1 /**************************************************************************
3 * Copyright 2009-2010 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
33 * LLVM IR doesn't support all basic arithmetic operations we care about (most
34 * notably min/max and saturated operations), and it is often necessary to
35 * resort machine-specific intrinsics directly. The functions here hide all
36 * these implementation details from the other modules.
38 * We also do simple expressions simplification here. Reasons are:
39 * - it is very easy given we have all necessary information readily available
40 * - LLVM optimization passes fail to simplify several vector expressions
41 * - We often know value constraints which the optimization passes have no way
42 * of knowing, such as when source arguments are known to be in [0, 1] range.
44 * @author Jose Fonseca <jfonseca@vmware.com>
50 #include "util/u_memory.h"
51 #include "util/u_debug.h"
52 #include "util/u_math.h"
53 #include "util/u_string.h"
54 #include "util/u_cpu_detect.h"
56 #include "lp_bld_type.h"
57 #include "lp_bld_const.h"
58 #include "lp_bld_init.h"
59 #include "lp_bld_intr.h"
60 #include "lp_bld_logic.h"
61 #include "lp_bld_pack.h"
62 #include "lp_bld_debug.h"
63 #include "lp_bld_bitarit.h"
64 #include "lp_bld_arit.h"
65 #include "lp_bld_flow.h"
68 #define EXP_POLY_DEGREE 5
70 #define LOG_POLY_DEGREE 4
75 * No checks for special case values of a or b = 1 or 0 are done.
76 * NaN's are handled according to the behavior specified by the
77 * nan_behavior argument.
80 lp_build_min_simple(struct lp_build_context
*bld
,
83 enum gallivm_nan_behavior nan_behavior
)
85 const struct lp_type type
= bld
->type
;
86 const char *intrinsic
= NULL
;
87 unsigned intr_size
= 0;
90 assert(lp_check_value(type
, a
));
91 assert(lp_check_value(type
, b
));
93 /* TODO: optimize the constant case */
95 if (type
.floating
&& util_cpu_caps
.has_sse
) {
96 if (type
.width
== 32) {
97 if (type
.length
== 1) {
98 intrinsic
= "llvm.x86.sse.min.ss";
101 else if (type
.length
<= 4 || !util_cpu_caps
.has_avx
) {
102 intrinsic
= "llvm.x86.sse.min.ps";
106 intrinsic
= "llvm.x86.avx.min.ps.256";
110 if (type
.width
== 64 && util_cpu_caps
.has_sse2
) {
111 if (type
.length
== 1) {
112 intrinsic
= "llvm.x86.sse2.min.sd";
115 else if (type
.length
== 2 || !util_cpu_caps
.has_avx
) {
116 intrinsic
= "llvm.x86.sse2.min.pd";
120 intrinsic
= "llvm.x86.avx.min.pd.256";
125 else if (type
.floating
&& util_cpu_caps
.has_altivec
) {
126 if (nan_behavior
== GALLIVM_NAN_RETURN_NAN
) {
127 debug_printf("%s: altivec doesn't support nan return nan behavior\n",
130 if (type
.width
== 32 && type
.length
== 4) {
131 intrinsic
= "llvm.ppc.altivec.vminfp";
134 } else if (util_cpu_caps
.has_sse2
&& type
.length
>= 2) {
136 if ((type
.width
== 8 || type
.width
== 16) &&
137 (type
.width
* type
.length
<= 64) &&
138 (gallivm_debug
& GALLIVM_DEBUG_PERF
)) {
139 debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
142 if (type
.width
== 8 && !type
.sign
) {
143 intrinsic
= "llvm.x86.sse2.pminu.b";
145 else if (type
.width
== 16 && type
.sign
) {
146 intrinsic
= "llvm.x86.sse2.pmins.w";
148 if (util_cpu_caps
.has_sse4_1
) {
149 if (type
.width
== 8 && type
.sign
) {
150 intrinsic
= "llvm.x86.sse41.pminsb";
152 if (type
.width
== 16 && !type
.sign
) {
153 intrinsic
= "llvm.x86.sse41.pminuw";
155 if (type
.width
== 32 && !type
.sign
) {
156 intrinsic
= "llvm.x86.sse41.pminud";
158 if (type
.width
== 32 && type
.sign
) {
159 intrinsic
= "llvm.x86.sse41.pminsd";
162 } else if (util_cpu_caps
.has_altivec
) {
164 if (type
.width
== 8) {
166 intrinsic
= "llvm.ppc.altivec.vminub";
168 intrinsic
= "llvm.ppc.altivec.vminsb";
170 } else if (type
.width
== 16) {
172 intrinsic
= "llvm.ppc.altivec.vminuh";
174 intrinsic
= "llvm.ppc.altivec.vminsh";
176 } else if (type
.width
== 32) {
178 intrinsic
= "llvm.ppc.altivec.vminuw";
180 intrinsic
= "llvm.ppc.altivec.vminsw";
186 /* We need to handle nan's for floating point numbers. If one of the
187 * inputs is nan the other should be returned (required by both D3D10+
189 * The sse intrinsics return the second operator in case of nan by
190 * default so we need to special code to handle those.
192 if (util_cpu_caps
.has_sse
&& type
.floating
&&
193 nan_behavior
!= GALLIVM_NAN_BEHAVIOR_UNDEFINED
&&
194 nan_behavior
!= GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
) {
195 LLVMValueRef isnan
, max
;
196 max
= lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
199 if (nan_behavior
== GALLIVM_NAN_RETURN_OTHER
) {
200 isnan
= lp_build_isnan(bld
, b
);
201 return lp_build_select(bld
, isnan
, a
, max
);
203 assert(nan_behavior
== GALLIVM_NAN_RETURN_NAN
);
204 isnan
= lp_build_isnan(bld
, a
);
205 return lp_build_select(bld
, isnan
, a
, max
);
208 return lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
215 switch (nan_behavior
) {
216 case GALLIVM_NAN_RETURN_NAN
: {
217 LLVMValueRef isnan
= lp_build_isnan(bld
, b
);
218 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
219 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
220 return lp_build_select(bld
, cond
, a
, b
);
223 case GALLIVM_NAN_RETURN_OTHER
: {
224 LLVMValueRef isnan
= lp_build_isnan(bld
, a
);
225 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
226 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
227 return lp_build_select(bld
, cond
, a
, b
);
230 case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
:
231 cond
= lp_build_cmp_ordered(bld
, PIPE_FUNC_LESS
, a
, b
);
232 return lp_build_select(bld
, cond
, a
, b
);
233 case GALLIVM_NAN_BEHAVIOR_UNDEFINED
:
234 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
235 return lp_build_select(bld
, cond
, a
, b
);
239 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
240 return lp_build_select(bld
, cond
, a
, b
);
243 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
244 return lp_build_select(bld
, cond
, a
, b
);
251 * No checks for special case values of a or b = 1 or 0 are done.
252 * NaN's are handled according to the behavior specified by the
253 * nan_behavior argument.
256 lp_build_max_simple(struct lp_build_context
*bld
,
259 enum gallivm_nan_behavior nan_behavior
)
261 const struct lp_type type
= bld
->type
;
262 const char *intrinsic
= NULL
;
263 unsigned intr_size
= 0;
266 assert(lp_check_value(type
, a
));
267 assert(lp_check_value(type
, b
));
269 /* TODO: optimize the constant case */
271 if (type
.floating
&& util_cpu_caps
.has_sse
) {
272 if (type
.width
== 32) {
273 if (type
.length
== 1) {
274 intrinsic
= "llvm.x86.sse.max.ss";
277 else if (type
.length
<= 4 || !util_cpu_caps
.has_avx
) {
278 intrinsic
= "llvm.x86.sse.max.ps";
282 intrinsic
= "llvm.x86.avx.max.ps.256";
286 if (type
.width
== 64 && util_cpu_caps
.has_sse2
) {
287 if (type
.length
== 1) {
288 intrinsic
= "llvm.x86.sse2.max.sd";
291 else if (type
.length
== 2 || !util_cpu_caps
.has_avx
) {
292 intrinsic
= "llvm.x86.sse2.max.pd";
296 intrinsic
= "llvm.x86.avx.max.pd.256";
301 else if (type
.floating
&& util_cpu_caps
.has_altivec
) {
302 if (nan_behavior
== GALLIVM_NAN_RETURN_NAN
) {
303 debug_printf("%s: altivec doesn't support nan return nan behavior\n",
306 if (type
.width
== 32 || type
.length
== 4) {
307 intrinsic
= "llvm.ppc.altivec.vmaxfp";
310 } else if (util_cpu_caps
.has_sse2
&& type
.length
>= 2) {
312 if ((type
.width
== 8 || type
.width
== 16) &&
313 (type
.width
* type
.length
<= 64) &&
314 (gallivm_debug
& GALLIVM_DEBUG_PERF
)) {
315 debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
318 if (type
.width
== 8 && !type
.sign
) {
319 intrinsic
= "llvm.x86.sse2.pmaxu.b";
322 else if (type
.width
== 16 && type
.sign
) {
323 intrinsic
= "llvm.x86.sse2.pmaxs.w";
325 if (util_cpu_caps
.has_sse4_1
) {
326 if (type
.width
== 8 && type
.sign
) {
327 intrinsic
= "llvm.x86.sse41.pmaxsb";
329 if (type
.width
== 16 && !type
.sign
) {
330 intrinsic
= "llvm.x86.sse41.pmaxuw";
332 if (type
.width
== 32 && !type
.sign
) {
333 intrinsic
= "llvm.x86.sse41.pmaxud";
335 if (type
.width
== 32 && type
.sign
) {
336 intrinsic
= "llvm.x86.sse41.pmaxsd";
339 } else if (util_cpu_caps
.has_altivec
) {
341 if (type
.width
== 8) {
343 intrinsic
= "llvm.ppc.altivec.vmaxub";
345 intrinsic
= "llvm.ppc.altivec.vmaxsb";
347 } else if (type
.width
== 16) {
349 intrinsic
= "llvm.ppc.altivec.vmaxuh";
351 intrinsic
= "llvm.ppc.altivec.vmaxsh";
353 } else if (type
.width
== 32) {
355 intrinsic
= "llvm.ppc.altivec.vmaxuw";
357 intrinsic
= "llvm.ppc.altivec.vmaxsw";
363 if (util_cpu_caps
.has_sse
&& type
.floating
&&
364 nan_behavior
!= GALLIVM_NAN_BEHAVIOR_UNDEFINED
&&
365 nan_behavior
!= GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
) {
366 LLVMValueRef isnan
, min
;
367 min
= lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
370 if (nan_behavior
== GALLIVM_NAN_RETURN_OTHER
) {
371 isnan
= lp_build_isnan(bld
, b
);
372 return lp_build_select(bld
, isnan
, a
, min
);
374 assert(nan_behavior
== GALLIVM_NAN_RETURN_NAN
);
375 isnan
= lp_build_isnan(bld
, a
);
376 return lp_build_select(bld
, isnan
, a
, min
);
379 return lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
386 switch (nan_behavior
) {
387 case GALLIVM_NAN_RETURN_NAN
: {
388 LLVMValueRef isnan
= lp_build_isnan(bld
, b
);
389 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
390 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
391 return lp_build_select(bld
, cond
, a
, b
);
394 case GALLIVM_NAN_RETURN_OTHER
: {
395 LLVMValueRef isnan
= lp_build_isnan(bld
, a
);
396 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
397 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
398 return lp_build_select(bld
, cond
, a
, b
);
401 case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
:
402 cond
= lp_build_cmp_ordered(bld
, PIPE_FUNC_GREATER
, a
, b
);
403 return lp_build_select(bld
, cond
, a
, b
);
404 case GALLIVM_NAN_BEHAVIOR_UNDEFINED
:
405 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
406 return lp_build_select(bld
, cond
, a
, b
);
410 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
411 return lp_build_select(bld
, cond
, a
, b
);
414 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
415 return lp_build_select(bld
, cond
, a
, b
);
421 * Generate 1 - a, or ~a depending on bld->type.
