1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
31 * Helper functions for packing/unpacking.
33 * Pack/unpacking is necessary for conversion between types of different
36 * They are also commonly used when an computation needs higher
37 * precision for the intermediate values. For example, if one needs the
42 * to use more precision for intermediate results then one should implement it
46 * compute(LLVMBuilderRef builder struct lp_type type, LLVMValueRef a, LLVMValueRef b)
48 * struct lp_type wide_type = lp_wider_type(type);
49 * LLVMValueRef al, ah, bl, bh, cl, ch, c;
51 * lp_build_unpack2(builder, type, wide_type, a, &al, &ah);
52 * lp_build_unpack2(builder, type, wide_type, b, &bl, &bh);
54 * cl = compute_half(al, bl);
55 * ch = compute_half(ah, bh);
57 * c = lp_build_pack2(bld->builder, wide_type, type, cl, ch);
62 * where compute_half() would do the computation for half the elements with
63 * twice the precision.
65 * @author Jose Fonseca <jfonseca@vmware.com>
69 #include "util/u_debug.h"
70 #include "util/u_math.h"
71 #include "util/u_cpu_detect.h"
72 #include "util/u_memory.h"
74 #include "lp_bld_type.h"
75 #include "lp_bld_const.h"
76 #include "lp_bld_init.h"
77 #include "lp_bld_intr.h"
78 #include "lp_bld_arit.h"
79 #include "lp_bld_pack.h"
80 #include "lp_bld_swizzle.h"
84 * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions.
87 lp_build_const_unpack_shuffle(struct gallivm_state
*gallivm
,
88 unsigned n
, unsigned lo_hi
)
90 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
93 assert(n
<= LP_MAX_VECTOR_LENGTH
);
96 /* TODO: cache results in a static table */
98 for(i
= 0, j
= lo_hi
*n
/2; i
< n
; i
+= 2, ++j
) {
99 elems
[i
+ 0] = lp_build_const_int32(gallivm
, 0 + j
);
100 elems
[i
+ 1] = lp_build_const_int32(gallivm
, n
+ j
);
103 return LLVMConstVector(elems
, n
);
107 * Similar to lp_build_const_unpack_shuffle but for special AVX 256bit unpack.
108 * See comment above lp_build_interleave2_half for more details.
111 lp_build_const_unpack_shuffle_half(struct gallivm_state
*gallivm
,
112 unsigned n
, unsigned lo_hi
)
114 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
117 assert(n
<= LP_MAX_VECTOR_LENGTH
);
120 for (i
= 0, j
= lo_hi
*(n
/4); i
< n
; i
+= 2, ++j
) {
124 elems
[i
+ 0] = lp_build_const_int32(gallivm
, 0 + j
);
125 elems
[i
+ 1] = lp_build_const_int32(gallivm
, n
+ j
);
128 return LLVMConstVector(elems
, n
);
132 * Build shuffle vectors that match PACKxx (SSE) instructions or
136 lp_build_const_pack_shuffle(struct gallivm_state
*gallivm
, unsigned n
)
138 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
141 assert(n
<= LP_MAX_VECTOR_LENGTH
);
143 for(i
= 0; i
< n
; ++i
)
144 #ifdef PIPE_ARCH_LITTLE_ENDIAN
145 elems
[i
] = lp_build_const_int32(gallivm
, 2*i
);
147 elems
[i
] = lp_build_const_int32(gallivm
, 2*i
+1);
150 return LLVMConstVector(elems
, n
);
154 * Return a vector with elements src[start:start+size]
155 * Most useful for getting half the values out of a 256bit sized vector,
156 * otherwise may cause data rearrangement to happen.
159 lp_build_extract_range(struct gallivm_state
*gallivm
,
164 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
167 assert(size
<= Elements(elems
));
169 for (i
= 0; i
< size
; ++i
)
170 elems
[i
] = lp_build_const_int32(gallivm
, i
+ start
);
173 return LLVMBuildExtractElement(gallivm
->builder
, src
, elems
[0], "");
176 return LLVMBuildShuffleVector(gallivm
->builder
, src
, src
,
177 LLVMConstVector(elems
, size
), "");
182 * Concatenates several (must be a power of 2) vectors (of same type)
184 * Most useful for building up a 256bit sized vector out of two 128bit ones.
