2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
29 #include <llvm/Config/llvm-config.h>
31 #include "c11/threads.h"
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
45 #include "shader_enums.h"
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block
;
54 LLVMBasicBlockRef loop_entry_block
;
57 /* Initialize module-independent parts of the context.
59 * The caller is responsible for initializing ctx::module and ctx::builder.
62 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
63 struct ac_llvm_compiler
*compiler
,
64 enum chip_class chip_class
, enum radeon_family family
,
65 enum ac_float_mode float_mode
, unsigned wave_size
,
66 unsigned ballot_mask_bits
)
70 ctx
->context
= LLVMContextCreate();
72 ctx
->chip_class
= chip_class
;
74 ctx
->wave_size
= wave_size
;
75 ctx
->ballot_mask_bits
= ballot_mask_bits
;
76 ctx
->float_mode
= float_mode
;
77 ctx
->module
= ac_create_module(wave_size
== 32 ? compiler
->tm_wave32
80 ctx
->builder
= ac_create_builder(ctx
->context
, float_mode
);
82 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
83 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
84 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
85 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
86 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
87 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
88 ctx
->i128
= LLVMIntTypeInContext(ctx
->context
, 128);
89 ctx
->intptr
= ctx
->i32
;
90 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
91 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
92 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
93 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
94 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
95 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
96 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
97 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
98 ctx
->v3f32
= LLVMVectorType(ctx
->f32
, 3);
99 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
100 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
101 ctx
->iN_wavemask
= LLVMIntTypeInContext(ctx
->context
, ctx
->wave_size
);
102 ctx
->iN_ballotmask
= LLVMIntTypeInContext(ctx
->context
, ballot_mask_bits
);
104 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
105 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
106 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
107 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
108 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
109 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
110 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
111 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
112 ctx
->i128_0
= LLVMConstInt(ctx
->i128
, 0, false);
113 ctx
->i128_1
= LLVMConstInt(ctx
->i128
, 1, false);
114 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
115 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
116 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
117 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
118 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
119 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
121 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
122 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
124 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
127 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
128 "invariant.load", 14);
130 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
132 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
133 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
135 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
136 "amdgpu.uniform", 14);
138 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
139 ctx
->flow
= calloc(1, sizeof(*ctx
->flow
));
143 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
145 free(ctx
->flow
->stack
);
151 ac_get_llvm_num_components(LLVMValueRef value
)
153 LLVMTypeRef type
= LLVMTypeOf(value
);
154 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
155 ? LLVMGetVectorSize(type
)
157 return num_components
;
161 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
165 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
170 return LLVMBuildExtractElement(ac
->builder
, value
,
171 LLVMConstInt(ac
->i32
, index
, false), "");
175 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
177 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
178 type
= LLVMGetElementType(type
);
180 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
181 return LLVMGetIntTypeWidth(type
);
183 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
184 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_LDS
)
188 if (type
== ctx
->f16
)
190 if (type
== ctx
->f32
)
192 if (type
== ctx
->f64
)
195 unreachable("Unhandled type kind in get_elem_bits");
199 ac_get_type_size(LLVMTypeRef type
)
201 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
204 case LLVMIntegerTypeKind
:
205 return LLVMGetIntTypeWidth(type
) / 8;
206 case LLVMHalfTypeKind
:
208 case LLVMFloatTypeKind
:
210 case LLVMDoubleTypeKind
:
212 case LLVMPointerTypeKind
:
213 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
216 case LLVMVectorTypeKind
:
217 return LLVMGetVectorSize(type
) *
218 ac_get_type_size(LLVMGetElementType(type
));
219 case LLVMArrayTypeKind
:
220 return LLVMGetArrayLength(type
) *
221 ac_get_type_size(LLVMGetElementType(type
));
228 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
232 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
234 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
236 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
239 unreachable("Unhandled integer size");
243 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
245 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
246 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
247 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
248 LLVMGetVectorSize(t
));
250 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
251 switch (LLVMGetPointerAddressSpace(t
)) {
252 case AC_ADDR_SPACE_GLOBAL
:
254 case AC_ADDR_SPACE_CONST_32BIT
:
255 case AC_ADDR_SPACE_LDS
:
258 unreachable("unhandled address space");
261 return to_integer_type_scalar(ctx
, t
);
265 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
267 LLVMTypeRef type
= LLVMTypeOf(v
);
268 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
269 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
271 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
275 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
277 LLVMTypeRef type
= LLVMTypeOf(v
);
278 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
280 return ac_to_integer(ctx
, v
);
283 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
287 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
289 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
291 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
294 unreachable("Unhandled float size");
298 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
300 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
301 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
302 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
303 LLVMGetVectorSize(t
));
305 return to_float_type_scalar(ctx
, t
);
309 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
311 LLVMTypeRef type
= LLVMTypeOf(v
);
312 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
317 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
318 LLVMTypeRef return_type
, LLVMValueRef
*params
,
319 unsigned param_count
, unsigned attrib_mask
)
321 LLVMValueRef function
, call
;
322 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
324 function
= LLVMGetNamedFunction(ctx
->module
, name
);
326 LLVMTypeRef param_types
[32], function_type
;
329 assert(param_count
<= 32);
331 for (i
= 0; i
< param_count
; ++i
) {
333 param_types
[i
] = LLVMTypeOf(params
[i
]);
336 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
337 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
339 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
340 LLVMSetLinkage(function
, LLVMExternalLinkage
);
342 if (!set_callsite_attrs
)
343 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
346 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
347 if (set_callsite_attrs
)
348 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
353 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
356 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
358 LLVMTypeRef elem_type
= type
;
360 assert(bufsize
>= 8);
362 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
363 int ret
= snprintf(buf
, bufsize
, "v%u",
364 LLVMGetVectorSize(type
));
366 char *type_name
= LLVMPrintTypeToString(type
);
367 fprintf(stderr
, "Error building type name for: %s\n",
369 LLVMDisposeMessage(type_name
);
372 elem_type
= LLVMGetElementType(type
);
376 switch (LLVMGetTypeKind(elem_type
)) {
378 case LLVMIntegerTypeKind
:
379 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
381 case LLVMHalfTypeKind
:
382 snprintf(buf
, bufsize
, "f16");
384 case LLVMFloatTypeKind
:
385 snprintf(buf
, bufsize
, "f32");
387 case LLVMDoubleTypeKind
:
388 snprintf(buf
, bufsize
, "f64");
394 * Helper function that builds an LLVM IR PHI node and immediately adds
398 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
399 unsigned count_incoming
, LLVMValueRef
*values
,
400 LLVMBasicBlockRef
*blocks
)
402 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
403 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
407 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
409 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
410 0, AC_FUNC_ATTR_CONVERGENT
);
413 /* Prevent optimizations (at least of memory accesses) across the current
414 * point in the program by emitting empty inline assembly that is marked as
415 * having side effects.
417 * Optionally, a value can be passed through the inline assembly to prevent
418 * LLVM from hoisting calls to ReadNone functions.
421 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
424 static int counter
= 0;
426 LLVMBuilderRef builder
= ctx
->builder
;
429 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
432 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
433 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
434 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
436 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
437 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
438 LLVMTypeRef type
= LLVMTypeOf(*pvgpr
);
439 unsigned bitsize
= ac_get_elem_bits(ctx
, type
);
440 LLVMValueRef vgpr
= *pvgpr
;
441 LLVMTypeRef vgpr_type
;
446 vgpr
= LLVMBuildZExt(ctx
->builder
, vgpr
, ctx
->i32
, "");
448 vgpr_type
= LLVMTypeOf(vgpr
);
449 vgpr_size
= ac_get_type_size(vgpr_type
);
451 assert(vgpr_size
% 4 == 0);
453 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
454 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
455 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
456 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
457 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
460 vgpr
= LLVMBuildTrunc(builder
, vgpr
, type
, "");
467 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
469 const char *intr
= LLVM_VERSION_MAJOR
>= 9 && ctx
->chip_class
>= GFX8
?
470 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
471 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, intr
, ctx
->i64
, NULL
, 0, 0);
472 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
476 ac_build_ballot(struct ac_llvm_context
*ctx
,
481 if (LLVM_VERSION_MAJOR
>= 9) {
482 if (ctx
->wave_size
== 64)
483 name
= "llvm.amdgcn.icmp.i64.i32";
485 name
= "llvm.amdgcn.icmp.i32.i32";
487 name
= "llvm.amdgcn.icmp.i32";
489 LLVMValueRef args
[3] = {
492 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
495 /* We currently have no other way to prevent LLVM from lifting the icmp
496 * calls to a dominating basic block.
498 ac_build_optimization_barrier(ctx
, &args
[0]);
500 args
[0] = ac_to_integer(ctx
, args
[0]);
502 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
503 AC_FUNC_ATTR_NOUNWIND
|
504 AC_FUNC_ATTR_READNONE
|
505 AC_FUNC_ATTR_CONVERGENT
);
508 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
513 if (LLVM_VERSION_MAJOR
>= 9) {
514 if (ctx
->wave_size
== 64)
515 name
= "llvm.amdgcn.icmp.i64.i1";
517 name
= "llvm.amdgcn.icmp.i32.i1";
519 name
= "llvm.amdgcn.icmp.i1";
521 LLVMValueRef args
[3] = {
524 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0),
527 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
528 AC_FUNC_ATTR_NOUNWIND
|
529 AC_FUNC_ATTR_READNONE
|
530 AC_FUNC_ATTR_CONVERGENT
);
534 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
536 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
537 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
538 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
542 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
544 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
545 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
546 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
550 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
552 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
553 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
555 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
556 vote_set
, active_set
, "");
557 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
559 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
560 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
564 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
565 unsigned value_count
, unsigned component
)
567 LLVMValueRef vec
= NULL
;
569 if (value_count
== 1) {
570 return values
[component
];
571 } else if (!value_count
)
572 unreachable("value_count is 0");
574 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
575 LLVMValueRef value
= values
[i
];
578 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
579 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
580 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
586 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
587 LLVMValueRef
*values
,
588 unsigned value_count
,
589 unsigned value_stride
,
593 LLVMBuilderRef builder
= ctx
->builder
;
594 LLVMValueRef vec
= NULL
;
597 if (value_count
== 1 && !always_vector
) {
599 return LLVMBuildLoad(builder
, values
[0], "");
601 } else if (!value_count
)
602 unreachable("value_count is 0");
604 for (i
= 0; i
< value_count
; i
++) {
605 LLVMValueRef value
= values
[i
* value_stride
];
607 value
= LLVMBuildLoad(builder
, value
, "");
610 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
611 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
612 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
618 ac_build_gather_values(struct ac_llvm_context
*ctx
,
619 LLVMValueRef
*values
,
620 unsigned value_count
)
622 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
625 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
626 * channels with undef. Extract at most src_channels components from the input.
629 ac_build_expand(struct ac_llvm_context
*ctx
,
631 unsigned src_channels
,
632 unsigned dst_channels
)
634 LLVMTypeRef elemtype
;
635 LLVMValueRef chan
[dst_channels
];
637 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
638 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
640 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
643 src_channels
= MIN2(src_channels
, vec_size
);
645 for (unsigned i
= 0; i
< src_channels
; i
++)
646 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
648 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
651 assert(src_channels
== 1);
654 elemtype
= LLVMTypeOf(value
);
657 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
658 chan
[i
] = LLVMGetUndef(elemtype
);
660 return ac_build_gather_values(ctx
, chan
, dst_channels
);
663 /* Extract components [start, start + channels) from a vector.
666 ac_extract_components(struct ac_llvm_context
*ctx
,
671 LLVMValueRef chan
[channels
];
673 for (unsigned i
= 0; i
< channels
; i
++)
674 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
+ start
);
676 return ac_build_gather_values(ctx
, chan
, channels
);
679 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
680 * with undef. Extract at most num_channels components from the input.
682 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
684 unsigned num_channels
)
686 return ac_build_expand(ctx
, value
, num_channels
, 4);
689 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
691 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
695 name
= "llvm.rint.f16";
696 else if (type_size
== 4)
697 name
= "llvm.rint.f32";
699 name
= "llvm.rint.f64";
701 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
702 AC_FUNC_ATTR_READNONE
);
706 ac_build_fdiv(struct ac_llvm_context
*ctx
,
710 /* If we do (num / den), LLVM >= 7.0 does:
711 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
713 * If we do (num * (1 / den)), LLVM does:
714 * return num * v_rcp_f32(den);
716 LLVMValueRef one
= LLVMConstReal(LLVMTypeOf(num
), 1.0);
717 LLVMValueRef rcp
= LLVMBuildFDiv(ctx
->builder
, one
, den
, "");
718 /* Use v_rcp_f32 instead of precise division. */
719 if (!LLVMIsConstant(rcp
))
720 LLVMSetMetadata(rcp
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
722 return LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
725 /* See fast_idiv_by_const.h. */
726 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
727 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
729 LLVMValueRef multiplier
,
730 LLVMValueRef pre_shift
,
731 LLVMValueRef post_shift
,
732 LLVMValueRef increment
)
734 LLVMBuilderRef builder
= ctx
->builder
;
736 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
737 num
= LLVMBuildMul(builder
,
738 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
739 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
740 num
= LLVMBuildAdd(builder
, num
,
741 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
742 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
743 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
744 return LLVMBuildLShr(builder
, num
, post_shift
, "");
747 /* See fast_idiv_by_const.h. */
748 /* If num != UINT_MAX, this more efficient version can be used. */
749 /* Set: increment = util_fast_udiv_info::increment; */
750 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
752 LLVMValueRef multiplier
,
753 LLVMValueRef pre_shift
,
754 LLVMValueRef post_shift
,
755 LLVMValueRef increment
)
757 LLVMBuilderRef builder
= ctx
->builder
;
759 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
760 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
761 num
= LLVMBuildMul(builder
,
762 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
763 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
764 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
765 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
766 return LLVMBuildLShr(builder
, num
, post_shift
, "");
769 /* See fast_idiv_by_const.h. */
770 /* Both operands must fit in 31 bits and the divisor must not be 1. */
771 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
773 LLVMValueRef multiplier
,
774 LLVMValueRef post_shift
)
776 LLVMBuilderRef builder
= ctx
->builder
;
778 num
= LLVMBuildMul(builder
,
779 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
780 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
781 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
782 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
783 return LLVMBuildLShr(builder
, num
, post_shift
, "");
786 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
787 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
788 * already multiplied by two. id is the cube face number.