424 lp_build_comp(struct lp_build_context
*bld
,
427 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
428 const struct lp_type type
= bld
->type
;
430 assert(lp_check_value(type
, a
));
437 if(type
.norm
&& !type
.floating
&& !type
.fixed
&& !type
.sign
) {
438 if(LLVMIsConstant(a
))
439 return LLVMConstNot(a
);
441 return LLVMBuildNot(builder
, a
, "");
444 if(LLVMIsConstant(a
))
446 return LLVMConstFSub(bld
->one
, a
);
448 return LLVMConstSub(bld
->one
, a
);
451 return LLVMBuildFSub(builder
, bld
->one
, a
, "");
453 return LLVMBuildSub(builder
, bld
->one
, a
, "");
461 lp_build_add(struct lp_build_context
*bld
,
465 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
466 const struct lp_type type
= bld
->type
;
469 assert(lp_check_value(type
, a
));
470 assert(lp_check_value(type
, b
));
476 if(a
== bld
->undef
|| b
== bld
->undef
)
480 const char *intrinsic
= NULL
;
482 if(a
== bld
->one
|| b
== bld
->one
)
485 if (type
.width
* type
.length
== 128 &&
486 !type
.floating
&& !type
.fixed
) {
487 if(util_cpu_caps
.has_sse2
) {
489 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
491 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
492 } else if (util_cpu_caps
.has_altivec
) {
494 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddsbs" : "llvm.ppc.altivec.vaddubs";
496 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddshs" : "llvm.ppc.altivec.vadduhs";
501 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
504 /* TODO: handle signed case */
505 if(type
.norm
&& !type
.floating
&& !type
.fixed
&& !type
.sign
)
506 a
= lp_build_min_simple(bld
, a
, lp_build_comp(bld
, b
), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
508 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
510 res
= LLVMConstFAdd(a
, b
);
512 res
= LLVMConstAdd(a
, b
);
515 res
= LLVMBuildFAdd(builder
, a
, b
, "");
517 res
= LLVMBuildAdd(builder
, a
, b
, "");
519 /* clamp to ceiling of 1.0 */
520 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
521 res
= lp_build_min_simple(bld
, res
, bld
->one
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
523 /* XXX clamp to floor of -1 or 0??? */
529 /** Return the scalar sum of the elements of a.
530 * Should avoid this operation whenever possible.
533 lp_build_horizontal_add(struct lp_build_context
*bld
,
536 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
537 const struct lp_type type
= bld
->type
;
538 LLVMValueRef index
, res
;
540 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
/ 2];
541 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
/ 2];
542 LLVMValueRef vecres
, elem2
;
544 assert(lp_check_value(type
, a
));
546 if (type
.length
== 1) {
550 assert(!bld
->type
.norm
);
553 * for byte vectors can do much better with psadbw.
554 * Using repeated shuffle/adds here. Note with multiple vectors
555 * this can be done more efficiently as outlined in the intel
556 * optimization manual.
557 * Note: could cause data rearrangement if used with smaller element
562 length
= type
.length
/ 2;
564 LLVMValueRef vec1
, vec2
;
565 for (i
= 0; i
< length
; i
++) {
566 shuffles1
[i
] = lp_build_const_int32(bld
->gallivm
, i
);
567 shuffles2
[i
] = lp_build_const_int32(bld
->gallivm
, i
+ length
);
569 vec1
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
570 LLVMConstVector(shuffles1
, length
), "");
571 vec2
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
572 LLVMConstVector(shuffles2
, length
), "");
574 vecres
= LLVMBuildFAdd(builder
, vec1
, vec2
, "");
577 vecres
= LLVMBuildAdd(builder
, vec1
, vec2
, "");
579 length
= length
>> 1;
582 /* always have vector of size 2 here */
585 index
= lp_build_const_int32(bld
->gallivm
, 0);
586 res
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
587 index
= lp_build_const_int32(bld
->gallivm
, 1);
588 elem2
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
591 res
= LLVMBuildFAdd(builder
, res
, elem2
, "");
593 res
= LLVMBuildAdd(builder
, res
, elem2
, "");
599 * Return the horizontal sums of 4 float vectors as a float4 vector.
600 * This uses the technique as outlined in Intel Optimization Manual.
603 lp_build_horizontal_add4x4f(struct lp_build_context
*bld
,
606 struct gallivm_state
*gallivm
= bld
->gallivm
;
607 LLVMBuilderRef builder
= gallivm
->builder
;
608 LLVMValueRef shuffles
[4];
610 LLVMValueRef sumtmp
[2], shuftmp
[2];
612 /* lower half of regs */
613 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
614 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
615 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
616 shuffles
[3] = lp_build_const_int32(gallivm
, 5);
617 tmp
[0] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
618 LLVMConstVector(shuffles
, 4), "");
619 tmp
[2] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
620 LLVMConstVector(shuffles
, 4), "");
622 /* upper half of regs */
623 shuffles
[0] = lp_build_const_int32(gallivm
, 2);
624 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
625 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
626 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
627 tmp
[1] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
628 LLVMConstVector(shuffles
, 4), "");
629 tmp
[3] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
630 LLVMConstVector(shuffles
, 4), "");
632 sumtmp
[0] = LLVMBuildFAdd(builder
, tmp
[0], tmp
[1], "");
633 sumtmp
[1] = LLVMBuildFAdd(builder
, tmp
[2], tmp
[3], "");
635 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
636 shuffles
[1] = lp_build_const_int32(gallivm
, 2);
637 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
638 shuffles
[3] = lp_build_const_int32(gallivm
, 6);
639 shuftmp
[0] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
640 LLVMConstVector(shuffles
, 4), "");
642 shuffles
[0] = lp_build_const_int32(gallivm
, 1);
643 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
644 shuffles
[2] = lp_build_const_int32(gallivm
, 5);
645 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
646 shuftmp
[1] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
647 LLVMConstVector(shuffles
, 4), "");
649 return LLVMBuildFAdd(builder
, shuftmp
[0], shuftmp
[1], "");
654 * partially horizontally add 2-4 float vectors with length nx4,
655 * i.e. only four adjacent values in each vector will be added,
656 * assuming values are really grouped in 4 which also determines
659 * Return a vector of the same length as the initial vectors,
660 * with the excess elements (if any) being undefined.
661 * The element order is independent of number of input vectors.
662 * For 3 vectors x0x1x2x3x4x5x6x7, y0y1y2y3y4y5y6y7, z0z1z2z3z4z5z6z7
663 * the output order thus will be
664 * sumx0-x3,sumy0-y3,sumz0-z3,undef,sumx4-x7,sumy4-y7,sumz4z7,undef
667 lp_build_hadd_partial4(struct lp_build_context
*bld
,
668 LLVMValueRef vectors
[],
671 struct gallivm_state
*gallivm
= bld
->gallivm
;
672 LLVMBuilderRef builder
= gallivm
->builder
;
673 LLVMValueRef ret_vec
;
675 const char *intrinsic
= NULL
;
677 assert(num_vecs
>= 2 && num_vecs
<= 4);
678 assert(bld
->type
.floating
);
680 /* only use this with at least 2 vectors, as it is sort of expensive
681 * (depending on cpu) and we always need two horizontal adds anyway,
682 * so a shuffle/add approach might be better.
688 tmp
[2] = num_vecs
> 2 ? vectors
[2] : vectors
[0];
689 tmp
[3] = num_vecs
> 3 ? vectors
[3] : vectors
[0];
691 if (util_cpu_caps
.has_sse3
&& bld
->type
.width
== 32 &&
692 bld
->type
.length
== 4) {
693 intrinsic
= "llvm.x86.sse3.hadd.ps";
695 else if (util_cpu_caps
.has_avx
&& bld
->type
.width
== 32 &&
696 bld
->type
.length
== 8) {
697 intrinsic
= "llvm.x86.avx.hadd.ps.256";
700 tmp
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
701 lp_build_vec_type(gallivm
, bld
->type
),
704 tmp
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
705 lp_build_vec_type(gallivm
, bld
->type
),
711 return lp_build_intrinsic_binary(builder
, intrinsic
,
712 lp_build_vec_type(gallivm
, bld
->type
),
716 if (bld
->type
.length
== 4) {
717 ret_vec
= lp_build_horizontal_add4x4f(bld
, tmp
);
720 LLVMValueRef partres
[LP_MAX_VECTOR_LENGTH
/4];
722 unsigned num_iter
= bld
->type
.length
/ 4;
723 struct lp_type parttype
= bld
->type
;
725 for (j
= 0; j
< num_iter
; j
++) {
726 LLVMValueRef partsrc
[4];
728 for (i
= 0; i
< 4; i
++) {
729 partsrc
[i
] = lp_build_extract_range(gallivm
, tmp
[i
], j
*4, 4);
731 partres
[j
] = lp_build_horizontal_add4x4f(bld
, partsrc
);
733 ret_vec
= lp_build_concat(gallivm
, partres
, parttype
, num_iter
);
742 lp_build_sub(struct lp_build_context
*bld
,
746 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
747 const struct lp_type type
= bld
->type
;
750 assert(lp_check_value(type
, a
));
751 assert(lp_check_value(type
, b
));
755 if(a
== bld
->undef
|| b
== bld
->undef
)
761 const char *intrinsic
= NULL
;
766 if (type
.width
* type
.length
== 128 &&
767 !type
.floating
&& !type
.fixed
) {
768 if (util_cpu_caps
.has_sse2
) {
770 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
772 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
773 } else if (util_cpu_caps
.has_altivec
) {
775 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubsbs" : "llvm.ppc.altivec.vsububs";
777 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubshs" : "llvm.ppc.altivec.vsubuhs";
782 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
785 /* TODO: handle signed case */
786 if(type
.norm
&& !type
.floating
&& !type
.fixed
&& !type
.sign
)
787 a
= lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
789 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
791 res
= LLVMConstFSub(a
, b
);
793 res
= LLVMConstSub(a
, b
);
796 res
= LLVMBuildFSub(builder
, a
, b
, "");
798 res
= LLVMBuildSub(builder
, a
, b
, "");
800 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
801 res
= lp_build_max_simple(bld
, res
, bld
->zero
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
809 * Normalized multiplication.
811 * There are several approaches for (using 8-bit normalized multiplication as
816 * makes the following approximation to the division (Sree)
818 * a*b/255 ~= (a*(b + 1)) >> 256
820 * which is the fastest method that satisfies the following OpenGL criteria of
822 * 0*0 = 0 and 255*255 = 255
826 * takes the geometric series approximation to the division
828 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
830 * in this case just the first two terms to fit in 16bit arithmetic
832 * t/255 ~= (t + (t >> 8)) >> 8
834 * note that just by itself it doesn't satisfies the OpenGL criteria, as
835 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
838 * - geometric series plus rounding
840 * when using a geometric series division instead of truncating the result
841 * use roundoff in the approximation (Jim Blinn)
843 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
845 * achieving the exact results.
