2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
30 #include "c11/threads.h"
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "util/u_math.h"
43 #include "shader_enums.h"
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block
;
52 LLVMBasicBlockRef loop_entry_block
;
55 /* Initialize module-independent parts of the context.
57 * The caller is responsible for initializing ctx::module and ctx::builder.
60 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
61 enum chip_class chip_class
, enum radeon_family family
)
65 ctx
->context
= LLVMContextCreate();
67 ctx
->chip_class
= chip_class
;
72 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
73 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
74 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
75 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
76 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
77 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
78 ctx
->intptr
= ctx
->i32
;
79 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
80 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
81 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
82 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
83 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
84 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
85 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
86 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
87 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
88 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
90 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
91 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
92 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
93 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
94 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
95 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
96 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
97 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
98 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
99 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
100 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
101 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
102 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
103 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
105 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
106 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
108 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
111 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
112 "invariant.load", 14);
114 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
116 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
117 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
119 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
120 "amdgpu.uniform", 14);
122 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
126 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
130 ctx
->flow_depth_max
= 0;
134 ac_get_llvm_num_components(LLVMValueRef value
)
136 LLVMTypeRef type
= LLVMTypeOf(value
);
137 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
138 ? LLVMGetVectorSize(type
)
140 return num_components
;
144 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
148 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
153 return LLVMBuildExtractElement(ac
->builder
, value
,
154 LLVMConstInt(ac
->i32
, index
, false), "");
158 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
160 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
161 type
= LLVMGetElementType(type
);
163 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
164 return LLVMGetIntTypeWidth(type
);
166 if (type
== ctx
->f16
)
168 if (type
== ctx
->f32
)
170 if (type
== ctx
->f64
)
173 unreachable("Unhandled type kind in get_elem_bits");
177 ac_get_type_size(LLVMTypeRef type
)
179 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
182 case LLVMIntegerTypeKind
:
183 return LLVMGetIntTypeWidth(type
) / 8;
184 case LLVMHalfTypeKind
:
186 case LLVMFloatTypeKind
:
188 case LLVMDoubleTypeKind
:
190 case LLVMPointerTypeKind
:
191 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
194 case LLVMVectorTypeKind
:
195 return LLVMGetVectorSize(type
) *
196 ac_get_type_size(LLVMGetElementType(type
));
197 case LLVMArrayTypeKind
:
198 return LLVMGetArrayLength(type
) *
199 ac_get_type_size(LLVMGetElementType(type
));
206 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
210 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
212 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
214 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
217 unreachable("Unhandled integer size");
221 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
223 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
224 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
225 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
226 LLVMGetVectorSize(t
));
228 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
229 switch (LLVMGetPointerAddressSpace(t
)) {
230 case AC_ADDR_SPACE_GLOBAL
:
232 case AC_ADDR_SPACE_LDS
:
235 unreachable("unhandled address space");
238 return to_integer_type_scalar(ctx
, t
);
242 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
244 LLVMTypeRef type
= LLVMTypeOf(v
);
245 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
246 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
248 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
252 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
254 LLVMTypeRef type
= LLVMTypeOf(v
);
255 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
257 return ac_to_integer(ctx
, v
);
260 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
264 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
266 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
268 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
271 unreachable("Unhandled float size");
275 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
277 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
278 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
279 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
280 LLVMGetVectorSize(t
));
282 return to_float_type_scalar(ctx
, t
);
286 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
288 LLVMTypeRef type
= LLVMTypeOf(v
);
289 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
294 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
295 LLVMTypeRef return_type
, LLVMValueRef
*params
,
296 unsigned param_count
, unsigned attrib_mask
)
298 LLVMValueRef function
, call
;
299 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
301 function
= LLVMGetNamedFunction(ctx
->module
, name
);
303 LLVMTypeRef param_types
[32], function_type
;
306 assert(param_count
<= 32);
308 for (i
= 0; i
< param_count
; ++i
) {
310 param_types
[i
] = LLVMTypeOf(params
[i
]);
313 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
314 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
316 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
317 LLVMSetLinkage(function
, LLVMExternalLinkage
);
319 if (!set_callsite_attrs
)
320 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
323 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
324 if (set_callsite_attrs
)
325 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
330 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
333 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
335 LLVMTypeRef elem_type
= type
;
337 assert(bufsize
>= 8);
339 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
340 int ret
= snprintf(buf
, bufsize
, "v%u",
341 LLVMGetVectorSize(type
));
343 char *type_name
= LLVMPrintTypeToString(type
);
344 fprintf(stderr
, "Error building type name for: %s\n",
348 elem_type
= LLVMGetElementType(type
);
352 switch (LLVMGetTypeKind(elem_type
)) {
354 case LLVMIntegerTypeKind
:
355 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
357 case LLVMHalfTypeKind
:
358 snprintf(buf
, bufsize
, "f16");
360 case LLVMFloatTypeKind
:
361 snprintf(buf
, bufsize
, "f32");
363 case LLVMDoubleTypeKind
:
364 snprintf(buf
, bufsize
, "f64");
370 * Helper function that builds an LLVM IR PHI node and immediately adds
374 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
375 unsigned count_incoming
, LLVMValueRef
*values
,
376 LLVMBasicBlockRef
*blocks
)
378 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
379 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
383 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
385 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
386 0, AC_FUNC_ATTR_CONVERGENT
);
389 /* Prevent optimizations (at least of memory accesses) across the current
390 * point in the program by emitting empty inline assembly that is marked as
391 * having side effects.
393 * Optionally, a value can be passed through the inline assembly to prevent
394 * LLVM from hoisting calls to ReadNone functions.
397 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
400 static int counter
= 0;
402 LLVMBuilderRef builder
= ctx
->builder
;
405 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
408 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
409 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
410 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
412 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
413 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
414 LLVMValueRef vgpr
= *pvgpr
;
415 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
416 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
419 assert(vgpr_size
% 4 == 0);
421 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
422 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
423 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
424 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
425 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
432 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
434 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, "llvm.readcyclecounter",
435 ctx
->i64
, NULL
, 0, 0);
436 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
440 ac_build_ballot(struct ac_llvm_context
*ctx
,
443 LLVMValueRef args
[3] = {
446 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
449 /* We currently have no other way to prevent LLVM from lifting the icmp
450 * calls to a dominating basic block.
452 ac_build_optimization_barrier(ctx
, &args
[0]);
454 args
[0] = ac_to_integer(ctx
, args
[0]);
456 return ac_build_intrinsic(ctx
,
457 "llvm.amdgcn.icmp.i32",
459 AC_FUNC_ATTR_NOUNWIND
|
460 AC_FUNC_ATTR_READNONE
|
461 AC_FUNC_ATTR_CONVERGENT
);
465 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
467 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
468 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
469 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
473 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
475 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
476 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
477 LLVMConstInt(ctx
->i64
, 0, 0), "");
481 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
483 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
484 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
486 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
487 vote_set
, active_set
, "");
488 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
490 LLVMConstInt(ctx
->i64
, 0, 0), "");
491 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
495 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
496 unsigned value_count
, unsigned component
)
498 LLVMValueRef vec
= NULL
;
500 if (value_count
== 1) {
501 return values
[component
];
502 } else if (!value_count
)
503 unreachable("value_count is 0");
505 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
506 LLVMValueRef value
= values
[i
];
509 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
510 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
511 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
517 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
518 LLVMValueRef
*values
,
519 unsigned value_count
,
520 unsigned value_stride
,
524 LLVMBuilderRef builder
= ctx
->builder
;
525 LLVMValueRef vec
= NULL
;
528 if (value_count
== 1 && !always_vector
) {
530 return LLVMBuildLoad(builder
, values
[0], "");
532 } else if (!value_count
)
533 unreachable("value_count is 0");
535 for (i
= 0; i
< value_count
; i
++) {
536 LLVMValueRef value
= values
[i
* value_stride
];
538 value
= LLVMBuildLoad(builder
, value
, "");
541 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
542 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
543 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
549 ac_build_gather_values(struct ac_llvm_context
*ctx
,
550 LLVMValueRef
*values
,
551 unsigned value_count
)
553 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
556 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
557 * channels with undef. Extract at most src_channels components from the input.
560 ac_build_expand(struct ac_llvm_context
*ctx
,
562 unsigned src_channels
,
563 unsigned dst_channels
)
565 LLVMTypeRef elemtype
;
566 LLVMValueRef chan
[dst_channels
];
568 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
569 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
571 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
574 src_channels
= MIN2(src_channels
, vec_size
);
576 for (unsigned i
= 0; i
< src_channels
; i
++)
577 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
579 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
582 assert(src_channels
== 1);
585 elemtype
= LLVMTypeOf(value
);
588 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
589 chan
[i
] = LLVMGetUndef(elemtype
);
591 return ac_build_gather_values(ctx
, chan
, dst_channels
);
594 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
595 * with undef. Extract at most num_channels components from the input.
