1 /**************************************************************************
3 * Copyright 2010-2018 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 SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
18 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
19 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
26 **************************************************************************/
31 * s3tc pixel format manipulation.
33 * @author Roland Scheidegger <sroland@vmware.com>
37 #include "util/u_format.h"
38 #include "util/u_math.h"
39 #include "util/u_string.h"
40 #include "util/u_cpu_detect.h"
41 #include "util/u_debug.h"
43 #include "lp_bld_arit.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_gather.h"
48 #include "lp_bld_format.h"
49 #include "lp_bld_logic.h"
50 #include "lp_bld_pack.h"
51 #include "lp_bld_flow.h"
52 #include "lp_bld_printf.h"
53 #include "lp_bld_struct.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_init.h"
56 #include "lp_bld_debug.h"
57 #include "lp_bld_intr.h"
61 * Reverse an interleave2_half
62 * (ie. pick every second element, independent lower/upper halfs)
63 * sse2 can only do that with 32bit (shufps) or larger elements
64 * natively. (Otherwise, and/pack (even) or shift/pack (odd)
65 * could be used, ideally llvm would do that for us.)
66 * XXX: Unfortunately, this does NOT translate to a shufps if those
67 * are int vectors (and casting will not help, llvm needs to recognize it
68 * as "real" float). Instead, llvm will use a pshufd/pshufd/punpcklqdq
69 * sequence which I'm pretty sure is a lot worse despite domain transition
70 * penalties with shufps (except maybe on Nehalem).
73 lp_build_uninterleave2_half(struct gallivm_state
*gallivm
,
79 LLVMValueRef shuffle
, elems
[LP_MAX_VECTOR_LENGTH
];
82 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
85 if (type
.length
* type
.width
== 256) {
86 assert(type
.length
== 8);
87 assert(type
.width
== 32);
88 static const unsigned shufvals
[8] = {0, 2, 8, 10, 4, 6, 12, 14};
89 for (i
= 0; i
< type
.length
; ++i
) {
90 elems
[i
] = lp_build_const_int32(gallivm
, shufvals
[i
] + lo_hi
);
93 for (i
= 0; i
< type
.length
; ++i
) {
94 elems
[i
] = lp_build_const_int32(gallivm
, 2*i
+ lo_hi
);
98 shuffle
= LLVMConstVector(elems
, type
.length
);
100 return LLVMBuildShuffleVector(gallivm
->builder
, a
, b
, shuffle
, "");
106 * Build shuffle for extending vectors.
109 lp_build_const_extend_shuffle(struct gallivm_state
*gallivm
,
110 unsigned n
, unsigned length
)
112 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
116 assert(length
<= LP_MAX_VECTOR_LENGTH
);
118 /* TODO: cache results in a static table */
120 for(i
= 0; i
< n
; i
++) {
121 elems
[i
] = lp_build_const_int32(gallivm
, i
);
123 for (i
= n
; i
< length
; i
++) {
124 elems
[i
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
127 return LLVMConstVector(elems
, length
);
131 lp_build_const_unpackx2_shuffle(struct gallivm_state
*gallivm
, unsigned n
)
133 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
136 assert(n
<= LP_MAX_VECTOR_LENGTH
);
138 /* TODO: cache results in a static table */
140 for(i
= 0, j
= 0; i
< n
; i
+= 2, ++j
) {
141 elems
[i
+ 0] = lp_build_const_int32(gallivm
, 0 + j
);
142 elems
[i
+ 1] = lp_build_const_int32(gallivm
, n
+ j
);
143 elems
[n
+ i
+ 0] = lp_build_const_int32(gallivm
, 0 + n
/2 + j
);
144 elems
[n
+ i
+ 1] = lp_build_const_int32(gallivm
, n
+ n
/2 + j
);
147 return LLVMConstVector(elems
, n
* 2);
151 * broadcast 1 element to all elements
154 lp_build_const_shuffle1(struct gallivm_state
*gallivm
,
155 unsigned index
, unsigned n
)
157 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
160 assert(n
<= LP_MAX_VECTOR_LENGTH
);
162 /* TODO: cache results in a static table */
164 for (i
= 0; i
< n
; i
++) {
165 elems
[i
] = lp_build_const_int32(gallivm
, index
);
168 return LLVMConstVector(elems
, n
);
172 * move 1 element to pos 0, rest undef
175 lp_build_shuffle1undef(struct gallivm_state
*gallivm
,
176 LLVMValueRef a
, unsigned index
, unsigned n
)
178 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
], shuf
;
181 assert(n
<= LP_MAX_VECTOR_LENGTH
);
183 elems
[0] = lp_build_const_int32(gallivm
, index
);
185 for (i
= 1; i
< n
; i
++) {
186 elems
[i
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
188 shuf
= LLVMConstVector(elems
, n
);
190 return LLVMBuildShuffleVector(gallivm
->builder
, a
, a
, shuf
, "");
194 format_dxt1_variant(enum pipe_format format
)
196 return format
== PIPE_FORMAT_DXT1_RGB
||
197 format
== PIPE_FORMAT_DXT1_RGBA
||
198 format
== PIPE_FORMAT_DXT1_SRGB
||
199 format
== PIPE_FORMAT_DXT1_SRGBA
;
204 * Gather elements from scatter positions in memory into vectors.
205 * This is customised for fetching texels from s3tc textures.
206 * For SSE, typical value is length=4.
208 * @param length length of the offsets
209 * @param colors the stored colors of the blocks will be extracted into this.
210 * @param codewords the codewords of the blocks will be extracted into this.
211 * @param alpha_lo used for storing lower 32bit of alpha components for dxt3/5
212 * @param alpha_hi used for storing higher 32bit of alpha components for dxt3/5
213 * @param base_ptr base pointer, should be a i8 pointer type.
214 * @param offsets vector with offsets
217 lp_build_gather_s3tc(struct gallivm_state
*gallivm
,
219 const struct util_format_description
*format_desc
,
220 LLVMValueRef
*colors
,
221 LLVMValueRef
*codewords
,
222 LLVMValueRef
*alpha_lo
,
223 LLVMValueRef
*alpha_hi
,
224 LLVMValueRef base_ptr
,
225 LLVMValueRef offsets
)
227 LLVMBuilderRef builder
= gallivm
->builder
;
228 unsigned block_bits
= format_desc
->block
.bits
;
230 LLVMValueRef elems
[8];
231 LLVMTypeRef type32
= LLVMInt32TypeInContext(gallivm
->context
);
232 LLVMTypeRef type64
= LLVMInt64TypeInContext(gallivm
->context
);
233 LLVMTypeRef type32dxt
;
234 struct lp_type lp_type32dxt
;
236 memset(&lp_type32dxt
, 0, sizeof lp_type32dxt
);
237 lp_type32dxt
.width
= 32;
238 lp_type32dxt
.length
= block_bits
/ 32;
239 type32dxt
= lp_build_vec_type(gallivm
, lp_type32dxt
);
241 assert(block_bits
== 64 || block_bits
== 128);
242 assert(length
== 1 || length
== 4 || length
== 8);
244 for (i
= 0; i
< length
; ++i
) {
245 elems
[i
] = lp_build_gather_elem(gallivm
, length
,
246 block_bits
, block_bits
, TRUE
,
247 base_ptr
, offsets
, i
, FALSE
);
248 elems
[i
] = LLVMBuildBitCast(builder
, elems
[i
], type32dxt
, "");
251 LLVMValueRef elem
= elems
[0];
252 if (block_bits
== 128) {
253 *alpha_lo
= LLVMBuildExtractElement(builder
, elem
,
254 lp_build_const_int32(gallivm
, 0), "");
255 *alpha_hi
= LLVMBuildExtractElement(builder
, elem
,
256 lp_build_const_int32(gallivm
, 1), "");
257 *colors
= LLVMBuildExtractElement(builder
, elem
,
258 lp_build_const_int32(gallivm
, 2), "");
259 *codewords
= LLVMBuildExtractElement(builder
, elem
,
260 lp_build_const_int32(gallivm
, 3), "");
263 *alpha_lo
= LLVMGetUndef(type32
);
264 *alpha_hi
= LLVMGetUndef(type32
);
265 *colors
= LLVMBuildExtractElement(builder
, elem
,
266 lp_build_const_int32(gallivm
, 0), "");
267 *codewords
= LLVMBuildExtractElement(builder
, elem
,
268 lp_build_const_int32(gallivm
, 1), "");
272 LLVMValueRef tmp
[4], cc01
, cc23
;
273 struct lp_type lp_type32
, lp_type64
;
274 memset(&lp_type32
, 0, sizeof lp_type32
);
275 lp_type32
.width
= 32;
276 lp_type32
.length
= length
;
277 memset(&lp_type64
, 0, sizeof lp_type64
);
278 lp_type64
.width
= 64;
279 lp_type64
.length
= length
/2;
281 if (block_bits
== 128) {
283 for (i
= 0; i
< 4; ++i
) {
286 elems
[i
] = lp_build_concat(gallivm
, tmp
, lp_type32dxt
, 2);
289 lp_build_transpose_aos(gallivm
, lp_type32
, elems
, tmp
);
295 LLVMTypeRef type64_vec
= LLVMVectorType(type64
, length
/2);
296 LLVMTypeRef type32_vec
= LLVMVectorType(type32
, length
);
298 for (i
= 0; i
< length
; ++i
) {
300 elems
[i
] = LLVMBuildShuffleVector(builder
, elems
[i
],
301 LLVMGetUndef(type32dxt
),
302 lp_build_const_extend_shuffle(gallivm
, 2, 4), "");
305 struct lp_type lp_type32_4
= {0};
306 lp_type32_4
.width
= 32;
307 lp_type32_4
.length
= 4;
308 for (i
= 0; i
< 4; ++i
) {
311 elems
[i
] = lp_build_concat(gallivm
, tmp
, lp_type32_4
, 2);
314 cc01
= lp_build_interleave2_half(gallivm
, lp_type32
, elems
[0], elems
[1], 0);
315 cc23
= lp_build_interleave2_half(gallivm
, lp_type32
, elems
[2], elems
[3], 0);
316 cc01
= LLVMBuildBitCast(builder
, cc01
, type64_vec
, "");
317 cc23
= LLVMBuildBitCast(builder
, cc23
, type64_vec
, "");
318 *colors
= lp_build_interleave2_half(gallivm
, lp_type64
, cc01
, cc23
, 0);
319 *codewords
= lp_build_interleave2_half(gallivm
, lp_type64
, cc01
, cc23
, 1);
320 *colors
= LLVMBuildBitCast(builder
, *colors
, type32_vec
, "");
321 *codewords
= LLVMBuildBitCast(builder
, *codewords
, type32_vec
, "");
326 /** Convert from <n x i32> containing 2 x n rgb565 colors
327 * to 2 <n x i32> rgba8888 colors
328 * This is the most optimized version I can think of
329 * should be nearly as fast as decoding only one color
330 * NOTE: alpha channel will be set to 0
331 * @param colors is a <n x i32> vector containing the rgb565 colors
334 color_expand2_565_to_8888(struct gallivm_state
*gallivm
,
337 LLVMValueRef
*color0
,
338 LLVMValueRef
*color1
)
340 LLVMBuilderRef builder
= gallivm
->builder
;
341 LLVMValueRef r
, g
, b
, rblo
, glo
;
342 LLVMValueRef rgblomask
, rb
, rgb0
, rgb1
;
343 struct lp_type type
, type16
, type8
;
347 memset(&type
, 0, sizeof type
);
351 memset(&type16
, 0, sizeof type16
);
353 type16
.length
= 2 * n
;
355 memset(&type8
, 0, sizeof type8
);
357 type8
.length
= 4 * n
;
359 rgblomask
= lp_build_const_int_vec(gallivm
, type16
, 0x0707);
360 colors
= LLVMBuildBitCast(builder
, colors
,
361 lp_build_vec_type(gallivm
, type16
), "");
362 /* move r into low 8 bits, b into high 8 bits, g into another reg (low bits)
363 * make sure low bits of r are zero - could use AND but requires constant */
364 r
= LLVMBuildLShr(builder
, colors
, lp_build_const_int_vec(gallivm
, type16
, 11), "");
365 r
= LLVMBuildShl(builder
, r
, lp_build_const_int_vec(gallivm
, type16
, 3), "");
366 b
= LLVMBuildShl(builder
, colors
, lp_build_const_int_vec(gallivm
, type16
, 11), "");
367 rb
= LLVMBuildOr(builder
, r
, b
, "");
368 rblo
= LLVMBuildLShr(builder
, rb
, lp_build_const_int_vec(gallivm
, type16
, 5), "");
369 /* don't have byte shift hence need mask */
370 rblo
= LLVMBuildAnd(builder
, rblo
, rgblomask
, "");
371 rb
= LLVMBuildOr(builder
, rb
, rblo
, "");
373 /* make sure low bits of g are zero */
374 g
= LLVMBuildAnd(builder
, colors
, lp_build_const_int_vec(gallivm
, type16
, 0x07e0), "");
375 g
= LLVMBuildLShr(builder
, g
, lp_build_const_int_vec(gallivm
, type16
, 3), "");
376 glo
= LLVMBuildLShr(builder
, g
, lp_build_const_int_vec(gallivm
, type16
, 6), "");
377 g
= LLVMBuildOr(builder
, g
, glo
, "");
379 rb
= LLVMBuildBitCast(builder
, rb
, lp_build_vec_type(gallivm
, type8
), "");
380 g
= LLVMBuildBitCast(builder
, g
, lp_build_vec_type(gallivm
, type8
), "");
381 rgb0
= lp_build_interleave2_half(gallivm
, type8
, rb
, g
, 0);
382 rgb1
= lp_build_interleave2_half(gallivm
, type8
, rb
, g
, 1);
384 rgb0
= LLVMBuildBitCast(builder
, rgb0
, lp_build_vec_type(gallivm
, type
), "");
385 rgb1
= LLVMBuildBitCast(builder
, rgb1
, lp_build_vec_type(gallivm
, type
), "");
387 /* rgb0 is rgb00, rgb01, rgb10, rgb11
388 * instead of rgb00, rgb10, rgb20, rgb30 hence need reshuffle
389 * on x86 this _should_ just generate one shufps...
