1 /**************************************************************************
2 *
3 * Copyright 2010-2018 VMware, Inc.
4 * All Rights Reserved.
5 *
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:
13 *
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.
21 *
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
24 * of the Software.
25 *
26 **************************************************************************/
27
28
29 /**
30 * @file
31 * s3tc pixel format manipulation.
32 *
33 * @author Roland Scheidegger <sroland@vmware.com>
34 */
35
36
37 #include <llvm/Config/llvm-config.h>
38
39 #include "util/format/u_format.h"
40 #include "util/u_math.h"
41 #include "util/u_string.h"
42 #include "util/u_cpu_detect.h"
43 #include "util/u_debug.h"
44
45 #include "lp_bld_arit.h"
46 #include "lp_bld_type.h"
47 #include "lp_bld_const.h"
48 #include "lp_bld_conv.h"
49 #include "lp_bld_gather.h"
50 #include "lp_bld_format.h"
51 #include "lp_bld_logic.h"
52 #include "lp_bld_pack.h"
53 #include "lp_bld_flow.h"
54 #include "lp_bld_printf.h"
55 #include "lp_bld_struct.h"
56 #include "lp_bld_swizzle.h"
57 #include "lp_bld_init.h"
58 #include "lp_bld_debug.h"
59 #include "lp_bld_intr.h"
60
61
62 /**
63 * Reverse an interleave2_half
64 * (ie. pick every second element, independent lower/upper halfs)
65 * sse2 can only do that with 32bit (shufps) or larger elements
66 * natively. (Otherwise, and/pack (even) or shift/pack (odd)
67 * could be used, ideally llvm would do that for us.)
68 * XXX: Unfortunately, this does NOT translate to a shufps if those
69 * are int vectors (and casting will not help, llvm needs to recognize it
70 * as "real" float). Instead, llvm will use a pshufd/pshufd/punpcklqdq
71 * sequence which I'm pretty sure is a lot worse despite domain transition
72 * penalties with shufps (except maybe on Nehalem).
73 */
74 static LLVMValueRef
lp_build_uninterleave2_half(struct gallivm_state * gallivm,struct lp_type type,LLVMValueRef a,LLVMValueRef b,unsigned lo_hi)75 lp_build_uninterleave2_half(struct gallivm_state *gallivm,
76 struct lp_type type,
77 LLVMValueRef a,
78 LLVMValueRef b,
79 unsigned lo_hi)
80 {
81 LLVMValueRef shuffle, elems[LP_MAX_VECTOR_LENGTH];
82 unsigned i;
83
84 assert(type.length <= LP_MAX_VECTOR_LENGTH);
85 assert(lo_hi < 2);
86
87 if (type.length * type.width == 256) {
88 assert(type.length == 8);
89 assert(type.width == 32);
90 static const unsigned shufvals[8] = {0, 2, 8, 10, 4, 6, 12, 14};
91 for (i = 0; i < type.length; ++i) {
92 elems[i] = lp_build_const_int32(gallivm, shufvals[i] + lo_hi);
93 }
94 } else {
95 for (i = 0; i < type.length; ++i) {
96 elems[i] = lp_build_const_int32(gallivm, 2*i + lo_hi);
97 }
98 }
99
100 shuffle = LLVMConstVector(elems, type.length);
101
102 return LLVMBuildShuffleVector(gallivm->builder, a, b, shuffle, "");
103
104 }
105
106
107 /**
108 * Build shuffle for extending vectors.