597 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
599 unsigned num_channels
)
601 return ac_build_expand(ctx
, value
, num_channels
, 4);
604 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
606 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
610 name
= "llvm.rint.f16";
611 else if (type_size
== 4)
612 name
= "llvm.rint.f32";
614 name
= "llvm.rint.f64";
616 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
617 AC_FUNC_ATTR_READNONE
);
621 ac_build_fdiv(struct ac_llvm_context
*ctx
,
625 /* If we do (num / den), LLVM >= 7.0 does:
626 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
628 * If we do (num * (1 / den)), LLVM does:
629 * return num * v_rcp_f32(den);
631 LLVMValueRef one
= LLVMConstReal(LLVMTypeOf(num
), 1.0);
632 LLVMValueRef rcp
= LLVMBuildFDiv(ctx
->builder
, one
, den
, "");
633 LLVMValueRef ret
= LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
635 /* Use v_rcp_f32 instead of precise division. */
636 if (!LLVMIsConstant(ret
))
637 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
641 /* See fast_idiv_by_const.h. */
642 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
643 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
645 LLVMValueRef multiplier
,
646 LLVMValueRef pre_shift
,
647 LLVMValueRef post_shift
,
648 LLVMValueRef increment
)
650 LLVMBuilderRef builder
= ctx
->builder
;
652 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
653 num
= LLVMBuildMul(builder
,
654 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
655 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
656 num
= LLVMBuildAdd(builder
, num
,
657 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
658 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
659 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
660 return LLVMBuildLShr(builder
, num
, post_shift
, "");
663 /* See fast_idiv_by_const.h. */
664 /* If num != UINT_MAX, this more efficient version can be used. */
665 /* Set: increment = util_fast_udiv_info::increment; */
666 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
668 LLVMValueRef multiplier
,
669 LLVMValueRef pre_shift
,
670 LLVMValueRef post_shift
,
671 LLVMValueRef increment
)
673 LLVMBuilderRef builder
= ctx
->builder
;
675 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
676 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
677 num
= LLVMBuildMul(builder
,
678 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
679 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
680 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
681 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
682 return LLVMBuildLShr(builder
, num
, post_shift
, "");
685 /* See fast_idiv_by_const.h. */
686 /* Both operands must fit in 31 bits and the divisor must not be 1. */
687 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
689 LLVMValueRef multiplier
,
690 LLVMValueRef post_shift
)
692 LLVMBuilderRef builder
= ctx
->builder
;
694 num
= LLVMBuildMul(builder
,
695 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
696 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
697 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
698 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
699 return LLVMBuildLShr(builder
, num
, post_shift
, "");
702 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
703 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
704 * already multiplied by two. id is the cube face number.
706 struct cube_selection_coords
{
713 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
715 struct cube_selection_coords
*out
)
717 LLVMTypeRef f32
= ctx
->f32
;
719 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
720 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
721 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
722 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
723 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
724 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
725 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
726 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
730 * Build a manual selection sequence for cube face sc/tc coordinates and
731 * major axis vector (multiplied by 2 for consistency) for the given
732 * vec3 \p coords, for the face implied by \p selcoords.
734 * For the major axis, we always adjust the sign to be in the direction of
735 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
736 * the selcoords major axis.
738 static void build_cube_select(struct ac_llvm_context
*ctx
,
739 const struct cube_selection_coords
*selcoords
,
740 const LLVMValueRef
*coords
,
741 LLVMValueRef
*out_st
,
742 LLVMValueRef
*out_ma
)
744 LLVMBuilderRef builder
= ctx
->builder
;
745 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
746 LLVMValueRef is_ma_positive
;
748 LLVMValueRef is_ma_z
, is_not_ma_z
;
749 LLVMValueRef is_ma_y
;
750 LLVMValueRef is_ma_x
;
754 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
755 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
756 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
757 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
759 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
760 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
761 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
762 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
763 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
766 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
767 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
768 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
769 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
770 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
773 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
774 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
775 LLVMConstReal(f32
, -1.0), "");
776 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
779 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
780 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
781 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
782 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
783 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
787 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
788 bool is_deriv
, bool is_array
, bool is_lod
,
789 LLVMValueRef
*coords_arg
,
790 LLVMValueRef
*derivs_arg
)
793 LLVMBuilderRef builder
= ctx
->builder
;
794 struct cube_selection_coords selcoords
;
795 LLVMValueRef coords
[3];
798 if (is_array
&& !is_lod
) {
799 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
801 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
803 * "For Array forms, the array layer used will be
805 * max(0, min(d−1, floor(layer+0.5)))
807 * where d is the depth of the texture array and layer
808 * comes from the component indicated in the tables below.
809 * Workaroudn for an issue where the layer is taken from a
810 * helper invocation which happens to fall on a different
811 * layer due to extrapolation."
813 * VI and earlier attempt to implement this in hardware by
814 * clamping the value of coords[2] = (8 * layer) + face.
815 * Unfortunately, this means that the we end up with the wrong
816 * face when clamping occurs.
818 * Clamp the layer earlier to work around the issue.
820 if (ctx
->chip_class
<= VI
) {
822 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
823 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
829 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
831 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
832 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
833 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
835 for (int i
= 0; i
< 2; ++i
)
836 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
838 coords
[2] = selcoords
.id
;
840 if (is_deriv
&& derivs_arg
) {
841 LLVMValueRef derivs
[4];
844 /* Convert cube derivatives to 2D derivatives. */
845 for (axis
= 0; axis
< 2; axis
++) {
846 LLVMValueRef deriv_st
[2];
847 LLVMValueRef deriv_ma
;
849 /* Transform the derivative alongside the texture
850 * coordinate. Mathematically, the correct formula is
851 * as follows. Assume we're projecting onto the +Z face
852 * and denote by dx/dh the derivative of the (original)
853 * X texture coordinate with respect to horizontal
854 * window coordinates. The projection onto the +Z face
859 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
860 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
862 * This motivatives the implementation below.
864 * Whether this actually gives the expected results for
865 * apps that might feed in derivatives obtained via
866 * finite differences is anyone's guess. The OpenGL spec
867 * seems awfully quiet about how textureGrad for cube
868 * maps should be handled.
870 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
871 deriv_st
, &deriv_ma
);
873 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
875 for (int i
= 0; i
< 2; ++i
)
876 derivs
[axis
* 2 + i
] =
877 LLVMBuildFSub(builder
,
878 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
879 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
882 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
885 /* Shift the texture coordinate. This must be applied after the
886 * derivative calculation.
888 for (int i
= 0; i
< 2; ++i
)
889 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
892 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
893 /* coords_arg.w component - array_index for cube arrays */
894 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
897 memcpy(coords_arg
, coords
, sizeof(coords
));
902 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
903 LLVMValueRef llvm_chan
,
904 LLVMValueRef attr_number
,
909 LLVMValueRef args
[5];
914 args
[2] = attr_number
;
917 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
918 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
923 args
[3] = attr_number
;
926 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
927 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
931 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
932 LLVMValueRef llvm_chan
,
933 LLVMValueRef attr_number
,
938 LLVMValueRef args
[6];
943 args
[2] = attr_number
;
944 args
[3] = ctx
->i1false
;
947 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
948 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
953 args
[3] = attr_number
;
954 args
[4] = ctx
->i1false
;
957 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
958 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
962 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
963 LLVMValueRef parameter
,
964 LLVMValueRef llvm_chan
,
965 LLVMValueRef attr_number
,
968 LLVMValueRef args
[4];
972 args
[2] = attr_number
;
975 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
976 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
980 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
981 LLVMValueRef base_ptr
,
984 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
988 ac_build_gep0(struct ac_llvm_context
*ctx
,
989 LLVMValueRef base_ptr
,
992 LLVMValueRef indices
[2] = {
996 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
999 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1002 return LLVMBuildPointerCast(ctx
->builder
,
1003 ac_build_gep0(ctx
, ptr
, index
),
1004 LLVMTypeOf(ptr
), "");
1008 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1009 LLVMValueRef base_ptr
, LLVMValueRef index
,
1012 LLVMBuildStore(ctx
->builder
, value
,
1013 ac_build_gep0(ctx
, base_ptr
, index
));
1017 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1018 * It's equivalent to doing a load from &base_ptr[index].
1020 * \param base_ptr Where the array starts.
1021 * \param index The element index into the array.
1022 * \param uniform Whether the base_ptr and index can be assumed to be
1023 * dynamically uniform (i.e. load to an SGPR)
1024 * \param invariant Whether the load is invariant (no other opcodes affect it)
1025 * \param no_unsigned_wraparound
1026 * For all possible re-associations and re-distributions of an expression
1027 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1028 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1029 * does not result in an unsigned integer wraparound. This is used for
1030 * optimal code generation of 32-bit pointer arithmetic.
1032 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1033 * integer wraparound can't be an imm offset in s_load_dword, because
1034 * the instruction performs "addr + offset" in 64 bits.
