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 if(((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
1376 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) &&
1378 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
1381 return LLVMBuildFDiv(builder
, a
, b
, "");
1383 return LLVMBuildSDiv(builder
, a
, b
, "");
1385 return LLVMBuildUDiv(builder
, a
, b
, "");
1390 * Linear interpolation helper.
1392 * @param normalized whether we are interpolating normalized values,
1393 * encoded in normalized integers, twice as wide.
1395 * @sa http://www.stereopsis.com/doubleblend.html
1397 static inline LLVMValueRef
1398 lp_build_lerp_simple(struct lp_build_context
*bld
,
1404 unsigned half_width
= bld
->type
.width
/2;
1405 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1409 assert(lp_check_value(bld
->type
, x
));
1410 assert(lp_check_value(bld
->type
, v0
));
1411 assert(lp_check_value(bld
->type
, v1
));
1413 delta
= lp_build_sub(bld
, v1
, v0
);
1415 if (bld
->type
.floating
) {
1417 return lp_build_mad(bld
, x
, delta
, v0
);
1420 if (flags
& LP_BLD_LERP_WIDE_NORMALIZED
) {
1421 if (!bld
->type
.sign
) {
1422 if (!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
)) {
1424 * Scale x from [0, 2**n - 1] to [0, 2**n] by adding the
1425 * most-significant-bit to the lowest-significant-bit, so that
1426 * later we can just divide by 2**n instead of 2**n - 1.
1429 x
= lp_build_add(bld
, x
, lp_build_shr_imm(bld
, x
, half_width
- 1));
1432 /* (x * delta) >> n */
1433 res
= lp_build_mul(bld
, x
, delta
);
1434 res
= lp_build_shr_imm(bld
, res
, half_width
);
1437 * The rescaling trick above doesn't work for signed numbers, so
1438 * use the 2**n - 1 divison approximation in lp_build_mul_norm
1441 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1442 res
= lp_build_mul_norm(bld
->gallivm
, bld
->type
, x
, delta
);
1445 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1446 res
= lp_build_mul(bld
, x
, delta
);
1449 if ((flags
& LP_BLD_LERP_WIDE_NORMALIZED
) && !bld
->type
.sign
) {
1451 * At this point both res and v0 only use the lower half of the bits,
1452 * the rest is zero. Instead of add / mask, do add with half wide type.
1454 struct lp_type narrow_type
;
1455 struct lp_build_context narrow_bld
;
1457 memset(&narrow_type
, 0, sizeof narrow_type
);
1458 narrow_type
.sign
= bld
->type
.sign
;
1459 narrow_type
.width
= bld
->type
.width
/2;
1460 narrow_type
.length
= bld
->type
.length
*2;
1462 lp_build_context_init(&narrow_bld
, bld
->gallivm
, narrow_type
);
1463 res
= LLVMBuildBitCast(builder
, res
, narrow_bld
.vec_type
, "");
1464 v0
= LLVMBuildBitCast(builder
, v0
, narrow_bld
.vec_type
, "");
1465 res
= lp_build_add(&narrow_bld
, v0
, res
);
1466 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
1468 res
= lp_build_add(bld
, v0
, res
);
1470 if (bld
->type
.fixed
) {
1472 * We need to mask out the high order bits when lerping 8bit
1473 * normalized colors stored on 16bits
1475 /* XXX: This step is necessary for lerping 8bit colors stored on
1476 * 16bits, but it will be wrong for true fixed point use cases.
1477 * Basically we need a more powerful lp_type, capable of further
1478 * distinguishing the values interpretation from the value storage.
1480 LLVMValueRef low_bits
;
1481 low_bits
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (1 << half_width
) - 1);
1482 res
= LLVMBuildAnd(builder
, res
, low_bits
, "");
1491 * Linear interpolation.
1494 lp_build_lerp(struct lp_build_context
*bld
,
1500 const struct lp_type type
= bld
->type
;
1503 assert(lp_check_value(type
, x
));
1504 assert(lp_check_value(type
, v0
));
1505 assert(lp_check_value(type
, v1
));
1507 assert(!(flags
& LP_BLD_LERP_WIDE_NORMALIZED
));
1510 struct lp_type wide_type
;
1511 struct lp_build_context wide_bld
;
1512 LLVMValueRef xl
, xh
, v0l
, v0h
, v1l
, v1h
, resl
, resh
;
1514 assert(type
.length
>= 2);
1517 * Create a wider integer type, enough to hold the
1518 * intermediate result of the multiplication.
1520 memset(&wide_type
, 0, sizeof wide_type
);
1521 wide_type
.sign
= type
.sign
;
1522 wide_type
.width
= type
.width
*2;
1523 wide_type
.length
= type
.length
/2;
1525 lp_build_context_init(&wide_bld
, bld
->gallivm
, wide_type
);
1527 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, x
, &xl
, &xh
);
1528 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v0
, &v0l
, &v0h
);
1529 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v1
, &v1l
, &v1h
);
1535 flags
|= LP_BLD_LERP_WIDE_NORMALIZED
;
1537 resl
= lp_build_lerp_simple(&wide_bld
, xl
, v0l
, v1l
, flags
);
1538 resh
= lp_build_lerp_simple(&wide_bld
, xh
, v0h
, v1h
, flags
);
1540 res
= lp_build_pack2_native(bld
->gallivm
, wide_type
, type
, resl
, resh
);
1542 res
= lp_build_lerp_simple(bld
, x
, v0
, v1
, flags
);
1550 * Bilinear interpolation.
1552 * Values indices are in v_{yx}.
1555 lp_build_lerp_2d(struct lp_build_context
*bld
,
1564 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
, flags
);
1565 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
, flags
);
1566 return lp_build_lerp(bld
, y
, v0
, v1
, flags
);
1571 lp_build_lerp_3d(struct lp_build_context
*bld
,
1585 LLVMValueRef v0
= lp_build_lerp_2d(bld
, x
, y
, v000
, v001
, v010
, v011
, flags
);
1586 LLVMValueRef v1
= lp_build_lerp_2d(bld
, x
, y
, v100
, v101
, v110
, v111
, flags
);
1587 return lp_build_lerp(bld
, z
, v0
, v1
, flags
);
1592 * Generate min(a, b)
1593 * Do checks for special cases but not for nans.
1596 lp_build_min(struct lp_build_context
*bld
,
1600 assert(lp_check_value(bld
->type
, a
));
1601 assert(lp_check_value(bld
->type
, b
));
1603 if(a
== bld
->undef
|| b
== bld
->undef
)
1609 if (bld
->type
.norm
) {
1610 if (!bld
->type
.sign
) {
1611 if (a
== bld
->zero
|| b
== bld
->zero
) {
1621 return lp_build_min_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1626 * Generate min(a, b)
1627 * NaN's are handled according to the behavior specified by the
1628 * nan_behavior argument.
1631 lp_build_min_ext(struct lp_build_context
*bld
,
1634 enum gallivm_nan_behavior nan_behavior
)
1636 assert(lp_check_value(bld
->type
, a
));
1637 assert(lp_check_value(bld
->type
, b
));
1639 if(a
== bld
->undef
|| b
== bld
->undef
)
1645 if (bld
->type
.norm
) {
1646 if (!bld
->type
.sign
) {
1647 if (a
== bld
->zero
|| b
== bld
->zero
) {
1657 return lp_build_min_simple(bld
, a
, b
, nan_behavior
);
1661 * Generate max(a, b)
1662 * Do checks for special cases, but NaN behavior is undefined.
1665 lp_build_max(struct lp_build_context
*bld
,
1669 assert(lp_check_value(bld
->type
, a
));
1670 assert(lp_check_value(bld
->type
, b
));
1672 if(a
== bld
->undef
|| b
== bld
->undef
)
1678 if(bld
->type
.norm
) {
1679 if(a
== bld
->one
|| b
== bld
->one
)
1681 if (!bld
->type
.sign
) {
1682 if (a
== bld
->zero
) {
1685 if (b
== bld
->zero
) {
1691 return lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1696 * Generate max(a, b)
1697 * Checks for special cases.
1698 * NaN's are handled according to the behavior specified by the
1699 * nan_behavior argument.
