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_debug.h"
60 #include "lp_bld_arit.h"
65 * No checks for special case values of a or b = 1 or 0 are done.
68 lp_build_min_simple(struct lp_build_context *bld,
72 const struct lp_type type = bld->type;
73 const char *intrinsic = NULL;
76 /* TODO: optimize the constant case */
78 if(type.width * type.length == 128) {
80 if(type.width == 32 && util_cpu_caps.has_sse)
81 intrinsic = "llvm.x86.sse.min.ps";
82 if(type.width == 64 && util_cpu_caps.has_sse2)
83 intrinsic = "llvm.x86.sse2.min.pd";
86 if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
87 intrinsic = "llvm.x86.sse2.pminu.b";
88 if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
89 intrinsic = "llvm.x86.sse41.pminsb";
90 if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
91 intrinsic = "llvm.x86.sse41.pminuw";
92 if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
93 intrinsic = "llvm.x86.sse2.pmins.w";
94 if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
95 intrinsic = "llvm.x86.sse41.pminud";
96 if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
97 intrinsic = "llvm.x86.sse41.pminsd";
102 return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
104 cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
105 return lp_build_select(bld, cond, a, b);
111 * No checks for special case values of a or b = 1 or 0 are done.
114 lp_build_max_simple(struct lp_build_context *bld,
118 const struct lp_type type = bld->type;
119 const char *intrinsic = NULL;
122 /* TODO: optimize the constant case */
124 if(type.width * type.length == 128) {
126 if(type.width == 32 && util_cpu_caps.has_sse)
127 intrinsic = "llvm.x86.sse.max.ps";
128 if(type.width == 64 && util_cpu_caps.has_sse2)
129 intrinsic = "llvm.x86.sse2.max.pd";
132 if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
133 intrinsic = "llvm.x86.sse2.pmaxu.b";
134 if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
135 intrinsic = "llvm.x86.sse41.pmaxsb";
136 if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
137 intrinsic = "llvm.x86.sse41.pmaxuw";
138 if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
139 intrinsic = "llvm.x86.sse2.pmaxs.w";
140 if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
141 intrinsic = "llvm.x86.sse41.pmaxud";
142 if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
143 intrinsic = "llvm.x86.sse41.pmaxsd";
148 return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
150 cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
151 return lp_build_select(bld, cond, a, b);
156 * Generate 1 - a, or ~a depending on bld->type.
159 lp_build_comp(struct lp_build_context *bld,
162 const struct lp_type type = bld->type;
169 if(type.norm && !type.floating && !type.fixed && !type.sign) {
170 if(LLVMIsConstant(a))
171 return LLVMConstNot(a);
173 return LLVMBuildNot(bld->builder, a, "");
176 if(LLVMIsConstant(a))
177 return LLVMConstSub(bld->one, a);
179 return LLVMBuildSub(bld->builder, bld->one, a, "");
187 lp_build_add(struct lp_build_context *bld,
191 const struct lp_type type = bld->type;
198 if(a == bld->undef || b == bld->undef)
202 const char *intrinsic = NULL;
204 if(a == bld->one || b == bld->one)
207 if(util_cpu_caps.has_sse2 &&
208 type.width * type.length == 128 &&
209 !type.floating && !type.fixed) {
211 intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
213 intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
217 return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
220 if(LLVMIsConstant(a) && LLVMIsConstant(b))
221 res = LLVMConstAdd(a, b);
223 res = LLVMBuildAdd(bld->builder, a, b, "");
225 /* clamp to ceiling of 1.0 */
226 if(bld->type.norm && (bld->type.floating || bld->type.fixed))
227 res = lp_build_min_simple(bld, res, bld->one);
229 /* XXX clamp to floor of -1 or 0??? */
239 lp_build_sub(struct lp_build_context *bld,
243 const struct lp_type type = bld->type;
248 if(a == bld->undef || b == bld->undef)
254 const char *intrinsic = NULL;
259 if(util_cpu_caps.has_sse2 &&
260 type.width * type.length == 128 &&
261 !type.floating && !type.fixed) {
263 intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
265 intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
269 return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
272 if(LLVMIsConstant(a) && LLVMIsConstant(b))
273 res = LLVMConstSub(a, b);
275 res = LLVMBuildSub(bld->builder, a, b, "");
277 if(bld->type.norm && (bld->type.floating || bld->type.fixed))
278 res = lp_build_max_simple(bld, res, bld->zero);
285 * Normalized 8bit multiplication.
