2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
30 #include "c11/threads.h"
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "util/u_math.h"
43 #include "shader_enums.h"
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block
;
52 LLVMBasicBlockRef loop_entry_block
;
55 /* Initialize module-independent parts of the context.
57 * The caller is responsible for initializing ctx::module and ctx::builder.
60 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
61 enum chip_class chip_class
, enum radeon_family family
)
65 ctx
->context
= LLVMContextCreate();
67 ctx
->chip_class
= chip_class
;
72 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
73 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
74 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
75 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
76 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
77 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
78 ctx
->intptr
= HAVE_32BIT_POINTERS
? ctx
->i32
: ctx
->i64
;
79 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
80 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
81 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
82 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
83 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
84 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
85 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
86 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
87 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
88 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
90 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
91 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
92 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
93 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
94 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
95 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
96 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
97 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
99 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
100 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
102 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
105 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
106 "invariant.load", 14);
108 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
110 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
111 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
113 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
114 "amdgpu.uniform", 14);
116 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
120 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
124 ctx
->flow_depth_max
= 0;
128 ac_get_llvm_num_components(LLVMValueRef value
)
130 LLVMTypeRef type
= LLVMTypeOf(value
);
131 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
132 ? LLVMGetVectorSize(type
)
134 return num_components
;
138 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
142 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
147 return LLVMBuildExtractElement(ac
->builder
, value
,
148 LLVMConstInt(ac
->i32
, index
, false), "");
152 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
154 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
155 type
= LLVMGetElementType(type
);
157 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
158 return LLVMGetIntTypeWidth(type
);
160 if (type
== ctx
->f16
)
162 if (type
== ctx
->f32
)
164 if (type
== ctx
->f64
)
167 unreachable("Unhandled type kind in get_elem_bits");
171 ac_get_type_size(LLVMTypeRef type
)
173 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
176 case LLVMIntegerTypeKind
:
177 return LLVMGetIntTypeWidth(type
) / 8;
178 case LLVMFloatTypeKind
:
180 case LLVMDoubleTypeKind
:
182 case LLVMPointerTypeKind
:
183 if (LLVMGetPointerAddressSpace(type
) == AC_CONST_32BIT_ADDR_SPACE
)
186 case LLVMVectorTypeKind
:
187 return LLVMGetVectorSize(type
) *
188 ac_get_type_size(LLVMGetElementType(type
));
189 case LLVMArrayTypeKind
:
190 return LLVMGetArrayLength(type
) *
191 ac_get_type_size(LLVMGetElementType(type
));
198 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
200 if (t
== ctx
->f16
|| t
== ctx
->i16
)
202 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
204 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
207 unreachable("Unhandled integer size");
211 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
213 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
214 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
215 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
216 LLVMGetVectorSize(t
));
218 return to_integer_type_scalar(ctx
, t
);
222 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
224 LLVMTypeRef type
= LLVMTypeOf(v
);
225 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
228 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
230 if (t
== ctx
->i16
|| t
== ctx
->f16
)
232 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
234 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
237 unreachable("Unhandled float size");
241 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
243 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
244 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
245 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
246 LLVMGetVectorSize(t
));
248 return to_float_type_scalar(ctx
, t
);
252 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
254 LLVMTypeRef type
= LLVMTypeOf(v
);
255 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
260 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
261 LLVMTypeRef return_type
, LLVMValueRef
*params
,
262 unsigned param_count
, unsigned attrib_mask
)
264 LLVMValueRef function
, call
;
265 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
267 function
= LLVMGetNamedFunction(ctx
->module
, name
);
269 LLVMTypeRef param_types
[32], function_type
;
272 assert(param_count
<= 32);
274 for (i
= 0; i
< param_count
; ++i
) {
276 param_types
[i
] = LLVMTypeOf(params
[i
]);
279 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
280 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
282 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
283 LLVMSetLinkage(function
, LLVMExternalLinkage
);
285 if (!set_callsite_attrs
)
286 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
289 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
290 if (set_callsite_attrs
)
291 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
296 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
299 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
301 LLVMTypeRef elem_type
= type
;
303 assert(bufsize
>= 8);
305 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
306 int ret
= snprintf(buf
, bufsize
, "v%u",
307 LLVMGetVectorSize(type
));
309 char *type_name
= LLVMPrintTypeToString(type
);
310 fprintf(stderr
, "Error building type name for: %s\n",
314 elem_type
= LLVMGetElementType(type
);
318 switch (LLVMGetTypeKind(elem_type
)) {
320 case LLVMIntegerTypeKind
:
321 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
323 case LLVMFloatTypeKind
:
324 snprintf(buf
, bufsize
, "f32");
326 case LLVMDoubleTypeKind
:
327 snprintf(buf
, bufsize
, "f64");
333 * Helper function that builds an LLVM IR PHI node and immediately adds
337 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
338 unsigned count_incoming
, LLVMValueRef
*values
,
339 LLVMBasicBlockRef
*blocks
)
341 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
342 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
346 /* Prevent optimizations (at least of memory accesses) across the current
347 * point in the program by emitting empty inline assembly that is marked as
348 * having side effects.
350 * Optionally, a value can be passed through the inline assembly to prevent
351 * LLVM from hoisting calls to ReadNone functions.
354 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
357 static int counter
= 0;
359 LLVMBuilderRef builder
= ctx
->builder
;
362 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
365 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
366 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
367 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
369 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
370 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
371 LLVMValueRef vgpr
= *pvgpr
;
372 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
373 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
376 assert(vgpr_size
% 4 == 0);
378 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
379 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
380 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
381 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
382 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
389 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
391 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, "llvm.readcyclecounter",
392 ctx
->i64
, NULL
, 0, 0);
393 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
397 ac_build_ballot(struct ac_llvm_context
*ctx
,
400 LLVMValueRef args
[3] = {
403 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
406 /* We currently have no other way to prevent LLVM from lifting the icmp
407 * calls to a dominating basic block.
409 ac_build_optimization_barrier(ctx
, &args
[0]);
411 args
[0] = ac_to_integer(ctx
, args
[0]);
413 return ac_build_intrinsic(ctx
,
414 "llvm.amdgcn.icmp.i32",
416 AC_FUNC_ATTR_NOUNWIND
|
417 AC_FUNC_ATTR_READNONE
|
418 AC_FUNC_ATTR_CONVERGENT
);
422 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
424 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
425 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
426 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
430 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
432 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
433 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
434 LLVMConstInt(ctx
->i64
, 0, 0), "");
438 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
440 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
441 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
443 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
444 vote_set
, active_set
, "");
445 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
447 LLVMConstInt(ctx
->i64
, 0, 0), "");
448 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
452 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
453 unsigned value_count
, unsigned component
)
455 LLVMValueRef vec
= NULL
;
457 if (value_count
== 1) {
458 return values
[component
];
459 } else if (!value_count
)
460 unreachable("value_count is 0");
462 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
463 LLVMValueRef value
= values
[i
];
466 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
467 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
468 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
474 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
475 LLVMValueRef
*values
,
476 unsigned value_count
,
477 unsigned value_stride
,
481 LLVMBuilderRef builder
= ctx
->builder
;
482 LLVMValueRef vec
= NULL
;
485 if (value_count
== 1 && !always_vector
) {
487 return LLVMBuildLoad(builder
, values
[0], "");
489 } else if (!value_count
)
490 unreachable("value_count is 0");
492 for (i
= 0; i
< value_count
; i
++) {
493 LLVMValueRef value
= values
[i
* value_stride
];
495 value
= LLVMBuildLoad(builder
, value
, "");
498 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
499 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
500 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
506 ac_build_gather_values(struct ac_llvm_context
*ctx
,
507 LLVMValueRef
*values
,
508 unsigned value_count
)
510 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
513 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
514 * with undef. Extract at most num_channels components from the input.
516 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
518 unsigned num_channels
)
520 LLVMTypeRef elemtype
;
521 LLVMValueRef chan
[4];
523 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
524 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
525 num_channels
= MIN2(num_channels
, vec_size
);
527 if (num_channels
>= 4)
530 for (unsigned i
= 0; i
< num_channels
; i
++)
531 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
533 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
536 assert(num_channels
== 1);
539 elemtype
= LLVMTypeOf(value
);
542 while (num_channels
< 4)
543 chan
[num_channels
++] = LLVMGetUndef(elemtype
);
545 return ac_build_gather_values(ctx
, chan
, 4);
549 ac_build_fdiv(struct ac_llvm_context
*ctx
,
553 LLVMValueRef ret
= LLVMBuildFDiv(ctx
->builder
, num
, den
, "");
555 /* Use v_rcp_f32 instead of precise division. */
556 if (!LLVMIsConstant(ret
))
557 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
561 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
562 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
563 * already multiplied by two. id is the cube face number.
565 struct cube_selection_coords
{
572 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
574 struct cube_selection_coords
*out
)
576 LLVMTypeRef f32
= ctx
->f32
;
578 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
579 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
580 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
581 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
582 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
583 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
584 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
585 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
589 * Build a manual selection sequence for cube face sc/tc coordinates and
590 * major axis vector (multiplied by 2 for consistency) for the given
591 * vec3 \p coords, for the face implied by \p selcoords.
