
->ac
.builder
, address
, ptr_type
, "");
2550 for (unsigned chan
= 0; chan
< 4; chan
++) {
2551 if (!(writemask
& (1 << chan
)))
2554 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, chan
* stride
/ natural_stride
, 0);
2556 LLVMValueRef ptr
= ac_build_gep_ptr(&ctx
->ac
, address
, offset
);
2557 LLVMValueRef src
= ac_llvm_extract_elem(&ctx
->ac
, val
,
2559 src
= LLVMBuildBitCast(ctx
->ac
.builder
, src
,
2560 LLVMGetElementType(LLVMTypeOf(ptr
)), "");
2561 LLVMValueRef store
= LLVMBuildStore(ctx
->ac
.builder
, src
, ptr
);
2563 if (nir_intrinsic_access(instr
) & (ACCESS_COHERENT
| ACCESS_VOLATILE
))
2564 LLVMSetOrdering(store
, LLVMAtomicOrderingMonotonic
);
2574 if (ctx
->ac
.postponed_kill
)
2575 ac_build_endif(&ctx
->ac
, 7002);
2578 static int image_type_to_components_count(enum glsl_sampler_dim dim
, bool array
)
2581 case GLSL_SAMPLER_DIM_BUF
:
2583 case GLSL_SAMPLER_DIM_1D
:
2584 return array
? 2 : 1;
2585 case GLSL_SAMPLER_DIM_2D
:
2586 return array
? 3 : 2;
2587 case GLSL_SAMPLER_DIM_MS
:
2588 return array
? 4 : 3;
2589 case GLSL_SAMPLER_DIM_3D
:
2590 case GLSL_SAMPLER_DIM_CUBE
:
2592 case GLSL_SAMPLER_DIM_RECT
:
2593 case GLSL_SAMPLER_DIM_SUBPASS
:
2595 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
2603 static LLVMValueRef
adjust_sample_index_using_fmask(struct ac_llvm_context
*ctx
,
2604 LLVMValueRef coord_x
, LLVMValueRef coord_y
,
2605 LLVMValueRef coord_z
,
2606 LLVMValueRef sample_index
,
2607 LLVMValueRef fmask_desc_ptr
)
2609 unsigned sample_chan
= coord_z
? 3 : 2;
2610 LLVMValueRef addr
[4] = {coord_x
, coord_y
, coord_z
};
2611 addr
[sample_chan
] = sample_index
;
2613 ac_apply_fmask_to_sample(ctx
, fmask_desc_ptr
, addr
, coord_z
!= NULL
);
2614 return addr
[sample_chan
];
2617 static nir_deref_instr
*get_image_deref(const nir_intrinsic_instr
*instr
)
2619 assert(instr
->src
[0].is_ssa
);
2620 return nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2623 static LLVMValueRef
get_image_descriptor(struct ac_nir_context
*ctx
,
2624 const nir_intrinsic_instr
*instr
,
2625 LLVMValueRef dynamic_index
,
2626 enum ac_descriptor_type desc_type
,
2629 nir_deref_instr
*deref_instr
=
2630 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_deref
?
2631 nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
) : NULL
;
2633 return get_sampler_desc(ctx
, deref_instr
, desc_type
, &instr
->instr
, dynamic_index
, true, write
);
2636 static void get_image_coords(struct ac_nir_context
*ctx
,
2637 const nir_intrinsic_instr
*instr
,
2638 LLVMValueRef dynamic_desc_index
,
2639 struct ac_image_args
*args
,
2640 enum glsl_sampler_dim dim
,
2643 LLVMValueRef src0
= get_src(ctx
, instr
->src
[1]);
2644 LLVMValueRef masks
[] = {
2645 LLVMConstInt(ctx
->ac
.i32
, 0, false), LLVMConstInt(ctx
->ac
.i32
, 1, false),
2646 LLVMConstInt(ctx
->ac
.i32
, 2, false), LLVMConstInt(ctx
->ac
.i32
, 3, false),
2648 LLVMValueRef sample_index
= ac_llvm_extract_elem(&ctx
->ac
, get_src(ctx
, instr
->src
[2]), 0);
2651 ASSERTED
bool add_frag_pos
= (dim
== GLSL_SAMPLER_DIM_SUBPASS
||
2652 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2653 bool is_ms
= (dim
== GLSL_SAMPLER_DIM_MS
||
2654 dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
);
2655 bool gfx9_1d
= ctx
->ac
.chip_class
== GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
;
2656 assert(!add_frag_pos
&& "Input attachments should be lowered by this point.");
2657 count
= image_type_to_components_count(dim
, is_array
);
2659 if (is_ms
&& (instr
->intrinsic
== nir_intrinsic_image_deref_load
||
2660 instr
->intrinsic
== nir_intrinsic_bindless_image_load
)) {
2661 LLVMValueRef fmask_load_address
[3];
2663 fmask_load_address
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2664 fmask_load_address
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[1], "");
2666 fmask_load_address
[2] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[2], "");
2668 fmask_load_address
[2] = NULL
;
2670 sample_index
= adjust_sample_index_using_fmask(&ctx
->ac
,
2671 fmask_load_address
[0],
2672 fmask_load_address
[1],
2673 fmask_load_address
[2],
2675 get_sampler_desc(ctx
, nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
),
2676 AC_DESC_FMASK
, &instr
->instr
, dynamic_desc_index
, true, false));
2678 if (count
== 1 && !gfx9_1d
) {
2679 if (instr
->src
[1].ssa
->num_components
)
2680 args
->coords
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
, src0
, masks
[0], "");
2682 args
->coords
[0] = src0
;
2687 for (chan
= 0; chan
< count
; ++chan
) {
2688 args
->coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, src0
, chan
);
2693 args
->coords
[2] = args
->coords
[1];
2694 args
->coords
[1] = ctx
->ac
.i32_0
;
2696 args
->coords
[1] = ctx
->ac
.i32_0
;
2699 if (ctx
->ac
.chip_class
== GFX9
&&
2700 dim
== GLSL_SAMPLER_DIM_2D
&&
2702 /* The hw can't bind a slice of a 3D image as a 2D
2703 * image, because it ignores BASE_ARRAY if the target
2704 * is 3D. The workaround is to read BASE_ARRAY and set
2705 * it as the 3rd address operand for all 2D images.
2707 LLVMValueRef first_layer
, const5
, mask
;
2709 const5
= LLVMConstInt(ctx
->ac
.i32
, 5, 0);
2710 mask
= LLVMConstInt(ctx
->ac
.i32
, S_008F24_BASE_ARRAY(~0), 0);
2711 first_layer
= LLVMBuildExtractElement(ctx
->ac
.builder
, args
->resource
, const5
, "");
2712 first_layer
= LLVMBuildAnd(ctx
->ac
.builder
, first_layer
, mask
, "");
2714 args
->coords
[count
] = first_layer
;
2720 args
->coords
[count
] = sample_index
;
2726 static LLVMValueRef
get_image_buffer_descriptor(struct ac_nir_context
*ctx
,
2727 const nir_intrinsic_instr
*instr
,
2728 LLVMValueRef dynamic_index
,
2729 bool write
, bool atomic
)
2731 LLVMValueRef rsrc
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_BUFFER
, write
);
2732 if (ctx
->ac
.chip_class
== GFX9
&& LLVM_VERSION_MAJOR
< 9 && atomic
) {
2733 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2734 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->ac
.builder
, rsrc
, LLVMConstInt(ctx
->ac
.i32
, 1, 0), "");
2735 stride
= LLVMBuildLShr(ctx
->ac
.builder
, stride
, LLVMConstInt(ctx
->ac
.i32
, 16, 0), "");
2737 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->ac
.builder
,
2738 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntUGT
, elem_count
, stride
, ""),
2739 elem_count
, stride
, "");
2741 rsrc
= LLVMBuildInsertElement(ctx
->ac
.builder
, rsrc
, new_elem_count
,
2742 LLVMConstInt(ctx
->ac
.i32
, 2, 0), "");
2747 static LLVMValueRef
enter_waterfall_image(struct ac_nir_context
*ctx
,
2748 struct waterfall_context
*wctx
,
2749 const nir_intrinsic_instr
*instr
)
2751 nir_deref_instr
*deref_instr
= NULL
;
2753 if (instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_deref
)
2754 deref_instr
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
2756 LLVMValueRef value
= get_sampler_desc_index(ctx
, deref_instr
, &instr
->instr
, true);
2757 return enter_waterfall(ctx
, wctx
, value
, nir_intrinsic_access(instr
) & ACCESS_NON_UNIFORM
);
2760 static LLVMValueRef
visit_image_load(struct ac_nir_context
*ctx
,
2761 const nir_intrinsic_instr
*instr
,
2766 enum glsl_sampler_dim dim
;
2767 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
2770 dim
= nir_intrinsic_image_dim(instr
);
2771 is_array
= nir_intrinsic_image_array(instr
);
2773 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2774 const struct glsl_type
*type
= image_deref
->type
;
2775 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2776 dim
= glsl_get_sampler_dim(type
);
2777 access
|= var
->data
.access
;
2778 is_array
= glsl_sampler_type_is_array(type
);
2781 struct waterfall_context wctx
;
2782 LLVMValueRef dynamic_index
= enter_waterfall_image(ctx
, &wctx
, instr
);
2784 struct ac_image_args args
= {};
2786 args
.cache_policy
= get_cache_policy(ctx
, access
, false, false);
2788 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2789 unsigned mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
2790 unsigned num_channels
= util_last_bit(mask
);
2791 LLVMValueRef rsrc
, vindex
;
2793 rsrc
= get_image_buffer_descriptor(ctx
, instr
, dynamic_index
, false, false);
2794 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
2797 assert(instr
->dest
.is_ssa
);
2798 bool can_speculate
= access
& ACCESS_CAN_REORDER
;
2799 res
= ac_build_buffer_load_format(&ctx
->ac
, rsrc
, vindex
,
2800 ctx
->ac
.i32_0
, num_channels
,
2803 instr
->dest
.ssa
.bit_size
== 16);
2804 res
= ac_build_expand_to_vec4(&ctx
->ac
, res
, num_channels
);
2806 res
= ac_trim_vector(&ctx
->ac
, res
, instr
->dest
.ssa
.num_components
);
2807 res
= ac_to_integer(&ctx
->ac
, res
);
2809 bool level_zero
= nir_src_is_const(instr
->src
[3]) && nir_src_as_uint(instr
->src
[3]) == 0;
2811 args
.opcode
= level_zero
? ac_image_load
: ac_image_load_mip
;
2812 args
.resource
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_IMAGE
, false);
2813 get_image_coords(ctx
, instr
, dynamic_index
, &args
, dim
, is_array
);
2814 args
.dim
= ac_get_image_dim(ctx
->ac
.chip_class
, dim
, is_array
);
2816 args
.lod
= get_src(ctx
, instr
->src
[3]);
2818 args
.attributes
= AC_FUNC_ATTR_READONLY
;
2820 assert(instr
->dest
.is_ssa
);
2821 args
.d16
= instr
->dest
.ssa
.bit_size
== 16;
2823 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
2825 return exit_waterfall(ctx
, &wctx
, res
);
2828 static void visit_image_store(struct ac_nir_context
*ctx
,
2829 const nir_intrinsic_instr
*instr
,
2832 if (ctx
->ac
.postponed_kill
) {
2833 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
2834 ctx
->ac
.postponed_kill
, "");
2835 ac_build_ifcc(&ctx
->ac
, cond
, 7003);
2838 enum glsl_sampler_dim dim
;
2839 enum gl_access_qualifier access
= nir_intrinsic_access(instr
);
2843 dim
= nir_intrinsic_image_dim(instr
);
2844 is_array
= nir_intrinsic_image_array(instr
);
2846 const nir_deref_instr
*image_deref
= get_image_deref(instr
);
2847 const struct glsl_type
*type
= image_deref
->type
;
2848 const nir_variable
*var
= nir_deref_instr_get_variable(image_deref
);
2849 dim
= glsl_get_sampler_dim(type
);
2850 access
|= var
->data
.access
;
2851 is_array
= glsl_sampler_type_is_array(type
);
2854 struct waterfall_context wctx
;
2855 LLVMValueRef dynamic_index
= enter_waterfall_image(ctx
, &wctx
, instr
);
2857 bool writeonly_memory
= access
& ACCESS_NON_READABLE
;
2858 struct ac_image_args args
= {};
2860 args
.cache_policy
= get_cache_policy(ctx
, access
, true, writeonly_memory
);
2862 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
2863 LLVMValueRef rsrc
= get_image_buffer_descriptor(ctx
, instr
, dynamic_index
, true, false);
2864 LLVMValueRef src
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2865 unsigned src_channels
= ac_get_llvm_num_components(src
);
2866 LLVMValueRef vindex
;
2868 if (src_channels
== 3)
2869 src
= ac_build_expand_to_vec4(&ctx
->ac
, src
, 3);
2871 vindex
= LLVMBuildExtractElement(ctx
->ac
.builder
,
2872 get_src(ctx
, instr
->src
[1]),
2875 ac_build_buffer_store_format(&ctx
->ac
, rsrc
, src
, vindex
,
2876 ctx
->ac
.i32_0
, args
.cache_policy
);
2878 bool level_zero
= nir_src_is_const(instr
->src
[4]) && nir_src_as_uint(instr
->src
[4]) == 0;
2880 args
.opcode
= level_zero
? ac_image_store
: ac_image_store_mip
;
2881 args
.data
[0] = ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[3]));
2882 args
.resource
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_IMAGE
, true);
2883 get_image_coords(ctx
, instr
, dynamic_index
, &args
, dim
, is_array
);
2884 args
.dim
= ac_get_image_dim(ctx
->ac
.chip_class
, dim
, is_array
);
2886 args
.lod
= get_src(ctx
, instr
->src
[4]);
2888 args
.d16
= ac_get_elem_bits(&ctx
->ac
, LLVMTypeOf(args
.data
[0])) == 16;
2890 ac_build_image_opcode(&ctx
->ac
, &args
);
2893 exit_waterfall(ctx
, &wctx
, NULL
);
2894 if (ctx
->ac
.postponed_kill
)
2895 ac_build_endif(&ctx
->ac
, 7003);
2898 static LLVMValueRef
visit_image_atomic(struct ac_nir_context
*ctx
,
2899 const nir_intrinsic_instr
*instr
,
2902 if (ctx
->ac
.postponed_kill
) {
2903 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
2904 ctx
->ac
.postponed_kill
, "");
2905 ac_build_ifcc(&ctx
->ac
, cond
, 7004);
2908 LLVMValueRef params
[7];
2909 int param_count
= 0;
2911 bool cmpswap
= instr
->intrinsic
== nir_intrinsic_image_deref_atomic_comp_swap
||
2912 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_comp_swap
;
2913 const char *atomic_name
;
2914 char intrinsic_name
[64];
2915 enum ac_atomic_op atomic_subop
;
2916 ASSERTED
int length
;
2918 enum glsl_sampler_dim dim
;
2921 if (instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_imin
||
2922 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_umin
||
2923 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_imax
||
2924 instr
->intrinsic
== nir_intrinsic_bindless_image_atomic_umax
) {
2925 ASSERTED
const GLenum format
= nir_intrinsic_format(instr
);
2926 assert(format
== GL_R32UI
|| format
== GL_R32I
);
2928 dim
= nir_intrinsic_image_dim(instr
);
2929 is_array
= nir_intrinsic_image_array(instr
);
2931 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
2932 dim
= glsl_get_sampler_dim(type
);
2933 is_array
= glsl_sampler_type_is_array(type
);
2936 struct waterfall_context wctx
;
2937 LLVMValueRef dynamic_index
= enter_waterfall_image(ctx
, &wctx
, instr
);
2939 switch (instr
->intrinsic
) {
2940 case nir_intrinsic_bindless_image_atomic_add
:
2941 case nir_intrinsic_image_deref_atomic_add
:
2942 atomic_name
= "add";
2943 atomic_subop
= ac_atomic_add
;
2945 case nir_intrinsic_bindless_image_atomic_imin
:
2946 case nir_intrinsic_image_deref_atomic_imin
:
2947 atomic_name
= "smin";
2948 atomic_subop
= ac_atomic_smin
;
2950 case nir_intrinsic_bindless_image_atomic_umin
:
2951 case nir_intrinsic_image_deref_atomic_umin
:
2952 atomic_name
= "umin";
2953 atomic_subop
= ac_atomic_umin
;
2955 case nir_intrinsic_bindless_image_atomic_imax
:
2956 case nir_intrinsic_image_deref_atomic_imax
:
2957 atomic_name
= "smax";
2958 atomic_subop
= ac_atomic_smax
;
2960 case nir_intrinsic_bindless_image_atomic_umax
:
2961 case nir_intrinsic_image_deref_atomic_umax
:
2962 atomic_name
= "umax";
2963 atomic_subop
= ac_atomic_umax
;
2965 case nir_intrinsic_bindless_image_atomic_and
:
2966 case nir_intrinsic_image_deref_atomic_and
:
2967 atomic_name
= "and";
2968 atomic_subop
= ac_atomic_and
;
2970 case nir_intrinsic_bindless_image_atomic_or
:
2971 case nir_intrinsic_image_deref_atomic_or
:
2973 atomic_subop
= ac_atomic_or
;
2975 case nir_intrinsic_bindless_image_atomic_xor
:
2976 case nir_intrinsic_image_deref_atomic_xor
:
2977 atomic_name
= "xor";
2978 atomic_subop
= ac_atomic_xor
;
2980 case nir_intrinsic_bindless_image_atomic_exchange
:
2981 case nir_intrinsic_image_deref_atomic_exchange
:
2982 atomic_name
= "swap";
2983 atomic_subop
= ac_atomic_swap
;
2985 case nir_intrinsic_bindless_image_atomic_comp_swap
:
2986 case nir_intrinsic_image_deref_atomic_comp_swap
:
2987 atomic_name
= "cmpswap";
2988 atomic_subop
= 0; /* not used */
2990 case nir_intrinsic_bindless_image_atomic_inc_wrap
:
2991 case nir_intrinsic_image_deref_atomic_inc_wrap
: {
2992 atomic_name
= "inc";
2993 atomic_subop
= ac_atomic_inc_wrap
;
2996 case nir_intrinsic_bindless_image_atomic_dec_wrap
:
2997 case nir_intrinsic_image_deref_atomic_dec_wrap
:
2998 atomic_name
= "dec";
2999 atomic_subop
= ac_atomic_dec_wrap
;
3006 params
[param_count
++] = get_src(ctx
, instr
->src
[4]);
3007 params
[param_count
++] = get_src(ctx
, instr
->src
[3]);
3009 LLVMValueRef result
;
3010 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
3011 params
[param_count
++] = get_image_buffer_descriptor(ctx
, instr
, dynamic_index
, true, true);
3012 params
[param_count
++] = LLVMBuildExtractElement(ctx
->ac
.builder
, get_src(ctx
, instr
->src
[1]),
3013 ctx
->ac
.i32_0
, ""); /* vindex */
3014 params
[param_count
++] = ctx
->ac
.i32_0
; /* voffset */
3015 if (LLVM_VERSION_MAJOR
>= 9) {
3016 /* XXX: The new raw/struct atomic intrinsics are buggy
3017 * with LLVM 8, see r358579.