849 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
850 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
851 * @sa Michael Herf, The "double blend trick", May 2000,
852 * http://www.stereopsis.com/doubleblend.html
855 lp_build_mul_norm(struct gallivm_state
*gallivm
,
856 struct lp_type wide_type
,
857 LLVMValueRef a
, LLVMValueRef b
)
859 LLVMBuilderRef builder
= gallivm
->builder
;
860 struct lp_build_context bld
;
865 assert(!wide_type
.floating
);
866 assert(lp_check_value(wide_type
, a
));
867 assert(lp_check_value(wide_type
, b
));
869 lp_build_context_init(&bld
, gallivm
, wide_type
);
871 n
= wide_type
.width
/ 2;
872 if (wide_type
.sign
) {
877 * TODO: for 16bits normalized SSE2 vectors we could consider using PMULHUW
878 * http://ssp.impulsetrain.com/2011/07/03/multiplying-normalized-16-bit-numbers-with-sse2/
882 * a*b / (2**n - 1) ~= (a*b + (a*b >> n) + half) >> n
885 ab
= LLVMBuildMul(builder
, a
, b
, "");
886 ab
= LLVMBuildAdd(builder
, ab
, lp_build_shr_imm(&bld
, ab
, n
), "");
889 * half = sgn(ab) * 0.5 * (2 ** n) = sgn(ab) * (1 << (n - 1))
892 half
= lp_build_const_int_vec(gallivm
, wide_type
, 1 << (n
- 1));
893 if (wide_type
.sign
) {
894 LLVMValueRef minus_half
= LLVMBuildNeg(builder
, half
, "");
895 LLVMValueRef sign
= lp_build_shr_imm(&bld
, ab
, wide_type
.width
- 1);
896 half
= lp_build_select(&bld
, sign
, minus_half
, half
);
898 ab
= LLVMBuildAdd(builder
, ab
, half
, "");
901 ab
= lp_build_shr_imm(&bld
, ab
, n
);
910 lp_build_mul(struct lp_build_context
*bld
,
914 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
915 const struct lp_type type
= bld
->type
;
919 assert(lp_check_value(type
, a
));
920 assert(lp_check_value(type
, b
));
930 if(a
== bld
->undef
|| b
== bld
->undef
)
933 if (!type
.floating
&& !type
.fixed
&& type
.norm
) {
934 struct lp_type wide_type
= lp_wider_type(type
);
935 LLVMValueRef al
, ah
, bl
, bh
, abl
, abh
, ab
;
937 lp_build_unpack2(bld
->gallivm
, type
, wide_type
, a
, &al
, &ah
);
938 lp_build_unpack2(bld
->gallivm
, type
, wide_type
, b
, &bl
, &bh
);
940 /* PMULLW, PSRLW, PADDW */
941 abl
= lp_build_mul_norm(bld
->gallivm
, wide_type
, al
, bl
);
942 abh
= lp_build_mul_norm(bld
->gallivm
, wide_type
, ah
, bh
);
944 ab
= lp_build_pack2(bld
->gallivm
, wide_type
, type
, abl
, abh
);
950 shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
/2);
954 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
956 res
= LLVMConstFMul(a
, b
);
958 res
= LLVMConstMul(a
, b
);
961 res
= LLVMConstAShr(res
, shift
);
963 res
= LLVMConstLShr(res
, shift
);
968 res
= LLVMBuildFMul(builder
, a
, b
, "");
970 res
= LLVMBuildMul(builder
, a
, b
, "");
973 res
= LLVMBuildAShr(builder
, res
, shift
, "");
975 res
= LLVMBuildLShr(builder
, res
, shift
, "");
984 * Small vector x scale multiplication optimization.
987 lp_build_mul_imm(struct lp_build_context
*bld
,
991 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
994 assert(lp_check_value(bld
->type
, a
));
1003 return lp_build_negate(bld
, a
);
1005 if(b
== 2 && bld
->type
.floating
)
1006 return lp_build_add(bld
, a
, a
);
1008 if(util_is_power_of_two(b
)) {
1009 unsigned shift
= ffs(b
) - 1;
1011 if(bld
->type
.floating
) {
1014 * Power of two multiplication by directly manipulating the exponent.
1016 * XXX: This might not be always faster, it will introduce a small error
1017 * for multiplication by zero, and it will produce wrong results
1020 unsigned mantissa
= lp_mantissa(bld
->type
);
1021 factor
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (unsigned long long)shift
<< mantissa
);
1022 a
= LLVMBuildBitCast(builder
, a
, lp_build_int_vec_type(bld
->type
), "");
1023 a
= LLVMBuildAdd(builder
, a
, factor
, "");
1024 a
= LLVMBuildBitCast(builder
, a
, lp_build_vec_type(bld
->gallivm
, bld
->type
), "");
1029 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, shift
);
1030 return LLVMBuildShl(builder
, a
, factor
, "");
1034 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, (double)b
);
1035 return lp_build_mul(bld
, a
, factor
);
1043 lp_build_div(struct lp_build_context
*bld
,
1047 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1048 const struct lp_type type
= bld
->type
;
1050 assert(lp_check_value(type
, a
));
1051 assert(lp_check_value(type
, b
));
1056 return lp_build_rcp(bld
, b
);
1061 if(a
== bld
->undef
|| b
== bld
->undef
)
1064 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
1066 return LLVMConstFDiv(a
, b
);
1068 return LLVMConstSDiv(a
, b
);
1070 return LLVMConstUDiv(a
, b
);
1073 if(((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
1074 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) &&
1076 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
1079 return LLVMBuildFDiv(builder
, a
, b
, "");
1081 return LLVMBuildSDiv(builder
, a
, b
, "");
1083 return LLVMBuildUDiv(builder
, a
, b
, "");
1088 * Linear interpolation helper.
1090 * @param normalized whether we are interpolating normalized values,
1091 * encoded in normalized integers, twice as wide.
1093 * @sa http://www.stereopsis.com/doubleblend.html
1095 static INLINE LLVMValueRef
1096 lp_build_lerp_simple(struct lp_build_context
*bld
,
1102 unsigned half_width
= bld
->type
.width
/2;
1103 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1107 assert(lp_check_value(bld
->type
, x
));
1108 assert(lp_check_value(bld
->type
, v0
));
1109 assert(lp_check_value(bld
->type
, v1
));
1111 delta
= lp_build_sub(bld
, v1
, v0
);
1113 if (flags
& LP_BLD_LERP_WIDE_NORMALIZED
) {
1114 if (!bld
->type
.sign
) {
1115 if (!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
)) {
1117 * Scale x from [0, 2**n - 1] to [0, 2**n] by adding the
1118 * most-significant-bit to the lowest-significant-bit, so that
1119 * later we can just divide by 2**n instead of 2**n - 1.
1122 x
= lp_build_add(bld
, x
, lp_build_shr_imm(bld
, x
, half_width
- 1));
1125 /* (x * delta) >> n */
1126 res
= lp_build_mul(bld
, x
, delta
);
1127 res
= lp_build_shr_imm(bld
, res
, half_width
);
1130 * The rescaling trick above doesn't work for signed numbers, so
1131 * use the 2**n - 1 divison approximation in lp_build_mul_norm
1134 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1135 res
= lp_build_mul_norm(bld
->gallivm
, bld
->type
, x
, delta
);
1138 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1139 res
= lp_build_mul(bld
, x
, delta
);
1142 res
= lp_build_add(bld
, v0
, res
);
1144 if (((flags
& LP_BLD_LERP_WIDE_NORMALIZED
) && !bld
->type
.sign
) ||
1146 /* We need to mask out the high order bits when lerping 8bit normalized colors stored on 16bits */
1147 /* XXX: This step is necessary for lerping 8bit colors stored on 16bits,
1148 * but it will be wrong for true fixed point use cases. Basically we need
1149 * a more powerful lp_type, capable of further distinguishing the values
1150 * interpretation from the value storage. */
1151 res
= LLVMBuildAnd(builder
, res
, lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (1 << half_width
) - 1), "");
1159 * Linear interpolation.
1162 lp_build_lerp(struct lp_build_context
*bld
,
1168 const struct lp_type type
= bld
->type
;
1171 assert(lp_check_value(type
, x
));
1172 assert(lp_check_value(type
, v0
));
1173 assert(lp_check_value(type
, v1
));
1175 assert(!(flags
& LP_BLD_LERP_WIDE_NORMALIZED
));
1178 struct lp_type wide_type
;
1179 struct lp_build_context wide_bld
;
1180 LLVMValueRef xl
, xh
, v0l
, v0h
, v1l
, v1h
, resl
, resh
;
1182 assert(type
.length
>= 2);
1185 * Create a wider integer type, enough to hold the
1186 * intermediate result of the multiplication.
1188 memset(&wide_type
, 0, sizeof wide_type
);
1189 wide_type
.sign
= type
.sign
;
1190 wide_type
.width
= type
.width
*2;
1191 wide_type
.length
= type
.length
/2;
1193 lp_build_context_init(&wide_bld
, bld
->gallivm
, wide_type
);
1195 lp_build_unpack2(bld
->gallivm
, type
, wide_type
, x
, &xl
, &xh
);
1196 lp_build_unpack2(bld
->gallivm
, type
, wide_type
, v0
, &v0l
, &v0h
);
1197 lp_build_unpack2(bld
->gallivm
, type
, wide_type
, v1
, &v1l
, &v1h
);
1203 flags
|= LP_BLD_LERP_WIDE_NORMALIZED
;
1205 resl
= lp_build_lerp_simple(&wide_bld
, xl
, v0l
, v1l
, flags
);
1206 resh
= lp_build_lerp_simple(&wide_bld
, xh
, v0h
, v1h
, flags
);
1208 res
= lp_build_pack2(bld
->gallivm
, wide_type
, type
, resl
, resh
);
1210 res
= lp_build_lerp_simple(bld
, x
, v0
, v1
, flags
);
1218 * Bilinear interpolation.
1220 * Values indices are in v_{yx}.
1223 lp_build_lerp_2d(struct lp_build_context
*bld
,
1232 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
, flags
);
1233 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
, flags
);
1234 return lp_build_lerp(bld
, y
, v0
, v1
, flags
);
1239 lp_build_lerp_3d(struct lp_build_context
*bld
,
1253 LLVMValueRef v0
= lp_build_lerp_2d(bld
, x
, y
, v000
, v001
, v010
, v011
, flags
);
1254 LLVMValueRef v1
= lp_build_lerp_2d(bld
, x
, y
, v100
, v101
, v110
, v111
, flags
);
1255 return lp_build_lerp(bld
, z
, v0
, v1
, flags
);
1260 * Generate min(a, b)
1261 * Do checks for special cases but not for nans.
1264 lp_build_min(struct lp_build_context
*bld
,
1268 assert(lp_check_value(bld
->type
, a
));
1269 assert(lp_check_value(bld
->type
, b
));
1271 if(a
== bld
->undef
|| b
== bld
->undef
)
1277 if (bld
->type
.norm
) {
1278 if (!bld
->type
.sign
) {
1279 if (a
== bld
->zero
|| b
== bld
->zero
) {
1289 return lp_build_min_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1294 * Generate min(a, b)
1295 * NaN's are handled according to the behavior specified by the
1296 * nan_behavior argument.
1299 lp_build_min_ext(struct lp_build_context
*bld
,
1302 enum gallivm_nan_behavior nan_behavior
)
1304 assert(lp_check_value(bld
->type
, a
));
1305 assert(lp_check_value(bld
->type
, b
));
1307 if(a
== bld
->undef
|| b
== bld
->undef
)
1313 if (bld
->type
.norm
) {
1314 if (!bld
->type
.sign
) {
1315 if (a
== bld
->zero
|| b
== bld
->zero
) {
1325 return lp_build_min_simple(bld
, a
, b
, nan_behavior
);
1329 * Generate max(a, b)
1330 * Do checks for special cases, but NaN behavior is undefined.
1333 lp_build_max(struct lp_build_context
*bld
,
1337 assert(lp_check_value(bld
->type
, a
));
1338 assert(lp_check_value(bld
->type
, b
));
1340 if(a
== bld
->undef
|| b
== bld
->undef
)
1346 if(bld
->type
.norm
) {
1347 if(a
== bld
->one
|| b
== bld
->one
)
1349 if (!bld
->type
.sign
) {
1350 if (a
== bld
->zero
) {
1353 if (b
== bld
->zero
) {
1359 return lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1364 * Generate max(a, b)
1365 * Checks for special cases.