593 * For the major axis, we always adjust the sign to be in the direction of
594 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
595 * the selcoords major axis.
597 static void build_cube_select(struct ac_llvm_context
*ctx
,
598 const struct cube_selection_coords
*selcoords
,
599 const LLVMValueRef
*coords
,
600 LLVMValueRef
*out_st
,
601 LLVMValueRef
*out_ma
)
603 LLVMBuilderRef builder
= ctx
->builder
;
604 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
605 LLVMValueRef is_ma_positive
;
607 LLVMValueRef is_ma_z
, is_not_ma_z
;
608 LLVMValueRef is_ma_y
;
609 LLVMValueRef is_ma_x
;
613 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
614 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
615 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
616 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
618 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
619 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
620 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
621 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
622 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
625 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
626 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
627 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
628 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
629 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
632 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
633 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
634 LLVMConstReal(f32
, -1.0), "");
635 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
638 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
639 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
640 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
641 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
642 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
646 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
647 bool is_deriv
, bool is_array
, bool is_lod
,
648 LLVMValueRef
*coords_arg
,
649 LLVMValueRef
*derivs_arg
)
652 LLVMBuilderRef builder
= ctx
->builder
;
653 struct cube_selection_coords selcoords
;
654 LLVMValueRef coords
[3];
657 if (is_array
&& !is_lod
) {
658 LLVMValueRef tmp
= coords_arg
[3];
659 tmp
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &tmp
, 1, 0);
661 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
663 * "For Array forms, the array layer used will be
665 * max(0, min(d−1, floor(layer+0.5)))
667 * where d is the depth of the texture array and layer
668 * comes from the component indicated in the tables below.
669 * Workaroudn for an issue where the layer is taken from a
670 * helper invocation which happens to fall on a different
671 * layer due to extrapolation."
673 * VI and earlier attempt to implement this in hardware by
674 * clamping the value of coords[2] = (8 * layer) + face.
675 * Unfortunately, this means that the we end up with the wrong
676 * face when clamping occurs.
678 * Clamp the layer earlier to work around the issue.
680 if (ctx
->chip_class
<= VI
) {
682 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
683 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
689 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
691 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
692 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
693 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
695 for (int i
= 0; i
< 2; ++i
)
696 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
698 coords
[2] = selcoords
.id
;
700 if (is_deriv
&& derivs_arg
) {
701 LLVMValueRef derivs
[4];
704 /* Convert cube derivatives to 2D derivatives. */
705 for (axis
= 0; axis
< 2; axis
++) {
706 LLVMValueRef deriv_st
[2];
707 LLVMValueRef deriv_ma
;
709 /* Transform the derivative alongside the texture
710 * coordinate. Mathematically, the correct formula is
711 * as follows. Assume we're projecting onto the +Z face
712 * and denote by dx/dh the derivative of the (original)
713 * X texture coordinate with respect to horizontal
714 * window coordinates. The projection onto the +Z face
719 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
720 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
722 * This motivatives the implementation below.
724 * Whether this actually gives the expected results for
725 * apps that might feed in derivatives obtained via
726 * finite differences is anyone's guess. The OpenGL spec
727 * seems awfully quiet about how textureGrad for cube
728 * maps should be handled.
730 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
731 deriv_st
, &deriv_ma
);
733 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
735 for (int i
= 0; i
< 2; ++i
)
736 derivs
[axis
* 2 + i
] =
737 LLVMBuildFSub(builder
,
738 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
739 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
742 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
745 /* Shift the texture coordinate. This must be applied after the
746 * derivative calculation.
748 for (int i
= 0; i
< 2; ++i
)
749 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
752 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
753 /* coords_arg.w component - array_index for cube arrays */
754 LLVMValueRef tmp
= LLVMBuildFMul(ctx
->builder
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), "");
755 coords
[2] = LLVMBuildFAdd(ctx
->builder
, tmp
, coords
[2], "");
758 memcpy(coords_arg
, coords
, sizeof(coords
));
763 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
764 LLVMValueRef llvm_chan
,
765 LLVMValueRef attr_number
,
770 LLVMValueRef args
[5];
775 args
[2] = attr_number
;
778 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
779 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
784 args
[3] = attr_number
;
787 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
788 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
792 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
793 LLVMValueRef parameter
,
794 LLVMValueRef llvm_chan
,
795 LLVMValueRef attr_number
,
798 LLVMValueRef args
[4];
802 args
[2] = attr_number
;
805 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
806 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
810 ac_build_gep0(struct ac_llvm_context
*ctx
,
811 LLVMValueRef base_ptr
,
814 LLVMValueRef indices
[2] = {
815 LLVMConstInt(ctx
->i32
, 0, 0),
818 return LLVMBuildGEP(ctx
->builder
, base_ptr
,
823 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
824 LLVMValueRef base_ptr
, LLVMValueRef index
,
827 LLVMBuildStore(ctx
->builder
, value
,
828 ac_build_gep0(ctx
, base_ptr
, index
));
832 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
833 * It's equivalent to doing a load from &base_ptr[index].
835 * \param base_ptr Where the array starts.
836 * \param index The element index into the array.
837 * \param uniform Whether the base_ptr and index can be assumed to be
838 * dynamically uniform (i.e. load to an SGPR)
839 * \param invariant Whether the load is invariant (no other opcodes affect it)
842 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
843 LLVMValueRef index
, bool uniform
, bool invariant
)
845 LLVMValueRef pointer
, result
;
847 pointer
= ac_build_gep0(ctx
, base_ptr
, index
);
849 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
850 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
852 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
856 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
859 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false);
862 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
863 LLVMValueRef base_ptr
, LLVMValueRef index
)
865 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true);
868 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
869 LLVMValueRef base_ptr
, LLVMValueRef index
)
871 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true);
874 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
875 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
876 * or v4i32 (num_channels=3,4).
879 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
882 unsigned num_channels
,
883 LLVMValueRef voffset
,
884 LLVMValueRef soffset
,
885 unsigned inst_offset
,
888 bool writeonly_memory
,
889 bool swizzle_enable_hint
)
891 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
893 if (num_channels
== 3) {
894 LLVMValueRef v
[3], v01
;
896 for (int i
= 0; i
< 3; i
++) {
897 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
898 LLVMConstInt(ctx
->i32
, i
, 0), "");
900 v01
= ac_build_gather_values(ctx
, v
, 2);
902 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
903 soffset
, inst_offset
, glc
, slc
,
904 writeonly_memory
, swizzle_enable_hint
);
905 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
906 soffset
, inst_offset
+ 8,
908 writeonly_memory
, swizzle_enable_hint
);
912 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
913 * (voffset is swizzled, but soffset isn't swizzled).