1036 * Expected usage for bindless textures by chaining GEPs:
1037 * // possible unsigned wraparound, don't use InBounds:
1038 * ptr1 = LLVMBuildGEP(base_ptr, index);
1039 * image = load(ptr1); // becomes "s_load ptr1, 0"
1041 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1042 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1045 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1046 LLVMValueRef index
, bool uniform
, bool invariant
,
1047 bool no_unsigned_wraparound
)
1049 LLVMValueRef pointer
, result
;
1050 LLVMValueRef indices
[2] = {ctx
->i32_0
, index
};
1052 if (no_unsigned_wraparound
&&
1053 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1054 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1056 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1059 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1060 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1062 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1066 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1069 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1072 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1073 LLVMValueRef base_ptr
, LLVMValueRef index
)
1075 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1078 /* This assumes that there is no unsigned integer wraparound during the address
1079 * computation, excluding all GEPs within base_ptr. */
1080 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1081 LLVMValueRef base_ptr
, LLVMValueRef index
)
1083 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1086 /* See ac_build_load_custom() documentation. */
1087 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1088 LLVMValueRef base_ptr
, LLVMValueRef index
)
1090 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1094 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1097 LLVMValueRef vindex
,
1098 LLVMValueRef voffset
,
1099 unsigned num_channels
,
1102 bool writeonly_memory
,
1105 LLVMValueRef args
[] = {
1107 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1108 vindex
? vindex
: ctx
->i32_0
,
1110 LLVMConstInt(ctx
->i1
, glc
, 0),
1111 LLVMConstInt(ctx
->i1
, slc
, 0)
1113 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1115 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1119 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.format.%s",
1122 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
1126 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, ARRAY_SIZE(args
),
1127 ac_get_store_intr_attribs(writeonly_memory
));
1131 ac_build_llvm8_buffer_store_common(struct ac_llvm_context
*ctx
,
1134 LLVMValueRef vindex
,
1135 LLVMValueRef voffset
,
1136 LLVMValueRef soffset
,
1137 unsigned num_channels
,
1140 bool writeonly_memory
,
1144 LLVMValueRef args
[6];
1147 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1149 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1150 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1151 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1152 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1153 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1155 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1156 const char *indexing_kind
= structurized
? "struct" : "raw";
1160 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1161 indexing_kind
, type_names
[func
]);
1163 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1164 indexing_kind
, type_names
[func
]);
1167 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1168 ac_get_store_intr_attribs(writeonly_memory
));
1172 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1175 LLVMValueRef vindex
,
1176 LLVMValueRef voffset
,
1177 unsigned num_channels
,
1179 bool writeonly_memory
)
1181 if (HAVE_LLVM
>= 0x800) {
1182 ac_build_llvm8_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1183 voffset
, NULL
, num_channels
,
1184 glc
, false, writeonly_memory
,
1187 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
, voffset
,
1188 num_channels
, glc
, false,
1189 writeonly_memory
, true);
1193 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1194 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1195 * or v4i32 (num_channels=3,4).
1198 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1201 unsigned num_channels
,
1202 LLVMValueRef voffset
,
1203 LLVMValueRef soffset
,
1204 unsigned inst_offset
,
1207 bool writeonly_memory
,
1208 bool swizzle_enable_hint
)
1210 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
1212 if (num_channels
== 3) {
1213 LLVMValueRef v
[3], v01
;
1215 for (int i
= 0; i
< 3; i
++) {
1216 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1217 LLVMConstInt(ctx
->i32
, i
, 0), "");
1219 v01
= ac_build_gather_values(ctx
, v
, 2);
1221 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1222 soffset
, inst_offset
, glc
, slc
,
1223 writeonly_memory
, swizzle_enable_hint
);
1224 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1225 soffset
, inst_offset
+ 8,
1227 writeonly_memory
, swizzle_enable_hint
);
1231 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1232 * (voffset is swizzled, but soffset isn't swizzled).
1233 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1235 if (!swizzle_enable_hint
) {
1236 LLVMValueRef offset
= soffset
;
1239 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1240 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1242 if (HAVE_LLVM
>= 0x800) {
1243 ac_build_llvm8_buffer_store_common(ctx
, rsrc
,
1244 ac_to_float(ctx
, vdata
),
1253 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1255 ac_build_buffer_store_common(ctx
, rsrc
,
1256 ac_to_float(ctx
, vdata
),
1258 num_channels
, glc
, slc
,
1259 writeonly_memory
, false);
1264 static const unsigned dfmts
[] = {
1265 V_008F0C_BUF_DATA_FORMAT_32
,
1266 V_008F0C_BUF_DATA_FORMAT_32_32
,
1267 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1268 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1270 unsigned dfmt
= dfmts
[num_channels
- 1];
1271 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1272 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1274 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1275 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1276 slc
, writeonly_memory
);
1280 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1282 LLVMValueRef vindex
,
1283 LLVMValueRef voffset
,
1284 unsigned num_channels
,
1290 LLVMValueRef args
[] = {
1291 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1292 vindex
? vindex
: ctx
->i32_0
,
1294 LLVMConstInt(ctx
->i1
, glc
, 0),
1295 LLVMConstInt(ctx
->i1
, slc
, 0)
1297 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1299 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
1300 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1304 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.format.%s",
1307 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
1311 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1313 ac_get_load_intr_attribs(can_speculate
));
1317 ac_build_llvm8_buffer_load_common(struct ac_llvm_context
*ctx
,
1319 LLVMValueRef vindex
,
1320 LLVMValueRef voffset
,
1321 LLVMValueRef soffset
,
1322 unsigned num_channels
,
1329 LLVMValueRef args
[5];
1331 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1333 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1334 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1335 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1336 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1337 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1339 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
1340 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1341 const char *indexing_kind
= structurized
? "struct" : "raw";
1345 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1346 indexing_kind
, type_names
[func
]);
1348 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1349 indexing_kind
, type_names
[func
]);
1352 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1354 ac_get_load_intr_attribs(can_speculate
));
1358 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1361 LLVMValueRef vindex
,
1362 LLVMValueRef voffset
,
1363 LLVMValueRef soffset
,
1364 unsigned inst_offset
,
1370 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1372 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1374 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1376 if (allow_smem
&& !slc
&&
1377 (!glc
|| (HAVE_LLVM
>= 0x0800 && ctx
->chip_class
>= VI
))) {
1378 assert(vindex
== NULL
);
1380 LLVMValueRef result
[8];
1382 for (int i
= 0; i
< num_channels
; i
++) {
1384 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1385 LLVMConstInt(ctx
->i32
, 4, 0), "");
1387 const char *intrname
=
1388 HAVE_LLVM
>= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1389 : "llvm.SI.load.const.v4i32";
1390 unsigned num_args
= HAVE_LLVM
>= 0x0800 ? 3 : 2;
1391 LLVMValueRef args
[3] = {
1394 glc
? ctx
->i32_1
: ctx
->i32_0
,
1396 result
[i
] = ac_build_intrinsic(ctx
, intrname
,
1397 ctx
->f32
, args
, num_args
,
1398 AC_FUNC_ATTR_READNONE
|
1399 (HAVE_LLVM
< 0x0800 ? AC_FUNC_ATTR_LEGACY
: 0));
1401 if (num_channels
== 1)
1404 if (num_channels
== 3)
1405 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1406 return ac_build_gather_values(ctx
, result
, num_channels
);
1409 if (HAVE_LLVM
>= 0x0800) {
1410 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
,
1412 num_channels
, glc
, slc
,
1413 can_speculate
, false,
1417 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, offset
,
1418 num_channels
, glc
, slc
,
1419 can_speculate
, false);
1422 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1424 LLVMValueRef vindex
,
1425 LLVMValueRef voffset
,
1426 unsigned num_channels
,
1430 if (HAVE_LLVM
>= 0x800) {
1431 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1432 num_channels
, glc
, false,
1433 can_speculate
, true, true);
1435 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1436 num_channels
, glc
, false,
1437 can_speculate
, true);
1440 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
1442 LLVMValueRef vindex
,
1443 LLVMValueRef voffset
,
1444 unsigned num_channels
,
1448 if (HAVE_LLVM
>= 0x800) {
1449 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1450 num_channels
, glc
, false,
1451 can_speculate
, true, true);
1454 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, LLVMConstInt(ctx
->i32
, 2, 0), "");
1455 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, ctx
->i32_1
, "");
1456 stride
= LLVMBuildLShr(ctx
->builder
, stride
, LLVMConstInt(ctx
->i32
, 16, 0), "");
1458 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->builder
,
1459 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, elem_count
, stride
, ""),
1460 elem_count
, stride
, "");
1462 LLVMValueRef new_rsrc
= LLVMBuildInsertElement(ctx
->builder
, rsrc
, new_elem_count
,
1463 LLVMConstInt(ctx
->i32
, 2, 0), "");
1465 return ac_build_buffer_load_common(ctx
, new_rsrc
, vindex
, voffset
,
1466 num_channels
, glc
, false,
1467 can_speculate
, true);
1471 ac_build_llvm8_tbuffer_load(struct ac_llvm_context
*ctx
,
1473 LLVMValueRef vindex
,
1474 LLVMValueRef voffset
,
1475 LLVMValueRef soffset
,
1476 unsigned num_channels
,
1484 LLVMValueRef args
[6];
1486 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1488 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1489 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1490 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1491 args
[idx
++] = LLVMConstInt(ctx
->i32
, dfmt
| (nfmt
<< 4), 0);
1492 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1493 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1495 LLVMTypeRef types
[] = {ctx
->i32
, ctx
->v2i32
, ctx
->v4i32
};
1496 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
1497 const char *indexing_kind
= structurized
? "struct" : "raw";
1500 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1501 indexing_kind
, type_names
[func
]);
1503 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1505 ac_get_load_intr_attribs(can_speculate
));
1509 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1511 LLVMValueRef vindex
,
1512 LLVMValueRef voffset
,
1513 LLVMValueRef soffset
,
1514 LLVMValueRef immoffset
,
1515 unsigned num_channels
,
1521 bool structurized
) /* only matters for LLVM 8+ */
1523 if (HAVE_LLVM
>= 0x800) {
1524 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1526 return ac_build_llvm8_tbuffer_load(ctx
, rsrc
, vindex
, voffset
,
1527 soffset
, num_channels
,
1528 dfmt
, nfmt
, glc
, slc
,
1529 can_speculate
, structurized
);
1532 LLVMValueRef args
[] = {
1534 vindex
? vindex
: ctx
->i32_0
,
1538 LLVMConstInt(ctx
->i32
, dfmt
, false),
1539 LLVMConstInt(ctx
->i32
, nfmt
, false),
1540 LLVMConstInt(ctx
->i32
, glc
, false),
1541 LLVMConstInt(ctx
->i32
, slc
, false),
1543 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1544 LLVMTypeRef types
[] = {ctx
->i32
, ctx
->v2i32
, ctx
->v4i32
};
1545 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
1548 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.load.%s",