289 * makes the following approximation to the division (Sree)
291 * a*b/255 ~= (a*(b + 1)) >> 256
293 * which is the fastest method that satisfies the following OpenGL criteria
295 * 0*0 = 0 and 255*255 = 255
299 * takes the geometric series approximation to the division
301 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
303 * in this case just the first two terms to fit in 16bit arithmetic
305 * t/255 ~= (t + (t >> 8)) >> 8
307 * note that just by itself it doesn't satisfies the OpenGL criteria, as
308 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
311 * - geometric series plus rounding
313 * when using a geometric series division instead of truncating the result
314 * use roundoff in the approximation (Jim Blinn)
316 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
318 * achieving the exact results
320 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
321 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
322 * @sa Michael Herf, The "double blend trick", May 2000,
323 * http://www.stereopsis.com/doubleblend.html
326 lp_build_mul_u8n(LLVMBuilderRef builder,
327 struct lp_type i16_type,
328 LLVMValueRef a, LLVMValueRef b)
333 c8 = lp_build_int_const_scalar(i16_type, 8);
337 /* a*b/255 ~= (a*(b + 1)) >> 256 */
338 b = LLVMBuildAdd(builder, b, lp_build_int_const_scalar(i16_type, 1), "");
339 ab = LLVMBuildMul(builder, a, b, "");
343 /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
344 ab = LLVMBuildMul(builder, a, b, "");
345 ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c8, ""), "");
346 ab = LLVMBuildAdd(builder, ab, lp_build_int_const_scalar(i16_type, 0x80), "");
350 ab = LLVMBuildLShr(builder, ab, c8, "");
360 lp_build_mul(struct lp_build_context *bld,
364 const struct lp_type type = bld->type;
376 if(a == bld->undef || b == bld->undef)
379 if(!type.floating && !type.fixed && type.norm) {
380 if(type.width == 8) {
381 struct lp_type i16_type = lp_wider_type(type);
382 LLVMValueRef al, ah, bl, bh, abl, abh, ab;
384 lp_build_unpack2(bld->builder, type, i16_type, a, &al, &ah);
385 lp_build_unpack2(bld->builder, type, i16_type, b, &bl, &bh);
387 /* PMULLW, PSRLW, PADDW */
388 abl = lp_build_mul_u8n(bld->builder, i16_type, al, bl);
389 abh = lp_build_mul_u8n(bld->builder, i16_type, ah, bh);
391 ab = lp_build_pack2(bld->builder, i16_type, type, abl, abh);
401 shift = lp_build_int_const_scalar(type, type.width/2);
405 if(LLVMIsConstant(a) && LLVMIsConstant(b)) {
406 res = LLVMConstMul(a, b);
409 res = LLVMConstAShr(res, shift);
411 res = LLVMConstLShr(res, shift);
415 res = LLVMBuildMul(bld->builder, a, b, "");
418 res = LLVMBuildAShr(bld->builder, res, shift, "");
420 res = LLVMBuildLShr(bld->builder, res, shift, "");
429 * Small vector x scale multiplication optimization.
432 lp_build_mul_imm(struct lp_build_context *bld,
445 return LLVMBuildNeg(bld->builder, a, "");
447 if(b == 2 && bld->type.floating)
448 return lp_build_add(bld, a, a);
451 unsigned shift = ffs(b) - 1;
453 if(bld->type.floating) {
456 * Power of two multiplication by directly manipulating the mantissa.
458 * XXX: This might not be always faster, it will introduce a small error
459 * for multiplication by zero, and it will produce wrong results
462 unsigned mantissa = lp_mantissa(bld->type);
463 factor = lp_build_int_const_scalar(bld->type, (unsigned long long)shift << mantissa);
464 a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), "");
465 a = LLVMBuildAdd(bld->builder, a, factor, "");
466 a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), "");
471 factor = lp_build_const_scalar(bld->type, shift);
472 return LLVMBuildShl(bld->builder, a, factor, "");
476 factor = lp_build_const_scalar(bld->type, (double)b);
477 return lp_build_mul(bld, a, factor);
485 lp_build_div(struct lp_build_context *bld,
489 const struct lp_type type = bld->type;
494 return lp_build_rcp(bld, b);
499 if(a == bld->undef || b == bld->undef)
502 if(LLVMIsConstant(a) && LLVMIsConstant(b))
503 return LLVMConstFDiv(a, b);
505 if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
506 return lp_build_mul(bld, a, lp_build_rcp(bld, b));
508 return LLVMBuildFDiv(bld->builder, a, b, "");
513 * Linear interpolation.