790 struct cube_selection_coords
{
797 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
799 struct cube_selection_coords
*out
)
801 LLVMTypeRef f32
= ctx
->f32
;
803 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
804 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
805 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
806 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
807 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
808 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
809 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
810 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
814 * Build a manual selection sequence for cube face sc/tc coordinates and
815 * major axis vector (multiplied by 2 for consistency) for the given
816 * vec3 \p coords, for the face implied by \p selcoords.
818 * For the major axis, we always adjust the sign to be in the direction of
819 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
820 * the selcoords major axis.
822 static void build_cube_select(struct ac_llvm_context
*ctx
,
823 const struct cube_selection_coords
*selcoords
,
824 const LLVMValueRef
*coords
,
825 LLVMValueRef
*out_st
,
826 LLVMValueRef
*out_ma
)
828 LLVMBuilderRef builder
= ctx
->builder
;
829 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
830 LLVMValueRef is_ma_positive
;
832 LLVMValueRef is_ma_z
, is_not_ma_z
;
833 LLVMValueRef is_ma_y
;
834 LLVMValueRef is_ma_x
;
838 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
839 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
840 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
841 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
843 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
844 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
845 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
846 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
847 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
850 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
851 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
852 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
853 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
854 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
857 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
858 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
859 LLVMConstReal(f32
, -1.0), "");
860 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
863 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
864 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
865 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
866 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
867 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
871 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
872 bool is_deriv
, bool is_array
, bool is_lod
,
873 LLVMValueRef
*coords_arg
,
874 LLVMValueRef
*derivs_arg
)
877 LLVMBuilderRef builder
= ctx
->builder
;
878 struct cube_selection_coords selcoords
;
879 LLVMValueRef coords
[3];
882 if (is_array
&& !is_lod
) {
883 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
885 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
887 * "For Array forms, the array layer used will be
889 * max(0, min(d−1, floor(layer+0.5)))
891 * where d is the depth of the texture array and layer
892 * comes from the component indicated in the tables below.
893 * Workaroudn for an issue where the layer is taken from a
894 * helper invocation which happens to fall on a different
895 * layer due to extrapolation."
897 * GFX8 and earlier attempt to implement this in hardware by
898 * clamping the value of coords[2] = (8 * layer) + face.
899 * Unfortunately, this means that the we end up with the wrong
900 * face when clamping occurs.
902 * Clamp the layer earlier to work around the issue.
904 if (ctx
->chip_class
<= GFX8
) {
906 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
907 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
913 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
915 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
916 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
917 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
919 for (int i
= 0; i
< 2; ++i
)
920 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
922 coords
[2] = selcoords
.id
;
924 if (is_deriv
&& derivs_arg
) {
925 LLVMValueRef derivs
[4];
928 /* Convert cube derivatives to 2D derivatives. */
929 for (axis
= 0; axis
< 2; axis
++) {
930 LLVMValueRef deriv_st
[2];
931 LLVMValueRef deriv_ma
;
933 /* Transform the derivative alongside the texture
934 * coordinate. Mathematically, the correct formula is
935 * as follows. Assume we're projecting onto the +Z face
936 * and denote by dx/dh the derivative of the (original)
937 * X texture coordinate with respect to horizontal
938 * window coordinates. The projection onto the +Z face
943 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
944 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
946 * This motivatives the implementation below.
948 * Whether this actually gives the expected results for
949 * apps that might feed in derivatives obtained via
950 * finite differences is anyone's guess. The OpenGL spec
951 * seems awfully quiet about how textureGrad for cube
952 * maps should be handled.
954 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
955 deriv_st
, &deriv_ma
);
957 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
959 for (int i
= 0; i
< 2; ++i
)
960 derivs
[axis
* 2 + i
] =
961 LLVMBuildFSub(builder
,
962 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
963 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
966 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
969 /* Shift the texture coordinate. This must be applied after the
970 * derivative calculation.
972 for (int i
= 0; i
< 2; ++i
)
973 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
976 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
977 /* coords_arg.w component - array_index for cube arrays */
978 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
981 memcpy(coords_arg
, coords
, sizeof(coords
));
986 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
987 LLVMValueRef llvm_chan
,
988 LLVMValueRef attr_number
,
993 LLVMValueRef args
[5];
998 args
[2] = attr_number
;
1001 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
1002 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1006 args
[2] = llvm_chan
;
1007 args
[3] = attr_number
;
1010 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
1011 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1015 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
1016 LLVMValueRef llvm_chan
,
1017 LLVMValueRef attr_number
,
1018 LLVMValueRef params
,
1022 LLVMValueRef args
[6];
1026 args
[1] = llvm_chan
;
1027 args
[2] = attr_number
;
1028 args
[3] = ctx
->i1false
;
1031 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
1032 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1036 args
[2] = llvm_chan
;
1037 args
[3] = attr_number
;
1038 args
[4] = ctx
->i1false
;
1041 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
1042 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
1046 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
1047 LLVMValueRef parameter
,
1048 LLVMValueRef llvm_chan
,
1049 LLVMValueRef attr_number
,
1050 LLVMValueRef params
)
1052 LLVMValueRef args
[4];
1054 args
[0] = parameter
;
1055 args
[1] = llvm_chan
;
1056 args
[2] = attr_number
;
1059 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
1060 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1064 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
1065 LLVMValueRef base_ptr
,
1068 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1072 ac_build_gep0(struct ac_llvm_context
*ctx
,
1073 LLVMValueRef base_ptr
,
1076 LLVMValueRef indices
[2] = {
1080 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1083 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1086 return LLVMBuildPointerCast(ctx
->builder
,
1087 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1088 LLVMTypeOf(ptr
), "");
1092 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1093 LLVMValueRef base_ptr
, LLVMValueRef index
,
1096 LLVMBuildStore(ctx
->builder
, value
,
1097 ac_build_gep0(ctx
, base_ptr
, index
));
1101 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1102 * It's equivalent to doing a load from &base_ptr[index].
1104 * \param base_ptr Where the array starts.
1105 * \param index The element index into the array.
1106 * \param uniform Whether the base_ptr and index can be assumed to be
1107 * dynamically uniform (i.e. load to an SGPR)
1108 * \param invariant Whether the load is invariant (no other opcodes affect it)
1109 * \param no_unsigned_wraparound
1110 * For all possible re-associations and re-distributions of an expression
1111 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1112 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1113 * does not result in an unsigned integer wraparound. This is used for
1114 * optimal code generation of 32-bit pointer arithmetic.
1116 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1117 * integer wraparound can't be an imm offset in s_load_dword, because
1118 * the instruction performs "addr + offset" in 64 bits.
1120 * Expected usage for bindless textures by chaining GEPs:
1121 * // possible unsigned wraparound, don't use InBounds:
1122 * ptr1 = LLVMBuildGEP(base_ptr, index);
1123 * image = load(ptr1); // becomes "s_load ptr1, 0"
1125 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1126 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1129 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1130 LLVMValueRef index
, bool uniform
, bool invariant
,
1131 bool no_unsigned_wraparound
)
1133 LLVMValueRef pointer
, result
;
1135 if (no_unsigned_wraparound
&&
1136 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1137 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1139 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1142 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1143 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1145 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1149 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1152 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1155 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1156 LLVMValueRef base_ptr
, LLVMValueRef index
)
1158 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1161 /* This assumes that there is no unsigned integer wraparound during the address
1162 * computation, excluding all GEPs within base_ptr. */
1163 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1164 LLVMValueRef base_ptr
, LLVMValueRef index
)
1166 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1169 /* See ac_build_load_custom() documentation. */
1170 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1171 LLVMValueRef base_ptr
, LLVMValueRef index
)
1173 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1176 static unsigned get_load_cache_policy(struct ac_llvm_context
*ctx
,
1177 unsigned cache_policy
)
1179 return cache_policy
|
1180 (ctx
->chip_class
>= GFX10
&& cache_policy
& ac_glc
? ac_dlc
: 0);
1184 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1187 LLVMValueRef vindex
,
1188 LLVMValueRef voffset
,
1189 LLVMValueRef soffset
,
1190 unsigned num_channels
,
1191 LLVMTypeRef return_channel_type
,
1192 unsigned cache_policy
,
1196 LLVMValueRef args
[6];
1199 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1201 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1202 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1203 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1204 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1205 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1206 const char *indexing_kind
= structurized
? "struct" : "raw";
1207 char name
[256], type_name
[8];
1209 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(return_channel_type
, func
) : return_channel_type
;
1210 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1213 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1214 indexing_kind
, type_name
);
1216 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1217 indexing_kind
, type_name
);
1220 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1221 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1225 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1228 LLVMValueRef vindex
,
1229 LLVMValueRef voffset
,
1230 unsigned num_channels
,
1231 unsigned cache_policy
)
1233 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1234 voffset
, NULL
, num_channels
,
1235 ctx
->f32
, cache_policy
,
1239 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1240 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1241 * or v4i32 (num_channels=3,4).
1244 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1247 unsigned num_channels
,
1248 LLVMValueRef voffset
,
1249 LLVMValueRef soffset
,
1250 unsigned inst_offset
,
1251 unsigned cache_policy
)
1253 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1255 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1256 LLVMValueRef v
[3], v01
;
1258 for (int i
= 0; i
< 3; i
++) {
1259 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1260 LLVMConstInt(ctx
->i32
, i
, 0), "");
1262 v01
= ac_build_gather_values(ctx
, v
, 2);
1264 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1265 soffset
, inst_offset
, cache_policy
);
1266 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1267 soffset
, inst_offset
+ 8,
1272 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1273 * (voffset is swizzled, but soffset isn't swizzled).
1274 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1276 if (!(cache_policy
& ac_swizzled
)) {
1277 LLVMValueRef offset
= soffset
;
1280 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1281 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1283 ac_build_buffer_store_common(ctx
, rsrc
, ac_to_float(ctx
, vdata
),
1284 ctx
->i32_0
, voffset
, offset
,
1285 num_channels
, ctx
->f32
,
1286 cache_policy
, false, false);
1290 static const unsigned dfmts
[] = {
1291 V_008F0C_BUF_DATA_FORMAT_32
,
1292 V_008F0C_BUF_DATA_FORMAT_32_32
,
1293 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1294 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1296 unsigned dfmt
= dfmts
[num_channels
- 1];
1297 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1298 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1300 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1301 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
);
1305 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1307 LLVMValueRef vindex
,
1308 LLVMValueRef voffset
,
1309 LLVMValueRef soffset
,
1310 unsigned num_channels
,
1311 LLVMTypeRef channel_type
,
1312 unsigned cache_policy
,
1317 LLVMValueRef args
[5];
1319 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1321 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1322 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1323 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1324 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1325 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1326 const char *indexing_kind
= structurized
? "struct" : "raw";
1327 char name
[256], type_name
[8];
1329 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1330 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1333 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1334 indexing_kind
, type_name
);
1336 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1337 indexing_kind
, type_name
);
1340 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1341 ac_get_load_intr_attribs(can_speculate
));
1345 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1348 LLVMValueRef vindex
,
1349 LLVMValueRef voffset
,
1350 LLVMValueRef soffset
,
1351 unsigned inst_offset
,
1352 unsigned cache_policy
,
1356 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1358 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1360 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1362 if (allow_smem
&& !(cache_policy
& ac_slc
) &&
1363 (!(cache_policy
& ac_glc
) || ctx
->chip_class
>= GFX8
)) {
1364 assert(vindex
== NULL
);
1366 LLVMValueRef result
[8];
1368 for (int i
= 0; i
< num_channels
; i
++) {
1370 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1371 LLVMConstInt(ctx
->i32
, 4, 0), "");
1373 LLVMValueRef args
[3] = {
1376 LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0),
1378 result
[i
] = ac_build_intrinsic(ctx
,
1379 "llvm.amdgcn.s.buffer.load.f32",
1381 AC_FUNC_ATTR_READNONE
);
1383 if (num_channels
== 1)
1386 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1387 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1388 return ac_build_gather_values(ctx
, result
, num_channels
);
1391 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
,
1393 num_channels
, ctx
->f32
,
1395 can_speculate
, false, false);
1398 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1400 LLVMValueRef vindex
,
1401 LLVMValueRef voffset
,
1402 unsigned num_channels
,
1403 unsigned cache_policy
,
1406 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1407 ctx
->i32_0
, num_channels
, ctx
->f32
,
1408 cache_policy
, can_speculate
,
1413 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1415 LLVMValueRef vindex
,
1416 LLVMValueRef voffset
,
1417 LLVMValueRef soffset
,
1418 LLVMValueRef immoffset
,
1419 unsigned num_channels
,
1422 unsigned cache_policy
,
1426 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1428 LLVMValueRef args
[6];
1430 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1432 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1433 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1434 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1435 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1436 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1437 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1438 const char *indexing_kind
= structurized
? "struct" : "raw";
1439 char name
[256], type_name
[8];
1441 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1442 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1444 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1445 indexing_kind
, type_name
);
1447 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1448 ac_get_load_intr_attribs(can_speculate
));
1452 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1454 LLVMValueRef vindex
,
1455 LLVMValueRef voffset
,
1456 LLVMValueRef soffset
,
1457 LLVMValueRef immoffset
,
1458 unsigned num_channels
,
1461 unsigned cache_policy
,
1464 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1465 immoffset
, num_channels
, dfmt
, nfmt
,
1466 cache_policy
, can_speculate
, true);
1470 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1472 LLVMValueRef voffset
,
1473 LLVMValueRef soffset
,
1474 LLVMValueRef immoffset
,
1475 unsigned num_channels
,
1478 unsigned cache_policy
,
1481 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1482 immoffset
, num_channels
, dfmt
, nfmt
,
1483 cache_policy
, can_speculate
, false);
1487 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1489 LLVMValueRef voffset
,
1490 LLVMValueRef soffset
,
1491 LLVMValueRef immoffset
,
1492 unsigned cache_policy
)
1496 if (LLVM_VERSION_MAJOR
>= 9) {
1497 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1499 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1500 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1502 1, ctx
->i16
, cache_policy
,
1503 false, false, false);
1505 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1506 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1508 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1509 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1512 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1519 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1521 LLVMValueRef voffset
,
1522 LLVMValueRef soffset
,
1523 LLVMValueRef immoffset
,
1524 unsigned cache_policy
)
1528 if (LLVM_VERSION_MAJOR
>= 9) {
1529 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1531 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1532 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1534 1, ctx
->i8
, cache_policy
,
1535 false, false, false);
1537 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1538 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1540 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1541 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1544 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1551 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1553 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1554 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1557 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1559 assert(LLVMTypeOf(src
) == ctx
->i32
);
1562 LLVMValueRef mantissa
;
1563 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1565 /* Converting normal numbers is just a shift + correcting the exponent bias */
1566 unsigned normal_shift
= 23 - mant_bits
;
1567 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1568 LLVMValueRef shifted
, normal
;
1570 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1571 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1573 /* Converting nan/inf numbers is the same, but with a different exponent update */
1574 LLVMValueRef naninf
;
1575 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1577 /* Converting denormals is the complex case: determine the leading zeros of the
1578 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1580 LLVMValueRef denormal
;
1581 LLVMValueRef params
[2] = {
1583 ctx
->i1true
, /* result can be undef when arg is 0 */
1585 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1586 params
, 2, AC_FUNC_ATTR_READNONE
);
1588 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1589 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1590 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1592 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1593 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1594 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1595 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1597 /* Select the final result. */
1598 LLVMValueRef result
;
1600 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1601 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1602 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1604 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1605 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1606 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1608 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1609 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1611 return ac_to_float(ctx
, result
);
1615 * Generate a fully general open coded buffer format fetch with all required
1616 * fixups suitable for vertex fetch, using non-format buffer loads.
