1 /**************************************************************************
3 * Copyright 2009 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>
48 #include "util/u_memory.h"
49 #include "util/u_debug.h"
50 #include "util/u_string.h"
51 #include "util/u_cpu_detect.h"
53 #include "lp_bld_type.h"
54 #include "lp_bld_const.h"
55 #include "lp_bld_intr.h"
56 #include "lp_bld_logic.h"
57 #include "lp_bld_pack.h"
58 #include "lp_bld_debug.h"
59 #include "lp_bld_arit.h"
64 * No checks for special case values of a or b = 1 or 0 are done.
67 lp_build_min_simple(struct lp_build_context
*bld
,
71 const struct lp_type type
= bld
->type
;
72 const char *intrinsic
= NULL
;
75 /* TODO: optimize the constant case */
77 if(type
.width
* type
.length
== 128) {
79 if(type
.width
== 32 && util_cpu_caps
.has_sse
)
80 intrinsic
= "llvm.x86.sse.min.ps";
81 if(type
.width
== 64 && util_cpu_caps
.has_sse2
)
82 intrinsic
= "llvm.x86.sse2.min.pd";
85 if(type
.width
== 8 && !type
.sign
&& util_cpu_caps
.has_sse2
)
86 intrinsic
= "llvm.x86.sse2.pminu.b";
87 if(type
.width
== 8 && type
.sign
&& util_cpu_caps
.has_sse4_1
)
88 intrinsic
= "llvm.x86.sse41.pminsb";
89 if(type
.width
== 16 && !type
.sign
&& util_cpu_caps
.has_sse4_1
)
90 intrinsic
= "llvm.x86.sse41.pminuw";
91 if(type
.width
== 16 && type
.sign
&& util_cpu_caps
.has_sse2
)
92 intrinsic
= "llvm.x86.sse2.pmins.w";
93 if(type
.width
== 32 && !type
.sign
&& util_cpu_caps
.has_sse4_1
)
94 intrinsic
= "llvm.x86.sse41.pminud";
95 if(type
.width
== 32 && type
.sign
&& util_cpu_caps
.has_sse4_1
)
96 intrinsic
= "llvm.x86.sse41.pminsd";
101 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, lp_build_vec_type(bld
->type
), a
, b
);
103 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
104 return lp_build_select(bld
, cond
, a
, b
);
110 * No checks for special case values of a or b = 1 or 0 are done.
113 lp_build_max_simple(struct lp_build_context
*bld
,
117 const struct lp_type type
= bld
->type
;
118 const char *intrinsic
= NULL
;
121 /* TODO: optimize the constant case */
123 if(type
.width
* type
.length
== 128) {
125 if(type
.width
== 32 && util_cpu_caps
.has_sse
)
126 intrinsic
= "llvm.x86.sse.max.ps";
127 if(type
.width
== 64 && util_cpu_caps
.has_sse2
)
128 intrinsic
= "llvm.x86.sse2.max.pd";
131 if(type
.width
== 8 && !type
.sign
&& util_cpu_caps
.has_sse2
)
132 intrinsic
= "llvm.x86.sse2.pmaxu.b";
133 if(type
.width
== 8 && type
.sign
&& util_cpu_caps
.has_sse4_1
)
134 intrinsic
= "llvm.x86.sse41.pmaxsb";
135 if(type
.width
== 16 && !type
.sign
&& util_cpu_caps
.has_sse4_1
)
136 intrinsic
= "llvm.x86.sse41.pmaxuw";
137 if(type
.width
== 16 && type
.sign
&& util_cpu_caps
.has_sse2
)
138 intrinsic
= "llvm.x86.sse2.pmaxs.w";
139 if(type
.width
== 32 && !type
.sign
&& util_cpu_caps
.has_sse4_1
)
140 intrinsic
= "llvm.x86.sse41.pmaxud";
141 if(type
.width
== 32 && type
.sign
&& util_cpu_caps
.has_sse4_1
)
142 intrinsic
= "llvm.x86.sse41.pmaxsd";
147 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, lp_build_vec_type(bld
->type
), a
, b
);
149 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
150 return lp_build_select(bld
, cond
, a
, b
);
155 * Generate 1 - a, or ~a depending on bld->type.
