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 LLVMValueRef ret
= LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
720 /* Use v_rcp_f32 instead of precise division. */
721 if (!LLVMIsConstant(ret
))
722 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
726 /* See fast_idiv_by_const.h. */
727 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
728 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
730 LLVMValueRef multiplier
,
731 LLVMValueRef pre_shift
,
732 LLVMValueRef post_shift
,
733 LLVMValueRef increment
)
735 LLVMBuilderRef builder
= ctx
->builder
;
737 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
738 num
= LLVMBuildMul(builder
,
739 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
740 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
741 num
= LLVMBuildAdd(builder
, num
,
742 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
743 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
744 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
745 return LLVMBuildLShr(builder
, num
, post_shift
, "");
748 /* See fast_idiv_by_const.h. */
749 /* If num != UINT_MAX, this more efficient version can be used. */
750 /* Set: increment = util_fast_udiv_info::increment; */
751 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
753 LLVMValueRef multiplier
,
754 LLVMValueRef pre_shift
,
755 LLVMValueRef post_shift
,
756 LLVMValueRef increment
)
758 LLVMBuilderRef builder
= ctx
->builder
;
760 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
761 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
762 num
= LLVMBuildMul(builder
,
763 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
764 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
765 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
766 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
767 return LLVMBuildLShr(builder
, num
, post_shift
, "");
770 /* See fast_idiv_by_const.h. */
771 /* Both operands must fit in 31 bits and the divisor must not be 1. */
772 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
774 LLVMValueRef multiplier
,
775 LLVMValueRef post_shift
)
777 LLVMBuilderRef builder
= ctx
->builder
;
779 num
= LLVMBuildMul(builder
,
780 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
781 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
782 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
783 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
784 return LLVMBuildLShr(builder
, num
, post_shift
, "");
787 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
788 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
789 * already multiplied by two. id is the cube face number.
791 struct cube_selection_coords
{
798 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
800 struct cube_selection_coords
*out
)
802 LLVMTypeRef f32
= ctx
->f32
;
804 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
805 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
806 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
807 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
808 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
809 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
810 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
811 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
815 * Build a manual selection sequence for cube face sc/tc coordinates and
816 * major axis vector (multiplied by 2 for consistency) for the given
817 * vec3 \p coords, for the face implied by \p selcoords.
819 * For the major axis, we always adjust the sign to be in the direction of
820 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
821 * the selcoords major axis.
823 static void build_cube_select(struct ac_llvm_context
*ctx
,
824 const struct cube_selection_coords
*selcoords
,
825 const LLVMValueRef
*coords
,
826 LLVMValueRef
*out_st
,
827 LLVMValueRef
*out_ma
)
829 LLVMBuilderRef builder
= ctx
->builder
;
830 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
831 LLVMValueRef is_ma_positive
;
833 LLVMValueRef is_ma_z
, is_not_ma_z
;
834 LLVMValueRef is_ma_y
;
835 LLVMValueRef is_ma_x
;
839 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
840 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
841 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
842 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
844 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
845 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
846 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
847 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
848 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
851 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
852 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
853 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
854 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
855 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
858 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
859 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
860 LLVMConstReal(f32
, -1.0), "");
861 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
864 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
865 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
866 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
867 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
868 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
872 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
873 bool is_deriv
, bool is_array
, bool is_lod
,
874 LLVMValueRef
*coords_arg
,
875 LLVMValueRef
*derivs_arg
)
878 LLVMBuilderRef builder
= ctx
->builder
;
879 struct cube_selection_coords selcoords
;
880 LLVMValueRef coords
[3];
883 if (is_array
&& !is_lod
) {
884 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
886 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
888 * "For Array forms, the array layer used will be
890 * max(0, min(d−1, floor(layer+0.5)))
892 * where d is the depth of the texture array and layer
893 * comes from the component indicated in the tables below.
894 * Workaroudn for an issue where the layer is taken from a
895 * helper invocation which happens to fall on a different
896 * layer due to extrapolation."
898 * GFX8 and earlier attempt to implement this in hardware by
899 * clamping the value of coords[2] = (8 * layer) + face.
900 * Unfortunately, this means that the we end up with the wrong
901 * face when clamping occurs.
903 * Clamp the layer earlier to work around the issue.
905 if (ctx
->chip_class
<= GFX8
) {
907 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
908 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
914 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
916 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
917 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
918 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
920 for (int i
= 0; i
< 2; ++i
)
921 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
923 coords
[2] = selcoords
.id
;
925 if (is_deriv
&& derivs_arg
) {
926 LLVMValueRef derivs
[4];
929 /* Convert cube derivatives to 2D derivatives. */
930 for (axis
= 0; axis
< 2; axis
++) {
931 LLVMValueRef deriv_st
[2];
932 LLVMValueRef deriv_ma
;
934 /* Transform the derivative alongside the texture
935 * coordinate. Mathematically, the correct formula is
936 * as follows. Assume we're projecting onto the +Z face
937 * and denote by dx/dh the derivative of the (original)
938 * X texture coordinate with respect to horizontal
939 * window coordinates. The projection onto the +Z face
944 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
945 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
947 * This motivatives the implementation below.
949 * Whether this actually gives the expected results for
950 * apps that might feed in derivatives obtained via
951 * finite differences is anyone's guess. The OpenGL spec
952 * seems awfully quiet about how textureGrad for cube
953 * maps should be handled.
955 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
956 deriv_st
, &deriv_ma
);
958 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
960 for (int i
= 0; i
< 2; ++i
)
961 derivs
[axis
* 2 + i
] =
962 LLVMBuildFSub(builder
,
963 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
964 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
967 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
970 /* Shift the texture coordinate. This must be applied after the
971 * derivative calculation.
973 for (int i
= 0; i
< 2; ++i
)
974 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
977 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
978 /* coords_arg.w component - array_index for cube arrays */
979 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
982 memcpy(coords_arg
, coords
, sizeof(coords
));
987 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
988 LLVMValueRef llvm_chan
,
989 LLVMValueRef attr_number
,
994 LLVMValueRef args
[5];
999 args
[2] = attr_number
;
1002 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
1003 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1007 args
[2] = llvm_chan
;
1008 args
[3] = attr_number
;
1011 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
1012 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1016 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
1017 LLVMValueRef llvm_chan
,
1018 LLVMValueRef attr_number
,
1019 LLVMValueRef params
,
1023 LLVMValueRef args
[6];
1027 args
[1] = llvm_chan
;
1028 args
[2] = attr_number
;
1029 args
[3] = ctx
->i1false
;
1032 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
1033 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1037 args
[2] = llvm_chan
;
1038 args
[3] = attr_number
;
1039 args
[4] = ctx
->i1false
;
1042 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
1043 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
1047 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
1048 LLVMValueRef parameter
,
1049 LLVMValueRef llvm_chan
,
1050 LLVMValueRef attr_number
,
1051 LLVMValueRef params
)
1053 LLVMValueRef args
[4];
1055 args
[0] = parameter
;
1056 args
[1] = llvm_chan
;
1057 args
[2] = attr_number
;
1060 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
1061 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1065 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
1066 LLVMValueRef base_ptr
,
1069 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1073 ac_build_gep0(struct ac_llvm_context
*ctx
,
1074 LLVMValueRef base_ptr
,
1077 LLVMValueRef indices
[2] = {
1081 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1084 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1087 return LLVMBuildPointerCast(ctx
->builder
,
1088 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1089 LLVMTypeOf(ptr
), "");
1093 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1094 LLVMValueRef base_ptr
, LLVMValueRef index
,
1097 LLVMBuildStore(ctx
->builder
, value
,
1098 ac_build_gep0(ctx
, base_ptr
, index
));
1102 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1103 * It's equivalent to doing a load from &base_ptr[index].
1105 * \param base_ptr Where the array starts.
1106 * \param index The element index into the array.
1107 * \param uniform Whether the base_ptr and index can be assumed to be
1108 * dynamically uniform (i.e. load to an SGPR)
1109 * \param invariant Whether the load is invariant (no other opcodes affect it)
1110 * \param no_unsigned_wraparound
1111 * For all possible re-associations and re-distributions of an expression
1112 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1113 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1114 * does not result in an unsigned integer wraparound. This is used for
1115 * optimal code generation of 32-bit pointer arithmetic.
1117 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1118 * integer wraparound can't be an imm offset in s_load_dword, because
1119 * the instruction performs "addr + offset" in 64 bits.
1121 * Expected usage for bindless textures by chaining GEPs:
1122 * // possible unsigned wraparound, don't use InBounds:
1123 * ptr1 = LLVMBuildGEP(base_ptr, index);
1124 * image = load(ptr1); // becomes "s_load ptr1, 0"
1126 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1127 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1130 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1131 LLVMValueRef index
, bool uniform
, bool invariant
,
1132 bool no_unsigned_wraparound
)
1134 LLVMValueRef pointer
, result
;
1136 if (no_unsigned_wraparound
&&
1137 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1138 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1140 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1143 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1144 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1146 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1150 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1153 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1156 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1157 LLVMValueRef base_ptr
, LLVMValueRef index
)
1159 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1162 /* This assumes that there is no unsigned integer wraparound during the address
1163 * computation, excluding all GEPs within base_ptr. */
1164 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1165 LLVMValueRef base_ptr
, LLVMValueRef index
)
1167 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1170 /* See ac_build_load_custom() documentation. */
1171 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1172 LLVMValueRef base_ptr
, LLVMValueRef index
)
1174 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1177 static unsigned get_load_cache_policy(struct ac_llvm_context
*ctx
,
1178 unsigned cache_policy
)
1180 return cache_policy
|
1181 (ctx
->chip_class
>= GFX10
&& cache_policy
& ac_glc
? ac_dlc
: 0);
1185 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1188 LLVMValueRef vindex
,
1189 LLVMValueRef voffset
,
1190 LLVMValueRef soffset
,
1191 unsigned num_channels
,
1192 LLVMTypeRef return_channel_type
,
1193 unsigned cache_policy
,
1197 LLVMValueRef args
[6];
1200 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1202 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1203 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1204 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1205 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1206 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1207 const char *indexing_kind
= structurized
? "struct" : "raw";
1208 char name
[256], type_name
[8];
1210 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(return_channel_type
, func
) : return_channel_type
;
1211 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1214 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1215 indexing_kind
, type_name
);
1217 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1218 indexing_kind
, type_name
);
1221 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1222 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1226 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1229 LLVMValueRef vindex
,
1230 LLVMValueRef voffset
,
1231 unsigned num_channels
,
1232 unsigned cache_policy
)
1234 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1235 voffset
, NULL
, num_channels
,
1236 ctx
->f32
, cache_policy
,
1240 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1241 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1242 * or v4i32 (num_channels=3,4).
1245 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1248 unsigned num_channels
,
1249 LLVMValueRef voffset
,
1250 LLVMValueRef soffset
,
1251 unsigned inst_offset
,
1252 unsigned cache_policy
)
1254 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1256 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1257 LLVMValueRef v
[3], v01
;
1259 for (int i
= 0; i
< 3; i
++) {
1260 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1261 LLVMConstInt(ctx
->i32
, i
, 0), "");
1263 v01
= ac_build_gather_values(ctx
, v
, 2);
1265 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1266 soffset
, inst_offset
, cache_policy
);
1267 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1268 soffset
, inst_offset
+ 8,
1273 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1274 * (voffset is swizzled, but soffset isn't swizzled).
