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 (!type
.sign
&& (a
== bld
->one
|| b
== bld
->one
))
557 if (!type
.floating
&& !type
.fixed
) {
558 if (HAVE_LLVM
>= 0x0900) {
560 intrinsic
= type
.sign
? "llvm.sadd.sat" : "llvm.uadd.sat";
561 lp_format_intrinsic(intrin
, sizeof intrin
, intrinsic
, bld
->vec_type
);
562 return lp_build_intrinsic_binary(builder
, intrin
, bld
->vec_type
, a
, b
);
564 if (type
.width
* type
.length
== 128) {
565 if (util_cpu_caps
.has_sse2
) {
567 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.b" :
568 HAVE_LLVM
< 0x0800 ? "llvm.x86.sse2.paddus.b" : NULL
;
569 if (type
.width
== 16)
570 intrinsic
= type
.sign
? "llvm.x86.sse2.padds.w" :
571 HAVE_LLVM
< 0x0800 ? "llvm.x86.sse2.paddus.w" : NULL
;
572 } else if (util_cpu_caps
.has_altivec
) {
574 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddsbs" : "llvm.ppc.altivec.vaddubs";
575 if (type
.width
== 16)
576 intrinsic
= type
.sign
? "llvm.ppc.altivec.vaddshs" : "llvm.ppc.altivec.vadduhs";
579 if (type
.width
* type
.length
== 256) {
580 if (util_cpu_caps
.has_avx2
) {
582 intrinsic
= type
.sign
? "llvm.x86.avx2.padds.b" :
583 HAVE_LLVM
< 0x0800 ? "llvm.x86.avx2.paddus.b" : NULL
;
584 if (type
.width
== 16)
585 intrinsic
= type
.sign
? "llvm.x86.avx2.padds.w" :
586 HAVE_LLVM
< 0x0800 ? "llvm.x86.avx2.paddus.w" : NULL
;
592 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
595 if(type
.norm
&& !type
.floating
&& !type
.fixed
) {
597 uint64_t sign
= (uint64_t)1 << (type
.width
- 1);
598 LLVMValueRef max_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
- 1);
599 LLVMValueRef min_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
);
600 /* a_clamp_max is the maximum a for positive b,
601 a_clamp_min is the minimum a for negative b. */
602 LLVMValueRef a_clamp_max
= lp_build_min_simple(bld
, a
, LLVMBuildSub(builder
, max_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
603 LLVMValueRef a_clamp_min
= lp_build_max_simple(bld
, a
, LLVMBuildSub(builder
, min_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
604 a
= lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, b
, bld
->zero
), a_clamp_max
, a_clamp_min
);
608 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
610 res
= LLVMConstFAdd(a
, b
);
612 res
= LLVMConstAdd(a
, b
);
615 res
= LLVMBuildFAdd(builder
, a
, b
, "");
617 res
= LLVMBuildAdd(builder
, a
, b
, "");
619 /* clamp to ceiling of 1.0 */
620 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
621 res
= lp_build_min_simple(bld
, res
, bld
->one
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
623 if (type
.norm
&& !type
.floating
&& !type
.fixed
) {
626 * newer llvm versions no longer support the intrinsics, but recognize
627 * the pattern. Since auto-upgrade of intrinsics doesn't work for jit
628 * code, it is important we match the pattern llvm uses (and pray llvm
629 * doesn't change it - and hope they decide on the same pattern for
630 * all backends supporting it...).
631 * NOTE: cmp/select does sext/trunc of the mask. Does not seem to
632 * interfere with llvm's ability to recognize the pattern but seems
634 * NOTE: llvm 9+ always uses (non arch specific) intrinsic.
636 LLVMValueRef overflowed
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, res
);
637 res
= lp_build_select(bld
, overflowed
,
638 LLVMConstAllOnes(bld
->int_vec_type
), res
);
642 /* XXX clamp to floor of -1 or 0??? */
648 /** Return the scalar sum of the elements of a.
649 * Should avoid this operation whenever possible.
652 lp_build_horizontal_add(struct lp_build_context
*bld
,
655 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
656 const struct lp_type type
= bld
->type
;
657 LLVMValueRef index
, res
;
659 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
/ 2];
660 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
/ 2];
661 LLVMValueRef vecres
, elem2
;
663 assert(lp_check_value(type
, a
));
665 if (type
.length
== 1) {
669 assert(!bld
->type
.norm
);
672 * for byte vectors can do much better with psadbw.
673 * Using repeated shuffle/adds here. Note with multiple vectors
674 * this can be done more efficiently as outlined in the intel
675 * optimization manual.
676 * Note: could cause data rearrangement if used with smaller element
681 length
= type
.length
/ 2;
683 LLVMValueRef vec1
, vec2
;
684 for (i
= 0; i
< length
; i
++) {
685 shuffles1
[i
] = lp_build_const_int32(bld
->gallivm
, i
);
686 shuffles2
[i
] = lp_build_const_int32(bld
->gallivm
, i
+ length
);
688 vec1
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
689 LLVMConstVector(shuffles1
, length
), "");
690 vec2
= LLVMBuildShuffleVector(builder
, vecres
, vecres
,
691 LLVMConstVector(shuffles2
, length
), "");
693 vecres
= LLVMBuildFAdd(builder
, vec1
, vec2
, "");
696 vecres
= LLVMBuildAdd(builder
, vec1
, vec2
, "");
698 length
= length
>> 1;
701 /* always have vector of size 2 here */
704 index
= lp_build_const_int32(bld
->gallivm
, 0);
705 res
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
706 index
= lp_build_const_int32(bld
->gallivm
, 1);
707 elem2
= LLVMBuildExtractElement(builder
, vecres
, index
, "");
710 res
= LLVMBuildFAdd(builder
, res
, elem2
, "");
712 res
= LLVMBuildAdd(builder
, res
, elem2
, "");
718 * Return the horizontal sums of 4 float vectors as a float4 vector.
719 * This uses the technique as outlined in Intel Optimization Manual.
722 lp_build_horizontal_add4x4f(struct lp_build_context
*bld
,
725 struct gallivm_state
*gallivm
= bld
->gallivm
;
726 LLVMBuilderRef builder
= gallivm
->builder
;
727 LLVMValueRef shuffles
[4];
729 LLVMValueRef sumtmp
[2], shuftmp
[2];
731 /* lower half of regs */
732 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
733 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
734 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
735 shuffles
[3] = lp_build_const_int32(gallivm
, 5);
736 tmp
[0] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
737 LLVMConstVector(shuffles
, 4), "");
738 tmp
[2] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
739 LLVMConstVector(shuffles
, 4), "");
741 /* upper half of regs */
742 shuffles
[0] = lp_build_const_int32(gallivm
, 2);
743 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
744 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
745 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
746 tmp
[1] = LLVMBuildShuffleVector(builder
, src
[0], src
[1],
747 LLVMConstVector(shuffles
, 4), "");
748 tmp
[3] = LLVMBuildShuffleVector(builder
, src
[2], src
[3],
749 LLVMConstVector(shuffles
, 4), "");
751 sumtmp
[0] = LLVMBuildFAdd(builder
, tmp
[0], tmp
[1], "");
752 sumtmp
[1] = LLVMBuildFAdd(builder
, tmp
[2], tmp
[3], "");
754 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
755 shuffles
[1] = lp_build_const_int32(gallivm
, 2);
756 shuffles
[2] = lp_build_const_int32(gallivm
, 4);
757 shuffles
[3] = lp_build_const_int32(gallivm
, 6);
758 shuftmp
[0] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
759 LLVMConstVector(shuffles
, 4), "");
761 shuffles
[0] = lp_build_const_int32(gallivm
, 1);
762 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
763 shuffles
[2] = lp_build_const_int32(gallivm
, 5);
764 shuffles
[3] = lp_build_const_int32(gallivm
, 7);
765 shuftmp
[1] = LLVMBuildShuffleVector(builder
, sumtmp
[0], sumtmp
[1],
766 LLVMConstVector(shuffles
, 4), "");
768 return LLVMBuildFAdd(builder
, shuftmp
[0], shuftmp
[1], "");
773 * partially horizontally add 2-4 float vectors with length nx4,
774 * i.e. only four adjacent values in each vector will be added,
775 * assuming values are really grouped in 4 which also determines
778 * Return a vector of the same length as the initial vectors,
779 * with the excess elements (if any) being undefined.
780 * The element order is independent of number of input vectors.
781 * For 3 vectors x0x1x2x3x4x5x6x7, y0y1y2y3y4y5y6y7, z0z1z2z3z4z5z6z7
782 * the output order thus will be
783 * sumx0-x3,sumy0-y3,sumz0-z3,undef,sumx4-x7,sumy4-y7,sumz4z7,undef
786 lp_build_hadd_partial4(struct lp_build_context
*bld
,
787 LLVMValueRef vectors
[],
790 struct gallivm_state
*gallivm
= bld
->gallivm
;
791 LLVMBuilderRef builder
= gallivm
->builder
;
792 LLVMValueRef ret_vec
;
794 const char *intrinsic
= NULL
;
796 assert(num_vecs
>= 2 && num_vecs
<= 4);
797 assert(bld
->type
.floating
);
799 /* only use this with at least 2 vectors, as it is sort of expensive
800 * (depending on cpu) and we always need two horizontal adds anyway,
801 * so a shuffle/add approach might be better.
807 tmp
[2] = num_vecs
> 2 ? vectors
[2] : vectors
[0];
808 tmp
[3] = num_vecs
> 3 ? vectors
[3] : vectors
[0];
810 if (util_cpu_caps
.has_sse3
&& bld
->type
.width
== 32 &&
811 bld
->type
.length
== 4) {
812 intrinsic
= "llvm.x86.sse3.hadd.ps";
814 else if (util_cpu_caps
.has_avx
&& bld
->type
.width
== 32 &&
815 bld
->type
.length
== 8) {
816 intrinsic
= "llvm.x86.avx.hadd.ps.256";
819 tmp
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
820 lp_build_vec_type(gallivm
, bld
->type
),
823 tmp
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
824 lp_build_vec_type(gallivm
, bld
->type
),
830 return lp_build_intrinsic_binary(builder
, intrinsic
,
831 lp_build_vec_type(gallivm
, bld
->type
),
835 if (bld
->type
.length
== 4) {
836 ret_vec
= lp_build_horizontal_add4x4f(bld
, tmp
);
839 LLVMValueRef partres
[LP_MAX_VECTOR_LENGTH
/4];
841 unsigned num_iter
= bld
->type
.length
/ 4;
842 struct lp_type parttype
= bld
->type
;
844 for (j
= 0; j
< num_iter
; j
++) {
845 LLVMValueRef partsrc
[4];
847 for (i
= 0; i
< 4; i
++) {
848 partsrc
[i
] = lp_build_extract_range(gallivm
, tmp
[i
], j
*4, 4);
850 partres
[j
] = lp_build_horizontal_add4x4f(bld
, partsrc
);
852 ret_vec
= lp_build_concat(gallivm
, partres
, parttype
, num_iter
);
861 lp_build_sub(struct lp_build_context
*bld
,
865 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
866 const struct lp_type type
= bld
->type
;
869 assert(lp_check_value(type
, a
));
870 assert(lp_check_value(type
, b
));
874 if (a
== bld
->undef
|| b
== bld
->undef
)
880 const char *intrinsic
= NULL
;
882 if (!type
.sign
&& b
== bld
->one
)
885 if (!type
.floating
&& !type
.fixed
) {
886 if (HAVE_LLVM
>= 0x0900) {
888 intrinsic
= type
.sign
? "llvm.ssub.sat" : "llvm.usub.sat";
889 lp_format_intrinsic(intrin
, sizeof intrin
, intrinsic
, bld
->vec_type
);
890 return lp_build_intrinsic_binary(builder
, intrin
, bld
->vec_type
, a
, b
);
892 if (type
.width
* type
.length
== 128) {
893 if (util_cpu_caps
.has_sse2
) {
895 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.b" :
896 HAVE_LLVM
< 0x0800 ? "llvm.x86.sse2.psubus.b" : NULL
;
897 if (type
.width
== 16)
898 intrinsic
= type
.sign
? "llvm.x86.sse2.psubs.w" :
899 HAVE_LLVM
< 0x0800 ? "llvm.x86.sse2.psubus.w" : NULL
;
900 } else if (util_cpu_caps
.has_altivec
) {
902 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubsbs" : "llvm.ppc.altivec.vsububs";
903 if (type
.width
== 16)
904 intrinsic
= type
.sign
? "llvm.ppc.altivec.vsubshs" : "llvm.ppc.altivec.vsubuhs";
907 if (type
.width
* type
.length
== 256) {
908 if (util_cpu_caps
.has_avx2
) {
910 intrinsic
= type
.sign
? "llvm.x86.avx2.psubs.b" :
911 HAVE_LLVM
< 0x0800 ? "llvm.x86.avx2.psubus.b" : NULL
;
912 if (type
.width
== 16)
913 intrinsic
= type
.sign
? "llvm.x86.avx2.psubs.w" :
914 HAVE_LLVM
< 0x0800 ? "llvm.x86.avx2.psubus.w" : NULL
;
920 return lp_build_intrinsic_binary(builder
, intrinsic
, lp_build_vec_type(bld
->gallivm
, bld
->type
), a
, b
);
923 if(type
.norm
&& !type
.floating
&& !type
.fixed
) {
925 uint64_t sign
= (uint64_t)1 << (type
.width
- 1);
926 LLVMValueRef max_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
- 1);
927 LLVMValueRef min_val
= lp_build_const_int_vec(bld
->gallivm
, type
, sign
);
928 /* a_clamp_max is the maximum a for negative b,
929 a_clamp_min is the minimum a for positive b. */
930 LLVMValueRef a_clamp_max
= lp_build_min_simple(bld
, a
, LLVMBuildAdd(builder
, max_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
931 LLVMValueRef a_clamp_min
= lp_build_max_simple(bld
, a
, LLVMBuildAdd(builder
, min_val
, b
, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
932 a
= lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, b
, bld
->zero
), a_clamp_min
, a_clamp_max
);
935 * This must match llvm pattern for saturated unsigned sub.
