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
->v3f32
= LLVMVectorType(ctx
->f32
, 3);
88 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
89 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
91 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
92 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
93 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
94 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
95 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
96 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
97 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
98 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
99 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
100 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
101 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
102 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
103 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
104 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
106 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
107 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
109 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
112 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
113 "invariant.load", 14);
115 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
117 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
118 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
120 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
121 "amdgpu.uniform", 14);
123 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
127 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
131 ctx
->flow_depth_max
= 0;
135 ac_get_llvm_num_components(LLVMValueRef value
)
137 LLVMTypeRef type
= LLVMTypeOf(value
);
138 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
139 ? LLVMGetVectorSize(type
)
141 return num_components
;
145 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
149 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
154 return LLVMBuildExtractElement(ac
->builder
, value
,
155 LLVMConstInt(ac
->i32
, index
, false), "");
159 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
161 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
162 type
= LLVMGetElementType(type
);
164 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
165 return LLVMGetIntTypeWidth(type
);
167 if (type
== ctx
->f16
)
169 if (type
== ctx
->f32
)
171 if (type
== ctx
->f64
)
174 unreachable("Unhandled type kind in get_elem_bits");
178 ac_get_type_size(LLVMTypeRef type
)
180 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
183 case LLVMIntegerTypeKind
:
184 return LLVMGetIntTypeWidth(type
) / 8;
185 case LLVMHalfTypeKind
:
187 case LLVMFloatTypeKind
:
189 case LLVMDoubleTypeKind
:
191 case LLVMPointerTypeKind
:
192 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
195 case LLVMVectorTypeKind
:
196 return LLVMGetVectorSize(type
) *
197 ac_get_type_size(LLVMGetElementType(type
));
198 case LLVMArrayTypeKind
:
199 return LLVMGetArrayLength(type
) *
200 ac_get_type_size(LLVMGetElementType(type
));
207 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
211 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
213 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
215 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
218 unreachable("Unhandled integer size");
222 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
224 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
225 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
226 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
227 LLVMGetVectorSize(t
));
229 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
230 switch (LLVMGetPointerAddressSpace(t
)) {
231 case AC_ADDR_SPACE_GLOBAL
:
233 case AC_ADDR_SPACE_LDS
:
236 unreachable("unhandled address space");
239 return to_integer_type_scalar(ctx
, t
);
243 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
245 LLVMTypeRef type
= LLVMTypeOf(v
);
246 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
247 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
249 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
253 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
255 LLVMTypeRef type
= LLVMTypeOf(v
);
256 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
258 return ac_to_integer(ctx
, v
);
261 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
265 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
267 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
269 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
272 unreachable("Unhandled float size");
276 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
278 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
279 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
280 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
281 LLVMGetVectorSize(t
));
283 return to_float_type_scalar(ctx
, t
);
287 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
289 LLVMTypeRef type
= LLVMTypeOf(v
);
290 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
295 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
296 LLVMTypeRef return_type
, LLVMValueRef
*params
,
297 unsigned param_count
, unsigned attrib_mask
)
299 LLVMValueRef function
, call
;
300 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
302 function
= LLVMGetNamedFunction(ctx
->module
, name
);
304 LLVMTypeRef param_types
[32], function_type
;
307 assert(param_count
<= 32);
309 for (i
= 0; i
< param_count
; ++i
) {
311 param_types
[i
] = LLVMTypeOf(params
[i
]);
314 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
315 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
317 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
318 LLVMSetLinkage(function
, LLVMExternalLinkage
);
320 if (!set_callsite_attrs
)
321 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
324 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
325 if (set_callsite_attrs
)
326 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
331 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
334 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
336 LLVMTypeRef elem_type
= type
;
338 assert(bufsize
>= 8);
340 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
341 int ret
= snprintf(buf
, bufsize
, "v%u",
342 LLVMGetVectorSize(type
));
344 char *type_name
= LLVMPrintTypeToString(type
);
345 fprintf(stderr
, "Error building type name for: %s\n",
349 elem_type
= LLVMGetElementType(type
);
353 switch (LLVMGetTypeKind(elem_type
)) {
355 case LLVMIntegerTypeKind
:
356 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
358 case LLVMHalfTypeKind
:
359 snprintf(buf
, bufsize
, "f16");
361 case LLVMFloatTypeKind
:
362 snprintf(buf
, bufsize
, "f32");
364 case LLVMDoubleTypeKind
:
365 snprintf(buf
, bufsize
, "f64");
371 * Helper function that builds an LLVM IR PHI node and immediately adds
375 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
376 unsigned count_incoming
, LLVMValueRef
*values
,
377 LLVMBasicBlockRef
*blocks
)
379 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
380 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
384 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
386 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
387 0, AC_FUNC_ATTR_CONVERGENT
);
390 /* Prevent optimizations (at least of memory accesses) across the current
391 * point in the program by emitting empty inline assembly that is marked as
392 * having side effects.
394 * Optionally, a value can be passed through the inline assembly to prevent
395 * LLVM from hoisting calls to ReadNone functions.
398 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
401 static int counter
= 0;
403 LLVMBuilderRef builder
= ctx
->builder
;
406 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
409 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
410 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
411 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
413 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
414 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
415 LLVMValueRef vgpr
= *pvgpr
;
416 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
417 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
420 assert(vgpr_size
% 4 == 0);
422 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
423 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
424 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
425 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
426 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
433 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
435 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, "llvm.readcyclecounter",
436 ctx
->i64
, NULL
, 0, 0);
437 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
441 ac_build_ballot(struct ac_llvm_context
*ctx
,
444 const char *name
= HAVE_LLVM
>= 0x900 ? "llvm.amdgcn.icmp.i64.i32" : "llvm.amdgcn.icmp.i32";
445 LLVMValueRef args
[3] = {
448 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
451 /* We currently have no other way to prevent LLVM from lifting the icmp
452 * calls to a dominating basic block.
454 ac_build_optimization_barrier(ctx
, &args
[0]);
456 args
[0] = ac_to_integer(ctx
, args
[0]);
458 return ac_build_intrinsic(ctx
, name
,
460 AC_FUNC_ATTR_NOUNWIND
|
461 AC_FUNC_ATTR_READNONE
|
462 AC_FUNC_ATTR_CONVERGENT
);
465 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
468 const char *name
= HAVE_LLVM
>= 0x900 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
469 LLVMValueRef args
[3] = {
472 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0),
475 assert(HAVE_LLVM
>= 0x0800);
476 return ac_build_intrinsic(ctx
, name
, ctx
->i64
, args
, 3,
477 AC_FUNC_ATTR_NOUNWIND
|
478 AC_FUNC_ATTR_READNONE
|
479 AC_FUNC_ATTR_CONVERGENT
);
483 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
485 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
486 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
487 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
491 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
493 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
494 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
495 LLVMConstInt(ctx
->i64
, 0, 0), "");
499 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
501 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
502 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
504 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
505 vote_set
, active_set
, "");
506 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
508 LLVMConstInt(ctx
->i64
, 0, 0), "");
509 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
513 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
514 unsigned value_count
, unsigned component
)
516 LLVMValueRef vec
= NULL
;
518 if (value_count
== 1) {
519 return values
[component
];
520 } else if (!value_count
)
521 unreachable("value_count is 0");
523 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
524 LLVMValueRef value
= values
[i
];
527 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
528 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
529 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
535 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
536 LLVMValueRef
*values
,
537 unsigned value_count
,
538 unsigned value_stride
,
542 LLVMBuilderRef builder
= ctx
->builder
;
543 LLVMValueRef vec
= NULL
;
546 if (value_count
== 1 && !always_vector
) {
548 return LLVMBuildLoad(builder
, values
[0], "");
550 } else if (!value_count
)
551 unreachable("value_count is 0");
553 for (i
= 0; i
< value_count
; i
++) {
554 LLVMValueRef value
= values
[i
* value_stride
];
556 value
= LLVMBuildLoad(builder
, value
, "");
559 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
560 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
561 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
567 ac_build_gather_values(struct ac_llvm_context
*ctx
,
568 LLVMValueRef
*values
,
569 unsigned value_count
)
571 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
574 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
575 * channels with undef. Extract at most src_channels components from the input.
578 ac_build_expand(struct ac_llvm_context
*ctx
,
580 unsigned src_channels
,
581 unsigned dst_channels
)
583 LLVMTypeRef elemtype
;
584 LLVMValueRef chan
[dst_channels
];
586 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
587 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
589 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
592 src_channels
= MIN2(src_channels
, vec_size
);
594 for (unsigned i
= 0; i
< src_channels
; i
++)
595 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
597 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
600 assert(src_channels
== 1);
603 elemtype
= LLVMTypeOf(value
);
606 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
607 chan
[i
] = LLVMGetUndef(elemtype
);
609 return ac_build_gather_values(ctx
, chan
, dst_channels
);
612 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
613 * with undef. Extract at most num_channels components from the input.
615 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
617 unsigned num_channels
)
619 return ac_build_expand(ctx
, value
, num_channels
, 4);
622 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
624 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
628 name
= "llvm.rint.f16";
629 else if (type_size
== 4)
630 name
= "llvm.rint.f32";
632 name
= "llvm.rint.f64";
634 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
635 AC_FUNC_ATTR_READNONE
);
639 ac_build_fdiv(struct ac_llvm_context
*ctx
,
643 /* If we do (num / den), LLVM >= 7.0 does:
644 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
646 * If we do (num * (1 / den)), LLVM does:
647 * return num * v_rcp_f32(den);
649 LLVMValueRef one
= LLVMConstReal(LLVMTypeOf(num
), 1.0);
650 LLVMValueRef rcp
= LLVMBuildFDiv(ctx
->builder
, one
, den
, "");
651 LLVMValueRef ret
= LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
653 /* Use v_rcp_f32 instead of precise division. */
654 if (!LLVMIsConstant(ret
))
655 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
659 /* See fast_idiv_by_const.h. */
660 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
661 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
663 LLVMValueRef multiplier
,
664 LLVMValueRef pre_shift
,
665 LLVMValueRef post_shift
,
666 LLVMValueRef increment
)
668 LLVMBuilderRef builder
= ctx
->builder
;
670 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
671 num
= LLVMBuildMul(builder
,
672 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
673 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
674 num
= LLVMBuildAdd(builder
, num
,
675 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
676 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
677 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
678 return LLVMBuildLShr(builder
, num
, post_shift
, "");
681 /* See fast_idiv_by_const.h. */
682 /* If num != UINT_MAX, this more efficient version can be used. */
683 /* Set: increment = util_fast_udiv_info::increment; */
684 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
686 LLVMValueRef multiplier
,
687 LLVMValueRef pre_shift
,
688 LLVMValueRef post_shift
,
689 LLVMValueRef increment
)
691 LLVMBuilderRef builder
= ctx
->builder
;
693 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
694 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
695 num
= LLVMBuildMul(builder
,
696 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
697 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
698 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
699 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
700 return LLVMBuildLShr(builder
, num
, post_shift
, "");
703 /* See fast_idiv_by_const.h. */
704 /* Both operands must fit in 31 bits and the divisor must not be 1. */
705 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
707 LLVMValueRef multiplier
,
708 LLVMValueRef post_shift
)
710 LLVMBuilderRef builder
= ctx
->builder
;
712 num
= LLVMBuildMul(builder
,
713 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
714 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
715 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
716 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
717 return LLVMBuildLShr(builder
, num
, post_shift
, "");
720 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
721 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
722 * already multiplied by two. id is the cube face number.