1618 * Some combinations of argument values have special interpretations:
1619 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1620 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1622 * \param log_size log(size of channel in bytes)
1623 * \param num_channels number of channels (1 to 4)
1624 * \param format AC_FETCH_FORMAT_xxx value
1625 * \param reverse whether XYZ channels are reversed
1626 * \param known_aligned whether the source is known to be aligned to hardware's
1627 * effective element size for loading the given format
1628 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1629 * \param rsrc buffer resource descriptor
1630 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1633 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1635 unsigned num_channels
,
1640 LLVMValueRef vindex
,
1641 LLVMValueRef voffset
,
1642 LLVMValueRef soffset
,
1643 unsigned cache_policy
,
1647 unsigned load_log_size
= log_size
;
1648 unsigned load_num_channels
= num_channels
;
1649 if (log_size
== 3) {
1651 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1652 load_num_channels
= 2 * num_channels
;
1654 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1658 int log_recombine
= 0;
1659 if (ctx
->chip_class
== GFX6
&& !known_aligned
) {
1660 /* Avoid alignment restrictions by loading one byte at a time. */
1661 load_num_channels
<<= load_log_size
;
1662 log_recombine
= load_log_size
;
1664 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1665 log_recombine
= -util_logbase2(load_num_channels
);
1666 load_num_channels
= 1;
1667 load_log_size
+= -log_recombine
;
1670 assert(load_log_size
>= 2 || LLVM_VERSION_MAJOR
>= 9);
1672 LLVMValueRef loads
[32]; /* up to 32 bytes */
1673 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1674 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1675 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1676 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1677 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1678 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1679 loads
[i
] = ac_build_buffer_load_common(
1680 ctx
, rsrc
, vindex
, voffset
, tmp
,
1681 num_channels
, channel_type
, cache_policy
,
1682 can_speculate
, false, true);
1683 if (load_log_size
>= 2)
1684 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1687 if (log_recombine
> 0) {
1688 /* Recombine bytes if necessary (GFX6 only) */
1689 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1691 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1692 LLVMValueRef accum
= NULL
;
1693 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1694 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1698 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1699 LLVMConstInt(dst_type
, 8 * i
, false), "");
1700 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1705 } else if (log_recombine
< 0) {
1706 /* Split vectors of dwords */
1707 if (load_log_size
> 2) {
1708 assert(load_num_channels
== 1);
1709 LLVMValueRef loaded
= loads
[0];
1710 unsigned log_split
= load_log_size
- 2;
1711 log_recombine
+= log_split
;
1712 load_num_channels
= 1 << log_split
;
1714 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1715 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1716 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1720 /* Further split dwords and shorts if required */
1721 if (log_recombine
< 0) {
1722 for (unsigned src
= load_num_channels
,
1723 dst
= load_num_channels
<< -log_recombine
;
1725 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1726 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1727 LLVMValueRef loaded
= loads
[src
- 1];
1728 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1729 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1730 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1731 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1732 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1738 if (log_size
== 3) {
1739 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1740 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1741 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1742 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1744 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1745 /* 10_11_11_FLOAT */
1746 LLVMValueRef data
= loads
[0];
1747 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1748 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1749 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1750 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1751 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1753 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1754 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1755 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1759 format
= AC_FETCH_FORMAT_FLOAT
;
1761 /* 2_10_10_10 data formats */
1762 LLVMValueRef data
= loads
[0];
1763 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1764 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1765 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1766 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1767 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1768 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1769 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1770 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1771 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1777 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1778 if (log_size
!= 2) {
1779 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1780 tmp
= ac_to_float(ctx
, loads
[chan
]);
1782 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1783 else if (log_size
== 1)
1784 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1785 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1788 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1789 if (log_size
!= 2) {
1790 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1791 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1793 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1794 if (log_size
!= 2) {
1795 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1796 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1799 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1800 format
== AC_FETCH_FORMAT_USCALED
||
1801 format
== AC_FETCH_FORMAT_UINT
;
1803 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1805 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1807 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1810 LLVMValueRef scale
= NULL
;
1811 if (format
== AC_FETCH_FORMAT_FIXED
) {
1812 assert(log_size
== 2);
1813 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1814 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1815 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1816 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1817 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1818 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1819 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1822 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1824 if (format
== AC_FETCH_FORMAT_SNORM
) {
1825 /* Clamp to [-1, 1] */
1826 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1827 LLVMValueRef clamp
=
1828 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1829 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1832 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1836 while (num_channels
< 4) {
1837 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1838 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1840 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1847 loads
[0] = loads
[2];
1851 return ac_build_gather_values(ctx
, loads
, 4);
1855 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1858 LLVMValueRef vindex
,
1859 LLVMValueRef voffset
,
1860 LLVMValueRef soffset
,
1861 LLVMValueRef immoffset
,
1862 unsigned num_channels
,
1865 unsigned cache_policy
,
1868 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
? voffset
: ctx
->i32_0
,
1871 LLVMValueRef args
[7];
1873 args
[idx
++] = vdata
;
1874 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1876 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1877 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1878 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1879 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1880 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1881 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1882 const char *indexing_kind
= structurized
? "struct" : "raw";
1883 char name
[256], type_name
[8];
1885 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1886 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1888 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1889 indexing_kind
, type_name
);
1891 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1892 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1896 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1899 LLVMValueRef vindex
,
1900 LLVMValueRef voffset
,
1901 LLVMValueRef soffset
,
1902 LLVMValueRef immoffset
,
1903 unsigned num_channels
,
1906 unsigned cache_policy
)
1908 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1909 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1914 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1917 LLVMValueRef voffset
,
1918 LLVMValueRef soffset
,
1919 LLVMValueRef immoffset
,
1920 unsigned num_channels
,
1923 unsigned cache_policy
)
1925 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1926 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1931 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1934 LLVMValueRef voffset
,
1935 LLVMValueRef soffset
,
1936 unsigned cache_policy
)
1938 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1940 if (LLVM_VERSION_MAJOR
>= 9) {
1941 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1942 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1943 voffset
, soffset
, 1,
1944 ctx
->i16
, cache_policy
,
1947 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1948 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1950 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1952 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1953 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1958 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1961 LLVMValueRef voffset
,
1962 LLVMValueRef soffset
,
1963 unsigned cache_policy
)
1965 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1967 if (LLVM_VERSION_MAJOR
>= 9) {
1968 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1969 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1970 voffset
, soffset
, 1,
1971 ctx
->i8
, cache_policy
,
1974 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1975 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1977 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1979 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1980 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1984 * Set range metadata on an instruction. This can only be used on load and
1985 * call instructions. If you know an instruction can only produce the values
1986 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1987 * \p lo is the minimum value inclusive.
1988 * \p hi is the maximum value exclusive.
1990 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1991 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1993 LLVMValueRef range_md
, md_args
[2];
1994 LLVMTypeRef type
= LLVMTypeOf(value
);
1995 LLVMContextRef context
= LLVMGetTypeContext(type
);
1997 md_args
[0] = LLVMConstInt(type
, lo
, false);
1998 md_args
[1] = LLVMConstInt(type
, hi
, false);
1999 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
2000 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
2004 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2008 LLVMValueRef tid_args
[2];
2009 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2010 tid_args
[1] = ctx
->i32_0
;
2011 tid_args
[1] = ac_build_intrinsic(ctx
,
2012 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2013 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2015 if (ctx
->wave_size
== 32) {
2018 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2020 2, AC_FUNC_ATTR_READNONE
);
2022 set_range_metadata(ctx
, tid
, 0, ctx
->wave_size
);
2027 * AMD GCN implements derivatives using the local data store (LDS)
2028 * All writes to the LDS happen in all executing threads at
2029 * the same time. TID is the Thread ID for the current
2030 * thread and is a value between 0 and 63, representing
2031 * the thread's position in the wavefront.
2033 * For the pixel shader threads are grouped into quads of four pixels.
2034 * The TIDs of the pixels of a quad are:
2042 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2043 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2044 * the current pixel's column, and masking with 0xfffffffe yields the TID
2045 * of the left pixel of the current pixel's row.
2047 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2048 * adding 2 yields the TID of the pixel below the top pixel.