936 * (lp_build_max_simple actually does the job with its current
937 * definition but do it explicitly here.)
938 * NOTE: cmp/select does sext/trunc of the mask. Does not seem to
939 * interfere with llvm's ability to recognize the pattern but seems
941 * NOTE: llvm 9+ always uses (non arch specific) intrinsic.
943 LLVMValueRef no_ov
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, b
);
944 a
= lp_build_select(bld
, no_ov
, a
, b
);
948 if(LLVMIsConstant(a
) && LLVMIsConstant(b
))
950 res
= LLVMConstFSub(a
, b
);
952 res
= LLVMConstSub(a
, b
);
955 res
= LLVMBuildFSub(builder
, a
, b
, "");
957 res
= LLVMBuildSub(builder
, a
, b
, "");
959 if(bld
->type
.norm
&& (bld
->type
.floating
|| bld
->type
.fixed
))
960 res
= lp_build_max_simple(bld
, res
, bld
->zero
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
968 * Normalized multiplication.
970 * There are several approaches for (using 8-bit normalized multiplication as
975 * makes the following approximation to the division (Sree)
977 * a*b/255 ~= (a*(b + 1)) >> 256
979 * which is the fastest method that satisfies the following OpenGL criteria of
981 * 0*0 = 0 and 255*255 = 255
985 * takes the geometric series approximation to the division
987 * t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
989 * in this case just the first two terms to fit in 16bit arithmetic
991 * t/255 ~= (t + (t >> 8)) >> 8
993 * note that just by itself it doesn't satisfies the OpenGL criteria, as
994 * 255*255 = 254, so the special case b = 255 must be accounted or roundoff
997 * - geometric series plus rounding
999 * when using a geometric series division instead of truncating the result
1000 * use roundoff in the approximation (Jim Blinn)
1002 * t/255 ~= (t + (t >> 8) + 0x80) >> 8
1004 * achieving the exact results.
1008 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
1009 * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
1010 * @sa Michael Herf, The "double blend trick", May 2000,
1011 * http://www.stereopsis.com/doubleblend.html
1014 lp_build_mul_norm(struct gallivm_state
*gallivm
,
1015 struct lp_type wide_type
,
1016 LLVMValueRef a
, LLVMValueRef b
)
1018 LLVMBuilderRef builder
= gallivm
->builder
;
1019 struct lp_build_context bld
;
1024 assert(!wide_type
.floating
);
1025 assert(lp_check_value(wide_type
, a
));
1026 assert(lp_check_value(wide_type
, b
));
1028 lp_build_context_init(&bld
, gallivm
, wide_type
);
1030 n
= wide_type
.width
/ 2;
1031 if (wide_type
.sign
) {
1036 * TODO: for 16bits normalized SSE2 vectors we could consider using PMULHUW
1037 * http://ssp.impulsetrain.com/2011/07/03/multiplying-normalized-16-bit-numbers-with-sse2/
1041 * a*b / (2**n - 1) ~= (a*b + (a*b >> n) + half) >> n
1044 ab
= LLVMBuildMul(builder
, a
, b
, "");
1045 ab
= LLVMBuildAdd(builder
, ab
, lp_build_shr_imm(&bld
, ab
, n
), "");
1048 * half = sgn(ab) * 0.5 * (2 ** n) = sgn(ab) * (1 << (n - 1))
1051 half
= lp_build_const_int_vec(gallivm
, wide_type
, 1LL << (n
- 1));
1052 if (wide_type
.sign
) {
1053 LLVMValueRef minus_half
= LLVMBuildNeg(builder
, half
, "");
1054 LLVMValueRef sign
= lp_build_shr_imm(&bld
, ab
, wide_type
.width
- 1);
1055 half
= lp_build_select(&bld
, sign
, minus_half
, half
);
1057 ab
= LLVMBuildAdd(builder
, ab
, half
, "");
1059 /* Final division */
1060 ab
= lp_build_shr_imm(&bld
, ab
, n
);
1069 lp_build_mul(struct lp_build_context
*bld
,
1073 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1074 const struct lp_type type
= bld
->type
;
1078 assert(lp_check_value(type
, a
));
1079 assert(lp_check_value(type
, b
));
1089 if(a
== bld
->undef
|| b
== bld
->undef
)
1092 if (!type
.floating
&& !type
.fixed
&& type
.norm
) {
1093 struct lp_type wide_type
= lp_wider_type(type
);
1094 LLVMValueRef al
, ah
, bl
, bh
, abl
, abh
, ab
;
1096 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, a
, &al
, &ah
);
1097 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, b
, &bl
, &bh
);
1099 /* PMULLW, PSRLW, PADDW */
1100 abl
= lp_build_mul_norm(bld
->gallivm
, wide_type
, al
, bl
);
1101 abh
= lp_build_mul_norm(bld
->gallivm
, wide_type
, ah
, bh
);
1103 ab
= lp_build_pack2_native(bld
->gallivm
, wide_type
, type
, abl
, abh
);
1109 shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
/2);
1113 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
1115 res
= LLVMConstFMul(a
, b
);
1117 res
= LLVMConstMul(a
, b
);
1120 res
= LLVMConstAShr(res
, shift
);
1122 res
= LLVMConstLShr(res
, shift
);
1127 res
= LLVMBuildFMul(builder
, a
, b
, "");
1129 res
= LLVMBuildMul(builder
, a
, b
, "");
1132 res
= LLVMBuildAShr(builder
, res
, shift
, "");
1134 res
= LLVMBuildLShr(builder
, res
, shift
, "");
1142 * Widening mul, valid for 32x32 bit -> 64bit only.
1143 * Result is low 32bits, high bits returned in res_hi.
1145 * Emits code that is meant to be compiled for the host CPU.
1148 lp_build_mul_32_lohi_cpu(struct lp_build_context
*bld
,
1151 LLVMValueRef
*res_hi
)
1153 struct gallivm_state
*gallivm
= bld
->gallivm
;
1154 LLVMBuilderRef builder
= gallivm
->builder
;
1156 assert(bld
->type
.width
== 32);
1157 assert(bld
->type
.floating
== 0);
1158 assert(bld
->type
.fixed
== 0);
1159 assert(bld
->type
.norm
== 0);
1162 * XXX: for some reason, with zext/zext/mul/trunc the code llvm produces
1163 * for x86 simd is atrocious (even if the high bits weren't required),
1164 * trying to handle real 64bit inputs (which of course can't happen due
1165 * to using 64bit umul with 32bit numbers zero-extended to 64bit, but
1166 * apparently llvm does not recognize this widening mul). This includes 6
1167 * (instead of 2) pmuludq plus extra adds and shifts
1168 * The same story applies to signed mul, albeit fixing this requires sse41.
1169 * https://llvm.org/bugs/show_bug.cgi?id=30845
1170 * So, whip up our own code, albeit only for length 4 and 8 (which
1171 * should be good enough)...
1172 * FIXME: For llvm >= 7.0 we should match the autoupgrade pattern
1173 * (bitcast/and/mul/shuffle for unsigned, bitcast/shl/ashr/mul/shuffle
1174 * for signed), which the fallback code does not, without this llvm
1175 * will likely still produce atrocious code.
1177 if (HAVE_LLVM
< 0x0700 &&
1178 (bld
->type
.length
== 4 || bld
->type
.length
== 8) &&
1179 ((util_cpu_caps
.has_sse2
&& (bld
->type
.sign
== 0)) ||
1180 util_cpu_caps
.has_sse4_1
)) {
1181 const char *intrinsic
= NULL
;
1182 LLVMValueRef aeven
, aodd
, beven
, bodd
, muleven
, mulodd
;
1183 LLVMValueRef shuf
[LP_MAX_VECTOR_WIDTH
/ 32], shuf_vec
;
1184 struct lp_type type_wide
= lp_wider_type(bld
->type
);
1185 LLVMTypeRef wider_type
= lp_build_vec_type(gallivm
, type_wide
);
1187 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1188 shuf
[i
] = lp_build_const_int32(gallivm
, i
+1);
1189 shuf
[i
+1] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1191 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1194 aodd
= LLVMBuildShuffleVector(builder
, aeven
, bld
->undef
, shuf_vec
, "");
1195 bodd
= LLVMBuildShuffleVector(builder
, beven
, bld
->undef
, shuf_vec
, "");
1197 if (util_cpu_caps
.has_avx2
&& bld
->type
.length
== 8) {
1198 if (bld
->type
.sign
) {
1199 intrinsic
= "llvm.x86.avx2.pmul.dq";
1201 intrinsic
= "llvm.x86.avx2.pmulu.dq";
1203 muleven
= lp_build_intrinsic_binary(builder
, intrinsic
,
1204 wider_type
, aeven
, beven
);
1205 mulodd
= lp_build_intrinsic_binary(builder
, intrinsic
,
1206 wider_type
, aodd
, bodd
);
1209 /* for consistent naming look elsewhere... */
1210 if (bld
->type
.sign
) {
1211 intrinsic
= "llvm.x86.sse41.pmuldq";
1213 intrinsic
= "llvm.x86.sse2.pmulu.dq";
1216 * XXX If we only have AVX but not AVX2 this is a pain.
1217 * lp_build_intrinsic_binary_anylength() can't handle it
1218 * (due to src and dst type not being identical).
1220 if (bld
->type
.length
== 8) {
1221 LLVMValueRef aevenlo
, aevenhi
, bevenlo
, bevenhi
;
1222 LLVMValueRef aoddlo
, aoddhi
, boddlo
, boddhi
;
1223 LLVMValueRef muleven2
[2], mulodd2
[2];
1224 struct lp_type type_wide_half
= type_wide
;
1225 LLVMTypeRef wtype_half
;
1226 type_wide_half
.length
= 2;
1227 wtype_half
= lp_build_vec_type(gallivm
, type_wide_half
);
1228 aevenlo
= lp_build_extract_range(gallivm
, aeven
, 0, 4);
1229 aevenhi
= lp_build_extract_range(gallivm
, aeven
, 4, 4);
1230 bevenlo
= lp_build_extract_range(gallivm
, beven
, 0, 4);
1231 bevenhi
= lp_build_extract_range(gallivm
, beven
, 4, 4);
1232 aoddlo
= lp_build_extract_range(gallivm
, aodd
, 0, 4);
1233 aoddhi
= lp_build_extract_range(gallivm
, aodd
, 4, 4);
1234 boddlo
= lp_build_extract_range(gallivm
, bodd
, 0, 4);
1235 boddhi
= lp_build_extract_range(gallivm
, bodd
, 4, 4);
1236 muleven2
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
1237 wtype_half
, aevenlo
, bevenlo
);
1238 mulodd2
[0] = lp_build_intrinsic_binary(builder
, intrinsic
,
1239 wtype_half
, aoddlo
, boddlo
);
1240 muleven2
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
1241 wtype_half
, aevenhi
, bevenhi
);
1242 mulodd2
[1] = lp_build_intrinsic_binary(builder
, intrinsic
,
1243 wtype_half
, aoddhi
, boddhi
);
1244 muleven
= lp_build_concat(gallivm
, muleven2
, type_wide_half
, 2);
1245 mulodd
= lp_build_concat(gallivm
, mulodd2
, type_wide_half
, 2);
1249 muleven
= lp_build_intrinsic_binary(builder
, intrinsic
,
1250 wider_type
, aeven
, beven
);
1251 mulodd
= lp_build_intrinsic_binary(builder
, intrinsic
,
1252 wider_type
, aodd
, bodd
);
1255 muleven
= LLVMBuildBitCast(builder
, muleven
, bld
->vec_type
, "");
1256 mulodd
= LLVMBuildBitCast(builder
, mulodd
, bld
->vec_type
, "");
1258 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1259 shuf
[i
] = lp_build_const_int32(gallivm
, i
+ 1);
1260 shuf
[i
+1] = lp_build_const_int32(gallivm
, i
+ 1 + bld
->type
.length
);
1262 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1263 *res_hi
= LLVMBuildShuffleVector(builder
, muleven
, mulodd
, shuf_vec
, "");
1265 for (i
= 0; i
< bld
->type
.length
; i
+= 2) {
1266 shuf
[i
] = lp_build_const_int32(gallivm
, i
);
1267 shuf
[i
+1] = lp_build_const_int32(gallivm
, i
+ bld
->type
.length
);
1269 shuf_vec
= LLVMConstVector(shuf
, bld
->type
.length
);
1270 return LLVMBuildShuffleVector(builder
, muleven
, mulodd
, shuf_vec
, "");
1273 return lp_build_mul_32_lohi(bld
, a
, b
, res_hi
);
1279 * Widening mul, valid for 32x32 bit -> 64bit only.