1366 * NaN's are handled according to the behavior specified by the
1367 * nan_behavior argument.
1370 lp_build_max_ext(struct lp_build_context
*bld
,
1373 enum gallivm_nan_behavior nan_behavior
)
1375 assert(lp_check_value(bld
->type
, a
));
1376 assert(lp_check_value(bld
->type
, b
));
1378 if(a
== bld
->undef
|| b
== bld
->undef
)
1384 if(bld
->type
.norm
) {
1385 if(a
== bld
->one
|| b
== bld
->one
)
1387 if (!bld
->type
.sign
) {
1388 if (a
== bld
->zero
) {
1391 if (b
== bld
->zero
) {
1397 return lp_build_max_simple(bld
, a
, b
, nan_behavior
);
1401 * Generate clamp(a, min, max)
1402 * NaN behavior (for any of a, min, max) is undefined.
1403 * Do checks for special cases.
1406 lp_build_clamp(struct lp_build_context
*bld
,
1411 assert(lp_check_value(bld
->type
, a
));
1412 assert(lp_check_value(bld
->type
, min
));
1413 assert(lp_check_value(bld
->type
, max
));
1415 a
= lp_build_min(bld
, a
, max
);
1416 a
= lp_build_max(bld
, a
, min
);
1422 * Generate clamp(a, 0, 1)
1423 * A NaN will get converted to zero.
1426 lp_build_clamp_zero_one_nanzero(struct lp_build_context
*bld
,
1429 a
= lp_build_max_ext(bld
, a
, bld
->zero
, GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1430 a
= lp_build_min(bld
, a
, bld
->one
);
1439 lp_build_abs(struct lp_build_context
*bld
,
1442 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1443 const struct lp_type type
= bld
->type
;
1444 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1446 assert(lp_check_value(type
, a
));
1452 /* Mask out the sign bit */
1453 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1454 unsigned long long absMask
= ~(1ULL << (type
.width
- 1));
1455 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
, ((unsigned long long) absMask
));
1456 a
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1457 a
= LLVMBuildAnd(builder
, a
, mask
, "");
1458 a
= LLVMBuildBitCast(builder
, a
, vec_type
, "");
1462 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
) {
1463 switch(type
.width
) {
1465 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
1467 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
1469 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
1472 else if (type
.width
*type
.length
== 256 && util_cpu_caps
.has_ssse3
&&
1473 (gallivm_debug
& GALLIVM_DEBUG_PERF
) &&
1474 (type
.width
== 8 || type
.width
== 16 || type
.width
== 32)) {
1475 debug_printf("%s: inefficient code, should split vectors manually\n",
1479 return lp_build_max(bld
, a
, LLVMBuildNeg(builder
, a
, ""));
1484 lp_build_negate(struct lp_build_context
*bld
,
1487 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1489 assert(lp_check_value(bld
->type
, a
));
1491 #if HAVE_LLVM >= 0x0207
1492 if (bld
->type
.floating
)
1493 a
= LLVMBuildFNeg(builder
, a
, "");
1496 a
= LLVMBuildNeg(builder
, a
, "");
1502 /** Return -1, 0 or +1 depending on the sign of a */
1504 lp_build_sgn(struct lp_build_context
*bld
,
1507 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1508 const struct lp_type type
= bld
->type
;
1512 assert(lp_check_value(type
, a
));
1514 /* Handle non-zero case */
1516 /* if not zero then sign must be positive */
1519 else if(type
.floating
) {
1520 LLVMTypeRef vec_type
;
1521 LLVMTypeRef int_type
;
1525 unsigned long long maskBit
= (unsigned long long)1 << (type
.width
- 1);
1527 int_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1528 vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1529 mask
= lp_build_const_int_vec(bld
->gallivm
, type
, maskBit
);
1531 /* Take the sign bit and add it to 1 constant */
1532 sign
= LLVMBuildBitCast(builder
, a
, int_type
, "");
1533 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
1534 one
= LLVMConstBitCast(bld
->one
, int_type
);
1535 res
= LLVMBuildOr(builder
, sign
, one
, "");
1536 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1540 /* signed int/norm/fixed point */
1541 /* could use psign with sse3 and appropriate vectors here */
1542 LLVMValueRef minus_one
= lp_build_const_vec(bld
->gallivm
, type
, -1.0);
1543 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
1544 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
1548 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
1549 res
= lp_build_select(bld
, cond
, bld
->zero
, res
);
1556 * Set the sign of float vector 'a' according to 'sign'.
1557 * If sign==0, return abs(a).
1558 * If sign==1, return -abs(a);
1559 * Other values for sign produce undefined results.
1562 lp_build_set_sign(struct lp_build_context
*bld
,
1563 LLVMValueRef a
, LLVMValueRef sign
)
1565 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1566 const struct lp_type type
= bld
->type
;
1567 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1568 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1569 LLVMValueRef shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
- 1);
1570 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
1571 ~((unsigned long long) 1 << (type
.width
- 1)));
1572 LLVMValueRef val
, res
;
1574 assert(type
.floating
);
1575 assert(lp_check_value(type
, a
));
1577 /* val = reinterpret_cast<int>(a) */
1578 val
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1579 /* val = val & mask */
1580 val
= LLVMBuildAnd(builder
, val
, mask
, "");
1581 /* sign = sign << shift */
1582 sign
= LLVMBuildShl(builder
, sign
, shift
, "");
1583 /* res = val | sign */
1584 res
= LLVMBuildOr(builder
, val
, sign
, "");
1585 /* res = reinterpret_cast<float>(res) */
1586 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1593 * Convert vector of (or scalar) int to vector of (or scalar) float.
1596 lp_build_int_to_float(struct lp_build_context
*bld
,
1599 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1600 const struct lp_type type
= bld
->type
;
1601 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1603 assert(type
.floating
);
1605 return LLVMBuildSIToFP(builder
, a
, vec_type
, "");
1609 arch_rounding_available(const struct lp_type type
)
1611 if ((util_cpu_caps
.has_sse4_1
&&
1612 (type
.length
== 1 || type
.width
*type
.length
== 128)) ||
1613 (util_cpu_caps
.has_avx
&& type
.width
*type
.length
== 256))
1615 else if ((util_cpu_caps
.has_altivec
&&
1616 (type
.width
== 32 && type
.length
== 4)))
1622 enum lp_build_round_mode
1624 LP_BUILD_ROUND_NEAREST
= 0,
1625 LP_BUILD_ROUND_FLOOR
= 1,
1626 LP_BUILD_ROUND_CEIL
= 2,
1627 LP_BUILD_ROUND_TRUNCATE
= 3
1631 * Helper for SSE4.1's ROUNDxx instructions.
1633 * NOTE: In the SSE4.1's nearest mode, if two values are equally close, the
1634 * result is the even value. That is, rounding 2.5 will be 2.0, and not 3.0.
1636 static INLINE LLVMValueRef
1637 lp_build_round_sse41(struct lp_build_context
*bld
,
1639 enum lp_build_round_mode mode
)
1641 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1642 const struct lp_type type
= bld
->type
;
1643 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
1644 const char *intrinsic
;
1647 assert(type
.floating
);
1649 assert(lp_check_value(type
, a
));
1650 assert(util_cpu_caps
.has_sse4_1
);
1652 if (type
.length
== 1) {
1653 LLVMTypeRef vec_type
;
1655 LLVMValueRef args
[3];
1656 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
1658 switch(type
.width
) {
1660 intrinsic
= "llvm.x86.sse41.round.ss";
1663 intrinsic
= "llvm.x86.sse41.round.sd";
1670 vec_type
= LLVMVectorType(bld
->elem_type
, 4);
1672 undef
= LLVMGetUndef(vec_type
);
1675 args
[1] = LLVMBuildInsertElement(builder
, undef
, a
, index0
, "");
1676 args
[2] = LLVMConstInt(i32t
, mode
, 0);
1678 res
= lp_build_intrinsic(builder
, intrinsic
,
1679 vec_type
, args
, Elements(args
));
1681 res
= LLVMBuildExtractElement(builder
, res
, index0
, "");
1684 if (type
.width
* type
.length
== 128) {
1685 switch(type
.width
) {
1687 intrinsic
= "llvm.x86.sse41.round.ps";
1690 intrinsic
= "llvm.x86.sse41.round.pd";
1698 assert(type
.width
* type
.length
== 256);
1699 assert(util_cpu_caps
.has_avx
);
1701 switch(type
.width
) {
1703 intrinsic
= "llvm.x86.avx.round.ps.256";
1706 intrinsic
= "llvm.x86.avx.round.pd.256";
1714 res
= lp_build_intrinsic_binary(builder
, intrinsic
,
1716 LLVMConstInt(i32t
, mode
, 0));
1723 static INLINE LLVMValueRef
1724 lp_build_iround_nearest_sse2(struct lp_build_context
*bld
,
1727 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1728 const struct lp_type type
= bld
->type
;
1729 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
1730 LLVMTypeRef ret_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1731 const char *intrinsic
;
1734 assert(type
.floating
);
1735 /* using the double precision conversions is a bit more complicated */
1736 assert(type
.width
== 32);
1738 assert(lp_check_value(type
, a
));
1739 assert(util_cpu_caps
.has_sse2
);
1741 /* This is relying on MXCSR rounding mode, which should always be nearest. */
1742 if (type
.length
== 1) {
1743 LLVMTypeRef vec_type
;
1746 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
1748 vec_type
= LLVMVectorType(bld
->elem_type
, 4);
1750 intrinsic
= "llvm.x86.sse.cvtss2si";
1752 undef
= LLVMGetUndef(vec_type
);
1754 arg
= LLVMBuildInsertElement(builder
, undef
, a
, index0
, "");
1756 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
1760 if (type
.width
* type
.length
== 128) {
1761 intrinsic
= "llvm.x86.sse2.cvtps2dq";
1764 assert(type
.width
*type
.length
== 256);
1765 assert(util_cpu_caps
.has_avx
);
1767 intrinsic
= "llvm.x86.avx.cvt.ps2dq.256";
1769 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
1779 static INLINE LLVMValueRef
1780 lp_build_round_altivec(struct lp_build_context
*bld
,
1782 enum lp_build_round_mode mode
)
1784 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1785 const struct lp_type type
= bld
->type
;
1786 const char *intrinsic
= NULL
;
1788 assert(type
.floating
);
1790 assert(lp_check_value(type
, a
));
1791 assert(util_cpu_caps
.has_altivec
);
1794 case LP_BUILD_ROUND_NEAREST
:
1795 intrinsic
= "llvm.ppc.altivec.vrfin";
1797 case LP_BUILD_ROUND_FLOOR
:
1798 intrinsic
= "llvm.ppc.altivec.vrfim";
1800 case LP_BUILD_ROUND_CEIL
:
1801 intrinsic
= "llvm.ppc.altivec.vrfip";
1803 case LP_BUILD_ROUND_TRUNCATE
:
1804 intrinsic
= "llvm.ppc.altivec.vrfiz";
1808 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
1811 static INLINE LLVMValueRef
1812 lp_build_round_arch(struct lp_build_context
*bld
,
1814 enum lp_build_round_mode mode
)
1816 if (util_cpu_caps
.has_sse4_1
)
1817 return lp_build_round_sse41(bld
, a
, mode
);
1818 else /* (util_cpu_caps.has_altivec) */
1819 return lp_build_round_altivec(bld
, a
, mode
);
1823 * Return the integer part of a float (vector) value (== round toward zero).
1824 * The returned value is a float (vector).