391 *color0
= lp_build_uninterleave2_half(gallivm
, type
, rgb0
, rgb1
, 0);
392 *color1
= lp_build_uninterleave2_half(gallivm
, type
, rgb0
, rgb1
, 1);
396 /** Convert from <n x i32> containing rgb565 colors
397 * (in first 16 bits) to <n x i32> rgba8888 colors
399 * NOTE: alpha channel will be set to 0
400 * @param colors is a <n x i32> vector containing the rgb565 colors
403 color_expand_565_to_8888(struct gallivm_state
*gallivm
,
407 LLVMBuilderRef builder
= gallivm
->builder
;
408 LLVMValueRef rgba
, r
, g
, b
, rgblo
, glo
;
409 LLVMValueRef rbhimask
, g6mask
, rgblomask
;
411 memset(&type
, 0, sizeof type
);
416 * first extract and shift colors into their final locations
417 * (high bits - low bits zero at this point)
418 * then replicate highest bits to the lowest bits
419 * note rb replication can be done in parallel but not g
421 * r5mask = 0xf800, g6mask = 0x07e0, b5mask = 0x001f
422 * rhigh = 8, ghigh = 5, bhigh = 19
423 * rblow = 5, glow = 6
424 * rgblowmask = 0x00070307
425 * r = colors >> rhigh
426 * b = colors << bhigh
427 * g = (colors & g6mask) << ghigh
428 * rb = (r | b) rbhimask
429 * rbtmp = rb >> rblow
431 * rbtmp = rbtmp | gtmp
432 * rbtmp = rbtmp & rgblowmask
433 * rgb = rb | g | rbtmp
435 g6mask
= lp_build_const_int_vec(gallivm
, type
, 0x07e0);
436 rbhimask
= lp_build_const_int_vec(gallivm
, type
, 0x00f800f8);
437 rgblomask
= lp_build_const_int_vec(gallivm
, type
, 0x00070307);
439 r
= LLVMBuildLShr(builder
, colors
, lp_build_const_int_vec(gallivm
, type
, 8), "");
440 b
= LLVMBuildShl(builder
, colors
, lp_build_const_int_vec(gallivm
, type
, 19), "");
441 g
= LLVMBuildAnd(builder
, colors
, g6mask
, "");
442 g
= LLVMBuildShl(builder
, g
, lp_build_const_int_vec(gallivm
, type
, 5), "");
443 rgba
= LLVMBuildOr(builder
, r
, b
, "");
444 rgba
= LLVMBuildAnd(builder
, rgba
, rbhimask
, "");
445 rgblo
= LLVMBuildLShr(builder
, rgba
, lp_build_const_int_vec(gallivm
, type
, 5), "");
446 glo
= LLVMBuildLShr(builder
, g
, lp_build_const_int_vec(gallivm
, type
, 6), "");
447 rgblo
= LLVMBuildOr(builder
, rgblo
, glo
, "");
448 rgblo
= LLVMBuildAnd(builder
, rgblo
, rgblomask
, "");
449 rgba
= LLVMBuildOr(builder
, rgba
, g
, "");
450 rgba
= LLVMBuildOr(builder
, rgba
, rgblo
, "");
457 * Average two byte vectors. (Will always round up.)
460 lp_build_pavgb(struct lp_build_context
*bld8
,
464 struct gallivm_state
*gallivm
= bld8
->gallivm
;
465 LLVMBuilderRef builder
= gallivm
->builder
;
466 assert(bld8
->type
.width
== 8);
467 assert(bld8
->type
.length
== 16 || bld8
->type
.length
== 32);
468 if (HAVE_LLVM
< 0x0600) {
469 LLVMValueRef intrargs
[2];
470 char *intr_name
= bld8
->type
.length
== 32 ? "llvm.x86.avx2.pavg.b" :
471 "llvm.x86.sse2.pavg.b";
474 return lp_build_intrinsic(builder
, intr_name
,
475 bld8
->vec_type
, intrargs
, 2, 0);
478 * Must match llvm's autoupgrade of pavg.b intrinsic to be useful.
479 * You better hope the backend code manages to detect the pattern, and
480 * the pattern doesn't change there...
482 struct lp_type type_ext
= bld8
->type
;
483 LLVMTypeRef vec_type_ext
;
485 LLVMValueRef ext_one
;
487 vec_type_ext
= lp_build_vec_type(gallivm
, type_ext
);
488 ext_one
= lp_build_const_vec(gallivm
, type_ext
, 1);
490 v0
= LLVMBuildZExt(builder
, v0
, vec_type_ext
, "");
491 v1
= LLVMBuildZExt(builder
, v1
, vec_type_ext
, "");
492 res
= LLVMBuildAdd(builder
, v0
, v1
, "");
493 res
= LLVMBuildAdd(builder
, res
, ext_one
, "");
494 res
= LLVMBuildLShr(builder
, res
, ext_one
, "");
495 res
= LLVMBuildTrunc(builder
, res
, bld8
->vec_type
, "");
501 * Calculate 1/3(v1-v0) + v0
502 * and 2*1/3(v1-v0) + v0
505 lp_build_lerp23(struct lp_build_context
*bld
,
511 struct gallivm_state
*gallivm
= bld
->gallivm
;
512 LLVMValueRef x
, x_lo
, x_hi
, delta_lo
, delta_hi
;
513 LLVMValueRef mul_lo
, mul_hi
, v0_lo
, v0_hi
, v1_lo
, v1_hi
, tmp
;
514 const struct lp_type type
= bld
->type
;
515 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
516 struct lp_type i16_type
= lp_wider_type(type
);
517 struct lp_build_context bld2
;
519 assert(lp_check_value(type
, v0
));
520 assert(lp_check_value(type
, v1
));
521 assert(!type
.floating
&& !type
.fixed
&& !type
.norm
&& type
.width
== 8);
523 lp_build_context_init(&bld2
, gallivm
, i16_type
);
524 bld2
.type
.sign
= TRUE
;
525 x
= lp_build_const_int_vec(gallivm
, bld
->type
, 255*1/3);
527 /* FIXME: use native avx256 unpack/pack */
528 lp_build_unpack2(gallivm
, type
, i16_type
, x
, &x_lo
, &x_hi
);
529 lp_build_unpack2(gallivm
, type
, i16_type
, v0
, &v0_lo
, &v0_hi
);
530 lp_build_unpack2(gallivm
, type
, i16_type
, v1
, &v1_lo
, &v1_hi
);
531 delta_lo
= lp_build_sub(&bld2
, v1_lo
, v0_lo
);
532 delta_hi
= lp_build_sub(&bld2
, v1_hi
, v0_hi
);
534 mul_lo
= LLVMBuildMul(builder
, x_lo
, delta_lo
, "");
535 mul_hi
= LLVMBuildMul(builder
, x_hi
, delta_hi
, "");
537 x_lo
= LLVMBuildLShr(builder
, mul_lo
, lp_build_const_int_vec(gallivm
, i16_type
, 8), "");
538 x_hi
= LLVMBuildLShr(builder
, mul_hi
, lp_build_const_int_vec(gallivm
, i16_type
, 8), "");
539 /* lerp optimization: pack now, do add afterwards */
540 tmp
= lp_build_pack2(gallivm
, i16_type
, type
, x_lo
, x_hi
);
541 *res0
= lp_build_add(bld
, tmp
, v0
);
543 x_lo
= LLVMBuildLShr(builder
, mul_lo
, lp_build_const_int_vec(gallivm
, i16_type
, 7), "");
544 x_hi
= LLVMBuildLShr(builder
, mul_hi
, lp_build_const_int_vec(gallivm
, i16_type
, 7), "");
545 /* unlike above still need mask (but add still afterwards). */
546 x_lo
= LLVMBuildAnd(builder
, x_lo
, lp_build_const_int_vec(gallivm
, i16_type
, 0xff), "");
547 x_hi
= LLVMBuildAnd(builder
, x_hi
, lp_build_const_int_vec(gallivm
, i16_type
, 0xff), "");
548 tmp
= lp_build_pack2(gallivm
, i16_type
, type
, x_lo
, x_hi
);
549 *res1
= lp_build_add(bld
, tmp
, v0
);
553 * Convert from <n x i64> s3tc dxt1 to <4n x i8> RGBA AoS
554 * @param colors is a <n x i32> vector with n x 2x16bit colors
555 * @param codewords is a <n x i32> vector containing the codewords
556 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
557 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
560 s3tc_dxt1_full_to_rgba_aos(struct gallivm_state
*gallivm
,
562 enum pipe_format format
,
564 LLVMValueRef codewords
,
568 LLVMBuilderRef builder
= gallivm
->builder
;
569 LLVMValueRef color0
, color1
, color2
, color3
, color2_2
, color3_2
;
570 LLVMValueRef rgba
, a
, colors0
, colors1
, col0
, col1
, const2
;
571 LLVMValueRef bit_pos
, sel_mask
, sel_lo
, sel_hi
, indices
;
572 struct lp_type type
, type8
;
573 struct lp_build_context bld8
, bld32
;
574 boolean is_dxt1_variant
= format_dxt1_variant(format
);
576 memset(&type
, 0, sizeof type
);
580 memset(&type8
, 0, sizeof type8
);
584 assert(lp_check_value(type
, i
));
585 assert(lp_check_value(type
, j
));
587 a
= lp_build_const_int_vec(gallivm
, type
, 0xff000000);
589 lp_build_context_init(&bld32
, gallivm
, type
);
590 lp_build_context_init(&bld8
, gallivm
, type8
);
594 * - expand color0/color1 to rgba8888
595 * - calculate color2/3 (interpolation) according to color0 < color1 rules
596 * - calculate color2/3 according to color0 >= color1 rules
597 * - do selection of color2/3 according to comparison of color0/1
598 * - extract indices (vector shift).
599 * - use compare/select to select the correct color. Since we have 2bit
600 * indices (and 4 colors), needs at least three compare/selects.
603 * expand the two colors
605 col0
= LLVMBuildAnd(builder
, colors
, lp_build_const_int_vec(gallivm
, type
, 0x0000ffff), "");
606 col1
= LLVMBuildLShr(builder
, colors
, lp_build_const_int_vec(gallivm
, type
, 16), "");
608 color_expand2_565_to_8888(gallivm
, n
, colors
, &color0
, &color1
);
611 color0
= color_expand_565_to_8888(gallivm
, n
, col0
);
612 color1
= color_expand_565_to_8888(gallivm
, n
, col1
);
617 * color2_1 is 2/3 color0 + 1/3 color1
618 * color3_1 is 1/3 color0 + 2/3 color1
619 * color2_2 is 1/2 color0 + 1/2 color1
623 colors0
= LLVMBuildBitCast(builder
, color0
, bld8
.vec_type
, "");
624 colors1
= LLVMBuildBitCast(builder
, color1
, bld8
.vec_type
, "");
625 /* can combine 2 lerps into one mostly - still looks expensive enough. */
626 lp_build_lerp23(&bld8
, colors0
, colors1
, &color2
, &color3
);
627 color2
= LLVMBuildBitCast(builder
, color2
, bld32
.vec_type
, "");
628 color3
= LLVMBuildBitCast(builder
, color3
, bld32
.vec_type
, "");
630 /* dxt3/5 always use 4-color encoding */
631 if (is_dxt1_variant
) {
633 if (format
== PIPE_FORMAT_DXT1_RGBA
||
634 format
== PIPE_FORMAT_DXT1_SRGBA
) {
635 color0
= LLVMBuildOr(builder
, color0
, a
, "");
636 color1
= LLVMBuildOr(builder
, color1
, a
, "");
637 color3
= LLVMBuildOr(builder
, color3
, a
, "");
640 * XXX with sse2 and 16x8 vectors, should use pavgb even when n == 1.
641 * Much cheaper (but we don't care that much if n == 1).
643 if ((util_cpu_caps
.has_sse2
&& n
== 4) ||
644 (util_cpu_caps
.has_avx2
&& n
== 8)) {
645 color2_2
= lp_build_pavgb(&bld8
, colors0
, colors1
);
646 color2_2
= LLVMBuildBitCast(builder
, color2_2
, bld32
.vec_type
, "");
649 struct lp_type i16_type
= lp_wider_type(type8
);
650 struct lp_build_context bld2
;
651 LLVMValueRef v0_lo
, v0_hi
, v1_lo
, v1_hi
, addlo
, addhi
;
653 lp_build_context_init(&bld2
, gallivm
, i16_type
);
654 bld2
.type
.sign
= TRUE
;
657 * This isn't as expensive as it looks (the unpack is the same as
658 * for lerp23), with correct rounding.
659 * (Note that while rounding is correct, this will always round down,
660 * whereas pavgb will always round up.)