1702 lp_build_max_ext(struct lp_build_context
*bld
,
1705 enum gallivm_nan_behavior nan_behavior
)
1707 assert(lp_check_value(bld
->type
, a
));
1708 assert(lp_check_value(bld
->type
, b
));
1710 if(a
== bld
->undef
|| b
== bld
->undef
)
1716 if(bld
->type
.norm
) {
1717 if(a
== bld
->one
|| b
== bld
->one
)
1719 if (!bld
->type
.sign
) {
1720 if (a
== bld
->zero
) {
1723 if (b
== bld
->zero
) {
1729 return lp_build_max_simple(bld
, a
, b
, nan_behavior
);
1733 * Generate clamp(a, min, max)
1734 * NaN behavior (for any of a, min, max) is undefined.
1735 * Do checks for special cases.
1738 lp_build_clamp(struct lp_build_context
*bld
,
1743 assert(lp_check_value(bld
->type
, a
));
1744 assert(lp_check_value(bld
->type
, min
));
1745 assert(lp_check_value(bld
->type
, max
));
1747 a
= lp_build_min(bld
, a
, max
);
1748 a
= lp_build_max(bld
, a
, min
);
1754 * Generate clamp(a, 0, 1)
1755 * A NaN will get converted to zero.
1758 lp_build_clamp_zero_one_nanzero(struct lp_build_context
*bld
,
1761 a
= lp_build_max_ext(bld
, a
, bld
->zero
, GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1762 a
= lp_build_min(bld
, a
, bld
->one
);
1771 lp_build_abs(struct lp_build_context
*bld
,
1774 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1775 const struct lp_type type
= bld
->type
;
1776 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1778 assert(lp_check_value(type
, a
));
1784 if (0x0306 <= HAVE_LLVM
&& HAVE_LLVM
< 0x0309) {
1785 /* Workaround llvm.org/PR27332 */
1786 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1787 unsigned long long absMask
= ~(1ULL << (type
.width
- 1));
1788 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
, ((unsigned long long) absMask
));
1789 a
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1790 a
= LLVMBuildAnd(builder
, a
, mask
, "");
1791 a
= LLVMBuildBitCast(builder
, a
, vec_type
, "");
1795 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.fabs", vec_type
);
1796 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
1800 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
) {
1801 switch(type
.width
) {
1803 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
1805 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
1807 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
1810 else if (type
.width
*type
.length
== 256 && util_cpu_caps
.has_avx2
) {
1811 switch(type
.width
) {
1813 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.b", vec_type
, a
);
1815 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.w", vec_type
, a
);
1817 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.d", vec_type
, a
);
1820 else if (type
.width
*type
.length
== 256 && util_cpu_caps
.has_ssse3
&&
1821 (gallivm_debug
& GALLIVM_DEBUG_PERF
) &&
1822 (type
.width
== 8 || type
.width
== 16 || type
.width
== 32)) {
1823 debug_printf("%s: inefficient code, should split vectors manually\n",
1827 return lp_build_max(bld
, a
, LLVMBuildNeg(builder
, a
, ""));
1832 lp_build_negate(struct lp_build_context
*bld
,
1835 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1837 assert(lp_check_value(bld
->type
, a
));
1839 if (bld
->type
.floating
)
1840 a
= LLVMBuildFNeg(builder
, a
, "");
1842 a
= LLVMBuildNeg(builder
, a
, "");
1848 /** Return -1, 0 or +1 depending on the sign of a */
1850 lp_build_sgn(struct lp_build_context
*bld
,
1853 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1854 const struct lp_type type
= bld
->type
;
1858 assert(lp_check_value(type
, a
));
1860 /* Handle non-zero case */
1862 /* if not zero then sign must be positive */
1865 else if(type
.floating
) {
1866 LLVMTypeRef vec_type
;
1867 LLVMTypeRef int_type
;
1871 unsigned long long maskBit
= (unsigned long long)1 << (type
.width
- 1);
1873 int_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1874 vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1875 mask
= lp_build_const_int_vec(bld
->gallivm
, type
, maskBit
);
1877 /* Take the sign bit and add it to 1 constant */
1878 sign
= LLVMBuildBitCast(builder
, a
, int_type
, "");
1879 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
1880 one
= LLVMConstBitCast(bld
->one
, int_type
);
1881 res
= LLVMBuildOr(builder
, sign
, one
, "");
1882 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1886 /* signed int/norm/fixed point */
1887 /* could use psign with sse3 and appropriate vectors here */
1888 LLVMValueRef minus_one
= lp_build_const_vec(bld
->gallivm
, type
, -1.0);
1889 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
1890 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
1894 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
1895 res
= lp_build_select(bld
, cond
, bld
->zero
, res
);
1902 * Set the sign of float vector 'a' according to 'sign'.
1903 * If sign==0, return abs(a).
1904 * If sign==1, return -abs(a);
1905 * Other values for sign produce undefined results.
1908 lp_build_set_sign(struct lp_build_context
*bld
,
1909 LLVMValueRef a
, LLVMValueRef sign
)
1911 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1912 const struct lp_type type
= bld
->type
;
1913 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1914 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1915 LLVMValueRef shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
- 1);
1916 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
1917 ~((unsigned long long) 1 << (type
.width
- 1)));
1918 LLVMValueRef val
, res
;
1920 assert(type
.floating
);
1921 assert(lp_check_value(type
, a
));
1923 /* val = reinterpret_cast<int>(a) */
1924 val
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1925 /* val = val & mask */
1926 val
= LLVMBuildAnd(builder
, val
, mask
, "");
1927 /* sign = sign << shift */
1928 sign
= LLVMBuildShl(builder
, sign
, shift
, "");
1929 /* res = val | sign */
1930 res
= LLVMBuildOr(builder
, val
, sign
, "");
1931 /* res = reinterpret_cast<float>(res) */
1932 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1939 * Convert vector of (or scalar) int to vector of (or scalar) float.
1942 lp_build_int_to_float(struct lp_build_context
*bld
,
1945 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1946 const struct lp_type type
= bld
->type
;
1947 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1949 assert(type
.floating
);
1951 return LLVMBuildSIToFP(builder
, a
, vec_type
, "");
1955 arch_rounding_available(const struct lp_type type
)
1957 if ((util_cpu_caps
.has_sse4_1
&&
1958 (type
.length
== 1 || type
.width
*type
.length
== 128)) ||
1959 (util_cpu_caps
.has_avx
&& type
.width
*type
.length
== 256))
1961 else if ((util_cpu_caps
.has_altivec
&&
1962 (type
.width
== 32 && type
.length
== 4)))
1968 enum lp_build_round_mode
1970 LP_BUILD_ROUND_NEAREST
= 0,
1971 LP_BUILD_ROUND_FLOOR
= 1,
1972 LP_BUILD_ROUND_CEIL
= 2,
1973 LP_BUILD_ROUND_TRUNCATE
= 3
1976 static inline LLVMValueRef
1977 lp_build_iround_nearest_sse2(struct lp_build_context
*bld
,
1980 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1981 const struct lp_type type
= bld
->type
;
1982 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
1983 LLVMTypeRef ret_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1984 const char *intrinsic
;
1987 assert(type
.floating
);
1988 /* using the double precision conversions is a bit more complicated */
1989 assert(type
.width
== 32);
1991 assert(lp_check_value(type
, a
));
1992 assert(util_cpu_caps
.has_sse2
);
1994 /* This is relying on MXCSR rounding mode, which should always be nearest. */
1995 if (type
.length
== 1) {
1996 LLVMTypeRef vec_type
;
1999 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
2001 vec_type
= LLVMVectorType(bld
->elem_type
, 4);
2003 intrinsic
= "llvm.x86.sse.cvtss2si";
2005 undef
= LLVMGetUndef(vec_type
);
2007 arg
= LLVMBuildInsertElement(builder
, undef
, a
, index0
, "");
2009 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2013 if (type
.width
* type
.length
== 128) {
2014 intrinsic
= "llvm.x86.sse2.cvtps2dq";
2017 assert(type
.width
*type
.length
== 256);
2018 assert(util_cpu_caps
.has_avx
);
2020 intrinsic
= "llvm.x86.avx.cvt.ps2dq.256";
2022 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2032 static inline LLVMValueRef
2033 lp_build_round_altivec(struct lp_build_context
*bld
,
2035 enum lp_build_round_mode mode
)
2037 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2038 const struct lp_type type
= bld
->type
;
2039 const char *intrinsic
= NULL
;
2041 assert(type
.floating
);
2043 assert(lp_check_value(type
, a
));
2044 assert(util_cpu_caps
.has_altivec
);
2049 case LP_BUILD_ROUND_NEAREST
:
2050 intrinsic
= "llvm.ppc.altivec.vrfin";
2052 case LP_BUILD_ROUND_FLOOR
:
2053 intrinsic
= "llvm.ppc.altivec.vrfim";
2055 case LP_BUILD_ROUND_CEIL
:
2056 intrinsic
= "llvm.ppc.altivec.vrfip";
2058 case LP_BUILD_ROUND_TRUNCATE
:
2059 intrinsic
= "llvm.ppc.altivec.vrfiz";
2063 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2066 static inline LLVMValueRef
2067 lp_build_round_arch(struct lp_build_context
*bld
,
2069 enum lp_build_round_mode mode
)
2071 if (util_cpu_caps
.has_sse4_1
) {
2072 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2073 const struct lp_type type
= bld
->type
;
2074 const char *intrinsic_root
;
2077 assert(type
.floating
);
2078 assert(lp_check_value(type
, a
));
2082 case LP_BUILD_ROUND_NEAREST
:
2083 intrinsic_root
= "llvm.nearbyint";
2085 case LP_BUILD_ROUND_FLOOR
:
2086 intrinsic_root
= "llvm.floor";
2088 case LP_BUILD_ROUND_CEIL
:
2089 intrinsic_root
= "llvm.ceil";
2091 case LP_BUILD_ROUND_TRUNCATE
:
2092 intrinsic_root
= "llvm.trunc";
2096 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, intrinsic_root
, bld
->vec_type
);
2097 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2099 else /* (util_cpu_caps.has_altivec) */
2100 return lp_build_round_altivec(bld
, a
, mode
);
2104 * Return the integer part of a float (vector) value (== round toward zero).