109 */
110 static LLVMValueRef
lp_build_const_extend_shuffle(struct gallivm_state * gallivm,unsigned n,unsigned length)111 lp_build_const_extend_shuffle(struct gallivm_state *gallivm,
112 unsigned n, unsigned length)
113 {
114 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
115 unsigned i;
116
117 assert(n <= length);
118 assert(length <= LP_MAX_VECTOR_LENGTH);
119
120 /* TODO: cache results in a static table */
121
122 for(i = 0; i < n; i++) {
123 elems[i] = lp_build_const_int32(gallivm, i);
124 }
125 for (i = n; i < length; i++) {
126 elems[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
127 }
128
129 return LLVMConstVector(elems, length);
130 }
131
132 static LLVMValueRef
lp_build_const_unpackx2_shuffle(struct gallivm_state * gallivm,unsigned n)133 lp_build_const_unpackx2_shuffle(struct gallivm_state *gallivm, unsigned n)
134 {
135 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
136 unsigned i, j;
137
138 assert(n <= LP_MAX_VECTOR_LENGTH);
139
140 /* TODO: cache results in a static table */
141
142 for(i = 0, j = 0; i < n; i += 2, ++j) {
143 elems[i + 0] = lp_build_const_int32(gallivm, 0 + j);
144 elems[i + 1] = lp_build_const_int32(gallivm, n + j);
145 elems[n + i + 0] = lp_build_const_int32(gallivm, 0 + n/2 + j);
146 elems[n + i + 1] = lp_build_const_int32(gallivm, n + n/2 + j);
147 }
148
149 return LLVMConstVector(elems, n * 2);
150 }
151
152 /*
153 * broadcast 1 element to all elements
154 */
155 static LLVMValueRef
lp_build_const_shuffle1(struct gallivm_state * gallivm,unsigned index,unsigned n)156 lp_build_const_shuffle1(struct gallivm_state *gallivm,
157 unsigned index, unsigned n)
158 {
159 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
160 unsigned i;
161
162 assert(n <= LP_MAX_VECTOR_LENGTH);
163
164 /* TODO: cache results in a static table */
165
166 for (i = 0; i < n; i++) {
167 elems[i] = lp_build_const_int32(gallivm, index);
168 }
169
170 return LLVMConstVector(elems, n);
171 }
172
173 /*
174 * move 1 element to pos 0, rest undef
175 */
176 static LLVMValueRef
lp_build_shuffle1undef(struct gallivm_state * gallivm,LLVMValueRef a,unsigned index,unsigned n)177 lp_build_shuffle1undef(struct gallivm_state *gallivm,
178 LLVMValueRef a, unsigned index, unsigned n)
179 {
180 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH], shuf;
181 unsigned i;
182
183 assert(n <= LP_MAX_VECTOR_LENGTH);
184
185 elems[0] = lp_build_const_int32(gallivm, index);
186
187 for (i = 1; i < n; i++) {
188 elems[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
189 }
190 shuf = LLVMConstVector(elems, n);
191
192 return LLVMBuildShuffleVector(gallivm->builder, a, a, shuf, "");
193 }
194
195 static boolean
format_dxt1_variant(enum pipe_format format)196 format_dxt1_variant(enum pipe_format format)
197 {
198 return format == PIPE_FORMAT_DXT1_RGB ||
199 format == PIPE_FORMAT_DXT1_RGBA ||
200 format == PIPE_FORMAT_DXT1_SRGB ||
201 format == PIPE_FORMAT_DXT1_SRGBA;
202
203 }
204
205 /**
206 * Gather elements from scatter positions in memory into vectors.
207 * This is customised for fetching texels from s3tc textures.
208 * For SSE, typical value is length=4.
209 *
210 * @param length length of the offsets
211 * @param colors the stored colors of the blocks will be extracted into this.
212 * @param codewords the codewords of the blocks will be extracted into this.
213 * @param alpha_lo used for storing lower 32bit of alpha components for dxt3/5
214 * @param alpha_hi used for storing higher 32bit of alpha components for dxt3/5
215 * @param base_ptr base pointer, should be a i8 pointer type.
216 * @param offsets vector with offsets
217 */
218 static void
lp_build_gather_s3tc(struct gallivm_state * gallivm,unsigned length,const struct util_format_description * format_desc,LLVMValueRef * colors,LLVMValueRef * codewords,LLVMValueRef * alpha_lo,LLVMValueRef * alpha_hi,LLVMValueRef base_ptr,LLVMValueRef offsets)219 lp_build_gather_s3tc(struct gallivm_state *gallivm,
220 unsigned length,
221 const struct util_format_description *format_desc,
222 LLVMValueRef *colors,
223 LLVMValueRef *codewords,
224 LLVMValueRef *alpha_lo,
225 LLVMValueRef *alpha_hi,
226 LLVMValueRef base_ptr,
227 LLVMValueRef offsets)
228 {
229 LLVMBuilderRef builder = gallivm->builder;
230 unsigned block_bits = format_desc->block.bits;
231 unsigned i;
232 LLVMValueRef elems[8];
233 LLVMTypeRef type32 = LLVMInt32TypeInContext(gallivm->context);