662 /* FIXME: use native avx256 unpack/pack */
663 lp_build_unpack2(gallivm
, type8
, i16_type
, colors0
, &v0_lo
, &v0_hi
);
664 lp_build_unpack2(gallivm
, type8
, i16_type
, colors1
, &v1_lo
, &v1_hi
);
666 addlo
= lp_build_add(&bld2
, v0_lo
, v1_lo
);
667 addhi
= lp_build_add(&bld2
, v0_hi
, v1_hi
);
668 addlo
= LLVMBuildLShr(builder
, addlo
,
669 lp_build_const_int_vec(gallivm
, i16_type
, 1), "");
670 addhi
= LLVMBuildLShr(builder
, addhi
,
671 lp_build_const_int_vec(gallivm
, i16_type
, 1), "");
672 color2_2
= lp_build_pack2(gallivm
, i16_type
, type8
, addlo
, addhi
);
673 color2_2
= LLVMBuildBitCast(builder
, color2_2
, bld32
.vec_type
, "");
675 color3_2
= lp_build_const_int_vec(gallivm
, type
, 0);
677 /* select between colors2/3 */
678 /* signed compare is faster saves some xors */
680 sel_mask
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, col0
, col1
);
681 color2
= lp_build_select(&bld32
, sel_mask
, color2
, color2_2
);
682 color3
= lp_build_select(&bld32
, sel_mask
, color3
, color3_2
);
685 if (format
== PIPE_FORMAT_DXT1_RGBA
||
686 format
== PIPE_FORMAT_DXT1_SRGBA
) {
687 color2
= LLVMBuildOr(builder
, color2
, a
, "");
691 const2
= lp_build_const_int_vec(gallivm
, type
, 2);
692 /* extract 2-bit index values */
693 bit_pos
= LLVMBuildShl(builder
, j
, const2
, "");
694 bit_pos
= LLVMBuildAdd(builder
, bit_pos
, i
, "");
695 bit_pos
= LLVMBuildAdd(builder
, bit_pos
, bit_pos
, "");
697 * NOTE: This innocent looking shift is very expensive with x86/ssex.
698 * Shifts with per-elemnent shift count get roughly translated to
699 * extract (count), extract (value), shift, move (back to xmm), unpack
701 * So about 20 instructions here for 4xi32.
702 * Newer llvm versions (3.7+) will not do extract/insert but use a
703 * a couple constant count vector shifts plus shuffles. About same
704 * amount of instructions unfortunately...
705 * Would get much worse with 8xi16 even...
706 * We could actually do better here:
707 * - subtract bit_pos from 128+30, shl 23, convert float to int...
708 * - now do mul with codewords followed by shr 30...
709 * But requires 32bit->32bit mul, sse41 only (well that's emulatable
710 * with 2 32bit->64bit muls...) and not exactly cheap
711 * AVX2, of course, fixes this nonsense.
713 indices
= LLVMBuildLShr(builder
, codewords
, bit_pos
, "");
715 /* finally select the colors */
716 sel_lo
= LLVMBuildAnd(builder
, indices
, bld32
.one
, "");
717 sel_lo
= lp_build_compare(gallivm
, type
, PIPE_FUNC_EQUAL
, sel_lo
, bld32
.one
);
718 color0
= lp_build_select(&bld32
, sel_lo
, color1
, color0
);
719 color2
= lp_build_select(&bld32
, sel_lo
, color3
, color2
);
720 sel_hi
= LLVMBuildAnd(builder
, indices
, const2
, "");
721 sel_hi
= lp_build_compare(gallivm
, type
, PIPE_FUNC_EQUAL
, sel_hi
, const2
);
722 rgba
= lp_build_select(&bld32
, sel_hi
, color2
, color0
);
725 if (format
== PIPE_FORMAT_DXT1_RGB
||
726 format
== PIPE_FORMAT_DXT1_SRGB
) {
727 rgba
= LLVMBuildOr(builder
, rgba
, a
, "");
729 return LLVMBuildBitCast(builder
, rgba
, bld8
.vec_type
, "");
734 s3tc_dxt1_to_rgba_aos(struct gallivm_state
*gallivm
,
736 enum pipe_format format
,
738 LLVMValueRef codewords
,
742 return s3tc_dxt1_full_to_rgba_aos(gallivm
, n
, format
,
743 colors
, codewords
, i
, j
);
748 * Convert from <n x i128> s3tc dxt3 to <4n x i8> RGBA AoS
749 * @param colors is a <n x i32> vector with n x 2x16bit colors
750 * @param codewords is a <n x i32> vector containing the codewords
751 * @param alphas is a <n x i64> vector containing the alpha values
752 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
753 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
756 s3tc_dxt3_to_rgba_aos(struct gallivm_state
*gallivm
,
758 enum pipe_format format
,
760 LLVMValueRef codewords
,
761 LLVMValueRef alpha_low
,
762 LLVMValueRef alpha_hi
,
766 LLVMBuilderRef builder
= gallivm
->builder
;
767 LLVMValueRef rgba
, tmp
, tmp2
;
768 LLVMValueRef bit_pos
, sel_mask
;
769 struct lp_type type
, type8
;
770 struct lp_build_context bld
;
772 memset(&type
, 0, sizeof type
);
776 memset(&type8
, 0, sizeof type8
);
780 assert(lp_check_value(type
, i
));
781 assert(lp_check_value(type
, j
));
783 lp_build_context_init(&bld
, gallivm
, type
);
785 rgba
= s3tc_dxt1_to_rgba_aos(gallivm
, n
, format
,
786 colors
, codewords
, i
, j
);
788 rgba
= LLVMBuildBitCast(builder
, rgba
, bld
.vec_type
, "");
791 * Extract alpha values. Since we now need to select from
792 * which 32bit vector values are fetched, construct selection
793 * mask from highest bit of bit_pos, and use select, then shift
794 * according to the bit_pos (without the highest bit).
795 * Note this is pointless for n == 1 case. Could just
796 * directly use 64bit arithmetic if we'd extract 64bit
797 * alpha value instead of 2x32...
800 bit_pos
= LLVMBuildShl(builder
, j
, lp_build_const_int_vec(gallivm
, type
, 2), "");
801 bit_pos
= LLVMBuildAdd(builder
, bit_pos
, i
, "");
802 bit_pos
= LLVMBuildShl(builder
, bit_pos
,
803 lp_build_const_int_vec(gallivm
, type
, 2), "");
804 sel_mask
= LLVMBuildLShr(builder
, bit_pos
,
805 lp_build_const_int_vec(gallivm
, type
, 5), "");
806 sel_mask
= LLVMBuildSub(builder
, sel_mask
, bld
.one
, "");
807 tmp
= lp_build_select(&bld
, sel_mask
, alpha_low
, alpha_hi
);
808 bit_pos
= LLVMBuildAnd(builder
, bit_pos
,
809 lp_build_const_int_vec(gallivm
, type
, 0xffffffdf), "");
810 /* Warning: slow shift with per element count (without avx2) */
812 * Could do pshufb here as well - just use appropriate 2 bits in bit_pos
813 * to select the right byte with pshufb. Then for the remaining one bit
814 * just do shift/select.
816 tmp
= LLVMBuildLShr(builder
, tmp
, bit_pos
, "");
818 /* combined expand from a4 to a8 and shift into position */
819 tmp
= LLVMBuildShl(builder
, tmp
, lp_build_const_int_vec(gallivm
, type
, 28), "");
820 tmp2
= LLVMBuildLShr(builder
, tmp
, lp_build_const_int_vec(gallivm
, type
, 4), "");
821 tmp
= LLVMBuildOr(builder
, tmp
, tmp2
, "");
823 rgba
= LLVMBuildOr(builder
, tmp
, rgba
, "");
825 return LLVMBuildBitCast(builder
, rgba
, lp_build_vec_type(gallivm
, type8
), "");
829 lp_build_lerpdxta(struct gallivm_state
*gallivm
,
833 LLVMValueRef sel_mask
,
837 * note we're doing lerp in 16bit since 32bit pmulld is only available in sse41
838 * (plus pmullw is actually faster...)
839 * we just pretend our 32bit values (which are really only 8bit) are 16bits.
840 * Note that this is obviously a disaster for the scalar case.
842 LLVMBuilderRef builder
= gallivm
->builder
;
843 LLVMValueRef delta
, ainterp
;
844 LLVMValueRef weight5
, weight7
, weight
;
845 struct lp_type type32
, type16
, type8
;
846 struct lp_build_context bld16
;
848 memset(&type32
, 0, sizeof type32
);
851 memset(&type16
, 0, sizeof type16
);
855 memset(&type8
, 0, sizeof type8
);
859 lp_build_context_init(&bld16
, gallivm
, type16
);
860 /* 255/7 is a bit off - increase accuracy at the expense of shift later */
861 sel_mask
= LLVMBuildBitCast(builder
, sel_mask
, bld16
.vec_type
, "");
862 weight5
= lp_build_const_int_vec(gallivm
, type16
, 255*64/5);
863 weight7
= lp_build_const_int_vec(gallivm
, type16
, 255*64/7);
864 weight
= lp_build_select(&bld16
, sel_mask
, weight7
, weight5
);
866 alpha0
= LLVMBuildBitCast(builder
, alpha0
, bld16
.vec_type
, "");
867 alpha1
= LLVMBuildBitCast(builder
, alpha1
, bld16
.vec_type
, "");
868 code
= LLVMBuildBitCast(builder
, code
, bld16
.vec_type
, "");
869 /* we'll get garbage in the elements which had code 0 (or larger than 5 or 7)
871 code
= LLVMBuildSub(builder
, code
, bld16
.one
, "");
873 weight
= LLVMBuildMul(builder
, weight
, code
, "");
874 weight
= LLVMBuildLShr(builder
, weight
,
875 lp_build_const_int_vec(gallivm
, type16
, 6), "");
877 delta
= LLVMBuildSub(builder
, alpha1
, alpha0
, "");
879 ainterp
= LLVMBuildMul(builder
, delta
, weight
, "");
880 ainterp
= LLVMBuildLShr(builder
, ainterp
,
881 lp_build_const_int_vec(gallivm
, type16
, 8), "");
883 ainterp
= LLVMBuildBitCast(builder
, ainterp
, lp_build_vec_type(gallivm
, type8
), "");
884 alpha0
= LLVMBuildBitCast(builder
, alpha0
, lp_build_vec_type(gallivm
, type8
), "");
885 ainterp
= LLVMBuildAdd(builder
, alpha0
, ainterp
, "");
886 ainterp
= LLVMBuildBitCast(builder
, ainterp
, lp_build_vec_type(gallivm
, type32
), "");
892 * Convert from <n x i128> s3tc dxt5 to <4n x i8> RGBA AoS
893 * @param colors is a <n x i32> vector with n x 2x16bit colors
894 * @param codewords is a <n x i32> vector containing the codewords
895 * @param alphas is a <n x i64> vector containing the alpha values
896 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
897 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
900 s3tc_dxt5_full_to_rgba_aos(struct gallivm_state
*gallivm
,
902 enum pipe_format format
,
904 LLVMValueRef codewords
,
905 LLVMValueRef alpha_lo
,
906 LLVMValueRef alpha_hi
,
910 LLVMBuilderRef builder
= gallivm
->builder
;
911 LLVMValueRef rgba
, tmp
, alpha0
, alpha1
, alphac
, alphac0
, bit_pos
, shift
;
912 LLVMValueRef sel_mask
, tmp_mask
, alpha
, alpha64
, code_s
;
913 LLVMValueRef mask6
, mask7
, ainterp
;
914 LLVMTypeRef i64t
= LLVMInt64TypeInContext(gallivm
->context
);
915 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
916 struct lp_type type
, type8
;
917 struct lp_build_context bld32
;
919 memset(&type
, 0, sizeof type
);
923 memset(&type8
, 0, sizeof type8
);
927 assert(lp_check_value(type
, i
));
928 assert(lp_check_value(type
, j
));
930 lp_build_context_init(&bld32
, gallivm
, type
);
932 assert(lp_check_value(type
, i
));
933 assert(lp_check_value(type
, j
));
935 rgba
= s3tc_dxt1_to_rgba_aos(gallivm
, n
, format
,
936 colors
, codewords
, i
, j
);
938 rgba
= LLVMBuildBitCast(builder
, rgba
, bld32
.vec_type
, "");
940 /* this looks pretty complex for vectorization:
941 * extract a0/a1 values
943 * select weights for interpolation depending on a0 > a1
944 * mul weights by code - 1
946 * use selects for getting either a0, a1, interp a, interp a/0.0, interp a/1.0
949 alpha0
= LLVMBuildAnd(builder
, alpha_lo
,
950 lp_build_const_int_vec(gallivm
, type
, 0xff), "");
951 alpha1
= LLVMBuildLShr(builder
, alpha_lo
,
952 lp_build_const_int_vec(gallivm
, type
, 8), "");
953 alpha1
= LLVMBuildAnd(builder
, alpha1
,
954 lp_build_const_int_vec(gallivm
, type
, 0xff), "");
957 bit_pos
= LLVMBuildShl(builder
, j
, lp_build_const_int_vec(gallivm
, type
, 2), "");
958 bit_pos
= LLVMBuildAdd(builder
, bit_pos
, i
, "");
959 tmp
= LLVMBuildAdd(builder
, bit_pos
, bit_pos
, "");
960 bit_pos
= LLVMBuildAdd(builder
, bit_pos
, tmp
, "");
961 /* get rid of first 2 bytes - saves shifts of alpha_lo/hi */
962 bit_pos
= LLVMBuildAdd(builder
, bit_pos
,
963 lp_build_const_int_vec(gallivm
, type
, 16), "");
966 struct lp_type type64
;
967 memset(&type64
, 0, sizeof type64
);
970 /* This is pretty pointless could avoid by just directly extracting
971 64bit in the first place but makes it more complicated elsewhere */
972 alpha_lo
= LLVMBuildZExt(builder
, alpha_lo
, i64t
, "");
973 alpha_hi
= LLVMBuildZExt(builder
, alpha_hi
, i64t
, "");
974 alphac0
= LLVMBuildShl(builder
, alpha_hi
,
975 lp_build_const_int_vec(gallivm
, type64
, 32), "");
976 alphac0
= LLVMBuildOr(builder
, alpha_lo
, alphac0
, "");
978 shift
= LLVMBuildZExt(builder
, bit_pos
, i64t
, "");
979 alphac0
= LLVMBuildLShr(builder
, alphac0
, shift
, "");
980 alphac0
= LLVMBuildTrunc(builder
, alphac0
, i32t
, "");
981 alphac
= LLVMBuildAnd(builder
, alphac0
,
982 lp_build_const_int_vec(gallivm
, type
, 0x7), "");
986 * Using non-native vector length here (actually, with avx2 and
987 * n == 4 llvm will indeed expand to ymm regs...)