724 struct cube_selection_coords
{
731 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
733 struct cube_selection_coords
*out
)
735 LLVMTypeRef f32
= ctx
->f32
;
737 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
738 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
739 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
740 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
741 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
742 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
743 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
744 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
748 * Build a manual selection sequence for cube face sc/tc coordinates and
749 * major axis vector (multiplied by 2 for consistency) for the given
750 * vec3 \p coords, for the face implied by \p selcoords.
752 * For the major axis, we always adjust the sign to be in the direction of
753 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
754 * the selcoords major axis.
756 static void build_cube_select(struct ac_llvm_context
*ctx
,
757 const struct cube_selection_coords
*selcoords
,
758 const LLVMValueRef
*coords
,
759 LLVMValueRef
*out_st
,
760 LLVMValueRef
*out_ma
)
762 LLVMBuilderRef builder
= ctx
->builder
;
763 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
764 LLVMValueRef is_ma_positive
;
766 LLVMValueRef is_ma_z
, is_not_ma_z
;
767 LLVMValueRef is_ma_y
;
768 LLVMValueRef is_ma_x
;
772 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
773 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
774 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
775 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
777 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
778 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
779 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
780 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
781 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
784 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
785 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
786 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
787 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
788 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
791 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
792 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
793 LLVMConstReal(f32
, -1.0), "");
794 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
797 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
798 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
799 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
800 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
801 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
805 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
806 bool is_deriv
, bool is_array
, bool is_lod
,
807 LLVMValueRef
*coords_arg
,
808 LLVMValueRef
*derivs_arg
)
811 LLVMBuilderRef builder
= ctx
->builder
;
812 struct cube_selection_coords selcoords
;
813 LLVMValueRef coords
[3];
816 if (is_array
&& !is_lod
) {
817 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
819 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
821 * "For Array forms, the array layer used will be
823 * max(0, min(d−1, floor(layer+0.5)))
825 * where d is the depth of the texture array and layer
826 * comes from the component indicated in the tables below.
827 * Workaroudn for an issue where the layer is taken from a
828 * helper invocation which happens to fall on a different
829 * layer due to extrapolation."
831 * GFX8 and earlier attempt to implement this in hardware by
832 * clamping the value of coords[2] = (8 * layer) + face.
833 * Unfortunately, this means that the we end up with the wrong
834 * face when clamping occurs.
836 * Clamp the layer earlier to work around the issue.
838 if (ctx
->chip_class
<= GFX8
) {
840 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
841 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
847 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
849 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
850 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
851 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
853 for (int i
= 0; i
< 2; ++i
)
854 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
856 coords
[2] = selcoords
.id
;
858 if (is_deriv
&& derivs_arg
) {
859 LLVMValueRef derivs
[4];
862 /* Convert cube derivatives to 2D derivatives. */
863 for (axis
= 0; axis
< 2; axis
++) {
864 LLVMValueRef deriv_st
[2];
865 LLVMValueRef deriv_ma
;
867 /* Transform the derivative alongside the texture
868 * coordinate. Mathematically, the correct formula is
869 * as follows. Assume we're projecting onto the +Z face
870 * and denote by dx/dh the derivative of the (original)
871 * X texture coordinate with respect to horizontal
872 * window coordinates. The projection onto the +Z face
877 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
878 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
880 * This motivatives the implementation below.
882 * Whether this actually gives the expected results for
883 * apps that might feed in derivatives obtained via
884 * finite differences is anyone's guess. The OpenGL spec
885 * seems awfully quiet about how textureGrad for cube
886 * maps should be handled.
888 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
889 deriv_st
, &deriv_ma
);
891 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
893 for (int i
= 0; i
< 2; ++i
)
894 derivs
[axis
* 2 + i
] =
895 LLVMBuildFSub(builder
,
896 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
897 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
900 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
903 /* Shift the texture coordinate. This must be applied after the
904 * derivative calculation.
906 for (int i
= 0; i
< 2; ++i
)
907 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
910 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
911 /* coords_arg.w component - array_index for cube arrays */
912 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
915 memcpy(coords_arg
, coords
, sizeof(coords
));
920 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
921 LLVMValueRef llvm_chan
,
922 LLVMValueRef attr_number
,
927 LLVMValueRef args
[5];
932 args
[2] = attr_number
;
935 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
936 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
941 args
[3] = attr_number
;
944 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
945 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
949 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
950 LLVMValueRef llvm_chan
,
951 LLVMValueRef attr_number
,
956 LLVMValueRef args
[6];
961 args
[2] = attr_number
;
962 args
[3] = ctx
->i1false
;
965 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
966 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
971 args
[3] = attr_number
;
972 args
[4] = ctx
->i1false
;
975 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
976 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
980 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
981 LLVMValueRef parameter
,
982 LLVMValueRef llvm_chan
,
983 LLVMValueRef attr_number
,
986 LLVMValueRef args
[4];
990 args
[2] = attr_number
;
993 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
994 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
998 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
999 LLVMValueRef base_ptr
,
1002 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1006 ac_build_gep0(struct ac_llvm_context
*ctx
,
1007 LLVMValueRef base_ptr
,
1010 LLVMValueRef indices
[2] = {
1014 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1017 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1020 return LLVMBuildPointerCast(ctx
->builder
,
1021 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1022 LLVMTypeOf(ptr
), "");
1026 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1027 LLVMValueRef base_ptr
, LLVMValueRef index
,
1030 LLVMBuildStore(ctx
->builder
, value
,
1031 ac_build_gep0(ctx
, base_ptr
, index
));
1035 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1036 * It's equivalent to doing a load from &base_ptr[index].
1038 * \param base_ptr Where the array starts.
1039 * \param index The element index into the array.
1040 * \param uniform Whether the base_ptr and index can be assumed to be
1041 * dynamically uniform (i.e. load to an SGPR)
1042 * \param invariant Whether the load is invariant (no other opcodes affect it)
1043 * \param no_unsigned_wraparound
1044 * For all possible re-associations and re-distributions of an expression
1045 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1046 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1047 * does not result in an unsigned integer wraparound. This is used for
1048 * optimal code generation of 32-bit pointer arithmetic.
1050 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1051 * integer wraparound can't be an imm offset in s_load_dword, because
1052 * the instruction performs "addr + offset" in 64 bits.
1054 * Expected usage for bindless textures by chaining GEPs:
1055 * // possible unsigned wraparound, don't use InBounds:
1056 * ptr1 = LLVMBuildGEP(base_ptr, index);
1057 * image = load(ptr1); // becomes "s_load ptr1, 0"
1059 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1060 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1063 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1064 LLVMValueRef index
, bool uniform
, bool invariant
,
1065 bool no_unsigned_wraparound
)
1067 LLVMValueRef pointer
, result
;
1069 if (no_unsigned_wraparound
&&
1070 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1071 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1073 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1076 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1077 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1079 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1083 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1086 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1089 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1090 LLVMValueRef base_ptr
, LLVMValueRef index
)
1092 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1095 /* This assumes that there is no unsigned integer wraparound during the address
1096 * computation, excluding all GEPs within base_ptr. */
1097 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1098 LLVMValueRef base_ptr
, LLVMValueRef index
)
1100 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1103 /* See ac_build_load_custom() documentation. */
1104 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1105 LLVMValueRef base_ptr
, LLVMValueRef index
)
1107 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1111 ac_build_llvm7_buffer_store_common(struct ac_llvm_context
*ctx
,
1114 LLVMValueRef vindex
,
1115 LLVMValueRef voffset
,
1116 unsigned num_channels
,
1121 LLVMValueRef args
[] = {
1123 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1124 vindex
? vindex
: ctx
->i32_0
,
1126 LLVMConstInt(ctx
->i1
, glc
, 0),
1127 LLVMConstInt(ctx
->i1
, slc
, 0)
1129 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1131 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1135 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.format.%s",
1138 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
1142 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, ARRAY_SIZE(args
),
1143 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1147 ac_build_llvm8_buffer_store_common(struct ac_llvm_context
*ctx
,
1150 LLVMValueRef vindex
,
1151 LLVMValueRef voffset
,
1152 LLVMValueRef soffset
,
1153 unsigned num_channels
,
1154 LLVMTypeRef return_channel_type
,
1160 LLVMValueRef args
[6];
1163 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1165 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1166 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1167 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1168 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1169 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1170 const char *indexing_kind
= structurized
? "struct" : "raw";
1171 char name
[256], type_name
[8];
1173 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(return_channel_type
, func
) : return_channel_type
;
1174 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1177 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1178 indexing_kind
, type_name
);
1180 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1181 indexing_kind
, type_name
);
1184 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1185 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1189 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1192 LLVMValueRef vindex
,
1193 LLVMValueRef voffset
,
1194 unsigned num_channels
,
1198 if (HAVE_LLVM
>= 0x800) {
1199 ac_build_llvm8_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1200 voffset
, NULL
, num_channels
,
1204 ac_build_llvm7_buffer_store_common(ctx
, rsrc
, data
, vindex
, voffset
,
1205 num_channels
, glc
, slc
,
1210 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1211 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1212 * or v4i32 (num_channels=3,4).
1215 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1218 unsigned num_channels
,
1219 LLVMValueRef voffset
,
1220 LLVMValueRef soffset
,
1221 unsigned inst_offset
,
1224 bool swizzle_enable_hint
)
1226 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1228 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1229 LLVMValueRef v
[3], v01
;
1231 for (int i
= 0; i
< 3; i
++) {
1232 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1233 LLVMConstInt(ctx
->i32
, i
, 0), "");
1235 v01
= ac_build_gather_values(ctx
, v
, 2);
1237 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1238 soffset
, inst_offset
, glc
, slc
,
1239 swizzle_enable_hint
);
1240 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1241 soffset
, inst_offset
+ 8,
1243 swizzle_enable_hint
);
1247 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1248 * (voffset is swizzled, but soffset isn't swizzled).