1280 * Result is low 32bits, high bits returned in res_hi.
1282 * Emits generic code.
1285 lp_build_mul_32_lohi(struct lp_build_context
*bld
,
1288 LLVMValueRef
*res_hi
)
1290 struct gallivm_state
*gallivm
= bld
->gallivm
;
1291 LLVMBuilderRef builder
= gallivm
->builder
;
1292 LLVMValueRef tmp
, shift
, res_lo
;
1293 struct lp_type type_tmp
;
1294 LLVMTypeRef wide_type
, narrow_type
;
1296 type_tmp
= bld
->type
;
1297 narrow_type
= lp_build_vec_type(gallivm
, type_tmp
);
1298 type_tmp
.width
*= 2;
1299 wide_type
= lp_build_vec_type(gallivm
, type_tmp
);
1300 shift
= lp_build_const_vec(gallivm
, type_tmp
, 32);
1302 if (bld
->type
.sign
) {
1303 a
= LLVMBuildSExt(builder
, a
, wide_type
, "");
1304 b
= LLVMBuildSExt(builder
, b
, wide_type
, "");
1306 a
= LLVMBuildZExt(builder
, a
, wide_type
, "");
1307 b
= LLVMBuildZExt(builder
, b
, wide_type
, "");
1309 tmp
= LLVMBuildMul(builder
, a
, b
, "");
1311 res_lo
= LLVMBuildTrunc(builder
, tmp
, narrow_type
, "");
1313 /* Since we truncate anyway, LShr and AShr are equivalent. */
1314 tmp
= LLVMBuildLShr(builder
, tmp
, shift
, "");
1315 *res_hi
= LLVMBuildTrunc(builder
, tmp
, narrow_type
, "");
1323 lp_build_mad(struct lp_build_context
*bld
,
1328 const struct lp_type type
= bld
->type
;
1329 if (type
.floating
) {
1330 return lp_build_fmuladd(bld
->gallivm
->builder
, a
, b
, c
);
1332 return lp_build_add(bld
, lp_build_mul(bld
, a
, b
), c
);
1338 * Small vector x scale multiplication optimization.
1341 lp_build_mul_imm(struct lp_build_context
*bld
,
1345 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1346 LLVMValueRef factor
;
1348 assert(lp_check_value(bld
->type
, a
));
1357 return lp_build_negate(bld
, a
);
1359 if(b
== 2 && bld
->type
.floating
)
1360 return lp_build_add(bld
, a
, a
);
1362 if(util_is_power_of_two_or_zero(b
)) {
1363 unsigned shift
= ffs(b
) - 1;
1365 if(bld
->type
.floating
) {
1368 * Power of two multiplication by directly manipulating the exponent.
1370 * XXX: This might not be always faster, it will introduce a small error
1371 * for multiplication by zero, and it will produce wrong results
1374 unsigned mantissa
= lp_mantissa(bld
->type
);
1375 factor
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (unsigned long long)shift
<< mantissa
);
1376 a
= LLVMBuildBitCast(builder
, a
, lp_build_int_vec_type(bld
->type
), "");
1377 a
= LLVMBuildAdd(builder
, a
, factor
, "");
1378 a
= LLVMBuildBitCast(builder
, a
, lp_build_vec_type(bld
->gallivm
, bld
->type
), "");
1383 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, shift
);
1384 return LLVMBuildShl(builder
, a
, factor
, "");
1388 factor
= lp_build_const_vec(bld
->gallivm
, bld
->type
, (double)b
);
1389 return lp_build_mul(bld
, a
, factor
);
1397 lp_build_div(struct lp_build_context
*bld
,
1401 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1402 const struct lp_type type
= bld
->type
;
1404 assert(lp_check_value(type
, a
));
1405 assert(lp_check_value(type
, b
));
1409 if(a
== bld
->one
&& type
.floating
)
1410 return lp_build_rcp(bld
, b
);
1415 if(a
== bld
->undef
|| b
== bld
->undef
)
1418 if(LLVMIsConstant(a
) && LLVMIsConstant(b
)) {
1420 return LLVMConstFDiv(a
, b
);
1422 return LLVMConstSDiv(a
, b
);
1424 return LLVMConstUDiv(a
, b
);
1427 /* fast rcp is disabled (just uses div), so makes no sense to try that */
1429 ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
1430 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) &&
1432 return lp_build_mul(bld
, a
, lp_build_rcp(bld
, b
));
1435 return LLVMBuildFDiv(builder
, a
, b
, "");
1437 return LLVMBuildSDiv(builder
, a
, b
, "");
1439 return LLVMBuildUDiv(builder
, a
, b
, "");
1444 * Linear interpolation helper.
1446 * @param normalized whether we are interpolating normalized values,
1447 * encoded in normalized integers, twice as wide.
1449 * @sa http://www.stereopsis.com/doubleblend.html
1451 static inline LLVMValueRef
1452 lp_build_lerp_simple(struct lp_build_context
*bld
,
1458 unsigned half_width
= bld
->type
.width
/2;
1459 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1463 assert(lp_check_value(bld
->type
, x
));
1464 assert(lp_check_value(bld
->type
, v0
));
1465 assert(lp_check_value(bld
->type
, v1
));
1467 delta
= lp_build_sub(bld
, v1
, v0
);
1469 if (bld
->type
.floating
) {
1471 return lp_build_mad(bld
, x
, delta
, v0
);
1474 if (flags
& LP_BLD_LERP_WIDE_NORMALIZED
) {
1475 if (!bld
->type
.sign
) {
1476 if (!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
)) {
1478 * Scale x from [0, 2**n - 1] to [0, 2**n] by adding the
1479 * most-significant-bit to the lowest-significant-bit, so that
1480 * later we can just divide by 2**n instead of 2**n - 1.
1483 x
= lp_build_add(bld
, x
, lp_build_shr_imm(bld
, x
, half_width
- 1));
1486 /* (x * delta) >> n */
1487 res
= lp_build_mul(bld
, x
, delta
);
1488 res
= lp_build_shr_imm(bld
, res
, half_width
);
1491 * The rescaling trick above doesn't work for signed numbers, so
1492 * use the 2**n - 1 divison approximation in lp_build_mul_norm
1495 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1496 res
= lp_build_mul_norm(bld
->gallivm
, bld
->type
, x
, delta
);
1499 assert(!(flags
& LP_BLD_LERP_PRESCALED_WEIGHTS
));
1500 res
= lp_build_mul(bld
, x
, delta
);
1503 if ((flags
& LP_BLD_LERP_WIDE_NORMALIZED
) && !bld
->type
.sign
) {
1505 * At this point both res and v0 only use the lower half of the bits,
1506 * the rest is zero. Instead of add / mask, do add with half wide type.
1508 struct lp_type narrow_type
;
1509 struct lp_build_context narrow_bld
;
1511 memset(&narrow_type
, 0, sizeof narrow_type
);
1512 narrow_type
.sign
= bld
->type
.sign
;
1513 narrow_type
.width
= bld
->type
.width
/2;
1514 narrow_type
.length
= bld
->type
.length
*2;
1516 lp_build_context_init(&narrow_bld
, bld
->gallivm
, narrow_type
);
1517 res
= LLVMBuildBitCast(builder
, res
, narrow_bld
.vec_type
, "");
1518 v0
= LLVMBuildBitCast(builder
, v0
, narrow_bld
.vec_type
, "");
1519 res
= lp_build_add(&narrow_bld
, v0
, res
);
1520 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
1522 res
= lp_build_add(bld
, v0
, res
);
1524 if (bld
->type
.fixed
) {
1526 * We need to mask out the high order bits when lerping 8bit
1527 * normalized colors stored on 16bits
1529 /* XXX: This step is necessary for lerping 8bit colors stored on
1530 * 16bits, but it will be wrong for true fixed point use cases.
1531 * Basically we need a more powerful lp_type, capable of further
1532 * distinguishing the values interpretation from the value storage.
1534 LLVMValueRef low_bits
;
1535 low_bits
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, (1 << half_width
) - 1);
1536 res
= LLVMBuildAnd(builder
, res
, low_bits
, "");
1545 * Linear interpolation.
1548 lp_build_lerp(struct lp_build_context
*bld
,
1554 const struct lp_type type
= bld
->type
;
1557 assert(lp_check_value(type
, x
));
1558 assert(lp_check_value(type
, v0
));
1559 assert(lp_check_value(type
, v1
));
1561 assert(!(flags
& LP_BLD_LERP_WIDE_NORMALIZED
));
1564 struct lp_type wide_type
;
1565 struct lp_build_context wide_bld
;
1566 LLVMValueRef xl
, xh
, v0l
, v0h
, v1l
, v1h
, resl
, resh
;
1568 assert(type
.length
>= 2);
1571 * Create a wider integer type, enough to hold the
1572 * intermediate result of the multiplication.
1574 memset(&wide_type
, 0, sizeof wide_type
);
1575 wide_type
.sign
= type
.sign
;
1576 wide_type
.width
= type
.width
*2;
1577 wide_type
.length
= type
.length
/2;
1579 lp_build_context_init(&wide_bld
, bld
->gallivm
, wide_type
);
1581 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, x
, &xl
, &xh
);
1582 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v0
, &v0l
, &v0h
);
1583 lp_build_unpack2_native(bld
->gallivm
, type
, wide_type
, v1
, &v1l
, &v1h
);
1589 flags
|= LP_BLD_LERP_WIDE_NORMALIZED
;
1591 resl
= lp_build_lerp_simple(&wide_bld
, xl
, v0l
, v1l
, flags
);
1592 resh
= lp_build_lerp_simple(&wide_bld
, xh
, v0h
, v1h
, flags
);
1594 res
= lp_build_pack2_native(bld
->gallivm
, wide_type
, type
, resl
, resh
);
1596 res
= lp_build_lerp_simple(bld
, x
, v0
, v1
, flags
);
1604 * Bilinear interpolation.
1606 * Values indices are in v_{yx}.
1609 lp_build_lerp_2d(struct lp_build_context
*bld
,
1618 LLVMValueRef v0
= lp_build_lerp(bld
, x
, v00
, v01
, flags
);
1619 LLVMValueRef v1
= lp_build_lerp(bld
, x
, v10
, v11
, flags
);
1620 return lp_build_lerp(bld
, y
, v0
, v1
, flags
);
1625 lp_build_lerp_3d(struct lp_build_context
*bld
,
1639 LLVMValueRef v0
= lp_build_lerp_2d(bld
, x
, y
, v000
, v001
, v010
, v011
, flags
);
1640 LLVMValueRef v1
= lp_build_lerp_2d(bld
, x
, y
, v100
, v101
, v110
, v111
, flags
);
1641 return lp_build_lerp(bld
, z
, v0
, v1
, flags
);
1646 * Generate min(a, b)
1647 * Do checks for special cases but not for nans.
1650 lp_build_min(struct lp_build_context
*bld
,
1654 assert(lp_check_value(bld
->type
, a
));
1655 assert(lp_check_value(bld
->type
, b
));
1657 if(a
== bld
->undef
|| b
== bld
->undef
)
1663 if (bld
->type
.norm
) {
1664 if (!bld
->type
.sign
) {
1665 if (a
== bld
->zero
|| b
== bld
->zero
) {
1675 return lp_build_min_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1680 * Generate min(a, b)
1681 * NaN's are handled according to the behavior specified by the
1682 * nan_behavior argument.
1685 lp_build_min_ext(struct lp_build_context
*bld
,
1688 enum gallivm_nan_behavior nan_behavior
)
1690 assert(lp_check_value(bld
->type
, a
));
1691 assert(lp_check_value(bld
->type
, b
));
1693 if(a
== bld
->undef
|| b
== bld
->undef
)
1699 if (bld
->type
.norm
) {
1700 if (!bld
->type
.sign
) {
1701 if (a
== bld
->zero
|| b
== bld
->zero
) {
1711 return lp_build_min_simple(bld
, a
, b
, nan_behavior
);
1715 * Generate max(a, b)
1716 * Do checks for special cases, but NaN behavior is undefined.