3019 params
[param_count
++] = ctx
->ac
.i32_0
; /* soffset */
3020 params
[param_count
++] = ctx
->ac
.i32_0
; /* slc */
3022 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3023 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name
);
3025 params
[param_count
++] = ctx
->ac
.i1false
; /* slc */
3027 length
= snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3028 "llvm.amdgcn.buffer.atomic.%s", atomic_name
);
3031 assert(length
< sizeof(intrinsic_name
));
3032 result
= ac_build_intrinsic(&ctx
->ac
, intrinsic_name
, ctx
->ac
.i32
,
3033 params
, param_count
, 0);
3035 struct ac_image_args args
= {};
3036 args
.opcode
= cmpswap
? ac_image_atomic_cmpswap
: ac_image_atomic
;
3037 args
.atomic
= atomic_subop
;
3038 args
.data
[0] = params
[0];
3040 args
.data
[1] = params
[1];
3041 args
.resource
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_IMAGE
, true);
3042 get_image_coords(ctx
, instr
, dynamic_index
, &args
, dim
, is_array
);
3043 args
.dim
= ac_get_image_dim(ctx
->ac
.chip_class
, dim
, is_array
);
3045 result
= ac_build_image_opcode(&ctx
->ac
, &args
);
3048 result
= exit_waterfall(ctx
, &wctx
, result
);
3049 if (ctx
->ac
.postponed_kill
)
3050 ac_build_endif(&ctx
->ac
, 7004);
3054 static LLVMValueRef
visit_image_samples(struct ac_nir_context
*ctx
,
3055 nir_intrinsic_instr
*instr
)
3057 struct waterfall_context wctx
;
3058 LLVMValueRef dynamic_index
= enter_waterfall_image(ctx
, &wctx
, instr
);
3059 LLVMValueRef rsrc
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_IMAGE
, false);
3061 LLVMValueRef ret
= ac_build_image_get_sample_count(&ctx
->ac
, rsrc
);
3063 return exit_waterfall(ctx
, &wctx
, ret
);
3066 static LLVMValueRef
visit_image_size(struct ac_nir_context
*ctx
,
3067 const nir_intrinsic_instr
*instr
,
3072 enum glsl_sampler_dim dim
;
3075 dim
= nir_intrinsic_image_dim(instr
);
3076 is_array
= nir_intrinsic_image_array(instr
);
3078 const struct glsl_type
*type
= get_image_deref(instr
)->type
;
3079 dim
= glsl_get_sampler_dim(type
);
3080 is_array
= glsl_sampler_type_is_array(type
);
3083 struct waterfall_context wctx
;
3084 LLVMValueRef dynamic_index
= enter_waterfall_image(ctx
, &wctx
, instr
);
3086 if (dim
== GLSL_SAMPLER_DIM_BUF
) {
3087 res
= get_buffer_size(ctx
, get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_BUFFER
, false), true);
3090 struct ac_image_args args
= { 0 };
3092 args
.dim
= ac_get_image_dim(ctx
->ac
.chip_class
, dim
, is_array
);
3094 args
.resource
= get_image_descriptor(ctx
, instr
, dynamic_index
, AC_DESC_IMAGE
, false);
3095 args
.opcode
= ac_image_get_resinfo
;
3096 assert(nir_src_as_uint(instr
->src
[1]) == 0);
3097 args
.lod
= ctx
->ac
.i32_0
;
3098 args
.attributes
= AC_FUNC_ATTR_READNONE
;
3100 res
= ac_build_image_opcode(&ctx
->ac
, &args
);
3102 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
3104 if (dim
== GLSL_SAMPLER_DIM_CUBE
&& is_array
) {
3105 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
3106 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3107 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
3108 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, z
, two
, "");
3111 if (ctx
->ac
.chip_class
== GFX9
&& dim
== GLSL_SAMPLER_DIM_1D
&& is_array
) {
3112 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
, two
, "");
3113 res
= LLVMBuildInsertElement(ctx
->ac
.builder
, res
, layers
,
3117 return exit_waterfall(ctx
, &wctx
, res
);
3120 static void emit_membar(struct ac_llvm_context
*ac
,
3121 const nir_intrinsic_instr
*instr
)
3123 unsigned wait_flags
= 0;
3125 switch (instr
->intrinsic
) {
3126 case nir_intrinsic_memory_barrier
:
3127 case nir_intrinsic_group_memory_barrier
:
3128 wait_flags
= AC_WAIT_LGKM
| AC_WAIT_VLOAD
| AC_WAIT_VSTORE
;
3130 case nir_intrinsic_memory_barrier_buffer
:
3131 case nir_intrinsic_memory_barrier_image
:
3132 wait_flags
= AC_WAIT_VLOAD
| AC_WAIT_VSTORE
;
3134 case nir_intrinsic_memory_barrier_shared
:
3135 wait_flags
= AC_WAIT_LGKM
;
3141 ac_build_waitcnt(ac
, wait_flags
);
3144 void ac_emit_barrier(struct ac_llvm_context
*ac
, gl_shader_stage stage
)
3146 /* GFX6 only (thanks to a hw bug workaround):
3147 * The real barrier instruction isn’t needed, because an entire patch
3148 * always fits into a single wave.
3150 if (ac
->chip_class
== GFX6
&& stage
== MESA_SHADER_TESS_CTRL
) {
3151 ac_build_waitcnt(ac
, AC_WAIT_LGKM
| AC_WAIT_VLOAD
| AC_WAIT_VSTORE
);
3154 ac_build_s_barrier(ac
);
3157 static void emit_discard(struct ac_nir_context
*ctx
,
3158 const nir_intrinsic_instr
*instr
)
3162 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
3163 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3164 get_src(ctx
, instr
->src
[0]),
3167 assert(instr
->intrinsic
== nir_intrinsic_discard
);
3168 cond
= ctx
->ac
.i1false
;
3171 ac_build_kill_if_false(&ctx
->ac
, cond
);
3174 static void emit_demote(struct ac_nir_context
*ctx
,
3175 const nir_intrinsic_instr
*instr
)
3179 if (instr
->intrinsic
== nir_intrinsic_demote_if
) {
3180 cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3181 get_src(ctx
, instr
->src
[0]),
3184 assert(instr
->intrinsic
== nir_intrinsic_demote
);
3185 cond
= ctx
->ac
.i1false
;
3188 /* Kill immediately while maintaining WQM. */
3189 ac_build_kill_if_false(&ctx
->ac
, ac_build_wqm_vote(&ctx
->ac
, cond
));
3191 LLVMValueRef mask
= LLVMBuildLoad(ctx
->ac
.builder
, ctx
->ac
.postponed_kill
, "");
3192 mask
= LLVMBuildAnd(ctx
->ac
.builder
, mask
, cond
, "");
3193 LLVMBuildStore(ctx
->ac
.builder
, mask
, ctx
->ac
.postponed_kill
);
3198 visit_load_local_invocation_index(struct ac_nir_context
*ctx
)
3200 LLVMValueRef result
;
3201 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
->ac
);
3202 result
= LLVMBuildAnd(ctx
->ac
.builder
,
3203 ac_get_arg(&ctx
->ac
, ctx
->args
->tg_size
),
3204 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
3206 if (ctx
->ac
.wave_size
== 32)
3207 result
= LLVMBuildLShr(ctx
->ac
.builder
, result
,
3208 LLVMConstInt(ctx
->ac
.i32
, 1, false), "");
3210 return LLVMBuildAdd(ctx
->ac
.builder
, result
, thread_id
, "");
3214 visit_load_subgroup_id(struct ac_nir_context
*ctx
)
3216 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
3217 LLVMValueRef result
;
3218 result
= LLVMBuildAnd(ctx
->ac
.builder
,
3219 ac_get_arg(&ctx
->ac
, ctx
->args
->tg_size
),
3220 LLVMConstInt(ctx
->ac
.i32
, 0xfc0, false), "");
3221 return LLVMBuildLShr(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 6, false), "");
3223 return LLVMConstInt(ctx
->ac
.i32
, 0, false);
3228 visit_load_num_subgroups(struct ac_nir_context
*ctx
)
3230 if (ctx
->stage
== MESA_SHADER_COMPUTE
) {
3231 return LLVMBuildAnd(ctx
->ac
.builder
,
3232 ac_get_arg(&ctx
->ac
, ctx
->args
->tg_size
),
3233 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
3235 return LLVMConstInt(ctx
->ac
.i32
, 1, false);
3240 visit_first_invocation(struct ac_nir_context
*ctx
)
3242 LLVMValueRef active_set
= ac_build_ballot(&ctx
->ac
, ctx
->ac
.i32_1
);
3243 const char *intr
= ctx
->ac
.wave_size
== 32 ? "llvm.cttz.i32" : "llvm.cttz.i64";
3245 /* The second argument is whether cttz(0) should be defined, but we do not care. */
3246 LLVMValueRef args
[] = {active_set
, ctx
->ac
.i1false
};
3247 LLVMValueRef result
= ac_build_intrinsic(&ctx
->ac
, intr
,
3248 ctx
->ac
.iN_wavemask
, args
, 2,
3249 AC_FUNC_ATTR_NOUNWIND
|
3250 AC_FUNC_ATTR_READNONE
);
3252 return LLVMBuildTrunc(ctx
->ac
.builder
, result
, ctx
->ac
.i32
, "");
3256 visit_load_shared(struct ac_nir_context
*ctx
,
3257 const nir_intrinsic_instr
*instr
)
3259 LLVMValueRef values
[4], derived_ptr
, index
, ret
;
3261 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0],
3262 instr
->dest
.ssa
.bit_size
);
3264 for (int chan
= 0; chan
< instr
->num_components
; chan
++) {
3265 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
3266 derived_ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &index
, 1, "");
3267 values
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, derived_ptr
, "");
3270 ret
= ac_build_gather_values(&ctx
->ac
, values
, instr
->num_components
);
3271 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, get_def_type(ctx
, &instr
->dest
.ssa
), "");
3275 visit_store_shared(struct ac_nir_context
*ctx
,
3276 const nir_intrinsic_instr
*instr
)
3278 LLVMValueRef derived_ptr
, data
,index
;
3279 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3281 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[1],
3282 instr
->src
[0].ssa
->bit_size
);
3283 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
3285 int writemask
= nir_intrinsic_write_mask(instr
);
3286 for (int chan
= 0; chan
< 4; chan
++) {
3287 if (!(writemask
& (1 << chan
))) {
3290 data
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
);
3291 index
= LLVMConstInt(ctx
->ac
.i32
, chan
, 0);
3292 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3293 LLVMBuildStore(builder
, data
, derived_ptr
);
3297 static LLVMValueRef
visit_var_atomic(struct ac_nir_context
*ctx
,
3298 const nir_intrinsic_instr
*instr
,
3299 LLVMValueRef ptr
, int src_idx
)
3301 if (ctx
->ac
.postponed_kill
) {
3302 LLVMValueRef cond
= LLVMBuildLoad(ctx
->ac
.builder
,
3303 ctx
->ac
.postponed_kill
, "");
3304 ac_build_ifcc(&ctx
->ac
, cond
, 7005);
3307 LLVMValueRef result
;
3308 LLVMValueRef src
= get_src(ctx
, instr
->src
[src_idx
]);
3310 const char *sync_scope
= LLVM_VERSION_MAJOR
>= 9 ? "workgroup-one-as" : "workgroup";
3312 if (instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_deref
) {
3313 nir_deref_instr
*deref
= nir_instr_as_deref(instr
->src
[0].ssa
->parent_instr
);
3314 if (deref
->mode
== nir_var_mem_global
) {
3315 /* use "singlethread" sync scope to implement relaxed ordering */
3316 sync_scope
= LLVM_VERSION_MAJOR
>= 9 ? "singlethread-one-as" : "singlethread";
3318 LLVMTypeRef ptr_type
= LLVMPointerType(LLVMTypeOf(src
), LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
)));
3319 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
, ptr_type
, "");
3323 if (instr
->intrinsic
== nir_intrinsic_shared_atomic_comp_swap
||
3324 instr
->intrinsic
== nir_intrinsic_deref_atomic_comp_swap
) {
3325 LLVMValueRef src1
= get_src(ctx
, instr
->src
[src_idx
+ 1]);
3326 result
= ac_build_atomic_cmp_xchg(&ctx
->ac
, ptr
, src
, src1
, sync_scope
);
3327 result
= LLVMBuildExtractValue(ctx
->ac
.builder
, result
, 0, "");
3329 LLVMAtomicRMWBinOp op
;
3330 switch (instr
->intrinsic
) {
3331 case nir_intrinsic_shared_atomic_add
:
3332 case nir_intrinsic_deref_atomic_add
:
3333 op
= LLVMAtomicRMWBinOpAdd
;
3335 case nir_intrinsic_shared_atomic_umin
:
3336 case nir_intrinsic_deref_atomic_umin
:
3337 op
= LLVMAtomicRMWBinOpUMin
;
3339 case nir_intrinsic_shared_atomic_umax
:
3340 case nir_intrinsic_deref_atomic_umax
:
3341 op
= LLVMAtomicRMWBinOpUMax
;
3343 case nir_intrinsic_shared_atomic_imin
:
3344 case nir_intrinsic_deref_atomic_imin
:
3345 op
= LLVMAtomicRMWBinOpMin
;
3347 case nir_intrinsic_shared_atomic_imax
:
3348 case nir_intrinsic_deref_atomic_imax
:
3349 op
= LLVMAtomicRMWBinOpMax
;
3351 case nir_intrinsic_shared_atomic_and
:
3352 case nir_intrinsic_deref_atomic_and
:
3353 op
= LLVMAtomicRMWBinOpAnd
;
3355 case nir_intrinsic_shared_atomic_or
:
3356 case nir_intrinsic_deref_atomic_or
:
3357 op
= LLVMAtomicRMWBinOpOr
;
3359 case nir_intrinsic_shared_atomic_xor
:
3360 case nir_intrinsic_deref_atomic_xor
:
3361 op
= LLVMAtomicRMWBinOpXor
;
3363 case nir_intrinsic_shared_atomic_exchange
:
3364 case nir_intrinsic_deref_atomic_exchange
:
3365 op
= LLVMAtomicRMWBinOpXchg
;
3367 #if LLVM_VERSION_MAJOR >= 10
3368 case nir_intrinsic_shared_atomic_fadd
:
3369 case nir_intrinsic_deref_atomic_fadd
:
3370 op
= LLVMAtomicRMWBinOpFAdd
;
3379 if (instr
->intrinsic
== nir_intrinsic_shared_atomic_fadd
||
3380 instr
->intrinsic
== nir_intrinsic_deref_atomic_fadd
) {
3381 val
= ac_to_float(&ctx
->ac
, src
);
3383 val
= ac_to_integer(&ctx
->ac
, src
);
3386 result
= ac_build_atomic_rmw(&ctx
->ac
, op
, ptr
, val
, sync_scope
);
3389 if (ctx
->ac
.postponed_kill
)
3390 ac_build_endif(&ctx
->ac
, 7005);
3394 static LLVMValueRef
load_sample_pos(struct ac_nir_context
*ctx
)
3396 LLVMValueRef values
[2];
3397 LLVMValueRef pos
[2];
3399 pos
[0] = ac_to_float(&ctx
->ac
,
3400 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[0]));
3401 pos
[1] = ac_to_float(&ctx
->ac
,
3402 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[1]));
3404 values
[0] = ac_build_fract(&ctx
->ac
, pos
[0], 32);
3405 values
[1] = ac_build_fract(&ctx
->ac
, pos
[1], 32);
3406 return ac_build_gather_values(&ctx
->ac
, values
, 2);
3409 static LLVMValueRef
lookup_interp_param(struct ac_nir_context
*ctx
,
3410 enum glsl_interp_mode interp
, unsigned location
)
3413 case INTERP_MODE_FLAT
:
3416 case INTERP_MODE_SMOOTH
:
3417 case INTERP_MODE_NONE
:
3418 if (location
== INTERP_CENTER
)
3419 return ac_get_arg(&ctx
->ac
, ctx
->args
->persp_center
);
3420 else if (location
== INTERP_CENTROID
)
3421 return ctx
->abi
->persp_centroid
;
3422 else if (location
== INTERP_SAMPLE
)
3423 return ac_get_arg(&ctx
->ac
, ctx
->args
->persp_sample
);
3425 case INTERP_MODE_NOPERSPECTIVE
:
3426 if (location
== INTERP_CENTER
)
3427 return ac_get_arg(&ctx
->ac
, ctx
->args
->linear_center
);
3428 else if (location
== INTERP_CENTROID
)
3429 return ctx
->abi
->linear_centroid
;
3430 else if (location
== INTERP_SAMPLE
)
3431 return ac_get_arg(&ctx
->ac
, ctx
->args
->linear_sample
);
3437 static LLVMValueRef
barycentric_center(struct ac_nir_context
*ctx
,
3440 LLVMValueRef interp_param
= lookup_interp_param(ctx
, mode
, INTERP_CENTER
);
3441 return LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
, ctx
->ac
.v2i32
, "");
3444 static LLVMValueRef
barycentric_offset(struct ac_nir_context
*ctx
,
3446 LLVMValueRef offset
)
3448 LLVMValueRef interp_param
= lookup_interp_param(ctx
, mode
, INTERP_CENTER
);
3449 LLVMValueRef src_c0
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, offset
, ctx
->ac
.i32_0
, ""));
3450 LLVMValueRef src_c1
= ac_to_float(&ctx
->ac
, LLVMBuildExtractElement(ctx
->ac
.builder
, offset
, ctx
->ac
.i32_1
, ""));
3452 LLVMValueRef ij_out
[2];
3453 LLVMValueRef ddxy_out
= ac_build_ddxy_interp(&ctx
->ac
, interp_param
);
3456 * take the I then J parameters, and the DDX/Y for it, and
3457 * calculate the IJ inputs for the interpolator.
3458 * temp1 = ddx * offset/sample.x + I;
3459 * interp_param.I = ddy * offset/sample.y + temp1;
3460 * temp1 = ddx * offset/sample.x + J;
3461 * interp_param.J = ddy * offset/sample.y + temp1;
3463 for (unsigned i
= 0; i
< 2; i
++) {
3464 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->ac
.i32
, i
, false);
3465 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->ac
.i32
, i
+ 2, false);
3466 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3467 ddxy_out
, ix_ll
, "");
3468 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3469 ddxy_out
, iy_ll
, "");
3470 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3471 interp_param
, ix_ll
, "");
3472 LLVMValueRef temp1
, temp2
;
3474 interp_el
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_el
,
3477 temp1
= ac_build_fmad(&ctx
->ac
, ddx_el
, src_c0
, interp_el
);
3478 temp2
= ac_build_fmad(&ctx
->ac
, ddy_el
, src_c1
, temp1
);
3480 ij_out
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
,
3481 temp2
, ctx
->ac
.i32
, "");
3483 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
3484 return LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
, ctx
->ac
.v2i32
, "");
3487 static LLVMValueRef
barycentric_centroid(struct ac_nir_context
*ctx
,
3490 LLVMValueRef interp_param
= lookup_interp_param(ctx
, mode
, INTERP_CENTROID
);
3491 return LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
, ctx
->ac
.v2i32
, "");
3494 static LLVMValueRef
barycentric_at_sample(struct ac_nir_context
*ctx
,
3496 LLVMValueRef sample_id
)
3498 if (ctx
->abi
->interp_at_sample_force_center
)
3499 return barycentric_center(ctx
, mode
);
3501 LLVMValueRef halfval
= LLVMConstReal(ctx
->ac
.f32
, 0.5f
);
3503 /* fetch sample ID */
3504 LLVMValueRef sample_pos
= ctx
->abi
->load_sample_position(ctx
->abi
, sample_id
);
3506 LLVMValueRef src_c0
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_pos
, ctx
->ac
.i32_0
, "");
3507 src_c0
= LLVMBuildFSub(ctx
->ac
.builder
, src_c0
, halfval
, "");
3508 LLVMValueRef src_c1
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_pos
, ctx
->ac
.i32_1
, "");
3509 src_c1
= LLVMBuildFSub(ctx
->ac
.builder
, src_c1
, halfval
, "");
3510 LLVMValueRef coords
[] = { src_c0
, src_c1
};
3511 LLVMValueRef offset
= ac_build_gather_values(&ctx
->ac
, coords
, 2);
3513 return barycentric_offset(ctx
, mode
, offset
);
3517 static LLVMValueRef
barycentric_sample(struct ac_nir_context
*ctx
,
3520 LLVMValueRef interp_param
= lookup_interp_param(ctx
, mode
, INTERP_SAMPLE
);
3521 return LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
, ctx
->ac
.v2i32
, "");
3524 static LLVMValueRef
barycentric_model(struct ac_nir_context
*ctx
)
3526 return LLVMBuildBitCast(ctx
->ac
.builder
,
3527 ac_get_arg(&ctx
->ac
, ctx
->args
->pull_model
),
3531 static LLVMValueRef
load_interpolated_input(struct ac_nir_context
*ctx
,
3532 LLVMValueRef interp_param
,
3533 unsigned index
, unsigned comp_start
,
3534 unsigned num_components
,
3537 LLVMValueRef attr_number
= LLVMConstInt(ctx
->ac
.i32
, index
, false);
3538 LLVMValueRef interp_param_f
;
3540 interp_param_f
= LLVMBuildBitCast(ctx
->ac
.builder
,
3541 interp_param
, ctx
->ac
.v2f32
, "");
3542 LLVMValueRef i
= LLVMBuildExtractElement(
3543 ctx
->ac
.builder
, interp_param_f
, ctx
->ac
.i32_0
, "");
3544 LLVMValueRef j
= LLVMBuildExtractElement(
3545 ctx
->ac
.builder
, interp_param_f
, ctx
->ac
.i32_1
, "");
3547 /* Workaround for issue 2647: kill threads with infinite interpolation coeffs */
3548 if (ctx
->verified_interp
&&
3549 !_mesa_hash_table_search(ctx
->verified_interp
, interp_param
)) {
3550 LLVMValueRef args
[2];
3552 args
[1] = LLVMConstInt(ctx
->ac
.i32
, S_NAN
| Q_NAN
| N_INFINITY
| P_INFINITY
, false);
3553 LLVMValueRef cond
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.class.f32", ctx
->ac
.i1
,
3554 args
, 2, AC_FUNC_ATTR_READNONE
);
3555 ac_build_kill_if_false(&ctx
->ac
, LLVMBuildNot(ctx
->ac
.builder
, cond
, ""));
3556 _mesa_hash_table_insert(ctx
->verified_interp
, interp_param
, interp_param
);
3559 LLVMValueRef values
[4];
3560 assert(bitsize
== 16 || bitsize
== 32);
3561 for (unsigned comp
= 0; comp
< num_components
; comp
++) {
3562 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, comp_start
+ comp
, false);
3563 if (bitsize
== 16) {
3564 values
[comp
] = ac_build_fs_interp_f16(&ctx
->ac
, llvm_chan
, attr_number
,
3565 ac_get_arg(&ctx
->ac
, ctx
->args
->prim_mask
), i
, j
);
3567 values
[comp
] = ac_build_fs_interp(&ctx
->ac
, llvm_chan
, attr_number
,
3568 ac_get_arg(&ctx
->ac
, ctx
->args
->prim_mask
), i
, j
);
3572 return ac_to_integer(&ctx
->ac
, ac_build_gather_values(&ctx
->ac
, values
, num_components
));
3575 static LLVMValueRef
load_input(struct ac_nir_context
*ctx
,
3576 nir_intrinsic_instr
*instr
)
3578 unsigned offset_idx
= instr
->intrinsic
== nir_intrinsic_load_input
? 0 : 1;
3580 /* We only lower inputs for fragment shaders ATM */
3581 ASSERTED nir_const_value
*offset
= nir_src_as_const_value(instr
->src
[offset_idx
]);
3583 assert(offset
[0].i32
== 0);
3585 unsigned component
= nir_intrinsic_component(instr
);
3586 unsigned index
= nir_intrinsic_base(instr
);
3587 unsigned vertex_id
= 2; /* P0 */
3589 if (instr
->intrinsic
== nir_intrinsic_load_input_vertex
) {
3590 nir_const_value
*src0
= nir_src_as_const_value(instr
->src
[0]);
3592 switch (src0
[0].i32
) {
3603 unreachable("Invalid vertex index");
3607 LLVMValueRef attr_number
= LLVMConstInt(ctx
->ac
.i32
, index
, false);
3608 LLVMValueRef values
[8];
3610 /* Each component of a 64-bit value takes up two GL-level channels. */
3611 unsigned num_components
= instr
->dest
.ssa
.num_components
;
3612 unsigned bit_size
= instr
->dest
.ssa
.bit_size
;
3614 bit_size
== 64 ? num_components
* 2 : num_components
;
3616 for (unsigned chan
= 0; chan
< channels
; chan
++) {
3617 if (component
+ chan
> 4)
3618 attr_number
= LLVMConstInt(ctx
->ac
.i32
, index
+ 1, false);
3619 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, (component
+ chan
) % 4, false);
3620 values
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
3621 LLVMConstInt(ctx
->ac
.i32
, vertex_id
, false),
3624 ac_get_arg(&ctx
->ac
, ctx
->args
->prim_mask
));
3625 values
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i32
, "");
3626 values
[chan
] = LLVMBuildTruncOrBitCast(ctx
->ac
.builder
, values
[chan
],
3627 bit_size
== 16 ? ctx
->ac
.i16
: ctx
->ac
.i32
, "");
3630 LLVMValueRef result
= ac_build_gather_values(&ctx
->ac
, values
, channels
);
3631 if (bit_size
== 64) {
3632 LLVMTypeRef type
= num_components
== 1 ? ctx
->ac
.i64
:
3633 LLVMVectorType(ctx
->ac
.i64
, num_components
);
3634 result
= LLVMBuildBitCast(ctx
->ac
.builder
, result
, type
, "");
3639 static void visit_intrinsic(struct ac_nir_context
*ctx
,
3640 nir_intrinsic_instr
*instr
)
3642 LLVMValueRef result
= NULL
;
3644 switch (instr
->intrinsic
) {
3645 case nir_intrinsic_ballot
:
3646 result
= ac_build_ballot(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3647 if (ctx
->ac
.ballot_mask_bits
> ctx
->ac
.wave_size
)
3648 result
= LLVMBuildZExt(ctx
->ac
.builder
, result
, ctx
->ac
.iN_ballotmask
, "");
3650 case nir_intrinsic_read_invocation
:
3651 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
3652 get_src(ctx
, instr
->src
[1]));
3654 case nir_intrinsic_read_first_invocation
:
3655 result
= ac_build_readlane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), NULL
);
3657 case nir_intrinsic_load_subgroup_invocation
:
3658 result
= ac_get_thread_id(&ctx
->ac
);
3660 case nir_intrinsic_load_work_group_id
: {
3661 LLVMValueRef values
[3];
3663 for (int i
= 0; i
< 3; i
++) {
3664 values
[i
] = ctx
->args
->workgroup_ids
[i
].used
?