1275 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1277 if (!(cache_policy
& ac_swizzled
)) {
1278 LLVMValueRef offset
= soffset
;
1281 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1282 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1284 ac_build_buffer_store_common(ctx
, rsrc
, ac_to_float(ctx
, vdata
),
1285 ctx
->i32_0
, voffset
, offset
,
1286 num_channels
, ctx
->f32
,
1287 cache_policy
, false, false);
1291 static const unsigned dfmts
[] = {
1292 V_008F0C_BUF_DATA_FORMAT_32
,
1293 V_008F0C_BUF_DATA_FORMAT_32_32
,
1294 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1295 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1297 unsigned dfmt
= dfmts
[num_channels
- 1];
1298 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1299 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1301 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1302 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
);
1306 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1308 LLVMValueRef vindex
,
1309 LLVMValueRef voffset
,
1310 LLVMValueRef soffset
,
1311 unsigned num_channels
,
1312 LLVMTypeRef channel_type
,
1313 unsigned cache_policy
,
1318 LLVMValueRef args
[5];
1320 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1322 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1323 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1324 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1325 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1326 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1327 const char *indexing_kind
= structurized
? "struct" : "raw";
1328 char name
[256], type_name
[8];
1330 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1331 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1334 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1335 indexing_kind
, type_name
);
1337 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1338 indexing_kind
, type_name
);
1341 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1342 ac_get_load_intr_attribs(can_speculate
));
1346 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1349 LLVMValueRef vindex
,
1350 LLVMValueRef voffset
,
1351 LLVMValueRef soffset
,
1352 unsigned inst_offset
,
1353 unsigned cache_policy
,
1357 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1359 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1361 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1363 if (allow_smem
&& !(cache_policy
& ac_slc
) &&
1364 (!(cache_policy
& ac_glc
) || ctx
->chip_class
>= GFX8
)) {
1365 assert(vindex
== NULL
);
1367 LLVMValueRef result
[8];
1369 for (int i
= 0; i
< num_channels
; i
++) {
1371 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1372 LLVMConstInt(ctx
->i32
, 4, 0), "");
1374 LLVMValueRef args
[3] = {
1377 LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0),
1379 result
[i
] = ac_build_intrinsic(ctx
,
1380 "llvm.amdgcn.s.buffer.load.f32",
1382 AC_FUNC_ATTR_READNONE
);
1384 if (num_channels
== 1)
1387 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1388 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1389 return ac_build_gather_values(ctx
, result
, num_channels
);
1392 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
,
1394 num_channels
, ctx
->f32
,
1396 can_speculate
, false, false);
1399 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1401 LLVMValueRef vindex
,
1402 LLVMValueRef voffset
,
1403 unsigned num_channels
,
1404 unsigned cache_policy
,
1407 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1408 ctx
->i32_0
, num_channels
, ctx
->f32
,
1409 cache_policy
, can_speculate
,
1414 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1416 LLVMValueRef vindex
,
1417 LLVMValueRef voffset
,
1418 LLVMValueRef soffset
,
1419 LLVMValueRef immoffset
,
1420 unsigned num_channels
,
1423 unsigned cache_policy
,
1427 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1429 LLVMValueRef args
[6];
1431 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1433 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1434 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1435 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1436 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1437 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1438 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1439 const char *indexing_kind
= structurized
? "struct" : "raw";
1440 char name
[256], type_name
[8];
1442 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1443 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1445 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1446 indexing_kind
, type_name
);
1448 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1449 ac_get_load_intr_attribs(can_speculate
));
1453 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1455 LLVMValueRef vindex
,
1456 LLVMValueRef voffset
,
1457 LLVMValueRef soffset
,
1458 LLVMValueRef immoffset
,
1459 unsigned num_channels
,
1462 unsigned cache_policy
,
1465 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1466 immoffset
, num_channels
, dfmt
, nfmt
,
1467 cache_policy
, can_speculate
, true);
1471 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1473 LLVMValueRef voffset
,
1474 LLVMValueRef soffset
,
1475 LLVMValueRef immoffset
,
1476 unsigned num_channels
,
1479 unsigned cache_policy
,
1482 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1483 immoffset
, num_channels
, dfmt
, nfmt
,
1484 cache_policy
, can_speculate
, false);
1488 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1490 LLVMValueRef voffset
,
1491 LLVMValueRef soffset
,
1492 LLVMValueRef immoffset
,
1493 unsigned cache_policy
)
1497 if (LLVM_VERSION_MAJOR
>= 9) {
1498 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1500 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1501 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1503 1, ctx
->i16
, cache_policy
,
1504 false, false, false);
1506 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1507 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1509 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1510 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1513 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1520 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1522 LLVMValueRef voffset
,
1523 LLVMValueRef soffset
,
1524 LLVMValueRef immoffset
,
1525 unsigned cache_policy
)
1529 if (LLVM_VERSION_MAJOR
>= 9) {
1530 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1532 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1533 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1535 1, ctx
->i8
, cache_policy
,
1536 false, false, false);
1538 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1539 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1541 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1542 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1545 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1552 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1554 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1555 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1558 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1560 assert(LLVMTypeOf(src
) == ctx
->i32
);
1563 LLVMValueRef mantissa
;
1564 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1566 /* Converting normal numbers is just a shift + correcting the exponent bias */
1567 unsigned normal_shift
= 23 - mant_bits
;
1568 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1569 LLVMValueRef shifted
, normal
;
1571 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1572 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1574 /* Converting nan/inf numbers is the same, but with a different exponent update */
1575 LLVMValueRef naninf
;
1576 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1578 /* Converting denormals is the complex case: determine the leading zeros of the
1579 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1581 LLVMValueRef denormal
;
1582 LLVMValueRef params
[2] = {
1584 ctx
->i1true
, /* result can be undef when arg is 0 */
1586 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1587 params
, 2, AC_FUNC_ATTR_READNONE
);
1589 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1590 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1591 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1593 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1594 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1595 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1596 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1598 /* Select the final result. */
1599 LLVMValueRef result
;
1601 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1602 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1603 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1605 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1606 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1607 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1609 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1610 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1612 return ac_to_float(ctx
, result
);
1616 * Generate a fully general open coded buffer format fetch with all required
1617 * fixups suitable for vertex fetch, using non-format buffer loads.
1619 * Some combinations of argument values have special interpretations:
1620 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1621 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1623 * \param log_size log(size of channel in bytes)
1624 * \param num_channels number of channels (1 to 4)
1625 * \param format AC_FETCH_FORMAT_xxx value
1626 * \param reverse whether XYZ channels are reversed
1627 * \param known_aligned whether the source is known to be aligned to hardware's
1628 * effective element size for loading the given format
1629 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1630 * \param rsrc buffer resource descriptor
1631 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1634 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1636 unsigned num_channels
,
1641 LLVMValueRef vindex
,
1642 LLVMValueRef voffset
,
1643 LLVMValueRef soffset
,
1644 unsigned cache_policy
,
1648 unsigned load_log_size
= log_size
;
1649 unsigned load_num_channels
= num_channels
;
1650 if (log_size
== 3) {
1652 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1653 load_num_channels
= 2 * num_channels
;
1655 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1659 int log_recombine
= 0;
1660 if (ctx
->chip_class
== GFX6
&& !known_aligned
) {
1661 /* Avoid alignment restrictions by loading one byte at a time. */
1662 load_num_channels
<<= load_log_size
;
1663 log_recombine
= load_log_size
;
1665 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1666 log_recombine
= -util_logbase2(load_num_channels
);
1667 load_num_channels
= 1;
1668 load_log_size
+= -log_recombine
;
1671 assert(load_log_size
>= 2 || LLVM_VERSION_MAJOR
>= 9);
1673 LLVMValueRef loads
[32]; /* up to 32 bytes */
1674 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1675 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1676 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1677 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1678 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1679 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1680 loads
[i
] = ac_build_buffer_load_common(
1681 ctx
, rsrc
, vindex
, voffset
, tmp
,
1682 num_channels
, channel_type
, cache_policy
,
1683 can_speculate
, false, true);
1684 if (load_log_size
>= 2)
1685 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1688 if (log_recombine
> 0) {
1689 /* Recombine bytes if necessary (GFX6 only) */
1690 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1692 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1693 LLVMValueRef accum
= NULL
;
1694 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1695 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1699 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1700 LLVMConstInt(dst_type
, 8 * i
, false), "");
1701 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1706 } else if (log_recombine
< 0) {
1707 /* Split vectors of dwords */
1708 if (load_log_size
> 2) {
1709 assert(load_num_channels
== 1);
1710 LLVMValueRef loaded
= loads
[0];
1711 unsigned log_split
= load_log_size
- 2;
1712 log_recombine
+= log_split
;
1713 load_num_channels
= 1 << log_split
;
1715 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1716 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1717 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1721 /* Further split dwords and shorts if required */
1722 if (log_recombine
< 0) {
1723 for (unsigned src
= load_num_channels
,
1724 dst
= load_num_channels
<< -log_recombine
;
1726 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1727 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1728 LLVMValueRef loaded
= loads
[src
- 1];
1729 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1730 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1731 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1732 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1733 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1739 if (log_size
== 3) {
1740 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1741 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1742 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1743 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1745 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1746 /* 10_11_11_FLOAT */
1747 LLVMValueRef data
= loads
[0];
1748 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1749 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1750 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1751 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1752 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1754 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1755 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1756 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1760 format
= AC_FETCH_FORMAT_FLOAT
;
1762 /* 2_10_10_10 data formats */
1763 LLVMValueRef data
= loads
[0];
1764 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1765 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1766 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1767 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1768 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1769 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1770 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1771 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1772 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1778 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1779 if (log_size
!= 2) {
1780 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1781 tmp
= ac_to_float(ctx
, loads
[chan
]);
1783 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1784 else if (log_size
== 1)
1785 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1786 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1789 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1790 if (log_size
!= 2) {
1791 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1792 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1794 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1795 if (log_size
!= 2) {
1796 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1797 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1800 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1801 format
== AC_FETCH_FORMAT_USCALED
||
1802 format
== AC_FETCH_FORMAT_UINT
;
1804 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1806 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1808 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1811 LLVMValueRef scale
= NULL
;
1812 if (format
== AC_FETCH_FORMAT_FIXED
) {
1813 assert(log_size
== 2);
1814 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1815 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1816 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1817 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1818 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1819 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1820 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1823 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1825 if (format
== AC_FETCH_FORMAT_SNORM
) {
1826 /* Clamp to [-1, 1] */
1827 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1828 LLVMValueRef clamp
=
1829 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1830 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1833 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1837 while (num_channels
< 4) {
1838 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1839 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1841 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1848 loads
[0] = loads
[2];
1852 return ac_build_gather_values(ctx
, loads
, 4);
1856 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1859 LLVMValueRef vindex
,
1860 LLVMValueRef voffset
,
1861 LLVMValueRef soffset
,
1862 LLVMValueRef immoffset
,
1863 unsigned num_channels
,
1866 unsigned cache_policy
,
1869 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
? voffset
: ctx
->i32_0
,
1872 LLVMValueRef args
[7];
1874 args
[idx
++] = vdata
;
1875 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1877 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1878 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1879 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1880 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1881 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1882 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1883 const char *indexing_kind
= structurized
? "struct" : "raw";
1884 char name
[256], type_name
[8];
1886 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1887 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1889 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1890 indexing_kind
, type_name
);
1892 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1893 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1897 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1900 LLVMValueRef vindex
,
1901 LLVMValueRef voffset
,
1902 LLVMValueRef soffset
,
1903 LLVMValueRef immoffset
,
1904 unsigned num_channels
,
1907 unsigned cache_policy
)
1909 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1910 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1915 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1918 LLVMValueRef voffset
,
1919 LLVMValueRef soffset
,
1920 LLVMValueRef immoffset
,
1921 unsigned num_channels
,
1924 unsigned cache_policy
)
1926 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1927 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1932 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1935 LLVMValueRef voffset
,
1936 LLVMValueRef soffset
,
1937 unsigned cache_policy
)
1939 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1941 if (LLVM_VERSION_MAJOR
>= 9) {
1942 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1943 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1944 voffset
, soffset
, 1,
1945 ctx
->i16
, cache_policy
,
1948 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1949 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1951 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1953 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1954 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1959 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1962 LLVMValueRef voffset
,
1963 LLVMValueRef soffset
,
1964 unsigned cache_policy
)
1966 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1968 if (LLVM_VERSION_MAJOR
>= 9) {
1969 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1970 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1971 voffset
, soffset
, 1,
1972 ctx
->i8
, cache_policy
,
1975 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1976 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1978 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1980 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1981 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1985 * Set range metadata on an instruction. This can only be used on load and
1986 * call instructions. If you know an instruction can only produce the values
1987 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1988 * \p lo is the minimum value inclusive.