187 lp_build_concat(struct gallivm_state
*gallivm
,
189 struct lp_type src_type
,
190 unsigned num_vectors
)
192 unsigned new_length
, i
;
193 LLVMValueRef tmp
[LP_MAX_VECTOR_LENGTH
/2];
194 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
196 assert(src_type
.length
* num_vectors
<= Elements(shuffles
));
197 assert(util_is_power_of_two(num_vectors
));
199 new_length
= src_type
.length
;
201 for (i
= 0; i
< num_vectors
; i
++)
204 while (num_vectors
> 1) {
207 for (i
= 0; i
< new_length
; i
++) {
208 shuffles
[i
] = lp_build_const_int32(gallivm
, i
);
210 for (i
= 0; i
< num_vectors
; i
++) {
211 tmp
[i
] = LLVMBuildShuffleVector(gallivm
->builder
, tmp
[i
*2], tmp
[i
*2 + 1],
212 LLVMConstVector(shuffles
, new_length
), "");
221 * Combines vectors to reduce from num_srcs to num_dsts.
222 * Returns the number of src vectors concatenated in a single dst.
224 * num_srcs must be exactly divisible by num_dsts.
226 * e.g. For num_srcs = 4 and src = [x, y, z, w]
227 * num_dsts = 1 dst = [xyzw] return = 4
228 * num_dsts = 2 dst = [xy, zw] return = 2
231 lp_build_concat_n(struct gallivm_state
*gallivm
,
232 struct lp_type src_type
,
238 int size
= num_srcs
/ num_dsts
;
241 assert(num_srcs
>= num_dsts
);
242 assert((num_srcs
% size
) == 0);
244 if (num_srcs
== num_dsts
) {
245 for (i
= 0; i
< num_dsts
; ++i
) {
251 for (i
= 0; i
< num_dsts
; ++i
) {
252 dst
[i
] = lp_build_concat(gallivm
, &src
[i
* size
], src_type
, size
);
260 * Interleave vector elements.
262 * Matches the PUNPCKLxx and PUNPCKHxx SSE instructions
263 * (but not for 256bit AVX vectors).
266 lp_build_interleave2(struct gallivm_state
*gallivm
,
272 LLVMValueRef shuffle
;
274 if (type
.length
== 2 && type
.width
== 128 && util_cpu_caps
.has_avx
) {
276 * XXX: This is a workaround for llvm code generation deficiency. Strangely
277 * enough, while this needs vinsertf128/vextractf128 instructions (hence
278 * a natural match when using 2x128bit vectors) the "normal" unpack shuffle
279 * generates code ranging from atrocious (llvm 3.1) to terrible (llvm 3.2, 3.3).
280 * So use some different shuffles instead (the exact shuffles don't seem to
281 * matter, as long as not using 128bit wide vectors, works with 8x32 or 4x64).
283 struct lp_type tmp_type
= type
;
284 LLVMValueRef srchalf
[2], tmpdst
;
287 a
= LLVMBuildBitCast(gallivm
->builder
, a
, lp_build_vec_type(gallivm
, tmp_type
), "");
288 b
= LLVMBuildBitCast(gallivm
->builder
, b
, lp_build_vec_type(gallivm
, tmp_type
), "");
289 srchalf
[0] = lp_build_extract_range(gallivm
, a
, lo_hi
* 2, 2);
290 srchalf
[1] = lp_build_extract_range(gallivm
, b
, lo_hi
* 2, 2);
292 tmpdst
= lp_build_concat(gallivm
, srchalf
, tmp_type
, 2);
293 return LLVMBuildBitCast(gallivm
->builder
, tmpdst
, lp_build_vec_type(gallivm
, type
), "");
296 shuffle
= lp_build_const_unpack_shuffle(gallivm
, type
.length
, lo_hi
);
298 return LLVMBuildShuffleVector(gallivm
->builder
, a
, b
, shuffle
, "");
302 * Interleave vector elements but with 256 bit,
303 * treats it as interleave with 2 concatenated 128 bit vectors.