3665 ac_get_arg(&ctx
->ac
, ctx
->args
->workgroup_ids
[i
]) : ctx
->ac
.i32_0
;
3668 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
3671 case nir_intrinsic_load_base_vertex
:
3672 case nir_intrinsic_load_first_vertex
:
3673 result
= ctx
->abi
->load_base_vertex(ctx
->abi
);
3675 case nir_intrinsic_load_local_group_size
:
3676 result
= ctx
->abi
->load_local_group_size(ctx
->abi
);
3678 case nir_intrinsic_load_vertex_id
:
3679 result
= LLVMBuildAdd(ctx
->ac
.builder
,
3680 ac_get_arg(&ctx
->ac
, ctx
->args
->vertex_id
),
3681 ac_get_arg(&ctx
->ac
, ctx
->args
->base_vertex
), "");
3683 case nir_intrinsic_load_vertex_id_zero_base
: {
3684 result
= ctx
->abi
->vertex_id
;
3687 case nir_intrinsic_load_local_invocation_id
: {
3688 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->local_invocation_ids
);
3691 case nir_intrinsic_load_base_instance
:
3692 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->start_instance
);
3694 case nir_intrinsic_load_draw_id
:
3695 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->draw_id
);
3697 case nir_intrinsic_load_view_index
:
3698 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->view_index
);
3700 case nir_intrinsic_load_invocation_id
:
3701 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3702 result
= ac_unpack_param(&ctx
->ac
,
3703 ac_get_arg(&ctx
->ac
, ctx
->args
->tcs_rel_ids
),
3706 if (ctx
->ac
.chip_class
>= GFX10
) {
3707 result
= LLVMBuildAnd(ctx
->ac
.builder
,
3708 ac_get_arg(&ctx
->ac
, ctx
->args
->gs_invocation_id
),
3709 LLVMConstInt(ctx
->ac
.i32
, 127, 0), "");
3711 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->gs_invocation_id
);
3715 case nir_intrinsic_load_primitive_id
:
3716 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3717 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->gs_prim_id
);
3718 } else if (ctx
->stage
== MESA_SHADER_TESS_CTRL
) {
3719 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->tcs_patch_id
);
3720 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3721 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->tes_patch_id
);
3723 fprintf(stderr
, "Unknown primitive id intrinsic: %d", ctx
->stage
);
3725 case nir_intrinsic_load_sample_id
:
3726 result
= ac_unpack_param(&ctx
->ac
,
3727 ac_get_arg(&ctx
->ac
, ctx
->args
->ancillary
),
3730 case nir_intrinsic_load_sample_pos
:
3731 result
= load_sample_pos(ctx
);
3733 case nir_intrinsic_load_sample_mask_in
:
3734 result
= ctx
->abi
->load_sample_mask_in(ctx
->abi
);
3736 case nir_intrinsic_load_frag_coord
: {
3737 LLVMValueRef values
[4] = {
3738 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[0]),
3739 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[1]),
3740 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[2]),
3741 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
3742 ac_get_arg(&ctx
->ac
, ctx
->args
->frag_pos
[3]))
3744 result
= ac_to_integer(&ctx
->ac
,
3745 ac_build_gather_values(&ctx
->ac
, values
, 4));
3748 case nir_intrinsic_load_layer_id
:
3749 result
= ctx
->abi
->inputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
3751 case nir_intrinsic_load_front_face
:
3752 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->front_face
);
3754 case nir_intrinsic_load_helper_invocation
:
3755 result
= ac_build_load_helper_invocation(&ctx
->ac
);
3757 case nir_intrinsic_is_helper_invocation
:
3758 result
= ac_build_is_helper_invocation(&ctx
->ac
);
3760 case nir_intrinsic_load_color0
:
3761 result
= ctx
->abi
->color0
;
3763 case nir_intrinsic_load_color1
:
3764 result
= ctx
->abi
->color1
;
3766 case nir_intrinsic_load_user_data_amd
:
3767 assert(LLVMTypeOf(ctx
->abi
->user_data
) == ctx
->ac
.v4i32
);
3768 result
= ctx
->abi
->user_data
;
3770 case nir_intrinsic_load_instance_id
:
3771 result
= ctx
->abi
->instance_id
;
3773 case nir_intrinsic_load_num_work_groups
:
3774 result
= ac_get_arg(&ctx
->ac
, ctx
->args
->num_work_groups
);
3776 case nir_intrinsic_load_local_invocation_index
:
3777 result
= visit_load_local_invocation_index(ctx
);
3779 case nir_intrinsic_load_subgroup_id
:
3780 result
= visit_load_subgroup_id(ctx
);
3782 case nir_intrinsic_load_num_subgroups
:
3783 result
= visit_load_num_subgroups(ctx
);
3785 case nir_intrinsic_first_invocation
:
3786 result
= visit_first_invocation(ctx
);
3788 case nir_intrinsic_load_push_constant
:
3789 result
= visit_load_push_constant(ctx
, instr
);
3791 case nir_intrinsic_vulkan_resource_index
: {
3792 LLVMValueRef index
= get_src(ctx
, instr
->src
[0]);
3793 unsigned desc_set
= nir_intrinsic_desc_set(instr
);
3794 unsigned binding
= nir_intrinsic_binding(instr
);
3796 result
= ctx
->abi
->load_resource(ctx
->abi
, index
, desc_set
,
3800 case nir_intrinsic_vulkan_resource_reindex
:
3801 result
= visit_vulkan_resource_reindex(ctx
, instr
);
3803 case nir_intrinsic_store_ssbo
:
3804 visit_store_ssbo(ctx
, instr
);
3806 case nir_intrinsic_load_ssbo
:
3807 result
= visit_load_buffer(ctx
, instr
);
3809 case nir_intrinsic_ssbo_atomic_add
:
3810 case nir_intrinsic_ssbo_atomic_imin
:
3811 case nir_intrinsic_ssbo_atomic_umin
:
3812 case nir_intrinsic_ssbo_atomic_imax
:
3813 case nir_intrinsic_ssbo_atomic_umax
:
3814 case nir_intrinsic_ssbo_atomic_and
:
3815 case nir_intrinsic_ssbo_atomic_or
:
3816 case nir_intrinsic_ssbo_atomic_xor
:
3817 case nir_intrinsic_ssbo_atomic_exchange
:
3818 case nir_intrinsic_ssbo_atomic_comp_swap
:
3819 result
= visit_atomic_ssbo(ctx
, instr
);
3821 case nir_intrinsic_load_ubo
:
3822 result
= visit_load_ubo_buffer(ctx
, instr
);
3824 case nir_intrinsic_get_buffer_size
:
3825 result
= visit_get_buffer_size(ctx
, instr
);
3827 case nir_intrinsic_load_deref
:
3828 result
= visit_load_var(ctx
, instr
);
3830 case nir_intrinsic_store_deref
:
3831 visit_store_var(ctx
, instr
);
3833 case nir_intrinsic_load_shared
:
3834 result
= visit_load_shared(ctx
, instr
);
3836 case nir_intrinsic_store_shared
:
3837 visit_store_shared(ctx
, instr
);
3839 case nir_intrinsic_bindless_image_samples
:
3840 case nir_intrinsic_image_deref_samples
:
3841 result
= visit_image_samples(ctx
, instr
);
3843 case nir_intrinsic_bindless_image_load
:
3844 result
= visit_image_load(ctx
, instr
, true);
3846 case nir_intrinsic_image_deref_load
:
3847 result
= visit_image_load(ctx
, instr
, false);
3849 case nir_intrinsic_bindless_image_store
:
3850 visit_image_store(ctx
, instr
, true);
3852 case nir_intrinsic_image_deref_store
:
3853 visit_image_store(ctx
, instr
, false);
3855 case nir_intrinsic_bindless_image_atomic_add
:
3856 case nir_intrinsic_bindless_image_atomic_imin
:
3857 case nir_intrinsic_bindless_image_atomic_umin
:
3858 case nir_intrinsic_bindless_image_atomic_imax
:
3859 case nir_intrinsic_bindless_image_atomic_umax
:
3860 case nir_intrinsic_bindless_image_atomic_and
:
3861 case nir_intrinsic_bindless_image_atomic_or
:
3862 case nir_intrinsic_bindless_image_atomic_xor
:
3863 case nir_intrinsic_bindless_image_atomic_exchange
:
3864 case nir_intrinsic_bindless_image_atomic_comp_swap
:
3865 case nir_intrinsic_bindless_image_atomic_inc_wrap
:
3866 case nir_intrinsic_bindless_image_atomic_dec_wrap
:
3867 result
= visit_image_atomic(ctx
, instr
, true);
3869 case nir_intrinsic_image_deref_atomic_add
:
3870 case nir_intrinsic_image_deref_atomic_imin
:
3871 case nir_intrinsic_image_deref_atomic_umin
:
3872 case nir_intrinsic_image_deref_atomic_imax
:
3873 case nir_intrinsic_image_deref_atomic_umax
:
3874 case nir_intrinsic_image_deref_atomic_and
:
3875 case nir_intrinsic_image_deref_atomic_or
:
3876 case nir_intrinsic_image_deref_atomic_xor
:
3877 case nir_intrinsic_image_deref_atomic_exchange
:
3878 case nir_intrinsic_image_deref_atomic_comp_swap
:
3879 case nir_intrinsic_image_deref_atomic_inc_wrap
:
3880 case nir_intrinsic_image_deref_atomic_dec_wrap
:
3881 result
= visit_image_atomic(ctx
, instr
, false);
3883 case nir_intrinsic_bindless_image_size
:
3884 result
= visit_image_size(ctx
, instr
, true);
3886 case nir_intrinsic_image_deref_size
:
3887 result
= visit_image_size(ctx
, instr
, false);
3889 case nir_intrinsic_shader_clock
:
3890 result
= ac_build_shader_clock(&ctx
->ac
,
3891 nir_intrinsic_memory_scope(instr
));
3893 case nir_intrinsic_discard
:
3894 case nir_intrinsic_discard_if
:
3895 emit_discard(ctx
, instr
);
3897 case nir_intrinsic_demote
:
3898 case nir_intrinsic_demote_if
:
3899 emit_demote(ctx
, instr
);
3901 case nir_intrinsic_memory_barrier
:
3902 case nir_intrinsic_group_memory_barrier
:
3903 case nir_intrinsic_memory_barrier_buffer
:
3904 case nir_intrinsic_memory_barrier_image
:
3905 case nir_intrinsic_memory_barrier_shared
:
3906 emit_membar(&ctx
->ac
, instr
);
3908 case nir_intrinsic_scoped_barrier
: {
3909 assert(!(nir_intrinsic_memory_semantics(instr
) &
3910 (NIR_MEMORY_MAKE_AVAILABLE
| NIR_MEMORY_MAKE_VISIBLE
)));
3912 nir_variable_mode modes
= nir_intrinsic_memory_modes(instr
);
3914 unsigned wait_flags
= 0;
3915 if (modes
& (nir_var_mem_global
| nir_var_mem_ssbo
))
3916 wait_flags
|= AC_WAIT_VLOAD
| AC_WAIT_VSTORE
;
3917 if (modes
& nir_var_mem_shared
)
3918 wait_flags
|= AC_WAIT_LGKM
;
3921 ac_build_waitcnt(&ctx
->ac
, wait_flags
);
3923 if (nir_intrinsic_execution_scope(instr
) == NIR_SCOPE_WORKGROUP
)
3924 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
3927 case nir_intrinsic_memory_barrier_tcs_patch
:
3929 case nir_intrinsic_control_barrier
:
3930 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
3932 case nir_intrinsic_shared_atomic_add
:
3933 case nir_intrinsic_shared_atomic_imin
:
3934 case nir_intrinsic_shared_atomic_umin
:
3935 case nir_intrinsic_shared_atomic_imax
:
3936 case nir_intrinsic_shared_atomic_umax
:
3937 case nir_intrinsic_shared_atomic_and
:
3938 case nir_intrinsic_shared_atomic_or
:
3939 case nir_intrinsic_shared_atomic_xor
:
3940 case nir_intrinsic_shared_atomic_exchange
:
3941 case nir_intrinsic_shared_atomic_comp_swap
:
3942 case nir_intrinsic_shared_atomic_fadd
: {
3943 LLVMValueRef ptr
= get_memory_ptr(ctx
, instr
->src
[0],
3944 instr
->src
[1].ssa
->bit_size
);
3945 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3948 case nir_intrinsic_deref_atomic_add
:
3949 case nir_intrinsic_deref_atomic_imin
:
3950 case nir_intrinsic_deref_atomic_umin
:
3951 case nir_intrinsic_deref_atomic_imax
:
3952 case nir_intrinsic_deref_atomic_umax
:
3953 case nir_intrinsic_deref_atomic_and
:
3954 case nir_intrinsic_deref_atomic_or
:
3955 case nir_intrinsic_deref_atomic_xor
:
3956 case nir_intrinsic_deref_atomic_exchange
:
3957 case nir_intrinsic_deref_atomic_comp_swap
:
3958 case nir_intrinsic_deref_atomic_fadd
: {
3959 LLVMValueRef ptr
= get_src(ctx
, instr
->src
[0]);
3960 result
= visit_var_atomic(ctx
, instr
, ptr
, 1);
3963 case nir_intrinsic_load_barycentric_pixel
:
3964 result
= barycentric_center(ctx
, nir_intrinsic_interp_mode(instr
));
3966 case nir_intrinsic_load_barycentric_centroid
:
3967 result
= barycentric_centroid(ctx
, nir_intrinsic_interp_mode(instr
));
3969 case nir_intrinsic_load_barycentric_sample
:
3970 result
= barycentric_sample(ctx
, nir_intrinsic_interp_mode(instr
));
3972 case nir_intrinsic_load_barycentric_model
:
3973 result
= barycentric_model(ctx
);
3975 case nir_intrinsic_load_barycentric_at_offset
: {
3976 LLVMValueRef offset
= ac_to_float(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
3977 result
= barycentric_offset(ctx
, nir_intrinsic_interp_mode(instr
), offset
);
3980 case nir_intrinsic_load_barycentric_at_sample
: {
3981 LLVMValueRef sample_id
= get_src(ctx
, instr
->src
[0]);
3982 result
= barycentric_at_sample(ctx
, nir_intrinsic_interp_mode(instr
), sample_id
);
3985 case nir_intrinsic_load_interpolated_input
: {
3986 /* We assume any indirect loads have been lowered away */
3987 ASSERTED nir_const_value
*offset
= nir_src_as_const_value(instr
->src
[1]);
3989 assert(offset
[0].i32
== 0);
3991 LLVMValueRef interp_param
= get_src(ctx
, instr
->src
[0]);
3992 unsigned index
= nir_intrinsic_base(instr
);
3993 unsigned component
= nir_intrinsic_component(instr
);
3994 result
= load_interpolated_input(ctx
, interp_param
, index
,
3996 instr
->dest
.ssa
.num_components
,
3997 instr
->dest
.ssa
.bit_size
);
4000 case nir_intrinsic_load_input
:
4001 case nir_intrinsic_load_input_vertex
:
4002 result
= load_input(ctx
, instr
);
4004 case nir_intrinsic_emit_vertex
:
4005 ctx
->abi
->emit_vertex(ctx
->abi
, nir_intrinsic_stream_id(instr
), ctx
->abi
->outputs
);
4007 case nir_intrinsic_emit_vertex_with_counter
: {
4008 unsigned stream
= nir_intrinsic_stream_id(instr
);
4009 LLVMValueRef next_vertex
= get_src(ctx
, instr
->src
[0]);
4010 ctx
->abi
->emit_vertex_with_counter(ctx
->abi
, stream
,
4015 case nir_intrinsic_end_primitive
:
4016 case nir_intrinsic_end_primitive_with_counter
:
4017 ctx
->abi
->emit_primitive(ctx
->abi
, nir_intrinsic_stream_id(instr
));
4019 case nir_intrinsic_load_tess_coord
:
4020 result
= ctx
->abi
->load_tess_coord(ctx
->abi
);
4022 case nir_intrinsic_load_tess_level_outer
:
4023 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
, false);
4025 case nir_intrinsic_load_tess_level_inner
:
4026 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
, false);
4028 case nir_intrinsic_load_tess_level_outer_default
:
4029 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_OUTER
, true);
4031 case nir_intrinsic_load_tess_level_inner_default
:
4032 result
= ctx
->abi
->load_tess_level(ctx
->abi
, VARYING_SLOT_TESS_LEVEL_INNER
, true);
4034 case nir_intrinsic_load_patch_vertices_in
:
4035 result
= ctx
->abi
->load_patch_vertices_in(ctx
->abi
