2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
30 #include "c11/threads.h"
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "util/u_math.h"
43 #include "shader_enums.h"
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block
;
52 LLVMBasicBlockRef loop_entry_block
;
55 /* Initialize module-independent parts of the context.
57 * The caller is responsible for initializing ctx::module and ctx::builder.
60 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
61 struct ac_llvm_compiler
*compiler
,
62 enum chip_class chip_class
, enum radeon_family family
,
63 enum ac_float_mode float_mode
, unsigned wave_size
,
64 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
->module
= ac_create_module(wave_size
== 32 ? compiler
->tm_wave32
77 ctx
->builder
= ac_create_builder(ctx
->context
, float_mode
);
79 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
80 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
81 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
82 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
83 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
84 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
85 ctx
->intptr
= ctx
->i32
;
86 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
87 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
88 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
89 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
90 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
91 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
92 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
93 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
94 ctx
->v3f32
= LLVMVectorType(ctx
->f32
, 3);
95 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
96 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
97 ctx
->iN_wavemask
= LLVMIntTypeInContext(ctx
->context
, ctx
->wave_size
);
98 ctx
->iN_ballotmask
= LLVMIntTypeInContext(ctx
->context
, ballot_mask_bits
);
100 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
101 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
102 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
103 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
104 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
105 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
106 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
107 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
108 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
109 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
110 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
111 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
112 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
113 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
115 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
116 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
118 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
121 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
122 "invariant.load", 14);
124 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
126 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
127 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
129 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
130 "amdgpu.uniform", 14);
132 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
133 ctx
->flow
= calloc(1, sizeof(*ctx
->flow
));
137 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
139 free(ctx
->flow
->stack
);
145 ac_get_llvm_num_components(LLVMValueRef value
)
147 LLVMTypeRef type
= LLVMTypeOf(value
);
148 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
149 ? LLVMGetVectorSize(type
)
151 return num_components
;
155 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
159 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
164 return LLVMBuildExtractElement(ac
->builder
, value
,
165 LLVMConstInt(ac
->i32
, index
, false), "");
169 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
171 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
172 type
= LLVMGetElementType(type
);
174 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
175 return LLVMGetIntTypeWidth(type
);
177 if (type
== ctx
->f16
)
179 if (type
== ctx
->f32
)
181 if (type
== ctx
->f64
)
184 unreachable("Unhandled type kind in get_elem_bits");
188 ac_get_type_size(LLVMTypeRef type
)
190 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
193 case LLVMIntegerTypeKind
:
194 return LLVMGetIntTypeWidth(type
) / 8;
195 case LLVMHalfTypeKind
:
197 case LLVMFloatTypeKind
:
199 case LLVMDoubleTypeKind
:
201 case LLVMPointerTypeKind
:
202 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
205 case LLVMVectorTypeKind
:
206 return LLVMGetVectorSize(type
) *
207 ac_get_type_size(LLVMGetElementType(type
));
208 case LLVMArrayTypeKind
:
209 return LLVMGetArrayLength(type
) *
210 ac_get_type_size(LLVMGetElementType(type
));
217 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
221 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
223 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
225 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
228 unreachable("Unhandled integer size");
232 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
234 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
235 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
236 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
237 LLVMGetVectorSize(t
));
239 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
240 switch (LLVMGetPointerAddressSpace(t
)) {
241 case AC_ADDR_SPACE_GLOBAL
:
243 case AC_ADDR_SPACE_LDS
:
246 unreachable("unhandled address space");
249 return to_integer_type_scalar(ctx
, t
);
253 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
255 LLVMTypeRef type
= LLVMTypeOf(v
);
256 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
257 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
259 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
263 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
265 LLVMTypeRef type
= LLVMTypeOf(v
);
266 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
268 return ac_to_integer(ctx
, v
);
271 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
275 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
277 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
279 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
282 unreachable("Unhandled float size");
286 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
288 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
289 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
290 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
291 LLVMGetVectorSize(t
));
293 return to_float_type_scalar(ctx
, t
);
297 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
299 LLVMTypeRef type
= LLVMTypeOf(v
);
300 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
305 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
306 LLVMTypeRef return_type
, LLVMValueRef
*params
,
307 unsigned param_count
, unsigned attrib_mask
)
309 LLVMValueRef function
, call
;
310 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
312 function
= LLVMGetNamedFunction(ctx
->module
, name
);
314 LLVMTypeRef param_types
[32], function_type
;
317 assert(param_count
<= 32);
319 for (i
= 0; i
< param_count
; ++i
) {
321 param_types
[i
] = LLVMTypeOf(params
[i
]);
324 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
325 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
327 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
328 LLVMSetLinkage(function
, LLVMExternalLinkage
);
330 if (!set_callsite_attrs
)
331 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
334 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
335 if (set_callsite_attrs
)
336 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
341 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
344 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
346 LLVMTypeRef elem_type
= type
;
348 assert(bufsize
>= 8);
350 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
351 int ret
= snprintf(buf
, bufsize
, "v%u",
352 LLVMGetVectorSize(type
));
354 char *type_name
= LLVMPrintTypeToString(type
);
355 fprintf(stderr
, "Error building type name for: %s\n",
357 LLVMDisposeMessage(type_name
);
360 elem_type
= LLVMGetElementType(type
);
364 switch (LLVMGetTypeKind(elem_type
)) {
366 case LLVMIntegerTypeKind
:
367 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
369 case LLVMHalfTypeKind
:
370 snprintf(buf
, bufsize
, "f16");
372 case LLVMFloatTypeKind
:
373 snprintf(buf
, bufsize
, "f32");
375 case LLVMDoubleTypeKind
:
376 snprintf(buf
, bufsize
, "f64");
382 * Helper function that builds an LLVM IR PHI node and immediately adds
386 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
387 unsigned count_incoming
, LLVMValueRef
*values
,
388 LLVMBasicBlockRef
*blocks
)
390 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
391 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
395 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
397 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
398 0, AC_FUNC_ATTR_CONVERGENT
);
401 /* Prevent optimizations (at least of memory accesses) across the current
402 * point in the program by emitting empty inline assembly that is marked as
403 * having side effects.
405 * Optionally, a value can be passed through the inline assembly to prevent
406 * LLVM from hoisting calls to ReadNone functions.
409 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
412 static int counter
= 0;
414 LLVMBuilderRef builder
= ctx
->builder
;
417 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
420 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
421 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
422 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
424 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
425 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
426 LLVMValueRef vgpr
= *pvgpr
;
427 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
428 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
431 assert(vgpr_size
% 4 == 0);
433 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
434 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
435 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
436 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
437 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
444 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
446 const char *intr
= HAVE_LLVM
>= 0x0900 && ctx
->chip_class
>= GFX8
?
447 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
448 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, intr
, ctx
->i64
, NULL
, 0, 0);
449 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
453 ac_build_ballot(struct ac_llvm_context
*ctx
,
458 if (HAVE_LLVM
>= 0x900) {
459 if (ctx
->wave_size
== 64)
460 name
= "llvm.amdgcn.icmp.i64.i32";
462 name
= "llvm.amdgcn.icmp.i32.i32";
464 name
= "llvm.amdgcn.icmp.i32";
466 LLVMValueRef args
[3] = {
469 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
472 /* We currently have no other way to prevent LLVM from lifting the icmp
473 * calls to a dominating basic block.
475 ac_build_optimization_barrier(ctx
, &args
[0]);
477 args
[0] = ac_to_integer(ctx
, args
[0]);
479 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
480 AC_FUNC_ATTR_NOUNWIND
|
481 AC_FUNC_ATTR_READNONE
|
482 AC_FUNC_ATTR_CONVERGENT
);
485 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
488 const char *name
= HAVE_LLVM
>= 0x900 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
489 LLVMValueRef args
[3] = {
492 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0),
495 return ac_build_intrinsic(ctx
, name
, ctx
->i64
, args
, 3,
496 AC_FUNC_ATTR_NOUNWIND
|
497 AC_FUNC_ATTR_READNONE
|
498 AC_FUNC_ATTR_CONVERGENT
);
502 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
504 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
505 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
506 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
510 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
512 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
513 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
514 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
518 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
520 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
521 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
523 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
524 vote_set
, active_set
, "");
525 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
527 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
528 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
532 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
533 unsigned value_count
, unsigned component
)
535 LLVMValueRef vec
= NULL
;
537 if (value_count
== 1) {
538 return values
[component
];
539 } else if (!value_count
)
540 unreachable("value_count is 0");
542 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
543 LLVMValueRef value
= values
[i
];
546 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
547 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
548 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
554 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
555 LLVMValueRef
*values
,
556 unsigned value_count
,
557 unsigned value_stride
,
561 LLVMBuilderRef builder
= ctx
->builder
;
562 LLVMValueRef vec
= NULL
;
565 if (value_count
== 1 && !always_vector
) {
567 return LLVMBuildLoad(builder
, values
[0], "");
569 } else if (!value_count
)
570 unreachable("value_count is 0");
572 for (i
= 0; i
< value_count
; i
++) {
573 LLVMValueRef value
= values
[i
* value_stride
];
575 value
= LLVMBuildLoad(builder
, value
, "");
578 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
579 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
580 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
586 ac_build_gather_values(struct ac_llvm_context
*ctx
,
587 LLVMValueRef
*values
,
588 unsigned value_count
)
590 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
593 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
594 * channels with undef. Extract at most src_channels components from the input.
597 ac_build_expand(struct ac_llvm_context
*ctx
,
599 unsigned src_channels
,
600 unsigned dst_channels
)
602 LLVMTypeRef elemtype
;
603 LLVMValueRef chan
[dst_channels
];
605 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
606 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
608 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
611 src_channels
= MIN2(src_channels
, vec_size
);
613 for (unsigned i
= 0; i
< src_channels
; i
++)
614 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
616 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
619 assert(src_channels
== 1);
622 elemtype
= LLVMTypeOf(value
);
625 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
626 chan
[i
] = LLVMGetUndef(elemtype
);
628 return ac_build_gather_values(ctx
, chan
, dst_channels
);
631 /* Extract components [start, start + channels) from a vector.
634 ac_extract_components(struct ac_llvm_context
*ctx
,
639 LLVMValueRef chan
[channels
];
641 for (unsigned i
= 0; i
< channels
; i
++)
642 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
+ start
);
644 return ac_build_gather_values(ctx
, chan
, channels
);
647 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
648 * with undef. Extract at most num_channels components from the input.