515 * This also works for integer values with a few caveats.
517 * @sa http://www.stereopsis.com/doubleblend.html
520 lp_build_lerp(struct lp_build_context *bld,
528 delta = lp_build_sub(bld, v1, v0);
530 res = lp_build_mul(bld, x, delta);
532 res = lp_build_add(bld, v0, res);
535 /* XXX: This step is necessary for lerping 8bit colors stored on 16bits,
536 * but it will be wrong for other uses. Basically we need a more
537 * powerful lp_type, capable of further distinguishing the values
538 * interpretation from the value storage. */
539 res = LLVMBuildAnd(bld->builder, res, lp_build_int_const_scalar(bld->type, (1 << bld->type.width/2) - 1), "");
546 lp_build_lerp_2d(struct lp_build_context *bld,
554 LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01);
555 LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11);
556 return lp_build_lerp(bld, y, v0, v1);
562 * Do checks for special cases.
565 lp_build_min(struct lp_build_context *bld,
569 if(a == bld->undef || b == bld->undef)
576 if(a == bld->zero || b == bld->zero)
584 return lp_build_min_simple(bld, a, b);
590 * Do checks for special cases.
593 lp_build_max(struct lp_build_context *bld,
597 if(a == bld->undef || b == bld->undef)
604 if(a == bld->one || b == bld->one)
612 return lp_build_max_simple(bld, a, b);
620 lp_build_abs(struct lp_build_context *bld,
623 const struct lp_type type = bld->type;
624 LLVMTypeRef vec_type = lp_build_vec_type(type);
630 /* Mask out the sign bit */
631 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
632 LLVMValueRef mask = lp_build_int_const_scalar(type, ((unsigned long long)1 << type.width) - 1);
633 a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
634 a = LLVMBuildAnd(bld->builder, a, mask, "");
635 a = LLVMBuildBitCast(bld->builder, a, vec_type, "");
639 if(type.width*type.length == 128 && util_cpu_caps.has_ssse3) {
642 return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
644 return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
646 return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
650 return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, ""));
655 lp_build_sgn(struct lp_build_context *bld,
658 const struct lp_type type = bld->type;
659 LLVMTypeRef vec_type = lp_build_vec_type(type);
663 /* Handle non-zero case */
665 /* if not zero then sign must be positive */
668 else if(type.floating) {
669 /* Take the sign bit and add it to 1 constant */
670 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
671 LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
674 sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
675 sign = LLVMBuildAnd(bld->builder, sign, mask, "");
676 one = LLVMConstBitCast(bld->one, int_vec_type);
677 res = LLVMBuildOr(bld->builder, sign, one, "");
678 res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
682 LLVMValueRef minus_one = lp_build_const_scalar(type, -1.0);
683 cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero);
684 res = lp_build_select(bld, cond, bld->one, minus_one);
688 cond = lp_build_cmp(bld, PIPE_FUNC_EQUAL, a, bld->zero);
689 res = lp_build_select(bld, cond, bld->zero, bld->one);
695 enum lp_build_round_sse41_mode
697 LP_BUILD_ROUND_SSE41_NEAREST = 0,
698 LP_BUILD_ROUND_SSE41_FLOOR = 1,
699 LP_BUILD_ROUND_SSE41_CEIL = 2,
700 LP_BUILD_ROUND_SSE41_TRUNCATE = 3
704 static INLINE LLVMValueRef
705 lp_build_round_sse41(struct lp_build_context *bld,
707 enum lp_build_round_sse41_mode mode)
709 const struct lp_type type = bld->type;
710 LLVMTypeRef vec_type = lp_build_vec_type(type);
711 const char *intrinsic;
713 assert(type.floating);
714 assert(type.width*type.length == 128);
715 assert(lp_check_value(type, a));
716 assert(util_cpu_caps.has_sse4_1);
720 intrinsic = "llvm.x86.sse41.round.ps";
723 intrinsic = "llvm.x86.sse41.round.pd";
730 return lp_build_intrinsic_binary(bld->builder, intrinsic, vec_type, a,
731 LLVMConstInt(LLVMInt32Type(), mode, 0));
736 lp_build_trunc(struct lp_build_context *bld,
739 const struct lp_type type = bld->type;
741 assert(type.floating);
742 assert(lp_check_value(type, a));
744 if(util_cpu_caps.has_sse4_1)
745 return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE);