158 lp_build_comp(struct lp_build_context
*bld
,
161 const struct lp_type type
= bld
->type
;
168 if(type
.norm
&& !type
.floating
&& !type
.fixed
&& !type
.sign
) {
169 if(LLVMIsConstant(a
))
170 return LLVMConstNot(a
);
172 return LLVMBuildNot(bld
->builder
, a
, "");
175 if(LLVMIsConstant(a
))
176 return LLVMConstSub(bld
->one
, a
);
178 return LLVMBuildSub(bld
->builder
, bld
->one
, a
, "");
186 lp_build_add(struct lp_build_context
*bld
,
190 const struct lp_type type
= bld
->type
;
197 if(a
== bld
->undef
|| b
== bld
->undef
)
201 const char *intrinsic
= NULL
;
203 if(a
== bld
->one
|| b
== bld
->one
)
206 if(util_cpu_caps
.has_sse2
&&
207 type
.width
* type
.length
== 128 &&
208 !type
.floating
&& !type
.fixed
) {
210 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
212 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
216 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, lp_build_vec_type(bld
->type
), a
, b
);
219 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
220 res
= LLVMConstAdd(a
, b
);
222 res
= LLVMBuildAdd(bld
->builder
, a
, b
, "");
224 /* clamp to ceiling of 1.0 */
225 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
226 res
= lp_build_min_simple(bld
, res
, bld
->one
);
228 /* XXX clamp to floor of -1 or 0??? */
238 lp_build_sub(struct lp_build_context
*bld
,
242 const struct lp_type type
= bld
->type
;
247 if(a
== bld
->undef
|| b
== bld
->undef
)
253 const char *intrinsic
= NULL
;
258 if(util_cpu_caps
.has_sse2
&&
259 type
.width
* type
.length
== 128 &&
260 !type
.floating
&& !type
.fixed
) {
262 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
264 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
268 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, lp_build_vec_type(bld
->type
), a
, b
);
271 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
272 res
= LLVMConstSub(a
, b
);
274 res
= LLVMBuildSub(bld
->builder
, a
, b
, "");
276 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
277 res
= lp_build_max_simple(bld
, res
, bld
->zero
);
284 * Normalized 8bit multiplication.
288 * makes the following approximation to the division (Sree)
290 * a*b/255 ~= (a*(b + 1)) >> 256
292 * which is the fastest method that satisfies the following OpenGL criteria
294 * 0*0 = 0 and 255*255 = 255
298 * takes the geometric series approximation to the division
300 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
302 * in this case just the first two terms to fit in 16bit arithmetic
304 * t/255 ~= (t + (t >> 8)) >> 8
306 * note that just by itself it doesn't satisfies the OpenGL criteria, as
307 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
310 * - geometric series plus rounding
312 * when using a geometric series division instead of truncating the result
313 * use roundoff in the approximation (Jim Blinn)
315 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
317 * achieving the exact results
319 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
320 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
321 * @sa Michael Herf, The "double blend trick", May 2000,
322 * http://www.stereopsis.com/doubleblend.html
325 lp_build_mul_u8n(LLVMBuilderRef builder
,
326 struct lp_type i16_type
,
327 LLVMValueRef a
, LLVMValueRef b
)
332 c8
= lp_build_int_const_scalar(i16_type
, 8);
336 /* a*b/255 ~= (a*(b + 1)) >> 256 */
337 b
= LLVMBuildAdd(builder
, b
, lp_build_int_const_scalar(i16_type
, 1), "");
338 ab
= LLVMBuildMul(builder
, a
, b
, "");
342 /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
343 ab
= LLVMBuildMul(builder
, a
, b
, "");
344 ab
= LLVMBuildAdd(builder
, ab
, LLVMBuildLShr(builder
, ab
, c8
, ""), "");
345 ab
= LLVMBuildAdd(builder
, ab
, lp_build_int_const_scalar(i16_type
, 0x80), "");
349 ab
= LLVMBuildLShr(builder
, ab
, c8
, "");
359 lp_build_mul(struct lp_build_context
*bld
,
363 const struct lp_type type
= bld
->type
;
373 if(a
== bld
->undef
|| b
== bld
->undef
)
376 if(!type
.floating
&& !type
.fixed
&& type
.norm
) {
377 if(type
.width
== 8) {
378 struct lp_type i16_type