1825 * Ex: trunc(-1.5) = -1.0
1828 lp_build_trunc(struct lp_build_context
*bld
,
1831 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1832 const struct lp_type type
= bld
->type
;
1834 assert(type
.floating
);
1835 assert(lp_check_value(type
, a
));
1837 if (arch_rounding_available(type
)) {
1838 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_TRUNCATE
);
1841 const struct lp_type type
= bld
->type
;
1842 struct lp_type inttype
;
1843 struct lp_build_context intbld
;
1844 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 2^24);
1845 LLVMValueRef trunc
, res
, anosign
, mask
;
1846 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
1847 LLVMTypeRef vec_type
= bld
->vec_type
;
1849 assert(type
.width
== 32); /* might want to handle doubles at some point */
1852 inttype
.floating
= 0;
1853 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
1855 /* round by truncation */
1856 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
1857 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
1859 /* mask out sign bit */
1860 anosign
= lp_build_abs(bld
, a
);
1862 * mask out all values if anosign > 2^24
1863 * This should work both for large ints (all rounding is no-op for them
1864 * because such floats are always exact) as well as special cases like
1865 * NaNs, Infs (taking advantage of the fact they use max exponent).
1866 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
1868 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
1869 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
1870 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
1871 return lp_build_select(bld
, mask
, a
, res
);
1877 * Return float (vector) rounded to nearest integer (vector). The returned
1878 * value is a float (vector).
1879 * Ex: round(0.9) = 1.0
1880 * Ex: round(-1.5) = -2.0
1883 lp_build_round(struct lp_build_context
*bld
,
1886 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1887 const struct lp_type type
= bld
->type
;
1889 assert(type
.floating
);
1890 assert(lp_check_value(type
, a
));
1892 if (arch_rounding_available(type
)) {
1893 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
1896 const struct lp_type type
= bld
->type
;
1897 struct lp_type inttype
;
1898 struct lp_build_context intbld
;
1899 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 2^24);
1900 LLVMValueRef res
, anosign
, mask
;
1901 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
1902 LLVMTypeRef vec_type
= bld
->vec_type
;
1904 assert(type
.width
== 32); /* might want to handle doubles at some point */
1907 inttype
.floating
= 0;
1908 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
1910 res
= lp_build_iround(bld
, a
);
1911 res
= LLVMBuildSIToFP(builder
, res
, vec_type
, "");
1913 /* mask out sign bit */
1914 anosign
= lp_build_abs(bld
, a
);
1916 * mask out all values if anosign > 2^24
1917 * This should work both for large ints (all rounding is no-op for them
1918 * because such floats are always exact) as well as special cases like
1919 * NaNs, Infs (taking advantage of the fact they use max exponent).
1920 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
1922 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
1923 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
1924 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
1925 return lp_build_select(bld
, mask
, a
, res
);
1931 * Return floor of float (vector), result is a float (vector)
1932 * Ex: floor(1.1) = 1.0
1933 * Ex: floor(-1.1) = -2.0
1936 lp_build_floor(struct lp_build_context
*bld
,
1939 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1940 const struct lp_type type
= bld
->type
;
1942 assert(type
.floating
);
1943 assert(lp_check_value(type
, a
));
1945 if (arch_rounding_available(type
)) {
1946 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
1949 const struct lp_type type
= bld
->type
;
1950 struct lp_type inttype
;
1951 struct lp_build_context intbld
;
1952 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 2^24);
1953 LLVMValueRef trunc
, res
, anosign
, mask
;
1954 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
1955 LLVMTypeRef vec_type
= bld
->vec_type
;
1957 assert(type
.width
== 32); /* might want to handle doubles at some point */
1960 inttype
.floating
= 0;
1961 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
1963 /* round by truncation */
1964 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
1965 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
1971 * fix values if rounding is wrong (for non-special cases)
1972 * - this is the case if trunc > a
1974 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, res
, a
);
1975 /* tmp = trunc > a ? 1.0 : 0.0 */
1976 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
1977 tmp
= lp_build_and(&intbld
, mask
, tmp
);
1978 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
1979 res
= lp_build_sub(bld
, res
, tmp
);
1982 /* mask out sign bit */
1983 anosign
= lp_build_abs(bld
, a
);
1985 * mask out all values if anosign > 2^24
1986 * This should work both for large ints (all rounding is no-op for them
1987 * because such floats are always exact) as well as special cases like
1988 * NaNs, Infs (taking advantage of the fact they use max exponent).
1989 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
1991 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
1992 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
1993 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
1994 return lp_build_select(bld
, mask
, a
, res
);
2000 * Return ceiling of float (vector), returning float (vector).
2001 * Ex: ceil( 1.1) = 2.0
2002 * Ex: ceil(-1.1) = -1.0
2005 lp_build_ceil(struct lp_build_context
*bld
,
2008 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2009 const struct lp_type type
= bld
->type
;
2011 assert(type
.floating
);
2012 assert(lp_check_value(type
, a
));
2014 if (arch_rounding_available(type
)) {
2015 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2018 const struct lp_type type
= bld
->type
;
2019 struct lp_type inttype
;
2020 struct lp_build_context intbld
;
2021 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 2^24);
2022 LLVMValueRef trunc
, res
, anosign
, mask
, tmp
;
2023 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2024 LLVMTypeRef vec_type
= bld
->vec_type
;
2026 assert(type
.width
== 32); /* might want to handle doubles at some point */
2029 inttype
.floating
= 0;
2030 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2032 /* round by truncation */
2033 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2034 trunc
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "ceil.trunc");
2037 * fix values if rounding is wrong (for non-special cases)
2038 * - this is the case if trunc < a
2040 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2041 /* tmp = trunc < a ? 1.0 : 0.0 */
2042 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2043 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2044 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2045 res
= lp_build_add(bld
, trunc
, tmp
);
2047 /* mask out sign bit */
2048 anosign
= lp_build_abs(bld
, a
);
2050 * mask out all values if anosign > 2^24
2051 * This should work both for large ints (all rounding is no-op for them
2052 * because such floats are always exact) as well as special cases like
2053 * NaNs, Infs (taking advantage of the fact they use max exponent).
2054 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2056 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2057 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2058 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2059 return lp_build_select(bld
, mask
, a
, res
);
2065 * Return fractional part of 'a' computed as a - floor(a)
2066 * Typically used in texture coord arithmetic.
2069 lp_build_fract(struct lp_build_context
*bld
,
2072 assert(bld
->type
.floating
);
2073 return lp_build_sub(bld
, a
, lp_build_floor(bld
, a
));
2078 * Prevent returning a fractional part of 1.0 for very small negative values of
2079 * 'a' by clamping against 0.99999(9).
2081 static inline LLVMValueRef
2082 clamp_fract(struct lp_build_context
*bld
, LLVMValueRef fract
)
2086 /* this is the largest number smaller than 1.0 representable as float */
2087 max
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2088 1.0 - 1.0/(1LL << (lp_mantissa(bld
->type
) + 1)));
2089 return lp_build_min(bld
, fract
, max
);
2094 * Same as lp_build_fract, but guarantees that the result is always smaller
2098 lp_build_fract_safe(struct lp_build_context
*bld
,
2101 return clamp_fract(bld
, lp_build_fract(bld
, a
));
2106 * Return the integer part of a float (vector) value (== round toward zero).
2107 * The returned value is an integer (vector).
2108 * Ex: itrunc(-1.5) = -1
2111 lp_build_itrunc(struct lp_build_context
*bld
,
2114 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2115 const struct lp_type type
= bld
->type
;
2116 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
2118 assert(type
.floating
);
2119 assert(lp_check_value(type
, a
));
2121 return LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2126 * Return float (vector) rounded to nearest integer (vector). The returned
2127 * value is an integer (vector).
2128 * Ex: iround(0.9) = 1
2129 * Ex: iround(-1.5) = -2
2132 lp_build_iround(struct lp_build_context
*bld
,
2135 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2136 const struct lp_type type
= bld
->type
;
2137 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2140 assert(type
.floating
);
2142 assert(lp_check_value(type
, a
));
2144 if ((util_cpu_caps
.has_sse2
&&
2145 ((type
.width
== 32) && (type
.length
== 1 || type
.length
== 4))) ||
2146 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2147 return lp_build_iround_nearest_sse2(bld
, a
);
2149 if (arch_rounding_available(type
)) {
2150 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2155 half
= lp_build_const_vec(bld
->gallivm
, type
, 0.5);
2158 LLVMTypeRef vec_type
= bld
->vec_type
;
2159 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
2160 (unsigned long long)1 << (type
.width
- 1));
2164 sign
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
2165 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
2168 half
= LLVMBuildBitCast(builder
, half
, int_vec_type
, "");
2169 half
= LLVMBuildOr(builder
, sign
, half
, "");
2170 half
= LLVMBuildBitCast(builder
, half
, vec_type
, "");
2173 res
= LLVMBuildFAdd(builder
, a
, half
, "");
2176 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "");
2183 * Return floor of float (vector), result is an int (vector)
2184 * Ex: ifloor(1.1) = 1.0
2185 * Ex: ifloor(-1.1) = -2.0
2188 lp_build_ifloor(struct lp_build_context
*bld
,
2191 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2192 const struct lp_type type
= bld
->type
;
2193 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2196 assert(type
.floating
);
2197 assert(lp_check_value(type
, a
));
2201 if (arch_rounding_available(type
)) {
2202 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2205 struct lp_type inttype
;
2206 struct lp_build_context intbld
;
2207 LLVMValueRef trunc
, itrunc
, mask
;
2209 assert(type
.floating
);
2210 assert(lp_check_value(type
, a
));
2213 inttype
.floating
= 0;
2214 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2216 /* round by truncation */
2217 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2218 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "ifloor.trunc");
2221 * fix values if rounding is wrong (for non-special cases)
2222 * - this is the case if trunc > a
2223 * The results of doing this with NaNs, very large values etc.
2224 * are undefined but this seems to be the case anyway.
2226 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, trunc
, a
);
2227 /* cheapie minus one with mask since the mask is minus one / zero */
2228 return lp_build_add(&intbld
, itrunc
, mask
);
2232 /* round to nearest (toward zero) */
2233 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "ifloor.res");
2240 * Return ceiling of float (vector), returning int (vector).
2241 * Ex: iceil( 1.1) = 2
2242 * Ex: iceil(-1.1) = -1
2245 lp_build_iceil(struct lp_build_context
*bld
,
2248 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2249 const struct lp_type type
= bld
->type
;
2250 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2253 assert(type
.floating
);
2254 assert(lp_check_value(type
, a
));
2256 if (arch_rounding_available(type
)) {
2257 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2260 struct lp_type inttype
;
2261 struct lp_build_context intbld
;
2262 LLVMValueRef trunc
, itrunc
, mask
;
2264 assert(type
.floating
);
2265 assert(lp_check_value(type
, a
));
2268 inttype
.floating
= 0;
2269 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2271 /* round by truncation */
2272 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2273 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "iceil.trunc");
2276 * fix values if rounding is wrong (for non-special cases)
2277 * - this is the case if trunc < a
2278 * The results of doing this with NaNs, very large values etc.
2279 * are undefined but this seems to be the case anyway.
2281 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2282 /* cheapie plus one with mask since the mask is minus one / zero */
2283 return lp_build_sub(&intbld
, itrunc
, mask
);
2286 /* round to nearest (toward zero) */
2287 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "iceil.res");
2294 * Combined ifloor() & fract().
2296 * Preferred to calling the functions separately, as it will ensure that the
2297 * strategy (floor() vs ifloor()) that results in less redundant work is used.
2300 lp_build_ifloor_fract(struct lp_build_context
*bld
,
2302 LLVMValueRef
*out_ipart
,
2303 LLVMValueRef
*out_fpart
)
2305 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2306 const struct lp_type type
= bld
->type
;
2309 assert(type
.floating
);
2310 assert(lp_check_value(type
, a
));
2312 if (arch_rounding_available(type
)) {
2314 * floor() is easier.
2317 ipart
= lp_build_floor(bld
, a
);
2318 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2319 *out_ipart
= LLVMBuildFPToSI(builder
, ipart
, bld
->int_vec_type
, "ipart");
2323 * ifloor() is easier.