305 * This differs to lp_build_interleave2 as that function would do the following (for lo):
306 * a0 b0 a1 b1 a2 b2 a3 b3, and this does not compile into an AVX unpack instruction.
309 * An example interleave 8x float with 8x float on AVX 256bit unpack:
310 * a0 a1 a2 a3 a4 a5 a6 a7 <-> b0 b1 b2 b3 b4 b5 b6 b7
312 * Equivalent to interleaving 2x 128 bit vectors
313 * a0 a1 a2 a3 <-> b0 b1 b2 b3 concatenated with a4 a5 a6 a7 <-> b4 b5 b6 b7
315 * So interleave-lo would result in:
316 * a0 b0 a1 b1 a4 b4 a5 b5
318 * And interleave-hi would result in:
319 * a2 b2 a3 b3 a6 b6 a7 b7
322 lp_build_interleave2_half(struct gallivm_state
*gallivm
,
328 if (type
.length
* type
.width
== 256) {
329 LLVMValueRef shuffle
= lp_build_const_unpack_shuffle_half(gallivm
, type
.length
, lo_hi
);
330 return LLVMBuildShuffleVector(gallivm
->builder
, a
, b
, shuffle
, "");
332 return lp_build_interleave2(gallivm
, type
, a
, b
, lo_hi
);
337 * Double the bit width.
339 * This will only change the number of bits the values are represented, not the
343 lp_build_unpack2(struct gallivm_state
*gallivm
,
344 struct lp_type src_type
,
345 struct lp_type dst_type
,
347 LLVMValueRef
*dst_lo
,
348 LLVMValueRef
*dst_hi
)
350 LLVMBuilderRef builder
= gallivm
->builder
;
352 LLVMTypeRef dst_vec_type
;
354 assert(!src_type
.floating
);
355 assert(!dst_type
.floating
);
356 assert(dst_type
.width
== src_type
.width
* 2);
357 assert(dst_type
.length
* 2 == src_type
.length
);
359 if(dst_type
.sign
&& src_type
.sign
) {
360 /* Replicate the sign bit in the most significant bits */
361 msb
= LLVMBuildAShr(builder
, src
, lp_build_const_int_vec(gallivm
, src_type
, src_type
.width
- 1), "");
364 /* Most significant bits always zero */
365 msb
= lp_build_zero(gallivm
, src_type
);
367 /* Interleave bits */
368 #ifdef PIPE_ARCH_LITTLE_ENDIAN
369 *dst_lo
= lp_build_interleave2(gallivm
, src_type
, src
, msb
, 0);
370 *dst_hi
= lp_build_interleave2(gallivm
, src_type
, src
, msb
, 1);
372 *dst_lo
= lp_build_interleave2(gallivm
, src_type
, msb
, src
, 0);
373 *dst_hi
= lp_build_interleave2(gallivm
, src_type
, msb
, src
, 1);
376 /* Cast the result into the new type (twice as wide) */
378 dst_vec_type
= lp_build_vec_type(gallivm
, dst_type
);
380 *dst_lo
= LLVMBuildBitCast(builder
, *dst_lo
, dst_vec_type
, "");
381 *dst_hi
= LLVMBuildBitCast(builder
, *dst_hi
, dst_vec_type
, "");
386 * Expand the bit width.