2105 * The returned value is a float (vector).
2106 * Ex: trunc(-1.5) = -1.0
2109 lp_build_trunc(struct lp_build_context
*bld
,
2112 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2113 const struct lp_type type
= bld
->type
;
2115 assert(type
.floating
);
2116 assert(lp_check_value(type
, a
));
2118 if (arch_rounding_available(type
)) {
2119 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_TRUNCATE
);
2122 const struct lp_type type
= bld
->type
;
2123 struct lp_type inttype
;
2124 struct lp_build_context intbld
;
2125 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2126 LLVMValueRef trunc
, res
, anosign
, mask
;
2127 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2128 LLVMTypeRef vec_type
= bld
->vec_type
;
2130 assert(type
.width
== 32); /* might want to handle doubles at some point */
2133 inttype
.floating
= 0;
2134 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2136 /* round by truncation */
2137 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2138 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2140 /* mask out sign bit */
2141 anosign
= lp_build_abs(bld
, a
);
2143 * mask out all values if anosign > 2^24
2144 * This should work both for large ints (all rounding is no-op for them
2145 * because such floats are always exact) as well as special cases like
2146 * NaNs, Infs (taking advantage of the fact they use max exponent).
2147 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2149 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2150 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2151 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2152 return lp_build_select(bld
, mask
, a
, res
);
2158 * Return float (vector) rounded to nearest integer (vector). The returned
2159 * value is a float (vector).
2160 * Ex: round(0.9) = 1.0
2161 * Ex: round(-1.5) = -2.0
2164 lp_build_round(struct lp_build_context
*bld
,
2167 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2168 const struct lp_type type
= bld
->type
;
2170 assert(type
.floating
);
2171 assert(lp_check_value(type
, a
));
2173 if (arch_rounding_available(type
)) {
2174 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2177 const struct lp_type type
= bld
->type
;
2178 struct lp_type inttype
;
2179 struct lp_build_context intbld
;
2180 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2181 LLVMValueRef res
, anosign
, mask
;
2182 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2183 LLVMTypeRef vec_type
= bld
->vec_type
;
2185 assert(type
.width
== 32); /* might want to handle doubles at some point */
2188 inttype
.floating
= 0;
2189 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2191 res
= lp_build_iround(bld
, a
);
2192 res
= LLVMBuildSIToFP(builder
, res
, vec_type
, "");
2194 /* mask out sign bit */
2195 anosign
= lp_build_abs(bld
, a
);
2197 * mask out all values if anosign > 2^24
2198 * This should work both for large ints (all rounding is no-op for them
2199 * because such floats are always exact) as well as special cases like
2200 * NaNs, Infs (taking advantage of the fact they use max exponent).
2201 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2203 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2204 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2205 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2206 return lp_build_select(bld
, mask
, a
, res
);
2212 * Return floor of float (vector), result is a float (vector)
2213 * Ex: floor(1.1) = 1.0
2214 * Ex: floor(-1.1) = -2.0
2217 lp_build_floor(struct lp_build_context
*bld
,
2220 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2221 const struct lp_type type
= bld
->type
;
2223 assert(type
.floating
);
2224 assert(lp_check_value(type
, a
));
2226 if (arch_rounding_available(type
)) {
2227 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2230 const struct lp_type type
= bld
->type
;
2231 struct lp_type inttype
;
2232 struct lp_build_context intbld
;
2233 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2234 LLVMValueRef trunc
, res
, anosign
, mask
;
2235 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2236 LLVMTypeRef vec_type
= bld
->vec_type
;
2238 if (type
.width
!= 32) {
2240 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.floor", vec_type
);
2241 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2244 assert(type
.width
== 32); /* might want to handle doubles at some point */
2247 inttype
.floating
= 0;
2248 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2250 /* round by truncation */
2251 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2252 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2258 * fix values if rounding is wrong (for non-special cases)
2259 * - this is the case if trunc > a
2261 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, res
, a
);
2262 /* tmp = trunc > a ? 1.0 : 0.0 */
2263 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2264 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2265 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2266 res
= lp_build_sub(bld
, res
, tmp
);
2269 /* mask out sign bit */
2270 anosign
= lp_build_abs(bld
, a
);
2272 * mask out all values if anosign > 2^24
2273 * This should work both for large ints (all rounding is no-op for them
2274 * because such floats are always exact) as well as special cases like
2275 * NaNs, Infs (taking advantage of the fact they use max exponent).
2276 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2278 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2279 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2280 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2281 return lp_build_select(bld
, mask
, a
, res
);
2287 * Return ceiling of float (vector), returning float (vector).
2288 * Ex: ceil( 1.1) = 2.0
2289 * Ex: ceil(-1.1) = -1.0
2292 lp_build_ceil(struct lp_build_context
*bld
,
2295 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2296 const struct lp_type type
= bld
->type
;
2298 assert(type
.floating
);
2299 assert(lp_check_value(type
, a
));
2301 if (arch_rounding_available(type
)) {
2302 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2305 const struct lp_type type
= bld
->type
;
2306 struct lp_type inttype
;
2307 struct lp_build_context intbld
;
2308 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2309 LLVMValueRef trunc
, res
, anosign
, mask
, tmp
;
2310 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2311 LLVMTypeRef vec_type
= bld
->vec_type
;
2313 if (type
.width
!= 32) {
2315 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.ceil", vec_type
);
2316 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2319 assert(type
.width
== 32); /* might want to handle doubles at some point */
2322 inttype
.floating
= 0;
2323 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2325 /* round by truncation */
2326 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2327 trunc
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "ceil.trunc");
2330 * fix values if rounding is wrong (for non-special cases)
2331 * - this is the case if trunc < a
2333 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2334 /* tmp = trunc < a ? 1.0 : 0.0 */
2335 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2336 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2337 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2338 res
= lp_build_add(bld
, trunc
, tmp
);
2340 /* mask out sign bit */
2341 anosign
= lp_build_abs(bld
, a
);
2343 * mask out all values if anosign > 2^24
2344 * This should work both for large ints (all rounding is no-op for them
2345 * because such floats are always exact) as well as special cases like
2346 * NaNs, Infs (taking advantage of the fact they use max exponent).
2347 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2349 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2350 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2351 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2352 return lp_build_select(bld
, mask
, a
, res
);
2358 * Return fractional part of 'a' computed as a - floor(a)
2359 * Typically used in texture coord arithmetic.
2362 lp_build_fract(struct lp_build_context
*bld
,
2365 assert(bld
->type
.floating
);
2366 return lp_build_sub(bld
, a
, lp_build_floor(bld
, a
));
2371 * Prevent returning 1.0 for very small negative values of 'a' by clamping
2372 * against 0.99999(9). (Will also return that value for NaNs.)
