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>
29 #include <llvm/Config/llvm-config.h>
31 #include "c11/threads.h"
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
45 #include "shader_enums.h"
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block
;
54 LLVMBasicBlockRef loop_entry_block
;
57 /* Initialize module-independent parts of the context.
59 * The caller is responsible for initializing ctx::module and ctx::builder.
62 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
63 struct ac_llvm_compiler
*compiler
,
64 enum chip_class chip_class
, enum radeon_family family
,
65 enum ac_float_mode float_mode
, unsigned wave_size
,
66 unsigned ballot_mask_bits
)
68 ctx
->context
= LLVMContextCreate();
70 ctx
->chip_class
= chip_class
;
72 ctx
->wave_size
= wave_size
;
73 ctx
->ballot_mask_bits
= ballot_mask_bits
;
74 ctx
->float_mode
= float_mode
;
75 ctx
->module
= ac_create_module(wave_size
== 32 ? compiler
->tm_wave32
78 ctx
->builder
= ac_create_builder(ctx
->context
, float_mode
);
80 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
81 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
82 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
83 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
84 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
85 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
86 ctx
->i128
= LLVMIntTypeInContext(ctx
->context
, 128);
87 ctx
->intptr
= ctx
->i32
;
88 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
89 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
90 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
91 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
92 ctx
->v4i16
= LLVMVectorType(ctx
->i16
, 4);
93 ctx
->v2f16
= LLVMVectorType(ctx
->f16
, 2);
94 ctx
->v4f16
= LLVMVectorType(ctx
->f16
, 4);
95 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
96 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
97 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
98 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
99 ctx
->v3f32
= LLVMVectorType(ctx
->f32
, 3);
100 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
101 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
102 ctx
->iN_wavemask
= LLVMIntTypeInContext(ctx
->context
, ctx
->wave_size
);
103 ctx
->iN_ballotmask
= LLVMIntTypeInContext(ctx
->context
, ballot_mask_bits
);
105 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
106 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
107 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
108 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
109 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
110 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
111 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
112 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
113 ctx
->i128_0
= LLVMConstInt(ctx
->i128
, 0, false);
114 ctx
->i128_1
= LLVMConstInt(ctx
->i128
, 1, false);
115 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
116 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
117 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
118 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
119 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
120 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
122 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
123 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
125 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
128 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
129 "invariant.load", 14);
131 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
132 "amdgpu.uniform", 14);
134 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
135 ctx
->flow
= calloc(1, sizeof(*ctx
->flow
));
139 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
141 free(ctx
->flow
->stack
);
147 ac_get_llvm_num_components(LLVMValueRef value
)
149 LLVMTypeRef type
= LLVMTypeOf(value
);
150 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
151 ? LLVMGetVectorSize(type
)
153 return num_components
;
157 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
161 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
166 return LLVMBuildExtractElement(ac
->builder
, value
,
167 LLVMConstInt(ac
->i32
, index
, false), "");
171 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
173 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
174 type
= LLVMGetElementType(type
);
176 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
177 return LLVMGetIntTypeWidth(type
);
179 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
180 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_LDS
)
184 if (type
== ctx
->f16
)
186 if (type
== ctx
->f32
)
188 if (type
== ctx
->f64
)
191 unreachable("Unhandled type kind in get_elem_bits");
195 ac_get_type_size(LLVMTypeRef type
)
197 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
200 case LLVMIntegerTypeKind
:
201 return LLVMGetIntTypeWidth(type
) / 8;
202 case LLVMHalfTypeKind
:
204 case LLVMFloatTypeKind
:
206 case LLVMDoubleTypeKind
:
208 case LLVMPointerTypeKind
:
209 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
212 case LLVMVectorTypeKind
:
213 return LLVMGetVectorSize(type
) *
214 ac_get_type_size(LLVMGetElementType(type
));
215 case LLVMArrayTypeKind
:
216 return LLVMGetArrayLength(type
) *
217 ac_get_type_size(LLVMGetElementType(type
));
224 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
228 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
230 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
232 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
235 unreachable("Unhandled integer size");
239 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
241 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
242 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
243 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
244 LLVMGetVectorSize(t
));
246 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
247 switch (LLVMGetPointerAddressSpace(t
)) {
248 case AC_ADDR_SPACE_GLOBAL
:
250 case AC_ADDR_SPACE_CONST_32BIT
:
251 case AC_ADDR_SPACE_LDS
:
254 unreachable("unhandled address space");
257 return to_integer_type_scalar(ctx
, t
);
261 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
263 LLVMTypeRef type
= LLVMTypeOf(v
);
264 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
265 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
267 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
271 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
273 LLVMTypeRef type
= LLVMTypeOf(v
);
274 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
276 return ac_to_integer(ctx
, v
);
279 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
283 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
285 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
287 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
290 unreachable("Unhandled float size");
294 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
296 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
297 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
298 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
299 LLVMGetVectorSize(t
));
301 return to_float_type_scalar(ctx
, t
);
305 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
307 LLVMTypeRef type
= LLVMTypeOf(v
);
308 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
313 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
314 LLVMTypeRef return_type
, LLVMValueRef
*params
,
315 unsigned param_count
, unsigned attrib_mask
)
317 LLVMValueRef function
, call
;
318 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
320 function
= LLVMGetNamedFunction(ctx
->module
, name
);
322 LLVMTypeRef param_types
[32], function_type
;
325 assert(param_count
<= 32);
327 for (i
= 0; i
< param_count
; ++i
) {
329 param_types
[i
] = LLVMTypeOf(params
[i
]);
332 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
333 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
335 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
336 LLVMSetLinkage(function
, LLVMExternalLinkage
);
338 if (!set_callsite_attrs
)
339 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
342 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
343 if (set_callsite_attrs
)
344 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
349 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
352 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
354 LLVMTypeRef elem_type
= type
;
356 assert(bufsize
>= 8);
358 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
359 int ret
= snprintf(buf
, bufsize
, "v%u",
360 LLVMGetVectorSize(type
));
362 char *type_name
= LLVMPrintTypeToString(type
);
363 fprintf(stderr
, "Error building type name for: %s\n",
365 LLVMDisposeMessage(type_name
);
368 elem_type
= LLVMGetElementType(type
);
372 switch (LLVMGetTypeKind(elem_type
)) {
374 case LLVMIntegerTypeKind
:
375 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
377 case LLVMHalfTypeKind
:
378 snprintf(buf
, bufsize
, "f16");
380 case LLVMFloatTypeKind
:
381 snprintf(buf
, bufsize
, "f32");
383 case LLVMDoubleTypeKind
:
384 snprintf(buf
, bufsize
, "f64");
390 * Helper function that builds an LLVM IR PHI node and immediately adds
394 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
395 unsigned count_incoming
, LLVMValueRef
*values
,
396 LLVMBasicBlockRef
*blocks
)
398 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
399 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
403 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
405 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
406 0, AC_FUNC_ATTR_CONVERGENT
);
409 /* Prevent optimizations (at least of memory accesses) across the current
410 * point in the program by emitting empty inline assembly that is marked as
411 * having side effects.
413 * Optionally, a value can be passed through the inline assembly to prevent
414 * LLVM from hoisting calls to ReadNone functions.
417 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
420 static int counter
= 0;
422 LLVMBuilderRef builder
= ctx
->builder
;
425 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
428 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
429 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
430 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
432 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
433 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
434 LLVMTypeRef type
= LLVMTypeOf(*pvgpr
);
435 unsigned bitsize
= ac_get_elem_bits(ctx
, type
);
436 LLVMValueRef vgpr
= *pvgpr
;
437 LLVMTypeRef vgpr_type
;
442 vgpr
= LLVMBuildZExt(ctx
->builder
, vgpr
, ctx
->i32
, "");
444 vgpr_type
= LLVMTypeOf(vgpr
);
445 vgpr_size
= ac_get_type_size(vgpr_type
);
447 assert(vgpr_size
% 4 == 0);
449 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
450 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
451 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
452 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
453 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
456 vgpr
= LLVMBuildTrunc(builder
, vgpr
, type
, "");
463 ac_build_shader_clock(struct ac_llvm_context
*ctx
, nir_scope scope
)
465 const char *name
= scope
== NIR_SCOPE_DEVICE
? "llvm.amdgcn.s.memrealtime" : "llvm.amdgcn.s.memtime";
466 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, name
, ctx
->i64
, NULL
, 0, 0);
467 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
471 ac_build_ballot(struct ac_llvm_context
*ctx
,
476 if (LLVM_VERSION_MAJOR
>= 9) {
477 if (ctx
->wave_size
== 64)
478 name
= "llvm.amdgcn.icmp.i64.i32";
480 name
= "llvm.amdgcn.icmp.i32.i32";
482 name
= "llvm.amdgcn.icmp.i32";
484 LLVMValueRef args
[3] = {
487 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
490 /* We currently have no other way to prevent LLVM from lifting the icmp
491 * calls to a dominating basic block.
493 ac_build_optimization_barrier(ctx
, &args
[0]);
495 args
[0] = ac_to_integer(ctx
, args
[0]);
497 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
498 AC_FUNC_ATTR_NOUNWIND
|
499 AC_FUNC_ATTR_READNONE
|
500 AC_FUNC_ATTR_CONVERGENT
);
503 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
508 if (LLVM_VERSION_MAJOR
>= 9) {
509 if (ctx
->wave_size
== 64)
510 name
= "llvm.amdgcn.icmp.i64.i1";
512 name
= "llvm.amdgcn.icmp.i32.i1";
514 name
= "llvm.amdgcn.icmp.i1";
516 LLVMValueRef args
[3] = {
519 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0),
522 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
523 AC_FUNC_ATTR_NOUNWIND
|
524 AC_FUNC_ATTR_READNONE
|
525 AC_FUNC_ATTR_CONVERGENT
);
529 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
531 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
532 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
533 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
537 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
539 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
540 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
541 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
545 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
547 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
548 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
550 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
551 vote_set
, active_set
, "");
552 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
554 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
555 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
559 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
560 unsigned value_count
, unsigned component
)
562 LLVMValueRef vec
= NULL
;
564 if (value_count
== 1) {
565 return values
[component
];
566 } else if (!value_count
)
567 unreachable("value_count is 0");
569 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
570 LLVMValueRef value
= values
[i
];
573 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
574 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
575 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
581 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
582 LLVMValueRef
*values
,
583 unsigned value_count
,
584 unsigned value_stride
,
588 LLVMBuilderRef builder
= ctx
->builder
;
589 LLVMValueRef vec
= NULL
;
592 if (value_count
== 1 && !always_vector
) {
594 return LLVMBuildLoad(builder
, values
[0], "");
596 } else if (!value_count
)
597 unreachable("value_count is 0");
599 for (i
= 0; i
< value_count
; i
++) {
600 LLVMValueRef value
= values
[i
* value_stride
];
602 value
= LLVMBuildLoad(builder
, value
, "");
605 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
606 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
607 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
613 ac_build_gather_values(struct ac_llvm_context
*ctx
,
614 LLVMValueRef
*values
,
615 unsigned value_count
)
617 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
620 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
621 * channels with undef. Extract at most src_channels components from the input.
624 ac_build_expand(struct ac_llvm_context
*ctx
,
626 unsigned src_channels
,
627 unsigned dst_channels
)
629 LLVMTypeRef elemtype
;
630 LLVMValueRef chan
[dst_channels
];
632 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
633 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
635 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
638 src_channels
= MIN2(src_channels
, vec_size
);
640 for (unsigned i
= 0; i
< src_channels
; i
++)
641 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
643 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
646 assert(src_channels
== 1);
649 elemtype
= LLVMTypeOf(value
);
652 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
653 chan
[i
] = LLVMGetUndef(elemtype
);
655 return ac_build_gather_values(ctx
, chan
, dst_channels
);
658 /* Extract components [start, start + channels) from a vector.
661 ac_extract_components(struct ac_llvm_context
*ctx
,
666 LLVMValueRef chan
[channels
];
668 for (unsigned i
= 0; i
< channels
; i
++)
669 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
+ start
);
671 return ac_build_gather_values(ctx
, chan
, channels
);
674 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
675 * with undef. Extract at most num_channels components from the input.
677 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
679 unsigned num_channels
)
681 return ac_build_expand(ctx
, value
, num_channels
, 4);
684 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
686 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
690 name
= "llvm.rint.f16";
691 else if (type_size
== 4)
692 name
= "llvm.rint.f32";
694 name
= "llvm.rint.f64";
696 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
697 AC_FUNC_ATTR_READNONE
);
701 ac_build_fdiv(struct ac_llvm_context
*ctx
,
705 unsigned type_size
= ac_get_type_size(LLVMTypeOf(den
));
708 /* For doubles, we need precise division to pass GLCTS. */
709 if (ctx
->float_mode
== AC_FLOAT_MODE_DEFAULT_OPENGL
&&
711 return LLVMBuildFDiv(ctx
->builder
, num
, den
, "");
714 name
= "llvm.amdgcn.rcp.f16";
715 else if (type_size
== 4)
716 name
= "llvm.amdgcn.rcp.f32";
718 name
= "llvm.amdgcn.rcp.f64";
720 LLVMValueRef rcp
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(den
),
721 &den
, 1, AC_FUNC_ATTR_READNONE
);
723 return LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
726 /* See fast_idiv_by_const.h. */
727 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
728 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
730 LLVMValueRef multiplier
,
731 LLVMValueRef pre_shift
,
732 LLVMValueRef post_shift
,
733 LLVMValueRef increment
)
735 LLVMBuilderRef builder
= ctx
->builder
;
737 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
738 num
= LLVMBuildMul(builder
,
739 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
740 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
741 num
= LLVMBuildAdd(builder
, num
,
742 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
743 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
744 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
745 return LLVMBuildLShr(builder
, num
, post_shift
, "");
748 /* See fast_idiv_by_const.h. */
749 /* If num != UINT_MAX, this more efficient version can be used. */
750 /* Set: increment = util_fast_udiv_info::increment; */
751 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
753 LLVMValueRef multiplier
,
754 LLVMValueRef pre_shift
,
755 LLVMValueRef post_shift
,
756 LLVMValueRef increment
)
758 LLVMBuilderRef builder
= ctx
->builder
;
760 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
761 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
762 num
= LLVMBuildMul(builder
,
763 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
764 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
765 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
766 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
767 return LLVMBuildLShr(builder
, num
, post_shift
, "");
770 /* See fast_idiv_by_const.h. */
771 /* Both operands must fit in 31 bits and the divisor must not be 1. */
772 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
774 LLVMValueRef multiplier
,
775 LLVMValueRef post_shift
)
777 LLVMBuilderRef builder
= ctx
->builder
;
779 num
= LLVMBuildMul(builder
,
780 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
781 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
782 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
783 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
784 return LLVMBuildLShr(builder
, num
, post_shift
, "");
787 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
788 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
789 * already multiplied by two. id is the cube face number.