988 * At least newer llvm versions handle that ok.
989 * llvm 3.7+ will even handle the emulated 64bit shift with variable
990 * shift count without extraction (and it's actually easier to
991 * emulate than the 32bit one).
993 alpha64
= LLVMBuildShuffleVector(builder
, alpha_lo
, alpha_hi
,
994 lp_build_const_unpackx2_shuffle(gallivm
, n
), "");
996 alpha64
= LLVMBuildBitCast(builder
, alpha64
, LLVMVectorType(i64t
, n
), "");
997 shift
= LLVMBuildZExt(builder
, bit_pos
, LLVMVectorType(i64t
, n
), "");
998 alphac
= LLVMBuildLShr(builder
, alpha64
, shift
, "");
999 alphac
= LLVMBuildTrunc(builder
, alphac
, bld32
.vec_type
, "");
1001 alphac
= LLVMBuildAnd(builder
, alphac
,
1002 lp_build_const_int_vec(gallivm
, type
, 0x7), "");
1005 /* signed compare is faster saves some xors */
1007 /* alpha0 > alpha1 selection */
1008 sel_mask
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
,
1010 ainterp
= lp_build_lerpdxta(gallivm
, alpha0
, alpha1
, alphac
, sel_mask
, n
);
1013 * if a0 > a1 then we select a0 for case 0, a1 for case 1, interp otherwise.
1014 * else we select a0 for case 0, a1 for case 1,
1015 * interp for case 2-5, 00 for 6 and 0xff(ffffff) for 7
1016 * a = (c == 0) ? a0 : a1
1017 * a = (c > 1) ? ainterp : a
1018 * Finally handle case 6/7 for !(a0 > a1)
1019 * a = (!(a0 > a1) && c == 6) ? 0 : a (andnot with mask)
1020 * a = (!(a0 > a1) && c == 7) ? 0xffffffff : a (or with mask)
1022 tmp_mask
= lp_build_compare(gallivm
, type
, PIPE_FUNC_EQUAL
,
1023 alphac
, bld32
.zero
);
1024 alpha
= lp_build_select(&bld32
, tmp_mask
, alpha0
, alpha1
);
1025 tmp_mask
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
,
1027 alpha
= lp_build_select(&bld32
, tmp_mask
, ainterp
, alpha
);
1029 code_s
= LLVMBuildAnd(builder
, alphac
,
1030 LLVMBuildNot(builder
, sel_mask
, ""), "");
1031 mask6
= lp_build_compare(gallivm
, type
, PIPE_FUNC_EQUAL
,
1032 code_s
, lp_build_const_int_vec(gallivm
, type
, 6));
1033 mask7
= lp_build_compare(gallivm
, type
, PIPE_FUNC_EQUAL
,
1034 code_s
, lp_build_const_int_vec(gallivm
, type
, 7));
1035 alpha
= LLVMBuildAnd(builder
, alpha
, LLVMBuildNot(builder
, mask6
, ""), "");
1036 alpha
= LLVMBuildOr(builder
, alpha
, mask7
, "");
1038 alpha
= LLVMBuildShl(builder
, alpha
, lp_build_const_int_vec(gallivm
, type
, 24), "");
1039 rgba
= LLVMBuildOr(builder
, alpha
, rgba
, "");
1041 return LLVMBuildBitCast(builder
, rgba
, lp_build_vec_type(gallivm
, type8
), "");
1046 lp_build_gather_s3tc_simple_scalar(struct gallivm_state
*gallivm
,
1047 const struct util_format_description
*format_desc
,
1048 LLVMValueRef
*dxt_block
,
1051 LLVMBuilderRef builder
= gallivm
->builder
;
1052 unsigned block_bits
= format_desc
->block
.bits
;
1053 LLVMValueRef elem
, shuf
;
1054 LLVMTypeRef type32
= LLVMIntTypeInContext(gallivm
->context
, 32);
1055 LLVMTypeRef src_type
= LLVMIntTypeInContext(gallivm
->context
, block_bits
);
1056 LLVMTypeRef src_ptr_type
= LLVMPointerType(src_type
, 0);
1057 LLVMTypeRef type32_4
= LLVMVectorType(type32
, 4);
1059 assert(block_bits
== 64 || block_bits
== 128);
1061 ptr
= LLVMBuildBitCast(builder
, ptr
, src_ptr_type
, "");
1062 elem
= LLVMBuildLoad(builder
, ptr
, "");
1064 if (block_bits
== 128) {
1065 /* just return block as is */
1066 *dxt_block
= LLVMBuildBitCast(builder
, elem
, type32_4
, "");
1069 LLVMTypeRef type32_2
= LLVMVectorType(type32
, 2);
1070 shuf
= lp_build_const_extend_shuffle(gallivm
, 2, 4);
1071 elem
= LLVMBuildBitCast(builder
, elem
, type32_2
, "");
1072 *dxt_block
= LLVMBuildShuffleVector(builder
, elem
,
1073 LLVMGetUndef(type32_2
), shuf
, "");
1079 s3tc_store_cached_block(struct gallivm_state
*gallivm
,
1081 LLVMValueRef tag_value
,
1082 LLVMValueRef hash_index
,
1085 LLVMBuilderRef builder
= gallivm
->builder
;
1086 LLVMValueRef ptr
, indices
[3];
1087 LLVMTypeRef type_ptr4x32
;
1090 type_ptr4x32
= LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4), 0);
1091 indices
[0] = lp_build_const_int32(gallivm
, 0);
1092 indices
[1] = lp_build_const_int32(gallivm
, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS
);
1093 indices
[2] = hash_index
;
1094 ptr
= LLVMBuildGEP(builder
, cache
, indices
, ARRAY_SIZE(indices
), "");
1095 LLVMBuildStore(builder
, tag_value
, ptr
);
1097 indices
[1] = lp_build_const_int32(gallivm
, LP_BUILD_FORMAT_CACHE_MEMBER_DATA
);
1098 hash_index
= LLVMBuildMul(builder
, hash_index
,
1099 lp_build_const_int32(gallivm
, 16), "");
1100 for (count
= 0; count
< 4; count
++) {
1101 indices
[2] = hash_index
;
1102 ptr
= LLVMBuildGEP(builder
, cache
, indices
, ARRAY_SIZE(indices
), "");
1103 ptr
= LLVMBuildBitCast(builder
, ptr
, type_ptr4x32
, "");
1104 LLVMBuildStore(builder
, col
[count
], ptr
);
1105 hash_index
= LLVMBuildAdd(builder
, hash_index
,
1106 lp_build_const_int32(gallivm
, 4), "");
1111 s3tc_lookup_cached_pixel(struct gallivm_state
*gallivm
,
1115 LLVMBuilderRef builder
= gallivm
->builder
;
1116 LLVMValueRef member_ptr
, indices
[3];
1118 indices
[0] = lp_build_const_int32(gallivm
, 0);
1119 indices
[1] = lp_build_const_int32(gallivm
, LP_BUILD_FORMAT_CACHE_MEMBER_DATA
);
1121 member_ptr
= LLVMBuildGEP(builder
, ptr
, indices
, ARRAY_SIZE(indices
), "");
1122 return LLVMBuildLoad(builder
, member_ptr
, "cache_data");
1126 s3tc_lookup_tag_data(struct gallivm_state
*gallivm
,
1130 LLVMBuilderRef builder
= gallivm
->builder
;
1131 LLVMValueRef member_ptr
, indices
[3];
1133 indices
[0] = lp_build_const_int32(gallivm
, 0);
1134 indices
[1] = lp_build_const_int32(gallivm
, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS
);
1136 member_ptr
= LLVMBuildGEP(builder
, ptr
, indices
, ARRAY_SIZE(indices
), "");
1137 return LLVMBuildLoad(builder
, member_ptr
, "tag_data");
1140 #if LP_BUILD_FORMAT_CACHE_DEBUG
1142 s3tc_update_cache_access(struct gallivm_state
*gallivm
,
1147 LLVMBuilderRef builder
= gallivm
->builder
;
1148 LLVMValueRef member_ptr
, cache_access
;
1150 assert(index
== LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL
||
1151 index
== LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS
);
1153 member_ptr
= lp_build_struct_get_ptr(gallivm
, ptr
, index
, "");
1154 cache_access
= LLVMBuildLoad(builder
, member_ptr
, "cache_access");
1155 cache_access
= LLVMBuildAdd(builder
, cache_access
,
1156 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
),
1158 LLVMBuildStore(builder
, cache_access
, member_ptr
);
1163 * Calculate 1/3(v1-v0) + v0 and 2*1/3(v1-v0) + v0.
1164 * The lerp is performed between the first 2 32bit colors
1165 * in the source vector, both results are returned packed in result vector.
1168 lp_build_lerp23_single(struct lp_build_context
*bld
,
1171 struct gallivm_state
*gallivm
= bld
->gallivm
;
1172 LLVMValueRef x
, mul
, delta
, res
, v0
, v1
, elems
[8];
1173 const struct lp_type type
= bld
->type
;
1174 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1175 struct lp_type i16_type
= lp_wider_type(type
);
1176 struct lp_type i32_type
= lp_wider_type(i16_type
);
1177 struct lp_build_context bld2
;
1179 assert(!type
.floating
&& !type
.fixed
&& !type
.norm
&& type
.width
== 8);
1181 lp_build_context_init(&bld2
, gallivm
, i16_type
);
1182 bld2
.type
.sign
= TRUE
;
1184 /* weights 256/3, 256*2/3, with correct rounding */
1185 elems
[0] = elems
[1] = elems
[2] = elems
[3] =
1186 lp_build_const_elem(gallivm
, i16_type
, 255*1/3);
1187 elems
[4] = elems
[5] = elems
[6] = elems
[7] =
1188 lp_build_const_elem(gallivm
, i16_type
, 171);
1189 x
= LLVMConstVector(elems
, 8);
1192 * v01 has col0 in 32bit elem 0, col1 in elem 1.
1193 * Interleave/unpack will give us separate v0/v1 vectors.
1195 v01
= lp_build_interleave2(gallivm
, i32_type
, v01
, v01
, 0);
1196 v01
= LLVMBuildBitCast(builder
, v01
, bld
->vec_type
, "");
1198 lp_build_unpack2(gallivm
, type
, i16_type
, v01
, &v0
, &v1
);
1199 delta
= lp_build_sub(&bld2
, v1
, v0
);
1201 mul
= LLVMBuildMul(builder
, x
, delta
, "");
1203 mul
= LLVMBuildLShr(builder
, mul
, lp_build_const_int_vec(gallivm
, i16_type
, 8), "");
1204 /* lerp optimization: pack now, do add afterwards */
1205 res
= lp_build_pack2(gallivm
, i16_type
, type
, mul
, bld2
.undef
);
1206 /* only lower 2 elems are valid - for these v0 is really v0 */
1207 return lp_build_add(bld
, res
, v01
);
1211 * decode one dxt1 block.