2051 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2056 unsigned tl_lanes
[4], trbl_lanes
[4];
2057 char name
[32], type
[8];
2058 LLVMValueRef tl
, trbl
;
2059 LLVMTypeRef result_type
;
2060 LLVMValueRef result
;
2062 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2064 if (result_type
== ctx
->f16
)
2065 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2067 for (unsigned i
= 0; i
< 4; ++i
) {
2068 tl_lanes
[i
] = i
& mask
;
2069 trbl_lanes
[i
] = (i
& mask
) + idx
;
2072 tl
= ac_build_quad_swizzle(ctx
, val
,
2073 tl_lanes
[0], tl_lanes
[1],
2074 tl_lanes
[2], tl_lanes
[3]);
2075 trbl
= ac_build_quad_swizzle(ctx
, val
,
2076 trbl_lanes
[0], trbl_lanes
[1],
2077 trbl_lanes
[2], trbl_lanes
[3]);
2079 if (result_type
== ctx
->f16
) {
2080 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2081 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2084 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2085 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2086 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2088 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2089 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2091 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2095 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2097 LLVMValueRef wave_id
)
2099 LLVMValueRef args
[2];
2100 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2102 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2106 ac_build_imsb(struct ac_llvm_context
*ctx
,
2108 LLVMTypeRef dst_type
)
2110 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2112 AC_FUNC_ATTR_READNONE
);
2114 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2115 * the index from LSB. Invert it by doing "31 - msb". */
2116 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2119 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2120 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2121 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2122 arg
, ctx
->i32_0
, ""),
2123 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2124 arg
, all_ones
, ""), "");
2126 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2130 ac_build_umsb(struct ac_llvm_context
*ctx
,
2132 LLVMTypeRef dst_type
)
2134 const char *intrin_name
;
2136 LLVMValueRef highest_bit
;
2140 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2143 intrin_name
= "llvm.ctlz.i64";
2145 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2149 intrin_name
= "llvm.ctlz.i32";
2151 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2155 intrin_name
= "llvm.ctlz.i16";
2157 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2161 intrin_name
= "llvm.ctlz.i8";
2163 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2167 unreachable(!"invalid bitsize");
2171 LLVMValueRef params
[2] = {
2176 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2178 AC_FUNC_ATTR_READNONE
);
2180 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2181 * the index from LSB. Invert it by doing "31 - msb". */
2182 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2184 if (bitsize
== 64) {
2185 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2186 } else if (bitsize
< 32) {
2187 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2190 /* check for zero */
2191 return LLVMBuildSelect(ctx
->builder
,
2192 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2193 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2196 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2200 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2201 LLVMValueRef args
[2] = {a
, b
};
2202 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2203 AC_FUNC_ATTR_READNONE
);
2206 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2210 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2211 LLVMValueRef args
[2] = {a
, b
};
2212 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2213 AC_FUNC_ATTR_READNONE
);
2216 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2219 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2220 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2223 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2226 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2227 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2230 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2233 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2234 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2237 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2240 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2241 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2244 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2246 LLVMTypeRef t
= LLVMTypeOf(value
);
2247 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2248 LLVMConstReal(t
, 1.0));
2251 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2253 LLVMValueRef args
[9];
2255 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2256 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2259 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2260 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2262 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2264 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2266 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2267 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2269 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2270 ctx
->voidt
, args
, 6, 0);
2272 args
[2] = a
->out
[0];
2273 args
[3] = a
->out
[1];
2274 args
[4] = a
->out
[2];
2275 args
[5] = a
->out
[3];
2276 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2277 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2279 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2280 ctx
->voidt
, args
, 8, 0);
2284 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2286 struct ac_export_args args
;
2288 args
.enabled_channels
= 0x0; /* enabled channels */
2289 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2290 args
.done
= 1; /* DONE bit */
2291 args
.target
= V_008DFC_SQ_EXP_NULL
;
2292 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2293 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2294 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2295 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2296 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2298 ac_build_export(ctx
, &args
);
2301 static unsigned ac_num_coords(enum ac_image_dim dim
)
2307 case ac_image_1darray
:
2311 case ac_image_2darray
:
2312 case ac_image_2dmsaa
:
2314 case ac_image_2darraymsaa
:
2317 unreachable("ac_num_coords: bad dim");
2321 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2325 case ac_image_1darray
:
2328 case ac_image_2darray
:
2333 case ac_image_2dmsaa
:
2334 case ac_image_2darraymsaa
:
2336 unreachable("derivatives not supported");
2340 static const char *get_atomic_name(enum ac_atomic_op op
)
2343 case ac_atomic_swap
: return "swap";
2344 case ac_atomic_add
: return "add";
2345 case ac_atomic_sub
: return "sub";
2346 case ac_atomic_smin
: return "smin";
2347 case ac_atomic_umin
: return "umin";
2348 case ac_atomic_smax
: return "smax";
2349 case ac_atomic_umax
: return "umax";
2350 case ac_atomic_and
: return "and";
2351 case ac_atomic_or
: return "or";
2352 case ac_atomic_xor
: return "xor";
2353 case ac_atomic_inc_wrap
: return "inc";
2354 case ac_atomic_dec_wrap
: return "dec";
2356 unreachable("bad atomic op");
2359 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2360 struct ac_image_args
*a
)
2362 const char *overload
[3] = { "", "", "" };
2363 unsigned num_overloads
= 0;
2364 LLVMValueRef args
[18];
2365 unsigned num_args
= 0;
2366 enum ac_image_dim dim
= a
->dim
;
2368 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2370 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2371 a
->opcode
!= ac_image_store_mip
) ||
2373 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2374 (!a
->compare
&& !a
->offset
));
2375 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2376 a
->opcode
== ac_image_get_lod
) ||
2378 assert((a
->bias
? 1 : 0) +
2380 (a
->level_zero
? 1 : 0) +
2381 (a
->derivs
[0] ? 1 : 0) <= 1);
2383 if (a
->opcode
== ac_image_get_lod
) {
2385 case ac_image_1darray
:
2388 case ac_image_2darray
:
2397 bool sample
= a
->opcode
== ac_image_sample
||
2398 a
->opcode
== ac_image_gather4
||
2399 a
->opcode
== ac_image_get_lod
;
2400 bool atomic
= a
->opcode
== ac_image_atomic
||
2401 a
->opcode
== ac_image_atomic_cmpswap
;
2402 bool load
= a
->opcode
== ac_image_sample
||
2403 a
->opcode
== ac_image_gather4
||
2404 a
->opcode
== ac_image_load
||
2405 a
->opcode
== ac_image_load_mip
;
2406 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2408 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2409 args
[num_args
++] = a
->data
[0];
2410 if (a
->opcode
== ac_image_atomic_cmpswap
)
2411 args
[num_args
++] = a
->data
[1];
2415 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2418 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2420 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2421 overload
[num_overloads
++] = ".f32";
2424 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2426 unsigned count
= ac_num_derivs(dim
);
2427 for (unsigned i
= 0; i
< count
; ++i
)
2428 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2429 overload
[num_overloads
++] = ".f32";
2431 unsigned num_coords
=
2432 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2433 for (unsigned i
= 0; i
< num_coords
; ++i
)
2434 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2436 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2437 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2439 args
[num_args
++] = a
->resource
;
2441 args
[num_args
++] = a
->sampler
;
2442 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2445 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2446 args
[num_args
++] = LLVMConstInt(ctx
->i32
,
2447 load
? get_load_cache_policy(ctx
, a
->cache_policy
) :
2448 a
->cache_policy
, false);
2451 const char *atomic_subop
= "";
2452 switch (a
->opcode
) {
2453 case ac_image_sample
: name
= "sample"; break;
2454 case ac_image_gather4
: name
= "gather4"; break;
2455 case ac_image_load
: name
= "load"; break;
2456 case ac_image_load_mip
: name
= "load.mip"; break;
2457 case ac_image_store
: name
= "store"; break;
2458 case ac_image_store_mip
: name
= "store.mip"; break;
2459 case ac_image_atomic
:
2461 atomic_subop
= get_atomic_name(a
->atomic
);
2463 case ac_image_atomic_cmpswap
:
2465 atomic_subop
= "cmpswap";
2467 case ac_image_get_lod
: name
= "getlod"; break;
2468 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2469 default: unreachable("invalid image opcode");
2472 const char *dimname
;
2474 case ac_image_1d
: dimname
= "1d"; break;
2475 case ac_image_2d
: dimname
= "2d"; break;
2476 case ac_image_3d
: dimname
= "3d"; break;
2477 case ac_image_cube
: dimname
= "cube"; break;
2478 case ac_image_1darray
: dimname
= "1darray"; break;
2479 case ac_image_2darray
: dimname
= "2darray"; break;
2480 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2481 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2482 default: unreachable("invalid dim");
2486 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2488 snprintf(intr_name
, sizeof(intr_name
),
2489 "llvm.amdgcn.image.%s%s" /* base name */
2490 "%s%s%s" /* sample/gather modifiers */
2491 ".%s.%s%s%s%s", /* dimension and type overloads */
2493 a
->compare
? ".c" : "",
2496 a
->derivs
[0] ? ".d" :
2497 a
->level_zero
? ".lz" : "",
2498 a
->offset
? ".o" : "",
2500 atomic
? "i32" : "v4f32",
2501 overload
[0], overload
[1], overload
[2]);
2506 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2511 LLVMValueRef result
=
2512 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2514 if (!sample
&& retty
== ctx
->v4f32
) {
2515 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2521 LLVMValueRef
ac_build_image_get_sample_count(struct ac_llvm_context
*ctx
,
2524 LLVMValueRef samples
;
2526 /* Read the samples from the descriptor directly.
2527 * Hardware doesn't have any instruction for this.
2529 samples
= LLVMBuildExtractElement(ctx
->builder
, rsrc
,
2530 LLVMConstInt(ctx
->i32
, 3, 0), "");
2531 samples
= LLVMBuildLShr(ctx
->builder
, samples
,
2532 LLVMConstInt(ctx
->i32
, 16, 0), "");
2533 samples
= LLVMBuildAnd(ctx
->builder
, samples
,
2534 LLVMConstInt(ctx
->i32
, 0xf, 0), "");
2535 samples
= LLVMBuildShl(ctx
->builder
, ctx
->i32_1
,
2540 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2541 LLVMValueRef args
[2])
2544 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2546 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2547 args
, 2, AC_FUNC_ATTR_READNONE
);
2550 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2551 LLVMValueRef args
[2])
2554 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2555 ctx
->v2i16
, args
, 2,
2556 AC_FUNC_ATTR_READNONE
);
2557 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2560 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2561 LLVMValueRef args
[2])
2564 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2565 ctx
->v2i16
, args
, 2,
2566 AC_FUNC_ATTR_READNONE
);
2567 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2570 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2571 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2572 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2574 assert(bits
== 8 || bits
== 10 || bits
== 16);
2576 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2577 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2578 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2579 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2580 LLVMValueRef max_alpha
=
2581 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2582 LLVMValueRef min_alpha
=
2583 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2587 for (int i
= 0; i
< 2; i
++) {
2588 bool alpha
= hi
&& i
== 1;
2589 args
[i
] = ac_build_imin(ctx
, args
[i
],
2590 alpha
? max_alpha
: max_rgb
);
2591 args
[i
] = ac_build_imax(ctx
, args
[i
],
2592 alpha
? min_alpha
: min_rgb
);
2597 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2598 ctx
->v2i16
, args
, 2,
2599 AC_FUNC_ATTR_READNONE
);
2600 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2603 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2604 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2605 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2607 assert(bits
== 8 || bits
== 10 || bits
== 16);
2609 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2610 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2611 LLVMValueRef max_alpha
=
2612 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2616 for (int i
= 0; i
< 2; i
++) {
2617 bool alpha
= hi
&& i
== 1;
2618 args
[i
] = ac_build_umin(ctx
, args
[i
],
2619 alpha
? max_alpha
: max_rgb
);
2624 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2625 ctx
->v2i16
, args
, 2,
2626 AC_FUNC_ATTR_READNONE
);
2627 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2630 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2632 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2633 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2636 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2638 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2642 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2643 LLVMValueRef offset
, LLVMValueRef width
,
2646 LLVMValueRef args
[] = {
2652 return ac_build_intrinsic(ctx
, is_signed
? "llvm.amdgcn.sbfe.i32" :
2653 "llvm.amdgcn.ubfe.i32",
2654 ctx
->i32
, args
, 3, AC_FUNC_ATTR_READNONE
);
2658 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2659 LLVMValueRef s1
, LLVMValueRef s2
)
2661 return LLVMBuildAdd(ctx
->builder
,
2662 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2665 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2666 LLVMValueRef s1
, LLVMValueRef s2
)
2668 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2669 if (ctx
->chip_class
>= GFX10
) {
2670 return ac_build_intrinsic(ctx
, "llvm.fma.f32", ctx
->f32
,
2671 (LLVMValueRef
[]) {s0
, s1
, s2
}, 3,
2672 AC_FUNC_ATTR_READNONE
);
2675 return LLVMBuildFAdd(ctx
->builder
,
2676 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2679 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned wait_flags
)
2684 unsigned lgkmcnt
= 63;
2685 unsigned vmcnt
= ctx
->chip_class
>= GFX9
? 63 : 15;
2686 unsigned vscnt
= 63;
2688 if (wait_flags
& AC_WAIT_LGKM
)
2690 if (wait_flags
& AC_WAIT_VLOAD
)
2693 if (wait_flags
& AC_WAIT_VSTORE
) {
2694 if (ctx
->chip_class
>= GFX10
)
2700 /* There is no intrinsic for vscnt(0), so use a fence. */
2701 if ((wait_flags
& AC_WAIT_LGKM
&&
2702 wait_flags
& AC_WAIT_VLOAD
&&
2703 wait_flags
& AC_WAIT_VSTORE
) ||
2705 LLVMBuildFence(ctx
->builder
, LLVMAtomicOrderingRelease
, false, "");
2709 unsigned simm16
= (lgkmcnt
<< 8) |
2710 (7 << 4) | /* expcnt */
2712 ((vmcnt
>> 4) << 14);
2714 LLVMValueRef args
[1] = {
2715 LLVMConstInt(ctx
->i32
, simm16
, false),
2717 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2718 ctx
->voidt
, args
, 1, 0);
2721 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2722 LLVMValueRef src1
, LLVMValueRef src2
,
2725 LLVMValueRef result
;
2727 if (bitsize
== 64 || (bitsize
== 16 && ctx
->chip_class
<= GFX8
)) {
2728 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2729 * or lower 16-bit fmed because it's only supported on GFX9+.