);
4037 case nir_intrinsic_vote_all
: {
4038 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4039 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4042 case nir_intrinsic_vote_any
: {
4043 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4044 result
= LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->ac
.i32
, "");
4047 case nir_intrinsic_shuffle
:
4048 if (ctx
->ac
.chip_class
== GFX8
||
4049 ctx
->ac
.chip_class
== GFX9
||
4050 (ctx
->ac
.chip_class
>= GFX10
&& ctx
->ac
.wave_size
== 32)) {
4051 result
= ac_build_shuffle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
4052 get_src(ctx
, instr
->src
[1]));
4054 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
4055 LLVMValueRef index
= get_src(ctx
, instr
->src
[1]);
4056 LLVMTypeRef type
= LLVMTypeOf(src
);
4057 struct waterfall_context wctx
;
4058 LLVMValueRef index_val
;
4060 index_val
= enter_waterfall(ctx
, &wctx
, index
, true);
4062 src
= LLVMBuildZExt(ctx
->ac
.builder
, src
,
4065 result
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.readlane",
4067 (LLVMValueRef
[]) { src
, index_val
}, 2,
4068 AC_FUNC_ATTR_READNONE
|
4069 AC_FUNC_ATTR_CONVERGENT
);
4071 result
= LLVMBuildTrunc(ctx
->ac
.builder
, result
, type
, "");
4073 result
= exit_waterfall(ctx
, &wctx
, result
);
4076 case nir_intrinsic_reduce
:
4077 result
= ac_build_reduce(&ctx
->ac
,
4078 get_src(ctx
, instr
->src
[0]),
4079 instr
->const_index
[0],
4080 instr
->const_index
[1]);
4082 case nir_intrinsic_inclusive_scan
:
4083 result
= ac_build_inclusive_scan(&ctx
->ac
,
4084 get_src(ctx
, instr
->src
[0]),
4085 instr
->const_index
[0]);
4087 case nir_intrinsic_exclusive_scan
:
4088 result
= ac_build_exclusive_scan(&ctx
->ac
,
4089 get_src(ctx
, instr
->src
[0]),
4090 instr
->const_index
[0]);
4092 case nir_intrinsic_quad_broadcast
: {
4093 unsigned lane
= nir_src_as_uint(instr
->src
[1]);
4094 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
4095 lane
, lane
, lane
, lane
);
4098 case nir_intrinsic_quad_swap_horizontal
:
4099 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 1, 0, 3 ,2);
4101 case nir_intrinsic_quad_swap_vertical
:
4102 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 2, 3, 0 ,1);
4104 case nir_intrinsic_quad_swap_diagonal
:
4105 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), 3, 2, 1 ,0);
4107 case nir_intrinsic_quad_swizzle_amd
: {
4108 uint32_t mask
= nir_intrinsic_swizzle_mask(instr
);
4109 result
= ac_build_quad_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
4110 mask
& 0x3, (mask
>> 2) & 0x3,
4111 (mask
>> 4) & 0x3, (mask
>> 6) & 0x3);
4114 case nir_intrinsic_masked_swizzle_amd
: {
4115 uint32_t mask
= nir_intrinsic_swizzle_mask(instr
);
4116 result
= ac_build_ds_swizzle(&ctx
->ac
, get_src(ctx
, instr
->src
[0]), mask
);
4119 case nir_intrinsic_write_invocation_amd
:
4120 result
= ac_build_writelane(&ctx
->ac
, get_src(ctx
, instr
->src
[0]),
4121 get_src(ctx
, instr
->src
[1]),
4122 get_src(ctx
, instr
->src
[2]));
4124 case nir_intrinsic_mbcnt_amd
:
4125 result
= ac_build_mbcnt(&ctx
->ac
, get_src(ctx
, instr
->src
[0]));
4127 case nir_intrinsic_load_scratch
: {
4128 LLVMValueRef offset
= get_src(ctx
, instr
->src
[0]);
4129 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->scratch
,
4131 LLVMTypeRef comp_type
=
4132 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
4133 LLVMTypeRef vec_type
=
4134 instr
->dest
.ssa
.num_components
== 1 ? comp_type
:
4135 LLVMVectorType(comp_type
, instr
->dest
.ssa
.num_components
);
4136 unsigned addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
4137 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
4138 LLVMPointerType(vec_type
, addr_space
), "");
4139 result
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
4142 case nir_intrinsic_store_scratch
: {
4143 LLVMValueRef offset
= get_src(ctx
, instr
->src
[1]);
4144 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->scratch
,
4146 LLVMTypeRef comp_type
=
4147 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->src
[0].ssa
->bit_size
);
4148 unsigned addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
4149 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
4150 LLVMPointerType(comp_type
, addr_space
), "");
4151 LLVMValueRef src
= get_src(ctx
, instr
->src
[0]);
4152 unsigned wrmask
= nir_intrinsic_write_mask(instr
);
4155 u_bit_scan_consecutive_range(&wrmask
, &start
, &count
);
4157 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, start
, false);
4158 LLVMValueRef offset_ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
, &offset
, 1, "");
4159 LLVMTypeRef vec_type
=
4160 count
== 1 ? comp_type
: LLVMVectorType(comp_type
, count
);
4161 offset_ptr
= LLVMBuildBitCast(ctx
->ac
.builder
,
4163 LLVMPointerType(vec_type
, addr_space
),
4165 LLVMValueRef offset_src
=
4166 ac_extract_components(&ctx
->ac
, src
, start
, count
);
4167 LLVMBuildStore(ctx
->ac
.builder
, offset_src
, offset_ptr
);
4171 case nir_intrinsic_load_constant
: {
4172 unsigned base
= nir_intrinsic_base(instr
);
4173 unsigned range
= nir_intrinsic_range(instr
);
4175 LLVMValueRef offset
= get_src(ctx
, instr
->src
[0]);
4176 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
,
4177 LLVMConstInt(ctx
->ac
.i32
, base
, false), "");
4179 /* Clamp the offset to avoid out-of-bound access because global
4180 * instructions can't handle them.
4182 LLVMValueRef size
= LLVMConstInt(ctx
->ac
.i32
, base
+ range
, false);
4183 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
4185 offset
= LLVMBuildSelect(ctx
->ac
.builder
, cond
, offset
, size
, "");
4187 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->constant_data
,
4189 LLVMTypeRef comp_type
=
4190 LLVMIntTypeInContext(ctx
->ac
.context
, instr
->dest
.ssa
.bit_size
);
4191 LLVMTypeRef vec_type
=
4192 instr
->dest
.ssa
.num_components
== 1 ? comp_type
:
4193 LLVMVectorType(comp_type
, instr
->dest
.ssa
.num_components
);
4194 unsigned addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
4195 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
4196 LLVMPointerType(vec_type
, addr_space
), "");
4197 result
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
4201 fprintf(stderr
, "Unknown intrinsic: ");
4202 nir_print_instr(&instr
->instr
, stderr
);
4203 fprintf(stderr
, "\n");
4207 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
4211 static LLVMValueRef
get_bindless_index_from_uniform(struct ac_nir_context
*ctx
,
4212 unsigned base_index
,
4213 unsigned constant_index
,
4214 LLVMValueRef dynamic_index
)
4216 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, base_index
* 4, 0);
4217 LLVMValueRef index
= LLVMBuildAdd(ctx
->ac
.builder
, dynamic_index
,
4218 LLVMConstInt(ctx
->ac
.i32
, constant_index
, 0), "");
4220 /* Bindless uniforms are 64bit so multiple index by 8 */
4221 index
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i32
, 8, 0), "");
4222 offset
= LLVMBuildAdd(ctx
->ac
.builder
, offset
, index
, "");
4224 LLVMValueRef ubo_index
= ctx
->abi
->load_ubo(ctx
->abi
, ctx
->ac
.i32_0
);
4226 LLVMValueRef ret
= ac_build_buffer_load(&ctx
->ac
, ubo_index
, 1, NULL
, offset
,
4227 NULL
, 0, 0, true, true);
4229 return LLVMBuildBitCast(ctx
->ac
.builder
, ret
, ctx
->ac
.i32
, "");
4232 struct sampler_desc_address
{
4233 unsigned descriptor_set
;
4234 unsigned base_index
; /* binding in vulkan */
4235 unsigned constant_index
;
4236 LLVMValueRef dynamic_index
;
4241 static struct sampler_desc_address
4242 get_sampler_desc_internal(struct ac_nir_context
*ctx
,
4243 nir_deref_instr
*deref_instr
,
4244 const nir_instr
*instr
,
4247 LLVMValueRef index
= NULL
;
4248 unsigned constant_index
= 0;
4249 unsigned descriptor_set
;
4250 unsigned base_index
;
4251 bool bindless
= false;
4256 nir_intrinsic_instr
*img_instr
= nir_instr_as_intrinsic(instr
);
4259 index
= get_src(ctx
, img_instr
->src
[0]);
4261 nir_tex_instr
*tex_instr
= nir_instr_as_tex(instr
);
4262 int sampSrcIdx
= nir_tex_instr_src_index(tex_instr
,
4263 nir_tex_src_sampler_handle
);
4264 if (sampSrcIdx
!= -1) {
4267 index
= get_src(ctx
, tex_instr
->src
[sampSrcIdx
].src
);
4269 assert(tex_instr
&& !image
);
4270 base_index
= tex_instr
->sampler_index
;
4274 while(deref_instr
->deref_type
!= nir_deref_type_var
) {
4275 if (deref_instr
->deref_type
== nir_deref_type_array
) {
4276 unsigned array_size
= glsl_get_aoa_size(deref_instr
->type
);
4280 if (nir_src_is_const(deref_instr
->arr
.index
)) {
4281 constant_index
+= array_size
* nir_src_as_uint(deref_instr
->arr
.index
);
4283 LLVMValueRef indirect
= get_src(ctx
, deref_instr
->arr
.index
);
4285 indirect
= LLVMBuildMul(ctx
->ac
.builder
, indirect
,
4286 LLVMConstInt(ctx
->ac
.i32
, array_size
, false), "");
4291 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
, indirect
, "");
4294 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
4295 } else if (deref_instr
->deref_type
== nir_deref_type_struct
) {
4296 unsigned sidx
= deref_instr
->strct
.index
;
4297 deref_instr
= nir_src_as_deref(deref_instr
->parent
);
4298 constant_index
+= glsl_get_struct_location_offset(deref_instr
->type
, sidx
);
4300 unreachable("Unsupported deref type");
4303 descriptor_set
= deref_instr
->var
->data
.descriptor_set
;
4305 if (deref_instr
->var
->data
.bindless
) {
4306 /* For now just assert on unhandled variable types */
4307 assert(deref_instr
->var
->data
.mode
== nir_var_uniform
);
4309 base_index
= deref_instr
->var
->data
.driver_location
;
4312 index
= index
? index
: ctx
->ac
.i32_0
;
4313 index
= get_bindless_index_from_uniform(ctx
, base_index
,
4314 constant_index
, index
);
4316 base_index
= deref_instr
->var
->data
.binding
;
4318 return (struct sampler_desc_address
) {
4319 .descriptor_set
= descriptor_set
,
4320 .base_index
= base_index
,
4321 .constant_index
= constant_index
,
4322 .dynamic_index
= index
,
4324 .bindless
= bindless
,
4328 /* Extract any possibly divergent index into a separate value that can be fed
4329 * into get_sampler_desc with the same arguments. */
4330 static LLVMValueRef
get_sampler_desc_index(struct ac_nir_context
*ctx
,
4331 nir_deref_instr
*deref_instr
,
4332 const nir_instr
*instr
,
4335 struct sampler_desc_address addr
= get_sampler_desc_internal(ctx
, deref_instr
, instr
, image
);
4336 return addr
.dynamic_index
;
4339 static LLVMValueRef
get_sampler_desc(struct ac_nir_context
*ctx
,
4340 nir_deref_instr
*deref_instr
,
4341 enum ac_descriptor_type desc_type
,
4342 const nir_instr
*instr
,
4344 bool image
, bool write
)
4346 struct sampler_desc_address addr
= get_sampler_desc_internal(ctx
, deref_instr
, instr
, image
);
4347 return ctx
->abi
->load_sampler_desc(ctx
->abi
,
4348 addr
.descriptor_set
,
4350 addr
.constant_index
, index
,
4351 desc_type
, addr
.image
, write
, addr
.bindless
);
4354 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4357 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4358 * filtering manually. The driver sets img7 to a mask clearing
4359 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4360 * s_and_b32 samp0, samp0, img7
4363 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4365 static LLVMValueRef
sici_fix_sampler_aniso(struct ac_nir_context
*ctx
,
4366 LLVMValueRef res
, LLVMValueRef samp
)
4368 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4369 LLVMValueRef img7
, samp0
;
4371 if (ctx
->ac
.chip_class
>= GFX8
)
4374 img7
= LLVMBuildExtractElement(builder
, res
,
4375 LLVMConstInt(ctx
->ac
.i32
, 7, 0), "");
4376 samp0
= LLVMBuildExtractElement(builder
, samp
,
4377 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4378 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4379 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4380 LLVMConstInt(ctx
->ac
.i32
, 0, 0), "");
4383 static void tex_fetch_ptrs(struct ac_nir_context
*ctx
,
4384 nir_tex_instr
*instr
,
4385 struct waterfall_context
*wctx
,
4386 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
,
4387 LLVMValueRef
*fmask_ptr
)
4389 nir_deref_instr
*texture_deref_instr
= NULL
;
4390 nir_deref_instr
*sampler_deref_instr
= NULL
;
4393 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4394 switch (instr
->src
[i
].src_type
) {
4395 case nir_tex_src_texture_deref
:
4396 texture_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
4398 case nir_tex_src_sampler_deref
:
4399 sampler_deref_instr
= nir_src_as_deref(instr
->src
[i
].src
);
4401 case nir_tex_src_plane
:
4402 plane
= nir_src_as_int(instr
->src
[i
].src
);
4409 LLVMValueRef texture_dynamic_index
= get_sampler_desc_index(ctx
, texture_deref_instr
,
4410 &instr
->instr
, false);
4411 if (!sampler_deref_instr
)
4412 sampler_deref_instr
= texture_deref_instr
;
4414 LLVMValueRef sampler_dynamic_index
= get_sampler_desc_index(ctx
, sampler_deref_instr
,
4415 &instr
->instr
, false);
4416 if (instr
->texture_non_uniform
)
4417 texture_dynamic_index
= enter_waterfall(ctx
, wctx
+ 0, texture_dynamic_index
, true);
4419 if (instr
->sampler_non_uniform
)
4420 sampler_dynamic_index
= enter_waterfall(ctx
, wctx
+ 1, sampler_dynamic_index
, true);
4422 enum ac_descriptor_type main_descriptor
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
? AC_DESC_BUFFER
: AC_DESC_IMAGE
;
4425 assert(instr
->op
!= nir_texop_txf_ms
&&
4426 instr
->op
!= nir_texop_samples_identical
);
4427 assert(instr
->sampler_dim
!= GLSL_SAMPLER_DIM_BUF
);
4429 main_descriptor
= AC_DESC_PLANE_0
+ plane
;
4432 if (instr
->op
== nir_texop_fragment_mask_fetch
) {
4433 /* The fragment mask is fetched from the compressed
4434 * multisampled surface.