1719 lp_build_max(struct lp_build_context
*bld
,
1723 assert(lp_check_value(bld
->type
, a
));
1724 assert(lp_check_value(bld
->type
, b
));
1726 if(a
== bld
->undef
|| b
== bld
->undef
)
1732 if(bld
->type
.norm
) {
1733 if(a
== bld
->one
|| b
== bld
->one
)
1735 if (!bld
->type
.sign
) {
1736 if (a
== bld
->zero
) {
1739 if (b
== bld
->zero
) {
1745 return lp_build_max_simple(bld
, a
, b
, GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1750 * Generate max(a, b)
1751 * Checks for special cases.
1752 * NaN's are handled according to the behavior specified by the
1753 * nan_behavior argument.
1756 lp_build_max_ext(struct lp_build_context
*bld
,
1759 enum gallivm_nan_behavior nan_behavior
)
1761 assert(lp_check_value(bld
->type
, a
));
1762 assert(lp_check_value(bld
->type
, b
));
1764 if(a
== bld
->undef
|| b
== bld
->undef
)
1770 if(bld
->type
.norm
) {
1771 if(a
== bld
->one
|| b
== bld
->one
)
1773 if (!bld
->type
.sign
) {
1774 if (a
== bld
->zero
) {
1777 if (b
== bld
->zero
) {
1783 return lp_build_max_simple(bld
, a
, b
, nan_behavior
);
1787 * Generate clamp(a, min, max)
1788 * NaN behavior (for any of a, min, max) is undefined.
1789 * Do checks for special cases.
1792 lp_build_clamp(struct lp_build_context
*bld
,
1797 assert(lp_check_value(bld
->type
, a
));
1798 assert(lp_check_value(bld
->type
, min
));
1799 assert(lp_check_value(bld
->type
, max
));
1801 a
= lp_build_min(bld
, a
, max
);
1802 a
= lp_build_max(bld
, a
, min
);
1808 * Generate clamp(a, 0, 1)
1809 * A NaN will get converted to zero.
1812 lp_build_clamp_zero_one_nanzero(struct lp_build_context
*bld
,
1815 a
= lp_build_max_ext(bld
, a
, bld
->zero
, GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1816 a
= lp_build_min(bld
, a
, bld
->one
);
1825 lp_build_abs(struct lp_build_context
*bld
,
1828 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1829 const struct lp_type type
= bld
->type
;
1830 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1832 assert(lp_check_value(type
, a
));
1838 if (0x0306 <= HAVE_LLVM
&& HAVE_LLVM
< 0x0309) {
1839 /* Workaround llvm.org/PR27332 */
1840 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1841 unsigned long long absMask
= ~(1ULL << (type
.width
- 1));
1842 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
, ((unsigned long long) absMask
));
1843 a
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1844 a
= LLVMBuildAnd(builder
, a
, mask
, "");
1845 a
= LLVMBuildBitCast(builder
, a
, vec_type
, "");
1849 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.fabs", vec_type
);
1850 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
1854 if(type
.width
*type
.length
== 128 && util_cpu_caps
.has_ssse3
&& HAVE_LLVM
< 0x0600) {
1855 switch(type
.width
) {
1857 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.b.128", vec_type
, a
);
1859 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.w.128", vec_type
, a
);
1861 return lp_build_intrinsic_unary(builder
, "llvm.x86.ssse3.pabs.d.128", vec_type
, a
);
1864 else if (type
.width
*type
.length
== 256 && util_cpu_caps
.has_avx2
&& HAVE_LLVM
< 0x0600) {
1865 switch(type
.width
) {
1867 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.b", vec_type
, a
);
1869 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.w", vec_type
, a
);
1871 return lp_build_intrinsic_unary(builder
, "llvm.x86.avx2.pabs.d", vec_type
, a
);
1875 return lp_build_select(bld
, lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
),
1876 a
, LLVMBuildNeg(builder
, a
, ""));
1881 lp_build_negate(struct lp_build_context
*bld
,
1884 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1886 assert(lp_check_value(bld
->type
, a
));
1888 if (bld
->type
.floating
)
1889 a
= LLVMBuildFNeg(builder
, a
, "");
1891 a
= LLVMBuildNeg(builder
, a
, "");
1897 /** Return -1, 0 or +1 depending on the sign of a */
1899 lp_build_sgn(struct lp_build_context
*bld
,
1902 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1903 const struct lp_type type
= bld
->type
;
1907 assert(lp_check_value(type
, a
));
1909 /* Handle non-zero case */
1911 /* if not zero then sign must be positive */
1914 else if(type
.floating
) {
1915 LLVMTypeRef vec_type
;
1916 LLVMTypeRef int_type
;
1920 unsigned long long maskBit
= (unsigned long long)1 << (type
.width
- 1);
1922 int_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1923 vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1924 mask
= lp_build_const_int_vec(bld
->gallivm
, type
, maskBit
);
1926 /* Take the sign bit and add it to 1 constant */
1927 sign
= LLVMBuildBitCast(builder
, a
, int_type
, "");
1928 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
1929 one
= LLVMConstBitCast(bld
->one
, int_type
);
1930 res
= LLVMBuildOr(builder
, sign
, one
, "");
1931 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1935 /* signed int/norm/fixed point */
1936 /* could use psign with sse3 and appropriate vectors here */
1937 LLVMValueRef minus_one
= lp_build_const_vec(bld
->gallivm
, type
, -1.0);
1938 cond
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, a
, bld
->zero
);
1939 res
= lp_build_select(bld
, cond
, bld
->one
, minus_one
);
1943 cond
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, a
, bld
->zero
);
1944 res
= lp_build_select(bld
, cond
, bld
->zero
, res
);
1951 * Set the sign of float vector 'a' according to 'sign'.
1952 * If sign==0, return abs(a).
1953 * If sign==1, return -abs(a);
1954 * Other values for sign produce undefined results.
1957 lp_build_set_sign(struct lp_build_context
*bld
,
1958 LLVMValueRef a
, LLVMValueRef sign
)
1960 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1961 const struct lp_type type
= bld
->type
;
1962 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
1963 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1964 LLVMValueRef shift
= lp_build_const_int_vec(bld
->gallivm
, type
, type
.width
- 1);
1965 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
1966 ~((unsigned long long) 1 << (type
.width
- 1)));
1967 LLVMValueRef val
, res
;
1969 assert(type
.floating
);
1970 assert(lp_check_value(type
, a
));
1972 /* val = reinterpret_cast<int>(a) */
1973 val
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
1974 /* val = val & mask */
1975 val
= LLVMBuildAnd(builder
, val
, mask
, "");
1976 /* sign = sign << shift */
1977 sign
= LLVMBuildShl(builder
, sign
, shift
, "");
1978 /* res = val | sign */
1979 res
= LLVMBuildOr(builder
, val
, sign
, "");
1980 /* res = reinterpret_cast<float>(res) */
1981 res
= LLVMBuildBitCast(builder
, res
, vec_type
, "");
1988 * Convert vector of (or scalar) int to vector of (or scalar) float.
1991 lp_build_int_to_float(struct lp_build_context
*bld
,
1994 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1995 const struct lp_type type
= bld
->type
;
1996 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
1998 assert(type
.floating
);
2000 return LLVMBuildSIToFP(builder
, a
, vec_type
, "");
2004 arch_rounding_available(const struct lp_type type
)
2006 if ((util_cpu_caps
.has_sse4_1
&&
2007 (type
.length
== 1 || type
.width
*type
.length
== 128)) ||
2008 (util_cpu_caps
.has_avx
&& type
.width
*type
.length
== 256) ||
2009 (util_cpu_caps
.has_avx512f
&& type
.width
*type
.length
== 512))
2011 else if ((util_cpu_caps
.has_altivec
&&
2012 (type
.width
== 32 && type
.length
== 4)))
2014 else if (util_cpu_caps
.has_neon
)
2020 enum lp_build_round_mode
2022 LP_BUILD_ROUND_NEAREST
= 0,
2023 LP_BUILD_ROUND_FLOOR
= 1,
2024 LP_BUILD_ROUND_CEIL
= 2,
2025 LP_BUILD_ROUND_TRUNCATE
= 3
2028 static inline LLVMValueRef
2029 lp_build_iround_nearest_sse2(struct lp_build_context
*bld
,
2032 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2033 const struct lp_type type
= bld
->type
;
2034 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
2035 LLVMTypeRef ret_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
2036 const char *intrinsic
;
2039 assert(type
.floating
);
2040 /* using the double precision conversions is a bit more complicated */
2041 assert(type
.width
== 32);
2043 assert(lp_check_value(type
, a
));
2044 assert(util_cpu_caps
.has_sse2
);
2046 /* This is relying on MXCSR rounding mode, which should always be nearest. */
2047 if (type
.length
== 1) {
2048 LLVMTypeRef vec_type
;
2051 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
2053 vec_type
= LLVMVectorType(bld
->elem_type
, 4);
2055 intrinsic
= "llvm.x86.sse.cvtss2si";
2057 undef
= LLVMGetUndef(vec_type
);
2059 arg
= LLVMBuildInsertElement(builder
, undef
, a
, index0
, "");
2061 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2065 if (type
.width
* type
.length
== 128) {
2066 intrinsic
= "llvm.x86.sse2.cvtps2dq";
2069 assert(type
.width
*type
.length
== 256);
2070 assert(util_cpu_caps
.has_avx
);
2072 intrinsic
= "llvm.x86.avx.cvt.ps2dq.256";
2074 res
= lp_build_intrinsic_unary(builder
, intrinsic
,
2084 static inline LLVMValueRef
2085 lp_build_round_altivec(struct lp_build_context
*bld
,
2087 enum lp_build_round_mode mode
)
2089 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2090 const struct lp_type type
= bld
->type
;
2091 const char *intrinsic
= NULL
;
2093 assert(type
.floating
);
2095 assert(lp_check_value(type
, a
));
2096 assert(util_cpu_caps
.has_altivec
);
2101 case LP_BUILD_ROUND_NEAREST
:
2102 intrinsic
= "llvm.ppc.altivec.vrfin";
2104 case LP_BUILD_ROUND_FLOOR
:
2105 intrinsic
= "llvm.ppc.altivec.vrfim";
2107 case LP_BUILD_ROUND_CEIL
:
2108 intrinsic
= "llvm.ppc.altivec.vrfip";
2110 case LP_BUILD_ROUND_TRUNCATE
:
2111 intrinsic
= "llvm.ppc.altivec.vrfiz";
2115 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2118 static inline LLVMValueRef
2119 lp_build_round_arch(struct lp_build_context
*bld
,
2121 enum lp_build_round_mode mode
)
2123 if (util_cpu_caps
.has_sse4_1
|| util_cpu_caps
.has_neon
) {
2124 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2125 const struct lp_type type
= bld
->type
;
2126 const char *intrinsic_root
;
2129 assert(type
.floating
);
2130 assert(lp_check_value(type
, a
));
2134 case LP_BUILD_ROUND_NEAREST
:
2135 intrinsic_root
= "llvm.nearbyint";
2137 case LP_BUILD_ROUND_FLOOR
:
2138 intrinsic_root
= "llvm.floor";
2140 case LP_BUILD_ROUND_CEIL
:
2141 intrinsic_root
= "llvm.ceil";
2143 case LP_BUILD_ROUND_TRUNCATE
:
2144 intrinsic_root
= "llvm.trunc";
2148 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, intrinsic_root
, bld
->vec_type
);
2149 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2151 else /* (util_cpu_caps.has_altivec) */
2152 return lp_build_round_altivec(bld
, a
, mode
);
2156 * Return the integer part of a float (vector) value (== round toward zero).
2157 * The returned value is a float (vector).
2158 * Ex: trunc(-1.5) = -1.0
2161 lp_build_trunc(struct lp_build_context
*bld
,
2164 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2165 const struct lp_type type
= bld
->type
;
2167 assert(type
.floating
);
2168 assert(lp_check_value(type
, a
));
2170 if (arch_rounding_available(type
)) {
2171 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_TRUNCATE
);
2174 const struct lp_type type
= bld
->type
;
2175 struct lp_type inttype
;
2176 struct lp_build_context intbld
;
2177 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2178 LLVMValueRef trunc
, res
, anosign
, mask
;
2179 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2180 LLVMTypeRef vec_type
= bld
->vec_type
;
2182 assert(type
.width
== 32); /* might want to handle doubles at some point */
2185 inttype
.floating
= 0;
2186 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2188 /* round by truncation */
2189 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2190 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2192 /* mask out sign bit */
2193 anosign
= lp_build_abs(bld
, a
);
2195 * mask out all values if anosign > 2^24
2196 * This should work both for large ints (all rounding is no-op for them
2197 * because such floats are always exact) as well as special cases like
2198 * NaNs, Infs (taking advantage of the fact they use max exponent).
2199 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2201 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2202 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2203 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2204 return lp_build_select(bld
, mask
, a
, res
);
2210 * Return float (vector) rounded to nearest integer (vector). The returned
2211 * value is a float (vector).