1214 s3tc_decode_block_dxt1(struct gallivm_state
*gallivm
,
1215 enum pipe_format format
,
1216 LLVMValueRef dxt_block
,
1219 LLVMBuilderRef builder
= gallivm
->builder
;
1220 LLVMValueRef color01
, color23
, color01_16
, color0123
;
1221 LLVMValueRef rgba
, tmp
, a
, sel_mask
, indices
, code
, const2
;
1222 struct lp_type type8
, type32
, type16
, type64
;
1223 struct lp_build_context bld8
, bld32
, bld16
, bld64
;
1225 boolean is_dxt1_variant
= format_dxt1_variant(format
);
1227 memset(&type32
, 0, sizeof type32
);
1232 memset(&type8
, 0, sizeof type8
);
1236 memset(&type16
, 0, sizeof type16
);
1240 memset(&type64
, 0, sizeof type64
);
1244 a
= lp_build_const_int_vec(gallivm
, type32
, 0xff000000);
1245 const2
= lp_build_const_int_vec(gallivm
, type32
, 2);
1247 lp_build_context_init(&bld32
, gallivm
, type32
);
1248 lp_build_context_init(&bld16
, gallivm
, type16
);
1249 lp_build_context_init(&bld8
, gallivm
, type8
);
1250 lp_build_context_init(&bld64
, gallivm
, type64
);
1252 if (is_dxt1_variant
) {
1253 color01
= lp_build_shuffle1undef(gallivm
, dxt_block
, 0, 4);
1254 code
= lp_build_shuffle1undef(gallivm
, dxt_block
, 1, 4);
1256 color01
= lp_build_shuffle1undef(gallivm
, dxt_block
, 2, 4);
1257 code
= lp_build_shuffle1undef(gallivm
, dxt_block
, 3, 4);
1259 code
= LLVMBuildBitCast(builder
, code
, bld8
.vec_type
, "");
1260 /* expand bytes to dwords */
1261 code
= lp_build_interleave2(gallivm
, type8
, code
, code
, 0);
1262 code
= lp_build_interleave2(gallivm
, type8
, code
, code
, 0);
1267 * - expand color0/color1 to rgba8888
1268 * - calculate color2/3 (interpolation) according to color0 < color1 rules
1269 * - calculate color2/3 according to color0 >= color1 rules
1270 * - do selection of color2/3 according to comparison of color0/1
1271 * - extract indices.
1272 * - use compare/select to select the correct color. Since we have 2bit
1273 * indices (and 4 colors), needs at least three compare/selects.
1277 * expand the two colors
1279 color01
= LLVMBuildBitCast(builder
, color01
, bld16
.vec_type
, "");
1280 color01
= lp_build_interleave2(gallivm
, type16
, color01
,
1282 color01_16
= LLVMBuildBitCast(builder
, color01
, bld32
.vec_type
, "");
1283 color01
= color_expand_565_to_8888(gallivm
, 4, color01_16
);
1286 * interpolate colors
1287 * color2_1 is 2/3 color0 + 1/3 color1
1288 * color3_1 is 1/3 color0 + 2/3 color1
1289 * color2_2 is 1/2 color0 + 1/2 color1
1293 /* TODO: since this is now always scalar, should
1294 * probably just use control flow here instead of calculating
1295 * both cases and then selection
1297 if (format
== PIPE_FORMAT_DXT1_RGBA
||
1298 format
== PIPE_FORMAT_DXT1_SRGBA
) {
1299 color01
= LLVMBuildOr(builder
, color01
, a
, "");
1301 /* can combine 2 lerps into one mostly */
1302 color23
= lp_build_lerp23_single(&bld8
, color01
);
1303 color23
= LLVMBuildBitCast(builder
, color23
, bld32
.vec_type
, "");
1305 /* dxt3/5 always use 4-color encoding */
1306 if (is_dxt1_variant
) {
1307 LLVMValueRef color23_2
, color2_2
;
1309 if (util_cpu_caps
.has_sse2
) {
1310 LLVMValueRef intrargs
[2];
1311 intrargs
[0] = LLVMBuildBitCast(builder
, color01
, bld8
.vec_type
, "");
1312 /* same interleave as for lerp23 - correct result in 2nd element */
1313 intrargs
[1] = lp_build_interleave2(gallivm
, type32
, color01
, color01
, 0);
1314 intrargs
[1] = LLVMBuildBitCast(builder
, intrargs
[1], bld8
.vec_type
, "");
1315 color2_2
= lp_build_pavgb(&bld8
, intrargs
[0], intrargs
[1]);
1318 LLVMValueRef v01
, v0
, v1
, vhalf
;
1320 * This isn't as expensive as it looks (the unpack is the same as
1321 * for lerp23, which is the reason why we do the pointless
1322 * interleave2 too), with correct rounding (the two lower elements
1323 * will be the same).
1325 v01
= lp_build_interleave2(gallivm
, type32
, color01
, color01
, 0);
1326 v01
= LLVMBuildBitCast(builder
, v01
, bld8
.vec_type
, "");
1327 lp_build_unpack2(gallivm
, type8
, type16
, v01
, &v0
, &v1
);
1328 vhalf
= lp_build_add(&bld16
, v0
, v1
);
1329 vhalf
= LLVMBuildLShr(builder
, vhalf
, bld16
.one
, "");
1330 color2_2
= lp_build_pack2(gallivm
, type16
, type8
, vhalf
, bld16
.undef
);
1332 /* shuffle in color 3 as elem 2 zero, color 2 elem 1 */
1333 color23_2
= LLVMBuildBitCast(builder
, color2_2
, bld64
.vec_type
, "");
1334 color23_2
= LLVMBuildLShr(builder
, color23_2
,
1335 lp_build_const_int_vec(gallivm
, type64
, 32), "");
1336 color23_2
= LLVMBuildBitCast(builder
, color23_2
, bld32
.vec_type
, "");
1338 tmp
= LLVMBuildBitCast(builder
, color01_16
, bld64
.vec_type
, "");
1339 tmp
= LLVMBuildLShr(builder
, tmp
,
1340 lp_build_const_int_vec(gallivm
, type64
, 32), "");
1341 tmp
= LLVMBuildBitCast(builder
, tmp
, bld32
.vec_type
, "");
1342 sel_mask
= lp_build_compare(gallivm
, type32
, PIPE_FUNC_GREATER
,
1344 sel_mask
= lp_build_interleave2(gallivm
, type32
, sel_mask
, sel_mask
, 0);
1345 color23
= lp_build_select(&bld32
, sel_mask
, color23
, color23_2
);
1348 if (util_cpu_caps
.has_ssse3
) {
1350 * Use pshufb as mini-lut. (Only doable with intrinsics as the
1351 * final shuffles are non-constant. pshufb is awesome!)
1353 LLVMValueRef shuf
[16], low2mask
;
1354 LLVMValueRef intrargs
[2], lut_ind
, lut_adj
;
1356 color01
= LLVMBuildBitCast(builder
, color01
, bld64
.vec_type
, "");
1357 color23
= LLVMBuildBitCast(builder
, color23
, bld64
.vec_type
, "");
1358 color0123
= lp_build_interleave2(gallivm
, type64
, color01
, color23
, 0);
1359 color0123
= LLVMBuildBitCast(builder
, color0123
, bld32
.vec_type
, "");
1361 if (format
== PIPE_FORMAT_DXT1_RGB
||
1362 format
== PIPE_FORMAT_DXT1_SRGB
) {
1363 color0123
= LLVMBuildOr(builder
, color0123
, a
, "");
1366 /* shuffle as r0r1r2r3g0g1... */
1367 for (i
= 0; i
< 4; i
++) {
1368 shuf
[4*i
] = lp_build_const_int32(gallivm
, 0 + i
);
1369 shuf
[4*i
+1] = lp_build_const_int32(gallivm
, 4 + i
);
1370 shuf
[4*i
+2] = lp_build_const_int32(gallivm
, 8 + i
);
1371 shuf
[4*i
+3] = lp_build_const_int32(gallivm
, 12 + i
);
1373 color0123
= LLVMBuildBitCast(builder
, color0123
, bld8
.vec_type
, "");
1374 color0123
= LLVMBuildShuffleVector(builder
, color0123
, bld8
.undef
,
1375 LLVMConstVector(shuf
, 16), "");
1377 /* lowest 2 bits of each 8 bit value contain index into "LUT" */
1378 low2mask
= lp_build_const_int_vec(gallivm
, type8
, 3);
1379 /* add 0/4/8/12 for r/g/b/a */
1380 lut_adj
= lp_build_const_int_vec(gallivm
, type32
, 0x0c080400);
1381 lut_adj
= LLVMBuildBitCast(builder
, lut_adj
, bld8
.vec_type
, "");
1382 intrargs
[0] = color0123
;
1383 for (i
= 0; i
< 4; i
++) {
1384 lut_ind
= LLVMBuildAnd(builder
, code
, low2mask
, "");
1385 lut_ind
= LLVMBuildOr(builder
, lut_ind
, lut_adj
, "");
1386 intrargs
[1] = lut_ind
;
1387 col
[i
] = lp_build_intrinsic(builder
, "llvm.x86.ssse3.pshuf.b.128",
1388 bld8
.vec_type
, intrargs
, 2, 0);
1389 col
[i
] = LLVMBuildBitCast(builder
, col
[i
], bld32
.vec_type
, "");
1390 code
= LLVMBuildBitCast(builder
, code
, bld32
.vec_type
, "");
1391 code
= LLVMBuildLShr(builder
, code
, const2
, "");
1392 code
= LLVMBuildBitCast(builder
, code
, bld8
.vec_type
, "");
1396 /* Thanks to vectorization can do 4 texels in parallel */
1397 LLVMValueRef color0
, color1
, color2
, color3
;
1398 if (format
== PIPE_FORMAT_DXT1_RGB
||
1399 format
== PIPE_FORMAT_DXT1_SRGB
) {
1400 color01
= LLVMBuildOr(builder
, color01
, a
, "");
1401 color23
= LLVMBuildOr(builder
, color23
, a
, "");
1403 color0
= LLVMBuildShuffleVector(builder
, color01
, bld32
.undef
,
1404 lp_build_const_shuffle1(gallivm
, 0, 4), "");
1405 color1
= LLVMBuildShuffleVector(builder
, color01
, bld32
.undef
,
1406 lp_build_const_shuffle1(gallivm
, 1, 4), "");
1407 color2
= LLVMBuildShuffleVector(builder
, color23
, bld32
.undef
,
1408 lp_build_const_shuffle1(gallivm
, 0, 4), "");
1409 color3
= LLVMBuildShuffleVector(builder
, color23
, bld32
.undef
,
1410 lp_build_const_shuffle1(gallivm
, 1, 4), "");
1411 code
= LLVMBuildBitCast(builder
, code
, bld32
.vec_type
, "");
1413 for (i
= 0; i
< 4; i
++) {
1414 /* select the colors */
1415 LLVMValueRef selmasklo
, rgba01
, rgba23
, bitlo
;
1417 indices
= LLVMBuildAnd(builder
, code
, bitlo
, "");
1418 selmasklo
= lp_build_compare(gallivm
, type32
, PIPE_FUNC_EQUAL
,
1420 rgba01
= lp_build_select(&bld32
, selmasklo
, color1
, color0
);
1422 LLVMValueRef selmaskhi
;
1423 indices
= LLVMBuildAnd(builder
, code
, const2
, "");
1424 selmaskhi
= lp_build_compare(gallivm
, type32
, PIPE_FUNC_EQUAL
,
1426 rgba23
= lp_build_select(&bld32
, selmasklo
, color3
, color2
);
1427 rgba
= lp_build_select(&bld32
, selmaskhi
, rgba23
, rgba01
);
1430 * Note that this will give "wrong" order.
1431 * col0 will be rgba0, rgba4, rgba8, rgba12, col1 rgba1, rgba5, ...
1432 * This would be easily fixable by using different shuffle, bitlo/hi
1433 * vectors above (and different shift), but seems slightly easier to
1434 * deal with for dxt3/dxt5 alpha too. So instead change lookup.
1437 code
= LLVMBuildLShr(builder
, code
, const2
, "");
1443 * decode one dxt3 block.
1446 s3tc_decode_block_dxt3(struct gallivm_state
*gallivm
,
1447 enum pipe_format format
,
1448 LLVMValueRef dxt_block
,
1451 LLVMBuilderRef builder
= gallivm
->builder
;
1452 LLVMValueRef alpha
, alphas0
, alphas1
, shift4_16
, a
[4], mask8hi
;
1453 struct lp_type type32
, type8
, type16
;
1456 memset(&type32
, 0, sizeof type32
);
1460 memset(&type8
, 0, sizeof type8
);
1464 memset(&type16
, 0, sizeof type16
);
1468 s3tc_decode_block_dxt1(gallivm
, format
, dxt_block
, col
);
1470 shift4_16
= lp_build_const_int_vec(gallivm
, type16
, 4);
1471 mask8hi
= lp_build_const_int_vec(gallivm
, type32
, 0xff000000);
1473 alpha
= LLVMBuildBitCast(builder
, dxt_block
,
1474 lp_build_vec_type(gallivm
, type8
), "");
1475 alpha
= lp_build_interleave2(gallivm
, type8
, alpha
, alpha
, 0);
1476 alpha
= LLVMBuildBitCast(builder
, alpha
,
1477 lp_build_vec_type(gallivm
, type16
), "");
1478 alpha
= LLVMBuildAnd(builder
, alpha
,
1479 lp_build_const_int_vec(gallivm
, type16
, 0xf00f), "");
1480 alphas0
= LLVMBuildLShr(builder
, alpha
, shift4_16
, "");
1481 alphas1
= LLVMBuildShl(builder
, alpha
, shift4_16
, "");
1482 alpha
= LLVMBuildOr(builder
, alphas0
, alpha
, "");
1483 alpha
= LLVMBuildOr(builder
, alphas1
, alpha
, "");
1484 alpha
= LLVMBuildBitCast(builder
, alpha
,
1485 lp_build_vec_type(gallivm
, type32
), "");
1487 * alpha now contains elems 0,1,2,3,... (ubytes)
1488 * we need 0,4,8,12, 1,5,9,13 etc. in dwords to match color (which
1489 * is just as easy as "natural" order - 3 shift/and instead of 6 unpack).
