1 /**************************************************************************
3 * Copyright 2009-2010 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
33 * LLVM IR doesn't support all basic arithmetic operations we care about (most
34 * notably min/max and saturated operations), and it is often necessary to
35 * resort machine-specific intrinsics directly. The functions here hide all
36 * these implementation details from the other modules.
38 * We also do simple expressions simplification here. Reasons are:
39 * - it is very easy given we have all necessary information readily available
40 * - LLVM optimization passes fail to simplify several vector expressions
41 * - We often know value constraints which the optimization passes have no way
42 * of knowing, such as when source arguments are known to be in [0, 1] range.
44 * @author Jose Fonseca <jfonseca@vmware.com>
50 #include "util/u_memory.h"
51 #include "util/u_debug.h"
52 #include "util/u_math.h"
53 #include "util/u_cpu_detect.h"
55 #include "lp_bld_type.h"
56 #include "lp_bld_const.h"
57 #include "lp_bld_init.h"
58 #include "lp_bld_intr.h"
59 #include "lp_bld_logic.h"
60 #include "lp_bld_pack.h"
61 #include "lp_bld_debug.h"
62 #include "lp_bld_bitarit.h"
63 #include "lp_bld_arit.h"
64 #include "lp_bld_flow.h"
66 #if defined(PIPE_ARCH_SSE)
67 #include <xmmintrin.h>
70 #ifndef _MM_DENORMALS_ZERO_MASK
71 #define _MM_DENORMALS_ZERO_MASK 0x0040
74 #ifndef _MM_FLUSH_ZERO_MASK
75 #define _MM_FLUSH_ZERO_MASK 0x8000
78 #define EXP_POLY_DEGREE 5
80 #define LOG_POLY_DEGREE 4
85 * No checks for special case values of a or b = 1 or 0 are done.
86 * NaN's are handled according to the behavior specified by the
87 * nan_behavior argument.
90 lp_build_min_simple(struct lp_build_context
*bld
,
93 enum gallivm_nan_behavior nan_behavior
)
95 const struct lp_type type
= bld
->type
;
96 const char *intrinsic
= NULL
;
97 unsigned intr_size
= 0;
100 assert(lp_check_value(type
, a
));
101 assert(lp_check_value(type
, b
));
103 /* TODO: optimize the constant case */
105 if (type
.floating
&& util_cpu_caps
.has_sse
) {
106 if (type
.width
== 32) {
107 if (type
.length
== 1) {
108 intrinsic
= "llvm.x86.sse.min.ss";
111 else if (type
.length
<= 4 || !util_cpu_caps
.has_avx
) {
112 intrinsic
= "llvm.x86.sse.min.ps";
116 intrinsic
= "llvm.x86.avx.min.ps.256";
120 if (type
.width
== 64 && util_cpu_caps
.has_sse2
) {
121 if (type
.length
== 1) {
122 intrinsic
= "llvm.x86.sse2.min.sd";
125 else if (type
.length
== 2 || !util_cpu_caps
.has_avx
) {
126 intrinsic
= "llvm.x86.sse2.min.pd";
130 intrinsic
= "llvm.x86.avx.min.pd.256";
135 else if (type
.floating
&& util_cpu_caps
.has_altivec
) {
136 if (nan_behavior
== GALLIVM_NAN_RETURN_NAN
||
137 nan_behavior
== GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
) {
138 debug_printf("%s: altivec doesn't support nan return nan behavior\n",
141 if (type
.width
== 32 && type
.length
== 4) {
142 intrinsic
= "llvm.ppc.altivec.vminfp";
145 } else if (HAVE_LLVM
< 0x0309 &&
146 util_cpu_caps
.has_avx2
&& type
.length
> 4) {
148 switch (type
.width
) {
150 intrinsic
= type
.sign
? "llvm.x86.avx2.pmins.b" : "llvm.x86.avx2.pminu.b";
153 intrinsic
= type
.sign
? "llvm.x86.avx2.pmins.w" : "llvm.x86.avx2.pminu.w";
156 intrinsic
= type
.sign
? "llvm.x86.avx2.pmins.d" : "llvm.x86.avx2.pminu.d";
159 } else if (HAVE_LLVM
< 0x0309 &&
160 util_cpu_caps
.has_sse2
&& type
.length
>= 2) {
162 if ((type
.width
== 8 || type
.width
== 16) &&
163 (type
.width
* type
.length
<= 64) &&
164 (gallivm_debug
& GALLIVM_DEBUG_PERF
)) {
165 debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
168 if (type
.width
== 8 && !type
.sign
) {
169 intrinsic
= "llvm.x86.sse2.pminu.b";
171 else if (type
.width
== 16 && type
.sign
) {
172 intrinsic
= "llvm.x86.sse2.pmins.w";
174 if (util_cpu_caps
.has_sse4_1
) {
175 if (type
.width
== 8 && type
.sign
) {
176 intrinsic
= "llvm.x86.sse41.pminsb";
178 if (type
.width
== 16 && !type
.sign
) {
179 intrinsic
= "llvm.x86.sse41.pminuw";
181 if (type
.width
== 32 && !type
.sign
) {
182 intrinsic
= "llvm.x86.sse41.pminud";
184 if (type
.width
== 32 && type
.sign
) {
185 intrinsic
= "llvm.x86.sse41.pminsd";
188 } else if (util_cpu_caps
.has_altivec
) {
190 if (type
.width
== 8) {
192 intrinsic
= "llvm.ppc.altivec.vminub";
194 intrinsic
= "llvm.ppc.altivec.vminsb";
196 } else if (type
.width
== 16) {
198 intrinsic
= "llvm.ppc.altivec.vminuh";
200 intrinsic
= "llvm.ppc.altivec.vminsh";
202 } else if (type
.width
== 32) {
204 intrinsic
= "llvm.ppc.altivec.vminuw";
206 intrinsic
= "llvm.ppc.altivec.vminsw";
212 /* We need to handle nan's for floating point numbers. If one of the
213 * inputs is nan the other should be returned (required by both D3D10+
215 * The sse intrinsics return the second operator in case of nan by
216 * default so we need to special code to handle those.
218 if (util_cpu_caps
.has_sse
&& type
.floating
&&
219 nan_behavior
!= GALLIVM_NAN_BEHAVIOR_UNDEFINED
&&
220 nan_behavior
!= GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
&&
221 nan_behavior
!= GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
) {
222 LLVMValueRef isnan
, min
;
223 min
= lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
226 if (nan_behavior
== GALLIVM_NAN_RETURN_OTHER
) {
227 isnan
= lp_build_isnan(bld
, b
);
228 return lp_build_select(bld
, isnan
, a
, min
);
230 assert(nan_behavior
== GALLIVM_NAN_RETURN_NAN
);
231 isnan
= lp_build_isnan(bld
, a
);
232 return lp_build_select(bld
, isnan
, a
, min
);
235 return lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
242 switch (nan_behavior
) {
243 case GALLIVM_NAN_RETURN_NAN
: {
244 LLVMValueRef isnan
= lp_build_isnan(bld
, b
);
245 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
246 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
247 return lp_build_select(bld
, cond
, a
, b
);
250 case GALLIVM_NAN_RETURN_OTHER
: {
251 LLVMValueRef isnan
= lp_build_isnan(bld
, a
);
252 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
253 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
254 return lp_build_select(bld
, cond
, a
, b
);
257 case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
:
258 cond
= lp_build_cmp_ordered(bld
, PIPE_FUNC_LESS
, a
, b
);
259 return lp_build_select(bld
, cond
, a
, b
);
260 case GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
:
261 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, b
, a
);
262 return lp_build_select(bld
, cond
, b
, a
);
263 case GALLIVM_NAN_BEHAVIOR_UNDEFINED
:
264 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
265 return lp_build_select(bld
, cond
, a
, b
);
269 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
270 return lp_build_select(bld
, cond
, a
, b
);
273 cond
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, a
, b
);
274 return lp_build_select(bld
, cond
, a
, b
);
280 lp_build_fmuladd(LLVMBuilderRef builder
,
285 LLVMTypeRef type
= LLVMTypeOf(a
);
286 assert(type
== LLVMTypeOf(b
));
287 assert(type
== LLVMTypeOf(c
));
288 if (HAVE_LLVM
< 0x0304) {
289 /* XXX: LLVM 3.3 does not breakdown llvm.fmuladd into mul+add when FMA is
290 * not supported, and instead it falls-back to a C function.
292 return LLVMBuildFAdd(builder
, LLVMBuildFMul(builder
, a
, b
, ""), c
, "");
295 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.fmuladd", type
);
296 LLVMValueRef args
[] = { a
, b
, c
};
297 return lp_build_intrinsic(builder
, intrinsic
, type
, args
, 3, 0);
303 * No checks for special case values of a or b = 1 or 0 are done.
304 * NaN's are handled according to the behavior specified by the
305 * nan_behavior argument.
308 lp_build_max_simple(struct lp_build_context
*bld
,
311 enum gallivm_nan_behavior nan_behavior
)
313 const struct lp_type type
= bld
->type
;
314 const char *intrinsic
= NULL
;
315 unsigned intr_size
= 0;
318 assert(lp_check_value(type
, a
));
319 assert(lp_check_value(type
, b
));
321 /* TODO: optimize the constant case */
323 if (type
.floating
&& util_cpu_caps
.has_sse
) {
324 if (type
.width
== 32) {
325 if (type
.length
== 1) {
326 intrinsic
= "llvm.x86.sse.max.ss";
329 else if (type
.length
<= 4 || !util_cpu_caps
.has_avx
) {
330 intrinsic
= "llvm.x86.sse.max.ps";
334 intrinsic
= "llvm.x86.avx.max.ps.256";
338 if (type
.width
== 64 && util_cpu_caps
.has_sse2
) {
339 if (type
.length
== 1) {
340 intrinsic
= "llvm.x86.sse2.max.sd";
343 else if (type
.length
== 2 || !util_cpu_caps
.has_avx
) {
344 intrinsic
= "llvm.x86.sse2.max.pd";
348 intrinsic
= "llvm.x86.avx.max.pd.256";
353 else if (type
.floating
&& util_cpu_caps
.has_altivec
) {
354 if (nan_behavior
== GALLIVM_NAN_RETURN_NAN
||
355 nan_behavior
== GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
) {
356 debug_printf("%s: altivec doesn't support nan return nan behavior\n",
359 if (type
.width
== 32 || type
.length
== 4) {
360 intrinsic
= "llvm.ppc.altivec.vmaxfp";
363 } else if (HAVE_LLVM
< 0x0309 &&
364 util_cpu_caps
.has_avx2
&& type
.length
> 4) {
366 switch (type
.width
) {
368 intrinsic
= type
.sign
? "llvm.x86.avx2.pmaxs.b" : "llvm.x86.avx2.pmaxu.b";
371 intrinsic
= type
.sign
? "llvm.x86.avx2.pmaxs.w" : "llvm.x86.avx2.pmaxu.w";
374 intrinsic
= type
.sign
? "llvm.x86.avx2.pmaxs.d" : "llvm.x86.avx2.pmaxu.d";
377 } else if (HAVE_LLVM
< 0x0309 &&
378 util_cpu_caps
.has_sse2
&& type
.length
>= 2) {
380 if ((type
.width
== 8 || type
.width
== 16) &&
381 (type
.width
* type
.length
<= 64) &&
382 (gallivm_debug
& GALLIVM_DEBUG_PERF
)) {
383 debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
386 if (type
.width
== 8 && !type
.sign
) {
387 intrinsic
= "llvm.x86.sse2.pmaxu.b";
390 else if (type
.width
== 16 && type
.sign
) {
391 intrinsic
= "llvm.x86.sse2.pmaxs.w";
393 if (util_cpu_caps
.has_sse4_1
) {
394 if (type
.width
== 8 && type
.sign
) {
395 intrinsic
= "llvm.x86.sse41.pmaxsb";
397 if (type
.width
== 16 && !type
.sign
) {
398 intrinsic
= "llvm.x86.sse41.pmaxuw";
400 if (type
.width
== 32 && !type
.sign
) {
401 intrinsic
= "llvm.x86.sse41.pmaxud";
403 if (type
.width
== 32 && type
.sign
) {
404 intrinsic
= "llvm.x86.sse41.pmaxsd";
407 } else if (util_cpu_caps
.has_altivec
) {
409 if (type
.width
== 8) {
411 intrinsic
= "llvm.ppc.altivec.vmaxub";
413 intrinsic
= "llvm.ppc.altivec.vmaxsb";
415 } else if (type
.width
== 16) {
417 intrinsic
= "llvm.ppc.altivec.vmaxuh";
419 intrinsic
= "llvm.ppc.altivec.vmaxsh";
421 } else if (type
.width
== 32) {
423 intrinsic
= "llvm.ppc.altivec.vmaxuw";
425 intrinsic
= "llvm.ppc.altivec.vmaxsw";
431 if (util_cpu_caps
.has_sse
&& type
.floating
&&
432 nan_behavior
!= GALLIVM_NAN_BEHAVIOR_UNDEFINED
&&
433 nan_behavior
!= GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
&&
434 nan_behavior
!= GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
) {
435 LLVMValueRef isnan
, max
;
436 max
= lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
439 if (nan_behavior
== GALLIVM_NAN_RETURN_OTHER
) {
440 isnan
= lp_build_isnan(bld
, b
);
441 return lp_build_select(bld
, isnan
, a
, max
);
443 assert(nan_behavior
== GALLIVM_NAN_RETURN_NAN
);
444 isnan
= lp_build_isnan(bld
, a
);
445 return lp_build_select(bld
, isnan
, a
, max
);
448 return lp_build_intrinsic_binary_anylength(bld
->gallivm
, intrinsic
,
455 switch (nan_behavior
) {
456 case GALLIVM_NAN_RETURN_NAN
: {
457 LLVMValueRef isnan
= lp_build_isnan(bld
, b
);
458 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
459 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
460 return lp_build_select(bld
, cond
, a
, b
);
463 case GALLIVM_NAN_RETURN_OTHER
: {
464 LLVMValueRef isnan
= lp_build_isnan(bld
, a
);
465 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
466 cond
= LLVMBuildXor(bld
->gallivm
->builder
, cond
, isnan
, "");
467 return lp_build_select(bld
, cond
, a
, b
);
470 case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
:
471 cond
= lp_build_cmp_ordered(bld
, PIPE_FUNC_GREATER
, a
, b
);
472 return lp_build_select(bld
, cond
, a
, b
);
473 case GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
:
474 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, b
, a
);
475 return lp_build_select(bld
, cond
, b
, a
);
476 case GALLIVM_NAN_BEHAVIOR_UNDEFINED
:
477 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
478 return lp_build_select(bld
, cond
, a
, b
);
482 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
483 return lp_build_select(bld
, cond
, a
, b
);
486 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
487 return lp_build_select(bld
, cond
, a
, b
);
493 * Generate 1 - a, or ~a depending on bld->type.