4436 main_descriptor
= AC_DESC_FMASK
;
4439 *res_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, main_descriptor
, &instr
->instr
,
4440 texture_dynamic_index
, false, false);
4443 *samp_ptr
= get_sampler_desc(ctx
, sampler_deref_instr
, AC_DESC_SAMPLER
, &instr
->instr
,
4444 sampler_dynamic_index
, false, false);
4445 if (instr
->sampler_dim
< GLSL_SAMPLER_DIM_RECT
)
4446 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4448 if (fmask_ptr
&& (instr
->op
== nir_texop_txf_ms
||
4449 instr
->op
== nir_texop_samples_identical
))
4450 *fmask_ptr
= get_sampler_desc(ctx
, texture_deref_instr
, AC_DESC_FMASK
,
4451 &instr
->instr
, texture_dynamic_index
, false, false);
4454 static LLVMValueRef
apply_round_slice(struct ac_llvm_context
*ctx
,
4457 coord
= ac_to_float(ctx
, coord
);
4458 coord
= ac_build_round(ctx
, coord
);
4459 coord
= ac_to_integer(ctx
, coord
);
4463 static void visit_tex(struct ac_nir_context
*ctx
, nir_tex_instr
*instr
)
4465 LLVMValueRef result
= NULL
;
4466 struct ac_image_args args
= { 0 };
4467 LLVMValueRef fmask_ptr
= NULL
, sample_index
= NULL
;
4468 LLVMValueRef ddx
= NULL
, ddy
= NULL
;
4469 unsigned offset_src
= 0;
4470 struct waterfall_context wctx
[2] = {{{0}}};
4472 tex_fetch_ptrs(ctx
, instr
, wctx
, &args
.resource
, &args
.sampler
, &fmask_ptr
);
4474 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4475 switch (instr
->src
[i
].src_type
) {
4476 case nir_tex_src_coord
: {
4477 LLVMValueRef coord
= get_src(ctx
, instr
->src
[i
].src
);
4478 for (unsigned chan
= 0; chan
< instr
->coord_components
; ++chan
)
4479 args
.coords
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, coord
, chan
);
4482 case nir_tex_src_projector
:
4484 case nir_tex_src_comparator
:
4485 if (instr
->is_shadow
) {
4486 args
.compare
= get_src(ctx
, instr
->src
[i
].src
);
4487 args
.compare
= ac_to_float(&ctx
->ac
, args
.compare
);
4490 case nir_tex_src_offset
:
4491 args
.offset
= get_src(ctx
, instr
->src
[i
].src
);
4494 case nir_tex_src_bias
:
4495 args
.bias
= get_src(ctx
, instr
->src
[i
].src
);
4497 case nir_tex_src_lod
: {
4498 if (nir_src_is_const(instr
->src
[i
].src
) && nir_src_as_uint(instr
->src
[i
].src
) == 0)
4499 args
.level_zero
= true;
4501 args
.lod
= get_src(ctx
, instr
->src
[i
].src
);
4504 case nir_tex_src_ms_index
:
4505 sample_index
= get_src(ctx
, instr
->src
[i
].src
);
4507 case nir_tex_src_ms_mcs
:
4509 case nir_tex_src_ddx
:
4510 ddx
= get_src(ctx
, instr
->src
[i
].src
);
4512 case nir_tex_src_ddy
:
4513 ddy
= get_src(ctx
, instr
->src
[i
].src
);
4515 case nir_tex_src_min_lod
:
4516 args
.min_lod
= get_src(ctx
, instr
->src
[i
].src
);
4518 case nir_tex_src_texture_offset
:
4519 case nir_tex_src_sampler_offset
:
4520 case nir_tex_src_plane
:
4526 if (instr
->op
== nir_texop_txs
&& instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
) {
4527 result
= get_buffer_size(ctx
, args
.resource
, true);
4531 if (instr
->op
== nir_texop_texture_samples
) {
4532 LLVMValueRef res
, samples
, is_msaa
;
4533 LLVMValueRef default_sample
;
4535 res
= LLVMBuildBitCast(ctx
->ac
.builder
, args
.resource
, ctx
->ac
.v8i32
, "");
4536 samples
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4537 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4538 is_msaa
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4539 LLVMConstInt(ctx
->ac
.i32
, 28, false), "");
4540 is_msaa
= LLVMBuildAnd(ctx
->ac
.builder
, is_msaa
,
4541 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4542 is_msaa
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, is_msaa
,
4543 LLVMConstInt(ctx
->ac
.i32
, 0xe, false), "");
4545 samples
= LLVMBuildLShr(ctx
->ac
.builder
, samples
,
4546 LLVMConstInt(ctx
->ac
.i32
, 16, false), "");
4547 samples
= LLVMBuildAnd(ctx
->ac
.builder
, samples
,
4548 LLVMConstInt(ctx
->ac
.i32
, 0xf, false), "");
4549 samples
= LLVMBuildShl(ctx
->ac
.builder
, ctx
->ac
.i32_1
,
4552 if (ctx
->abi
->robust_buffer_access
) {
4553 LLVMValueRef dword1
, is_null_descriptor
;
4555 /* Extract the second dword of the descriptor, if it's
4556 * all zero, then it's a null descriptor.
4558 dword1
= LLVMBuildExtractElement(ctx
->ac
.builder
, res
,
4559 LLVMConstInt(ctx
->ac
.i32
, 1, false), "");
4560 is_null_descriptor
=
4561 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, dword1
,
4562 LLVMConstInt(ctx
->ac
.i32
, 0, false), "");
4564 LLVMBuildSelect(ctx
->ac
.builder
, is_null_descriptor
,
4565 ctx
->ac
.i32_0
, ctx
->ac
.i32_1
, "");
4567 default_sample
= ctx
->ac
.i32_1
;
4570 samples
= LLVMBuildSelect(ctx
->ac
.builder
, is_msaa
, samples
,
4571 default_sample
, "");
4576 if (args
.offset
&& instr
->op
!= nir_texop_txf
&& instr
->op
!= nir_texop_txf_ms
) {
4577 LLVMValueRef offset
[3], pack
;
4578 for (unsigned chan
= 0; chan
< 3; ++chan
)
4579 offset
[chan
] = ctx
->ac
.i32_0
;
4581 unsigned num_components
= ac_get_llvm_num_components(args
.offset
);
4582 for (unsigned chan
= 0; chan
< num_components
; chan
++) {
4583 offset
[chan
] = ac_llvm_extract_elem(&ctx
->ac
, args
.offset
, chan
);
4584 offset
[chan
] = LLVMBuildAnd(ctx
->ac
.builder
, offset
[chan
],
4585 LLVMConstInt(ctx
->ac
.i32
, 0x3f, false), "");
4587 offset
[chan
] = LLVMBuildShl(ctx
->ac
.builder
, offset
[chan
],
4588 LLVMConstInt(ctx
->ac
.i32
, chan
* 8, false), "");
4590 pack
= LLVMBuildOr(ctx
->ac
.builder
, offset
[0], offset
[1], "");
4591 pack
= LLVMBuildOr(ctx
->ac
.builder
, pack
, offset
[2], "");
4595 /* Section 8.23.1 (Depth Texture Comparison Mode) of the
4596 * OpenGL 4.5 spec says:
4598 * "If the texture’s internal format indicates a fixed-point
4599 * depth texture, then D_t and D_ref are clamped to the
4600 * range [0, 1]; otherwise no clamping is performed."
4602 * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4603 * so the depth comparison value isn't clamped for Z16 and
4604 * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
4605 * an explicitly clamped 32-bit float format.
4608 ctx
->ac
.chip_class
>= GFX8
&&
4609 ctx
->ac
.chip_class
<= GFX9
&&
4610 ctx
->abi
->clamp_shadow_reference
) {
4611 LLVMValueRef upgraded
, clamped
;
4613 upgraded
= LLVMBuildExtractElement(ctx
->ac
.builder
, args
.sampler
,
4614 LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4615 upgraded
= LLVMBuildLShr(ctx
->ac
.builder
, upgraded
,
4616 LLVMConstInt(ctx
->ac
.i32
, 29, false), "");
4617 upgraded
= LLVMBuildTrunc(ctx
->ac
.builder
, upgraded
, ctx
->ac
.i1
, "");
4618 clamped
= ac_build_clamp(&ctx
->ac
, args
.compare
);
4619 args
.compare
= LLVMBuildSelect(ctx
->ac
.builder
, upgraded
, clamped
,
4623 /* pack derivatives */
4625 int num_src_deriv_channels
, num_dest_deriv_channels
;
4626 switch (instr
->sampler_dim
) {
4627 case GLSL_SAMPLER_DIM_3D
:
4628 case GLSL_SAMPLER_DIM_CUBE
:
4629 num_src_deriv_channels
= 3;
4630 num_dest_deriv_channels
= 3;
4632 case GLSL_SAMPLER_DIM_2D
:
4634 num_src_deriv_channels
= 2;
4635 num_dest_deriv_channels
= 2;
4637 case GLSL_SAMPLER_DIM_1D
:
4638 num_src_deriv_channels
= 1;
4639 if (ctx
->ac
.chip_class
== GFX9
) {
4640 num_dest_deriv_channels
= 2;
4642 num_dest_deriv_channels
= 1;
4647 for (unsigned i
= 0; i
< num_src_deriv_channels
; i
++) {
4648 args
.derivs
[i
] = ac_to_float(&ctx
->ac
,
4649 ac_llvm_extract_elem(&ctx
->ac
, ddx
, i
));
4650 args
.derivs
[num_dest_deriv_channels
+ i
] = ac_to_float(&ctx
->ac
,
4651 ac_llvm_extract_elem(&ctx
->ac
, ddy
, i
));
4653 for (unsigned i
= num_src_deriv_channels
; i
< num_dest_deriv_channels
; i
++) {
4654 args
.derivs
[i
] = ctx
->ac
.f32_0
;
4655 args
.derivs
[num_dest_deriv_channels
+ i
] = ctx
->ac
.f32_0
;
4659 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&& args
.coords
[0]) {
4660 for (unsigned chan
= 0; chan
< instr
->coord_components
; chan
++)
4661 args
.coords
[chan
] = ac_to_float(&ctx
->ac
, args
.coords
[chan
]);
4662 if (instr
->coord_components
== 3)
4663 args
.coords
[3] = LLVMGetUndef(ctx
->ac
.f32
);
4664 ac_prepare_cube_coords(&ctx
->ac
,
4665 instr
->op
== nir_texop_txd
, instr
->is_array
,
4666 instr
->op
== nir_texop_lod
, args
.coords
, args
.derivs
);
4669 /* Texture coordinates fixups */
4670 if (instr
->coord_components
> 1 &&
4671 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
4673 instr
->op
!= nir_texop_txf
) {
4674 args
.coords
[1] = apply_round_slice(&ctx
->ac
, args
.coords
[1]);
4677 if (instr
->coord_components
> 2 &&
4678 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
||
4679 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
||
4680 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS
||
4681 instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
) &&
4683 instr
->op
!= nir_texop_txf
&&
4684 instr
->op
!= nir_texop_txf_ms
&&
4685 instr
->op
!= nir_texop_fragment_fetch
&&
4686 instr
->op
!= nir_texop_fragment_mask_fetch
) {
4687 args
.coords
[2] = apply_round_slice(&ctx
->ac
, args
.coords
[2]);
4690 if (ctx
->ac
.chip_class
== GFX9
&&
4691 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
4692 instr
->op
!= nir_texop_lod
) {
4693 LLVMValueRef filler
;
4694 if (instr
->op
== nir_texop_txf
)
4695 filler
= ctx
->ac
.i32_0
;
4697 filler
= LLVMConstReal(ctx
->ac
.f32
, 0.5);
4699 if (instr
->is_array
)
4700 args
.coords
[2] = args
.coords
[1];
4701 args
.coords
[1] = filler
;
4704 /* Pack sample index */
4705 if (sample_index
&& (instr
->op
== nir_texop_txf_ms
||
4706 instr
->op
== nir_texop_fragment_fetch
))
4707 args
.coords
[instr
->coord_components
] = sample_index
;
4709 if (instr
->op
== nir_texop_samples_identical
) {
4710 struct ac_image_args txf_args
= { 0 };
4711 memcpy(txf_args
.coords
, args
.coords
, sizeof(txf_args
.coords
));
4713 txf_args
.dmask
= 0xf;
4714 txf_args
.resource
= fmask_ptr
;
4715 txf_args
.dim
= instr
->is_array
? ac_image_2darray
: ac_image_2d
;
4716 result
= build_tex_intrinsic(ctx
, instr
, &txf_args
);
4718 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4719 result
= emit_int_cmp(&ctx
->ac
, LLVMIntEQ
, result
, ctx
->ac
.i32_0
);
4723 if ((instr
->sampler_dim
== GLSL_SAMPLER_DIM_SUBPASS_MS
||
4724 instr
->sampler_dim
== GLSL_SAMPLER_DIM_MS
) &&
4725 instr
->op
!= nir_texop_txs
&&
4726 instr
->op
!= nir_texop_fragment_fetch
&&
4727 instr
->op
!= nir_texop_fragment_mask_fetch
) {
4728 unsigned sample_chan
= instr
->is_array
? 3 : 2;
4729 args
.coords
[sample_chan
] = adjust_sample_index_using_fmask(
4730 &ctx
->ac
, args
.coords
[0], args
.coords
[1],
4731 instr
->is_array
? args
.coords
[2] : NULL
,
4732 args
.coords
[sample_chan
], fmask_ptr
);
4735 if (args
.offset
&& (instr
->op
== nir_texop_txf
|| instr
->op
== nir_texop_txf_ms
)) {
4736 int num_offsets
= instr
->src
[offset_src
].src
.ssa
->num_components
;
4737 num_offsets
= MIN2(num_offsets
, instr
->coord_components
);
4738 for (unsigned i
= 0; i
< num_offsets
; ++i
) {
4739 args
.coords
[i
] = LLVMBuildAdd(
4740 ctx
->ac
.builder
, args
.coords
[i
],
4741 LLVMConstInt(ctx
->ac
.i32
, nir_src_comp_as_uint(instr
->src
[offset_src
].src
, i
), false), "");
4746 /* DMASK was repurposed for GATHER4. 4 components are always
4747 * returned and DMASK works like a swizzle - it selects
4748 * the component to fetch. The only valid DMASK values are
4749 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4750 * (red,red,red,red) etc.) The ISA document doesn't mention
4754 if (instr
->op
== nir_texop_tg4
) {
4755 if (instr
->is_shadow
)
4758 args
.dmask
= 1 << instr
->component
;
4761 if (instr
->sampler_dim
!= GLSL_SAMPLER_DIM_BUF
) {
4762 args
.dim
= ac_get_sampler_dim(ctx
->ac
.chip_class
, instr
->sampler_dim
, instr
->is_array
);
4763 args
.unorm
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
;
4766 /* Adjust the number of coordinates because we only need (x,y) for 2D
4767 * multisampled images and (x,y,layer) for 2D multisampled layered
4768 * images or for multisampled input attachments.