1491 a
[0] = LLVMBuildShl(builder
, alpha
,
1492 lp_build_const_int_vec(gallivm
, type32
, 24), "");
1493 a
[1] = LLVMBuildShl(builder
, alpha
,
1494 lp_build_const_int_vec(gallivm
, type32
, 16), "");
1495 a
[1] = LLVMBuildAnd(builder
, a
[1], mask8hi
, "");
1496 a
[2] = LLVMBuildShl(builder
, alpha
,
1497 lp_build_const_int_vec(gallivm
, type32
, 8), "");
1498 a
[2] = LLVMBuildAnd(builder
, a
[2], mask8hi
, "");
1499 a
[3] = LLVMBuildAnd(builder
, alpha
, mask8hi
, "");
1501 for (i
= 0; i
< 4; i
++) {
1502 col
[i
] = LLVMBuildOr(builder
, col
[i
], a
[i
], "");
1508 lp_build_lerpdxta_block(struct gallivm_state
*gallivm
,
1509 LLVMValueRef alpha0
,
1510 LLVMValueRef alpha1
,
1512 LLVMValueRef sel_mask
)
1514 LLVMBuilderRef builder
= gallivm
->builder
;
1515 LLVMValueRef delta
, ainterp
;
1516 LLVMValueRef weight5
, weight7
, weight
;
1517 struct lp_type type16
;
1518 struct lp_build_context bld
;
1520 memset(&type16
, 0, sizeof type16
);
1525 lp_build_context_init(&bld
, gallivm
, type16
);
1527 * 256/7 is only 36.57 so we'd lose quite some precision. Since it would
1528 * actually be desirable to do this here with even higher accuracy than
1529 * even 8 bit (more or less required for rgtc, albeit that's not handled
1530 * here right now), shift the weights after multiplication by code.
1532 weight5
= lp_build_const_int_vec(gallivm
, type16
, 256*64/5);
1533 weight7
= lp_build_const_int_vec(gallivm
, type16
, 256*64/7);
1534 weight
= lp_build_select(&bld
, sel_mask
, weight7
, weight5
);
1537 * we'll get garbage in the elements which had code 0 (or larger than
1538 * 5 or 7) but we don't care (or rather, need to fix up anyway).
1540 code
= LLVMBuildSub(builder
, code
, bld
.one
, "");
1542 weight
= LLVMBuildMul(builder
, weight
, code
, "");
1543 weight
= LLVMBuildLShr(builder
, weight
,
1544 lp_build_const_int_vec(gallivm
, type16
, 6), "");
1546 delta
= LLVMBuildSub(builder
, alpha1
, alpha0
, "");
1548 ainterp
= LLVMBuildMul(builder
, delta
, weight
, "");
1549 ainterp
= LLVMBuildLShr(builder
, ainterp
,
1550 lp_build_const_int_vec(gallivm
, type16
, 8), "");
1552 /* lerp is done later (with packed values) */
1559 * decode one dxt5 block.
1562 s3tc_decode_block_dxt5(struct gallivm_state
*gallivm
,
1563 enum pipe_format format
,
1564 LLVMValueRef dxt_block
,
1567 LLVMBuilderRef builder
= gallivm
->builder
;
1568 LLVMValueRef alpha
, alpha0
, alpha1
, ares
;
1569 LLVMValueRef ainterp
, ainterp0
, ainterp1
, shuffle1
, sel_mask
, sel_mask2
;
1570 LLVMValueRef a
[4], acode
, tmp0
, tmp1
;
1571 LLVMTypeRef i64t
, i32t
;
1572 struct lp_type type32
, type64
, type8
, type16
;
1573 struct lp_build_context bld16
, bld8
;
1576 memset(&type32
, 0, sizeof type32
);
1580 memset(&type64
, 0, sizeof type64
);
1584 memset(&type8
, 0, sizeof type8
);
1588 memset(&type16
, 0, sizeof type16
);
1592 lp_build_context_init(&bld16
, gallivm
, type16
);
1593 lp_build_context_init(&bld8
, gallivm
, type8
);
1595 i64t
= lp_build_vec_type(gallivm
, type64
);
1596 i32t
= lp_build_vec_type(gallivm
, type32
);
1598 s3tc_decode_block_dxt1(gallivm
, format
, dxt_block
, col
);
1601 * three possible strategies for vectorizing alpha:
1602 * 1) compute all 8 values then use scalar extraction
1603 * (i.e. have all 8 alpha values packed in one 64bit scalar
1604 * and do something like ax = vals >> (codex * 8) followed
1605 * by inserting these values back into color)
1606 * 2) same as 8 but just use pshufb as a mini-LUT for selection.
1607 * (without pshufb would need boatloads of cmp/selects trying to
1608 * keep things vectorized for essentially scalar selection).
1609 * 3) do something similar to the uncached case
1610 * needs more calculations (need to calc 16 values instead of 8 though
1611 * that's only an issue for the lerp which we need to do twice otherwise
1612 * everything still fits into 128bit) but keeps things vectorized mostly.
1613 * Trying 3) here though not sure it's really faster...
1614 * With pshufb, we try 2) (cheaper and more accurate)
1618 * Ideally, we'd use 2 variable 16bit shifts here (byte shifts wouldn't
1619 * help since code crosses 8bit boundaries). But variable shifts are
1620 * AVX2 only, and even then only dword/quadword (intel _really_ hates
1621 * shifts!). Instead, emulate by 16bit muls.
1622 * Also, the required byte shuffles are essentially non-emulatable, so
1623 * require ssse3 (albeit other archs might do them fine).
1624 * This is not directly tied to ssse3 - just need sane byte shuffles.
1625 * But ordering is going to be different below so use same condition.
1629 /* vectorize alpha */
1630 alpha
= LLVMBuildBitCast(builder
, dxt_block
, i64t
, "");
1631 alpha0
= LLVMBuildAnd(builder
, alpha
,
1632 lp_build_const_int_vec(gallivm
, type64
, 0xff), "");
1633 alpha0
= LLVMBuildBitCast(builder
, alpha0
, bld16
.vec_type
, "");
1634 alpha
= LLVMBuildBitCast(builder
, alpha
, bld16
.vec_type
, "");
1635 alpha1
= LLVMBuildLShr(builder
, alpha
,
1636 lp_build_const_int_vec(gallivm
, type16
, 8), "");
1637 alpha
= LLVMBuildBitCast(builder
, alpha
, i64t
, "");
1638 shuffle1
= lp_build_const_shuffle1(gallivm
, 0, 8);
1639 alpha0
= LLVMBuildShuffleVector(builder
, alpha0
, alpha0
, shuffle1
, "");
1640 alpha1
= LLVMBuildShuffleVector(builder
, alpha1
, alpha1
, shuffle1
, "");
1643 sel_mask
= lp_build_compare(gallivm
, type16
, PIPE_FUNC_GREATER
,
1645 type16
.sign
= FALSE
;
1646 sel_mask
= LLVMBuildBitCast(builder
, sel_mask
, bld8
.vec_type
, "");
1648 if (!util_cpu_caps
.has_ssse3
) {
1649 LLVMValueRef acodeg
, mask1
, acode0
, acode1
;
1651 /* extraction of the 3 bit values into something more useful is HARD */
1652 /* first steps are actually scalar */
1653 acode
= LLVMBuildLShr(builder
, alpha
,
1654 lp_build_const_int_vec(gallivm
, type64
, 16), "");
1655 tmp0
= LLVMBuildAnd(builder
, acode
,
1656 lp_build_const_int_vec(gallivm
, type64
, 0xffffff), "");
1657 tmp1
= LLVMBuildLShr(builder
, acode
,
1658 lp_build_const_int_vec(gallivm
, type64
, 24), "");
1659 tmp0
= LLVMBuildBitCast(builder
, tmp0
, i32t
, "");
1660 tmp1
= LLVMBuildBitCast(builder
, tmp1
, i32t
, "");
1661 acode
= lp_build_interleave2(gallivm
, type32
, tmp0
, tmp1
, 0);
1662 /* now have 2x24bit in 4x32bit, order 01234567, 89..., undef, undef */
1663 tmp0
= LLVMBuildAnd(builder
, acode
,
1664 lp_build_const_int_vec(gallivm
, type32
, 0xfff), "");
1665 tmp1
= LLVMBuildLShr(builder
, acode
,
1666 lp_build_const_int_vec(gallivm
, type32
, 12), "");
1667 acode
= lp_build_interleave2(gallivm
, type32
, tmp0
, tmp1
, 0);
1668 /* now have 4x12bit in 4x32bit, order 0123, 4567, ,,, */
1669 tmp0
= LLVMBuildAnd(builder
, acode
,
1670 lp_build_const_int_vec(gallivm
, type32
, 0x3f), "");
1671 tmp1
= LLVMBuildLShr(builder
, acode
,
1672 lp_build_const_int_vec(gallivm
, type32
, 6), "");
1673 /* use signed pack doesn't matter and otherwise need sse41 */
1674 type32
.sign
= type16
.sign
= TRUE
;
1675 acode
= lp_build_pack2(gallivm
, type32
, type16
, tmp0
, tmp1
);
1676 type32
.sign
= type16
.sign
= FALSE
;
1677 /* now have 8x6bit in 8x16bit, 01, 45, 89, ..., 23, 67, ... */
1678 acode0
= LLVMBuildAnd(builder
, acode
,
1679 lp_build_const_int_vec(gallivm
, type16
, 0x7), "");
1680 acode1
= LLVMBuildLShr(builder
, acode
,
1681 lp_build_const_int_vec(gallivm
, type16
, 3), "");
1682 acode
= lp_build_pack2(gallivm
, type16
, type8
, acode0
, acode1
);
1683 /* acode0 contains elems 0,4,8,12,2,6,10,14, acode1 1,5,9,... */
1685 acodeg
= LLVMBuildAnd(builder
, acode
,
1686 LLVMBuildNot(builder
, sel_mask
, ""), "");
1687 mask1
= lp_build_compare(gallivm
, type8
, PIPE_FUNC_EQUAL
,
1690 sel_mask
= LLVMBuildBitCast(builder
, sel_mask
, bld16
.vec_type
, "");
1691 ainterp0
= lp_build_lerpdxta_block(gallivm
, alpha0
, alpha1
, acode0
, sel_mask
);
1692 ainterp1
= lp_build_lerpdxta_block(gallivm
, alpha0
, alpha1
, acode1
, sel_mask
);
1693 sel_mask
= LLVMBuildBitCast(builder
, sel_mask
, bld8
.vec_type
, "");
1694 ainterp
= lp_build_pack2(gallivm
, type16
, type8
, ainterp0
, ainterp1
);
1695 alpha0
= lp_build_pack2(gallivm
, type16
, type8
, alpha0
, alpha0
);
1696 alpha1
= lp_build_pack2(gallivm
, type16
, type8
, alpha1
, alpha1
);
1697 ainterp
= LLVMBuildAdd(builder
, ainterp
, alpha0
, "");
1699 sel_mask2
= lp_build_compare(gallivm
, type8
, PIPE_FUNC_EQUAL
,
1701 ainterp
= lp_build_select(&bld8
, sel_mask2
, alpha0
, ainterp
);
1702 ainterp
= lp_build_select(&bld8
, mask1
, alpha1
, ainterp
);
1704 /* fix up val67 if a0 <= a1 */
1705 sel_mask2
= lp_build_compare(gallivm
, type8
, PIPE_FUNC_EQUAL
,
1706 acodeg
, lp_build_const_int_vec(gallivm
, type8
, 6));
1707 ares
= LLVMBuildAnd(builder
, ainterp
, LLVMBuildNot(builder
, sel_mask2
, ""), "");
1708 sel_mask2
= lp_build_compare(gallivm
, type8
, PIPE_FUNC_EQUAL
,
1709 acodeg
, lp_build_const_int_vec(gallivm
, type8
, 7));
1710 ares
= LLVMBuildOr(builder
, ares
, sel_mask2
, "");
1712 /* unpack in right order (0,4,8,12,1,5,..) */
1713 /* this gives us zero, a0, zero, a4, zero, a8, ... for tmp0 */
1714 tmp0
= lp_build_interleave2(gallivm
, type8
, bld8
.zero
, ares
, 0);
1715 tmp1
= lp_build_interleave2(gallivm
, type8
, bld8
.zero
, ares
, 1);
1716 tmp0
= LLVMBuildBitCast(builder
, tmp0
, bld16
.vec_type
, "");
1717 tmp1
= LLVMBuildBitCast(builder
, tmp1
, bld16
.vec_type
, "");
1719 a
[0] = lp_build_interleave2(gallivm
, type16
, bld16
.zero
, tmp0
, 0);
1720 a
[1] = lp_build_interleave2(gallivm
, type16
, bld16
.zero
, tmp1
, 0);
1721 a
[2] = lp_build_interleave2(gallivm
, type16
, bld16
.zero
, tmp0
, 1);
1722 a
[3] = lp_build_interleave2(gallivm
, type16
, bld16
.zero
, tmp1
, 1);
1725 LLVMValueRef elems
[16], intrargs
[2], shufa
, mulclo
, mulchi
, mask8hi
;
1726 LLVMTypeRef type16s
= LLVMInt16TypeInContext(gallivm
->context
);
1727 LLVMTypeRef type8s
= LLVMInt8TypeInContext(gallivm
->context
);
1730 * Ideally, we'd use 2 variable 16bit shifts here (byte shifts wouldn't
1731 * help since code crosses 8bit boundaries). But variable shifts are
1732 * AVX2 only, and even then only dword/quadword (intel _really_ hates
1733 * shifts!). Instead, emulate by 16bit muls.