496 lp_build_comp(struct lp_build_context
*bld
,
499 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
500 const struct lp_type type
= bld
->type
;
502 assert(lp_check_value(type
, a
));
509 if(type
.norm
&& !type
.floating
&& !type
.fixed
&& !type
.sign
) {
510 if(LLVMIsConstant(a
))
511 return LLVMConstNot(a
);
513 return LLVMBuildNot(builder
, a
, "");
516 if(LLVMIsConstant(a
))
518 return LLVMConstFSub(bld
->one
, a
);
520 return LLVMConstSub(bld
->one
, a
);
523 return LLVMBuildFSub(builder
, bld
->one
, a
, "");
525 return LLVMBuildSub(builder
, bld
->one
, a
, "");
533 lp_build_add(struct lp_build_context
*bld
,
537 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
538 const struct lp_type type
= bld
->type
;
541 assert(lp_check_value(type
, a
));
542 assert(lp_check_value(type
, b
));
548 if(a
== bld
->undef
|| b
== bld
->undef
)
552 const char *intrinsic
= NULL
;
554 if(a
== bld
->one
|| b
== bld
->one
)
557 if (!type
.floating
&& !type
.fixed
) {
558 if (type
.width
* type
.length
== 128) {
559 if(util_cpu_caps
.has_sse2
) {
561 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
563 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
564 } else if (util_cpu_caps
.has_altivec
) {
566 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddsbs" : "llvm.ppc.altivec.vaddubs";
568 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddshs" : "llvm.ppc.altivec.vadduhs";
571 if (type
.width
* type
.length
== 256) {
572 if(util_cpu_caps
.has_avx2
) {
574 intrinsic
= type
.sign
? "llvm.x86.avx2.padds.b" : "llvm.x86.avx2.paddus.b";
576 intrinsic
= type
.sign
? "llvm.x86.avx2.padds.w" : "llvm.x86.avx2.paddus.w";
582 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
585 if(type
.norm
&& !type
.floating
&& !type
.fixed
) {
587 uint64_t sign
= (uint64_t)1 << (type
.width
- 1);
588 LLVMValueRef max_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
- 1);
589 LLVMValueRef min_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
);
590 /* a_clamp_max is the maximum a for positive b,
591 a_clamp_min is the minimum a for negative b. */
592 LLVMValueRef a_clamp_max
= lp_build_min_simple(bld
, a
, LLVMBuildSub(builder
, max_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
593 LLVMValueRef a_clamp_min
= lp_build_max_simple(bld
, a
, LLVMBuildSub(builder
, min_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
594 a
= lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, b
, bld
->zero
), a_clamp_max
, a_clamp_min
);
596 a
= lp_build_min_simple(bld
, a
, lp_build_comp(bld
, b
), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
600 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
602 res
= LLVMConstFAdd(a
, b
);
604 res
= LLVMConstAdd(a
, b
);
607 res
= LLVMBuildFAdd(builder
, a
, b
, "");
609 res
= LLVMBuildAdd(builder
, a
, b
, "");
611 /* clamp to ceiling of 1.0 */
612 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
613 res
= lp_build_min_simple(bld
, res
, bld
->one
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
615 /* XXX clamp to floor of -1 or 0??? */
621 /** Return the scalar sum of the elements of a.
622 * Should avoid this operation whenever possible.
625 lp_build_horizontal_add(struct lp_build_context
*bld
,
628 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
629 const struct lp_type type
= bld
->type
;
630 LLVMValueRef index
, res
;
632 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
/ 2];
633 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
/ 2];
634 LLVMValueRef vecres
, elem2
;
636 assert(lp_check_value(type
, a
));
638 if (type
.length
== 1) {
642 assert(!bld
->type
.norm
);
645 * for byte vectors can do much better with psadbw.
646 * Using repeated shuffle/adds here. Note with multiple vectors
647 * this can be done more efficiently as outlined in the intel
648 * optimization manual.
649 * Note: could cause data rearrangement if used with smaller element
654 length
= type
.length
/ 2;
656 LLVMValueRef vec1
, vec2
;
657 for (i
= 0; i
< length
; i
++) {
658 shuffles1
[i
] = lp_build_const_int32(bld
->gallivm
, i
);
659 shuffles2
[i
] = lp_build_const_int32(bld
->gallivm
, i
+ length
);
661 vec1
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
662 LLVMConstVector(shuffles1
, length
), "");
663 vec2
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
664 LLVMConstVector(shuffles2
, length
), "");
666 vecres
= LLVMBuildFAdd(builder
, vec1
, vec2
, "");
669 vecres
= LLVMBuildAdd(builder
, vec1
, vec2
, "");
671 length
= length
>> 1;
674 /* always have vector of size 2 here */
677 index
= lp_build_const_int32(bld
->gallivm
, 0);
678 res
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
679 index
= lp_build_const_int32(bld
->gallivm
, 1);
680 elem2
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
683 res
= LLVMBuildFAdd(builder
, res
, elem2
, "");
685 res
= LLVMBuildAdd(builder
, res
, elem2
, "");
691 * Return the horizontal sums of 4 float vectors as a float4 vector.
692 * This uses the technique as outlined in Intel Optimization Manual.
695 lp_build_horizontal_add4x4f(struct lp_build_context
*bld
,
698 struct gallivm_state
*gallivm
= bld
->gallivm
;
699 LLVMBuilderRef builder
= gallivm
->builder
;
700 LLVMValueRef shuffles
[4];
702 LLVMValueRef sumtmp
[2], shuftmp
[2];
704 /* lower half of regs */
705 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
706 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
707 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
708 shuffles
[3] = lp_build_const_int32(gallivm
, 5);
709 tmp
[0] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
710 LLVMConstVector(shuffles
, 4), "");
711 tmp
[2] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
712 LLVMConstVector(shuffles
, 4), "");
714 /* upper half of regs */
715 shuffles
[0] = lp_build_const_int32(gallivm
, 2);
716 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
717 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
718 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
719 tmp
[1] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
720 LLVMConstVector(shuffles
, 4), "");
721 tmp
[3] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
722 LLVMConstVector(shuffles
, 4), "");
724 sumtmp
[0] = LLVMBuildFAdd(builder
, tmp
[0], tmp
[1], "");
725 sumtmp
[1] = LLVMBuildFAdd(builder
, tmp
[2], tmp
[3], "");
727 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
728 shuffles
[1] = lp_build_const_int32(gallivm
, 2);
729 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
730 shuffles
[3] = lp_build_const_int32(gallivm
, 6);
731 shuftmp
[0] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
732 LLVMConstVector(shuffles
, 4), "");
734 shuffles
[0] = lp_build_const_int32(gallivm
, 1);
735 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
736 shuffles
[2] = lp_build_const_int32(gallivm
, 5);
737 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
738 shuftmp
[1] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
739 LLVMConstVector(shuffles
, 4), "");
741 return LLVMBuildFAdd(builder
, shuftmp
[0], shuftmp
[1], "");
746 * partially horizontally add 2-4 float vectors with length nx4,
747 * i.e. only four adjacent values in each vector will be added,
748 * assuming values are really grouped in 4 which also determines
751 * Return a vector of the same length as the initial vectors,
752 * with the excess elements (if any) being undefined.
753 * The element order is independent of number of input vectors.
754 * For 3 vectors x0x1x2x3x4x5x6x7, y0y1y2y3y4y5y6y7, z0z1z2z3z4z5z6z7
755 * the output order thus will be
756 * sumx0-x3,sumy0-y3,sumz0-z3,undef,sumx4-x7,sumy4-y7,sumz4z7,undef
759 lp_build_hadd_partial4(struct lp_build_context
*bld
,
760 LLVMValueRef vectors
[],
763 struct gallivm_state
*gallivm
= bld
->gallivm
;
764 LLVMBuilderRef builder
= gallivm
->builder
;
765 LLVMValueRef ret_vec
;
767 const char *intrinsic
= NULL
;
769 assert(num_vecs
>= 2 && num_vecs
<= 4);
770 assert(bld
->type
.floating
);
772 /* only use this with at least 2 vectors, as it is sort of expensive
773 * (depending on cpu) and we always need two horizontal adds anyway,
774 * so a shuffle/add approach might be better.
780 tmp
[2] = num_vecs
> 2 ? vectors
[2] : vectors
[0];
781 tmp
[3] = num_vecs
> 3 ? vectors
[3] : vectors
[0];
783 if (util_cpu_caps
.has_sse3
&& bld
->type
.width
== 32 &&
784 bld
->type
.length
== 4) {
785 intrinsic
= "llvm.x86.sse3.hadd.ps";
787 else if (util_cpu_caps
.has_avx
&& bld
->type
.width
== 32 &&
788 bld
->type
.length
== 8) {
789 intrinsic
= "llvm.x86.avx.hadd.ps.256";
792 tmp
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
793 lp_build_vec_type(gallivm
, bld
->type
),
796 tmp
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
797 lp_build_vec_type(gallivm
, bld
->type
),
803 return lp_build_intrinsic_binary(builder
, intrinsic
,
804 lp_build_vec_type(gallivm
, bld
->type
),
808 if (bld
->type
.length
== 4) {
809 ret_vec
= lp_build_horizontal_add4x4f(bld
, tmp
);
812 LLVMValueRef partres
[LP_MAX_VECTOR_LENGTH
/4];
814 unsigned num_iter
= bld
->type
.length
/ 4;
815 struct lp_type parttype
= bld
->type
;
817 for (j
= 0; j
< num_iter
; j
++) {
818 LLVMValueRef partsrc
[4];
820 for (i
= 0; i
< 4; i
++) {
821 partsrc
[i
] = lp_build_extract_range(gallivm
, tmp
[i
], j
*4, 4);
823 partres
[j
] = lp_build_horizontal_add4x4f(bld
, partsrc
);
825 ret_vec
= lp_build_concat(gallivm
, partres
, parttype
, num_iter
);
834 lp_build_sub(struct lp_build_context
*bld
,
838 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
839 const struct lp_type type
= bld
->type
;
842 assert(lp_check_value(type
, a
));
843 assert(lp_check_value(type
, b
));
847 if(a
== bld
->undef
|| b
== bld
->undef
)
853 const char *intrinsic
= NULL
;
858 if (!type
.floating
&& !type
.fixed
) {
859 if (type
.width
* type
.length
== 128) {
860 if (util_cpu_caps
.has_sse2
) {
862 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
864 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
865 } else if (util_cpu_caps
.has_altivec
) {
867 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubsbs" : "llvm.ppc.altivec.vsububs";
869 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubshs" : "llvm.ppc.altivec.vsubuhs";
872 if (type
.width
* type
.length
== 256) {
873 if (util_cpu_caps
.has_avx2
) {
875 intrinsic
= type
.sign
? "llvm.x86.avx2.psubs.b" : "llvm.x86.avx2.psubus.b";
877 intrinsic
= type
.sign
? "llvm.x86.avx2.psubs.w" : "llvm.x86.avx2.psubus.w";
883 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
886 if(type
.norm
&& !type
.floating
&& !type
.fixed
) {
888 uint64_t sign
= (uint64_t)1 << (type
.width
- 1);
889 LLVMValueRef max_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
- 1);
890 LLVMValueRef min_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
);
891 /* a_clamp_max is the maximum a for negative b,
892 a_clamp_min is the minimum a for positive b. */
893 LLVMValueRef a_clamp_max
= lp_build_min_simple(bld
, a
, LLVMBuildAdd(builder
, max_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
894 LLVMValueRef a_clamp_min
= lp_build_max_simple(bld
, a
, LLVMBuildAdd(builder
, min_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
895 a
= lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, b
, bld
->zero
), a_clamp_min
, a_clamp_max
);
897 a
= lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
901 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
903 res
= LLVMConstFSub(a
, b
);
905 res
= LLVMConstSub(a
, b
);
908 res
= LLVMBuildFSub(builder
, a
, b
, "");
910 res
= LLVMBuildSub(builder
, a
, b
, "");
912 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
913 res
= lp_build_max_simple(bld
, res
, bld
->zero
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
921 * Normalized multiplication.
