2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
29 #include <llvm/Config/llvm-config.h>
31 #include "c11/threads.h"
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
45 #include "shader_enums.h"
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block
;
54 LLVMBasicBlockRef loop_entry_block
;
57 /* Initialize module-independent parts of the context.
59 * The caller is responsible for initializing ctx::module and ctx::builder.
62 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
63 struct ac_llvm_compiler
*compiler
,
64 enum chip_class chip_class
, enum radeon_family family
,
65 enum ac_float_mode float_mode
, unsigned wave_size
,
66 unsigned ballot_mask_bits
)
68 ctx
->context
= LLVMContextCreate();
70 ctx
->chip_class
= chip_class
;
72 ctx
->wave_size
= wave_size
;
73 ctx
->ballot_mask_bits
= ballot_mask_bits
;
74 ctx
->float_mode
= float_mode
;
75 ctx
->module
= ac_create_module(wave_size
== 32 ? compiler
->tm_wave32
78 ctx
->builder
= ac_create_builder(ctx
->context
, float_mode
);
80 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
81 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
82 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
83 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
84 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
85 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
86 ctx
->i128
= LLVMIntTypeInContext(ctx
->context
, 128);
87 ctx
->intptr
= ctx
->i32
;
88 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
89 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
90 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
91 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
92 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
93 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
94 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
95 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
96 ctx
->v3f32
= LLVMVectorType(ctx
->f32
, 3);
97 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
98 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
99 ctx
->iN_wavemask
= LLVMIntTypeInContext(ctx
->context
, ctx
->wave_size
);
100 ctx
->iN_ballotmask
= LLVMIntTypeInContext(ctx
->context
, ballot_mask_bits
);
102 ctx
->i8_0
= LLVMConstInt(ctx
->i8
, 0, false);
103 ctx
->i8_1
= LLVMConstInt(ctx
->i8
, 1, false);
104 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
105 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
106 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
107 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
108 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
109 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
110 ctx
->i128_0
= LLVMConstInt(ctx
->i128
, 0, false);
111 ctx
->i128_1
= LLVMConstInt(ctx
->i128
, 1, false);
112 ctx
->f16_0
= LLVMConstReal(ctx
->f16
, 0.0);
113 ctx
->f16_1
= LLVMConstReal(ctx
->f16
, 1.0);
114 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
115 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
116 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
117 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
119 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
120 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
122 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
125 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
126 "invariant.load", 14);
128 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
129 "amdgpu.uniform", 14);
131 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
132 ctx
->flow
= calloc(1, sizeof(*ctx
->flow
));
136 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
138 free(ctx
->flow
->stack
);
144 ac_get_llvm_num_components(LLVMValueRef value
)
146 LLVMTypeRef type
= LLVMTypeOf(value
);
147 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMFixedVectorTypeKind
148 ? LLVMGetVectorSize(type
)
150 return num_components
;
154 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
158 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMFixedVectorTypeKind
) {
163 return LLVMBuildExtractElement(ac
->builder
, value
,
164 LLVMConstInt(ac
->i32
, index
, false), "");
168 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
170 if (LLVMGetTypeKind(type
) == LLVMFixedVectorTypeKind
)
171 type
= LLVMGetElementType(type
);
173 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
174 return LLVMGetIntTypeWidth(type
);
176 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
177 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_LDS
)
181 if (type
== ctx
->f16
)
183 if (type
== ctx
->f32
)
185 if (type
== ctx
->f64
)
188 unreachable("Unhandled type kind in get_elem_bits");
192 ac_get_type_size(LLVMTypeRef type
)
194 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
197 case LLVMIntegerTypeKind
:
198 return LLVMGetIntTypeWidth(type
) / 8;
199 case LLVMHalfTypeKind
:
201 case LLVMFloatTypeKind
:
203 case LLVMDoubleTypeKind
:
205 case LLVMPointerTypeKind
:
206 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
209 case LLVMFixedVectorTypeKind
:
210 return LLVMGetVectorSize(type
) *
211 ac_get_type_size(LLVMGetElementType(type
));
212 case LLVMArrayTypeKind
:
213 return LLVMGetArrayLength(type
) *
214 ac_get_type_size(LLVMGetElementType(type
));
221 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
225 else if (t
== ctx
->f16
|| t
== ctx
->i16
)
227 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
229 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
232 unreachable("Unhandled integer size");
236 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
238 if (LLVMGetTypeKind(t
) == LLVMFixedVectorTypeKind
) {
239 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
240 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
241 LLVMGetVectorSize(t
));
243 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
244 switch (LLVMGetPointerAddressSpace(t
)) {
245 case AC_ADDR_SPACE_GLOBAL
:
247 case AC_ADDR_SPACE_CONST_32BIT
:
248 case AC_ADDR_SPACE_LDS
:
251 unreachable("unhandled address space");
254 return to_integer_type_scalar(ctx
, t
);
258 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
260 LLVMTypeRef type
= LLVMTypeOf(v
);
261 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
262 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
264 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
268 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
270 LLVMTypeRef type
= LLVMTypeOf(v
);
271 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
273 return ac_to_integer(ctx
, v
);
276 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
280 else if (t
== ctx
->i16
|| t
== ctx
->f16
)
282 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
284 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
287 unreachable("Unhandled float size");
291 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
293 if (LLVMGetTypeKind(t
) == LLVMFixedVectorTypeKind
) {
294 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
295 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
296 LLVMGetVectorSize(t
));
298 return to_float_type_scalar(ctx
, t
);
302 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
304 LLVMTypeRef type
= LLVMTypeOf(v
);
305 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
310 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
311 LLVMTypeRef return_type
, LLVMValueRef
*params
,
312 unsigned param_count
, unsigned attrib_mask
)
314 LLVMValueRef function
, call
;
315 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
317 function
= LLVMGetNamedFunction(ctx
->module
, name
);
319 LLVMTypeRef param_types
[32], function_type
;
322 assert(param_count
<= 32);
324 for (i
= 0; i
< param_count
; ++i
) {
326 param_types
[i
] = LLVMTypeOf(params
[i
]);
329 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
330 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
332 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
333 LLVMSetLinkage(function
, LLVMExternalLinkage
);
335 if (!set_callsite_attrs
)
336 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
339 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
340 if (set_callsite_attrs
)
341 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
346 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
349 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
351 LLVMTypeRef elem_type
= type
;
353 assert(bufsize
>= 8);
355 if (LLVMGetTypeKind(type
) == LLVMFixedVectorTypeKind
) {
356 int ret
= snprintf(buf
, bufsize
, "v%u",
357 LLVMGetVectorSize(type
));
359 char *type_name
= LLVMPrintTypeToString(type
);
360 fprintf(stderr
, "Error building type name for: %s\n",
362 LLVMDisposeMessage(type_name
);
365 elem_type
= LLVMGetElementType(type
);
369 switch (LLVMGetTypeKind(elem_type
)) {
371 case LLVMIntegerTypeKind
:
372 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
374 case LLVMHalfTypeKind
:
375 snprintf(buf
, bufsize
, "f16");
377 case LLVMFloatTypeKind
:
378 snprintf(buf
, bufsize
, "f32");
380 case LLVMDoubleTypeKind
:
381 snprintf(buf
, bufsize
, "f64");
387 * Helper function that builds an LLVM IR PHI node and immediately adds
391 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
392 unsigned count_incoming
, LLVMValueRef
*values
,
393 LLVMBasicBlockRef
*blocks
)
395 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
396 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
400 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
402 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
403 0, AC_FUNC_ATTR_CONVERGENT
);
406 /* Prevent optimizations (at least of memory accesses) across the current
407 * point in the program by emitting empty inline assembly that is marked as
408 * having side effects.
410 * Optionally, a value can be passed through the inline assembly to prevent
411 * LLVM from hoisting calls to ReadNone functions.
414 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
417 static int counter
= 0;
419 LLVMBuilderRef builder
= ctx
->builder
;
422 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
425 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
426 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
427 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
429 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
430 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
431 LLVMTypeRef type
= LLVMTypeOf(*pvgpr
);
432 unsigned bitsize
= ac_get_elem_bits(ctx
, type
);
433 LLVMValueRef vgpr
= *pvgpr
;
434 LLVMTypeRef vgpr_type
;
439 vgpr
= LLVMBuildZExt(ctx
->builder
, vgpr
, ctx
->i32
, "");
441 vgpr_type
= LLVMTypeOf(vgpr
);
442 vgpr_size
= ac_get_type_size(vgpr_type
);
444 assert(vgpr_size
% 4 == 0);
446 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
447 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
448 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
449 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
450 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
453 vgpr
= LLVMBuildTrunc(builder
, vgpr
, type
, "");
460 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
462 const char *intr
= LLVM_VERSION_MAJOR
>= 9 && ctx
->chip_class
>= GFX8
?
463 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
464 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, intr
, ctx
->i64
, NULL
, 0, 0);
465 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
469 ac_build_ballot(struct ac_llvm_context
*ctx
,
474 if (LLVM_VERSION_MAJOR
>= 9) {
475 if (ctx
->wave_size
== 64)
476 name
= "llvm.amdgcn.icmp.i64.i32";
478 name
= "llvm.amdgcn.icmp.i32.i32";
480 name
= "llvm.amdgcn.icmp.i32";
482 LLVMValueRef args
[3] = {
485 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
488 /* We currently have no other way to prevent LLVM from lifting the icmp
489 * calls to a dominating basic block.
491 ac_build_optimization_barrier(ctx
, &args
[0]);
493 args
[0] = ac_to_integer(ctx
, args
[0]);
495 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
496 AC_FUNC_ATTR_NOUNWIND
|
497 AC_FUNC_ATTR_READNONE
|
498 AC_FUNC_ATTR_CONVERGENT
);
501 LLVMValueRef
ac_get_i1_sgpr_mask(struct ac_llvm_context
*ctx
,
506 if (LLVM_VERSION_MAJOR
>= 9) {
507 if (ctx
->wave_size
== 64)
508 name
= "llvm.amdgcn.icmp.i64.i1";
510 name
= "llvm.amdgcn.icmp.i32.i1";
512 name
= "llvm.amdgcn.icmp.i1";
514 LLVMValueRef args
[3] = {
517 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0),
520 return ac_build_intrinsic(ctx
, name
, ctx
->iN_wavemask
, args
, 3,
521 AC_FUNC_ATTR_NOUNWIND
|
522 AC_FUNC_ATTR_READNONE
|
523 AC_FUNC_ATTR_CONVERGENT
);
527 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
529 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
530 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
531 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
535 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
537 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
538 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
539 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
543 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
545 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
546 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
548 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
549 vote_set
, active_set
, "");
550 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
552 LLVMConstInt(ctx
->iN_wavemask
, 0, 0), "");
553 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
557 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
558 unsigned value_count
, unsigned component
)
560 LLVMValueRef vec
= NULL
;
562 if (value_count
== 1) {
563 return values
[component
];
564 } else if (!value_count
)
565 unreachable("value_count is 0");
567 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
568 LLVMValueRef value
= values
[i
];
571 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
572 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
573 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
579 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
580 LLVMValueRef
*values
,
581 unsigned value_count
,
582 unsigned value_stride
,
586 LLVMBuilderRef builder
= ctx
->builder
;
587 LLVMValueRef vec
= NULL
;
590 if (value_count
== 1 && !always_vector
) {
592 return LLVMBuildLoad(builder
, values
[0], "");
594 } else if (!value_count
)
595 unreachable("value_count is 0");
597 for (i
= 0; i
< value_count
; i
++) {
598 LLVMValueRef value
= values
[i
* value_stride
];
600 value
= LLVMBuildLoad(builder
, value
, "");
603 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
604 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
605 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
611 ac_build_gather_values(struct ac_llvm_context
*ctx
,
612 LLVMValueRef
*values
,
613 unsigned value_count
)
615 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
618 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
619 * channels with undef. Extract at most src_channels components from the input.
622 ac_build_expand(struct ac_llvm_context
*ctx
,
624 unsigned src_channels
,
625 unsigned dst_channels
)
627 LLVMTypeRef elemtype
;
628 LLVMValueRef chan
[dst_channels
];
630 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFixedVectorTypeKind
) {
631 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
633 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
636 src_channels
= MIN2(src_channels
, vec_size
);
638 for (unsigned i
= 0; i
< src_channels
; i
++)
639 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
641 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
644 assert(src_channels
== 1);
647 elemtype
= LLVMTypeOf(value
);
650 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
651 chan
[i
] = LLVMGetUndef(elemtype
);
653 return ac_build_gather_values(ctx
, chan
, dst_channels
);
656 /* Extract components [start, start + channels) from a vector.
659 ac_extract_components(struct ac_llvm_context
*ctx
,
664 LLVMValueRef chan
[channels
];
666 for (unsigned i
= 0; i
< channels
; i
++)
667 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
+ start
);
669 return ac_build_gather_values(ctx
, chan
, channels
);
672 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
673 * with undef. Extract at most num_channels components from the input.
675 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
677 unsigned num_channels
)
679 return ac_build_expand(ctx
, value
, num_channels
, 4);
682 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
684 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
688 name
= "llvm.rint.f16";
689 else if (type_size
== 4)
690 name
= "llvm.rint.f32";
692 name
= "llvm.rint.f64";
694 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
695 AC_FUNC_ATTR_READNONE
);
699 ac_build_fdiv(struct ac_llvm_context
*ctx
,
703 unsigned type_size
= ac_get_type_size(LLVMTypeOf(den
));
707 name
= "llvm.amdgcn.rcp.f16";
708 else if (type_size
== 4)
709 name
= "llvm.amdgcn.rcp.f32";
711 name
= "llvm.amdgcn.rcp.f64";
713 LLVMValueRef rcp
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(den
),
714 &den
, 1, AC_FUNC_ATTR_READNONE
);
716 return LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
719 /* See fast_idiv_by_const.h. */
720 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
721 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
723 LLVMValueRef multiplier
,
724 LLVMValueRef pre_shift
,
725 LLVMValueRef post_shift
,
726 LLVMValueRef increment
)
728 LLVMBuilderRef builder
= ctx
->builder
;
730 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
731 num
= LLVMBuildMul(builder
,
732 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
733 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
734 num
= LLVMBuildAdd(builder
, num
,
735 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
736 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
737 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
738 return LLVMBuildLShr(builder
, num
, post_shift
, "");
741 /* See fast_idiv_by_const.h. */
742 /* If num != UINT_MAX, this more efficient version can be used. */
743 /* Set: increment = util_fast_udiv_info::increment; */
744 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
746 LLVMValueRef multiplier
,
747 LLVMValueRef pre_shift
,
748 LLVMValueRef post_shift
,
749 LLVMValueRef increment
)
751 LLVMBuilderRef builder
= ctx
->builder
;
753 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
754 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
755 num
= LLVMBuildMul(builder
,
756 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
757 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
758 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
759 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
760 return LLVMBuildLShr(builder
, num
, post_shift
, "");
763 /* See fast_idiv_by_const.h. */
764 /* Both operands must fit in 31 bits and the divisor must not be 1. */
765 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
767 LLVMValueRef multiplier
,
768 LLVMValueRef post_shift
)
770 LLVMBuilderRef builder
= ctx
->builder
;
772 num
= LLVMBuildMul(builder
,
773 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
774 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
775 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
776 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
777 return LLVMBuildLShr(builder
, num
, post_shift
, "");
780 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
781 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
782 * already multiplied by two. id is the cube face number.
784 struct cube_selection_coords
{
791 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
793 struct cube_selection_coords
*out
)
795 LLVMTypeRef f32
= ctx
->f32
;
797 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
798 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
799 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
800 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
801 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
802 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
803 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
804 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
808 * Build a manual selection sequence for cube face sc/tc coordinates and
809 * major axis vector (multiplied by 2 for consistency) for the given
810 * vec3 \p coords, for the face implied by \p selcoords.
812 * For the major axis, we always adjust the sign to be in the direction of
813 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
814 * the selcoords major axis.