1249 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1251 if (!swizzle_enable_hint
) {
1252 LLVMValueRef offset
= soffset
;
1255 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1256 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1258 if (HAVE_LLVM
>= 0x800) {
1259 ac_build_llvm8_buffer_store_common(ctx
, rsrc
,
1260 ac_to_float(ctx
, vdata
),
1269 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1271 ac_build_llvm7_buffer_store_common(ctx
, rsrc
,
1272 ac_to_float(ctx
, vdata
),
1274 num_channels
, glc
, slc
,
1280 static const unsigned dfmts
[] = {
1281 V_008F0C_BUF_DATA_FORMAT_32
,
1282 V_008F0C_BUF_DATA_FORMAT_32_32
,
1283 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1284 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1286 unsigned dfmt
= dfmts
[num_channels
- 1];
1287 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1288 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1290 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1291 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1296 ac_build_llvm7_buffer_load_common(struct ac_llvm_context
*ctx
,
1298 LLVMValueRef vindex
,
1299 LLVMValueRef voffset
,
1300 unsigned num_channels
,
1306 LLVMValueRef args
[] = {
1307 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1308 vindex
? vindex
: ctx
->i32_0
,
1310 LLVMConstInt(ctx
->i1
, glc
, 0),
1311 LLVMConstInt(ctx
->i1
, slc
, 0)
1313 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1315 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
1316 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1320 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.format.%s",
1323 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
1327 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1329 ac_get_load_intr_attribs(can_speculate
));
1333 ac_build_llvm8_buffer_load_common(struct ac_llvm_context
*ctx
,
1335 LLVMValueRef vindex
,
1336 LLVMValueRef voffset
,
1337 LLVMValueRef soffset
,
1338 unsigned num_channels
,
1339 LLVMTypeRef channel_type
,
1346 LLVMValueRef args
[5];
1348 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1350 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1351 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1352 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1353 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1354 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1355 const char *indexing_kind
= structurized
? "struct" : "raw";
1356 char name
[256], type_name
[8];
1358 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1359 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1362 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1363 indexing_kind
, type_name
);
1365 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1366 indexing_kind
, type_name
);
1369 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1370 ac_get_load_intr_attribs(can_speculate
));
1374 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1377 LLVMValueRef vindex
,
1378 LLVMValueRef voffset
,
1379 LLVMValueRef soffset
,
1380 unsigned inst_offset
,
1386 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1388 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1390 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1392 if (allow_smem
&& !slc
&&
1393 (!glc
|| (HAVE_LLVM
>= 0x0800 && ctx
->chip_class
>= GFX8
))) {
1394 assert(vindex
== NULL
);
1396 LLVMValueRef result
[8];
1398 for (int i
= 0; i
< num_channels
; i
++) {
1400 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1401 LLVMConstInt(ctx
->i32
, 4, 0), "");
1403 const char *intrname
=
1404 HAVE_LLVM
>= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1405 : "llvm.SI.load.const.v4i32";
1406 unsigned num_args
= HAVE_LLVM
>= 0x0800 ? 3 : 2;
1407 LLVMValueRef args
[3] = {
1410 glc
? ctx
->i32_1
: ctx
->i32_0
,
1412 result
[i
] = ac_build_intrinsic(ctx
, intrname
,
1413 ctx
->f32
, args
, num_args
,
1414 AC_FUNC_ATTR_READNONE
|
1415 (HAVE_LLVM
< 0x0800 ? AC_FUNC_ATTR_LEGACY
: 0));
1417 if (num_channels
== 1)
1420 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1421 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1422 return ac_build_gather_values(ctx
, result
, num_channels
);
1425 if (HAVE_LLVM
>= 0x0800) {
1426 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
,
1428 num_channels
, ctx
->f32
,
1430 can_speculate
, false,
1434 return ac_build_llvm7_buffer_load_common(ctx
, rsrc
, vindex
, offset
,
1435 num_channels
, glc
, slc
,
1436 can_speculate
, false);
1439 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1441 LLVMValueRef vindex
,
1442 LLVMValueRef voffset
,
1443 unsigned num_channels
,
1447 if (HAVE_LLVM
>= 0x800) {
1448 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1449 num_channels
, ctx
->f32
,
1451 can_speculate
, true, true);
1453 return ac_build_llvm7_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1454 num_channels
, glc
, false,
1455 can_speculate
, true);
1458 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
1460 LLVMValueRef vindex
,
1461 LLVMValueRef voffset
,
1462 unsigned num_channels
,
1466 if (HAVE_LLVM
>= 0x800) {
1467 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1468 num_channels
, ctx
->f32
,
1470 can_speculate
, true, true);
1473 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, LLVMConstInt(ctx
->i32
, 2, 0), "");
1474 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, ctx
->i32_1
, "");
1475 stride
= LLVMBuildLShr(ctx
->builder
, stride
, LLVMConstInt(ctx
->i32
, 16, 0), "");
1477 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->builder
,
1478 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, elem_count
, stride
, ""),
1479 elem_count
, stride
, "");
1481 LLVMValueRef new_rsrc
= LLVMBuildInsertElement(ctx
->builder
, rsrc
, new_elem_count
,
1482 LLVMConstInt(ctx
->i32
, 2, 0), "");
1484 return ac_build_llvm7_buffer_load_common(ctx
, new_rsrc
, vindex
, voffset
,
1485 num_channels
, glc
, false,
1486 can_speculate
, true);
1490 ac_build_llvm8_tbuffer_load(struct ac_llvm_context
*ctx
,
1492 LLVMValueRef vindex
,
1493 LLVMValueRef voffset
,
1494 LLVMValueRef soffset
,
1495 unsigned num_channels
,
1503 LLVMValueRef args
[6];
1505 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1507 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1508 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1509 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1510 args
[idx
++] = LLVMConstInt(ctx
->i32
, dfmt
| (nfmt
<< 4), 0);
1511 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1512 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1513 const char *indexing_kind
= structurized
? "struct" : "raw";
1514 char name
[256], type_name
[8];
1516 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1517 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1519 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1520 indexing_kind
, type_name
);
1522 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1523 ac_get_load_intr_attribs(can_speculate
));
1527 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1529 LLVMValueRef vindex
,
1530 LLVMValueRef voffset
,
1531 LLVMValueRef soffset
,
1532 LLVMValueRef immoffset
,
1533 unsigned num_channels
,
1539 bool structurized
) /* only matters for LLVM 8+ */
1541 if (HAVE_LLVM
>= 0x800) {
1542 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1544 return ac_build_llvm8_tbuffer_load(ctx
, rsrc
, vindex
, voffset
,
1545 soffset
, num_channels
,
1546 dfmt
, nfmt
, glc
, slc
,
1547 can_speculate
, structurized
);
1550 LLVMValueRef args
[] = {
1552 vindex
? vindex
: ctx
->i32_0
,
1556 LLVMConstInt(ctx
->i32
, dfmt
, false),
1557 LLVMConstInt(ctx
->i32
, nfmt
, false),
1558 LLVMConstInt(ctx
->i1
, glc
, false),
1559 LLVMConstInt(ctx
->i1
, slc
, false),
1561 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1562 LLVMTypeRef types
[] = {ctx
->i32
, ctx
->v2i32
, ctx
->v4i32
};
1563 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
1566 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.load.%s",
1569 return ac_build_intrinsic(ctx
, name
, types
[func
], args
, 9,
1570 ac_get_load_intr_attribs(can_speculate
));
1574 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1576 LLVMValueRef vindex
,
1577 LLVMValueRef voffset
,
1578 LLVMValueRef soffset
,
1579 LLVMValueRef immoffset
,
1580 unsigned num_channels
,
1587 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1588 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1589 slc
, can_speculate
, true);
1593 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1595 LLVMValueRef voffset
,
1596 LLVMValueRef soffset
,
1597 LLVMValueRef immoffset
,
1598 unsigned num_channels
,
1605 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1606 immoffset
, num_channels
, dfmt
, nfmt
, glc
,
1607 slc
, can_speculate
, false);
1611 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1613 LLVMValueRef voffset
,
1614 LLVMValueRef soffset
,
1615 LLVMValueRef immoffset
,
1620 if (HAVE_LLVM
>= 0x900) {
1621 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1623 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1624 res
= ac_build_llvm8_buffer_load_common(ctx
, rsrc
, NULL
,
1626 1, ctx
->i16
, glc
, false,
1627 false, false, false);
1629 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1630 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1632 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1633 immoffset
, 1, dfmt
, nfmt
, glc
, false,
1636 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1643 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1645 LLVMValueRef voffset
,
1646 LLVMValueRef soffset
,
1647 LLVMValueRef immoffset
,
1652 if (HAVE_LLVM
>= 0x900) {
1653 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1655 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1656 res
= ac_build_llvm8_buffer_load_common(ctx
, rsrc
, NULL
,
1658 1, ctx
->i8
, glc
, false,
1659 false, false, false);
1661 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1662 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1664 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1665 immoffset
, 1, dfmt
, nfmt
, glc
, false,
1668 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1675 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1677 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1678 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1681 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1683 assert(LLVMTypeOf(src
) == ctx
->i32
);
1686 LLVMValueRef mantissa
;
1687 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1689 /* Converting normal numbers is just a shift + correcting the exponent bias */
1690 unsigned normal_shift
= 23 - mant_bits
;
1691 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1692 LLVMValueRef shifted
, normal
;
1694 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1695 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1697 /* Converting nan/inf numbers is the same, but with a different exponent update */
1698 LLVMValueRef naninf
;
1699 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1701 /* Converting denormals is the complex case: determine the leading zeros of the
1702 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1704 LLVMValueRef denormal
;
1705 LLVMValueRef params
[2] = {
1707 ctx
->i1true
, /* result can be undef when arg is 0 */
1709 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1710 params
, 2, AC_FUNC_ATTR_READNONE
);
1712 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1713 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1714 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1716 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1717 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1718 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1719 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1721 /* Select the final result. */
1722 LLVMValueRef result
;
1724 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1725 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1726 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1728 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1729 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1730 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1732 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1733 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1735 return ac_to_float(ctx
, result
);
1739 * Generate a fully general open coded buffer format fetch with all required
1740 * fixups suitable for vertex fetch, using non-format buffer loads.