2731 LLVMValueRef min1
, min2
, max1
;
2733 min1
= ac_build_fmin(ctx
, src0
, src1
);
2734 max1
= ac_build_fmax(ctx
, src0
, src1
);
2735 min2
= ac_build_fmin(ctx
, max1
, src2
);
2737 result
= ac_build_fmax(ctx
, min2
, min1
);
2742 if (bitsize
== 16) {
2743 intr
= "llvm.amdgcn.fmed3.f16";
2746 assert(bitsize
== 32);
2747 intr
= "llvm.amdgcn.fmed3.f32";
2751 LLVMValueRef params
[] = {
2757 result
= ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2758 AC_FUNC_ATTR_READNONE
);
2761 if (ctx
->chip_class
< GFX9
&& bitsize
== 32) {
2762 /* Only pre-GFX9 chips do not flush denorms. */
2763 result
= ac_build_canonicalize(ctx
, result
, bitsize
);
2769 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2775 if (bitsize
== 16) {
2776 intr
= "llvm.amdgcn.fract.f16";
2778 } else if (bitsize
== 32) {
2779 intr
= "llvm.amdgcn.fract.f32";
2782 intr
= "llvm.amdgcn.fract.f64";
2786 LLVMValueRef params
[] = {
2789 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2790 AC_FUNC_ATTR_READNONE
);
2793 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2796 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2797 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2798 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2800 LLVMValueRef cmp
, val
;
2801 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2802 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2803 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2804 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2808 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2811 LLVMValueRef cmp
, val
, zero
, one
;
2814 if (bitsize
== 16) {
2818 } else if (bitsize
== 32) {
2828 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2829 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2830 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2831 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2835 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2837 LLVMValueRef result
;
2840 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2844 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i128", ctx
->i128
,
2845 (LLVMValueRef
[]) { src0
}, 1,
2846 AC_FUNC_ATTR_READNONE
);
2847 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2850 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2851 (LLVMValueRef
[]) { src0
}, 1,
2852 AC_FUNC_ATTR_READNONE
);
2854 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2857 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2858 (LLVMValueRef
[]) { src0
}, 1,
2859 AC_FUNC_ATTR_READNONE
);
2862 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2863 (LLVMValueRef
[]) { src0
}, 1,
2864 AC_FUNC_ATTR_READNONE
);
2866 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2869 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2870 (LLVMValueRef
[]) { src0
}, 1,
2871 AC_FUNC_ATTR_READNONE
);
2873 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2876 unreachable(!"invalid bitsize");
2883 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2886 LLVMValueRef result
;
2889 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2893 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2894 (LLVMValueRef
[]) { src0
}, 1,
2895 AC_FUNC_ATTR_READNONE
);
2897 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2900 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2901 (LLVMValueRef
[]) { src0
}, 1,
2902 AC_FUNC_ATTR_READNONE
);
2905 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2906 (LLVMValueRef
[]) { src0
}, 1,
2907 AC_FUNC_ATTR_READNONE
);
2909 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2912 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
2913 (LLVMValueRef
[]) { src0
}, 1,
2914 AC_FUNC_ATTR_READNONE
);
2916 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2919 unreachable(!"invalid bitsize");
2926 #define AC_EXP_TARGET 0
2927 #define AC_EXP_ENABLED_CHANNELS 1
2928 #define AC_EXP_OUT0 2
2936 struct ac_vs_exp_chan
2940 enum ac_ir_type type
;
2943 struct ac_vs_exp_inst
{
2946 struct ac_vs_exp_chan chan
[4];
2949 struct ac_vs_exports
{
2951 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2954 /* Return true if the PARAM export has been eliminated. */
2955 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2956 uint32_t num_outputs
,
2957 struct ac_vs_exp_inst
*exp
)
2959 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2960 bool is_zero
[4] = {}, is_one
[4] = {};
2962 for (i
= 0; i
< 4; i
++) {
2963 /* It's a constant expression. Undef outputs are eliminated too. */
2964 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2967 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2968 if (exp
->chan
[i
].const_float
== 0)
2970 else if (exp
->chan
[i
].const_float
== 1)
2973 return false; /* other constant */
2978 /* Only certain combinations of 0 and 1 can be eliminated. */
2979 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2980 default_val
= is_zero
[3] ? 0 : 1;
2981 else if (is_one
[0] && is_one
[1] && is_one
[2])
2982 default_val
= is_zero
[3] ? 2 : 3;
2986 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2987 LLVMInstructionEraseFromParent(exp
->inst
);
2989 /* Change OFFSET to DEFAULT_VAL. */
2990 for (i
= 0; i
< num_outputs
; i
++) {
2991 if (vs_output_param_offset
[i
] == exp
->offset
) {
2992 vs_output_param_offset
[i
] =
2993 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
3000 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
3001 uint8_t *vs_output_param_offset
,
3002 uint32_t num_outputs
,
3003 struct ac_vs_exports
*processed
,
3004 struct ac_vs_exp_inst
*exp
)
3006 unsigned p
, copy_back_channels
= 0;
3008 /* See if the output is already in the list of processed outputs.
3009 * The LLVMValueRef comparison relies on SSA.
3011 for (p
= 0; p
< processed
->num
; p
++) {
3012 bool different
= false;
3014 for (unsigned j
= 0; j
< 4; j
++) {
3015 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
3016 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
3018 /* Treat undef as a match. */
3019 if (c2
->type
== AC_IR_UNDEF
)
3022 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3023 * and consider the instruction duplicated.
3025 if (c1
->type
== AC_IR_UNDEF
) {
3026 copy_back_channels
|= 1 << j
;
3030 /* Test whether the channels are not equal. */
3031 if (c1
->type
!= c2
->type
||
3032 (c1
->type
== AC_IR_CONST
&&
3033 c1
->const_float
!= c2
->const_float
) ||
3034 (c1
->type
== AC_IR_VALUE
&&
3035 c1
->value
!= c2
->value
)) {
3043 copy_back_channels
= 0;
3045 if (p
== processed
->num
)
3048 /* If a match was found, but the matching export has undef where the new
3049 * one has a normal value, copy the normal value to the undef channel.
3051 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3053 /* Get current enabled channels mask. */
3054 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3055 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3057 while (copy_back_channels
) {
3058 unsigned chan
= u_bit_scan(©_back_channels
);
3060 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3061 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3062 exp
->chan
[chan
].value
);
3063 match
->chan
[chan
] = exp
->chan
[chan
];
3065 /* Update number of enabled channels because the original mask
3066 * is not always 0xf.
3068 enabled_channels
|= (1 << chan
);
3069 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3070 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3073 /* The PARAM export is duplicated. Kill it. */
3074 LLVMInstructionEraseFromParent(exp
->inst
);
3076 /* Change OFFSET to the matching export. */
3077 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3078 if (vs_output_param_offset
[i
] == exp
->offset
) {
3079 vs_output_param_offset
[i
] = match
->offset
;
3086 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3087 LLVMValueRef main_fn
,
3088 uint8_t *vs_output_param_offset
,
3089 uint32_t num_outputs
,
3090 uint8_t *num_param_exports
)
3092 LLVMBasicBlockRef bb
;
3093 bool removed_any
= false;
3094 struct ac_vs_exports exports
;
3098 /* Process all LLVM instructions. */
3099 bb
= LLVMGetFirstBasicBlock(main_fn
);
3101 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3104 LLVMValueRef cur
= inst
;
3105 inst
= LLVMGetNextInstruction(inst
);
3106 struct ac_vs_exp_inst exp
;
3108 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3111 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3113 if (!ac_llvm_is_function(callee
))
3116 const char *name
= LLVMGetValueName(callee
);
3117 unsigned num_args
= LLVMCountParams(callee
);
3119 /* Check if this is an export instruction. */
3120 if ((num_args
!= 9 && num_args
!= 8) ||
3121 (strcmp(name
, "llvm.SI.export") &&
3122 strcmp(name
, "llvm.amdgcn.exp.f32")))
3125 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3126 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3128 if (target
< V_008DFC_SQ_EXP_PARAM
)
3131 target
-= V_008DFC_SQ_EXP_PARAM
;
3133 /* Parse the instruction. */
3134 memset(&exp
, 0, sizeof(exp
));
3135 exp
.offset
= target
;
3138 for (unsigned i
= 0; i
< 4; i
++) {
3139 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3141 exp
.chan
[i
].value
= v
;
3143 if (LLVMIsUndef(v
)) {
3144 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3145 } else if (LLVMIsAConstantFP(v
)) {
3146 LLVMBool loses_info
;
3147 exp
.chan
[i
].type
= AC_IR_CONST
;
3148 exp
.chan
[i
].const_float
=
3149 LLVMConstRealGetDouble(v
, &loses_info
);
3151 exp
.chan
[i
].type
= AC_IR_VALUE
;
3155 /* Eliminate constant and duplicated PARAM exports. */
3156 if (ac_eliminate_const_output(vs_output_param_offset
,
3157 num_outputs
, &exp
) ||
3158 ac_eliminate_duplicated_output(ctx
,
3159 vs_output_param_offset
,
3160 num_outputs
, &exports
,
3164 exports
.exp
[exports
.num
++] = exp
;
3167 bb
= LLVMGetNextBasicBlock(bb
);
3170 /* Remove holes in export memory due to removed PARAM exports.
3171 * This is done by renumbering all PARAM exports.
3174 uint8_t old_offset
[VARYING_SLOT_MAX
];
3177 /* Make a copy of the offsets. We need the old version while
3178 * we are modifying some of them. */
3179 memcpy(old_offset
, vs_output_param_offset
,
3180 sizeof(old_offset
));
3182 for (i
= 0; i
< exports
.num
; i
++) {
3183 unsigned offset
= exports
.exp
[i
].offset
;
3185 /* Update vs_output_param_offset. Multiple outputs can
3186 * have the same offset.
3188 for (out
= 0; out
< num_outputs
; out
++) {
3189 if (old_offset
[out
] == offset
)
3190 vs_output_param_offset
[out
] = i
;
3193 /* Change the PARAM offset in the instruction. */
3194 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3195 LLVMConstInt(ctx
->i32
,
3196 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3198 *num_param_exports
= exports
.num
;
3202 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3204 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3205 ac_build_intrinsic(ctx
,
3206 "llvm.amdgcn.init.exec", ctx
->voidt
,
3207 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3210 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3212 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3213 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3214 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3218 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3219 LLVMValueRef dw_addr
)
3221 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3224 void ac_lds_store(struct ac_llvm_context
*ctx
,
3225 LLVMValueRef dw_addr
,
3228 value
= ac_to_integer(ctx
, value
);
3229 ac_build_indexed_store(ctx
, ctx
->lds
,
3233 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3234 LLVMTypeRef dst_type
,
3237 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3238 const char *intrin_name
;
3242 switch (src0_bitsize
) {
3244 intrin_name
= "llvm.cttz.i64";
3249 intrin_name
= "llvm.cttz.i32";
3254 intrin_name
= "llvm.cttz.i16";
3259 intrin_name
= "llvm.cttz.i8";
3264 unreachable(!"invalid bitsize");
3267 LLVMValueRef params
[2] = {
3270 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3271 * add special code to check for x=0. The reason is that
3272 * the LLVM behavior for x=0 is different from what we
3273 * need here. However, LLVM also assumes that ffs(x) is
3274 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3275 * a conditional assignment to handle 0 is still required.
3277 * The hardware already implements the correct behavior.
3282 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3284 AC_FUNC_ATTR_READNONE
);
3286 if (src0_bitsize
== 64) {
3287 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3288 } else if (src0_bitsize
< 32) {
3289 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3292 /* TODO: We need an intrinsic to skip this conditional. */
3293 /* Check for zero: */
3294 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3297 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3300 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3302 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3305 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3307 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3310 static struct ac_llvm_flow
*
3311 get_current_flow(struct ac_llvm_context
*ctx
)
3313 if (ctx
->flow
->depth
> 0)
3314 return &ctx
->flow
->stack
[ctx
->flow
->depth
- 1];
3318 static struct ac_llvm_flow
*
3319 get_innermost_loop(struct ac_llvm_context
*ctx
)
3321 for (unsigned i
= ctx
->flow
->depth
; i
> 0; --i
) {
3322 if (ctx
->flow
->stack
[i
- 1].loop_entry_block
)
3323 return &ctx
->flow
->stack
[i
- 1];
3328 static struct ac_llvm_flow
*
3329 push_flow(struct ac_llvm_context
*ctx
)
3331 struct ac_llvm_flow
*flow
;
3333 if (ctx
->flow
->depth
>= ctx
->flow
->depth_max
) {
3334 unsigned new_max
= MAX2(ctx
->flow
->depth
<< 1,
3335 AC_LLVM_INITIAL_CF_DEPTH
);
3337 ctx
->flow
->stack
= realloc(ctx
->flow
->stack
, new_max
* sizeof(*ctx
->flow
->stack
));
3338 ctx
->flow
->depth_max
= new_max
;
3341 flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
];
3344 flow
->next_block
= NULL
;
3345 flow
->loop_entry_block
= NULL
;
3349 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3353 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3354 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3357 /* Append a basic block at the level of the parent flow.
3359 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3362 assert(ctx
->flow
->depth
>= 1);
3364 if (ctx
->flow
->depth
>= 2) {
3365 struct ac_llvm_flow
*flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
- 2];
3367 return LLVMInsertBasicBlockInContext(ctx
->context
,
3368 flow
->next_block
, name
);
3371 LLVMValueRef main_fn
=
3372 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3373 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3376 /* Emit a branch to the given default target for the current block if
3377 * applicable -- that is, if the current block does not already contain a
3378 * branch from a break or continue.
3380 static void emit_default_branch(LLVMBuilderRef builder
,
3381 LLVMBasicBlockRef target
)
3383 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3384 LLVMBuildBr(builder
, target
);
3387 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3389 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3390 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3391 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3392 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3393 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3394 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3397 void ac_build_break(struct ac_llvm_context
*ctx
)
3399 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3400 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3403 void ac_build_continue(struct ac_llvm_context
*ctx
)
3405 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3406 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3409 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3411 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3412 LLVMBasicBlockRef endif_block
;
3414 assert(!current_branch
->loop_entry_block
);
3416 endif_block
= append_basic_block(ctx
, "ENDIF");
3417 emit_default_branch(ctx
->builder
, endif_block
);
3419 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3420 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3422 current_branch
->next_block
= endif_block
;
3425 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3427 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3429 assert(!current_branch
->loop_entry_block
);
3431 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3432 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3433 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3438 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3440 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3442 assert(current_loop
->loop_entry_block
);
3444 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3446 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3447 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3451 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3453 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3454 LLVMBasicBlockRef if_block
;
3456 if_block
= append_basic_block(ctx
, "IF");
3457 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3458 set_basicblock_name(if_block
, "if", label_id
);
3459 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3460 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3463 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3466 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3467 value
, ctx
->f32_0
, "");
3468 ac_build_ifcc(ctx
, cond
, label_id
);
3471 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3474 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3475 ac_to_integer(ctx
, value
),
3477 ac_build_ifcc(ctx
, cond
, label_id
);
3480 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3483 LLVMBuilderRef builder
= ac
->builder
;
3484 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3485 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3486 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3487 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3488 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3492 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3494 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3497 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3498 LLVMDisposeBuilder(first_builder
);
3502 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3503 LLVMTypeRef type
, const char *name
)
3505 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3506 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3510 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3513 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3514 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3515 LLVMPointerType(type
, addr_space
), "");
3518 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3521 unsigned num_components
= ac_get_llvm_num_components(value
);
3522 if (count
== num_components
)
3525 LLVMValueRef masks
[MAX2(count
, 2)];
3526 masks
[0] = ctx
->i32_0
;
3527 masks
[1] = ctx
->i32_1
;
3528 for (unsigned i
= 2; i
< count
; i
++)
3529 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3532 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3535 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3536 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3539 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3540 unsigned rshift
, unsigned bitwidth
)
3542 LLVMValueRef value
= param
;
3544 value
= LLVMBuildLShr(ctx
->builder
, value
,
3545 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3547 if (rshift
+ bitwidth
< 32) {
3548 unsigned mask
= (1 << bitwidth
) - 1;
3549 value
= LLVMBuildAnd(ctx
->builder
, value
,
3550 LLVMConstInt(ctx
->i32
, mask
, false), "");
3555 /* Adjust the sample index according to FMASK.