816 static void build_cube_select(struct ac_llvm_context
*ctx
,
817 const struct cube_selection_coords
*selcoords
,
818 const LLVMValueRef
*coords
,
819 LLVMValueRef
*out_st
,
820 LLVMValueRef
*out_ma
)
822 LLVMBuilderRef builder
= ctx
->builder
;
823 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
824 LLVMValueRef is_ma_positive
;
826 LLVMValueRef is_ma_z
, is_not_ma_z
;
827 LLVMValueRef is_ma_y
;
828 LLVMValueRef is_ma_x
;
832 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
833 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
834 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
835 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
837 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
838 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
839 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
840 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
841 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
844 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
845 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
846 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
847 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
848 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
851 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
852 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
853 LLVMConstReal(f32
, -1.0), "");
854 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
857 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
858 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
859 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
860 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
861 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
865 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
866 bool is_deriv
, bool is_array
, bool is_lod
,
867 LLVMValueRef
*coords_arg
,
868 LLVMValueRef
*derivs_arg
)
871 LLVMBuilderRef builder
= ctx
->builder
;
872 struct cube_selection_coords selcoords
;
873 LLVMValueRef coords
[3];
876 if (is_array
&& !is_lod
) {
877 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
879 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
881 * "For Array forms, the array layer used will be
883 * max(0, min(d−1, floor(layer+0.5)))
885 * where d is the depth of the texture array and layer
886 * comes from the component indicated in the tables below.
887 * Workaroudn for an issue where the layer is taken from a
888 * helper invocation which happens to fall on a different
889 * layer due to extrapolation."
891 * GFX8 and earlier attempt to implement this in hardware by
892 * clamping the value of coords[2] = (8 * layer) + face.
893 * Unfortunately, this means that the we end up with the wrong
894 * face when clamping occurs.
896 * Clamp the layer earlier to work around the issue.
898 if (ctx
->chip_class
<= GFX8
) {
900 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
901 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
907 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
909 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
910 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
911 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
913 for (int i
= 0; i
< 2; ++i
)
914 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
916 coords
[2] = selcoords
.id
;
918 if (is_deriv
&& derivs_arg
) {
919 LLVMValueRef derivs
[4];
922 /* Convert cube derivatives to 2D derivatives. */
923 for (axis
= 0; axis
< 2; axis
++) {
924 LLVMValueRef deriv_st
[2];
925 LLVMValueRef deriv_ma
;
927 /* Transform the derivative alongside the texture
928 * coordinate. Mathematically, the correct formula is
929 * as follows. Assume we're projecting onto the +Z face
930 * and denote by dx/dh the derivative of the (original)
931 * X texture coordinate with respect to horizontal
932 * window coordinates. The projection onto the +Z face
937 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
938 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
940 * This motivatives the implementation below.
942 * Whether this actually gives the expected results for
943 * apps that might feed in derivatives obtained via
944 * finite differences is anyone's guess. The OpenGL spec
945 * seems awfully quiet about how textureGrad for cube
946 * maps should be handled.
948 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
949 deriv_st
, &deriv_ma
);
951 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
953 for (int i
= 0; i
< 2; ++i
)
954 derivs
[axis
* 2 + i
] =
955 LLVMBuildFSub(builder
,
956 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
957 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
960 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
963 /* Shift the texture coordinate. This must be applied after the
964 * derivative calculation.
966 for (int i
= 0; i
< 2; ++i
)
967 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
970 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
971 /* coords_arg.w component - array_index for cube arrays */
972 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
975 memcpy(coords_arg
, coords
, sizeof(coords
));
980 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
981 LLVMValueRef llvm_chan
,
982 LLVMValueRef attr_number
,
987 LLVMValueRef args
[5];
992 args
[2] = attr_number
;
995 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
996 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1000 args
[2] = llvm_chan
;
1001 args
[3] = attr_number
;
1004 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
1005 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1009 ac_build_fs_interp_f16(struct ac_llvm_context
*ctx
,
1010 LLVMValueRef llvm_chan
,
1011 LLVMValueRef attr_number
,
1012 LLVMValueRef params
,
1016 LLVMValueRef args
[6];
1020 args
[1] = llvm_chan
;
1021 args
[2] = attr_number
;
1022 args
[3] = ctx
->i1false
;
1025 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1.f16",
1026 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
1030 args
[2] = llvm_chan
;
1031 args
[3] = attr_number
;
1032 args
[4] = ctx
->i1false
;
1035 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2.f16",
1036 ctx
->f16
, args
, 6, AC_FUNC_ATTR_READNONE
);
1040 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
1041 LLVMValueRef parameter
,
1042 LLVMValueRef llvm_chan
,
1043 LLVMValueRef attr_number
,
1044 LLVMValueRef params
)
1046 LLVMValueRef args
[4];
1048 args
[0] = parameter
;
1049 args
[1] = llvm_chan
;
1050 args
[2] = attr_number
;
1053 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
1054 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
1058 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
1059 LLVMValueRef base_ptr
,
1062 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1066 ac_build_gep0(struct ac_llvm_context
*ctx
,
1067 LLVMValueRef base_ptr
,
1070 LLVMValueRef indices
[2] = {
1074 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1077 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
1080 return LLVMBuildPointerCast(ctx
->builder
,
1081 LLVMBuildGEP(ctx
->builder
, ptr
, &index
, 1, ""),
1082 LLVMTypeOf(ptr
), "");
1086 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
1087 LLVMValueRef base_ptr
, LLVMValueRef index
,
1090 LLVMBuildStore(ctx
->builder
, value
,
1091 ac_build_gep0(ctx
, base_ptr
, index
));
1095 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1096 * It's equivalent to doing a load from &base_ptr[index].
1098 * \param base_ptr Where the array starts.
1099 * \param index The element index into the array.
1100 * \param uniform Whether the base_ptr and index can be assumed to be
1101 * dynamically uniform (i.e. load to an SGPR)
1102 * \param invariant Whether the load is invariant (no other opcodes affect it)
1103 * \param no_unsigned_wraparound
1104 * For all possible re-associations and re-distributions of an expression
1105 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1106 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1107 * does not result in an unsigned integer wraparound. This is used for
1108 * optimal code generation of 32-bit pointer arithmetic.
1110 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1111 * integer wraparound can't be an imm offset in s_load_dword, because
1112 * the instruction performs "addr + offset" in 64 bits.
1114 * Expected usage for bindless textures by chaining GEPs:
1115 * // possible unsigned wraparound, don't use InBounds:
1116 * ptr1 = LLVMBuildGEP(base_ptr, index);
1117 * image = load(ptr1); // becomes "s_load ptr1, 0"
1119 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1120 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1123 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1124 LLVMValueRef index
, bool uniform
, bool invariant
,
1125 bool no_unsigned_wraparound
)
1127 LLVMValueRef pointer
, result
;
1129 if (no_unsigned_wraparound
&&
1130 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1131 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1133 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
1136 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1137 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1139 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1143 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1146 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1149 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1150 LLVMValueRef base_ptr
, LLVMValueRef index
)
1152 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1155 /* This assumes that there is no unsigned integer wraparound during the address
1156 * computation, excluding all GEPs within base_ptr. */
1157 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1158 LLVMValueRef base_ptr
, LLVMValueRef index
)
1160 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1163 /* See ac_build_load_custom() documentation. */
1164 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1165 LLVMValueRef base_ptr
, LLVMValueRef index
)
1167 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1170 static unsigned get_load_cache_policy(struct ac_llvm_context
*ctx
,
1171 unsigned cache_policy
)
1173 return cache_policy
|
1174 (ctx
->chip_class
>= GFX10
&& cache_policy
& ac_glc
? ac_dlc
: 0);
1178 ac_build_buffer_store_common(struct ac_llvm_context
*ctx
,
1181 LLVMValueRef vindex
,
1182 LLVMValueRef voffset
,
1183 LLVMValueRef soffset
,
1184 unsigned num_channels
,
1185 LLVMTypeRef return_channel_type
,
1186 unsigned cache_policy
,
1190 LLVMValueRef args
[6];
1193 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1195 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1196 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1197 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1198 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1199 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1200 const char *indexing_kind
= structurized
? "struct" : "raw";
1201 char name
[256], type_name
[8];
1203 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(return_channel_type
, func
) : return_channel_type
;
1204 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1207 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.format.%s",
1208 indexing_kind
, type_name
);
1210 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.store.%s",
1211 indexing_kind
, type_name
);
1214 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1215 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1219 ac_build_buffer_store_format(struct ac_llvm_context
*ctx
,
1222 LLVMValueRef vindex
,
1223 LLVMValueRef voffset
,
1224 unsigned num_channels
,
1225 unsigned cache_policy
)
1227 ac_build_buffer_store_common(ctx
, rsrc
, data
, vindex
,
1228 voffset
, NULL
, num_channels
,
1229 ctx
->f32
, cache_policy
,
1233 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1234 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1235 * or v4i32 (num_channels=3,4).
1238 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1241 unsigned num_channels
,
1242 LLVMValueRef voffset
,
1243 LLVMValueRef soffset
,
1244 unsigned inst_offset
,
1245 unsigned cache_policy
)
1247 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1249 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false)) {
1250 LLVMValueRef v
[3], v01
;
1252 for (int i
= 0; i
< 3; i
++) {
1253 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1254 LLVMConstInt(ctx
->i32
, i
, 0), "");
1256 v01
= ac_build_gather_values(ctx
, v
, 2);
1258 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1259 soffset
, inst_offset
, cache_policy
);
1260 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1261 soffset
, inst_offset
+ 8,
1266 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1267 * (voffset is swizzled, but soffset isn't swizzled).
1268 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1270 if (!(cache_policy
& ac_swizzled
)) {
1271 LLVMValueRef offset
= soffset
;
1274 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1275 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1277 ac_build_buffer_store_common(ctx
, rsrc
, ac_to_float(ctx
, vdata
),
1278 ctx
->i32_0
, voffset
, offset
,
1279 num_channels
, ctx
->f32
,
1280 cache_policy
, false, false);
1284 static const unsigned dfmts
[] = {
1285 V_008F0C_BUF_DATA_FORMAT_32
,
1286 V_008F0C_BUF_DATA_FORMAT_32_32
,
1287 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1288 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1290 unsigned dfmt
= dfmts
[num_channels
- 1];
1291 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1292 LLVMValueRef immoffset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1294 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1295 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
);
1299 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1301 LLVMValueRef vindex
,
1302 LLVMValueRef voffset
,
1303 LLVMValueRef soffset
,
1304 unsigned num_channels
,
1305 LLVMTypeRef channel_type
,
1306 unsigned cache_policy
,
1311 LLVMValueRef args
[5];
1313 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1315 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1316 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1317 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1318 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1319 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, use_format
) && num_channels
== 3 ? 4 : num_channels
;
1320 const char *indexing_kind
= structurized
? "struct" : "raw";
1321 char name
[256], type_name
[8];
1323 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(channel_type
, func
) : channel_type
;
1324 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1327 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1328 indexing_kind
, type_name
);
1330 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1331 indexing_kind
, type_name
);
1334 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1335 ac_get_load_intr_attribs(can_speculate
));
1339 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1342 LLVMValueRef vindex
,
1343 LLVMValueRef voffset
,
1344 LLVMValueRef soffset
,
1345 unsigned inst_offset
,
1346 unsigned cache_policy
,
1350 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1352 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1354 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1356 if (allow_smem
&& !(cache_policy
& ac_slc
) &&
1357 (!(cache_policy
& ac_glc
) || ctx
->chip_class
>= GFX8
)) {
1358 assert(vindex
== NULL
);
1360 LLVMValueRef result
[8];
1362 for (int i
= 0; i
< num_channels
; i
++) {
1364 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1365 LLVMConstInt(ctx
->i32
, 4, 0), "");
1367 LLVMValueRef args
[3] = {
1370 LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0),
1372 result
[i
] = ac_build_intrinsic(ctx
,
1373 "llvm.amdgcn.s.buffer.load.f32",
1375 AC_FUNC_ATTR_READNONE
);
1377 if (num_channels
== 1)
1380 if (num_channels
== 3 && !ac_has_vec3_support(ctx
->chip_class
, false))
1381 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1382 return ac_build_gather_values(ctx
, result
, num_channels
);
1385 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
,
1387 num_channels
, ctx
->f32
,
1389 can_speculate
, false, false);
1392 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1394 LLVMValueRef vindex
,
1395 LLVMValueRef voffset
,
1396 unsigned num_channels
,
1397 unsigned cache_policy
,
1400 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1401 ctx
->i32_0
, num_channels
, ctx
->f32
,
1402 cache_policy
, can_speculate
,
1407 ac_build_tbuffer_load(struct ac_llvm_context
*ctx
,
1409 LLVMValueRef vindex
,
1410 LLVMValueRef voffset
,
1411 LLVMValueRef soffset
,
1412 LLVMValueRef immoffset
,
1413 unsigned num_channels
,
1416 unsigned cache_policy
,
1420 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1422 LLVMValueRef args
[6];
1424 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1426 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1427 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1428 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1429 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1430 args
[idx
++] = LLVMConstInt(ctx
->i32
, get_load_cache_policy(ctx
, cache_policy
), 0);
1431 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1432 const char *indexing_kind
= structurized
? "struct" : "raw";
1433 char name
[256], type_name
[8];
1435 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1436 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1438 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.load.%s",
1439 indexing_kind
, type_name
);
1441 return ac_build_intrinsic(ctx
, name
, type
, args
, idx
,
1442 ac_get_load_intr_attribs(can_speculate
));
1446 ac_build_struct_tbuffer_load(struct ac_llvm_context
*ctx
,
1448 LLVMValueRef vindex
,
1449 LLVMValueRef voffset
,
1450 LLVMValueRef soffset
,
1451 LLVMValueRef immoffset
,
1452 unsigned num_channels
,
1455 unsigned cache_policy
,
1458 return ac_build_tbuffer_load(ctx
, rsrc
, vindex
, voffset
, soffset
,
1459 immoffset
, num_channels
, dfmt
, nfmt
,
1460 cache_policy
, can_speculate
, true);
1464 ac_build_raw_tbuffer_load(struct ac_llvm_context
*ctx
,
1466 LLVMValueRef voffset
,
1467 LLVMValueRef soffset
,
1468 LLVMValueRef immoffset
,
1469 unsigned num_channels
,
1472 unsigned cache_policy
,
1475 return ac_build_tbuffer_load(ctx
, rsrc
, NULL
, voffset
, soffset
,
1476 immoffset
, num_channels
, dfmt
, nfmt
,
1477 cache_policy
, can_speculate
, false);
1481 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1483 LLVMValueRef voffset
,
1484 LLVMValueRef soffset
,
1485 LLVMValueRef immoffset
,
1486 unsigned cache_policy
)
1490 if (LLVM_VERSION_MAJOR
>= 9) {
1491 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1493 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1494 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1496 1, ctx
->i16
, cache_policy
,
1497 false, false, false);
1499 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1500 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1502 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1503 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1506 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1513 ac_build_tbuffer_load_byte(struct ac_llvm_context
*ctx
,
1515 LLVMValueRef voffset
,
1516 LLVMValueRef soffset
,
1517 LLVMValueRef immoffset
,
1518 unsigned cache_policy
)
1522 if (LLVM_VERSION_MAJOR
>= 9) {
1523 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
, immoffset
, "");
1525 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1526 res
= ac_build_buffer_load_common(ctx
, rsrc
, NULL
,
1528 1, ctx
->i8
, cache_policy
,
1529 false, false, false);
1531 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1532 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1534 res
= ac_build_raw_tbuffer_load(ctx
, rsrc
, voffset
, soffset
,
1535 immoffset
, 1, dfmt
, nfmt
, cache_policy
,
1538 res
= LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i8
, "");
1545 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1547 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1548 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1551 ac_ufN_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned exp_bits
, unsigned mant_bits
)
1553 assert(LLVMTypeOf(src
) == ctx
->i32
);
1556 LLVMValueRef mantissa
;
1557 mantissa
= LLVMBuildAnd(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, (1 << mant_bits
) - 1, false), "");
1559 /* Converting normal numbers is just a shift + correcting the exponent bias */
1560 unsigned normal_shift
= 23 - mant_bits
;
1561 unsigned bias_shift
= 127 - ((1 << (exp_bits
- 1)) - 1);
1562 LLVMValueRef shifted
, normal
;
1564 shifted
= LLVMBuildShl(ctx
->builder
, src
, LLVMConstInt(ctx
->i32
, normal_shift
, false), "");
1565 normal
= LLVMBuildAdd(ctx
->builder
, shifted
, LLVMConstInt(ctx
->i32
, bias_shift
<< 23, false), "");
1567 /* Converting nan/inf numbers is the same, but with a different exponent update */
1568 LLVMValueRef naninf
;
1569 naninf
= LLVMBuildOr(ctx
->builder
, normal
, LLVMConstInt(ctx
->i32
, 0xff << 23, false), "");
1571 /* Converting denormals is the complex case: determine the leading zeros of the
1572 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1574 LLVMValueRef denormal
;
1575 LLVMValueRef params
[2] = {
1577 ctx
->i1true
, /* result can be undef when arg is 0 */
1579 LLVMValueRef ctlz
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32", ctx
->i32
,
1580 params
, 2, AC_FUNC_ATTR_READNONE
);
1582 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1583 tmp
= LLVMBuildSub(ctx
->builder
, ctlz
, LLVMConstInt(ctx
->i32
, 8, false), "");
1584 denormal
= LLVMBuildShl(ctx
->builder
, mantissa
, tmp
, "");
1586 unsigned denormal_exp
= bias_shift
+ (32 - mant_bits
) - 1;
1587 tmp
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, denormal_exp
, false), ctlz
, "");
1588 tmp
= LLVMBuildShl(ctx
->builder
, tmp
, LLVMConstInt(ctx
->i32
, 23, false), "");
1589 denormal
= LLVMBuildAdd(ctx
->builder
, denormal
, tmp
, "");
1591 /* Select the final result. */
1592 LLVMValueRef result
;
1594 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1595 LLVMConstInt(ctx
->i32
, ((1 << exp_bits
) - 1) << mant_bits
, false), "");
1596 result
= LLVMBuildSelect(ctx
->builder
, tmp
, naninf
, normal
, "");
1598 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, src
,
1599 LLVMConstInt(ctx
->i32
, 1 << mant_bits
, false), "");
1600 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, denormal
, "");
1602 tmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, src
, ctx
->i32_0
, "");
1603 result
= LLVMBuildSelect(ctx
->builder
, tmp
, result
, ctx
->i32_0
, "");
1605 return ac_to_float(ctx
, result
);
1609 * Generate a fully general open coded buffer format fetch with all required
1610 * fixups suitable for vertex fetch, using non-format buffer loads.