791 struct cube_selection_coords
{
798 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
800 struct cube_selection_coords
*out
)
802 LLVMTypeRef f32
= ctx
->f32
;
804 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
805 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
806 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
807 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
808 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
809 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
810 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
811 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
815 * Build a manual selection sequence for cube face sc/tc coordinates and
816 * major axis vector (multiplied by 2 for consistency) for the given
817 * vec3 \p coords, for the face implied by \p selcoords.
819 * For the major axis, we always adjust the sign to be in the direction of
820 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
821 * the selcoords major axis.
823 static void build_cube_select(struct ac_llvm_context
*ctx
,
824 const struct cube_selection_coords
*selcoords
,
825 const LLVMValueRef
*coords
,
826 LLVMValueRef
*out_st
,
827 LLVMValueRef
*out_ma
)
829 LLVMBuilderRef builder
= ctx
->builder
;
830 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
831 LLVMValueRef is_ma_positive
;
833 LLVMValueRef is_ma_z
, is_not_ma_z
;
834 LLVMValueRef is_ma_y
;
835 LLVMValueRef is_ma_x
;
839 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
840 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
841 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
842 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
844 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
845 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
846 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
847 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
848 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
851 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
852 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
853 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
854 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
855 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
858 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
859 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
860 LLVMConstReal(f32
, -1.0), "");
861 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
864 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
865 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
866 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
867 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
868 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
872 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
873 bool is_deriv
, bool is_array
, bool is_lod
,
874 LLVMValueRef
*coords_arg
,
875 LLVMValueRef
*derivs_arg
)
878 LLVMBuilderRef builder
= ctx
->builder
;
879 struct cube_selection_coords selcoords
;
880 LLVMValueRef coords
[3];
883 if (is_array
&& !is_lod
) {
884 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
886 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
888 * "For Array forms, the array layer used will be
890 * max(0, min(d−1, floor(layer+0.5)))
892 * where d is the depth of the texture array and layer
893 * comes from the component indicated in the tables below.
894 * Workaroudn for an issue where the layer is taken from a
895 * helper invocation which happens to fall on a different
896 * layer due to extrapolation."
898 * GFX8 and earlier attempt to implement this in hardware by
899 * clamping the value of coords[2] = (8 * layer) + face.
900 * Unfortunately, this means that the we end up with the wrong
901 * face when clamping occurs.
903 * Clamp the layer earlier to work around the issue.
905 if (ctx
->chip_class
<= GFX8
) {
907 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
908 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
914 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
916 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
917 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
918 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
920 for (int i
= 0; i
< 2; ++i
)
921 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
923 coords
[2] = selcoords
.id
;
925 if (is_deriv
&& derivs_arg
) {
926 LLVMValueRef derivs
[4];
929 /* Convert cube derivatives to 2D derivatives. */
930 for (axis
= 0; axis
< 2; axis
++) {
931 LLVMValueRef deriv_st
[2];
932 LLVMValueRef deriv_ma
;
934 /* Transform the derivative alongside the texture
935 * coordinate. Mathematically, the correct formula is
936 * as follows. Assume we're projecting onto the +Z face
937 * and denote by dx/dh the derivative of the (original)
938 * X texture coordinate with respect to horizontal
939 * window coordinates. The projection onto the +Z face
944 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
945 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
947 * This motivatives the implementation below.
949 * Whether this actually gives the expected results for
950 * apps that might feed in derivatives obtained via
951 * finite differences is anyone's guess. The OpenGL spec
952 * seems awfully quiet about how textureGrad for cube
953 * maps should be handled.
955 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
956 deriv_st
, &deriv_ma
);
958 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
960 for (int i
= 0; i
< 2; ++i
)
961 derivs
[axis
* 2 + i
] =
962 LLVMBuildFSub(builder
,
963 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
964 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
967 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
970 /* Shift the texture coordinate. This must be applied after the
971 * derivative calculation.
973 for (int i
= 0; i
< 2; ++i
)
974 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
977 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
978 /* coords_arg.w component - array_index for cube arrays */
979 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
982 memcpy(coords_arg
, coords
, sizeof(coords
));
987 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
988 LLVMValueRef llvm_chan
,
989 LLVMValueRef attr_number
,
994 LLVMValueRef args
[5];
999 args
[2] = attr_number
;
1002 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
1003 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1007 args
[2] = llvm_chan
;
1008 args
[3] = attr_number
;
1011 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
1012 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1016 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
1017 LLVMValueRef llvm_chan
,
1018 LLVMValueRef attr_number
,
1019 LLVMValueRef params
,
1023 LLVMValueRef args
[6];
1027 args
[1] = llvm_chan
;
1028 args
[2] = attr_number
;
1029 args
[3] = ctx
->i1false
;
1032 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
1033 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1037 args
[2] = llvm_chan
;
1038 args
[3] = attr_number
;
1039 args
[4] = ctx
->i1false
;
1042 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
1043 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
1047 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
1048 LLVMValueRef parameter
,
1049 LLVMValueRef llvm_chan
,
1050 LLVMValueRef attr_number
,
1051 LLVMValueRef params
)
1053 LLVMValueRef args
[4];
1055 args
[0] = parameter
;
1056 args
[1] = llvm_chan
;
1057 args
[2] = attr_number
;
1060 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
1061 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1065 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
1066 LLVMValueRef base_ptr
,
1069 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1073 ac_build_gep0(struct ac_llvm_context
*ctx
,
1074 LLVMValueRef base_ptr
,
1077 LLVMValueRef indices
[2] = {
1081 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1084 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1087 return LLVMBuildPointerCast(ctx
->builder
,
1088 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1089 LLVMTypeOf(ptr
), "");
1093 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1094 LLVMValueRef base_ptr
, LLVMValueRef index
,
1097 LLVMBuildStore(ctx
->builder
, value
,
1098 ac_build_gep0(ctx
, base_ptr
, index
));
1102 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1103 * It's equivalent to doing a load from &base_ptr[index].
1105 * \param base_ptr Where the array starts.
1106 * \param index The element index into the array.
1107 * \param uniform Whether the base_ptr and index can be assumed to be
1108 * dynamically uniform (i.e. load to an SGPR)
1109 * \param invariant Whether the load is invariant (no other opcodes affect it)
1110 * \param no_unsigned_wraparound
1111 * For all possible re-associations and re-distributions of an expression
1112 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1113 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1114 * does not result in an unsigned integer wraparound. This is used for
1115 * optimal code generation of 32-bit pointer arithmetic.
1117 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1118 * integer wraparound can't be an imm offset in s_load_dword, because
1119 * the instruction performs "addr + offset" in 64 bits.
1121 * Expected usage for bindless textures by chaining GEPs:
1122 * // possible unsigned wraparound, don't use InBounds:
1123 * ptr1 = LLVMBuildGEP(base_ptr, index);
1124 * image = load(ptr1); // becomes "s_load ptr1, 0"
1126 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1127 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1130 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1131 LLVMValueRef index
, bool uniform
, bool invariant
,
1132 bool no_unsigned_wraparound
)
1134 LLVMValueRef pointer
, result
;
1136 if (no_unsigned_wraparound
&&
1137 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1138 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1140 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1143 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1144 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1146 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1150 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1153 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1156 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1157 LLVMValueRef base_ptr
, LLVMValueRef index
)
1159 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1162 /* This assumes that there is no unsigned integer wraparound during the address
1163 * computation, excluding all GEPs within base_ptr. */
1164 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1165 LLVMValueRef base_ptr
, LLVMValueRef index
)
1167 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1170 /* See ac_build_load_custom() documentation. */
1171 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1172 LLVMValueRef base_ptr
, LLVMValueRef index
)
1174 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1177 static unsigned get_load_cache_policy(struct ac_llvm_context
*ctx
,
1178 unsigned cache_policy
)
1180 return cache_policy
|
1181 (ctx
->chip_class
>= GFX10
&& cache_policy
& ac_glc
? ac_dlc
: 0);
1185 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1188 LLVMValueRef vindex
,
1189 LLVMValueRef voffset
,
1190 LLVMValueRef soffset
,
1191 unsigned cache_policy
,
1195 LLVMValueRef args
[6];
1198 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1200 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1201 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1202 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1203 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1204 const char *indexing_kind
= structurized
? "struct" : "raw";
1205 char name
[256], type_name
[8];
1207 ac_build_type_name_for_intr(LLVMTypeOf(data
), type_name
, sizeof(type_name
));
1210 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1211 indexing_kind
, type_name
);
1213 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1214 indexing_kind
, type_name
);
1217 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1218 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1222 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1225 LLVMValueRef vindex
,
1226 LLVMValueRef voffset
,
1227 unsigned cache_policy
)
1229 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
, voffset
, NULL
,
1230 cache_policy
, true, true);
1233 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1234 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1235 * or v4i32 (num_channels=3,4).
1238 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1241 unsigned num_channels
,
1242 LLVMValueRef voffset
,
1243 LLVMValueRef soffset
,
1244 unsigned inst_offset
,
1245 unsigned cache_policy
)
1247 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1249 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1250 LLVMValueRef v
[3], v01
;
1252 for (int i
= 0; i
< 3; i
++) {
1253 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1254 LLVMConstInt(ctx
->i32
, i
, 0), "");
1256 v01
= ac_build_gather_values(ctx
, v
, 2);
1258 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1259 soffset
, inst_offset
, cache_policy
);
1260 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1261 soffset
, inst_offset
+ 8,
1266 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1267 * (voffset is swizzled, but soffset isn't swizzled).
1268 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1270 if (!(cache_policy
& ac_swizzled
)) {
1271 LLVMValueRef offset
= soffset
;
1274 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1275 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1277 ac_build_buffer_store_common(ctx
, rsrc
, ac_to_float(ctx
, vdata
),
1278 ctx
->i32_0
, voffset
, offset
,
1279 cache_policy
, false, false);
1283 static const unsigned dfmts
[] = {
1284 V_008F0C_BUF_DATA_FORMAT_32
,
1285 V_008F0C_BUF_DATA_FORMAT_32_32
,
1286 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1287 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1289 unsigned dfmt
= dfmts
[num_channels
- 1];
1290 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1291 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1293 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1294 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
);
1298 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1300 LLVMValueRef vindex
,
1301 LLVMValueRef voffset
,
1302 LLVMValueRef soffset
,
1303 unsigned num_channels
,
1304 LLVMTypeRef channel_type
,
1305 unsigned cache_policy
,
1310 LLVMValueRef args
[5];
1312 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1314 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1315 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1316 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1317 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1318 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1319 const char *indexing_kind
= structurized
? "struct" : "raw";
1320 char name
[256], type_name
[8];
1322 /* D16 is only supported on gfx8+ */
1323 assert(!use_format
||
1324 (channel_type
!= ctx
->f16
&& channel_type
!= ctx
->i16
) ||
1325 ctx
->chip_class
>= GFX8
);
1327 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1328 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1331 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1332 indexing_kind
, type_name
);
1334 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1335 indexing_kind
, type_name
);
1338 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1339 ac_get_load_intr_attribs(can_speculate
));
1343 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1346 LLVMValueRef vindex
,
1347 LLVMValueRef voffset
,
1348 LLVMValueRef soffset
,
1349 unsigned inst_offset
,
1350 unsigned cache_policy
,
1354 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1356 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1358 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1360 if (allow_smem
&& !(cache_policy
& ac_slc
) &&
1361 (!(cache_policy
& ac_glc
) || ctx
->chip_class
>= GFX8
)) {
1362 assert(vindex
== NULL
);
1364 LLVMValueRef result
[8];
1366 for (int i
= 0; i
< num_channels
; i
++) {
1368 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1369 LLVMConstInt(ctx
->i32
, 4, 0), "");
1371 LLVMValueRef args
[3] = {
1374 LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0),
1376 result
[i
] = ac_build_intrinsic(ctx
,
1377 "llvm.amdgcn.s.buffer.load.f32",
1379 AC_FUNC_ATTR_READNONE
);
1381 if (num_channels
== 1)
1384 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1385 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1386 return ac_build_gather_values(ctx
, result
, num_channels
);
1389 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
,
1391 num_channels
, ctx
->f32
,
1393 can_speculate
, false, false);
1396 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1398 LLVMValueRef vindex
,
1399 LLVMValueRef voffset
,
1400 unsigned num_channels
,
1401 unsigned cache_policy
,
1405 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1406 ctx
->i32_0
, num_channels
,
1407 d16
? ctx
->f16
: ctx
->f32
,
1408 cache_policy
, can_speculate
,
1413 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1415 LLVMValueRef vindex
,
1416 LLVMValueRef voffset
,
1417 LLVMValueRef soffset
,
1418 LLVMValueRef immoffset
,
1419 unsigned num_channels
,
1422 unsigned cache_policy
,
1426 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1428 LLVMValueRef args
[6];
1430 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1432 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1433 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1434 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1435 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1436 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1437 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1438 const char *indexing_kind
= structurized
? "struct" : "raw";
1439 char name
[256], type_name
[8];
1441 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1442 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1444 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1445 indexing_kind
, type_name
);
1447 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1448 ac_get_load_intr_attribs(can_speculate
));
1452 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1454 LLVMValueRef vindex
,
1455 LLVMValueRef voffset
,
1456 LLVMValueRef soffset
,
1457 LLVMValueRef immoffset
,
1458 unsigned num_channels
,
1461 unsigned cache_policy
,
1464 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1465 immoffset
, num_channels
, dfmt
, nfmt
,
1466 cache_policy
, can_speculate
, true);
1470 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1472 LLVMValueRef voffset
,
1473 LLVMValueRef soffset
,
1474 LLVMValueRef immoffset
,
1475 unsigned num_channels
,
1478 unsigned cache_policy
,
1481 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1482 immoffset
, num_channels
, dfmt
, nfmt
,
1483 cache_policy
, can_speculate
, false);
1487 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1489 LLVMValueRef voffset
,
1490 LLVMValueRef soffset
,
1491 LLVMValueRef immoffset
,
1492 unsigned cache_policy
)
1496 if (LLVM_VERSION_MAJOR
>= 9) {
1497 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1499 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1500 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1502 1, ctx
->i16
, cache_policy
,
1503 false, false, false);
1505 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1506 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1508 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1509 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1512 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1519 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1521 LLVMValueRef voffset
,
1522 LLVMValueRef soffset
,
1523 LLVMValueRef immoffset
,
1524 unsigned cache_policy
)
1528 if (LLVM_VERSION_MAJOR
>= 9) {
1529 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1531 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1532 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1534 1, ctx
->i8
, cache_policy
,
1535 false, false, false);
1537 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1538 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1540 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1541 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1544 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1551 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1553 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1554 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1557 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1559 assert(LLVMTypeOf(src
) == ctx
->i32
);
1562 LLVMValueRef mantissa
;
1563 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1565 /* Converting normal numbers is just a shift + correcting the exponent bias */
1566 unsigned normal_shift
= 23 - mant_bits
;
1567 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1568 LLVMValueRef shifted
, normal
;
1570 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1571 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1573 /* Converting nan/inf numbers is the same, but with a different exponent update */
1574 LLVMValueRef naninf
;
1575 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1577 /* Converting denormals is the complex case: determine the leading zeros of the
1578 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1580 LLVMValueRef denormal
;
1581 LLVMValueRef params
[2] = {
1583 ctx
->i1true
, /* result can be undef when arg is 0 */
1585 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1586 params
, 2, AC_FUNC_ATTR_READNONE
);
1588 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1589 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1590 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1592 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1593 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1594 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1595 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1597 /* Select the final result. */
1598 LLVMValueRef result
;
1600 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1601 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1602 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1604 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1605 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1606 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1608 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1609 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1611 return ac_to_float(ctx
, result
);
1615 * Generate a fully general open coded buffer format fetch with all required
1616 * fixups suitable for vertex fetch, using non-format buffer loads.