4770 if (instr
->op
== nir_texop_fragment_mask_fetch
) {
4771 if (args
.dim
== ac_image_2dmsaa
) {
4772 args
.dim
= ac_image_2d
;
4774 assert(args
.dim
== ac_image_2darraymsaa
);
4775 args
.dim
= ac_image_2darray
;
4779 assert(instr
->dest
.is_ssa
);
4780 args
.d16
= instr
->dest
.ssa
.bit_size
== 16;
4782 result
= build_tex_intrinsic(ctx
, instr
, &args
);
4784 if (instr
->op
== nir_texop_query_levels
)
4785 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
4786 else if (instr
->is_shadow
&& instr
->is_new_style_shadow
&&
4787 instr
->op
!= nir_texop_txs
&& instr
->op
!= nir_texop_lod
&&
4788 instr
->op
!= nir_texop_tg4
)
4789 result
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, ctx
->ac
.i32_0
, "");
4790 else if (instr
->op
== nir_texop_txs
&&
4791 instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
4793 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
4794 LLVMValueRef six
= LLVMConstInt(ctx
->ac
.i32
, 6, false);
4795 LLVMValueRef z
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
4796 z
= LLVMBuildSDiv(ctx
->ac
.builder
, z
, six
, "");
4797 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, z
, two
, "");
4798 } else if (ctx
->ac
.chip_class
== GFX9
&&
4799 instr
->op
== nir_texop_txs
&&
4800 instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
&&
4802 LLVMValueRef two
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
4803 LLVMValueRef layers
= LLVMBuildExtractElement(ctx
->ac
.builder
, result
, two
, "");
4804 result
= LLVMBuildInsertElement(ctx
->ac
.builder
, result
, layers
,
4806 } else if (instr
->dest
.ssa
.num_components
!= 4)
4807 result
= ac_trim_vector(&ctx
->ac
, result
, instr
->dest
.ssa
.num_components
);
4811 assert(instr
->dest
.is_ssa
);
4812 result
= ac_to_integer(&ctx
->ac
, result
);
4814 for (int i
= ARRAY_SIZE(wctx
); --i
>= 0;) {
4815 result
= exit_waterfall(ctx
, wctx
+ i
, result
);
4818 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
4822 static void visit_phi(struct ac_nir_context
*ctx
, nir_phi_instr
*instr
)
4824 LLVMTypeRef type
= get_def_type(ctx
, &instr
->dest
.ssa
);
4825 LLVMValueRef result
= LLVMBuildPhi(ctx
->ac
.builder
, type
, "");
4827 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
4828 _mesa_hash_table_insert(ctx
->phis
, instr
, result
);
4831 static void visit_post_phi(struct ac_nir_context
*ctx
,
4832 nir_phi_instr
*instr
,
4833 LLVMValueRef llvm_phi
)
4835 nir_foreach_phi_src(src
, instr
) {
4836 LLVMBasicBlockRef block
= get_block(ctx
, src
->pred
);
4837 LLVMValueRef llvm_src
= get_src(ctx
, src
->src
);
4839 LLVMAddIncoming(llvm_phi
, &llvm_src
, &block
, 1);
4843 static void phi_post_pass(struct ac_nir_context
*ctx
)
4845 hash_table_foreach(ctx
->phis
, entry
) {
4846 visit_post_phi(ctx
, (nir_phi_instr
*)entry
->key
,
4847 (LLVMValueRef
)entry
->data
);
4852 static bool is_def_used_in_an_export(const nir_ssa_def
* def
) {
4853 nir_foreach_use(use_src
, def
) {
4854 if (use_src
->parent_instr
->type
== nir_instr_type_intrinsic
) {
4855 nir_intrinsic_instr
*instr
= nir_instr_as_intrinsic(use_src
->parent_instr
);
4856 if (instr
->intrinsic
== nir_intrinsic_store_deref
)
4858 } else if (use_src
->parent_instr
->type
== nir_instr_type_alu
) {
4859 nir_alu_instr
*instr
= nir_instr_as_alu(use_src
->parent_instr
);
4860 if (instr
->op
== nir_op_vec4
&&
4861 is_def_used_in_an_export(&instr
->dest
.dest
.ssa
)) {
4869 static void visit_ssa_undef(struct ac_nir_context
*ctx
,
4870 const nir_ssa_undef_instr
*instr
)
4872 unsigned num_components
= instr
->def
.num_components
;
4873 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->ac
.context
, instr
->def
.bit_size
);
4875 if (!ctx
->abi
->convert_undef_to_zero
|| is_def_used_in_an_export(&instr
->def
)) {
4878 if (num_components
== 1)
4879 undef
= LLVMGetUndef(type
);
4881 undef
= LLVMGetUndef(LLVMVectorType(type
, num_components
));
4883 ctx
->ssa_defs
[instr
->def
.index
] = undef
;
4885 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
4886 if (num_components
> 1) {
4887 zero
= ac_build_gather_values_extended(
4888 &ctx
->ac
, &zero
, 4, 0, false, false);
4890 ctx
->ssa_defs
[instr
->def
.index
] = zero
;
4894 static void visit_jump(struct ac_llvm_context
*ctx
,
4895 const nir_jump_instr
*instr
)
4897 switch (instr
->type
) {
4898 case nir_jump_break
:
4899 ac_build_break(ctx
);
4901 case nir_jump_continue
:
4902 ac_build_continue(ctx
);
4905 fprintf(stderr
, "Unknown NIR jump instr: ");
4906 nir_print_instr(&instr
->instr
, stderr
);
4907 fprintf(stderr
, "\n");
4913 glsl_base_to_llvm_type(struct ac_llvm_context
*ac
,
4914 enum glsl_base_type type
)
4918 case GLSL_TYPE_UINT
:
4919 case GLSL_TYPE_BOOL
:
4920 case GLSL_TYPE_SUBROUTINE
:
4922 case GLSL_TYPE_INT8
:
4923 case GLSL_TYPE_UINT8
:
4925 case GLSL_TYPE_INT16
:
4926 case GLSL_TYPE_UINT16
:
4928 case GLSL_TYPE_FLOAT
:
4930 case GLSL_TYPE_FLOAT16
:
4932 case GLSL_TYPE_INT64
:
4933 case GLSL_TYPE_UINT64
:
4935 case GLSL_TYPE_DOUBLE
:
4938 unreachable("unknown GLSL type");
4943 glsl_to_llvm_type(struct ac_llvm_context
*ac
,
4944 const struct glsl_type
*type
)
4946 if (glsl_type_is_scalar(type
)) {
4947 return glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
));
4950 if (glsl_type_is_vector(type
)) {
4951 return LLVMVectorType(
4952 glsl_base_to_llvm_type(ac
, glsl_get_base_type(type
)),
4953 glsl_get_vector_elements(type
));
4956 if (glsl_type_is_matrix(type
)) {
4957 return LLVMArrayType(
4958 glsl_to_llvm_type(ac
, glsl_get_column_type(type
)),
4959 glsl_get_matrix_columns(type
));
4962 if (glsl_type_is_array(type
)) {
4963 return LLVMArrayType(
4964 glsl_to_llvm_type(ac
, glsl_get_array_element(type
)),
4965 glsl_get_length(type
));
4968 assert(glsl_type_is_struct_or_ifc(type
));
4970 LLVMTypeRef member_types
[glsl_get_length(type
)];
4972 for (unsigned i
= 0; i
< glsl_get_length(type
); i
++) {
4974 glsl_to_llvm_type(ac
,
4975 glsl_get_struct_field(type
, i
));
4978 return LLVMStructTypeInContext(ac
->context
, member_types
,
4979 glsl_get_length(type
), false);
4982 static void visit_deref(struct ac_nir_context
*ctx
,
4983 nir_deref_instr
*instr
)
4985 if (instr
->mode
!= nir_var_mem_shared
&&
4986 instr
->mode
!= nir_var_mem_global
)
4989 LLVMValueRef result
= NULL
;
4990 switch(instr
->deref_type
) {
4991 case nir_deref_type_var
: {
4992 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->vars
, instr
->var
);
4993 result
= entry
->data
;
4996 case nir_deref_type_struct
:
4997 if (instr
->mode
== nir_var_mem_global
) {
4998 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
4999 uint64_t offset
= glsl_get_struct_field_offset(parent
->type
,
5000 instr
->strct
.index
);
5001 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
5002 LLVMConstInt(ctx
->ac
.i32
, offset
, 0));
5004 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
5005 LLVMConstInt(ctx
->ac
.i32
, instr
->strct
.index
, 0));
5008 case nir_deref_type_array
:
5009 if (instr
->mode
== nir_var_mem_global
) {
5010 nir_deref_instr
*parent
= nir_deref_instr_parent(instr
);
5011 unsigned stride
= glsl_get_explicit_stride(parent
->type
);
5013 if ((glsl_type_is_matrix(parent
->type
) &&
5014 glsl_matrix_type_is_row_major(parent
->type
)) ||
5015 (glsl_type_is_vector(parent
->type
) && stride
== 0))
5016 stride
= type_scalar_size_bytes(parent
->type
);
5019 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
5020 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
5021 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
5023 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
5025 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
5027 result
= ac_build_gep0(&ctx
->ac
, get_src(ctx
, instr
->parent
),
5028 get_src(ctx
, instr
->arr
.index
));
5031 case nir_deref_type_ptr_as_array
:
5032 if (instr
->mode
== nir_var_mem_global
) {
5033 unsigned stride
= nir_deref_instr_ptr_as_array_stride(instr
);
5035 LLVMValueRef index
= get_src(ctx
, instr
->arr
.index
);
5036 if (LLVMTypeOf(index
) != ctx
->ac
.i64
)
5037 index
= LLVMBuildZExt(ctx
->ac
.builder
, index
, ctx
->ac
.i64
, "");
5039 LLVMValueRef offset
= LLVMBuildMul(ctx
->ac
.builder
, index
, LLVMConstInt(ctx
->ac
.i64
, stride
, 0), "");
5041 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
), offset
);
5043 result
= ac_build_gep_ptr(&ctx
->ac
, get_src(ctx
, instr
->parent
),
5044 get_src(ctx
, instr
->arr
.index
));
5047 case nir_deref_type_cast
: {
5048 result
= get_src(ctx
, instr
->parent
);
5050 /* We can't use the structs from LLVM because the shader
5051 * specifies its own offsets. */
5052 LLVMTypeRef pointee_type
= ctx
->ac
.i8
;
5053 if (instr
->mode
== nir_var_mem_shared
)
5054 pointee_type
= glsl_to_llvm_type(&ctx
->ac
, instr
->type
);
5056 unsigned address_space
;
5058 switch(instr
->mode
) {
5059 case nir_var_mem_shared
:
5060 address_space
= AC_ADDR_SPACE_LDS
;
5062 case nir_var_mem_global
:
5063 address_space
= AC_ADDR_SPACE_GLOBAL
;
5066 unreachable("Unhandled address space");
5069 LLVMTypeRef type
= LLVMPointerType(pointee_type
, address_space
);
5071 if (LLVMTypeOf(result
) != type
) {
5072 if (LLVMGetTypeKind(LLVMTypeOf(result
)) == LLVMVectorTypeKind
) {
5073 result
= LLVMBuildBitCast(ctx
->ac
.builder
, result
,
5076 result
= LLVMBuildIntToPtr(ctx
->ac
.builder
, result
,
5083 unreachable("Unhandled deref_instr deref type");
5086 ctx
->ssa_defs
[instr
->dest
.ssa
.index
] = result
;
5089 static void visit_cf_list(struct ac_nir_context
*ctx
,
5090 struct exec_list
*list
);
5092 static void visit_block(struct ac_nir_context
*ctx
, nir_block
*block
)
5094 nir_foreach_instr(instr
, block
)
5096 switch (instr
->type
) {
5097 case nir_instr_type_alu
:
5098 visit_alu(ctx
, nir_instr_as_alu(instr
));
5100 case nir_instr_type_load_const
:
5101 visit_load_const(ctx
, nir_instr_as_load_const(instr
));
5103 case nir_instr_type_intrinsic
:
5104 visit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
5106 case nir_instr_type_tex
:
5107 visit_tex(ctx
, nir_instr_as_tex(instr
));
5109 case nir_instr_type_phi
:
5110 visit_phi(ctx
, nir_instr_as_phi(instr
));
5112 case nir_instr_type_ssa_undef
:
5113 visit_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
5115 case nir_instr_type_jump
:
5116 visit_jump(&ctx
->ac
, nir_instr_as_jump(instr
));
5118 case nir_instr_type_deref
:
5119 visit_deref(ctx
, nir_instr_as_deref(instr
));
5122 fprintf(stderr
, "Unknown NIR instr type: ");
5123 nir_print_instr(instr
, stderr
);
5124 fprintf(stderr
, "\n");
5129 _mesa_hash_table_insert(ctx
->defs
, block
,
5130 LLVMGetInsertBlock(ctx
->ac
.builder
));
5133 static void visit_if(struct ac_nir_context
*ctx
, nir_if
*if_stmt
)
5135 LLVMValueRef value
= get_src(ctx
, if_stmt
->condition
);
5137 nir_block
*then_block
=
5138 (nir_block
*) exec_list_get_head(&if_stmt
->then_list
);
5140 ac_build_uif(&ctx
->ac
, value
, then_block
->index
);
5142 visit_cf_list(ctx
, &if_stmt
->then_list
);
5144 if (!exec_list_is_empty(&if_stmt
->else_list
)) {
5145 nir_block
*else_block
=
5146 (nir_block
*) exec_list_get_head(&if_stmt
->else_list
);
5148 ac_build_else(&ctx
->ac
, else_block
->index
);
5149 visit_cf_list(ctx
, &if_stmt
->else_list
);
5152 ac_build_endif(&ctx
->ac
, then_block
->index
);
5155 static void visit_loop(struct ac_nir_context
*ctx
, nir_loop
*loop
)
5157 nir_block
*first_loop_block
=
5158 (nir_block
*) exec_list_get_head(&loop
->body
);
5160 ac_build_bgnloop(&ctx
->ac
, first_loop_block
->index
);
5162 visit_cf_list(ctx
, &loop
->body
);
5164 ac_build_endloop(&ctx
->ac
, first_loop_block
->index
);
5167 static void visit_cf_list(struct ac_nir_context
*ctx
,
5168 struct exec_list
*list
)
5170 foreach_list_typed(nir_cf_node
, node
, node
, list
)
5172 switch (node
->type
) {
5173 case nir_cf_node_block
:
5174 visit_block(ctx
, nir_cf_node_as_block(node
));
5177 case nir_cf_node_if
:
5178 visit_if(ctx
, nir_cf_node_as_if(node
));
5181 case nir_cf_node_loop
:
5182 visit_loop(ctx
, nir_cf_node_as_loop(node
));
5192 ac_handle_shader_output_decl(struct ac_llvm_context
*ctx
,
5193 struct ac_shader_abi
*abi
,
5194 struct nir_shader
*nir
,
5195 struct nir_variable
*variable
,
5196 gl_shader_stage stage
)
5198 unsigned output_loc
= variable
->data
.driver_location
/ 4;
5199 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5201 /* tess ctrl has it's own load/store paths for outputs */
5202 if (stage
== MESA_SHADER_TESS_CTRL
)
5205 if (stage
== MESA_SHADER_VERTEX
||
5206 stage
== MESA_SHADER_TESS_EVAL
||
5207 stage
== MESA_SHADER_GEOMETRY
) {
5208 int idx
= variable
->data
.location
+ variable
->data
.index
;
5209 if (idx
== VARYING_SLOT_CLIP_DIST0
) {
5210 int length
= nir
->info
.clip_distance_array_size
+
5211 nir
->info
.cull_distance_array_size
;
5220 bool is_16bit
= glsl_type_is_16bit(glsl_without_array(variable
->type
));
5221 LLVMTypeRef type
= is_16bit
? ctx
->f16
: ctx
->f32
;
5222 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
5223 for (unsigned chan
= 0; chan
< 4; chan
++) {
5224 abi
->outputs
[ac_llvm_reg_index_soa(output_loc
+ i
, chan
)] =
5225 ac_build_alloca_undef(ctx
, type
, "");
5231 setup_locals(struct ac_nir_context
*ctx
,
5232 struct nir_function
*func
)
5235 ctx
->num_locals
= 0;
5236 nir_foreach_function_temp_variable(variable
, func
->impl
) {
5237 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
5238 variable
->data
.driver_location
= ctx
->num_locals
* 4;
5239 variable
->data
.location_frac
= 0;
5240 ctx
->num_locals
+= attrib_count
;
5242 ctx
->locals
= malloc(4 * ctx
->num_locals
* sizeof(LLVMValueRef
));
5246 for (i
= 0; i
< ctx
->num_locals
; i
++) {
5247 for (j
= 0; j
< 4; j
++) {
5248 ctx
->locals
[i
* 4 + j
] =
5249 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "temp");
5255 setup_scratch(struct ac_nir_context
*ctx
,
5256 struct nir_shader
*shader
)
5258 if (shader
->scratch_size
== 0)
5261 ctx
->scratch
= ac_build_alloca_undef(&ctx
->ac
,
5262 LLVMArrayType(ctx
->ac
.i8
, shader
->scratch_size
),
5267 setup_constant_data(struct ac_nir_context
*ctx
,
5268 struct nir_shader
*shader
)
5270 if (!shader
->constant_data
)
5274 LLVMConstStringInContext(ctx
->ac
.context
,
5275 shader
->constant_data
,
5276 shader
->constant_data_size
,
5278 LLVMTypeRef type
= LLVMArrayType(ctx
->ac
.i8
, shader
->constant_data_size
);
5280 /* We want to put the constant data in the CONST address space so that
5281 * we can use scalar loads. However, LLVM versions before 10 put these
5282 * variables in the same section as the code, which is unacceptable
5283 * for RadeonSI as it needs to relocate all the data sections after
5284 * the code sections. See https://reviews.llvm.org/D65813.