747 LLVMTypeRef vec_type = lp_build_vec_type(type);
748 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
750 res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
751 res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
758 lp_build_round(struct lp_build_context *bld,
761 const struct lp_type type = bld->type;
763 assert(type.floating);
764 assert(lp_check_value(type, a));
766 if(util_cpu_caps.has_sse4_1)
767 return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
769 LLVMTypeRef vec_type = lp_build_vec_type(type);
771 res = lp_build_iround(bld, a);
772 res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
779 lp_build_floor(struct lp_build_context *bld,
782 const struct lp_type type = bld->type;
784 assert(type.floating);
786 if(util_cpu_caps.has_sse4_1)
787 return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
789 LLVMTypeRef vec_type = lp_build_vec_type(type);
791 res = lp_build_ifloor(bld, a);
792 res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
799 lp_build_ceil(struct lp_build_context *bld,
802 const struct lp_type type = bld->type;
804 assert(type.floating);
805 assert(lp_check_value(type, a));
807 if(util_cpu_caps.has_sse4_1)
808 return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
810 LLVMTypeRef vec_type = lp_build_vec_type(type);
812 res = lp_build_iceil(bld, a);
813 res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
820 * Convert to integer, through whichever rounding method that's fastest,
821 * typically truncating to zero.
824 lp_build_itrunc(struct lp_build_context *bld,
827 const struct lp_type type = bld->type;
828 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
830 assert(type.floating);
831 assert(lp_check_value(type, a));
833 return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
838 lp_build_iround(struct lp_build_context *bld,
841 const struct lp_type type = bld->type;
842 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
845 assert(type.floating);
846 assert(lp_check_value(type, a));
848 if(util_cpu_caps.has_sse4_1) {
849 res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
852 LLVMTypeRef vec_type = lp_build_vec_type(type);
853 LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
858 sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
859 sign = LLVMBuildAnd(bld->builder, sign, mask, "");
862 half = lp_build_const_scalar(type, 0.5);
863 half = LLVMBuildBitCast(bld->builder, half, int_vec_type, "");
864 half = LLVMBuildOr(bld->builder, sign, half, "");
865 half = LLVMBuildBitCast(bld->builder, half, vec_type, "");
867 res = LLVMBuildAdd(bld->builder, a, half, "");
870 res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
877 lp_build_ifloor(struct lp_build_context *bld,
880 const struct lp_type type = bld->type;
881 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
884 assert(type.floating);
885 assert(lp_check_value(type, a));
887 if(util_cpu_caps.has_sse4_1) {
888 res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
891 /* Take the sign bit and add it to 1 constant */
892 LLVMTypeRef vec_type = lp_build_vec_type(type);
893 unsigned mantissa = lp_mantissa(type);
894 LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
898 /* sign = a < 0 ? ~0 : 0 */
899 sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
900 sign = LLVMBuildAnd(bld->builder, sign, mask, "");
901 sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), "");
903 /* offset = -0.99999(9)f */
904 offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
905 offset = LLVMConstBitCast(offset, int_vec_type);
907 /* offset = a < 0 ? -0.99999(9)f : 0.0f */
908 offset = LLVMBuildAnd(bld->builder, offset, sign, "");
909 offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");
911 res = LLVMBuildAdd(bld->builder, a, offset, "");
914 res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
921 lp_build_iceil(struct lp_build_context *bld,
924 const struct lp_type type = bld->type;
925 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
928 assert(type.floating);
929 assert(lp_check_value(type, a));
931 if(util_cpu_caps.has_sse4_1) {
932 res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
939 res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