650 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
652 unsigned num_channels
)
654 return ac_build_expand(ctx
, value
, num_channels
, 4);
657 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
659 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
663 name
= "llvm.rint.f16";
664 else if (type_size
== 4)
665 name
= "llvm.rint.f32";
667 name
= "llvm.rint.f64";
669 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
670 AC_FUNC_ATTR_READNONE
);
674 ac_build_fdiv(struct ac_llvm_context
*ctx
,
678 /* If we do (num / den), LLVM >= 7.0 does:
679 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
681 * If we do (num * (1 / den)), LLVM does:
682 * return num * v_rcp_f32(den);
684 LLVMValueRef one
= LLVMConstReal(LLVMTypeOf(num
), 1.0);
685 LLVMValueRef rcp
= LLVMBuildFDiv(ctx
->builder
, one
, den
, "");
686 LLVMValueRef ret
= LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
688 /* Use v_rcp_f32 instead of precise division. */
689 if (!LLVMIsConstant(ret
))
690 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
694 /* See fast_idiv_by_const.h. */
695 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
696 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
698 LLVMValueRef multiplier
,
699 LLVMValueRef pre_shift
,
700 LLVMValueRef post_shift
,
701 LLVMValueRef increment
)
703 LLVMBuilderRef builder
= ctx
->builder
;
705 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
706 num
= LLVMBuildMul(builder
,
707 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
708 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
709 num
= LLVMBuildAdd(builder
, num
,
710 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
711 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
712 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
713 return LLVMBuildLShr(builder
, num
, post_shift
, "");
716 /* See fast_idiv_by_const.h. */
717 /* If num != UINT_MAX, this more efficient version can be used. */
718 /* Set: increment = util_fast_udiv_info::increment; */
719 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
721 LLVMValueRef multiplier
,
722 LLVMValueRef pre_shift
,
723 LLVMValueRef post_shift
,
724 LLVMValueRef increment
)
726 LLVMBuilderRef builder
= ctx
->builder
;
728 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
729 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
730 num
= LLVMBuildMul(builder
,
731 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
732 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
733 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
734 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
735 return LLVMBuildLShr(builder
, num
, post_shift
, "");
738 /* See fast_idiv_by_const.h. */
739 /* Both operands must fit in 31 bits and the divisor must not be 1. */
740 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
742 LLVMValueRef multiplier
,
743 LLVMValueRef post_shift
)
745 LLVMBuilderRef builder
= ctx
->builder
;
747 num
= LLVMBuildMul(builder
,
748 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
749 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
750 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
751 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
752 return LLVMBuildLShr(builder
, num
, post_shift
, "");
755 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
756 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
757 * already multiplied by two. id is the cube face number.
759 struct cube_selection_coords
{
766 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
768 struct cube_selection_coords
*out
)
770 LLVMTypeRef f32
= ctx
->f32
;
772 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
773 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
774 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
775 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
776 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
777 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
778 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
779 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
783 * Build a manual selection sequence for cube face sc/tc coordinates and
784 * major axis vector (multiplied by 2 for consistency) for the given
785 * vec3 \p coords, for the face implied by \p selcoords.
787 * For the major axis, we always adjust the sign to be in the direction of
788 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
789 * the selcoords major axis.
791 static void build_cube_select(struct ac_llvm_context
*ctx
,
792 const struct cube_selection_coords
*selcoords
,
793 const LLVMValueRef
*coords
,
794 LLVMValueRef
*out_st
,
795 LLVMValueRef
*out_ma
)
797 LLVMBuilderRef builder
= ctx
->builder
;
798 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
799 LLVMValueRef is_ma_positive
;
801 LLVMValueRef is_ma_z
, is_not_ma_z
;
802 LLVMValueRef is_ma_y
;
803 LLVMValueRef is_ma_x
;
807 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
808 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
809 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
810 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
812 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
813 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
814 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
815 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
816 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
819 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
820 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
821 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
822 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
823 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
826 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
827 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
828 LLVMConstReal(f32
, -1.0), "");
829 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
832 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
833 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
834 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
835 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
836 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
840 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
841 bool is_deriv
, bool is_array
, bool is_lod
,
842 LLVMValueRef
*coords_arg
,
843 LLVMValueRef
*derivs_arg
)
846 LLVMBuilderRef builder
= ctx
->builder
;
847 struct cube_selection_coords selcoords
;
848 LLVMValueRef coords
[3];
851 if (is_array
&& !is_lod
) {
852 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
854 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
856 * "For Array forms, the array layer used will be
858 * max(0, min(d−1, floor(layer+0.5)))
860 * where d is the depth of the texture array and layer
861 * comes from the component indicated in the tables below.
862 * Workaroudn for an issue where the layer is taken from a
863 * helper invocation which happens to fall on a different
864 * layer due to extrapolation."
866 * GFX8 and earlier attempt to implement this in hardware by
867 * clamping the value of coords[2] = (8 * layer) + face.
868 * Unfortunately, this means that the we end up with the wrong
869 * face when clamping occurs.
871 * Clamp the layer earlier to work around the issue.
873 if (ctx
->chip_class
<= GFX8
) {
875 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
876 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
882 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
884 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
885 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
886 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
888 for (int i
= 0; i
< 2; ++i
)
889 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
891 coords
[2] = selcoords
.id
;
893 if (is_deriv
&& derivs_arg
) {
894 LLVMValueRef derivs
[4];
897 /* Convert cube derivatives to 2D derivatives. */
898 for (axis
= 0; axis
< 2; axis
++) {
899 LLVMValueRef deriv_st
[2];
900 LLVMValueRef deriv_ma
;
902 /* Transform the derivative alongside the texture
903 * coordinate. Mathematically, the correct formula is
904 * as follows. Assume we're projecting onto the +Z face
905 * and denote by dx/dh the derivative of the (original)
906 * X texture coordinate with respect to horizontal
907 * window coordinates. The projection onto the +Z face
912 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
913 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
915 * This motivatives the implementation below.
917 * Whether this actually gives the expected results for
918 * apps that might feed in derivatives obtained via
919 * finite differences is anyone's guess. The OpenGL spec
920 * seems awfully quiet about how textureGrad for cube
921 * maps should be handled.
923 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
924 deriv_st
, &deriv_ma
);
926 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
928 for (int i
= 0; i
< 2; ++i
)
929 derivs
[axis
* 2 + i
] =
930 LLVMBuildFSub(builder
,
931 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
932 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
935 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
938 /* Shift the texture coordinate. This must be applied after the
939 * derivative calculation.
941 for (int i
= 0; i
< 2; ++i
)
942 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
945 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
946 /* coords_arg.w component - array_index for cube arrays */
947 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
950 memcpy(coords_arg
, coords
, sizeof(coords
));
955 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
956 LLVMValueRef llvm_chan
,
957 LLVMValueRef attr_number
,
962 LLVMValueRef args
[5];
967 args
[2] = attr_number
;
970 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
971 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
976 args
[3] = attr_number
;
979 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
980 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
984 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
985 LLVMValueRef llvm_chan
,
986 LLVMValueRef attr_number
,
991 LLVMValueRef args
[6];
996 args
[2] = attr_number
;
997 args
[3] = ctx
->i1false
;
1000 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
1001 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1005 args
[2] = llvm_chan
;
1006 args
[3] = attr_number
;
1007 args
[4] = ctx
->i1false
;
1010 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
1011 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
1015 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
1016 LLVMValueRef parameter
,
1017 LLVMValueRef llvm_chan
,
1018 LLVMValueRef attr_number
,
1019 LLVMValueRef params
)
1021 LLVMValueRef args
[4];
1023 args
[0] = parameter
;
1024 args
[1] = llvm_chan
;
1025 args
[2] = attr_number
;
1028 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
1029 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1033 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
1034 LLVMValueRef base_ptr
,
1037 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1041 ac_build_gep0(struct ac_llvm_context
*ctx
,
1042 LLVMValueRef base_ptr
,
1045 LLVMValueRef indices
[2] = {
1049 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1052 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1055 return LLVMBuildPointerCast(ctx
->builder
,
1056 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1057 LLVMTypeOf(ptr
), "");
1061 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1062 LLVMValueRef base_ptr
, LLVMValueRef index
,
1065 LLVMBuildStore(ctx
->builder
, value
,
1066 ac_build_gep0(ctx
, base_ptr
, index
));
1070 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1071 * It's equivalent to doing a load from &base_ptr[index].
1073 * \param base_ptr Where the array starts.
1074 * \param index The element index into the array.
1075 * \param uniform Whether the base_ptr and index can be assumed to be
1076 * dynamically uniform (i.e. load to an SGPR)
1077 * \param invariant Whether the load is invariant (no other opcodes affect it)
1078 * \param no_unsigned_wraparound
1079 * For all possible re-associations and re-distributions of an expression
1080 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1081 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1082 * does not result in an unsigned integer wraparound. This is used for
1083 * optimal code generation of 32-bit pointer arithmetic.
1085 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1086 * integer wraparound can't be an imm offset in s_load_dword, because
1087 * the instruction performs "addr + offset" in 64 bits.
1089 * Expected usage for bindless textures by chaining GEPs:
1090 * // possible unsigned wraparound, don't use InBounds:
1091 * ptr1 = LLVMBuildGEP(base_ptr, index);
1092 * image = load(ptr1); // becomes "s_load ptr1, 0"
1094 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1095 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1098 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1099 LLVMValueRef index
, bool uniform
, bool invariant
,
1100 bool no_unsigned_wraparound
)
1102 LLVMValueRef pointer
, result
;
1104 if (no_unsigned_wraparound
&&
1105 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1106 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1108 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1111 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1112 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1114 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1118 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1121 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1124 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1125 LLVMValueRef base_ptr
, LLVMValueRef index
)
1127 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1130 /* This assumes that there is no unsigned integer wraparound during the address
1131 * computation, excluding all GEPs within base_ptr. */
1132 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1133 LLVMValueRef base_ptr
, LLVMValueRef index
)
1135 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1138 /* See ac_build_load_custom() documentation. */
1139 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1140 LLVMValueRef base_ptr
, LLVMValueRef index
)
1142 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1145 static unsigned get_load_cache_policy(struct ac_llvm_context
*ctx
,
1146 unsigned cache_policy
)
1148 return cache_policy
|
1149 (ctx
->chip_class
>= GFX10
&& cache_policy
& ac_glc
? ac_dlc
: 0);
1153 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1156 LLVMValueRef vindex
,
1157 LLVMValueRef voffset
,
1158 LLVMValueRef soffset
,
1159 unsigned num_channels
,
1160 LLVMTypeRef return_channel_type
,
1161 unsigned cache_policy
,
1165 LLVMValueRef args
[6];
1168 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1170 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1171 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1172 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1173 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1174 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1175 const char *indexing_kind
= structurized
? "struct" : "raw";
1176 char name
[256], type_name
[8];
1178 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(return_channel_type
, func
) : return_channel_type
;
1179 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1182 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1183 indexing_kind
, type_name
);
1185 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1186 indexing_kind
, type_name
);
1189 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1190 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1194 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1197 LLVMValueRef vindex
,
1198 LLVMValueRef voffset
,
1199 unsigned num_channels
,
1200 unsigned cache_policy
)
1202 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1203 voffset
, NULL
, num_channels
,
1204 ctx
->f32
, cache_policy
,
1208 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1209 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1210 * or v4i32 (num_channels=3,4).