= lp_wider_type(type
);
379 LLVMValueRef al
, ah
, bl
, bh
, abl
, abh
, ab
;
381 lp_build_unpack2(bld
->builder
, type
, i16_type
, a
, &al
, &ah
);
382 lp_build_unpack2(bld
->builder
, type
, i16_type
, b
, &bl
, &bh
);
384 /* PMULLW, PSRLW, PADDW */
385 abl
= lp_build_mul_u8n(bld
->builder
, i16_type
, al
, bl
);
386 abh
= lp_build_mul_u8n(bld
->builder
, i16_type
, ah
, bh
);
388 ab
= lp_build_pack2(bld
->builder
, i16_type
, type
, abl
, abh
);
397 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
398 return LLVMConstMul(a
, b
);
400 return LLVMBuildMul(bld
->builder
, a
, b
, "");
408 lp_build_div(struct lp_build_context
*bld
,
412 const struct lp_type type
= bld
->type
;
417 return lp_build_rcp(bld
, b
);
422 if(a
== bld
->undef
|| b
== bld
->undef
)
425 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
426 return LLVMConstFDiv(a
, b
);
428 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4)
429 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
431 return LLVMBuildFDiv(bld
->builder
, a
, b
, "");
436 lp_build_lerp(struct lp_build_context
*bld
,
441 return lp_build_add(bld
, v0
, lp_build_mul(bld
, x
, lp_build_sub(bld
, v1
, v0
)));
446 lp_build_lerp_2d(struct lp_build_context
*bld
,
454 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
);
455 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
);
456 return lp_build_lerp(bld
, y
, v0
, v1
);
462 * Do checks for special cases.
465 lp_build_min(struct lp_build_context
*bld
,
469 if(a
== bld
->undef
|| b
== bld
->undef
)
476 if(a
== bld
->zero
|| b
== bld
->zero
)
484 return lp_build_min_simple(bld
, a
, b
);
490 * Do checks for special cases.
493 lp_build_max(struct lp_build_context
*bld
,
497 if(a
== bld
->undef
|| b
== bld
->undef
)
504 if(a
== bld
->one
|| b
== bld
->one
)
512 return lp_build_max_simple(bld
, a
, b
);
520 lp_build_abs(struct lp_build_context
*bld
,
523 const struct lp_type type
= bld
->type
;
524 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
530 /* Mask out the sign bit */
531 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
532 LLVMValueRef mask
= lp_build_int_const_scalar(type
, ((unsigned long long)1 << type
.width
) - 1);
533 a
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
534 a
= LLVMBuildAnd(bld
->builder
, a
, mask
, "");
535 a
= LLVMBuildBitCast(bld
->builder
, a
, vec_type
, "");
539 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
) {
542 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
544 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
546 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
550 return lp_build_max(bld
, a
, LLVMBuildNeg(bld
->builder
, a
, ""));
555 lp_build_sgn(struct lp_build_context
*bld
,
558 const struct lp_type type
= bld
->type
;
559 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
563 /* Handle non-zero case */
565 /* if not zero then sign must be positive */
568 else if(type
.floating
) {
569 /* Take the sign bit and add it to 1 constant */
570 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
571 LLVMValueRef mask
= lp_build_int_const_scalar(type
, (unsigned long long)1 << (type
.width
- 1));
574 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
575 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
576 one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
577 res
= LLVMBuildOr(bld
->builder
, sign
, one
, "");
578 res
= LLVMBuildBitCast(bld
->builder
, res
, vec_type
, "");
582 LLVMValueRef minus_one
= lp_build_const_scalar(type
, -1.0);
583 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
584 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
588 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
589 res
= lp_build_select(bld
, cond
, bld
->zero
, bld
->one
);
595 enum lp_build_round_sse41_mode
597 LP_BUILD_ROUND_SSE41_NEAREST
= 0,
598 LP_BUILD_ROUND_SSE41_FLOOR
= 1,
599 LP_BUILD_ROUND_SSE41_CEIL
= 2,
600 LP_BUILD_ROUND_SSE41_TRUNCATE
= 3
604 static INLINE LLVMValueRef
605 lp_build_round_sse41(struct lp_build_context
*bld
,
607 enum lp_build_round_sse41_mode mode