914 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
916 if (!swizzle_enable_hint
) {
917 LLVMValueRef offset
= soffset
;
919 static const char *types
[] = {"f32", "v2f32", "v4f32"};
922 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
923 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
925 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
927 LLVMValueRef args
[] = {
928 ac_to_float(ctx
, vdata
),
929 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
930 LLVMConstInt(ctx
->i32
, 0, 0),
932 LLVMConstInt(ctx
->i1
, glc
, 0),
933 LLVMConstInt(ctx
->i1
, slc
, 0),
937 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
938 types
[CLAMP(num_channels
, 1, 3) - 1]);
940 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
941 args
, ARRAY_SIZE(args
),
943 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
944 AC_FUNC_ATTR_WRITEONLY
);
948 static const unsigned dfmt
[] = {
949 V_008F0C_BUF_DATA_FORMAT_32
,
950 V_008F0C_BUF_DATA_FORMAT_32_32
,
951 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
952 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
954 static const char *types
[] = {"i32", "v2i32", "v4i32"};
955 LLVMValueRef args
[] = {
957 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
958 LLVMConstInt(ctx
->i32
, 0, 0),
959 voffset
? voffset
: LLVMConstInt(ctx
->i32
, 0, 0),
961 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
962 LLVMConstInt(ctx
->i32
, dfmt
[num_channels
- 1], 0),
963 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, 0),
964 LLVMConstInt(ctx
->i1
, glc
, 0),
965 LLVMConstInt(ctx
->i1
, slc
, 0),
968 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.store.%s",
969 types
[CLAMP(num_channels
, 1, 3) - 1]);
971 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
972 args
, ARRAY_SIZE(args
),
974 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
975 AC_FUNC_ATTR_WRITEONLY
);
979 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
982 LLVMValueRef voffset
,
983 unsigned num_channels
,
989 LLVMValueRef args
[] = {
990 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
991 vindex
? vindex
: LLVMConstInt(ctx
->i32
, 0, 0),
993 LLVMConstInt(ctx
->i1
, glc
, 0),
994 LLVMConstInt(ctx
->i1
, slc
, 0)
996 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
998 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
999 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1003 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.format.%s",
1006 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
1010 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1012 ac_get_load_intr_attribs(can_speculate
));
1016 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1019 LLVMValueRef vindex
,
1020 LLVMValueRef voffset
,
1021 LLVMValueRef soffset
,
1022 unsigned inst_offset
,
1028 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1030 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1032 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1034 /* TODO: VI and later generations can use SMEM with GLC=1.*/
1035 if (allow_smem
&& !glc
&& !slc
) {
1036 assert(vindex
== NULL
);
1038 LLVMValueRef result
[8];
1040 for (int i
= 0; i
< num_channels
; i
++) {
1042 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1043 LLVMConstInt(ctx
->i32
, 4, 0), "");
1045 LLVMValueRef args
[2] = {rsrc
, offset
};
1046 result
[i
] = ac_build_intrinsic(ctx
, "llvm.SI.load.const.v4i32",
1048 AC_FUNC_ATTR_READNONE
|
1049 AC_FUNC_ATTR_LEGACY
);
1051 if (num_channels
== 1)
1054 if (num_channels
== 3)
1055 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1056 return ac_build_gather_values(ctx
, result
, num_channels
);
1059 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, offset
,
1060 num_channels
, glc
, slc
,
1061 can_speculate
, false);
1064 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1066 LLVMValueRef vindex
,
1067 LLVMValueRef voffset
,
1068 unsigned num_channels
,
1072 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1073 num_channels
, glc
, false,
1074 can_speculate
, true);
1077 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
1079 LLVMValueRef vindex
,
1080 LLVMValueRef voffset
,
1081 unsigned num_channels
,
1085 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, LLVMConstInt(ctx
->i32
, 2, 0), "");
1086 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, LLVMConstInt(ctx
->i32
, 1, 0), "");
1087 stride
= LLVMBuildLShr(ctx
->builder
, stride
, LLVMConstInt(ctx
->i32
, 16, 0), "");
1089 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->builder
,
1090 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, elem_count
, stride
, ""),
1091 elem_count
, stride
, "");
1093 LLVMValueRef new_rsrc
= LLVMBuildInsertElement(ctx
->builder
, rsrc
, new_elem_count
,
1094 LLVMConstInt(ctx
->i32
, 2, 0), "");
1096 return ac_build_buffer_load_common(ctx
, new_rsrc
, vindex
, voffset
,
1097 num_channels
, glc
, false,
1098 can_speculate
, true);
1102 * Set range metadata on an instruction. This can only be used on load and
1103 * call instructions. If you know an instruction can only produce the values
1104 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1105 * \p lo is the minimum value inclusive.
1106 * \p hi is the maximum value exclusive.
1108 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1109 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1111 LLVMValueRef range_md
, md_args
[2];
1112 LLVMTypeRef type
= LLVMTypeOf(value
);
1113 LLVMContextRef context
= LLVMGetTypeContext(type
);
1115 md_args
[0] = LLVMConstInt(type
, lo
, false);
1116 md_args
[1] = LLVMConstInt(type
, hi
, false);
1117 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1118 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1122 ac_get_thread_id(struct ac_llvm_context
*ctx
)
1126 LLVMValueRef tid_args
[2];
1127 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
1128 tid_args
[1] = LLVMConstInt(ctx
->i32
, 0, false);
1129 tid_args
[1] = ac_build_intrinsic(ctx
,
1130 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1131 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
1133 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
1135 2, AC_FUNC_ATTR_READNONE
);
1136 set_range_metadata(ctx
, tid
, 0, 64);
1141 * SI implements derivatives using the local data store (LDS)
1142 * All writes to the LDS happen in all executing threads at
1143 * the same time. TID is the Thread ID for the current
1144 * thread and is a value between 0 and 63, representing
1145 * the thread's position in the wavefront.
1147 * For the pixel shader threads are grouped into quads of four pixels.
1148 * The TIDs of the pixels of a quad are:
1156 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1157 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1158 * the current pixel's column, and masking with 0xfffffffe yields the TID
1159 * of the left pixel of the current pixel's row.
1161 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1162 * adding 2 yields the TID of the pixel below the top pixel.
1165 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1170 LLVMValueRef tl
, trbl
, args
[2];
1171 LLVMValueRef result
;
1173 if (HAVE_LLVM
>= 0x0700) {
1174 unsigned tl_lanes
[4], trbl_lanes
[4];
1176 for (unsigned i
= 0; i
< 4; ++i
) {
1177 tl_lanes
[i
] = i
& mask
;
1178 trbl_lanes
[i
] = (i
& mask
) + idx
;
1181 tl
= ac_build_quad_swizzle(ctx
, val
,
1182 tl_lanes
[0], tl_lanes
[1],
1183 tl_lanes
[2], tl_lanes
[3]);
1184 trbl
= ac_build_quad_swizzle(ctx
, val
,
1185 trbl_lanes
[0], trbl_lanes
[1],
1186 trbl_lanes
[2], trbl_lanes
[3]);
1187 } else if (ctx
->chip_class
>= VI
) {
1188 LLVMValueRef thread_id
, tl_tid
, trbl_tid
;
1189 thread_id
= ac_get_thread_id(ctx
);
1191 tl_tid
= LLVMBuildAnd(ctx
->builder
, thread_id
,
1192 LLVMConstInt(ctx
->i32
, mask
, false), "");
1194 trbl_tid
= LLVMBuildAdd(ctx
->builder
, tl_tid
,
1195 LLVMConstInt(ctx
->i32
, idx
, false), "");
1197 args
[0] = LLVMBuildMul(ctx
->builder
, tl_tid
,
1198 LLVMConstInt(ctx
->i32
, 4, false), "");
1200 tl
= ac_build_intrinsic(ctx
,
1201 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1203 AC_FUNC_ATTR_READNONE
|
1204 AC_FUNC_ATTR_CONVERGENT
);
1206 args
[0] = LLVMBuildMul(ctx
->builder
, trbl_tid
,
1207 LLVMConstInt(ctx
->i32
, 4, false), "");
1208 trbl
= ac_build_intrinsic(ctx
,
1209 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1211 AC_FUNC_ATTR_READNONE
|
1212 AC_FUNC_ATTR_CONVERGENT
);
1214 uint32_t masks
[2] = {};
1217 case AC_TID_MASK_TOP_LEFT
:
1225 case AC_TID_MASK_TOP
:
1229 case AC_TID_MASK_LEFT
:
1238 args
[1] = LLVMConstInt(ctx
->i32
, masks
[0], false);
1240 tl
= ac_build_intrinsic(ctx
,
1241 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1243 AC_FUNC_ATTR_READNONE
|
1244 AC_FUNC_ATTR_CONVERGENT
);
1246 args
[1] = LLVMConstInt(ctx
->i32
, masks
[1], false);
1247 trbl
= ac_build_intrinsic(ctx
,
1248 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1250 AC_FUNC_ATTR_READNONE
|
1251 AC_FUNC_ATTR_CONVERGENT
);
1254 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1255 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1256 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1258 if (HAVE_LLVM
>= 0x0700) {
1259 result
= ac_build_intrinsic(ctx
,
1260 "llvm.amdgcn.wqm.f32", ctx
->f32
,
1268 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1270 LLVMValueRef wave_id
)
1272 LLVMValueRef args
[2];
1273 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1275 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
1279 ac_build_imsb(struct ac_llvm_context
*ctx
,
1281 LLVMTypeRef dst_type
)
1283 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
1285 AC_FUNC_ATTR_READNONE
);
1287 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1288 * the index from LSB. Invert it by doing "31 - msb". */