1989 * \p hi is the maximum value exclusive.
1991 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1992 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1994 LLVMValueRef range_md
, md_args
[2];
1995 LLVMTypeRef type
= LLVMTypeOf(value
);
1996 LLVMContextRef context
= LLVMGetTypeContext(type
);
1998 md_args
[0] = LLVMConstInt(type
, lo
, false);
1999 md_args
[1] = LLVMConstInt(type
, hi
, false);
2000 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
2001 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
2005 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2009 LLVMValueRef tid_args
[2];
2010 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2011 tid_args
[1] = ctx
->i32_0
;
2012 tid_args
[1] = ac_build_intrinsic(ctx
,
2013 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2014 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2016 if (ctx
->wave_size
== 32) {
2019 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2021 2, AC_FUNC_ATTR_READNONE
);
2023 set_range_metadata(ctx
, tid
, 0, ctx
->wave_size
);
2028 * AMD GCN implements derivatives using the local data store (LDS)
2029 * All writes to the LDS happen in all executing threads at
2030 * the same time. TID is the Thread ID for the current
2031 * thread and is a value between 0 and 63, representing
2032 * the thread's position in the wavefront.
2034 * For the pixel shader threads are grouped into quads of four pixels.
2035 * The TIDs of the pixels of a quad are:
2043 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2044 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2045 * the current pixel's column, and masking with 0xfffffffe yields the TID
2046 * of the left pixel of the current pixel's row.
2048 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2049 * adding 2 yields the TID of the pixel below the top pixel.
2052 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2057 unsigned tl_lanes
[4], trbl_lanes
[4];
2058 char name
[32], type
[8];
2059 LLVMValueRef tl
, trbl
;
2060 LLVMTypeRef result_type
;
2061 LLVMValueRef result
;
2063 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2065 if (result_type
== ctx
->f16
)
2066 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2068 for (unsigned i
= 0; i
< 4; ++i
) {
2069 tl_lanes
[i
] = i
& mask
;
2070 trbl_lanes
[i
] = (i
& mask
) + idx
;
2073 tl
= ac_build_quad_swizzle(ctx
, val
,
2074 tl_lanes
[0], tl_lanes
[1],
2075 tl_lanes
[2], tl_lanes
[3]);
2076 trbl
= ac_build_quad_swizzle(ctx
, val
,
2077 trbl_lanes
[0], trbl_lanes
[1],
2078 trbl_lanes
[2], trbl_lanes
[3]);
2080 if (result_type
== ctx
->f16
) {
2081 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2082 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2085 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2086 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2087 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2089 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2090 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2092 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2096 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2098 LLVMValueRef wave_id
)
2100 LLVMValueRef args
[2];
2101 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2103 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2107 ac_build_imsb(struct ac_llvm_context
*ctx
,
2109 LLVMTypeRef dst_type
)
2111 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2113 AC_FUNC_ATTR_READNONE
);
2115 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2116 * the index from LSB. Invert it by doing "31 - msb". */
2117 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2120 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2121 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2122 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2123 arg
, ctx
->i32_0
, ""),
2124 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2125 arg
, all_ones
, ""), "");
2127 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2131 ac_build_umsb(struct ac_llvm_context
*ctx
,
2133 LLVMTypeRef dst_type
)
2135 const char *intrin_name
;
2137 LLVMValueRef highest_bit
;
2141 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2144 intrin_name
= "llvm.ctlz.i64";
2146 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2150 intrin_name
= "llvm.ctlz.i32";
2152 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2156 intrin_name
= "llvm.ctlz.i16";
2158 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2162 intrin_name
= "llvm.ctlz.i8";
2164 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2168 unreachable(!"invalid bitsize");
2172 LLVMValueRef params
[2] = {
2177 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2179 AC_FUNC_ATTR_READNONE
);
2181 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2182 * the index from LSB. Invert it by doing "31 - msb". */
2183 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2185 if (bitsize
== 64) {
2186 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2187 } else if (bitsize
< 32) {
2188 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2191 /* check for zero */
2192 return LLVMBuildSelect(ctx
->builder
,
2193 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2194 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2197 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2201 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2202 LLVMValueRef args
[2] = {a
, b
};
2203 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2204 AC_FUNC_ATTR_READNONE
);
2207 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2211 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2212 LLVMValueRef args
[2] = {a
, b
};
2213 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2214 AC_FUNC_ATTR_READNONE
);
2217 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2220 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2221 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2224 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2227 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2228 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2231 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2234 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2235 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2238 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2241 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2242 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2245 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2247 LLVMTypeRef t
= LLVMTypeOf(value
);
2248 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2249 LLVMConstReal(t
, 1.0));
2252 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2254 LLVMValueRef args
[9];
2256 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2257 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2260 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2261 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2263 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2265 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2267 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2268 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2270 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2271 ctx
->voidt
, args
, 6, 0);
2273 args
[2] = a
->out
[0];
2274 args
[3] = a
->out
[1];
2275 args
[4] = a
->out
[2];
2276 args
[5] = a
->out
[3];
2277 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2278 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2280 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2281 ctx
->voidt
, args
, 8, 0);
2285 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2287 struct ac_export_args args
;
2289 args
.enabled_channels
= 0x0; /* enabled channels */
2290 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2291 args
.done
= 1; /* DONE bit */
2292 args
.target
= V_008DFC_SQ_EXP_NULL
;
2293 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2294 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2295 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2296 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2297 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2299 ac_build_export(ctx
, &args
);
2302 static unsigned ac_num_coords(enum ac_image_dim dim
)
2308 case ac_image_1darray
:
2312 case ac_image_2darray
:
2313 case ac_image_2dmsaa
:
2315 case ac_image_2darraymsaa
:
2318 unreachable("ac_num_coords: bad dim");
2322 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2326 case ac_image_1darray
:
2329 case ac_image_2darray
:
2334 case ac_image_2dmsaa
:
2335 case ac_image_2darraymsaa
:
2337 unreachable("derivatives not supported");
2341 static const char *get_atomic_name(enum ac_atomic_op op
)
2344 case ac_atomic_swap
: return "swap";
2345 case ac_atomic_add
: return "add";
2346 case ac_atomic_sub
: return "sub";
2347 case ac_atomic_smin
: return "smin";
2348 case ac_atomic_umin
: return "umin";
2349 case ac_atomic_smax
: return "smax";
2350 case ac_atomic_umax
: return "umax";
2351 case ac_atomic_and
: return "and";
2352 case ac_atomic_or
: return "or";
2353 case ac_atomic_xor
: return "xor";
2354 case ac_atomic_inc_wrap
: return "inc";
2355 case ac_atomic_dec_wrap
: return "dec";
2357 unreachable("bad atomic op");
2360 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2361 struct ac_image_args
*a
)
2363 const char *overload
[3] = { "", "", "" };
2364 unsigned num_overloads
= 0;
2365 LLVMValueRef args
[18];
2366 unsigned num_args
= 0;
2367 enum ac_image_dim dim
= a
->dim
;
2369 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2371 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2372 a
->opcode
!= ac_image_store_mip
) ||
2374 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2375 (!a
->compare
&& !a
->offset
));
2376 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2377 a
->opcode
== ac_image_get_lod
) ||
2379 assert((a
->bias
? 1 : 0) +
2381 (a
->level_zero
? 1 : 0) +
2382 (a
->derivs
[0] ? 1 : 0) <= 1);
2384 if (a
->opcode
== ac_image_get_lod
) {
2386 case ac_image_1darray
:
2389 case ac_image_2darray
:
2398 bool sample
= a
->opcode
== ac_image_sample
||
2399 a
->opcode
== ac_image_gather4
||
2400 a
->opcode
== ac_image_get_lod
;
2401 bool atomic
= a
->opcode
== ac_image_atomic
||
2402 a
->opcode
== ac_image_atomic_cmpswap
;
2403 bool load
= a
->opcode
== ac_image_sample
||
2404 a
->opcode
== ac_image_gather4
||
2405 a
->opcode
== ac_image_load
||
2406 a
->opcode
== ac_image_load_mip
;
2407 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2409 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2410 args
[num_args
++] = a
->data
[0];
2411 if (a
->opcode
== ac_image_atomic_cmpswap
)
2412 args
[num_args
++] = a
->data
[1];
2416 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2419 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2421 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2422 overload
[num_overloads
++] = ".f32";
2425 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2427 unsigned count
= ac_num_derivs(dim
);
2428 for (unsigned i
= 0; i
< count
; ++i
)
2429 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2430 overload
[num_overloads
++] = ".f32";
2432 unsigned num_coords
=
2433 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2434 for (unsigned i
= 0; i
< num_coords
; ++i
)
2435 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2437 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2438 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2440 args
[num_args
++] = a
->resource
;
2442 args
[num_args
++] = a
->sampler
;
2443 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2446 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2447 args
[num_args
++] = LLVMConstInt(ctx
->i32
,
2448 load
? get_load_cache_policy(ctx
, a
->cache_policy
) :
2449 a
->cache_policy
, false);
2452 const char *atomic_subop
= "";
2453 switch (a
->opcode
) {
2454 case ac_image_sample
: name
= "sample"; break;
2455 case ac_image_gather4
: name
= "gather4"; break;
2456 case ac_image_load
: name
= "load"; break;
2457 case ac_image_load_mip
: name
= "load.mip"; break;
2458 case ac_image_store
: name
= "store"; break;
2459 case ac_image_store_mip
: name
= "store.mip"; break;
2460 case ac_image_atomic
:
2462 atomic_subop
= get_atomic_name(a
->atomic
);
2464 case ac_image_atomic_cmpswap
:
2466 atomic_subop
= "cmpswap";
2468 case ac_image_get_lod
: name
= "getlod"; break;
2469 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2470 default: unreachable("invalid image opcode");
2473 const char *dimname
;
2475 case ac_image_1d
: dimname
= "1d"; break;
2476 case ac_image_2d
: dimname
= "2d"; break;
2477 case ac_image_3d
: dimname
= "3d"; break;
2478 case ac_image_cube
: dimname
= "cube"; break;
2479 case ac_image_1darray
: dimname
= "1darray"; break;
2480 case ac_image_2darray
: dimname
= "2darray"; break;
2481 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2482 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2483 default: unreachable("invalid dim");
2487 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2489 snprintf(intr_name
, sizeof(intr_name
),
2490 "llvm.amdgcn.image.%s%s" /* base name */
2491 "%s%s%s" /* sample/gather modifiers */
2492 ".%s.%s%s%s%s", /* dimension and type overloads */
2494 a
->compare
? ".c" : "",
2497 a
->derivs
[0] ? ".d" :
2498 a
->level_zero
? ".lz" : "",
2499 a
->offset
? ".o" : "",
2501 atomic
? "i32" : "v4f32",
2502 overload
[0], overload
[1], overload
[2]);
2507 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2512 LLVMValueRef result
=
2513 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2515 if (!sample
&& retty
== ctx
->v4f32
) {
2516 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2522 LLVMValueRef
ac_build_image_get_sample_count(struct ac_llvm_context
*ctx
,
2525 LLVMValueRef samples
;
2527 /* Read the samples from the descriptor directly.
2528 * Hardware doesn't have any instruction for this.