1618 * Some combinations of argument values have special interpretations:
1619 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1620 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1622 * \param log_size log(size of channel in bytes)
1623 * \param num_channels number of channels (1 to 4)
1624 * \param format AC_FETCH_FORMAT_xxx value
1625 * \param reverse whether XYZ channels are reversed
1626 * \param known_aligned whether the source is known to be aligned to hardware's
1627 * effective element size for loading the given format
1628 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1629 * \param rsrc buffer resource descriptor
1630 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1633 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1635 unsigned num_channels
,
1640 LLVMValueRef vindex
,
1641 LLVMValueRef voffset
,
1642 LLVMValueRef soffset
,
1643 unsigned cache_policy
,
1647 unsigned load_log_size
= log_size
;
1648 unsigned load_num_channels
= num_channels
;
1649 if (log_size
== 3) {
1651 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1652 load_num_channels
= 2 * num_channels
;
1654 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1658 int log_recombine
= 0;
1659 if ((ctx
->chip_class
== GFX6
|| ctx
->chip_class
== GFX10
) && !known_aligned
) {
1660 /* Avoid alignment restrictions by loading one byte at a time. */
1661 load_num_channels
<<= load_log_size
;
1662 log_recombine
= load_log_size
;
1664 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1665 log_recombine
= -util_logbase2(load_num_channels
);
1666 load_num_channels
= 1;
1667 load_log_size
+= -log_recombine
;
1670 assert(load_log_size
>= 2 || LLVM_VERSION_MAJOR
>= 9);
1672 LLVMValueRef loads
[32]; /* up to 32 bytes */
1673 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1674 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1675 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1676 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1677 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1678 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1679 loads
[i
] = ac_build_buffer_load_common(
1680 ctx
, rsrc
, vindex
, voffset
, tmp
,
1681 num_channels
, channel_type
, cache_policy
,
1682 can_speculate
, false, true);
1683 if (load_log_size
>= 2)
1684 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1687 if (log_recombine
> 0) {
1688 /* Recombine bytes if necessary (GFX6 only) */
1689 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1691 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1692 LLVMValueRef accum
= NULL
;
1693 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1694 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1698 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1699 LLVMConstInt(dst_type
, 8 * i
, false), "");
1700 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1705 } else if (log_recombine
< 0) {
1706 /* Split vectors of dwords */
1707 if (load_log_size
> 2) {
1708 assert(load_num_channels
== 1);
1709 LLVMValueRef loaded
= loads
[0];
1710 unsigned log_split
= load_log_size
- 2;
1711 log_recombine
+= log_split
;
1712 load_num_channels
= 1 << log_split
;
1714 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1715 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1716 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1720 /* Further split dwords and shorts if required */
1721 if (log_recombine
< 0) {
1722 for (unsigned src
= load_num_channels
,
1723 dst
= load_num_channels
<< -log_recombine
;
1725 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1726 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1727 LLVMValueRef loaded
= loads
[src
- 1];
1728 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1729 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1730 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1731 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1732 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1738 if (log_size
== 3) {
1739 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1740 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1741 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1742 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1744 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1745 /* 10_11_11_FLOAT */
1746 LLVMValueRef data
= loads
[0];
1747 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1748 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1749 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1750 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1751 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1753 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1754 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1755 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1759 format
= AC_FETCH_FORMAT_FLOAT
;
1761 /* 2_10_10_10 data formats */
1762 LLVMValueRef data
= loads
[0];
1763 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1764 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1765 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1766 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1767 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1768 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1769 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1770 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1771 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1777 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1778 if (log_size
!= 2) {
1779 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1780 tmp
= ac_to_float(ctx
, loads
[chan
]);
1782 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1783 else if (log_size
== 1)
1784 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1785 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1788 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1789 if (log_size
!= 2) {
1790 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1791 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1793 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1794 if (log_size
!= 2) {
1795 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1796 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1799 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1800 format
== AC_FETCH_FORMAT_USCALED
||
1801 format
== AC_FETCH_FORMAT_UINT
;
1803 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1805 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1807 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1810 LLVMValueRef scale
= NULL
;
1811 if (format
== AC_FETCH_FORMAT_FIXED
) {
1812 assert(log_size
== 2);
1813 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1814 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1815 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1816 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1817 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1818 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1819 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1822 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1824 if (format
== AC_FETCH_FORMAT_SNORM
) {
1825 /* Clamp to [-1, 1] */
1826 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1827 LLVMValueRef clamp
=
1828 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1829 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1832 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1836 while (num_channels
< 4) {
1837 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1838 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1840 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1847 loads
[0] = loads
[2];
1851 return ac_build_gather_values(ctx
, loads
, 4);
1855 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1858 LLVMValueRef vindex
,
1859 LLVMValueRef voffset
,
1860 LLVMValueRef soffset
,
1861 LLVMValueRef immoffset
,
1862 unsigned num_channels
,
1865 unsigned cache_policy
,
1868 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
? voffset
: ctx
->i32_0
,
1871 LLVMValueRef args
[7];
1873 args
[idx
++] = vdata
;
1874 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1876 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1877 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1878 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1879 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1880 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1881 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1882 const char *indexing_kind
= structurized
? "struct" : "raw";
1883 char name
[256], type_name
[8];
1885 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1886 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1888 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1889 indexing_kind
, type_name
);
1891 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1892 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1896 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1899 LLVMValueRef vindex
,
1900 LLVMValueRef voffset
,
1901 LLVMValueRef soffset
,
1902 LLVMValueRef immoffset
,
1903 unsigned num_channels
,
1906 unsigned cache_policy
)
1908 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1909 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1914 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1917 LLVMValueRef voffset
,
1918 LLVMValueRef soffset
,
1919 LLVMValueRef immoffset
,
1920 unsigned num_channels
,
1923 unsigned cache_policy
)
1925 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1926 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1931 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1934 LLVMValueRef voffset
,
1935 LLVMValueRef soffset
,
1936 unsigned cache_policy
)
1938 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1940 if (LLVM_VERSION_MAJOR
>= 9) {
1941 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1942 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1943 voffset
, soffset
, cache_policy
,
1946 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1947 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1949 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1951 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1952 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1957 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1960 LLVMValueRef voffset
,
1961 LLVMValueRef soffset
,
1962 unsigned cache_policy
)
1964 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1966 if (LLVM_VERSION_MAJOR
>= 9) {
1967 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1968 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1969 voffset
, soffset
, cache_policy
,
1972 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1973 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1975 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1977 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1978 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1982 * Set range metadata on an instruction. This can only be used on load and
1983 * call instructions. If you know an instruction can only produce the values
1984 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1985 * \p lo is the minimum value inclusive.
1986 * \p hi is the maximum value exclusive.
1988 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1989 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1991 LLVMValueRef range_md
, md_args
[2];
1992 LLVMTypeRef type
= LLVMTypeOf(value
);
1993 LLVMContextRef context
= LLVMGetTypeContext(type
);
1995 md_args
[0] = LLVMConstInt(type
, lo
, false);
1996 md_args
[1] = LLVMConstInt(type
, hi
, false);
1997 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1998 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
2002 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2006 LLVMValueRef tid_args
[2];
2007 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2008 tid_args
[1] = ctx
->i32_0
;
2009 tid_args
[1] = ac_build_intrinsic(ctx
,
2010 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2011 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2013 if (ctx
->wave_size
== 32) {
2016 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2018 2, AC_FUNC_ATTR_READNONE
);
2020 set_range_metadata(ctx
, tid
, 0, ctx
->wave_size
);
2025 * AMD GCN implements derivatives using the local data store (LDS)
2026 * All writes to the LDS happen in all executing threads at
2027 * the same time. TID is the Thread ID for the current
2028 * thread and is a value between 0 and 63, representing
2029 * the thread's position in the wavefront.
2031 * For the pixel shader threads are grouped into quads of four pixels.
2032 * The TIDs of the pixels of a quad are:
2040 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2041 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2042 * the current pixel's column, and masking with 0xfffffffe yields the TID
2043 * of the left pixel of the current pixel's row.
2045 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2046 * adding 2 yields the TID of the pixel below the top pixel.