1289 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1292 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1293 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1294 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1295 arg
, LLVMConstInt(ctx
->i32
, 0, 0), ""),
1296 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1297 arg
, all_ones
, ""), "");
1299 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1303 ac_build_umsb(struct ac_llvm_context
*ctx
,
1305 LLVMTypeRef dst_type
)
1307 const char *intrin_name
;
1309 LLVMValueRef highest_bit
;
1312 if (ac_get_elem_bits(ctx
, LLVMTypeOf(arg
)) == 64) {
1313 intrin_name
= "llvm.ctlz.i64";
1315 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
1318 intrin_name
= "llvm.ctlz.i32";
1320 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
1324 LLVMValueRef params
[2] = {
1329 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
1331 AC_FUNC_ATTR_READNONE
);
1333 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1334 * the index from LSB. Invert it by doing "31 - msb". */
1335 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
1336 msb
= LLVMBuildTruncOrBitCast(ctx
->builder
, msb
, ctx
->i32
, "");
1338 /* check for zero */
1339 return LLVMBuildSelect(ctx
->builder
,
1340 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
1341 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1344 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1347 LLVMValueRef args
[2] = {a
, b
};
1348 return ac_build_intrinsic(ctx
, "llvm.minnum.f32", ctx
->f32
, args
, 2,
1349 AC_FUNC_ATTR_READNONE
);
1352 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1355 LLVMValueRef args
[2] = {a
, b
};
1356 return ac_build_intrinsic(ctx
, "llvm.maxnum.f32", ctx
->f32
, args
, 2,
1357 AC_FUNC_ATTR_READNONE
);
1360 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1363 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
1364 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1367 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1370 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
1371 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1374 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1377 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
1378 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1381 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1383 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, ctx
->f32_0
),
1387 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
1389 LLVMValueRef args
[9];
1391 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1392 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1395 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
1396 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
1398 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
1400 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
1402 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1403 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1405 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
1406 ctx
->voidt
, args
, 6, 0);
1408 args
[2] = a
->out
[0];
1409 args
[3] = a
->out
[1];
1410 args
[4] = a
->out
[2];
1411 args
[5] = a
->out
[3];
1412 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1413 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1415 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
1416 ctx
->voidt
, args
, 8, 0);
1420 void ac_build_export_null(struct ac_llvm_context
*ctx
)
1422 struct ac_export_args args
;
1424 args
.enabled_channels
= 0x0; /* enabled channels */
1425 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
1426 args
.done
= 1; /* DONE bit */
1427 args
.target
= V_008DFC_SQ_EXP_NULL
;
1428 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
1429 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
1430 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
1431 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
1432 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
1434 ac_build_export(ctx
, &args
);
1437 static unsigned ac_num_coords(enum ac_image_dim dim
)
1443 case ac_image_1darray
:
1447 case ac_image_2darray
:
1448 case ac_image_2dmsaa
:
1450 case ac_image_2darraymsaa
:
1453 unreachable("ac_num_coords: bad dim");
1457 static unsigned ac_num_derivs(enum ac_image_dim dim
)
1461 case ac_image_1darray
:
1464 case ac_image_2darray
:
1469 case ac_image_2dmsaa
:
1470 case ac_image_2darraymsaa
:
1472 unreachable("derivatives not supported");
1476 static const char *get_atomic_name(enum ac_atomic_op op
)
1479 case ac_atomic_swap
: return "swap";
1480 case ac_atomic_add
: return "add";
1481 case ac_atomic_sub
: return "sub";
1482 case ac_atomic_smin
: return "smin";
1483 case ac_atomic_umin
: return "umin";
1484 case ac_atomic_smax
: return "smax";
1485 case ac_atomic_umax
: return "umax";
1486 case ac_atomic_and
: return "and";
1487 case ac_atomic_or
: return "or";
1488 case ac_atomic_xor
: return "xor";
1490 unreachable("bad atomic op");
1493 /* LLVM 6 and older */
1494 static LLVMValueRef
ac_build_image_opcode_llvm6(struct ac_llvm_context
*ctx
,
1495 struct ac_image_args
*a
)
1497 LLVMValueRef args
[16];
1498 LLVMTypeRef retty
= ctx
->v4f32
;
1499 const char *name
= NULL
;
1500 const char *atomic_subop
= "";
1501 char intr_name
[128], coords_type
[64];
1503 bool sample
= a
->opcode
== ac_image_sample
||
1504 a
->opcode
== ac_image_gather4
||
1505 a
->opcode
== ac_image_get_lod
;
1506 bool atomic
= a
->opcode
== ac_image_atomic
||
1507 a
->opcode
== ac_image_atomic_cmpswap
;
1508 bool da
= a
->dim
== ac_image_cube
||
1509 a
->dim
== ac_image_1darray
||
1510 a
->dim
== ac_image_2darray
||
1511 a
->dim
== ac_image_2darraymsaa
;
1512 if (a
->opcode
== ac_image_get_lod
)
1515 unsigned num_coords
=
1516 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(a
->dim
) : 0;
1518 unsigned num_addr
= 0;
1520 if (a
->opcode
== ac_image_get_lod
) {
1522 case ac_image_1darray
:
1525 case ac_image_2darray
:
1535 args
[num_addr
++] = ac_to_integer(ctx
, a
->offset
);
1537 args
[num_addr
++] = ac_to_integer(ctx
, a
->bias
);
1539 args
[num_addr
++] = ac_to_integer(ctx
, a
->compare
);
1541 unsigned num_derivs
= ac_num_derivs(a
->dim
);
1542 for (unsigned i
= 0; i
< num_derivs
; ++i
)
1543 args
[num_addr
++] = ac_to_integer(ctx
, a
->derivs
[i
]);
1545 for (unsigned i
= 0; i
< num_coords
; ++i
)
1546 args
[num_addr
++] = ac_to_integer(ctx
, a
->coords
[i
]);
1548 args
[num_addr
++] = ac_to_integer(ctx
, a
->lod
);
1550 unsigned pad_goal
= util_next_power_of_two(num_addr
);
1551 while (num_addr
< pad_goal
)
1552 args
[num_addr
++] = LLVMGetUndef(ctx
->i32
);
1554 addr
= ac_build_gather_values(ctx
, args
, num_addr
);
1556 unsigned num_args
= 0;
1557 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
1558 args
[num_args
++] = a
->data
[0];
1559 if (a
->opcode
== ac_image_atomic_cmpswap
)
1560 args
[num_args
++] = a
->data
[1];
1563 unsigned coords_arg
= num_args
;
1565 args
[num_args
++] = ac_to_float(ctx
, addr
);
1567 args
[num_args
++] = ac_to_integer(ctx
, addr
);
1569 args
[num_args
++] = a
->resource
;
1571 args
[num_args
++] = a
->sampler
;
1573 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1575 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, 0);
1576 args
[num_args
++] = a
->cache_policy
& ac_glc
? ctx
->i1true
: ctx
->i1false
;
1577 args
[num_args
++] = a
->cache_policy
& ac_slc
? ctx
->i1true
: ctx
->i1false
;
1578 args
[num_args
++] = ctx
->i1false
; /* lwe */
1579 args
[num_args
++] = LLVMConstInt(ctx
->i1
, da
, 0);
1581 args
[num_args
++] = ctx
->i1false
; /* r128 */
1582 args
[num_args
++] = LLVMConstInt(ctx
->i1
, da
, 0);
1583 args
[num_args
++] = a
->cache_policy
& ac_slc
? ctx
->i1true
: ctx
->i1false
;
1586 switch (a
->opcode
) {
1587 case ac_image_sample
:
1588 name
= "llvm.amdgcn.image.sample";
1590 case ac_image_gather4
:
1591 name
= "llvm.amdgcn.image.gather4";
1594 name
= "llvm.amdgcn.image.load";
1596 case ac_image_load_mip
:
1597 name
= "llvm.amdgcn.image.load.mip";
1599 case ac_image_store
:
1600 name
= "llvm.amdgcn.image.store";
1603 case ac_image_store_mip
:
1604 name
= "llvm.amdgcn.image.store.mip";
1607 case ac_image_atomic
:
1608 case ac_image_atomic_cmpswap
:
1609 name
= "llvm.amdgcn.image.atomic.";
1611 if (a
->opcode
== ac_image_atomic_cmpswap
) {
1612 atomic_subop
= "cmpswap";
1614 atomic_subop
= get_atomic_name(a
->atomic
);
1617 case ac_image_get_lod
:
1618 name
= "llvm.amdgcn.image.getlod";
1620 case ac_image_get_resinfo
:
1621 name
= "llvm.amdgcn.image.getresinfo";
1624 unreachable("invalid image opcode");
1627 ac_build_type_name_for_intr(LLVMTypeOf(args
[coords_arg
]), coords_type
,
1628 sizeof(coords_type
));
1631 snprintf(intr_name
, sizeof(intr_name
), "llvm.amdgcn.image.atomic.%s.%s",
1632 atomic_subop
, coords_type
);
1635 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
1637 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.v4f32.%s.v8i32",
1639 a
->compare
? ".c" : "",
1642 a
->derivs
[0] ? ".d" :
1643 a
->level_zero
? ".lz" : "",
1644 a
->offset
? ".o" : "",
1648 LLVMValueRef result
=
1649 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
1651 if (!sample
&& retty
== ctx
->v4f32
) {
1652 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1658 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
1659 struct ac_image_args
*a
)
1661 const char *overload
[3] = { "", "", "" };
1662 unsigned num_overloads
= 0;
1663 LLVMValueRef args
[18];
1664 unsigned num_args
= 0;
1665 enum ac_image_dim dim
= a
->dim
;
1667 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
1669 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
1670 a
->opcode