2530 samples
= LLVMBuildExtractElement(ctx
->builder
, rsrc
,
2531 LLVMConstInt(ctx
->i32
, 3, 0), "");
2532 samples
= LLVMBuildLShr(ctx
->builder
, samples
,
2533 LLVMConstInt(ctx
->i32
, 16, 0), "");
2534 samples
= LLVMBuildAnd(ctx
->builder
, samples
,
2535 LLVMConstInt(ctx
->i32
, 0xf, 0), "");
2536 samples
= LLVMBuildShl(ctx
->builder
, ctx
->i32_1
,
2541 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2542 LLVMValueRef args
[2])
2545 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2547 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2548 args
, 2, AC_FUNC_ATTR_READNONE
);
2551 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2552 LLVMValueRef args
[2])
2555 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2556 ctx
->v2i16
, args
, 2,
2557 AC_FUNC_ATTR_READNONE
);
2558 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2561 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2562 LLVMValueRef args
[2])
2565 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2566 ctx
->v2i16
, args
, 2,
2567 AC_FUNC_ATTR_READNONE
);
2568 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2571 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2572 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2573 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2575 assert(bits
== 8 || bits
== 10 || bits
== 16);
2577 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2578 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2579 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2580 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2581 LLVMValueRef max_alpha
=
2582 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2583 LLVMValueRef min_alpha
=
2584 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2588 for (int i
= 0; i
< 2; i
++) {
2589 bool alpha
= hi
&& i
== 1;
2590 args
[i
] = ac_build_imin(ctx
, args
[i
],
2591 alpha
? max_alpha
: max_rgb
);
2592 args
[i
] = ac_build_imax(ctx
, args
[i
],
2593 alpha
? min_alpha
: min_rgb
);
2598 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2599 ctx
->v2i16
, args
, 2,
2600 AC_FUNC_ATTR_READNONE
);
2601 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2604 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2605 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2606 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2608 assert(bits
== 8 || bits
== 10 || bits
== 16);
2610 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2611 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2612 LLVMValueRef max_alpha
=
2613 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2617 for (int i
= 0; i
< 2; i
++) {
2618 bool alpha
= hi
&& i
== 1;
2619 args
[i
] = ac_build_umin(ctx
, args
[i
],
2620 alpha
? max_alpha
: max_rgb
);
2625 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2626 ctx
->v2i16
, args
, 2,
2627 AC_FUNC_ATTR_READNONE
);
2628 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2631 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2633 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2634 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2637 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2639 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2643 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2644 LLVMValueRef offset
, LLVMValueRef width
,
2647 LLVMValueRef args
[] = {
2653 return ac_build_intrinsic(ctx
, is_signed
? "llvm.amdgcn.sbfe.i32" :
2654 "llvm.amdgcn.ubfe.i32",
2655 ctx
->i32
, args
, 3, AC_FUNC_ATTR_READNONE
);
2659 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2660 LLVMValueRef s1
, LLVMValueRef s2
)
2662 return LLVMBuildAdd(ctx
->builder
,
2663 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2666 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2667 LLVMValueRef s1
, LLVMValueRef s2
)
2669 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2670 if (ctx
->chip_class
>= GFX10
) {
2671 return ac_build_intrinsic(ctx
, "llvm.fma.f32", ctx
->f32
,
2672 (LLVMValueRef
[]) {s0
, s1
, s2
}, 3,
2673 AC_FUNC_ATTR_READNONE
);
2676 return LLVMBuildFAdd(ctx
->builder
,
2677 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2680 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned wait_flags
)
2685 unsigned lgkmcnt
= 63;
2686 unsigned vmcnt
= ctx
->chip_class
>= GFX9
? 63 : 15;
2687 unsigned vscnt
= 63;
2689 if (wait_flags
& AC_WAIT_LGKM
)
2691 if (wait_flags
& AC_WAIT_VLOAD
)
2694 if (wait_flags
& AC_WAIT_VSTORE
) {
2695 if (ctx
->chip_class
>= GFX10
)
2701 /* There is no intrinsic for vscnt(0), so use a fence. */
2702 if ((wait_flags
& AC_WAIT_LGKM
&&
2703 wait_flags
& AC_WAIT_VLOAD
&&
2704 wait_flags
& AC_WAIT_VSTORE
) ||
2706 LLVMBuildFence(ctx
->builder
, LLVMAtomicOrderingRelease
, false, "");
2710 unsigned simm16
= (lgkmcnt
<< 8) |
2711 (7 << 4) | /* expcnt */
2713 ((vmcnt
>> 4) << 14);
2715 LLVMValueRef args
[1] = {
2716 LLVMConstInt(ctx
->i32
, simm16
, false),
2718 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2719 ctx
->voidt
, args
, 1, 0);
2722 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2723 LLVMValueRef src1
, LLVMValueRef src2
,
2726 LLVMValueRef result
;
2728 if (bitsize
== 64 || (bitsize
== 16 && ctx
->chip_class
<= GFX8
)) {
2729 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2730 * or lower 16-bit fmed because it's only supported on GFX9+.
2732 LLVMValueRef min1
, min2
, max1
;
2734 min1
= ac_build_fmin(ctx
, src0
, src1
);
2735 max1
= ac_build_fmax(ctx
, src0
, src1
);
2736 min2
= ac_build_fmin(ctx
, max1
, src2
);
2738 result
= ac_build_fmax(ctx
, min2
, min1
);
2743 if (bitsize
== 16) {
2744 intr
= "llvm.amdgcn.fmed3.f16";
2747 assert(bitsize
== 32);
2748 intr
= "llvm.amdgcn.fmed3.f32";
2752 LLVMValueRef params
[] = {
2758 result
= ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2759 AC_FUNC_ATTR_READNONE
);
2762 if (ctx
->chip_class
< GFX9
&& bitsize
== 32) {
2763 /* Only pre-GFX9 chips do not flush denorms. */
2764 result
= ac_build_canonicalize(ctx
, result
, bitsize
);
2770 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2776 if (bitsize
== 16) {
2777 intr
= "llvm.amdgcn.fract.f16";
2779 } else if (bitsize
== 32) {
2780 intr
= "llvm.amdgcn.fract.f32";
2783 intr
= "llvm.amdgcn.fract.f64";
2787 LLVMValueRef params
[] = {
2790 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2791 AC_FUNC_ATTR_READNONE
);
2794 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2797 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2798 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2799 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2801 LLVMValueRef cmp
, val
;
2802 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2803 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2804 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2805 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2809 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2812 LLVMValueRef cmp
, val
, zero
, one
;
2815 if (bitsize
== 16) {
2819 } else if (bitsize
== 32) {
2829 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2830 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2831 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2832 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2836 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2838 LLVMValueRef result
;
2841 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2845 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i128", ctx
->i128
,
2846 (LLVMValueRef
[]) { src0
}, 1,
2847 AC_FUNC_ATTR_READNONE
);
2848 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2851 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2852 (LLVMValueRef
[]) { src0
}, 1,
2853 AC_FUNC_ATTR_READNONE
);
2855 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2858 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2859 (LLVMValueRef
[]) { src0
}, 1,
2860 AC_FUNC_ATTR_READNONE
);
2863 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2864 (LLVMValueRef
[]) { src0
}, 1,
2865 AC_FUNC_ATTR_READNONE
);
2867 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2870 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2871 (LLVMValueRef
[]) { src0
}, 1,
2872 AC_FUNC_ATTR_READNONE
);
2874 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2877 unreachable(!"invalid bitsize");
2884 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2887 LLVMValueRef result
;
2890 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2894 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2895 (LLVMValueRef
[]) { src0
}, 1,
2896 AC_FUNC_ATTR_READNONE
);
2898 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2901 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2902 (LLVMValueRef
[]) { src0
}, 1,
2903 AC_FUNC_ATTR_READNONE
);
2906 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2907 (LLVMValueRef
[]) { src0
}, 1,
2908 AC_FUNC_ATTR_READNONE
);
2910 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2913 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
2914 (LLVMValueRef
[]) { src0
}, 1,
2915 AC_FUNC_ATTR_READNONE
);
2917 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2920 unreachable(!"invalid bitsize");
2927 #define AC_EXP_TARGET 0
2928 #define AC_EXP_ENABLED_CHANNELS 1
2929 #define AC_EXP_OUT0 2
2937 struct ac_vs_exp_chan
2941 enum ac_ir_type type
;
2944 struct ac_vs_exp_inst
{
2947 struct ac_vs_exp_chan chan
[4];
2950 struct ac_vs_exports
{
2952 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2955 /* Return true if the PARAM export has been eliminated. */
2956 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2957 uint32_t num_outputs
,
2958 struct ac_vs_exp_inst
*exp
)
2960 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2961 bool is_zero
[4] = {}, is_one
[4] = {};
2963 for (i
= 0; i
< 4; i
++) {
2964 /* It's a constant expression. Undef outputs are eliminated too. */
2965 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2968 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2969 if (exp
->chan
[i
].const_float
== 0)
2971 else if (exp
->chan
[i
].const_float
== 1)
2974 return false; /* other constant */
2979 /* Only certain combinations of 0 and 1 can be eliminated. */
2980 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2981 default_val
= is_zero
[3] ? 0 : 1;
2982 else if (is_one
[0] && is_one
[1] && is_one
[2])
2983 default_val
= is_zero
[3] ? 2 : 3;
2987 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2988 LLVMInstructionEraseFromParent(exp
->inst
);
2990 /* Change OFFSET to DEFAULT_VAL. */
2991 for (i
= 0; i
< num_outputs
; i
++) {
2992 if (vs_output_param_offset
[i
] == exp
->offset
) {
2993 vs_output_param_offset
[i
] =
2994 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
3001 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
3002 uint8_t *vs_output_param_offset
,
3003 uint32_t num_outputs
,
3004 struct ac_vs_exports
*processed
,
3005 struct ac_vs_exp_inst
*exp
)
3007 unsigned p
, copy_back_channels
= 0;
3009 /* See if the output is already in the list of processed outputs.
3010 * The LLVMValueRef comparison relies on SSA.
3012 for (p
= 0; p
< processed
->num
; p
++) {
3013 bool different
= false;
3015 for (unsigned j
= 0; j
< 4; j
++) {
3016 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
3017 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
3019 /* Treat undef as a match. */
3020 if (c2
->type
== AC_IR_UNDEF
)
3023 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3024 * and consider the instruction duplicated.
3026 if (c1
->type
== AC_IR_UNDEF
) {
3027 copy_back_channels
|= 1 << j
;
3031 /* Test whether the channels are not equal. */
3032 if (c1
->type
!= c2
->type
||
3033 (c1
->type
== AC_IR_CONST
&&
3034 c1
->const_float
!= c2
->const_float
) ||
3035 (c1
->type
== AC_IR_VALUE
&&
3036 c1
->value
!= c2
->value
)) {
3044 copy_back_channels
= 0;
3046 if (p
== processed
->num
)
3049 /* If a match was found, but the matching export has undef where the new
3050 * one has a normal value, copy the normal value to the undef channel.
3052 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3054 /* Get current enabled channels mask. */
3055 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3056 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3058 while (copy_back_channels
) {
3059 unsigned chan
= u_bit_scan(©_back_channels
);
3061 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3062 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3063 exp
->chan
[chan
].value
);
3064 match
->chan
[chan
] = exp
->chan
[chan
];
3066 /* Update number of enabled channels because the original mask
3067 * is not always 0xf.