1213 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1216 unsigned num_channels
,
1217 LLVMValueRef voffset
,
1218 LLVMValueRef soffset
,
1219 unsigned inst_offset
,
1220 unsigned cache_policy
,
1221 bool swizzle_enable_hint
)
1223 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1225 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1226 LLVMValueRef v
[3], v01
;
1228 for (int i
= 0; i
< 3; i
++) {
1229 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1230 LLVMConstInt(ctx
->i32
, i
, 0), "");
1232 v01
= ac_build_gather_values(ctx
, v
, 2);
1234 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1235 soffset
, inst_offset
, cache_policy
,
1236 swizzle_enable_hint
);
1237 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1238 soffset
, inst_offset
+ 8,
1240 swizzle_enable_hint
);
1244 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1245 * (voffset is swizzled, but soffset isn't swizzled).
1246 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1248 if (!swizzle_enable_hint
) {
1249 LLVMValueRef offset
= soffset
;
1252 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1253 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1255 ac_build_buffer_store_common(ctx
, rsrc
, ac_to_float(ctx
, vdata
),
1256 ctx
->i32_0
, voffset
, offset
,
1257 num_channels
, ctx
->f32
,
1258 cache_policy
, false, false);
1262 static const unsigned dfmts
[] = {
1263 V_008F0C_BUF_DATA_FORMAT_32
,
1264 V_008F0C_BUF_DATA_FORMAT_32_32
,
1265 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1266 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1268 unsigned dfmt
= dfmts
[num_channels
- 1];
1269 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1270 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1272 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1273 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
);
1277 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1279 LLVMValueRef vindex
,
1280 LLVMValueRef voffset
,
1281 LLVMValueRef soffset
,
1282 unsigned num_channels
,
1283 LLVMTypeRef channel_type
,
1284 unsigned cache_policy
,
1289 LLVMValueRef args
[5];
1291 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1293 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1294 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1295 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1296 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1297 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1298 const char *indexing_kind
= structurized
? "struct" : "raw";
1299 char name
[256], type_name
[8];
1301 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1302 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1305 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1306 indexing_kind
, type_name
);
1308 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1309 indexing_kind
, type_name
);
1312 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1313 ac_get_load_intr_attribs(can_speculate
));
1317 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1320 LLVMValueRef vindex
,
1321 LLVMValueRef voffset
,
1322 LLVMValueRef soffset
,
1323 unsigned inst_offset
,
1324 unsigned cache_policy
,
1328 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1330 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1332 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1334 if (allow_smem
&& !(cache_policy
& ac_slc
) &&
1335 (!(cache_policy
& ac_glc
) || ctx
->chip_class
>= GFX8
)) {
1336 assert(vindex
== NULL
);
1338 LLVMValueRef result
[8];
1340 for (int i
= 0; i
< num_channels
; i
++) {
1342 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1343 LLVMConstInt(ctx
->i32
, 4, 0), "");
1345 LLVMValueRef args
[3] = {
1348 LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0),
1350 result
[i
] = ac_build_intrinsic(ctx
,
1351 "llvm.amdgcn.s.buffer.load.f32",
1353 AC_FUNC_ATTR_READNONE
);
1355 if (num_channels
== 1)
1358 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1359 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1360 return ac_build_gather_values(ctx
, result
, num_channels
);
1363 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
,
1365 num_channels
, ctx
->f32
,
1367 can_speculate
, false, false);
1370 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1372 LLVMValueRef vindex
,
1373 LLVMValueRef voffset
,
1374 unsigned num_channels
,
1375 unsigned cache_policy
,
1378 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1379 ctx
->i32_0
, num_channels
, ctx
->f32
,
1380 cache_policy
, can_speculate
,
1384 /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
1385 /// value for LLVM8+ tbuffer intrinsics.
1387 ac_get_tbuffer_format(struct ac_llvm_context
*ctx
,
1388 unsigned dfmt
, unsigned nfmt
)
1390 if (ctx
->chip_class
>= GFX10
) {
1393 default: unreachable("bad dfmt");
1394 case V_008F0C_BUF_DATA_FORMAT_INVALID
: format
= V_008F0C_IMG_FORMAT_INVALID
; break;
1395 case V_008F0C_BUF_DATA_FORMAT_8
: format
= V_008F0C_IMG_FORMAT_8_UINT
; break;
1396 case V_008F0C_BUF_DATA_FORMAT_8_8
: format
= V_008F0C_IMG_FORMAT_8_8_UINT
; break;
1397 case V_008F0C_BUF_DATA_FORMAT_8_8_8_8
: format
= V_008F0C_IMG_FORMAT_8_8_8_8_UINT
; break;
1398 case V_008F0C_BUF_DATA_FORMAT_16
: format
= V_008F0C_IMG_FORMAT_16_UINT
; break;
1399 case V_008F0C_BUF_DATA_FORMAT_16_16
: format
= V_008F0C_IMG_FORMAT_16_16_UINT
; break;
1400 case V_008F0C_BUF_DATA_FORMAT_16_16_16_16
: format
= V_008F0C_IMG_FORMAT_16_16_16_16_UINT
; break;
1401 case V_008F0C_BUF_DATA_FORMAT_32
: format
= V_008F0C_IMG_FORMAT_32_UINT
; break;
1402 case V_008F0C_BUF_DATA_FORMAT_32_32
: format
= V_008F0C_IMG_FORMAT_32_32_UINT
; break;
1403 case V_008F0C_BUF_DATA_FORMAT_32_32_32
: format
= V_008F0C_IMG_FORMAT_32_32_32_UINT
; break;
1404 case V_008F0C_BUF_DATA_FORMAT_32_32_32_32
: format
= V_008F0C_IMG_FORMAT_32_32_32_32_UINT
; break;
1405 case V_008F0C_BUF_DATA_FORMAT_2_10_10_10
: format
= V_008F0C_IMG_FORMAT_2_10_10_10_UINT
; break;
1408 // Use the regularity properties of the combined format enum.
1410 // Note: float is incompatible with 8-bit data formats,
1411 // [us]{norm,scaled} are incomparible with 32-bit data formats.
1412 // [us]scaled are not writable.
1414 case V_008F0C_BUF_NUM_FORMAT_UNORM
: format
-= 4; break;
1415 case V_008F0C_BUF_NUM_FORMAT_SNORM
: format
-= 3; break;
1416 case V_008F0C_BUF_NUM_FORMAT_USCALED
: format
-= 2; break;
1417 case V_008F0C_BUF_NUM_FORMAT_SSCALED
: format
-= 1; break;
1418 default: unreachable("bad nfmt");
1419 case V_008F0C_BUF_NUM_FORMAT_UINT
: break;
1420 case V_008F0C_BUF_NUM_FORMAT_SINT
: format
+= 1; break;
1421 case V_008F0C_BUF_NUM_FORMAT_FLOAT
: format
+= 2; break;
1426 return dfmt
| (nfmt
<< 4);
1431 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1433 LLVMValueRef vindex
,
1434 LLVMValueRef voffset
,
1435 LLVMValueRef soffset
,
1436 LLVMValueRef immoffset
,
1437 unsigned num_channels
,
1440 unsigned cache_policy
,
1444 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1446 LLVMValueRef args
[6];
1448 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1450 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1451 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1452 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1453 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
, dfmt
, nfmt
), 0);
1454 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1455 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1456 const char *indexing_kind
= structurized
? "struct" : "raw";
1457 char name
[256], type_name
[8];
1459 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1460 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1462 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1463 indexing_kind
, type_name
);
1465 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1466 ac_get_load_intr_attribs(can_speculate
));
1470 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1472 LLVMValueRef vindex
,
1473 LLVMValueRef voffset
,
1474 LLVMValueRef soffset
,
1475 LLVMValueRef immoffset
,
1476 unsigned num_channels
,
1479 unsigned cache_policy
,
1482 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1483 immoffset
, num_channels
, dfmt
, nfmt
,
1484 cache_policy
, can_speculate
, true);
1488 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1490 LLVMValueRef voffset
,
1491 LLVMValueRef soffset
,
1492 LLVMValueRef immoffset
,
1493 unsigned num_channels
,
1496 unsigned cache_policy
,
1499 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1500 immoffset
, num_channels
, dfmt
, nfmt
,
1501 cache_policy
, can_speculate
, false);
1505 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1507 LLVMValueRef voffset
,
1508 LLVMValueRef soffset
,
1509 LLVMValueRef immoffset
,
1510 unsigned cache_policy
)
1514 if (HAVE_LLVM
>= 0x900) {
1515 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1517 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1518 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1520 1, ctx
->i16
, cache_policy
,
1521 false, false, false);
1523 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1524 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1526 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1527 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1530 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1537 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1539 LLVMValueRef voffset
,
1540 LLVMValueRef soffset
,
1541 LLVMValueRef immoffset
,
1542 unsigned cache_policy
)
1546 if (HAVE_LLVM
>= 0x900) {
1547 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1549 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1550 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1552 1, ctx
->i8
, cache_policy
,
1553 false, false, false);
1555 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1556 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1558 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1559 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1562 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1569 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1571 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1572 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1575 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1577 assert(LLVMTypeOf(src
) == ctx
->i32
);
1580 LLVMValueRef mantissa
;
1581 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1583 /* Converting normal numbers is just a shift + correcting the exponent bias */
1584 unsigned normal_shift
= 23 - mant_bits
;
1585 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1586 LLVMValueRef shifted
, normal
;
1588 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1589 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1591 /* Converting nan/inf numbers is the same, but with a different exponent update */
1592 LLVMValueRef naninf
;
1593 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1595 /* Converting denormals is the complex case: determine the leading zeros of the
1596 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1598 LLVMValueRef denormal
;
1599 LLVMValueRef params
[2] = {
1601 ctx
->i1true
, /* result can be undef when arg is 0 */
1603 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1604 params
, 2, AC_FUNC_ATTR_READNONE
);
1606 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1607 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1608 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1610 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1611 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1612 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1613 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1615 /* Select the final result. */
1616 LLVMValueRef result
;
1618 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1619 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1620 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1622 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1623 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1624 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1626 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1627 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1629 return ac_to_float(ctx
, result
);
1633 * Generate a fully general open coded buffer format fetch with all required
1634 * fixups suitable for vertex fetch, using non-format buffer loads.