)
609 const struct lp_type type
= bld
->type
;
610 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
611 const char *intrinsic
;
613 assert(type
.floating
);
614 assert(type
.width
*type
.length
== 128);
615 assert(lp_check_value(type
, a
));
616 assert(util_cpu_caps
.has_sse4_1
);
620 intrinsic
= "llvm.x86.sse41.round.ps";
623 intrinsic
= "llvm.x86.sse41.round.pd";
630 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, vec_type
, a
,
631 LLVMConstInt(LLVMInt32Type(), mode
, 0));
636 lp_build_trunc(struct lp_build_context
*bld
,
639 const struct lp_type type
= bld
->type
;
641 assert(type
.floating
);
642 assert(lp_check_value(type
, a
));
644 if(util_cpu_caps
.has_sse4_1
)
645 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_TRUNCATE
);
647 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
648 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
650 res
= LLVMBuildFPToSI(bld
->builder
, a
, int_vec_type
, "");
651 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
658 lp_build_round(struct lp_build_context
*bld
,
661 const struct lp_type type
= bld
->type
;
663 assert(type
.floating
);
664 assert(lp_check_value(type
, a
));
666 if(util_cpu_caps
.has_sse4_1
)
667 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_NEAREST
);
669 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
671 res
= lp_build_iround(bld
, a
);
672 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
679 lp_build_floor(struct lp_build_context
*bld
,
682 const struct lp_type type
= bld
->type
;
684 assert(type
.floating
);
686 if(util_cpu_caps
.has_sse4_1
)
687 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_FLOOR
);
689 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
691 res
= lp_build_ifloor(bld
, a
);
692 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
699 lp_build_ceil(struct lp_build_context
*bld
,
702 const struct lp_type type
= bld
->type
;
704 assert(type
.floating
);
705 assert(lp_check_value(type
, a
));
707 if(util_cpu_caps
.has_sse4_1
)
708 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_CEIL
);
710 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
712 res
= lp_build_iceil(bld
, a
);
713 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
720 * Convert to integer, through whichever rounding method that's fastest,
721 * typically truncating to zero.
724 lp_build_itrunc(struct lp_build_context
*bld
,
727 const struct lp_type type
= bld
->type
;
728 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
730 assert(type
.floating
);
731 assert(lp_check_value(type
, a
));
733 return LLVMBuildFPToSI(bld
->builder
, a
, int_vec_type
, "");
738 lp_build_iround(struct lp_build_context
*bld
,
741 const struct lp_type type
= bld
->type
;
742 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
745 assert(type
.floating
);
746 assert(lp_check_value(type
, a
));
748 if(util_cpu_caps
.has_sse4_1
) {
749 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_NEAREST
);
752 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
753 LLVMValueRef mask
= lp_build_int_const_scalar(type
, (unsigned long long)1 << (type
.width
- 1));
758 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
759 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
762 half
= lp_build_const_scalar(type
, 0.5);
763 half
= LLVMBuildBitCast(bld
->builder
, half
, int_vec_type
, "");
764 half
= LLVMBuildOr(bld
->builder
, sign
, half
, "");
765 half
= LLVMBuildBitCast(bld
->builder
, half
, vec_type
, "");
767 res
= LLVMBuildAdd(bld
->builder
, a
, half
, "");
770 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "");
777 lp_build_ifloor(struct lp_build_context
*bld
,
780 const struct lp_type type
= bld
->type
;
781 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
784 assert(type
.floating
);
785 assert(lp_check_value(type
, a
));
787 if(util_cpu_caps
.has_sse4_1
) {
788 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_FLOOR
);
791 /* Take the sign bit and add it to 1 constant */
792 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
793 unsigned mantissa