1551 return ac_build_intrinsic(ctx
, name
, types
[func
], args
, 9,
1552 ac_get_load_intr_attribs(can_speculate
));
1556 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1558 LLVMValueRef vindex
,
1559 LLVMValueRef voffset
,
1560 LLVMValueRef soffset
,
1561 LLVMValueRef immoffset
,
1562 unsigned num_channels
,
1569 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1570 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1571 slc
, can_speculate
, true);
1575 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1577 LLVMValueRef voffset
,
1578 LLVMValueRef soffset
,
1579 LLVMValueRef immoffset
,
1580 unsigned num_channels
,
1587 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1588 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1589 slc
, can_speculate
, false);
1593 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1595 LLVMValueRef voffset
,
1596 LLVMValueRef soffset
,
1597 LLVMValueRef immoffset
,
1600 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1601 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1604 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1605 immoffset
, 1, dfmt
, nfmt
, glc
, false,
1608 return LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1612 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1614 LLVMValueRef voffset
,
1615 LLVMValueRef soffset
,
1616 LLVMValueRef immoffset
,
1619 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1620 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1623 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1624 immoffset
, 1, dfmt
, nfmt
, glc
, false,
1627 return LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1630 ac_build_llvm8_tbuffer_store(struct ac_llvm_context
*ctx
,
1633 LLVMValueRef vindex
,
1634 LLVMValueRef voffset
,
1635 LLVMValueRef soffset
,
1636 unsigned num_channels
,
1641 bool writeonly_memory
,
1644 LLVMValueRef args
[7];
1646 args
[idx
++] = vdata
;
1647 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1649 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1650 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1651 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1652 args
[idx
++] = LLVMConstInt(ctx
->i32
, dfmt
| (nfmt
<< 4), 0);
1653 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1654 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1656 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
1657 const char *indexing_kind
= structurized
? "struct" : "raw";
1660 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1661 indexing_kind
, type_names
[func
]);
1663 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1664 ac_get_store_intr_attribs(writeonly_memory
));
1668 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1671 LLVMValueRef vindex
,
1672 LLVMValueRef voffset
,
1673 LLVMValueRef soffset
,
1674 LLVMValueRef immoffset
,
1675 unsigned num_channels
,
1680 bool writeonly_memory
,
1681 bool structurized
) /* only matters for LLVM 8+ */
1683 if (HAVE_LLVM
>= 0x800) {
1684 voffset
= LLVMBuildAdd(ctx
->builder
,
1685 voffset
? voffset
: ctx
->i32_0
,
1688 ac_build_llvm8_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
,
1689 soffset
, num_channels
, dfmt
, nfmt
,
1690 glc
, slc
, writeonly_memory
,
1693 LLVMValueRef params
[] = {
1696 vindex
? vindex
: ctx
->i32_0
,
1697 voffset
? voffset
: ctx
->i32_0
,
1698 soffset
? soffset
: ctx
->i32_0
,
1700 LLVMConstInt(ctx
->i32
, dfmt
, false),
1701 LLVMConstInt(ctx
->i32
, nfmt
, false),
1702 LLVMConstInt(ctx
->i32
, glc
, false),
1703 LLVMConstInt(ctx
->i32
, slc
, false),
1705 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1706 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
1709 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.store.%s",
1712 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, params
, 10,
1713 ac_get_store_intr_attribs(writeonly_memory
));
1718 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1721 LLVMValueRef vindex
,
1722 LLVMValueRef voffset
,
1723 LLVMValueRef soffset
,
1724 LLVMValueRef immoffset
,
1725 unsigned num_channels
,
1730 bool writeonly_memory
)
1732 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1733 immoffset
, num_channels
, dfmt
, nfmt
, glc
, slc
,
1734 writeonly_memory
, true);
1738 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1741 LLVMValueRef voffset
,
1742 LLVMValueRef soffset
,
1743 LLVMValueRef immoffset
,
1744 unsigned num_channels
,
1749 bool writeonly_memory
)
1751 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1752 immoffset
, num_channels
, dfmt
, nfmt
, glc
, slc
,
1753 writeonly_memory
, false);
1757 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1760 LLVMValueRef voffset
,
1761 LLVMValueRef soffset
,
1763 bool writeonly_memory
)
1765 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1766 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1768 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1769 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1771 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1772 ctx
->i32_0
, 1, dfmt
, nfmt
, glc
, false,
1777 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1780 LLVMValueRef voffset
,
1781 LLVMValueRef soffset
,
1783 bool writeonly_memory
)
1785 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1786 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1788 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1789 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1791 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1792 ctx
->i32_0
, 1, dfmt
, nfmt
, glc
, false,
1796 * Set range metadata on an instruction. This can only be used on load and
1797 * call instructions. If you know an instruction can only produce the values
1798 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1799 * \p lo is the minimum value inclusive.
1800 * \p hi is the maximum value exclusive.
1802 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1803 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1805 LLVMValueRef range_md
, md_args
[2];
1806 LLVMTypeRef type
= LLVMTypeOf(value
);
1807 LLVMContextRef context
= LLVMGetTypeContext(type
);
1809 md_args
[0] = LLVMConstInt(type
, lo
, false);
1810 md_args
[1] = LLVMConstInt(type
, hi
, false);
1811 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1812 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1816 ac_get_thread_id(struct ac_llvm_context
*ctx
)
1820 LLVMValueRef tid_args
[2];
1821 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
1822 tid_args
[1] = ctx
->i32_0
;
1823 tid_args
[1] = ac_build_intrinsic(ctx
,
1824 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1825 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
1827 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
1829 2, AC_FUNC_ATTR_READNONE
);
1830 set_range_metadata(ctx
, tid
, 0, 64);
1835 * SI implements derivatives using the local data store (LDS)
1836 * All writes to the LDS happen in all executing threads at
1837 * the same time. TID is the Thread ID for the current
1838 * thread and is a value between 0 and 63, representing
1839 * the thread's position in the wavefront.
1841 * For the pixel shader threads are grouped into quads of four pixels.
1842 * The TIDs of the pixels of a quad are:
1850 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1851 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1852 * the current pixel's column, and masking with 0xfffffffe yields the TID
1853 * of the left pixel of the current pixel's row.
1855 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1856 * adding 2 yields the TID of the pixel below the top pixel.
1859 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1864 unsigned tl_lanes
[4], trbl_lanes
[4];
1865 LLVMValueRef tl
, trbl
;
1866 LLVMValueRef result
;
1868 for (unsigned i
= 0; i
< 4; ++i
) {
1869 tl_lanes
[i
] = i
& mask
;
1870 trbl_lanes
[i
] = (i
& mask
) + idx
;
1873 tl
= ac_build_quad_swizzle(ctx
, val
,
1874 tl_lanes
[0], tl_lanes
[1],
1875 tl_lanes
[2], tl_lanes
[3]);
1876 trbl
= ac_build_quad_swizzle(ctx
, val
,
1877 trbl_lanes
[0], trbl_lanes
[1],
1878 trbl_lanes
[2], trbl_lanes
[3]);
1880 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1881 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1882 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1884 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.f32", ctx
->f32
,
1891 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1893 LLVMValueRef wave_id
)
1895 LLVMValueRef args
[2];
1896 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1898 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
1902 ac_build_imsb(struct ac_llvm_context
*ctx
,
1904 LLVMTypeRef dst_type
)
1906 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
1908 AC_FUNC_ATTR_READNONE
);
1910 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1911 * the index from LSB. Invert it by doing "31 - msb". */
1912 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1915 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1916 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1917 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1918 arg
, ctx
->i32_0
, ""),
1919 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1920 arg
, all_ones
, ""), "");
1922 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1926 ac_build_umsb(struct ac_llvm_context
*ctx
,
1928 LLVMTypeRef dst_type
)
1930 const char *intrin_name
;
1932 LLVMValueRef highest_bit
;
1936 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
1939 intrin_name
= "llvm.ctlz.i64";
1941 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
1945 intrin_name
= "llvm.ctlz.i32";
1947 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
1951 intrin_name
= "llvm.ctlz.i16";
1953 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
1957 unreachable(!"invalid bitsize");
1961 LLVMValueRef params
[2] = {
1966 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
1968 AC_FUNC_ATTR_READNONE
);
1970 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1971 * the index from LSB. Invert it by doing "31 - msb". */
1972 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
1973 msb
= LLVMBuildTruncOrBitCast(ctx
->builder
, msb
, ctx
->i32
, "");
1975 /* check for zero */
1976 return LLVMBuildSelect(ctx
->builder
,
1977 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
1978 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1981 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1985 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
1986 LLVMValueRef args
[2] = {a
, b
};
1987 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
1988 AC_FUNC_ATTR_READNONE
);
1991 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1995 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
1996 LLVMValueRef args
[2] = {a
, b
};
1997 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
1998 AC_FUNC_ATTR_READNONE
);
2001 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2004 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2005 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2008 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2011 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2012 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2015 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2018 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2019 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2022 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2024 LLVMTypeRef t
= LLVMTypeOf(value
);
2025 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2026 LLVMConstReal(t
, 1.0));
2029 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2031 LLVMValueRef args
[9];
2033 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2034 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2037 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2038 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2040 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2042 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2044 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2045 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2047 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2048 ctx
->voidt
, args
, 6, 0);
2050 args
[2] = a
->out
[0];
2051 args
[3] = a
->out
[1];
2052 args
[4] = a
->out
[2];
2053 args
[5] = a
->out
[3];
2054 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2055 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2057 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2058 ctx
->voidt
, args
, 8, 0);
2062 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2064 struct ac_export_args args
;
2066 args
.enabled_channels
= 0x0; /* enabled channels */
2067 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2068 args
.done
= 1; /* DONE bit */
2069 args
.target
= V_008DFC_SQ_EXP_NULL
;
2070 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2071 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2072 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2073 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2074 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2076 ac_build_export(ctx