3557 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3558 * which is the identity mapping. Each nibble says which physical sample
3559 * should be fetched to get that sample.
3561 * For example, 0x11111100 means there are only 2 samples stored and
3562 * the second sample covers 3/4 of the pixel. When reading samples 0
3563 * and 1, return physical sample 0 (determined by the first two 0s
3564 * in FMASK), otherwise return physical sample 1.
3566 * The sample index should be adjusted as follows:
3567 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3569 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3570 LLVMValueRef
*addr
, bool is_array_tex
)
3572 struct ac_image_args fmask_load
= {};
3573 fmask_load
.opcode
= ac_image_load
;
3574 fmask_load
.resource
= fmask
;
3575 fmask_load
.dmask
= 0xf;
3576 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3577 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3579 fmask_load
.coords
[0] = addr
[0];
3580 fmask_load
.coords
[1] = addr
[1];
3582 fmask_load
.coords
[2] = addr
[2];
3584 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3585 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3588 /* Apply the formula. */
3589 unsigned sample_chan
= is_array_tex
? 3 : 2;
3590 LLVMValueRef final_sample
;
3591 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3592 LLVMConstInt(ac
->i32
, 4, 0), "");
3593 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3594 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3595 * with EQAA, so those will map to 0. */
3596 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3597 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3599 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3600 * resource descriptor is 0 (invalid).
3603 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3604 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3605 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3607 /* Replace the MSAA sample index. */
3608 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3609 addr
[sample_chan
], "");
3613 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3614 LLVMValueRef lane
, bool with_opt_barrier
)
3616 LLVMTypeRef type
= LLVMTypeOf(src
);
3617 LLVMValueRef result
;
3619 if (with_opt_barrier
)
3620 ac_build_optimization_barrier(ctx
, &src
);
3622 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3624 lane
= LLVMBuildZExt(ctx
->builder
, lane
, ctx
->i32
, "");
3626 result
= ac_build_intrinsic(ctx
,
3627 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3628 ctx
->i32
, (LLVMValueRef
[]) { src
, lane
},
3629 lane
== NULL
? 1 : 2,
3630 AC_FUNC_ATTR_READNONE
|
3631 AC_FUNC_ATTR_CONVERGENT
);
3633 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3637 ac_build_readlane_common(struct ac_llvm_context
*ctx
,
3638 LLVMValueRef src
, LLVMValueRef lane
,
3639 bool with_opt_barrier
)
3641 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3642 src
= ac_to_integer(ctx
, src
);
3643 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3647 assert(bits
% 32 == 0);
3648 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3649 LLVMValueRef src_vector
=
3650 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3651 ret
= LLVMGetUndef(vec_type
);
3652 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3653 LLVMValueRef ret_comp
;
3655 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3656 LLVMConstInt(ctx
->i32
, i
, 0), "");
3658 ret_comp
= _ac_build_readlane(ctx
, src
, lane
,
3661 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3662 LLVMConstInt(ctx
->i32
, i
, 0), "");
3665 ret
= _ac_build_readlane(ctx
, src
, lane
, with_opt_barrier
);
3668 if (LLVMGetTypeKind(src_type
) == LLVMPointerTypeKind
)
3669 return LLVMBuildIntToPtr(ctx
->builder
, ret
, src_type
, "");
3670 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3674 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3676 * The optimization barrier is not needed if the value is the same in all lanes
3677 * or if this is called in the outermost block.
3681 * @param lane - id of the lane or NULL for the first active lane
3682 * @return value of the lane
3684 LLVMValueRef
ac_build_readlane_no_opt_barrier(struct ac_llvm_context
*ctx
,
3685 LLVMValueRef src
, LLVMValueRef lane
)
3687 return ac_build_readlane_common(ctx
, src
, lane
, false);
3692 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3694 return ac_build_readlane_common(ctx
, src
, lane
, true);
3698 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3700 return ac_build_intrinsic(ctx
, "llvm.amdgcn.writelane", ctx
->i32
,
3701 (LLVMValueRef
[]) {value
, lane
, src
}, 3,
3702 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3706 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3708 if (ctx
->wave_size
== 32) {
3709 return ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3710 (LLVMValueRef
[]) { mask
, ctx
->i32_0
},
3711 2, AC_FUNC_ATTR_READNONE
);
3713 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3714 LLVMVectorType(ctx
->i32
, 2),
3716 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3718 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3721 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3722 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3723 2, AC_FUNC_ATTR_READNONE
);
3724 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3725 (LLVMValueRef
[]) { mask_hi
, val
},
3726 2, AC_FUNC_ATTR_READNONE
);
3731 _dpp_quad_perm
= 0x000,
3732 _dpp_row_sl
= 0x100,
3733 _dpp_row_sr
= 0x110,
3734 _dpp_row_rr
= 0x120,
3739 dpp_row_mirror
= 0x140,
3740 dpp_row_half_mirror
= 0x141,
3741 dpp_row_bcast15
= 0x142,
3742 dpp_row_bcast31
= 0x143
3745 static inline enum dpp_ctrl
3746 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3748 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3749 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3752 static inline enum dpp_ctrl
3753 dpp_row_sl(unsigned amount
)
3755 assert(amount
> 0 && amount
< 16);
3756 return _dpp_row_sl
| amount
;
3759 static inline enum dpp_ctrl
3760 dpp_row_sr(unsigned amount
)
3762 assert(amount
> 0 && amount
< 16);
3763 return _dpp_row_sr
| amount
;
3767 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3768 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3771 LLVMTypeRef type
= LLVMTypeOf(src
);
3774 old
= LLVMBuildZExt(ctx
->builder
, old
, ctx
->i32
, "");
3775 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3777 res
= ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32", ctx
->i32
,
3780 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3781 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3782 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3783 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3784 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3786 return LLVMBuildTrunc(ctx
->builder
, res
, type
, "");
3790 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3791 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3794 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3795 src
= ac_to_integer(ctx
, src
);
3796 old
= ac_to_integer(ctx
, old
);
3797 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3800 assert(bits
% 32 == 0);
3801 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3802 LLVMValueRef src_vector
=
3803 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3804 LLVMValueRef old_vector
=
3805 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3806 ret
= LLVMGetUndef(vec_type
);
3807 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3808 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3809 LLVMConstInt(ctx
->i32
, i
,
3811 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3812 LLVMConstInt(ctx
->i32
, i
,
3814 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3819 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3821 LLVMConstInt(ctx
->i32
, i
,
3825 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3826 bank_mask
, bound_ctrl
);
3828 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3832 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3833 bool exchange_rows
, bool bound_ctrl
)
3835 LLVMTypeRef type
= LLVMTypeOf(src
);
3836 LLVMValueRef result
;
3838 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3840 LLVMValueRef args
[6] = {
3843 LLVMConstInt(ctx
->i32
, sel
, false),
3844 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3845 ctx
->i1true
, /* fi */
3846 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3849 result
= ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3850 : "llvm.amdgcn.permlane16",
3852 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3854 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3858 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3859 bool exchange_rows
, bool bound_ctrl
)
3861 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3862 src
= ac_to_integer(ctx
, src
);
3863 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3866 assert(bits
% 32 == 0);
3867 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3868 LLVMValueRef src_vector
=
3869 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3870 ret
= LLVMGetUndef(vec_type
);
3871 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3872 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3873 LLVMConstInt(ctx
->i32
, i
,
3875 LLVMValueRef ret_comp
=
3876 _ac_build_permlane16(ctx
, src
, sel
,
3879 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3881 LLVMConstInt(ctx
->i32
, i
,
3885 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3888 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3891 static inline unsigned
3892 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3894 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3895 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3899 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3901 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3904 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3906 ret
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle", ctx
->i32
,
3908 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3909 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3911 return LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3915 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3917 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3918 src
= ac_to_integer(ctx
, src
);
3919 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3922 assert(bits
% 32 == 0);
3923 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3924 LLVMValueRef src_vector
=
3925 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3926 ret
= LLVMGetUndef(vec_type
);
3927 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3928 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3929 LLVMConstInt(ctx
->i32
, i
,
3931 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3933 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3935 LLVMConstInt(ctx
->i32
, i
,
3939 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3941 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3945 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3947 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3948 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3949 char name
[32], type
[8];
3952 src
= ac_to_integer(ctx
, src
);
3955 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3957 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3958 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3959 ret
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3960 (LLVMValueRef
[]) { src
}, 1,
3961 AC_FUNC_ATTR_READNONE
);
3964 ret
= LLVMBuildTrunc(ctx
->builder
, ret
,
3965 ac_to_integer_type(ctx
, src_type
), "");
3967 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3971 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3972 LLVMValueRef inactive
)
3974 char name
[33], type
[8];
3975 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3976 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3977 src
= ac_to_integer(ctx
, src
);
3978 inactive
= ac_to_integer(ctx
, inactive
);
3981 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3982 inactive
= LLVMBuildZExt(ctx
->builder
, inactive
, ctx
->i32
, "");
3985 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3986 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3988 ac_build_intrinsic(ctx
, name
,
3989 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3991 AC_FUNC_ATTR_READNONE
|
3992 AC_FUNC_ATTR_CONVERGENT
);
3994 ret
= LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
4000 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
4002 if (type_size
== 1) {
4004 case nir_op_iadd
: return ctx
->i8_0
;
4005 case nir_op_imul
: return ctx
->i8_1
;
4006 case nir_op_imin
: return LLVMConstInt(ctx
->i8
, INT8_MAX
, 0);
4007 case nir_op_umin
: return LLVMConstInt(ctx
->i8
, UINT8_MAX
, 0);
4008 case nir_op_imax
: return LLVMConstInt(ctx
->i8
, INT8_MIN
, 0);
4009 case nir_op_umax
: return ctx
->i8_0
;
4010 case nir_op_iand
: return LLVMConstInt(ctx
->i8
, -1, 0);
4011 case nir_op_ior
: return ctx
->i8_0
;
4012 case nir_op_ixor
: return ctx
->i8_0
;
4014 unreachable("bad reduction intrinsic");
4016 } else if (type_size
== 2) {
4018 case nir_op_iadd
: return ctx
->i16_0
;
4019 case nir_op_fadd
: return ctx
->f16_0
;
4020 case nir_op_imul
: return ctx
->i16_1
;
4021 case nir_op_fmul
: return ctx
->f16_1
;
4022 case nir_op_imin
: return LLVMConstInt(ctx
->i16
, INT16_MAX
, 0);
4023 case nir_op_umin
: return LLVMConstInt(ctx
->i16
, UINT16_MAX
, 0);
4024 case nir_op_fmin
: return LLVMConstReal(ctx
->f16
, INFINITY
);
4025 case nir_op_imax
: return LLVMConstInt(ctx
->i16
, INT16_MIN
, 0);
4026 case nir_op_umax
: return ctx
->i16_0
;
4027 case nir_op_fmax
: return LLVMConstReal(ctx
->f16
, -INFINITY
);
4028 case nir_op_iand
: return LLVMConstInt(ctx
->i16
, -1, 0);
4029 case nir_op_ior
: return ctx
->i16_0
;
4030 case nir_op_ixor
: return ctx
->i16_0
;
4032 unreachable("bad reduction intrinsic");
4034 } else if (type_size
== 4) {
4036 case nir_op_iadd
: return ctx
->i32_0
;
4037 case nir_op_fadd
: return ctx
->f32_0
;
4038 case nir_op_imul
: return ctx
->i32_1
;
4039 case nir_op_fmul
: return ctx
->f32_1
;
4040 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
4041 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
4042 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
4043 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
4044 case nir_op_umax
: return ctx
->i32_0
;
4045 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
4046 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
4047 case nir_op_ior
: return ctx
->i32_0
;
4048 case nir_op_ixor
: return ctx
->i32_0
;
4050 unreachable("bad reduction intrinsic");
4052 } else { /* type_size == 64bit */
4054 case nir_op_iadd
: return ctx
->i64_0
;
4055 case nir_op_fadd
: return ctx
->f64_0
;
4056 case nir_op_imul
: return ctx
->i64_1
;
4057 case nir_op_fmul
: return ctx
->f64_1
;
4058 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
4059 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4060 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
4061 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
4062 case nir_op_umax
: return ctx
->i64_0
;
4063 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
4064 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
4065 case nir_op_ior
: return ctx
->i64_0
;
4066 case nir_op_ixor
: return ctx
->i64_0
;
4068 unreachable("bad reduction intrinsic");
4074 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
4076 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
4077 bool _32bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 4;
4079 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
4080 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
4081 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
4082 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
4083 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
4084 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
4086 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
4087 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
4089 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
4090 _64bit
? "llvm.minnum.f64" : _32bit
? "llvm.minnum.f32" : "llvm.minnum.f16",
4091 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4092 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4093 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
4094 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
4096 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
4097 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
4099 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
4100 _64bit
? "llvm.maxnum.f64" : _32bit
? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4101 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4102 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4103 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
4104 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
4105 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
4107 unreachable("bad reduction intrinsic");
4112 * \param src The value to shift.
4113 * \param identity The value to use the first lane.