1742 * Some combinations of argument values have special interpretations:
1743 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1744 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1746 * \param log_size log(size of channel in bytes)
1747 * \param num_channels number of channels (1 to 4)
1748 * \param format AC_FETCH_FORMAT_xxx value
1749 * \param reverse whether XYZ channels are reversed
1750 * \param known_aligned whether the source is known to be aligned to hardware's
1751 * effective element size for loading the given format
1752 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1753 * \param rsrc buffer resource descriptor
1754 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1757 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1759 unsigned num_channels
,
1764 LLVMValueRef vindex
,
1765 LLVMValueRef voffset
,
1766 LLVMValueRef soffset
,
1772 unsigned load_log_size
= log_size
;
1773 unsigned load_num_channels
= num_channels
;
1774 if (log_size
== 3) {
1776 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1777 load_num_channels
= 2 * num_channels
;
1779 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1783 int log_recombine
= 0;
1784 if (ctx
->chip_class
== GFX6
&& !known_aligned
) {
1785 /* Avoid alignment restrictions by loading one byte at a time. */
1786 load_num_channels
<<= load_log_size
;
1787 log_recombine
= load_log_size
;
1789 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1790 log_recombine
= -util_logbase2(load_num_channels
);
1791 load_num_channels
= 1;
1792 load_log_size
+= -log_recombine
;
1795 assert(load_log_size
>= 2 || HAVE_LLVM
>= 0x0900);
1797 LLVMValueRef loads
[32]; /* up to 32 bytes */
1798 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1799 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1800 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1801 if (HAVE_LLVM
>= 0x0800) {
1802 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1803 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1804 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1805 loads
[i
] = ac_build_llvm8_buffer_load_common(
1806 ctx
, rsrc
, vindex
, voffset
, tmp
,
1807 num_channels
, channel_type
, glc
, slc
,
1808 can_speculate
, false, true);
1810 tmp
= LLVMBuildAdd(ctx
->builder
, voffset
, tmp
, "");
1811 loads
[i
] = ac_build_llvm7_buffer_load_common(
1812 ctx
, rsrc
, vindex
, tmp
,
1813 1 << (load_log_size
- 2), glc
, slc
, can_speculate
, false);
1815 if (load_log_size
>= 2)
1816 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1819 if (log_recombine
> 0) {
1820 /* Recombine bytes if necessary (GFX6 only) */
1821 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1823 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1824 LLVMValueRef accum
= NULL
;
1825 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1826 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1830 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1831 LLVMConstInt(dst_type
, 8 * i
, false), "");
1832 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1837 } else if (log_recombine
< 0) {
1838 /* Split vectors of dwords */
1839 if (load_log_size
> 2) {
1840 assert(load_num_channels
== 1);
1841 LLVMValueRef loaded
= loads
[0];
1842 unsigned log_split
= load_log_size
- 2;
1843 log_recombine
+= log_split
;
1844 load_num_channels
= 1 << log_split
;
1846 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1847 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1848 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1852 /* Further split dwords and shorts if required */
1853 if (log_recombine
< 0) {
1854 for (unsigned src
= load_num_channels
,
1855 dst
= load_num_channels
<< -log_recombine
;
1857 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1858 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1859 LLVMValueRef loaded
= loads
[src
- 1];
1860 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1861 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1862 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1863 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1864 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1870 if (log_size
== 3) {
1871 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1872 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1873 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1874 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1876 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1877 /* 10_11_11_FLOAT */
1878 LLVMValueRef data
= loads
[0];
1879 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1880 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1881 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1882 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1883 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1885 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1886 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1887 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1891 format
= AC_FETCH_FORMAT_FLOAT
;
1893 /* 2_10_10_10 data formats */
1894 LLVMValueRef data
= loads
[0];
1895 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1896 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1897 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1898 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1899 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1900 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1901 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1902 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1903 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1909 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1910 if (log_size
!= 2) {
1911 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1912 tmp
= ac_to_float(ctx
, loads
[chan
]);
1914 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1915 else if (log_size
== 1)
1916 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1917 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1920 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1921 if (log_size
!= 2) {
1922 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1923 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1925 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1926 if (log_size
!= 2) {
1927 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1928 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1931 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1932 format
== AC_FETCH_FORMAT_USCALED
||
1933 format
== AC_FETCH_FORMAT_UINT
;
1935 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1937 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1939 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1942 LLVMValueRef scale
= NULL
;
1943 if (format
== AC_FETCH_FORMAT_FIXED
) {
1944 assert(log_size
== 2);
1945 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1946 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1947 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1948 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1949 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1950 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1951 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1954 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1956 if (format
== AC_FETCH_FORMAT_SNORM
) {
1957 /* Clamp to [-1, 1] */
1958 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1959 LLVMValueRef clamp
=
1960 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1961 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1964 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1968 while (num_channels
< 4) {
1969 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1970 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1972 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1979 loads
[0] = loads
[2];
1983 return ac_build_gather_values(ctx
, loads
, 4);
1987 ac_build_llvm8_tbuffer_store(struct ac_llvm_context
*ctx
,
1990 LLVMValueRef vindex
,
1991 LLVMValueRef voffset
,
1992 LLVMValueRef soffset
,
1993 unsigned num_channels
,
2000 LLVMValueRef args
[7];
2002 args
[idx
++] = vdata
;
2003 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
2005 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
2006 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
2007 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
2008 args
[idx
++] = LLVMConstInt(ctx
->i32
, dfmt
| (nfmt
<< 4), 0);
2009 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
2010 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
2011 const char *indexing_kind
= structurized
? "struct" : "raw";
2012 char name
[256], type_name
[8];
2014 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
2015 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
2017 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
2018 indexing_kind
, type_name
);
2020 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
2021 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
2025 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
2028 LLVMValueRef vindex
,
2029 LLVMValueRef voffset
,
2030 LLVMValueRef soffset
,
2031 LLVMValueRef immoffset
,
2032 unsigned num_channels
,
2037 bool structurized
) /* only matters for LLVM 8+ */
2039 if (HAVE_LLVM
>= 0x800) {
2040 voffset
= LLVMBuildAdd(ctx
->builder
,
2041 voffset
? voffset
: ctx
->i32_0
,
2044 ac_build_llvm8_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
,
2045 soffset
, num_channels
, dfmt
, nfmt
,
2046 glc
, slc
, structurized
);
2048 LLVMValueRef params
[] = {
2051 vindex
? vindex
: ctx
->i32_0
,
2052 voffset
? voffset
: ctx
->i32_0
,
2053 soffset
? soffset
: ctx
->i32_0
,
2055 LLVMConstInt(ctx
->i32
, dfmt
, false),
2056 LLVMConstInt(ctx
->i32
, nfmt
, false),
2057 LLVMConstInt(ctx
->i1
, glc
, false),
2058 LLVMConstInt(ctx
->i1
, slc
, false),
2060 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
2061 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
2064 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.store.%s",
2067 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, params
, 10,
2068 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
2073 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
2076 LLVMValueRef vindex
,
2077 LLVMValueRef voffset
,
2078 LLVMValueRef soffset
,
2079 LLVMValueRef immoffset
,
2080 unsigned num_channels
,
2086 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
2087 immoffset
, num_channels
, dfmt
, nfmt
, glc
, slc
,
2092 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
2095 LLVMValueRef voffset
,
2096 LLVMValueRef soffset
,
2097 LLVMValueRef immoffset
,
2098 unsigned num_channels
,
2104 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
2105 immoffset
, num_channels
, dfmt
, nfmt
, glc
, slc
,
2110 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
2113 LLVMValueRef voffset
,
2114 LLVMValueRef soffset
,
2117 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
2119 if (HAVE_LLVM
>= 0x900) {
2120 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2121 ac_build_llvm8_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
2122 voffset
, soffset
, 1,
2123 ctx
->i16
, glc
, false,
2126 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
2127 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
2129 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
2131 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
2132 ctx
->i32_0
, 1, dfmt
, nfmt
, glc
,
2138 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
2141 LLVMValueRef voffset
,
2142 LLVMValueRef soffset
,
2145 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
2147 if (HAVE_LLVM
>= 0x900) {
2148 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2149 ac_build_llvm8_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
2150 voffset
, soffset
, 1,
2151 ctx
->i8
, glc
, false,
2154 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
2155 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
2157 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
2159 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
2160 ctx
->i32_0
, 1, dfmt
, nfmt
, glc
, false);
2164 * Set range metadata on an instruction. This can only be used on load and
2165 * call instructions. If you know an instruction can only produce the values
2166 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
2167 * \p lo is the minimum value inclusive.
2168 * \p hi is the maximum value exclusive.
2170 static void set_range_metadata(struct ac_llvm_context
*ctx
,
2171 LLVMValueRef value
, unsigned lo
, unsigned hi
)
2173 LLVMValueRef range_md
, md_args
[2];
2174 LLVMTypeRef type
= LLVMTypeOf(value
);
2175 LLVMContextRef context
= LLVMGetTypeContext(type
);
2177 md_args
[0] = LLVMConstInt(type
, lo
, false);
2178 md_args
[1] = LLVMConstInt(type
, hi
, false);
2179 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
2180 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
2184 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2188 LLVMValueRef tid_args
[2];
2189 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2190 tid_args
[1] = ctx
->i32_0
;
2191 tid_args
[1] = ac_build_intrinsic(ctx
,
2192 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2193 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2195 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2197 2, AC_FUNC_ATTR_READNONE
);
2198 set_range_metadata(ctx
, tid
, 0, 64);
2203 * AMD GCN implements derivatives using the local data store (LDS)
2204 * All writes to the LDS happen in all executing threads at
2205 * the same time. TID is the Thread ID for the current
2206 * thread and is a value between 0 and 63, representing
2207 * the thread's position in the wavefront.
2209 * For the pixel shader threads are grouped into quads of four pixels.
2210 * The TIDs of the pixels of a quad are:
2218 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2219 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2220 * the current pixel's column, and masking with 0xfffffffe yields the TID
2221 * of the left pixel of the current pixel's row.
2223 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2224 * adding 2 yields the TID of the pixel below the top pixel.