1636 * Some combinations of argument values have special interpretations:
1637 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1638 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1640 * \param log_size log(size of channel in bytes)
1641 * \param num_channels number of channels (1 to 4)
1642 * \param format AC_FETCH_FORMAT_xxx value
1643 * \param reverse whether XYZ channels are reversed
1644 * \param known_aligned whether the source is known to be aligned to hardware's
1645 * effective element size for loading the given format
1646 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1647 * \param rsrc buffer resource descriptor
1648 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1651 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1653 unsigned num_channels
,
1658 LLVMValueRef vindex
,
1659 LLVMValueRef voffset
,
1660 LLVMValueRef soffset
,
1661 unsigned cache_policy
,
1665 unsigned load_log_size
= log_size
;
1666 unsigned load_num_channels
= num_channels
;
1667 if (log_size
== 3) {
1669 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1670 load_num_channels
= 2 * num_channels
;
1672 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1676 int log_recombine
= 0;
1677 if (ctx
->chip_class
== GFX6
&& !known_aligned
) {
1678 /* Avoid alignment restrictions by loading one byte at a time. */
1679 load_num_channels
<<= load_log_size
;
1680 log_recombine
= load_log_size
;
1682 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1683 log_recombine
= -util_logbase2(load_num_channels
);
1684 load_num_channels
= 1;
1685 load_log_size
+= -log_recombine
;
1688 assert(load_log_size
>= 2 || HAVE_LLVM
>= 0x0900);
1690 LLVMValueRef loads
[32]; /* up to 32 bytes */
1691 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1692 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1693 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1694 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1695 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1696 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1697 loads
[i
] = ac_build_buffer_load_common(
1698 ctx
, rsrc
, vindex
, voffset
, tmp
,
1699 num_channels
, channel_type
, cache_policy
,
1700 can_speculate
, false, true);
1701 if (load_log_size
>= 2)
1702 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1705 if (log_recombine
> 0) {
1706 /* Recombine bytes if necessary (GFX6 only) */
1707 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1709 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1710 LLVMValueRef accum
= NULL
;
1711 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1712 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1716 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1717 LLVMConstInt(dst_type
, 8 * i
, false), "");
1718 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1723 } else if (log_recombine
< 0) {
1724 /* Split vectors of dwords */
1725 if (load_log_size
> 2) {
1726 assert(load_num_channels
== 1);
1727 LLVMValueRef loaded
= loads
[0];
1728 unsigned log_split
= load_log_size
- 2;
1729 log_recombine
+= log_split
;
1730 load_num_channels
= 1 << log_split
;
1732 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1733 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1734 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1738 /* Further split dwords and shorts if required */
1739 if (log_recombine
< 0) {
1740 for (unsigned src
= load_num_channels
,
1741 dst
= load_num_channels
<< -log_recombine
;
1743 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1744 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1745 LLVMValueRef loaded
= loads
[src
- 1];
1746 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1747 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1748 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1749 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1750 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1756 if (log_size
== 3) {
1757 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1758 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1759 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1760 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1762 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1763 /* 10_11_11_FLOAT */
1764 LLVMValueRef data
= loads
[0];
1765 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1766 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1767 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1768 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1769 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1771 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1772 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1773 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1777 format
= AC_FETCH_FORMAT_FLOAT
;
1779 /* 2_10_10_10 data formats */
1780 LLVMValueRef data
= loads
[0];
1781 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1782 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1783 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1784 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1785 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1786 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1787 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1788 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1789 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1795 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1796 if (log_size
!= 2) {
1797 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1798 tmp
= ac_to_float(ctx
, loads
[chan
]);
1800 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1801 else if (log_size
== 1)
1802 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1803 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1806 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1807 if (log_size
!= 2) {
1808 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1809 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1811 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1812 if (log_size
!= 2) {
1813 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1814 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1817 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1818 format
== AC_FETCH_FORMAT_USCALED
||
1819 format
== AC_FETCH_FORMAT_UINT
;
1821 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1823 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1825 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1828 LLVMValueRef scale
= NULL
;
1829 if (format
== AC_FETCH_FORMAT_FIXED
) {
1830 assert(log_size
== 2);
1831 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1832 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1833 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1834 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1835 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1836 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1837 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1840 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1842 if (format
== AC_FETCH_FORMAT_SNORM
) {
1843 /* Clamp to [-1, 1] */
1844 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1845 LLVMValueRef clamp
=
1846 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1847 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1850 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1854 while (num_channels
< 4) {
1855 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1856 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1858 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1865 loads
[0] = loads
[2];
1869 return ac_build_gather_values(ctx
, loads
, 4);
1873 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1876 LLVMValueRef vindex
,
1877 LLVMValueRef voffset
,
1878 LLVMValueRef soffset
,
1879 LLVMValueRef immoffset
,
1880 unsigned num_channels
,
1883 unsigned cache_policy
,
1886 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
? voffset
: ctx
->i32_0
,
1889 LLVMValueRef args
[7];
1891 args
[idx
++] = vdata
;
1892 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1894 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1895 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1896 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1897 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
, dfmt
, nfmt
), 0);
1898 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1899 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1900 const char *indexing_kind
= structurized
? "struct" : "raw";
1901 char name
[256], type_name
[8];
1903 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1904 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1906 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1907 indexing_kind
, type_name
);
1909 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1910 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1914 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1917 LLVMValueRef vindex
,
1918 LLVMValueRef voffset
,
1919 LLVMValueRef soffset
,
1920 LLVMValueRef immoffset
,
1921 unsigned num_channels
,
1924 unsigned cache_policy
)
1926 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1927 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1932 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1935 LLVMValueRef voffset
,
1936 LLVMValueRef soffset
,
1937 LLVMValueRef immoffset
,
1938 unsigned num_channels
,
1941 unsigned cache_policy
)
1943 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1944 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1949 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1952 LLVMValueRef voffset
,
1953 LLVMValueRef soffset
,
1954 unsigned cache_policy
)
1956 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1958 if (HAVE_LLVM
>= 0x900) {
1959 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1960 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1961 voffset
, soffset
, 1,
1962 ctx
->i16
, cache_policy
,
1965 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1966 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1968 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1970 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1971 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1976 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1979 LLVMValueRef voffset
,
1980 LLVMValueRef soffset
,
1981 unsigned cache_policy
)
1983 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1985 if (HAVE_LLVM
>= 0x900) {
1986 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1987 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1988 voffset
, soffset
, 1,
1989 ctx
->i8
, cache_policy
,
1992 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1993 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1995 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1997 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1998 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
2002 * Set range metadata on an instruction. This can only be used on load and
2003 * call instructions. If you know an instruction can only produce the values
2004 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
2005 * \p lo is the minimum value inclusive.
2006 * \p hi is the maximum value exclusive.
2008 static void set_range_metadata(struct ac_llvm_context
*ctx
,
2009 LLVMValueRef value
, unsigned lo
, unsigned hi
)
2011 LLVMValueRef range_md
, md_args
[2];
2012 LLVMTypeRef type
= LLVMTypeOf(value
);
2013 LLVMContextRef context
= LLVMGetTypeContext(type
);
2015 md_args
[0] = LLVMConstInt(type
, lo
, false);
2016 md_args
[1] = LLVMConstInt(type
, hi
, false);
2017 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
2018 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
2022 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2026 LLVMValueRef tid_args
[2];
2027 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2028 tid_args
[1] = ctx
->i32_0
;
2029 tid_args
[1] = ac_build_intrinsic(ctx
,
2030 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2031 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2033 if (ctx
->wave_size
== 32) {
2036 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2038 2, AC_FUNC_ATTR_READNONE
);
2040 set_range_metadata(ctx
, tid
, 0, ctx
->wave_size
);
2045 * AMD GCN implements derivatives using the local data store (LDS)
2046 * All writes to the LDS happen in all executing threads at
2047 * the same time. TID is the Thread ID for the current
2048 * thread and is a value between 0 and 63, representing
2049 * the thread's position in the wavefront.
2051 * For the pixel shader threads are grouped into quads of four pixels.
2052 * The TIDs of the pixels of a quad are:
2060 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2061 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2062 * the current pixel's column, and masking with 0xfffffffe yields the TID
2063 * of the left pixel of the current pixel's row.
2065 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2066 * adding 2 yields the TID of the pixel below the top pixel.
2069 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2074 unsigned tl_lanes
[4], trbl_lanes
[4];
2075 char name
[32], type
[8];
2076 LLVMValueRef tl
, trbl
;
2077 LLVMTypeRef result_type
;
2078 LLVMValueRef result
;
2080 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2082 if (result_type
== ctx
->f16
)
2083 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2085 for (unsigned i
= 0; i
< 4; ++i
) {
2086 tl_lanes
[i
] = i
& mask
;
2087 trbl_lanes
[i
] = (i
& mask
) + idx
;
2090 tl
= ac_build_quad_swizzle(ctx
, val
,
2091 tl_lanes
[0], tl_lanes
[1],
2092 tl_lanes
[2], tl_lanes
[3]);
2093 trbl
= ac_build_quad_swizzle(ctx
, val
,
2094 trbl_lanes
[0], trbl_lanes
[1],
2095 trbl_lanes
[2], trbl_lanes
[3]);
2097 if (result_type
== ctx
->f16
) {
2098 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2099 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2102 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2103 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2104 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2106 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2107 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2109 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2113 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2115 LLVMValueRef wave_id
)
2117 LLVMValueRef args
[2];
2118 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2120 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2124 ac_build_imsb(struct ac_llvm_context
*ctx
,
2126 LLVMTypeRef dst_type
)
2128 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2130 AC_FUNC_ATTR_READNONE
);
2132 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2133 * the index from LSB. Invert it by doing "31 - msb". */
2134 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2137 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2138 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2139 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2140 arg
, ctx
->i32_0
, ""),
2141 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2142 arg
, all_ones
, ""), "");
2144 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2148 ac_build_umsb(struct ac_llvm_context
*ctx
,
2150 LLVMTypeRef dst_type
)
2152 const char *intrin_name
;
2154 LLVMValueRef highest_bit
;
2158 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2161 intrin_name
= "llvm.ctlz.i64";
2163 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2167 intrin_name
= "llvm.ctlz.i32";
2169 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2173 intrin_name
= "llvm.ctlz.i16";
2175 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2179 intrin_name
= "llvm.ctlz.i8";
2181 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2185 unreachable(!"invalid bitsize");
2189 LLVMValueRef params
[2] = {
2194 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2196 AC_FUNC_ATTR_READNONE
);
2198 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2199 * the index from LSB. Invert it by doing "31 - msb". */
2200 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2202 if (bitsize
== 64) {
2203 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2204 } else if (bitsize
< 32) {
2205 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2208 /* check for zero */
2209 return LLVMBuildSelect(ctx
->builder
,
2210 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2211 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2214 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2218 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2219 LLVMValueRef args
[2] = {a
, b
};
2220 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2221 AC_FUNC_ATTR_READNONE
);
2224 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2228 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2229 LLVMValueRef args
[2] = {a
, b
};
2230 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2231 AC_FUNC_ATTR_READNONE
);
