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_math.h"
51 #include "util/u_string.h"
52 #include "util/u_cpu_detect.h"
54 #include "lp_bld_type.h"
55 #include "lp_bld_const.h"
56 #include "lp_bld_intr.h"
57 #include "lp_bld_logic.h"
58 #include "lp_bld_pack.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??? */
234 /** Return the sum of the elements of a */
236 lp_build_sum_vector(struct lp_build_context
*bld
,
239 const struct lp_type type
= bld
->type
;
240 LLVMValueRef index
, res
;
247 assert(type
.length
> 1);
249 assert(!bld
->type
.norm
);
251 index
= LLVMConstInt(LLVMInt32Type(), 0, 0);
252 res
= LLVMBuildExtractElement(bld
->builder
, a
, index
, "");
254 for (i
= 1; i
< type
.length
; i
++) {
255 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
256 res
= LLVMBuildAdd(bld
->builder
, res
,
257 LLVMBuildExtractElement(bld
->builder
, a
, index
, ""),
269 lp_build_sub(struct lp_build_context
*bld
,
273 const struct lp_type type
= bld
->type
;
278 if(a
== bld
->undef
|| b
== bld
->undef
)
284 const char *intrinsic
= NULL
;
289 if(util_cpu_caps
.has_sse2
&&
290 type
.width
* type
.length
== 128 &&
291 !type
.floating
&& !type
.fixed
) {
293 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
295 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
299 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, lp_build_vec_type(bld
->type
), a
, b
);
302 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
303 res
= LLVMConstSub(a
, b
);
305 res
= LLVMBuildSub(bld
->builder
, a
, b
, "");
307 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
308 res
= lp_build_max_simple(bld
, res
, bld
->zero
);
315 * Normalized 8bit multiplication.
319 * makes the following approximation to the division (Sree)
321 * a*b/255 ~= (a*(b + 1)) >> 256
323 * which is the fastest method that satisfies the following OpenGL criteria
325 * 0*0 = 0 and 255*255 = 255
329 * takes the geometric series approximation to the division
331 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
333 * in this case just the first two terms to fit in 16bit arithmetic
335 * t/255 ~= (t + (t >> 8)) >> 8
337 * note that just by itself it doesn't satisfies the OpenGL criteria, as
338 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
341 * - geometric series plus rounding
343 * when using a geometric series division instead of truncating the result
344 * use roundoff in the approximation (Jim Blinn)
346 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
348 * achieving the exact results
350 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
351 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
352 * @sa Michael Herf, The "double blend trick", May 2000,
353 * http://www.stereopsis.com/doubleblend.html
356 lp_build_mul_u8n(LLVMBuilderRef builder
,
357 struct lp_type i16_type
,
358 LLVMValueRef a
, LLVMValueRef b
)
363 c8
= lp_build_const_int_vec(i16_type
, 8);
367 /* a*b/255 ~= (a*(b + 1)) >> 256 */
368 b
= LLVMBuildAdd(builder
, b
, lp_build_const_int_vec(i16_type
, 1), "");
369 ab
= LLVMBuildMul(builder
, a
, b
, "");
373 /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
374 ab
= LLVMBuildMul(builder
, a
, b
, "");
375 ab
= LLVMBuildAdd(builder
, ab
, LLVMBuildLShr(builder
, ab
, c8
, ""), "");
376 ab
= LLVMBuildAdd(builder
, ab
, lp_build_const_int_vec(i16_type
, 0x80), "");
380 ab
= LLVMBuildLShr(builder
, ab
, c8
, "");
390 lp_build_mul(struct lp_build_context
*bld
,
394 const struct lp_type type
= bld
->type
;
406 if(a
== bld
->undef
|| b
== bld
->undef
)
409 if(!type
.floating
&& !type
.fixed
&& type
.norm
) {
410 if(type
.width
== 8) {
411 struct lp_type i16_type
= lp_wider_type(type
);
412 LLVMValueRef al
, ah
, bl
, bh
, abl
, abh
, ab
;
414 lp_build_unpack2(bld
->builder
, type
, i16_type
, a
, &al
, &ah
);
415 lp_build_unpack2(bld
->builder
, type
, i16_type
, b
, &bl
, &bh
);
417 /* PMULLW, PSRLW, PADDW */
418 abl
= lp_build_mul_u8n(bld
->builder
, i16_type
, al
, bl
);
419 abh
= lp_build_mul_u8n(bld
->builder
, i16_type
, ah
, bh
);
421 ab
= lp_build_pack2(bld
->builder
, i16_type
, type
, abl
, abh
);
431 shift
= lp_build_const_int_vec(type
, type
.width
/2);
435 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
436 res
= LLVMConstMul(a
, b
);
439 res
= LLVMConstAShr(res
, shift
);
441 res
= LLVMConstLShr(res
, shift
);
445 res
= LLVMBuildMul(bld
->builder
, a
, b
, "");
448 res
= LLVMBuildAShr(bld
->builder
, res
, shift
, "");
450 res
= LLVMBuildLShr(bld
->builder
, res
, shift
, "");
459 * Small vector x scale multiplication optimization.
462 lp_build_mul_imm(struct lp_build_context
*bld
,
475 return LLVMBuildNeg(bld
->builder
, a
, "");
477 if(b
== 2 && bld
->type
.floating
)
478 return lp_build_add(bld
, a
, a
);
481 unsigned shift
= ffs(b
) - 1;
483 if(bld
->type
.floating
) {
486 * Power of two multiplication by directly manipulating the mantissa.
488 * XXX: This might not be always faster, it will introduce a small error
489 * for multiplication by zero, and it will produce wrong results
492 unsigned mantissa
= lp_mantissa(bld
->type
);
493 factor
= lp_build_const_int_vec(bld
->type
, (unsigned long long)shift
<< mantissa
);
494 a
= LLVMBuildBitCast(bld
->builder
, a
, lp_build_int_vec_type(bld
->type
), "");
495 a
= LLVMBuildAdd(bld
->builder
, a
, factor
, "");
496 a
= LLVMBuildBitCast(bld
->builder
, a
, lp_build_vec_type(bld
->type
), "");
501 factor
= lp_build_const_vec(bld
->type
, shift
);
502 return LLVMBuildShl(bld
->builder
, a
, factor
, "");
506 factor
= lp_build_const_vec(bld
->type
, (double)b
);
507 return lp_build_mul(bld
, a
, factor
);
515 lp_build_div(struct lp_build_context
*bld
,
519 const struct lp_type type
= bld
->type
;
524 return lp_build_rcp(bld
, b
);
529 if(a
== bld
->undef
|| b
== bld
->undef
)
532 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
533 return LLVMConstFDiv(a
, b
);
535 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4)
536 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
538 return LLVMBuildFDiv(bld
->builder
, a
, b
, "");
543 * Linear interpolation.