1612 * Some combinations of argument values have special interpretations:
1613 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1614 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1616 * \param log_size log(size of channel in bytes)
1617 * \param num_channels number of channels (1 to 4)
1618 * \param format AC_FETCH_FORMAT_xxx value
1619 * \param reverse whether XYZ channels are reversed
1620 * \param known_aligned whether the source is known to be aligned to hardware's
1621 * effective element size for loading the given format
1622 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1623 * \param rsrc buffer resource descriptor
1624 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1627 ac_build_opencoded_load_format(struct ac_llvm_context
*ctx
,
1629 unsigned num_channels
,
1634 LLVMValueRef vindex
,
1635 LLVMValueRef voffset
,
1636 LLVMValueRef soffset
,
1637 unsigned cache_policy
,
1641 unsigned load_log_size
= log_size
;
1642 unsigned load_num_channels
= num_channels
;
1643 if (log_size
== 3) {
1645 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1646 load_num_channels
= 2 * num_channels
;
1648 load_num_channels
= 1; /* 10_11_11 or 2_10_10_10 */
1652 int log_recombine
= 0;
1653 if (ctx
->chip_class
== GFX6
&& !known_aligned
) {
1654 /* Avoid alignment restrictions by loading one byte at a time. */
1655 load_num_channels
<<= load_log_size
;
1656 log_recombine
= load_log_size
;
1658 } else if (load_num_channels
== 2 || load_num_channels
== 4) {
1659 log_recombine
= -util_logbase2(load_num_channels
);
1660 load_num_channels
= 1;
1661 load_log_size
+= -log_recombine
;
1664 assert(load_log_size
>= 2 || LLVM_VERSION_MAJOR
>= 9);
1666 LLVMValueRef loads
[32]; /* up to 32 bytes */
1667 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1668 tmp
= LLVMBuildAdd(ctx
->builder
, soffset
,
1669 LLVMConstInt(ctx
->i32
, i
<< load_log_size
, false), "");
1670 LLVMTypeRef channel_type
= load_log_size
== 0 ? ctx
->i8
:
1671 load_log_size
== 1 ? ctx
->i16
: ctx
->i32
;
1672 unsigned num_channels
= 1 << (MAX2(load_log_size
, 2) - 2);
1673 loads
[i
] = ac_build_buffer_load_common(
1674 ctx
, rsrc
, vindex
, voffset
, tmp
,
1675 num_channels
, channel_type
, cache_policy
,
1676 can_speculate
, false, true);
1677 if (load_log_size
>= 2)
1678 loads
[i
] = ac_to_integer(ctx
, loads
[i
]);
1681 if (log_recombine
> 0) {
1682 /* Recombine bytes if necessary (GFX6 only) */
1683 LLVMTypeRef dst_type
= log_recombine
== 2 ? ctx
->i32
: ctx
->i16
;
1685 for (unsigned src
= 0, dst
= 0; src
< load_num_channels
; ++dst
) {
1686 LLVMValueRef accum
= NULL
;
1687 for (unsigned i
= 0; i
< (1 << log_recombine
); ++i
, ++src
) {
1688 tmp
= LLVMBuildZExt(ctx
->builder
, loads
[src
], dst_type
, "");
1692 tmp
= LLVMBuildShl(ctx
->builder
, tmp
,
1693 LLVMConstInt(dst_type
, 8 * i
, false), "");
1694 accum
= LLVMBuildOr(ctx
->builder
, accum
, tmp
, "");
1699 } else if (log_recombine
< 0) {
1700 /* Split vectors of dwords */
1701 if (load_log_size
> 2) {
1702 assert(load_num_channels
== 1);
1703 LLVMValueRef loaded
= loads
[0];
1704 unsigned log_split
= load_log_size
- 2;
1705 log_recombine
+= log_split
;
1706 load_num_channels
= 1 << log_split
;
1708 for (unsigned i
= 0; i
< load_num_channels
; ++i
) {
1709 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
1710 loads
[i
] = LLVMBuildExtractElement(ctx
->builder
, loaded
, tmp
, "");
1714 /* Further split dwords and shorts if required */
1715 if (log_recombine
< 0) {
1716 for (unsigned src
= load_num_channels
,
1717 dst
= load_num_channels
<< -log_recombine
;
1719 unsigned dst_bits
= 1 << (3 + load_log_size
+ log_recombine
);
1720 LLVMTypeRef dst_type
= LLVMIntTypeInContext(ctx
->context
, dst_bits
);
1721 LLVMValueRef loaded
= loads
[src
- 1];
1722 LLVMTypeRef loaded_type
= LLVMTypeOf(loaded
);
1723 for (unsigned i
= 1 << -log_recombine
; i
> 0; --i
, --dst
) {
1724 tmp
= LLVMConstInt(loaded_type
, dst_bits
* (i
- 1), false);
1725 tmp
= LLVMBuildLShr(ctx
->builder
, loaded
, tmp
, "");
1726 loads
[dst
- 1] = LLVMBuildTrunc(ctx
->builder
, tmp
, dst_type
, "");
1732 if (log_size
== 3) {
1733 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1734 for (unsigned i
= 0; i
< num_channels
; ++i
) {
1735 tmp
= ac_build_gather_values(ctx
, &loads
[2 * i
], 2);
1736 loads
[i
] = LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->f64
, "");
1738 } else if (format
== AC_FETCH_FORMAT_FIXED
) {
1739 /* 10_11_11_FLOAT */
1740 LLVMValueRef data
= loads
[0];
1741 LLVMValueRef i32_2047
= LLVMConstInt(ctx
->i32
, 2047, false);
1742 LLVMValueRef r
= LLVMBuildAnd(ctx
->builder
, data
, i32_2047
, "");
1743 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 11, false), "");
1744 LLVMValueRef g
= LLVMBuildAnd(ctx
->builder
, tmp
, i32_2047
, "");
1745 LLVMValueRef b
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 22, false), "");
1747 loads
[0] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, r
, 5, 6));
1748 loads
[1] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, g
, 5, 6));
1749 loads
[2] = ac_to_integer(ctx
, ac_ufN_to_float(ctx
, b
, 5, 5));
1753 format
= AC_FETCH_FORMAT_FLOAT
;
1755 /* 2_10_10_10 data formats */
1756 LLVMValueRef data
= loads
[0];
1757 LLVMTypeRef i10
= LLVMIntTypeInContext(ctx
->context
, 10);
1758 LLVMTypeRef i2
= LLVMIntTypeInContext(ctx
->context
, 2);
1759 loads
[0] = LLVMBuildTrunc(ctx
->builder
, data
, i10
, "");
1760 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 10, false), "");
1761 loads
[1] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1762 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 20, false), "");
1763 loads
[2] = LLVMBuildTrunc(ctx
->builder
, tmp
, i10
, "");
1764 tmp
= LLVMBuildLShr(ctx
->builder
, data
, LLVMConstInt(ctx
->i32
, 30, false), "");
1765 loads
[3] = LLVMBuildTrunc(ctx
->builder
, tmp
, i2
, "");
1771 if (format
== AC_FETCH_FORMAT_FLOAT
) {
1772 if (log_size
!= 2) {
1773 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1774 tmp
= ac_to_float(ctx
, loads
[chan
]);
1776 tmp
= LLVMBuildFPTrunc(ctx
->builder
, tmp
, ctx
->f32
, "");
1777 else if (log_size
== 1)
1778 tmp
= LLVMBuildFPExt(ctx
->builder
, tmp
, ctx
->f32
, "");
1779 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1782 } else if (format
== AC_FETCH_FORMAT_UINT
) {
1783 if (log_size
!= 2) {
1784 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1785 loads
[chan
] = LLVMBuildZExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1787 } else if (format
== AC_FETCH_FORMAT_SINT
) {
1788 if (log_size
!= 2) {
1789 for (unsigned chan
= 0; chan
< num_channels
; ++chan
)
1790 loads
[chan
] = LLVMBuildSExt(ctx
->builder
, loads
[chan
], ctx
->i32
, "");
1793 bool unsign
= format
== AC_FETCH_FORMAT_UNORM
||
1794 format
== AC_FETCH_FORMAT_USCALED
||
1795 format
== AC_FETCH_FORMAT_UINT
;
1797 for (unsigned chan
= 0; chan
< num_channels
; ++chan
) {
1799 tmp
= LLVMBuildUIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1801 tmp
= LLVMBuildSIToFP(ctx
->builder
, loads
[chan
], ctx
->f32
, "");
1804 LLVMValueRef scale
= NULL
;
1805 if (format
== AC_FETCH_FORMAT_FIXED
) {
1806 assert(log_size
== 2);
1807 scale
= LLVMConstReal(ctx
->f32
, 1.0 / 0x10000);
1808 } else if (format
== AC_FETCH_FORMAT_UNORM
) {
1809 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1810 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << bits
) - 1));
1811 } else if (format
== AC_FETCH_FORMAT_SNORM
) {
1812 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(loads
[chan
]));
1813 scale
= LLVMConstReal(ctx
->f32
, 1.0 / (((uint64_t)1 << (bits
- 1)) - 1));
1816 tmp
= LLVMBuildFMul(ctx
->builder
, tmp
, scale
, "");
1818 if (format
== AC_FETCH_FORMAT_SNORM
) {
1819 /* Clamp to [-1, 1] */
1820 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
1821 LLVMValueRef clamp
=
1822 LLVMBuildFCmp(ctx
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
1823 tmp
= LLVMBuildSelect(ctx
->builder
, clamp
, neg_one
, tmp
, "");
1826 loads
[chan
] = ac_to_integer(ctx
, tmp
);
1830 while (num_channels
< 4) {
1831 if (format
== AC_FETCH_FORMAT_UINT
|| format
== AC_FETCH_FORMAT_SINT
) {
1832 loads
[num_channels
] = num_channels
== 3 ? ctx
->i32_1
: ctx
->i32_0
;
1834 loads
[num_channels
] = ac_to_integer(ctx
, num_channels
== 3 ? ctx
->f32_1
: ctx
->f32_0
);
1841 loads
[0] = loads
[2];
1845 return ac_build_gather_values(ctx
, loads
, 4);
1849 ac_build_tbuffer_store(struct ac_llvm_context
*ctx
,
1852 LLVMValueRef vindex
,
1853 LLVMValueRef voffset
,
1854 LLVMValueRef soffset
,
1855 LLVMValueRef immoffset
,
1856 unsigned num_channels
,
1859 unsigned cache_policy
,
1862 voffset
= LLVMBuildAdd(ctx
->builder
, voffset
? voffset
: ctx
->i32_0
,
1865 LLVMValueRef args
[7];
1867 args
[idx
++] = vdata
;
1868 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1870 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1871 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1872 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1873 args
[idx
++] = LLVMConstInt(ctx
->i32
, ac_get_tbuffer_format(ctx
->chip_class
, dfmt
, nfmt
), 0);
1874 args
[idx
++] = LLVMConstInt(ctx
->i32
, cache_policy
, 0);
1875 unsigned func
= !ac_has_vec3_support(ctx
->chip_class
, true) && num_channels
== 3 ? 4 : num_channels
;
1876 const char *indexing_kind
= structurized
? "struct" : "raw";
1877 char name
[256], type_name
[8];
1879 LLVMTypeRef type
= func
> 1 ? LLVMVectorType(ctx
->i32
, func
) : ctx
->i32
;
1880 ac_build_type_name_for_intr(type
, type_name
, sizeof(type_name
));
1882 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.tbuffer.store.%s",
1883 indexing_kind
, type_name
);
1885 ac_build_intrinsic(ctx
, name
, ctx
->voidt
, args
, idx
,
1886 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
);
1890 ac_build_struct_tbuffer_store(struct ac_llvm_context
*ctx
,
1893 LLVMValueRef vindex
,
1894 LLVMValueRef voffset
,
1895 LLVMValueRef soffset
,
1896 LLVMValueRef immoffset
,
1897 unsigned num_channels
,
1900 unsigned cache_policy
)
1902 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, vindex
, voffset
, soffset
,
1903 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1908 ac_build_raw_tbuffer_store(struct ac_llvm_context
*ctx
,
1911 LLVMValueRef voffset
,
1912 LLVMValueRef soffset
,
1913 LLVMValueRef immoffset
,
1914 unsigned num_channels
,
1917 unsigned cache_policy
)
1919 ac_build_tbuffer_store(ctx
, rsrc
, vdata
, NULL
, voffset
, soffset
,
1920 immoffset
, num_channels
, dfmt
, nfmt
, cache_policy
,
1925 ac_build_tbuffer_store_short(struct ac_llvm_context
*ctx
,
1928 LLVMValueRef voffset
,
1929 LLVMValueRef soffset
,
1930 unsigned cache_policy
)
1932 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i16
, "");
1934 if (LLVM_VERSION_MAJOR
>= 9) {
1935 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1936 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1937 voffset
, soffset
, 1,
1938 ctx
->i16
, cache_policy
,
1941 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_16
;
1942 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1944 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1946 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1947 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1952 ac_build_tbuffer_store_byte(struct ac_llvm_context
*ctx
,
1955 LLVMValueRef voffset
,
1956 LLVMValueRef soffset
,
1957 unsigned cache_policy
)
1959 vdata
= LLVMBuildBitCast(ctx
->builder
, vdata
, ctx
->i8
, "");
1961 if (LLVM_VERSION_MAJOR
>= 9) {
1962 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1963 ac_build_buffer_store_common(ctx
, rsrc
, vdata
, NULL
,
1964 voffset
, soffset
, 1,
1965 ctx
->i8
, cache_policy
,
1968 unsigned dfmt
= V_008F0C_BUF_DATA_FORMAT_8
;
1969 unsigned nfmt
= V_008F0C_BUF_NUM_FORMAT_UINT
;
1971 vdata
= LLVMBuildZExt(ctx
->builder
, vdata
, ctx
->i32
, "");
1973 ac_build_raw_tbuffer_store(ctx
, rsrc
, vdata
, voffset
, soffset
,
1974 ctx
->i32_0
, 1, dfmt
, nfmt
, cache_policy
);
1978 * Set range metadata on an instruction. This can only be used on load and
1979 * call instructions. If you know an instruction can only produce the values
1980 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1981 * \p lo is the minimum value inclusive.