1734 * Also, the required byte shuffles are essentially non-emulatable, so
1735 * require ssse3 (albeit other archs might do them fine, but the
1736 * complete path is ssse3 only for now).
1738 for (i
= 0, j
= 0; i
< 16; i
+= 8, j
+= 3) {
1739 elems
[i
+0] = elems
[i
+1] = elems
[i
+2] = lp_build_const_int32(gallivm
, j
+2);
1740 elems
[i
+3] = elems
[i
+4] = lp_build_const_int32(gallivm
, j
+3);
1741 elems
[i
+5] = elems
[i
+6] = elems
[i
+7] = lp_build_const_int32(gallivm
, j
+4);
1743 shufa
= LLVMConstVector(elems
, 16);
1744 alpha
= LLVMBuildBitCast(builder
, alpha
, bld8
.vec_type
, "");
1745 acode
= LLVMBuildShuffleVector(builder
, alpha
, bld8
.undef
, shufa
, "");
1746 acode
= LLVMBuildBitCast(builder
, acode
, bld16
.vec_type
, "");
1748 * Put 0/2/4/6 into high 3 bits of 16 bits (save AND mask)
1749 * Do the same for 1/3/5/7 (albeit still need mask there - ideally
1750 * we'd place them into bits 4-7 so could save shift but impossible.)
1752 for (i
= 0; i
< 8; i
+= 4) {
1753 elems
[i
+0] = LLVMConstInt(type16s
, 1 << (13-0), 0);
1754 elems
[i
+1] = LLVMConstInt(type16s
, 1 << (13-6), 0);
1755 elems
[i
+2] = LLVMConstInt(type16s
, 1 << (13-4), 0);
1756 elems
[i
+3] = LLVMConstInt(type16s
, 1 << (13-2), 0);
1758 mulclo
= LLVMConstVector(elems
, 8);
1759 for (i
= 0; i
< 8; i
+= 4) {
1760 elems
[i
+0] = LLVMConstInt(type16s
, 1 << (13-3), 0);
1761 elems
[i
+1] = LLVMConstInt(type16s
, 1 << (13-9), 0);
1762 elems
[i
+2] = LLVMConstInt(type16s
, 1 << (13-7), 0);
1763 elems
[i
+3] = LLVMConstInt(type16s
, 1 << (13-5), 0);
1765 mulchi
= LLVMConstVector(elems
, 8);
1767 tmp0
= LLVMBuildMul(builder
, acode
, mulclo
, "");
1768 tmp1
= LLVMBuildMul(builder
, acode
, mulchi
, "");
1769 tmp0
= LLVMBuildLShr(builder
, tmp0
,
1770 lp_build_const_int_vec(gallivm
, type16
, 13), "");
1771 tmp1
= LLVMBuildLShr(builder
, tmp1
,
1772 lp_build_const_int_vec(gallivm
, type16
, 5), "");
1773 tmp1
= LLVMBuildAnd(builder
, tmp1
,
1774 lp_build_const_int_vec(gallivm
, type16
, 0x700), "");
1775 acode
= LLVMBuildOr(builder
, tmp0
, tmp1
, "");
1776 acode
= LLVMBuildBitCast(builder
, acode
, bld8
.vec_type
, "");
1779 * Note that ordering is different here to non-ssse3 path:
1783 LLVMValueRef weight0
, weight1
, weight
, delta
;
1784 LLVMValueRef constff_elem7
, const0_elem6
;
1785 /* weights, correctly rounded (round(256*x/7)) */
1786 elems
[0] = LLVMConstInt(type16s
, 256, 0);
1787 elems
[1] = LLVMConstInt(type16s
, 0, 0);
1788 elems
[2] = LLVMConstInt(type16s
, 219, 0);
1789 elems
[3] = LLVMConstInt(type16s
, 183, 0);
1790 elems
[4] = LLVMConstInt(type16s
, 146, 0);
1791 elems
[5] = LLVMConstInt(type16s
, 110, 0);
1792 elems
[6] = LLVMConstInt(type16s
, 73, 0);
1793 elems
[7] = LLVMConstInt(type16s
, 37, 0);
1794 weight0
= LLVMConstVector(elems
, 8);
1796 elems
[0] = LLVMConstInt(type16s
, 256, 0);
1797 elems
[1] = LLVMConstInt(type16s
, 0, 0);
1798 elems
[2] = LLVMConstInt(type16s
, 205, 0);
1799 elems
[3] = LLVMConstInt(type16s
, 154, 0);
1800 elems
[4] = LLVMConstInt(type16s
, 102, 0);
1801 elems
[5] = LLVMConstInt(type16s
, 51, 0);
1802 elems
[6] = LLVMConstInt(type16s
, 0, 0);
1803 elems
[7] = LLVMConstInt(type16s
, 0, 0);
1804 weight1
= LLVMConstVector(elems
, 8);
1806 weight0
= LLVMBuildBitCast(builder
, weight0
, bld8
.vec_type
, "");
1807 weight1
= LLVMBuildBitCast(builder
, weight1
, bld8
.vec_type
, "");
1808 weight
= lp_build_select(&bld8
, sel_mask
, weight0
, weight1
);
1809 weight
= LLVMBuildBitCast(builder
, weight
, bld16
.vec_type
, "");
1811 for (i
= 0; i
< 16; i
++) {
1812 elems
[i
] = LLVMConstNull(type8s
);
1814 elems
[7] = LLVMConstInt(type8s
, 255, 0);
1815 constff_elem7
= LLVMConstVector(elems
, 16);
1817 for (i
= 0; i
< 16; i
++) {
1818 elems
[i
] = LLVMConstInt(type8s
, 255, 0);
1820 elems
[6] = LLVMConstInt(type8s
, 0, 0);
1821 const0_elem6
= LLVMConstVector(elems
, 16);
1823 /* standard simple lerp - but the version we need isn't available */
1824 delta
= LLVMBuildSub(builder
, alpha0
, alpha1
, "");
1825 ainterp
= LLVMBuildMul(builder
, delta
, weight
, "");
1826 ainterp
= LLVMBuildLShr(builder
, ainterp
,
1827 lp_build_const_int_vec(gallivm
, type16
, 8), "");
1828 ainterp
= LLVMBuildBitCast(builder
, ainterp
, bld8
.vec_type
, "");
1829 alpha1
= LLVMBuildBitCast(builder
, alpha1
, bld8
.vec_type
, "");
1830 ainterp
= LLVMBuildAdd(builder
, ainterp
, alpha1
, "");
1831 ainterp
= LLVMBuildBitCast(builder
, ainterp
, bld16
.vec_type
, "");
1832 ainterp
= lp_build_pack2(gallivm
, type16
, type8
, ainterp
, bld16
.undef
);
1834 /* fixing 0/0xff case is slightly more complex */
1835 constff_elem7
= LLVMBuildAnd(builder
, constff_elem7
,
1836 LLVMBuildNot(builder
, sel_mask
, ""), "");
1837 const0_elem6
= LLVMBuildOr(builder
, const0_elem6
, sel_mask
, "");
1838 ainterp
= LLVMBuildOr(builder
, ainterp
, constff_elem7
, "");
1839 ainterp
= LLVMBuildAnd(builder
, ainterp
, const0_elem6
, "");
1841 /* now pick all 16 elements at once! */
1842 intrargs
[0] = ainterp
;
1843 intrargs
[1] = acode
;
1844 ares
= lp_build_intrinsic(builder
, "llvm.x86.ssse3.pshuf.b.128",
1845 bld8
.vec_type
, intrargs
, 2, 0);
1847 ares
= LLVMBuildBitCast(builder
, ares
, i32t
, "");
1848 mask8hi
= lp_build_const_int_vec(gallivm
, type32
, 0xff000000);
1849 a
[0] = LLVMBuildShl(builder
, ares
,
1850 lp_build_const_int_vec(gallivm
, type32
, 24), "");
1851 a
[1] = LLVMBuildShl(builder
, ares
,
1852 lp_build_const_int_vec(gallivm
, type32
, 16), "");
1853 a
[1] = LLVMBuildAnd(builder
, a
[1], mask8hi
, "");
1854 a
[2] = LLVMBuildShl(builder
, ares
,
1855 lp_build_const_int_vec(gallivm
, type32
, 8), "");
1856 a
[2] = LLVMBuildAnd(builder
, a
[2], mask8hi
, "");
1857 a
[3] = LLVMBuildAnd(builder
, ares
, mask8hi
, "");
1860 for (i
= 0; i
< 4; i
++) {
1861 a
[i
] = LLVMBuildBitCast(builder
, a
[i
], i32t
, "");
1862 col
[i
] = LLVMBuildOr(builder
, col
[i
], a
[i
], "");
1868 generate_update_cache_one_block(struct gallivm_state
*gallivm
,
1869 LLVMValueRef function
,
1870 const struct util_format_description
*format_desc
)
1872 LLVMBasicBlockRef block
;
1873 LLVMBuilderRef old_builder
;
1874 LLVMValueRef ptr_addr
;
1875 LLVMValueRef hash_index
;
1877 LLVMValueRef dxt_block
, tag_value
;
1878 LLVMValueRef col
[LP_MAX_VECTOR_LENGTH
];
1880 ptr_addr
= LLVMGetParam(function
, 0);
1881 hash_index
= LLVMGetParam(function
, 1);
1882 cache
= LLVMGetParam(function
, 2);
1884 lp_build_name(ptr_addr
, "ptr_addr" );
1885 lp_build_name(hash_index
, "hash_index");
1886 lp_build_name(cache
, "cache_addr");
1892 old_builder
= gallivm
->builder
;
1893 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
1894 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
1895 LLVMPositionBuilderAtEnd(gallivm
->builder
, block
);
1897 lp_build_gather_s3tc_simple_scalar(gallivm
, format_desc
, &dxt_block
,
1900 switch (format_desc
->format
) {
1901 case PIPE_FORMAT_DXT1_RGB
:
1902 case PIPE_FORMAT_DXT1_RGBA
:
1903 case PIPE_FORMAT_DXT1_SRGB
:
1904 case PIPE_FORMAT_DXT1_SRGBA
:
1905 s3tc_decode_block_dxt1(gallivm
, format_desc
->format
, dxt_block
, col
);
1907 case PIPE_FORMAT_DXT3_RGBA
:
1908 case PIPE_FORMAT_DXT3_SRGBA
:
1909 s3tc_decode_block_dxt3(gallivm
, format_desc
->format
, dxt_block
, col
);
1911 case PIPE_FORMAT_DXT5_RGBA
:
1912 case PIPE_FORMAT_DXT5_SRGBA
:
1913 s3tc_decode_block_dxt5(gallivm
, format_desc
->format
, dxt_block
, col
);
1917 s3tc_decode_block_dxt1(gallivm
, format_desc
->format
, dxt_block
, col
);
1921 tag_value
= LLVMBuildPtrToInt(gallivm
->builder
, ptr_addr
,
1922 LLVMInt64TypeInContext(gallivm
->context
), "");
1923 s3tc_store_cached_block(gallivm
, col
, tag_value
, hash_index
, cache
);
1925 LLVMBuildRetVoid(gallivm
->builder
);
1927 LLVMDisposeBuilder(gallivm
->builder
);
1928 gallivm
->builder
= old_builder
;
1930 gallivm_verify_function(gallivm
, function
);
1935 update_cached_block(struct gallivm_state
*gallivm
,
1936 const struct util_format_description
*format_desc
,
1937 LLVMValueRef ptr_addr
,
1938 LLVMValueRef hash_index
,
1942 LLVMBuilderRef builder
= gallivm
->builder
;
1943 LLVMModuleRef module
= gallivm
->module
;
1945 LLVMTypeRef i8t
= LLVMInt8TypeInContext(gallivm
->context
);
1946 LLVMTypeRef pi8t
= LLVMPointerType(i8t
, 0);
1947 LLVMValueRef function
, inst
;
1948 LLVMBasicBlockRef bb
;
1949 LLVMValueRef args
[3];
1951 util_snprintf(name
, sizeof name
, "%s_update_cache_one_block",
1952 format_desc
->short_name
);
1953 function
= LLVMGetNamedFunction(module
, name
);
1956 LLVMTypeRef ret_type
;
1957 LLVMTypeRef arg_types
[3];
1958 LLVMTypeRef function_type
;
1962 * Generate the function prototype.