388 * This will only change the number of bits the values are represented, not the
392 lp_build_unpack(struct gallivm_state
*gallivm
,
393 struct lp_type src_type
,
394 struct lp_type dst_type
,
396 LLVMValueRef
*dst
, unsigned num_dsts
)
401 /* Register width must remain constant */
402 assert(src_type
.width
* src_type
.length
== dst_type
.width
* dst_type
.length
);
404 /* We must not loose or gain channels. Only precision */
405 assert(src_type
.length
== dst_type
.length
* num_dsts
);
410 while(src_type
.width
< dst_type
.width
) {
411 struct lp_type tmp_type
= src_type
;
414 tmp_type
.length
/= 2;
416 for(i
= num_tmps
; i
--; ) {
417 lp_build_unpack2(gallivm
, src_type
, tmp_type
, dst
[i
], &dst
[2*i
+ 0], &dst
[2*i
+ 1]);
425 assert(num_tmps
== num_dsts
);
430 * Non-interleaved pack.
432 * This will move values as
434 * lo = l0 __ l1 __ l2 __.. __ ln __
435 * hi = h0 __ h1 __ h2 __.. __ hn __
436 * res = l0 l1 l2 .. ln h0 h1 h2 .. hn
438 * This will only change the number of bits the values are represented, not the
441 * It is assumed the values are already clamped into the destination type range.
442 * Values outside that range will produce undefined results. Use
443 * lp_build_packs2 instead.
446 lp_build_pack2(struct gallivm_state
*gallivm
,
447 struct lp_type src_type
,
448 struct lp_type dst_type
,
452 LLVMBuilderRef builder
= gallivm
->builder
;
453 LLVMTypeRef dst_vec_type
= lp_build_vec_type(gallivm
, dst_type
);
454 LLVMValueRef shuffle
;
455 LLVMValueRef res
= NULL
;
456 struct lp_type intr_type
= dst_type
;
458 #if HAVE_LLVM < 0x0207
459 intr_type
= src_type
;
462 assert(!src_type
.floating
);
463 assert(!dst_type
.floating
);
464 assert(src_type
.width
== dst_type
.width
* 2);
465 assert(src_type
.length
* 2 == dst_type
.length
);
467 /* Check for special cases first */
468 if((util_cpu_caps
.has_sse2
|| util_cpu_caps
.has_altivec
) &&
469 src_type
.width
* src_type
.length
>= 128) {
470 const char *intrinsic
= NULL
;
472 switch(src_type
.width
) {
474 if (util_cpu_caps
.has_sse2
) {
476 intrinsic
= "llvm.x86.sse2.packssdw.128";
479 if (util_cpu_caps
.has_sse4_1
) {
480 intrinsic
= "llvm.x86.sse41.packusdw";
481 #if HAVE_LLVM < 0x0207
482 /* llvm < 2.7 has inconsistent signatures except for packusdw */
483 intr_type
= dst_type
;
487 } else if (util_cpu_caps
.has_altivec
) {
489 intrinsic
= "llvm.ppc.altivec.vpkswus";
491 intrinsic
= "llvm.ppc.altivec.vpkuwus";
497 if (util_cpu_caps
.has_sse2
) {
498 intrinsic
= "llvm.x86.sse2.packsswb.128";
499 } else if (util_cpu_caps
.has_altivec
) {
500 intrinsic
= "llvm.ppc.altivec.vpkshss";
503 if (util_cpu_caps
.has_sse2
) {
504 intrinsic
= "llvm.x86.sse2.packuswb.128";
505 } else if (util_cpu_caps
.has_altivec
) {
506 intrinsic
= "llvm.ppc.altivec.vpkshus";
510 /* default uses generic shuffle below */
513 if (src_type
.width
* src_type
.length
== 128) {
514 LLVMTypeRef intr_vec_type
= lp_build_vec_type(gallivm
, intr_type
);
515 res
= lp_build_intrinsic_binary(builder
, intrinsic
, intr_vec_type
, lo
, hi
);
516 if (dst_vec_type