= lp_mantissa(type
);
794 LLVMValueRef mask
= lp_build_int_const_scalar(type
, (unsigned long long)1 << (type
.width
- 1));
798 /* sign = a < 0 ? ~0 : 0 */
799 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
800 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
801 sign
= LLVMBuildAShr(bld
->builder
, sign
, lp_build_int_const_scalar(type
, type
.width
- 1), "");
803 /* offset = -0.99999(9)f */
804 offset
= lp_build_const_scalar(type
, -(double)(((unsigned long long)1 << mantissa
) - 1)/((unsigned long long)1 << mantissa
));
805 offset
= LLVMConstBitCast(offset
, int_vec_type
);
807 /* offset = a < 0 ? -0.99999(9)f : 0.0f */
808 offset
= LLVMBuildAnd(bld
->builder
, offset
, sign
, "");
809 offset
= LLVMBuildBitCast(bld
->builder
, offset
, vec_type
, "");
811 res
= LLVMBuildAdd(bld
->builder
, a
, offset
, "");
814 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "");
821 lp_build_iceil(struct lp_build_context
*bld
,
824 const struct lp_type type
= bld
->type
;
825 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
828 assert(type
.floating
);
829 assert(lp_check_value(type
, a
));
831 if(util_cpu_caps
.has_sse4_1
) {
832 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_CEIL
);
839 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "");
846 lp_build_sqrt(struct lp_build_context
*bld
,
849 const struct lp_type type
= bld
->type
;
850 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
853 /* TODO: optimize the constant case */
854 /* TODO: optimize the constant case */
856 assert(type
.floating
);
857 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.sqrt.v%uf%u", type
.length
, type
.width
);
859 return lp_build_intrinsic_unary(bld
->builder
, intrinsic
, vec_type
, a
);
864 lp_build_rcp(struct lp_build_context
*bld
,
867 const struct lp_type type
= bld
->type
;
876 assert(type
.floating
);
878 if(LLVMIsConstant(a
))
879 return LLVMConstFDiv(bld
->one
, a
);
881 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4)
882 /* FIXME: improve precision */
883 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type
), a
);
885 return LLVMBuildFDiv(bld
->builder
, bld
->one
, a
, "");
893 lp_build_rsqrt(struct lp_build_context
*bld
,
896 const struct lp_type type
= bld
->type
;
898 assert(type
.floating
);
900 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4)
901 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type
), a
);
903 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
911 lp_build_cos(struct lp_build_context
*bld
,
914 const struct lp_type type
= bld
->type
;
915 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
918 /* TODO: optimize the constant case */
920 assert(type
.floating
);
921 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.cos.v%uf%u", type
.length
, type
.width
);
923 return lp_build_intrinsic_unary(bld
->builder
, intrinsic
, vec_type
, a
);
931 lp_build_sin(struct lp_build_context
*bld
,
934 const struct lp_type type
= bld
->type
;
935 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
938 /* TODO: optimize the constant case */
940 assert(type
.floating
);
941 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.sin.v%uf%u", type
.length
, type
.width
);
943 return lp_build_intrinsic_unary(bld
->builder
, intrinsic
, vec_type
, a
);
951 lp_build_pow(struct lp_build_context
*bld
,
955 /* TODO: optimize the constant case */
956 if(LLVMIsConstant(x
) && LLVMIsConstant(y
))
957 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
960 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
968 lp_build_exp(struct lp_build_context
*bld
,
971 /* log2(e) = 1/log(2) */
972 LLVMValueRef log2e
= lp_build_const_scalar(bld
->type
, 1.4426950408889634);
974 return lp_build_mul(bld
, log2e
, lp_build_exp2(bld
, x
));
982 lp_build_log(struct lp_build_context
*bld
,
986 LLVMValueRef log2
= lp_build_const_scalar(bld
->type
, 1.4426950408889634);
988 return lp_build_mul(bld
, log2
, lp_build_exp2(bld
, x
));
992 #define EXP_POLY_DEGREE 3
993 #define LOG_POLY_DEGREE 5
997 * Generate polynomial.