2049 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2054 unsigned tl_lanes
[4], trbl_lanes
[4];
2055 char name
[32], type
[8];
2056 LLVMValueRef tl
, trbl
;
2057 LLVMTypeRef result_type
;
2058 LLVMValueRef result
;
2060 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2062 if (result_type
== ctx
->f16
)
2063 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2064 else if (result_type
== ctx
->v2f16
)
2065 val
= LLVMBuildBitCast(ctx
->builder
, val
, ctx
->i32
, "");
2067 for (unsigned i
= 0; i
< 4; ++i
) {
2068 tl_lanes
[i
] = i
& mask
;
2069 trbl_lanes
[i
] = (i
& mask
) + idx
;
2072 tl
= ac_build_quad_swizzle(ctx
, val
,
2073 tl_lanes
[0], tl_lanes
[1],
2074 tl_lanes
[2], tl_lanes
[3]);
2075 trbl
= ac_build_quad_swizzle(ctx
, val
,
2076 trbl_lanes
[0], trbl_lanes
[1],
2077 trbl_lanes
[2], trbl_lanes
[3]);
2079 if (result_type
== ctx
->f16
) {
2080 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2081 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2084 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2085 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2086 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2088 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2089 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2091 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2095 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2097 LLVMValueRef wave_id
)
2099 LLVMValueRef args
[2];
2100 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2102 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2106 ac_build_imsb(struct ac_llvm_context
*ctx
,
2108 LLVMTypeRef dst_type
)
2110 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2112 AC_FUNC_ATTR_READNONE
);
2114 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2115 * the index from LSB. Invert it by doing "31 - msb". */
2116 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2119 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2120 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2121 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2122 arg
, ctx
->i32_0
, ""),
2123 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2124 arg
, all_ones
, ""), "");
2126 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2130 ac_build_umsb(struct ac_llvm_context
*ctx
,
2132 LLVMTypeRef dst_type
)
2134 const char *intrin_name
;
2136 LLVMValueRef highest_bit
;
2140 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2143 intrin_name
= "llvm.ctlz.i64";
2145 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2149 intrin_name
= "llvm.ctlz.i32";
2151 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2155 intrin_name
= "llvm.ctlz.i16";
2157 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2161 intrin_name
= "llvm.ctlz.i8";
2163 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2167 unreachable(!"invalid bitsize");
2171 LLVMValueRef params
[2] = {
2176 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2178 AC_FUNC_ATTR_READNONE
);
2180 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2181 * the index from LSB. Invert it by doing "31 - msb". */
2182 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2184 if (bitsize
== 64) {
2185 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2186 } else if (bitsize
< 32) {
2187 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2190 /* check for zero */
2191 return LLVMBuildSelect(ctx
->builder
,
2192 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2193 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2196 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2199 char name
[64], type
[64];
2201 ac_build_type_name_for_intr(LLVMTypeOf(a
), type
, sizeof(type
));
2202 snprintf(name
, sizeof(name
), "llvm.minnum.%s", type
);
2203 LLVMValueRef args
[2] = {a
, b
};
2204 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2205 AC_FUNC_ATTR_READNONE
);
2208 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2211 char name
[64], type
[64];
2213 ac_build_type_name_for_intr(LLVMTypeOf(a
), type
, sizeof(type
));
2214 snprintf(name
, sizeof(name
), "llvm.maxnum.%s", type
);
2215 LLVMValueRef args
[2] = {a
, b
};
2216 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2217 AC_FUNC_ATTR_READNONE
);
2220 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2223 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2224 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2227 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2230 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2231 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2234 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2237 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2238 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2241 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2244 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2245 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2248 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2250 LLVMTypeRef t
= LLVMTypeOf(value
);
2251 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2252 LLVMConstReal(t
, 1.0));
2255 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2257 LLVMValueRef args
[9];
2259 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2260 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2263 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2265 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2267 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2268 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2270 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2271 ctx
->voidt
, args
, 6, 0);
2273 args
[2] = a
->out
[0];
2274 args
[3] = a
->out
[1];
2275 args
[4] = a
->out
[2];
2276 args
[5] = a
->out
[3];
2277 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2278 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2280 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2281 ctx
->voidt
, args
, 8, 0);
2285 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2287 struct ac_export_args args
;
2289 args
.enabled_channels
= 0x0; /* enabled channels */
2290 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2291 args
.done
= 1; /* DONE bit */
2292 args
.target
= V_008DFC_SQ_EXP_NULL
;
2293 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2294 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2295 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2296 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2297 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2299 ac_build_export(ctx
, &args
);
2302 static unsigned ac_num_coords(enum ac_image_dim dim
)
2308 case ac_image_1darray
:
2312 case ac_image_2darray
:
2313 case ac_image_2dmsaa
:
2315 case ac_image_2darraymsaa
:
2318 unreachable("ac_num_coords: bad dim");
2322 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2326 case ac_image_1darray
:
2329 case ac_image_2darray
:
2334 case ac_image_2dmsaa
:
2335 case ac_image_2darraymsaa
:
2337 unreachable("derivatives not supported");
2341 static const char *get_atomic_name(enum ac_atomic_op op
)
2344 case ac_atomic_swap
: return "swap";
2345 case ac_atomic_add
: return "add";
2346 case ac_atomic_sub
: return "sub";
2347 case ac_atomic_smin
: return "smin";
2348 case ac_atomic_umin
: return "umin";
2349 case ac_atomic_smax
: return "smax";
2350 case ac_atomic_umax
: return "umax";
2351 case ac_atomic_and
: return "and";
2352 case ac_atomic_or
: return "or";
2353 case ac_atomic_xor
: return "xor";
2354 case ac_atomic_inc_wrap
: return "inc";
2355 case ac_atomic_dec_wrap
: return "dec";
2357 unreachable("bad atomic op");
2360 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2361 struct ac_image_args
*a
)
2363 const char *overload
[3] = { "", "", "" };
2364 unsigned num_overloads
= 0;
2365 LLVMValueRef args
[18];
2366 unsigned num_args
= 0;
2367 enum ac_image_dim dim
= a
->dim
;
2369 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2371 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2372 a
->opcode
!= ac_image_store_mip
) ||
2374 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2375 (!a
->compare
&& !a
->offset
));
2376 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2377 a
->opcode
== ac_image_get_lod
) ||
2379 assert((a
->bias
? 1 : 0) +
2381 (a
->level_zero
? 1 : 0) +
2382 (a
->derivs
[0] ? 1 : 0) <= 1);
2383 assert((a
->min_lod
? 1 : 0) +
2385 (a
->level_zero
? 1 : 0) <= 1);
2386 assert(!a
->d16
|| (ctx
->chip_class
>= GFX8
&&
2387 a
->opcode
!= ac_image_atomic
&&
2388 a
->opcode
!= ac_image_atomic_cmpswap
&&
2389 a
->opcode
!= ac_image_get_lod
&&
2390 a
->opcode
!= ac_image_get_resinfo
));
2392 if (a
->opcode
== ac_image_get_lod
) {
2394 case ac_image_1darray
:
2397 case ac_image_2darray
:
2406 bool sample
= a
->opcode
== ac_image_sample
||
2407 a
->opcode
== ac_image_gather4
||
2408 a
->opcode
== ac_image_get_lod
;
2409 bool atomic
= a
->opcode
== ac_image_atomic
||
2410 a
->opcode
== ac_image_atomic_cmpswap
;
2411 bool load
= a
->opcode
== ac_image_sample
||
2412 a
->opcode
== ac_image_gather4
||
2413 a
->opcode
== ac_image_load
||
2414 a
->opcode
== ac_image_load_mip
;
2415 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2417 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2418 args
[num_args
++] = a
->data
[0];
2419 if (a
->opcode
== ac_image_atomic_cmpswap
)
2420 args
[num_args
++] = a
->data
[1];
2424 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2427 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2429 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2430 overload
[num_overloads
++] = ".f32";
2433 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2435 unsigned count
= ac_num_derivs(dim
);
2436 for (unsigned i
= 0; i
< count
; ++i
)
2437 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2438 overload
[num_overloads
++] = ".f32";
2440 unsigned num_coords
=
2441 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2442 for (unsigned i
= 0; i
< num_coords
; ++i
)
2443 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2445 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2447 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->min_lod
, coord_type
, "");
2449 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2451 args
[num_args
++] = a
->resource
;
2453 args
[num_args
++] = a
->sampler
;
2454 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2457 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2458 args
[num_args
++] = LLVMConstInt(ctx
->i32
,
2459 load
? get_load_cache_policy(ctx
, a
->cache_policy
) :
2460 a
->cache_policy
, false);
2463 const char *atomic_subop
= "";
2464 switch (a
->opcode
) {
2465 case ac_image_sample
: name
= "sample"; break;
2466 case ac_image_gather4
: name
= "gather4"; break;
2467 case ac_image_load
: name
= "load"; break;
2468 case ac_image_load_mip
: name
= "load.mip"; break;
2469 case ac_image_store
: name
= "store"; break;
2470 case ac_image_store_mip
: name
= "store.mip"; break;
2471 case ac_image_atomic
:
2473 atomic_subop
= get_atomic_name(a
->atomic
);
2475 case ac_image_atomic_cmpswap
:
2477 atomic_subop
= "cmpswap";
2479 case ac_image_get_lod
: name
= "getlod"; break;
2480 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2481 default: unreachable("invalid image opcode");
2484 const char *dimname
;
2486 case ac_image_1d
: dimname
= "1d"; break;
2487 case ac_image_2d
: dimname
= "2d"; break;
2488 case ac_image_3d
: dimname
= "3d"; break;
2489 case ac_image_cube
: dimname
= "cube"; break;
2490 case ac_image_1darray
: dimname
= "1darray"; break;
2491 case ac_image_2darray
: dimname
= "2darray"; break;
2492 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2493 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2494 default: unreachable("invalid dim");
2498 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2500 snprintf(intr_name
, sizeof(intr_name
),
2501 "llvm.amdgcn.image.%s%s" /* base name */
2502 "%s%s%s%s" /* sample/gather modifiers */
2503 ".%s.%s%s%s%s", /* dimension and type overloads */
2505 a
->compare
? ".c" : "",
2508 a
->derivs
[0] ? ".d" :
2509 a
->level_zero
? ".lz" : "",
2510 a
->min_lod
? ".cl" : "",
2511 a
->offset
? ".o" : "",
2513 atomic
? "i32" : (a
->d16
? "v4f16" : "v4f32"),
2514 overload
[0], overload
[1], overload
[2]);
2519 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2522 retty
= a
->d16
? ctx
->v4f16
: ctx
->v4f32
;
2524 LLVMValueRef result
=
2525 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2527 if (!sample
&& !atomic
&& retty
!= ctx
->voidt
)
2528 result
= ac_to_integer(ctx
, result
);
2533 LLVMValueRef
ac_build_image_get_sample_count(struct ac_llvm_context
*ctx
,
2536 LLVMValueRef samples
;
2538 /* Read the samples from the descriptor directly.
2539 * Hardware doesn't have any instruction for this.
2541 samples
= LLVMBuildExtractElement(ctx
->builder
, rsrc
,
2542 LLVMConstInt(ctx
->i32
, 3, 0), "");
2543 samples
= LLVMBuildLShr(ctx
->builder
, samples
,
2544 LLVMConstInt(ctx
->i32
, 16, 0), "");
2545 samples
= LLVMBuildAnd(ctx
->builder
, samples
,
2546 LLVMConstInt(ctx
->i32
, 0xf, 0), "");
2547 samples
= LLVMBuildShl(ctx
->builder
, ctx
->i32_1
,
2552 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2553 LLVMValueRef args
[2])
2555 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", ctx
->v2f16
,
2556 args
, 2, AC_FUNC_ATTR_READNONE
);
2559 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2560 LLVMValueRef args
[2])
2563 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2564 ctx
->v2i16
, args
, 2,
2565 AC_FUNC_ATTR_READNONE
);
2566 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2569 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2570 LLVMValueRef args
[2])
2573 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2574 ctx
->v2i16
, args
, 2,
2575 AC_FUNC_ATTR_READNONE
);
2576 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2579 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2580 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2581 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2583 assert(bits
== 8 || bits
== 10 || bits
== 16);
2585 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2586 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2587 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2588 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2589 LLVMValueRef max_alpha
=
2590 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2591 LLVMValueRef min_alpha
=
2592 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2596 for (int i
= 0; i
< 2; i
++) {
2597 bool alpha
= hi
&& i
== 1;
2598 args
[i
] = ac_build_imin(ctx
, args
[i
],
2599 alpha
? max_alpha
: max_rgb
);
2600 args
[i
] = ac_build_imax(ctx
, args
[i
],
2601 alpha
? min_alpha
: min_rgb
);
2606 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2607 ctx
->v2i16
, args
, 2,
2608 AC_FUNC_ATTR_READNONE
);
2609 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2612 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2613 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2614 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2616 assert(bits
== 8 || bits
== 10 || bits
== 16);
2618 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2619 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2620 LLVMValueRef max_alpha
=
2621 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2625 for (int i
= 0; i
< 2; i
++) {
2626 bool alpha
= hi
&& i
== 1;
2627 args
[i
] = ac_build_umin(ctx
, args
[i
],
2628 alpha
? max_alpha
: max_rgb
);
2633 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2634 ctx
->v2i16
, args
, 2,
2635 AC_FUNC_ATTR_READNONE
);
2636 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2639 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2641 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2642 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2645 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2647 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2651 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2652 LLVMValueRef offset
, LLVMValueRef width
,
2655 LLVMValueRef args
[] = {
2661 return ac_build_intrinsic(ctx
, is_signed
? "llvm.amdgcn.sbfe.i32" :
2662 "llvm.amdgcn.ubfe.i32",
2663 ctx
->i32
, args
, 3, AC_FUNC_ATTR_READNONE
);
2667 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2668 LLVMValueRef s1
, LLVMValueRef s2
)
2670 return LLVMBuildAdd(ctx
->builder
,
2671 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2674 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2675 LLVMValueRef s1
, LLVMValueRef s2
)
2677 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2678 if (ctx
->chip_class
>= GFX10
) {
2679 return ac_build_intrinsic(ctx
, "llvm.fma.f32", ctx
->f32
,
2680 (LLVMValueRef
[]) {s0
, s1
, s2
}, 3,
2681 AC_FUNC_ATTR_READNONE
);
2684 return LLVMBuildFAdd(ctx
->builder
,
2685 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2688 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned wait_flags
)
2693 unsigned lgkmcnt
= 63;
2694 unsigned vmcnt
= ctx
->chip_class
>= GFX9
? 63 : 15;
2695 unsigned vscnt
= 63;
2697 if (wait_flags
& AC_WAIT_LGKM
)
2699 if (wait_flags
& AC_WAIT_VLOAD
)
2702 if (wait_flags
& AC_WAIT_VSTORE
) {
2703 if (ctx
->chip_class
>= GFX10
)
2709 /* There is no intrinsic for vscnt(0), so use a fence. */
2710 if ((wait_flags
& AC_WAIT_LGKM
&&
2711 wait_flags
& AC_WAIT_VLOAD
&&
2712 wait_flags
& AC_WAIT_VSTORE
) ||
2714 LLVMBuildFence(ctx
->builder
, LLVMAtomicOrderingRelease
, false, "");
2718 unsigned simm16
= (lgkmcnt
<< 8) |
2719 (7 << 4) | /* expcnt */
2721 ((vmcnt
>> 4) << 14);
2723 LLVMValueRef args
[1] = {
2724 LLVMConstInt(ctx
->i32
, simm16
, false),
2726 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2727 ctx
->voidt
, args
, 1, 0);
2730 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2731 LLVMValueRef src1
, LLVMValueRef src2
,
2734 LLVMValueRef result
;
2736 if (bitsize
== 64 || (bitsize
== 16 && ctx
->chip_class
<= GFX8
)) {
2737 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2738 * or lower 16-bit fmed because it's only supported on GFX9+.