234 LLVMTypeRef type64 = LLVMInt64TypeInContext(gallivm->context);
235 LLVMTypeRef type32dxt;
236 struct lp_type lp_type32dxt;
237
238 memset(&lp_type32dxt, 0, sizeof lp_type32dxt);
239 lp_type32dxt.width = 32;
240 lp_type32dxt.length = block_bits / 32;
241 type32dxt = lp_build_vec_type(gallivm, lp_type32dxt);
242
243 assert(block_bits == 64 || block_bits == 128);
244 assert(length == 1 || length == 4 || length == 8);
245
246 for (i = 0; i < length; ++i) {
247 elems[i] = lp_build_gather_elem(gallivm, length,
248 block_bits, block_bits, TRUE,
249 base_ptr, offsets, i, FALSE);
250 elems[i] = LLVMBuildBitCast(builder, elems[i], type32dxt, "");
251 }
252 if (length == 1) {
253 LLVMValueRef elem = elems[0];
254 if (block_bits == 128) {
255 *alpha_lo = LLVMBuildExtractElement(builder, elem,
256 lp_build_const_int32(gallivm, 0), "");
257 *alpha_hi = LLVMBuildExtractElement(builder, elem,
258 lp_build_const_int32(gallivm, 1), "");
259 *colors = LLVMBuildExtractElement(builder, elem,
260 lp_build_const_int32(gallivm, 2), "");
261 *codewords = LLVMBuildExtractElement(builder, elem,
262 lp_build_const_int32(gallivm, 3), "");
263 }
264 else {
265 *alpha_lo = LLVMGetUndef(type32);
266 *alpha_hi = LLVMGetUndef(type32);
267 *colors = LLVMBuildExtractElement(builder, elem,
268 lp_build_const_int32(gallivm, 0), "");
269 *codewords = LLVMBuildExtractElement(builder, elem,
270 lp_build_const_int32(gallivm, 1), "");
271 }
272 }
273 else {
274 LLVMValueRef tmp[4], cc01, cc23;
275 struct lp_type lp_type32, lp_type64;
276 memset(&lp_type32, 0, sizeof lp_type32);
277 lp_type32.width = 32;
278 lp_type32.length = length;
279 memset(&lp_type64, 0, sizeof lp_type64);
280 lp_type64.width = 64;
281 lp_type64.length = length/2;
282
283 if (block_bits == 128) {
284 if (length == 8) {
285 for (i = 0; i < 4; ++i) {
286 tmp[0] = elems[i];
287 tmp[1] = elems[i+4];
288 elems[i] = lp_build_concat(gallivm, tmp, lp_type32dxt, 2);
289 }
290 }
291 lp_build_transpose_aos(gallivm, lp_type32, elems, tmp);
292 *colors = tmp[2];
293 *codewords = tmp[3];
294 *alpha_lo = tmp[0];
295 *alpha_hi = tmp[1];
296 } else {
297 LLVMTypeRef type64_vec = LLVMVectorType(type64, length/2);
298 LLVMTypeRef type32_vec = LLVMVectorType(type32, length);
299
300 for (i = 0; i < length; ++i) {
301 /* no-op shuffle */
302 elems[i] = LLVMBuildShuffleVector(builder, elems[i],
303 LLVMGetUndef(type32dxt),
304 lp_build_const_extend_shuffle(gallivm, 2, 4), "");
305 }
306 if (length == 8) {
307 struct lp_type lp_type32_4 = {0};
308 lp_type32_4.width = 32;
309 lp_type32_4.length = 4;
310 for (i = 0; i < 4; ++i) {
311 tmp[0] = elems[i];
312 tmp[1] = elems[i+4];
313 elems[i] = lp_build_concat(gallivm, tmp, lp_type32_4, 2);
314 }
315 }
316 cc01 = lp_build_interleave2_half(gallivm, lp_type32, elems[0], elems[1], 0);
317 cc23 = lp_build_interleave2_half(gallivm, lp_type32, elems[2], elems[3], 0);
318 cc01 = LLVMBuildBitCast(builder, cc01, type64_vec, "");
319 cc23 = LLVMBuildBitCast(builder, cc23, type64_vec, "");
320 *colors = lp_build_interleave2_half(gallivm, lp_type64, cc01, cc23, 0);
321 *codewords = lp_build_interleave2_half(gallivm, lp_type64, cc01, cc23, 1);
322 *colors = LLVMBuildBitCast(builder, *colors, type32_vec, "");
323 *codewords = LLVMBuildBitCast(builder, *codewords, type32_vec, "");
324 }
325 }
326 }
327
328 /** Convert from <n x i32> containing 2 x n rgb565 colors
329 * to 2 <n x i32> rgba8888 colors
330 * This is the most optimized version I can think of
331 * should be nearly as fast as decoding only one color
332 * NOTE: alpha channel will be set to 0
333 * @param colors is a <n x i32> vector containing the rgb565 colors
334 */
335 static void
color_expand2_565_to_8888(struct gallivm_state * gallivm,unsigned n,LLVMValueRef colors,LLVMValueRef * color0,LLVMValueRef * color1)336 color_expand2_565_to_8888(struct gallivm_state *gallivm,
337 unsigned n,
338 LLVMValueRef colors,
339 LLVMValueRef *color0,
340 LLVMValueRef *color1)
341 {
342 LLVMBuilderRef builder = gallivm->builder;
343 LLVMValueRef r, g, b, rblo, glo;
344 LLVMValueRef rgblomask, rb, rgb0, rgb1;
345 struct lp_type type, type16, type8;
346
347 assert(n > 1);
348
349 memset(&type, 0, sizeof type);
350 type.width = 32;
351 type.length = n;
352
353 memset(&type16, 0, sizeof type16);
354 type16.width = 16;
355 type16.length = 2 * n;
356
357 memset(&type8, 0, sizeof type8);
358 type8.width = 8;
359 type8.length = 4 * n;
360
361 rgblomask = lp_build_const_int_vec(gallivm, type16, 0x0707);