923 * There are several approaches for (using 8-bit normalized multiplication as
928 * makes the following approximation to the division (Sree)
930 * a*b/255 ~= (a*(b + 1)) >> 256
932 * which is the fastest method that satisfies the following OpenGL criteria of
934 * 0*0 = 0 and 255*255 = 255
938 * takes the geometric series approximation to the division
940 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
942 * in this case just the first two terms to fit in 16bit arithmetic
944 * t/255 ~= (t + (t >> 8)) >> 8
946 * note that just by itself it doesn't satisfies the OpenGL criteria, as
947 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
950 * - geometric series plus rounding
952 * when using a geometric series division instead of truncating the result
953 * use roundoff in the approximation (Jim Blinn)
955 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
957 * achieving the exact results.
961 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
962 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
963 * @sa Michael Herf, The "double blend trick", May 2000,
964 * http://www.stereopsis.com/doubleblend.html
967 lp_build_mul_norm(struct gallivm_state
*gallivm
,
968 struct lp_type wide_type
,
969 LLVMValueRef a
, LLVMValueRef b
)
971 LLVMBuilderRef builder
= gallivm
->builder
;
972 struct lp_build_context bld
;
977 assert(!wide_type
.floating
);
978 assert(lp_check_value(wide_type
, a
));
979 assert(lp_check_value(wide_type
, b
));
981 lp_build_context_init(&bld
, gallivm
, wide_type
);
983 n
= wide_type
.width
/ 2;
984 if (wide_type
.sign
) {
989 * TODO: for 16bits normalized SSE2 vectors we could consider using PMULHUW
990 * http://ssp.impulsetrain.com/2011/07/03/multiplying-normalized-16-bit-numbers-with-sse2/
994 * a*b / (2**n - 1) ~= (a*b + (a*b >> n) + half) >> n
997 ab
= LLVMBuildMul(builder
, a
, b
, "");
998 ab
= LLVMBuildAdd(builder
, ab
, lp_build_shr_imm(&bld
, ab
, n
), "");
1001 * half = sgn(ab) * 0.5 * (2 ** n) = sgn(ab) * (1 << (n - 1))
1004 half
= lp_build_const_int_vec(gallivm
, wide_type
, 1LL << (n
- 1));
1005 if (wide_type
.sign
) {
1006 LLVMValueRef minus_half
= LLVMBuildNeg(builder
, half
, "");
1007 LLVMValueRef sign
= lp_build_shr_imm(&bld
, ab
, wide_type
.width
- 1);
1008 half
= lp_build_select(&bld
, sign
, minus_half
, half
);
1010 ab
= LLVMBuildAdd(builder
, ab
, half
, "");
1012 /* Final division */
1013 ab
= lp_build_shr_imm(&bld
, ab
, n
);
1022 lp_build_mul(struct lp_build_context
*bld
,
1026 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1027 const struct lp_type type
= bld
->type
;
1031 assert(lp_check_value(type
, a
));
1032 assert(lp_check_value(type
, b
));
1042 if(a
== bld
->undef
|| b
== bld
->undef
)
1045 if (!type
.floating
&& !type
.fixed
&& type
.norm
) {
1046 struct lp_type wide_type
= lp_wider_type(type
);
1047 LLVMValueRef al
, ah
, bl
, bh
, abl
, abh
, ab
;
1049 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, a
, &al
, &ah
);
1050 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, b
, &bl
, &bh
);
1052 /* PMULLW, PSRLW, PADDW */
1053 abl
= lp_build_mul_norm(bld
->gallivm
, wide_type
, al
, bl
);
1054 abh
= lp_build_mul_norm(bld
->gallivm
, wide_type
, ah
, bh
);
1056 ab
= lp_build_pack2_native(bld
->gallivm
, wide_type
, type
, abl
, abh
);
1062 shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
/2);
1066 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
1068 res
= LLVMConstFMul(a
, b
);
1070 res
= LLVMConstMul(a
, b
);
1073 res
= LLVMConstAShr(res
, shift
);
1075 res
= LLVMConstLShr(res
, shift
);
1080 res
= LLVMBuildFMul(builder
, a
, b
, "");
1082 res
= LLVMBuildMul(builder
, a
, b
, "");
1085 res
= LLVMBuildAShr(builder
, res
, shift
, "");
1087 res
= LLVMBuildLShr(builder
, res
, shift
, "");
1095 * Widening mul, valid for 32x32 bit -> 64bit only.
1096 * Result is low 32bits, high bits returned in res_hi.
1098 * Emits code that is meant to be compiled for the host CPU.
1101 lp_build_mul_32_lohi_cpu(struct lp_build_context
*bld
,
1104 LLVMValueRef
*res_hi
)
1106 struct gallivm_state
*gallivm
= bld
->gallivm
;
1107 LLVMBuilderRef builder
= gallivm
->builder
;
1109 assert(bld
->type
.width
== 32);
1110 assert(bld
->type
.floating
== 0);
1111 assert(bld
->type
.fixed
== 0);
1112 assert(bld
->type
.norm
== 0);
1115 * XXX: for some reason, with zext/zext/mul/trunc the code llvm produces
1116 * for x86 simd is atrocious (even if the high bits weren't required),
1117 * trying to handle real 64bit inputs (which of course can't happen due
1118 * to using 64bit umul with 32bit numbers zero-extended to 64bit, but
1119 * apparently llvm does not recognize this widening mul). This includes 6
1120 * (instead of 2) pmuludq plus extra adds and shifts
1121 * The same story applies to signed mul, albeit fixing this requires sse41.
1122 * https://llvm.org/bugs/show_bug.cgi?id=30845
1123 * So, whip up our own code, albeit only for length 4 and 8 (which
1124 * should be good enough)...
1126 if ((bld
->type
.length
== 4 || bld
->type
.length
== 8) &&
1127 ((util_cpu_caps
.has_sse2
&& (bld
->type
.sign
== 0)) ||
1128 util_cpu_caps
.has_sse4_1
)) {
1129 const char *intrinsic
= NULL
;
1130 LLVMValueRef aeven
, aodd
, beven
, bodd
, muleven
, mulodd
;
1131 LLVMValueRef shuf
[LP_MAX_VECTOR_WIDTH
/ 32], shuf_vec
;
1132 struct lp_type type_wide
= lp_wider_type(bld
->type
);
1133 LLVMTypeRef wider_type
= lp_build_vec_type(gallivm
, type_wide
);
1135 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1136 shuf
[i
] = lp_build_const_int32(gallivm
, i
+1);
1137 shuf
[i
+1] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1139 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1142 aodd
= LLVMBuildShuffleVector(builder
, aeven
, bld
->undef
, shuf_vec
, "");
1143 bodd
= LLVMBuildShuffleVector(builder
, beven
, bld
->undef
, shuf_vec
, "");
1145 if (util_cpu_caps
.has_avx2
&& bld
->type
.length
== 8) {
1146 if (bld
->type
.sign
) {
1147 intrinsic
= "llvm.x86.avx2.pmul.dq";
1149 intrinsic
= "llvm.x86.avx2.pmulu.dq";
1151 muleven
= lp_build_intrinsic_binary(builder
, intrinsic
,
1152 wider_type
, aeven
, beven
);
1153 mulodd
= lp_build_intrinsic_binary(builder
, intrinsic
,
1154 wider_type
, aodd
, bodd
);
1157 /* for consistent naming look elsewhere... */
1158 if (bld
->type
.sign
) {
1159 intrinsic
= "llvm.x86.sse41.pmuldq";
1161 intrinsic
= "llvm.x86.sse2.pmulu.dq";
1164 * XXX If we only have AVX but not AVX2 this is a pain.
1165 * lp_build_intrinsic_binary_anylength() can't handle it
1166 * (due to src and dst type not being identical).
1168 if (bld
->type
.length
== 8) {
1169 LLVMValueRef aevenlo
, aevenhi
, bevenlo
, bevenhi
;
1170 LLVMValueRef aoddlo
, aoddhi
, boddlo
, boddhi
;
1171 LLVMValueRef muleven2
[2], mulodd2
[2];
1172 struct lp_type type_wide_half
= type_wide
;
1173 LLVMTypeRef wtype_half
;
1174 type_wide_half
.length
= 2;
1175 wtype_half
= lp_build_vec_type(gallivm
, type_wide_half
);
1176 aevenlo
= lp_build_extract_range(gallivm
, aeven
, 0, 4);
1177 aevenhi
= lp_build_extract_range(gallivm
, aeven
, 4, 4);
1178 bevenlo
= lp_build_extract_range(gallivm
, beven
, 0, 4);
1179 bevenhi
= lp_build_extract_range(gallivm
, beven
, 4, 4);
1180 aoddlo
= lp_build_extract_range(gallivm
, aodd
, 0, 4);
1181 aoddhi
= lp_build_extract_range(gallivm
, aodd
, 4, 4);
1182 boddlo
= lp_build_extract_range(gallivm
, bodd
, 0, 4);
1183 boddhi
= lp_build_extract_range(gallivm
, bodd
, 4, 4);
1184 muleven2
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
1185 wtype_half
, aevenlo
, bevenlo
);
1186 mulodd2
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
1187 wtype_half
, aoddlo
, boddlo
);
1188 muleven2
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
1189 wtype_half
, aevenhi
, bevenhi
);
1190 mulodd2
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
1191 wtype_half
, aoddhi
, boddhi
);
1192 muleven
= lp_build_concat(gallivm
, muleven2
, type_wide_half
, 2);
1193 mulodd
= lp_build_concat(gallivm
, mulodd2
, type_wide_half
, 2);
1197 muleven
= lp_build_intrinsic_binary(builder
, intrinsic
,
1198 wider_type
, aeven
, beven
);
1199 mulodd
= lp_build_intrinsic_binary(builder
, intrinsic
,
1200 wider_type
, aodd
, bodd
);
1203 muleven
= LLVMBuildBitCast(builder
, muleven
, bld
->vec_type
, "");
1204 mulodd
= LLVMBuildBitCast(builder
, mulodd
, bld
->vec_type
, "");
1206 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1207 shuf
[i
] = lp_build_const_int32(gallivm
, i
+ 1);
1208 shuf
[i
+1] = lp_build_const_int32(gallivm
, i
+ 1 + bld
->type
.length
);
1210 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1211 *res_hi
= LLVMBuildShuffleVector(builder
, muleven
, mulodd
, shuf_vec
, "");
1213 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1214 shuf
[i
] = lp_build_const_int32(gallivm
, i
);
1215 shuf
[i
+1] = lp_build_const_int32(gallivm
, i
+ bld
->type
.length
);
1217 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1218 return LLVMBuildShuffleVector(builder
, muleven
, mulodd
, shuf_vec
, "");
1221 return lp_build_mul_32_lohi(bld
, a
, b
, res_hi
);
1227 * Widening mul, valid for 32x32 bit -> 64bit only.
1228 * Result is low 32bits, high bits returned in res_hi.
1230 * Emits generic code.
1233 lp_build_mul_32_lohi(struct lp_build_context
*bld
,
1236 LLVMValueRef
*res_hi
)
1238 struct gallivm_state
*gallivm
= bld
->gallivm
;
1239 LLVMBuilderRef builder
= gallivm
->builder
;
1240 LLVMValueRef tmp
, shift
, res_lo
;
1241 struct lp_type type_tmp
;
1242 LLVMTypeRef wide_type
, narrow_type
;
1244 type_tmp
= bld
->type
;
1245 narrow_type
= lp_build_vec_type(gallivm
, type_tmp
);
1246 type_tmp
.width
*= 2;
1247 wide_type
= lp_build_vec_type(gallivm
, type_tmp
);
1248 shift
= lp_build_const_vec(gallivm
, type_tmp
, 32);
1250 if (bld
->type
.sign
) {
1251 a
= LLVMBuildSExt(builder
, a
, wide_type
, "");
1252 b
= LLVMBuildSExt(builder
, b
, wide_type
, "");
1254 a
= LLVMBuildZExt(builder
, a
, wide_type
, "");
1255 b
= LLVMBuildZExt(builder
, b
, wide_type
, "");
1257 tmp
= LLVMBuildMul(builder
, a
, b
, "");
1259 res_lo
= LLVMBuildTrunc(builder
, tmp
, narrow_type
, "");
1261 /* Since we truncate anyway, LShr and AShr are equivalent. */
1262 tmp
= LLVMBuildLShr(builder
, tmp
, shift
, "");
1263 *res_hi
= LLVMBuildTrunc(builder
, tmp
, narrow_type
, "");
1271 lp_build_mad(struct lp_build_context
*bld
,
1276 const struct lp_type type
= bld
->type
;
1277 if (type
.floating
) {
1278 return lp_build_fmuladd(bld
->gallivm
->builder
, a
, b
, c
);
1280 return lp_build_add(bld
, lp_build_mul(bld
, a
, b
), c
);
1286 * Small vector x scale multiplication optimization.
1289 lp_build_mul_imm(struct lp_build_context
*bld
,
1293 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1294 LLVMValueRef factor
;
1296 assert(lp_check_value(bld
->type
, a
));
1305 return lp_build_negate(bld
, a
);
1307 if(b
== 2 && bld
->type
.floating
)
1308 return lp_build_add(bld
, a
, a
);
1310 if(util_is_power_of_two(b
)) {
1311 unsigned shift
= ffs(b
) - 1;
1313 if(bld
->type
.floating
) {
1316 * Power of two multiplication by directly manipulating the exponent.