!= ac_image_store_mip
) ||
1672 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
1673 (!a
->compare
&& !a
->offset
));
1674 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
1675 a
->opcode
== ac_image_get_lod
) ||
1677 assert((a
->bias
? 1 : 0) +
1679 (a
->level_zero
? 1 : 0) +
1680 (a
->derivs
[0] ? 1 : 0) <= 1);
1682 if (HAVE_LLVM
< 0x0700)
1683 return ac_build_image_opcode_llvm6(ctx
, a
);
1685 if (a
->opcode
== ac_image_get_lod
) {
1687 case ac_image_1darray
:
1690 case ac_image_2darray
:
1699 bool sample
= a
->opcode
== ac_image_sample
||
1700 a
->opcode
== ac_image_gather4
||
1701 a
->opcode
== ac_image_get_lod
;
1702 bool atomic
= a
->opcode
== ac_image_atomic
||
1703 a
->opcode
== ac_image_atomic_cmpswap
;
1704 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
1706 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
1707 args
[num_args
++] = a
->data
[0];
1708 if (a
->opcode
== ac_image_atomic_cmpswap
)
1709 args
[num_args
++] = a
->data
[1];
1713 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
1716 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
1718 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
1719 overload
[num_overloads
++] = ".f32";
1722 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
1724 unsigned count
= ac_num_derivs(dim
);
1725 for (unsigned i
= 0; i
< count
; ++i
)
1726 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
1727 overload
[num_overloads
++] = ".f32";
1729 unsigned num_coords
=
1730 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
1731 for (unsigned i
= 0; i
< num_coords
; ++i
)
1732 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
1734 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
1735 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
1737 args
[num_args
++] = a
->resource
;
1739 args
[num_args
++] = a
->sampler
;
1740 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
1743 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
1744 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->cache_policy
, false);
1747 const char *atomic_subop
= "";
1748 switch (a
->opcode
) {
1749 case ac_image_sample
: name
= "sample"; break;
1750 case ac_image_gather4
: name
= "gather4"; break;
1751 case ac_image_load
: name
= "load"; break;
1752 case ac_image_load_mip
: name
= "load.mip"; break;
1753 case ac_image_store
: name
= "store"; break;
1754 case ac_image_store_mip
: name
= "store.mip"; break;
1755 case ac_image_atomic
:
1757 atomic_subop
= get_atomic_name(a
->atomic
);
1759 case ac_image_atomic_cmpswap
:
1761 atomic_subop
= "cmpswap";
1763 case ac_image_get_lod
: name
= "getlod"; break;
1764 case ac_image_get_resinfo
: name
= "getresinfo"; break;
1765 default: unreachable("invalid image opcode");
1768 const char *dimname
;
1770 case ac_image_1d
: dimname
= "1d"; break;
1771 case ac_image_2d
: dimname
= "2d"; break;
1772 case ac_image_3d
: dimname
= "3d"; break;
1773 case ac_image_cube
: dimname
= "cube"; break;
1774 case ac_image_1darray
: dimname
= "1darray"; break;
1775 case ac_image_2darray
: dimname
= "2darray"; break;
1776 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
1777 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
1778 default: unreachable("invalid dim");
1782 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
1784 snprintf(intr_name
, sizeof(intr_name
),
1785 "llvm.amdgcn.image.%s%s" /* base name */
1786 "%s%s%s" /* sample/gather modifiers */
1787 ".%s.%s%s%s%s", /* dimension and type overloads */
1789 a
->compare
? ".c" : "",
1792 a
->derivs
[0] ? ".d" :
1793 a
->level_zero
? ".lz" : "",
1794 a
->offset
? ".o" : "",
1796 atomic
? "i32" : "v4f32",
1797 overload
[0], overload
[1], overload
[2]);
1802 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
1807 LLVMValueRef result
=
1808 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
1810 if (!sample
&& retty
== ctx
->v4f32
) {
1811 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1817 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
1818 LLVMValueRef args
[2])
1821 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
1823 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz",
1825 AC_FUNC_ATTR_READNONE
);
1826 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1829 /* Upper 16 bits must be zero. */
1830 static LLVMValueRef
ac_llvm_pack_two_int16(struct ac_llvm_context
*ctx
,
1831 LLVMValueRef val
[2])
1833 return LLVMBuildOr(ctx
->builder
, val
[0],
1834 LLVMBuildShl(ctx
->builder
, val
[1],
1835 LLVMConstInt(ctx
->i32
, 16, 0),
1839 /* Upper 16 bits are ignored and will be dropped. */
1840 static LLVMValueRef
ac_llvm_pack_two_int32_as_int16(struct ac_llvm_context
*ctx
,
1841 LLVMValueRef val
[2])
1843 LLVMValueRef v
[2] = {
1844 LLVMBuildAnd(ctx
->builder
, val
[0],
1845 LLVMConstInt(ctx
->i32
, 0xffff, 0), ""),
1848 return ac_llvm_pack_two_int16(ctx
, v
);
1851 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
1852 LLVMValueRef args
[2])
1854 if (HAVE_LLVM
>= 0x0600) {
1856 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
1857 ctx
->v2i16
, args
, 2,
1858 AC_FUNC_ATTR_READNONE
);
1859 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1862 LLVMValueRef val
[2];
1864 for (int chan
= 0; chan
< 2; chan
++) {
1865 /* Clamp between [-1, 1]. */
1866 val
[chan
] = ac_build_fmin(ctx
, args
[chan
], ctx
->f32_1
);
1867 val
[chan
] = ac_build_fmax(ctx
, val
[chan
], LLVMConstReal(ctx
->f32
, -1));
1868 /* Convert to a signed integer in [-32767, 32767]. */
1869 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
1870 LLVMConstReal(ctx
->f32
, 32767), "");
1871 /* If positive, add 0.5, else add -0.5. */
1872 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
1873 LLVMBuildSelect(ctx
->builder
,
1874 LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
,
1875 val
[chan
], ctx
->f32_0
, ""),
1876 LLVMConstReal(ctx
->f32
, 0.5),
1877 LLVMConstReal(ctx
->f32
, -0.5), ""), "");
1878 val
[chan
] = LLVMBuildFPToSI(ctx
->builder
, val
[chan
], ctx
->i32
, "");
1880 return ac_llvm_pack_two_int32_as_int16(ctx
, val
);
1883 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
1884 LLVMValueRef args
[2])
1886 if (HAVE_LLVM
>= 0x0600) {
1888 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
1889 ctx
->v2i16
, args
, 2,
1890 AC_FUNC_ATTR_READNONE
);
1891 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1894 LLVMValueRef val
[2];
1896 for (int chan
= 0; chan
< 2; chan
++) {
1897 val
[chan
] = ac_build_clamp(ctx
, args
[chan
]);
1898 val
[chan
] = LLVMBuildFMul(ctx
->builder
, val
[chan
],
1899 LLVMConstReal(ctx
->f32
, 65535), "");
1900 val
[chan
] = LLVMBuildFAdd(ctx
->builder
, val
[chan
],
1901 LLVMConstReal(ctx
->f32
, 0.5), "");
1902 val
[chan
] = LLVMBuildFPToUI(ctx
->builder
, val
[chan
],
1905 return ac_llvm_pack_two_int32_as_int16(ctx
, val
);
1908 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1909 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
1910 LLVMValueRef args
[2], unsigned bits
, bool hi
)
1912 assert(bits
== 8 || bits
== 10 || bits
== 16);
1914 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
1915 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
1916 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
1917 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
1918 LLVMValueRef max_alpha
=
1919 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
1920 LLVMValueRef min_alpha
=
1921 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
1922 bool has_intrinsic
= HAVE_LLVM
>= 0x0600;
1925 if (!has_intrinsic
|| bits
!= 16) {
1926 for (int i
= 0; i
< 2; i
++) {
1927 bool alpha
= hi
&& i
== 1;
1928 args
[i
] = ac_build_imin(ctx
, args
[i
],
1929 alpha
? max_alpha
: max_rgb
);
1930 args
[i
] = ac_build_imax(ctx
, args
[i
],
1931 alpha
? min_alpha
: min_rgb
);
1935 if (has_intrinsic
) {
1937 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
1938 ctx
->v2i16
, args
, 2,
1939 AC_FUNC_ATTR_READNONE
);
1940 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1943 return ac_llvm_pack_two_int32_as_int16(ctx
, args
);
1946 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1947 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
1948 LLVMValueRef args
[2], unsigned bits
, bool hi
)
1950 assert(bits
== 8 || bits
== 10 || bits
== 16);
1952 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
1953 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
1954 LLVMValueRef max_alpha
=
1955 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
1956 bool has_intrinsic
= HAVE_LLVM
>= 0x0600;
1959 if (!has_intrinsic
|| bits
!= 16) {
1960 for (int i
= 0; i
< 2; i
++) {
1961 bool alpha
= hi
&& i
== 1;
1962 args
[i
] = ac_build_umin(ctx
, args
[i
],
1963 alpha
? max_alpha
: max_rgb
);
1967 if (has_intrinsic
) {
1969 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
1970 ctx
->v2i16
, args
, 2,
1971 AC_FUNC_ATTR_READNONE
);
1972 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1975 return ac_llvm_pack_two_int16(ctx
, args
);
1978 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1980 assert(HAVE_LLVM
>= 0x0600);
1981 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
1982 &i1
, 1, AC_FUNC_ATTR_READNONE
);
1985 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1987 if (HAVE_LLVM
>= 0x0600) {
1988 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
1993 LLVMValueRef value
= LLVMBuildSelect(ctx
->builder
, i1
,
1994 LLVMConstReal(ctx
->f32
, 1),
1995 LLVMConstReal(ctx
->f32
, -1), "");
1996 ac_build_intrinsic(ctx
, "llvm.AMDGPU.kill", ctx
->voidt
,
1997 &value
, 1, AC_FUNC_ATTR_LEGACY
);
2000 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
2001 LLVMValueRef offset
, LLVMValueRef width
,
2004 LLVMValueRef args
[] = {
2010 return ac_build_intrinsic(ctx
,
2011 is_signed
? "llvm.amdgcn.sbfe.i32" :