2227 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2232 unsigned tl_lanes
[4], trbl_lanes
[4];
2233 char name
[32], type
[8];
2234 LLVMValueRef tl
, trbl
;
2235 LLVMTypeRef result_type
;
2236 LLVMValueRef result
;
2238 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2240 if (result_type
== ctx
->f16
)
2241 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2243 for (unsigned i
= 0; i
< 4; ++i
) {
2244 tl_lanes
[i
] = i
& mask
;
2245 trbl_lanes
[i
] = (i
& mask
) + idx
;
2248 tl
= ac_build_quad_swizzle(ctx
, val
,
2249 tl_lanes
[0], tl_lanes
[1],
2250 tl_lanes
[2], tl_lanes
[3]);
2251 trbl
= ac_build_quad_swizzle(ctx
, val
,
2252 trbl_lanes
[0], trbl_lanes
[1],
2253 trbl_lanes
[2], trbl_lanes
[3]);
2255 if (result_type
== ctx
->f16
) {
2256 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2257 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2260 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2261 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2262 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2264 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2265 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2267 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2271 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2273 LLVMValueRef wave_id
)
2275 LLVMValueRef args
[2];
2276 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2278 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2282 ac_build_imsb(struct ac_llvm_context
*ctx
,
2284 LLVMTypeRef dst_type
)
2286 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2288 AC_FUNC_ATTR_READNONE
);
2290 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2291 * the index from LSB. Invert it by doing "31 - msb". */
2292 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2295 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2296 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2297 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2298 arg
, ctx
->i32_0
, ""),
2299 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2300 arg
, all_ones
, ""), "");
2302 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2306 ac_build_umsb(struct ac_llvm_context
*ctx
,
2308 LLVMTypeRef dst_type
)
2310 const char *intrin_name
;
2312 LLVMValueRef highest_bit
;
2316 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2319 intrin_name
= "llvm.ctlz.i64";
2321 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2325 intrin_name
= "llvm.ctlz.i32";
2327 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2331 intrin_name
= "llvm.ctlz.i16";
2333 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2337 intrin_name
= "llvm.ctlz.i8";
2339 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2343 unreachable(!"invalid bitsize");
2347 LLVMValueRef params
[2] = {
2352 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2354 AC_FUNC_ATTR_READNONE
);
2356 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2357 * the index from LSB. Invert it by doing "31 - msb". */
2358 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2360 if (bitsize
== 64) {
2361 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2362 } else if (bitsize
< 32) {
2363 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2366 /* check for zero */
2367 return LLVMBuildSelect(ctx
->builder
,
2368 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2369 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2372 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2376 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2377 LLVMValueRef args
[2] = {a
, b
};
2378 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2379 AC_FUNC_ATTR_READNONE
);
2382 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2386 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2387 LLVMValueRef args
[2] = {a
, b
};
2388 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2389 AC_FUNC_ATTR_READNONE
);
2392 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2395 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2396 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2399 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2402 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2403 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2406 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2409 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2410 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2413 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2416 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2417 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2420 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2422 LLVMTypeRef t
= LLVMTypeOf(value
);
2423 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2424 LLVMConstReal(t
, 1.0));
2427 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2429 LLVMValueRef args
[9];
2431 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2432 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2435 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2436 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2438 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2440 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2442 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2443 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2445 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2446 ctx
->voidt
, args
, 6, 0);
2448 args
[2] = a
->out
[0];
2449 args
[3] = a
->out
[1];
2450 args
[4] = a
->out
[2];
2451 args
[5] = a
->out
[3];
2452 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2453 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2455 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2456 ctx
->voidt
, args
, 8, 0);
2460 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2462 struct ac_export_args args
;
2464 args
.enabled_channels
= 0x0; /* enabled channels */
2465 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2466 args
.done
= 1; /* DONE bit */
2467 args
.target
= V_008DFC_SQ_EXP_NULL
;
2468 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2469 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2470 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2471 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2472 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2474 ac_build_export(ctx
, &args
);
2477 static unsigned ac_num_coords(enum ac_image_dim dim
)
2483 case ac_image_1darray
:
2487 case ac_image_2darray
:
2488 case ac_image_2dmsaa
:
2490 case ac_image_2darraymsaa
:
2493 unreachable("ac_num_coords: bad dim");
2497 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2501 case ac_image_1darray
:
2504 case ac_image_2darray
:
2509 case ac_image_2dmsaa
:
2510 case ac_image_2darraymsaa
:
2512 unreachable("derivatives not supported");
2516 static const char *get_atomic_name(enum ac_atomic_op op
)
2519 case ac_atomic_swap
: return "swap";
2520 case ac_atomic_add
: return "add";
2521 case ac_atomic_sub
: return "sub";
2522 case ac_atomic_smin
: return "smin";
2523 case ac_atomic_umin
: return "umin";
2524 case ac_atomic_smax
: return "smax";
2525 case ac_atomic_umax
: return "umax";
2526 case ac_atomic_and
: return "and";
2527 case ac_atomic_or
: return "or";
2528 case ac_atomic_xor
: return "xor";
2530 unreachable("bad atomic op");
2533 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2534 struct ac_image_args
*a
)
2536 const char *overload
[3] = { "", "", "" };
2537 unsigned num_overloads
= 0;
2538 LLVMValueRef args
[18];
2539 unsigned num_args
= 0;
2540 enum ac_image_dim dim
= a
->dim
;
2542 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2544 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2545 a
->opcode
!= ac_image_store_mip
) ||
2547 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2548 (!a
->compare
&& !a
->offset
));
2549 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2550 a
->opcode
== ac_image_get_lod
) ||
2552 assert((a
->bias
? 1 : 0) +
2554 (a
->level_zero
? 1 : 0) +
2555 (a
->derivs
[0] ? 1 : 0) <= 1);
2557 if (a
->opcode
== ac_image_get_lod
) {
2559 case ac_image_1darray
:
2562 case ac_image_2darray
:
2571 bool sample
= a
->opcode
== ac_image_sample
||
2572 a
->opcode
== ac_image_gather4
||
2573 a
->opcode
== ac_image_get_lod
;
2574 bool atomic
= a
->opcode
== ac_image_atomic
||
2575 a
->opcode
== ac_image_atomic_cmpswap
;
2576 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2578 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2579 args
[num_args
++] = a
->data
[0];
2580 if (a
->opcode
== ac_image_atomic_cmpswap
)
2581 args
[num_args
++] = a
->data
[1];
2585 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2588 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2590 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2591 overload
[num_overloads
++] = ".f32";
2594 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2596 unsigned count
= ac_num_derivs(dim
);
2597 for (unsigned i
= 0; i
< count
; ++i
)
2598 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2599 overload
[num_overloads
++] = ".f32";
2601 unsigned num_coords
=
2602 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2603 for (unsigned i
= 0; i
< num_coords
; ++i
)
2604 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2606 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2607 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2609 args
[num_args
++] = a
->resource
;
2611 args
[num_args
++] = a
->sampler
;
2612 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2615 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2616 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->cache_policy
, false);
2619 const char *atomic_subop
= "";
2620 switch (a
->opcode
) {
2621 case ac_image_sample
: name
= "sample"; break;
2622 case ac_image_gather4
: name
= "gather4"; break;
2623 case ac_image_load
: name
= "load"; break;
2624 case ac_image_load_mip
: name
= "load.mip"; break;
2625 case ac_image_store
: name
= "store"; break;
2626 case ac_image_store_mip
: name
= "store.mip"; break;
2627 case ac_image_atomic
:
2629 atomic_subop
= get_atomic_name(a
->atomic
);
2631 case ac_image_atomic_cmpswap
:
2633 atomic_subop
= "cmpswap";
2635 case ac_image_get_lod
: name
= "getlod"; break;
2636 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2637 default: unreachable("invalid image opcode");
2640 const char *dimname
;
2642 case ac_image_1d
: dimname
= "1d"; break;
2643 case ac_image_2d
: dimname
= "2d"; break;
2644 case ac_image_3d
: dimname
= "3d"; break;
2645 case ac_image_cube
: dimname
= "cube"; break;
2646 case ac_image_1darray
: dimname
= "1darray"; break;
2647 case ac_image_2darray
: dimname
= "2darray"; break;
2648 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2649 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2650 default: unreachable("invalid dim");
2654 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2656 snprintf(intr_name
, sizeof(intr_name
),
2657 "llvm.amdgcn.image.%s%s" /* base name */
2658 "%s%s%s" /* sample/gather modifiers */
2659 ".%s.%s%s%s%s", /* dimension and type overloads */
2661 a
->compare
? ".c" : "",
2664 a
->derivs
[0] ? ".d" :
2665 a
->level_zero
? ".lz" : "",
2666 a
->offset
? ".o" : "",
2668 atomic
? "i32" : "v4f32",
2669 overload
[0], overload
[1], overload
[2]);
2674 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2679 LLVMValueRef result
=
2680 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2682 if (!sample
&& retty
== ctx
->v4f32
) {
2683 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2689 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2690 LLVMValueRef args
[2])
2693 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2695 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2696 args
, 2, AC_FUNC_ATTR_READNONE
);
2699 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2700 LLVMValueRef args
[2])
2703 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2704 ctx
->v2i16
, args
, 2,
2705 AC_FUNC_ATTR_READNONE
);
2706 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2709 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2710 LLVMValueRef args
[2])
2713 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2714 ctx
->v2i16
, args
, 2,
2715 AC_FUNC_ATTR_READNONE
);
2716 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2719 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2720 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2721 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2723 assert(bits
== 8 || bits
== 10 || bits
== 16);
2725 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2726 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2727 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2728 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2729 LLVMValueRef max_alpha
=
2730 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2731 LLVMValueRef min_alpha
=
2732 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2736 for (int i
= 0; i
< 2; i
++) {
2737 bool alpha
= hi
&& i
== 1;
2738 args
[i
] = ac_build_imin(ctx
, args
[i
],
2739 alpha
? max_alpha
: max_rgb
);
2740 args
[i
] = ac_build_imax(ctx
, args
[i
],
2741 alpha
? min_alpha
: min_rgb
);
2746 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2747 ctx
->v2i16
, args
, 2,
2748 AC_FUNC_ATTR_READNONE
);
2749 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2752 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2753 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2754 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2756 assert(bits
== 8 || bits
== 10 || bits
== 16);
2758 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2759 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2760 LLVMValueRef max_alpha
=
2761 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2765 for (int i
= 0; i
< 2; i
++) {
2766 bool alpha
= hi
&& i
== 1;
2767 args
[i
] = ac_build_umin(ctx
, args
[i
],
2768 alpha
? max_alpha
: max_rgb
);
2773 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2774 ctx
->v2i16
, args
, 2,
2775 AC_FUNC_ATTR_READNONE
);
2776 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2779 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2781 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2782 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2785 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2787 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2791 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2792 LLVMValueRef offset
, LLVMValueRef width
,
2795 LLVMValueRef args
[] = {
2801 LLVMValueRef result
= ac_build_intrinsic(ctx
,
2802 is_signed
? "llvm.amdgcn.sbfe.i32" :
2803 "llvm.amdgcn.ubfe.i32",
2805 AC_FUNC_ATTR_READNONE
);
2807 if (HAVE_LLVM
< 0x0800) {
2808 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
2809 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
2811 LLVMValueRef zero
= ctx
->i32_0
;
2812 LLVMValueRef icond
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, width
, zero
, "");
2813 result
= LLVMBuildSelect(ctx
->builder
, icond
, zero
, result
, "");
2819 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2820 LLVMValueRef s1
, LLVMValueRef s2
)
2822 return LLVMBuildAdd(ctx
->builder
,
2823 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2826 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2827 LLVMValueRef s1
, LLVMValueRef s2
)
2829 return LLVMBuildFAdd(ctx
->builder
,
2830 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2833 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
2835 LLVMValueRef args
[1] = {
2836 LLVMConstInt(ctx
->i32
, simm16
, false),
2838 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2839 ctx
->voidt
, args
, 1, 0);
2842 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2843 LLVMValueRef src1
, LLVMValueRef src2
,
2849 if (bitsize
== 16) {
2850 intr
= "llvm.amdgcn.fmed3.f16";
2852 } else if (bitsize
== 32) {
2853 intr
= "llvm.amdgcn.fmed3.f32";
2856 intr
= "llvm.amdgcn.fmed3.f64";
2860 LLVMValueRef params
[] = {
2865 return ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2866 AC_FUNC_ATTR_READNONE
);
2869 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2875 if (bitsize
== 16) {
2876 intr
= "llvm.amdgcn.fract.f16";
2878 } else if (bitsize
== 32) {
2879 intr
= "llvm.amdgcn.fract.f32";
2882 intr
= "llvm.amdgcn.fract.f64";
2886 LLVMValueRef params
[] = {
2889 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2890 AC_FUNC_ATTR_READNONE
);
2893 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2896 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2897 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2898 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2900 LLVMValueRef cmp
, val
;
2901 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2902 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2903 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2904 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2908 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2911 LLVMValueRef cmp
, val
, zero
, one
;
2914 if (bitsize
== 16) {
2918 } else if (bitsize
== 32) {
2928 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2929 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2930 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2931 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2935 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2937 LLVMValueRef result
;
2940 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2944 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2945 (LLVMValueRef
[]) { src0
}, 1,
2946 AC_FUNC_ATTR_READNONE
);
2948 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2951 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2952 (LLVMValueRef
[]) { src0
}, 1,
2953 AC_FUNC_ATTR_READNONE
);
2956 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2957 (LLVMValueRef
[]) { src0
}, 1,
2958 AC_FUNC_ATTR_READNONE
);
2960 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2963 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2964 (LLVMValueRef
[]) { src0
}, 1,
2965 AC_FUNC_ATTR_READNONE
);
2967 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2970 unreachable(!"invalid bitsize");
2977 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2980 LLVMValueRef result
;
2983 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2987 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2988 (LLVMValueRef
[]) { src0
}, 1,
2989 AC_FUNC_ATTR_READNONE
);
2991 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2994 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2995 (LLVMValueRef
[]) { src0
}, 1,
2996 AC_FUNC_ATTR_READNONE
);
2999 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
3000 (LLVMValueRef
[]) { src0
}, 1,
3001 AC_FUNC_ATTR_READNONE
);
3003 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
3006 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
3007 (LLVMValueRef
[]) { src0
}, 1,
3008 AC_FUNC_ATTR_READNONE
);
3010 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
3013 unreachable(!"invalid bitsize");
3020 #define AC_EXP_TARGET 0
3021 #define AC_EXP_ENABLED_CHANNELS 1
3022 #define AC_EXP_OUT0 2
3030 struct ac_vs_exp_chan
3034 enum ac_ir_type type
;
3037 struct ac_vs_exp_inst
{
3040 struct ac_vs_exp_chan chan
[4];
3043 struct ac_vs_exports
{
3045 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
3048 /* Return true if the PARAM export has been eliminated. */
3049 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
3050 uint32_t num_outputs
,
3051 struct ac_vs_exp_inst
*exp
)
3053 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
3054 bool is_zero
[4] = {}, is_one
[4] = {};
3056 for (i
= 0; i
< 4; i
++) {
3057 /* It's a constant expression. Undef outputs are eliminated too. */
3058 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
3061 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
3062 if (exp
->chan
[i
].const_float
== 0)
3064 else if (exp
->chan
[i
].const_float
== 1)
3067 return false; /* other constant */
3072 /* Only certain combinations of 0 and 1 can be eliminated. */
3073 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
3074 default_val
= is_zero
[3] ? 0 : 1;
3075 else if (is_one
[0] && is_one
[1] && is_one
[2])
3076 default_val
= is_zero
[3] ? 2 : 3;
3080 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
3081 LLVMInstructionEraseFromParent(exp
->inst
);
3083 /* Change OFFSET to DEFAULT_VAL. */
3084 for (i
= 0; i
< num_outputs
; i
++) {
3085 if (vs_output_param_offset
[i
] == exp
->offset
) {
3086 vs_output_param_offset
[i
] =
3087 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
3094 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
3095 uint8_t *vs_output_param_offset
,
3096 uint32_t num_outputs
,
3097 struct ac_vs_exports
*processed
,
3098 struct ac_vs_exp_inst
*exp
)
3100 unsigned p
, copy_back_channels
= 0;
3102 /* See if the output is already in the list of processed outputs.