2374 static inline LLVMValueRef
2375 clamp_fract(struct lp_build_context
*bld
, LLVMValueRef fract
)
2379 /* this is the largest number smaller than 1.0 representable as float */
2380 max
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2381 1.0 - 1.0/(1LL << (lp_mantissa(bld
->type
) + 1)));
2382 return lp_build_min_ext(bld
, fract
, max
,
2383 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
2388 * Same as lp_build_fract, but guarantees that the result is always smaller
2389 * than one. Will also return the smaller-than-one value for infs, NaNs.
2392 lp_build_fract_safe(struct lp_build_context
*bld
,
2395 return clamp_fract(bld
, lp_build_fract(bld
, a
));
2400 * Return the integer part of a float (vector) value (== round toward zero).
2401 * The returned value is an integer (vector).
2402 * Ex: itrunc(-1.5) = -1
2405 lp_build_itrunc(struct lp_build_context
*bld
,
2408 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2409 const struct lp_type type
= bld
->type
;
2410 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
2412 assert(type
.floating
);
2413 assert(lp_check_value(type
, a
));
2415 return LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2420 * Return float (vector) rounded to nearest integer (vector). The returned
2421 * value is an integer (vector).
2422 * Ex: iround(0.9) = 1
2423 * Ex: iround(-1.5) = -2
2426 lp_build_iround(struct lp_build_context
*bld
,
2429 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2430 const struct lp_type type
= bld
->type
;
2431 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2434 assert(type
.floating
);
2436 assert(lp_check_value(type
, a
));
2438 if ((util_cpu_caps
.has_sse2
&&
2439 ((type
.width
== 32) && (type
.length
== 1 || type
.length
== 4))) ||
2440 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2441 return lp_build_iround_nearest_sse2(bld
, a
);
2443 if (arch_rounding_available(type
)) {
2444 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2449 half
= lp_build_const_vec(bld
->gallivm
, type
, 0.5);
2452 LLVMTypeRef vec_type
= bld
->vec_type
;
2453 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
2454 (unsigned long long)1 << (type
.width
- 1));
2458 sign
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
2459 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
2462 half
= LLVMBuildBitCast(builder
, half
, int_vec_type
, "");
2463 half
= LLVMBuildOr(builder
, sign
, half
, "");
2464 half
= LLVMBuildBitCast(builder
, half
, vec_type
, "");
2467 res
= LLVMBuildFAdd(builder
, a
, half
, "");
2470 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "");
2477 * Return floor of float (vector), result is an int (vector)
2478 * Ex: ifloor(1.1) = 1.0
2479 * Ex: ifloor(-1.1) = -2.0
2482 lp_build_ifloor(struct lp_build_context
*bld
,
2485 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2486 const struct lp_type type
= bld
->type
;
2487 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2490 assert(type
.floating
);
2491 assert(lp_check_value(type
, a
));
2495 if (arch_rounding_available(type
)) {
2496 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2499 struct lp_type inttype
;
2500 struct lp_build_context intbld
;
2501 LLVMValueRef trunc
, itrunc
, mask
;
2503 assert(type
.floating
);
2504 assert(lp_check_value(type
, a
));
2507 inttype
.floating
= 0;
2508 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2510 /* round by truncation */
2511 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2512 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "ifloor.trunc");
2515 * fix values if rounding is wrong (for non-special cases)
2516 * - this is the case if trunc > a
2517 * The results of doing this with NaNs, very large values etc.
2518 * are undefined but this seems to be the case anyway.
2520 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, trunc
, a
);
2521 /* cheapie minus one with mask since the mask is minus one / zero */
2522 return lp_build_add(&intbld
, itrunc
, mask
);
2526 /* round to nearest (toward zero) */
2527 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "ifloor.res");
2534 * Return ceiling of float (vector), returning int (vector).
2535 * Ex: iceil( 1.1) = 2
2536 * Ex: iceil(-1.1) = -1
2539 lp_build_iceil(struct lp_build_context
*bld
,
2542 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2543 const struct lp_type type
= bld
->type
;
2544 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2547 assert(type
.floating
);
2548 assert(lp_check_value(type
, a
));
2550 if (arch_rounding_available(type
)) {
2551 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2554 struct lp_type inttype
;
2555 struct lp_build_context intbld
;
2556 LLVMValueRef trunc
, itrunc
, mask
;
2558 assert(type
.floating
);
2559 assert(lp_check_value(type
, a
));
2562 inttype
.floating
= 0;
2563 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2565 /* round by truncation */
2566 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2567 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "iceil.trunc");
2570 * fix values if rounding is wrong (for non-special cases)
2571 * - this is the case if trunc < a
2572 * The results of doing this with NaNs, very large values etc.
2573 * are undefined but this seems to be the case anyway.
2575 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2576 /* cheapie plus one with mask since the mask is minus one / zero */
2577 return lp_build_sub(&intbld
, itrunc
, mask
);
2580 /* round to nearest (toward zero) */
2581 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "iceil.res");
2588 * Combined ifloor() & fract().
2590 * Preferred to calling the functions separately, as it will ensure that the
2591 * strategy (floor() vs ifloor()) that results in less redundant work is used.
2594 lp_build_ifloor_fract(struct lp_build_context
*bld
,
2596 LLVMValueRef
*out_ipart
,
2597 LLVMValueRef
*out_fpart
)
2599 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2600 const struct lp_type type
= bld
->type
;
2603 assert(type
.floating
);
2604 assert(lp_check_value(type
, a
));
2606 if (arch_rounding_available(type
)) {
2608 * floor() is easier.
2611 ipart
= lp_build_floor(bld
, a
);
2612 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2613 *out_ipart
= LLVMBuildFPToSI(builder
, ipart
, bld
->int_vec_type
, "ipart");
2617 * ifloor() is easier.
2620 *out_ipart
= lp_build_ifloor(bld
, a
);
2621 ipart
= LLVMBuildSIToFP(builder
, *out_ipart
, bld
->vec_type
, "ipart");
2622 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2628 * Same as lp_build_ifloor_fract, but guarantees that the fractional part is
2629 * always smaller than one.
2632 lp_build_ifloor_fract_safe(struct lp_build_context
*bld
,
2634 LLVMValueRef
*out_ipart
,
2635 LLVMValueRef
*out_fpart
)
2637 lp_build_ifloor_fract(bld
, a
, out_ipart
, out_fpart
);
2638 *out_fpart
= clamp_fract(bld
, *out_fpart
);
2643 lp_build_sqrt(struct lp_build_context
*bld
,
2646 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2647 const struct lp_type type
= bld
->type
;
2648 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
2651 assert(lp_check_value(type
, a
));
2653 assert(type
.floating
);
2654 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.sqrt", vec_type
);
2656 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2661 * Do one Newton-Raphson step to improve reciprocate precision:
2663 * x_{i+1} = x_i * (2 - a * x_i)
2665 * XXX: Unfortunately this won't give IEEE-754 conformant results for 0 or
2666 * +/-Inf, giving NaN instead. Certain applications rely on this behavior,
2667 * such as Google Earth, which does RCP(RSQRT(0.0) when drawing the Earth's
2668 * halo. It would be necessary to clamp the argument to prevent this.
2671 * - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division
2672 * - http://softwarecommunity.intel.com/articles/eng/1818.htm
2674 static inline LLVMValueRef
2675 lp_build_rcp_refine(struct lp_build_context
*bld
,
2679 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2680 LLVMValueRef two
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 2.0);
2683 res
= LLVMBuildFMul(builder
, a
, rcp_a
, "");
2684 res
= LLVMBuildFSub(builder
, two
, res
, "");
2685 res
= LLVMBuildFMul(builder
, rcp_a
, res
, "");
2692 lp_build_rcp(struct lp_build_context
*bld
,
2695 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2696 const struct lp_type type
= bld
->type
;
2698 assert(lp_check_value(type
, a
));
2707 assert(type
.floating
);
2709 if(LLVMIsConstant(a
))
2710 return LLVMConstFDiv(bld
->one
, a
);
2713 * We don't use RCPPS because:
2714 * - it only has 10bits of precision
2715 * - it doesn't even get the reciprocate of 1.0 exactly
2716 * - doing Newton-Rapshon steps yields wrong (NaN) values for 0.0 or Inf
2717 * - for recent processors the benefit over DIVPS is marginal, a case
2720 * We could still use it on certain processors if benchmarks show that the
2721 * RCPPS plus necessary workarounds are still preferrable to DIVPS; or for
2722 * particular uses that require less workarounds.