2012 "llvm.amdgcn.ubfe.i32",
2014 AC_FUNC_ATTR_READNONE
);
2017 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
2019 LLVMValueRef args
[1] = {
2020 LLVMConstInt(ctx
->i32
, simm16
, false),
2022 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
2023 ctx
->voidt
, args
, 1, 0);
2026 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2032 if (bitsize
== 32) {
2033 intr
= "llvm.floor.f32";
2036 intr
= "llvm.floor.f64";
2040 LLVMValueRef params
[] = {
2043 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2044 AC_FUNC_ATTR_READNONE
);
2045 return LLVMBuildFSub(ctx
->builder
, src0
, floor
, "");
2048 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2051 LLVMValueRef cmp
, val
, zero
, one
;
2054 if (bitsize
== 32) {
2064 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2065 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2066 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2067 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2071 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2074 LLVMValueRef cmp
, val
, zero
, one
;
2077 if (bitsize
== 32) {
2087 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2088 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2089 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2090 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2094 #define AC_EXP_TARGET 0
2095 #define AC_EXP_ENABLED_CHANNELS 1
2096 #define AC_EXP_OUT0 2
2104 struct ac_vs_exp_chan
2108 enum ac_ir_type type
;
2111 struct ac_vs_exp_inst
{
2114 struct ac_vs_exp_chan chan
[4];
2117 struct ac_vs_exports
{
2119 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2122 /* Return true if the PARAM export has been eliminated. */
2123 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2124 uint32_t num_outputs
,
2125 struct ac_vs_exp_inst
*exp
)
2127 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2128 bool is_zero
[4] = {}, is_one
[4] = {};
2130 for (i
= 0; i
< 4; i
++) {
2131 /* It's a constant expression. Undef outputs are eliminated too. */
2132 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2135 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2136 if (exp
->chan
[i
].const_float
== 0)
2138 else if (exp
->chan
[i
].const_float
== 1)
2141 return false; /* other constant */
2146 /* Only certain combinations of 0 and 1 can be eliminated. */
2147 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2148 default_val
= is_zero
[3] ? 0 : 1;
2149 else if (is_one
[0] && is_one
[1] && is_one
[2])
2150 default_val
= is_zero
[3] ? 2 : 3;
2154 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2155 LLVMInstructionEraseFromParent(exp
->inst
);
2157 /* Change OFFSET to DEFAULT_VAL. */
2158 for (i
= 0; i
< num_outputs
; i
++) {
2159 if (vs_output_param_offset
[i
] == exp
->offset
) {
2160 vs_output_param_offset
[i
] =
2161 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
2168 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
2169 uint8_t *vs_output_param_offset
,
2170 uint32_t num_outputs
,
2171 struct ac_vs_exports
*processed
,
2172 struct ac_vs_exp_inst
*exp
)
2174 unsigned p
, copy_back_channels
= 0;
2176 /* See if the output is already in the list of processed outputs.
2177 * The LLVMValueRef comparison relies on SSA.
2179 for (p
= 0; p
< processed
->num
; p
++) {
2180 bool different
= false;
2182 for (unsigned j
= 0; j
< 4; j
++) {
2183 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
2184 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
2186 /* Treat undef as a match. */
2187 if (c2
->type
== AC_IR_UNDEF
)
2190 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2191 * and consider the instruction duplicated.
2193 if (c1
->type
== AC_IR_UNDEF
) {
2194 copy_back_channels
|= 1 << j
;
2198 /* Test whether the channels are not equal. */
2199 if (c1
->type
!= c2
->type
||
2200 (c1
->type
== AC_IR_CONST
&&
2201 c1
->const_float
!= c2
->const_float
) ||
2202 (c1
->type
== AC_IR_VALUE
&&
2203 c1
->value
!= c2
->value
)) {
2211 copy_back_channels
= 0;
2213 if (p
== processed
->num
)
2216 /* If a match was found, but the matching export has undef where the new
2217 * one has a normal value, copy the normal value to the undef channel.
2219 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
2221 /* Get current enabled channels mask. */
2222 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
2223 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
2225 while (copy_back_channels
) {
2226 unsigned chan
= u_bit_scan(©_back_channels
);
2228 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
2229 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
2230 exp
->chan
[chan
].value
);
2231 match
->chan
[chan
] = exp
->chan
[chan
];
2233 /* Update number of enabled channels because the original mask
2234 * is not always 0xf.
2236 enabled_channels
|= (1 << chan
);
2237 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
2238 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
2241 /* The PARAM export is duplicated. Kill it. */
2242 LLVMInstructionEraseFromParent(exp
->inst
);
2244 /* Change OFFSET to the matching export. */
2245 for (unsigned i
= 0; i
< num_outputs
; i
++) {
2246 if (vs_output_param_offset
[i
] == exp
->offset
) {
2247 vs_output_param_offset
[i
] = match
->offset
;
2254 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
2255 LLVMValueRef main_fn
,
2256 uint8_t *vs_output_param_offset
,
2257 uint32_t num_outputs
,
2258 uint8_t *num_param_exports
)
2260 LLVMBasicBlockRef bb
;
2261 bool removed_any
= false;
2262 struct ac_vs_exports exports
;
2266 /* Process all LLVM instructions. */
2267 bb
= LLVMGetFirstBasicBlock(main_fn
);
2269 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
2272 LLVMValueRef cur
= inst
;
2273 inst
= LLVMGetNextInstruction(inst
);
2274 struct ac_vs_exp_inst exp
;
2276 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
2279 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
2281 if (!ac_llvm_is_function(callee
))
2284 const char *name
= LLVMGetValueName(callee
);
2285 unsigned num_args
= LLVMCountParams(callee
);
2287 /* Check if this is an export instruction. */
2288 if ((num_args
!= 9 && num_args
!= 8) ||
2289 (strcmp(name
, "llvm.SI.export") &&
2290 strcmp(name
, "llvm.amdgcn.exp.f32")))
2293 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
2294 unsigned target
= LLVMConstIntGetZExtValue(arg
);
2296 if (target
< V_008DFC_SQ_EXP_PARAM
)
2299 target
-= V_008DFC_SQ_EXP_PARAM
;
2301 /* Parse the instruction. */
2302 memset(&exp
, 0, sizeof(exp
));
2303 exp
.offset
= target
;
2306 for (unsigned i
= 0; i
< 4; i
++) {
2307 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
2309 exp
.chan
[i
].value
= v
;
2311 if (LLVMIsUndef(v
)) {
2312 exp
.chan
[i
].type
= AC_IR_UNDEF
;
2313 } else if (LLVMIsAConstantFP(v
)) {
2314 LLVMBool loses_info
;
2315 exp
.chan
[i
].type
= AC_IR_CONST
;
2316 exp
.chan
[i
].const_float
=
2317 LLVMConstRealGetDouble(v
, &loses_info
);
2319 exp
.chan
[i
].type
= AC_IR_VALUE
;
2323 /* Eliminate constant and duplicated PARAM exports. */
2324 if (ac_eliminate_const_output(vs_output_param_offset
,
2325 num_outputs
, &exp
) ||
2326 ac_eliminate_duplicated_output(ctx
,
2327 vs_output_param_offset
,
2328 num_outputs
, &exports
,
2332 exports
.exp
[exports
.num
++] = exp
;
2335 bb
= LLVMGetNextBasicBlock(bb
);
2338 /* Remove holes in export memory due to removed PARAM exports.
2339 * This is done by renumbering all PARAM exports.
2342 uint8_t old_offset
[VARYING_SLOT_MAX
];
2345 /* Make a copy of the offsets. We need the old version while
2346 * we are modifying some of them. */
2347 memcpy(old_offset
, vs_output_param_offset
,
2348 sizeof(old_offset
));
2350 for (i
= 0; i
< exports
.num
; i
++) {
2351 unsigned offset
= exports
.exp
[i
].offset
;
2353 /* Update vs_output_param_offset. Multiple outputs can
2354 * have the same offset.
2356 for (out
= 0; out
< num_outputs
; out
++) {
2357 if (old_offset
[out
] == offset
)
2358 vs_output_param_offset
[out
] = i
;
2361 /* Change the PARAM offset in the instruction. */
2362 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
2363 LLVMConstInt(ctx
->i32
,
2364 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
2366 *num_param_exports
= exports
.num
;
2370 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
2372 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
2373 ac_build_intrinsic(ctx
,
2374 "llvm.amdgcn.init.exec", ctx
->voidt
,
2375 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
2378 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
2380 unsigned lds_size
= ctx
->chip_class
>= CIK
? 65536 : 32768;
2381 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
2382 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_LOCAL_ADDR_SPACE
),
2386 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
2387 LLVMValueRef dw_addr
)
2389 return ac_build_load(ctx
, ctx
->lds
, dw_addr
);
2392 void ac_lds_store(struct ac_llvm_context
*ctx
,
2393 LLVMValueRef dw_addr
,
2396 value
= ac_to_integer(ctx
, value
);
2397 ac_build_indexed_store(ctx
, ctx
->lds
,
2401 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
2402 LLVMTypeRef dst_type
,
2405 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2406 const char *intrin_name
;
2409 if (src0_bitsize
== 64) {
2410 intrin_name
= "llvm.cttz.i64";
2414 intrin_name
= "llvm.cttz.i32";
2419 LLVMValueRef params
[2] = {
2422 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2423 * add special code to check for x=0. The reason is that
2424 * the LLVM behavior for x=0 is different from what we
2425 * need here. However, LLVM also assumes that ffs(x) is
2426 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2427 * a conditional assignment to handle 0 is still required.