3069 enabled_channels
|= (1 << chan
);
3070 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3071 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3074 /* The PARAM export is duplicated. Kill it. */
3075 LLVMInstructionEraseFromParent(exp
->inst
);
3077 /* Change OFFSET to the matching export. */
3078 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3079 if (vs_output_param_offset
[i
] == exp
->offset
) {
3080 vs_output_param_offset
[i
] = match
->offset
;
3087 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3088 LLVMValueRef main_fn
,
3089 uint8_t *vs_output_param_offset
,
3090 uint32_t num_outputs
,
3091 uint8_t *num_param_exports
)
3093 LLVMBasicBlockRef bb
;
3094 bool removed_any
= false;
3095 struct ac_vs_exports exports
;
3099 /* Process all LLVM instructions. */
3100 bb
= LLVMGetFirstBasicBlock(main_fn
);
3102 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3105 LLVMValueRef cur
= inst
;
3106 inst
= LLVMGetNextInstruction(inst
);
3107 struct ac_vs_exp_inst exp
;
3109 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3112 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3114 if (!ac_llvm_is_function(callee
))
3117 const char *name
= LLVMGetValueName(callee
);
3118 unsigned num_args
= LLVMCountParams(callee
);
3120 /* Check if this is an export instruction. */
3121 if ((num_args
!= 9 && num_args
!= 8) ||
3122 (strcmp(name
, "llvm.SI.export") &&
3123 strcmp(name
, "llvm.amdgcn.exp.f32")))
3126 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3127 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3129 if (target
< V_008DFC_SQ_EXP_PARAM
)
3132 target
-= V_008DFC_SQ_EXP_PARAM
;
3134 /* Parse the instruction. */
3135 memset(&exp
, 0, sizeof(exp
));
3136 exp
.offset
= target
;
3139 for (unsigned i
= 0; i
< 4; i
++) {
3140 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3142 exp
.chan
[i
].value
= v
;
3144 if (LLVMIsUndef(v
)) {
3145 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3146 } else if (LLVMIsAConstantFP(v
)) {
3147 LLVMBool loses_info
;
3148 exp
.chan
[i
].type
= AC_IR_CONST
;
3149 exp
.chan
[i
].const_float
=
3150 LLVMConstRealGetDouble(v
, &loses_info
);
3152 exp
.chan
[i
].type
= AC_IR_VALUE
;
3156 /* Eliminate constant and duplicated PARAM exports. */
3157 if (ac_eliminate_const_output(vs_output_param_offset
,
3158 num_outputs
, &exp
) ||
3159 ac_eliminate_duplicated_output(ctx
,
3160 vs_output_param_offset
,
3161 num_outputs
, &exports
,
3165 exports
.exp
[exports
.num
++] = exp
;
3168 bb
= LLVMGetNextBasicBlock(bb
);
3171 /* Remove holes in export memory due to removed PARAM exports.
3172 * This is done by renumbering all PARAM exports.
3175 uint8_t old_offset
[VARYING_SLOT_MAX
];
3178 /* Make a copy of the offsets. We need the old version while
3179 * we are modifying some of them. */
3180 memcpy(old_offset
, vs_output_param_offset
,
3181 sizeof(old_offset
));
3183 for (i
= 0; i
< exports
.num
; i
++) {
3184 unsigned offset
= exports
.exp
[i
].offset
;
3186 /* Update vs_output_param_offset. Multiple outputs can
3187 * have the same offset.
3189 for (out
= 0; out
< num_outputs
; out
++) {
3190 if (old_offset
[out
] == offset
)
3191 vs_output_param_offset
[out
] = i
;
3194 /* Change the PARAM offset in the instruction. */
3195 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3196 LLVMConstInt(ctx
->i32
,
3197 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3199 *num_param_exports
= exports
.num
;
3203 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3205 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3206 ac_build_intrinsic(ctx
,
3207 "llvm.amdgcn.init.exec", ctx
->voidt
,
3208 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3211 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3213 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3214 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3215 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3219 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3220 LLVMValueRef dw_addr
)
3222 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3225 void ac_lds_store(struct ac_llvm_context
*ctx
,
3226 LLVMValueRef dw_addr
,
3229 value
= ac_to_integer(ctx
, value
);
3230 ac_build_indexed_store(ctx
, ctx
->lds
,
3234 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3235 LLVMTypeRef dst_type
,
3238 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3239 const char *intrin_name
;
3243 switch (src0_bitsize
) {
3245 intrin_name
= "llvm.cttz.i64";
3250 intrin_name
= "llvm.cttz.i32";
3255 intrin_name
= "llvm.cttz.i16";
3260 intrin_name
= "llvm.cttz.i8";
3265 unreachable(!"invalid bitsize");
3268 LLVMValueRef params
[2] = {
3271 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3272 * add special code to check for x=0. The reason is that
3273 * the LLVM behavior for x=0 is different from what we
3274 * need here. However, LLVM also assumes that ffs(x) is
3275 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3276 * a conditional assignment to handle 0 is still required.
3278 * The hardware already implements the correct behavior.
3283 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3285 AC_FUNC_ATTR_READNONE
);
3287 if (src0_bitsize
== 64) {
3288 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3289 } else if (src0_bitsize
< 32) {
3290 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3293 /* TODO: We need an intrinsic to skip this conditional. */
3294 /* Check for zero: */
3295 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3298 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3301 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3303 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3306 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3308 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3311 static struct ac_llvm_flow
*
3312 get_current_flow(struct ac_llvm_context
*ctx
)
3314 if (ctx
->flow
->depth
> 0)
3315 return &ctx
->flow
->stack
[ctx
->flow
->depth
- 1];
3319 static struct ac_llvm_flow
*
3320 get_innermost_loop(struct ac_llvm_context
*ctx
)
3322 for (unsigned i
= ctx
->flow
->depth
; i
> 0; --i
) {
3323 if (ctx
->flow
->stack
[i
- 1].loop_entry_block
)
3324 return &ctx
->flow
->stack
[i
- 1];
3329 static struct ac_llvm_flow
*
3330 push_flow(struct ac_llvm_context
*ctx
)
3332 struct ac_llvm_flow
*flow
;
3334 if (ctx
->flow
->depth
>= ctx
->flow
->depth_max
) {
3335 unsigned new_max
= MAX2(ctx
->flow
->depth
<< 1,
3336 AC_LLVM_INITIAL_CF_DEPTH
);
3338 ctx
->flow
->stack
= realloc(ctx
->flow
->stack
, new_max
* sizeof(*ctx
->flow
->stack
));
3339 ctx
->flow
->depth_max
= new_max
;
3342 flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
];
3345 flow
->next_block
= NULL
;
3346 flow
->loop_entry_block
= NULL
;
3350 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3354 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3355 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3358 /* Append a basic block at the level of the parent flow.
3360 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3363 assert(ctx
->flow
->depth
>= 1);
3365 if (ctx
->flow
->depth
>= 2) {
3366 struct ac_llvm_flow
*flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
- 2];
3368 return LLVMInsertBasicBlockInContext(ctx
->context
,
3369 flow
->next_block
, name
);
3372 LLVMValueRef main_fn
=
3373 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3374 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3377 /* Emit a branch to the given default target for the current block if
3378 * applicable -- that is, if the current block does not already contain a
3379 * branch from a break or continue.
3381 static void emit_default_branch(LLVMBuilderRef builder
,
3382 LLVMBasicBlockRef target
)
3384 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3385 LLVMBuildBr(builder
, target
);
3388 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3390 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3391 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3392 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3393 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3394 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3395 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3398 void ac_build_break(struct ac_llvm_context
*ctx
)
3400 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3401 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3404 void ac_build_continue(struct ac_llvm_context
*ctx
)
3406 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3407 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3410 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3412 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3413 LLVMBasicBlockRef endif_block
;
3415 assert(!current_branch
->loop_entry_block
);
3417 endif_block
= append_basic_block(ctx
, "ENDIF");
3418 emit_default_branch(ctx
->builder
, endif_block
);
3420 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3421 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3423 current_branch
->next_block
= endif_block
;
3426 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3428 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3430 assert(!current_branch
->loop_entry_block
);
3432 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3433 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3434 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3439 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3441 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3443 assert(current_loop
->loop_entry_block
);
3445 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3447 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3448 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3452 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3454 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3455 LLVMBasicBlockRef if_block
;
3457 if_block
= append_basic_block(ctx
, "IF");
3458 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3459 set_basicblock_name(if_block
, "if", label_id
);
3460 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3461 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3464 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3467 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3468 value
, ctx
->f32_0
, "");
3469 ac_build_ifcc(ctx
, cond
, label_id
);
3472 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3475 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3476 ac_to_integer(ctx
, value
),
3478 ac_build_ifcc(ctx
, cond
, label_id
);
3481 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3484 LLVMBuilderRef builder
= ac
->builder
;
3485 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3486 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3487 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3488 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3489 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3493 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3495 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3498 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3499 LLVMDisposeBuilder(first_builder
);
3503 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3504 LLVMTypeRef type
, const char *name
)
3506 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3507 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3511 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3514 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3515 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3516 LLVMPointerType(type
, addr_space
), "");
3519 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3522 unsigned num_components
= ac_get_llvm_num_components(value
);
3523 if (count
== num_components
)
3526 LLVMValueRef masks
[MAX2(count
, 2)];
3527 masks
[0] = ctx
->i32_0
;
3528 masks
[1] = ctx
->i32_1
;
3529 for (unsigned i
= 2; i
< count
; i
++)
3530 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3533 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3536 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3537 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3540 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3541 unsigned rshift
, unsigned bitwidth
)
3543 LLVMValueRef value
= param
;
3545 value
= LLVMBuildLShr(ctx
->builder
, value
,
3546 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3548 if (rshift
+ bitwidth
< 32) {
3549 unsigned mask
= (1 << bitwidth
) - 1;
3550 value
= LLVMBuildAnd(ctx
->builder
, value
,
3551 LLVMConstInt(ctx
->i32
, mask
, false), "");
3556 /* Adjust the sample index according to FMASK.
3558 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3559 * which is the identity mapping. Each nibble says which physical sample
3560 * should be fetched to get that sample.
3562 * For example, 0x11111100 means there are only 2 samples stored and
3563 * the second sample covers 3/4 of the pixel. When reading samples 0
3564 * and 1, return physical sample 0 (determined by the first two 0s
3565 * in FMASK), otherwise return physical sample 1.
3567 * The sample index should be adjusted as follows:
3568 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3570 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3571 LLVMValueRef
*addr
, bool is_array_tex
)
3573 struct ac_image_args fmask_load
= {};
3574 fmask_load
.opcode
= ac_image_load
;
3575 fmask_load
.resource
= fmask
;
3576 fmask_load
.dmask
= 0xf;
3577 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3578 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3580 fmask_load
.coords
[0] = addr
[0];
3581 fmask_load
.coords
[1] = addr
[1];
3583 fmask_load
.coords
[2] = addr
[2];
3585 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3586 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3589 /* Apply the formula. */
3590 unsigned sample_chan
= is_array_tex
? 3 : 2;
3591 LLVMValueRef final_sample
;
3592 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3593 LLVMConstInt(ac
->i32
, 4, 0), "");
3594 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3595 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3596 * with EQAA, so those will map to 0. */
3597 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3598 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3600 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3601 * resource descriptor is 0 (invalid).
3604 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3605 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3606 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3608 /* Replace the MSAA sample index. */
3609 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3610 addr
[sample_chan
], "");
3614 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3616 LLVMTypeRef type
= LLVMTypeOf(src
);
3617 LLVMValueRef result
;
3619 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3621 lane
= LLVMBuildZExt(ctx
->builder
, lane
, ctx
->i32
, "");
3623 result
= ac_build_intrinsic(ctx
,
3624 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3625 ctx
->i32
, (LLVMValueRef
[]) { src
, lane
},
3626 lane
== NULL
? 1 : 2,
3627 AC_FUNC_ATTR_READNONE
|
3628 AC_FUNC_ATTR_CONVERGENT
);
3630 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3634 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3636 * The optimization barrier is not needed if the value is the same in all lanes
3637 * or if this is called in the outermost block.