2326 *out_ipart
= lp_build_ifloor(bld
, a
);
2327 ipart
= LLVMBuildSIToFP(builder
, *out_ipart
, bld
->vec_type
, "ipart");
2328 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2334 * Same as lp_build_ifloor_fract, but guarantees that the fractional part is
2335 * always smaller than one.
2338 lp_build_ifloor_fract_safe(struct lp_build_context
*bld
,
2340 LLVMValueRef
*out_ipart
,
2341 LLVMValueRef
*out_fpart
)
2343 lp_build_ifloor_fract(bld
, a
, out_ipart
, out_fpart
);
2344 *out_fpart
= clamp_fract(bld
, *out_fpart
);
2349 lp_build_sqrt(struct lp_build_context
*bld
,
2352 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2353 const struct lp_type type
= bld
->type
;
2354 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
2357 assert(lp_check_value(type
, a
));
2359 /* TODO: optimize the constant case */
2361 assert(type
.floating
);
2362 if (type
.length
== 1) {
2363 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.sqrt.f%u", type
.width
);
2366 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.sqrt.v%uf%u", type
.length
, type
.width
);
2369 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2374 * Do one Newton-Raphson step to improve reciprocate precision:
2376 * x_{i+1} = x_i * (2 - a * x_i)
2378 * XXX: Unfortunately this won't give IEEE-754 conformant results for 0 or
2379 * +/-Inf, giving NaN instead. Certain applications rely on this behavior,
2380 * such as Google Earth, which does RCP(RSQRT(0.0) when drawing the Earth's
2381 * halo. It would be necessary to clamp the argument to prevent this.
2384 * - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division
2385 * - http://softwarecommunity.intel.com/articles/eng/1818.htm
2387 static INLINE LLVMValueRef
2388 lp_build_rcp_refine(struct lp_build_context
*bld
,
2392 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2393 LLVMValueRef two
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 2.0);
2396 res
= LLVMBuildFMul(builder
, a
, rcp_a
, "");
2397 res
= LLVMBuildFSub(builder
, two
, res
, "");
2398 res
= LLVMBuildFMul(builder
, rcp_a
, res
, "");
2405 lp_build_rcp(struct lp_build_context
*bld
,
2408 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2409 const struct lp_type type
= bld
->type
;
2411 assert(lp_check_value(type
, a
));
2420 assert(type
.floating
);
2422 if(LLVMIsConstant(a
))
2423 return LLVMConstFDiv(bld
->one
, a
);
2426 * We don't use RCPPS because:
2427 * - it only has 10bits of precision
2428 * - it doesn't even get the reciprocate of 1.0 exactly
2429 * - doing Newton-Rapshon steps yields wrong (NaN) values for 0.0 or Inf
2430 * - for recent processors the benefit over DIVPS is marginal, a case
2433 * We could still use it on certain processors if benchmarks show that the
2434 * RCPPS plus necessary workarounds are still preferrable to DIVPS; or for
2435 * particular uses that require less workarounds.
2438 if (FALSE
&& ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2439 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8))){
2440 const unsigned num_iterations
= 0;
2443 const char *intrinsic
= NULL
;
2445 if (type
.length
== 4) {
2446 intrinsic
= "llvm.x86.sse.rcp.ps";
2449 intrinsic
= "llvm.x86.avx.rcp.ps.256";
2452 res
= lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2454 for (i
= 0; i
< num_iterations
; ++i
) {
2455 res
= lp_build_rcp_refine(bld
, a
, res
);
2461 return LLVMBuildFDiv(builder
, bld
->one
, a
, "");
2466 * Do one Newton-Raphson step to improve rsqrt precision:
2468 * x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
2470 * See also Intel 64 and IA-32 Architectures Optimization Manual.
2472 static INLINE LLVMValueRef
2473 lp_build_rsqrt_refine(struct lp_build_context
*bld
,
2475 LLVMValueRef rsqrt_a
)
2477 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2478 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 0.5);
2479 LLVMValueRef three
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 3.0);
2482 res
= LLVMBuildFMul(builder
, rsqrt_a
, rsqrt_a
, "");
2483 res
= LLVMBuildFMul(builder
, a
, res
, "");
2484 res
= LLVMBuildFSub(builder
, three
, res
, "");
2485 res
= LLVMBuildFMul(builder
, rsqrt_a
, res
, "");
2486 res
= LLVMBuildFMul(builder
, half
, res
, "");
2493 * Generate 1/sqrt(a).
2494 * Result is undefined for values < 0, infinity for +0.
2497 lp_build_rsqrt(struct lp_build_context
*bld
,
2500 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2501 const struct lp_type type
= bld
->type
;
2503 assert(lp_check_value(type
, a
));
2505 assert(type
.floating
);
2508 * This should be faster but all denormals will end up as infinity.
2510 if (0 && lp_build_fast_rsqrt_available(type
)) {
2511 const unsigned num_iterations
= 1;
2515 /* rsqrt(1.0) != 1.0 here */
2516 res
= lp_build_fast_rsqrt(bld
, a
);
2518 if (num_iterations
) {
2520 * Newton-Raphson will result in NaN instead of infinity for zero,
2521 * and NaN instead of zero for infinity.
2522 * Also, need to ensure rsqrt(1.0) == 1.0.
2523 * All numbers smaller than FLT_MIN will result in +infinity
2524 * (rsqrtps treats all denormals as zero).
2527 * Certain non-c99 compilers don't know INFINITY and might not support
2528 * hacks to evaluate it at compile time neither.
2530 const unsigned posinf_int
= 0x7F800000;
2532 LLVMValueRef flt_min
= lp_build_const_vec(bld
->gallivm
, type
, FLT_MIN
);
2533 LLVMValueRef inf
= lp_build_const_int_vec(bld
->gallivm
, type
, posinf_int
);
2535 inf
= LLVMBuildBitCast(builder
, inf
, lp_build_vec_type(bld
->gallivm
, type
), "");
2537 for (i
= 0; i
< num_iterations
; ++i
) {
2538 res
= lp_build_rsqrt_refine(bld
, a
, res
);
2540 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_LESS
, a
, flt_min
);
2541 res
= lp_build_select(bld
, cmp
, inf
, res
);
2542 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, inf
);
2543 res
= lp_build_select(bld
, cmp
, bld
->zero
, res
);
2544 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, bld
->one
);
2545 res
= lp_build_select(bld
, cmp
, bld
->one
, res
);
2551 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2555 * If there's a fast (inaccurate) rsqrt instruction available
2556 * (caller may want to avoid to call rsqrt_fast if it's not available,
2557 * i.e. for calculating x^0.5 it may do rsqrt_fast(x) * x but if
2558 * unavailable it would result in sqrt/div/mul so obviously
2559 * much better to just call sqrt, skipping both div and mul).
2562 lp_build_fast_rsqrt_available(struct lp_type type
)
2564 assert(type
.floating
);
2566 if ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2567 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2575 * Generate 1/sqrt(a).
2576 * Result is undefined for values < 0, infinity for +0.
2577 * Precision is limited, only ~10 bits guaranteed
2578 * (rsqrt 1.0 may not be 1.0, denorms may be flushed to 0).
2581 lp_build_fast_rsqrt(struct lp_build_context
*bld
,
2584 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2585 const struct lp_type type
= bld
->type
;
2587 assert(lp_check_value(type
, a
));
2589 if (lp_build_fast_rsqrt_available(type
)) {
2590 const char *intrinsic
= NULL
;
2592 if (type
.length
== 4) {
2593 intrinsic
= "llvm.x86.sse.rsqrt.ps";
2596 intrinsic
= "llvm.x86.avx.rsqrt.ps.256";
2598 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2601 debug_printf("%s: emulating fast rsqrt with rcp/sqrt\n", __FUNCTION__
);
2603 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2608 * Generate sin(a) or cos(a) using polynomial approximation.
2609 * TODO: it might be worth recognizing sin and cos using same source
2610 * (i.e. d3d10 sincos opcode). Obviously doing both at the same time
2611 * would be way cheaper than calculating (nearly) everything twice...
2612 * Not sure it's common enough to be worth bothering however, scs
2613 * opcode could also benefit from calculating both though.
2616 lp_build_sin_or_cos(struct lp_build_context
*bld
,
2620 struct gallivm_state
*gallivm
= bld
->gallivm
;
2621 LLVMBuilderRef b
= gallivm
->builder
;
2622 struct lp_type int_type
= lp_int_type(bld
->type
);
2625 * take the absolute value,
2626 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
2629 LLVMValueRef inv_sig_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, ~0x80000000);
2630 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, bld
->int_vec_type
, "a_v4si");
2632 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
2633 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, bld
->vec_type
, "x_abs");
2637 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
2640 LLVMValueRef FOPi
= lp_build_const_vec(gallivm
, bld
->type
, 1.27323954473516);
2641 LLVMValueRef scale_y
= LLVMBuildFMul(b
, x_abs
, FOPi
, "scale_y");
2644 * store the integer part of y in mm0
2645 * emm2 = _mm_cvttps_epi32(y);
2648 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, bld
->int_vec_type
, "emm2_i");
2651 * j=(j+1) & (~1) (see the cephes sources)
2652 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
2655 LLVMValueRef all_one
= lp_build_const_int_vec(gallivm
, bld
->type
, 1);
2656 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
2658 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
2660 LLVMValueRef inv_one
= lp_build_const_int_vec(gallivm
, bld
->type
, ~1);
2661 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
2664 * y = _mm_cvtepi32_ps(emm2);
2666 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, bld
->vec_type
, "y_2");
2668 LLVMValueRef const_2
= lp_build_const_int_vec(gallivm
, bld
->type
, 2);
2669 LLVMValueRef const_4
= lp_build_const_int_vec(gallivm
, bld
->type
, 4);
2670 LLVMValueRef const_29
= lp_build_const_int_vec(gallivm
, bld
->type
, 29);
2671 LLVMValueRef sign_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, 0x80000000);
2674 * Argument used for poly selection and sign bit determination
2675 * is different for sin vs. cos.
2677 LLVMValueRef emm2_2
= cos
? LLVMBuildSub(b
, emm2_and
, const_2
, "emm2_2") :
2680 LLVMValueRef sign_bit
= cos
? LLVMBuildShl(b
, LLVMBuildAnd(b
, const_4
,
2681 LLVMBuildNot(b
, emm2_2
, ""), ""),
2682 const_29
, "sign_bit") :
2683 LLVMBuildAnd(b
, LLVMBuildXor(b
, a_v4si
,
2684 LLVMBuildShl(b
, emm2_add
,
2686 sign_mask
, "sign_bit");
2689 * get the polynom selection mask
2690 * there is one polynom for 0 <= x <= Pi/4
2691 * and another one for Pi/4<x<=Pi/2
2692 * Both branches will be computed.