1318 * XXX: This might not be always faster, it will introduce a small error
1319 * for multiplication by zero, and it will produce wrong results
1322 unsigned mantissa
= lp_mantissa(bld
->type
);
1323 factor
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (unsigned long long)shift
<< mantissa
);
1324 a
= LLVMBuildBitCast(builder
, a
, lp_build_int_vec_type(bld
->type
), "");
1325 a
= LLVMBuildAdd(builder
, a
, factor
, "");
1326 a
= LLVMBuildBitCast(builder
, a
, lp_build_vec_type(bld
->gallivm
, bld
->type
), "");
1331 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, shift
);
1332 return LLVMBuildShl(builder
, a
, factor
, "");
1336 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, (double)b
);
1337 return lp_build_mul(bld
, a
, factor
);
1345 lp_build_div(struct lp_build_context
*bld
,
1349 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1350 const struct lp_type type
= bld
->type
;
1352 assert(lp_check_value(type
, a
));
1353 assert(lp_check_value(type
, b
));
1357 if(a
== bld
->one
&& type
.floating
)
1358 return lp_build_rcp(bld
, b
);
1363 if(a
== bld
->undef
|| b
== bld
->undef
)
1366 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
1368 return LLVMConstFDiv(a
, b
);
1370 return LLVMConstSDiv(a
, b
);
1372 return LLVMConstUDiv(a
, b
);
1375 /* fast rcp is disabled (just uses div), so makes no sense to try that */
1377 ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
1378 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) &&
1380 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
1383 return LLVMBuildFDiv(builder
, a
, b
, "");
1385 return LLVMBuildSDiv(builder
, a
, b
, "");
1387 return LLVMBuildUDiv(builder
, a
, b
, "");
1392 * Linear interpolation helper.
1394 * @param normalized whether we are interpolating normalized values,
1395 * encoded in normalized integers, twice as wide.
1397 * @sa http://www.stereopsis.com/doubleblend.html
1399 static inline LLVMValueRef
1400 lp_build_lerp_simple(struct lp_build_context
*bld
,
1406 unsigned half_width
= bld
->type
.width
/2;
1407 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1411 assert(lp_check_value(bld
->type
, x
));
1412 assert(lp_check_value(bld
->type
, v0
));
1413 assert(lp_check_value(bld
->type
, v1
));
1415 delta
= lp_build_sub(bld
, v1
, v0
);
1417 if (bld
->type
.floating
) {
1419 return lp_build_mad(bld
, x
, delta
, v0
);
1422 if (flags
& LP_BLD_LERP_WIDE_NORMALIZED
) {
1423 if (!bld
->type
.sign
) {
1424 if (!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
)) {
1426 * Scale x from [0, 2**n - 1] to [0, 2**n] by adding the
1427 * most-significant-bit to the lowest-significant-bit, so that
1428 * later we can just divide by 2**n instead of 2**n - 1.
1431 x
= lp_build_add(bld
, x
, lp_build_shr_imm(bld
, x
, half_width
- 1));
1434 /* (x * delta) >> n */
1435 res
= lp_build_mul(bld
, x
, delta
);
1436 res
= lp_build_shr_imm(bld
, res
, half_width
);
1439 * The rescaling trick above doesn't work for signed numbers, so
1440 * use the 2**n - 1 divison approximation in lp_build_mul_norm
1443 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1444 res
= lp_build_mul_norm(bld
->gallivm
, bld
->type
, x
, delta
);
1447 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1448 res
= lp_build_mul(bld
, x
, delta
);
1451 if ((flags
& LP_BLD_LERP_WIDE_NORMALIZED
) && !bld
->type
.sign
) {
1453 * At this point both res and v0 only use the lower half of the bits,
1454 * the rest is zero. Instead of add / mask, do add with half wide type.
1456 struct lp_type narrow_type
;
1457 struct lp_build_context narrow_bld
;
1459 memset(&narrow_type
, 0, sizeof narrow_type
);
1460 narrow_type
.sign
= bld
->type
.sign
;
1461 narrow_type
.width
= bld
->type
.width
/2;
1462 narrow_type
.length
= bld
->type
.length
*2;
1464 lp_build_context_init(&narrow_bld
, bld
->gallivm
, narrow_type
);
1465 res
= LLVMBuildBitCast(builder
, res
, narrow_bld
.vec_type
, "");
1466 v0
= LLVMBuildBitCast(builder
, v0
, narrow_bld
.vec_type
, "");
1467 res
= lp_build_add(&narrow_bld
, v0
, res
);
1468 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
1470 res
= lp_build_add(bld
, v0
, res
);
1472 if (bld
->type
.fixed
) {
1474 * We need to mask out the high order bits when lerping 8bit
1475 * normalized colors stored on 16bits
1477 /* XXX: This step is necessary for lerping 8bit colors stored on
1478 * 16bits, but it will be wrong for true fixed point use cases.
1479 * Basically we need a more powerful lp_type, capable of further
1480 * distinguishing the values interpretation from the value storage.
1482 LLVMValueRef low_bits
;
1483 low_bits
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (1 << half_width
) - 1);
1484 res
= LLVMBuildAnd(builder
, res
, low_bits
, "");
1493 * Linear interpolation.
1496 lp_build_lerp(struct lp_build_context
*bld
,
1502 const struct lp_type type
= bld
->type
;
1505 assert(lp_check_value(type
, x
));
1506 assert(lp_check_value(type
, v0
));
1507 assert(lp_check_value(type
, v1
));
1509 assert(!(flags
& LP_BLD_LERP_WIDE_NORMALIZED
));
1512 struct lp_type wide_type
;
1513 struct lp_build_context wide_bld
;
1514 LLVMValueRef xl
, xh
, v0l
, v0h
, v1l
, v1h
, resl
, resh
;
1516 assert(type
.length
>= 2);
1519 * Create a wider integer type, enough to hold the
1520 * intermediate result of the multiplication.
1522 memset(&wide_type
, 0, sizeof wide_type
);
1523 wide_type
.sign
= type
.sign
;
1524 wide_type
.width
= type
.width
*2;
1525 wide_type
.length
= type
.length
/2;
1527 lp_build_context_init(&wide_bld
, bld
->gallivm
, wide_type
);
1529 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, x
, &xl
, &xh
);
1530 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v0
, &v0l
, &v0h
);
1531 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v1
, &v1l
, &v1h
);
1537 flags
|= LP_BLD_LERP_WIDE_NORMALIZED
;
1539 resl
= lp_build_lerp_simple(&wide_bld
, xl
, v0l
, v1l
, flags
);
1540 resh
= lp_build_lerp_simple(&wide_bld
, xh
, v0h
, v1h
, flags
);
1542 res
= lp_build_pack2_native(bld
->gallivm
, wide_type
, type
, resl
, resh
);
1544 res
= lp_build_lerp_simple(bld
, x
, v0
, v1
, flags
);
1552 * Bilinear interpolation.
1554 * Values indices are in v_{yx}.
1557 lp_build_lerp_2d(struct lp_build_context
*bld
,
1566 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
, flags
);
1567 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
, flags
);
1568 return lp_build_lerp(bld
, y
, v0
, v1
, flags
);
1573 lp_build_lerp_3d(struct lp_build_context
*bld
,
1587 LLVMValueRef v0
= lp_build_lerp_2d(bld
, x
, y
, v000
, v001
, v010
, v011
, flags
);
1588 LLVMValueRef v1
= lp_build_lerp_2d(bld
, x
, y
, v100
, v101
, v110
, v111
, flags
);
1589 return lp_build_lerp(bld
, z
, v0
, v1
, flags
);
1594 * Generate min(a, b)
1595 * Do checks for special cases but not for nans.
1598 lp_build_min(struct lp_build_context
*bld
,
1602 assert(lp_check_value(bld
->type
, a
));
1603 assert(lp_check_value(bld
->type
, b
));
1605 if(a
== bld
->undef
|| b
== bld
->undef
)
1611 if (bld
->type
.norm
) {
1612 if (!bld
->type
.sign
) {
1613 if (a
== bld
->zero
|| b
== bld
->zero
) {
1623 return lp_build_min_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1628 * Generate min(a, b)
1629 * NaN's are handled according to the behavior specified by the
1630 * nan_behavior argument.
1633 lp_build_min_ext(struct lp_build_context
*bld
,
1636 enum gallivm_nan_behavior nan_behavior
)
1638 assert(lp_check_value(bld
->type
, a
));
1639 assert(lp_check_value(bld
->type
, b
));
1641 if(a
== bld
->undef
|| b
== bld
->undef
)
1647 if (bld
->type
.norm
) {
1648 if (!bld
->type
.sign
) {
1649 if (a
== bld
->zero
|| b
== bld
->zero
) {
1659 return lp_build_min_simple(bld
, a
, b
, nan_behavior
);
1663 * Generate max(a, b)
1664 * Do checks for special cases, but NaN behavior is undefined.
1667 lp_build_max(struct lp_build_context
*bld
,
1671 assert(lp_check_value(bld
->type
, a
));
1672 assert(lp_check_value(bld
->type
, b
));
1674 if(a
== bld
->undef
|| b
== bld
->undef
)
1680 if(bld
->type
.norm
) {
1681 if(a
== bld
->one
|| b
== bld
->one
)
1683 if (!bld
->type
.sign
) {
1684 if (a
== bld
->zero
) {
1687 if (b
== bld
->zero
) {
1693 return lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1698 * Generate max(a, b)
1699 * Checks for special cases.
1700 * NaN's are handled according to the behavior specified by the
1701 * nan_behavior argument.
1704 lp_build_max_ext(struct lp_build_context
*bld
,
1707 enum gallivm_nan_behavior nan_behavior
)
1709 assert(lp_check_value(bld
->type
, a
));
1710 assert(lp_check_value(bld
->type
, b
));
1712 if(a
== bld
->undef
|| b
== bld
->undef
)
1718 if(bld
->type
.norm
) {
1719 if(a
== bld
->one
|| b
== bld
->one
)
1721 if (!bld
->type
.sign
) {
1722 if (a
== bld
->zero
) {
1725 if (b
== bld
->zero
) {
1731 return lp_build_max_simple(bld
, a
, b
, nan_behavior
);
1735 * Generate clamp(a, min, max)
1736 * NaN behavior (for any of a, min, max) is undefined.
1737 * Do checks for special cases.
1740 lp_build_clamp(struct lp_build_context
*bld
,
1745 assert(lp_check_value(bld
->type
, a
));
1746 assert(lp_check_value(bld
->type
, min
));
1747 assert(lp_check_value(bld
->type
, max
));
1749 a
= lp_build_min(bld
, a
, max
);
1750 a
= lp_build_max(bld
, a
, min
);
1756 * Generate clamp(a, 0, 1)
1757 * A NaN will get converted to zero.
1760 lp_build_clamp_zero_one_nanzero(struct lp_build_context
*bld
,
1763 a
= lp_build_max_ext(bld
, a
, bld
->zero
, GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1764 a
= lp_build_min(bld
, a
, bld
->one
);
1773 lp_build_abs(struct lp_build_context
*bld
,
1776 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1777 const struct lp_type type
= bld
->type
;
1778 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1780 assert(lp_check_value(type
, a
));
1786 if (0x0306 <= HAVE_LLVM
&& HAVE_LLVM
< 0x0309) {
1787 /* Workaround llvm.org/PR27332 */
1788 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1789 unsigned long long absMask
= ~(1ULL << (type
.width
- 1));
1790 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
, ((unsigned long long) absMask
));
1791 a
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1792 a
= LLVMBuildAnd(builder
, a
, mask
, "");
1793 a
= LLVMBuildBitCast(builder
, a
, vec_type
, "");
1797 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.fabs", vec_type
);
1798 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
1802 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
&& HAVE_LLVM
< 0x0600) {
1803 switch(type
.width
) {
1805 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
1807 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
1809 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
1812 else if (type
.width
*type
.length
== 256 && util_cpu_caps
.has_avx2
&& HAVE_LLVM
< 0x0600) {
1813 switch(type
.width
) {
1815 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.b", vec_type
, a
);
1817 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.w", vec_type
, a
);
1819 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.d", vec_type
, a
);
1823 return lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
),
1824 a
, LLVMBuildNeg(builder
, a
, ""));
1829 lp_build_negate(struct lp_build_context
*bld
,
1832 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1834 assert(lp_check_value(bld
->type
, a
));
1836 if (bld
->type
.floating
)
1837 a
= LLVMBuildFNeg(builder
, a
, "");
1839 a
= LLVMBuildNeg(builder
, a
, "");
1845 /** Return -1, 0 or +1 depending on the sign of a */
1847 lp_build_sgn(struct lp_build_context
*bld
,
1850 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1851 const struct lp_type type
= bld
->type
;
1855 assert(lp_check_value(type
, a
));
1857 /* Handle non-zero case */
1859 /* if not zero then sign must be positive */
1862 else if(type
.floating
) {
1863 LLVMTypeRef vec_type
;
1864 LLVMTypeRef int_type
;
1868 unsigned long long maskBit
= (unsigned long long)1 << (type
.width
- 1);
1870 int_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1871 vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1872 mask
= lp_build_const_int_vec(bld
->gallivm
, type
, maskBit
);
1874 /* Take the sign bit and add it to 1 constant */
1875 sign
= LLVMBuildBitCast(builder
, a
, int_type
, "");
1876 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
1877 one
= LLVMConstBitCast(bld
->one
, int_type
);
1878 res
= LLVMBuildOr(builder
, sign
, one
, "");
1879 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1883 /* signed int/norm/fixed point */
1884 /* could use psign with sse3 and appropriate vectors here */
1885 LLVMValueRef minus_one
= lp_build_const_vec(bld
->gallivm
, type
, -1.0);
1886 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
1887 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
1891 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
1892 res
= lp_build_select(bld
, cond
, bld
->zero
, res
);
1899 * Set the sign of float vector 'a' according to 'sign'.