2429 * The hardware already implements the correct behavior.
2431 LLVMConstInt(ctx
->i1
, 1, false),
2434 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2436 AC_FUNC_ATTR_READNONE
);
2438 if (src0_bitsize
== 64) {
2439 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
2442 /* TODO: We need an intrinsic to skip this conditional. */
2443 /* Check for zero: */
2444 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
2447 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
2450 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
2452 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2453 AC_CONST_ADDR_SPACE
);
2456 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
2458 if (!HAVE_32BIT_POINTERS
)
2459 return ac_array_in_const_addr_space(elem_type
);
2461 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2462 AC_CONST_32BIT_ADDR_SPACE
);
2465 static struct ac_llvm_flow
*
2466 get_current_flow(struct ac_llvm_context
*ctx
)
2468 if (ctx
->flow_depth
> 0)
2469 return &ctx
->flow
[ctx
->flow_depth
- 1];
2473 static struct ac_llvm_flow
*
2474 get_innermost_loop(struct ac_llvm_context
*ctx
)
2476 for (unsigned i
= ctx
->flow_depth
; i
> 0; --i
) {
2477 if (ctx
->flow
[i
- 1].loop_entry_block
)
2478 return &ctx
->flow
[i
- 1];
2483 static struct ac_llvm_flow
*
2484 push_flow(struct ac_llvm_context
*ctx
)
2486 struct ac_llvm_flow
*flow
;
2488 if (ctx
->flow_depth
>= ctx
->flow_depth_max
) {
2489 unsigned new_max
= MAX2(ctx
->flow_depth
<< 1,
2490 AC_LLVM_INITIAL_CF_DEPTH
);
2492 ctx
->flow
= realloc(ctx
->flow
, new_max
* sizeof(*ctx
->flow
));
2493 ctx
->flow_depth_max
= new_max
;
2496 flow
= &ctx
->flow
[ctx
->flow_depth
];
2499 flow
->next_block
= NULL
;
2500 flow
->loop_entry_block
= NULL
;
2504 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
2508 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
2509 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
2512 /* Append a basic block at the level of the parent flow.
2514 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
2517 assert(ctx
->flow_depth
>= 1);
2519 if (ctx
->flow_depth
>= 2) {
2520 struct ac_llvm_flow
*flow
= &ctx
->flow
[ctx
->flow_depth
- 2];
2522 return LLVMInsertBasicBlockInContext(ctx
->context
,
2523 flow
->next_block
, name
);
2526 LLVMValueRef main_fn
=
2527 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
2528 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
2531 /* Emit a branch to the given default target for the current block if
2532 * applicable -- that is, if the current block does not already contain a
2533 * branch from a break or continue.
2535 static void emit_default_branch(LLVMBuilderRef builder
,
2536 LLVMBasicBlockRef target
)
2538 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
2539 LLVMBuildBr(builder
, target
);
2542 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
2544 struct ac_llvm_flow
*flow
= push_flow(ctx
);
2545 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
2546 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
2547 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
2548 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
2549 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
2552 void ac_build_break(struct ac_llvm_context
*ctx
)
2554 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
2555 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
2558 void ac_build_continue(struct ac_llvm_context
*ctx
)
2560 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
2561 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
2564 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
2566 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
2567 LLVMBasicBlockRef endif_block
;
2569 assert(!current_branch
->loop_entry_block
);
2571 endif_block
= append_basic_block(ctx
, "ENDIF");
2572 emit_default_branch(ctx
->builder
, endif_block
);
2574 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
2575 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
2577 current_branch
->next_block
= endif_block
;
2580 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
2582 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
2584 assert(!current_branch
->loop_entry_block
);
2586 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
2587 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
2588 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
2593 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
2595 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
2597 assert(current_loop
->loop_entry_block
);
2599 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
2601 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
2602 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
2606 static void if_cond_emit(struct ac_llvm_context
*ctx
, LLVMValueRef cond
,
2609 struct ac_llvm_flow
*flow
= push_flow(ctx
);
2610 LLVMBasicBlockRef if_block
;
2612 if_block
= append_basic_block(ctx
, "IF");
2613 flow
->next_block
= append_basic_block(ctx
, "ELSE");
2614 set_basicblock_name(if_block
, "if", label_id
);
2615 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
2616 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
2619 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2622 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
2623 value
, ctx
->f32_0
, "");
2624 if_cond_emit(ctx
, cond
, label_id
);
2627 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2630 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
2631 ac_to_integer(ctx
, value
),
2633 if_cond_emit(ctx
, cond
, label_id
);
2636 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
2639 LLVMBuilderRef builder
= ac
->builder
;
2640 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
2641 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
2642 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
2643 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
2644 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
2648 LLVMPositionBuilderBefore(first_builder
, first_instr
);
2650 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
2653 res
= LLVMBuildAlloca(first_builder
, type
, name
);
2654 LLVMBuildStore(builder
, LLVMConstNull(type
), res
);
2656 LLVMDisposeBuilder(first_builder
);
2661 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
,
2662 LLVMTypeRef type
, const char *name
)
2664 LLVMValueRef ptr
= ac_build_alloca(ac
, type
, name
);
2665 LLVMBuildStore(ac
->builder
, LLVMGetUndef(type
), ptr
);
2669 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
2672 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
2673 return LLVMBuildBitCast(ctx
->builder
, ptr
,
2674 LLVMPointerType(type
, addr_space
), "");
2677 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2680 unsigned num_components
= ac_get_llvm_num_components(value
);
2681 if (count
== num_components
)
2684 LLVMValueRef masks
[] = {
2685 LLVMConstInt(ctx
->i32
, 0, false), LLVMConstInt(ctx
->i32
, 1, false),
2686 LLVMConstInt(ctx
->i32
, 2, false), LLVMConstInt(ctx
->i32
, 3, false)};
2689 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
2692 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
2693 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
2696 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
2697 unsigned rshift
, unsigned bitwidth
)
2699 LLVMValueRef value
= param
;
2701 value
= LLVMBuildLShr(ctx
->builder
, value
,
2702 LLVMConstInt(ctx
->i32
, rshift
, false), "");
2704 if (rshift
+ bitwidth
< 32) {
2705 unsigned mask
= (1 << bitwidth
) - 1;
2706 value
= LLVMBuildAnd(ctx
->builder
, value
,
2707 LLVMConstInt(ctx
->i32
, mask
, false), "");
2712 /* Adjust the sample index according to FMASK.
2714 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2715 * which is the identity mapping. Each nibble says which physical sample
2716 * should be fetched to get that sample.
2718 * For example, 0x11111100 means there are only 2 samples stored and
2719 * the second sample covers 3/4 of the pixel. When reading samples 0
2720 * and 1, return physical sample 0 (determined by the first two 0s
2721 * in FMASK), otherwise return physical sample 1.
2723 * The sample index should be adjusted as follows:
2724 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2726 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
2727 LLVMValueRef
*addr
, bool is_array_tex
)
2729 struct ac_image_args fmask_load
= {};
2730 fmask_load
.opcode
= ac_image_load
;
2731 fmask_load
.resource
= fmask
;
2732 fmask_load
.dmask
= 0xf;
2733 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
2735 fmask_load
.coords
[0] = addr
[0];
2736 fmask_load
.coords
[1] = addr
[1];
2738 fmask_load
.coords
[2] = addr
[2];
2740 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
2741 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
2744 /* Apply the formula. */
2745 unsigned sample_chan
= is_array_tex
? 3 : 2;
2746 LLVMValueRef final_sample
;
2747 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
2748 LLVMConstInt(ac
->i32
, 4, 0), "");
2749 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
2750 /* Mask the sample index by 0x7, because 0x8 means an unknown value
2751 * with EQAA, so those will map to 0. */
2752 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
2753 LLVMConstInt(ac
->i32
, 0x7, 0), "");
2755 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2756 * resource descriptor is 0 (invalid).