545 * This also works for integer values with a few caveats.
547 * @sa http://www.stereopsis.com/doubleblend.html
550 lp_build_lerp(struct lp_build_context
*bld
,
558 delta
= lp_build_sub(bld
, v1
, v0
);
560 res
= lp_build_mul(bld
, x
, delta
);
562 res
= lp_build_add(bld
, v0
, res
);
565 /* XXX: This step is necessary for lerping 8bit colors stored on 16bits,
566 * but it will be wrong for other uses. Basically we need a more
567 * powerful lp_type, capable of further distinguishing the values
568 * interpretation from the value storage. */
569 res
= LLVMBuildAnd(bld
->builder
, res
, lp_build_const_int_vec(bld
->type
, (1 << bld
->type
.width
/2) - 1), "");
576 lp_build_lerp_2d(struct lp_build_context
*bld
,
584 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
);
585 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
);
586 return lp_build_lerp(bld
, y
, v0
, v1
);
592 * Do checks for special cases.
595 lp_build_min(struct lp_build_context
*bld
,
599 if(a
== bld
->undef
|| b
== bld
->undef
)
606 if(a
== bld
->zero
|| b
== bld
->zero
)
614 return lp_build_min_simple(bld
, a
, b
);
620 * Do checks for special cases.
623 lp_build_max(struct lp_build_context
*bld
,
627 if(a
== bld
->undef
|| b
== bld
->undef
)
634 if(a
== bld
->one
|| b
== bld
->one
)
642 return lp_build_max_simple(bld
, a
, b
);
647 * Generate clamp(a, min, max)
648 * Do checks for special cases.
651 lp_build_clamp(struct lp_build_context
*bld
,
656 a
= lp_build_min(bld
, a
, max
);
657 a
= lp_build_max(bld
, a
, min
);
666 lp_build_abs(struct lp_build_context
*bld
,
669 const struct lp_type type
= bld
->type
;
670 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
676 /* Mask out the sign bit */
677 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
678 unsigned long long absMask
= ~(1ULL << (type
.width
- 1));
679 LLVMValueRef mask
= lp_build_const_int_vec(type
, ((unsigned long long) absMask
));
680 a
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
681 a
= LLVMBuildAnd(bld
->builder
, a
, mask
, "");
682 a
= LLVMBuildBitCast(bld
->builder
, a
, vec_type
, "");
686 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
) {
689 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
691 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
693 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
697 return lp_build_max(bld
, a
, LLVMBuildNeg(bld
->builder
, a
, ""));
702 lp_build_negate(struct lp_build_context
*bld
,
705 return LLVMBuildNeg(bld
->builder
, a
, "");
709 /** Return -1, 0 or +1 depending on the sign of a */
711 lp_build_sgn(struct lp_build_context
*bld
,
714 const struct lp_type type
= bld
->type
;
718 /* Handle non-zero case */
720 /* if not zero then sign must be positive */
723 else if(type
.floating
) {
724 LLVMTypeRef vec_type
;
725 LLVMTypeRef int_type
;
729 unsigned long long maskBit
= (unsigned long long)1 << (type
.width
- 1);
731 int_type
= lp_build_int_vec_type(type
);
732 vec_type
= lp_build_vec_type(type
);
733 mask
= lp_build_const_int_vec(type
, maskBit
);
735 /* Take the sign bit and add it to 1 constant */
736 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_type
, "");
737 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
738 one
= LLVMConstBitCast(bld
->one
, int_type
);
739 res
= LLVMBuildOr(bld
->builder
, sign
, one
, "");
740 res
= LLVMBuildBitCast(bld
->builder
, res
, vec_type
, "");
744 LLVMValueRef minus_one
= lp_build_const_vec(type
, -1.0);
745 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
746 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
750 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
751 res
= lp_build_select(bld
, cond
, bld
->zero
, res
);
758 * Set the sign of float vector 'a' according to 'sign'.
759 * If sign==0, return abs(a).
760 * If sign==1, return -abs(a);
761 * Other values for sign produce undefined results.
764 lp_build_set_sign(struct lp_build_context
*bld
,
765 LLVMValueRef a
, LLVMValueRef sign
)
767 const struct lp_type type
= bld
->type
;
768 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
769 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
770 LLVMValueRef shift
= lp_build_const_int_vec(type
, type
.width
- 1);
771 LLVMValueRef mask
= lp_build_const_int_vec(type
,
772 ~((unsigned long long) 1 << (type
.width
- 1)));
773 LLVMValueRef val
, res
;
775 assert(type
.floating
);
777 /* val = reinterpret_cast<int>(a) */
778 val
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
779 /* val = val & mask */
780 val
= LLVMBuildAnd(bld
->builder
, val
, mask
, "");
781 /* sign = sign << shift */
782 sign
= LLVMBuildShl(bld
->builder
, sign
, shift
, "");
783 /* res = val | sign */
784 res
= LLVMBuildOr(bld
->builder
, val
, sign
, "");
785 /* res = reinterpret_cast<float>(res) */
786 res
= LLVMBuildBitCast(bld
->builder
, res
, vec_type
, "");
793 * Convert vector of (or scalar) int to vector of (or scalar) float.