3641 * @param lane - id of the lane or NULL for the first active lane
3642 * @return value of the lane
3644 LLVMValueRef
ac_build_readlane_no_opt_barrier(struct ac_llvm_context
*ctx
,
3645 LLVMValueRef src
, LLVMValueRef lane
)
3647 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3651 assert(bits
% 32 == 0);
3652 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3653 LLVMValueRef src_vector
=
3654 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3655 ret
= LLVMGetUndef(vec_type
);
3656 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3657 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3658 LLVMConstInt(ctx
->i32
, i
, 0), "");
3659 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
3660 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3661 LLVMConstInt(ctx
->i32
, i
, 0), "");
3664 ret
= _ac_build_readlane(ctx
, src
, lane
);
3671 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3673 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3674 src
= ac_to_integer(ctx
, src
);
3677 ac_build_optimization_barrier(ctx
, &src
);
3679 ret
= ac_build_readlane_no_opt_barrier(ctx
, src
, lane
);
3680 if (LLVMGetTypeKind(src_type
) == LLVMPointerTypeKind
)
3681 return LLVMBuildIntToPtr(ctx
->builder
, ret
, src_type
, "");
3682 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3686 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3688 return ac_build_intrinsic(ctx
, "llvm.amdgcn.writelane", ctx
->i32
,
3689 (LLVMValueRef
[]) {value
, lane
, src
}, 3,
3690 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3694 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3696 if (ctx
->wave_size
== 32) {
3697 return ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3698 (LLVMValueRef
[]) { mask
, ctx
->i32_0
},
3699 2, AC_FUNC_ATTR_READNONE
);
3701 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3702 LLVMVectorType(ctx
->i32
, 2),
3704 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3706 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3709 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3710 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3711 2, AC_FUNC_ATTR_READNONE
);
3712 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3713 (LLVMValueRef
[]) { mask_hi
, val
},
3714 2, AC_FUNC_ATTR_READNONE
);
3719 _dpp_quad_perm
= 0x000,
3720 _dpp_row_sl
= 0x100,
3721 _dpp_row_sr
= 0x110,
3722 _dpp_row_rr
= 0x120,
3727 dpp_row_mirror
= 0x140,
3728 dpp_row_half_mirror
= 0x141,
3729 dpp_row_bcast15
= 0x142,
3730 dpp_row_bcast31
= 0x143
3733 static inline enum dpp_ctrl
3734 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3736 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3737 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3740 static inline enum dpp_ctrl
3741 dpp_row_sl(unsigned amount
)
3743 assert(amount
> 0 && amount
< 16);
3744 return _dpp_row_sl
| amount
;
3747 static inline enum dpp_ctrl
3748 dpp_row_sr(unsigned amount
)
3750 assert(amount
> 0 && amount
< 16);
3751 return _dpp_row_sr
| amount
;
3755 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3756 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3759 LLVMTypeRef type
= LLVMTypeOf(src
);
3762 old
= LLVMBuildZExt(ctx
->builder
, old
, ctx
->i32
, "");
3763 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3765 res
= ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32", ctx
->i32
,
3768 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3769 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3770 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3771 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3772 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3774 return LLVMBuildTrunc(ctx
->builder
, res
, type
, "");
3778 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3779 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3782 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3783 src
= ac_to_integer(ctx
, src
);
3784 old
= ac_to_integer(ctx
, old
);
3785 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3788 assert(bits
% 32 == 0);
3789 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3790 LLVMValueRef src_vector
=
3791 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3792 LLVMValueRef old_vector
=
3793 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3794 ret
= LLVMGetUndef(vec_type
);
3795 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3796 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3797 LLVMConstInt(ctx
->i32
, i
,
3799 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3800 LLVMConstInt(ctx
->i32
, i
,
3802 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3807 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3809 LLVMConstInt(ctx
->i32
, i
,
3813 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3814 bank_mask
, bound_ctrl
);
3816 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3820 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3821 bool exchange_rows
, bool bound_ctrl
)
3823 LLVMTypeRef type
= LLVMTypeOf(src
);
3824 LLVMValueRef result
;
3826 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3828 LLVMValueRef args
[6] = {
3831 LLVMConstInt(ctx
->i32
, sel
, false),
3832 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3833 ctx
->i1true
, /* fi */
3834 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3837 result
= ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3838 : "llvm.amdgcn.permlane16",
3840 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3842 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3846 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3847 bool exchange_rows
, bool bound_ctrl
)
3849 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3850 src
= ac_to_integer(ctx
, src
);
3851 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3854 assert(bits
% 32 == 0);
3855 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3856 LLVMValueRef src_vector
=
3857 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3858 ret
= LLVMGetUndef(vec_type
);
3859 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3860 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3861 LLVMConstInt(ctx
->i32
, i
,
3863 LLVMValueRef ret_comp
=
3864 _ac_build_permlane16(ctx
, src
, sel
,
3867 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3869 LLVMConstInt(ctx
->i32
, i
,
3873 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3876 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3879 static inline unsigned
3880 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3882 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3883 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3887 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3889 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3892 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3894 ret
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle", ctx
->i32
,
3896 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3897 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3899 return LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3903 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3905 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3906 src
= ac_to_integer(ctx
, src
);
3907 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3910 assert(bits
% 32 == 0);
3911 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3912 LLVMValueRef src_vector
=
3913 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3914 ret
= LLVMGetUndef(vec_type
);
3915 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3916 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3917 LLVMConstInt(ctx
->i32
, i
,
3919 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3921 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3923 LLVMConstInt(ctx
->i32
, i
,
3927 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3929 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3933 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3935 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3936 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3937 char name
[32], type
[8];
3940 src
= ac_to_integer(ctx
, src
);
3943 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3945 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3946 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3947 ret
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3948 (LLVMValueRef
[]) { src
}, 1,
3949 AC_FUNC_ATTR_READNONE
);
3952 ret
= LLVMBuildTrunc(ctx
->builder
, ret
,
3953 ac_to_integer_type(ctx
, src_type
), "");
3955 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3959 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3960 LLVMValueRef inactive
)
3962 char name
[33], type
[8];
3963 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3964 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3965 src
= ac_to_integer(ctx
, src
);
3966 inactive
= ac_to_integer(ctx
, inactive
);
3969 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3970 inactive
= LLVMBuildZExt(ctx
->builder
, inactive
, ctx
->i32
, "");
3973 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3974 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3976 ac_build_intrinsic(ctx
, name
,
3977 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3979 AC_FUNC_ATTR_READNONE
|
3980 AC_FUNC_ATTR_CONVERGENT
);
3982 ret
= LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3988 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3990 if (type_size
== 1) {
3992 case nir_op_iadd
: return ctx
->i8_0
;
3993 case nir_op_imul
: return ctx
->i8_1
;
3994 case nir_op_imin
: return LLVMConstInt(ctx
->i8
, INT8_MAX
, 0);
3995 case nir_op_umin
: return LLVMConstInt(ctx
->i8
, UINT8_MAX
, 0);
3996 case nir_op_imax
: return LLVMConstInt(ctx
->i8
, INT8_MIN
, 0);
3997 case nir_op_umax
: return ctx
->i8_0
;
3998 case nir_op_iand
: return LLVMConstInt(ctx
->i8
, -1, 0);
3999 case nir_op_ior
: return ctx
->i8_0
;
4000 case nir_op_ixor
: return ctx
->i8_0
;
4002 unreachable("bad reduction intrinsic");
4004 } else if (type_size
== 2) {
4006 case nir_op_iadd
: return ctx
->i16_0
;
4007 case nir_op_fadd
: return ctx
->f16_0
;
4008 case nir_op_imul
: return ctx
->i16_1
;
4009 case nir_op_fmul
: return ctx
->f16_1
;
4010 case nir_op_imin
: return LLVMConstInt(ctx
->i16
, INT16_MAX
, 0);
4011 case nir_op_umin
: return LLVMConstInt(ctx
->i16
, UINT16_MAX
, 0);
4012 case nir_op_fmin
: return LLVMConstReal(ctx
->f16
, INFINITY
);
4013 case nir_op_imax
: return LLVMConstInt(ctx
->i16
, INT16_MIN
, 0);
4014 case nir_op_umax
: return ctx
->i16_0
;
4015 case nir_op_fmax
: return LLVMConstReal(ctx
->f16
, -INFINITY
);
4016 case nir_op_iand
: return LLVMConstInt(ctx
->i16
, -1, 0);
4017 case nir_op_ior
: return ctx
->i16_0
;
4018 case nir_op_ixor
: return ctx
->i16_0
;
4020 unreachable("bad reduction intrinsic");
4022 } else if (type_size
== 4) {
4024 case nir_op_iadd
: return ctx
->i32_0
;
4025 case nir_op_fadd
: return ctx
->f32_0
;
4026 case nir_op_imul
: return ctx
->i32_1
;
4027 case nir_op_fmul
: return ctx
->f32_1
;
4028 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
4029 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
4030 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
4031 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
4032 case nir_op_umax
: return ctx
->i32_0
;
4033 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
4034 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
4035 case nir_op_ior
: return ctx
->i32_0
;
4036 case nir_op_ixor
: return ctx
->i32_0
;
4038 unreachable("bad reduction intrinsic");
4040 } else { /* type_size == 64bit */
4042 case nir_op_iadd
: return ctx
->i64_0
;
4043 case nir_op_fadd
: return ctx
->f64_0
;
4044 case nir_op_imul
: return ctx
->i64_1
;
4045 case nir_op_fmul
: return ctx
->f64_1
;
4046 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
4047 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4048 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
4049 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
4050 case nir_op_umax
: return ctx
->i64_0
;
4051 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
4052 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
4053 case nir_op_ior
: return ctx
->i64_0
;
4054 case nir_op_ixor
: return ctx
->i64_0
;
4056 unreachable("bad reduction intrinsic");
4062 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
4064 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
4065 bool _32bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 4;
4067 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
4068 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
4069 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
4070 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
4071 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
4072 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
4074 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
4075 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
4077 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
4078 _64bit
? "llvm.minnum.f64" : _32bit
? "llvm.minnum.f32" : "llvm.minnum.f16",
4079 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4080 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4081 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
4082 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
4084 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
4085 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
4087 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
4088 _64bit
? "llvm.maxnum.f64" : _32bit
? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4089 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4090 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4091 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
4092 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
4093 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
4095 unreachable("bad reduction intrinsic");
4100 * \param src The value to shift.
4101 * \param identity The value to use the first lane.