2234 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2237 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2238 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2241 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2244 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2245 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2248 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2251 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2252 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2255 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2258 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2259 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2262 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2264 LLVMTypeRef t
= LLVMTypeOf(value
);
2265 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2266 LLVMConstReal(t
, 1.0));
2269 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2271 LLVMValueRef args
[9];
2273 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2274 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2277 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2278 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2280 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2282 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2284 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2285 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2287 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2288 ctx
->voidt
, args
, 6, 0);
2290 args
[2] = a
->out
[0];
2291 args
[3] = a
->out
[1];
2292 args
[4] = a
->out
[2];
2293 args
[5] = a
->out
[3];
2294 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2295 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2297 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2298 ctx
->voidt
, args
, 8, 0);
2302 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2304 struct ac_export_args args
;
2306 args
.enabled_channels
= 0x0; /* enabled channels */
2307 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2308 args
.done
= 1; /* DONE bit */
2309 args
.target
= V_008DFC_SQ_EXP_NULL
;
2310 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2311 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2312 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2313 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2314 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2316 ac_build_export(ctx
, &args
);
2319 static unsigned ac_num_coords(enum ac_image_dim dim
)
2325 case ac_image_1darray
:
2329 case ac_image_2darray
:
2330 case ac_image_2dmsaa
:
2332 case ac_image_2darraymsaa
:
2335 unreachable("ac_num_coords: bad dim");
2339 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2343 case ac_image_1darray
:
2346 case ac_image_2darray
:
2351 case ac_image_2dmsaa
:
2352 case ac_image_2darraymsaa
:
2354 unreachable("derivatives not supported");
2358 static const char *get_atomic_name(enum ac_atomic_op op
)
2361 case ac_atomic_swap
: return "swap";
2362 case ac_atomic_add
: return "add";
2363 case ac_atomic_sub
: return "sub";
2364 case ac_atomic_smin
: return "smin";
2365 case ac_atomic_umin
: return "umin";
2366 case ac_atomic_smax
: return "smax";
2367 case ac_atomic_umax
: return "umax";
2368 case ac_atomic_and
: return "and";
2369 case ac_atomic_or
: return "or";
2370 case ac_atomic_xor
: return "xor";
2371 case ac_atomic_inc_wrap
: return "inc";
2372 case ac_atomic_dec_wrap
: return "dec";
2374 unreachable("bad atomic op");
2377 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2378 struct ac_image_args
*a
)
2380 const char *overload
[3] = { "", "", "" };
2381 unsigned num_overloads
= 0;
2382 LLVMValueRef args
[18];
2383 unsigned num_args
= 0;
2384 enum ac_image_dim dim
= a
->dim
;
2386 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2388 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2389 a
->opcode
!= ac_image_store_mip
) ||
2391 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2392 (!a
->compare
&& !a
->offset
));
2393 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2394 a
->opcode
== ac_image_get_lod
) ||
2396 assert((a
->bias
? 1 : 0) +
2398 (a
->level_zero
? 1 : 0) +
2399 (a
->derivs
[0] ? 1 : 0) <= 1);
2401 if (a
->opcode
== ac_image_get_lod
) {
2403 case ac_image_1darray
:
2406 case ac_image_2darray
:
2415 bool sample
= a
->opcode
== ac_image_sample
||
2416 a
->opcode
== ac_image_gather4
||
2417 a
->opcode
== ac_image_get_lod
;
2418 bool atomic
= a
->opcode
== ac_image_atomic
||
2419 a
->opcode
== ac_image_atomic_cmpswap
;
2420 bool load
= a
->opcode
== ac_image_sample
||
2421 a
->opcode
== ac_image_gather4
||
2422 a
->opcode
== ac_image_load
||
2423 a
->opcode
== ac_image_load_mip
;
2424 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2426 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2427 args
[num_args
++] = a
->data
[0];
2428 if (a
->opcode
== ac_image_atomic_cmpswap
)
2429 args
[num_args
++] = a
->data
[1];
2433 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2436 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2438 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2439 overload
[num_overloads
++] = ".f32";
2442 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2444 unsigned count
= ac_num_derivs(dim
);
2445 for (unsigned i
= 0; i
< count
; ++i
)
2446 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2447 overload
[num_overloads
++] = ".f32";
2449 unsigned num_coords
=
2450 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2451 for (unsigned i
= 0; i
< num_coords
; ++i
)
2452 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2454 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2455 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2457 args
[num_args
++] = a
->resource
;
2459 args
[num_args
++] = a
->sampler
;
2460 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2463 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2464 args
[num_args
++] = LLVMConstInt(ctx
->i32
,
2465 load
? get_load_cache_policy(ctx
, a
->cache_policy
) :
2466 a
->cache_policy
, false);
2469 const char *atomic_subop
= "";
2470 switch (a
->opcode
) {
2471 case ac_image_sample
: name
= "sample"; break;
2472 case ac_image_gather4
: name
= "gather4"; break;
2473 case ac_image_load
: name
= "load"; break;
2474 case ac_image_load_mip
: name
= "load.mip"; break;
2475 case ac_image_store
: name
= "store"; break;
2476 case ac_image_store_mip
: name
= "store.mip"; break;
2477 case ac_image_atomic
:
2479 atomic_subop
= get_atomic_name(a
->atomic
);
2481 case ac_image_atomic_cmpswap
:
2483 atomic_subop
= "cmpswap";
2485 case ac_image_get_lod
: name
= "getlod"; break;
2486 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2487 default: unreachable("invalid image opcode");
2490 const char *dimname
;
2492 case ac_image_1d
: dimname
= "1d"; break;
2493 case ac_image_2d
: dimname
= "2d"; break;
2494 case ac_image_3d
: dimname
= "3d"; break;
2495 case ac_image_cube
: dimname
= "cube"; break;
2496 case ac_image_1darray
: dimname
= "1darray"; break;
2497 case ac_image_2darray
: dimname
= "2darray"; break;
2498 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2499 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2500 default: unreachable("invalid dim");
2504 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2506 snprintf(intr_name
, sizeof(intr_name
),
2507 "llvm.amdgcn.image.%s%s" /* base name */
2508 "%s%s%s" /* sample/gather modifiers */
2509 ".%s.%s%s%s%s", /* dimension and type overloads */
2511 a
->compare
? ".c" : "",
2514 a
->derivs
[0] ? ".d" :
2515 a
->level_zero
? ".lz" : "",
2516 a
->offset
? ".o" : "",
2518 atomic
? "i32" : "v4f32",
2519 overload
[0], overload
[1], overload
[2]);
2524 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2529 LLVMValueRef result
=
2530 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2532 if (!sample
&& retty
== ctx
->v4f32
) {
2533 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2539 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2540 LLVMValueRef args
[2])
2543 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2545 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2546 args
, 2, AC_FUNC_ATTR_READNONE
);
2549 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2550 LLVMValueRef args
[2])
2553 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2554 ctx
->v2i16
, args
, 2,
2555 AC_FUNC_ATTR_READNONE
);
2556 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2559 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2560 LLVMValueRef args
[2])
2563 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2564 ctx
->v2i16
, args
, 2,
2565 AC_FUNC_ATTR_READNONE
);
2566 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2569 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2570 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2571 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2573 assert(bits
== 8 || bits
== 10 || bits
== 16);
2575 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2576 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2577 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2578 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2579 LLVMValueRef max_alpha
=
2580 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2581 LLVMValueRef min_alpha
=
2582 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2586 for (int i
= 0; i
< 2; i
++) {
2587 bool alpha
= hi
&& i
== 1;
2588 args
[i
] = ac_build_imin(ctx
, args
[i
],
2589 alpha
? max_alpha
: max_rgb
);
2590 args
[i
] = ac_build_imax(ctx
, args
[i
],
2591 alpha
? min_alpha
: min_rgb
);
2596 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2597 ctx
->v2i16
, args
, 2,
2598 AC_FUNC_ATTR_READNONE
);
2599 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2602 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2603 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2604 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2606 assert(bits
== 8 || bits
== 10 || bits
== 16);
2608 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2609 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2610 LLVMValueRef max_alpha
=
2611 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2615 for (int i
= 0; i
< 2; i
++) {
2616 bool alpha
= hi
&& i
== 1;
2617 args
[i
] = ac_build_umin(ctx
, args
[i
],
2618 alpha
? max_alpha
: max_rgb
);
2623 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2624 ctx
->v2i16
, args
, 2,
2625 AC_FUNC_ATTR_READNONE
);
2626 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2629 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2631 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2632 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2635 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2637 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2641 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2642 LLVMValueRef offset
, LLVMValueRef width
,
2645 LLVMValueRef args
[] = {
2651 return ac_build_intrinsic(ctx
, is_signed
? "llvm.amdgcn.sbfe.i32" :
2652 "llvm.amdgcn.ubfe.i32",
2653 ctx
->i32
, args
, 3, AC_FUNC_ATTR_READNONE
);
2657 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2658 LLVMValueRef s1
, LLVMValueRef s2
)
2660 return LLVMBuildAdd(ctx
->builder
,
2661 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2664 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2665 LLVMValueRef s1
, LLVMValueRef s2
)
2667 return LLVMBuildFAdd(ctx
->builder
,
2668 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2671 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned wait_flags
)
2676 unsigned lgkmcnt
= 63;
2677 unsigned vmcnt
= ctx
->chip_class
>= GFX9
? 63 : 15;
2678 unsigned vscnt
= 63;
2680 if (wait_flags
& AC_WAIT_LGKM
)
2682 if (wait_flags
& AC_WAIT_VLOAD
)
2685 if (wait_flags
& AC_WAIT_VSTORE
) {
2686 if (ctx
->chip_class
>= GFX10
)
2692 /* There is no intrinsic for vscnt(0), so use a fence. */
2693 if ((wait_flags
& AC_WAIT_LGKM
&&
2694 wait_flags
& AC_WAIT_VLOAD
&&
2695 wait_flags
& AC_WAIT_VSTORE
) ||
2697 LLVMBuildFence(ctx
->builder
, LLVMAtomicOrderingRelease
, false, "");
2701 unsigned simm16
= (lgkmcnt
<< 8) |
2702 (7 << 4) | /* expcnt */
2704 ((vmcnt
>> 4) << 14);
2706 LLVMValueRef args
[1] = {
2707 LLVMConstInt(ctx
->i32
, simm16
, false),
2709 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2710 ctx
->voidt
, args
, 1, 0);
2713 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2714 LLVMValueRef src1
, LLVMValueRef src2
,
2720 if (bitsize
== 16) {
2721 intr
= "llvm.amdgcn.fmed3.f16";
2723 } else if (bitsize
== 32) {
2724 intr
= "llvm.amdgcn.fmed3.f32";
2727 intr
= "llvm.amdgcn.fmed3.f64";
2731 LLVMValueRef params
[] = {
2736 return ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2737 AC_FUNC_ATTR_READNONE
);
2740 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2746 if (bitsize
== 16) {
2747 intr
= "llvm.amdgcn.fract.f16";
2749 } else if (bitsize
== 32) {
2750 intr
= "llvm.amdgcn.fract.f32";
2753 intr
= "llvm.amdgcn.fract.f64";
2757 LLVMValueRef params
[] = {
2760 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2761 AC_FUNC_ATTR_READNONE
);
2764 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2767 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2768 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2769 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2771 LLVMValueRef cmp
, val
;
2772 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2773 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2774 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2775 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2779 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2782 LLVMValueRef cmp
, val
, zero
, one
;
2785 if (bitsize
== 16) {
2789 } else if (bitsize
== 32) {
2799 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2800 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2801 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2802 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2806 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2808 LLVMValueRef result
;
2811 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2815 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2816 (LLVMValueRef
[]) { src0
}, 1,
2817 AC_FUNC_ATTR_READNONE
);
2819 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2822 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2823 (LLVMValueRef
[]) { src0
}, 1,
2824 AC_FUNC_ATTR_READNONE
);
2827 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2828 (LLVMValueRef
[]) { src0
}, 1,
2829 AC_FUNC_ATTR_READNONE
);
2831 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2834 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2835 (LLVMValueRef
[]) { src0
}, 1,
2836 AC_FUNC_ATTR_READNONE
);
2838 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2841 unreachable(!"invalid bitsize");
2848 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2851 LLVMValueRef result
;
2854 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2858 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2859 (LLVMValueRef
[]) { src0
}, 1,
2860 AC_FUNC_ATTR_READNONE
);
2862 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2865 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2866 (LLVMValueRef
[]) { src0
}, 1,
2867 AC_FUNC_ATTR_READNONE
);
2870 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2871 (LLVMValueRef
[]) { src0
}, 1,
2872 AC_FUNC_ATTR_READNONE
);
2874 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2877 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
2878 (LLVMValueRef
[]) { src0
}, 1,
2879 AC_FUNC_ATTR_READNONE
);
2881 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2884 unreachable(!"invalid bitsize");
2891 #define AC_EXP_TARGET 0
2892 #define AC_EXP_ENABLED_CHANNELS 1
2893 #define AC_EXP_OUT0 2
2901 struct ac_vs_exp_chan
2905 enum ac_ir_type type
;
2908 struct ac_vs_exp_inst
{
2911 struct ac_vs_exp_chan chan
[4];
2914 struct ac_vs_exports
{
2916 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2919 /* Return true if the PARAM export has been eliminated. */
2920 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2921 uint32_t num_outputs
,
2922 struct ac_vs_exp_inst
*exp
)
2924 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2925 bool is_zero
[4] = {}, is_one
[4] = {};
2927 for (i
= 0; i
< 4; i
++) {
2928 /* It's a constant expression. Undef outputs are eliminated too. */
2929 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2932 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2933 if (exp
->chan
[i
].const_float
== 0)
2935 else if (exp
->chan
[i
].const_float
== 1)
2938 return false; /* other constant */
2943 /* Only certain combinations of 0 and 1 can be eliminated. */
2944 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2945 default_val
= is_zero
[3] ? 0 : 1;
2946 else if (is_one
[0] && is_one
[1] && is_one
[2])
2947 default_val
= is_zero
[3] ? 2 : 3;
2951 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2952 LLVMInstructionEraseFromParent(exp
->inst
);
2954 /* Change OFFSET to DEFAULT_VAL. */
2955 for (i
= 0; i
< num_outputs
; i
++) {
2956 if (vs_output_param_offset
[i
] == exp
->offset
) {
2957 vs_output_param_offset
[i
] =
2958 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
2965 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
2966 uint8_t *vs_output_param_offset
,
2967 uint32_t num_outputs
,
2968 struct ac_vs_exports
*processed
,
2969 struct ac_vs_exp_inst
*exp
)
2971 unsigned p
, copy_back_channels
= 0;
2973 /* See if the output is already in the list of processed outputs.