796 lp_build_int_to_float(struct lp_build_context
*bld
,
799 const struct lp_type type
= bld
->type
;
800 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
802 assert(type
.floating
);
804 return LLVMBuildSIToFP(bld
->builder
, a
, vec_type
, "");
809 enum lp_build_round_sse41_mode
811 LP_BUILD_ROUND_SSE41_NEAREST
= 0,
812 LP_BUILD_ROUND_SSE41_FLOOR
= 1,
813 LP_BUILD_ROUND_SSE41_CEIL
= 2,
814 LP_BUILD_ROUND_SSE41_TRUNCATE
= 3
818 static INLINE LLVMValueRef
819 lp_build_round_sse41(struct lp_build_context
*bld
,
821 enum lp_build_round_sse41_mode mode
)
823 const struct lp_type type
= bld
->type
;
824 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
825 const char *intrinsic
;
827 assert(type
.floating
);
828 assert(type
.width
*type
.length
== 128);
829 assert(lp_check_value(type
, a
));
830 assert(util_cpu_caps
.has_sse4_1
);
834 intrinsic
= "llvm.x86.sse41.round.ps";
837 intrinsic
= "llvm.x86.sse41.round.pd";
844 return lp_build_intrinsic_binary(bld
->builder
, intrinsic
, vec_type
, a
,
845 LLVMConstInt(LLVMInt32Type(), mode
, 0));
850 * Return the integer part of a float (vector) value. The returned value is
852 * Ex: trunc(-1.5) = 1.0
855 lp_build_trunc(struct lp_build_context
*bld
,
858 const struct lp_type type
= bld
->type
;
860 assert(type
.floating
);
861 assert(lp_check_value(type
, a
));
863 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128)
864 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_TRUNCATE
);
866 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
867 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
869 res
= LLVMBuildFPToSI(bld
->builder
, a
, int_vec_type
, "");
870 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
877 * Return float (vector) rounded to nearest integer (vector). The returned
878 * value is a float (vector).
879 * Ex: round(0.9) = 1.0
880 * Ex: round(-1.5) = -2.0
883 lp_build_round(struct lp_build_context
*bld
,
886 const struct lp_type type
= bld
->type
;
888 assert(type
.floating
);
889 assert(lp_check_value(type
, a
));
891 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128)
892 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_NEAREST
);
894 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
896 res
= lp_build_iround(bld
, a
);
897 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
904 * Return floor of float (vector), result is a float (vector)
905 * Ex: floor(1.1) = 1.0
906 * Ex: floor(-1.1) = -2.0
909 lp_build_floor(struct lp_build_context
*bld
,
912 const struct lp_type type
= bld
->type
;
914 assert(type
.floating
);
915 assert(lp_check_value(type
, a
));
917 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128)
918 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_FLOOR
);
920 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
922 res
= lp_build_ifloor(bld
, a
);
923 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
930 * Return ceiling of float (vector), returning float (vector).
931 * Ex: ceil( 1.1) = 2.0
932 * Ex: ceil(-1.1) = -1.0
935 lp_build_ceil(struct lp_build_context
*bld
,
938 const struct lp_type type
= bld
->type
;
940 assert(type
.floating
);
941 assert(lp_check_value(type
, a
));
943 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128)
944 return lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_CEIL
);
946 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
948 res
= lp_build_iceil(bld
, a
);
949 res
= LLVMBuildSIToFP(bld
->builder
, res
, vec_type
, "");
956 * Return fractional part of 'a' computed as a - floor(a)
957 * Typically used in texture coord arithmetic.
960 lp_build_fract(struct lp_build_context
*bld
,
963 assert(bld
->type
.floating
);
964 return lp_build_sub(bld
, a
, lp_build_floor(bld
, a
));
969 * Return the integer part of a float (vector) value. The returned value is
970 * an integer (vector).
971 * Ex: itrunc(-1.5) = 1
974 lp_build_itrunc(struct lp_build_context
*bld
,
977 const struct lp_type type
= bld
->type
;
978 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
980 assert(type
.floating
);
981 assert(lp_check_value(type
, a
));
983 return LLVMBuildFPToSI(bld
->builder
, a
, int_vec_type
, "");
988 * Return float (vector) rounded to nearest integer (vector). The returned
989 * value is an integer (vector).
990 * Ex: iround(0.9) = 1
991 * Ex: iround(-1.5) = -2
994 lp_build_iround(struct lp_build_context
*bld
,
997 const struct lp_type type
= bld
->type
;
998 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1001 assert(type
.floating
);
1003 assert(lp_check_value(type
, a
));
1005 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128) {
1006 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_NEAREST
);
1009 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1010 LLVMValueRef mask
= lp_build_const_int_vec(type
, (unsigned long long)1 << (type
.width
- 1));
1015 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
1016 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
1019 half
= lp_build_const_vec(type
, 0.5);
1020 half
= LLVMBuildBitCast(bld
->builder
, half
, int_vec_type
, "");
1021 half
= LLVMBuildOr(bld
->builder
, sign
, half
, "");
1022 half
= LLVMBuildBitCast(bld
->builder
, half
, vec_type
, "");
1024 res
= LLVMBuildAdd(bld
->builder
, a
, half
, "");
1027 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "");
1034 * Return floor of float (vector), result is an int (vector)
1035 * Ex: ifloor(1.1) = 1.0
1036 * Ex: ifloor(-1.1) = -2.0
1039 lp_build_ifloor(struct lp_build_context
*bld
,
1042 const struct lp_type type
= bld
->type
;
1043 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1046 assert(type
.floating
);
1047 assert(lp_check_value(type
, a
));
1049 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128) {
1050 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_FLOOR
);
1053 /* Take the sign bit and add it to 1 constant */
1054 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1055 unsigned mantissa
= lp_mantissa(type
);
1056 LLVMValueRef mask
= lp_build_const_int_vec(type
, (unsigned long long)1 << (type
.width
- 1));
1058 LLVMValueRef offset
;
1060 /* sign = a < 0 ? ~0 : 0 */
1061 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
1062 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
1063 sign
= LLVMBuildAShr(bld
->builder
, sign
, lp_build_const_int_vec(type
, type
.width
- 1), "ifloor.sign");
1065 /* offset = -0.99999(9)f */
1066 offset
= lp_build_const_vec(type
, -(double)(((unsigned long long)1 << mantissa
) - 10)/((unsigned long long)1 << mantissa
));
1067 offset
= LLVMConstBitCast(offset
, int_vec_type
);
1069 /* offset = a < 0 ? offset : 0.0f */
1070 offset
= LLVMBuildAnd(bld
->builder
, offset
, sign
, "");
1071 offset
= LLVMBuildBitCast(bld
->builder
, offset
, vec_type
, "ifloor.offset");
1073 res
= LLVMBuildAdd(bld
->builder
, a
, offset
, "ifloor.res");
1076 /* round to nearest (toward zero) */
1077 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "ifloor.res");
1084 * Return ceiling of float (vector), returning int (vector).