2725 if (FALSE
&& ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2726 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8))){
2727 const unsigned num_iterations
= 0;
2730 const char *intrinsic
= NULL
;
2732 if (type
.length
== 4) {
2733 intrinsic
= "llvm.x86.sse.rcp.ps";
2736 intrinsic
= "llvm.x86.avx.rcp.ps.256";
2739 res
= lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2741 for (i
= 0; i
< num_iterations
; ++i
) {
2742 res
= lp_build_rcp_refine(bld
, a
, res
);
2748 return LLVMBuildFDiv(builder
, bld
->one
, a
, "");
2753 * Do one Newton-Raphson step to improve rsqrt precision:
2755 * x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
2757 * See also Intel 64 and IA-32 Architectures Optimization Manual.
2759 static inline LLVMValueRef
2760 lp_build_rsqrt_refine(struct lp_build_context
*bld
,
2762 LLVMValueRef rsqrt_a
)
2764 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2765 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 0.5);
2766 LLVMValueRef three
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 3.0);
2769 res
= LLVMBuildFMul(builder
, rsqrt_a
, rsqrt_a
, "");
2770 res
= LLVMBuildFMul(builder
, a
, res
, "");
2771 res
= LLVMBuildFSub(builder
, three
, res
, "");
2772 res
= LLVMBuildFMul(builder
, rsqrt_a
, res
, "");
2773 res
= LLVMBuildFMul(builder
, half
, res
, "");
2780 * Generate 1/sqrt(a).
2781 * Result is undefined for values < 0, infinity for +0.
2784 lp_build_rsqrt(struct lp_build_context
*bld
,
2787 const struct lp_type type
= bld
->type
;
2789 assert(lp_check_value(type
, a
));
2791 assert(type
.floating
);
2794 * This should be faster but all denormals will end up as infinity.
2796 if (0 && lp_build_fast_rsqrt_available(type
)) {
2797 const unsigned num_iterations
= 1;
2801 /* rsqrt(1.0) != 1.0 here */
2802 res
= lp_build_fast_rsqrt(bld
, a
);
2804 if (num_iterations
) {
2806 * Newton-Raphson will result in NaN instead of infinity for zero,
2807 * and NaN instead of zero for infinity.
2808 * Also, need to ensure rsqrt(1.0) == 1.0.
2809 * All numbers smaller than FLT_MIN will result in +infinity
2810 * (rsqrtps treats all denormals as zero).
2813 LLVMValueRef flt_min
= lp_build_const_vec(bld
->gallivm
, type
, FLT_MIN
);
2814 LLVMValueRef inf
= lp_build_const_vec(bld
->gallivm
, type
, INFINITY
);
2816 for (i
= 0; i
< num_iterations
; ++i
) {
2817 res
= lp_build_rsqrt_refine(bld
, a
, res
);
2819 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_LESS
, a
, flt_min
);
2820 res
= lp_build_select(bld
, cmp
, inf
, res
);
2821 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, inf
);
2822 res
= lp_build_select(bld
, cmp
, bld
->zero
, res
);
2823 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, bld
->one
);
2824 res
= lp_build_select(bld
, cmp
, bld
->one
, res
);
2830 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2834 * If there's a fast (inaccurate) rsqrt instruction available
2835 * (caller may want to avoid to call rsqrt_fast if it's not available,
2836 * i.e. for calculating x^0.5 it may do rsqrt_fast(x) * x but if
2837 * unavailable it would result in sqrt/div/mul so obviously
2838 * much better to just call sqrt, skipping both div and mul).
2841 lp_build_fast_rsqrt_available(struct lp_type type
)
2843 assert(type
.floating
);
2845 if ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2846 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2854 * Generate 1/sqrt(a).
2855 * Result is undefined for values < 0, infinity for +0.
2856 * Precision is limited, only ~10 bits guaranteed
2857 * (rsqrt 1.0 may not be 1.0, denorms may be flushed to 0).
2860 lp_build_fast_rsqrt(struct lp_build_context
*bld
,
2863 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2864 const struct lp_type type
= bld
->type
;
2866 assert(lp_check_value(type
, a
));
2868 if (lp_build_fast_rsqrt_available(type
)) {
2869 const char *intrinsic
= NULL
;
2871 if (type
.length
== 4) {
2872 intrinsic
= "llvm.x86.sse.rsqrt.ps";
2875 intrinsic
= "llvm.x86.avx.rsqrt.ps.256";
2877 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2880 debug_printf("%s: emulating fast rsqrt with rcp/sqrt\n", __FUNCTION__
);
2882 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2887 * Generate sin(a) or cos(a) using polynomial approximation.
2888 * TODO: it might be worth recognizing sin and cos using same source
2889 * (i.e. d3d10 sincos opcode). Obviously doing both at the same time
2890 * would be way cheaper than calculating (nearly) everything twice...
2891 * Not sure it's common enough to be worth bothering however, scs
2892 * opcode could also benefit from calculating both though.
2895 lp_build_sin_or_cos(struct lp_build_context
*bld
,
2899 struct gallivm_state
*gallivm
= bld
->gallivm
;
2900 LLVMBuilderRef b
= gallivm
->builder
;
2901 struct lp_type int_type
= lp_int_type(bld
->type
);
2904 * take the absolute value,
2905 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
2908 LLVMValueRef inv_sig_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, ~0x80000000);
2909 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, bld
->int_vec_type
, "a_v4si");
2911 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
2912 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, bld
->vec_type
, "x_abs");
2916 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
2919 LLVMValueRef FOPi
= lp_build_const_vec(gallivm
, bld
->type
, 1.27323954473516);
2920 LLVMValueRef scale_y
= LLVMBuildFMul(b
, x_abs
, FOPi
, "scale_y");
2923 * store the integer part of y in mm0
2924 * emm2 = _mm_cvttps_epi32(y);
2927 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, bld
->int_vec_type
, "emm2_i");
2930 * j=(j+1) & (~1) (see the cephes sources)
2931 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
2934 LLVMValueRef all_one
= lp_build_const_int_vec(gallivm
, bld
->type
, 1);
2935 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
2937 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
2939 LLVMValueRef inv_one
= lp_build_const_int_vec(gallivm
, bld
->type
, ~1);
2940 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
2943 * y = _mm_cvtepi32_ps(emm2);
2945 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, bld
->vec_type
, "y_2");
2947 LLVMValueRef const_2
= lp_build_const_int_vec(gallivm
, bld
->type
, 2);
2948 LLVMValueRef const_4
= lp_build_const_int_vec(gallivm
, bld
->type
, 4);
2949 LLVMValueRef const_29
= lp_build_const_int_vec(gallivm
, bld
->type
, 29);
2950 LLVMValueRef sign_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, 0x80000000);
2953 * Argument used for poly selection and sign bit determination
2954 * is different for sin vs. cos.
2956 LLVMValueRef emm2_2
= cos
? LLVMBuildSub(b
, emm2_and
, const_2
, "emm2_2") :
2959 LLVMValueRef sign_bit
= cos
? LLVMBuildShl(b
, LLVMBuildAnd(b
, const_4
,
2960 LLVMBuildNot(b
, emm2_2
, ""), ""),
2961 const_29
, "sign_bit") :
2962 LLVMBuildAnd(b
, LLVMBuildXor(b
, a_v4si
,
2963 LLVMBuildShl(b
, emm2_add
,
2965 sign_mask
, "sign_bit");
2968 * get the polynom selection mask
2969 * there is one polynom for 0 <= x <= Pi/4
2970 * and another one for Pi/4<x<=Pi/2
2971 * Both branches will be computed.