2974 * The LLVMValueRef comparison relies on SSA.
2976 for (p
= 0; p
< processed
->num
; p
++) {
2977 bool different
= false;
2979 for (unsigned j
= 0; j
< 4; j
++) {
2980 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
2981 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
2983 /* Treat undef as a match. */
2984 if (c2
->type
== AC_IR_UNDEF
)
2987 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2988 * and consider the instruction duplicated.
2990 if (c1
->type
== AC_IR_UNDEF
) {
2991 copy_back_channels
|= 1 << j
;
2995 /* Test whether the channels are not equal. */
2996 if (c1
->type
!= c2
->type
||
2997 (c1
->type
== AC_IR_CONST
&&
2998 c1
->const_float
!= c2
->const_float
) ||
2999 (c1
->type
== AC_IR_VALUE
&&
3000 c1
->value
!= c2
->value
)) {
3008 copy_back_channels
= 0;
3010 if (p
== processed
->num
)
3013 /* If a match was found, but the matching export has undef where the new
3014 * one has a normal value, copy the normal value to the undef channel.
3016 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3018 /* Get current enabled channels mask. */
3019 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3020 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3022 while (copy_back_channels
) {
3023 unsigned chan
= u_bit_scan(©_back_channels
);
3025 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3026 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3027 exp
->chan
[chan
].value
);
3028 match
->chan
[chan
] = exp
->chan
[chan
];
3030 /* Update number of enabled channels because the original mask
3031 * is not always 0xf.
3033 enabled_channels
|= (1 << chan
);
3034 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3035 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3038 /* The PARAM export is duplicated. Kill it. */
3039 LLVMInstructionEraseFromParent(exp
->inst
);
3041 /* Change OFFSET to the matching export. */
3042 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3043 if (vs_output_param_offset
[i
] == exp
->offset
) {
3044 vs_output_param_offset
[i
] = match
->offset
;
3051 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3052 LLVMValueRef main_fn
,
3053 uint8_t *vs_output_param_offset
,
3054 uint32_t num_outputs
,
3055 uint8_t *num_param_exports
)
3057 LLVMBasicBlockRef bb
;
3058 bool removed_any
= false;
3059 struct ac_vs_exports exports
;
3063 /* Process all LLVM instructions. */
3064 bb
= LLVMGetFirstBasicBlock(main_fn
);
3066 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3069 LLVMValueRef cur
= inst
;
3070 inst
= LLVMGetNextInstruction(inst
);
3071 struct ac_vs_exp_inst exp
;
3073 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3076 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3078 if (!ac_llvm_is_function(callee
))
3081 const char *name
= LLVMGetValueName(callee
);
3082 unsigned num_args
= LLVMCountParams(callee
);
3084 /* Check if this is an export instruction. */
3085 if ((num_args
!= 9 && num_args
!= 8) ||
3086 (strcmp(name
, "llvm.SI.export") &&
3087 strcmp(name
, "llvm.amdgcn.exp.f32")))
3090 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3091 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3093 if (target
< V_008DFC_SQ_EXP_PARAM
)
3096 target
-= V_008DFC_SQ_EXP_PARAM
;
3098 /* Parse the instruction. */
3099 memset(&exp
, 0, sizeof(exp
));
3100 exp
.offset
= target
;
3103 for (unsigned i
= 0; i
< 4; i
++) {
3104 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3106 exp
.chan
[i
].value
= v
;
3108 if (LLVMIsUndef(v
)) {
3109 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3110 } else if (LLVMIsAConstantFP(v
)) {
3111 LLVMBool loses_info
;
3112 exp
.chan
[i
].type
= AC_IR_CONST
;
3113 exp
.chan
[i
].const_float
=
3114 LLVMConstRealGetDouble(v
, &loses_info
);
3116 exp
.chan
[i
].type
= AC_IR_VALUE
;
3120 /* Eliminate constant and duplicated PARAM exports. */
3121 if (ac_eliminate_const_output(vs_output_param_offset
,
3122 num_outputs
, &exp
) ||
3123 ac_eliminate_duplicated_output(ctx
,
3124 vs_output_param_offset
,
3125 num_outputs
, &exports
,
3129 exports
.exp
[exports
.num
++] = exp
;
3132 bb
= LLVMGetNextBasicBlock(bb
);
3135 /* Remove holes in export memory due to removed PARAM exports.
3136 * This is done by renumbering all PARAM exports.
3139 uint8_t old_offset
[VARYING_SLOT_MAX
];
3142 /* Make a copy of the offsets. We need the old version while
3143 * we are modifying some of them. */
3144 memcpy(old_offset
, vs_output_param_offset
,
3145 sizeof(old_offset
));
3147 for (i
= 0; i
< exports
.num
; i
++) {
3148 unsigned offset
= exports
.exp
[i
].offset
;
3150 /* Update vs_output_param_offset. Multiple outputs can
3151 * have the same offset.
3153 for (out
= 0; out
< num_outputs
; out
++) {
3154 if (old_offset
[out
] == offset
)
3155 vs_output_param_offset
[out
] = i
;
3158 /* Change the PARAM offset in the instruction. */
3159 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3160 LLVMConstInt(ctx
->i32
,
3161 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3163 *num_param_exports
= exports
.num
;
3167 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3169 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3170 ac_build_intrinsic(ctx
,
3171 "llvm.amdgcn.init.exec", ctx
->voidt
,
3172 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3175 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3177 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3178 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3179 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3183 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3184 LLVMValueRef dw_addr
)
3186 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3189 void ac_lds_store(struct ac_llvm_context
*ctx
,
3190 LLVMValueRef dw_addr
,
3193 value
= ac_to_integer(ctx
, value
);
3194 ac_build_indexed_store(ctx
, ctx
->lds
,
3198 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3199 LLVMTypeRef dst_type
,
3202 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3203 const char *intrin_name
;
3207 switch (src0_bitsize
) {
3209 intrin_name
= "llvm.cttz.i64";
3214 intrin_name
= "llvm.cttz.i32";
3219 intrin_name
= "llvm.cttz.i16";
3224 intrin_name
= "llvm.cttz.i8";
3229 unreachable(!"invalid bitsize");
3232 LLVMValueRef params
[2] = {
3235 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3236 * add special code to check for x=0. The reason is that
3237 * the LLVM behavior for x=0 is different from what we
3238 * need here. However, LLVM also assumes that ffs(x) is
3239 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3240 * a conditional assignment to handle 0 is still required.
3242 * The hardware already implements the correct behavior.
3247 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3249 AC_FUNC_ATTR_READNONE
);
3251 if (src0_bitsize
== 64) {
3252 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3253 } else if (src0_bitsize
< 32) {
3254 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3257 /* TODO: We need an intrinsic to skip this conditional. */
3258 /* Check for zero: */
3259 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3262 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3265 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3267 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3270 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3272 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3275 static struct ac_llvm_flow
*
3276 get_current_flow(struct ac_llvm_context
*ctx
)
3278 if (ctx
->flow
->depth
> 0)
3279 return &ctx
->flow
->stack
[ctx
->flow
->depth
- 1];
3283 static struct ac_llvm_flow
*
3284 get_innermost_loop(struct ac_llvm_context
*ctx
)
3286 for (unsigned i
= ctx
->flow
->depth
; i
> 0; --i
) {
3287 if (ctx
->flow
->stack
[i
- 1].loop_entry_block
)
3288 return &ctx
->flow
->stack
[i
- 1];
3293 static struct ac_llvm_flow
*
3294 push_flow(struct ac_llvm_context
*ctx
)
3296 struct ac_llvm_flow
*flow
;
3298 if (ctx
->flow
->depth
>= ctx
->flow
->depth_max
) {
3299 unsigned new_max
= MAX2(ctx
->flow
->depth
<< 1,
3300 AC_LLVM_INITIAL_CF_DEPTH
);
3302 ctx
->flow
->stack
= realloc(ctx
->flow
->stack
, new_max
* sizeof(*ctx
->flow
->stack
));
3303 ctx
->flow
->depth_max
= new_max
;
3306 flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
];
3309 flow
->next_block
= NULL
;
3310 flow
->loop_entry_block
= NULL
;
3314 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3318 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3319 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3322 /* Append a basic block at the level of the parent flow.