2740 LLVMValueRef min1
, min2
, max1
;
2742 min1
= ac_build_fmin(ctx
, src0
, src1
);
2743 max1
= ac_build_fmax(ctx
, src0
, src1
);
2744 min2
= ac_build_fmin(ctx
, max1
, src2
);
2746 result
= ac_build_fmax(ctx
, min2
, min1
);
2751 if (bitsize
== 16) {
2752 intr
= "llvm.amdgcn.fmed3.f16";
2755 assert(bitsize
== 32);
2756 intr
= "llvm.amdgcn.fmed3.f32";
2760 LLVMValueRef params
[] = {
2766 result
= ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2767 AC_FUNC_ATTR_READNONE
);
2770 if (ctx
->chip_class
< GFX9
&& bitsize
== 32) {
2771 /* Only pre-GFX9 chips do not flush denorms. */
2772 result
= ac_build_canonicalize(ctx
, result
, bitsize
);
2778 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2784 if (bitsize
== 16) {
2785 intr
= "llvm.amdgcn.fract.f16";
2787 } else if (bitsize
== 32) {
2788 intr
= "llvm.amdgcn.fract.f32";
2791 intr
= "llvm.amdgcn.fract.f64";
2795 LLVMValueRef params
[] = {
2798 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2799 AC_FUNC_ATTR_READNONE
);
2802 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2805 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2806 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2807 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2809 LLVMValueRef cmp
, val
;
2810 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2811 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2812 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2813 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2817 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2820 LLVMValueRef cmp
, val
, zero
, one
;
2823 if (bitsize
== 16) {
2827 } else if (bitsize
== 32) {
2837 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2838 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2839 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2840 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2844 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2846 LLVMValueRef result
;
2849 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2853 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i128", ctx
->i128
,
2854 (LLVMValueRef
[]) { src0
}, 1,
2855 AC_FUNC_ATTR_READNONE
);
2856 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2859 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2860 (LLVMValueRef
[]) { src0
}, 1,
2861 AC_FUNC_ATTR_READNONE
);
2863 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2866 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2867 (LLVMValueRef
[]) { src0
}, 1,
2868 AC_FUNC_ATTR_READNONE
);
2871 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2872 (LLVMValueRef
[]) { src0
}, 1,
2873 AC_FUNC_ATTR_READNONE
);
2875 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2878 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2879 (LLVMValueRef
[]) { src0
}, 1,
2880 AC_FUNC_ATTR_READNONE
);
2882 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2885 unreachable(!"invalid bitsize");
2892 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2895 LLVMValueRef result
;
2898 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2902 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2903 (LLVMValueRef
[]) { src0
}, 1,
2904 AC_FUNC_ATTR_READNONE
);
2906 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2909 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2910 (LLVMValueRef
[]) { src0
}, 1,
2911 AC_FUNC_ATTR_READNONE
);
2914 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2915 (LLVMValueRef
[]) { src0
}, 1,
2916 AC_FUNC_ATTR_READNONE
);
2918 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2921 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
2922 (LLVMValueRef
[]) { src0
}, 1,
2923 AC_FUNC_ATTR_READNONE
);
2925 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2928 unreachable(!"invalid bitsize");
2935 #define AC_EXP_TARGET 0
2936 #define AC_EXP_ENABLED_CHANNELS 1
2937 #define AC_EXP_OUT0 2
2945 struct ac_vs_exp_chan
2949 enum ac_ir_type type
;
2952 struct ac_vs_exp_inst
{
2955 struct ac_vs_exp_chan chan
[4];
2958 struct ac_vs_exports
{
2960 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2963 /* Return true if the PARAM export has been eliminated. */
2964 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2965 uint32_t num_outputs
,
2966 struct ac_vs_exp_inst
*exp
)
2968 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2969 bool is_zero
[4] = {}, is_one
[4] = {};
2971 for (i
= 0; i
< 4; i
++) {
2972 /* It's a constant expression. Undef outputs are eliminated too. */
2973 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2976 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2977 if (exp
->chan
[i
].const_float
== 0)
2979 else if (exp
->chan
[i
].const_float
== 1)
2982 return false; /* other constant */
2987 /* Only certain combinations of 0 and 1 can be eliminated. */
2988 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2989 default_val
= is_zero
[3] ? 0 : 1;
2990 else if (is_one
[0] && is_one
[1] && is_one
[2])
2991 default_val
= is_zero
[3] ? 2 : 3;
2995 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2996 LLVMInstructionEraseFromParent(exp
->inst
);
2998 /* Change OFFSET to DEFAULT_VAL. */
2999 for (i
= 0; i
< num_outputs
; i
++) {
3000 if (vs_output_param_offset
[i
] == exp
->offset
) {
3001 vs_output_param_offset
[i
] =
3002 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
3009 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
3010 uint8_t *vs_output_param_offset
,
3011 uint32_t num_outputs
,
3012 struct ac_vs_exports
*processed
,
3013 struct ac_vs_exp_inst
*exp
)
3015 unsigned p
, copy_back_channels
= 0;
3017 /* See if the output is already in the list of processed outputs.
3018 * The LLVMValueRef comparison relies on SSA.
3020 for (p
= 0; p
< processed
->num
; p
++) {
3021 bool different
= false;
3023 for (unsigned j
= 0; j
< 4; j
++) {
3024 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
3025 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
3027 /* Treat undef as a match. */
3028 if (c2
->type
== AC_IR_UNDEF
)
3031 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3032 * and consider the instruction duplicated.
3034 if (c1
->type
== AC_IR_UNDEF
) {
3035 copy_back_channels
|= 1 << j
;
3039 /* Test whether the channels are not equal. */
3040 if (c1
->type
!= c2
->type
||
3041 (c1
->type
== AC_IR_CONST
&&
3042 c1
->const_float
!= c2
->const_float
) ||
3043 (c1
->type
== AC_IR_VALUE
&&
3044 c1
->value
!= c2
->value
)) {
3052 copy_back_channels
= 0;
3054 if (p
== processed
->num
)
3057 /* If a match was found, but the matching export has undef where the new
3058 * one has a normal value, copy the normal value to the undef channel.
3060 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3062 /* Get current enabled channels mask. */
3063 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3064 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3066 while (copy_back_channels
) {
3067 unsigned chan
= u_bit_scan(©_back_channels
);
3069 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3070 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3071 exp
->chan
[chan
].value
);
3072 match
->chan
[chan
] = exp
->chan
[chan
];
3074 /* Update number of enabled channels because the original mask
3075 * is not always 0xf.
3077 enabled_channels
|= (1 << chan
);
3078 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3079 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3082 /* The PARAM export is duplicated. Kill it. */
3083 LLVMInstructionEraseFromParent(exp
->inst
);
3085 /* Change OFFSET to the matching export. */
3086 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3087 if (vs_output_param_offset
[i
] == exp
->offset
) {
3088 vs_output_param_offset
[i
] = match
->offset
;
3095 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3096 LLVMValueRef main_fn
,
3097 uint8_t *vs_output_param_offset
,
3098 uint32_t num_outputs
,
3099 uint32_t skip_output_mask
,
3100 uint8_t *num_param_exports
)
3102 LLVMBasicBlockRef bb
;
3103 bool removed_any
= false;
3104 struct ac_vs_exports exports
;
3108 /* Process all LLVM instructions. */
3109 bb
= LLVMGetFirstBasicBlock(main_fn
);
3111 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3114 LLVMValueRef cur
= inst
;
3115 inst
= LLVMGetNextInstruction(inst
);
3116 struct ac_vs_exp_inst exp
;
3118 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3121 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3123 if (!ac_llvm_is_function(callee
))
3126 const char *name
= LLVMGetValueName(callee
);
3127 unsigned num_args
= LLVMCountParams(callee
);
3129 /* Check if this is an export instruction. */
3130 if ((num_args
!= 9 && num_args
!= 8) ||
3131 (strcmp(name
, "llvm.SI.export") &&
3132 strcmp(name
, "llvm.amdgcn.exp.f32")))
3135 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3136 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3138 if (target
< V_008DFC_SQ_EXP_PARAM
)
3141 target
-= V_008DFC_SQ_EXP_PARAM
;
3143 /* Parse the instruction. */
3144 memset(&exp
, 0, sizeof(exp
));
3145 exp
.offset
= target
;
3148 for (unsigned i
= 0; i
< 4; i
++) {
3149 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3151 exp
.chan
[i
].value
= v
;
3153 if (LLVMIsUndef(v
)) {
3154 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3155 } else if (LLVMIsAConstantFP(v
)) {
3156 LLVMBool loses_info
;
3157 exp
.chan
[i
].type
= AC_IR_CONST
;
3158 exp
.chan
[i
].const_float
=
3159 LLVMConstRealGetDouble(v
, &loses_info
);
3161 exp
.chan
[i
].type
= AC_IR_VALUE
;
3165 /* Eliminate constant and duplicated PARAM exports. */
3166 if (!((1u << target
) & skip_output_mask
) &&
3167 (ac_eliminate_const_output(vs_output_param_offset
,
3168 num_outputs
, &exp
) ||
3169 ac_eliminate_duplicated_output(ctx
,
3170 vs_output_param_offset
,
3171 num_outputs
, &exports
,
3175 exports
.exp
[exports
.num
++] = exp
;
3178 bb
= LLVMGetNextBasicBlock(bb
);
3181 /* Remove holes in export memory due to removed PARAM exports.
3182 * This is done by renumbering all PARAM exports.
3185 uint8_t old_offset
[VARYING_SLOT_MAX
];
3188 /* Make a copy of the offsets. We need the old version while
3189 * we are modifying some of them. */
3190 memcpy(old_offset
, vs_output_param_offset
,
3191 sizeof(old_offset
));
3193 for (i
= 0; i
< exports
.num
; i
++) {
3194 unsigned offset
= exports
.exp
[i
].offset
;
3196 /* Update vs_output_param_offset. Multiple outputs can
3197 * have the same offset.
3199 for (out
= 0; out
< num_outputs
; out
++) {
3200 if (old_offset
[out
] == offset
)
3201 vs_output_param_offset
[out
] = i
;
3204 /* Change the PARAM offset in the instruction. */
3205 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3206 LLVMConstInt(ctx
->i32
,
3207 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3209 *num_param_exports
= exports
.num
;
3213 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3215 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3216 ac_build_intrinsic(ctx
,
3217 "llvm.amdgcn.init.exec", ctx
->voidt
,
3218 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3221 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3223 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3224 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3225 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3229 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3230 LLVMValueRef dw_addr
)
3232 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3235 void ac_lds_store(struct ac_llvm_context
*ctx
,
3236 LLVMValueRef dw_addr
,
3239 value
= ac_to_integer(ctx
, value
);
3240 ac_build_indexed_store(ctx
, ctx
->lds
,
3244 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3245 LLVMTypeRef dst_type
,
3248 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3249 const char *intrin_name
;
3253 switch (src0_bitsize
) {
3255 intrin_name
= "llvm.cttz.i64";
3260 intrin_name
= "llvm.cttz.i32";
3265 intrin_name
= "llvm.cttz.i16";
3270 intrin_name
= "llvm.cttz.i8";
3275 unreachable(!"invalid bitsize");
3278 LLVMValueRef params
[2] = {
3281 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3282 * add special code to check for x=0. The reason is that
3283 * the LLVM behavior for x=0 is different from what we
3284 * need here. However, LLVM also assumes that ffs(x) is
3285 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3286 * a conditional assignment to handle 0 is still required.
3288 * The hardware already implements the correct behavior.
3293 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3295 AC_FUNC_ATTR_READNONE
);
3297 if (src0_bitsize
== 64) {
3298 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3299 } else if (src0_bitsize
< 32) {
3300 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3303 /* TODO: We need an intrinsic to skip this conditional. */
3304 /* Check for zero: */
3305 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3308 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3311 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3313 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3316 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3318 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3321 static struct ac_llvm_flow
*
3322 get_current_flow(struct ac_llvm_context
*ctx
)
3324 if (ctx
->flow
->depth
> 0)
3325 return &ctx
->flow
->stack
[ctx
->flow
->depth
- 1];
3329 static struct ac_llvm_flow
*
3330 get_innermost_loop(struct ac_llvm_context
*ctx
)
3332 for (unsigned i
= ctx
->flow
->depth
; i
> 0; --i
) {
3333 if (ctx
->flow
->stack
[i
- 1].loop_entry_block
)
3334 return &ctx
->flow
->stack
[i
- 1];
3339 static struct ac_llvm_flow
*
3340 push_flow(struct ac_llvm_context
*ctx
)
3342 struct ac_llvm_flow
*flow
;
3344 if (ctx
->flow
->depth
>= ctx
->flow
->depth_max
) {
3345 unsigned new_max
= MAX2(ctx
->flow
->depth
<< 1,
3346 AC_LLVM_INITIAL_CF_DEPTH
);
3348 ctx
->flow
->stack
= realloc(ctx
->flow
->stack
, new_max
* sizeof(*ctx
->flow
->stack
));
3349 ctx
->flow
->depth_max
= new_max
;
3352 flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
];
3355 flow
->next_block
= NULL
;
3356 flow
->loop_entry_block
= NULL
;
3360 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3364 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3365 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3368 /* Append a basic block at the level of the parent flow.
3370 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3373 assert(ctx
->flow
->depth
>= 1);
3375 if (ctx
->flow
->depth
>= 2) {
3376 struct ac_llvm_flow
*flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
- 2];
3378 return LLVMInsertBasicBlockInContext(ctx
->context
,
3379 flow
->next_block
, name
);
3382 LLVMValueRef main_fn
=
3383 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3384 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3387 /* Emit a branch to the given default target for the current block if
3388 * applicable -- that is, if the current block does not already contain a
3389 * branch from a break or continue.
3391 static void emit_default_branch(LLVMBuilderRef builder
,
3392 LLVMBasicBlockRef target
)
3394 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3395 LLVMBuildBr(builder
, target
);
3398 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3400 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3401 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3402 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3403 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3404 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3405 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3408 void ac_build_break(struct ac_llvm_context
*ctx
)
3410 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3411 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3414 void ac_build_continue(struct ac_llvm_context
*ctx
)
3416 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3417 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3420 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3422 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3423 LLVMBasicBlockRef endif_block
;
3425 assert(!current_branch
->loop_entry_block
);
3427 endif_block
= append_basic_block(ctx
, "ENDIF");
3428 emit_default_branch(ctx
->builder
, endif_block
);
3430 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3431 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3433 current_branch
->next_block
= endif_block
;
3436 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3438 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3440 assert(!current_branch
->loop_entry_block
);
3442 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3443 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3444 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3449 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3451 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3453 assert(current_loop
->loop_entry_block
);
3455 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3457 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3458 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3462 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3464 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3465 LLVMBasicBlockRef if_block
;
3467 if_block
= append_basic_block(ctx
, "IF");
3468 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3469 set_basicblock_name(if_block
, "if", label_id
);
3470 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3471 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3474 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3477 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3478 value
, ctx
->f32_0
, "");
3479 ac_build_ifcc(ctx
, cond
, label_id
);
3482 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3485 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3486 ac_to_integer(ctx
, value
),
3488 ac_build_ifcc(ctx
, cond
, label_id
);
3491 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3494 LLVMBuilderRef builder
= ac
->builder
;
3495 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3496 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3497 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3498 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3499 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3503 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3505 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3508 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3509 LLVMDisposeBuilder(first_builder
);
3513 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3514 LLVMTypeRef type
, const char *name
)
3516 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3517 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3521 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3524 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3525 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3526 LLVMPointerType(type
, addr_space
), "");
3529 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3532 unsigned num_components
= ac_get_llvm_num_components(value
);
3533 if (count
== num_components
)
3536 LLVMValueRef masks
[MAX2(count
, 2)];
3537 masks
[0] = ctx
->i32_0
;
3538 masks
[1] = ctx
->i32_1
;
3539 for (unsigned i
= 2; i
< count
; i
++)
3540 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3543 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3546 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3547 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3550 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3551 unsigned rshift
, unsigned bitwidth
)
3553 LLVMValueRef value
= param
;
3555 value
= LLVMBuildLShr(ctx
->builder
, value
,
3556 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3558 if (rshift
+ bitwidth
< 32) {
3559 unsigned mask
= (1 << bitwidth
) - 1;
3560 value
= LLVMBuildAnd(ctx
->builder
, value
,
3561 LLVMConstInt(ctx
->i32
, mask
, false), "");
3566 /* Adjust the sample index according to FMASK.