1085 * Ex: iceil( 1.1) = 2
1086 * Ex: iceil(-1.1) = -1
1089 lp_build_iceil(struct lp_build_context
*bld
,
1092 const struct lp_type type
= bld
->type
;
1093 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1096 assert(type
.floating
);
1097 assert(lp_check_value(type
, a
));
1099 if (util_cpu_caps
.has_sse4_1
&& type
.width
*type
.length
== 128) {
1100 res
= lp_build_round_sse41(bld
, a
, LP_BUILD_ROUND_SSE41_CEIL
);
1103 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1104 unsigned mantissa
= lp_mantissa(type
);
1105 LLVMValueRef mask
= lp_build_const_int_vec(type
, (unsigned long long)1 << (type
.width
- 1));
1107 LLVMValueRef offset
;
1109 /* sign = a < 0 ? 0 : ~0 */
1110 sign
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
1111 sign
= LLVMBuildAnd(bld
->builder
, sign
, mask
, "");
1112 sign
= LLVMBuildAShr(bld
->builder
, sign
, lp_build_const_int_vec(type
, type
.width
- 1), "iceil.sign");
1113 sign
= LLVMBuildNot(bld
->builder
, sign
, "iceil.not");
1115 /* offset = 0.99999(9)f */
1116 offset
= lp_build_const_vec(type
, (double)(((unsigned long long)1 << mantissa
) - 10)/((unsigned long long)1 << mantissa
));
1117 offset
= LLVMConstBitCast(offset
, int_vec_type
);
1119 /* offset = a < 0 ? 0.0 : offset */
1120 offset
= LLVMBuildAnd(bld
->builder
, offset
, sign
, "");
1121 offset
= LLVMBuildBitCast(bld
->builder
, offset
, vec_type
, "iceil.offset");
1123 res
= LLVMBuildAdd(bld
->builder
, a
, offset
, "iceil.res");
1126 /* round to nearest (toward zero) */
1127 res
= LLVMBuildFPToSI(bld
->builder
, res
, int_vec_type
, "iceil.res");
1134 lp_build_sqrt(struct lp_build_context
*bld
,
1137 const struct lp_type type
= bld
->type
;
1138 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1141 /* TODO: optimize the constant case */
1142 /* TODO: optimize the constant case */
1144 assert(type
.floating
);
1145 util_snprintf(intrinsic
, sizeof intrinsic
, "llvm.sqrt.v%uf%u", type
.length
, type
.width
);
1147 return lp_build_intrinsic_unary(bld
->builder
, intrinsic
, vec_type
, a
);
1152 lp_build_rcp(struct lp_build_context
*bld
,
1155 const struct lp_type type
= bld
->type
;
1164 assert(type
.floating
);
1166 if(LLVMIsConstant(a
))
1167 return LLVMConstFDiv(bld
->one
, a
);
1169 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) {
1171 * XXX: Added precision is not always necessary, so only enable this
1172 * when we have a better system in place to track minimum precision.
1177 * Do one Newton-Raphson step to improve precision:
1179 * x1 = (2 - a * rcp(a)) * rcp(a)
1182 LLVMValueRef two
= lp_build_const_vec(bld
->type
, 2.0);
1186 rcp_a
= lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type
), a
);
1188 res
= LLVMBuildMul(bld
->builder
, a
, rcp_a
, "");
1189 res
= LLVMBuildSub(bld
->builder
, two
, res
, "");
1190 res
= LLVMBuildMul(bld
->builder
, res
, rcp_a
, "");
1194 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type
), a
);
1198 return LLVMBuildFDiv(bld
->builder
, bld
->one
, a
, "");
1203 * Generate 1/sqrt(a)
1206 lp_build_rsqrt(struct lp_build_context
*bld
,
1209 const struct lp_type type
= bld
->type
;
1211 assert(type
.floating
);
1213 if(util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4)
1214 return lp_build_intrinsic_unary(bld
->builder
, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type
), a
);
1216 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
1220 static inline LLVMValueRef
1221 lp_build_const_v4si(unsigned long value
)
1223 LLVMValueRef element
= LLVMConstInt(LLVMInt32Type(), value
, 0);
1224 LLVMValueRef elements
[4] = { element
, element
, element
, element
};
1225 return LLVMConstVector(elements
, 4);
1228 static inline LLVMValueRef
1229 lp_build_const_v4sf(float value
)
1231 LLVMValueRef element
= LLVMConstReal(LLVMFloatType(), value
);
1232 LLVMValueRef elements
[4] = { element
, element
, element
, element
};
1233 return LLVMConstVector(elements
, 4);
1238 * Generate sin(a) using SSE2
1241 lp_build_sin(struct lp_build_context
*bld
,
1244 struct lp_type int_type
= lp_int_type(bld
->type
);
1245 LLVMBuilderRef b
= bld
->builder
;
1246 LLVMTypeRef v4sf
= LLVMVectorType(LLVMFloatType(), 4);
1247 LLVMTypeRef v4si
= LLVMVectorType(LLVMInt32Type(), 4);
1250 * take the absolute value,
1251 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
1254 LLVMValueRef inv_sig_mask
= lp_build_const_v4si(~0x80000000);
1255 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, v4si
, "a_v4si");
1257 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
1258 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, v4sf
, "x_abs");
1261 * extract the sign bit (upper one)
1262 * sign_bit = _mm_and_ps(sign_bit, *(v4sf*)_ps_sign_mask);
1264 LLVMValueRef sig_mask
= lp_build_const_v4si(0x80000000);
1265 LLVMValueRef sign_bit_i
= LLVMBuildAnd(b
, a_v4si
, sig_mask
, "sign_bit_i");
1269 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
1272 LLVMValueRef FOPi
= lp_build_const_v4sf(1.27323954473516);
1273 LLVMValueRef scale_y
= LLVMBuildMul(b
, x_abs
, FOPi
, "scale_y");
1276 * store the integer part of y in mm0
1277 * emm2 = _mm_cvttps_epi32(y);
1280 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, v4si
, "emm2_i");
1283 * j=(j+1) & (~1) (see the cephes sources)
1284 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
1287 LLVMValueRef all_one
= lp_build_const_v4si(1);
1288 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
1290 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
1292 LLVMValueRef inv_one
= lp_build_const_v4si(~1);
1293 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
1296 * y = _mm_cvtepi32_ps(emm2);
1298 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, v4sf
, "y_2");
1300 /* get the swap sign flag
1301 * emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
1303 LLVMValueRef pi32_4
= lp_build_const_v4si(4);
1304 LLVMValueRef emm0_and
= LLVMBuildAnd(b
, emm2_add
, pi32_4
, "emm0_and");
1307 * emm2 = _mm_slli_epi32(emm0, 29);
1309 LLVMValueRef const_29
= lp_build_const_v4si(29);
1310 LLVMValueRef swap_sign_bit
= LLVMBuildShl(b
, emm0_and
, const_29
, "swap_sign_bit");
1313 * get the polynom selection mask
1314 * there is one polynom for 0 <= x <= Pi/4
1315 * and another one for Pi/4<x<=Pi/2
1316 * Both branches will be computed.