998 * Ex: x^2 * coeffs[0] + x * coeffs[1] + coeffs[2].
1001 lp_build_polynomial(struct lp_build_context
*bld
,
1003 const double *coeffs
,
1004 unsigned num_coeffs
)
1006 const struct lp_type type
= bld
->type
;
1007 LLVMValueRef res
= NULL
;
1010 /* TODO: optimize the constant case */
1011 if(LLVMIsConstant(x
))
1012 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1015 for (i
= num_coeffs
; i
--; ) {
1016 LLVMValueRef coeff
= lp_build_const_scalar(type
, coeffs
[i
]);
1018 res
= lp_build_add(bld
, coeff
, lp_build_mul(bld
, x
, res
));
1031 * Minimax polynomial fit of 2**x, in range [-0.5, 0.5[
1033 const double lp_build_exp2_polynomial
[] = {
1034 #if EXP_POLY_DEGREE == 5
1035 9.9999994e-1, 6.9315308e-1, 2.4015361e-1, 5.5826318e-2, 8.9893397e-3, 1.8775767e-3
1036 #elif EXP_POLY_DEGREE == 4
1037 1.0000026, 6.9300383e-1, 2.4144275e-1, 5.2011464e-2, 1.3534167e-2
1038 #elif EXP_POLY_DEGREE == 3
1039 9.9992520e-1, 6.9583356e-1, 2.2606716e-1, 7.8024521e-2
1040 #elif EXP_POLY_DEGREE == 2
1041 1.0017247, 6.5763628e-1, 3.3718944e-1
1049 lp_build_exp2_approx(struct lp_build_context
*bld
,
1051 LLVMValueRef
*p_exp2_int_part
,
1052 LLVMValueRef
*p_frac_part
,
1053 LLVMValueRef
*p_exp2
)
1055 const struct lp_type type
= bld
->type
;
1056 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1057 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1058 LLVMValueRef ipart
= NULL
;
1059 LLVMValueRef fpart
= NULL
;
1060 LLVMValueRef expipart
= NULL
;
1061 LLVMValueRef expfpart
= NULL
;
1062 LLVMValueRef res
= NULL
;
1064 if(p_exp2_int_part
|| p_frac_part
|| p_exp2
) {
1065 /* TODO: optimize the constant case */
1066 if(LLVMIsConstant(x
))
1067 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1070 assert(type
.floating
&& type
.width
== 32);
1072 x
= lp_build_min(bld
, x
, lp_build_const_scalar(type
, 129.0));
1073 x
= lp_build_max(bld
, x
, lp_build_const_scalar(type
, -126.99999));
1075 /* ipart = int(x - 0.5) */
1076 ipart
= LLVMBuildSub(bld
->builder
, x
, lp_build_const_scalar(type
, 0.5f
), "");
1077 ipart
= LLVMBuildFPToSI(bld
->builder
, ipart
, int_vec_type
, "");
1079 /* fpart = x - ipart */
1080 fpart
= LLVMBuildSIToFP(bld
->builder
, ipart
, vec_type
, "");
1081 fpart
= LLVMBuildSub(bld
->builder
, x
, fpart
, "");
1084 if(p_exp2_int_part
|| p_exp2
) {
1085 /* expipart = (float) (1 << ipart) */
1086 expipart
= LLVMBuildAdd(bld
->builder
, ipart
, lp_build_int_const_scalar(type
, 127), "");
1087 expipart
= LLVMBuildShl(bld
->builder
, expipart
, lp_build_int_const_scalar(type
, 23), "");
1088 expipart
= LLVMBuildBitCast(bld
->builder
, expipart
, vec_type
, "");
1092 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
1093 Elements(lp_build_exp2_polynomial
));
1095 res
= LLVMBuildMul(bld
->builder
, expipart
, expfpart
, "");
1099 *p_exp2_int_part
= expipart
;
1102 *p_frac_part
= fpart
;
1110 lp_build_exp2(struct lp_build_context
*bld
,
1114 lp_build_exp2_approx(bld
, x
, NULL
, NULL
, &res
);