1900 * If sign==0, return abs(a).
1901 * If sign==1, return -abs(a);
1902 * Other values for sign produce undefined results.
1905 lp_build_set_sign(struct lp_build_context
*bld
,
1906 LLVMValueRef a
, LLVMValueRef sign
)
1908 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1909 const struct lp_type type
= bld
->type
;
1910 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1911 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1912 LLVMValueRef shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
- 1);
1913 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
1914 ~((unsigned long long) 1 << (type
.width
- 1)));
1915 LLVMValueRef val
, res
;
1917 assert(type
.floating
);
1918 assert(lp_check_value(type
, a
));
1920 /* val = reinterpret_cast<int>(a) */
1921 val
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1922 /* val = val & mask */
1923 val
= LLVMBuildAnd(builder
, val
, mask
, "");
1924 /* sign = sign << shift */
1925 sign
= LLVMBuildShl(builder
, sign
, shift
, "");
1926 /* res = val | sign */
1927 res
= LLVMBuildOr(builder
, val
, sign
, "");
1928 /* res = reinterpret_cast<float>(res) */
1929 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1936 * Convert vector of (or scalar) int to vector of (or scalar) float.
1939 lp_build_int_to_float(struct lp_build_context
*bld
,
1942 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1943 const struct lp_type type
= bld
->type
;
1944 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1946 assert(type
.floating
);
1948 return LLVMBuildSIToFP(builder
, a
, vec_type
, "");
1952 arch_rounding_available(const struct lp_type type
)
1954 if ((util_cpu_caps
.has_sse4_1
&&
1955 (type
.length
== 1 || type
.width
*type
.length
== 128)) ||
1956 (util_cpu_caps
.has_avx
&& type
.width
*type
.length
== 256))
1958 else if ((util_cpu_caps
.has_altivec
&&
1959 (type
.width
== 32 && type
.length
== 4)))
1965 enum lp_build_round_mode
1967 LP_BUILD_ROUND_NEAREST
= 0,
1968 LP_BUILD_ROUND_FLOOR
= 1,
1969 LP_BUILD_ROUND_CEIL
= 2,
1970 LP_BUILD_ROUND_TRUNCATE
= 3
1973 static inline LLVMValueRef
1974 lp_build_iround_nearest_sse2(struct lp_build_context
*bld
,
1977 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1978 const struct lp_type type
= bld
->type
;
1979 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
1980 LLVMTypeRef ret_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1981 const char *intrinsic
;
1984 assert(type
.floating
);
1985 /* using the double precision conversions is a bit more complicated */
1986 assert(type
.width
== 32);
1988 assert(lp_check_value(type
, a
));
1989 assert(util_cpu_caps
.has_sse2
);
1991 /* This is relying on MXCSR rounding mode, which should always be nearest. */
1992 if (type
.length
== 1) {
1993 LLVMTypeRef vec_type
;
1996 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
1998 vec_type
= LLVMVectorType(bld
->elem_type
, 4);
2000 intrinsic
= "llvm.x86.sse.cvtss2si";
2002 undef
= LLVMGetUndef(vec_type
);
2004 arg
= LLVMBuildInsertElement(builder
, undef
, a
, index0
, "");
2006 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2010 if (type
.width
* type
.length
== 128) {
2011 intrinsic
= "llvm.x86.sse2.cvtps2dq";
2014 assert(type
.width
*type
.length
== 256);
2015 assert(util_cpu_caps
.has_avx
);
2017 intrinsic
= "llvm.x86.avx.cvt.ps2dq.256";
2019 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2029 static inline LLVMValueRef
2030 lp_build_round_altivec(struct lp_build_context
*bld
,
2032 enum lp_build_round_mode mode
)
2034 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2035 const struct lp_type type
= bld
->type
;
2036 const char *intrinsic
= NULL
;
2038 assert(type
.floating
);
2040 assert(lp_check_value(type
, a
));
2041 assert(util_cpu_caps
.has_altivec
);
2046 case LP_BUILD_ROUND_NEAREST
:
2047 intrinsic
= "llvm.ppc.altivec.vrfin";
2049 case LP_BUILD_ROUND_FLOOR
:
2050 intrinsic
= "llvm.ppc.altivec.vrfim";
2052 case LP_BUILD_ROUND_CEIL
:
2053 intrinsic
= "llvm.ppc.altivec.vrfip";
2055 case LP_BUILD_ROUND_TRUNCATE
:
2056 intrinsic
= "llvm.ppc.altivec.vrfiz";
2060 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2063 static inline LLVMValueRef
2064 lp_build_round_arch(struct lp_build_context
*bld
,
2066 enum lp_build_round_mode mode
)
2068 if (util_cpu_caps
.has_sse4_1
) {
2069 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2070 const struct lp_type type
= bld
->type
;
2071 const char *intrinsic_root
;
2074 assert(type
.floating
);
2075 assert(lp_check_value(type
, a
));
2079 case LP_BUILD_ROUND_NEAREST
:
2080 intrinsic_root
= "llvm.nearbyint";
2082 case LP_BUILD_ROUND_FLOOR
:
2083 intrinsic_root
= "llvm.floor";
2085 case LP_BUILD_ROUND_CEIL
:
2086 intrinsic_root
= "llvm.ceil";
2088 case LP_BUILD_ROUND_TRUNCATE
:
2089 intrinsic_root
= "llvm.trunc";
2093 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, intrinsic_root
, bld
->vec_type
);
2094 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2096 else /* (util_cpu_caps.has_altivec) */
2097 return lp_build_round_altivec(bld
, a
, mode
);
2101 * Return the integer part of a float (vector) value (== round toward zero).
2102 * The returned value is a float (vector).
2103 * Ex: trunc(-1.5) = -1.0
2106 lp_build_trunc(struct lp_build_context
*bld
,
2109 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2110 const struct lp_type type
= bld
->type
;
2112 assert(type
.floating
);
2113 assert(lp_check_value(type
, a
));
2115 if (arch_rounding_available(type
)) {
2116 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_TRUNCATE
);
2119 const struct lp_type type
= bld
->type
;
2120 struct lp_type inttype
;
2121 struct lp_build_context intbld
;
2122 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2123 LLVMValueRef trunc
, res
, anosign
, mask
;
2124 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2125 LLVMTypeRef vec_type
= bld
->vec_type
;
2127 assert(type
.width
== 32); /* might want to handle doubles at some point */
2130 inttype
.floating
= 0;
2131 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2133 /* round by truncation */
2134 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2135 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2137 /* mask out sign bit */
2138 anosign
= lp_build_abs(bld
, a
);
2140 * mask out all values if anosign > 2^24
2141 * This should work both for large ints (all rounding is no-op for them
2142 * because such floats are always exact) as well as special cases like
2143 * NaNs, Infs (taking advantage of the fact they use max exponent).
2144 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2146 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2147 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2148 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2149 return lp_build_select(bld
, mask
, a
, res
);
2155 * Return float (vector) rounded to nearest integer (vector). The returned
2156 * value is a float (vector).
2157 * Ex: round(0.9) = 1.0
2158 * Ex: round(-1.5) = -2.0
2161 lp_build_round(struct lp_build_context
*bld
,
2164 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2165 const struct lp_type type
= bld
->type
;
2167 assert(type
.floating
);
2168 assert(lp_check_value(type
, a
));
2170 if (arch_rounding_available(type
)) {
2171 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2174 const struct lp_type type
= bld
->type
;
2175 struct lp_type inttype
;
2176 struct lp_build_context intbld
;
2177 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2178 LLVMValueRef res
, anosign
, mask
;
2179 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2180 LLVMTypeRef vec_type
= bld
->vec_type
;
2182 assert(type
.width
== 32); /* might want to handle doubles at some point */
2185 inttype
.floating
= 0;
2186 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2188 res
= lp_build_iround(bld
, a
);
2189 res
= LLVMBuildSIToFP(builder
, res
, vec_type
, "");
2191 /* mask out sign bit */
2192 anosign
= lp_build_abs(bld
, a
);
2194 * mask out all values if anosign > 2^24
2195 * This should work both for large ints (all rounding is no-op for them
2196 * because such floats are always exact) as well as special cases like
2197 * NaNs, Infs (taking advantage of the fact they use max exponent).
2198 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2200 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2201 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2202 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2203 return lp_build_select(bld
, mask
, a
, res
);
2209 * Return floor of float (vector), result is a float (vector)
2210 * Ex: floor(1.1) = 1.0
2211 * Ex: floor(-1.1) = -2.0
2214 lp_build_floor(struct lp_build_context
*bld
,
2217 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2218 const struct lp_type type
= bld
->type
;
2220 assert(type
.floating
);
2221 assert(lp_check_value(type
, a
));
2223 if (arch_rounding_available(type
)) {
2224 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2227 const struct lp_type type
= bld
->type
;
2228 struct lp_type inttype
;
2229 struct lp_build_context intbld
;
2230 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2231 LLVMValueRef trunc
, res
, anosign
, mask
;
2232 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2233 LLVMTypeRef vec_type
= bld
->vec_type
;
2235 if (type
.width
!= 32) {
2237 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.floor", vec_type
);
2238 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2241 assert(type
.width
== 32); /* might want to handle doubles at some point */
2244 inttype
.floating
= 0;
2245 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2247 /* round by truncation */
2248 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2249 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2255 * fix values if rounding is wrong (for non-special cases)
2256 * - this is the case if trunc > a
2258 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, res
, a
);
2259 /* tmp = trunc > a ? 1.0 : 0.0 */
2260 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2261 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2262 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2263 res
= lp_build_sub(bld
, res
, tmp
);
2266 /* mask out sign bit */
2267 anosign
= lp_build_abs(bld
, a
);
2269 * mask out all values if anosign > 2^24
2270 * This should work both for large ints (all rounding is no-op for them
2271 * because such floats are always exact) as well as special cases like
2272 * NaNs, Infs (taking advantage of the fact they use max exponent).
2273 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2275 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2276 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2277 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2278 return lp_build_select(bld
, mask
, a
, res
);
2284 * Return ceiling of float (vector), returning float (vector).
2285 * Ex: ceil( 1.1) = 2.0
2286 * Ex: ceil(-1.1) = -1.0
2289 lp_build_ceil(struct lp_build_context
*bld
,
2292 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2293 const struct lp_type type
= bld
->type
;
2295 assert(type
.floating
);
2296 assert(lp_check_value(type
, a
));
2298 if (arch_rounding_available(type
)) {
2299 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2302 const struct lp_type type
= bld
->type
;
2303 struct lp_type inttype
;
2304 struct lp_build_context intbld
;
2305 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2306 LLVMValueRef trunc
, res
, anosign
, mask
, tmp
;
2307 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2308 LLVMTypeRef vec_type
= bld
->vec_type
;
2310 if (type
.width
!= 32) {
2312 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.ceil", vec_type
);
2313 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2316 assert(type
.width
== 32); /* might want to handle doubles at some point */
2319 inttype
.floating
= 0;
2320 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2322 /* round by truncation */
2323 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2324 trunc
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "ceil.trunc");
2327 * fix values if rounding is wrong (for non-special cases)
2328 * - this is the case if trunc < a
2330 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2331 /* tmp = trunc < a ? 1.0 : 0.0 */
2332 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2333 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2334 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2335 res
= lp_build_add(bld
, trunc
, tmp
);
2337 /* mask out sign bit */
2338 anosign
= lp_build_abs(bld
, a
);
2340 * mask out all values if anosign > 2^24
2341 * This should work both for large ints (all rounding is no-op for them
2342 * because such floats are always exact) as well as special cases like
2343 * NaNs, Infs (taking advantage of the fact they use max exponent).
2344 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2346 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2347 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2348 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2349 return lp_build_select(bld
, mask
, a
, res
);
2355 * Return fractional part of 'a' computed as a - floor(a)
2356 * Typically used in texture coord arithmetic.
2359 lp_build_fract(struct lp_build_context
*bld
,
2362 assert(bld
->type
.floating
);
2363 return lp_build_sub(bld
, a
, lp_build_floor(bld
, a
));
2368 * Prevent returning 1.0 for very small negative values of 'a' by clamping
2369 * against 0.99999(9). (Will also return that value for NaNs.)
2371 static inline LLVMValueRef
2372 clamp_fract(struct lp_build_context
*bld
, LLVMValueRef fract
)
2376 /* this is the largest number smaller than 1.0 representable as float */
2377 max
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2378 1.0 - 1.0/(1LL << (lp_mantissa(bld
->type
) + 1)));
2379 return lp_build_min_ext(bld
, fract
, max
,
2380 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
2385 * Same as lp_build_fract, but guarantees that the result is always smaller
2386 * than one. Will also return the smaller-than-one value for infs, NaNs.
2389 lp_build_fract_safe(struct lp_build_context
*bld
,
2392 return clamp_fract(bld
, lp_build_fract(bld
, a
));
2397 * Return the integer part of a float (vector) value (== round toward zero).
2398 * The returned value is an integer (vector).