2212 * Ex: round(0.9) = 1.0
2213 * Ex: round(-1.5) = -2.0
2216 lp_build_round(struct lp_build_context
*bld
,
2219 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2220 const struct lp_type type
= bld
->type
;
2222 assert(type
.floating
);
2223 assert(lp_check_value(type
, a
));
2225 if (arch_rounding_available(type
)) {
2226 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2229 const struct lp_type type
= bld
->type
;
2230 struct lp_type inttype
;
2231 struct lp_build_context intbld
;
2232 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2233 LLVMValueRef res
, anosign
, mask
;
2234 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2235 LLVMTypeRef vec_type
= bld
->vec_type
;
2237 assert(type
.width
== 32); /* might want to handle doubles at some point */
2240 inttype
.floating
= 0;
2241 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2243 res
= lp_build_iround(bld
, a
);
2244 res
= LLVMBuildSIToFP(builder
, res
, vec_type
, "");
2246 /* mask out sign bit */
2247 anosign
= lp_build_abs(bld
, a
);
2249 * mask out all values if anosign > 2^24
2250 * This should work both for large ints (all rounding is no-op for them
2251 * because such floats are always exact) as well as special cases like
2252 * NaNs, Infs (taking advantage of the fact they use max exponent).
2253 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2255 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2256 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2257 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2258 return lp_build_select(bld
, mask
, a
, res
);
2264 * Return floor of float (vector), result is a float (vector)
2265 * Ex: floor(1.1) = 1.0
2266 * Ex: floor(-1.1) = -2.0
2269 lp_build_floor(struct lp_build_context
*bld
,
2272 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2273 const struct lp_type type
= bld
->type
;
2275 assert(type
.floating
);
2276 assert(lp_check_value(type
, a
));
2278 if (arch_rounding_available(type
)) {
2279 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2282 const struct lp_type type
= bld
->type
;
2283 struct lp_type inttype
;
2284 struct lp_build_context intbld
;
2285 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2286 LLVMValueRef trunc
, res
, anosign
, mask
;
2287 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2288 LLVMTypeRef vec_type
= bld
->vec_type
;
2290 if (type
.width
!= 32) {
2292 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.floor", vec_type
);
2293 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2296 assert(type
.width
== 32); /* might want to handle doubles at some point */
2299 inttype
.floating
= 0;
2300 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2302 /* round by truncation */
2303 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2304 res
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "floor.trunc");
2310 * fix values if rounding is wrong (for non-special cases)
2311 * - this is the case if trunc > a
2313 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, res
, a
);
2314 /* tmp = trunc > a ? 1.0 : 0.0 */
2315 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2316 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2317 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2318 res
= lp_build_sub(bld
, res
, tmp
);
2321 /* mask out sign bit */
2322 anosign
= lp_build_abs(bld
, a
);
2324 * mask out all values if anosign > 2^24
2325 * This should work both for large ints (all rounding is no-op for them
2326 * because such floats are always exact) as well as special cases like
2327 * NaNs, Infs (taking advantage of the fact they use max exponent).
2328 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2330 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2331 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2332 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2333 return lp_build_select(bld
, mask
, a
, res
);
2339 * Return ceiling of float (vector), returning float (vector).
2340 * Ex: ceil( 1.1) = 2.0
2341 * Ex: ceil(-1.1) = -1.0
2344 lp_build_ceil(struct lp_build_context
*bld
,
2347 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2348 const struct lp_type type
= bld
->type
;
2350 assert(type
.floating
);
2351 assert(lp_check_value(type
, a
));
2353 if (arch_rounding_available(type
)) {
2354 return lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2357 const struct lp_type type
= bld
->type
;
2358 struct lp_type inttype
;
2359 struct lp_build_context intbld
;
2360 LLVMValueRef cmpval
= lp_build_const_vec(bld
->gallivm
, type
, 1<<24);
2361 LLVMValueRef trunc
, res
, anosign
, mask
, tmp
;
2362 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2363 LLVMTypeRef vec_type
= bld
->vec_type
;
2365 if (type
.width
!= 32) {
2367 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.ceil", vec_type
);
2368 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2371 assert(type
.width
== 32); /* might want to handle doubles at some point */
2374 inttype
.floating
= 0;
2375 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2377 /* round by truncation */
2378 trunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2379 trunc
= LLVMBuildSIToFP(builder
, trunc
, vec_type
, "ceil.trunc");
2382 * fix values if rounding is wrong (for non-special cases)
2383 * - this is the case if trunc < a
2385 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2386 /* tmp = trunc < a ? 1.0 : 0.0 */
2387 tmp
= LLVMBuildBitCast(builder
, bld
->one
, int_vec_type
, "");
2388 tmp
= lp_build_and(&intbld
, mask
, tmp
);
2389 tmp
= LLVMBuildBitCast(builder
, tmp
, vec_type
, "");
2390 res
= lp_build_add(bld
, trunc
, tmp
);
2392 /* mask out sign bit */
2393 anosign
= lp_build_abs(bld
, a
);
2395 * mask out all values if anosign > 2^24
2396 * This should work both for large ints (all rounding is no-op for them
2397 * because such floats are always exact) as well as special cases like
2398 * NaNs, Infs (taking advantage of the fact they use max exponent).
2399 * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
2401 anosign
= LLVMBuildBitCast(builder
, anosign
, int_vec_type
, "");
2402 cmpval
= LLVMBuildBitCast(builder
, cmpval
, int_vec_type
, "");
2403 mask
= lp_build_cmp(&intbld
, PIPE_FUNC_GREATER
, anosign
, cmpval
);
2404 return lp_build_select(bld
, mask
, a
, res
);
2410 * Return fractional part of 'a' computed as a - floor(a)
2411 * Typically used in texture coord arithmetic.
2414 lp_build_fract(struct lp_build_context
*bld
,
2417 assert(bld
->type
.floating
);
2418 return lp_build_sub(bld
, a
, lp_build_floor(bld
, a
));
2423 * Prevent returning 1.0 for very small negative values of 'a' by clamping
2424 * against 0.99999(9). (Will also return that value for NaNs.)
2426 static inline LLVMValueRef
2427 clamp_fract(struct lp_build_context
*bld
, LLVMValueRef fract
)
2431 /* this is the largest number smaller than 1.0 representable as float */
2432 max
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
2433 1.0 - 1.0/(1LL << (lp_mantissa(bld
->type
) + 1)));
2434 return lp_build_min_ext(bld
, fract
, max
,
2435 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
2440 * Same as lp_build_fract, but guarantees that the result is always smaller
2441 * than one. Will also return the smaller-than-one value for infs, NaNs.
2444 lp_build_fract_safe(struct lp_build_context
*bld
,
2447 return clamp_fract(bld
, lp_build_fract(bld
, a
));
2452 * Return the integer part of a float (vector) value (== round toward zero).
2453 * The returned value is an integer (vector).
2454 * Ex: itrunc(-1.5) = -1
2457 lp_build_itrunc(struct lp_build_context
*bld
,
2460 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2461 const struct lp_type type
= bld
->type
;
2462 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
2464 assert(type
.floating
);
2465 assert(lp_check_value(type
, a
));
2467 return LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2472 * Return float (vector) rounded to nearest integer (vector). The returned
2473 * value is an integer (vector).
2474 * Ex: iround(0.9) = 1
2475 * Ex: iround(-1.5) = -2
2478 lp_build_iround(struct lp_build_context
*bld
,
2481 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2482 const struct lp_type type
= bld
->type
;
2483 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2486 assert(type
.floating
);
2488 assert(lp_check_value(type
, a
));
2490 if ((util_cpu_caps
.has_sse2
&&
2491 ((type
.width
== 32) && (type
.length
== 1 || type
.length
== 4))) ||
2492 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2493 return lp_build_iround_nearest_sse2(bld
, a
);
2495 if (arch_rounding_available(type
)) {
2496 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_NEAREST
);
2501 half
= lp_build_const_vec(bld
->gallivm
, type
, nextafterf(0.5, 0.0));
2504 LLVMTypeRef vec_type
= bld
->vec_type
;
2505 LLVMValueRef mask
= lp_build_const_int_vec(bld
->gallivm
, type
,
2506 (unsigned long long)1 << (type
.width
- 1));
2510 sign
= LLVMBuildBitCast(builder
, a
, int_vec_type
, "");
2511 sign
= LLVMBuildAnd(builder
, sign
, mask
, "");
2514 half
= LLVMBuildBitCast(builder
, half
, int_vec_type
, "");
2515 half
= LLVMBuildOr(builder
, sign
, half
, "");
2516 half
= LLVMBuildBitCast(builder
, half
, vec_type
, "");
2519 res
= LLVMBuildFAdd(builder
, a
, half
, "");
2522 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "");
2529 * Return floor of float (vector), result is an int (vector)
2530 * Ex: ifloor(1.1) = 1.0
2531 * Ex: ifloor(-1.1) = -2.0
2534 lp_build_ifloor(struct lp_build_context
*bld
,
2537 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2538 const struct lp_type type
= bld
->type
;
2539 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2542 assert(type
.floating
);
2543 assert(lp_check_value(type
, a
));
2547 if (arch_rounding_available(type
)) {
2548 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_FLOOR
);
2551 struct lp_type inttype
;
2552 struct lp_build_context intbld
;
2553 LLVMValueRef trunc
, itrunc
, mask
;
2555 assert(type
.floating
);
2556 assert(lp_check_value(type
, a
));
2559 inttype
.floating
= 0;
2560 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2562 /* round by truncation */
2563 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2564 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "ifloor.trunc");
2567 * fix values if rounding is wrong (for non-special cases)
2568 * - this is the case if trunc > a
2569 * The results of doing this with NaNs, very large values etc.
2570 * are undefined but this seems to be the case anyway.
2572 mask
= lp_build_cmp(bld
, PIPE_FUNC_GREATER
, trunc
, a
);
2573 /* cheapie minus one with mask since the mask is minus one / zero */
2574 return lp_build_add(&intbld
, itrunc
, mask
);
2578 /* round to nearest (toward zero) */
2579 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "ifloor.res");
2586 * Return ceiling of float (vector), returning int (vector).
2587 * Ex: iceil( 1.1) = 2
2588 * Ex: iceil(-1.1) = -1
2591 lp_build_iceil(struct lp_build_context
*bld
,
2594 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2595 const struct lp_type type
= bld
->type
;
2596 LLVMTypeRef int_vec_type
= bld
->int_vec_type
;
2599 assert(type
.floating
);
2600 assert(lp_check_value(type
, a
));
2602 if (arch_rounding_available(type
)) {
2603 res
= lp_build_round_arch(bld
, a
, LP_BUILD_ROUND_CEIL
);
2606 struct lp_type inttype
;
2607 struct lp_build_context intbld
;
2608 LLVMValueRef trunc
, itrunc
, mask
;
2610 assert(type
.floating
);
2611 assert(lp_check_value(type
, a
));
2614 inttype
.floating
= 0;
2615 lp_build_context_init(&intbld
, bld
->gallivm
, inttype
);
2617 /* round by truncation */
2618 itrunc
= LLVMBuildFPToSI(builder
, a
, int_vec_type
, "");
2619 trunc
= LLVMBuildSIToFP(builder
, itrunc
, bld
->vec_type
, "iceil.trunc");
2622 * fix values if rounding is wrong (for non-special cases)
2623 * - this is the case if trunc < a
2624 * The results of doing this with NaNs, very large values etc.
2625 * are undefined but this seems to be the case anyway.
2627 mask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, trunc
, a
);
2628 /* cheapie plus one with mask since the mask is minus one / zero */
2629 return lp_build_sub(&intbld
, itrunc
, mask
);
2632 /* round to nearest (toward zero) */
2633 res
= LLVMBuildFPToSI(builder
, res
, int_vec_type
, "iceil.res");
2640 * Combined ifloor() & fract().
2642 * Preferred to calling the functions separately, as it will ensure that the
2643 * strategy (floor() vs ifloor()) that results in less redundant work is used.
2646 lp_build_ifloor_fract(struct lp_build_context
*bld
,
2648 LLVMValueRef
*out_ipart
,
2649 LLVMValueRef
*out_fpart
)
2651 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2652 const struct lp_type type
= bld
->type
;
2655 assert(type
.floating
);
2656 assert(lp_check_value(type
, a
));
2658 if (arch_rounding_available(type
)) {
2660 * floor() is easier.
2663 ipart
= lp_build_floor(bld
, a
);
2664 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2665 *out_ipart
= LLVMBuildFPToSI(builder
, ipart
, bld
->int_vec_type
, "ipart");
2669 * ifloor() is easier.
2672 *out_ipart
= lp_build_ifloor(bld
, a
);
2673 ipart
= LLVMBuildSIToFP(builder
, *out_ipart
, bld
->vec_type
, "ipart");
2674 *out_fpart
= LLVMBuildFSub(builder
, a
, ipart
, "fpart");
2680 * Same as lp_build_ifloor_fract, but guarantees that the fractional part is
2681 * always smaller than one.