, &args
);
2079 static unsigned ac_num_coords(enum ac_image_dim dim
)
2085 case ac_image_1darray
:
2089 case ac_image_2darray
:
2090 case ac_image_2dmsaa
:
2092 case ac_image_2darraymsaa
:
2095 unreachable("ac_num_coords: bad dim");
2099 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2103 case ac_image_1darray
:
2106 case ac_image_2darray
:
2111 case ac_image_2dmsaa
:
2112 case ac_image_2darraymsaa
:
2114 unreachable("derivatives not supported");
2118 static const char *get_atomic_name(enum ac_atomic_op op
)
2121 case ac_atomic_swap
: return "swap";
2122 case ac_atomic_add
: return "add";
2123 case ac_atomic_sub
: return "sub";
2124 case ac_atomic_smin
: return "smin";
2125 case ac_atomic_umin
: return "umin";
2126 case ac_atomic_smax
: return "smax";
2127 case ac_atomic_umax
: return "umax";
2128 case ac_atomic_and
: return "and";
2129 case ac_atomic_or
: return "or";
2130 case ac_atomic_xor
: return "xor";
2132 unreachable("bad atomic op");
2135 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2136 struct ac_image_args
*a
)
2138 const char *overload
[3] = { "", "", "" };
2139 unsigned num_overloads
= 0;
2140 LLVMValueRef args
[18];
2141 unsigned num_args
= 0;
2142 enum ac_image_dim dim
= a
->dim
;
2144 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2146 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2147 a
->opcode
!= ac_image_store_mip
) ||
2149 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2150 (!a
->compare
&& !a
->offset
));
2151 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2152 a
->opcode
== ac_image_get_lod
) ||
2154 assert((a
->bias
? 1 : 0) +
2156 (a
->level_zero
? 1 : 0) +
2157 (a
->derivs
[0] ? 1 : 0) <= 1);
2159 if (a
->opcode
== ac_image_get_lod
) {
2161 case ac_image_1darray
:
2164 case ac_image_2darray
:
2173 bool sample
= a
->opcode
== ac_image_sample
||
2174 a
->opcode
== ac_image_gather4
||
2175 a
->opcode
== ac_image_get_lod
;
2176 bool atomic
= a
->opcode
== ac_image_atomic
||
2177 a
->opcode
== ac_image_atomic_cmpswap
;
2178 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2180 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2181 args
[num_args
++] = a
->data
[0];
2182 if (a
->opcode
== ac_image_atomic_cmpswap
)
2183 args
[num_args
++] = a
->data
[1];
2187 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2190 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2192 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2193 overload
[num_overloads
++] = ".f32";
2196 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2198 unsigned count
= ac_num_derivs(dim
);
2199 for (unsigned i
= 0; i
< count
; ++i
)
2200 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2201 overload
[num_overloads
++] = ".f32";
2203 unsigned num_coords
=
2204 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2205 for (unsigned i
= 0; i
< num_coords
; ++i
)
2206 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2208 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2209 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2211 args
[num_args
++] = a
->resource
;
2213 args
[num_args
++] = a
->sampler
;
2214 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2217 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2218 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->cache_policy
, false);
2221 const char *atomic_subop
= "";
2222 switch (a
->opcode
) {
2223 case ac_image_sample
: name
= "sample"; break;
2224 case ac_image_gather4
: name
= "gather4"; break;
2225 case ac_image_load
: name
= "load"; break;
2226 case ac_image_load_mip
: name
= "load.mip"; break;
2227 case ac_image_store
: name
= "store"; break;
2228 case ac_image_store_mip
: name
= "store.mip"; break;
2229 case ac_image_atomic
:
2231 atomic_subop
= get_atomic_name(a
->atomic
);
2233 case ac_image_atomic_cmpswap
:
2235 atomic_subop
= "cmpswap";
2237 case ac_image_get_lod
: name
= "getlod"; break;
2238 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2239 default: unreachable("invalid image opcode");
2242 const char *dimname
;
2244 case ac_image_1d
: dimname
= "1d"; break;
2245 case ac_image_2d
: dimname
= "2d"; break;
2246 case ac_image_3d
: dimname
= "3d"; break;
2247 case ac_image_cube
: dimname
= "cube"; break;
2248 case ac_image_1darray
: dimname
= "1darray"; break;
2249 case ac_image_2darray
: dimname
= "2darray"; break;
2250 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2251 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2252 default: unreachable("invalid dim");
2256 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2258 snprintf(intr_name
, sizeof(intr_name
),
2259 "llvm.amdgcn.image.%s%s" /* base name */
2260 "%s%s%s" /* sample/gather modifiers */
2261 ".%s.%s%s%s%s", /* dimension and type overloads */
2263 a
->compare
? ".c" : "",
2266 a
->derivs
[0] ? ".d" :
2267 a
->level_zero
? ".lz" : "",
2268 a
->offset
? ".o" : "",
2270 atomic
? "i32" : "v4f32",
2271 overload
[0], overload
[1], overload
[2]);
2276 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2281 LLVMValueRef result
=
2282 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2284 if (!sample
&& retty
== ctx
->v4f32
) {
2285 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2291 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2292 LLVMValueRef args
[2])
2295 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2297 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2298 args
, 2, AC_FUNC_ATTR_READNONE
);
2301 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2302 LLVMValueRef args
[2])
2305 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2306 ctx
->v2i16
, args
, 2,
2307 AC_FUNC_ATTR_READNONE
);
2308 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2311 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2312 LLVMValueRef args
[2])
2315 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2316 ctx
->v2i16
, args
, 2,
2317 AC_FUNC_ATTR_READNONE
);
2318 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2321 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2322 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2323 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2325 assert(bits
== 8 || bits
== 10 || bits
== 16);
2327 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2328 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2329 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2330 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2331 LLVMValueRef max_alpha
=
2332 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2333 LLVMValueRef min_alpha
=
2334 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2338 for (int i
= 0; i
< 2; i
++) {
2339 bool alpha
= hi
&& i
== 1;
2340 args
[i
] = ac_build_imin(ctx
, args
[i
],
2341 alpha
? max_alpha
: max_rgb
);
2342 args
[i
] = ac_build_imax(ctx
, args
[i
],
2343 alpha
? min_alpha
: min_rgb
);
2348 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2349 ctx
->v2i16
, args
, 2,
2350 AC_FUNC_ATTR_READNONE
);
2351 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2354 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2355 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2356 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2358 assert(bits
== 8 || bits
== 10 || bits
== 16);
2360 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2361 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2362 LLVMValueRef max_alpha
=
2363 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2367 for (int i
= 0; i
< 2; i
++) {
2368 bool alpha
= hi
&& i
== 1;
2369 args
[i
] = ac_build_umin(ctx
, args
[i
],
2370 alpha
? max_alpha
: max_rgb
);
2375 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2376 ctx
->v2i16
, args
, 2,
2377 AC_FUNC_ATTR_READNONE
);
2378 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2381 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2383 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2384 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2387 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2389 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2393 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2394 LLVMValueRef offset
, LLVMValueRef width
,
2397 LLVMValueRef args
[] = {
2403 return ac_build_intrinsic(ctx
,
2404 is_signed
? "llvm.amdgcn.sbfe.i32" :
2405 "llvm.amdgcn.ubfe.i32",
2407 AC_FUNC_ATTR_READNONE
);
2410 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2411 LLVMValueRef s1
, LLVMValueRef s2
)
2413 return LLVMBuildAdd(ctx
->builder
,
2414 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2417 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2418 LLVMValueRef s1
, LLVMValueRef s2
)
2420 return LLVMBuildFAdd(ctx
->builder
,
2421 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2424 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
2426 LLVMValueRef args
[1] = {
2427 LLVMConstInt(ctx
->i32
, simm16
, false),
2429 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2430 ctx
->voidt
, args
, 1, 0);
2433 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2439 if (bitsize
== 32) {
2440 intr
= "llvm.amdgcn.fract.f32";
2443 intr
= "llvm.amdgcn.fract.f64";
2447 LLVMValueRef params
[] = {
2450 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2451 AC_FUNC_ATTR_READNONE
);
2454 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2457 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2458 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2459 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2461 LLVMValueRef cmp
, val
;
2462 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2463 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2464 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2465 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2469 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2472 LLVMValueRef cmp
, val
, zero
, one
;
2475 if (bitsize
== 32) {
2485 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2486 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2487 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2488 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2492 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2494 LLVMValueRef result
;
2497 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2501 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2502 (LLVMValueRef
[]) { src0
}, 1,
2503 AC_FUNC_ATTR_READNONE
);
2505 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2508 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2509 (LLVMValueRef
[]) { src0
}, 1,
2510 AC_FUNC_ATTR_READNONE
);
2513 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2514 (LLVMValueRef
[]) { src0
}, 1,
2515 AC_FUNC_ATTR_READNONE
);
2518 unreachable(!"invalid bitsize");
2525 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2528 LLVMValueRef result
;
2531 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2535 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2536 (LLVMValueRef
[]) { src0
}, 1,
2537 AC_FUNC_ATTR_READNONE
);
2540 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2541 (LLVMValueRef
[]) { src0
}, 1,
2542 AC_FUNC_ATTR_READNONE
);
2545 unreachable(!"invalid bitsize");
2552 #define AC_EXP_TARGET 0
2553 #define AC_EXP_ENABLED_CHANNELS 1
2554 #define AC_EXP_OUT0 2
2562 struct ac_vs_exp_chan
2566 enum ac_ir_type type
;
2569 struct ac_vs_exp_inst
{
2572 struct ac_vs_exp_chan chan
[4];
2575 struct ac_vs_exports
{
2577 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2580 /* Return true if the PARAM export has been eliminated. */
2581 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2582 uint32_t num_outputs
,
2583 struct ac_vs_exp_inst
*exp
)
2585 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2586 bool is_zero
[4] = {}, is_one
[4] = {};
2588 for (i
= 0; i
< 4; i
++) {
2589 /* It's a constant expression. Undef outputs are eliminated too. */
2590 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2593 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2594 if (exp
->chan
[i
].const_float
== 0)
2596 else if (exp
->chan
[i
].const_float
== 1)
2599 return false; /* other constant */
2604 /* Only certain combinations of 0 and 1 can be eliminated. */
2605 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2606 default_val
= is_zero
[3] ? 0 : 1;
2607 else if (is_one
[0] && is_one
[1] && is_one
[2])
2608 default_val
= is_zero
[3] ? 2 : 3;
2612 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2613 LLVMInstructionEraseFromParent(exp
->inst
);
2615 /* Change OFFSET to DEFAULT_VAL. */
2616 for (i
= 0; i
< num_outputs
; i
++) {
2617 if (vs_output_param_offset
[i
] == exp
->offset
) {
2618 vs_output_param_offset
[i
] =
2619 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
2626 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
2627 uint8_t *vs_output_param_offset
,
2628 uint32_t num_outputs
,
2629 struct ac_vs_exports
*processed
,
2630 struct ac_vs_exp_inst
*exp
)
2632 unsigned p
, copy_back_channels
= 0;
2634 /* See if the output is already in the list of processed outputs.