362 colors = LLVMBuildBitCast(builder, colors,
363 lp_build_vec_type(gallivm, type16), "");
364 /* move r into low 8 bits, b into high 8 bits, g into another reg (low bits)
365 * make sure low bits of r are zero - could use AND but requires constant */
366 r = LLVMBuildLShr(builder, colors, lp_build_const_int_vec(gallivm, type16, 11), "");
367 r = LLVMBuildShl(builder, r, lp_build_const_int_vec(gallivm, type16, 3), "");
368 b = LLVMBuildShl(builder, colors, lp_build_const_int_vec(gallivm, type16, 11), "");
369 rb = LLVMBuildOr(builder, r, b, "");
370 rblo = LLVMBuildLShr(builder, rb, lp_build_const_int_vec(gallivm, type16, 5), "");
371 /* don't have byte shift hence need mask */
372 rblo = LLVMBuildAnd(builder, rblo, rgblomask, "");
373 rb = LLVMBuildOr(builder, rb, rblo, "");
374
375 /* make sure low bits of g are zero */
376 g = LLVMBuildAnd(builder, colors, lp_build_const_int_vec(gallivm, type16, 0x07e0), "");
377 g = LLVMBuildLShr(builder, g, lp_build_const_int_vec(gallivm, type16, 3), "");
378 glo = LLVMBuildLShr(builder, g, lp_build_const_int_vec(gallivm, type16, 6), "");
379 g = LLVMBuildOr(builder, g, glo, "");
380
381 rb = LLVMBuildBitCast(builder, rb, lp_build_vec_type(gallivm, type8), "");
382 g = LLVMBuildBitCast(builder, g, lp_build_vec_type(gallivm, type8), "");
383 rgb0 = lp_build_interleave2_half(gallivm, type8, rb, g, 0);
384 rgb1 = lp_build_interleave2_half(gallivm, type8, rb, g, 1);
385
386 rgb0 = LLVMBuildBitCast(builder, rgb0, lp_build_vec_type(gallivm, type), "");
387 rgb1 = LLVMBuildBitCast(builder, rgb1, lp_build_vec_type(gallivm, type), "");
388
389 /* rgb0 is rgb00, rgb01, rgb10, rgb11
390 * instead of rgb00, rgb10, rgb20, rgb30 hence need reshuffle
391 * on x86 this _should_ just generate one shufps...
392 */
393 *color0 = lp_build_uninterleave2_half(gallivm, type, rgb0, rgb1, 0);
394 *color1 = lp_build_uninterleave2_half(gallivm, type, rgb0, rgb1, 1);
395 }
396
397
398 /** Convert from <n x i32> containing rgb565 colors
399 * (in first 16 bits) to <n x i32> rgba8888 colors
400 * bits 16-31 MBZ
401 * NOTE: alpha channel will be set to 0
402 * @param colors is a <n x i32> vector containing the rgb565 colors
403 */
404 static LLVMValueRef
color_expand_565_to_8888(struct gallivm_state * gallivm,unsigned n,LLVMValueRef colors)405 color_expand_565_to_8888(struct gallivm_state *gallivm,
406 unsigned n,
407 LLVMValueRef colors)
408 {
409 LLVMBuilderRef builder = gallivm->builder;
410 LLVMValueRef rgba, r, g, b, rgblo, glo;
411 LLVMValueRef rbhimask, g6mask, rgblomask;
412 struct lp_type type;
413 memset(&type, 0, sizeof type);
414 type.width = 32;
415 type.length = n;
416
417 /* color expansion:
418 * first extract and shift colors into their final locations
419 * (high bits - low bits zero at this point)
420 * then replicate highest bits to the lowest bits
421 * note rb replication can be done in parallel but not g
422 * (different shift)
423 * r5mask = 0xf800, g6mask = 0x07e0, b5mask = 0x001f
424 * rhigh = 8, ghigh = 5, bhigh = 19
425 * rblow = 5, glow = 6
426 * rgblowmask = 0x00070307
427 * r = colors >> rhigh
428 * b = colors << bhigh
429 * g = (colors & g6mask) << ghigh
430 * rb = (r | b) rbhimask
431 * rbtmp = rb >> rblow
432 * gtmp = rb >> glow
433 * rbtmp = rbtmp | gtmp
434 * rbtmp = rbtmp & rgblowmask
435 * rgb = rb | g | rbtmp
436 */
437 g6mask = lp_build_const_int_vec(gallivm, type, 0x07e0);
438 rbhimask = lp_build_const_int_vec(gallivm, type, 0x00f800f8);
439 rgblomask = lp_build_const_int_vec(gallivm, type, 0x00070307);
440
441 r = LLVMBuildLShr(builder, colors, lp_build_const_int_vec(gallivm, type, 8), "");
442 b = LLVMBuildShl(builder, colors, lp_build_const_int_vec(gallivm, type, 19), "");
443 g = LLVMBuildAnd(builder, colors, g6mask, "");
444 g = LLVMBuildShl(builder, g, lp_build_const_int_vec(gallivm, type, 5), "");
445 rgba = LLVMBuildOr(builder, r, b, "");
446 rgba = LLVMBuildAnd(builder, rgba, rbhimask, "");
447 rgblo = LLVMBuildLShr(builder, rgba, lp_build_const_int_vec(gallivm, type, 5), "");
448 glo = LLVMBuildLShr(builder, g, lp_build_const_int_vec(gallivm, type, 6), "");
449 rgblo = LLVMBuildOr(builder, rgblo, glo, "");
450 rgblo = LLVMBuildAnd(builder, rgblo, rgblomask, "");
451 rgba = LLVMBuildOr(builder, rgba, g, "");
452 rgba = LLVMBuildOr(builder, rgba, rgblo, "");
453
454 return rgba;
455 }
456
457
458 /*
459 * Average two byte vectors. (Will always round up.)