3568 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3569 * which is the identity mapping. Each nibble says which physical sample
3570 * should be fetched to get that sample.
3572 * For example, 0x11111100 means there are only 2 samples stored and
3573 * the second sample covers 3/4 of the pixel. When reading samples 0
3574 * and 1, return physical sample 0 (determined by the first two 0s
3575 * in FMASK), otherwise return physical sample 1.
3577 * The sample index should be adjusted as follows:
3578 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3580 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3581 LLVMValueRef
*addr
, bool is_array_tex
)
3583 struct ac_image_args fmask_load
= {};
3584 fmask_load
.opcode
= ac_image_load
;
3585 fmask_load
.resource
= fmask
;
3586 fmask_load
.dmask
= 0xf;
3587 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3588 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3590 fmask_load
.coords
[0] = addr
[0];
3591 fmask_load
.coords
[1] = addr
[1];
3593 fmask_load
.coords
[2] = addr
[2];
3595 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3596 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3599 /* Apply the formula. */
3600 unsigned sample_chan
= is_array_tex
? 3 : 2;
3601 LLVMValueRef final_sample
;
3602 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3603 LLVMConstInt(ac
->i32
, 4, 0), "");
3604 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3605 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3606 * with EQAA, so those will map to 0. */
3607 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3608 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3610 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3611 * resource descriptor is 0 (invalid).
3614 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3615 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3616 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3618 /* Replace the MSAA sample index. */
3619 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3620 addr
[sample_chan
], "");
3624 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3625 LLVMValueRef lane
, bool with_opt_barrier
)
3627 LLVMTypeRef type
= LLVMTypeOf(src
);
3628 LLVMValueRef result
;
3630 if (with_opt_barrier
)
3631 ac_build_optimization_barrier(ctx
, &src
);
3633 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3635 lane
= LLVMBuildZExt(ctx
->builder
, lane
, ctx
->i32
, "");
3637 result
= ac_build_intrinsic(ctx
,
3638 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3639 ctx
->i32
, (LLVMValueRef
[]) { src
, lane
},
3640 lane
== NULL
? 1 : 2,
3641 AC_FUNC_ATTR_READNONE
|
3642 AC_FUNC_ATTR_CONVERGENT
);
3644 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3648 ac_build_readlane_common(struct ac_llvm_context
*ctx
,
3649 LLVMValueRef src
, LLVMValueRef lane
,
3650 bool with_opt_barrier
)
3652 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3653 src
= ac_to_integer(ctx
, src
);
3654 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3658 assert(bits
% 32 == 0);
3659 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3660 LLVMValueRef src_vector
=
3661 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3662 ret
= LLVMGetUndef(vec_type
);
3663 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3664 LLVMValueRef ret_comp
;
3666 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3667 LLVMConstInt(ctx
->i32
, i
, 0), "");
3669 ret_comp
= _ac_build_readlane(ctx
, src
, lane
,
3672 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3673 LLVMConstInt(ctx
->i32
, i
, 0), "");
3676 ret
= _ac_build_readlane(ctx
, src
, lane
, with_opt_barrier
);
3679 if (LLVMGetTypeKind(src_type
) == LLVMPointerTypeKind
)
3680 return LLVMBuildIntToPtr(ctx
->builder
, ret
, src_type
, "");
3681 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3685 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3687 * The optimization barrier is not needed if the value is the same in all lanes
3688 * or if this is called in the outermost block.
3692 * @param lane - id of the lane or NULL for the first active lane
3693 * @return value of the lane
3695 LLVMValueRef
ac_build_readlane_no_opt_barrier(struct ac_llvm_context
*ctx
,
3696 LLVMValueRef src
, LLVMValueRef lane
)
3698 return ac_build_readlane_common(ctx
, src
, lane
, false);
3703 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3705 return ac_build_readlane_common(ctx
, src
, lane
, true);
3709 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3711 return ac_build_intrinsic(ctx
, "llvm.amdgcn.writelane", ctx
->i32
,
3712 (LLVMValueRef
[]) {value
, lane
, src
}, 3,
3713 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3717 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3719 if (ctx
->wave_size
== 32) {
3720 return ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3721 (LLVMValueRef
[]) { mask
, ctx
->i32_0
},
3722 2, AC_FUNC_ATTR_READNONE
);
3724 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
, ctx
->v2i32
, "");
3725 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3727 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3730 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3731 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3732 2, AC_FUNC_ATTR_READNONE
);
3733 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3734 (LLVMValueRef
[]) { mask_hi
, val
},
3735 2, AC_FUNC_ATTR_READNONE
);
3740 _dpp_quad_perm
= 0x000,
3741 _dpp_row_sl
= 0x100,
3742 _dpp_row_sr
= 0x110,
3743 _dpp_row_rr
= 0x120,
3748 dpp_row_mirror
= 0x140,
3749 dpp_row_half_mirror
= 0x141,
3750 dpp_row_bcast15
= 0x142,
3751 dpp_row_bcast31
= 0x143
3754 static inline enum dpp_ctrl
3755 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3757 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3758 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3761 static inline enum dpp_ctrl
3762 dpp_row_sl(unsigned amount
)
3764 assert(amount
> 0 && amount
< 16);
3765 return _dpp_row_sl
| amount
;
3768 static inline enum dpp_ctrl
3769 dpp_row_sr(unsigned amount
)
3771 assert(amount
> 0 && amount
< 16);
3772 return _dpp_row_sr
| amount
;
3776 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3777 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3780 LLVMTypeRef type
= LLVMTypeOf(src
);
3783 old
= LLVMBuildZExt(ctx
->builder
, old
, ctx
->i32
, "");
3784 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3786 res
= ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32", ctx
->i32
,
3789 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3790 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3791 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3792 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3793 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3795 return LLVMBuildTrunc(ctx
->builder
, res
, type
, "");
3799 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3800 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3803 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3804 src
= ac_to_integer(ctx
, src
);
3805 old
= ac_to_integer(ctx
, old
);
3806 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3809 assert(bits
% 32 == 0);
3810 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3811 LLVMValueRef src_vector
=
3812 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3813 LLVMValueRef old_vector
=
3814 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3815 ret
= LLVMGetUndef(vec_type
);
3816 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3817 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3818 LLVMConstInt(ctx
->i32
, i
,
3820 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3821 LLVMConstInt(ctx
->i32
, i
,
3823 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3828 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3830 LLVMConstInt(ctx
->i32
, i
,
3834 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3835 bank_mask
, bound_ctrl
);
3837 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3841 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3842 bool exchange_rows
, bool bound_ctrl
)
3844 LLVMTypeRef type
= LLVMTypeOf(src
);
3845 LLVMValueRef result
;
3847 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3849 LLVMValueRef args
[6] = {
3852 LLVMConstInt(ctx
->i32
, sel
, false),
3853 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3854 ctx
->i1true
, /* fi */
3855 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3858 result
= ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3859 : "llvm.amdgcn.permlane16",
3861 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3863 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3867 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3868 bool exchange_rows
, bool bound_ctrl
)
3870 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3871 src
= ac_to_integer(ctx
, src
);
3872 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3875 assert(bits
% 32 == 0);
3876 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3877 LLVMValueRef src_vector
=
3878 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3879 ret
= LLVMGetUndef(vec_type
);
3880 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3881 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3882 LLVMConstInt(ctx
->i32
, i
,
3884 LLVMValueRef ret_comp
=
3885 _ac_build_permlane16(ctx
, src
, sel
,
3888 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3890 LLVMConstInt(ctx
->i32
, i
,
3894 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3897 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3900 static inline unsigned
3901 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3903 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3904 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3908 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3910 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3913 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3915 ret
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle", ctx
->i32
,
3917 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3918 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3920 return LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3924 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3926 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3927 src
= ac_to_integer(ctx
, src
);
3928 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3931 assert(bits
% 32 == 0);
3932 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3933 LLVMValueRef src_vector
=
3934 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3935 ret
= LLVMGetUndef(vec_type
);
3936 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3937 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3938 LLVMConstInt(ctx
->i32
, i
,
3940 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3942 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3944 LLVMConstInt(ctx
->i32
, i
,
3948 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3950 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3954 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3956 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3957 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3958 char name
[32], type
[8];
3961 src
= ac_to_integer(ctx
, src
);
3964 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3966 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3967 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3968 ret
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3969 (LLVMValueRef
[]) { src
}, 1,
3970 AC_FUNC_ATTR_READNONE
);
3973 ret
= LLVMBuildTrunc(ctx
->builder
, ret
,
3974 ac_to_integer_type(ctx
, src_type
), "");
3976 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3980 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3981 LLVMValueRef inactive
)
3983 char name
[33], type
[8];
3984 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3985 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3986 src
= ac_to_integer(ctx
, src
);
3987 inactive
= ac_to_integer(ctx
, inactive
);
3990 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3991 inactive
= LLVMBuildZExt(ctx
->builder
, inactive
, ctx
->i32
, "");
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
);
4003 ret
= LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
4009 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
4011 if (type_size
== 1) {
4013 case nir_op_iadd
: return ctx
->i8_0
;
4014 case nir_op_imul
: return ctx
->i8_1
;
4015 case nir_op_imin
: return LLVMConstInt(ctx
->i8
, INT8_MAX
, 0);
4016 case nir_op_umin
: return LLVMConstInt(ctx
->i8
, UINT8_MAX
, 0);
4017 case nir_op_imax
: return LLVMConstInt(ctx
->i8
, INT8_MIN
, 0);
4018 case nir_op_umax
: return ctx
->i8_0
;
4019 case nir_op_iand
: return LLVMConstInt(ctx
->i8
, -1, 0);
4020 case nir_op_ior
: return ctx
->i8_0
;
4021 case nir_op_ixor
: return ctx
->i8_0
;
4023 unreachable("bad reduction intrinsic");
4025 } else if (type_size
== 2) {
4027 case nir_op_iadd
: return ctx
->i16_0
;
4028 case nir_op_fadd
: return ctx
->f16_0
;
4029 case nir_op_imul
: return ctx
->i16_1
;
4030 case nir_op_fmul
: return ctx
->f16_1
;
4031 case nir_op_imin
: return LLVMConstInt(ctx
->i16
, INT16_MAX
, 0);
4032 case nir_op_umin
: return LLVMConstInt(ctx
->i16
, UINT16_MAX
, 0);
4033 case nir_op_fmin
: return LLVMConstReal(ctx
->f16
, INFINITY
);
4034 case nir_op_imax
: return LLVMConstInt(ctx
->i16
, INT16_MIN
, 0);
4035 case nir_op_umax
: return ctx
->i16_0
;
4036 case nir_op_fmax
: return LLVMConstReal(ctx
->f16
, -INFINITY
);
4037 case nir_op_iand
: return LLVMConstInt(ctx
->i16
, -1, 0);
4038 case nir_op_ior
: return ctx
->i16_0
;
4039 case nir_op_ixor
: return ctx
->i16_0
;
4041 unreachable("bad reduction intrinsic");
4043 } else if (type_size
== 4) {
4045 case nir_op_iadd
: return ctx
->i32_0
;
4046 case nir_op_fadd
: return ctx
->f32_0
;
4047 case nir_op_imul
: return ctx
->i32_1
;
4048 case nir_op_fmul
: return ctx
->f32_1
;
4049 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
4050 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
4051 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
4052 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
4053 case nir_op_umax
: return ctx
->i32_0
;
4054 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
4055 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
4056 case nir_op_ior
: return ctx
->i32_0
;
4057 case nir_op_ixor
: return ctx
->i32_0
;
4059 unreachable("bad reduction intrinsic");
4061 } else { /* type_size == 64bit */
4063 case nir_op_iadd
: return ctx
->i64_0
;
4064 case nir_op_fadd
: return ctx
->f64_0
;
4065 case nir_op_imul
: return ctx
->i64_1
;
4066 case nir_op_fmul
: return ctx
->f64_1
;
4067 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
4068 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4069 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
4070 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
4071 case nir_op_umax
: return ctx
->i64_0
;
4072 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
4073 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
4074 case nir_op_ior
: return ctx
->i64_0
;
4075 case nir_op_ixor
: return ctx
->i64_0
;
4077 unreachable("bad reduction intrinsic");
4083 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
4085 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
4086 bool _32bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 4;
4088 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
4089 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
4090 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
4091 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
4092 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
4093 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
4095 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
4096 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
4098 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
4099 _64bit
? "llvm.minnum.f64" : _32bit
? "llvm.minnum.f32" : "llvm.minnum.f16",
4100 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4101 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4102 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
4103 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
4105 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
4106 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
4108 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
4109 _64bit
? "llvm.maxnum.f64" : _32bit
? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4110 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4111 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4112 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
4113 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
4114 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
4116 unreachable("bad reduction intrinsic");
4121 * \param src The value to shift.