3103 * The LLVMValueRef comparison relies on SSA.
3105 for (p
= 0; p
< processed
->num
; p
++) {
3106 bool different
= false;
3108 for (unsigned j
= 0; j
< 4; j
++) {
3109 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
3110 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
3112 /* Treat undef as a match. */
3113 if (c2
->type
== AC_IR_UNDEF
)
3116 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3117 * and consider the instruction duplicated.
3119 if (c1
->type
== AC_IR_UNDEF
) {
3120 copy_back_channels
|= 1 << j
;
3124 /* Test whether the channels are not equal. */
3125 if (c1
->type
!= c2
->type
||
3126 (c1
->type
== AC_IR_CONST
&&
3127 c1
->const_float
!= c2
->const_float
) ||
3128 (c1
->type
== AC_IR_VALUE
&&
3129 c1
->value
!= c2
->value
)) {
3137 copy_back_channels
= 0;
3139 if (p
== processed
->num
)
3142 /* If a match was found, but the matching export has undef where the new
3143 * one has a normal value, copy the normal value to the undef channel.
3145 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3147 /* Get current enabled channels mask. */
3148 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3149 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3151 while (copy_back_channels
) {
3152 unsigned chan
= u_bit_scan(©_back_channels
);
3154 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3155 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3156 exp
->chan
[chan
].value
);
3157 match
->chan
[chan
] = exp
->chan
[chan
];
3159 /* Update number of enabled channels because the original mask
3160 * is not always 0xf.
3162 enabled_channels
|= (1 << chan
);
3163 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3164 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3167 /* The PARAM export is duplicated. Kill it. */
3168 LLVMInstructionEraseFromParent(exp
->inst
);
3170 /* Change OFFSET to the matching export. */
3171 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3172 if (vs_output_param_offset
[i
] == exp
->offset
) {
3173 vs_output_param_offset
[i
] = match
->offset
;
3180 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3181 LLVMValueRef main_fn
,
3182 uint8_t *vs_output_param_offset
,
3183 uint32_t num_outputs
,
3184 uint8_t *num_param_exports
)
3186 LLVMBasicBlockRef bb
;
3187 bool removed_any
= false;
3188 struct ac_vs_exports exports
;
3192 /* Process all LLVM instructions. */
3193 bb
= LLVMGetFirstBasicBlock(main_fn
);
3195 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3198 LLVMValueRef cur
= inst
;
3199 inst
= LLVMGetNextInstruction(inst
);
3200 struct ac_vs_exp_inst exp
;
3202 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3205 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3207 if (!ac_llvm_is_function(callee
))
3210 const char *name
= LLVMGetValueName(callee
);
3211 unsigned num_args
= LLVMCountParams(callee
);
3213 /* Check if this is an export instruction. */
3214 if ((num_args
!= 9 && num_args
!= 8) ||
3215 (strcmp(name
, "llvm.SI.export") &&
3216 strcmp(name
, "llvm.amdgcn.exp.f32")))
3219 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3220 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3222 if (target
< V_008DFC_SQ_EXP_PARAM
)
3225 target
-= V_008DFC_SQ_EXP_PARAM
;
3227 /* Parse the instruction. */
3228 memset(&exp
, 0, sizeof(exp
));
3229 exp
.offset
= target
;
3232 for (unsigned i
= 0; i
< 4; i
++) {
3233 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3235 exp
.chan
[i
].value
= v
;
3237 if (LLVMIsUndef(v
)) {
3238 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3239 } else if (LLVMIsAConstantFP(v
)) {
3240 LLVMBool loses_info
;
3241 exp
.chan
[i
].type
= AC_IR_CONST
;
3242 exp
.chan
[i
].const_float
=
3243 LLVMConstRealGetDouble(v
, &loses_info
);
3245 exp
.chan
[i
].type
= AC_IR_VALUE
;
3249 /* Eliminate constant and duplicated PARAM exports. */
3250 if (ac_eliminate_const_output(vs_output_param_offset
,
3251 num_outputs
, &exp
) ||
3252 ac_eliminate_duplicated_output(ctx
,
3253 vs_output_param_offset
,
3254 num_outputs
, &exports
,
3258 exports
.exp
[exports
.num
++] = exp
;
3261 bb
= LLVMGetNextBasicBlock(bb
);
3264 /* Remove holes in export memory due to removed PARAM exports.
3265 * This is done by renumbering all PARAM exports.
3268 uint8_t old_offset
[VARYING_SLOT_MAX
];
3271 /* Make a copy of the offsets. We need the old version while
3272 * we are modifying some of them. */
3273 memcpy(old_offset
, vs_output_param_offset
,
3274 sizeof(old_offset
));
3276 for (i
= 0; i
< exports
.num
; i
++) {
3277 unsigned offset
= exports
.exp
[i
].offset
;
3279 /* Update vs_output_param_offset. Multiple outputs can
3280 * have the same offset.
3282 for (out
= 0; out
< num_outputs
; out
++) {
3283 if (old_offset
[out
] == offset
)
3284 vs_output_param_offset
[out
] = i
;
3287 /* Change the PARAM offset in the instruction. */
3288 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3289 LLVMConstInt(ctx
->i32
,
3290 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3292 *num_param_exports
= exports
.num
;
3296 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3298 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3299 ac_build_intrinsic(ctx
,
3300 "llvm.amdgcn.init.exec", ctx
->voidt
,
3301 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3304 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3306 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3307 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3308 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3312 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3313 LLVMValueRef dw_addr
)
3315 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3318 void ac_lds_store(struct ac_llvm_context
*ctx
,
3319 LLVMValueRef dw_addr
,
3322 value
= ac_to_integer(ctx
, value
);
3323 ac_build_indexed_store(ctx
, ctx
->lds
,
3327 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3328 LLVMTypeRef dst_type
,
3331 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3332 const char *intrin_name
;
3336 switch (src0_bitsize
) {
3338 intrin_name
= "llvm.cttz.i64";
3343 intrin_name
= "llvm.cttz.i32";
3348 intrin_name
= "llvm.cttz.i16";
3353 intrin_name
= "llvm.cttz.i8";
3358 unreachable(!"invalid bitsize");
3361 LLVMValueRef params
[2] = {
3364 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3365 * add special code to check for x=0. The reason is that
3366 * the LLVM behavior for x=0 is different from what we
3367 * need here. However, LLVM also assumes that ffs(x) is
3368 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3369 * a conditional assignment to handle 0 is still required.
3371 * The hardware already implements the correct behavior.
3376 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3378 AC_FUNC_ATTR_READNONE
);
3380 if (src0_bitsize
== 64) {
3381 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3382 } else if (src0_bitsize
< 32) {
3383 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3386 /* TODO: We need an intrinsic to skip this conditional. */
3387 /* Check for zero: */
3388 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3391 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3394 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3396 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3399 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3401 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3404 static struct ac_llvm_flow
*
3405 get_current_flow(struct ac_llvm_context
*ctx
)
3407 if (ctx
->flow_depth
> 0)
3408 return &ctx
->flow
[ctx
->flow_depth
- 1];
3412 static struct ac_llvm_flow
*
3413 get_innermost_loop(struct ac_llvm_context
*ctx
)
3415 for (unsigned i
= ctx
->flow_depth
; i
> 0; --i
) {
3416 if (ctx
->flow
[i
- 1].loop_entry_block
)
3417 return &ctx
->flow
[i
- 1];
3422 static struct ac_llvm_flow
*
3423 push_flow(struct ac_llvm_context
*ctx
)
3425 struct ac_llvm_flow
*flow
;
3427 if (ctx
->flow_depth
>= ctx
->flow_depth_max
) {
3428 unsigned new_max
= MAX2(ctx
->flow_depth
<< 1,
3429 AC_LLVM_INITIAL_CF_DEPTH
);
3431 ctx
->flow
= realloc(ctx
->flow
, new_max
* sizeof(*ctx
->flow
));
3432 ctx
->flow_depth_max
= new_max
;
3435 flow
= &ctx
->flow
[ctx
->flow_depth
];
3438 flow
->next_block
= NULL
;
3439 flow
->loop_entry_block
= NULL
;
3443 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3447 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3448 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3451 /* Append a basic block at the level of the parent flow.