460 */
461 static LLVMValueRef
lp_build_pavgb(struct lp_build_context * bld8,LLVMValueRef v0,LLVMValueRef v1)462 lp_build_pavgb(struct lp_build_context *bld8,
463 LLVMValueRef v0,
464 LLVMValueRef v1)
465 {
466 struct gallivm_state *gallivm = bld8->gallivm;
467 LLVMBuilderRef builder = gallivm->builder;
468 assert(bld8->type.width == 8);
469 assert(bld8->type.length == 16 || bld8->type.length == 32);
470 if (LLVM_VERSION_MAJOR < 6) {
471 LLVMValueRef intrargs[2];
472 char *intr_name = bld8->type.length == 32 ? "llvm.x86.avx2.pavg.b" :
473 "llvm.x86.sse2.pavg.b";
474 intrargs[0] = v0;
475 intrargs[1] = v1;
476 return lp_build_intrinsic(builder, intr_name,
477 bld8->vec_type, intrargs, 2, 0);
478 } else {
479 /*
480 * Must match llvm's autoupgrade of pavg.b intrinsic to be useful.
481 * You better hope the backend code manages to detect the pattern, and
482 * the pattern doesn't change there...
483 */
484 struct lp_type type_ext = bld8->type;
485 LLVMTypeRef vec_type_ext;
486 LLVMValueRef res;
487 LLVMValueRef ext_one;
488 type_ext.width = 16;
489 vec_type_ext = lp_build_vec_type(gallivm, type_ext);
490 ext_one = lp_build_const_vec(gallivm, type_ext, 1);
491
492 v0 = LLVMBuildZExt(builder, v0, vec_type_ext, "");
493 v1 = LLVMBuildZExt(builder, v1, vec_type_ext, "");
494 res = LLVMBuildAdd(builder, v0, v1, "");
495 res = LLVMBuildAdd(builder, res, ext_one, "");
496 res = LLVMBuildLShr(builder, res, ext_one, "");
497 res = LLVMBuildTrunc(builder, res, bld8->vec_type, "");
498 return res;
499 }
500 }
501
502 /**
503 * Calculate 1/3(v1-v0) + v0
504 * and 2*1/3(v1-v0) + v0
505 */
506 static void
lp_build_lerp23(struct lp_build_context * bld,LLVMValueRef v0,LLVMValueRef v1,LLVMValueRef * res0,LLVMValueRef * res1)507 lp_build_lerp23(struct lp_build_context *bld,
508 LLVMValueRef v0,
509 LLVMValueRef v1,
510 LLVMValueRef *res0,
511 LLVMValueRef *res1)
512 {
513 struct gallivm_state *gallivm = bld->gallivm;
514 LLVMValueRef x, x_lo, x_hi, delta_lo, delta_hi;
515 LLVMValueRef mul_lo, mul_hi, v0_lo, v0_hi, v1_lo, v1_hi, tmp;
516 const struct lp_type type = bld->type;
517 LLVMBuilderRef builder = bld->gallivm->builder;
518 struct lp_type i16_type = lp_wider_type(type);
519 struct lp_build_context bld2;
520
521 assert(lp_check_value(type, v0));
522 assert(lp_check_value(type, v1));
523 assert(!type.floating && !type.fixed && !type.norm && type.width == 8);
524
525 lp_build_context_init(&bld2, gallivm, i16_type);
526 bld2.type.sign = TRUE;
527 x = lp_build_const_int_vec(gallivm, bld->type, 255*1/3);
528
529 /* FIXME: use native avx256 unpack/pack */
530 lp_build_unpack2(gallivm, type, i16_type, x, &x_lo, &x_hi);
531 lp_build_unpack2(gallivm, type, i16_type, v0, &v0_lo, &v0_hi);
532 lp_build_unpack2(gallivm, type, i16_type, v1, &v1_lo, &v1_hi);
533 delta_lo = lp_build_sub(&bld2, v1_lo, v0_lo);
534 delta_hi = lp_build_sub(&bld2, v1_hi, v0_hi);
535
536 mul_lo = LLVMBuildMul(builder, x_lo, delta_lo, "");