4114 * \param maxprefix specifies that the result only needs to be correct for a
4115 * prefix of this many threads
4116 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4119 ac_wavefront_shift_right_1(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4120 LLVMValueRef identity
, unsigned maxprefix
)
4122 if (ctx
->chip_class
>= GFX10
) {
4123 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4124 LLVMValueRef active
, tmp1
, tmp2
;
4125 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4127 tmp1
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4129 tmp2
= ac_build_permlane16(ctx
, src
, (uint64_t)~0, true, false);
4131 if (maxprefix
> 32) {
4132 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4133 LLVMConstInt(ctx
->i32
, 32, false), "");
4135 tmp2
= LLVMBuildSelect(ctx
->builder
, active
,
4136 ac_build_readlane(ctx
, src
,
4137 LLVMConstInt(ctx
->i32
, 31, false)),
4140 active
= LLVMBuildOr(ctx
->builder
, active
,
4141 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4142 LLVMBuildAnd(ctx
->builder
, tid
,
4143 LLVMConstInt(ctx
->i32
, 0x1f, false), ""),
4144 LLVMConstInt(ctx
->i32
, 0x10, false), ""), "");
4145 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4146 } else if (maxprefix
> 16) {
4147 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4148 LLVMConstInt(ctx
->i32
, 16, false), "");
4150 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4152 } else if (ctx
->chip_class
>= GFX8
) {
4153 return ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
4156 /* wavefront shift_right by 1 on SI/CI */
4157 LLVMValueRef active
, tmp1
, tmp2
;
4158 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4159 tmp1
= ac_build_ds_swizzle(ctx
, src
, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4160 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4161 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4162 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x7, 0), ""),
4163 LLVMConstInt(ctx
->i32
, 0x4, 0), "");
4164 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4165 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4166 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4167 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0xf, 0), ""),
4168 LLVMConstInt(ctx
->i32
, 0x8, 0), "");
4169 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4170 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4171 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4172 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x1f, 0), ""),
4173 LLVMConstInt(ctx
->i32
, 0x10, 0), "");
4174 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4175 tmp2
= ac_build_readlane(ctx
, src
, LLVMConstInt(ctx
->i32
, 31, 0));
4176 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), "");
4177 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4178 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 0, 0), "");
4179 return LLVMBuildSelect(ctx
->builder
, active
, identity
, tmp1
, "");
4183 * \param maxprefix specifies that the result only needs to be correct for a
4184 * prefix of this many threads
4187 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
4188 unsigned maxprefix
, bool inclusive
)
4190 LLVMValueRef result
, tmp
;
4193 src
= ac_wavefront_shift_right_1(ctx
, src
, identity
, maxprefix
);
4197 if (ctx
->chip_class
<= GFX7
) {
4198 assert(maxprefix
== 64);
4199 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4200 LLVMValueRef active
;
4201 tmp
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4202 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4203 LLVMBuildAnd(ctx
->builder
, tid
, ctx
->i32_1
, ""),
4205 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4206 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4207 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4208 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4209 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 2, 0), ""),
4211 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4212 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4213 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4214 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4215 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 4, 0), ""),
4217 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4218 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4219 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4220 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4221 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 8, 0), ""),
4223 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4224 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4225 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4226 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4227 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 16, 0), ""),
4229 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4230 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4231 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, 0));
4232 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4233 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), ""),
4235 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4236 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4242 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4243 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4246 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
4247 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4250 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
4251 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4254 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
4255 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4258 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
4259 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4260 if (maxprefix
<= 16)
4263 if (ctx
->chip_class
>= GFX10
) {
4264 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4265 LLVMValueRef active
;
4267 tmp
= ac_build_permlane16(ctx
, result
, ~(uint64_t)0, true, false);
4269 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4270 LLVMBuildAnd(ctx
->builder
, tid
,
4271 LLVMConstInt(ctx
->i32
, 16, false), ""),
4274 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4276 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4278 if (maxprefix
<= 32)
4281 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4283 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, tid
,
4284 LLVMConstInt(ctx
->i32
, 32, false), "");
4286 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4288 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4292 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4293 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4294 if (maxprefix
<= 32)
4296 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4297 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4302 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4304 LLVMValueRef result
;
4306 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4307 LLVMBuilderRef builder
= ctx
->builder
;
4308 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4309 result
= ac_build_ballot(ctx
, src
);
4310 result
= ac_build_mbcnt(ctx
, result
);
4311 result
= LLVMBuildAdd(builder
, result
, src
, "");
4315 ac_build_optimization_barrier(ctx
, &src
);
4317 LLVMValueRef identity
=
4318 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4319 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4320 LLVMTypeOf(identity
), "");
4321 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, true);
4323 return ac_build_wwm(ctx
, result
);
4327 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4329 LLVMValueRef result
;
4331 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4332 LLVMBuilderRef builder
= ctx
->builder
;
4333 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4334 result
= ac_build_ballot(ctx
, src
);
4335 result
= ac_build_mbcnt(ctx
, result
);
4339 ac_build_optimization_barrier(ctx
, &src
);
4341 LLVMValueRef identity
=
4342 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4343 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4344 LLVMTypeOf(identity
), "");
4345 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, false);
4347 return ac_build_wwm(ctx
, result
);
4351 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4353 if (cluster_size
== 1) return src
;
4354 ac_build_optimization_barrier(ctx
, &src
);
4355 LLVMValueRef result
, swap
;
4356 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4357 ac_get_type_size(LLVMTypeOf(src
)));
4358 result
= LLVMBuildBitCast(ctx
->builder
,
4359 ac_build_set_inactive(ctx
, src
, identity
),
4360 LLVMTypeOf(identity
), "");
4361 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4362 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4363 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4365 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4366 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4367 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4369 if (ctx
->chip_class
>= GFX8
)
4370 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4372 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4373 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4374 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4376 if (ctx
->chip_class
>= GFX8
)
4377 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4379 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4380 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4381 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4383 if (ctx
->chip_class
>= GFX10
)
4384 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4385 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4386 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4388 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4389 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4390 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4392 if (ctx
->chip_class
>= GFX8
) {
4393 if (ctx
->wave_size
== 64) {
4394 if (ctx
->chip_class
>= GFX10
)
4395 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4397 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4398 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4399 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4402 return ac_build_wwm(ctx
, result
);
4404 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4405 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4406 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4407 return ac_build_wwm(ctx
, result
);
4412 * "Top half" of a scan that reduces per-wave values across an entire
4415 * The source value must be present in the highest lane of the wave, and the
4416 * highest lane must be live.
4419 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4421 if (ws
->maxwaves
<= 1)
4424 const LLVMValueRef last_lane
= LLVMConstInt(ctx
->i32
, ctx
->wave_size
- 1, false);
4425 LLVMBuilderRef builder
= ctx
->builder
;
4426 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4429 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, last_lane
, "");
4430 ac_build_ifcc(ctx
, tmp
, 1000);
4431 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4432 ac_build_endif(ctx
, 1000);
4436 * "Bottom half" of a scan that reduces per-wave values across an entire
4439 * The caller must place a barrier between the top and bottom halves.
4442 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4444 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4445 const LLVMValueRef identity
=
4446 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4448 if (ws
->maxwaves
<= 1) {
4449 ws
->result_reduce
= ws
->src
;
4450 ws
->result_inclusive
= ws
->src
;
4451 ws
->result_exclusive
= identity
;
4454 assert(ws
->maxwaves
<= 32);
4456 LLVMBuilderRef builder
= ctx
->builder
;
4457 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4458 LLVMBasicBlockRef bbs
[2];
4459 LLVMValueRef phivalues_scan
[2];
4460 LLVMValueRef tmp
, tmp2
;
4462 bbs
[0] = LLVMGetInsertBlock(builder
);
4463 phivalues_scan
[0] = LLVMGetUndef(type
);
4465 if (ws
->enable_reduce
)
4466 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4467 else if (ws
->enable_inclusive
)
4468 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4470 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4471 ac_build_ifcc(ctx
, tmp
, 1001);
4473 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4475 ac_build_optimization_barrier(ctx
, &tmp
);
4477 bbs
[1] = LLVMGetInsertBlock(builder
);
4478 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4480 ac_build_endif(ctx
, 1001);
4482 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4484 if (ws
->enable_reduce
) {
4485 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4486 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4488 if (ws
->enable_inclusive
)
4489 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4490 if (ws
->enable_exclusive
) {
4491 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4492 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4493 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4494 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4499 * Inclusive scan of a per-wave value across an entire workgroup.
4501 * This implies an s_barrier instruction.
4503 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4504 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4505 * useful manner because of the barrier in the algorithm.)
4508 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4510 ac_build_wg_wavescan_top(ctx
, ws
);
4511 ac_build_s_barrier(ctx
);
4512 ac_build_wg_wavescan_bottom(ctx
, ws
);
4516 * "Top half" of a scan that reduces per-thread values across an entire
4519 * All lanes must be active when this code runs.
4522 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4524 if (ws
->enable_exclusive
) {
4525 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4526 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4527 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4528 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4530 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4533 bool enable_inclusive
= ws
->enable_inclusive
;
4534 bool enable_exclusive
= ws
->enable_exclusive
;
4535 ws
->enable_inclusive
= false;
4536 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4537 ac_build_wg_wavescan_top(ctx
, ws
);
4538 ws
->enable_inclusive
= enable_inclusive
;
4539 ws
->enable_exclusive
= enable_exclusive
;
4543 * "Bottom half" of a scan that reduces per-thread values across an entire
4546 * The caller must place a barrier between the top and bottom halves.
4549 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4551 bool enable_inclusive
= ws
->enable_inclusive
;
4552 bool enable_exclusive
= ws
->enable_exclusive
;
4553 ws
->enable_inclusive
= false;
4554 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4555 ac_build_wg_wavescan_bottom(ctx
, ws
);
4556 ws
->enable_inclusive
= enable_inclusive
;
4557 ws
->enable_exclusive
= enable_exclusive
;
4559 /* ws->result_reduce is already the correct value */
4560 if (ws
->enable_inclusive
)
4561 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4562 if (ws
->enable_exclusive
)
4563 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4567 * A scan that reduces per-thread values across an entire workgroup.
4569 * The caller must ensure that all lanes are active when this code runs
4570 * (WWM is insufficient!), because there is an implied barrier.
4573 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4575 ac_build_wg_scan_top(ctx
, ws
);
4576 ac_build_s_barrier(ctx
);
4577 ac_build_wg_scan_bottom(ctx
, ws
);
4581 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4582 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4584 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4585 if (ctx
->chip_class
>= GFX8
) {
4586 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4588 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4593 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4595 LLVMTypeRef type
= LLVMTypeOf(src
);
4596 LLVMValueRef result
;
4598 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4599 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
4601 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4602 (LLVMValueRef
[]) {index
, src
}, 2,
4603 AC_FUNC_ATTR_READNONE
|
4604 AC_FUNC_ATTR_CONVERGENT
);
4605 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
4609 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4615 if (bitsize
== 16) {
4616 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4618 } else if (bitsize
== 32) {
4619 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4622 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4626 LLVMValueRef params
[] = {
4629 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4630 AC_FUNC_ATTR_READNONE
);
4633 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4639 if (bitsize
== 16) {
4640 intr
= "llvm.amdgcn.frexp.mant.f16";
4642 } else if (bitsize
== 32) {
4643 intr
= "llvm.amdgcn.frexp.mant.f32";
4646 intr
= "llvm.amdgcn.frexp.mant.f64";
4650 LLVMValueRef params
[] = {
4653 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4654 AC_FUNC_ATTR_READNONE
);
4658 ac_build_canonicalize(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4664 if (bitsize
== 16) {
4665 intr
= "llvm.canonicalize.f16";
4667 } else if (bitsize
== 32) {
4668 intr
= "llvm.canonicalize.f32";
4671 intr
= "llvm.canonicalize.f64";
4675 LLVMValueRef params
[] = {
4678 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4679 AC_FUNC_ATTR_READNONE
);
4683 * this takes an I,J coordinate pair,
4684 * and works out the X and Y derivatives.
4685 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4688 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4690 LLVMValueRef result
[4], a
;
4693 for (i
= 0; i
< 2; i
++) {
4694 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4695 LLVMConstInt(ctx
->i32
, i
, false), "");
4696 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4697 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4699 return ac_build_gather_values(ctx
, result
, 4);
4703 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4705 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4707 AC_FUNC_ATTR_READNONE
);
4708 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4709 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4713 ac_build_is_helper_invocation(struct ac_llvm_context
*ctx
)
4715 if (!ctx
->postponed_kill
)
4716 return ac_build_load_helper_invocation(ctx
);
4718 /* !(exact && postponed) */
4719 LLVMValueRef exact
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4721 AC_FUNC_ATTR_READNONE
);
4723 LLVMValueRef postponed
= LLVMBuildLoad(ctx
->builder
, ctx
->postponed_kill
, "");
4724 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, exact
, postponed
, "");
4726 return LLVMBuildSelect(ctx
->builder
, result
, ctx
->i32_0
,
4727 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false), "");
4730 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4731 LLVMValueRef
*args
, unsigned num_args
)
4733 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4734 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));
4739 ac_export_mrt_z(struct ac_llvm_context
*ctx
, LLVMValueRef depth
,
4740 LLVMValueRef stencil
, LLVMValueRef samplemask
,
4741 struct ac_export_args
*args
)
4744 unsigned format
= ac_get_spi_shader_z_format(depth
!= NULL
,
4746 samplemask
!= NULL
);
4748 assert(depth
|| stencil
|| samplemask
);
4750 memset(args
, 0, sizeof(*args
));
4752 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
4753 args
->done
= 1; /* DONE bit */
4755 /* Specify the target we are exporting */
4756 args
->target
= V_008DFC_SQ_EXP_MRTZ
;
4758 args
->compr
= 0; /* COMP flag */
4759 args
->out
[0] = LLVMGetUndef(ctx
->f32
); /* R, depth */
4760 args
->out
[1] = LLVMGetUndef(ctx
->f32
); /* G, stencil test val[0:7], stencil op val[8:15] */
4761 args
->out
[2] = LLVMGetUndef(ctx
->f32
); /* B, sample mask */
4762 args
->out
[3] = LLVMGetUndef(ctx
->f32
); /* A, alpha to mask */
4764 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
4766 args
->compr
= 1; /* COMPR flag */
4769 /* Stencil should be in X[23:16]. */
4770 stencil
= ac_to_integer(ctx
, stencil
);
4771 stencil
= LLVMBuildShl(ctx
->builder
, stencil
,
4772 LLVMConstInt(ctx
->i32
, 16, 0), "");
4773 args
->out
[0] = ac_to_float(ctx
, stencil
);
4777 /* SampleMask should be in Y[15:0]. */
4778 args
->out
[1] = samplemask
;
4783 args
->out
[0] = depth
;
4787 args
->out
[1] = stencil
;
4791 args
->out
[2] = samplemask
;
4796 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4797 * at the X writemask component. */
4798 if (ctx
->chip_class
== GFX6
&&
4799 ctx
->family
!= CHIP_OLAND
&&
4800 ctx
->family
!= CHIP_HAINAN
)
4803 /* Specify which components to enable */
4804 args
->enabled_channels
= mask
;
4807 /* Send GS Alloc Req message from the first wave of the group to SPI.