!= intr_vec_type
) {
517 res
= LLVMBuildBitCast(builder
, res
, dst_vec_type
, "");
521 int num_split
= src_type
.width
* src_type
.length
/ 128;
523 int nlen
= 128 / src_type
.width
;
524 struct lp_type ndst_type
= lp_type_unorm(dst_type
.width
, 128);
525 struct lp_type nintr_type
= lp_type_unorm(intr_type
.width
, 128);
526 LLVMValueRef tmpres
[LP_MAX_VECTOR_WIDTH
/ 128];
527 LLVMValueRef tmplo
, tmphi
;
528 LLVMTypeRef ndst_vec_type
= lp_build_vec_type(gallivm
, ndst_type
);
529 LLVMTypeRef nintr_vec_type
= lp_build_vec_type(gallivm
, nintr_type
);
531 assert(num_split
<= LP_MAX_VECTOR_WIDTH
/ 128);
533 for (i
= 0; i
< num_split
/ 2; i
++) {
534 tmplo
= lp_build_extract_range(gallivm
,
536 tmphi
= lp_build_extract_range(gallivm
,
537 lo
, i
*nlen
*2 + nlen
, nlen
);
538 tmpres
[i
] = lp_build_intrinsic_binary(builder
, intrinsic
,
539 nintr_vec_type
, tmplo
, tmphi
);
540 if (ndst_vec_type
!= nintr_vec_type
) {
541 tmpres
[i
] = LLVMBuildBitCast(builder
, tmpres
[i
], ndst_vec_type
, "");
544 for (i
= 0; i
< num_split
/ 2; i
++) {
545 tmplo
= lp_build_extract_range(gallivm
,
547 tmphi
= lp_build_extract_range(gallivm
,
548 hi
, i
*nlen
*2 + nlen
, nlen
);
549 tmpres
[i
+num_split
/2] = lp_build_intrinsic_binary(builder
, intrinsic
,
552 if (ndst_vec_type
!= nintr_vec_type
) {
553 tmpres
[i
+num_split
/2] = LLVMBuildBitCast(builder
, tmpres
[i
+num_split
/2],
557 res
= lp_build_concat(gallivm
, tmpres
, ndst_type
, num_split
);
563 /* generic shuffle */
564 lo
= LLVMBuildBitCast(builder
, lo
, dst_vec_type
, "");
565 hi
= LLVMBuildBitCast(builder
, hi
, dst_vec_type
, "");
567 shuffle
= lp_build_const_pack_shuffle(gallivm
, dst_type
.length
);
569 res
= LLVMBuildShuffleVector(builder
, lo
, hi
, shuffle
, "");
577 * Non-interleaved pack and saturate.
579 * Same as lp_build_pack2 but will saturate values so that they fit into the
583 lp_build_packs2(struct gallivm_state
*gallivm
,
584 struct lp_type src_type
,
585 struct lp_type dst_type
,
591 assert(!src_type
.floating
);
592 assert(!dst_type
.floating
);
593 assert(src_type
.sign
== dst_type
.sign
);
594 assert(src_type
.width
== dst_type
.width
* 2);
595 assert(src_type
.length
* 2 == dst_type
.length
);
599 /* All X86 SSE non-interleaved pack instructions take signed inputs and
600 * saturate them, so no need to clamp for those cases. */
601 if(util_cpu_caps
.has_sse2
&&
602 src_type
.width
* src_type
.length
>= 128 &&
604 (src_type
.width
== 32 || src_type
.width
== 16))
608 struct lp_build_context bld
;
609 unsigned dst_bits
= dst_type
.sign
? dst_type
.width
- 1 : dst_type
.width
;
610 LLVMValueRef dst_max
= lp_build_const_int_vec(gallivm
, src_type
, ((unsigned long long)1 << dst_bits
) - 1);
611 lp_build_context_init(&bld
, gallivm
, src_type
);
612 lo
= lp_build_min(&bld
, lo
, dst_max
);
613 hi
= lp_build_min(&bld
, hi
, dst_max
);
614 /* FIXME: What about lower bound? */
617 return lp_build_pack2(gallivm
, src_type
, dst_type
, lo
, hi
);
622 * Truncate the bit width.
624 * TODO: Handle saturation consistently.