1982 * \p hi is the maximum value exclusive.
1984 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1985 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1987 LLVMValueRef range_md
, md_args
[2];
1988 LLVMTypeRef type
= LLVMTypeOf(value
);
1989 LLVMContextRef context
= LLVMGetTypeContext(type
);
1991 md_args
[0] = LLVMConstInt(type
, lo
, false);
1992 md_args
[1] = LLVMConstInt(type
, hi
, false);
1993 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1994 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1998 ac_get_thread_id(struct ac_llvm_context
*ctx
)
2002 LLVMValueRef tid_args
[2];
2003 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
2004 tid_args
[1] = ctx
->i32_0
;
2005 tid_args
[1] = ac_build_intrinsic(ctx
,
2006 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2007 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
2009 if (ctx
->wave_size
== 32) {
2012 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
2014 2, AC_FUNC_ATTR_READNONE
);
2016 set_range_metadata(ctx
, tid
, 0, ctx
->wave_size
);
2021 * AMD GCN implements derivatives using the local data store (LDS)
2022 * All writes to the LDS happen in all executing threads at
2023 * the same time. TID is the Thread ID for the current
2024 * thread and is a value between 0 and 63, representing
2025 * the thread's position in the wavefront.
2027 * For the pixel shader threads are grouped into quads of four pixels.
2028 * The TIDs of the pixels of a quad are:
2036 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2037 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2038 * the current pixel's column, and masking with 0xfffffffe yields the TID
2039 * of the left pixel of the current pixel's row.
2041 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2042 * adding 2 yields the TID of the pixel below the top pixel.
2045 ac_build_ddxy(struct ac_llvm_context
*ctx
,
2050 unsigned tl_lanes
[4], trbl_lanes
[4];
2051 char name
[32], type
[8];
2052 LLVMValueRef tl
, trbl
;
2053 LLVMTypeRef result_type
;
2054 LLVMValueRef result
;
2056 result_type
= ac_to_float_type(ctx
, LLVMTypeOf(val
));
2058 if (result_type
== ctx
->f16
)
2059 val
= LLVMBuildZExt(ctx
->builder
, val
, ctx
->i32
, "");
2061 for (unsigned i
= 0; i
< 4; ++i
) {
2062 tl_lanes
[i
] = i
& mask
;
2063 trbl_lanes
[i
] = (i
& mask
) + idx
;
2066 tl
= ac_build_quad_swizzle(ctx
, val
,
2067 tl_lanes
[0], tl_lanes
[1],
2068 tl_lanes
[2], tl_lanes
[3]);
2069 trbl
= ac_build_quad_swizzle(ctx
, val
,
2070 trbl_lanes
[0], trbl_lanes
[1],
2071 trbl_lanes
[2], trbl_lanes
[3]);
2073 if (result_type
== ctx
->f16
) {
2074 tl
= LLVMBuildTrunc(ctx
->builder
, tl
, ctx
->i16
, "");
2075 trbl
= LLVMBuildTrunc(ctx
->builder
, trbl
, ctx
->i16
, "");
2078 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, result_type
, "");
2079 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, result_type
, "");
2080 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
2082 ac_build_type_name_for_intr(result_type
, type
, sizeof(type
));
2083 snprintf(name
, sizeof(name
), "llvm.amdgcn.wqm.%s", type
);
2085 return ac_build_intrinsic(ctx
, name
, result_type
, &result
, 1, 0);
2089 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
2091 LLVMValueRef wave_id
)
2093 LLVMValueRef args
[2];
2094 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
2096 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
2100 ac_build_imsb(struct ac_llvm_context
*ctx
,
2102 LLVMTypeRef dst_type
)
2104 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
2106 AC_FUNC_ATTR_READNONE
);
2108 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2109 * the index from LSB. Invert it by doing "31 - msb". */
2110 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
2113 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
2114 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
2115 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2116 arg
, ctx
->i32_0
, ""),
2117 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
2118 arg
, all_ones
, ""), "");
2120 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
2124 ac_build_umsb(struct ac_llvm_context
*ctx
,
2126 LLVMTypeRef dst_type
)
2128 const char *intrin_name
;
2130 LLVMValueRef highest_bit
;
2134 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
2137 intrin_name
= "llvm.ctlz.i64";
2139 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
2143 intrin_name
= "llvm.ctlz.i32";
2145 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
2149 intrin_name
= "llvm.ctlz.i16";
2151 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
2155 intrin_name
= "llvm.ctlz.i8";
2157 highest_bit
= LLVMConstInt(ctx
->i8
, 7, false);
2161 unreachable(!"invalid bitsize");
2165 LLVMValueRef params
[2] = {
2170 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2172 AC_FUNC_ATTR_READNONE
);
2174 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2175 * the index from LSB. Invert it by doing "31 - msb". */
2176 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
2178 if (bitsize
== 64) {
2179 msb
= LLVMBuildTrunc(ctx
->builder
, msb
, ctx
->i32
, "");
2180 } else if (bitsize
< 32) {
2181 msb
= LLVMBuildSExt(ctx
->builder
, msb
, ctx
->i32
, "");
2184 /* check for zero */
2185 return LLVMBuildSelect(ctx
->builder
,
2186 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
2187 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
2190 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2194 snprintf(name
, sizeof(name
), "llvm.minnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2195 LLVMValueRef args
[2] = {a
, b
};
2196 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2197 AC_FUNC_ATTR_READNONE
);
2200 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2204 snprintf(name
, sizeof(name
), "llvm.maxnum.f%d", ac_get_elem_bits(ctx
, LLVMTypeOf(a
)));
2205 LLVMValueRef args
[2] = {a
, b
};
2206 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(a
), args
, 2,
2207 AC_FUNC_ATTR_READNONE
);
2210 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2213 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
2214 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2217 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2220 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
2221 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2224 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2227 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
2228 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2231 LLVMValueRef
ac_build_umax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
2234 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, a
, b
, "");
2235 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
2238 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
2240 LLVMTypeRef t
= LLVMTypeOf(value
);
2241 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, LLVMConstReal(t
, 0.0)),
2242 LLVMConstReal(t
, 1.0));
2245 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
2247 LLVMValueRef args
[9];
2249 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
2250 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
2253 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
2254 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
2256 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
2258 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
2260 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2261 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2263 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
2264 ctx
->voidt
, args
, 6, 0);
2266 args
[2] = a
->out
[0];
2267 args
[3] = a
->out
[1];
2268 args
[4] = a
->out
[2];
2269 args
[5] = a
->out
[3];
2270 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
2271 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
2273 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
2274 ctx
->voidt
, args
, 8, 0);
2278 void ac_build_export_null(struct ac_llvm_context
*ctx
)
2280 struct ac_export_args args
;
2282 args
.enabled_channels
= 0x0; /* enabled channels */
2283 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2284 args
.done
= 1; /* DONE bit */
2285 args
.target
= V_008DFC_SQ_EXP_NULL
;
2286 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2287 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
2288 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
2289 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
2290 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
2292 ac_build_export(ctx
, &args
);
2295 static unsigned ac_num_coords(enum ac_image_dim dim
)
2301 case ac_image_1darray
:
2305 case ac_image_2darray
:
2306 case ac_image_2dmsaa
:
2308 case ac_image_2darraymsaa
:
2311 unreachable("ac_num_coords: bad dim");
2315 static unsigned ac_num_derivs(enum ac_image_dim dim
)
2319 case ac_image_1darray
:
2322 case ac_image_2darray
:
2327 case ac_image_2dmsaa
:
2328 case ac_image_2darraymsaa
:
2330 unreachable("derivatives not supported");
2334 static const char *get_atomic_name(enum ac_atomic_op op
)
2337 case ac_atomic_swap
: return "swap";
2338 case ac_atomic_add
: return "add";
2339 case ac_atomic_sub
: return "sub";
2340 case ac_atomic_smin
: return "smin";
2341 case ac_atomic_umin
: return "umin";
2342 case ac_atomic_smax
: return "smax";
2343 case ac_atomic_umax
: return "umax";
2344 case ac_atomic_and
: return "and";
2345 case ac_atomic_or
: return "or";
2346 case ac_atomic_xor
: return "xor";
2347 case ac_atomic_inc_wrap
: return "inc";
2348 case ac_atomic_dec_wrap
: return "dec";
2350 unreachable("bad atomic op");
2353 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
2354 struct ac_image_args
*a
)
2356 const char *overload
[3] = { "", "", "" };
2357 unsigned num_overloads
= 0;
2358 LLVMValueRef args
[18];
2359 unsigned num_args
= 0;
2360 enum ac_image_dim dim
= a
->dim
;
2362 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
2364 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
2365 a
->opcode
!= ac_image_store_mip
) ||
2367 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2368 (!a
->compare
&& !a
->offset
));
2369 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
2370 a
->opcode
== ac_image_get_lod
) ||
2372 assert((a
->bias
? 1 : 0) +
2374 (a
->level_zero
? 1 : 0) +
2375 (a
->derivs
[0] ? 1 : 0) <= 1);
2377 if (a
->opcode
== ac_image_get_lod
) {
2379 case ac_image_1darray
:
2382 case ac_image_2darray
:
2391 bool sample
= a
->opcode
== ac_image_sample
||
2392 a
->opcode
== ac_image_gather4
||
2393 a
->opcode
== ac_image_get_lod
;
2394 bool atomic
= a
->opcode
== ac_image_atomic
||
2395 a
->opcode
== ac_image_atomic_cmpswap
;
2396 bool load
= a
->opcode
== ac_image_sample
||
2397 a
->opcode
== ac_image_gather4
||
2398 a
->opcode
== ac_image_load
||
2399 a
->opcode
== ac_image_load_mip
;
2400 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
2402 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
2403 args
[num_args
++] = a
->data
[0];
2404 if (a
->opcode
== ac_image_atomic_cmpswap
)
2405 args
[num_args
++] = a
->data
[1];
2409 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
2412 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
2414 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
2415 overload
[num_overloads
++] = ".f32";
2418 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
2420 unsigned count
= ac_num_derivs(dim
);
2421 for (unsigned i
= 0; i
< count
; ++i
)
2422 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
2423 overload
[num_overloads
++] = ".f32";
2425 unsigned num_coords
=
2426 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
2427 for (unsigned i
= 0; i
< num_coords
; ++i
)
2428 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
2430 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
2431 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
2433 args
[num_args
++] = a
->resource
;
2435 args
[num_args
++] = a
->sampler
;
2436 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
2439 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
2440 args
[num_args
++] = LLVMConstInt(ctx
->i32
,
2441 load
? get_load_cache_policy(ctx
, a
->cache_policy
) :
2442 a
->cache_policy
, false);
2445 const char *atomic_subop
= "";
2446 switch (a
->opcode
) {
2447 case ac_image_sample
: name
= "sample"; break;
2448 case ac_image_gather4
: name
= "gather4"; break;
2449 case ac_image_load
: name
= "load"; break;
2450 case ac_image_load_mip
: name
= "load.mip"; break;
2451 case ac_image_store
: name
= "store"; break;
2452 case ac_image_store_mip
: name
= "store.mip"; break;
2453 case ac_image_atomic
:
2455 atomic_subop
= get_atomic_name(a
->atomic
);
2457 case ac_image_atomic_cmpswap
:
2459 atomic_subop
= "cmpswap";
2461 case ac_image_get_lod
: name
= "getlod"; break;
2462 case ac_image_get_resinfo
: name
= "getresinfo"; break;
2463 default: unreachable("invalid image opcode");
2466 const char *dimname
;
2468 case ac_image_1d
: dimname
= "1d"; break;
2469 case ac_image_2d
: dimname
= "2d"; break;
2470 case ac_image_3d
: dimname
= "3d"; break;
2471 case ac_image_cube
: dimname
= "cube"; break;
2472 case ac_image_1darray
: dimname
= "1darray"; break;
2473 case ac_image_2darray
: dimname
= "2darray"; break;
2474 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
2475 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
2476 default: unreachable("invalid dim");
2480 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
2482 snprintf(intr_name
, sizeof(intr_name
),
2483 "llvm.amdgcn.image.%s%s" /* base name */
2484 "%s%s%s" /* sample/gather modifiers */
2485 ".%s.%s%s%s%s", /* dimension and type overloads */
2487 a
->compare
? ".c" : "",
2490 a
->derivs
[0] ? ".d" :
2491 a
->level_zero
? ".lz" : "",
2492 a
->offset
? ".o" : "",
2494 atomic
? "i32" : "v4f32",
2495 overload
[0], overload
[1], overload
[2]);
2500 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
2505 LLVMValueRef result
=
2506 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
2508 if (!sample
&& retty
== ctx
->v4f32
) {
2509 result
= LLVMBuildBitCast(ctx
->builder
, result
,
2515 LLVMValueRef
ac_build_image_get_sample_count(struct ac_llvm_context
*ctx
,
2518 LLVMValueRef samples
;
2520 /* Read the samples from the descriptor directly.
2521 * Hardware doesn't have any instruction for this.
2523 samples
= LLVMBuildExtractElement(ctx
->builder
, rsrc
,
2524 LLVMConstInt(ctx
->i32
, 3, 0), "");
2525 samples
= LLVMBuildLShr(ctx
->builder
, samples
,
2526 LLVMConstInt(ctx
->i32
, 16, 0), "");
2527 samples
= LLVMBuildAnd(ctx
->builder
, samples
,
2528 LLVMConstInt(ctx
->i32
, 0xf, 0), "");
2529 samples
= LLVMBuildShl(ctx
->builder
, ctx
->i32_1
,
2534 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
2535 LLVMValueRef args
[2])
2538 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
2540 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
2541 args
, 2, AC_FUNC_ATTR_READNONE
);
2544 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
2545 LLVMValueRef args
[2])
2548 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
2549 ctx
->v2i16
, args
, 2,
2550 AC_FUNC_ATTR_READNONE
);
2551 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2554 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
2555 LLVMValueRef args
[2])
2558 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
2559 ctx
->v2i16
, args
, 2,
2560 AC_FUNC_ATTR_READNONE
);
2561 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2564 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2565 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
2566 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2568 assert(bits
== 8 || bits
== 10 || bits
== 16);
2570 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2571 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
2572 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2573 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
2574 LLVMValueRef max_alpha
=
2575 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
2576 LLVMValueRef min_alpha
=
2577 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2581 for (int i
= 0; i
< 2; i
++) {
2582 bool alpha
= hi
&& i
== 1;
2583 args
[i
] = ac_build_imin(ctx
, args
[i
],
2584 alpha
? max_alpha
: max_rgb
);
2585 args
[i
] = ac_build_imax(ctx
, args
[i
],
2586 alpha
? min_alpha
: min_rgb
);
2591 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
2592 ctx
->v2i16
, args
, 2,
2593 AC_FUNC_ATTR_READNONE
);
2594 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2597 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2598 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
2599 LLVMValueRef args
[2], unsigned bits
, bool hi
)
2601 assert(bits
== 8 || bits
== 10 || bits
== 16);
2603 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2604 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
2605 LLVMValueRef max_alpha
=
2606 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2610 for (int i
= 0; i
< 2; i
++) {
2611 bool alpha
= hi
&& i
== 1;
2612 args
[i
] = ac_build_umin(ctx
, args
[i
],
2613 alpha
? max_alpha
: max_rgb
);
2618 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
2619 ctx
->v2i16
, args
, 2,
2620 AC_FUNC_ATTR_READNONE
);
2621 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
2624 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2626 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
2627 &i1
, 1, AC_FUNC_ATTR_READNONE
);
2630 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
2632 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
2636 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2637 LLVMValueRef offset
, LLVMValueRef width
,
2640 LLVMValueRef args
[] = {
2646 return ac_build_intrinsic(ctx
, is_signed
? "llvm.amdgcn.sbfe.i32" :
2647 "llvm.amdgcn.ubfe.i32",
2648 ctx
->i32
, args
, 3, AC_FUNC_ATTR_READNONE
);
2652 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2653 LLVMValueRef s1
, LLVMValueRef s2
)
2655 return LLVMBuildAdd(ctx
->builder
,
2656 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2659 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
2660 LLVMValueRef s1
, LLVMValueRef s2
)
2662 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2663 if (ctx
->chip_class
>= GFX10
) {
2664 return ac_build_intrinsic(ctx
, "llvm.fma.f32", ctx
->f32
,
2665 (LLVMValueRef
[]) {s0
, s1
, s2
}, 3,
2666 AC_FUNC_ATTR_READNONE
);
2669 return LLVMBuildFAdd(ctx
->builder
,
2670 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
2673 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned wait_flags
)
2678 unsigned lgkmcnt
= 63;
2679 unsigned vmcnt
= ctx
->chip_class
>= GFX9
? 63 : 15;
2680 unsigned vscnt
= 63;
2682 if (wait_flags
& AC_WAIT_LGKM
)
2684 if (wait_flags
& AC_WAIT_VLOAD
)
2687 if (wait_flags
& AC_WAIT_VSTORE
) {
2688 if (ctx
->chip_class
>= GFX10
)
2694 /* There is no intrinsic for vscnt(0), so use a fence. */
2695 if ((wait_flags
& AC_WAIT_LGKM
&&
2696 wait_flags
& AC_WAIT_VLOAD
&&
2697 wait_flags
& AC_WAIT_VSTORE
) ||
2699 LLVMBuildFence(ctx
->builder
, LLVMAtomicOrderingRelease
, false, "");
2703 unsigned simm16
= (lgkmcnt
<< 8) |
2704 (7 << 4) | /* expcnt */
2706 ((vmcnt
>> 4) << 14);
2708 LLVMValueRef args
[1] = {
2709 LLVMConstInt(ctx
->i32
, simm16
, false),
2711 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2712 ctx
->voidt
, args
, 1, 0);
2715 LLVMValueRef
ac_build_fmed3(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2716 LLVMValueRef src1
, LLVMValueRef src2
,
2719 LLVMValueRef result
;
2721 if (bitsize
== 64 || (bitsize
== 16 && ctx
->chip_class
<= GFX8
)) {
2722 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2723 * or lower 16-bit fmed because it's only supported on GFX9+.