3324 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3327 assert(ctx
->flow
->depth
>= 1);
3329 if (ctx
->flow
->depth
>= 2) {
3330 struct ac_llvm_flow
*flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
- 2];
3332 return LLVMInsertBasicBlockInContext(ctx
->context
,
3333 flow
->next_block
, name
);
3336 LLVMValueRef main_fn
=
3337 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3338 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3341 /* Emit a branch to the given default target for the current block if
3342 * applicable -- that is, if the current block does not already contain a
3343 * branch from a break or continue.
3345 static void emit_default_branch(LLVMBuilderRef builder
,
3346 LLVMBasicBlockRef target
)
3348 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3349 LLVMBuildBr(builder
, target
);
3352 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3354 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3355 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3356 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3357 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3358 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3359 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3362 void ac_build_break(struct ac_llvm_context
*ctx
)
3364 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3365 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3368 void ac_build_continue(struct ac_llvm_context
*ctx
)
3370 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3371 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3374 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3376 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3377 LLVMBasicBlockRef endif_block
;
3379 assert(!current_branch
->loop_entry_block
);
3381 endif_block
= append_basic_block(ctx
, "ENDIF");
3382 emit_default_branch(ctx
->builder
, endif_block
);
3384 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3385 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3387 current_branch
->next_block
= endif_block
;
3390 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3392 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3394 assert(!current_branch
->loop_entry_block
);
3396 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3397 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3398 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3403 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3405 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3407 assert(current_loop
->loop_entry_block
);
3409 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3411 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3412 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3416 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3418 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3419 LLVMBasicBlockRef if_block
;
3421 if_block
= append_basic_block(ctx
, "IF");
3422 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3423 set_basicblock_name(if_block
, "if", label_id
);
3424 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3425 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3428 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3431 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3432 value
, ctx
->f32_0
, "");
3433 ac_build_ifcc(ctx
, cond
, label_id
);
3436 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3439 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3440 ac_to_integer(ctx
, value
),
3442 ac_build_ifcc(ctx
, cond
, label_id
);
3445 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3448 LLVMBuilderRef builder
= ac
->builder
;
3449 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3450 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3451 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3452 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3453 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3457 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3459 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3462 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3463 LLVMDisposeBuilder(first_builder
);
3467 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3468 LLVMTypeRef type
, const char *name
)
3470 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3471 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3475 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3478 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3479 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3480 LLVMPointerType(type
, addr_space
), "");
3483 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3486 unsigned num_components
= ac_get_llvm_num_components(value
);
3487 if (count
== num_components
)
3490 LLVMValueRef masks
[MAX2(count
, 2)];
3491 masks
[0] = ctx
->i32_0
;
3492 masks
[1] = ctx
->i32_1
;
3493 for (unsigned i
= 2; i
< count
; i
++)
3494 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3497 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3500 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3501 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3504 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3505 unsigned rshift
, unsigned bitwidth
)
3507 LLVMValueRef value
= param
;
3509 value
= LLVMBuildLShr(ctx
->builder
, value
,
3510 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3512 if (rshift
+ bitwidth
< 32) {
3513 unsigned mask
= (1 << bitwidth
) - 1;
3514 value
= LLVMBuildAnd(ctx
->builder
, value
,
3515 LLVMConstInt(ctx
->i32
, mask
, false), "");
3520 /* Adjust the sample index according to FMASK.
3522 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3523 * which is the identity mapping. Each nibble says which physical sample
3524 * should be fetched to get that sample.
3526 * For example, 0x11111100 means there are only 2 samples stored and
3527 * the second sample covers 3/4 of the pixel. When reading samples 0
3528 * and 1, return physical sample 0 (determined by the first two 0s
3529 * in FMASK), otherwise return physical sample 1.
3531 * The sample index should be adjusted as follows:
3532 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3534 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3535 LLVMValueRef
*addr
, bool is_array_tex
)
3537 struct ac_image_args fmask_load
= {};
3538 fmask_load
.opcode
= ac_image_load
;
3539 fmask_load
.resource
= fmask
;
3540 fmask_load
.dmask
= 0xf;
3541 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3542 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3544 fmask_load
.coords
[0] = addr
[0];
3545 fmask_load
.coords
[1] = addr
[1];
3547 fmask_load
.coords
[2] = addr
[2];
3549 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3550 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3553 /* Apply the formula. */
3554 unsigned sample_chan
= is_array_tex
? 3 : 2;
3555 LLVMValueRef final_sample
;
3556 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3557 LLVMConstInt(ac
->i32
, 4, 0), "");
3558 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3559 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3560 * with EQAA, so those will map to 0. */
3561 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3562 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3564 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3565 * resource descriptor is 0 (invalid).
3568 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3569 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3570 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3572 /* Replace the MSAA sample index. */
3573 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3574 addr
[sample_chan
], "");
3578 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3580 ac_build_optimization_barrier(ctx
, &src
);
3581 return ac_build_intrinsic(ctx
,
3582 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3583 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3585 lane
== NULL
? 1 : 2,
3586 AC_FUNC_ATTR_READNONE
|
3587 AC_FUNC_ATTR_CONVERGENT
);
3591 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3594 * @param lane - id of the lane or NULL for the first active lane
3595 * @return value of the lane
3598 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3600 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3601 src
= ac_to_integer(ctx
, src
);
3602 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3606 ret
= _ac_build_readlane(ctx
, src
, lane
);
3608 assert(bits
% 32 == 0);
3609 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3610 LLVMValueRef src_vector
=
3611 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3612 ret
= LLVMGetUndef(vec_type
);
3613 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3614 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3615 LLVMConstInt(ctx
->i32
, i
, 0), "");
3616 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
3617 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3618 LLVMConstInt(ctx
->i32
, i
, 0), "");
3621 if (LLVMGetTypeKind(src_type
) == LLVMPointerTypeKind
)
3622 return LLVMBuildIntToPtr(ctx
->builder
, ret
, src_type
, "");
3623 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3627 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3629 return ac_build_intrinsic(ctx
, "llvm.amdgcn.writelane", ctx
->i32
,
3630 (LLVMValueRef
[]) {value
, lane
, src
}, 3,
3631 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3635 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3637 if (ctx
->wave_size
== 32) {
3638 return ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3639 (LLVMValueRef
[]) { mask
, ctx
->i32_0
},
3640 2, AC_FUNC_ATTR_READNONE
);
3642 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3643 LLVMVectorType(ctx
->i32
, 2),
3645 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3647 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3650 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3651 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3652 2, AC_FUNC_ATTR_READNONE
);
3653 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3654 (LLVMValueRef
[]) { mask_hi
, val
},
3655 2, AC_FUNC_ATTR_READNONE
);
3660 _dpp_quad_perm
= 0x000,
3661 _dpp_row_sl
= 0x100,
3662 _dpp_row_sr
= 0x110,
3663 _dpp_row_rr
= 0x120,
3668 dpp_row_mirror
= 0x140,
3669 dpp_row_half_mirror
= 0x141,
3670 dpp_row_bcast15
= 0x142,
3671 dpp_row_bcast31
= 0x143
3674 static inline enum dpp_ctrl
3675 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3677 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3678 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3681 static inline enum dpp_ctrl
3682 dpp_row_sl(unsigned amount
)
3684 assert(amount
> 0 && amount
< 16);
3685 return _dpp_row_sl
| amount
;
3688 static inline enum dpp_ctrl
3689 dpp_row_sr(unsigned amount
)
3691 assert(amount
> 0 && amount
< 16);
3692 return _dpp_row_sr
| amount
;
3696 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3697 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3700 return ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32",
3704 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3705 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3706 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3707 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3708 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3712 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3713 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3716 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3717 src
= ac_to_integer(ctx
, src
);
3718 old
= ac_to_integer(ctx
, old
);
3719 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3722 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3723 bank_mask
, bound_ctrl
);
3725 assert(bits
% 32 == 0);
3726 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3727 LLVMValueRef src_vector
=
3728 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3729 LLVMValueRef old_vector
=
3730 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3731 ret
= LLVMGetUndef(vec_type
);
3732 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3733 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3734 LLVMConstInt(ctx
->i32
, i
,
3736 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3737 LLVMConstInt(ctx
->i32
, i
,
3739 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3744 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3746 LLVMConstInt(ctx
->i32
, i
,
3750 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3754 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3755 bool exchange_rows
, bool bound_ctrl
)
3757 LLVMValueRef args
[6] = {
3760 LLVMConstInt(ctx
->i32
, sel
, false),
3761 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3762 ctx
->i1true
, /* fi */
3763 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3765 return ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3766 : "llvm.amdgcn.permlane16",
3768 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3772 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3773 bool exchange_rows
, bool bound_ctrl
)
3775 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3776 src
= ac_to_integer(ctx
, src
);
3777 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3780 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3783 assert(bits
% 32 == 0);
3784 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3785 LLVMValueRef src_vector
=
3786 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3787 ret
= LLVMGetUndef(vec_type
);
3788 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3789 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3790 LLVMConstInt(ctx
->i32
, i
,
3792 LLVMValueRef ret_comp
=
3793 _ac_build_permlane16(ctx
, src
, sel
,
3796 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3798 LLVMConstInt(ctx
->i32
, i
,
3802 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3805 static inline unsigned
3806 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3808 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3809 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3813 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3815 return ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle",
3816 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3817 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3818 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3822 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3824 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3825 src
= ac_to_integer(ctx
, src
);
3826 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3829 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3831 assert(bits
% 32 == 0);
3832 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3833 LLVMValueRef src_vector
=
3834 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3835 ret
= LLVMGetUndef(vec_type
);
3836 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3837 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3838 LLVMConstInt(ctx
->i32
, i
,
3840 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3842 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3844 LLVMConstInt(ctx
->i32
, i
,
3848 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3852 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3854 char name
[32], type
[8];
3855 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3856 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3857 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3858 (LLVMValueRef
[]) { src
}, 1,
3859 AC_FUNC_ATTR_READNONE
);
3863 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3864 LLVMValueRef inactive
)
3866 char name
[33], type
[8];
3867 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3868 src
= ac_to_integer(ctx
, src
);
3869 inactive
= ac_to_integer(ctx
, inactive
);
3870 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3871 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3873 ac_build_intrinsic(ctx
, name
,
3874 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3876 AC_FUNC_ATTR_READNONE
|
3877 AC_FUNC_ATTR_CONVERGENT
);
3878 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3882 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3884 if (type_size
== 4) {
3886 case nir_op_iadd
: return ctx
->i32_0
;
3887 case nir_op_fadd
: return ctx
->f32_0
;
3888 case nir_op_imul
: return ctx
->i32_1
;
3889 case nir_op_fmul
: return ctx
->f32_1
;
3890 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
3891 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
3892 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
3893 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
3894 case nir_op_umax
: return ctx
->i32_0
;
3895 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
3896 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
3897 case nir_op_ior
: return ctx
->i32_0
;
3898 case nir_op_ixor
: return ctx
->i32_0
;
3900 unreachable("bad reduction intrinsic");
3902 } else { /* type_size == 64bit */
3904 case nir_op_iadd
: return ctx
->i64_0
;
3905 case nir_op_fadd
: return ctx
->f64_0
;
3906 case nir_op_imul
: return ctx
->i64_1
;
3907 case nir_op_fmul
: return ctx
->f64_1
;
3908 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
3909 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
3910 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
3911 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
3912 case nir_op_umax
: return ctx
->i64_0
;
3913 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
3914 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
3915 case nir_op_ior
: return ctx
->i64_0
;
3916 case nir_op_ixor
: return ctx
->i64_0
;
3918 unreachable("bad reduction intrinsic");
3924 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
3926 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
3928 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
3929 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
3930 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
3931 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
3932 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
3933 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
3935 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
3936 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
3938 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
3939 _64bit
? "llvm.minnum.f64" : "llvm.minnum.f32",
3940 _64bit
? ctx
->f64
: ctx
->f32
,
3941 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3942 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
3943 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
3945 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
3946 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
3948 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
3949 _64bit
? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3950 _64bit
? ctx
->f64
: ctx
->f32
,
3951 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3952 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
3953 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
3954 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
3956 unreachable("bad reduction intrinsic");
3961 * \param maxprefix specifies that the result only needs to be correct for a
3962 * prefix of this many threads
3964 * TODO: add inclusive and excluse scan functions for GFX6.