627 lp_build_pack(struct gallivm_state
*gallivm
,
628 struct lp_type src_type
,
629 struct lp_type dst_type
,
631 const LLVMValueRef
*src
, unsigned num_srcs
)
633 LLVMValueRef (*pack2
)(struct gallivm_state
*gallivm
,
634 struct lp_type src_type
,
635 struct lp_type dst_type
,
638 LLVMValueRef tmp
[LP_MAX_VECTOR_LENGTH
];
641 /* Register width must remain constant */
642 assert(src_type
.width
* src_type
.length
== dst_type
.width
* dst_type
.length
);
644 /* We must not loose or gain channels. Only precision */
645 assert(src_type
.length
* num_srcs
== dst_type
.length
);
648 pack2
= &lp_build_pack2
;
650 pack2
= &lp_build_packs2
;
652 for(i
= 0; i
< num_srcs
; ++i
)
655 while(src_type
.width
> dst_type
.width
) {
656 struct lp_type tmp_type
= src_type
;
659 tmp_type
.length
*= 2;
661 /* Take in consideration the sign changes only in the last step */
662 if(tmp_type
.width
== dst_type
.width
)
663 tmp_type
.sign
= dst_type
.sign
;
667 for(i
= 0; i
< num_srcs
; ++i
)
668 tmp
[i
] = pack2(gallivm
, src_type
, tmp_type
,
669 tmp
[2*i
+ 0], tmp
[2*i
+ 1]);
674 assert(num_srcs
== 1);
681 * Truncate or expand the bitwidth.
683 * NOTE: Getting the right sign flags is crucial here, as we employ some
684 * intrinsics that do saturation.
687 lp_build_resize(struct gallivm_state
*gallivm
,
688 struct lp_type src_type
,
689 struct lp_type dst_type
,
690 const LLVMValueRef
*src
, unsigned num_srcs
,
691 LLVMValueRef
*dst
, unsigned num_dsts
)
693 LLVMBuilderRef builder
= gallivm
->builder
;
694 LLVMValueRef tmp
[LP_MAX_VECTOR_LENGTH
];
698 * We don't support float <-> int conversion here. That must be done
699 * before/after calling this function.
701 assert(src_type
.floating
== dst_type
.floating
);
704 * We don't support double <-> float conversion yet, although it could be
705 * added with little effort.
707 assert((!src_type
.floating
&& !dst_type
.floating
) ||
708 src_type
.width
== dst_type
.width
);
710 /* We must not loose or gain channels. Only precision */
711 assert(src_type
.length
* num_srcs
== dst_type
.length
* num_dsts
);
713 /* We don't support M:N conversion, only 1:N, M:1, or 1:1 */
714 assert(num_srcs
== 1 || num_dsts
== 1);
716 assert(src_type
.length
<= LP_MAX_VECTOR_LENGTH
);
717 assert(dst_type
.length
<= LP_MAX_VECTOR_LENGTH
);
718 assert(num_srcs
<= LP_MAX_VECTOR_LENGTH
);
719 assert(num_dsts
<= LP_MAX_VECTOR_LENGTH
);
721 if (src_type
.width
> dst_type
.width
) {
723 * Truncate bit width.
726 assert(num_dsts
== 1);
728 if (src_type
.width
* src_type
.length
== dst_type
.width
* dst_type
.length
) {
730 * Register width remains constant -- use vector packing intrinsics
732 tmp
[0] = lp_build_pack(gallivm
, src_type
, dst_type
, TRUE
, src
, num_srcs
);
735 if (src_type
.width
/ dst_type
.width
> num_srcs
) {
737 * First change src vectors size (with shuffle) so they have the
738 * same size as the destination vector, then pack normally.
739 * Note: cannot use cast/extract because llvm generates atrocious code.
741 unsigned size_ratio
= (src_type
.width
* src_type
.length
) /
742 (dst_type
.length
* dst_type
.width
);
743 unsigned new_length
= src_type
.length
/ size_ratio
;
745 for (i
= 0; i
< size_ratio
* num_srcs
; i
++) {
746 unsigned start_index
= (i
% size_ratio
) * new_length
;
747 tmp
[i
] = lp_build_extract_range(gallivm
, src
[i
/ size_ratio
],
748 start_index
, new_length
);
750 num_srcs
*= size_ratio
;
751 src_type
.length
= new_length
;
752 tmp
[0] = lp_build_pack(gallivm
, src_type
, dst_type
, TRUE
, tmp
, num_srcs
);
756 * Truncate bit width but expand vector size - first pack
757 * then expand simply because this should be more AVX-friendly
758 * for the cases we probably hit.