2635 * The LLVMValueRef comparison relies on SSA.
2637 for (p
= 0; p
< processed
->num
; p
++) {
2638 bool different
= false;
2640 for (unsigned j
= 0; j
< 4; j
++) {
2641 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
2642 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
2644 /* Treat undef as a match. */
2645 if (c2
->type
== AC_IR_UNDEF
)
2648 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2649 * and consider the instruction duplicated.
2651 if (c1
->type
== AC_IR_UNDEF
) {
2652 copy_back_channels
|= 1 << j
;
2656 /* Test whether the channels are not equal. */
2657 if (c1
->type
!= c2
->type
||
2658 (c1
->type
== AC_IR_CONST
&&
2659 c1
->const_float
!= c2
->const_float
) ||
2660 (c1
->type
== AC_IR_VALUE
&&
2661 c1
->value
!= c2
->value
)) {
2669 copy_back_channels
= 0;
2671 if (p
== processed
->num
)
2674 /* If a match was found, but the matching export has undef where the new
2675 * one has a normal value, copy the normal value to the undef channel.
2677 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
2679 /* Get current enabled channels mask. */
2680 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
2681 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
2683 while (copy_back_channels
) {
2684 unsigned chan
= u_bit_scan(©_back_channels
);
2686 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
2687 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
2688 exp
->chan
[chan
].value
);
2689 match
->chan
[chan
] = exp
->chan
[chan
];
2691 /* Update number of enabled channels because the original mask
2692 * is not always 0xf.
2694 enabled_channels
|= (1 << chan
);
2695 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
2696 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
2699 /* The PARAM export is duplicated. Kill it. */
2700 LLVMInstructionEraseFromParent(exp
->inst
);
2702 /* Change OFFSET to the matching export. */
2703 for (unsigned i
= 0; i
< num_outputs
; i
++) {
2704 if (vs_output_param_offset
[i
] == exp
->offset
) {
2705 vs_output_param_offset
[i
] = match
->offset
;
2712 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
2713 LLVMValueRef main_fn
,
2714 uint8_t *vs_output_param_offset
,
2715 uint32_t num_outputs
,
2716 uint8_t *num_param_exports
)
2718 LLVMBasicBlockRef bb
;
2719 bool removed_any
= false;
2720 struct ac_vs_exports exports
;
2724 /* Process all LLVM instructions. */
2725 bb
= LLVMGetFirstBasicBlock(main_fn
);
2727 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
2730 LLVMValueRef cur
= inst
;
2731 inst
= LLVMGetNextInstruction(inst
);
2732 struct ac_vs_exp_inst exp
;
2734 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
2737 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
2739 if (!ac_llvm_is_function(callee
))
2742 const char *name
= LLVMGetValueName(callee
);
2743 unsigned num_args
= LLVMCountParams(callee
);
2745 /* Check if this is an export instruction. */
2746 if ((num_args
!= 9 && num_args
!= 8) ||
2747 (strcmp(name
, "llvm.SI.export") &&
2748 strcmp(name
, "llvm.amdgcn.exp.f32")))
2751 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
2752 unsigned target
= LLVMConstIntGetZExtValue(arg
);
2754 if (target
< V_008DFC_SQ_EXP_PARAM
)
2757 target
-= V_008DFC_SQ_EXP_PARAM
;
2759 /* Parse the instruction. */
2760 memset(&exp
, 0, sizeof(exp
));
2761 exp
.offset
= target
;
2764 for (unsigned i
= 0; i
< 4; i
++) {
2765 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
2767 exp
.chan
[i
].value
= v
;
2769 if (LLVMIsUndef(v
)) {
2770 exp
.chan
[i
].type
= AC_IR_UNDEF
;
2771 } else if (LLVMIsAConstantFP(v
)) {
2772 LLVMBool loses_info
;
2773 exp
.chan
[i
].type
= AC_IR_CONST
;
2774 exp
.chan
[i
].const_float
=
2775 LLVMConstRealGetDouble(v
, &loses_info
);
2777 exp
.chan
[i
].type
= AC_IR_VALUE
;
2781 /* Eliminate constant and duplicated PARAM exports. */
2782 if (ac_eliminate_const_output(vs_output_param_offset
,
2783 num_outputs
, &exp
) ||
2784 ac_eliminate_duplicated_output(ctx
,
2785 vs_output_param_offset
,
2786 num_outputs
, &exports
,
2790 exports
.exp
[exports
.num
++] = exp
;
2793 bb
= LLVMGetNextBasicBlock(bb
);
2796 /* Remove holes in export memory due to removed PARAM exports.
2797 * This is done by renumbering all PARAM exports.
2800 uint8_t old_offset
[VARYING_SLOT_MAX
];
2803 /* Make a copy of the offsets. We need the old version while
2804 * we are modifying some of them. */
2805 memcpy(old_offset
, vs_output_param_offset
,
2806 sizeof(old_offset
));
2808 for (i
= 0; i
< exports
.num
; i
++) {
2809 unsigned offset
= exports
.exp
[i
].offset
;
2811 /* Update vs_output_param_offset. Multiple outputs can
2812 * have the same offset.
2814 for (out
= 0; out
< num_outputs
; out
++) {
2815 if (old_offset
[out
] == offset
)
2816 vs_output_param_offset
[out
] = i
;
2819 /* Change the PARAM offset in the instruction. */
2820 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
2821 LLVMConstInt(ctx
->i32
,
2822 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
2824 *num_param_exports
= exports
.num
;
2828 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
2830 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
2831 ac_build_intrinsic(ctx
,
2832 "llvm.amdgcn.init.exec", ctx
->voidt
,
2833 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
2836 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
2838 unsigned lds_size
= ctx
->chip_class
>= CIK
? 65536 : 32768;
2839 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
2840 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
2844 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
2845 LLVMValueRef dw_addr
)
2847 return ac_build_load(ctx
, ctx
->lds
, dw_addr
);
2850 void ac_lds_store(struct ac_llvm_context
*ctx
,
2851 LLVMValueRef dw_addr
,
2854 value
= ac_to_integer(ctx
, value
);
2855 ac_build_indexed_store(ctx
, ctx
->lds
,
2859 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
2860 LLVMTypeRef dst_type
,
2863 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2864 const char *intrin_name
;
2868 switch (src0_bitsize
) {
2870 intrin_name
= "llvm.cttz.i64";
2875 intrin_name
= "llvm.cttz.i32";
2880 intrin_name
= "llvm.cttz.i16";
2885 unreachable(!"invalid bitsize");
2888 LLVMValueRef params
[2] = {
2891 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2892 * add special code to check for x=0. The reason is that
2893 * the LLVM behavior for x=0 is different from what we
2894 * need here. However, LLVM also assumes that ffs(x) is
2895 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2896 * a conditional assignment to handle 0 is still required.
2898 * The hardware already implements the correct behavior.
2903 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2905 AC_FUNC_ATTR_READNONE
);
2907 if (src0_bitsize
== 64) {
2908 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
2911 /* TODO: We need an intrinsic to skip this conditional. */
2912 /* Check for zero: */
2913 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
2916 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
2919 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
2921 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2922 AC_ADDR_SPACE_CONST
);
2925 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
2927 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2928 AC_ADDR_SPACE_CONST_32BIT
);
2931 static struct ac_llvm_flow
*
2932 get_current_flow(struct ac_llvm_context
*ctx
)
2934 if (ctx
->flow_depth
> 0)
2935 return &ctx
->flow
[ctx
->flow_depth
- 1];
2939 static struct ac_llvm_flow
*
2940 get_innermost_loop(struct ac_llvm_context
*ctx
)
2942 for (unsigned i
= ctx
->flow_depth
; i
> 0; --i
) {
2943 if (ctx
->flow
[i
- 1].loop_entry_block
)
2944 return &ctx
->flow
[i
- 1];
2949 static struct ac_llvm_flow
*
2950 push_flow(struct ac_llvm_context
*ctx
)
2952 struct ac_llvm_flow
*flow
;
2954 if (ctx
->flow_depth
>= ctx
->flow_depth_max
) {
2955 unsigned new_max
= MAX2(ctx
->flow_depth
<< 1,
2956 AC_LLVM_INITIAL_CF_DEPTH
);
2958 ctx
->flow
= realloc(ctx
->flow
, new_max
* sizeof(*ctx
->flow
));
2959 ctx
->flow_depth_max
= new_max
;
2962 flow
= &ctx
->flow
[ctx
->flow_depth
];
2965 flow
->next_block
= NULL
;
2966 flow
->loop_entry_block
= NULL
;
2970 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
2974 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
2975 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
2978 /* Append a basic block at the level of the parent flow.