4808 * Message payload is:
4809 * - bits 0..10: vertices in group
4810 * - bits 12..22: primitives in group
4812 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context
*ctx
, LLVMValueRef wave_id
,
4813 LLVMValueRef vtx_cnt
, LLVMValueRef prim_cnt
)
4815 LLVMBuilderRef builder
= ctx
->builder
;
4817 bool export_dummy_prim
= false;
4819 /* HW workaround for a GPU hang with 100% culling.
4820 * We always have to export at least 1 primitive.
4821 * Export a degenerate triangle using vertex 0 for all 3 vertices.
4823 if (prim_cnt
== ctx
->i32_0
&&
4824 (ctx
->family
== CHIP_NAVI10
||
4825 ctx
->family
== CHIP_NAVI12
||
4826 ctx
->family
== CHIP_NAVI14
)) {
4827 assert(vtx_cnt
== ctx
->i32_0
);
4828 prim_cnt
= ctx
->i32_1
;
4829 vtx_cnt
= ctx
->i32_1
;
4830 export_dummy_prim
= true;
4833 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, wave_id
, ctx
->i32_0
, ""), 5020);
4835 tmp
= LLVMBuildShl(builder
, prim_cnt
, LLVMConstInt(ctx
->i32
, 12, false),"");
4836 tmp
= LLVMBuildOr(builder
, tmp
, vtx_cnt
, "");
4837 ac_build_sendmsg(ctx
, AC_SENDMSG_GS_ALLOC_REQ
, tmp
);
4839 if (export_dummy_prim
) {
4840 struct ac_ngg_prim prim
= {};
4841 /* The vertex indices are 0,0,0. */
4842 prim
.passthrough
= ctx
->i32_0
;
4844 struct ac_export_args pos
= {};
4845 pos
.out
[0] = pos
.out
[1] = pos
.out
[2] = pos
.out
[3] = ctx
->f32_0
;
4846 pos
.target
= V_008DFC_SQ_EXP_POS
;
4847 pos
.enabled_channels
= 0xf;
4850 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(ctx
),
4851 ctx
->i32_0
, ""), 5021);
4852 ac_build_export_prim(ctx
, &prim
);
4853 ac_build_export(ctx
, &pos
);
4854 ac_build_endif(ctx
, 5021);
4857 ac_build_endif(ctx
, 5020);
4860 LLVMValueRef
ac_pack_prim_export(struct ac_llvm_context
*ctx
,
4861 const struct ac_ngg_prim
*prim
)
4863 /* The prim export format is:
4864 * - bits 0..8: index 0
4865 * - bit 9: edge flag 0
4866 * - bits 10..18: index 1
4867 * - bit 19: edge flag 1
4868 * - bits 20..28: index 2
4869 * - bit 29: edge flag 2
4870 * - bit 31: null primitive (skip)
4872 LLVMBuilderRef builder
= ctx
->builder
;
4873 LLVMValueRef tmp
= LLVMBuildZExt(builder
, prim
->isnull
, ctx
->i32
, "");
4874 LLVMValueRef result
= LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->i32
, 31, false), "");
4876 for (unsigned i
= 0; i
< prim
->num_vertices
; ++i
) {
4877 tmp
= LLVMBuildShl(builder
, prim
->index
[i
],
4878 LLVMConstInt(ctx
->i32
, 10 * i
, false), "");
4879 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4880 tmp
= LLVMBuildZExt(builder
, prim
->edgeflag
[i
], ctx
->i32
, "");
4881 tmp
= LLVMBuildShl(builder
, tmp
,
4882 LLVMConstInt(ctx
->i32
, 10 * i
+ 9, false), "");
4883 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4888 void ac_build_export_prim(struct ac_llvm_context
*ctx
,
4889 const struct ac_ngg_prim
*prim
)
4891 struct ac_export_args args
;
4893 if (prim
->passthrough
) {
4894 args
.out
[0] = prim
->passthrough
;
4896 args
.out
[0] = ac_pack_prim_export(ctx
, prim
);
4899 args
.out
[0] = LLVMBuildBitCast(ctx
->builder
, args
.out
[0], ctx
->f32
, "");
4900 args
.out
[1] = LLVMGetUndef(ctx
->f32
);
4901 args
.out
[2] = LLVMGetUndef(ctx
->f32
);
4902 args
.out
[3] = LLVMGetUndef(ctx
->f32
);
4904 args
.target
= V_008DFC_SQ_EXP_PRIM
;
4905 args
.enabled_channels
= 1;
4907 args
.valid_mask
= false;
4910 ac_build_export(ctx
, &args
);
4914 arg_llvm_type(enum ac_arg_type type
, unsigned size
, struct ac_llvm_context
*ctx
)
4916 if (type
== AC_ARG_FLOAT
) {
4917 return size
== 1 ? ctx
->f32
: LLVMVectorType(ctx
->f32
, size
);
4918 } else if (type
== AC_ARG_INT
) {
4919 return size
== 1 ? ctx
->i32
: LLVMVectorType(ctx
->i32
, size
);
4921 LLVMTypeRef ptr_type
;
4923 case AC_ARG_CONST_PTR
:
4926 case AC_ARG_CONST_FLOAT_PTR
:
4927 ptr_type
= ctx
->f32
;
4929 case AC_ARG_CONST_PTR_PTR
:
4930 ptr_type
= ac_array_in_const32_addr_space(ctx
->i8
);
4932 case AC_ARG_CONST_DESC_PTR
:
4933 ptr_type
= ctx
->v4i32
;
4935 case AC_ARG_CONST_IMAGE_PTR
:
4936 ptr_type
= ctx
->v8i32
;
4939 unreachable("unknown arg type");
4942 return ac_array_in_const32_addr_space(ptr_type
);
4945 return ac_array_in_const_addr_space(ptr_type
);
4951 ac_build_main(const struct ac_shader_args
*args
,
4952 struct ac_llvm_context
*ctx
,
4953 enum ac_llvm_calling_convention convention
,
4954 const char *name
, LLVMTypeRef ret_type
,
4955 LLVMModuleRef module
)
4957 LLVMTypeRef arg_types
[AC_MAX_ARGS
];
4959 for (unsigned i
= 0; i
< args
->arg_count
; i
++) {
4960 arg_types
[i
] = arg_llvm_type(args
->args
[i
].type
,
4961 args
->args
[i
].size
, ctx
);
4964 LLVMTypeRef main_function_type
=
4965 LLVMFunctionType(ret_type
, arg_types
, args
->arg_count
, 0);
4967 LLVMValueRef main_function
=
4968 LLVMAddFunction(module
, name
, main_function_type
);
4969 LLVMBasicBlockRef main_function_body
=
4970 LLVMAppendBasicBlockInContext(ctx
->context
, main_function
, "main_body");
4971 LLVMPositionBuilderAtEnd(ctx
->builder
, main_function_body
);
4973 LLVMSetFunctionCallConv(main_function
, convention
);
4974 for (unsigned i
= 0; i
< args
->arg_count
; ++i
) {
4975 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
4977 if (args
->args
[i
].file
!= AC_ARG_SGPR
)
4980 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
4982 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4983 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
4984 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4988 ctx
->main_function
= main_function
;
4990 if (LLVM_VERSION_MAJOR
>= 11) {
4991 /* Enable denormals for FP16 and FP64: */
4992 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math",
4994 /* Disable denormals for FP32: */
4995 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math-f32",
4996 "preserve-sign,preserve-sign");
4998 return main_function
;
5001 void ac_build_s_endpgm(struct ac_llvm_context
*ctx
)
5003 LLVMTypeRef calltype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
5004 LLVMValueRef code
= LLVMConstInlineAsm(calltype
, "s_endpgm", "", true, false);
5005 LLVMBuildCall(ctx
->builder
, code
, NULL
, 0, "");
5008 LLVMValueRef
ac_prefix_bitcount(struct ac_llvm_context
*ctx
,
5009 LLVMValueRef mask
, LLVMValueRef index
)
5011 LLVMBuilderRef builder
= ctx
->builder
;
5012 LLVMTypeRef type
= LLVMTypeOf(mask
);
5014 LLVMValueRef bit
= LLVMBuildShl(builder
, LLVMConstInt(type
, 1, 0),
5015 LLVMBuildZExt(builder
, index
, type
, ""), "");
5016 LLVMValueRef prefix_bits
= LLVMBuildSub(builder
, bit
, LLVMConstInt(type
, 1, 0), "");
5017 LLVMValueRef prefix_mask
= LLVMBuildAnd(builder
, mask
, prefix_bits
, "");
5018 return ac_build_bit_count(ctx
, prefix_mask
);
5021 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
5022 LLVMValueRef
ac_prefix_bitcount_2x64(struct ac_llvm_context
*ctx
,
5023 LLVMValueRef mask
[2], LLVMValueRef index
)
5025 LLVMBuilderRef builder
= ctx
->builder
;
5027 /* Reference version using i128. */
5028 LLVMValueRef input_mask
=
5029 LLVMBuildBitCast(builder
, ac_build_gather_values(ctx
, mask
, 2), ctx
->i128
, "");
5031 return ac_prefix_bitcount(ctx
, input_mask
, index
);
5033 /* Optimized version using 2 64-bit masks. */
5034 LLVMValueRef is_hi
, is_0
, c64
, c128
, all_bits
;
5035 LLVMValueRef prefix_mask
[2], shift
[2], mask_bcnt0
, prefix_bcnt
[2];
5037 /* Compute the 128-bit prefix mask. */
5038 c64
= LLVMConstInt(ctx
->i32
, 64, 0);
5039 c128
= LLVMConstInt(ctx
->i32
, 128, 0);
5040 all_bits
= LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
5041 /* The first index that can have non-zero high bits in the prefix mask is 65. */
5042 is_hi
= LLVMBuildICmp(builder
, LLVMIntUGT
, index
, c64
, "");
5043 is_0
= LLVMBuildICmp(builder
, LLVMIntEQ
, index
, ctx
->i32_0
, "");
5044 mask_bcnt0
= ac_build_bit_count(ctx
, mask
[0]);
5046 for (unsigned i
= 0; i
< 2; i
++) {
5047 shift
[i
] = LLVMBuildSub(builder
, i
? c128
: c64
, index
, "");
5048 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
5049 * so we handle it by the is_0 select.
5050 * For i==1, index==64, same story, so we handle it by the last is_hi select.
5051 * For i==0, index==64, we shift by 0, which is what we want.
5053 prefix_mask
[i
] = LLVMBuildLShr(builder
, all_bits
,
5054 LLVMBuildZExt(builder
, shift
[i
], ctx
->i64
, ""), "");
5055 prefix_mask
[i
] = LLVMBuildAnd(builder
, mask
[i
], prefix_mask
[i
], "");
5056 prefix_bcnt
[i
] = ac_build_bit_count(ctx
, prefix_mask
[i
]);
5059 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_0
, ctx
->i32_0
, prefix_bcnt
[0], "");
5060 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_hi
, mask_bcnt0
, prefix_bcnt
[0], "");
5061 prefix_bcnt
[1] = LLVMBuildSelect(builder
, is_hi
, prefix_bcnt
[1], ctx
->i32_0
, "");
5063 return LLVMBuildAdd(builder
, prefix_bcnt
[0], prefix_bcnt
[1], "");
5068 * Convert triangle strip indices to triangle indices. This is used to decompose
5069 * triangle strips into triangles.
5071 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context
*ctx
,
5072 LLVMValueRef is_odd
,
5073 LLVMValueRef flatshade_first
,
5074 LLVMValueRef index
[3])
5076 LLVMBuilderRef builder
= ctx
->builder
;
5077 LLVMValueRef out
[3];
5079 /* We need to change the vertex order for odd triangles to get correct
5080 * front/back facing by swapping 2 vertex indices, but we also have to
5081 * keep the provoking vertex in the same place.
5083 * If the first vertex is provoking, swap index 1 and 2.
5084 * If the last vertex is provoking, swap index 0 and 1.
5086 out
[0] = LLVMBuildSelect(builder
, flatshade_first
,
5088 LLVMBuildSelect(builder
, is_odd
,
5089 index
[1], index
[0], ""), "");
5090 out
[1] = LLVMBuildSelect(builder
, flatshade_first
,
5091 LLVMBuildSelect(builder
, is_odd
,
5092 index
[2], index
[1], ""),
5093 LLVMBuildSelect(builder
, is_odd
,
5094 index
[0], index
[1], ""), "");
5095 out
[2] = LLVMBuildSelect(builder
, flatshade_first
,
5096 LLVMBuildSelect(builder
, is_odd
,
5097 index
[1], index
[2], ""),
5099 memcpy(index
, out
, sizeof(out
));