2725 LLVMValueRef min1
, min2
, max1
;
2727 min1
= ac_build_fmin(ctx
, src0
, src1
);
2728 max1
= ac_build_fmax(ctx
, src0
, src1
);
2729 min2
= ac_build_fmin(ctx
, max1
, src2
);
2731 result
= ac_build_fmax(ctx
, min2
, min1
);
2736 if (bitsize
== 16) {
2737 intr
= "llvm.amdgcn.fmed3.f16";
2740 assert(bitsize
== 32);
2741 intr
= "llvm.amdgcn.fmed3.f32";
2745 LLVMValueRef params
[] = {
2751 result
= ac_build_intrinsic(ctx
, intr
, type
, params
, 3,
2752 AC_FUNC_ATTR_READNONE
);
2755 if (ctx
->chip_class
< GFX9
&& bitsize
== 32) {
2756 /* Only pre-GFX9 chips do not flush denorms. */
2757 result
= ac_build_canonicalize(ctx
, result
, bitsize
);
2763 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2769 if (bitsize
== 16) {
2770 intr
= "llvm.amdgcn.fract.f16";
2772 } else if (bitsize
== 32) {
2773 intr
= "llvm.amdgcn.fract.f32";
2776 intr
= "llvm.amdgcn.fract.f64";
2780 LLVMValueRef params
[] = {
2783 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2784 AC_FUNC_ATTR_READNONE
);
2787 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2790 LLVMTypeRef type
= LLVMIntTypeInContext(ctx
->context
, bitsize
);
2791 LLVMValueRef zero
= LLVMConstInt(type
, 0, false);
2792 LLVMValueRef one
= LLVMConstInt(type
, 1, false);
2794 LLVMValueRef cmp
, val
;
2795 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2796 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2797 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2798 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2802 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2805 LLVMValueRef cmp
, val
, zero
, one
;
2808 if (bitsize
== 16) {
2812 } else if (bitsize
== 32) {
2822 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2823 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2824 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2825 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2829 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2831 LLVMValueRef result
;
2834 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2838 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i128", ctx
->i128
,
2839 (LLVMValueRef
[]) { src0
}, 1,
2840 AC_FUNC_ATTR_READNONE
);
2841 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2844 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2845 (LLVMValueRef
[]) { src0
}, 1,
2846 AC_FUNC_ATTR_READNONE
);
2848 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2851 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2852 (LLVMValueRef
[]) { src0
}, 1,
2853 AC_FUNC_ATTR_READNONE
);
2856 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2857 (LLVMValueRef
[]) { src0
}, 1,
2858 AC_FUNC_ATTR_READNONE
);
2860 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2863 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i8", ctx
->i8
,
2864 (LLVMValueRef
[]) { src0
}, 1,
2865 AC_FUNC_ATTR_READNONE
);
2867 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2870 unreachable(!"invalid bitsize");
2877 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2880 LLVMValueRef result
;
2883 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2887 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i64", ctx
->i64
,
2888 (LLVMValueRef
[]) { src0
}, 1,
2889 AC_FUNC_ATTR_READNONE
);
2891 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2894 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2895 (LLVMValueRef
[]) { src0
}, 1,
2896 AC_FUNC_ATTR_READNONE
);
2899 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2900 (LLVMValueRef
[]) { src0
}, 1,
2901 AC_FUNC_ATTR_READNONE
);
2903 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2906 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i8", ctx
->i8
,
2907 (LLVMValueRef
[]) { src0
}, 1,
2908 AC_FUNC_ATTR_READNONE
);
2910 result
= LLVMBuildZExt(ctx
->builder
, result
, ctx
->i32
, "");
2913 unreachable(!"invalid bitsize");
2920 #define AC_EXP_TARGET 0
2921 #define AC_EXP_ENABLED_CHANNELS 1
2922 #define AC_EXP_OUT0 2
2930 struct ac_vs_exp_chan
2934 enum ac_ir_type type
;
2937 struct ac_vs_exp_inst
{
2940 struct ac_vs_exp_chan chan
[4];
2943 struct ac_vs_exports
{
2945 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2948 /* Return true if the PARAM export has been eliminated. */
2949 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2950 uint32_t num_outputs
,
2951 struct ac_vs_exp_inst
*exp
)
2953 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2954 bool is_zero
[4] = {}, is_one
[4] = {};
2956 for (i
= 0; i
< 4; i
++) {
2957 /* It's a constant expression. Undef outputs are eliminated too. */
2958 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2961 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2962 if (exp
->chan
[i
].const_float
== 0)
2964 else if (exp
->chan
[i
].const_float
== 1)
2967 return false; /* other constant */
2972 /* Only certain combinations of 0 and 1 can be eliminated. */
2973 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2974 default_val
= is_zero
[3] ? 0 : 1;
2975 else if (is_one
[0] && is_one
[1] && is_one
[2])
2976 default_val
= is_zero
[3] ? 2 : 3;
2980 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2981 LLVMInstructionEraseFromParent(exp
->inst
);
2983 /* Change OFFSET to DEFAULT_VAL. */
2984 for (i
= 0; i
< num_outputs
; i
++) {
2985 if (vs_output_param_offset
[i
] == exp
->offset
) {
2986 vs_output_param_offset
[i
] =
2987 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
2994 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
2995 uint8_t *vs_output_param_offset
,
2996 uint32_t num_outputs
,
2997 struct ac_vs_exports
*processed
,
2998 struct ac_vs_exp_inst
*exp
)
3000 unsigned p
, copy_back_channels
= 0;
3002 /* See if the output is already in the list of processed outputs.
3003 * The LLVMValueRef comparison relies on SSA.
3005 for (p
= 0; p
< processed
->num
; p
++) {
3006 bool different
= false;
3008 for (unsigned j
= 0; j
< 4; j
++) {
3009 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
3010 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
3012 /* Treat undef as a match. */
3013 if (c2
->type
== AC_IR_UNDEF
)
3016 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3017 * and consider the instruction duplicated.
3019 if (c1
->type
== AC_IR_UNDEF
) {
3020 copy_back_channels
|= 1 << j
;
3024 /* Test whether the channels are not equal. */
3025 if (c1
->type
!= c2
->type
||
3026 (c1
->type
== AC_IR_CONST
&&
3027 c1
->const_float
!= c2
->const_float
) ||
3028 (c1
->type
== AC_IR_VALUE
&&
3029 c1
->value
!= c2
->value
)) {
3037 copy_back_channels
= 0;
3039 if (p
== processed
->num
)
3042 /* If a match was found, but the matching export has undef where the new
3043 * one has a normal value, copy the normal value to the undef channel.
3045 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
3047 /* Get current enabled channels mask. */
3048 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
3049 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
3051 while (copy_back_channels
) {
3052 unsigned chan
= u_bit_scan(©_back_channels
);
3054 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
3055 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
3056 exp
->chan
[chan
].value
);
3057 match
->chan
[chan
] = exp
->chan
[chan
];
3059 /* Update number of enabled channels because the original mask
3060 * is not always 0xf.
3062 enabled_channels
|= (1 << chan
);
3063 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
3064 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
3067 /* The PARAM export is duplicated. Kill it. */
3068 LLVMInstructionEraseFromParent(exp
->inst
);
3070 /* Change OFFSET to the matching export. */
3071 for (unsigned i
= 0; i
< num_outputs
; i
++) {
3072 if (vs_output_param_offset
[i
] == exp
->offset
) {
3073 vs_output_param_offset
[i
] = match
->offset
;
3080 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
3081 LLVMValueRef main_fn
,
3082 uint8_t *vs_output_param_offset
,
3083 uint32_t num_outputs
,
3084 uint32_t skip_output_mask
,
3085 uint8_t *num_param_exports
)
3087 LLVMBasicBlockRef bb
;
3088 bool removed_any
= false;
3089 struct ac_vs_exports exports
;
3093 /* Process all LLVM instructions. */
3094 bb
= LLVMGetFirstBasicBlock(main_fn
);
3096 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
3099 LLVMValueRef cur
= inst
;
3100 inst
= LLVMGetNextInstruction(inst
);
3101 struct ac_vs_exp_inst exp
;
3103 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
3106 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
3108 if (!ac_llvm_is_function(callee
))
3111 const char *name
= LLVMGetValueName(callee
);
3112 unsigned num_args
= LLVMCountParams(callee
);
3114 /* Check if this is an export instruction. */
3115 if ((num_args
!= 9 && num_args
!= 8) ||
3116 (strcmp(name
, "llvm.SI.export") &&
3117 strcmp(name
, "llvm.amdgcn.exp.f32")))
3120 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
3121 unsigned target
= LLVMConstIntGetZExtValue(arg
);
3123 if (target
< V_008DFC_SQ_EXP_PARAM
)
3126 target
-= V_008DFC_SQ_EXP_PARAM
;
3128 /* Parse the instruction. */
3129 memset(&exp
, 0, sizeof(exp
));
3130 exp
.offset
= target
;
3133 for (unsigned i
= 0; i
< 4; i
++) {
3134 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
3136 exp
.chan
[i
].value
= v
;
3138 if (LLVMIsUndef(v
)) {
3139 exp
.chan
[i
].type
= AC_IR_UNDEF
;
3140 } else if (LLVMIsAConstantFP(v
)) {
3141 LLVMBool loses_info
;
3142 exp
.chan
[i
].type
= AC_IR_CONST
;
3143 exp
.chan
[i
].const_float
=
3144 LLVMConstRealGetDouble(v
, &loses_info
);
3146 exp
.chan
[i
].type
= AC_IR_VALUE
;
3150 /* Eliminate constant and duplicated PARAM exports. */
3151 if (!((1u << target
) & skip_output_mask
) &&
3152 (ac_eliminate_const_output(vs_output_param_offset
,
3153 num_outputs
, &exp
) ||
3154 ac_eliminate_duplicated_output(ctx
,
3155 vs_output_param_offset
,
3156 num_outputs
, &exports
,
3160 exports
.exp
[exports
.num
++] = exp
;
3163 bb
= LLVMGetNextBasicBlock(bb
);
3166 /* Remove holes in export memory due to removed PARAM exports.
3167 * This is done by renumbering all PARAM exports.
3170 uint8_t old_offset
[VARYING_SLOT_MAX
];
3173 /* Make a copy of the offsets. We need the old version while
3174 * we are modifying some of them. */
3175 memcpy(old_offset
, vs_output_param_offset
,
3176 sizeof(old_offset
));
3178 for (i
= 0; i
< exports
.num
; i
++) {
3179 unsigned offset
= exports
.exp
[i
].offset
;
3181 /* Update vs_output_param_offset. Multiple outputs can
3182 * have the same offset.
3184 for (out
= 0; out
< num_outputs
; out
++) {
3185 if (old_offset
[out
] == offset
)
3186 vs_output_param_offset
[out
] = i
;
3189 /* Change the PARAM offset in the instruction. */
3190 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
3191 LLVMConstInt(ctx
->i32
,
3192 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
3194 *num_param_exports
= exports
.num
;
3198 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
3200 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
3201 ac_build_intrinsic(ctx
,
3202 "llvm.amdgcn.init.exec", ctx
->voidt
,
3203 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
3206 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
3208 unsigned lds_size
= ctx
->chip_class
>= GFX7
? 65536 : 32768;
3209 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
3210 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
3214 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
3215 LLVMValueRef dw_addr
)
3217 return LLVMBuildLoad(ctx
->builder
, ac_build_gep0(ctx
, ctx
->lds
, dw_addr
), "");
3220 void ac_lds_store(struct ac_llvm_context
*ctx
,
3221 LLVMValueRef dw_addr
,
3224 value
= ac_to_integer(ctx
, value
);
3225 ac_build_indexed_store(ctx
, ctx
->lds
,
3229 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
3230 LLVMTypeRef dst_type
,
3233 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
3234 const char *intrin_name
;
3238 switch (src0_bitsize
) {
3240 intrin_name
= "llvm.cttz.i64";
3245 intrin_name
= "llvm.cttz.i32";
3250 intrin_name
= "llvm.cttz.i16";
3255 intrin_name
= "llvm.cttz.i8";
3260 unreachable(!"invalid bitsize");
3263 LLVMValueRef params
[2] = {
3266 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3267 * add special code to check for x=0. The reason is that
3268 * the LLVM behavior for x=0 is different from what we
3269 * need here. However, LLVM also assumes that ffs(x) is
3270 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3271 * a conditional assignment to handle 0 is still required.
3273 * The hardware already implements the correct behavior.
3278 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
3280 AC_FUNC_ATTR_READNONE
);
3282 if (src0_bitsize
== 64) {
3283 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
3284 } else if (src0_bitsize
< 32) {
3285 lsb
= LLVMBuildSExt(ctx
->builder
, lsb
, ctx
->i32
, "");
3288 /* TODO: We need an intrinsic to skip this conditional. */
3289 /* Check for zero: */
3290 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
3293 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
3296 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
3298 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST
);
3301 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
3303 return LLVMPointerType(elem_type
, AC_ADDR_SPACE_CONST_32BIT
);
3306 static struct ac_llvm_flow
*
3307 get_current_flow(struct ac_llvm_context
*ctx
)
3309 if (ctx
->flow
->depth
> 0)
3310 return &ctx
->flow
->stack
[ctx
->flow
->depth
- 1];
3314 static struct ac_llvm_flow
*
3315 get_innermost_loop(struct ac_llvm_context
*ctx
)
3317 for (unsigned i
= ctx
->flow
->depth
; i
> 0; --i
) {
3318 if (ctx
->flow
->stack
[i
- 1].loop_entry_block
)
3319 return &ctx
->flow
->stack
[i
- 1];
3324 static struct ac_llvm_flow
*
3325 push_flow(struct ac_llvm_context
*ctx
)
3327 struct ac_llvm_flow
*flow
;
3329 if (ctx
->flow
->depth
>= ctx
->flow
->depth_max
) {
3330 unsigned new_max
= MAX2(ctx
->flow
->depth
<< 1,
3331 AC_LLVM_INITIAL_CF_DEPTH
);
3333 ctx
->flow
->stack
= realloc(ctx
->flow
->stack
, new_max
* sizeof(*ctx
->flow
->stack
));
3334 ctx
->flow
->depth_max
= new_max
;
3337 flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
];
3340 flow
->next_block
= NULL
;
3341 flow
->loop_entry_block
= NULL
;
3345 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
3349 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
3350 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
3353 /* Append a basic block at the level of the parent flow.