946 lp_build_sqrt(struct lp_build_context *bld,
949 const struct lp_type type = bld->type;
950 LLVMTypeRef vec_type = lp_build_vec_type(type);
953 /* TODO: optimize the constant case */
954 /* TODO: optimize the constant case */
956 assert(type.floating);
957 util_snprintf(intrinsic, sizeof intrinsic, "llvm.sqrt.v%uf%u", type.length, type.width);
959 return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
964 lp_build_rcp(struct lp_build_context *bld,
967 const struct lp_type type = bld->type;
976 assert(type.floating);
978 if(LLVMIsConstant(a))
979 return LLVMConstFDiv(bld->one, a);
981 if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
982 /* FIXME: improve precision */
983 return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a);
985 return LLVMBuildFDiv(bld->builder, bld->one, a, "");
993 lp_build_rsqrt(struct lp_build_context *bld,
996 const struct lp_type type = bld->type;
998 assert(type.floating);
1000 if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
1001 return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type), a);
1003 return lp_build_rcp(bld, lp_build_sqrt(bld, a));
1011 lp_build_cos(struct lp_build_context *bld,
1014 const struct lp_type type = bld->type;
1015 LLVMTypeRef vec_type = lp_build_vec_type(type);
1018 /* TODO: optimize the constant case */
1020 assert(type.floating);
1021 util_snprintf(intrinsic, sizeof intrinsic, "llvm.cos.v%uf%u", type.length, type.width);
1023 return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
1031 lp_build_sin(struct lp_build_context *bld,
1034 const struct lp_type type = bld->type;
1035 LLVMTypeRef vec_type = lp_build_vec_type(type);
1038 /* TODO: optimize the constant case */
1040 assert(type.floating);
1041 util_snprintf(intrinsic, sizeof intrinsic, "llvm.sin.v%uf%u", type.length, type.width);
1043 return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
1048 * Generate pow(x, y)
1051 lp_build_pow(struct lp_build_context *bld,
1055 /* TODO: optimize the constant case */
1056 if(LLVMIsConstant(x) && LLVMIsConstant(y))
1057 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1060 return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y));
1068 lp_build_exp(struct lp_build_context *bld,
1071 /* log2(e) = 1/log(2) */
1072 LLVMValueRef log2e = lp_build_const_scalar(bld->type, 1.4426950408889634);
1074 return lp_build_mul(bld, log2e, lp_build_exp2(bld, x));
1082 lp_build_log(struct lp_build_context *bld,
1086 LLVMValueRef log2 = lp_build_const_scalar(bld->type, 0.69314718055994529);
1088 return lp_build_mul(bld, log2, lp_build_exp2(bld, x));
1092 #define EXP_POLY_DEGREE 3
1093 #define LOG_POLY_DEGREE 5
1097 * Generate polynomial.
1098 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
1101 lp_build_polynomial(struct lp_build_context *bld,
1103 const double *coeffs,
1104 unsigned num_coeffs)
1106 const struct lp_type type = bld->type;
1107 LLVMValueRef res = NULL;
1110 /* TODO: optimize the constant case */
1111 if(LLVMIsConstant(x))
1112 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1115 for (i = num_coeffs; i--; ) {
1116 LLVMValueRef coeff = lp_build_const_scalar(type, coeffs[i]);
1118 res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res));
1131 * Minimax polynomial fit of 2**x, in range [-0.5, 0.5[
1133 const double lp_build_exp2_polynomial[] = {
1134 #if EXP_POLY_DEGREE == 5
1135 9.9999994e-1, 6.9315308e-1, 2.4015361e-1, 5.5826318e-2, 8.9893397e-3, 1.8775767e-3
1136 #elif EXP_POLY_DEGREE == 4
1137 1.0000026, 6.9300383e-1, 2.4144275e-1, 5.2011464e-2, 1.3534167e-2
1138 #elif EXP_POLY_DEGREE == 3
1139 9.9992520e-1, 6.9583356e-1, 2.2606716e-1, 7.8024521e-2
1140 #elif EXP_POLY_DEGREE == 2
1141 1.0017247, 6.5763628e-1, 3.3718944e-1
1149 lp_build_exp2_approx(struct lp_build_context *bld,
1151 LLVMValueRef *p_exp2_int_part,
1152 LLVMValueRef *p_frac_part,
1153 LLVMValueRef *p_exp2)
1155 const struct lp_type type = bld->type;
1156 LLVMTypeRef vec_type = lp_build_vec_type(type);
1157 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
1158 LLVMValueRef ipart = NULL;
1159 LLVMValueRef fpart = NULL;
1160 LLVMValueRef expipart = NULL;
1161 LLVMValueRef expfpart = NULL;