2759 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
2760 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
2761 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
2763 /* Replace the MSAA sample index. */
2764 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
2765 addr
[sample_chan
], "");
2769 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
2771 ac_build_optimization_barrier(ctx
, &src
);
2772 return ac_build_intrinsic(ctx
,
2773 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2774 LLVMTypeOf(src
), (LLVMValueRef
[]) {
2776 lane
== NULL
? 1 : 2,
2777 AC_FUNC_ATTR_READNONE
|
2778 AC_FUNC_ATTR_CONVERGENT
);
2782 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2785 * @param lane - id of the lane or NULL for the first active lane
2786 * @return value of the lane
2789 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
2791 LLVMTypeRef src_type
= LLVMTypeOf(src
);
2792 src
= ac_to_integer(ctx
, src
);
2793 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
2797 ret
= _ac_build_readlane(ctx
, src
, lane
);
2799 assert(bits
% 32 == 0);
2800 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
2801 LLVMValueRef src_vector
=
2802 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
2803 ret
= LLVMGetUndef(vec_type
);
2804 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
2805 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
2806 LLVMConstInt(ctx
->i32
, i
, 0), "");
2807 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
2808 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
2809 LLVMConstInt(ctx
->i32
, i
, 0), "");
2812 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
2816 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
2818 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
2820 LLVMValueRef pred
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, lane
,
2821 ac_get_thread_id(ctx
), "");
2822 return LLVMBuildSelect(ctx
->builder
, pred
, value
, src
, "");
2826 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
2828 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
2829 LLVMVectorType(ctx
->i32
, 2),
2831 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
2833 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
2836 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2837 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
2838 2, AC_FUNC_ATTR_READNONE
);
2839 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
2840 (LLVMValueRef
[]) { mask_hi
, val
},
2841 2, AC_FUNC_ATTR_READNONE
);
2846 _dpp_quad_perm
= 0x000,
2847 _dpp_row_sl
= 0x100,
2848 _dpp_row_sr
= 0x110,
2849 _dpp_row_rr
= 0x120,
2854 dpp_row_mirror
= 0x140,
2855 dpp_row_half_mirror
= 0x141,
2856 dpp_row_bcast15
= 0x142,
2857 dpp_row_bcast31
= 0x143
2860 static inline enum dpp_ctrl
2861 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
2863 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
2864 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
2867 static inline enum dpp_ctrl
2868 dpp_row_sl(unsigned amount
)
2870 assert(amount
> 0 && amount
< 16);
2871 return _dpp_row_sl
| amount
;
2874 static inline enum dpp_ctrl
2875 dpp_row_sr(unsigned amount
)
2877 assert(amount
> 0 && amount
< 16);
2878 return _dpp_row_sr
| amount
;
2882 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
2883 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
2886 return ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32",
2890 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
2891 LLVMConstInt(ctx
->i32
, row_mask
, 0),
2892 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
2893 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
2894 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
2898 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
2899 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
2902 LLVMTypeRef src_type
= LLVMTypeOf(src
);
2903 src
= ac_to_integer(ctx
, src
);
2904 old
= ac_to_integer(ctx
, old
);
2905 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
2908 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
2909 bank_mask
, bound_ctrl
);
2911 assert(bits
% 32 == 0);
2912 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
2913 LLVMValueRef src_vector
=
2914 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
2915 LLVMValueRef old_vector
=
2916 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
2917 ret
= LLVMGetUndef(vec_type
);
2918 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
2919 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
2920 LLVMConstInt(ctx
->i32
, i
,
2922 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
2923 LLVMConstInt(ctx
->i32
, i
,
2925 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
2930 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
2932 LLVMConstInt(ctx
->i32
, i
,
2936 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
2939 static inline unsigned
2940 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
2942 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
2943 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
2947 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
2949 return ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle",
2950 LLVMTypeOf(src
), (LLVMValueRef
[]) {
2951 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
2952 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
2956 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
2958 LLVMTypeRef src_type
= LLVMTypeOf(src
);
2959 src
= ac_to_integer(ctx
, src
);
2960 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
2963 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
2965 assert(bits
% 32 == 0);
2966 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
2967 LLVMValueRef src_vector
=
2968 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
2969 ret
= LLVMGetUndef(vec_type
);
2970 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
2971 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
2972 LLVMConstInt(ctx
->i32
, i
,
2974 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
2976 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
2978 LLVMConstInt(ctx
->i32
, i
,
2982 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
2986 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
2988 char name
[32], type
[8];
2989 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
2990 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
2991 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
2992 (LLVMValueRef
[]) { src
}, 1,
2993 AC_FUNC_ATTR_READNONE
);
2997 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
2998 LLVMValueRef inactive
)
3000 char name
[33], type
[8];
3001 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3002 src
= ac_to_integer(ctx
, src
);
3003 inactive
= ac_to_integer(ctx
, inactive
);
3004 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3005 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3007 ac_build_intrinsic(ctx
, name
,
3008 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3010 AC_FUNC_ATTR_READNONE
|
3011 AC_FUNC_ATTR_CONVERGENT
);
3012 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3016 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3018 if (type_size
== 4) {
3020 case nir_op_iadd
: return ctx
->i32_0
;
3021 case nir_op_fadd
: return ctx
->f32_0
;
3022 case nir_op_imul
: return ctx
->i32_1
;
3023 case nir_op_fmul
: return ctx
->f32_1
;
3024 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
3025 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
3026 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
3027 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
3028 case nir_op_umax
: return ctx
->i32_0
;
3029 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
3030 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
3031 case nir_op_ior
: return ctx
->i32_0
;
3032 case nir_op_ixor
: return ctx
->i32_0
;
3034 unreachable("bad reduction intrinsic");
3036 } else { /* type_size == 64bit */
3038 case nir_op_iadd
: return ctx
->i64_0
;
3039 case nir_op_fadd
: return ctx
->f64_0
;
3040 case nir_op_imul
: return ctx
->i64_1
;
3041 case nir_op_fmul
: return ctx
->f64_1
;
3042 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
3043 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
3044 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
3045 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
3046 case nir_op_umax
: return ctx
->i64_0
;
3047 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
3048 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
3049 case nir_op_ior
: return ctx
->i64_0
;
3050 case nir_op_ixor
: return ctx
->i64_0
;
3052 unreachable("bad reduction intrinsic");
3058 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
3060 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
3062 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
3063 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
3064 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
3065 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
3066 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
3067 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
3069 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
3070 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
3072 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
3073 _64bit
? "llvm.minnum.f64" : "llvm.minnum.f32",
3074 _64bit
? ctx
->f64
: ctx
->f32
,
3075 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3076 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
3077 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
3079 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
3080 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
3082 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
3083 _64bit
? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3084 _64bit
? ctx
->f64
: ctx
->f32
,
3085 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3086 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
3087 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
3088 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
3090 unreachable("bad reduction intrinsic");
3094 /* TODO: add inclusive and excluse scan functions for SI chip class. */
3096 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
)
3098 LLVMValueRef result
, tmp
;
3100 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
3101 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3102 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
3103 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3104 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
3105 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3106 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
3107 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3108 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
3109 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3110 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3111 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3112 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3113 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3118 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3120 ac_build_optimization_barrier(ctx
, &src
);
3121 LLVMValueRef result
;
3122 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
3123 ac_get_type_size(LLVMTypeOf(src
)));
3124 result
= LLVMBuildBitCast(ctx
->builder
,
3125 ac_build_set_inactive(ctx
, src
, identity
),
3126 LLVMTypeOf(identity
), "");
3127 result
= ac_build_scan(ctx
, op
, result
, identity
);
3129 return ac_build_wwm(ctx
, result
);
3133 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3135 ac_build_optimization_barrier(ctx
, &src
);
3136 LLVMValueRef result
;
3137 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
3138 ac_get_type_size(LLVMTypeOf(src
)));
3139 result
= LLVMBuildBitCast(ctx
->builder
,
3140 ac_build_set_inactive(ctx
, src
, identity
),
3141 LLVMTypeOf(identity
), "");
3142 result
= ac_build_dpp(ctx
, identity
, result
, dpp_wf_sr1
, 0xf, 0xf, false);
3143 result
= ac_build_scan(ctx
, op
, result
, identity
);
3145 return ac_build_wwm(ctx
, result
);
3149 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
3151 if (cluster_size
== 1) return src
;
3152 ac_build_optimization_barrier(ctx
, &src
);
3153 LLVMValueRef result
, swap
;
3154 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
3155 ac_get_type_size(LLVMTypeOf(src
)));
3156 result
= LLVMBuildBitCast(ctx
->builder
,
3157 ac_build_set_inactive(ctx
, src
, identity
),
3158 LLVMTypeOf(identity
), "");
3159 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
3160 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3161 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
3163 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
3164 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3165 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
3167 if (ctx
->chip_class
>= VI
)
3168 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
3170 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
3171 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3172 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
3174 if (ctx
->chip_class
>= VI
)
3175 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
3177 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
3178 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3179 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
3181 if (ctx
->chip_class
>= VI
&& cluster_size
!= 32)
3182 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3184 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
3185 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3186 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
3188 if (ctx
->chip_class
>= VI
) {
3189 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3190 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3191 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
3192 return ac_build_wwm(ctx
, result
);
3194 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
3195 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
3196 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3197 return ac_build_wwm(ctx
, result
);
3202 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3203 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3205 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
3206 if (ctx
->chip_class
>= VI
&& HAVE_LLVM
>= 0x0600) {
3207 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
3209 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
3214 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
3216 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
3217 return ac_build_intrinsic(ctx
,
3218 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
3219 (LLVMValueRef
[]) {index
, src
}, 2,
3220 AC_FUNC_ATTR_READNONE
|
3221 AC_FUNC_ATTR_CONVERGENT
);