2684 lp_build_ifloor_fract_safe(struct lp_build_context
*bld
,
2686 LLVMValueRef
*out_ipart
,
2687 LLVMValueRef
*out_fpart
)
2689 lp_build_ifloor_fract(bld
, a
, out_ipart
, out_fpart
);
2690 *out_fpart
= clamp_fract(bld
, *out_fpart
);
2695 lp_build_sqrt(struct lp_build_context
*bld
,
2698 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2699 const struct lp_type type
= bld
->type
;
2700 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
2703 assert(lp_check_value(type
, a
));
2705 assert(type
.floating
);
2706 lp_format_intrinsic(intrinsic
, sizeof intrinsic
, "llvm.sqrt", vec_type
);
2708 return lp_build_intrinsic_unary(builder
, intrinsic
, vec_type
, a
);
2713 * Do one Newton-Raphson step to improve reciprocate precision:
2715 * x_{i+1} = x_i + x_i * (1 - a * x_i)
2717 * XXX: Unfortunately this won't give IEEE-754 conformant results for 0 or
2718 * +/-Inf, giving NaN instead. Certain applications rely on this behavior,
2719 * such as Google Earth, which does RCP(RSQRT(0.0)) when drawing the Earth's
2720 * halo. It would be necessary to clamp the argument to prevent this.
2723 * - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division
2724 * - http://softwarecommunity.intel.com/articles/eng/1818.htm
2726 static inline LLVMValueRef
2727 lp_build_rcp_refine(struct lp_build_context
*bld
,
2731 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2735 neg_a
= LLVMBuildFNeg(builder
, a
, "");
2736 res
= lp_build_fmuladd(builder
, neg_a
, rcp_a
, bld
->one
);
2737 res
= lp_build_fmuladd(builder
, res
, rcp_a
, rcp_a
);
2744 lp_build_rcp(struct lp_build_context
*bld
,
2747 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2748 const struct lp_type type
= bld
->type
;
2750 assert(lp_check_value(type
, a
));
2759 assert(type
.floating
);
2761 if(LLVMIsConstant(a
))
2762 return LLVMConstFDiv(bld
->one
, a
);
2765 * We don't use RCPPS because:
2766 * - it only has 10bits of precision
2767 * - it doesn't even get the reciprocate of 1.0 exactly
2768 * - doing Newton-Rapshon steps yields wrong (NaN) values for 0.0 or Inf
2769 * - for recent processors the benefit over DIVPS is marginal, a case
2772 * We could still use it on certain processors if benchmarks show that the
2773 * RCPPS plus necessary workarounds are still preferrable to DIVPS; or for
2774 * particular uses that require less workarounds.
2777 if (FALSE
&& ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2778 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8))){
2779 const unsigned num_iterations
= 0;
2782 const char *intrinsic
= NULL
;
2784 if (type
.length
== 4) {
2785 intrinsic
= "llvm.x86.sse.rcp.ps";
2788 intrinsic
= "llvm.x86.avx.rcp.ps.256";
2791 res
= lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2793 for (i
= 0; i
< num_iterations
; ++i
) {
2794 res
= lp_build_rcp_refine(bld
, a
, res
);
2800 return LLVMBuildFDiv(builder
, bld
->one
, a
, "");
2805 * Do one Newton-Raphson step to improve rsqrt precision:
2807 * x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
2809 * See also Intel 64 and IA-32 Architectures Optimization Manual.
2811 static inline LLVMValueRef
2812 lp_build_rsqrt_refine(struct lp_build_context
*bld
,
2814 LLVMValueRef rsqrt_a
)
2816 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2817 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 0.5);
2818 LLVMValueRef three
= lp_build_const_vec(bld
->gallivm
, bld
->type
, 3.0);
2821 res
= LLVMBuildFMul(builder
, rsqrt_a
, rsqrt_a
, "");
2822 res
= LLVMBuildFMul(builder
, a
, res
, "");
2823 res
= LLVMBuildFSub(builder
, three
, res
, "");
2824 res
= LLVMBuildFMul(builder
, rsqrt_a
, res
, "");
2825 res
= LLVMBuildFMul(builder
, half
, res
, "");
2832 * Generate 1/sqrt(a).
2833 * Result is undefined for values < 0, infinity for +0.
2836 lp_build_rsqrt(struct lp_build_context
*bld
,
2839 const struct lp_type type
= bld
->type
;
2841 assert(lp_check_value(type
, a
));
2843 assert(type
.floating
);
2846 * This should be faster but all denormals will end up as infinity.
2848 if (0 && lp_build_fast_rsqrt_available(type
)) {
2849 const unsigned num_iterations
= 1;
2853 /* rsqrt(1.0) != 1.0 here */
2854 res
= lp_build_fast_rsqrt(bld
, a
);
2856 if (num_iterations
) {
2858 * Newton-Raphson will result in NaN instead of infinity for zero,
2859 * and NaN instead of zero for infinity.
2860 * Also, need to ensure rsqrt(1.0) == 1.0.
2861 * All numbers smaller than FLT_MIN will result in +infinity
2862 * (rsqrtps treats all denormals as zero).
2865 LLVMValueRef flt_min
= lp_build_const_vec(bld
->gallivm
, type
, FLT_MIN
);
2866 LLVMValueRef inf
= lp_build_const_vec(bld
->gallivm
, type
, INFINITY
);
2868 for (i
= 0; i
< num_iterations
; ++i
) {
2869 res
= lp_build_rsqrt_refine(bld
, a
, res
);
2871 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_LESS
, a
, flt_min
);
2872 res
= lp_build_select(bld
, cmp
, inf
, res
);
2873 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, inf
);
2874 res
= lp_build_select(bld
, cmp
, bld
->zero
, res
);
2875 cmp
= lp_build_compare(bld
->gallivm
, type
, PIPE_FUNC_EQUAL
, a
, bld
->one
);
2876 res
= lp_build_select(bld
, cmp
, bld
->one
, res
);
2882 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2886 * If there's a fast (inaccurate) rsqrt instruction available
2887 * (caller may want to avoid to call rsqrt_fast if it's not available,
2888 * i.e. for calculating x^0.5 it may do rsqrt_fast(x) * x but if
2889 * unavailable it would result in sqrt/div/mul so obviously
2890 * much better to just call sqrt, skipping both div and mul).
2893 lp_build_fast_rsqrt_available(struct lp_type type
)
2895 assert(type
.floating
);
2897 if ((util_cpu_caps
.has_sse
&& type
.width
== 32 && type
.length
== 4) ||
2898 (util_cpu_caps
.has_avx
&& type
.width
== 32 && type
.length
== 8)) {
2906 * Generate 1/sqrt(a).
2907 * Result is undefined for values < 0, infinity for +0.
2908 * Precision is limited, only ~10 bits guaranteed
2909 * (rsqrt 1.0 may not be 1.0, denorms may be flushed to 0).
2912 lp_build_fast_rsqrt(struct lp_build_context
*bld
,
2915 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2916 const struct lp_type type
= bld
->type
;
2918 assert(lp_check_value(type
, a
));
2920 if (lp_build_fast_rsqrt_available(type
)) {
2921 const char *intrinsic
= NULL
;
2923 if (type
.length
== 4) {
2924 intrinsic
= "llvm.x86.sse.rsqrt.ps";
2927 intrinsic
= "llvm.x86.avx.rsqrt.ps.256";
2929 return lp_build_intrinsic_unary(builder
, intrinsic
, bld
->vec_type
, a
);
2932 debug_printf("%s: emulating fast rsqrt with rcp/sqrt\n", __FUNCTION__
);
2934 return lp_build_rcp(bld
, lp_build_sqrt(bld
, a
));
2939 * Generate sin(a) or cos(a) using polynomial approximation.
2940 * TODO: it might be worth recognizing sin and cos using same source
2941 * (i.e. d3d10 sincos opcode). Obviously doing both at the same time
2942 * would be way cheaper than calculating (nearly) everything twice...
2943 * Not sure it's common enough to be worth bothering however, scs
2944 * opcode could also benefit from calculating both though.
2947 lp_build_sin_or_cos(struct lp_build_context
*bld
,
2951 struct gallivm_state
*gallivm
= bld
->gallivm
;
2952 LLVMBuilderRef b
= gallivm
->builder
;
2953 struct lp_type int_type
= lp_int_type(bld
->type
);
2956 * take the absolute value,
2957 * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
2960 LLVMValueRef inv_sig_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, ~0x80000000);
2961 LLVMValueRef a_v4si
= LLVMBuildBitCast(b
, a
, bld
->int_vec_type
, "a_v4si");
2963 LLVMValueRef absi
= LLVMBuildAnd(b
, a_v4si
, inv_sig_mask
, "absi");
2964 LLVMValueRef x_abs
= LLVMBuildBitCast(b
, absi
, bld
->vec_type
, "x_abs");
2968 * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
2971 LLVMValueRef FOPi
= lp_build_const_vec(gallivm
, bld
->type
, 1.27323954473516);
2972 LLVMValueRef scale_y
= LLVMBuildFMul(b
, x_abs
, FOPi
, "scale_y");
2975 * store the integer part of y in mm0
2976 * emm2 = _mm_cvttps_epi32(y);
2979 LLVMValueRef emm2_i
= LLVMBuildFPToSI(b
, scale_y
, bld
->int_vec_type
, "emm2_i");
2982 * j=(j+1) & (~1) (see the cephes sources)
2983 * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
2986 LLVMValueRef all_one
= lp_build_const_int_vec(gallivm
, bld
->type
, 1);
2987 LLVMValueRef emm2_add
= LLVMBuildAdd(b
, emm2_i
, all_one
, "emm2_add");
2989 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
2991 LLVMValueRef inv_one
= lp_build_const_int_vec(gallivm
, bld
->type
, ~1);
2992 LLVMValueRef emm2_and
= LLVMBuildAnd(b
, emm2_add
, inv_one
, "emm2_and");
2995 * y = _mm_cvtepi32_ps(emm2);
2997 LLVMValueRef y_2
= LLVMBuildSIToFP(b
, emm2_and
, bld
->vec_type
, "y_2");
2999 LLVMValueRef const_2
= lp_build_const_int_vec(gallivm
, bld
->type
, 2);
3000 LLVMValueRef const_4
= lp_build_const_int_vec(gallivm
, bld
->type
, 4);
3001 LLVMValueRef const_29
= lp_build_const_int_vec(gallivm
, bld
->type
, 29);
3002 LLVMValueRef sign_mask
= lp_build_const_int_vec(gallivm
, bld
->type
, 0x80000000);
3005 * Argument used for poly selection and sign bit determination
3006 * is different for sin vs. cos.
3008 LLVMValueRef emm2_2
= cos
? LLVMBuildSub(b
, emm2_and
, const_2
, "emm2_2") :
3011 LLVMValueRef sign_bit
= cos
? LLVMBuildShl(b
, LLVMBuildAnd(b
, const_4
,
3012 LLVMBuildNot(b
, emm2_2
, ""), ""),
3013 const_29
, "sign_bit") :
3014 LLVMBuildAnd(b
, LLVMBuildXor(b
, a_v4si
,
3015 LLVMBuildShl(b
, emm2_add
,
3017 sign_mask
, "sign_bit");
3020 * get the polynom selection mask
3021 * there is one polynom for 0 <= x <= Pi/4
3022 * and another one for Pi/4<x<=Pi/2
3023 * Both branches will be computed.