760 unsigned size_ratio
= (dst_type
.width
* dst_type
.length
) /
761 (src_type
.length
* src_type
.width
);
762 unsigned num_pack_srcs
= num_srcs
/ size_ratio
;
763 dst_type
.length
= dst_type
.length
/ size_ratio
;
765 for (i
= 0; i
< size_ratio
; i
++) {
766 tmp
[i
] = lp_build_pack(gallivm
, src_type
, dst_type
, TRUE
,
767 &src
[i
*num_pack_srcs
], num_pack_srcs
);
769 tmp
[0] = lp_build_concat(gallivm
, tmp
, dst_type
, size_ratio
);
773 else if (src_type
.width
< dst_type
.width
) {
778 assert(num_srcs
== 1);
780 if (src_type
.width
* src_type
.length
== dst_type
.width
* dst_type
.length
) {
782 * Register width remains constant -- use vector unpack intrinsics
784 lp_build_unpack(gallivm
, src_type
, dst_type
, src
[0], tmp
, num_dsts
);
788 * Do it element-wise.
790 assert(src_type
.length
* num_srcs
== dst_type
.length
* num_dsts
);
792 for (i
= 0; i
< num_dsts
; i
++) {
793 tmp
[i
] = lp_build_undef(gallivm
, dst_type
);
796 for (i
= 0; i
< src_type
.length
; ++i
) {
797 unsigned j
= i
/ dst_type
.length
;
798 LLVMValueRef srcindex
= lp_build_const_int32(gallivm
, i
);
799 LLVMValueRef dstindex
= lp_build_const_int32(gallivm
, i
% dst_type
.length
);
800 LLVMValueRef val
= LLVMBuildExtractElement(builder
, src
[0], srcindex
, "");
802 if (src_type
.sign
&& dst_type
.sign
) {
803 val
= LLVMBuildSExt(builder
, val
, lp_build_elem_type(gallivm
, dst_type
), "");
805 val
= LLVMBuildZExt(builder
, val
, lp_build_elem_type(gallivm
, dst_type
), "");
807 tmp
[j
] = LLVMBuildInsertElement(builder
, tmp
[j
], val
, dstindex
, "");
816 assert(num_srcs
== 1);
817 assert(num_dsts
== 1);
822 for(i
= 0; i
< num_dsts
; ++i
)
828 * Expands src vector from src.length to dst_length
831 lp_build_pad_vector(struct gallivm_state
*gallivm
,
835 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
838 unsigned i
, src_length
;
840 type
= LLVMTypeOf(src
);
842 if (LLVMGetTypeKind(type
) != LLVMVectorTypeKind
) {
843 /* Can't use ShuffleVector on non-vector type */
844 undef
= LLVMGetUndef(LLVMVectorType(type
, dst_length
));
845 return LLVMBuildInsertElement(gallivm
->builder
, undef
, src
, lp_build_const_int32(gallivm
, 0), "");
848 undef
= LLVMGetUndef(type
);
849 src_length
= LLVMGetVectorSize(type
);
851 assert(dst_length
<= Elements(elems
));
852 assert(dst_length
>= src_length
);
854 if (src_length
== dst_length
)
857 /* All elements from src vector */
858 for (i
= 0; i
< src_length
; ++i
)
859 elems
[i
] = lp_build_const_int32(gallivm
, i
);
861 /* Undef fill remaining space */
862 for (i
= src_length
; i
< dst_length
; ++i
)
863 elems
[i
] = lp_build_const_int32(gallivm
, src_length
);
865 /* Combine the two vectors */
866 return LLVMBuildShuffleVector(gallivm
->builder
, src
, undef
, LLVMConstVector(elems
, dst_length
), "");