3453 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3456 assert(ctx
->flow_depth
>= 1);
3458 if (ctx
->flow_depth
>= 2) {
3459 struct ac_llvm_flow
*flow
= &ctx
->flow
[ctx
->flow_depth
- 2];
3461 return LLVMInsertBasicBlockInContext(ctx
->context
,
3462 flow
->next_block
, name
);
3465 LLVMValueRef main_fn
=
3466 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3467 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3470 /* Emit a branch to the given default target for the current block if
3471 * applicable -- that is, if the current block does not already contain a
3472 * branch from a break or continue.
3474 static void emit_default_branch(LLVMBuilderRef builder
,
3475 LLVMBasicBlockRef target
)
3477 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3478 LLVMBuildBr(builder
, target
);
3481 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3483 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3484 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3485 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3486 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3487 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3488 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3491 void ac_build_break(struct ac_llvm_context
*ctx
)
3493 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3494 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3497 void ac_build_continue(struct ac_llvm_context
*ctx
)
3499 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3500 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3503 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3505 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3506 LLVMBasicBlockRef endif_block
;
3508 assert(!current_branch
->loop_entry_block
);
3510 endif_block
= append_basic_block(ctx
, "ENDIF");
3511 emit_default_branch(ctx
->builder
, endif_block
);
3513 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3514 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3516 current_branch
->next_block
= endif_block
;
3519 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3521 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3523 assert(!current_branch
->loop_entry_block
);
3525 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3526 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3527 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3532 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3534 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3536 assert(current_loop
->loop_entry_block
);
3538 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3540 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3541 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3545 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3547 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3548 LLVMBasicBlockRef if_block
;
3550 if_block
= append_basic_block(ctx
, "IF");
3551 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3552 set_basicblock_name(if_block
, "if", label_id
);
3553 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3554 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3557 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3560 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3561 value
, ctx
->f32_0
, "");
3562 ac_build_ifcc(ctx
, cond
, label_id
);
3565 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3568 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3569 ac_to_integer(ctx
, value
),
3571 ac_build_ifcc(ctx
, cond
, label_id
);
3574 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3577 LLVMBuilderRef builder
= ac
->builder
;
3578 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3579 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3580 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3581 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3582 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3586 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3588 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3591 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3592 LLVMDisposeBuilder(first_builder
);
3596 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3597 LLVMTypeRef type
, const char *name
)
3599 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3600 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3604 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3607 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3608 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3609 LLVMPointerType(type
, addr_space
), "");
3612 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3615 unsigned num_components
= ac_get_llvm_num_components(value
);
3616 if (count
== num_components
)
3619 LLVMValueRef masks
[MAX2(count
, 2)];
3620 masks
[0] = ctx
->i32_0
;
3621 masks
[1] = ctx
->i32_1
;
3622 for (unsigned i
= 2; i
< count
; i
++)
3623 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3626 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3629 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3630 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3633 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3634 unsigned rshift
, unsigned bitwidth
)
3636 LLVMValueRef value
= param
;
3638 value
= LLVMBuildLShr(ctx
->builder
, value
,
3639 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3641 if (rshift
+ bitwidth
< 32) {
3642 unsigned mask
= (1 << bitwidth
) - 1;
3643 value
= LLVMBuildAnd(ctx
->builder
, value
,
3644 LLVMConstInt(ctx
->i32
, mask
, false), "");
3649 /* Adjust the sample index according to FMASK.
3651 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3652 * which is the identity mapping. Each nibble says which physical sample
3653 * should be fetched to get that sample.
3655 * For example, 0x11111100 means there are only 2 samples stored and
3656 * the second sample covers 3/4 of the pixel. When reading samples 0
3657 * and 1, return physical sample 0 (determined by the first two 0s
3658 * in FMASK), otherwise return physical sample 1.
3660 * The sample index should be adjusted as follows:
3661 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3663 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3664 LLVMValueRef
*addr
, bool is_array_tex
)
3666 struct ac_image_args fmask_load
= {};
3667 fmask_load
.opcode
= ac_image_load
;
3668 fmask_load
.resource
= fmask
;
3669 fmask_load
.dmask
= 0xf;
3670 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3671 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3673 fmask_load
.coords
[0] = addr
[0];
3674 fmask_load
.coords
[1] = addr
[1];
3676 fmask_load
.coords
[2] = addr
[2];
3678 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3679 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3682 /* Apply the formula. */
3683 unsigned sample_chan
= is_array_tex
? 3 : 2;
3684 LLVMValueRef final_sample
;
3685 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3686 LLVMConstInt(ac
->i32
, 4, 0), "");
3687 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3688 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3689 * with EQAA, so those will map to 0. */
3690 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3691 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3693 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3694 * resource descriptor is 0 (invalid).
3697 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3698 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3699 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3701 /* Replace the MSAA sample index. */
3702 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3703 addr
[sample_chan
], "");
3707 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3709 ac_build_optimization_barrier(ctx
, &src
);
3710 return ac_build_intrinsic(ctx
,
3711 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3712 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3714 lane
== NULL
? 1 : 2,
3715 AC_FUNC_ATTR_READNONE
|
3716 AC_FUNC_ATTR_CONVERGENT
);
3720 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3723 * @param lane - id of the lane or NULL for the first active lane
3724 * @return value of the lane
3727 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3729 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3730 src
= ac_to_integer(ctx
, src
);
3731 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3735 ret
= _ac_build_readlane(ctx
, src
, lane
);
3737 assert(bits
% 32 == 0);
3738 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3739 LLVMValueRef src_vector
=
3740 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3741 ret
= LLVMGetUndef(vec_type
);
3742 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3743 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3744 LLVMConstInt(ctx
->i32
, i
, 0), "");
3745 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
3746 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3747 LLVMConstInt(ctx
->i32
, i
, 0), "");
3750 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3754 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3756 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
3758 LLVMValueRef pred
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, lane
,
3759 ac_get_thread_id(ctx
), "");
3760 return LLVMBuildSelect(ctx
->builder
, pred
, value
, src
, "");
3764 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3766 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3767 LLVMVectorType(ctx
->i32
, 2),
3769 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3771 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3774 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3775 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3776 2, AC_FUNC_ATTR_READNONE
);
3777 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3778 (LLVMValueRef
[]) { mask_hi
, val
},
3779 2, AC_FUNC_ATTR_READNONE
);
3784 _dpp_quad_perm
= 0x000,
3785 _dpp_row_sl
= 0x100,
3786 _dpp_row_sr
= 0x110,
3787 _dpp_row_rr
= 0x120,
3792 dpp_row_mirror
= 0x140,
3793 dpp_row_half_mirror
= 0x141,
3794 dpp_row_bcast15
= 0x142,
3795 dpp_row_bcast31
= 0x143
3798 static inline enum dpp_ctrl
3799 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3801 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3802 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3805 static inline enum dpp_ctrl
3806 dpp_row_sl(unsigned amount
)
3808 assert(amount
> 0 && amount
< 16);
3809 return _dpp_row_sl
| amount
;
3812 static inline enum dpp_ctrl
3813 dpp_row_sr(unsigned amount
)
3815 assert(amount
> 0 && amount
< 16);
3816 return _dpp_row_sr
| amount
;
3820 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3821 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3824 return ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32",
3828 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3829 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3830 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3831 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3832 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3836 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3837 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3840 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3841 src
= ac_to_integer(ctx
, src
);
3842 old
= ac_to_integer(ctx
, old
);
3843 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3846 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3847 bank_mask
, bound_ctrl
);
3849 assert(bits
% 32 == 0);
3850 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3851 LLVMValueRef src_vector
=
3852 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3853 LLVMValueRef old_vector
=
3854 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3855 ret
= LLVMGetUndef(vec_type
);
3856 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3857 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3858 LLVMConstInt(ctx
->i32
, i
,
3860 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3861 LLVMConstInt(ctx
->i32
, i
,
3863 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3868 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3870 LLVMConstInt(ctx
->i32
, i
,
3874 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3878 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3879 bool exchange_rows
, bool bound_ctrl
)
3881 LLVMValueRef args
[6] = {
3884 LLVMConstInt(ctx
->i32
, sel
, false),
3885 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3886 ctx
->i1true
, /* fi */
3887 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3889 return ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3890 : "llvm.amdgcn.permlane16",
3892 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3896 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3897 bool exchange_rows
, bool bound_ctrl
)
3899 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3900 src
= ac_to_integer(ctx
, src
);
3901 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3904 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3907 assert(bits
% 32 == 0);
3908 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3909 LLVMValueRef src_vector
=
3910 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3911 ret
= LLVMGetUndef(vec_type
);
3912 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3913 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3914 LLVMConstInt(ctx
->i32
, i
,
3916 LLVMValueRef ret_comp
=
3917 _ac_build_permlane16(ctx
, src
, sel
,
3920 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3922 LLVMConstInt(ctx
->i32
, i
,
3926 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3929 static inline unsigned
3930 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3932 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3933 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3937 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3939 return ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle",
3940 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3941 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3942 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3946 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3948 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3949 src
= ac_to_integer(ctx
, src
);
3950 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3953 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3955 assert(bits
% 32 == 0);
3956 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3957 LLVMValueRef src_vector
=
3958 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3959 ret
= LLVMGetUndef(vec_type
);
3960 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3961 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3962 LLVMConstInt(ctx
->i32
, i
,
3964 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3966 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3968 LLVMConstInt(ctx
->i32
, i
,
3972 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3976 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3978 char name
[32], type
[8];
3979 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3980 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3981 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3982 (LLVMValueRef
[]) { src
}, 1,
3983 AC_FUNC_ATTR_READNONE
);
3987 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3988 LLVMValueRef inactive
)
3990 char name
[33], type
[8];
3991 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3992 src
= ac_to_integer(ctx
, src
);
3993 inactive
= ac_to_integer(ctx
, inactive
);
3994 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3995 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3997 ac_build_intrinsic(ctx
, name
,
3998 LLVMTypeOf(src
), (LLVMValueRef
[]) {
4000 AC_FUNC_ATTR_READNONE
|
4001 AC_FUNC_ATTR_CONVERGENT
);
4002 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
4006 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
4008 if (type_size
== 4) {
4010 case nir_op_iadd
: return ctx
->i32_0
;
4011 case nir_op_fadd
: return ctx
->f32_0
;
4012 case nir_op_imul
: return ctx
->i32_1
;
4013 case nir_op_fmul
: return ctx
->f32_1
;
4014 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
4015 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
4016 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
4017 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
4018 case nir_op_umax
: return ctx
->i32_0
;
4019 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
4020 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
4021 case nir_op_ior
: return ctx
->i32_0
;
4022 case nir_op_ixor
: return ctx
->i32_0
;
4024 unreachable("bad reduction intrinsic");
4026 } else { /* type_size == 64bit */
4028 case nir_op_iadd
: return ctx
->i64_0
;
4029 case nir_op_fadd
: return ctx
->f64_0
;
4030 case nir_op_imul
: return ctx
->i64_1
;
4031 case nir_op_fmul
: return ctx
->f64_1
;
4032 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
4033 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4034 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
4035 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
4036 case nir_op_umax
: return ctx
->i64_0
;
4037 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
4038 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
4039 case nir_op_ior
: return ctx
->i64_0
;
4040 case nir_op_ixor
: return ctx
->i64_0
;
4042 unreachable("bad reduction intrinsic");
4048 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
4050 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
4052 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
4053 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
4054 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
4055 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
4056 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
4057 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
4059 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
4060 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
4062 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
4063 _64bit
? "llvm.minnum.f64" : "llvm.minnum.f32",
4064 _64bit
? ctx
->f64
: ctx
->f32
,
4065 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4066 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
4067 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
4069 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
4070 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
4072 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
4073 _64bit
? "llvm.maxnum.f64" : "llvm.maxnum.f32",
4074 _64bit
? ctx
->f64
: ctx
->f32
,
4075 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4076 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
4077 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
4078 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
4080 unreachable("bad reduction intrinsic");
4085 * \param maxprefix specifies that the result only needs to be correct for a
4086 * prefix of this many threads
4088 * TODO: add inclusive and excluse scan functions for GFX6.