1318 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
1319 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
1322 LLVMValueRef pi32_2
= lp_build_const_v4si(2);
1323 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_and
, pi32_2
, "emm2_3");
1324 LLVMValueRef poly_mask
= lp_build_compare(b
, int_type
, PIPE_FUNC_EQUAL
,
1325 emm2_3
, lp_build_const_v4si(0));
1327 * sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
1329 LLVMValueRef sign_bit_1
= LLVMBuildXor(b
, sign_bit_i
, swap_sign_bit
, "sign_bit");
1332 * _PS_CONST(minus_cephes_DP1, -0.78515625);
1333 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
1334 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
1336 LLVMValueRef DP1
= lp_build_const_v4sf(-0.78515625);
1337 LLVMValueRef DP2
= lp_build_const_v4sf(-2.4187564849853515625e-4);
1338 LLVMValueRef DP3
= lp_build_const_v4sf(-3.77489497744594108e-8);
1341 * The magic pass: "Extended precision modular arithmetic"
1342 * x = ((x - y * DP1) - y * DP2) - y * DP3;
1343 * xmm1 = _mm_mul_ps(y, xmm1);
1344 * xmm2 = _mm_mul_ps(y, xmm2);
1345 * xmm3 = _mm_mul_ps(y, xmm3);
1347 LLVMValueRef xmm1
= LLVMBuildMul(b
, y_2
, DP1
, "xmm1");
1348 LLVMValueRef xmm2
= LLVMBuildMul(b
, y_2
, DP2
, "xmm2");
1349 LLVMValueRef xmm3
= LLVMBuildMul(b
, y_2
, DP3
, "xmm3");
1352 * x = _mm_add_ps(x, xmm1);
1353 * x = _mm_add_ps(x, xmm2);
1354 * x = _mm_add_ps(x, xmm3);
1357 LLVMValueRef x_1
= LLVMBuildAdd(b
, x_abs
, xmm1
, "x_1");
1358 LLVMValueRef x_2
= LLVMBuildAdd(b
, x_1
, xmm2
, "x_2");
1359 LLVMValueRef x_3
= LLVMBuildAdd(b
, x_2
, xmm3
, "x_3");
1362 * Evaluate the first polynom (0 <= x <= Pi/4)
1364 * z = _mm_mul_ps(x,x);
1366 LLVMValueRef z
= LLVMBuildMul(b
, x_3
, x_3
, "z");
1369 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
1370 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
1371 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
1373 LLVMValueRef coscof_p0
= lp_build_const_v4sf(2.443315711809948E-005);
1374 LLVMValueRef coscof_p1
= lp_build_const_v4sf(-1.388731625493765E-003);
1375 LLVMValueRef coscof_p2
= lp_build_const_v4sf(4.166664568298827E-002);
1378 * y = *(v4sf*)_ps_coscof_p0;
1379 * y = _mm_mul_ps(y, z);
1381 LLVMValueRef y_3
= LLVMBuildMul(b
, z
, coscof_p0
, "y_3");
1382 LLVMValueRef y_4
= LLVMBuildAdd(b
, y_3
, coscof_p1
, "y_4");
1383 LLVMValueRef y_5
= LLVMBuildMul(b
, y_4
, z
, "y_5");
1384 LLVMValueRef y_6
= LLVMBuildAdd(b
, y_5
, coscof_p2
, "y_6");
1385 LLVMValueRef y_7
= LLVMBuildMul(b
, y_6
, z
, "y_7");
1386 LLVMValueRef y_8
= LLVMBuildMul(b
, y_7
, z
, "y_8");
1390 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
1391 * y = _mm_sub_ps(y, tmp);
1392 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
1394 LLVMValueRef half
= lp_build_const_v4sf(0.5);
1395 LLVMValueRef tmp
= LLVMBuildMul(b
, z
, half
, "tmp");
1396 LLVMValueRef y_9
= LLVMBuildSub(b
, y_8
, tmp
, "y_8");
1397 LLVMValueRef one
= lp_build_const_v4sf(1.0);
1398 LLVMValueRef y_10
= LLVMBuildAdd(b
, y_9
, one
, "y_9");
1401 * _PS_CONST(sincof_p0, -1.9515295891E-4);
1402 * _PS_CONST(sincof_p1, 8.3321608736E-3);
1403 * _PS_CONST(sincof_p2, -1.6666654611E-1);
1405 LLVMValueRef sincof_p0
= lp_build_const_v4sf(-1.9515295891E-4);
1406 LLVMValueRef sincof_p1
= lp_build_const_v4sf(8.3321608736E-3);
1407 LLVMValueRef sincof_p2
= lp_build_const_v4sf(-1.6666654611E-1);
1410 * Evaluate the second polynom (Pi/4 <= x <= 0)
1412 * y2 = *(v4sf*)_ps_sincof_p0;
1413 * y2 = _mm_mul_ps(y2, z);
1414 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
1415 * y2 = _mm_mul_ps(y2, z);
1416 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
1417 * y2 = _mm_mul_ps(y2, z);
1418 * y2 = _mm_mul_ps(y2, x);
1419 * y2 = _mm_add_ps(y2, x);
1422 LLVMValueRef y2_3
= LLVMBuildMul(b
, z
, sincof_p0
, "y2_3");
1423 LLVMValueRef y2_4
= LLVMBuildAdd(b
, y2_3
, sincof_p1
, "y2_4");
1424 LLVMValueRef y2_5
= LLVMBuildMul(b
, y2_4
, z
, "y2_5");
1425 LLVMValueRef y2_6
= LLVMBuildAdd(b
, y2_5
, sincof_p2
, "y2_6");
1426 LLVMValueRef y2_7
= LLVMBuildMul(b
, y2_6
, z
, "y2_7");
1427 LLVMValueRef y2_8
= LLVMBuildMul(b
, y2_7
, x_3
, "y2_8");
1428 LLVMValueRef y2_9
= LLVMBuildAdd(b
, y2_8
, x_3
, "y2_9");
1431 * select the correct result from the two polynoms
1433 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
1434 * y = _mm_andnot_ps(xmm3, y);
1435 * y = _mm_add_ps(y,y2);