537 mul_hi = LLVMBuildMul(builder, x_hi, delta_hi, "");
538
539 x_lo = LLVMBuildLShr(builder, mul_lo, lp_build_const_int_vec(gallivm, i16_type, 8), "");
540 x_hi = LLVMBuildLShr(builder, mul_hi, lp_build_const_int_vec(gallivm, i16_type, 8), "");
541 /* lerp optimization: pack now, do add afterwards */
542 tmp = lp_build_pack2(gallivm, i16_type, type, x_lo, x_hi);
543 *res0 = lp_build_add(bld, tmp, v0);
544
545 x_lo = LLVMBuildLShr(builder, mul_lo, lp_build_const_int_vec(gallivm, i16_type, 7), "");
546 x_hi = LLVMBuildLShr(builder, mul_hi, lp_build_const_int_vec(gallivm, i16_type, 7), "");
547 /* unlike above still need mask (but add still afterwards). */
548 x_lo = LLVMBuildAnd(builder, x_lo, lp_build_const_int_vec(gallivm, i16_type, 0xff), "");
549 x_hi = LLVMBuildAnd(builder, x_hi, lp_build_const_int_vec(gallivm, i16_type, 0xff), "");
550 tmp = lp_build_pack2(gallivm, i16_type, type, x_lo, x_hi);
551 *res1 = lp_build_add(bld, tmp, v0);
552 }
553
554 /**
555 * Convert from <n x i64> s3tc dxt1 to <4n x i8> RGBA AoS
556 * @param colors is a <n x i32> vector with n x 2x16bit colors
557 * @param codewords is a <n x i32> vector containing the codewords
558 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
559 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
560 */
561 static LLVMValueRef
s3tc_dxt1_full_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef colors,LLVMValueRef codewords,LLVMValueRef i,LLVMValueRef j)562 s3tc_dxt1_full_to_rgba_aos(struct gallivm_state *gallivm,
563 unsigned n,
564 enum pipe_format format,
565 LLVMValueRef colors,
566 LLVMValueRef codewords,
567 LLVMValueRef i,
568 LLVMValueRef j)
569 {
570 LLVMBuilderRef builder = gallivm->builder;
571 LLVMValueRef color0, color1, color2, color3, color2_2, color3_2;
572 LLVMValueRef rgba, a, colors0, colors1, col0, col1, const2;
573 LLVMValueRef bit_pos, sel_mask, sel_lo, sel_hi, indices;
574 struct lp_type type, type8;
575 struct lp_build_context bld8, bld32;
576 boolean is_dxt1_variant = format_dxt1_variant(format);
577
578 memset(&type, 0, sizeof type);
579 type.width = 32;
580 type.length = n;
581
582 memset(&type8, 0, sizeof type8);
583 type8.width = 8;
584 type8.length = 4*n;
585
586 assert(lp_check_value(type, i));
587 assert(lp_check_value(type, j));
588
589 a = lp_build_const_int_vec(gallivm, type, 0xff000000);
590
591 lp_build_context_init(&bld32, gallivm, type);
592 lp_build_context_init(&bld8, gallivm, type8);
593
594 /*
595 * works as follows:
596 * - expand color0/color1 to rgba8888
597 * - calculate color2/3 (interpolation) according to color0 < color1 rules
598 * - calculate color2/3 according to color0 >= color1 rules
599 * - do selection of color2/3 according to comparison of color0/1
600 * - extract indices (vector shift).
601 * - use compare/select to select the correct color. Since we have 2bit
602 * indices (and 4 colors), needs at least three compare/selects.