2694 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
2695 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
2698 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_2
, const_2
, "emm2_3");
2699 LLVMValueRef poly_mask
= lp_build_compare(gallivm
,
2700 int_type
, PIPE_FUNC_EQUAL
,
2701 emm2_3
, lp_build_const_int_vec(gallivm
, bld
->type
, 0));
2704 * _PS_CONST(minus_cephes_DP1, -0.78515625);
2705 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
2706 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
2708 LLVMValueRef DP1
= lp_build_const_vec(gallivm
, bld
->type
, -0.78515625);
2709 LLVMValueRef DP2
= lp_build_const_vec(gallivm
, bld
->type
, -2.4187564849853515625e-4);
2710 LLVMValueRef DP3
= lp_build_const_vec(gallivm
, bld
->type
, -3.77489497744594108e-8);
2713 * The magic pass: "Extended precision modular arithmetic"
2714 * x = ((x - y * DP1) - y * DP2) - y * DP3;
2715 * xmm1 = _mm_mul_ps(y, xmm1);
2716 * xmm2 = _mm_mul_ps(y, xmm2);
2717 * xmm3 = _mm_mul_ps(y, xmm3);
2719 LLVMValueRef xmm1
= LLVMBuildFMul(b
, y_2
, DP1
, "xmm1");
2720 LLVMValueRef xmm2
= LLVMBuildFMul(b
, y_2
, DP2
, "xmm2");
2721 LLVMValueRef xmm3
= LLVMBuildFMul(b
, y_2
, DP3
, "xmm3");
2724 * x = _mm_add_ps(x, xmm1);
2725 * x = _mm_add_ps(x, xmm2);
2726 * x = _mm_add_ps(x, xmm3);
2729 LLVMValueRef x_1
= LLVMBuildFAdd(b
, x_abs
, xmm1
, "x_1");
2730 LLVMValueRef x_2
= LLVMBuildFAdd(b
, x_1
, xmm2
, "x_2");
2731 LLVMValueRef x_3
= LLVMBuildFAdd(b
, x_2
, xmm3
, "x_3");
2734 * Evaluate the first polynom (0 <= x <= Pi/4)
2736 * z = _mm_mul_ps(x,x);
2738 LLVMValueRef z
= LLVMBuildFMul(b
, x_3
, x_3
, "z");
2741 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
2742 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
2743 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
2745 LLVMValueRef coscof_p0
= lp_build_const_vec(gallivm
, bld
->type
, 2.443315711809948E-005);
2746 LLVMValueRef coscof_p1
= lp_build_const_vec(gallivm
, bld
->type
, -1.388731625493765E-003);
2747 LLVMValueRef coscof_p2
= lp_build_const_vec(gallivm
, bld
->type
, 4.166664568298827E-002);
2750 * y = *(v4sf*)_ps_coscof_p0;
2751 * y = _mm_mul_ps(y, z);
2753 LLVMValueRef y_3
= LLVMBuildFMul(b
, z
, coscof_p0
, "y_3");
2754 LLVMValueRef y_4
= LLVMBuildFAdd(b
, y_3
, coscof_p1
, "y_4");
2755 LLVMValueRef y_5
= LLVMBuildFMul(b
, y_4
, z
, "y_5");
2756 LLVMValueRef y_6
= LLVMBuildFAdd(b
, y_5
, coscof_p2
, "y_6");
2757 LLVMValueRef y_7
= LLVMBuildFMul(b
, y_6
, z
, "y_7");
2758 LLVMValueRef y_8
= LLVMBuildFMul(b
, y_7
, z
, "y_8");
2762 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
2763 * y = _mm_sub_ps(y, tmp);
2764 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
2766 LLVMValueRef half
= lp_build_const_vec(gallivm
, bld
->type
, 0.5);
2767 LLVMValueRef tmp
= LLVMBuildFMul(b
, z
, half
, "tmp");
2768 LLVMValueRef y_9
= LLVMBuildFSub(b
, y_8
, tmp
, "y_8");
2769 LLVMValueRef one
= lp_build_const_vec(gallivm
, bld
->type
, 1.0);
2770 LLVMValueRef y_10
= LLVMBuildFAdd(b
, y_9
, one
, "y_9");
2773 * _PS_CONST(sincof_p0, -1.9515295891E-4);
2774 * _PS_CONST(sincof_p1, 8.3321608736E-3);
2775 * _PS_CONST(sincof_p2, -1.6666654611E-1);
2777 LLVMValueRef sincof_p0
= lp_build_const_vec(gallivm
, bld
->type
, -1.9515295891E-4);
2778 LLVMValueRef sincof_p1
= lp_build_const_vec(gallivm
, bld
->type
, 8.3321608736E-3);
2779 LLVMValueRef sincof_p2
= lp_build_const_vec(gallivm
, bld
->type
, -1.6666654611E-1);
2782 * Evaluate the second polynom (Pi/4 <= x <= 0)
2784 * y2 = *(v4sf*)_ps_sincof_p0;
2785 * y2 = _mm_mul_ps(y2, z);
2786 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
2787 * y2 = _mm_mul_ps(y2, z);
2788 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
2789 * y2 = _mm_mul_ps(y2, z);
2790 * y2 = _mm_mul_ps(y2, x);
2791 * y2 = _mm_add_ps(y2, x);
2794 LLVMValueRef y2_3
= LLVMBuildFMul(b
, z
, sincof_p0
, "y2_3");
2795 LLVMValueRef y2_4
= LLVMBuildFAdd(b
, y2_3
, sincof_p1
, "y2_4");
2796 LLVMValueRef y2_5
= LLVMBuildFMul(b
, y2_4
, z
, "y2_5");
2797 LLVMValueRef y2_6
= LLVMBuildFAdd(b
, y2_5
, sincof_p2
, "y2_6");
2798 LLVMValueRef y2_7
= LLVMBuildFMul(b
, y2_6
, z
, "y2_7");
2799 LLVMValueRef y2_8
= LLVMBuildFMul(b
, y2_7
, x_3
, "y2_8");
2800 LLVMValueRef y2_9
= LLVMBuildFAdd(b
, y2_8
, x_3
, "y2_9");
2803 * select the correct result from the two polynoms
2805 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
2806 * y = _mm_andnot_ps(xmm3, y);
2807 * y = _mm_or_ps(y,y2);
2809 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, bld
->int_vec_type
, "y2_i");
2810 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, bld
->int_vec_type
, "y_i");
2811 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
2812 LLVMValueRef poly_mask_inv
= LLVMBuildNot(b
, poly_mask
, "poly_mask_inv");
2813 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
2814 LLVMValueRef y_combine
= LLVMBuildOr(b
, y_and
, y2_and
, "y_combine");
2818 * y = _mm_xor_ps(y, sign_bit);
2820 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit
, "y_sign");
2821 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, bld
->vec_type
, "y_result");
2823 LLVMValueRef isfinite
= lp_build_isfinite(bld
, a
);
2825 /* clamp output to be within [-1, 1] */
2826 y_result
= lp_build_clamp(bld
, y_result
,
2827 lp_build_const_vec(bld
->gallivm
, bld
->type
, -1.f
),
2828 lp_build_const_vec(bld
->gallivm
, bld
->type
, 1.f
));
2829 /* If a is -inf, inf or NaN then return NaN */
2830 y_result
= lp_build_select(bld
, isfinite
, y_result
,
2831 lp_build_const_vec(bld
->gallivm
, bld
->type
, NAN
));
2840 lp_build_sin(struct lp_build_context
*bld
,
2843 return lp_build_sin_or_cos(bld
, a
, FALSE
);
2851 lp_build_cos(struct lp_build_context
*bld
,
2854 return lp_build_sin_or_cos(bld
, a
, TRUE
);
2859 * Generate pow(x, y)
2862 lp_build_pow(struct lp_build_context
*bld
,
2866 /* TODO: optimize the constant case */
2867 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
2868 LLVMIsConstant(x
) && LLVMIsConstant(y
)) {
2869 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
2873 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
2881 lp_build_exp(struct lp_build_context
*bld
,
2884 /* log2(e) = 1/log(2) */
2885 LLVMValueRef log2e
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2886 1.4426950408889634);
2888 assert(lp_check_value(bld
->type
, x
));
2890 return lp_build_exp2(bld
, lp_build_mul(bld
, log2e
, x
));
2896 * Behavior is undefined with infs, 0s and nans
2899 lp_build_log(struct lp_build_context
*bld
,
2903 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2904 0.69314718055994529);
2906 assert(lp_check_value(bld
->type
, x
));
2908 return lp_build_mul(bld
, log2
, lp_build_log2(bld
, x
));
2912 * Generate log(x) that handles edge cases (infs, 0s and nans)
2915 lp_build_log_safe(struct lp_build_context
*bld
,
2919 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2920 0.69314718055994529);
2922 assert(lp_check_value(bld
->type
, x
));
2924 return lp_build_mul(bld
, log2
, lp_build_log2_safe(bld
, x
));
2929 * Generate polynomial.
2930 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
2933 lp_build_polynomial(struct lp_build_context
*bld
,
2935 const double *coeffs
,
2936 unsigned num_coeffs
)
2938 const struct lp_type type
= bld
->type
;
2939 LLVMValueRef even
= NULL
, odd
= NULL
;
2943 assert(lp_check_value(bld
->type
, x
));
2945 /* TODO: optimize the constant case */
2946 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
2947 LLVMIsConstant(x
)) {
2948 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
2953 * Calculate odd and even terms seperately to decrease data dependency
2955 * c[0] + x^2 * c[2] + x^4 * c[4] ...
2956 * + x * (c[1] + x^2 * c[3] + x^4 * c[5]) ...
2958 x2
= lp_build_mul(bld
, x
, x
);
2960 for (i
= num_coeffs
; i
--; ) {
2963 coeff
= lp_build_const_vec(bld
->gallivm
, type
, coeffs
[i
]);
2967 even
= lp_build_add(bld
, coeff
, lp_build_mul(bld
, x2
, even
));
2972 odd
= lp_build_add(bld
, coeff
, lp_build_mul(bld
, x2
, odd
));
2979 return lp_build_add(bld
, lp_build_mul(bld
, odd
, x
), even
);
2988 * Minimax polynomial fit of 2**x, in range [0, 1[
2990 const double lp_build_exp2_polynomial
[] = {
2991 #if EXP_POLY_DEGREE == 5
2992 1.000000000000000000000, /*XXX: was 0.999999925063526176901, recompute others */
2993 0.693153073200168932794,
2994 0.240153617044375388211,
2995 0.0558263180532956664775,
2996 0.00898934009049466391101,
2997 0.00187757667519147912699
2998 #elif EXP_POLY_DEGREE == 4
2999 1.00000259337069434683,
3000 0.693003834469974940458,
3001 0.24144275689150793076,
3002 0.0520114606103070150235,
3003 0.0135341679161270268764
3004 #elif EXP_POLY_DEGREE == 3
3005 0.999925218562710312959,
3006 0.695833540494823811697,
3007 0.226067155427249155588,
3008 0.0780245226406372992967
3009 #elif EXP_POLY_DEGREE == 2
3010 1.00172476321474503578,
3011 0.657636275736077639316,
3012 0.33718943461968720704
3020 lp_build_exp2(struct lp_build_context
*bld
,
3023 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3024 const struct lp_type type
= bld
->type
;
3025 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3026 LLVMValueRef ipart
= NULL
;
3027 LLVMValueRef fpart
= NULL
;
3028 LLVMValueRef expipart
= NULL
;
3029 LLVMValueRef expfpart
= NULL
;
3030 LLVMValueRef res
= NULL
;
3032 assert(lp_check_value(bld
->type
, x
));
3035 /* TODO: optimize the constant case */
3036 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3037 LLVMIsConstant(x
)) {
3038 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3042 assert(type
.floating
&& type
.width
== 32);
3044 /* We want to preserve NaN and make sure than for exp2 if x > 128,
3045 * the result is INF and if it's smaller than -126.9 the result is 0 */
3046 x
= lp_build_min_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, 128.0), x
,
3047 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
3048 x
= lp_build_max(bld
, lp_build_const_vec(bld
->gallivm
, type
, -126.99999), x
);
3050 /* ipart = floor(x) */
3051 /* fpart = x - ipart */
3052 lp_build_ifloor_fract(bld
, x
, &ipart
, &fpart
);
3056 /* expipart = (float) (1 << ipart) */
3057 expipart
= LLVMBuildAdd(builder
, ipart
,
3058 lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3059 expipart
= LLVMBuildShl(builder
, expipart
,
3060 lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3061 expipart
= LLVMBuildBitCast(builder
, expipart
, vec_type
, "");
3064 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
3065 Elements(lp_build_exp2_polynomial
));
3067 res
= LLVMBuildFMul(builder
, expipart
, expfpart
, "");
3076 * Extract the exponent of a IEEE-754 floating point value.