3025 * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
3026 * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
3029 LLVMValueRef emm2_3
= LLVMBuildAnd(b
, emm2_2
, const_2
, "emm2_3");
3030 LLVMValueRef poly_mask
= lp_build_compare(gallivm
,
3031 int_type
, PIPE_FUNC_EQUAL
,
3032 emm2_3
, lp_build_const_int_vec(gallivm
, bld
->type
, 0));
3035 * _PS_CONST(minus_cephes_DP1, -0.78515625);
3036 * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
3037 * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
3039 LLVMValueRef DP1
= lp_build_const_vec(gallivm
, bld
->type
, -0.78515625);
3040 LLVMValueRef DP2
= lp_build_const_vec(gallivm
, bld
->type
, -2.4187564849853515625e-4);
3041 LLVMValueRef DP3
= lp_build_const_vec(gallivm
, bld
->type
, -3.77489497744594108e-8);
3044 * The magic pass: "Extended precision modular arithmetic"
3045 * x = ((x - y * DP1) - y * DP2) - y * DP3;
3047 LLVMValueRef x_1
= lp_build_fmuladd(b
, y_2
, DP1
, x_abs
);
3048 LLVMValueRef x_2
= lp_build_fmuladd(b
, y_2
, DP2
, x_1
);
3049 LLVMValueRef x_3
= lp_build_fmuladd(b
, y_2
, DP3
, x_2
);
3052 * Evaluate the first polynom (0 <= x <= Pi/4)
3054 * z = _mm_mul_ps(x,x);
3056 LLVMValueRef z
= LLVMBuildFMul(b
, x_3
, x_3
, "z");
3059 * _PS_CONST(coscof_p0, 2.443315711809948E-005);
3060 * _PS_CONST(coscof_p1, -1.388731625493765E-003);
3061 * _PS_CONST(coscof_p2, 4.166664568298827E-002);
3063 LLVMValueRef coscof_p0
= lp_build_const_vec(gallivm
, bld
->type
, 2.443315711809948E-005);
3064 LLVMValueRef coscof_p1
= lp_build_const_vec(gallivm
, bld
->type
, -1.388731625493765E-003);
3065 LLVMValueRef coscof_p2
= lp_build_const_vec(gallivm
, bld
->type
, 4.166664568298827E-002);
3068 * y = *(v4sf*)_ps_coscof_p0;
3069 * y = _mm_mul_ps(y, z);
3071 LLVMValueRef y_4
= lp_build_fmuladd(b
, z
, coscof_p0
, coscof_p1
);
3072 LLVMValueRef y_6
= lp_build_fmuladd(b
, y_4
, z
, coscof_p2
);
3073 LLVMValueRef y_7
= LLVMBuildFMul(b
, y_6
, z
, "y_7");
3074 LLVMValueRef y_8
= LLVMBuildFMul(b
, y_7
, z
, "y_8");
3078 * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
3079 * y = _mm_sub_ps(y, tmp);
3080 * y = _mm_add_ps(y, *(v4sf*)_ps_1);
3082 LLVMValueRef half
= lp_build_const_vec(gallivm
, bld
->type
, 0.5);
3083 LLVMValueRef tmp
= LLVMBuildFMul(b
, z
, half
, "tmp");
3084 LLVMValueRef y_9
= LLVMBuildFSub(b
, y_8
, tmp
, "y_8");
3085 LLVMValueRef one
= lp_build_const_vec(gallivm
, bld
->type
, 1.0);
3086 LLVMValueRef y_10
= LLVMBuildFAdd(b
, y_9
, one
, "y_9");
3089 * _PS_CONST(sincof_p0, -1.9515295891E-4);
3090 * _PS_CONST(sincof_p1, 8.3321608736E-3);
3091 * _PS_CONST(sincof_p2, -1.6666654611E-1);
3093 LLVMValueRef sincof_p0
= lp_build_const_vec(gallivm
, bld
->type
, -1.9515295891E-4);
3094 LLVMValueRef sincof_p1
= lp_build_const_vec(gallivm
, bld
->type
, 8.3321608736E-3);
3095 LLVMValueRef sincof_p2
= lp_build_const_vec(gallivm
, bld
->type
, -1.6666654611E-1);
3098 * Evaluate the second polynom (Pi/4 <= x <= 0)
3100 * y2 = *(v4sf*)_ps_sincof_p0;
3101 * y2 = _mm_mul_ps(y2, z);
3102 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
3103 * y2 = _mm_mul_ps(y2, z);
3104 * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
3105 * y2 = _mm_mul_ps(y2, z);
3106 * y2 = _mm_mul_ps(y2, x);
3107 * y2 = _mm_add_ps(y2, x);
3110 LLVMValueRef y2_4
= lp_build_fmuladd(b
, z
, sincof_p0
, sincof_p1
);
3111 LLVMValueRef y2_6
= lp_build_fmuladd(b
, y2_4
, z
, sincof_p2
);
3112 LLVMValueRef y2_7
= LLVMBuildFMul(b
, y2_6
, z
, "y2_7");
3113 LLVMValueRef y2_9
= lp_build_fmuladd(b
, y2_7
, x_3
, x_3
);
3116 * select the correct result from the two polynoms
3118 * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
3119 * y = _mm_andnot_ps(xmm3, y);
3120 * y = _mm_or_ps(y,y2);
3122 LLVMValueRef y2_i
= LLVMBuildBitCast(b
, y2_9
, bld
->int_vec_type
, "y2_i");
3123 LLVMValueRef y_i
= LLVMBuildBitCast(b
, y_10
, bld
->int_vec_type
, "y_i");
3124 LLVMValueRef y2_and
= LLVMBuildAnd(b
, y2_i
, poly_mask
, "y2_and");
3125 LLVMValueRef poly_mask_inv
= LLVMBuildNot(b
, poly_mask
, "poly_mask_inv");
3126 LLVMValueRef y_and
= LLVMBuildAnd(b
, y_i
, poly_mask_inv
, "y_and");
3127 LLVMValueRef y_combine
= LLVMBuildOr(b
, y_and
, y2_and
, "y_combine");
3131 * y = _mm_xor_ps(y, sign_bit);
3133 LLVMValueRef y_sign
= LLVMBuildXor(b
, y_combine
, sign_bit
, "y_sign");
3134 LLVMValueRef y_result
= LLVMBuildBitCast(b
, y_sign
, bld
->vec_type
, "y_result");
3136 LLVMValueRef isfinite
= lp_build_isfinite(bld
, a
);
3138 /* clamp output to be within [-1, 1] */
3139 y_result
= lp_build_clamp(bld
, y_result
,
3140 lp_build_const_vec(bld
->gallivm
, bld
->type
, -1.f
),
3141 lp_build_const_vec(bld
->gallivm
, bld
->type
, 1.f
));
3142 /* If a is -inf, inf or NaN then return NaN */
3143 y_result
= lp_build_select(bld
, isfinite
, y_result
,
3144 lp_build_const_vec(bld
->gallivm
, bld
->type
, NAN
));
3153 lp_build_sin(struct lp_build_context
*bld
,
3156 return lp_build_sin_or_cos(bld
, a
, FALSE
);
3164 lp_build_cos(struct lp_build_context
*bld
,
3167 return lp_build_sin_or_cos(bld
, a
, TRUE
);
3172 * Generate pow(x, y)
3175 lp_build_pow(struct lp_build_context
*bld
,
3179 /* TODO: optimize the constant case */
3180 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3181 LLVMIsConstant(x
) && LLVMIsConstant(y
)) {
3182 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3186 return lp_build_exp2(bld
, lp_build_mul(bld
, lp_build_log2(bld
, x
), y
));
3194 lp_build_exp(struct lp_build_context
*bld
,
3197 /* log2(e) = 1/log(2) */
3198 LLVMValueRef log2e
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3199 1.4426950408889634);
3201 assert(lp_check_value(bld
->type
, x
));
3203 return lp_build_exp2(bld
, lp_build_mul(bld
, log2e
, x
));
3209 * Behavior is undefined with infs, 0s and nans
3212 lp_build_log(struct lp_build_context
*bld
,
3216 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3217 0.69314718055994529);
3219 assert(lp_check_value(bld
->type
, x
));
3221 return lp_build_mul(bld
, log2
, lp_build_log2(bld
, x
));
3225 * Generate log(x) that handles edge cases (infs, 0s and nans)
3228 lp_build_log_safe(struct lp_build_context
*bld
,
3232 LLVMValueRef log2
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
3233 0.69314718055994529);
3235 assert(lp_check_value(bld
->type
, x
));
3237 return lp_build_mul(bld
, log2
, lp_build_log2_safe(bld
, x
));
3242 * Generate polynomial.
3243 * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
3246 lp_build_polynomial(struct lp_build_context
*bld
,
3248 const double *coeffs
,
3249 unsigned num_coeffs
)
3251 const struct lp_type type
= bld
->type
;
3252 LLVMValueRef even
= NULL
, odd
= NULL
;
3256 assert(lp_check_value(bld
->type
, x
));
3258 /* TODO: optimize the constant case */
3259 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3260 LLVMIsConstant(x
)) {
3261 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3266 * Calculate odd and even terms seperately to decrease data dependency
3268 * c[0] + x^2 * c[2] + x^4 * c[4] ...
3269 * + x * (c[1] + x^2 * c[3] + x^4 * c[5]) ...
3271 x2
= lp_build_mul(bld
, x
, x
);
3273 for (i
= num_coeffs
; i
--; ) {
3276 coeff
= lp_build_const_vec(bld
->gallivm
, type
, coeffs
[i
]);
3280 even
= lp_build_mad(bld
, x2
, even
, coeff
);
3285 odd
= lp_build_mad(bld
, x2
, odd
, coeff
);
3292 return lp_build_mad(bld
, odd
, x
, even
);
3301 * Minimax polynomial fit of 2**x, in range [0, 1[
3303 const double lp_build_exp2_polynomial
[] = {
3304 #if EXP_POLY_DEGREE == 5
3305 1.000000000000000000000, /*XXX: was 0.999999925063526176901, recompute others */
3306 0.693153073200168932794,
3307 0.240153617044375388211,
3308 0.0558263180532956664775,
3309 0.00898934009049466391101,
3310 0.00187757667519147912699
3311 #elif EXP_POLY_DEGREE == 4
3312 1.00000259337069434683,
3313 0.693003834469974940458,
3314 0.24144275689150793076,
3315 0.0520114606103070150235,
3316 0.0135341679161270268764
3317 #elif EXP_POLY_DEGREE == 3
3318 0.999925218562710312959,
3319 0.695833540494823811697,
3320 0.226067155427249155588,
3321 0.0780245226406372992967
3322 #elif EXP_POLY_DEGREE == 2
3323 1.00172476321474503578,
3324 0.657636275736077639316,
3325 0.33718943461968720704
3333 lp_build_exp2(struct lp_build_context
*bld
,
3336 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3337 const struct lp_type type
= bld
->type
;
3338 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3339 LLVMValueRef ipart
= NULL
;
3340 LLVMValueRef fpart
= NULL
;
3341 LLVMValueRef expipart
= NULL
;
3342 LLVMValueRef expfpart
= NULL
;
3343 LLVMValueRef res
= NULL
;
3345 assert(lp_check_value(bld
->type
, x
));
3347 /* TODO: optimize the constant case */
3348 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3349 LLVMIsConstant(x
)) {
3350 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3354 assert(type
.floating
&& type
.width
== 32);
3356 /* We want to preserve NaN and make sure than for exp2 if x > 128,
3357 * the result is INF and if it's smaller than -126.9 the result is 0 */
3358 x
= lp_build_min_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, 128.0), x
,
3359 GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3360 x
= lp_build_max_ext(bld
, lp_build_const_vec(bld
->gallivm
, type
, -126.99999),
3361 x
, GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN
);
3363 /* ipart = floor(x) */
3364 /* fpart = x - ipart */
3365 lp_build_ifloor_fract(bld
, x
, &ipart
, &fpart
);
3367 /* expipart = (float) (1 << ipart) */
3368 expipart
= LLVMBuildAdd(builder
, ipart
,
3369 lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3370 expipart
= LLVMBuildShl(builder
, expipart
,
3371 lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3372 expipart
= LLVMBuildBitCast(builder
, expipart
, vec_type
, "");
3374 expfpart
= lp_build_polynomial(bld
, fpart
, lp_build_exp2_polynomial
,
3375 ARRAY_SIZE(lp_build_exp2_polynomial
));
3377 res
= LLVMBuildFMul(builder
, expipart
, expfpart
, "");
3385 * Extract the exponent of a IEEE-754 floating point value.
3387 * Optionally apply an integer bias.
3389 * Result is an integer value with
3391 * ifloor(log2(x)) + bias
3394 lp_build_extract_exponent(struct lp_build_context
*bld
,
3398 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3399 const struct lp_type type
= bld
->type
;
3400 unsigned mantissa
= lp_mantissa(type
);
3403 assert(type
.floating
);
3405 assert(lp_check_value(bld
->type
, x
));
3407 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3409 res
= LLVMBuildLShr(builder
, x
,
3410 lp_build_const_int_vec(bld
->gallivm
, type
, mantissa
), "");
3411 res
= LLVMBuildAnd(builder
, res
,
3412 lp_build_const_int_vec(bld
->gallivm
, type
, 255), "");
3413 res
= LLVMBuildSub(builder
, res
,
3414 lp_build_const_int_vec(bld
->gallivm
, type
, 127 - bias
), "");
3421 * Extract the mantissa of the a floating.
3423 * Result is a floating point value with
3425 * x / floor(log2(x))
3428 lp_build_extract_mantissa(struct lp_build_context
*bld
,
3431 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3432 const struct lp_type type
= bld
->type
;
3433 unsigned mantissa
= lp_mantissa(type
);
3434 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
,
3435 (1ULL << mantissa
) - 1);
3436 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, bld
->int_vec_type
);
3439 assert(lp_check_value(bld
->type
, x
));
3441 assert(type
.floating
);
3443 x
= LLVMBuildBitCast(builder
, x
, bld
->int_vec_type
, "");
3445 /* res = x / 2**ipart */
3446 res
= LLVMBuildAnd(builder
, x
, mantmask
, "");
3447 res
= LLVMBuildOr(builder
, res
, one
, "");
3448 res
= LLVMBuildBitCast(builder
, res
, bld
->vec_type
, "");
3456 * Minimax polynomial fit of log2((1.0 + sqrt(x))/(1.0 - sqrt(x)))/sqrt(x) ,for x in range of [0, 1/9[
3457 * These coefficients can be generate with
3458 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
3460 const double lp_build_log2_polynomial
[] = {
3461 #if LOG_POLY_DEGREE == 5
3462 2.88539008148777786488L,
3463 0.961796878841293367824L,
3464 0.577058946784739859012L,
3465 0.412914355135828735411L,
3466 0.308591899232910175289L,
3467 0.352376952300281371868L,
3468 #elif LOG_POLY_DEGREE == 4
3469 2.88539009343309178325L,
3470 0.961791550404184197881L,
3471 0.577440339438736392009L,
3472 0.403343858251329912514L,
3473 0.406718052498846252698L,
3474 #elif LOG_POLY_DEGREE == 3
3475 2.88538959748872753838L,
3476 0.961932915889597772928L,
3477 0.571118517972136195241L,
3478 0.493997535084709500285L,
3485 * See http://www.devmaster.net/forums/showthread.php?p=43580
3486 * http://en.wikipedia.org/wiki/Logarithm#Calculation
3487 * http://www.nezumi.demon.co.uk/consult/logx.htm
3489 * If handle_edge_cases is true the function will perform computations
3490 * to match the required D3D10+ behavior for each of the edge cases.