3355 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
3358 assert(ctx
->flow
->depth
>= 1);
3360 if (ctx
->flow
->depth
>= 2) {
3361 struct ac_llvm_flow
*flow
= &ctx
->flow
->stack
[ctx
->flow
->depth
- 2];
3363 return LLVMInsertBasicBlockInContext(ctx
->context
,
3364 flow
->next_block
, name
);
3367 LLVMValueRef main_fn
=
3368 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
3369 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
3372 /* Emit a branch to the given default target for the current block if
3373 * applicable -- that is, if the current block does not already contain a
3374 * branch from a break or continue.
3376 static void emit_default_branch(LLVMBuilderRef builder
,
3377 LLVMBasicBlockRef target
)
3379 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
3380 LLVMBuildBr(builder
, target
);
3383 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
3385 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3386 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
3387 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
3388 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
3389 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3390 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
3393 void ac_build_break(struct ac_llvm_context
*ctx
)
3395 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3396 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
3399 void ac_build_continue(struct ac_llvm_context
*ctx
)
3401 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
3402 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
3405 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
3407 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3408 LLVMBasicBlockRef endif_block
;
3410 assert(!current_branch
->loop_entry_block
);
3412 endif_block
= append_basic_block(ctx
, "ENDIF");
3413 emit_default_branch(ctx
->builder
, endif_block
);
3415 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3416 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
3418 current_branch
->next_block
= endif_block
;
3421 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
3423 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
3425 assert(!current_branch
->loop_entry_block
);
3427 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
3428 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
3429 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
3434 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
3436 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
3438 assert(current_loop
->loop_entry_block
);
3440 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
3442 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
3443 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
3447 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
3449 struct ac_llvm_flow
*flow
= push_flow(ctx
);
3450 LLVMBasicBlockRef if_block
;
3452 if_block
= append_basic_block(ctx
, "IF");
3453 flow
->next_block
= append_basic_block(ctx
, "ELSE");
3454 set_basicblock_name(if_block
, "if", label_id
);
3455 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
3456 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
3459 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3462 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
3463 value
, ctx
->f32_0
, "");
3464 ac_build_ifcc(ctx
, cond
, label_id
);
3467 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3470 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
3471 ac_to_integer(ctx
, value
),
3473 ac_build_ifcc(ctx
, cond
, label_id
);
3476 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
3479 LLVMBuilderRef builder
= ac
->builder
;
3480 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
3481 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
3482 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
3483 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
3484 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
3488 LLVMPositionBuilderBefore(first_builder
, first_instr
);
3490 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
3493 res
= LLVMBuildAlloca(first_builder
, type
, name
);
3494 LLVMDisposeBuilder(first_builder
);
3498 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
3499 LLVMTypeRef type
, const char *name
)
3501 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
3502 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
3506 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
3509 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3510 return LLVMBuildBitCast(ctx
->builder
, ptr
,
3511 LLVMPointerType(type
, addr_space
), "");
3514 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
3517 unsigned num_components
= ac_get_llvm_num_components(value
);
3518 if (count
== num_components
)
3521 LLVMValueRef masks
[MAX2(count
, 2)];
3522 masks
[0] = ctx
->i32_0
;
3523 masks
[1] = ctx
->i32_1
;
3524 for (unsigned i
= 2; i
< count
; i
++)
3525 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
3528 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
3531 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
3532 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
3535 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
3536 unsigned rshift
, unsigned bitwidth
)
3538 LLVMValueRef value
= param
;
3540 value
= LLVMBuildLShr(ctx
->builder
, value
,
3541 LLVMConstInt(ctx
->i32
, rshift
, false), "");
3543 if (rshift
+ bitwidth
< 32) {
3544 unsigned mask
= (1 << bitwidth
) - 1;
3545 value
= LLVMBuildAnd(ctx
->builder
, value
,
3546 LLVMConstInt(ctx
->i32
, mask
, false), "");
3551 /* Adjust the sample index according to FMASK.
3553 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3554 * which is the identity mapping. Each nibble says which physical sample
3555 * should be fetched to get that sample.
3557 * For example, 0x11111100 means there are only 2 samples stored and
3558 * the second sample covers 3/4 of the pixel. When reading samples 0
3559 * and 1, return physical sample 0 (determined by the first two 0s
3560 * in FMASK), otherwise return physical sample 1.
3562 * The sample index should be adjusted as follows:
3563 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3565 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
3566 LLVMValueRef
*addr
, bool is_array_tex
)
3568 struct ac_image_args fmask_load
= {};
3569 fmask_load
.opcode
= ac_image_load
;
3570 fmask_load
.resource
= fmask
;
3571 fmask_load
.dmask
= 0xf;
3572 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
3573 fmask_load
.attributes
= AC_FUNC_ATTR_READNONE
;
3575 fmask_load
.coords
[0] = addr
[0];
3576 fmask_load
.coords
[1] = addr
[1];
3578 fmask_load
.coords
[2] = addr
[2];
3580 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
3581 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
3584 /* Apply the formula. */
3585 unsigned sample_chan
= is_array_tex
? 3 : 2;
3586 LLVMValueRef final_sample
;
3587 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
3588 LLVMConstInt(ac
->i32
, 4, 0), "");
3589 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
3590 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3591 * with EQAA, so those will map to 0. */
3592 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
3593 LLVMConstInt(ac
->i32
, 0x7, 0), "");
3595 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3596 * resource descriptor is 0 (invalid).
3599 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
3600 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
3601 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
3603 /* Replace the MSAA sample index. */
3604 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
3605 addr
[sample_chan
], "");
3609 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3610 LLVMValueRef lane
, bool with_opt_barrier
)
3612 LLVMTypeRef type
= LLVMTypeOf(src
);
3613 LLVMValueRef result
;
3615 if (with_opt_barrier
)
3616 ac_build_optimization_barrier(ctx
, &src
);
3618 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3620 lane
= LLVMBuildZExt(ctx
->builder
, lane
, ctx
->i32
, "");
3622 result
= ac_build_intrinsic(ctx
,
3623 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3624 ctx
->i32
, (LLVMValueRef
[]) { src
, lane
},
3625 lane
== NULL
? 1 : 2,
3626 AC_FUNC_ATTR_READNONE
|
3627 AC_FUNC_ATTR_CONVERGENT
);
3629 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3633 ac_build_readlane_common(struct ac_llvm_context
*ctx
,
3634 LLVMValueRef src
, LLVMValueRef lane
,
3635 bool with_opt_barrier
)
3637 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3638 src
= ac_to_integer(ctx
, src
);
3639 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3643 assert(bits
% 32 == 0);
3644 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3645 LLVMValueRef src_vector
=
3646 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3647 ret
= LLVMGetUndef(vec_type
);
3648 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3649 LLVMValueRef ret_comp
;
3651 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3652 LLVMConstInt(ctx
->i32
, i
, 0), "");
3654 ret_comp
= _ac_build_readlane(ctx
, src
, lane
,
3657 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
3658 LLVMConstInt(ctx
->i32
, i
, 0), "");
3661 ret
= _ac_build_readlane(ctx
, src
, lane
, with_opt_barrier
);
3664 if (LLVMGetTypeKind(src_type
) == LLVMPointerTypeKind
)
3665 return LLVMBuildIntToPtr(ctx
->builder
, ret
, src_type
, "");
3666 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3670 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3672 * The optimization barrier is not needed if the value is the same in all lanes
3673 * or if this is called in the outermost block.
3677 * @param lane - id of the lane or NULL for the first active lane
3678 * @return value of the lane
3680 LLVMValueRef
ac_build_readlane_no_opt_barrier(struct ac_llvm_context
*ctx
,
3681 LLVMValueRef src
, LLVMValueRef lane
)
3683 return ac_build_readlane_common(ctx
, src
, lane
, false);
3688 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
3690 return ac_build_readlane_common(ctx
, src
, lane
, true);
3694 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
3696 return ac_build_intrinsic(ctx
, "llvm.amdgcn.writelane", ctx
->i32
,
3697 (LLVMValueRef
[]) {value
, lane
, src
}, 3,
3698 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3702 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
3704 if (ctx
->wave_size
== 32) {
3705 return ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3706 (LLVMValueRef
[]) { mask
, ctx
->i32_0
},
3707 2, AC_FUNC_ATTR_READNONE
);
3709 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
3710 LLVMVectorType(ctx
->i32
, 2),
3712 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3714 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
3717 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
3718 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
3719 2, AC_FUNC_ATTR_READNONE
);
3720 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
3721 (LLVMValueRef
[]) { mask_hi
, val
},
3722 2, AC_FUNC_ATTR_READNONE
);
3727 _dpp_quad_perm
= 0x000,
3728 _dpp_row_sl
= 0x100,
3729 _dpp_row_sr
= 0x110,
3730 _dpp_row_rr
= 0x120,
3735 dpp_row_mirror
= 0x140,
3736 dpp_row_half_mirror
= 0x141,
3737 dpp_row_bcast15
= 0x142,
3738 dpp_row_bcast31
= 0x143
3741 static inline enum dpp_ctrl
3742 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3744 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
3745 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
3748 static inline enum dpp_ctrl
3749 dpp_row_sl(unsigned amount
)
3751 assert(amount
> 0 && amount
< 16);
3752 return _dpp_row_sl
| amount
;
3755 static inline enum dpp_ctrl
3756 dpp_row_sr(unsigned amount
)
3758 assert(amount
> 0 && amount
< 16);
3759 return _dpp_row_sr
| amount
;
3763 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3764 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3767 LLVMTypeRef type
= LLVMTypeOf(src
);
3770 old
= LLVMBuildZExt(ctx
->builder
, old
, ctx
->i32
, "");
3771 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3773 res
= ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32", ctx
->i32
,
3776 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
3777 LLVMConstInt(ctx
->i32
, row_mask
, 0),
3778 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
3779 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
3780 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3782 return LLVMBuildTrunc(ctx
->builder
, res
, type
, "");
3786 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
3787 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
3790 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3791 src
= ac_to_integer(ctx
, src
);
3792 old
= ac_to_integer(ctx
, old
);
3793 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3796 assert(bits
% 32 == 0);
3797 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3798 LLVMValueRef src_vector
=
3799 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3800 LLVMValueRef old_vector
=
3801 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
3802 ret
= LLVMGetUndef(vec_type
);
3803 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3804 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3805 LLVMConstInt(ctx
->i32
, i
,
3807 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
3808 LLVMConstInt(ctx
->i32
, i
,
3810 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
3815 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3817 LLVMConstInt(ctx
->i32
, i
,
3821 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
3822 bank_mask
, bound_ctrl
);
3824 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3828 _ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3829 bool exchange_rows
, bool bound_ctrl
)
3831 LLVMTypeRef type
= LLVMTypeOf(src
);
3832 LLVMValueRef result
;
3834 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3836 LLVMValueRef args
[6] = {
3839 LLVMConstInt(ctx
->i32
, sel
, false),
3840 LLVMConstInt(ctx
->i32
, sel
>> 32, false),
3841 ctx
->i1true
, /* fi */
3842 bound_ctrl
? ctx
->i1true
: ctx
->i1false
,
3845 result
= ac_build_intrinsic(ctx
, exchange_rows
? "llvm.amdgcn.permlanex16"
3846 : "llvm.amdgcn.permlane16",
3848 AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3850 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
3854 ac_build_permlane16(struct ac_llvm_context
*ctx
, LLVMValueRef src
, uint64_t sel
,
3855 bool exchange_rows
, bool bound_ctrl
)
3857 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3858 src
= ac_to_integer(ctx
, src
);
3859 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3862 assert(bits
% 32 == 0);
3863 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3864 LLVMValueRef src_vector
=
3865 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3866 ret
= LLVMGetUndef(vec_type
);
3867 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3868 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3869 LLVMConstInt(ctx
->i32
, i
,
3871 LLVMValueRef ret_comp
=
3872 _ac_build_permlane16(ctx
, src
, sel
,
3875 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3877 LLVMConstInt(ctx
->i32
, i
,
3881 ret
= _ac_build_permlane16(ctx
, src
, sel
, exchange_rows
,
3884 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3887 static inline unsigned
3888 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
3890 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
3891 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
3895 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3897 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3900 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3902 ret
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle", ctx
->i32
,
3904 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3905 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3907 return LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3911 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3913 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3914 src
= ac_to_integer(ctx
, src
);
3915 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3918 assert(bits
% 32 == 0);
3919 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3920 LLVMValueRef src_vector
=
3921 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3922 ret
= LLVMGetUndef(vec_type
);
3923 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3924 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3925 LLVMConstInt(ctx
->i32
, i
,
3927 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3929 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3931 LLVMConstInt(ctx
->i32
, i
,
3935 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3937 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3941 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3943 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3944 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3945 char name
[32], type
[8];
3948 src
= ac_to_integer(ctx
, src
);
3951 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3953 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3954 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3955 ret
= ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3956 (LLVMValueRef
[]) { src
}, 1,
3957 AC_FUNC_ATTR_READNONE
);
3960 ret
= LLVMBuildTrunc(ctx
->builder
, ret
,
3961 ac_to_integer_type(ctx
, src_type
), "");
3963 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3967 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3968 LLVMValueRef inactive
)
3970 char name
[33], type
[8];
3971 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3972 unsigned bitsize
= ac_get_elem_bits(ctx
, src_type
);
3973 src
= ac_to_integer(ctx
, src
);
3974 inactive
= ac_to_integer(ctx
, inactive
);
3977 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
3978 inactive
= LLVMBuildZExt(ctx
->builder
, inactive
, ctx
->i32
, "");
3981 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3982 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3984 ac_build_intrinsic(ctx
, name
,
3985 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3987 AC_FUNC_ATTR_READNONE
|
3988 AC_FUNC_ATTR_CONVERGENT
);
3990 ret
= LLVMBuildTrunc(ctx
->builder
, ret
, src_type
, "");
3996 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3998 if (type_size
== 1) {
4000 case nir_op_iadd
: return ctx
->i8_0
;
4001 case nir_op_imul
: return ctx
->i8_1
;
4002 case nir_op_imin
: return LLVMConstInt(ctx
->i8
, INT8_MAX
, 0);
4003 case nir_op_umin
: return LLVMConstInt(ctx
->i8
, UINT8_MAX
, 0);
4004 case nir_op_imax
: return LLVMConstInt(ctx
->i8
, INT8_MIN
, 0);
4005 case nir_op_umax
: return ctx
->i8_0
;
4006 case nir_op_iand
: return LLVMConstInt(ctx
->i8
, -1, 0);
4007 case nir_op_ior
: return ctx
->i8_0
;
4008 case nir_op_ixor
: return ctx
->i8_0
;
4010 unreachable("bad reduction intrinsic");
4012 } else if (type_size
== 2) {
4014 case nir_op_iadd
: return ctx
->i16_0
;
4015 case nir_op_fadd
: return ctx
->f16_0
;
4016 case nir_op_imul
: return ctx
->i16_1
;
4017 case nir_op_fmul
: return ctx
->f16_1
;
4018 case nir_op_imin
: return LLVMConstInt(ctx
->i16
, INT16_MAX
, 0);
4019 case nir_op_umin
: return LLVMConstInt(ctx
->i16
, UINT16_MAX
, 0);
4020 case nir_op_fmin
: return LLVMConstReal(ctx
->f16
, INFINITY
);
4021 case nir_op_imax
: return LLVMConstInt(ctx
->i16
, INT16_MIN
, 0);
4022 case nir_op_umax
: return ctx
->i16_0
;
4023 case nir_op_fmax
: return LLVMConstReal(ctx
->f16
, -INFINITY
);
4024 case nir_op_iand
: return LLVMConstInt(ctx
->i16
, -1, 0);
4025 case nir_op_ior
: return ctx
->i16_0
;
4026 case nir_op_ixor
: return ctx
->i16_0
;
4028 unreachable("bad reduction intrinsic");
4030 } else if (type_size
== 4) {
4032 case nir_op_iadd
: return ctx
->i32_0
;
4033 case nir_op_fadd
: return ctx
->f32_0
;
4034 case nir_op_imul
: return ctx
->i32_1
;
4035 case nir_op_fmul
: return ctx
->f32_1
;
4036 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
4037 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
4038 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
4039 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
4040 case nir_op_umax
: return ctx
->i32_0
;
4041 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
4042 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
4043 case nir_op_ior
: return ctx
->i32_0
;
4044 case nir_op_ixor
: return ctx
->i32_0
;
4046 unreachable("bad reduction intrinsic");
4048 } else { /* type_size == 64bit */
4050 case nir_op_iadd
: return ctx
->i64_0
;
4051 case nir_op_fadd
: return ctx
->f64_0
;
4052 case nir_op_imul
: return ctx
->i64_1
;
4053 case nir_op_fmul
: return ctx
->f64_1
;
4054 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
4055 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
4056 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
4057 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
4058 case nir_op_umax
: return ctx
->i64_0
;
4059 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
4060 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
4061 case nir_op_ior
: return ctx
->i64_0
;
4062 case nir_op_ixor
: return ctx
->i64_0
;
4064 unreachable("bad reduction intrinsic");
4070 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
4072 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
4073 bool _32bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 4;
4075 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
4076 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
4077 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
4078 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
4079 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
4080 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
4082 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
4083 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
4085 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
4086 _64bit
? "llvm.minnum.f64" : _32bit
? "llvm.minnum.f32" : "llvm.minnum.f16",
4087 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4088 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4089 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
4090 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
4092 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
4093 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
4095 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
4096 _64bit
? "llvm.maxnum.f64" : _32bit
? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4097 _64bit
? ctx
->f64
: _32bit
? ctx
->f32
: ctx
->f16
,
4098 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
4099 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
4100 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
4101 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
4103 unreachable("bad reduction intrinsic");
4108 * \param src The value to shift.