3078 * Optionally apply an integer bias.
3080 * Result is an integer value with
3082 * ifloor(log2(x)) + bias
3085 lp_build_extract_exponent(struct lp_build_context
*bld
,
3089 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3090 const struct lp_type type
= bld
->type
;
3091 unsigned mantissa
= lp_mantissa(type
);
3094 assert(type
.floating
);
3096 assert(lp_check_value(bld
->type
, x
));
3098 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3100 res
= LLVMBuildLShr(builder
, x
,
3101 lp_build_const_int_vec(bld
->gallivm
, type
, mantissa
), "");
3102 res
= LLVMBuildAnd(builder
, res
,
3103 lp_build_const_int_vec(bld
->gallivm
, type
, 255), "");
3104 res
= LLVMBuildSub(builder
, res
,
3105 lp_build_const_int_vec(bld
->gallivm
, type
, 127 - bias
), "");
3112 * Extract the mantissa of the a floating.
3114 * Result is a floating point value with
3116 * x / floor(log2(x))
3119 lp_build_extract_mantissa(struct lp_build_context
*bld
,
3122 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3123 const struct lp_type type
= bld
->type
;
3124 unsigned mantissa
= lp_mantissa(type
);
3125 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
,
3126 (1ULL << mantissa
) - 1);
3127 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, bld
->int_vec_type
);
3130 assert(lp_check_value(bld
->type
, x
));
3132 assert(type
.floating
);
3134 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3136 /* res = x / 2**ipart */
3137 res
= LLVMBuildAnd(builder
, x
, mantmask
, "");
3138 res
= LLVMBuildOr(builder
, res
, one
, "");
3139 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
3147 * Minimax polynomial fit of log2((1.0 + sqrt(x))/(1.0 - sqrt(x)))/sqrt(x) ,for x in range of [0, 1/9[
3148 * These coefficients can be generate with
3149 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
3151 const double lp_build_log2_polynomial
[] = {
3152 #if LOG_POLY_DEGREE == 5
3153 2.88539008148777786488L,
3154 0.961796878841293367824L,
3155 0.577058946784739859012L,
3156 0.412914355135828735411L,
3157 0.308591899232910175289L,
3158 0.352376952300281371868L,
3159 #elif LOG_POLY_DEGREE == 4
3160 2.88539009343309178325L,
3161 0.961791550404184197881L,
3162 0.577440339438736392009L,
3163 0.403343858251329912514L,
3164 0.406718052498846252698L,
3165 #elif LOG_POLY_DEGREE == 3
3166 2.88538959748872753838L,
3167 0.961932915889597772928L,
3168 0.571118517972136195241L,
3169 0.493997535084709500285L,
3176 * See http://www.devmaster.net/forums/showthread.php?p=43580
3177 * http://en.wikipedia.org/wiki/Logarithm#Calculation
3178 * http://www.nezumi.demon.co.uk/consult/logx.htm
3180 * If handle_edge_cases is true the function will perform computations
3181 * to match the required D3D10+ behavior for each of the edge cases.
3182 * That means that if input is:
3183 * - less than zero (to and including -inf) then NaN will be returned
3184 * - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
3185 * - +infinity, then +infinity will be returned
3186 * - NaN, then NaN will be returned
3188 * Those checks are fairly expensive so if you don't need them make sure
3189 * handle_edge_cases is false.
3192 lp_build_log2_approx(struct lp_build_context
*bld
,
3194 LLVMValueRef
*p_exp
,
3195 LLVMValueRef
*p_floor_log2
,
3196 LLVMValueRef
*p_log2
,
3197 boolean handle_edge_cases
)
3199 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3200 const struct lp_type type
= bld
->type
;
3201 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3202 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
3204 LLVMValueRef expmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x7f800000);
3205 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x007fffff);
3206 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
3208 LLVMValueRef i
= NULL
;
3209 LLVMValueRef y
= NULL
;
3210 LLVMValueRef z
= NULL
;
3211 LLVMValueRef exp
= NULL
;
3212 LLVMValueRef mant
= NULL
;
3213 LLVMValueRef logexp
= NULL
;
3214 LLVMValueRef logmant
= NULL
;
3215 LLVMValueRef res
= NULL
;
3217 assert(lp_check_value(bld
->type
, x
));
3219 if(p_exp
|| p_floor_log2
|| p_log2
) {
3220 /* TODO: optimize the constant case */
3221 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3222 LLVMIsConstant(x
)) {
3223 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3227 assert(type
.floating
&& type
.width
== 32);
3230 * We don't explicitly handle denormalized numbers. They will yield a
3231 * result in the neighbourhood of -127, which appears to be adequate
3235 i
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3237 /* exp = (float) exponent(x) */
3238 exp
= LLVMBuildAnd(builder
, i
, expmask
, "");
3241 if(p_floor_log2
|| p_log2
) {
3242 logexp
= LLVMBuildLShr(builder
, exp
, lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3243 logexp
= LLVMBuildSub(builder
, logexp
, lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3244 logexp
= LLVMBuildSIToFP(builder
, logexp
, vec_type
, "");
3248 /* mant = 1 + (float) mantissa(x) */
3249 mant
= LLVMBuildAnd(builder
, i
, mantmask
, "");
3250 mant
= LLVMBuildOr(builder
, mant
, one
, "");
3251 mant
= LLVMBuildBitCast(builder
, mant
, vec_type
, "");
3253 /* y = (mant - 1) / (mant + 1) */
3254 y
= lp_build_div(bld
,
3255 lp_build_sub(bld
, mant
, bld
->one
),
3256 lp_build_add(bld
, mant
, bld
->one
)
3260 z
= lp_build_mul(bld
, y
, y
);
3263 logmant
= lp_build_polynomial(bld
, z
, lp_build_log2_polynomial
,
3264 Elements(lp_build_log2_polynomial
));
3266 /* logmant = y * P(z) */
3267 logmant
= lp_build_mul(bld
, y
, logmant
);
3269 res
= lp_build_add(bld
, logmant
, logexp
);
3271 if (type
.floating
&& handle_edge_cases
) {
3272 LLVMValueRef negmask
, infmask
, zmask
;
3273 negmask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, x
,
3274 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3275 zmask
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, x
,
3276 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3277 infmask
= lp_build_cmp(bld
, PIPE_FUNC_GEQUAL
, x
,
3278 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
));
3280 /* If x is qual to inf make sure we return inf */
3281 res
= lp_build_select(bld
, infmask
,
3282 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
),
3284 /* If x is qual to 0, return -inf */
3285 res
= lp_build_select(bld
, zmask
,
3286 lp_build_const_vec(bld
->gallivm
, type
, -INFINITY
),
3288 /* If x is nan or less than 0, return nan */
3289 res
= lp_build_select(bld
, negmask
,
3290 lp_build_const_vec(bld
->gallivm
, type
, NAN
),
3296 exp
= LLVMBuildBitCast(builder
, exp
, vec_type
, "");
3301 *p_floor_log2
= logexp
;
3309 * log2 implementation which doesn't have special code to
3310 * handle edge cases (-inf, 0, inf, NaN). It's faster but
3311 * the results for those cases are undefined.
3314 lp_build_log2(struct lp_build_context
*bld
,
3318 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, FALSE
);
3323 * Version of log2 which handles all edge cases.
3324 * Look at documentation of lp_build_log2_approx for
3325 * description of the behavior for each of the edge cases.
3328 lp_build_log2_safe(struct lp_build_context
*bld
,
3332 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, TRUE
);
3338 * Faster (and less accurate) log2.
3340 * log2(x) = floor(log2(x)) - 1 + x / 2**floor(log2(x))
3342 * Piece-wise linear approximation, with exact results when x is a
3345 * See http://www.flipcode.com/archives/Fast_log_Function.shtml
3348 lp_build_fast_log2(struct lp_build_context
*bld
,
3351 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3355 assert(lp_check_value(bld
->type
, x
));
3357 assert(bld
->type
.floating
);
3359 /* ipart = floor(log2(x)) - 1 */
3360 ipart
= lp_build_extract_exponent(bld
, x
, -1);
3361 ipart
= LLVMBuildSIToFP(builder
, ipart
, bld
->vec_type
, "");
3363 /* fpart = x / 2**ipart */
3364 fpart
= lp_build_extract_mantissa(bld
, x
);
3367 return LLVMBuildFAdd(builder
, ipart
, fpart
, "");
3372 * Fast implementation of iround(log2(x)).
3374 * Not an approximation -- it should give accurate results all the time.
3377 lp_build_ilog2(struct lp_build_context
*bld
,
3380 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3381 LLVMValueRef sqrt2
= lp_build_const_vec(bld
->gallivm
, bld
->type
, M_SQRT2
);
3384 assert(bld
->type
.floating
);
3386 assert(lp_check_value(bld
->type
, x
));
3388 /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
3389 x
= LLVMBuildFMul(builder
, x
, sqrt2
, "");
3391 /* ipart = floor(log2(x) + 0.5) */
3392 ipart
= lp_build_extract_exponent(bld
, x
, 0);
3398 lp_build_mod(struct lp_build_context
*bld
,
3402 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3404 const struct lp_type type
= bld
->type
;
3406 assert(lp_check_value(type
, x
));
3407 assert(lp_check_value(type
, y
));
3410 res
= LLVMBuildFRem(builder
, x
, y
, "");
3412 res
= LLVMBuildSRem(builder
, x
, y
, "");
3414 res
= LLVMBuildURem(builder
, x
, y
, "");
3420 * For floating inputs it creates and returns a mask
3421 * which is all 1's for channels which are NaN.
3422 * Channels inside x which are not NaN will be 0.
3425 lp_build_isnan(struct lp_build_context
*bld
,
3429 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3431 assert(bld
->type
.floating
);
3432 assert(lp_check_value(bld
->type
, x
));
3434 mask
= LLVMBuildFCmp(bld
->gallivm
->builder
, LLVMRealOEQ
, x
, x
,
3436 mask
= LLVMBuildNot(bld
->gallivm
->builder
, mask
, "");
3437 mask
= LLVMBuildSExt(bld
->gallivm
->builder
, mask
, int_vec_type
, "isnan");
3441 /* Returns all 1's for floating point numbers that are
3442 * finite numbers and returns all zeros for -inf,
3445 lp_build_isfinite(struct lp_build_context
*bld
,
3448 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3449 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3450 struct lp_type int_type
= lp_int_type(bld
->type
);
3451 LLVMValueRef intx
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3452 LLVMValueRef infornan32
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
3455 if (!bld
->type
.floating
) {
3456 return lp_build_const_int_vec(bld
->gallivm
, bld
->type
, 0);
3458 assert(bld
->type
.floating
);
3459 assert(lp_check_value(bld
->type
, x
));
3460 assert(bld
->type
.width
== 32);
3462 intx
= LLVMBuildAnd(builder
, intx
, infornan32
, "");
3463 return lp_build_compare(bld
->gallivm
, int_type
, PIPE_FUNC_NOTEQUAL
,
3468 * Returns true if the number is nan or inf and false otherwise.
3469 * The input has to be a floating point vector.
3472 lp_build_is_inf_or_nan(struct gallivm_state
*gallivm
,
3473 const struct lp_type type
,
3476 LLVMBuilderRef builder
= gallivm
->builder
;
3477 struct lp_type int_type
= lp_int_type(type
);
3478 LLVMValueRef const0
= lp_build_const_int_vec(gallivm
, int_type
,
3482 assert(type
.floating
);
3484 ret
= LLVMBuildBitCast(builder
, x
, lp_build_vec_type(gallivm
, int_type
), "");
3485 ret
= LLVMBuildAnd(builder
, ret
, const0
, "");
3486 ret
= lp_build_compare(gallivm
, int_type
, PIPE_FUNC_EQUAL
,