2399 * Ex: itrunc(-1.5) = -1
2402 lp_build_itrunc(struct lp_build_context
*bld
,
2405 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2406 const struct lp_type type
= bld
->type
;
2407 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
2409 assert(type
.floating
);
2410 assert(lp_check_value(type
, a
));
2412 return LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2417 * Return float (vector) rounded to nearest integer (vector). The returned
2418 * value is an integer (vector).
2419 * Ex: iround(0.9) = 1
2420 * Ex: iround(-1.5) = -2
2423 lp_build_iround(struct lp_build_context
*bld
,
2426 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2427 const struct lp_type type
= bld
->type
;
2428 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2431 assert(type
.floating
);
2433 assert(lp_check_value(type
, a
));
2435 if ((util_cpu_caps
.has_sse2
&&
2436 ((type
.width
== 32) && (type
.length
== 1 || type
.length
== 4))) ||
2437 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2438 return lp_build_iround_nearest_sse2(bld
, a
);
2440 if (arch_rounding_available(type
)) {
2441 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2446 half
= lp_build_const_vec(bld
->gallivm
, type
, 0.5);
2449 LLVMTypeRef vec_type
= bld
->vec_type
;
2450 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
2451 (unsigned long long)1 << (type
.width
- 1));
2455 sign
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
2456 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
2459 half
= LLVMBuildBitCast(builder
, half
, int_vec_type
, "");
2460 half
= LLVMBuildOr(builder
, sign
, half
, "");
2461 half
= LLVMBuildBitCast(builder
, half
, vec_type
, "");
2464 res
= LLVMBuildFAdd(builder
, a
, half
, "");
2467 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "");
2474 * Return floor of float (vector), result is an int (vector)
2475 * Ex: ifloor(1.1) = 1.0
2476 * Ex: ifloor(-1.1) = -2.0
2479 lp_build_ifloor(struct lp_build_context
*bld
,
2482 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2483 const struct lp_type type
= bld
->type
;
2484 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2487 assert(type
.floating
);
2488 assert(lp_check_value(type
, a
));
2492 if (arch_rounding_available(type
)) {
2493 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2496 struct lp_type inttype
;
2497 struct lp_build_context intbld
;
2498 LLVMValueRef trunc
, itrunc
, mask
;
2500 assert(type
.floating
);
2501 assert(lp_check_value(type
, a
));
2504 inttype
.floating
= 0;
2505 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2507 /* round by truncation */
2508 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2509 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "ifloor.trunc");
2512 * fix values if rounding is wrong (for non-special cases)
2513 * - this is the case if trunc > a
2514 * The results of doing this with NaNs, very large values etc.
2515 * are undefined but this seems to be the case anyway.
2517 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, trunc
, a
);
2518 /* cheapie minus one with mask since the mask is minus one / zero */
2519 return lp_build_add(&intbld
, itrunc
, mask
);
2523 /* round to nearest (toward zero) */
2524 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "ifloor.res");
2531 * Return ceiling of float (vector), returning int (vector).
2532 * Ex: iceil( 1.1) = 2
2533 * Ex: iceil(-1.1) = -1
2536 lp_build_iceil(struct lp_build_context
*bld
,
2539 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2540 const struct lp_type type
= bld
->type
;
2541 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2544 assert(type
.floating
);
2545 assert(lp_check_value(type
, a
));
2547 if (arch_rounding_available(type
)) {
2548 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2551 struct lp_type inttype
;
2552 struct lp_build_context intbld
;
2553 LLVMValueRef trunc
, itrunc
, mask
;
2555 assert(type
.floating
);
2556 assert(lp_check_value(type
, a
));
2559 inttype
.floating
= 0;
2560 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2562 /* round by truncation */
2563 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2564 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "iceil.trunc");
2567 * fix values if rounding is wrong (for non-special cases)
2568 * - this is the case if trunc < a
2569 * The results of doing this with NaNs, very large values etc.
2570 * are undefined but this seems to be the case anyway.
2572 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2573 /* cheapie plus one with mask since the mask is minus one / zero */
2574 return lp_build_sub(&intbld
, itrunc
, mask
);
2577 /* round to nearest (toward zero) */
2578 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "iceil.res");
2585 * Combined ifloor() & fract().
2587 * Preferred to calling the functions separately, as it will ensure that the
2588 * strategy (floor() vs ifloor()) that results in less redundant work is used.
2591 lp_build_ifloor_fract(struct lp_build_context
*bld
,
2593 LLVMValueRef
*out_ipart
,
2594 LLVMValueRef
*out_fpart
)
2596 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2597 const struct lp_type type
= bld
->type
;
2600 assert(type
.floating
);
2601 assert(lp_check_value(type
, a
));
2603 if (arch_rounding_available(type
)) {
2605 * floor() is easier.
2608 ipart
= lp_build_floor(bld
, a
);
2609 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2610 *out_ipart
= LLVMBuildFPToSI(builder
, ipart
, bld
->int_vec_type
, "ipart");
2614 * ifloor() is easier.
2617 *out_ipart
= lp_build_ifloor(bld
, a
);
2618 ipart
= LLVMBuildSIToFP(builder
, *out_ipart
, bld
->vec_type
, "ipart");
2619 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2625 * Same as lp_build_ifloor_fract, but guarantees that the fractional part is
2626 * always smaller than one.
2629 lp_build_ifloor_fract_safe(struct lp_build_context
*bld
,
2631 LLVMValueRef
*out_ipart
,
2632 LLVMValueRef
*out_fpart
)
2634 lp_build_ifloor_fract(bld
, a
, out_ipart
, out_fpart
);
2635 *out_fpart
= clamp_fract(bld
, *out_fpart
);
2640 lp_build_sqrt(struct lp_build_context
*bld
,
2643 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2644 const struct lp_type type
= bld
->type
;
2645 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
2648 assert(lp_check_value(type
, a
));
2650 assert(type
.floating
);
2651 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.sqrt", vec_type
);
2653 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2658 * Do one Newton-Raphson step to improve reciprocate precision:
2660 * x_{i+1} = x_i * (2 - a * x_i)
2662 * XXX: Unfortunately this won't give IEEE-754 conformant results for 0 or
2663 * +/-Inf, giving NaN instead. Certain applications rely on this behavior,
2664 * such as Google Earth, which does RCP(RSQRT(0.0) when drawing the Earth's
2665 * halo. It would be necessary to clamp the argument to prevent this.
2668 * - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division
2669 * - http://softwarecommunity.intel.com/articles/eng/1818.htm
2671 static inline LLVMValueRef
2672 lp_build_rcp_refine(struct lp_build_context
*bld
,
2676 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2677 LLVMValueRef two
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 2.0);
2680 res
= LLVMBuildFMul(builder
, a
, rcp_a
, "");
2681 res
= LLVMBuildFSub(builder
, two
, res
, "");
2682 res
= LLVMBuildFMul(builder
, rcp_a
, res
, "");
2689 lp_build_rcp(struct lp_build_context
*bld
,
2692 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2693 const struct lp_type type
= bld
->type
;
2695 assert(lp_check_value(type
, a
));
2704 assert(type
.floating
);
2706 if(LLVMIsConstant(a
))
2707 return LLVMConstFDiv(bld
->one
, a
);
2710 * We don't use RCPPS because:
2711 * - it only has 10bits of precision
2712 * - it doesn't even get the reciprocate of 1.0 exactly
2713 * - doing Newton-Rapshon steps yields wrong (NaN) values for 0.0 or Inf
2714 * - for recent processors the benefit over DIVPS is marginal, a case
2717 * We could still use it on certain processors if benchmarks show that the
2718 * RCPPS plus necessary workarounds are still preferrable to DIVPS; or for
2719 * particular uses that require less workarounds.
2722 if (FALSE
&& ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2723 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8))){
2724 const unsigned num_iterations
= 0;
2727 const char *intrinsic
= NULL
;
2729 if (type
.length
== 4) {
2730 intrinsic
= "llvm.x86.sse.rcp.ps";
2733 intrinsic
= "llvm.x86.avx.rcp.ps.256";
2736 res
= lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2738 for (i
= 0; i
< num_iterations
; ++i
) {
2739 res
= lp_build_rcp_refine(bld
, a
, res
);
2745 return LLVMBuildFDiv(builder
, bld
->one
, a
, "");
2750 * Do one Newton-Raphson step to improve rsqrt precision:
2752 * x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
2754 * See also Intel 64 and IA-32 Architectures Optimization Manual.
2756 static inline LLVMValueRef
2757 lp_build_rsqrt_refine(struct lp_build_context
*bld
,
2759 LLVMValueRef rsqrt_a
)
2761 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2762 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 0.5);
2763 LLVMValueRef three
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 3.0);
2766 res
= LLVMBuildFMul(builder
, rsqrt_a
, rsqrt_a
, "");
2767 res
= LLVMBuildFMul(builder
, a
, res
, "");
2768 res
= LLVMBuildFSub(builder
, three
, res
, "");
2769 res
= LLVMBuildFMul(builder
, rsqrt_a
, res
, "");
2770 res
= LLVMBuildFMul(builder
, half
, res
, "");
2777 * Generate 1/sqrt(a).
2778 * Result is undefined for values < 0, infinity for +0.
2781 lp_build_rsqrt(struct lp_build_context
*bld
,
2784 const struct lp_type type
= bld
->type
;
2786 assert(lp_check_value(type
, a
));
2788 assert(type
.floating
);
2791 * This should be faster but all denormals will end up as infinity.
2793 if (0 && lp_build_fast_rsqrt_available(type
)) {
2794 const unsigned num_iterations
= 1;
2798 /* rsqrt(1.0) != 1.0 here */
2799 res
= lp_build_fast_rsqrt(bld
, a
);
2801 if (num_iterations
) {
2803 * Newton-Raphson will result in NaN instead of infinity for zero,
2804 * and NaN instead of zero for infinity.
2805 * Also, need to ensure rsqrt(1.0) == 1.0.
2806 * All numbers smaller than FLT_MIN will result in +infinity
2807 * (rsqrtps treats all denormals as zero).
2810 LLVMValueRef flt_min
= lp_build_const_vec(bld
->gallivm
, type
, FLT_MIN
);
2811 LLVMValueRef inf
= lp_build_const_vec(bld
->gallivm
, type
, INFINITY
);
2813 for (i
= 0; i
< num_iterations
; ++i
) {
2814 res
= lp_build_rsqrt_refine(bld
, a
, res
);
2816 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_LESS
, a
, flt_min
);
2817 res
= lp_build_select(bld
, cmp
, inf
, res
);
2818 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, inf
);
2819 res
= lp_build_select(bld
, cmp
, bld
->zero
, res
);
2820 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, bld
->one
);
2821 res
= lp_build_select(bld
, cmp
, bld
->one
, res
);
2827 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2831 * If there's a fast (inaccurate) rsqrt instruction available
2832 * (caller may want to avoid to call rsqrt_fast if it's not available,
2833 * i.e. for calculating x^0.5 it may do rsqrt_fast(x) * x but if
2834 * unavailable it would result in sqrt/div/mul so obviously
2835 * much better to just call sqrt, skipping both div and mul).
2838 lp_build_fast_rsqrt_available(struct lp_type type
)
2840 assert(type
.floating
);
2842 if ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2843 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2851 * Generate 1/sqrt(a).
2852 * Result is undefined for values < 0, infinity for +0.
2853 * Precision is limited, only ~10 bits guaranteed
2854 * (rsqrt 1.0 may not be 1.0, denorms may be flushed to 0).
2857 lp_build_fast_rsqrt(struct lp_build_context
*bld
,
2860 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2861 const struct lp_type type
= bld
->type
;
2863 assert(lp_check_value(type
, a
));
2865 if (lp_build_fast_rsqrt_available(type
)) {
2866 const char *intrinsic
= NULL
;
2868 if (type
.length
== 4) {
2869 intrinsic
= "llvm.x86.sse.rsqrt.ps";
2872 intrinsic
= "llvm.x86.avx.rsqrt.ps.256";
2874 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2877 debug_printf("%s: emulating fast rsqrt with rcp/sqrt\n", __FUNCTION__
);
2879 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2884 * Generate sin(a) or cos(a) using polynomial approximation.
2885 * TODO: it might be worth recognizing sin and cos using same source
2886 * (i.e. d3d10 sincos opcode). Obviously doing both at the same time
2887 * would be way cheaper than calculating (nearly) everything twice...
2888 * Not sure it's common enough to be worth bothering however, scs
2889 * opcode could also benefit from calculating both though.
2892 lp_build_sin_or_cos(struct lp_build_context
*bld
,
2896 struct gallivm_state
*gallivm
= bld
->gallivm
;
2897 LLVMBuilderRef b
= gallivm
->builder
;
2898 struct lp_type int_type
= lp_int_type(bld
->type
);
2901 * take the absolute value,
2902 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
2905 LLVMValueRef inv_sig_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, ~0x80000000);
2906 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, bld
->int_vec_type
, "a_v4si");
2908 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
2909 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, bld
->vec_type
, "x_abs");
2913 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
2916 LLVMValueRef FOPi
= lp_build_const_vec(gallivm
, bld
->type
, 1.27323954473516);
2917 LLVMValueRef scale_y
= LLVMBuildFMul(b
, x_abs
, FOPi
, "scale_y");
2920 * store the integer part of y in mm0
2921 * emm2 = _mm_cvttps_epi32(y);
2924 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, bld
->int_vec_type
, "emm2_i");
2927 * j=(j+1) & (~1) (see the cephes sources)
2928 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
2931 LLVMValueRef all_one
= lp_build_const_int_vec(gallivm
, bld
->type
, 1);
2932 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
2934 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
2936 LLVMValueRef inv_one
= lp_build_const_int_vec(gallivm
, bld
->type
, ~1);
2937 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
2940 * y = _mm_cvtepi32_ps(emm2);
2942 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, bld
->vec_type
, "y_2");
2944 LLVMValueRef const_2
= lp_build_const_int_vec(gallivm
, bld
->type
, 2);
2945 LLVMValueRef const_4
= lp_build_const_int_vec(gallivm
, bld
->type
, 4);
2946 LLVMValueRef const_29
= lp_build_const_int_vec(gallivm
, bld
->type
, 29);
2947 LLVMValueRef sign_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, 0x80000000);
2950 * Argument used for poly selection and sign bit determination
2951 * is different for sin vs. cos.