4109 * \param identity The value to use the first lane.
4110 * \param maxprefix specifies that the result only needs to be correct for a
4111 * prefix of this many threads
4112 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4115 ac_wavefront_shift_right_1(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4116 LLVMValueRef identity
, unsigned maxprefix
)
4118 if (ctx
->chip_class
>= GFX10
) {
4119 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4120 LLVMValueRef active
, tmp1
, tmp2
;
4121 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4123 tmp1
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4125 tmp2
= ac_build_permlane16(ctx
, src
, (uint64_t)~0, true, false);
4127 if (maxprefix
> 32) {
4128 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4129 LLVMConstInt(ctx
->i32
, 32, false), "");
4131 tmp2
= LLVMBuildSelect(ctx
->builder
, active
,
4132 ac_build_readlane(ctx
, src
,
4133 LLVMConstInt(ctx
->i32
, 31, false)),
4136 active
= LLVMBuildOr(ctx
->builder
, active
,
4137 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4138 LLVMBuildAnd(ctx
->builder
, tid
,
4139 LLVMConstInt(ctx
->i32
, 0x1f, false), ""),
4140 LLVMConstInt(ctx
->i32
, 0x10, false), ""), "");
4141 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4142 } else if (maxprefix
> 16) {
4143 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
,
4144 LLVMConstInt(ctx
->i32
, 16, false), "");
4146 return LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4148 } else if (ctx
->chip_class
>= GFX8
) {
4149 return ac_build_dpp(ctx
, identity
, src
, dpp_wf_sr1
, 0xf, 0xf, false);
4152 /* wavefront shift_right by 1 on SI/CI */
4153 LLVMValueRef active
, tmp1
, tmp2
;
4154 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4155 tmp1
= ac_build_ds_swizzle(ctx
, src
, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4156 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4157 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4158 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x7, 0), ""),
4159 LLVMConstInt(ctx
->i32
, 0x4, 0), "");
4160 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4161 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4162 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4163 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0xf, 0), ""),
4164 LLVMConstInt(ctx
->i32
, 0x8, 0), "");
4165 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4166 tmp2
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4167 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
4168 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 0x1f, 0), ""),
4169 LLVMConstInt(ctx
->i32
, 0x10, 0), "");
4170 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4171 tmp2
= ac_build_readlane(ctx
, src
, LLVMConstInt(ctx
->i32
, 31, 0));
4172 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), "");
4173 tmp1
= LLVMBuildSelect(ctx
->builder
, active
, tmp2
, tmp1
, "");
4174 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, tid
, LLVMConstInt(ctx
->i32
, 0, 0), "");
4175 return LLVMBuildSelect(ctx
->builder
, active
, identity
, tmp1
, "");
4179 * \param maxprefix specifies that the result only needs to be correct for a
4180 * prefix of this many threads
4183 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
4184 unsigned maxprefix
, bool inclusive
)
4186 LLVMValueRef result
, tmp
;
4189 src
= ac_wavefront_shift_right_1(ctx
, src
, identity
, maxprefix
);
4193 if (ctx
->chip_class
<= GFX7
) {
4194 assert(maxprefix
== 64);
4195 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4196 LLVMValueRef active
;
4197 tmp
= ac_build_ds_swizzle(ctx
, src
, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4198 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4199 LLVMBuildAnd(ctx
->builder
, tid
, ctx
->i32_1
, ""),
4201 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4202 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4203 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4204 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4205 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 2, 0), ""),
4207 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4208 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4209 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x18, 0x03, 0x00));
4210 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4211 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 4, 0), ""),
4213 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4214 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4215 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x10, 0x07, 0x00));
4216 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4217 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 8, 0), ""),
4219 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4220 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4221 tmp
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4222 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4223 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 16, 0), ""),
4225 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4226 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4227 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, 0));
4228 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4229 LLVMBuildAnd(ctx
->builder
, tid
, LLVMConstInt(ctx
->i32
, 32, 0), ""),
4231 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4232 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4238 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
4239 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4242 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
4243 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4246 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
4247 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4250 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
4251 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4254 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
4255 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4256 if (maxprefix
<= 16)
4259 if (ctx
->chip_class
>= GFX10
) {
4260 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4261 LLVMValueRef active
;
4263 tmp
= ac_build_permlane16(ctx
, result
, ~(uint64_t)0, true, false);
4265 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
4266 LLVMBuildAnd(ctx
->builder
, tid
,
4267 LLVMConstInt(ctx
->i32
, 16, false), ""),
4270 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4272 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4274 if (maxprefix
<= 32)
4277 tmp
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4279 active
= LLVMBuildICmp(ctx
->builder
, LLVMIntUGE
, tid
,
4280 LLVMConstInt(ctx
->i32
, 32, false), "");
4282 tmp
= LLVMBuildSelect(ctx
->builder
, active
, tmp
, identity
, "");
4284 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4288 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4289 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4290 if (maxprefix
<= 32)
4292 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4293 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
4298 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4300 LLVMValueRef result
;
4302 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4303 LLVMBuilderRef builder
= ctx
->builder
;
4304 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4305 result
= ac_build_ballot(ctx
, src
);
4306 result
= ac_build_mbcnt(ctx
, result
);
4307 result
= LLVMBuildAdd(builder
, result
, src
, "");
4311 ac_build_optimization_barrier(ctx
, &src
);
4313 LLVMValueRef identity
=
4314 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4315 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4316 LLVMTypeOf(identity
), "");
4317 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, true);
4319 return ac_build_wwm(ctx
, result
);
4323 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
4325 LLVMValueRef result
;
4327 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
4328 LLVMBuilderRef builder
= ctx
->builder
;
4329 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
4330 result
= ac_build_ballot(ctx
, src
);
4331 result
= ac_build_mbcnt(ctx
, result
);
4335 ac_build_optimization_barrier(ctx
, &src
);
4337 LLVMValueRef identity
=
4338 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
4339 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
4340 LLVMTypeOf(identity
), "");
4341 result
= ac_build_scan(ctx
, op
, result
, identity
, ctx
->wave_size
, false);
4343 return ac_build_wwm(ctx
, result
);
4347 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
4349 if (cluster_size
== 1) return src
;
4350 ac_build_optimization_barrier(ctx
, &src
);
4351 LLVMValueRef result
, swap
;
4352 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
4353 ac_get_type_size(LLVMTypeOf(src
)));
4354 result
= LLVMBuildBitCast(ctx
->builder
,
4355 ac_build_set_inactive(ctx
, src
, identity
),
4356 LLVMTypeOf(identity
), "");
4357 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
4358 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4359 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
4361 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
4362 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4363 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
4365 if (ctx
->chip_class
>= GFX8
)
4366 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
4368 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
4369 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4370 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
4372 if (ctx
->chip_class
>= GFX8
)
4373 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
4375 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
4376 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4377 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
4379 if (ctx
->chip_class
>= GFX10
)
4380 swap
= ac_build_permlane16(ctx
, result
, 0, true, false);
4381 else if (ctx
->chip_class
>= GFX8
&& cluster_size
!= 32)
4382 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
4384 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
4385 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4386 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
4388 if (ctx
->chip_class
>= GFX8
) {
4389 if (ctx
->wave_size
== 64) {
4390 if (ctx
->chip_class
>= GFX10
)
4391 swap
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 31, false));
4393 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
4394 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4395 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
4398 return ac_build_wwm(ctx
, result
);
4400 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
4401 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
4402 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
4403 return ac_build_wwm(ctx
, result
);
4408 * "Top half" of a scan that reduces per-wave values across an entire
4411 * The source value must be present in the highest lane of the wave, and the
4412 * highest lane must be live.
4415 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4417 if (ws
->maxwaves
<= 1)
4420 const LLVMValueRef last_lane
= LLVMConstInt(ctx
->i32
, ctx
->wave_size
- 1, false);
4421 LLVMBuilderRef builder
= ctx
->builder
;
4422 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4425 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, last_lane
, "");
4426 ac_build_ifcc(ctx
, tmp
, 1000);
4427 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
4428 ac_build_endif(ctx
, 1000);
4432 * "Bottom half" of a scan that reduces per-wave values across an entire
4435 * The caller must place a barrier between the top and bottom halves.
4438 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4440 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
4441 const LLVMValueRef identity
=
4442 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
4444 if (ws
->maxwaves
<= 1) {
4445 ws
->result_reduce
= ws
->src
;
4446 ws
->result_inclusive
= ws
->src
;
4447 ws
->result_exclusive
= identity
;
4450 assert(ws
->maxwaves
<= 32);
4452 LLVMBuilderRef builder
= ctx
->builder
;
4453 LLVMValueRef tid
= ac_get_thread_id(ctx
);
4454 LLVMBasicBlockRef bbs
[2];
4455 LLVMValueRef phivalues_scan
[2];
4456 LLVMValueRef tmp
, tmp2
;
4458 bbs
[0] = LLVMGetInsertBlock(builder
);
4459 phivalues_scan
[0] = LLVMGetUndef(type
);
4461 if (ws
->enable_reduce
)
4462 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
4463 else if (ws
->enable_inclusive
)
4464 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
4466 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
4467 ac_build_ifcc(ctx
, tmp
, 1001);
4469 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
4471 ac_build_optimization_barrier(ctx
, &tmp
);
4473 bbs
[1] = LLVMGetInsertBlock(builder
);
4474 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
, true);
4476 ac_build_endif(ctx
, 1001);
4478 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
4480 if (ws
->enable_reduce
) {
4481 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
4482 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
4484 if (ws
->enable_inclusive
)
4485 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
4486 if (ws
->enable_exclusive
) {
4487 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
4488 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
4489 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
4490 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
4495 * Inclusive scan of a per-wave value across an entire workgroup.
4497 * This implies an s_barrier instruction.
4499 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4500 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4501 * useful manner because of the barrier in the algorithm.)
4504 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4506 ac_build_wg_wavescan_top(ctx
, ws
);
4507 ac_build_s_barrier(ctx
);
4508 ac_build_wg_wavescan_bottom(ctx
, ws
);
4512 * "Top half" of a scan that reduces per-thread values across an entire
4515 * All lanes must be active when this code runs.
4518 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4520 if (ws
->enable_exclusive
) {
4521 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
4522 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
4523 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
4524 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
4526 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
4529 bool enable_inclusive
= ws
->enable_inclusive
;
4530 bool enable_exclusive
= ws
->enable_exclusive
;
4531 ws
->enable_inclusive
= false;
4532 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4533 ac_build_wg_wavescan_top(ctx
, ws
);
4534 ws
->enable_inclusive
= enable_inclusive
;
4535 ws
->enable_exclusive
= enable_exclusive
;
4539 * "Bottom half" of a scan that reduces per-thread values across an entire
4542 * The caller must place a barrier between the top and bottom halves.
4545 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4547 bool enable_inclusive
= ws
->enable_inclusive
;
4548 bool enable_exclusive
= ws
->enable_exclusive
;
4549 ws
->enable_inclusive
= false;
4550 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
4551 ac_build_wg_wavescan_bottom(ctx
, ws
);
4552 ws
->enable_inclusive
= enable_inclusive
;
4553 ws
->enable_exclusive
= enable_exclusive
;
4555 /* ws->result_reduce is already the correct value */
4556 if (ws
->enable_inclusive
)
4557 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_inclusive
, ws
->src
, ws
->op
);
4558 if (ws
->enable_exclusive
)
4559 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
4563 * A scan that reduces per-thread values across an entire workgroup.
4565 * The caller must ensure that all lanes are active when this code runs
4566 * (WWM is insufficient!), because there is an implied barrier.
4569 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
4571 ac_build_wg_scan_top(ctx
, ws
);
4572 ac_build_s_barrier(ctx
);
4573 ac_build_wg_scan_bottom(ctx
, ws
);
4577 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
4578 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
4580 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
4581 if (ctx
->chip_class
>= GFX8
) {
4582 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
4584 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
4589 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
4591 LLVMTypeRef type
= LLVMTypeOf(src
);
4592 LLVMValueRef result
;
4594 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4595 src
= LLVMBuildZExt(ctx
->builder
, src
, ctx
->i32
, "");
4597 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4598 (LLVMValueRef
[]) {index
, src
}, 2,
4599 AC_FUNC_ATTR_READNONE
|
4600 AC_FUNC_ATTR_CONVERGENT
);
4601 return LLVMBuildTrunc(ctx
->builder
, result
, type
, "");
4605 ac_build_frexp_exp(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4611 if (bitsize
== 16) {
4612 intr
= "llvm.amdgcn.frexp.exp.i16.f16";
4614 } else if (bitsize
== 32) {
4615 intr
= "llvm.amdgcn.frexp.exp.i32.f32";
4618 intr
= "llvm.amdgcn.frexp.exp.i32.f64";
4622 LLVMValueRef params
[] = {
4625 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4626 AC_FUNC_ATTR_READNONE
);
4629 ac_build_frexp_mant(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4635 if (bitsize
== 16) {
4636 intr
= "llvm.amdgcn.frexp.mant.f16";
4638 } else if (bitsize
== 32) {
4639 intr
= "llvm.amdgcn.frexp.mant.f32";
4642 intr
= "llvm.amdgcn.frexp.mant.f64";
4646 LLVMValueRef params
[] = {
4649 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4650 AC_FUNC_ATTR_READNONE
);
4654 ac_build_canonicalize(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
4660 if (bitsize
== 16) {
4661 intr
= "llvm.canonicalize.f16";
4663 } else if (bitsize
== 32) {
4664 intr
= "llvm.canonicalize.f32";
4667 intr
= "llvm.canonicalize.f64";
4671 LLVMValueRef params
[] = {
4674 return ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
4675 AC_FUNC_ATTR_READNONE
);
4679 * this takes an I,J coordinate pair,
4680 * and works out the X and Y derivatives.