4091 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
4092 unsigned maxprefix
, bool inclusive
)
4094 LLVMValueRef result
, tmp
;
4096 if (ctx
->chip_class
>= GFX10
) {
4097 result
= inclusive
? src
: identity
;
4102 result
= ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
4106 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4107 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4110 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
4111 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4114 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
4115 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4118 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
4119 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4122 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
4123 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4124 if (maxprefix
<= 16)
4127 if (ctx
->chip_class
>= GFX10
) {
4128 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4129 LLVMBuilderRef builder
= ctx
->builder
;
4130 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4132 /* TODO-GFX10: Can we get better code-gen by putting this into
4133 * a branch so that LLVM generates EXEC mask manipulations? */
4137 tmp
= ac_build_alu_op(ctx
, result
, src
, op
);
4138 tmp
= ac_build_permlane16(ctx
, tmp
, ~(uint64_t)0, true, false);
4139 tmp
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4140 cc
= LLVMBuildAnd(builder
, tid
, LLVMConstInt(ctx
->i32
, 16, false), "");
4141 cc
= LLVMBuildICmp(builder
, LLVMIntNE
, cc
, ctx
->i32_0
, "");
4142 result
= LLVMBuildSelect(builder
, cc
, tmp
, result
, "");
4143 if (maxprefix
<= 32)
4149 tmp
= ac_build_alu_op(ctx
, result
, src
, op
);
4150 tmp
= ac_build_readlane(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 31, false));
4151 tmp
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4152 cc
= LLVMBuildICmp(builder
, LLVMIntUGE
, tid
,
4153 LLVMConstInt(ctx
->i32
, 32, false), "");
4154 result
= LLVMBuildSelect(builder
, cc
, tmp
, result
, "");
4158 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4159 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4160 if (maxprefix
<= 32)
4162 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4163 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4168 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4170 LLVMValueRef result
;
4172 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4173 LLVMBuilderRef builder
= ctx
->builder
;
4174 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4175 result
= ac_build_ballot(ctx
, src
);
4176 result
= ac_build_mbcnt(ctx
, result
);
4177 result
= LLVMBuildAdd(builder
, result
, src
, "");
4181 ac_build_optimization_barrier(ctx
, &src
);
4183 LLVMValueRef identity
=
4184 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4185 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4186 LLVMTypeOf(identity
), "");
4187 result
= ac_build_scan(ctx
, op
, result
, identity
, 64, true);
4189 return ac_build_wwm(ctx
, result
);
4193 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4195 LLVMValueRef result
;
4197 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4198 LLVMBuilderRef builder
= ctx
->builder
;
4199 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4200 result
= ac_build_ballot(ctx
, src
);
4201 result
= ac_build_mbcnt(ctx
, result
);
4205 ac_build_optimization_barrier(ctx
, &src
);
4207 LLVMValueRef identity
=
4208 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4209 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4210 LLVMTypeOf(identity
), "");
4211 result
= ac_build_scan(ctx
, op
, result
, identity
, 64, false);
4213 return ac_build_wwm(ctx
, result
);
4217 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4219 if (cluster_size
== 1) return src
;
4220 ac_build_optimization_barrier(ctx
, &src
);
4221 LLVMValueRef result
, swap
;
4222 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4223 ac_get_type_size(LLVMTypeOf(src
)));
4224 result
= LLVMBuildBitCast(ctx
->builder
,
4225 ac_build_set_inactive(ctx
, src
, identity
),
4226 LLVMTypeOf(identity
), "");
4227 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4228 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4229 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4231 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4232 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4233 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4235 if (ctx
->chip_class
>= GFX8
)
4236 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4238 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4239 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4240 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4242 if (ctx
->chip_class
>= GFX8
)
4243 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4245 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4246 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4247 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4249 if (ctx
->chip_class
>= GFX10
)
4250 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4251 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4252 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4254 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4255 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4256 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4258 if (ctx
->chip_class
>= GFX8
) {
4259 if (ctx
->chip_class
>= GFX10
)
4260 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4262 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4263 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4264 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4265 return ac_build_wwm(ctx
, result
);
4267 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4268 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4269 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4270 return ac_build_wwm(ctx
, result
);
4275 * "Top half" of a scan that reduces per-wave values across an entire
4278 * The source value must be present in the highest lane of the wave, and the
4279 * highest lane must be live.
4282 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4284 if (ws
->maxwaves
<= 1)
4287 const LLVMValueRef i32_63
= LLVMConstInt(ctx
->i32
, 63, false);
4288 LLVMBuilderRef builder
= ctx
->builder
;
4289 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4292 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, i32_63
, "");
4293 ac_build_ifcc(ctx
, tmp
, 1000);
4294 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4295 ac_build_endif(ctx
, 1000);
4299 * "Bottom half" of a scan that reduces per-wave values across an entire
4302 * The caller must place a barrier between the top and bottom halves.
4305 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4307 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4308 const LLVMValueRef identity
=
4309 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4311 if (ws
->maxwaves
<= 1) {
4312 ws
->result_reduce
= ws
->src
;
4313 ws
->result_inclusive
= ws
->src
;
4314 ws
->result_exclusive
= identity
;
4317 assert(ws
->maxwaves
<= 32);
4319 LLVMBuilderRef builder
= ctx
->builder
;
4320 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4321 LLVMBasicBlockRef bbs
[2];
4322 LLVMValueRef phivalues_scan
[2];
4323 LLVMValueRef tmp
, tmp2
;
4325 bbs
[0] = LLVMGetInsertBlock(builder
);
4326 phivalues_scan
[0] = LLVMGetUndef(type
);
4328 if (ws
->enable_reduce
)
4329 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4330 else if (ws
->enable_inclusive
)
4331 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4333 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4334 ac_build_ifcc(ctx
, tmp
, 1001);
4336 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4338 ac_build_optimization_barrier(ctx
, &tmp
);
4340 bbs
[1] = LLVMGetInsertBlock(builder
);
4341 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4343 ac_build_endif(ctx
, 1001);
4345 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4347 if (ws
->enable_reduce
) {
4348 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4349 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4351 if (ws
->enable_inclusive
)
4352 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4353 if (ws
->enable_exclusive
) {
4354 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4355 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4356 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4357 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4362 * Inclusive scan of a per-wave value across an entire workgroup.
4364 * This implies an s_barrier instruction.
4366 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4367 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4368 * useful manner because of the barrier in the algorithm.)
4371 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4373 ac_build_wg_wavescan_top(ctx
, ws
);
4374 ac_build_s_barrier(ctx
);
4375 ac_build_wg_wavescan_bottom(ctx
, ws
);
4379 * "Top half" of a scan that reduces per-thread values across an entire
4382 * All lanes must be active when this code runs.
4385 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4387 if (ws
->enable_exclusive
) {
4388 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4389 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4390 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4391 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4393 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4396 bool enable_inclusive
= ws
->enable_inclusive
;
4397 bool enable_exclusive
= ws
->enable_exclusive
;
4398 ws
->enable_inclusive
= false;
4399 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4400 ac_build_wg_wavescan_top(ctx
, ws
);
4401 ws
->enable_inclusive
= enable_inclusive
;
4402 ws
->enable_exclusive
= enable_exclusive
;
4406 * "Bottom half" of a scan that reduces per-thread values across an entire
4409 * The caller must place a barrier between the top and bottom halves.
4412 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4414 bool enable_inclusive
= ws
->enable_inclusive
;
4415 bool enable_exclusive
= ws
->enable_exclusive
;
4416 ws
->enable_inclusive
= false;
4417 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4418 ac_build_wg_wavescan_bottom(ctx
, ws
);
4419 ws
->enable_inclusive
= enable_inclusive
;
4420 ws
->enable_exclusive
= enable_exclusive
;
4422 /* ws->result_reduce is already the correct value */
4423 if (ws
->enable_inclusive
)
4424 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4425 if (ws
->enable_exclusive
)
4426 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4430 * A scan that reduces per-thread values across an entire workgroup.
4432 * The caller must ensure that all lanes are active when this code runs
4433 * (WWM is insufficient!), because there is an implied barrier.
4436 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4438 ac_build_wg_scan_top(ctx
, ws
);
4439 ac_build_s_barrier(ctx
);
4440 ac_build_wg_scan_bottom(ctx
, ws
);
4444 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4445 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4447 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4448 if (ctx
->chip_class
>= GFX8
) {
4449 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4451 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4456 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4458 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4459 return ac_build_intrinsic(ctx
,
4460 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4461 (LLVMValueRef
[]) {index
, src
}, 2,
4462 AC_FUNC_ATTR_READNONE
|
4463 AC_FUNC_ATTR_CONVERGENT
);
4467 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4473 if (bitsize
== 16) {
4474 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4476 } else if (bitsize
== 32) {
4477 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4480 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4484 LLVMValueRef params
[] = {
4487 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4488 AC_FUNC_ATTR_READNONE
);
4491 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4497 if (bitsize
== 16) {
4498 intr
= "llvm.amdgcn.frexp.mant.f16";
4500 } else if (bitsize
== 32) {
4501 intr
= "llvm.amdgcn.frexp.mant.f32";
4504 intr
= "llvm.amdgcn.frexp.mant.f64";
4508 LLVMValueRef params
[] = {
4511 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4512 AC_FUNC_ATTR_READNONE
);
4516 * this takes an I,J coordinate pair,
4517 * and works out the X and Y derivatives.
4518 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4521 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4523 LLVMValueRef result
[4], a
;
4526 for (i
= 0; i
< 2; i
++) {
4527 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4528 LLVMConstInt(ctx
->i32
, i
, false), "");
4529 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4530 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4532 return ac_build_gather_values(ctx
, result
, 4);
4536 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4538 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4540 AC_FUNC_ATTR_READNONE
);
4541 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4542 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4545 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4546 LLVMValueRef
*args
, unsigned num_args
)
4548 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4549 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));