5286 unsigned address_space
=
5287 LLVM_VERSION_MAJOR
< 10 ? AC_ADDR_SPACE_GLOBAL
: AC_ADDR_SPACE_CONST
;
5289 LLVMValueRef global
=
5290 LLVMAddGlobalInAddressSpace(ctx
->ac
.module
, type
,
5294 LLVMSetInitializer(global
, data
);
5295 LLVMSetGlobalConstant(global
, true);
5296 LLVMSetVisibility(global
, LLVMHiddenVisibility
);
5297 ctx
->constant_data
= global
;
5301 setup_shared(struct ac_nir_context
*ctx
,
5302 struct nir_shader
*nir
)
5307 LLVMTypeRef type
= LLVMArrayType(ctx
->ac
.i8
,
5308 nir
->info
.cs
.shared_size
);
5311 LLVMAddGlobalInAddressSpace(ctx
->ac
.module
, type
,
5314 LLVMSetAlignment(lds
, 64 * 1024);
5316 ctx
->ac
.lds
= LLVMBuildBitCast(ctx
->ac
.builder
, lds
,
5317 LLVMPointerType(ctx
->ac
.i8
,
5318 AC_ADDR_SPACE_LDS
), "");
5321 void ac_nir_translate(struct ac_llvm_context
*ac
, struct ac_shader_abi
*abi
,
5322 const struct ac_shader_args
*args
, struct nir_shader
*nir
)
5324 struct ac_nir_context ctx
= {};
5325 struct nir_function
*func
;
5331 ctx
.stage
= nir
->info
.stage
;
5332 ctx
.info
= &nir
->info
;
5334 ctx
.main_function
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
5336 nir_foreach_shader_out_variable(variable
, nir
)
5337 ac_handle_shader_output_decl(&ctx
.ac
, ctx
.abi
, nir
, variable
,
5340 ctx
.defs
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
5341 _mesa_key_pointer_equal
);
5342 ctx
.phis
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
5343 _mesa_key_pointer_equal
);
5344 ctx
.vars
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
5345 _mesa_key_pointer_equal
);
5347 if (ctx
.abi
->kill_ps_if_inf_interp
)
5348 ctx
.verified_interp
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
5349 _mesa_key_pointer_equal
);
5351 func
= (struct nir_function
*)exec_list_get_head(&nir
->functions
);
5353 nir_index_ssa_defs(func
->impl
);
5354 ctx
.ssa_defs
= calloc(func
->impl
->ssa_alloc
, sizeof(LLVMValueRef
));
5356 setup_locals(&ctx
, func
);
5357 setup_scratch(&ctx
, nir
);
5358 setup_constant_data(&ctx
, nir
);
5360 if (gl_shader_stage_is_compute(nir
->info
.stage
))
5361 setup_shared(&ctx
, nir
);
5363 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&& nir
->info
.fs
.uses_demote
) {
5364 ctx
.ac
.postponed_kill
= ac_build_alloca_undef(&ctx
.ac
, ac
->i1
, "");
5365 /* true = don't kill. */
5366 LLVMBuildStore(ctx
.ac
.builder
, ctx
.ac
.i1true
, ctx
.ac
.postponed_kill
);
5369 visit_cf_list(&ctx
, &func
->impl
->body
);
5370 phi_post_pass(&ctx
);
5372 if (ctx
.ac
.postponed_kill
)
5373 ac_build_kill_if_false(&ctx
.ac
, LLVMBuildLoad(ctx
.ac
.builder
,
5374 ctx
.ac
.postponed_kill
, ""));
5376 if (!gl_shader_stage_is_compute(nir
->info
.stage
))
5377 ctx
.abi
->emit_outputs(ctx
.abi
, AC_LLVM_MAX_OUTPUTS
,
5382 ralloc_free(ctx
.defs
);
5383 ralloc_free(ctx
.phis
);
5384 ralloc_free(ctx
.vars
);
5385 if (ctx
.abi
->kill_ps_if_inf_interp
)
5386 ralloc_free(ctx
.verified_interp
);
5390 ac_lower_indirect_derefs(struct nir_shader
*nir
, enum chip_class chip_class
)
5392 bool progress
= false;
5394 /* Lower large variables to scratch first so that we won't bloat the
5395 * shader by generating large if ladders for them. We later lower
5396 * scratch to alloca's, assuming LLVM won't generate VGPR indexing.
5398 NIR_PASS(progress
, nir
, nir_lower_vars_to_scratch
,
5399 nir_var_function_temp
,
5401 glsl_get_natural_size_align_bytes
);
5403 /* While it would be nice not to have this flag, we are constrained
5404 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
5406 bool llvm_has_working_vgpr_indexing
= chip_class
!= GFX9
;
5408 /* TODO: Indirect indexing of GS inputs is unimplemented.
5410 * TCS and TES load inputs directly from LDS or offchip memory, so
5411 * indirect indexing is trivial.
5413 nir_variable_mode indirect_mask
= 0;
5414 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
5415 (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
&&
5416 nir
->info
.stage
!= MESA_SHADER_TESS_EVAL
&&
5417 !llvm_has_working_vgpr_indexing
)) {
5418 indirect_mask
|= nir_var_shader_in
;
5420 if (!llvm_has_working_vgpr_indexing
&&
5421 nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
5422 indirect_mask
|= nir_var_shader_out
;
5424 /* TODO: We shouldn't need to do this, however LLVM isn't currently
5425 * smart enough to handle indirects without causing excess spilling
5426 * causing the gpu to hang.
5428 * See the following thread for more details of the problem:
5429 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
5431 indirect_mask
|= nir_var_function_temp
;
5433 progress
|= nir_lower_indirect_derefs(nir
, indirect_mask
);
5438 get_inst_tessfactor_writemask(nir_intrinsic_instr
*intrin
)
5440 if (intrin
->intrinsic
!= nir_intrinsic_store_deref
)
5444 nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[0]));
5446 if (var
->data
.mode
!= nir_var_shader_out
)
5449 unsigned writemask
= 0;
5450 const int location
= var
->data
.location
;
5451 unsigned first_component
= var
->data
.location_frac
;
5452 unsigned num_comps
= intrin
->dest
.ssa
.num_components
;
5454 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
)
5455 writemask
= ((1 << (num_comps
+ 1)) - 1) << first_component
;
5456 else if (location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
5457 writemask
= (((1 << (num_comps
+ 1)) - 1) << first_component
) << 4;
5463 scan_tess_ctrl(nir_cf_node
*cf_node
, unsigned *upper_block_tf_writemask
,
5464 unsigned *cond_block_tf_writemask
,
5465 bool *tessfactors_are_def_in_all_invocs
, bool is_nested_cf
)
5467 switch (cf_node
->type
) {
5468 case nir_cf_node_block
: {
5469 nir_block
*block
= nir_cf_node_as_block(cf_node
);
5470 nir_foreach_instr(instr
, block
) {
5471 if (instr
->type
!= nir_instr_type_intrinsic
)
5474 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
5475 if (intrin
->intrinsic
== nir_intrinsic_control_barrier
) {
5477 /* If we find a barrier in nested control flow put this in the
5478 * too hard basket. In GLSL this is not possible but it is in
5482 *tessfactors_are_def_in_all_invocs
= false;
5486 /* The following case must be prevented:
5487 * gl_TessLevelInner = ...;
5489 * if (gl_InvocationID == 1)
5490 * gl_TessLevelInner = ...;
5492 * If you consider disjoint code segments separated by barriers, each
5493 * such segment that writes tess factor channels should write the same
5494 * channels in all codepaths within that segment.
5496 if (upper_block_tf_writemask
|| cond_block_tf_writemask
) {
5497 /* Accumulate the result: */
5498 *tessfactors_are_def_in_all_invocs
&=
5499 !(*cond_block_tf_writemask
& ~(*upper_block_tf_writemask
));
5501 /* Analyze the next code segment from scratch. */
5502 *upper_block_tf_writemask
= 0;
5503 *cond_block_tf_writemask
= 0;
5506 *upper_block_tf_writemask
|= get_inst_tessfactor_writemask(intrin
);
5511 case nir_cf_node_if
: {
5512 unsigned then_tessfactor_writemask
= 0;
5513 unsigned else_tessfactor_writemask
= 0;
5515 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
5516 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->then_list
) {
5517 scan_tess_ctrl(nested_node
, &then_tessfactor_writemask
,
5518 cond_block_tf_writemask
,
5519 tessfactors_are_def_in_all_invocs
, true);
5522 foreach_list_typed(nir_cf_node
, nested_node
, node
, &if_stmt
->else_list
) {
5523 scan_tess_ctrl(nested_node
, &else_tessfactor_writemask
,
5524 cond_block_tf_writemask
,
5525 tessfactors_are_def_in_all_invocs
, true);
5528 if (then_tessfactor_writemask
|| else_tessfactor_writemask
) {
5529 /* If both statements write the same tess factor channels,
5530 * we can say that the upper block writes them too.
5532 *upper_block_tf_writemask
|= then_tessfactor_writemask
&
5533 else_tessfactor_writemask
;
5534 *cond_block_tf_writemask
|= then_tessfactor_writemask
|
5535 else_tessfactor_writemask
;
5540 case nir_cf_node_loop
: {
5541 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
5542 foreach_list_typed(nir_cf_node
, nested_node
, node
, &loop
->body
) {
5543 scan_tess_ctrl(nested_node
, cond_block_tf_writemask
,
5544 cond_block_tf_writemask
,
5545 tessfactors_are_def_in_all_invocs
, true);
5551 unreachable("unknown cf node type");
5556 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader
*nir
)
5558 assert(nir
->info
.stage
== MESA_SHADER_TESS_CTRL
);
5560 /* The pass works as follows:
5561 * If all codepaths write tess factors, we can say that all
5562 * invocations define tess factors.
5564 * Each tess factor channel is tracked separately.
5566 unsigned main_block_tf_writemask
= 0; /* if main block writes tess factors */
5567 unsigned cond_block_tf_writemask
= 0; /* if cond block writes tess factors */
5569 /* Initial value = true. Here the pass will accumulate results from
5570 * multiple segments surrounded by barriers. If tess factors aren't
5571 * written at all, it's a shader bug and we don't care if this will be
5574 bool tessfactors_are_def_in_all_invocs
= true;
5576 nir_foreach_function(function
, nir
) {
5577 if (function
->impl
) {
5578 foreach_list_typed(nir_cf_node
, node
, node
, &function
->impl
->body
) {
5579 scan_tess_ctrl(node
, &main_block_tf_writemask
,
5580 &cond_block_tf_writemask
,
5581 &tessfactors_are_def_in_all_invocs
,
5587 /* Accumulate the result for the last code segment separated by a
5590 if (main_block_tf_writemask
|| cond_block_tf_writemask
) {
5591 tessfactors_are_def_in_all_invocs
&=
5592 !(cond_block_tf_writemask
& ~main_block_tf_writemask
);
5595 return tessfactors_are_def_in_all_invocs
;