4681 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4684 ac_build_ddxy_interp(struct ac_llvm_context
*ctx
, LLVMValueRef interp_ij
)
4686 LLVMValueRef result
[4], a
;
4689 for (i
= 0; i
< 2; i
++) {
4690 a
= LLVMBuildExtractElement(ctx
->builder
, interp_ij
,
4691 LLVMConstInt(ctx
->i32
, i
, false), "");
4692 result
[i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 1, a
);
4693 result
[2+i
] = ac_build_ddxy(ctx
, AC_TID_MASK_TOP_LEFT
, 2, a
);
4695 return ac_build_gather_values(ctx
, result
, 4);
4699 ac_build_load_helper_invocation(struct ac_llvm_context
*ctx
)
4701 LLVMValueRef result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4703 AC_FUNC_ATTR_READNONE
);
4704 result
= LLVMBuildNot(ctx
->builder
, result
, "");
4705 return LLVMBuildSExt(ctx
->builder
, result
, ctx
->i32
, "");
4709 ac_build_is_helper_invocation(struct ac_llvm_context
*ctx
)
4711 if (!ctx
->postponed_kill
)
4712 return ac_build_load_helper_invocation(ctx
);
4714 /* !(exact && postponed) */
4715 LLVMValueRef exact
= ac_build_intrinsic(ctx
, "llvm.amdgcn.ps.live",
4717 AC_FUNC_ATTR_READNONE
);
4719 LLVMValueRef postponed
= LLVMBuildLoad(ctx
->builder
, ctx
->postponed_kill
, "");
4720 LLVMValueRef result
= LLVMBuildAnd(ctx
->builder
, exact
, postponed
, "");
4722 return LLVMBuildSelect(ctx
->builder
, result
, ctx
->i32_0
,
4723 LLVMConstInt(ctx
->i32
, 0xFFFFFFFF, false), "");
4726 LLVMValueRef
ac_build_call(struct ac_llvm_context
*ctx
, LLVMValueRef func
,
4727 LLVMValueRef
*args
, unsigned num_args
)
4729 LLVMValueRef ret
= LLVMBuildCall(ctx
->builder
, func
, args
, num_args
, "");
4730 LLVMSetInstructionCallConv(ret
, LLVMGetFunctionCallConv(func
));
4735 ac_export_mrt_z(struct ac_llvm_context
*ctx
, LLVMValueRef depth
,
4736 LLVMValueRef stencil
, LLVMValueRef samplemask
,
4737 struct ac_export_args
*args
)
4740 unsigned format
= ac_get_spi_shader_z_format(depth
!= NULL
,
4742 samplemask
!= NULL
);
4744 assert(depth
|| stencil
|| samplemask
);
4746 memset(args
, 0, sizeof(*args
));
4748 args
->valid_mask
= 1; /* whether the EXEC mask is valid */
4749 args
->done
= 1; /* DONE bit */
4751 /* Specify the target we are exporting */
4752 args
->target
= V_008DFC_SQ_EXP_MRTZ
;
4754 args
->compr
= 0; /* COMP flag */
4755 args
->out
[0] = LLVMGetUndef(ctx
->f32
); /* R, depth */
4756 args
->out
[1] = LLVMGetUndef(ctx
->f32
); /* G, stencil test val[0:7], stencil op val[8:15] */
4757 args
->out
[2] = LLVMGetUndef(ctx
->f32
); /* B, sample mask */
4758 args
->out
[3] = LLVMGetUndef(ctx
->f32
); /* A, alpha to mask */
4760 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
4762 args
->compr
= 1; /* COMPR flag */
4765 /* Stencil should be in X[23:16]. */
4766 stencil
= ac_to_integer(ctx
, stencil
);
4767 stencil
= LLVMBuildShl(ctx
->builder
, stencil
,
4768 LLVMConstInt(ctx
->i32
, 16, 0), "");
4769 args
->out
[0] = ac_to_float(ctx
, stencil
);
4773 /* SampleMask should be in Y[15:0]. */
4774 args
->out
[1] = samplemask
;
4779 args
->out
[0] = depth
;
4783 args
->out
[1] = stencil
;
4787 args
->out
[2] = samplemask
;
4792 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4793 * at the X writemask component. */
4794 if (ctx
->chip_class
== GFX6
&&
4795 ctx
->family
!= CHIP_OLAND
&&
4796 ctx
->family
!= CHIP_HAINAN
)
4799 /* Specify which components to enable */
4800 args
->enabled_channels
= mask
;
4803 /* Send GS Alloc Req message from the first wave of the group to SPI.
4804 * Message payload is:
4805 * - bits 0..10: vertices in group
4806 * - bits 12..22: primitives in group
4808 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context
*ctx
, LLVMValueRef wave_id
,
4809 LLVMValueRef vtx_cnt
, LLVMValueRef prim_cnt
)
4811 LLVMBuilderRef builder
= ctx
->builder
;
4813 bool export_dummy_prim
= false;
4815 /* HW workaround for a GPU hang with 100% culling.
4816 * We always have to export at least 1 primitive.
4817 * Export a degenerate triangle using vertex 0 for all 3 vertices.
4819 if (prim_cnt
== ctx
->i32_0
&& ctx
->chip_class
== GFX10
) {
4820 assert(vtx_cnt
== ctx
->i32_0
);
4821 prim_cnt
= ctx
->i32_1
;
4822 vtx_cnt
= ctx
->i32_1
;
4823 export_dummy_prim
= true;
4826 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, wave_id
, ctx
->i32_0
, ""), 5020);
4828 tmp
= LLVMBuildShl(builder
, prim_cnt
, LLVMConstInt(ctx
->i32
, 12, false),"");
4829 tmp
= LLVMBuildOr(builder
, tmp
, vtx_cnt
, "");
4830 ac_build_sendmsg(ctx
, AC_SENDMSG_GS_ALLOC_REQ
, tmp
);
4832 if (export_dummy_prim
) {
4833 struct ac_ngg_prim prim
= {};
4834 /* The vertex indices are 0,0,0. */
4835 prim
.passthrough
= ctx
->i32_0
;
4837 struct ac_export_args pos
= {};
4838 pos
.out
[0] = pos
.out
[1] = pos
.out
[2] = pos
.out
[3] = ctx
->f32_0
;
4839 pos
.target
= V_008DFC_SQ_EXP_POS
;
4840 pos
.enabled_channels
= 0xf;
4843 ac_build_ifcc(ctx
, LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(ctx
),
4844 ctx
->i32_0
, ""), 5021);
4845 ac_build_export_prim(ctx
, &prim
);
4846 ac_build_export(ctx
, &pos
);
4847 ac_build_endif(ctx
, 5021);
4850 ac_build_endif(ctx
, 5020);
4853 LLVMValueRef
ac_pack_prim_export(struct ac_llvm_context
*ctx
,
4854 const struct ac_ngg_prim
*prim
)
4856 /* The prim export format is:
4857 * - bits 0..8: index 0
4858 * - bit 9: edge flag 0
4859 * - bits 10..18: index 1
4860 * - bit 19: edge flag 1
4861 * - bits 20..28: index 2
4862 * - bit 29: edge flag 2
4863 * - bit 31: null primitive (skip)
4865 LLVMBuilderRef builder
= ctx
->builder
;
4866 LLVMValueRef tmp
= LLVMBuildZExt(builder
, prim
->isnull
, ctx
->i32
, "");
4867 LLVMValueRef result
= LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->i32
, 31, false), "");
4869 for (unsigned i
= 0; i
< prim
->num_vertices
; ++i
) {
4870 tmp
= LLVMBuildShl(builder
, prim
->index
[i
],
4871 LLVMConstInt(ctx
->i32
, 10 * i
, false), "");
4872 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4873 tmp
= LLVMBuildZExt(builder
, prim
->edgeflag
[i
], ctx
->i32
, "");
4874 tmp
= LLVMBuildShl(builder
, tmp
,
4875 LLVMConstInt(ctx
->i32
, 10 * i
+ 9, false), "");
4876 result
= LLVMBuildOr(builder
, result
, tmp
, "");
4881 void ac_build_export_prim(struct ac_llvm_context
*ctx
,
4882 const struct ac_ngg_prim
*prim
)
4884 struct ac_export_args args
;
4886 if (prim
->passthrough
) {
4887 args
.out
[0] = prim
->passthrough
;
4889 args
.out
[0] = ac_pack_prim_export(ctx
, prim
);
4892 args
.out
[0] = LLVMBuildBitCast(ctx
->builder
, args
.out
[0], ctx
->f32
, "");
4893 args
.out
[1] = LLVMGetUndef(ctx
->f32
);
4894 args
.out
[2] = LLVMGetUndef(ctx
->f32
);
4895 args
.out
[3] = LLVMGetUndef(ctx
->f32
);
4897 args
.target
= V_008DFC_SQ_EXP_PRIM
;
4898 args
.enabled_channels
= 1;
4900 args
.valid_mask
= false;
4903 ac_build_export(ctx
, &args
);
4907 arg_llvm_type(enum ac_arg_type type
, unsigned size
, struct ac_llvm_context
*ctx
)
4909 if (type
== AC_ARG_FLOAT
) {
4910 return size
== 1 ? ctx
->f32
: LLVMVectorType(ctx
->f32
, size
);
4911 } else if (type
== AC_ARG_INT
) {
4912 return size
== 1 ? ctx
->i32
: LLVMVectorType(ctx
->i32
, size
);
4914 LLVMTypeRef ptr_type
;
4916 case AC_ARG_CONST_PTR
:
4919 case AC_ARG_CONST_FLOAT_PTR
:
4920 ptr_type
= ctx
->f32
;
4922 case AC_ARG_CONST_PTR_PTR
:
4923 ptr_type
= ac_array_in_const32_addr_space(ctx
->i8
);
4925 case AC_ARG_CONST_DESC_PTR
:
4926 ptr_type
= ctx
->v4i32
;
4928 case AC_ARG_CONST_IMAGE_PTR
:
4929 ptr_type
= ctx
->v8i32
;
4932 unreachable("unknown arg type");
4935 return ac_array_in_const32_addr_space(ptr_type
);
4938 return ac_array_in_const_addr_space(ptr_type
);
4944 ac_build_main(const struct ac_shader_args
*args
,
4945 struct ac_llvm_context
*ctx
,
4946 enum ac_llvm_calling_convention convention
,
4947 const char *name
, LLVMTypeRef ret_type
,
4948 LLVMModuleRef module
)
4950 LLVMTypeRef arg_types
[AC_MAX_ARGS
];
4952 for (unsigned i
= 0; i
< args
->arg_count
; i
++) {
4953 arg_types
[i
] = arg_llvm_type(args
->args
[i
].type
,
4954 args
->args
[i
].size
, ctx
);
4957 LLVMTypeRef main_function_type
=
4958 LLVMFunctionType(ret_type
, arg_types
, args
->arg_count
, 0);
4960 LLVMValueRef main_function
=
4961 LLVMAddFunction(module
, name
, main_function_type
);
4962 LLVMBasicBlockRef main_function_body
=
4963 LLVMAppendBasicBlockInContext(ctx
->context
, main_function
, "main_body");
4964 LLVMPositionBuilderAtEnd(ctx
->builder
, main_function_body
);
4966 LLVMSetFunctionCallConv(main_function
, convention
);
4967 for (unsigned i
= 0; i
< args
->arg_count
; ++i
) {
4968 LLVMValueRef P
= LLVMGetParam(main_function
, i
);
4970 if (args
->args
[i
].file
!= AC_ARG_SGPR
)
4973 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_INREG
);
4975 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4976 ac_add_function_attr(ctx
->context
, main_function
, i
+ 1, AC_FUNC_ATTR_NOALIAS
);
4977 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4981 ctx
->main_function
= main_function
;
4983 if (LLVM_VERSION_MAJOR
>= 11) {
4984 /* Enable denormals for FP16 and FP64: */
4985 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math",
4987 /* Disable denormals for FP32: */
4988 LLVMAddTargetDependentFunctionAttr(main_function
, "denormal-fp-math-f32",
4989 "preserve-sign,preserve-sign");
4991 return main_function
;
4994 void ac_build_s_endpgm(struct ac_llvm_context
*ctx
)
4996 LLVMTypeRef calltype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
4997 LLVMValueRef code
= LLVMConstInlineAsm(calltype
, "s_endpgm", "", true, false);
4998 LLVMBuildCall(ctx
->builder
, code
, NULL
, 0, "");
5001 LLVMValueRef
ac_prefix_bitcount(struct ac_llvm_context
*ctx
,
5002 LLVMValueRef mask
, LLVMValueRef index
)
5004 LLVMBuilderRef builder
= ctx
->builder
;
5005 LLVMTypeRef type
= LLVMTypeOf(mask
);
5007 LLVMValueRef bit
= LLVMBuildShl(builder
, LLVMConstInt(type
, 1, 0),
5008 LLVMBuildZExt(builder
, index
, type
, ""), "");
5009 LLVMValueRef prefix_bits
= LLVMBuildSub(builder
, bit
, LLVMConstInt(type
, 1, 0), "");
5010 LLVMValueRef prefix_mask
= LLVMBuildAnd(builder
, mask
, prefix_bits
, "");
5011 return ac_build_bit_count(ctx
, prefix_mask
);
5014 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
5015 LLVMValueRef
ac_prefix_bitcount_2x64(struct ac_llvm_context
*ctx
,
5016 LLVMValueRef mask
[2], LLVMValueRef index
)
5018 LLVMBuilderRef builder
= ctx
->builder
;
5020 /* Reference version using i128. */
5021 LLVMValueRef input_mask
=
5022 LLVMBuildBitCast(builder
, ac_build_gather_values(ctx
, mask
, 2), ctx
->i128
, "");
5024 return ac_prefix_bitcount(ctx
, input_mask
, index
);
5026 /* Optimized version using 2 64-bit masks. */
5027 LLVMValueRef is_hi
, is_0
, c64
, c128
, all_bits
;
5028 LLVMValueRef prefix_mask
[2], shift
[2], mask_bcnt0
, prefix_bcnt
[2];
5030 /* Compute the 128-bit prefix mask. */
5031 c64
= LLVMConstInt(ctx
->i32
, 64, 0);
5032 c128
= LLVMConstInt(ctx
->i32
, 128, 0);
5033 all_bits
= LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
5034 /* The first index that can have non-zero high bits in the prefix mask is 65. */
5035 is_hi
= LLVMBuildICmp(builder
, LLVMIntUGT
, index
, c64
, "");
5036 is_0
= LLVMBuildICmp(builder
, LLVMIntEQ
, index
, ctx
->i32_0
, "");
5037 mask_bcnt0
= ac_build_bit_count(ctx
, mask
[0]);
5039 for (unsigned i
= 0; i
< 2; i
++) {
5040 shift
[i
] = LLVMBuildSub(builder
, i
? c128
: c64
, index
, "");
5041 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
5042 * so we handle it by the is_0 select.
5043 * For i==1, index==64, same story, so we handle it by the last is_hi select.
5044 * For i==0, index==64, we shift by 0, which is what we want.
5046 prefix_mask
[i
] = LLVMBuildLShr(builder
, all_bits
,
5047 LLVMBuildZExt(builder
, shift
[i
], ctx
->i64
, ""), "");
5048 prefix_mask
[i
] = LLVMBuildAnd(builder
, mask
[i
], prefix_mask
[i
], "");
5049 prefix_bcnt
[i
] = ac_build_bit_count(ctx
, prefix_mask
[i
]);
5052 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_0
, ctx
->i32_0
, prefix_bcnt
[0], "");
5053 prefix_bcnt
[0] = LLVMBuildSelect(builder
, is_hi
, mask_bcnt0
, prefix_bcnt
[0], "");
5054 prefix_bcnt
[1] = LLVMBuildSelect(builder
, is_hi
, prefix_bcnt
[1], ctx
->i32_0
, "");
5056 return LLVMBuildAdd(builder
, prefix_bcnt
[0], prefix_bcnt
[1], "");
5061 * Convert triangle strip indices to triangle indices. This is used to decompose
5062 * triangle strips into triangles.
5064 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context
*ctx
,
5065 LLVMValueRef is_odd
,
5066 LLVMValueRef flatshade_first
,
5067 LLVMValueRef index
[3])
5069 LLVMBuilderRef builder
= ctx
->builder
;
5070 LLVMValueRef out
[3];
5072 /* We need to change the vertex order for odd triangles to get correct
5073 * front/back facing by swapping 2 vertex indices, but we also have to
5074 * keep the provoking vertex in the same place.
5076 * If the first vertex is provoking, swap index 1 and 2.
5077 * If the last vertex is provoking, swap index 0 and 1.
5079 out
[0] = LLVMBuildSelect(builder
, flatshade_first
,
5081 LLVMBuildSelect(builder
, is_odd
,
5082 index
[1], index
[0], ""), "");
5083 out
[1] = LLVMBuildSelect(builder
, flatshade_first
,
5084 LLVMBuildSelect(builder
, is_odd
,
5085 index
[2], index
[1], ""),
5086 LLVMBuildSelect(builder
, is_odd
,
5087 index
[0], index
[1], ""), "");
5088 out
[2] = LLVMBuildSelect(builder
, flatshade_first
,
5089 LLVMBuildSelect(builder
, is_odd
,
5090 index
[1], index
[2], ""),
5092 memcpy(index
, out
, sizeof(out
));