1965 ret_type
= LLVMVoidTypeInContext(gallivm
->context
);
1966 arg_types
[0] = pi8t
;
1967 arg_types
[1] = LLVMInt32TypeInContext(gallivm
->context
);
1968 arg_types
[2] = LLVMTypeOf(cache
); // XXX: put right type here
1969 function_type
= LLVMFunctionType(ret_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
1970 function
= LLVMAddFunction(module
, name
, function_type
);
1972 for (arg
= 0; arg
< ARRAY_SIZE(arg_types
); ++arg
)
1973 if (LLVMGetTypeKind(arg_types
[arg
]) == LLVMPointerTypeKind
)
1974 lp_add_function_attr(function
, arg
+ 1, LP_FUNC_ATTR_NOALIAS
);
1976 LLVMSetFunctionCallConv(function
, LLVMFastCallConv
);
1977 LLVMSetVisibility(function
, LLVMHiddenVisibility
);
1978 generate_update_cache_one_block(gallivm
, function
, format_desc
);
1982 args
[1] = hash_index
;
1985 LLVMBuildCall(builder
, function
, args
, ARRAY_SIZE(args
), "");
1986 bb
= LLVMGetInsertBlock(builder
);
1987 inst
= LLVMGetLastInstruction(bb
);
1988 LLVMSetInstructionCallConv(inst
, LLVMFastCallConv
);
1995 compressed_fetch_cached(struct gallivm_state
*gallivm
,
1996 const struct util_format_description
*format_desc
,
1998 LLVMValueRef base_ptr
,
1999 LLVMValueRef offset
,
2005 LLVMBuilderRef builder
= gallivm
->builder
;
2006 unsigned count
, low_bit
, log2size
;
2007 LLVMValueRef color
, offset_stored
, addr
, ptr_addrtrunc
, tmp
;
2008 LLVMValueRef ij_index
, hash_index
, hash_mask
, block_index
;
2009 LLVMTypeRef i8t
= LLVMInt8TypeInContext(gallivm
->context
);
2010 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
2011 LLVMTypeRef i64t
= LLVMInt64TypeInContext(gallivm
->context
);
2012 struct lp_type type
;
2013 struct lp_build_context bld32
;
2014 memset(&type
, 0, sizeof type
);
2018 lp_build_context_init(&bld32
, gallivm
, type
);
2021 * compute hash - we use direct mapped cache, the hash function could
2022 * be better but it needs to be simple
2024 * compare offset with offset stored at tag (hash)
2025 * if not equal extract block, store block, update tag
2026 * extract color from cache
2030 low_bit
= util_logbase2(format_desc
->block
.bits
/ 8);
2031 log2size
= util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE
);
2032 addr
= LLVMBuildPtrToInt(builder
, base_ptr
, i64t
, "");
2033 ptr_addrtrunc
= LLVMBuildPtrToInt(builder
, base_ptr
, i32t
, "");
2034 ptr_addrtrunc
= lp_build_broadcast_scalar(&bld32
, ptr_addrtrunc
);
2035 /* For the hash function, first mask off the unused lowest bits. Then just
2036 do some xor with address bits - only use lower 32bits */
2037 ptr_addrtrunc
= LLVMBuildAdd(builder
, offset
, ptr_addrtrunc
, "");
2038 ptr_addrtrunc
= LLVMBuildLShr(builder
, ptr_addrtrunc
,
2039 lp_build_const_int_vec(gallivm
, type
, low_bit
), "");
2040 /* This only really makes sense for size 64,128,256 */
2041 hash_index
= ptr_addrtrunc
;
2042 ptr_addrtrunc
= LLVMBuildLShr(builder
, ptr_addrtrunc
,
2043 lp_build_const_int_vec(gallivm
, type
, 2*log2size
), "");
2044 hash_index
= LLVMBuildXor(builder
, ptr_addrtrunc
, hash_index
, "");
2045 tmp
= LLVMBuildLShr(builder
, hash_index
,
2046 lp_build_const_int_vec(gallivm
, type
, log2size
), "");
2047 hash_index
= LLVMBuildXor(builder
, hash_index
, tmp
, "");
2049 hash_mask
= lp_build_const_int_vec(gallivm
, type
, LP_BUILD_FORMAT_CACHE_SIZE
- 1);
2050 hash_index
= LLVMBuildAnd(builder
, hash_index
, hash_mask
, "");
2051 ij_index
= LLVMBuildShl(builder
, i
, lp_build_const_int_vec(gallivm
, type
, 2), "");
2052 ij_index
= LLVMBuildAdd(builder
, ij_index
, j
, "");
2053 block_index
= LLVMBuildShl(builder
, hash_index
,
2054 lp_build_const_int_vec(gallivm
, type
, 4), "");
2055 block_index
= LLVMBuildAdd(builder
, ij_index
, block_index
, "");
2058 color
= bld32
.undef
;
2059 for (count
= 0; count
< n
; count
++) {
2060 LLVMValueRef index
, cond
, colorx
;
2061 LLVMValueRef block_indexx
, hash_indexx
, addrx
, offsetx
, ptr_addrx
;
2062 struct lp_build_if_state if_ctx
;
2064 index
= lp_build_const_int32(gallivm
, count
);
2065 offsetx
= LLVMBuildExtractElement(builder
, offset
, index
, "");
2066 addrx
= LLVMBuildZExt(builder
, offsetx
, i64t
, "");
2067 addrx
= LLVMBuildAdd(builder
, addrx
, addr
, "");
2068 block_indexx
= LLVMBuildExtractElement(builder
, block_index
, index
, "");
2069 hash_indexx
= LLVMBuildLShr(builder
, block_indexx
,
2070 lp_build_const_int32(gallivm
, 4), "");
2071 offset_stored
= s3tc_lookup_tag_data(gallivm
, cache
, hash_indexx
);
2072 cond
= LLVMBuildICmp(builder
, LLVMIntNE
, offset_stored
, addrx
, "");
2074 lp_build_if(&if_ctx
, gallivm
, cond
);
2076 ptr_addrx
= LLVMBuildIntToPtr(builder
, addrx
,
2077 LLVMPointerType(i8t
, 0), "");
2078 update_cached_block(gallivm
, format_desc
, ptr_addrx
, hash_indexx
, cache
);
2079 #if LP_BUILD_FORMAT_CACHE_DEBUG
2080 s3tc_update_cache_access(gallivm
, cache
, 1,
2081 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS
);
2084 lp_build_endif(&if_ctx
);
2086 colorx
= s3tc_lookup_cached_pixel(gallivm
, cache
, block_indexx
);
2088 color
= LLVMBuildInsertElement(builder
, color
, colorx
,
2089 lp_build_const_int32(gallivm
, count
), "");
2094 struct lp_build_if_state if_ctx
;
2096 tmp
= LLVMBuildZExt(builder
, offset
, i64t
, "");
2097 addr
= LLVMBuildAdd(builder
, tmp
, addr
, "");
2098 offset_stored
= s3tc_lookup_tag_data(gallivm
, cache
, hash_index
);
2099 cond
= LLVMBuildICmp(builder
, LLVMIntNE
, offset_stored
, addr
, "");
2101 lp_build_if(&if_ctx
, gallivm
, cond
);
2103 tmp
= LLVMBuildIntToPtr(builder
, addr
, LLVMPointerType(i8t
, 0), "");
2104 update_cached_block(gallivm
, format_desc
, tmp
, hash_index
, cache
);
2105 #if LP_BUILD_FORMAT_CACHE_DEBUG
2106 s3tc_update_cache_access(gallivm
, cache
, 1,
2107 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS
);
2110 lp_build_endif(&if_ctx
);
2112 color
= s3tc_lookup_cached_pixel(gallivm
, cache
, block_index
);
2114 #if LP_BUILD_FORMAT_CACHE_DEBUG
2115 s3tc_update_cache_access(gallivm
, cache
, n
,
2116 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL
);
2118 return LLVMBuildBitCast(builder
, color
, LLVMVectorType(i8t
, n
* 4), "");
2123 s3tc_dxt5_to_rgba_aos(struct gallivm_state
*gallivm
,
2125 enum pipe_format format
,
2126 LLVMValueRef colors
,
2127 LLVMValueRef codewords
,
2128 LLVMValueRef alpha_lo
,
2129 LLVMValueRef alpha_hi
,
2133 return s3tc_dxt5_full_to_rgba_aos(gallivm
, n
, format
, colors
,
2134 codewords
, alpha_lo
, alpha_hi
, i
, j
);
2139 * @param n number of pixels processed (usually n=4, but it should also work with n=1
2140 * and multiples of 4)
2141 * @param base_ptr base pointer (32bit or 64bit pointer depending on the architecture)
2142 * @param offset <n x i32> vector with the relative offsets of the S3TC blocks
2143 * @param i is a <n x i32> vector with the x subpixel coordinate (0..3)
2144 * @param j is a <n x i32> vector with the y subpixel coordinate (0..3)
2145 * @return a <4*n x i8> vector with the pixel RGBA values in AoS
2148 lp_build_fetch_s3tc_rgba_aos(struct gallivm_state
*gallivm
,
2149 const struct util_format_description
*format_desc
,
2151 LLVMValueRef base_ptr
,
2152 LLVMValueRef offset
,
2158 LLVMTypeRef i8t
= LLVMInt8TypeInContext(gallivm
->context
);
2159 LLVMBuilderRef builder
= gallivm
->builder
;
2161 assert(format_desc
->layout
== UTIL_FORMAT_LAYOUT_S3TC
);
2162 assert(format_desc
->block
.width
== 4);
2163 assert(format_desc
->block
.height
== 4);
2165 assert((n
== 1) || (n
% 4 == 0));
2167 /* debug_printf("format = %d\n", format_desc->format);*/
2169 rgba
= compressed_fetch_cached(gallivm
, format_desc
, n
,
2170 base_ptr
, offset
, i
, j
, cache
);
2175 * Could use n > 8 here with avx2, but doesn't seem faster.
2179 LLVMTypeRef i8_vectype
= LLVMVectorType(i8t
, 4 * n
);
2180 LLVMTypeRef i128_type
= LLVMIntTypeInContext(gallivm
->context
, 128);
2181 LLVMTypeRef i128_vectype
= LLVMVectorType(i128_type
, n
/ 4);
2182 LLVMTypeRef i324_vectype
= LLVMVectorType(LLVMInt32TypeInContext(
2183 gallivm
->context
), 4);
2184 LLVMValueRef offset4
, i4
, j4
, rgba4
[LP_MAX_VECTOR_LENGTH
/16];
2185 struct lp_type lp_324_vectype
= lp_type_uint_vec(32, 128);
2187 assert(n
/ 4 <= ARRAY_SIZE(rgba4
));
2189 rgba
= LLVMGetUndef(i128_vectype
);
2191 for (count
= 0; count
< n
/ 4; count
++) {
2192 LLVMValueRef colors
, codewords
, alpha_lo
= NULL
, alpha_hi
= NULL
;
2194 i4
= lp_build_extract_range(gallivm
, i
, count
* 4, 4);
2195 j4
= lp_build_extract_range(gallivm
, j
, count
* 4, 4);
2196 offset4
= lp_build_extract_range(gallivm
, offset
, count
* 4, 4);
2198 lp_build_gather_s3tc(gallivm
, 4, format_desc
, &colors
, &codewords
,
2199 &alpha_lo
, &alpha_hi
, base_ptr
, offset4
);
2201 switch (format_desc
->format
) {
2202 case PIPE_FORMAT_DXT1_RGB
:
2203 case PIPE_FORMAT_DXT1_RGBA
:
2204 case PIPE_FORMAT_DXT1_SRGB
:
2205 case PIPE_FORMAT_DXT1_SRGBA
:
2206 rgba4
[count
] = s3tc_dxt1_to_rgba_aos(gallivm
, 4, format_desc
->format
,
2207 colors
, codewords
, i4
, j4
);
2209 case PIPE_FORMAT_DXT3_RGBA
:
2210 case PIPE_FORMAT_DXT3_SRGBA
:
2211 rgba4
[count
] = s3tc_dxt3_to_rgba_aos(gallivm
, 4, format_desc
->format
, colors
,
2212 codewords
, alpha_lo
, alpha_hi
, i4
, j4
);
2214 case PIPE_FORMAT_DXT5_RGBA
:
2215 case PIPE_FORMAT_DXT5_SRGBA
:
2216 rgba4
[count
] = s3tc_dxt5_to_rgba_aos(gallivm
, 4, format_desc
->format
, colors
,
2217 codewords
, alpha_lo
, alpha_hi
, i4
, j4
);
2221 rgba4
[count
] = LLVMGetUndef(LLVMVectorType(i8t
, 4));
2224 /* shuffles typically give best results with dword elements...*/
2225 rgba4
[count
] = LLVMBuildBitCast(builder
, rgba4
[count
], i324_vectype
, "");
2227 rgba
= lp_build_concat(gallivm
, rgba4
, lp_324_vectype
, n
/ 4);
2228 rgba
= LLVMBuildBitCast(builder
, rgba
, i8_vectype
, "");
2231 LLVMValueRef colors
, codewords
, alpha_lo
= NULL
, alpha_hi
= NULL
;
2233 lp_build_gather_s3tc(gallivm
, n
, format_desc
, &colors
, &codewords
,
2234 &alpha_lo
, &alpha_hi
, base_ptr
, offset
);
2236 switch (format_desc
->format
) {
2237 case PIPE_FORMAT_DXT1_RGB
:
2238 case PIPE_FORMAT_DXT1_RGBA
:
2239 case PIPE_FORMAT_DXT1_SRGB
:
2240 case PIPE_FORMAT_DXT1_SRGBA
:
2241 rgba
= s3tc_dxt1_to_rgba_aos(gallivm
, n
, format_desc
->format
,
2242 colors
, codewords
, i
, j
);
2244 case PIPE_FORMAT_DXT3_RGBA
:
2245 case PIPE_FORMAT_DXT3_SRGBA
:
2246 rgba
= s3tc_dxt3_to_rgba_aos(gallivm
, n
, format_desc
->format
, colors
,
2247 codewords
, alpha_lo
, alpha_hi
, i
, j
);
2249 case PIPE_FORMAT_DXT5_RGBA
:
2250 case PIPE_FORMAT_DXT5_SRGBA
:
2251 rgba
= s3tc_dxt5_to_rgba_aos(gallivm
, n
, format_desc
->format
, colors
,
2252 codewords
, alpha_lo
, alpha_hi
, i
, j
);
2256 rgba
= LLVMGetUndef(LLVMVectorType(i8t
, 4*n
));
2261 /* always return just decompressed values - srgb conversion is done later */