1437 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, v4si
, "y2_i");
1438 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, v4si
, "y_i");
1439 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
1440 LLVMValueRef inv
= lp_build_const_v4si(~0);
1441 LLVMValueRef poly_mask_inv
= LLVMBuildXor(b
, poly_mask
, inv
, "poly_mask_inv");
1442 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
1443 LLVMValueRef y_combine
= LLVMBuildAdd(b
, y_and
, y2_and
, "y_combine");
1447 * y = _mm_xor_ps(y, sign_bit);
1449 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit_1
, "y_sin");
1450 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, v4sf
, "y_result");
1456 * Generate cos(a) using SSE2
1459 lp_build_cos(struct lp_build_context
*bld
,
1462 struct lp_type int_type
= lp_int_type(bld
->type
);
1463 LLVMBuilderRef b
= bld
->builder
;
1464 LLVMTypeRef v4sf
= LLVMVectorType(LLVMFloatType(), 4);
1465 LLVMTypeRef v4si
= LLVMVectorType(LLVMInt32Type(), 4);
1468 * take the absolute value,
1469 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
1472 LLVMValueRef inv_sig_mask
= lp_build_const_v4si(~0x80000000);
1473 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, v4si
, "a_v4si");
1475 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
1476 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, v4sf
, "x_abs");
1480 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
1483 LLVMValueRef FOPi
= lp_build_const_v4sf(1.27323954473516);
1484 LLVMValueRef scale_y
= LLVMBuildMul(b
, x_abs
, FOPi
, "scale_y");
1487 * store the integer part of y in mm0
1488 * emm2 = _mm_cvttps_epi32(y);
1491 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, v4si
, "emm2_i");
1494 * j=(j+1) & (~1) (see the cephes sources)
1495 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
1498 LLVMValueRef all_one
= lp_build_const_v4si(1);
1499 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
1501 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
1503 LLVMValueRef inv_one
= lp_build_const_v4si(~1);
1504 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
1507 * y = _mm_cvtepi32_ps(emm2);
1509 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, v4sf
, "y_2");
1513 * emm2 = _mm_sub_epi32(emm2, *(v4si*)_pi32_2);
1515 LLVMValueRef const_2
= lp_build_const_v4si(2);
1516 LLVMValueRef emm2_2
= LLVMBuildSub(b
, emm2_and
, const_2
, "emm2_2");
1519 /* get the swap sign flag
1520 * emm0 = _mm_andnot_si128(emm2, *(v4si*)_pi32_4);
1522 LLVMValueRef inv
= lp_build_const_v4si(~0);
1523 LLVMValueRef emm0_not
= LLVMBuildXor(b
, emm2_2
, inv
, "emm0_not");
1524 LLVMValueRef pi32_4
= lp_build_const_v4si(4);
1525 LLVMValueRef emm0_and
= LLVMBuildAnd(b
, emm0_not
, pi32_4
, "emm0_and");
1528 * emm2 = _mm_slli_epi32(emm0, 29);
1530 LLVMValueRef const_29
= lp_build_const_v4si(29);
1531 LLVMValueRef sign_bit
= LLVMBuildShl(b
, emm0_and
, const_29
, "sign_bit");
1534 * get the polynom selection mask
1535 * there is one polynom for 0 <= x <= Pi/4
1536 * and another one for Pi/4<x<=Pi/2
1537 * Both branches will be computed.
1539 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
1540 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
1543 LLVMValueRef pi32_2
= lp_build_const_v4si(2);
1544 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_2
, pi32_2
, "emm2_3");
1545 LLVMValueRef poly_mask
= lp_build_compare(b
, int_type
, PIPE_FUNC_EQUAL
,
1546 emm2_3
, lp_build_const_v4si(0));
1549 * _PS_CONST(minus_cephes_DP1, -0.78515625);
1550 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
1551 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
1553 LLVMValueRef DP1
= lp_build_const_v4sf(-0.78515625);
1554 LLVMValueRef DP2
= lp_build_const_v4sf(-2.4187564849853515625e-4);
1555 LLVMValueRef DP3
= lp_build_const_v4sf(-3.77489497744594108e-8);
1558 * The magic pass: "Extended precision modular arithmetic"
1559 * x = ((x - y * DP1) - y * DP2) - y * DP3;
1560 * xmm1 = _mm_mul_ps(y, xmm1);
1561 * xmm2 = _mm_mul_ps(y, xmm2);
1562 * xmm3 = _mm_mul_ps(y, xmm3);
1564 LLVMValueRef xmm1
= LLVMBuildMul(b
, y_2
, DP1
, "xmm1");
1565 LLVMValueRef xmm2
= LLVMBuildMul(b
, y_2
, DP2
, "xmm2");
1566 LLVMValueRef xmm3
= LLVMBuildMul(b
, y_2
, DP3
, "xmm3");
1569 * x = _mm_add_ps(x, xmm1);
1570 * x = _mm_add_ps(x, xmm2);
1571 * x = _mm_add_ps(x, xmm3);
1574 LLVMValueRef x_1