4122 * \param identity The value to use the first lane.
4123 * \param maxprefix specifies that the result only needs to be correct for a
4124 * prefix of this many threads
4125 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4128 ac_wavefront_shift_right_1(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4129 LLVMValueRef identity
, unsigned maxprefix
)
4131 if (ctx
->chip_class
>= GFX10
) {
4132 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4133 LLVMValueRef active
, tmp1
, tmp2
;
4134 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4136 tmp1
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4138 tmp2
= ac_build_permlane16(ctx
, src
, (uint64_t)~0, true, false);
4140 if (maxprefix
> 32) {
4141 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4142 LLVMConstInt(ctx
->i32
, 32, false), "");
4144 tmp2
= LLVMBuildSelect(ctx
->builder
, active
,
4145 ac_build_readlane(ctx
, src
,
4146 LLVMConstInt(ctx
->i32
, 31, false)),
4149 active
= LLVMBuildOr(ctx
->builder
, active
,
4150 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4151 LLVMBuildAnd(ctx
->builder
, tid
,
4152 LLVMConstInt(ctx
->i32
, 0x1f, false), ""),
4153 LLVMConstInt(ctx
->i32
, 0x10, false), ""), "");
4154 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4155 } else if (maxprefix
> 16) {
4156 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4157 LLVMConstInt(ctx
->i32
, 16, false), "");
4159 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4161 } else if (ctx
->chip_class
>= GFX8
) {
4162 return ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
4165 /* wavefront shift_right by 1 on SI/CI */
4166 LLVMValueRef active
, tmp1
, tmp2
;
4167 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4168 tmp1
= ac_build_ds_swizzle(ctx
, src
, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4169 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4170 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4171 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x7, 0), ""),
4172 LLVMConstInt(ctx
->i32
, 0x4, 0), "");
4173 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4174 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4175 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4176 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0xf, 0), ""),
4177 LLVMConstInt(ctx
->i32
, 0x8, 0), "");
4178 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4179 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4180 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4181 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x1f, 0), ""),
4182 LLVMConstInt(ctx
->i32
, 0x10, 0), "");
4183 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4184 tmp2
= ac_build_readlane(ctx
, src
, LLVMConstInt(ctx
->i32
, 31, 0));
4185 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), "");
4186 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4187 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 0, 0), "");
4188 return LLVMBuildSelect(ctx
->builder
, active
, identity
, tmp1
, "");
4192 * \param maxprefix specifies that the result only needs to be correct for a
4193 * prefix of this many threads
4196 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
4197 unsigned maxprefix
, bool inclusive
)
4199 LLVMValueRef result
, tmp
;
4202 src
= ac_wavefront_shift_right_1(ctx
, src
, identity
, maxprefix
);
4206 if (ctx
->chip_class
<= GFX7
) {
4207 assert(maxprefix
== 64);
4208 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4209 LLVMValueRef active
;
4210 tmp
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4211 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4212 LLVMBuildAnd(ctx
->builder
, tid
, ctx
->i32_1
, ""),
4214 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4215 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4216 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4217 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4218 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 2, 0), ""),
4220 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4221 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4222 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4223 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4224 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 4, 0), ""),
4226 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4227 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4228 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4229 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4230 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 8, 0), ""),
4232 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4233 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4234 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4235 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4236 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 16, 0), ""),
4238 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4239 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4240 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, 0));
4241 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4242 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), ""),
4244 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4245 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4251 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4252 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4255 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
4256 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4259 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
4260 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4263 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
4264 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4267 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
4268 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4269 if (maxprefix
<= 16)
4272 if (ctx
->chip_class
>= GFX10
) {
4273 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4274 LLVMValueRef active
;
4276 tmp
= ac_build_permlane16(ctx
, result
, ~(uint64_t)0, true, false);
4278 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4279 LLVMBuildAnd(ctx
->builder
, tid
,
4280 LLVMConstInt(ctx
->i32
, 16, false), ""),
4283 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4285 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4287 if (maxprefix
<= 32)
4290 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4292 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, tid
,
4293 LLVMConstInt(ctx
->i32
, 32, false), "");
4295 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4297 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4301 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4302 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4303 if (maxprefix
<= 32)
4305 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4306 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4311 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4313 LLVMValueRef result
;
4315 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4316 LLVMBuilderRef builder
= ctx
->builder
;
4317 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4318 result
= ac_build_ballot(ctx
, src
);
4319 result
= ac_build_mbcnt(ctx
, result
);
4320 result
= LLVMBuildAdd(builder
, result
, src
, "");
4324 ac_build_optimization_barrier(ctx
, &src
);
4326 LLVMValueRef identity
=
4327 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4328 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4329 LLVMTypeOf(identity
), "");
4330 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, true);
4332 return ac_build_wwm(ctx
, result
);
4336 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4338 LLVMValueRef result
;
4340 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4341 LLVMBuilderRef builder
= ctx
->builder
;
4342 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4343 result
= ac_build_ballot(ctx
, src
);
4344 result
= ac_build_mbcnt(ctx
, result
);
4348 ac_build_optimization_barrier(ctx
, &src
);
4350 LLVMValueRef identity
=
4351 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4352 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4353 LLVMTypeOf(identity
), "");
4354 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, false);
4356 return ac_build_wwm(ctx
, result
);
4360 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4362 if (cluster_size
== 1) return src
;
4363 ac_build_optimization_barrier(ctx
, &src
);
4364 LLVMValueRef result
, swap
;
4365 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4366 ac_get_type_size(LLVMTypeOf(src
)));
4367 result
= LLVMBuildBitCast(ctx
->builder
,
4368 ac_build_set_inactive(ctx
, src
, identity
),
4369 LLVMTypeOf(identity
), "");
4370 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4371 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4372 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4374 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4375 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4376 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4378 if (ctx
->chip_class
>= GFX8
)
4379 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4381 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4382 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4383 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4385 if (ctx
->chip_class
>= GFX8
)
4386 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4388 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4389 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4390 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4392 if (ctx
->chip_class
>= GFX10
)
4393 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4394 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4395 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4397 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4398 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4399 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4401 if (ctx
->chip_class
>= GFX8
) {
4402 if (ctx
->wave_size
== 64) {
4403 if (ctx
->chip_class
>= GFX10
)
4404 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4406 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4407 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4408 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4411 return ac_build_wwm(ctx
, result
);
4413 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4414 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4415 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4416 return ac_build_wwm(ctx
, result
);
4421 * "Top half" of a scan that reduces per-wave values across an entire
4424 * The source value must be present in the highest lane of the wave, and the
4425 * highest lane must be live.
4428 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4430 if (ws
->maxwaves
<= 1)
4433 const LLVMValueRef last_lane
= LLVMConstInt(ctx
->i32
, ctx
->wave_size
- 1, false);
4434 LLVMBuilderRef builder
= ctx
->builder
;
4435 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4438 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, last_lane
, "");
4439 ac_build_ifcc(ctx
, tmp
, 1000);
4440 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4441 ac_build_endif(ctx
, 1000);
4445 * "Bottom half" of a scan that reduces per-wave values across an entire
4448 * The caller must place a barrier between the top and bottom halves.
4451 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4453 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4454 const LLVMValueRef identity
=
4455 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4457 if (ws
->maxwaves
<= 1) {
4458 ws
->result_reduce
= ws
->src
;
4459 ws
->result_inclusive
= ws
->src
;
4460 ws
->result_exclusive
= identity
;
4463 assert(ws
->maxwaves
<= 32);
4465 LLVMBuilderRef builder
= ctx
->builder
;
4466 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4467 LLVMBasicBlockRef bbs
[2];
4468 LLVMValueRef phivalues_scan
[2];
4469 LLVMValueRef tmp
, tmp2
;
4471 bbs
[0] = LLVMGetInsertBlock(builder
);
4472 phivalues_scan
[0] = LLVMGetUndef(type
);
4474 if (ws
->enable_reduce
)
4475 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4476 else if (ws
->enable_inclusive
)
4477 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4479 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4480 ac_build_ifcc(ctx
, tmp
, 1001);
4482 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4484 ac_build_optimization_barrier(ctx
, &tmp
);
4486 bbs
[1] = LLVMGetInsertBlock(builder
);
4487 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4489 ac_build_endif(ctx
, 1001);
4491 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4493 if (ws
->enable_reduce
) {
4494 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4495 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4497 if (ws
->enable_inclusive
)
4498 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4499 if (ws
->enable_exclusive
) {
4500 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4501 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4502 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4503 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4508 * Inclusive scan of a per-wave value across an entire workgroup.
4510 * This implies an s_barrier instruction.
4512 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4513 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4514 * useful manner because of the barrier in the algorithm.)
4517 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4519 ac_build_wg_wavescan_top(ctx
, ws
);
4520 ac_build_s_barrier(ctx
);
4521 ac_build_wg_wavescan_bottom(ctx
, ws
);
4525 * "Top half" of a scan that reduces per-thread values across an entire
4528 * All lanes must be active when this code runs.
4531 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4533 if (ws
->enable_exclusive
) {
4534 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4535 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4536 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4537 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4539 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4542 bool enable_inclusive
= ws
->enable_inclusive
;
4543 bool enable_exclusive
= ws
->enable_exclusive
;
4544 ws
->enable_inclusive
= false;
4545 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4546 ac_build_wg_wavescan_top(ctx
, ws
);
4547 ws
->enable_inclusive
= enable_inclusive
;
4548 ws
->enable_exclusive
= enable_exclusive
;
4552 * "Bottom half" of a scan that reduces per-thread values across an entire
4555 * The caller must place a barrier between the top and bottom halves.
4558 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4560 bool enable_inclusive
= ws
->enable_inclusive
;
4561 bool enable_exclusive
= ws
->enable_exclusive
;
4562 ws
->enable_inclusive
= false;
4563 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4564 ac_build_wg_wavescan_bottom(ctx
, ws
);
4565 ws
->enable_inclusive
= enable_inclusive
;
4566 ws
->enable_exclusive
= enable_exclusive
;
4568 /* ws->result_reduce is already the correct value */
4569 if (ws
->enable_inclusive
)
4570 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4571 if (ws
->enable_exclusive
)
4572 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4576 * A scan that reduces per-thread values across an entire workgroup.
4578 * The caller must ensure that all lanes are active when this code runs
4579 * (WWM is insufficient!), because there is an implied barrier.
4582 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4584 ac_build_wg_scan_top(ctx
, ws
);
4585 ac_build_s_barrier(ctx
);
4586 ac_build_wg_scan_bottom(ctx
, ws
);
4590 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4591 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4593 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4594 if (ctx
->chip_class
>= GFX8
) {
4595 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4597 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4602 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4604 LLVMTypeRef type
= LLVMTypeOf(src
);
4605 LLVMValueRef result
;
4607 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4608 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
4610 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4611 (LLVMValueRef
[]) {index
, src
}, 2,
4612 AC_FUNC_ATTR_READNONE
|
4613 AC_FUNC_ATTR_CONVERGENT
);
4614 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
4618 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4624 if (bitsize
== 16) {
4625 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4627 } else if (bitsize
== 32) {
4628 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4631 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4635 LLVMValueRef params
[] = {
4638 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4639 AC_FUNC_ATTR_READNONE
);
4642 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4648 if (bitsize
== 16) {
4649 intr
= "llvm.amdgcn.frexp.mant.f16";
4651 } else if (bitsize
== 32) {
4652 intr
= "llvm.amdgcn.frexp.mant.f32";
4655 intr
= "llvm.amdgcn.frexp.mant.f64";
4659 LLVMValueRef params
[] = {
4662 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4663 AC_FUNC_ATTR_READNONE
);
4667 ac_build_canonicalize(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4673 if (bitsize
== 16) {
4674 intr
= "llvm.canonicalize.f16";
4676 } else if (bitsize
== 32) {
4677 intr
= "llvm.canonicalize.f32";
4680 intr
= "llvm.canonicalize.f64";
4684 LLVMValueRef params
[] = {
4687 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4688 AC_FUNC_ATTR_READNONE
);
4692 * this takes an I,J coordinate pair,
4693 * and works out the X and Y derivatives.
4694 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4697 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4699 LLVMValueRef result
[4], a
;
4702 for (i
= 0; i
< 2; i
++) {
4703 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4704 LLVMConstInt(ctx
->i32
, i
, false), "");
4705 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4706 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4708 return ac_build_gather_values(ctx
, result
, 4);
4712 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4714 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4716 AC_FUNC_ATTR_READNONE
);
4717 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4718 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4722 ac_build_is_helper_invocation(struct ac_llvm_context
*ctx
)
4724 if (!ctx
->postponed_kill
)
4725 return ac_build_load_helper_invocation(ctx
);
4727 /* !(exact && postponed) */
4728 LLVMValueRef exact
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4730 AC_FUNC_ATTR_READNONE
);
4732 LLVMValueRef postponed
= LLVMBuildLoad(ctx
->builder
, ctx
->postponed_kill
, "");
4733 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, exact
, postponed
, "");
4735 return LLVMBuildSelect(ctx
->builder
, result
, ctx
->i32_0
,
4736 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false), "");
4739 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4740 LLVMValueRef
*args
, unsigned num_args
)
4742 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4743 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));
4748 ac_export_mrt_z(struct ac_llvm_context
*ctx
, LLVMValueRef depth
,
4749 LLVMValueRef stencil
, LLVMValueRef samplemask
,
4750 struct ac_export_args
*args
)
4753 unsigned format
= ac_get_spi_shader_z_format(depth
!= NULL
,
4755 samplemask
!= NULL
);
4757 assert(depth
|| stencil
|| samplemask
);
4759 memset(args
, 0, sizeof(*args
));
4761 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
4762 args
->done
= 1; /* DONE bit */
4764 /* Specify the target we are exporting */
4765 args
->target
= V_008DFC_SQ_EXP_MRTZ
;
4767 args
->compr
= 0; /* COMP flag */
4768 args
->out
[0] = LLVMGetUndef(ctx
->f32
); /* R, depth */
4769 args
->out
[1] = LLVMGetUndef(ctx
->f32
); /* G, stencil test val[0:7], stencil op val[8:15] */
4770 args
->out
[2] = LLVMGetUndef(ctx
->f32
); /* B, sample mask */
4771 args
->out
[3] = LLVMGetUndef(ctx
->f32
); /* A, alpha to mask */
4773 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
4775 args
->compr
= 1; /* COMPR flag */
4778 /* Stencil should be in X[23:16]. */
4779 stencil
= ac_to_integer(ctx
, stencil
);
4780 stencil
= LLVMBuildShl(ctx
->builder
, stencil
,
4781 LLVMConstInt(ctx
->i32
, 16, 0), "");
4782 args
->out
[0] = ac_to_float(ctx
, stencil
);
4786 /* SampleMask should be in Y[15:0]. */
4787 args
->out
[1] = samplemask
;
4792 args
->out
[0] = depth
;
4796 args
->out
[1] = stencil
;
4800 args
->out
[2] = samplemask
;
4805 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4806 * at the X writemask component. */
4807 if (ctx
->chip_class
== GFX6
&&
4808 ctx
->family
!= CHIP_OLAND
&&
4809 ctx
->family
!= CHIP_HAINAN
)
4812 /* Specify which components to enable */
4813 args
->enabled_channels
= mask
;
4816 /* Send GS Alloc Req message from the first wave of the group to SPI.