1120 * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
1121 * These coefficients can be generate with
1122 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
1124 const double lp_build_log2_polynomial
[] = {
1125 #if LOG_POLY_DEGREE == 6
1126 3.11578814719469302614, -3.32419399085241980044, 2.59883907202499966007, -1.23152682416275988241, 0.318212422185251071475, -0.0344359067839062357313
1127 #elif LOG_POLY_DEGREE == 5
1128 2.8882704548164776201, -2.52074962577807006663, 1.48116647521213171641, -0.465725644288844778798, 0.0596515482674574969533
1129 #elif LOG_POLY_DEGREE == 4
1130 2.61761038894603480148, -1.75647175389045657003, 0.688243882994381274313, -0.107254423828329604454
1131 #elif LOG_POLY_DEGREE == 3
1132 2.28330284476918490682, -1.04913055217340124191, 0.204446009836232697516
1140 * See http://www.devmaster.net/forums/showthread.php?p=43580
1143 lp_build_log2_approx(struct lp_build_context
*bld
,
1145 LLVMValueRef
*p_exp
,
1146 LLVMValueRef
*p_floor_log2
,
1147 LLVMValueRef
*p_log2
)
1149 const struct lp_type type
= bld
->type
;
1150 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1151 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1153 LLVMValueRef expmask
= lp_build_int_const_scalar(type
, 0x7f800000);
1154 LLVMValueRef mantmask
= lp_build_int_const_scalar(type
, 0x007fffff);
1155 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
1157 LLVMValueRef i
= NULL
;
1158 LLVMValueRef exp
= NULL
;
1159 LLVMValueRef mant
= NULL
;
1160 LLVMValueRef logexp
= NULL
;
1161 LLVMValueRef logmant
= NULL
;
1162 LLVMValueRef res
= NULL
;
1164 if(p_exp
|| p_floor_log2
|| p_log2
) {
1165 /* TODO: optimize the constant case */
1166 if(LLVMIsConstant(x
))
1167 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1170 assert(type
.floating
&& type
.width
== 32);
1172 i
= LLVMBuildBitCast(bld
->builder
, x
, int_vec_type
, "");
1174 /* exp = (float) exponent(x) */
1175 exp
= LLVMBuildAnd(bld
->builder
, i
, expmask
, "");
1178 if(p_floor_log2
|| p_log2
) {
1179 logexp
= LLVMBuildLShr(bld
->builder
, exp
, lp_build_int_const_scalar(type
, 23), "");
1180 logexp
= LLVMBuildSub(bld
->builder
, logexp
, lp_build_int_const_scalar(type
, 127), "");
1181 logexp
= LLVMBuildSIToFP(bld
->builder
, logexp
, vec_type
, "");
1185 /* mant = (float) mantissa(x) */
1186 mant
= LLVMBuildAnd(bld
->builder
, i
, mantmask
, "");
1187 mant
= LLVMBuildOr(bld
->builder
, mant
, one
, "");
1188 mant
= LLVMBuildSIToFP(bld
->builder
, mant
, vec_type
, "");
1190 logmant
= lp_build_polynomial(bld
, mant
, lp_build_log2_polynomial
,
1191 Elements(lp_build_log2_polynomial
));
1193 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
1194 logmant
= LLVMBuildMul(bld
->builder
, logmant
, LLVMBuildMul(bld
->builder
, mant
, bld
->one
, ""), "");
1196 res
= LLVMBuildAdd(bld
->builder
, logmant
, logexp
, "");
1203 *p_floor_log2
= logexp
;
1211 lp_build_log2(struct lp_build_context
*bld
,
1215 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
);