2980 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
2983 assert(ctx
->flow_depth
>= 1);
2985 if (ctx
->flow_depth
>= 2) {
2986 struct ac_llvm_flow
*flow
= &ctx
->flow
[ctx
->flow_depth
- 2];
2988 return LLVMInsertBasicBlockInContext(ctx
->context
,
2989 flow
->next_block
, name
);
2992 LLVMValueRef main_fn
=
2993 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
2994 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
2997 /* Emit a branch to the given default target for the current block if
2998 * applicable -- that is, if the current block does not already contain a
2999 * branch from a break or continue.
3001 static void emit_default_branch(LLVMBuilderRef builder
,
3002 LLVMBasicBlockRef target
)
3004 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3005 LLVMBuildBr(builder
, target
);
3008 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3010 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3011 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3012 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3013 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3014 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3015 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3018 void ac_build_break(struct ac_llvm_context
*ctx
)
3020 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3021 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3024 void ac_build_continue(struct ac_llvm_context
*ctx
)
3026 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3027 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3030 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3032 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3033 LLVMBasicBlockRef endif_block
;
3035 assert(!current_branch
->loop_entry_block
);
3037 endif_block
= append_basic_block(ctx
, "ENDIF");
3038 emit_default_branch(ctx
->builder
, endif_block
);
3040 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3041 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3043 current_branch
->next_block
= endif_block
;
3046 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3048 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3050 assert(!current_branch
->loop_entry_block
);
3052 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3053 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3054 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3059 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3061 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3063 assert(current_loop
->loop_entry_block
);
3065 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3067 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3068 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3072 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3074 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3075 LLVMBasicBlockRef if_block
;
3077 if_block
= append_basic_block(ctx
, "IF");
3078 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3079 set_basicblock_name(if_block
, "if", label_id
);
3080 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3081 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3084 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3087 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3088 value
, ctx
->f32_0
, "");
3089 ac_build_ifcc(ctx
, cond
, label_id
);
3092 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3095 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3096 ac_to_integer(ctx
, value
),
3098 ac_build_ifcc(ctx
, cond
, label_id
);
3101 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3104 LLVMBuilderRef builder
= ac
->builder
;
3105 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3106 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3107 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3108 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3109 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3113 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3115 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3118 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3119 LLVMDisposeBuilder(first_builder
);
3123 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3124 LLVMTypeRef type
, const char *name
)
3126 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3127 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3131 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3134 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3135 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3136 LLVMPointerType(type
, addr_space
), "");
3139 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3142 unsigned num_components
= ac_get_llvm_num_components(value
);
3143 if (count
== num_components
)
3146 LLVMValueRef masks
[MAX2(count
, 2)];
3147 masks
[0] = ctx
->i32_0
;
3148 masks
[1] = ctx
->i32_1
;
3149 for (unsigned i
= 2; i
< count
; i
++)
3150 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3153 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3156 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3157 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3160 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3161 unsigned rshift
, unsigned bitwidth
)
3163 LLVMValueRef value
= param
;
3165 value
= LLVMBuildLShr(ctx
->builder
, value
,
3166 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3168 if (rshift
+ bitwidth
< 32) {
3169 unsigned mask
= (1 << bitwidth
) - 1;
3170 value
= LLVMBuildAnd(ctx
->builder
, value
,
3171 LLVMConstInt(ctx
->i32
, mask
, false), "");
3176 /* Adjust the sample index according to FMASK.
3178 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3179 * which is the identity mapping. Each nibble says which physical sample
3180 * should be fetched to get that sample.
3182 * For example, 0x11111100 means there are only 2 samples stored and
3183 * the second sample covers 3/4 of the pixel. When reading samples 0
3184 * and 1, return physical sample 0 (determined by the first two 0s
3185 * in FMASK), otherwise return physical sample 1.
3187 * The sample index should be adjusted as follows:
3188 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3190 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3191 LLVMValueRef
*addr
, bool is_array_tex
)
3193 struct ac_image_args fmask_load
= {};
3194 fmask_load
.opcode
= ac_image_load
;
3195 fmask_load
.resource
= fmask
;
3196 fmask_load
.dmask
= 0xf;
3197 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3199 fmask_load
.coords
[0] = addr
[0];
3200 fmask_load
.coords
[1] = addr
[1];
3202 fmask_load
.coords
[2] = addr
[2];
3204 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3205 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3208 /* Apply the formula. */
3209 unsigned sample_chan
= is_array_tex
? 3 : 2;
3210 LLVMValueRef final_sample
;
3211 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3212 LLVMConstInt(ac
->i32
, 4, 0), "");
3213 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3214 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3215 * with EQAA, so those will map to 0. */
3216 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3217 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3219 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3220 * resource descriptor is 0 (invalid).
3223 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3224 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3225 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3227 /* Replace the MSAA sample index. */
3228 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3229 addr
[sample_chan
], "");
3233 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3235 ac_build_optimization_barrier(ctx
, &src
);
3236 return ac_build_intrinsic(ctx
,
3237 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3238 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3240 lane
== NULL
? 1 : 2,
3241 AC_FUNC_ATTR_READNONE
|
3242 AC_FUNC_ATTR_CONVERGENT
);
3246 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3249 * @param lane - id of the lane or NULL for the first active lane
3250 * @return value of the lane
3253 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3255 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3256 src
= ac_to_integer(ctx
, src
);
3257 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3261 ret
= _ac_build_readlane(ctx
, src
, lane
);
3263 assert(bits
% 32 == 0);
3264 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3265 LLVMValueRef src_vector
=
3266 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3267 ret
= LLVMGetUndef(vec_type
);
3268 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3269 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3270 LLVMConstInt(ctx
->i32
, i
, 0), "");
3271 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
3272 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3273 LLVMConstInt(ctx
->i32
, i
, 0), "");
3276 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3280 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3282 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
3284 LLVMValueRef pred
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, lane
,
3285 ac_get_thread_id(ctx
), "");
3286 return LLVMBuildSelect(ctx
->builder
, pred
, value
, src
, "");
3290 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3292 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3293 LLVMVectorType(ctx
->i32
, 2),
3295 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3297 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3300 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3301 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3302 2, AC_FUNC_ATTR_READNONE
);
3303 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3304 (LLVMValueRef
[]) { mask_hi
, val
},
3305 2, AC_FUNC_ATTR_READNONE
);
3310 _dpp_quad_perm
= 0x000,
3311 _dpp_row_sl
= 0x100,
3312 _dpp_row_sr
= 0x110,
3313 _dpp_row_rr
= 0x120,
3318 dpp_row_mirror
= 0x140,
3319 dpp_row_half_mirror
= 0x141,
3320 dpp_row_bcast15
= 0x142,
3321 dpp_row_bcast31
= 0x143
3324 static inline enum dpp_ctrl
3325 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3327 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3328 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3331 static inline enum dpp_ctrl
3332 dpp_row_sl(unsigned amount
)
3334 assert(amount
> 0 && amount
< 16);
3335 return _dpp_row_sl
| amount
;
3338 static inline enum dpp_ctrl
3339 dpp_row_sr(unsigned amount
)
3341 assert(amount
> 0 && amount
< 16);
3342 return _dpp_row_sr
| amount
;
3346 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3347 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3350 return ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32",
3354 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3355 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3356 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3357 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3358 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3362 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3363 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3366 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3367 src
= ac_to_integer(ctx
, src
);
3368 old
= ac_to_integer(ctx
, old
);
3369 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3372 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3373 bank_mask
, bound_ctrl
);
3375 assert(bits
% 32 == 0);
3376 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3377 LLVMValueRef src_vector
=
3378 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3379 LLVMValueRef old_vector
=
3380 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3381 ret
= LLVMGetUndef(vec_type
);
3382 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3383 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3384 LLVMConstInt(ctx
->i32
, i
,
3386 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3387 LLVMConstInt(ctx
->i32
, i
,
3389 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3394 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3396 LLVMConstInt(ctx
->i32
, i
,
3400 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3403 static inline unsigned
3404 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3406 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3407 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3411 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3413 return ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle",
3414 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3415 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3416 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3420 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3422 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3423 src
= ac_to_integer(ctx
, src
);
3424 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3427 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3429 assert(bits
% 32 == 0);
3430 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3431 LLVMValueRef src_vector
=
3432 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3433 ret
= LLVMGetUndef(vec_type
);
3434 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3435 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3436 LLVMConstInt(ctx
->i32
, i
,
3438 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3440 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3442 LLVMConstInt(ctx
->i32
, i
,
3446 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3450 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3452 char name
[32], type
[8];
3453 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3454 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3455 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3456 (LLVMValueRef
[]) { src
}, 1,
3457 AC_FUNC_ATTR_READNONE
);
3461 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3462 LLVMValueRef inactive
)
3464 char name
[33], type
[8];
3465 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3466 src
= ac_to_integer(ctx
, src
);
3467 inactive
= ac_to_integer(ctx
, inactive
);
3468 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3469 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3471 ac_build_intrinsic(ctx
, name
,
3472 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3474 AC_FUNC_ATTR_READNONE
|
3475 AC_FUNC_ATTR_CONVERGENT
);
3476 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3480 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3482 if (type_size
== 4) {
3484 case nir_op_iadd
: return ctx
->i32_0
;
3485 case nir_op_fadd
: return ctx
->f32_0
;
3486 case nir_op_imul
: return ctx
->i32_1
;
3487 case nir_op_fmul
: return ctx
->f32_1
;
3488 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
3489 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
3490 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
3491 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
3492 case nir_op_umax
: return ctx
->i32_0
;
3493 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
3494 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
3495 case nir_op_ior
: return ctx
->i32_0
;
3496 case nir_op_ixor
: return ctx
->i32_0
;
3498 unreachable("bad reduction intrinsic");
3500 } else { /* type_size == 64bit */
3502 case nir_op_iadd
: return ctx
->i64_0
;
3503 case nir_op_fadd
: return ctx
->f64_0
;
3504 case nir_op_imul
: return ctx
->i64_1
;
3505 case nir_op_fmul
: return ctx
->f64_1
;
3506 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
3507 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
3508 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
3509 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
3510 case nir_op_umax
: return ctx
->i64_0
;
3511 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
3512 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
3513 case nir_op_ior
: return ctx
->i64_0
;
3514 case nir_op_ixor
: return ctx
->i64_0
;
3516 unreachable("bad reduction intrinsic");
3522 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
3524 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
3526 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
3527 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
3528 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
3529 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
3530 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
3531 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
3533 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
3534 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
3536 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
3537 _64bit
? "llvm.minnum.f64" : "llvm.minnum.f32",
3538 _64bit
? ctx
->f64
: ctx
->f32
,
3539 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3540 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
3541 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
3543 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
3544 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
3546 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
3547 _64bit
? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3548 _64bit
? ctx
->f64
: ctx
->f32
,
3549 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3550 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
3551 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
3552 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
3554 unreachable("bad reduction intrinsic");
3559 * \param maxprefix specifies that the result only needs to be correct for a
3560 * prefix of this many threads
3562 * TODO: add inclusive and excluse scan functions for SI chip class.