2973 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
2974 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
2977 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_2
, const_2
, "emm2_3");
2978 LLVMValueRef poly_mask
= lp_build_compare(gallivm
,
2979 int_type
, PIPE_FUNC_EQUAL
,
2980 emm2_3
, lp_build_const_int_vec(gallivm
, bld
->type
, 0));
2983 * _PS_CONST(minus_cephes_DP1, -0.78515625);
2984 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
2985 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
2987 LLVMValueRef DP1
= lp_build_const_vec(gallivm
, bld
->type
, -0.78515625);
2988 LLVMValueRef DP2
= lp_build_const_vec(gallivm
, bld
->type
, -2.4187564849853515625e-4);
2989 LLVMValueRef DP3
= lp_build_const_vec(gallivm
, bld
->type
, -3.77489497744594108e-8);
2992 * The magic pass: "Extended precision modular arithmetic"
2993 * x = ((x - y * DP1) - y * DP2) - y * DP3;
2995 LLVMValueRef x_1
= lp_build_fmuladd(b
, y_2
, DP1
, x_abs
);
2996 LLVMValueRef x_2
= lp_build_fmuladd(b
, y_2
, DP2
, x_1
);
2997 LLVMValueRef x_3
= lp_build_fmuladd(b
, y_2
, DP3
, x_2
);
3000 * Evaluate the first polynom (0 <= x <= Pi/4)
3002 * z = _mm_mul_ps(x,x);
3004 LLVMValueRef z
= LLVMBuildFMul(b
, x_3
, x_3
, "z");
3007 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
3008 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
3009 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
3011 LLVMValueRef coscof_p0
= lp_build_const_vec(gallivm
, bld
->type
, 2.443315711809948E-005);
3012 LLVMValueRef coscof_p1
= lp_build_const_vec(gallivm
, bld
->type
, -1.388731625493765E-003);
3013 LLVMValueRef coscof_p2
= lp_build_const_vec(gallivm
, bld
->type
, 4.166664568298827E-002);
3016 * y = *(v4sf*)_ps_coscof_p0;
3017 * y = _mm_mul_ps(y, z);
3019 LLVMValueRef y_4
= lp_build_fmuladd(b
, z
, coscof_p0
, coscof_p1
);
3020 LLVMValueRef y_6
= lp_build_fmuladd(b
, y_4
, z
, coscof_p2
);
3021 LLVMValueRef y_7
= LLVMBuildFMul(b
, y_6
, z
, "y_7");
3022 LLVMValueRef y_8
= LLVMBuildFMul(b
, y_7
, z
, "y_8");
3026 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
3027 * y = _mm_sub_ps(y, tmp);
3028 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
3030 LLVMValueRef half
= lp_build_const_vec(gallivm
, bld
->type
, 0.5);
3031 LLVMValueRef tmp
= LLVMBuildFMul(b
, z
, half
, "tmp");
3032 LLVMValueRef y_9
= LLVMBuildFSub(b
, y_8
, tmp
, "y_8");
3033 LLVMValueRef one
= lp_build_const_vec(gallivm
, bld
->type
, 1.0);
3034 LLVMValueRef y_10
= LLVMBuildFAdd(b
, y_9
, one
, "y_9");
3037 * _PS_CONST(sincof_p0, -1.9515295891E-4);
3038 * _PS_CONST(sincof_p1, 8.3321608736E-3);
3039 * _PS_CONST(sincof_p2, -1.6666654611E-1);
3041 LLVMValueRef sincof_p0
= lp_build_const_vec(gallivm
, bld
->type
, -1.9515295891E-4);
3042 LLVMValueRef sincof_p1
= lp_build_const_vec(gallivm
, bld
->type
, 8.3321608736E-3);
3043 LLVMValueRef sincof_p2
= lp_build_const_vec(gallivm
, bld
->type
, -1.6666654611E-1);
3046 * Evaluate the second polynom (Pi/4 <= x <= 0)
3048 * y2 = *(v4sf*)_ps_sincof_p0;
3049 * y2 = _mm_mul_ps(y2, z);
3050 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
3051 * y2 = _mm_mul_ps(y2, z);
3052 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
3053 * y2 = _mm_mul_ps(y2, z);
3054 * y2 = _mm_mul_ps(y2, x);
3055 * y2 = _mm_add_ps(y2, x);
3058 LLVMValueRef y2_4
= lp_build_fmuladd(b
, z
, sincof_p0
, sincof_p1
);
3059 LLVMValueRef y2_6
= lp_build_fmuladd(b
, y2_4
, z
, sincof_p2
);
3060 LLVMValueRef y2_7
= LLVMBuildFMul(b
, y2_6
, z
, "y2_7");
3061 LLVMValueRef y2_9
= lp_build_fmuladd(b
, y2_7
, x_3
, x_3
);
3064 * select the correct result from the two polynoms
3066 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
3067 * y = _mm_andnot_ps(xmm3, y);
3068 * y = _mm_or_ps(y,y2);
3070 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, bld
->int_vec_type
, "y2_i");
3071 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, bld
->int_vec_type
, "y_i");
3072 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
3073 LLVMValueRef poly_mask_inv
= LLVMBuildNot(b
, poly_mask
, "poly_mask_inv");
3074 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
3075 LLVMValueRef y_combine
= LLVMBuildOr(b
, y_and
, y2_and
, "y_combine");
3079 * y = _mm_xor_ps(y, sign_bit);
3081 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit
, "y_sign");
3082 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, bld
->vec_type
, "y_result");
3084 LLVMValueRef isfinite
= lp_build_isfinite(bld
, a
);
3086 /* clamp output to be within [-1, 1] */
3087 y_result
= lp_build_clamp(bld
, y_result
,
3088 lp_build_const_vec(bld
->gallivm
, bld
->type
, -1.f
),
3089 lp_build_const_vec(bld
->gallivm
, bld
->type
, 1.f
));
3090 /* If a is -inf, inf or NaN then return NaN */
3091 y_result
= lp_build_select(bld
, isfinite
, y_result
,
3092 lp_build_const_vec(bld
->gallivm
, bld
->type
, NAN
));
3101 lp_build_sin(struct lp_build_context
*bld
,
3104 return lp_build_sin_or_cos(bld
, a
, FALSE
);
3112 lp_build_cos(struct lp_build_context
*bld
,
3115 return lp_build_sin_or_cos(bld
, a
, TRUE
);
3120 * Generate pow(x, y)
3123 lp_build_pow(struct lp_build_context
*bld
,
3127 /* TODO: optimize the constant case */
3128 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3129 LLVMIsConstant(x
) && LLVMIsConstant(y
)) {
3130 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3134 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
3142 lp_build_exp(struct lp_build_context
*bld
,
3145 /* log2(e) = 1/log(2) */
3146 LLVMValueRef log2e
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3147 1.4426950408889634);
3149 assert(lp_check_value(bld
->type
, x
));
3151 return lp_build_exp2(bld
, lp_build_mul(bld
, log2e
, x
));
3157 * Behavior is undefined with infs, 0s and nans
3160 lp_build_log(struct lp_build_context
*bld
,
3164 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3165 0.69314718055994529);
3167 assert(lp_check_value(bld
->type
, x
));
3169 return lp_build_mul(bld
, log2
, lp_build_log2(bld
, x
));
3173 * Generate log(x) that handles edge cases (infs, 0s and nans)
3176 lp_build_log_safe(struct lp_build_context
*bld
,
3180 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3181 0.69314718055994529);
3183 assert(lp_check_value(bld
->type
, x
));
3185 return lp_build_mul(bld
, log2
, lp_build_log2_safe(bld
, x
));
3190 * Generate polynomial.
3191 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
3194 lp_build_polynomial(struct lp_build_context
*bld
,
3196 const double *coeffs
,
3197 unsigned num_coeffs
)
3199 const struct lp_type type
= bld
->type
;
3200 LLVMValueRef even
= NULL
, odd
= NULL
;
3204 assert(lp_check_value(bld
->type
, x
));
3206 /* TODO: optimize the constant case */
3207 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3208 LLVMIsConstant(x
)) {
3209 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3214 * Calculate odd and even terms seperately to decrease data dependency
3216 * c[0] + x^2 * c[2] + x^4 * c[4] ...
3217 * + x * (c[1] + x^2 * c[3] + x^4 * c[5]) ...