= LLVMBuildAdd(b
, x_abs
, xmm1
, "x_1");
1575 LLVMValueRef x_2
= LLVMBuildAdd(b
, x_1
, xmm2
, "x_2");
1576 LLVMValueRef x_3
= LLVMBuildAdd(b
, x_2
, xmm3
, "x_3");
1579 * Evaluate the first polynom (0 <= x <= Pi/4)
1581 * z = _mm_mul_ps(x,x);
1583 LLVMValueRef z
= LLVMBuildMul(b
, x_3
, x_3
, "z");
1586 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
1587 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
1588 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
1590 LLVMValueRef coscof_p0
= lp_build_const_v4sf(2.443315711809948E-005);
1591 LLVMValueRef coscof_p1
= lp_build_const_v4sf(-1.388731625493765E-003);
1592 LLVMValueRef coscof_p2
= lp_build_const_v4sf(4.166664568298827E-002);
1595 * y = *(v4sf*)_ps_coscof_p0;
1596 * y = _mm_mul_ps(y, z);
1598 LLVMValueRef y_3
= LLVMBuildMul(b
, z
, coscof_p0
, "y_3");
1599 LLVMValueRef y_4
= LLVMBuildAdd(b
, y_3
, coscof_p1
, "y_4");
1600 LLVMValueRef y_5
= LLVMBuildMul(b
, y_4
, z
, "y_5");
1601 LLVMValueRef y_6
= LLVMBuildAdd(b
, y_5
, coscof_p2
, "y_6");
1602 LLVMValueRef y_7
= LLVMBuildMul(b
, y_6
, z
, "y_7");
1603 LLVMValueRef y_8
= LLVMBuildMul(b
, y_7
, z
, "y_8");
1607 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
1608 * y = _mm_sub_ps(y, tmp);
1609 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
1611 LLVMValueRef half
= lp_build_const_v4sf(0.5);
1612 LLVMValueRef tmp
= LLVMBuildMul(b
, z
, half
, "tmp");
1613 LLVMValueRef y_9
= LLVMBuildSub(b
, y_8
, tmp
, "y_8");
1614 LLVMValueRef one
= lp_build_const_v4sf(1.0);
1615 LLVMValueRef y_10
= LLVMBuildAdd(b
, y_9
, one
, "y_9");
1618 * _PS_CONST(sincof_p0, -1.9515295891E-4);
1619 * _PS_CONST(sincof_p1, 8.3321608736E-3);
1620 * _PS_CONST(sincof_p2, -1.6666654611E-1);
1622 LLVMValueRef sincof_p0
= lp_build_const_v4sf(-1.9515295891E-4);
1623 LLVMValueRef sincof_p1
= lp_build_const_v4sf(8.3321608736E-3);
1624 LLVMValueRef sincof_p2
= lp_build_const_v4sf(-1.6666654611E-1);
1627 * Evaluate the second polynom (Pi/4 <= x <= 0)
1629 * y2 = *(v4sf*)_ps_sincof_p0;
1630 * y2 = _mm_mul_ps(y2, z);
1631 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
1632 * y2 = _mm_mul_ps(y2, z);
1633 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
1634 * y2 = _mm_mul_ps(y2, z);
1635 * y2 = _mm_mul_ps(y2, x);
1636 * y2 = _mm_add_ps(y2, x);
1639 LLVMValueRef y2_3
= LLVMBuildMul(b
, z
, sincof_p0
, "y2_3");
1640 LLVMValueRef y2_4
= LLVMBuildAdd(b
, y2_3
, sincof_p1
, "y2_4");
1641 LLVMValueRef y2_5
= LLVMBuildMul(b
, y2_4
, z
, "y2_5");
1642 LLVMValueRef y2_6
= LLVMBuildAdd(b
, y2_5
, sincof_p2
, "y2_6");
1643 LLVMValueRef y2_7
= LLVMBuildMul(b
, y2_6
, z
, "y2_7");
1644 LLVMValueRef y2_8
= LLVMBuildMul(b
, y2_7
, x_3
, "y2_8");
1645 LLVMValueRef y2_9
= LLVMBuildAdd(b
, y2_8
, x_3
, "y2_9");
1648 * select the correct result from the two polynoms
1650 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
1651 * y = _mm_andnot_ps(xmm3, y);
1652 * y = _mm_add_ps(y,y2);
1654 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, v4si
, "y2_i");
1655 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, v4si
, "y_i");
1656 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
1657 LLVMValueRef poly_mask_inv
= LLVMBuildXor(b
, poly_mask
, inv
, "poly_mask_inv");
1658 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
1659 LLVMValueRef y_combine
= LLVMBuildAdd(b
, y_and
, y2_and
, "y_combine");
1663 * y = _mm_xor_ps(y, sign_bit);
1665 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit
, "y_sin");
1666 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, v4sf
, "y_result");
1672 * Generate pow(x, y)
1675 lp_build_pow(struct lp_build_context
*bld
,
1679 /* TODO: optimize the constant case */
1680 if(LLVMIsConstant(x
) && LLVMIsConstant(y
))
1681 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1684 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
1692 lp_build_exp(struct lp_build_context
*bld
,
1695 /* log2(e) = 1/log(2) */
1696 LLVMValueRef log2e
= lp_build_const_vec(bld
->type
, 1.4426950408889634);
1698 return lp_build_mul(bld
, log2e
, lp_build_exp2(bld
, x
));
1706 lp_build_log(struct lp_build_context
*bld
,
1710 LLVMValueRef log2
= lp_build_const_vec(bld
->type
, 0.69314718055994529);
1712 return lp_build_mul(bld
, log2
, lp_build_exp2(bld
, x
));
1716 #define EXP_POLY_DEGREE 3
1717 #define LOG_POLY_DEGREE 5
1721 * Generate polynomial.