3565 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
3568 LLVMValueRef result
, tmp
;
3572 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
3573 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3576 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
3577 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3580 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
3581 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3584 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
3585 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3588 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
3589 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3590 if (maxprefix
<= 16)
3592 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3593 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3594 if (maxprefix
<= 32)
3596 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3597 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3602 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3604 LLVMValueRef result
;
3606 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
3607 LLVMBuilderRef builder
= ctx
->builder
;
3608 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
3609 result
= ac_build_ballot(ctx
, src
);
3610 result
= ac_build_mbcnt(ctx
, result
);
3611 result
= LLVMBuildAdd(builder
, result
, src
, "");
3615 ac_build_optimization_barrier(ctx
, &src
);
3617 LLVMValueRef identity
=
3618 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
3619 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
3620 LLVMTypeOf(identity
), "");
3621 result
= ac_build_scan(ctx
, op
, result
, identity
, 64);
3623 return ac_build_wwm(ctx
, result
);
3627 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3629 LLVMValueRef result
;
3631 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
3632 LLVMBuilderRef builder
= ctx
->builder
;
3633 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
3634 result
= ac_build_ballot(ctx
, src
);
3635 result
= ac_build_mbcnt(ctx
, result
);
3639 ac_build_optimization_barrier(ctx
, &src
);
3641 LLVMValueRef identity
=
3642 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
3643 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
3644 LLVMTypeOf(identity
), "");
3645 result
= ac_build_dpp(ctx
, identity
, result
, dpp_wf_sr1
, 0xf, 0xf, false);
3646 result
= ac_build_scan(ctx
, op
, result
, identity
, 64);
3648 return ac_build_wwm(ctx
, result
);
3652 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
3654 if (cluster_size
== 1) return src
;
3655 ac_build_optimization_barrier(ctx
, &src
);
3656 LLVMValueRef result
, swap
;
3657 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
3658 ac_get_type_size(LLVMTypeOf(src
)));
3659 result
= LLVMBuildBitCast(ctx
->builder
,
3660 ac_build_set_inactive(ctx
, src
, identity
),
3661 LLVMTypeOf(identity
), "");
3662 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
3663 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3664 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
3666 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
3667 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3668 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
3670 if (ctx
->chip_class
>= VI
)
3671 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
3673 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
3674 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3675 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
3677 if (ctx
->chip_class
>= VI
)
3678 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
3680 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
3681 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3682 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
3684 if (ctx
->chip_class
>= VI
&& cluster_size
!= 32)
3685 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3687 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
3688 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3689 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
3691 if (ctx
->chip_class
>= VI
) {
3692 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3693 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3694 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
3695 return ac_build_wwm(ctx
, result
);
3697 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
3698 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
3699 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3700 return ac_build_wwm(ctx
, result
);
3705 * "Top half" of a scan that reduces per-wave values across an entire
3708 * The source value must be present in the highest lane of the wave, and the
3709 * highest lane must be live.
3712 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3714 if (ws
->maxwaves
<= 1)
3717 const LLVMValueRef i32_63
= LLVMConstInt(ctx
->i32
, 63, false);
3718 LLVMBuilderRef builder
= ctx
->builder
;
3719 LLVMValueRef tid
= ac_get_thread_id(ctx
);
3722 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, i32_63
, "");
3723 ac_build_ifcc(ctx
, tmp
, 1000);
3724 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
3725 ac_build_endif(ctx
, 1000);
3729 * "Bottom half" of a scan that reduces per-wave values across an entire
3732 * The caller must place a barrier between the top and bottom halves.
3735 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3737 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
3738 const LLVMValueRef identity
=
3739 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
3741 if (ws
->maxwaves
<= 1) {
3742 ws
->result_reduce
= ws
->src
;
3743 ws
->result_inclusive
= ws
->src
;
3744 ws
->result_exclusive
= identity
;
3747 assert(ws
->maxwaves
<= 32);
3749 LLVMBuilderRef builder
= ctx
->builder
;
3750 LLVMValueRef tid
= ac_get_thread_id(ctx
);
3751 LLVMBasicBlockRef bbs
[2];
3752 LLVMValueRef phivalues_scan
[2];
3753 LLVMValueRef tmp
, tmp2
;
3755 bbs
[0] = LLVMGetInsertBlock(builder
);
3756 phivalues_scan
[0] = LLVMGetUndef(type
);
3758 if (ws
->enable_reduce
)
3759 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
3760 else if (ws
->enable_inclusive
)
3761 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
3763 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
3764 ac_build_ifcc(ctx
, tmp
, 1001);
3766 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
3768 ac_build_optimization_barrier(ctx
, &tmp
);
3770 bbs
[1] = LLVMGetInsertBlock(builder
);
3771 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
);
3773 ac_build_endif(ctx
, 1001);
3775 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
3777 if (ws
->enable_reduce
) {
3778 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
3779 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
3781 if (ws
->enable_inclusive
)
3782 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
3783 if (ws
->enable_exclusive
) {
3784 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
3785 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
3786 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
3787 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
3792 * Inclusive scan of a per-wave value across an entire workgroup.
3794 * This implies an s_barrier instruction.
3796 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3797 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3798 * useful manner because of the barrier in the algorithm.)
3801 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3803 ac_build_wg_wavescan_top(ctx
, ws
);
3804 ac_build_s_barrier(ctx
);
3805 ac_build_wg_wavescan_bottom(ctx
, ws
);
3809 * "Top half" of a scan that reduces per-thread values across an entire
3812 * All lanes must be active when this code runs.
3815 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3817 if (ws
->enable_exclusive
) {
3818 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
3819 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
3820 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
3821 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
3823 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
3826 bool enable_inclusive
= ws
->enable_inclusive
;
3827 bool enable_exclusive
= ws
->enable_exclusive
;
3828 ws
->enable_inclusive
= false;
3829 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
3830 ac_build_wg_wavescan_top(ctx
, ws
);
3831 ws
->enable_inclusive
= enable_inclusive
;
3832 ws
->enable_exclusive
= enable_exclusive
;
3836 * "Bottom half" of a scan that reduces per-thread values across an entire
3839 * The caller must place a barrier between the top and bottom halves.
3842 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3844 bool enable_inclusive
= ws
->enable_inclusive
;
3845 bool enable_exclusive
= ws
->enable_exclusive
;
3846 ws
->enable_inclusive
= false;
3847 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
3848 ac_build_wg_wavescan_bottom(ctx
, ws
);
3849 ws
->enable_inclusive
= enable_inclusive
;
3850 ws
->enable_exclusive
= enable_exclusive
;
3852 /* ws->result_reduce is already the correct value */
3853 if (ws
->enable_inclusive
)
3854 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->src
, ws
->op
);
3855 if (ws
->enable_exclusive
)
3856 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
3860 * A scan that reduces per-thread values across an entire workgroup.
3862 * The caller must ensure that all lanes are active when this code runs
3863 * (WWM is insufficient!), because there is an implied barrier.
3866 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3868 ac_build_wg_scan_top(ctx
, ws
);
3869 ac_build_s_barrier(ctx
);
3870 ac_build_wg_scan_bottom(ctx
, ws
);
3874 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3875 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3877 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
3878 if (ctx
->chip_class
>= VI
) {
3879 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
3881 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
3886 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
3888 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
3889 return ac_build_intrinsic(ctx
,
3890 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
3891 (LLVMValueRef
[]) {index
, src
}, 2,
3892 AC_FUNC_ATTR_READNONE
|
3893 AC_FUNC_ATTR_CONVERGENT
);