2953 LLVMValueRef emm2_2
= cos
? LLVMBuildSub(b
, emm2_and
, const_2
, "emm2_2") :
2956 LLVMValueRef sign_bit
= cos
? LLVMBuildShl(b
, LLVMBuildAnd(b
, const_4
,
2957 LLVMBuildNot(b
, emm2_2
, ""), ""),
2958 const_29
, "sign_bit") :
2959 LLVMBuildAnd(b
, LLVMBuildXor(b
, a_v4si
,
2960 LLVMBuildShl(b
, emm2_add
,
2962 sign_mask
, "sign_bit");
2965 * get the polynom selection mask
2966 * there is one polynom for 0 <= x <= Pi/4
2967 * and another one for Pi/4<x<=Pi/2
2968 * Both branches will be computed.
2970 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
2971 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
2974 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_2
, const_2
, "emm2_3");
2975 LLVMValueRef poly_mask
= lp_build_compare(gallivm
,
2976 int_type
, PIPE_FUNC_EQUAL
,
2977 emm2_3
, lp_build_const_int_vec(gallivm
, bld
->type
, 0));
2980 * _PS_CONST(minus_cephes_DP1, -0.78515625);
2981 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
2982 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
2984 LLVMValueRef DP1
= lp_build_const_vec(gallivm
, bld
->type
, -0.78515625);
2985 LLVMValueRef DP2
= lp_build_const_vec(gallivm
, bld
->type
, -2.4187564849853515625e-4);
2986 LLVMValueRef DP3
= lp_build_const_vec(gallivm
, bld
->type
, -3.77489497744594108e-8);
2989 * The magic pass: "Extended precision modular arithmetic"
2990 * x = ((x - y * DP1) - y * DP2) - y * DP3;
2992 LLVMValueRef x_1
= lp_build_fmuladd(b
, y_2
, DP1
, x_abs
);
2993 LLVMValueRef x_2
= lp_build_fmuladd(b
, y_2
, DP2
, x_1
);
2994 LLVMValueRef x_3
= lp_build_fmuladd(b
, y_2
, DP3
, x_2
);
2997 * Evaluate the first polynom (0 <= x <= Pi/4)
2999 * z = _mm_mul_ps(x,x);
3001 LLVMValueRef z
= LLVMBuildFMul(b
, x_3
, x_3
, "z");
3004 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
3005 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
3006 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
3008 LLVMValueRef coscof_p0
= lp_build_const_vec(gallivm
, bld
->type
, 2.443315711809948E-005);
3009 LLVMValueRef coscof_p1
= lp_build_const_vec(gallivm
, bld
->type
, -1.388731625493765E-003);
3010 LLVMValueRef coscof_p2
= lp_build_const_vec(gallivm
, bld
->type
, 4.166664568298827E-002);
3013 * y = *(v4sf*)_ps_coscof_p0;
3014 * y = _mm_mul_ps(y, z);
3016 LLVMValueRef y_4
= lp_build_fmuladd(b
, z
, coscof_p0
, coscof_p1
);
3017 LLVMValueRef y_6
= lp_build_fmuladd(b
, y_4
, z
, coscof_p2
);
3018 LLVMValueRef y_7
= LLVMBuildFMul(b
, y_6
, z
, "y_7");
3019 LLVMValueRef y_8
= LLVMBuildFMul(b
, y_7
, z
, "y_8");
3023 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
3024 * y = _mm_sub_ps(y, tmp);
3025 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
3027 LLVMValueRef half
= lp_build_const_vec(gallivm
, bld
->type
, 0.5);
3028 LLVMValueRef tmp
= LLVMBuildFMul(b
, z
, half
, "tmp");
3029 LLVMValueRef y_9
= LLVMBuildFSub(b
, y_8
, tmp
, "y_8");
3030 LLVMValueRef one
= lp_build_const_vec(gallivm
, bld
->type
, 1.0);
3031 LLVMValueRef y_10
= LLVMBuildFAdd(b
, y_9
, one
, "y_9");
3034 * _PS_CONST(sincof_p0, -1.9515295891E-4);
3035 * _PS_CONST(sincof_p1, 8.3321608736E-3);
3036 * _PS_CONST(sincof_p2, -1.6666654611E-1);
3038 LLVMValueRef sincof_p0
= lp_build_const_vec(gallivm
, bld
->type
, -1.9515295891E-4);
3039 LLVMValueRef sincof_p1
= lp_build_const_vec(gallivm
, bld
->type
, 8.3321608736E-3);
3040 LLVMValueRef sincof_p2
= lp_build_const_vec(gallivm
, bld
->type
, -1.6666654611E-1);
3043 * Evaluate the second polynom (Pi/4 <= x <= 0)
3045 * y2 = *(v4sf*)_ps_sincof_p0;
3046 * y2 = _mm_mul_ps(y2, z);
3047 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
3048 * y2 = _mm_mul_ps(y2, z);
3049 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
3050 * y2 = _mm_mul_ps(y2, z);
3051 * y2 = _mm_mul_ps(y2, x);
3052 * y2 = _mm_add_ps(y2, x);
3055 LLVMValueRef y2_4
= lp_build_fmuladd(b
, z
, sincof_p0
, sincof_p1
);
3056 LLVMValueRef y2_6
= lp_build_fmuladd(b
, y2_4
, z
, sincof_p2
);
3057 LLVMValueRef y2_7
= LLVMBuildFMul(b
, y2_6
, z
, "y2_7");
3058 LLVMValueRef y2_9
= lp_build_fmuladd(b
, y2_7
, x_3
, x_3
);
3061 * select the correct result from the two polynoms
3063 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
3064 * y = _mm_andnot_ps(xmm3, y);
3065 * y = _mm_or_ps(y,y2);
3067 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, bld
->int_vec_type
, "y2_i");
3068 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, bld
->int_vec_type
, "y_i");
3069 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
3070 LLVMValueRef poly_mask_inv
= LLVMBuildNot(b
, poly_mask
, "poly_mask_inv");
3071 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
3072 LLVMValueRef y_combine
= LLVMBuildOr(b
, y_and
, y2_and
, "y_combine");
3076 * y = _mm_xor_ps(y, sign_bit);
3078 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit
, "y_sign");
3079 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, bld
->vec_type
, "y_result");
3081 LLVMValueRef isfinite
= lp_build_isfinite(bld
, a
);
3083 /* clamp output to be within [-1, 1] */
3084 y_result
= lp_build_clamp(bld
, y_result
,
3085 lp_build_const_vec(bld
->gallivm
, bld
->type
, -1.f
),
3086 lp_build_const_vec(bld
->gallivm
, bld
->type
, 1.f
));
3087 /* If a is -inf, inf or NaN then return NaN */
3088 y_result
= lp_build_select(bld
, isfinite
, y_result
,
3089 lp_build_const_vec(bld
->gallivm
, bld
->type
, NAN
));
3098 lp_build_sin(struct lp_build_context
*bld
,
3101 return lp_build_sin_or_cos(bld
, a
, FALSE
);
3109 lp_build_cos(struct lp_build_context
*bld
,
3112 return lp_build_sin_or_cos(bld
, a
, TRUE
);
3117 * Generate pow(x, y)
3120 lp_build_pow(struct lp_build_context
*bld
,
3124 /* TODO: optimize the constant case */
3125 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3126 LLVMIsConstant(x
) && LLVMIsConstant(y
)) {
3127 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3131 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
3139 lp_build_exp(struct lp_build_context
*bld
,
3142 /* log2(e) = 1/log(2) */
3143 LLVMValueRef log2e
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3144 1.4426950408889634);
3146 assert(lp_check_value(bld
->type
, x
));
3148 return lp_build_exp2(bld
, lp_build_mul(bld
, log2e
, x
));
3154 * Behavior is undefined with infs, 0s and nans
3157 lp_build_log(struct lp_build_context
*bld
,
3161 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3162 0.69314718055994529);
3164 assert(lp_check_value(bld
->type
, x
));
3166 return lp_build_mul(bld
, log2
, lp_build_log2(bld
, x
));
3170 * Generate log(x) that handles edge cases (infs, 0s and nans)
3173 lp_build_log_safe(struct lp_build_context
*bld
,
3177 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3178 0.69314718055994529);
3180 assert(lp_check_value(bld
->type
, x
));
3182 return lp_build_mul(bld
, log2
, lp_build_log2_safe(bld
, x
));
3187 * Generate polynomial.
3188 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
3191 lp_build_polynomial(struct lp_build_context
*bld
,
3193 const double *coeffs
,
3194 unsigned num_coeffs
)
3196 const struct lp_type type
= bld
->type
;
3197 LLVMValueRef even
= NULL
, odd
= NULL
;
3201 assert(lp_check_value(bld
->type
, x
));
3203 /* TODO: optimize the constant case */
3204 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3205 LLVMIsConstant(x
)) {
3206 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3211 * Calculate odd and even terms seperately to decrease data dependency
3213 * c[0] + x^2 * c[2] + x^4 * c[4] ...
3214 * + x * (c[1] + x^2 * c[3] + x^4 * c[5]) ...
3216 x2
= lp_build_mul(bld
, x
, x
);
3218 for (i
= num_coeffs
; i
--; ) {
3221 coeff
= lp_build_const_vec(bld
->gallivm
, type
, coeffs
[i
]);
3225 even
= lp_build_mad(bld
, x2
, even
, coeff
);
3230 odd
= lp_build_mad(bld
, x2
, odd
, coeff
);
3237 return lp_build_mad(bld
, odd
, x
, even
);
3246 * Minimax polynomial fit of 2**x, in range [0, 1[
3248 const double lp_build_exp2_polynomial
[] = {
3249 #if EXP_POLY_DEGREE == 5
3250 1.000000000000000000000, /*XXX: was 0.999999925063526176901, recompute others */
3251 0.693153073200168932794,
3252 0.240153617044375388211,
3253 0.0558263180532956664775,
3254 0.00898934009049466391101,
3255 0.00187757667519147912699
3256 #elif EXP_POLY_DEGREE == 4
3257 1.00000259337069434683,
3258 0.693003834469974940458,
3259 0.24144275689150793076,
3260 0.0520114606103070150235,
3261 0.0135341679161270268764
3262 #elif EXP_POLY_DEGREE == 3
3263 0.999925218562710312959,
3264 0.695833540494823811697,
3265 0.226067155427249155588,
3266 0.0780245226406372992967
3267 #elif EXP_POLY_DEGREE == 2
3268 1.00172476321474503578,
3269 0.657636275736077639316,
3270 0.33718943461968720704
3278 lp_build_exp2(struct lp_build_context
*bld
,
3281 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3282 const struct lp_type type
= bld
->type
;
3283 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3284 LLVMValueRef ipart
= NULL
;
3285 LLVMValueRef fpart
= NULL
;
3286 LLVMValueRef expipart
= NULL
;
3287 LLVMValueRef expfpart
= NULL
;
3288 LLVMValueRef res
= NULL
;
3290 assert(lp_check_value(bld
->type
, x
));
3292 /* TODO: optimize the constant case */
3293 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3294 LLVMIsConstant(x
)) {
3295 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3299 assert(type
.floating
&& type
.width
== 32);
3301 /* We want to preserve NaN and make sure than for exp2 if x > 128,
3302 * the result is INF and if it's smaller than -126.9 the result is 0 */
3303 x
= lp_build_min_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, 128.0), x
,
3304 GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3305 x
= lp_build_max_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, -126.99999),
3306 x
, GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3308 /* ipart = floor(x) */
3309 /* fpart = x - ipart */
3310 lp_build_ifloor_fract(bld
, x
, &ipart
, &fpart
);
3312 /* expipart = (float) (1 << ipart) */
3313 expipart
= LLVMBuildAdd(builder
, ipart
,
3314 lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3315 expipart
= LLVMBuildShl(builder
, expipart
,
3316 lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3317 expipart
= LLVMBuildBitCast(builder
, expipart
, vec_type
, "");
3319 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
3320 ARRAY_SIZE(lp_build_exp2_polynomial
));
3322 res
= LLVMBuildFMul(builder
, expipart
, expfpart
, "");
3330 * Extract the exponent of a IEEE-754 floating point value.
3332 * Optionally apply an integer bias.
3334 * Result is an integer value with
3336 * ifloor(log2(x)) + bias
3339 lp_build_extract_exponent(struct lp_build_context
*bld
,
3343 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3344 const struct lp_type type
= bld
->type
;
3345 unsigned mantissa
= lp_mantissa(type
);
3348 assert(type
.floating
);
3350 assert(lp_check_value(bld
->type
, x
));
3352 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3354 res
= LLVMBuildLShr(builder
, x
,
3355 lp_build_const_int_vec(bld
->gallivm
, type
, mantissa
), "");
3356 res
= LLVMBuildAnd(builder
, res
,
3357 lp_build_const_int_vec(bld
->gallivm
, type
, 255), "");
3358 res
= LLVMBuildSub(builder
, res
,
3359 lp_build_const_int_vec(bld
->gallivm
, type
, 127 - bias
), "");
3366 * Extract the mantissa of the a floating.