4102 * \param maxprefix specifies that the result only needs to be correct for a
4103 * prefix of this many threads
4104 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4107 ac_wavefront_shift_right_1(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4108 LLVMValueRef identity
, unsigned maxprefix
)
4110 if (ctx
->chip_class
>= GFX10
) {
4111 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4112 LLVMValueRef active
, tmp1
, tmp2
;
4113 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4115 tmp1
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4117 tmp2
= ac_build_permlane16(ctx
, src
, (uint64_t)~0, true, false);
4119 if (maxprefix
> 32) {
4120 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4121 LLVMConstInt(ctx
->i32
, 32, false), "");
4123 tmp2
= LLVMBuildSelect(ctx
->builder
, active
,
4124 ac_build_readlane(ctx
, src
,
4125 LLVMConstInt(ctx
->i32
, 31, false)),
4128 active
= LLVMBuildOr(ctx
->builder
, active
,
4129 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4130 LLVMBuildAnd(ctx
->builder
, tid
,
4131 LLVMConstInt(ctx
->i32
, 0x1f, false), ""),
4132 LLVMConstInt(ctx
->i32
, 0x10, false), ""), "");
4133 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4134 } else if (maxprefix
> 16) {
4135 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4136 LLVMConstInt(ctx
->i32
, 16, false), "");
4138 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4140 } else if (ctx
->chip_class
>= GFX8
) {
4141 return ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
4144 /* wavefront shift_right by 1 on SI/CI */
4145 LLVMValueRef active
, tmp1
, tmp2
;
4146 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4147 tmp1
= ac_build_ds_swizzle(ctx
, src
, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4148 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4149 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4150 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x7, 0), ""),
4151 LLVMConstInt(ctx
->i32
, 0x4, 0), "");
4152 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4153 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4154 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4155 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0xf, 0), ""),
4156 LLVMConstInt(ctx
->i32
, 0x8, 0), "");
4157 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4158 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4159 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4160 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x1f, 0), ""),
4161 LLVMConstInt(ctx
->i32
, 0x10, 0), "");
4162 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4163 tmp2
= ac_build_readlane(ctx
, src
, LLVMConstInt(ctx
->i32
, 31, 0));
4164 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), "");
4165 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4166 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 0, 0), "");
4167 return LLVMBuildSelect(ctx
->builder
, active
, identity
, tmp1
, "");
4171 * \param maxprefix specifies that the result only needs to be correct for a
4172 * prefix of this many threads
4175 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
4176 unsigned maxprefix
, bool inclusive
)
4178 LLVMValueRef result
, tmp
;
4181 src
= ac_wavefront_shift_right_1(ctx
, src
, identity
, maxprefix
);
4185 if (ctx
->chip_class
<= GFX7
) {
4186 assert(maxprefix
== 64);
4187 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4188 LLVMValueRef active
;
4189 tmp
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4190 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4191 LLVMBuildAnd(ctx
->builder
, tid
, ctx
->i32_1
, ""),
4193 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4194 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4195 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4196 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4197 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 2, 0), ""),
4199 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4200 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4201 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4202 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4203 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 4, 0), ""),
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(0x10, 0x07, 0x00));
4208 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4209 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 8, 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(0x00, 0x0f, 0x00));
4214 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4215 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 16, 0), ""),
4217 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4218 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4219 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, 0));
4220 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4221 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), ""),
4223 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4224 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4230 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4231 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4234 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
4235 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4238 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
4239 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4242 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
4243 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4246 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
4247 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4248 if (maxprefix
<= 16)
4251 if (ctx
->chip_class
>= GFX10
) {
4252 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4253 LLVMValueRef active
;
4255 tmp
= ac_build_permlane16(ctx
, result
, ~(uint64_t)0, true, false);
4257 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4258 LLVMBuildAnd(ctx
->builder
, tid
,
4259 LLVMConstInt(ctx
->i32
, 16, false), ""),
4262 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4264 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4266 if (maxprefix
<= 32)
4269 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4271 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, tid
,
4272 LLVMConstInt(ctx
->i32
, 32, false), "");
4274 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4276 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4280 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4281 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4282 if (maxprefix
<= 32)
4284 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4285 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4290 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4292 LLVMValueRef result
;
4294 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4295 LLVMBuilderRef builder
= ctx
->builder
;
4296 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4297 result
= ac_build_ballot(ctx
, src
);
4298 result
= ac_build_mbcnt(ctx
, result
);
4299 result
= LLVMBuildAdd(builder
, result
, src
, "");
4303 ac_build_optimization_barrier(ctx
, &src
);
4305 LLVMValueRef identity
=
4306 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4307 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4308 LLVMTypeOf(identity
), "");
4309 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, true);
4311 return ac_build_wwm(ctx
, result
);
4315 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4317 LLVMValueRef result
;
4319 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4320 LLVMBuilderRef builder
= ctx
->builder
;
4321 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4322 result
= ac_build_ballot(ctx
, src
);
4323 result
= ac_build_mbcnt(ctx
, result
);
4327 ac_build_optimization_barrier(ctx
, &src
);
4329 LLVMValueRef identity
=
4330 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4331 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4332 LLVMTypeOf(identity
), "");
4333 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, false);
4335 return ac_build_wwm(ctx
, result
);
4339 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4341 if (cluster_size
== 1) return src
;
4342 ac_build_optimization_barrier(ctx
, &src
);
4343 LLVMValueRef result
, swap
;
4344 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4345 ac_get_type_size(LLVMTypeOf(src
)));
4346 result
= LLVMBuildBitCast(ctx
->builder
,
4347 ac_build_set_inactive(ctx
, src
, identity
),
4348 LLVMTypeOf(identity
), "");
4349 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4350 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4351 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4353 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4354 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4355 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4357 if (ctx
->chip_class
>= GFX8
)
4358 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4360 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4361 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4362 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4364 if (ctx
->chip_class
>= GFX8
)
4365 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4367 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4368 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4369 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4371 if (ctx
->chip_class
>= GFX10
)
4372 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4373 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4374 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4376 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4377 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4378 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4380 if (ctx
->chip_class
>= GFX8
) {
4381 if (ctx
->wave_size
== 64) {
4382 if (ctx
->chip_class
>= GFX10
)
4383 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4385 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4386 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4387 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4390 return ac_build_wwm(ctx
, result
);
4392 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4393 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4394 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4395 return ac_build_wwm(ctx
, result
);
4400 * "Top half" of a scan that reduces per-wave values across an entire
4403 * The source value must be present in the highest lane of the wave, and the
4404 * highest lane must be live.
4407 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4409 if (ws
->maxwaves
<= 1)
4412 const LLVMValueRef last_lane
= LLVMConstInt(ctx
->i32
, ctx
->wave_size
- 1, false);
4413 LLVMBuilderRef builder
= ctx
->builder
;
4414 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4417 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, last_lane
, "");
4418 ac_build_ifcc(ctx
, tmp
, 1000);
4419 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4420 ac_build_endif(ctx
, 1000);
4424 * "Bottom half" of a scan that reduces per-wave values across an entire
4427 * The caller must place a barrier between the top and bottom halves.
4430 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4432 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4433 const LLVMValueRef identity
=
4434 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4436 if (ws
->maxwaves
<= 1) {
4437 ws
->result_reduce
= ws
->src
;
4438 ws
->result_inclusive
= ws
->src
;
4439 ws
->result_exclusive
= identity
;
4442 assert(ws
->maxwaves
<= 32);
4444 LLVMBuilderRef builder
= ctx
->builder
;
4445 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4446 LLVMBasicBlockRef bbs
[2];
4447 LLVMValueRef phivalues_scan
[2];
4448 LLVMValueRef tmp
, tmp2
;
4450 bbs
[0] = LLVMGetInsertBlock(builder
);
4451 phivalues_scan
[0] = LLVMGetUndef(type
);
4453 if (ws
->enable_reduce
)
4454 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4455 else if (ws
->enable_inclusive
)
4456 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4458 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4459 ac_build_ifcc(ctx
, tmp
, 1001);
4461 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4463 ac_build_optimization_barrier(ctx
, &tmp
);
4465 bbs
[1] = LLVMGetInsertBlock(builder
);
4466 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4468 ac_build_endif(ctx
, 1001);
4470 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4472 if (ws
->enable_reduce
) {
4473 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4474 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4476 if (ws
->enable_inclusive
)
4477 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4478 if (ws
->enable_exclusive
) {
4479 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4480 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4481 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4482 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4487 * Inclusive scan of a per-wave value across an entire workgroup.
4489 * This implies an s_barrier instruction.
4491 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4492 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4493 * useful manner because of the barrier in the algorithm.)
4496 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4498 ac_build_wg_wavescan_top(ctx
, ws
);
4499 ac_build_s_barrier(ctx
);
4500 ac_build_wg_wavescan_bottom(ctx
, ws
);
4504 * "Top half" of a scan that reduces per-thread values across an entire
4507 * All lanes must be active when this code runs.
4510 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4512 if (ws
->enable_exclusive
) {
4513 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4514 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4515 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4516 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4518 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4521 bool enable_inclusive
= ws
->enable_inclusive
;
4522 bool enable_exclusive
= ws
->enable_exclusive
;
4523 ws
->enable_inclusive
= false;
4524 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4525 ac_build_wg_wavescan_top(ctx
, ws
);
4526 ws
->enable_inclusive
= enable_inclusive
;
4527 ws
->enable_exclusive
= enable_exclusive
;
4531 * "Bottom half" of a scan that reduces per-thread values across an entire
4534 * The caller must place a barrier between the top and bottom halves.
4537 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4539 bool enable_inclusive
= ws
->enable_inclusive
;
4540 bool enable_exclusive
= ws
->enable_exclusive
;
4541 ws
->enable_inclusive
= false;
4542 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4543 ac_build_wg_wavescan_bottom(ctx
, ws
);
4544 ws
->enable_inclusive
= enable_inclusive
;
4545 ws
->enable_exclusive
= enable_exclusive
;
4547 /* ws->result_reduce is already the correct value */
4548 if (ws
->enable_inclusive
)
4549 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4550 if (ws
->enable_exclusive
)
4551 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4555 * A scan that reduces per-thread values across an entire workgroup.
4557 * The caller must ensure that all lanes are active when this code runs
4558 * (WWM is insufficient!), because there is an implied barrier.
4561 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4563 ac_build_wg_scan_top(ctx
, ws
);
4564 ac_build_s_barrier(ctx
);
4565 ac_build_wg_scan_bottom(ctx
, ws
);
4569 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4570 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4572 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4573 if (ctx
->chip_class
>= GFX8
) {
4574 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4576 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4581 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4583 LLVMTypeRef type
= LLVMTypeOf(src
);
4584 LLVMValueRef result
;
4586 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4587 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
4589 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4590 (LLVMValueRef
[]) {index
, src
}, 2,
4591 AC_FUNC_ATTR_READNONE
|
4592 AC_FUNC_ATTR_CONVERGENT
);
4593 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
4597 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4603 if (bitsize
== 16) {
4604 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4606 } else if (bitsize
== 32) {
4607 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4610 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4614 LLVMValueRef params
[] = {
4617 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4618 AC_FUNC_ATTR_READNONE
);
4621 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4627 if (bitsize
== 16) {
4628 intr
= "llvm.amdgcn.frexp.mant.f16";
4630 } else if (bitsize
== 32) {
4631 intr
= "llvm.amdgcn.frexp.mant.f32";
4634 intr
= "llvm.amdgcn.frexp.mant.f64";
4638 LLVMValueRef params
[] = {
4641 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4642 AC_FUNC_ATTR_READNONE
);
4646 ac_build_canonicalize(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4652 if (bitsize
== 16) {
4653 intr
= "llvm.canonicalize.f16";
4655 } else if (bitsize
== 32) {
4656 intr
= "llvm.canonicalize.f32";
4659 intr
= "llvm.canonicalize.f64";
4663 LLVMValueRef params
[] = {
4666 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4667 AC_FUNC_ATTR_READNONE
);
4671 * this takes an I,J coordinate pair,
4672 * and works out the X and Y derivatives.
4673 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4676 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4678 LLVMValueRef result
[4], a
;
4681 for (i
= 0; i
< 2; i
++) {
4682 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4683 LLVMConstInt(ctx
->i32
, i
, false), "");
4684 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4685 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4687 return ac_build_gather_values(ctx
, result
, 4);
4691 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4693 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4695 AC_FUNC_ATTR_READNONE
);
4696 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4697 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4701 ac_build_is_helper_invocation(struct ac_llvm_context
*ctx
)
4703 /* !(exact && postponed) */
4704 LLVMValueRef exact
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4706 AC_FUNC_ATTR_READNONE
);
4708 LLVMValueRef postponed
= LLVMBuildLoad(ctx
->builder
, ctx
->postponed_kill
, "");
4709 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, exact
, postponed
, "");
4711 return LLVMBuildSelect(ctx
->builder
, result
, ctx
->i32_0
,
4712 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false), "");
4715 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4716 LLVMValueRef
*args
, unsigned num_args
)
4718 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4719 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));
4724 ac_export_mrt_z(struct ac_llvm_context
*ctx
, LLVMValueRef depth
,
4725 LLVMValueRef stencil
, LLVMValueRef samplemask
,
4726 struct ac_export_args
*args
)
4729 unsigned format
= ac_get_spi_shader_z_format(depth
!= NULL
,
4731 samplemask
!= NULL
);
4733 assert(depth
|| stencil
|| samplemask
);
4735 memset(args
, 0, sizeof(*args
));
4737 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
4738 args
->done
= 1; /* DONE bit */
4740 /* Specify the target we are exporting */
4741 args
->target
= V_008DFC_SQ_EXP_MRTZ
;
4743 args
->compr
= 0; /* COMP flag */
4744 args
->out
[0] = LLVMGetUndef(ctx
->f32
); /* R, depth */
4745 args
->out
[1] = LLVMGetUndef(ctx
->f32
); /* G, stencil test val[0:7], stencil op val[8:15] */
4746 args
->out
[2] = LLVMGetUndef(ctx
->f32
); /* B, sample mask */
4747 args
->out
[3] = LLVMGetUndef(ctx
->f32
); /* A, alpha to mask */
4749 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
4751 args
->compr
= 1; /* COMPR flag */
4754 /* Stencil should be in X[23:16]. */
4755 stencil
= ac_to_integer(ctx
, stencil
);
4756 stencil
= LLVMBuildShl(ctx
->builder
, stencil
,
4757 LLVMConstInt(ctx
->i32
, 16, 0), "");
4758 args
->out
[0] = ac_to_float(ctx
, stencil
);
4762 /* SampleMask should be in Y[15:0]. */
4763 args
->out
[1] = samplemask
;
4768 args
->out
[0] = depth
;
4772 args
->out
[1] = stencil
;
4776 args
->out
[2] = samplemask
;
4781 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4782 * at the X writemask component. */
4783 if (ctx
->chip_class
== GFX6
&&
4784 ctx
->family
!= CHIP_OLAND
&&
4785 ctx
->family
!= CHIP_HAINAN
)
4788 /* Specify which components to enable */
4789 args
->enabled_channels
= mask
;
4792 /* Send GS Alloc Req message from the first wave of the group to SPI.