603 */
604 /*
605 * expand the two colors
606 */
607 col0 = LLVMBuildAnd(builder, colors, lp_build_const_int_vec(gallivm, type, 0x0000ffff), "");
608 col1 = LLVMBuildLShr(builder, colors, lp_build_const_int_vec(gallivm, type, 16), "");
609 if (n > 1) {
610 color_expand2_565_to_8888(gallivm, n, colors, &color0, &color1);
611 }
612 else {
613 color0 = color_expand_565_to_8888(gallivm, n, col0);
614 color1 = color_expand_565_to_8888(gallivm, n, col1);
615 }
616
617 /*
618 * interpolate colors
619 * color2_1 is 2/3 color0 + 1/3 color1
620 * color3_1 is 1/3 color0 + 2/3 color1
621 * color2_2 is 1/2 color0 + 1/2 color1
622 * color3_2 is 0
623 */
624
625 colors0 = LLVMBuildBitCast(builder, color0, bld8.vec_type, "");
626 colors1 = LLVMBuildBitCast(builder, color1, bld8.vec_type, "");
627 /* can combine 2 lerps into one mostly - still looks expensive enough. */
628 lp_build_lerp23(&bld8, colors0, colors1, &color2, &color3);
629 color2 = LLVMBuildBitCast(builder, color2, bld32.vec_type, "");
630 color3 = LLVMBuildBitCast(builder, color3, bld32.vec_type, "");
631
632 /* dxt3/5 always use 4-color encoding */
633 if (is_dxt1_variant) {
634 /* fix up alpha */
635 if (format == PIPE_FORMAT_DXT1_RGBA ||
636 format == PIPE_FORMAT_DXT1_SRGBA) {
637 color0 = LLVMBuildOr(builder, color0, a, "");
638 color1 = LLVMBuildOr(builder, color1, a, "");
639 color3 = LLVMBuildOr(builder, color3, a, "");
640 }
641 /*
642 * XXX with sse2 and 16x8 vectors, should use pavgb even when n == 1.
643 * Much cheaper (but we don't care that much if n == 1).
644 */
645 if ((util_get_cpu_caps()->has_sse2 && n == 4) ||
646 (util_get_cpu_caps()->has_avx2 && n == 8)) {
647 color2_2 = lp_build_pavgb(&bld8, colors0, colors1);
648 color2_2 = LLVMBuildBitCast(builder, color2_2, bld32.vec_type, "");
649 }
650 else {
651 struct lp_type i16_type = lp_wider_type(type8);
652 struct lp_build_context bld2;
653 LLVMValueRef v0_lo, v0_hi, v1_lo, v1_hi, addlo, addhi;
654
655 lp_build_context_init(&bld2, gallivm, i16_type);
656 bld2.type.sign = TRUE;
657
658 /*
659 * This isn't as expensive as it looks (the unpack is the same as
660 * for lerp23), with correct rounding.
661 * (Note that while rounding is correct, this will always round down,
662 * whereas pavgb will always round up.)
663 */
664 /* FIXME: use native avx256 unpack/pack */
665 lp_build_unpack2(gallivm, type8, i16_type, colors0, &v0_lo, &v0_hi);
666 lp_build_unpack2(gallivm, type8, i16_type, colors1, &v1_lo, &v1_hi);
667
668 addlo = lp_build_add(&bld2, v0_lo, v1_lo);
669 addhi = lp_build_add(&bld2, v0_hi, v1_hi);
670 addlo = LLVMBuildLShr(builder, addlo,
671 lp_build_const_int_vec(gallivm, i16_type, 1), "");
672 addhi = LLVMBuildLShr(builder, addhi,
673 lp_build_const_int_vec(gallivm, i16_type, 1), "");
674 color2_2 = lp_build_pack2(gallivm, i16_type, type8, addlo, addhi);
675 color2_2 = LLVMBuildBitCast(builder, color2_2, bld32.vec_type, "");
676 }
677 color3_2 = lp_build_const_int_vec(gallivm, type, 0);
678
679 /* select between colors2/3 */
680 /* signed compare is faster saves some xors */
681 type.sign = TRUE;
682 sel_mask = lp_build_compare(gallivm, type, PIPE_FUNC_GREATER, col0, col1);
683 color2 = lp_build_select(&bld32, sel_mask, color2, color2_2);
684 color3 = lp_build_select(&bld32, sel_mask, color3, color3_2);
685 type.sign = FALSE;
686
687 if (format == PIPE_FORMAT_DXT1_RGBA ||
688 format == PIPE_FORMAT_DXT1_SRGBA) {
689 color2 = LLVMBuildOr(builder, color2, a, "");
690 }
691 }
692
693 const2 = lp_build_const_int_vec(gallivm, type, 2);
694 /* extract 2-bit index values */
695 bit_pos = LLVMBuildShl(builder, j, const2, "");
696 bit_pos = LLVMBuildAdd(builder, bit_pos, i, "");
697 bit_pos = LLVMBuildAdd(builder, bit_pos, bit_pos, "");
698 /*
699 * NOTE: This innocent looking shift is very expensive with x86/ssex.
700 * Shifts with per-elemnent shift count get roughly translated to
701 * extract (count), extract (value), shift, move (back to xmm), unpack
702 * per element!
703 * So about 20 instructions here for 4xi32.