4817 * Message payload is:
4818 * - bits 0..10: vertices in group
4819 * - bits 12..22: primitives in group
4821 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context
*ctx
, LLVMValueRef wave_id
,
4822 LLVMValueRef vtx_cnt
, LLVMValueRef prim_cnt
)
4824 LLVMBuilderRef builder
= ctx
->builder
;
4826 bool export_dummy_prim
= false;
4828 /* HW workaround for a GPU hang with 100% culling.
4829 * We always have to export at least 1 primitive.
4830 * Export a degenerate triangle using vertex 0 for all 3 vertices.
4832 if (prim_cnt
== ctx
->i32_0
&& ctx
->chip_class
== GFX10
) {
4833 assert(vtx_cnt
== ctx
->i32_0
);
4834 prim_cnt
= ctx
->i32_1
;
4835 vtx_cnt
= ctx
->i32_1
;
4836 export_dummy_prim
= true;
4839 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, wave_id
, ctx
->i32_0
, ""), 5020);
4841 tmp
= LLVMBuildShl(builder
, prim_cnt
, LLVMConstInt(ctx
->i32
, 12, false),"");
4842 tmp
= LLVMBuildOr(builder
, tmp
, vtx_cnt
, "");
4843 ac_build_sendmsg(ctx
, AC_SENDMSG_GS_ALLOC_REQ
, tmp
);
4845 if (export_dummy_prim
) {
4846 struct ac_ngg_prim prim
= {};
4847 /* The vertex indices are 0,0,0. */
4848 prim
.passthrough
= ctx
->i32_0
;
4850 struct ac_export_args pos
= {};
4851 pos
.out
[0] = pos
.out
[1] = pos
.out
[2] = pos
.out
[3] = ctx
->f32_0
;
4852 pos
.target
= V_008DFC_SQ_EXP_POS
;
4853 pos
.enabled_channels
= 0xf;
4856 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(ctx
),
4857 ctx
->i32_0
, ""), 5021);
4858 ac_build_export_prim(ctx
, &prim
);
4859 ac_build_export(ctx
, &pos
);
4860 ac_build_endif(ctx
, 5021);
4863 ac_build_endif(ctx
, 5020);
4866 LLVMValueRef
ac_pack_prim_export(struct ac_llvm_context
*ctx
,
4867 const struct ac_ngg_prim
*prim
)
4869 /* The prim export format is:
4870 * - bits 0..8: index 0
4871 * - bit 9: edge flag 0
4872 * - bits 10..18: index 1
4873 * - bit 19: edge flag 1
4874 * - bits 20..28: index 2
4875 * - bit 29: edge flag 2
4876 * - bit 31: null primitive (skip)
4878 LLVMBuilderRef builder
= ctx
->builder
;
4879 LLVMValueRef tmp
= LLVMBuildZExt(builder
, prim
->isnull
, ctx
->i32
, "");
4880 LLVMValueRef result
= LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->i32
, 31, false), "");
4882 for (unsigned i
= 0; i
< prim
->num_vertices
; ++i
) {
4883 tmp
= LLVMBuildShl(builder
, prim
->index
[i
],
4884 LLVMConstInt(ctx
->i32
, 10 * i
, false), "");
4885 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4886 tmp
= LLVMBuildZExt(builder
, prim
->edgeflag
[i
], ctx
->i32
, "");
4887 tmp
= LLVMBuildShl(builder
, tmp
,
4888 LLVMConstInt(ctx
->i32
, 10 * i
+ 9, false), "");
4889 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4894 void ac_build_export_prim(struct ac_llvm_context
*ctx
,
4895 const struct ac_ngg_prim
*prim
)
4897 struct ac_export_args args
;
4899 if (prim
->passthrough
) {
4900 args
.out
[0] = prim
->passthrough
;
4902 args
.out
[0] = ac_pack_prim_export(ctx
, prim
);
4905 args
.out
[0] = LLVMBuildBitCast(ctx
->builder
, args
.out
[0], ctx
->f32
, "");
4906 args
.out
[1] = LLVMGetUndef(ctx
->f32
);
4907 args
.out
[2] = LLVMGetUndef(ctx
->f32
);
4908 args
.out
[3] = LLVMGetUndef(ctx
->f32
);
4910 args
.target
= V_008DFC_SQ_EXP_PRIM
;
4911 args
.enabled_channels
= 1;
4913 args
.valid_mask
= false;
4916 ac_build_export(ctx
, &args
);
4920 arg_llvm_type(enum ac_arg_type type
, unsigned size
, struct ac_llvm_context
*ctx
)
4922 if (type
== AC_ARG_FLOAT
) {
4923 return size
== 1 ? ctx
->f32
: LLVMVectorType(ctx
->f32
, size
);
4924 } else if (type
== AC_ARG_INT
) {
4925 return size
== 1 ? ctx
->i32
: LLVMVectorType(ctx
->i32
, size
);
4927 LLVMTypeRef ptr_type
;
4929 case AC_ARG_CONST_PTR
:
4932 case AC_ARG_CONST_FLOAT_PTR
:
4933 ptr_type
= ctx
->f32
;
4935 case AC_ARG_CONST_PTR_PTR
:
4936 ptr_type
= ac_array_in_const32_addr_space(ctx
->i8
);
4938 case AC_ARG_CONST_DESC_PTR
:
4939 ptr_type
= ctx
->v4i32
;
4941 case AC_ARG_CONST_IMAGE_PTR
:
4942 ptr_type
= ctx
->v8i32
;
4945 unreachable("unknown arg type");
4948 return ac_array_in_const32_addr_space(ptr_type
);
4951 return ac_array_in_const_addr_space(ptr_type
);
4957 ac_build_main(const struct ac_shader_args
*args
,
4958 struct ac_llvm_context
*ctx
,
4959 enum ac_llvm_calling_convention convention
,
4960 const char *name
, LLVMTypeRef ret_type
,
4961 LLVMModuleRef module
)
4963 LLVMTypeRef arg_types
[AC_MAX_ARGS
];
4965 for (unsigned i
= 0; i
< args
->arg_count
; i
++) {
4966 arg_types
[i
] = arg_llvm_type(args
->args
[i
].type
,
4967 args
->args
[i
].size
, ctx
);
4970 LLVMTypeRef main_function_type
=
4971 LLVMFunctionType(ret_type
, arg_types
, args
->arg_count
, 0);
4973 LLVMValueRef main_function
=
4974 LLVMAddFunction(module
, name
, main_function_type
);
4975 LLVMBasicBlockRef main_function_body
=
4976 LLVMAppendBasicBlockInContext(ctx
->context
, main_function
, "main_body");
4977 LLVMPositionBuilderAtEnd(ctx
->builder
, main_function_body
);
4979 LLVMSetFunctionCallConv(main_function
, convention
);
4980 for (unsigned i
= 0; i
< args
->arg_count
; ++i
) {
4981 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
4983 if (args
->args
[i
].file
!= AC_ARG_SGPR
)
4986 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
4988 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4989 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
4990 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4991 ac_add_attr_alignment(P
, 32);
4995 ctx
->main_function
= main_function
;
4997 if (LLVM_VERSION_MAJOR
>= 11) {
4998 /* Enable denormals for FP16 and FP64: */
4999 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math",
5001 /* Disable denormals for FP32: */
5002 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math-f32",
5003 "preserve-sign,preserve-sign");
5005 return main_function
;
5008 void ac_build_s_endpgm(struct ac_llvm_context
*ctx
)
5010 LLVMTypeRef calltype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
5011 LLVMValueRef code
= LLVMConstInlineAsm(calltype
, "s_endpgm", "", true, false);
5012 LLVMBuildCall(ctx
->builder
, code
, NULL
, 0, "");
5015 LLVMValueRef
ac_prefix_bitcount(struct ac_llvm_context
*ctx
,
5016 LLVMValueRef mask
, LLVMValueRef index
)
5018 LLVMBuilderRef builder
= ctx
->builder
;
5019 LLVMTypeRef type
= LLVMTypeOf(mask
);
5021 LLVMValueRef bit
= LLVMBuildShl(builder
, LLVMConstInt(type
, 1, 0),
5022 LLVMBuildZExt(builder
, index
, type
, ""), "");
5023 LLVMValueRef prefix_bits
= LLVMBuildSub(builder
, bit
, LLVMConstInt(type
, 1, 0), "");
5024 LLVMValueRef prefix_mask
= LLVMBuildAnd(builder
, mask
, prefix_bits
, "");
5025 return ac_build_bit_count(ctx
, prefix_mask
);
5028 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
5029 LLVMValueRef
ac_prefix_bitcount_2x64(struct ac_llvm_context
*ctx
,
5030 LLVMValueRef mask
[2], LLVMValueRef index
)
5032 LLVMBuilderRef builder
= ctx
->builder
;
5034 /* Reference version using i128. */
5035 LLVMValueRef input_mask
=
5036 LLVMBuildBitCast(builder
, ac_build_gather_values(ctx
, mask
, 2), ctx
->i128
, "");
5038 return ac_prefix_bitcount(ctx
, input_mask
, index
);
5040 /* Optimized version using 2 64-bit masks. */
5041 LLVMValueRef is_hi
, is_0
, c64
, c128
, all_bits
;
5042 LLVMValueRef prefix_mask
[2], shift
[2], mask_bcnt0
, prefix_bcnt
[2];
5044 /* Compute the 128-bit prefix mask. */
5045 c64
= LLVMConstInt(ctx
->i32
, 64, 0);
5046 c128
= LLVMConstInt(ctx
->i32
, 128, 0);
5047 all_bits
= LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
5048 /* The first index that can have non-zero high bits in the prefix mask is 65. */
5049 is_hi
= LLVMBuildICmp(builder
, LLVMIntUGT
, index
, c64
, "");
5050 is_0
= LLVMBuildICmp(builder
, LLVMIntEQ
, index
, ctx
->i32_0
, "");
5051 mask_bcnt0
= ac_build_bit_count(ctx
, mask
[0]);
5053 for (unsigned i
= 0; i
< 2; i
++) {
5054 shift
[i
] = LLVMBuildSub(builder
, i
? c128
: c64
, index
, "");
5055 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
5056 * so we handle it by the is_0 select.
5057 * For i==1, index==64, same story, so we handle it by the last is_hi select.
5058 * For i==0, index==64, we shift by 0, which is what we want.
5060 prefix_mask
[i
] = LLVMBuildLShr(builder
, all_bits
,
5061 LLVMBuildZExt(builder
, shift
[i
], ctx
->i64
, ""), "");
5062 prefix_mask
[i
] = LLVMBuildAnd(builder
, mask
[i
], prefix_mask
[i
], "");
5063 prefix_bcnt
[i
] = ac_build_bit_count(ctx
, prefix_mask
[i
]);
5066 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_0
, ctx
->i32_0
, prefix_bcnt
[0], "");
5067 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_hi
, mask_bcnt0
, prefix_bcnt
[0], "");
5068 prefix_bcnt
[1] = LLVMBuildSelect(builder
, is_hi
, prefix_bcnt
[1], ctx
->i32_0
, "");
5070 return LLVMBuildAdd(builder
, prefix_bcnt
[0], prefix_bcnt
[1], "");
5075 * Convert triangle strip indices to triangle indices. This is used to decompose
5076 * triangle strips into triangles.
5078 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context
*ctx
,
5079 LLVMValueRef is_odd
,
5080 LLVMValueRef flatshade_first
,
5081 LLVMValueRef index
[3])
5083 LLVMBuilderRef builder
= ctx
->builder
;
5084 LLVMValueRef out
[3];
5086 /* We need to change the vertex order for odd triangles to get correct
5087 * front/back facing by swapping 2 vertex indices, but we also have to
5088 * keep the provoking vertex in the same place.
5090 * If the first vertex is provoking, swap index 1 and 2.
5091 * If the last vertex is provoking, swap index 0 and 1.
5093 out
[0] = LLVMBuildSelect(builder
, flatshade_first
,
5095 LLVMBuildSelect(builder
, is_odd
,
5096 index
[1], index
[0], ""), "");
5097 out
[1] = LLVMBuildSelect(builder
, flatshade_first
,
5098 LLVMBuildSelect(builder
, is_odd
,
5099 index
[2], index
[1], ""),
5100 LLVMBuildSelect(builder
, is_odd
,
5101 index
[0], index
[1], ""), "");
5102 out
[2] = LLVMBuildSelect(builder
, flatshade_first
,
5103 LLVMBuildSelect(builder
, is_odd
,
5104 index
[1], index
[2], ""),
5106 memcpy(index
, out
, sizeof(out
));