3491 * That means that if input is:
3492 * - less than zero (to and including -inf) then NaN will be returned
3493 * - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
3494 * - +infinity, then +infinity will be returned
3495 * - NaN, then NaN will be returned
3497 * Those checks are fairly expensive so if you don't need them make sure
3498 * handle_edge_cases is false.
3501 lp_build_log2_approx(struct lp_build_context
*bld
,
3503 LLVMValueRef
*p_exp
,
3504 LLVMValueRef
*p_floor_log2
,
3505 LLVMValueRef
*p_log2
,
3506 boolean handle_edge_cases
)
3508 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3509 const struct lp_type type
= bld
->type
;
3510 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->gallivm
, type
);
3511 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, type
);
3513 LLVMValueRef expmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x7f800000);
3514 LLVMValueRef mantmask
= lp_build_const_int_vec(bld
->gallivm
, type
, 0x007fffff);
3515 LLVMValueRef one
= LLVMConstBitCast(bld
->one
, int_vec_type
);
3517 LLVMValueRef i
= NULL
;
3518 LLVMValueRef y
= NULL
;
3519 LLVMValueRef z
= NULL
;
3520 LLVMValueRef exp
= NULL
;
3521 LLVMValueRef mant
= NULL
;
3522 LLVMValueRef logexp
= NULL
;
3523 LLVMValueRef p_z
= NULL
;
3524 LLVMValueRef res
= NULL
;
3526 assert(lp_check_value(bld
->type
, x
));
3528 if(p_exp
|| p_floor_log2
|| p_log2
) {
3529 /* TODO: optimize the constant case */
3530 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&&
3531 LLVMIsConstant(x
)) {
3532 debug_printf("%s: inefficient/imprecise constant arithmetic\n",
3536 assert(type
.floating
&& type
.width
== 32);
3539 * We don't explicitly handle denormalized numbers. They will yield a
3540 * result in the neighbourhood of -127, which appears to be adequate
3544 i
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3546 /* exp = (float) exponent(x) */
3547 exp
= LLVMBuildAnd(builder
, i
, expmask
, "");
3550 if(p_floor_log2
|| p_log2
) {
3551 logexp
= LLVMBuildLShr(builder
, exp
, lp_build_const_int_vec(bld
->gallivm
, type
, 23), "");
3552 logexp
= LLVMBuildSub(builder
, logexp
, lp_build_const_int_vec(bld
->gallivm
, type
, 127), "");
3553 logexp
= LLVMBuildSIToFP(builder
, logexp
, vec_type
, "");
3557 /* mant = 1 + (float) mantissa(x) */
3558 mant
= LLVMBuildAnd(builder
, i
, mantmask
, "");
3559 mant
= LLVMBuildOr(builder
, mant
, one
, "");
3560 mant
= LLVMBuildBitCast(builder
, mant
, vec_type
, "");
3562 /* y = (mant - 1) / (mant + 1) */
3563 y
= lp_build_div(bld
,
3564 lp_build_sub(bld
, mant
, bld
->one
),
3565 lp_build_add(bld
, mant
, bld
->one
)
3569 z
= lp_build_mul(bld
, y
, y
);
3572 p_z
= lp_build_polynomial(bld
, z
, lp_build_log2_polynomial
,
3573 ARRAY_SIZE(lp_build_log2_polynomial
));
3575 /* y * P(z) + logexp */
3576 res
= lp_build_mad(bld
, y
, p_z
, logexp
);
3578 if (type
.floating
&& handle_edge_cases
) {
3579 LLVMValueRef negmask
, infmask
, zmask
;
3580 negmask
= lp_build_cmp(bld
, PIPE_FUNC_LESS
, x
,
3581 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3582 zmask
= lp_build_cmp(bld
, PIPE_FUNC_EQUAL
, x
,
3583 lp_build_const_vec(bld
->gallivm
, type
, 0.0f
));
3584 infmask
= lp_build_cmp(bld
, PIPE_FUNC_GEQUAL
, x
,
3585 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
));
3587 /* If x is qual to inf make sure we return inf */
3588 res
= lp_build_select(bld
, infmask
,
3589 lp_build_const_vec(bld
->gallivm
, type
, INFINITY
),
3591 /* If x is qual to 0, return -inf */
3592 res
= lp_build_select(bld
, zmask
,
3593 lp_build_const_vec(bld
->gallivm
, type
, -INFINITY
),
3595 /* If x is nan or less than 0, return nan */
3596 res
= lp_build_select(bld
, negmask
,
3597 lp_build_const_vec(bld
->gallivm
, type
, NAN
),
3603 exp
= LLVMBuildBitCast(builder
, exp
, vec_type
, "");
3608 *p_floor_log2
= logexp
;
3616 * log2 implementation which doesn't have special code to
3617 * handle edge cases (-inf, 0, inf, NaN). It's faster but
3618 * the results for those cases are undefined.
3621 lp_build_log2(struct lp_build_context
*bld
,
3625 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, FALSE
);
3630 * Version of log2 which handles all edge cases.
3631 * Look at documentation of lp_build_log2_approx for
3632 * description of the behavior for each of the edge cases.
3635 lp_build_log2_safe(struct lp_build_context
*bld
,
3639 lp_build_log2_approx(bld
, x
, NULL
, NULL
, &res
, TRUE
);
3645 * Faster (and less accurate) log2.
3647 * log2(x) = floor(log2(x)) - 1 + x / 2**floor(log2(x))
3649 * Piece-wise linear approximation, with exact results when x is a
3652 * See http://www.flipcode.com/archives/Fast_log_Function.shtml
3655 lp_build_fast_log2(struct lp_build_context
*bld
,
3658 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3662 assert(lp_check_value(bld
->type
, x
));
3664 assert(bld
->type
.floating
);
3666 /* ipart = floor(log2(x)) - 1 */
3667 ipart
= lp_build_extract_exponent(bld
, x
, -1);
3668 ipart
= LLVMBuildSIToFP(builder
, ipart
, bld
->vec_type
, "");
3670 /* fpart = x / 2**ipart */
3671 fpart
= lp_build_extract_mantissa(bld
, x
);
3674 return LLVMBuildFAdd(builder
, ipart
, fpart
, "");
3679 * Fast implementation of iround(log2(x)).
3681 * Not an approximation -- it should give accurate results all the time.
3684 lp_build_ilog2(struct lp_build_context
*bld
,
3687 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3688 LLVMValueRef sqrt2
= lp_build_const_vec(bld
->gallivm
, bld
->type
, M_SQRT2
);
3691 assert(bld
->type
.floating
);
3693 assert(lp_check_value(bld
->type
, x
));
3695 /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
3696 x
= LLVMBuildFMul(builder
, x
, sqrt2
, "");
3698 /* ipart = floor(log2(x) + 0.5) */
3699 ipart
= lp_build_extract_exponent(bld
, x
, 0);
3705 lp_build_mod(struct lp_build_context
*bld
,
3709 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3711 const struct lp_type type
= bld
->type
;
3713 assert(lp_check_value(type
, x
));
3714 assert(lp_check_value(type
, y
));
3717 res
= LLVMBuildFRem(builder
, x
, y
, "");
3719 res
= LLVMBuildSRem(builder
, x
, y
, "");
3721 res
= LLVMBuildURem(builder
, x
, y
, "");
3727 * For floating inputs it creates and returns a mask
3728 * which is all 1's for channels which are NaN.
3729 * Channels inside x which are not NaN will be 0.
3732 lp_build_isnan(struct lp_build_context
*bld
,
3736 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3738 assert(bld
->type
.floating
);
3739 assert(lp_check_value(bld
->type
, x
));
3741 mask
= LLVMBuildFCmp(bld
->gallivm
->builder
, LLVMRealOEQ
, x
, x
,
3743 mask
= LLVMBuildNot(bld
->gallivm
->builder
, mask
, "");
3744 mask
= LLVMBuildSExt(bld
->gallivm
->builder
, mask
, int_vec_type
, "isnan");
3748 /* Returns all 1's for floating point numbers that are
3749 * finite numbers and returns all zeros for -inf,
3752 lp_build_isfinite(struct lp_build_context
*bld
,
3755 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
3756 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, bld
->type
);
3757 struct lp_type int_type
= lp_int_type(bld
->type
);
3758 LLVMValueRef intx
= LLVMBuildBitCast(builder
, x
, int_vec_type
, "");
3759 LLVMValueRef infornan32
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
3762 if (!bld
->type
.floating
) {
3763 return lp_build_const_int_vec(bld
->gallivm
, bld
->type
, 0);
3765 assert(bld
->type
.floating
);
3766 assert(lp_check_value(bld
->type
, x
));
3767 assert(bld
->type
.width
== 32);
3769 intx
= LLVMBuildAnd(builder
, intx
, infornan32
, "");
3770 return lp_build_compare(bld
->gallivm
, int_type
, PIPE_FUNC_NOTEQUAL
,
3775 * Returns true if the number is nan or inf and false otherwise.
3776 * The input has to be a floating point vector.
3779 lp_build_is_inf_or_nan(struct gallivm_state
*gallivm
,
3780 const struct lp_type type
,
3783 LLVMBuilderRef builder
= gallivm
->builder
;
3784 struct lp_type int_type
= lp_int_type(type
);
3785 LLVMValueRef const0
= lp_build_const_int_vec(gallivm
, int_type
,
3789 assert(type
.floating
);
3791 ret
= LLVMBuildBitCast(builder
, x
, lp_build_vec_type(gallivm
, int_type
), "");
3792 ret
= LLVMBuildAnd(builder
, ret
, const0
, "");
3793 ret
= lp_build_compare(gallivm
, int_type
, PIPE_FUNC_EQUAL
,
3801 lp_build_fpstate_get(struct gallivm_state
*gallivm
)
3803 if (util_cpu_caps
.has_sse
) {
3804 LLVMBuilderRef builder
= gallivm
->builder
;
3805 LLVMValueRef mxcsr_ptr
= lp_build_alloca(
3807 LLVMInt32TypeInContext(gallivm
->context
),
3809 LLVMValueRef mxcsr_ptr8
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3810 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3811 lp_build_intrinsic(builder
,
3812 "llvm.x86.sse.stmxcsr",
3813 LLVMVoidTypeInContext(gallivm
->context
),
3821 lp_build_fpstate_set_denorms_zero(struct gallivm_state
*gallivm
,
3824 if (util_cpu_caps
.has_sse
) {
3825 /* turn on DAZ (64) | FTZ (32768) = 32832 if available */
3826 int daz_ftz
= _MM_FLUSH_ZERO_MASK
;
3828 LLVMBuilderRef builder
= gallivm
->builder
;
3829 LLVMValueRef mxcsr_ptr
= lp_build_fpstate_get(gallivm
);
3830 LLVMValueRef mxcsr
=
3831 LLVMBuildLoad(builder
, mxcsr_ptr
, "mxcsr");
3833 if (util_cpu_caps
.has_daz
) {
3834 /* Enable denormals are zero mode */
3835 daz_ftz
|= _MM_DENORMALS_ZERO_MASK
;
3838 mxcsr
= LLVMBuildOr(builder
, mxcsr
,
3839 LLVMConstInt(LLVMTypeOf(mxcsr
), daz_ftz
, 0), "");
3841 mxcsr
= LLVMBuildAnd(builder
, mxcsr
,
3842 LLVMConstInt(LLVMTypeOf(mxcsr
), ~daz_ftz
, 0), "");
3845 LLVMBuildStore(builder
, mxcsr
, mxcsr_ptr
);
3846 lp_build_fpstate_set(gallivm
, mxcsr_ptr
);
3851 lp_build_fpstate_set(struct gallivm_state
*gallivm
,
3852 LLVMValueRef mxcsr_ptr
)
3854 if (util_cpu_caps
.has_sse
) {
3855 LLVMBuilderRef builder
= gallivm
->builder
;
3856 mxcsr_ptr
= LLVMBuildPointerCast(builder
, mxcsr_ptr
,
3857 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), "");
3858 lp_build_intrinsic(builder
,
3859 "llvm.x86.sse.ldmxcsr",
3860 LLVMVoidTypeInContext(gallivm
->context
),