3219 x2
= lp_build_mul(bld
, x
, x
);
3221 for (i
= num_coeffs
; i
--; ) {
3224 coeff
= lp_build_const_vec(bld
->gallivm
, type
, coeffs
[i
]);
3228 even
= lp_build_mad(bld
, x2
, even
, coeff
);
3233 odd
= lp_build_mad(bld
, x2
, odd
, coeff
);
3240 return lp_build_mad(bld
, odd
, x
, even
);
3249 * Minimax polynomial fit of 2**x, in range [0, 1[
3251 const double lp_build_exp2_polynomial
[] = {
3252 #if EXP_POLY_DEGREE == 5
3253 1.000000000000000000000, /*XXX: was 0.999999925063526176901, recompute others */
3254 0.693153073200168932794,
3255 0.240153617044375388211,
3256 0.0558263180532956664775,
3257 0.00898934009049466391101,
3258 0.00187757667519147912699
3259 #elif EXP_POLY_DEGREE == 4
3260 1.00000259337069434683,
3261 0.693003834469974940458,
3262 0.24144275689150793076,
3263 0.0520114606103070150235,
3264 0.0135341679161270268764
3265 #elif EXP_POLY_DEGREE == 3
3266 0.999925218562710312959,
3267 0.695833540494823811697,
3268 0.226067155427249155588,
3269 0.0780245226406372992967
3270 #elif EXP_POLY_DEGREE == 2
3271 1.00172476321474503578,
3272 0.657636275736077639316,
3273 0.33718943461968720704
3281 lp_build_exp2(struct lp_build_context
*bld
,
3284 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3285 const struct lp_type type
= bld
->type
;
3286 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3287 LLVMValueRef ipart
= NULL
;
3288 LLVMValueRef fpart
= NULL
;
3289 LLVMValueRef expipart
= NULL
;
3290 LLVMValueRef expfpart
= NULL
;
3291 LLVMValueRef res
= NULL
;
3293 assert(lp_check_value(bld
->type
, x
));
3295 /* TODO: optimize the constant case */
3296 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3297 LLVMIsConstant(x
)) {
3298 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3302 assert(type
.floating
&& type
.width
== 32);
3304 /* We want to preserve NaN and make sure than for exp2 if x > 128,
3305 * the result is INF and if it's smaller than -126.9 the result is 0 */
3306 x
= lp_build_min_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, 128.0), x
,
3307 GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3308 x
= lp_build_max_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, -126.99999),
3309 x
, GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3311 /* ipart = floor(x) */
3312 /* fpart = x - ipart */
3313 lp_build_ifloor_fract(bld
, x
, &ipart
, &fpart
);
3315 /* expipart = (float) (1 << ipart) */
3316 expipart
= LLVMBuildAdd(builder
, ipart
,
3317 lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3318 expipart
= LLVMBuildShl(builder
, expipart
,
3319 lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3320 expipart
= LLVMBuildBitCast(builder
, expipart
, vec_type
, "");
3322 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
3323 ARRAY_SIZE(lp_build_exp2_polynomial
));
3325 res
= LLVMBuildFMul(builder
, expipart
, expfpart
, "");
3333 * Extract the exponent of a IEEE-754 floating point value.
3335 * Optionally apply an integer bias.
3337 * Result is an integer value with
3339 * ifloor(log2(x)) + bias
3342 lp_build_extract_exponent(struct lp_build_context
*bld
,
3346 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3347 const struct lp_type type
= bld
->type
;
3348 unsigned mantissa
= lp_mantissa(type
);
3351 assert(type
.floating
);
3353 assert(lp_check_value(bld
->type
, x
));
3355 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3357 res
= LLVMBuildLShr(builder
, x
,
3358 lp_build_const_int_vec(bld
->gallivm
, type
, mantissa
), "");
3359 res
= LLVMBuildAnd(builder
, res
,
3360 lp_build_const_int_vec(bld
->gallivm
, type
, 255), "");
3361 res
= LLVMBuildSub(builder
, res
,
3362 lp_build_const_int_vec(bld
->gallivm
, type
, 127 - bias
), "");
3369 * Extract the mantissa of the a floating.
3371 * Result is a floating point value with
3373 * x / floor(log2(x))
3376 lp_build_extract_mantissa(struct lp_build_context
*bld
,
3379 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3380 const struct lp_type type
= bld
->type
;
3381 unsigned mantissa
= lp_mantissa(type
);
3382 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
,
3383 (1ULL << mantissa
) - 1);
3384 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, bld
->int_vec_type
);
3387 assert(lp_check_value(bld
->type
, x
));
3389 assert(type
.floating
);
3391 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3393 /* res = x / 2**ipart */
3394 res
= LLVMBuildAnd(builder
, x
, mantmask
, "");
3395 res
= LLVMBuildOr(builder
, res
, one
, "");
3396 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
3404 * Minimax polynomial fit of log2((1.0 + sqrt(x))/(1.0 - sqrt(x)))/sqrt(x) ,for x in range of [0, 1/9[
3405 * These coefficients can be generate with
3406 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
3408 const double lp_build_log2_polynomial
[] = {
3409 #if LOG_POLY_DEGREE == 5
3410 2.88539008148777786488L,
3411 0.961796878841293367824L,
3412 0.577058946784739859012L,
3413 0.412914355135828735411L,
3414 0.308591899232910175289L,
3415 0.352376952300281371868L,
3416 #elif LOG_POLY_DEGREE == 4
3417 2.88539009343309178325L,
3418 0.961791550404184197881L,
3419 0.577440339438736392009L,
3420 0.403343858251329912514L,
3421 0.406718052498846252698L,
3422 #elif LOG_POLY_DEGREE == 3
3423 2.88538959748872753838L,
3424 0.961932915889597772928L,
3425 0.571118517972136195241L,
3426 0.493997535084709500285L,
3433 * See http://www.devmaster.net/forums/showthread.php?p=43580
3434 * http://en.wikipedia.org/wiki/Logarithm#Calculation
3435 * http://www.nezumi.demon.co.uk/consult/logx.htm
3437 * If handle_edge_cases is true the function will perform computations
3438 * to match the required D3D10+ behavior for each of the edge cases.
3439 * That means that if input is:
3440 * - less than zero (to and including -inf) then NaN will be returned
3441 * - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
3442 * - +infinity, then +infinity will be returned
3443 * - NaN, then NaN will be returned
3445 * Those checks are fairly expensive so if you don't need them make sure
3446 * handle_edge_cases is false.
3449 lp_build_log2_approx(struct lp_build_context
*bld
,
3451 LLVMValueRef
*p_exp
,
3452 LLVMValueRef
*p_floor_log2
,
3453 LLVMValueRef
*p_log2
,
3454 boolean handle_edge_cases
)
3456 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3457 const struct lp_type type
= bld
->type
;
3458 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3459 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
3461 LLVMValueRef expmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x7f800000);
3462 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x007fffff);
3463 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
3465 LLVMValueRef i
= NULL
;
3466 LLVMValueRef y
= NULL
;
3467 LLVMValueRef z
= NULL
;
3468 LLVMValueRef exp
= NULL
;
3469 LLVMValueRef mant
= NULL
;
3470 LLVMValueRef logexp
= NULL
;
3471 LLVMValueRef p_z
= NULL
;
3472 LLVMValueRef res
= NULL
;
3474 assert(lp_check_value(bld
->type
, x
));
3476 if(p_exp
|| p_floor_log2
|| p_log2
) {
3477 /* TODO: optimize the constant case */
3478 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3479 LLVMIsConstant(x
)) {
3480 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3484 assert(type
.floating
&& type
.width
== 32);
3487 * We don't explicitly handle denormalized numbers. They will yield a
3488 * result in the neighbourhood of -127, which appears to be adequate
3492 i
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3494 /* exp = (float) exponent(x) */
3495 exp
= LLVMBuildAnd(builder
, i
, expmask
, "");
3498 if(p_floor_log2
|| p_log2
) {
3499 logexp
= LLVMBuildLShr(builder
, exp
, lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3500 logexp
= LLVMBuildSub(builder
, logexp
, lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3501 logexp
= LLVMBuildSIToFP(builder
, logexp
, vec_type
, "");
3505 /* mant = 1 + (float) mantissa(x) */
3506 mant
= LLVMBuildAnd(builder
, i
, mantmask
, "");
3507 mant
= LLVMBuildOr(builder
, mant
, one
, "");
3508 mant
= LLVMBuildBitCast(builder
, mant
, vec_type
, "");
3510 /* y = (mant - 1) / (mant + 1) */
3511 y
= lp_build_div(bld
,
3512 lp_build_sub(bld
, mant
, bld
->one
),
3513 lp_build_add(bld
, mant
, bld
->one
)
3517 z
= lp_build_mul(bld
, y
, y
);
3520 p_z
= lp_build_polynomial(bld
, z
, lp_build_log2_polynomial
,
3521 ARRAY_SIZE(lp_build_log2_polynomial
));
3523 /* y * P(z) + logexp */
3524 res
= lp_build_mad(bld
, y
, p_z
, logexp
);
3526 if (type
.floating
&& handle_edge_cases
) {
3527 LLVMValueRef negmask
, infmask
, zmask
;
3528 negmask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, x
,
3529 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3530 zmask
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, x
,
3531 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3532 infmask
= lp_build_cmp(bld
, PIPE_FUNC_GEQUAL
, x
,
3533 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
));
3535 /* If x is qual to inf make sure we return inf */
3536 res
= lp_build_select(bld
, infmask
,
3537 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
),
3539 /* If x is qual to 0, return -inf */
3540 res
= lp_build_select(bld
, zmask
,
3541 lp_build_const_vec(bld
->gallivm
, type
, -INFINITY
),
3543 /* If x is nan or less than 0, return nan */
3544 res
= lp_build_select(bld
, negmask
,
3545 lp_build_const_vec(bld
->gallivm
, type
, NAN
),
3551 exp
= LLVMBuildBitCast(builder
, exp
, vec_type
, "");
3556 *p_floor_log2
= logexp
;
3564 * log2 implementation which doesn't have special code to
3565 * handle edge cases (-inf, 0, inf, NaN). It's faster but
3566 * the results for those cases are undefined.