1162 LLVMValueRef res = NULL;
1164 if(p_exp2_int_part || p_frac_part || p_exp2) {
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 x = lp_build_min(bld, x, lp_build_const_scalar(type, 129.0));
1173 x = lp_build_max(bld, x, lp_build_const_scalar(type, -126.99999));
1175 /* ipart = int(x - 0.5) */
1176 ipart = LLVMBuildSub(bld->builder, x, lp_build_const_scalar(type, 0.5f), "");
1177 ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");
1179 /* fpart = x - ipart */
1180 fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
1181 fpart = LLVMBuildSub(bld->builder, x, fpart, "");
1184 if(p_exp2_int_part || p_exp2) {
1185 /* expipart = (float) (1 << ipart) */
1186 expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_int_const_scalar(type, 127), "");
1187 expipart = LLVMBuildShl(bld->builder, expipart, lp_build_int_const_scalar(type, 23), "");
1188 expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, "");
1192 expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
1193 Elements(lp_build_exp2_polynomial));
1195 res = LLVMBuildMul(bld->builder, expipart, expfpart, "");
1199 *p_exp2_int_part = expipart;
1202 *p_frac_part = fpart;
1210 lp_build_exp2(struct lp_build_context *bld,
1214 lp_build_exp2_approx(bld, x, NULL, NULL, &res);
1220 * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
1221 * These coefficients can be generate with
1222 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
1224 const double lp_build_log2_polynomial[] = {
1225 #if LOG_POLY_DEGREE == 6
1226 3.11578814719469302614, -3.32419399085241980044, 2.59883907202499966007, -1.23152682416275988241, 0.318212422185251071475, -0.0344359067839062357313
1227 #elif LOG_POLY_DEGREE == 5
1228 2.8882704548164776201, -2.52074962577807006663, 1.48116647521213171641, -0.465725644288844778798, 0.0596515482674574969533
1229 #elif LOG_POLY_DEGREE == 4
1230 2.61761038894603480148, -1.75647175389045657003, 0.688243882994381274313, -0.107254423828329604454
1231 #elif LOG_POLY_DEGREE == 3
1232 2.28330284476918490682, -1.04913055217340124191, 0.204446009836232697516
1240 * See http://www.devmaster.net/forums/showthread.php?p=43580
1243 lp_build_log2_approx(struct lp_build_context *bld,
1245 LLVMValueRef *p_exp,
1246 LLVMValueRef *p_floor_log2,
1247 LLVMValueRef *p_log2)
1249 const struct lp_type type = bld->type;
1250 LLVMTypeRef vec_type = lp_build_vec_type(type);
1251 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
1253 LLVMValueRef expmask = lp_build_int_const_scalar(type, 0x7f800000);
1254 LLVMValueRef mantmask = lp_build_int_const_scalar(type, 0x007fffff);
1255 LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type);
1257 LLVMValueRef i = NULL;
1258 LLVMValueRef exp = NULL;
1259 LLVMValueRef mant = NULL;
1260 LLVMValueRef logexp = NULL;
1261 LLVMValueRef logmant = NULL;
1262 LLVMValueRef res = NULL;
1264 if(p_exp || p_floor_log2 || p_log2) {
1265 /* TODO: optimize the constant case */
1266 if(LLVMIsConstant(x))
1267 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
1270 assert(type.floating && type.width == 32);
1272 i = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");
1274 /* exp = (float) exponent(x) */
1275 exp = LLVMBuildAnd(bld->builder, i, expmask, "");
1278 if(p_floor_log2 || p_log2) {
1279 logexp = LLVMBuildLShr(bld->builder, exp, lp_build_int_const_scalar(type, 23), "");
1280 logexp = LLVMBuildSub(bld->builder, logexp, lp_build_int_const_scalar(type, 127), "");
1281 logexp = LLVMBuildSIToFP(bld->builder, logexp, vec_type, "");
1285 /* mant = (float) mantissa(x) */
1286 mant = LLVMBuildAnd(bld->builder, i, mantmask, "");
1287 mant = LLVMBuildOr(bld->builder, mant, one, "");
1288 mant = LLVMBuildBitCast(bld->builder, mant, vec_type, "");
1290 logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial,
1291 Elements(lp_build_log2_polynomial));
1293 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
1294 logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildSub(bld->builder, mant, bld->one, ""), "");
1296 res = LLVMBuildAdd(bld->builder, logmant, logexp, "");
1303 *p_floor_log2 = logexp;
1311 lp_build_log2(struct lp_build_context *bld,
1315 lp_build_log2_approx(bld, x, NULL, NULL, &res);