4793 * Message payload is:
4794 * - bits 0..10: vertices in group
4795 * - bits 12..22: primitives in group
4797 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context
*ctx
, LLVMValueRef wave_id
,
4798 LLVMValueRef vtx_cnt
, LLVMValueRef prim_cnt
)
4800 LLVMBuilderRef builder
= ctx
->builder
;
4803 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, wave_id
, ctx
->i32_0
, ""), 5020);
4805 tmp
= LLVMBuildShl(builder
, prim_cnt
, LLVMConstInt(ctx
->i32
, 12, false),"");
4806 tmp
= LLVMBuildOr(builder
, tmp
, vtx_cnt
, "");
4807 ac_build_sendmsg(ctx
, AC_SENDMSG_GS_ALLOC_REQ
, tmp
);
4809 ac_build_endif(ctx
, 5020);
4812 LLVMValueRef
ac_pack_prim_export(struct ac_llvm_context
*ctx
,
4813 const struct ac_ngg_prim
*prim
)
4815 /* The prim export format is:
4816 * - bits 0..8: index 0
4817 * - bit 9: edge flag 0
4818 * - bits 10..18: index 1
4819 * - bit 19: edge flag 1
4820 * - bits 20..28: index 2
4821 * - bit 29: edge flag 2
4822 * - bit 31: null primitive (skip)
4824 LLVMBuilderRef builder
= ctx
->builder
;
4825 LLVMValueRef tmp
= LLVMBuildZExt(builder
, prim
->isnull
, ctx
->i32
, "");
4826 LLVMValueRef result
= LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->i32
, 31, false), "");
4828 for (unsigned i
= 0; i
< prim
->num_vertices
; ++i
) {
4829 tmp
= LLVMBuildShl(builder
, prim
->index
[i
],
4830 LLVMConstInt(ctx
->i32
, 10 * i
, false), "");
4831 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4832 tmp
= LLVMBuildZExt(builder
, prim
->edgeflag
[i
], ctx
->i32
, "");
4833 tmp
= LLVMBuildShl(builder
, tmp
,
4834 LLVMConstInt(ctx
->i32
, 10 * i
+ 9, false), "");
4835 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4840 void ac_build_export_prim(struct ac_llvm_context
*ctx
,
4841 const struct ac_ngg_prim
*prim
)
4843 struct ac_export_args args
;
4845 if (prim
->passthrough
) {
4846 args
.out
[0] = prim
->passthrough
;
4848 args
.out
[0] = ac_pack_prim_export(ctx
, prim
);
4851 args
.out
[0] = LLVMBuildBitCast(ctx
->builder
, args
.out
[0], ctx
->f32
, "");
4852 args
.out
[1] = LLVMGetUndef(ctx
->f32
);
4853 args
.out
[2] = LLVMGetUndef(ctx
->f32
);
4854 args
.out
[3] = LLVMGetUndef(ctx
->f32
);
4856 args
.target
= V_008DFC_SQ_EXP_PRIM
;
4857 args
.enabled_channels
= 1;
4859 args
.valid_mask
= false;
4862 ac_build_export(ctx
, &args
);
4866 arg_llvm_type(enum ac_arg_type type
, unsigned size
, struct ac_llvm_context
*ctx
)
4868 if (type
== AC_ARG_FLOAT
) {
4869 return size
== 1 ? ctx
->f32
: LLVMVectorType(ctx
->f32
, size
);
4870 } else if (type
== AC_ARG_INT
) {
4871 return size
== 1 ? ctx
->i32
: LLVMVectorType(ctx
->i32
, size
);
4873 LLVMTypeRef ptr_type
;
4875 case AC_ARG_CONST_PTR
:
4878 case AC_ARG_CONST_FLOAT_PTR
:
4879 ptr_type
= ctx
->f32
;
4881 case AC_ARG_CONST_PTR_PTR
:
4882 ptr_type
= ac_array_in_const32_addr_space(ctx
->i8
);
4884 case AC_ARG_CONST_DESC_PTR
:
4885 ptr_type
= ctx
->v4i32
;
4887 case AC_ARG_CONST_IMAGE_PTR
:
4888 ptr_type
= ctx
->v8i32
;
4891 unreachable("unknown arg type");
4894 return ac_array_in_const32_addr_space(ptr_type
);
4897 return ac_array_in_const_addr_space(ptr_type
);
4903 ac_build_main(const struct ac_shader_args
*args
,
4904 struct ac_llvm_context
*ctx
,
4905 enum ac_llvm_calling_convention convention
,
4906 const char *name
, LLVMTypeRef ret_type
,
4907 LLVMModuleRef module
)
4909 LLVMTypeRef arg_types
[AC_MAX_ARGS
];
4911 for (unsigned i
= 0; i
< args
->arg_count
; i
++) {
4912 arg_types
[i
] = arg_llvm_type(args
->args
[i
].type
,
4913 args
->args
[i
].size
, ctx
);
4916 LLVMTypeRef main_function_type
=
4917 LLVMFunctionType(ret_type
, arg_types
, args
->arg_count
, 0);
4919 LLVMValueRef main_function
=
4920 LLVMAddFunction(module
, name
, main_function_type
);
4921 LLVMBasicBlockRef main_function_body
=
4922 LLVMAppendBasicBlockInContext(ctx
->context
, main_function
, "main_body");
4923 LLVMPositionBuilderAtEnd(ctx
->builder
, main_function_body
);
4925 LLVMSetFunctionCallConv(main_function
, convention
);
4926 for (unsigned i
= 0; i
< args
->arg_count
; ++i
) {
4927 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
4929 if (args
->args
[i
].file
!= AC_ARG_SGPR
)
4932 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
4934 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4935 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
4936 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4940 ctx
->main_function
= main_function
;
4941 return main_function
;
4944 void ac_build_s_endpgm(struct ac_llvm_context
*ctx
)
4946 LLVMTypeRef calltype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
4947 LLVMValueRef code
= LLVMConstInlineAsm(calltype
, "s_endpgm", "", true, false);
4948 LLVMBuildCall(ctx
->builder
, code
, NULL
, 0, "");
4951 LLVMValueRef
ac_prefix_bitcount(struct ac_llvm_context
*ctx
,
4952 LLVMValueRef mask
, LLVMValueRef index
)
4954 LLVMBuilderRef builder
= ctx
->builder
;
4955 LLVMTypeRef type
= LLVMTypeOf(mask
);
4957 LLVMValueRef bit
= LLVMBuildShl(builder
, LLVMConstInt(type
, 1, 0),
4958 LLVMBuildZExt(builder
, index
, type
, ""), "");
4959 LLVMValueRef prefix_bits
= LLVMBuildSub(builder
, bit
, LLVMConstInt(type
, 1, 0), "");
4960 LLVMValueRef prefix_mask
= LLVMBuildAnd(builder
, mask
, prefix_bits
, "");
4961 return ac_build_bit_count(ctx
, prefix_mask
);
4964 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
4965 LLVMValueRef
ac_prefix_bitcount_2x64(struct ac_llvm_context
*ctx
,
4966 LLVMValueRef mask
[2], LLVMValueRef index
)
4968 LLVMBuilderRef builder
= ctx
->builder
;
4970 /* Reference version using i128. */
4971 LLVMValueRef input_mask
=
4972 LLVMBuildBitCast(builder
, ac_build_gather_values(ctx
, mask
, 2), ctx
->i128
, "");
4974 return ac_prefix_bitcount(ctx
, input_mask
, index
);
4976 /* Optimized version using 2 64-bit masks. */
4977 LLVMValueRef is_hi
, is_0
, c64
, c128
, all_bits
;
4978 LLVMValueRef prefix_mask
[2], shift
[2], mask_bcnt0
, prefix_bcnt
[2];
4980 /* Compute the 128-bit prefix mask. */
4981 c64
= LLVMConstInt(ctx
->i32
, 64, 0);
4982 c128
= LLVMConstInt(ctx
->i32
, 128, 0);
4983 all_bits
= LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4984 /* The first index that can have non-zero high bits in the prefix mask is 65. */
4985 is_hi
= LLVMBuildICmp(builder
, LLVMIntUGT
, index
, c64
, "");
4986 is_0
= LLVMBuildICmp(builder
, LLVMIntEQ
, index
, ctx
->i32_0
, "");
4987 mask_bcnt0
= ac_build_bit_count(ctx
, mask
[0]);
4989 for (unsigned i
= 0; i
< 2; i
++) {
4990 shift
[i
] = LLVMBuildSub(builder
, i
? c128
: c64
, index
, "");
4991 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
4992 * so we handle it by the is_0 select.
4993 * For i==1, index==64, same story, so we handle it by the last is_hi select.
4994 * For i==0, index==64, we shift by 0, which is what we want.
4996 prefix_mask
[i
] = LLVMBuildLShr(builder
, all_bits
,
4997 LLVMBuildZExt(builder
, shift
[i
], ctx
->i64
, ""), "");
4998 prefix_mask
[i
] = LLVMBuildAnd(builder
, mask
[i
], prefix_mask
[i
], "");
4999 prefix_bcnt
[i
] = ac_build_bit_count(ctx
, prefix_mask
[i
]);
5002 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_0
, ctx
->i32_0
, prefix_bcnt
[0], "");
5003 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_hi
, mask_bcnt0
, prefix_bcnt
[0], "");
5004 prefix_bcnt
[1] = LLVMBuildSelect(builder
, is_hi
, prefix_bcnt
[1], ctx
->i32_0
, "");
5006 return LLVMBuildAdd(builder
, prefix_bcnt
[0], prefix_bcnt
[1], "");
5011 * Convert triangle strip indices to triangle indices. This is used to decompose
5012 * triangle strips into triangles.
5014 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context
*ctx
,
5015 LLVMValueRef is_odd
,
5016 LLVMValueRef flatshade_first
,
5017 LLVMValueRef index
[3])
5019 LLVMBuilderRef builder
= ctx
->builder
;
5020 LLVMValueRef out
[3];
5022 /* We need to change the vertex order for odd triangles to get correct
5023 * front/back facing by swapping 2 vertex indices, but we also have to
5024 * keep the provoking vertex in the same place.
5026 * If the first vertex is provoking, swap index 1 and 2.
5027 * If the last vertex is provoking, swap index 0 and 1.
5029 out
[0] = LLVMBuildSelect(builder
, flatshade_first
,
5031 LLVMBuildSelect(builder
, is_odd
,
5032 index
[1], index
[0], ""), "");
5033 out
[1] = LLVMBuildSelect(builder
, flatshade_first
,
5034 LLVMBuildSelect(builder
, is_odd
,
5035 index
[2], index
[1], ""),
5036 LLVMBuildSelect(builder
, is_odd
,
5037 index
[0], index
[1], ""), "");
5038 out
[2] = LLVMBuildSelect(builder
, flatshade_first
,
5039 LLVMBuildSelect(builder
, is_odd
,
5040 index
[1], index
[2], ""),
5042 memcpy(index
, out
, sizeof(out
));