1722 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
1725 lp_build_polynomial(struct lp_build_context
*bld
,
1727 const double *coeffs
,
1728 unsigned num_coeffs
)
1730 const struct lp_type type
= bld
->type
;
1731 LLVMValueRef res
= NULL
;
1734 /* TODO: optimize the constant case */
1735 if(LLVMIsConstant(x
))
1736 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1739 for (i
= num_coeffs
; i
--; ) {
1742 coeff
= lp_build_const_vec(type
, coeffs
[i
]);
1745 res
= lp_build_add(bld
, coeff
, lp_build_mul(bld
, x
, res
));
1758 * Minimax polynomial fit of 2**x, in range [0, 1[
1760 const double lp_build_exp2_polynomial
[] = {
1761 #if EXP_POLY_DEGREE == 5
1762 0.999999999690134838155,
1763 0.583974334321735217258,
1764 0.164553105719676828492,
1765 0.0292811063701710962255,
1766 0.00354944426657875141846,
1767 0.000296253726543423377365
1768 #elif EXP_POLY_DEGREE == 4
1769 1.00000001502262084505,
1770 0.563586057338685991394,
1771 0.150436017652442413623,
1772 0.0243220604213317927308,
1773 0.0025359088446580436489
1774 #elif EXP_POLY_DEGREE == 3
1775 0.999925218562710312959,
1776 0.695833540494823811697,
1777 0.226067155427249155588,
1778 0.0780245226406372992967
1779 #elif EXP_POLY_DEGREE == 2
1780 1.00172476321474503578,
1781 0.657636275736077639316,
1782 0.33718943461968720704
1790 lp_build_exp2_approx(struct lp_build_context
*bld
,
1792 LLVMValueRef
*p_exp2_int_part
,
1793 LLVMValueRef
*p_frac_part
,
1794 LLVMValueRef
*p_exp2
)
1796 const struct lp_type type
= bld
->type
;
1797 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1798 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1799 LLVMValueRef ipart
= NULL
;
1800 LLVMValueRef fpart
= NULL
;
1801 LLVMValueRef expipart
= NULL
;
1802 LLVMValueRef expfpart
= NULL
;
1803 LLVMValueRef res
= NULL
;
1805 if(p_exp2_int_part
|| p_frac_part
|| p_exp2
) {
1806 /* TODO: optimize the constant case */
1807 if(LLVMIsConstant(x
))
1808 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1811 assert(type
.floating
&& type
.width
== 32);
1813 x
= lp_build_min(bld
, x
, lp_build_const_vec(type
, 129.0));
1814 x
= lp_build_max(bld
, x
, lp_build_const_vec(type
, -126.99999));
1816 /* ipart = floor(x) */
1817 ipart
= lp_build_floor(bld
, x
);
1819 /* fpart = x - ipart */
1820 fpart
= LLVMBuildSub(bld
->builder
, x
, ipart
, "");
1823 if(p_exp2_int_part
|| p_exp2
) {
1824 /* expipart = (float) (1 << ipart) */
1825 ipart
= LLVMBuildFPToSI(bld
->builder
, ipart
, int_vec_type
, "");
1826 expipart
= LLVMBuildAdd(bld
->builder
, ipart
, lp_build_const_int_vec(type
, 127), "");
1827 expipart
= LLVMBuildShl(bld
->builder
, expipart
, lp_build_const_int_vec(type
, 23), "");
1828 expipart
= LLVMBuildBitCast(bld
->builder
, expipart
, vec_type
, "");
1832 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
1833 Elements(lp_build_exp2_polynomial
));
1835 res
= LLVMBuildMul(bld
->builder
, expipart
, expfpart
, "");
1839 *p_exp2_int_part
= expipart
;
1842 *p_frac_part
= fpart
;
1850 lp_build_exp2(struct lp_build_context
*bld
,
1854 lp_build_exp2_approx(bld
, x
, NULL
, NULL
, &res
);
1860 * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
1861 * These coefficients can be generate with
1862 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
1864 const double lp_build_log2_polynomial
[] = {
1865 #if LOG_POLY_DEGREE == 6
1866 3.11578814719469302614,
1867 -3.32419399085241980044,
1868 2.59883907202499966007,
1869 -1.23152682416275988241,
1870 0.318212422185251071475,
1871 -0.0344359067839062357313
1872 #elif LOG_POLY_DEGREE == 5
1873 2.8882704548164776201,
1874 -2.52074962577807006663,
1875 1.48116647521213171641,
1876 -0.465725644288844778798,
1877 0.0596515482674574969533
1878 #elif LOG_POLY_DEGREE == 4
1879 2.61761038894603480148,
1880 -1.75647175389045657003,
1881 0.688243882994381274313,
1882 -0.107254423828329604454
1883 #elif LOG_POLY_DEGREE == 3
1884 2.28330284476918490682,
1885 -1.04913055217340124191,
1886 0.204446009836232697516
1894 * See http://www.devmaster.net/forums/showthread.php?p=43580
1897 lp_build_log2_approx(struct lp_build_context
*bld
,
1899 LLVMValueRef
*p_exp
,
1900 LLVMValueRef
*p_floor_log2
,
1901 LLVMValueRef
*p_log2
)
1903 const struct lp_type type
= bld
->type
;
1904 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
1905 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
1907 LLVMValueRef expmask
= lp_build_const_int_vec(type
, 0x7f800000);
1908 LLVMValueRef mantmask
= lp_build_const_int_vec(type
, 0x007fffff);
1909 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
1911 LLVMValueRef i
= NULL
;
1912 LLVMValueRef exp
= NULL
;
1913 LLVMValueRef mant
= NULL
;
1914 LLVMValueRef logexp
= NULL
;
1915 LLVMValueRef logmant
= NULL
;
1916 LLVMValueRef res
= NULL
;
1918 if(p_exp
|| p_floor_log2
|| p_log2
) {
1919 /* TODO: optimize the constant case */
1920 if(LLVMIsConstant(x
))
1921 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1924 assert(type
.floating
&& type
.width
== 32);
1926 i
= LLVMBuildBitCast(bld
->builder
, x
, int_vec_type
, "");
1928 /* exp = (float) exponent(x) */
1929 exp
= LLVMBuildAnd(bld
->builder
, i
, expmask
, "");
1932 if(p_floor_log2
|| p_log2
) {
1933 logexp
= LLVMBuildLShr(bld
->builder
, exp
, lp_build_const_int_vec(type
, 23), "");
1934 logexp
= LLVMBuildSub(bld
->builder
, logexp
, lp_build_const_int_vec(type
, 127), "");
1935 logexp
= LLVMBuildSIToFP(bld
->builder
, logexp
, vec_type
, "");
1939 /* mant = (float) mantissa(x) */
1940 mant
= LLVMBuildAnd(bld
->builder
, i
, mantmask
, "");
1941 mant
= LLVMBuildOr(bld
->builder
, mant
, one
, "");
1942 mant
= LLVMBuildBitCast(bld
->builder
, mant
, vec_type
, "");
1944 logmant
= lp_build_polynomial(bld
, mant
, lp_build_log2_polynomial
,
1945 Elements(lp_build_log2_polynomial
));
1947 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
1948 logmant
= LLVMBuildMul(bld
->builder
, logmant
, LLVMBuildSub(bld
->builder
, mant
, bld
->one
, ""), "");
1950 res
= LLVMBuildAdd(bld
->builder
, logmant
, logexp
, "");
1954 exp
= LLVMBuildBitCast(bld
->builder
, exp
, vec_type
, "");
1959 *p_floor_log2
= logexp
;
1967 lp_build_log2(struct lp_build_context
*bld
,
1971 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
);