3569 lp_build_log2(struct lp_build_context
*bld
,
3573 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, FALSE
);
3578 * Version of log2 which handles all edge cases.
3579 * Look at documentation of lp_build_log2_approx for
3580 * description of the behavior for each of the edge cases.
3583 lp_build_log2_safe(struct lp_build_context
*bld
,
3587 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, TRUE
);
3593 * Faster (and less accurate) log2.
3595 * log2(x) = floor(log2(x)) - 1 + x / 2**floor(log2(x))
3597 * Piece-wise linear approximation, with exact results when x is a
3600 * See http://www.flipcode.com/archives/Fast_log_Function.shtml
3603 lp_build_fast_log2(struct lp_build_context
*bld
,
3606 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3610 assert(lp_check_value(bld
->type
, x
));
3612 assert(bld
->type
.floating
);
3614 /* ipart = floor(log2(x)) - 1 */
3615 ipart
= lp_build_extract_exponent(bld
, x
, -1);
3616 ipart
= LLVMBuildSIToFP(builder
, ipart
, bld
->vec_type
, "");
3618 /* fpart = x / 2**ipart */
3619 fpart
= lp_build_extract_mantissa(bld
, x
);
3622 return LLVMBuildFAdd(builder
, ipart
, fpart
, "");
3627 * Fast implementation of iround(log2(x)).
3629 * Not an approximation -- it should give accurate results all the time.
3632 lp_build_ilog2(struct lp_build_context
*bld
,
3635 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3636 LLVMValueRef sqrt2
= lp_build_const_vec(bld
->gallivm
, bld
->type
, M_SQRT2
);
3639 assert(bld
->type
.floating
);
3641 assert(lp_check_value(bld
->type
, x
));
3643 /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
3644 x
= LLVMBuildFMul(builder
, x
, sqrt2
, "");
3646 /* ipart = floor(log2(x) + 0.5) */
3647 ipart
= lp_build_extract_exponent(bld
, x
, 0);
3653 lp_build_mod(struct lp_build_context
*bld
,
3657 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3659 const struct lp_type type
= bld
->type
;
3661 assert(lp_check_value(type
, x
));
3662 assert(lp_check_value(type
, y
));
3665 res
= LLVMBuildFRem(builder
, x
, y
, "");
3667 res
= LLVMBuildSRem(builder
, x
, y
, "");
3669 res
= LLVMBuildURem(builder
, x
, y
, "");
3675 * For floating inputs it creates and returns a mask
3676 * which is all 1's for channels which are NaN.
3677 * Channels inside x which are not NaN will be 0.
3680 lp_build_isnan(struct lp_build_context
*bld
,
3684 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3686 assert(bld
->type
.floating
);
3687 assert(lp_check_value(bld
->type
, x
));
3689 mask
= LLVMBuildFCmp(bld
->gallivm
->builder
, LLVMRealOEQ
, x
, x
,
3691 mask
= LLVMBuildNot(bld
->gallivm
->builder
, mask
, "");
3692 mask
= LLVMBuildSExt(bld
->gallivm
->builder
, mask
, int_vec_type
, "isnan");
3696 /* Returns all 1's for floating point numbers that are
3697 * finite numbers and returns all zeros for -inf,
3700 lp_build_isfinite(struct lp_build_context
*bld
,
3703 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3704 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3705 struct lp_type int_type
= lp_int_type(bld
->type
);
3706 LLVMValueRef intx
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3707 LLVMValueRef infornan32
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
3710 if (!bld
->type
.floating
) {
3711 return lp_build_const_int_vec(bld
->gallivm
, bld
->type
, 0);
3713 assert(bld
->type
.floating
);
3714 assert(lp_check_value(bld
->type
, x
));
3715 assert(bld
->type
.width
== 32);
3717 intx
= LLVMBuildAnd(builder
, intx
, infornan32
, "");
3718 return lp_build_compare(bld
->gallivm
, int_type
, PIPE_FUNC_NOTEQUAL
,
3723 * Returns true if the number is nan or inf and false otherwise.
3724 * The input has to be a floating point vector.
3727 lp_build_is_inf_or_nan(struct gallivm_state
*gallivm
,
3728 const struct lp_type type
,
3731 LLVMBuilderRef builder
= gallivm
->builder
;
3732 struct lp_type int_type
= lp_int_type(type
);
3733 LLVMValueRef const0
= lp_build_const_int_vec(gallivm
, int_type
,
3737 assert(type
.floating
);
3739 ret
= LLVMBuildBitCast(builder
, x
, lp_build_vec_type(gallivm
, int_type
), "");
3740 ret
= LLVMBuildAnd(builder
, ret
, const0
, "");
3741 ret
= lp_build_compare(gallivm
, int_type
, PIPE_FUNC_EQUAL
,
3749 lp_build_fpstate_get(struct gallivm_state
*gallivm
)
3751 if (util_cpu_caps
.has_sse
) {
3752 LLVMBuilderRef builder
= gallivm
->builder
;
3753 LLVMValueRef mxcsr_ptr
= lp_build_alloca(
3755 LLVMInt32TypeInContext(gallivm
->context
),
3757 LLVMValueRef mxcsr_ptr8
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3758 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3759 lp_build_intrinsic(builder
,
3760 "llvm.x86.sse.stmxcsr",
3761 LLVMVoidTypeInContext(gallivm
->context
),
3769 lp_build_fpstate_set_denorms_zero(struct gallivm_state
*gallivm
,
3772 if (util_cpu_caps
.has_sse
) {
3773 /* turn on DAZ (64) | FTZ (32768) = 32832 if available */
3774 int daz_ftz
= _MM_FLUSH_ZERO_MASK
;
3776 LLVMBuilderRef builder
= gallivm
->builder
;
3777 LLVMValueRef mxcsr_ptr
= lp_build_fpstate_get(gallivm
);
3778 LLVMValueRef mxcsr
=
3779 LLVMBuildLoad(builder
, mxcsr_ptr
, "mxcsr");
3781 if (util_cpu_caps
.has_daz
) {
3782 /* Enable denormals are zero mode */
3783 daz_ftz
|= _MM_DENORMALS_ZERO_MASK
;
3786 mxcsr
= LLVMBuildOr(builder
, mxcsr
,
3787 LLVMConstInt(LLVMTypeOf(mxcsr
), daz_ftz
, 0), "");
3789 mxcsr
= LLVMBuildAnd(builder
, mxcsr
,
3790 LLVMConstInt(LLVMTypeOf(mxcsr
), ~daz_ftz
, 0), "");
3793 LLVMBuildStore(builder
, mxcsr
, mxcsr_ptr
);
3794 lp_build_fpstate_set(gallivm
, mxcsr_ptr
);
3799 lp_build_fpstate_set(struct gallivm_state
*gallivm
,
3800 LLVMValueRef mxcsr_ptr
)
3802 if (util_cpu_caps
.has_sse
) {
3803 LLVMBuilderRef builder
= gallivm
->builder
;
3804 mxcsr_ptr
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3805 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3806 lp_build_intrinsic(builder
,
3807 "llvm.x86.sse.ldmxcsr",
3808 LLVMVoidTypeInContext(gallivm
->context
),