3368 * Result is a floating point value with
3370 * x / floor(log2(x))
3373 lp_build_extract_mantissa(struct lp_build_context
*bld
,
3376 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3377 const struct lp_type type
= bld
->type
;
3378 unsigned mantissa
= lp_mantissa(type
);
3379 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
,
3380 (1ULL << mantissa
) - 1);
3381 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, bld
->int_vec_type
);
3384 assert(lp_check_value(bld
->type
, x
));
3386 assert(type
.floating
);
3388 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3390 /* res = x / 2**ipart */
3391 res
= LLVMBuildAnd(builder
, x
, mantmask
, "");
3392 res
= LLVMBuildOr(builder
, res
, one
, "");
3393 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
3401 * Minimax polynomial fit of log2((1.0 + sqrt(x))/(1.0 - sqrt(x)))/sqrt(x) ,for x in range of [0, 1/9[
3402 * These coefficients can be generate with
3403 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
3405 const double lp_build_log2_polynomial
[] = {
3406 #if LOG_POLY_DEGREE == 5
3407 2.88539008148777786488L,
3408 0.961796878841293367824L,
3409 0.577058946784739859012L,
3410 0.412914355135828735411L,
3411 0.308591899232910175289L,
3412 0.352376952300281371868L,
3413 #elif LOG_POLY_DEGREE == 4
3414 2.88539009343309178325L,
3415 0.961791550404184197881L,
3416 0.577440339438736392009L,
3417 0.403343858251329912514L,
3418 0.406718052498846252698L,
3419 #elif LOG_POLY_DEGREE == 3
3420 2.88538959748872753838L,
3421 0.961932915889597772928L,
3422 0.571118517972136195241L,
3423 0.493997535084709500285L,
3430 * See http://www.devmaster.net/forums/showthread.php?p=43580
3431 * http://en.wikipedia.org/wiki/Logarithm#Calculation
3432 * http://www.nezumi.demon.co.uk/consult/logx.htm
3434 * If handle_edge_cases is true the function will perform computations
3435 * to match the required D3D10+ behavior for each of the edge cases.
3436 * That means that if input is:
3437 * - less than zero (to and including -inf) then NaN will be returned
3438 * - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
3439 * - +infinity, then +infinity will be returned
3440 * - NaN, then NaN will be returned
3442 * Those checks are fairly expensive so if you don't need them make sure
3443 * handle_edge_cases is false.
3446 lp_build_log2_approx(struct lp_build_context
*bld
,
3448 LLVMValueRef
*p_exp
,
3449 LLVMValueRef
*p_floor_log2
,
3450 LLVMValueRef
*p_log2
,
3451 boolean handle_edge_cases
)
3453 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3454 const struct lp_type type
= bld
->type
;
3455 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3456 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
3458 LLVMValueRef expmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x7f800000);
3459 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x007fffff);
3460 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
3462 LLVMValueRef i
= NULL
;
3463 LLVMValueRef y
= NULL
;
3464 LLVMValueRef z
= NULL
;
3465 LLVMValueRef exp
= NULL
;
3466 LLVMValueRef mant
= NULL
;
3467 LLVMValueRef logexp
= NULL
;
3468 LLVMValueRef p_z
= NULL
;
3469 LLVMValueRef res
= NULL
;
3471 assert(lp_check_value(bld
->type
, x
));
3473 if(p_exp
|| p_floor_log2
|| p_log2
) {
3474 /* TODO: optimize the constant case */
3475 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3476 LLVMIsConstant(x
)) {
3477 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3481 assert(type
.floating
&& type
.width
== 32);
3484 * We don't explicitly handle denormalized numbers. They will yield a
3485 * result in the neighbourhood of -127, which appears to be adequate
3489 i
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3491 /* exp = (float) exponent(x) */
3492 exp
= LLVMBuildAnd(builder
, i
, expmask
, "");
3495 if(p_floor_log2
|| p_log2
) {
3496 logexp
= LLVMBuildLShr(builder
, exp
, lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3497 logexp
= LLVMBuildSub(builder
, logexp
, lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3498 logexp
= LLVMBuildSIToFP(builder
, logexp
, vec_type
, "");
3502 /* mant = 1 + (float) mantissa(x) */
3503 mant
= LLVMBuildAnd(builder
, i
, mantmask
, "");
3504 mant
= LLVMBuildOr(builder
, mant
, one
, "");
3505 mant
= LLVMBuildBitCast(builder
, mant
, vec_type
, "");
3507 /* y = (mant - 1) / (mant + 1) */
3508 y
= lp_build_div(bld
,
3509 lp_build_sub(bld
, mant
, bld
->one
),
3510 lp_build_add(bld
, mant
, bld
->one
)
3514 z
= lp_build_mul(bld
, y
, y
);
3517 p_z
= lp_build_polynomial(bld
, z
, lp_build_log2_polynomial
,
3518 ARRAY_SIZE(lp_build_log2_polynomial
));
3520 /* y * P(z) + logexp */
3521 res
= lp_build_mad(bld
, y
, p_z
, logexp
);
3523 if (type
.floating
&& handle_edge_cases
) {
3524 LLVMValueRef negmask
, infmask
, zmask
;
3525 negmask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, x
,
3526 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3527 zmask
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, x
,
3528 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3529 infmask
= lp_build_cmp(bld
, PIPE_FUNC_GEQUAL
, x
,
3530 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
));
3532 /* If x is qual to inf make sure we return inf */
3533 res
= lp_build_select(bld
, infmask
,
3534 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
),
3536 /* If x is qual to 0, return -inf */
3537 res
= lp_build_select(bld
, zmask
,
3538 lp_build_const_vec(bld
->gallivm
, type
, -INFINITY
),
3540 /* If x is nan or less than 0, return nan */
3541 res
= lp_build_select(bld
, negmask
,
3542 lp_build_const_vec(bld
->gallivm
, type
, NAN
),
3548 exp
= LLVMBuildBitCast(builder
, exp
, vec_type
, "");
3553 *p_floor_log2
= logexp
;
3561 * log2 implementation which doesn't have special code to
3562 * handle edge cases (-inf, 0, inf, NaN). It's faster but
3563 * the results for those cases are undefined.
3566 lp_build_log2(struct lp_build_context
*bld
,
3570 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, FALSE
);
3575 * Version of log2 which handles all edge cases.
3576 * Look at documentation of lp_build_log2_approx for
3577 * description of the behavior for each of the edge cases.
3580 lp_build_log2_safe(struct lp_build_context
*bld
,
3584 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, TRUE
);
3590 * Faster (and less accurate) log2.
3592 * log2(x) = floor(log2(x)) - 1 + x / 2**floor(log2(x))
3594 * Piece-wise linear approximation, with exact results when x is a
3597 * See http://www.flipcode.com/archives/Fast_log_Function.shtml
3600 lp_build_fast_log2(struct lp_build_context
*bld
,
3603 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3607 assert(lp_check_value(bld
->type
, x
));
3609 assert(bld
->type
.floating
);
3611 /* ipart = floor(log2(x)) - 1 */
3612 ipart
= lp_build_extract_exponent(bld
, x
, -1);
3613 ipart
= LLVMBuildSIToFP(builder
, ipart
, bld
->vec_type
, "");
3615 /* fpart = x / 2**ipart */
3616 fpart
= lp_build_extract_mantissa(bld
, x
);
3619 return LLVMBuildFAdd(builder
, ipart
, fpart
, "");
3624 * Fast implementation of iround(log2(x)).
3626 * Not an approximation -- it should give accurate results all the time.
3629 lp_build_ilog2(struct lp_build_context
*bld
,
3632 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3633 LLVMValueRef sqrt2
= lp_build_const_vec(bld
->gallivm
, bld
->type
, M_SQRT2
);
3636 assert(bld
->type
.floating
);
3638 assert(lp_check_value(bld
->type
, x
));
3640 /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
3641 x
= LLVMBuildFMul(builder
, x
, sqrt2
, "");
3643 /* ipart = floor(log2(x) + 0.5) */
3644 ipart
= lp_build_extract_exponent(bld
, x
, 0);
3650 lp_build_mod(struct lp_build_context
*bld
,
3654 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3656 const struct lp_type type
= bld
->type
;
3658 assert(lp_check_value(type
, x
));
3659 assert(lp_check_value(type
, y
));
3662 res
= LLVMBuildFRem(builder
, x
, y
, "");
3664 res
= LLVMBuildSRem(builder
, x
, y
, "");
3666 res
= LLVMBuildURem(builder
, x
, y
, "");
3672 * For floating inputs it creates and returns a mask
3673 * which is all 1's for channels which are NaN.
3674 * Channels inside x which are not NaN will be 0.
3677 lp_build_isnan(struct lp_build_context
*bld
,
3681 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3683 assert(bld
->type
.floating
);
3684 assert(lp_check_value(bld
->type
, x
));
3686 mask
= LLVMBuildFCmp(bld
->gallivm
->builder
, LLVMRealOEQ
, x
, x
,
3688 mask
= LLVMBuildNot(bld
->gallivm
->builder
, mask
, "");
3689 mask
= LLVMBuildSExt(bld
->gallivm
->builder
, mask
, int_vec_type
, "isnan");
3693 /* Returns all 1's for floating point numbers that are
3694 * finite numbers and returns all zeros for -inf,
3697 lp_build_isfinite(struct lp_build_context
*bld
,
3700 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3701 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3702 struct lp_type int_type
= lp_int_type(bld
->type
);
3703 LLVMValueRef intx
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3704 LLVMValueRef infornan32
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
3707 if (!bld
->type
.floating
) {
3708 return lp_build_const_int_vec(bld
->gallivm
, bld
->type
, 0);
3710 assert(bld
->type
.floating
);
3711 assert(lp_check_value(bld
->type
, x
));
3712 assert(bld
->type
.width
== 32);
3714 intx
= LLVMBuildAnd(builder
, intx
, infornan32
, "");
3715 return lp_build_compare(bld
->gallivm
, int_type
, PIPE_FUNC_NOTEQUAL
,
3720 * Returns true if the number is nan or inf and false otherwise.
3721 * The input has to be a floating point vector.
3724 lp_build_is_inf_or_nan(struct gallivm_state
*gallivm
,
3725 const struct lp_type type
,
3728 LLVMBuilderRef builder
= gallivm
->builder
;
3729 struct lp_type int_type
= lp_int_type(type
);
3730 LLVMValueRef const0
= lp_build_const_int_vec(gallivm
, int_type
,
3734 assert(type
.floating
);
3736 ret
= LLVMBuildBitCast(builder
, x
, lp_build_vec_type(gallivm
, int_type
), "");
3737 ret
= LLVMBuildAnd(builder
, ret
, const0
, "");
3738 ret
= lp_build_compare(gallivm
, int_type
, PIPE_FUNC_EQUAL
,
3746 lp_build_fpstate_get(struct gallivm_state
*gallivm
)
3748 if (util_cpu_caps
.has_sse
) {
3749 LLVMBuilderRef builder
= gallivm
->builder
;
3750 LLVMValueRef mxcsr_ptr
= lp_build_alloca(
3752 LLVMInt32TypeInContext(gallivm
->context
),
3754 LLVMValueRef mxcsr_ptr8
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3755 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3756 lp_build_intrinsic(builder
,
3757 "llvm.x86.sse.stmxcsr",
3758 LLVMVoidTypeInContext(gallivm
->context
),
3766 lp_build_fpstate_set_denorms_zero(struct gallivm_state
*gallivm
,
3769 if (util_cpu_caps
.has_sse
) {
3770 /* turn on DAZ (64) | FTZ (32768) = 32832 if available */
3771 int daz_ftz
= _MM_FLUSH_ZERO_MASK
;
3773 LLVMBuilderRef builder
= gallivm
->builder
;
3774 LLVMValueRef mxcsr_ptr
= lp_build_fpstate_get(gallivm
);
3775 LLVMValueRef mxcsr
=
3776 LLVMBuildLoad(builder
, mxcsr_ptr
, "mxcsr");
3778 if (util_cpu_caps
.has_daz
) {
3779 /* Enable denormals are zero mode */
3780 daz_ftz
|= _MM_DENORMALS_ZERO_MASK
;
3783 mxcsr
= LLVMBuildOr(builder
, mxcsr
,
3784 LLVMConstInt(LLVMTypeOf(mxcsr
), daz_ftz
, 0), "");
3786 mxcsr
= LLVMBuildAnd(builder
, mxcsr
,
3787 LLVMConstInt(LLVMTypeOf(mxcsr
), ~daz_ftz
, 0), "");
3790 LLVMBuildStore(builder
, mxcsr
, mxcsr_ptr
);
3791 lp_build_fpstate_set(gallivm
, mxcsr_ptr
);
3796 lp_build_fpstate_set(struct gallivm_state
*gallivm
,
3797 LLVMValueRef mxcsr_ptr
)
3799 if (util_cpu_caps
.has_sse
) {
3800 LLVMBuilderRef builder
= gallivm
->builder
;
3801 mxcsr_ptr
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3802 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3803 lp_build_intrinsic(builder
,
3804 "llvm.x86.sse.ldmxcsr",
3805 LLVMVoidTypeInContext(gallivm
->context
),