3967 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
3968 unsigned maxprefix
, bool inclusive
)
3970 LLVMValueRef result
, tmp
;
3972 if (ctx
->chip_class
>= GFX10
) {
3973 result
= inclusive
? src
: identity
;
3976 src
= ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
3981 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
3982 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3985 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
3986 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3989 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
3990 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3993 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
3994 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3997 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
3998 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3999 if (maxprefix
<= 16)
4002 if (ctx
->chip_class
>= GFX10
) {
4003 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4004 LLVMBuilderRef builder
= ctx
->builder
;
4005 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4007 /* TODO-GFX10: Can we get better code-gen by putting this into
4008 * a branch so that LLVM generates EXEC mask manipulations? */
4012 tmp
= ac_build_alu_op(ctx
, result
, src
, op
);
4013 tmp
= ac_build_permlane16(ctx
, tmp
, ~(uint64_t)0, true, false);
4014 tmp
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4015 cc
= LLVMBuildAnd(builder
, tid
, LLVMConstInt(ctx
->i32
, 16, false), "");
4016 cc
= LLVMBuildICmp(builder
, LLVMIntNE
, cc
, ctx
->i32_0
, "");
4017 result
= LLVMBuildSelect(builder
, cc
, tmp
, result
, "");
4018 if (maxprefix
<= 32)
4024 tmp
= ac_build_alu_op(ctx
, result
, src
, op
);
4025 tmp
= ac_build_readlane(ctx
, tmp
, LLVMConstInt(ctx
->i32
, 31, false));
4026 tmp
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4027 cc
= LLVMBuildICmp(builder
, LLVMIntUGE
, tid
,
4028 LLVMConstInt(ctx
->i32
, 32, false), "");
4029 result
= LLVMBuildSelect(builder
, cc
, tmp
, result
, "");
4033 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4034 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4035 if (maxprefix
<= 32)
4037 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4038 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4043 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4045 LLVMValueRef result
;
4047 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4048 LLVMBuilderRef builder
= ctx
->builder
;
4049 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4050 result
= ac_build_ballot(ctx
, src
);
4051 result
= ac_build_mbcnt(ctx
, result
);
4052 result
= LLVMBuildAdd(builder
, result
, src
, "");
4056 ac_build_optimization_barrier(ctx
, &src
);
4058 LLVMValueRef identity
=
4059 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4060 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4061 LLVMTypeOf(identity
), "");
4062 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, true);
4064 return ac_build_wwm(ctx
, result
);
4068 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4070 LLVMValueRef result
;
4072 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4073 LLVMBuilderRef builder
= ctx
->builder
;
4074 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4075 result
= ac_build_ballot(ctx
, src
);
4076 result
= ac_build_mbcnt(ctx
, result
);
4080 ac_build_optimization_barrier(ctx
, &src
);
4082 LLVMValueRef identity
=
4083 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4084 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4085 LLVMTypeOf(identity
), "");
4086 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, false);
4088 return ac_build_wwm(ctx
, result
);
4092 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4094 if (cluster_size
== 1) return src
;
4095 ac_build_optimization_barrier(ctx
, &src
);
4096 LLVMValueRef result
, swap
;
4097 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4098 ac_get_type_size(LLVMTypeOf(src
)));
4099 result
= LLVMBuildBitCast(ctx
->builder
,
4100 ac_build_set_inactive(ctx
, src
, identity
),
4101 LLVMTypeOf(identity
), "");
4102 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4103 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4104 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4106 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4107 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4108 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4110 if (ctx
->chip_class
>= GFX8
)
4111 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4113 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4114 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4115 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4117 if (ctx
->chip_class
>= GFX8
)
4118 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4120 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4121 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4122 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4124 if (ctx
->chip_class
>= GFX10
)
4125 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4126 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4127 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4129 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4130 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4131 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4133 if (ctx
->chip_class
>= GFX8
) {
4134 if (ctx
->chip_class
>= GFX10
)
4135 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4137 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4138 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4139 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4140 return ac_build_wwm(ctx
, result
);
4142 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4143 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4144 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4145 return ac_build_wwm(ctx
, result
);
4150 * "Top half" of a scan that reduces per-wave values across an entire
4153 * The source value must be present in the highest lane of the wave, and the
4154 * highest lane must be live.
4157 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4159 if (ws
->maxwaves
<= 1)
4162 const LLVMValueRef last_lane
= LLVMConstInt(ctx
->i32
, ctx
->wave_size
- 1, false);
4163 LLVMBuilderRef builder
= ctx
->builder
;
4164 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4167 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, last_lane
, "");
4168 ac_build_ifcc(ctx
, tmp
, 1000);
4169 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4170 ac_build_endif(ctx
, 1000);
4174 * "Bottom half" of a scan that reduces per-wave values across an entire
4177 * The caller must place a barrier between the top and bottom halves.
4180 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4182 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4183 const LLVMValueRef identity
=
4184 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4186 if (ws
->maxwaves
<= 1) {
4187 ws
->result_reduce
= ws
->src
;
4188 ws
->result_inclusive
= ws
->src
;
4189 ws
->result_exclusive
= identity
;
4192 assert(ws
->maxwaves
<= 32);
4194 LLVMBuilderRef builder
= ctx
->builder
;
4195 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4196 LLVMBasicBlockRef bbs
[2];
4197 LLVMValueRef phivalues_scan
[2];
4198 LLVMValueRef tmp
, tmp2
;
4200 bbs
[0] = LLVMGetInsertBlock(builder
);
4201 phivalues_scan
[0] = LLVMGetUndef(type
);
4203 if (ws
->enable_reduce
)
4204 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4205 else if (ws
->enable_inclusive
)
4206 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4208 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4209 ac_build_ifcc(ctx
, tmp
, 1001);
4211 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4213 ac_build_optimization_barrier(ctx
, &tmp
);
4215 bbs
[1] = LLVMGetInsertBlock(builder
);
4216 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4218 ac_build_endif(ctx
, 1001);
4220 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4222 if (ws
->enable_reduce
) {
4223 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4224 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4226 if (ws
->enable_inclusive
)
4227 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4228 if (ws
->enable_exclusive
) {
4229 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4230 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4231 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4232 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4237 * Inclusive scan of a per-wave value across an entire workgroup.
4239 * This implies an s_barrier instruction.
4241 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4242 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4243 * useful manner because of the barrier in the algorithm.)
4246 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4248 ac_build_wg_wavescan_top(ctx
, ws
);
4249 ac_build_s_barrier(ctx
);
4250 ac_build_wg_wavescan_bottom(ctx
, ws
);
4254 * "Top half" of a scan that reduces per-thread values across an entire
4257 * All lanes must be active when this code runs.
4260 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4262 if (ws
->enable_exclusive
) {
4263 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4264 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4265 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4266 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4268 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4271 bool enable_inclusive
= ws
->enable_inclusive
;
4272 bool enable_exclusive
= ws
->enable_exclusive
;
4273 ws
->enable_inclusive
= false;
4274 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4275 ac_build_wg_wavescan_top(ctx
, ws
);
4276 ws
->enable_inclusive
= enable_inclusive
;
4277 ws
->enable_exclusive
= enable_exclusive
;
4281 * "Bottom half" of a scan that reduces per-thread values across an entire
4284 * The caller must place a barrier between the top and bottom halves.
4287 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4289 bool enable_inclusive
= ws
->enable_inclusive
;
4290 bool enable_exclusive
= ws
->enable_exclusive
;
4291 ws
->enable_inclusive
= false;
4292 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4293 ac_build_wg_wavescan_bottom(ctx
, ws
);
4294 ws
->enable_inclusive
= enable_inclusive
;
4295 ws
->enable_exclusive
= enable_exclusive
;
4297 /* ws->result_reduce is already the correct value */
4298 if (ws
->enable_inclusive
)
4299 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4300 if (ws
->enable_exclusive
)
4301 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4305 * A scan that reduces per-thread values across an entire workgroup.
4307 * The caller must ensure that all lanes are active when this code runs
4308 * (WWM is insufficient!), because there is an implied barrier.
4311 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4313 ac_build_wg_scan_top(ctx
, ws
);
4314 ac_build_s_barrier(ctx
);
4315 ac_build_wg_scan_bottom(ctx
, ws
);
4319 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4320 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4322 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4323 if (ctx
->chip_class
>= GFX8
) {
4324 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4326 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4331 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4333 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4334 return ac_build_intrinsic(ctx
,
4335 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4336 (LLVMValueRef
[]) {index
, src
}, 2,
4337 AC_FUNC_ATTR_READNONE
|
4338 AC_FUNC_ATTR_CONVERGENT
);
4342 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4348 if (bitsize
== 16) {
4349 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4351 } else if (bitsize
== 32) {
4352 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4355 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4359 LLVMValueRef params
[] = {
4362 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4363 AC_FUNC_ATTR_READNONE
);
4366 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4372 if (bitsize
== 16) {
4373 intr
= "llvm.amdgcn.frexp.mant.f16";
4375 } else if (bitsize
== 32) {
4376 intr
= "llvm.amdgcn.frexp.mant.f32";
4379 intr
= "llvm.amdgcn.frexp.mant.f64";
4383 LLVMValueRef params
[] = {
4386 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4387 AC_FUNC_ATTR_READNONE
);
4391 * this takes an I,J coordinate pair,
4392 * and works out the X and Y derivatives.
4393 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4396 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4398 LLVMValueRef result
[4], a
;
4401 for (i
= 0; i
< 2; i
++) {
4402 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4403 LLVMConstInt(ctx
->i32
, i
, false), "");
4404 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4405 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4407 return ac_build_gather_values(ctx
, result
, 4);
4411 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4413 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4415 AC_FUNC_ATTR_READNONE
);
4416 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4417 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4420 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4421 LLVMValueRef
*args
, unsigned num_args
)
4423 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4424 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));