704 * Newer llvm versions (3.7+) will not do extract/insert but use a
705 * a couple constant count vector shifts plus shuffles. About same
706 * amount of instructions unfortunately...
707 * Would get much worse with 8xi16 even...
708 * We could actually do better here:
709 * - subtract bit_pos from 128+30, shl 23, convert float to int...
710 * - now do mul with codewords followed by shr 30...
711 * But requires 32bit->32bit mul, sse41 only (well that's emulatable
712 * with 2 32bit->64bit muls...) and not exactly cheap
713 * AVX2, of course, fixes this nonsense.
714 */
715 indices = LLVMBuildLShr(builder, codewords, bit_pos, "");
716
717 /* finally select the colors */
718 sel_lo = LLVMBuildAnd(builder, indices, bld32.one, "");
719 sel_lo = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL, sel_lo, bld32.one);
720 color0 = lp_build_select(&bld32, sel_lo, color1, color0);
721 color2 = lp_build_select(&bld32, sel_lo, color3, color2);
722 sel_hi = LLVMBuildAnd(builder, indices, const2, "");
723 sel_hi = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL, sel_hi, const2);
724 rgba = lp_build_select(&bld32, sel_hi, color2, color0);
725
726 /* fix up alpha */
727 if (format == PIPE_FORMAT_DXT1_RGB ||
728 format == PIPE_FORMAT_DXT1_SRGB) {
729 rgba = LLVMBuildOr(builder, rgba, a, "");
730 }
731 return LLVMBuildBitCast(builder, rgba, bld8.vec_type, "");
732 }
733
734
735 static LLVMValueRef
s3tc_dxt1_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef colors,LLVMValueRef codewords,LLVMValueRef i,LLVMValueRef j)736 s3tc_dxt1_to_rgba_aos(struct gallivm_state *gallivm,
737 unsigned n,
738 enum pipe_format format,
739 LLVMValueRef colors,
740 LLVMValueRef codewords,
741 LLVMValueRef i,
742 LLVMValueRef j)
743 {
744 return s3tc_dxt1_full_to_rgba_aos(gallivm, n, format,
745 colors, codewords, i, j);
746 }
747
748
749 /**
750 * Convert from <n x i128> s3tc dxt3 to <4n x i8> RGBA AoS
751 * @param colors is a <n x i32> vector with n x 2x16bit colors
752 * @param codewords is a <n x i32> vector containing the codewords
753 * @param alphas is a <n x i64> vector containing the alpha values
754 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
755 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
756 */
757 static LLVMValueRef
s3tc_dxt3_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef colors,LLVMValueRef codewords,LLVMValueRef alpha_low,LLVMValueRef alpha_hi,LLVMValueRef i,LLVMValueRef j)758 s3tc_dxt3_to_rgba_aos(struct gallivm_state *gallivm,
759 unsigned n,
760 enum pipe_format format,
761 LLVMValueRef colors,
762 LLVMValueRef codewords,
763 LLVMValueRef alpha_low,
764 LLVMValueRef alpha_hi,
765 LLVMValueRef i,
766 LLVMValueRef j)
767 {
768 LLVMBuilderRef builder = gallivm->builder;
769 LLVMValueRef rgba, tmp, tmp2;
770 LLVMValueRef bit_pos, sel_mask;
771 struct lp_type type, type8;
772 struct lp_build_context bld;
773
774 memset(&type, 0, sizeof type);
775 type.width = 32;
776 type.length = n;
777
778 memset(&type8, 0, sizeof type8);
779 type8.width = 8;
780 type8.length = n*4;
781
782 assert(lp_check_value(type, i));
783 assert(lp_check_value(type, j));
784
785 lp_build_context_init(&bld, gallivm, type);
786
787 rgba = s3tc_dxt1_to_rgba_aos(gallivm, n, format,
788 colors, codewords, i, j);
789
790 rgba = LLVMBuildBitCast(builder, rgba, bld.vec_type, "");
791
792 /*
793 * Extract alpha values. Since we now need to select from
794 * which 32bit vector values are fetched, construct selection
795 * mask from highest bit of bit_pos, and use select, then shift
796 * according to the bit_pos (without the highest bit).
797 * Note this is pointless for n == 1 case. Could just
798 * directly use 64bit arithmetic if we'd extract 64bit
799 * alpha value instead of 2x32...
800 */
801 /* pos = 4*(4j+i) */
802 bit_pos = LLVMBuildShl(builder, j, lp_build_const_int_vec(gallivm, type, 2), "");
803 bit_pos = LLVMBuildAdd(builder, bit_pos, i, "");
804 bit_pos = LLVMBuildShl(builder, bit_pos,
805 lp_build_const_int_vec(gallivm, type, 2), "");
806 sel_mask = LLVMBuildLShr(builder, bit_pos,
807 lp_build_const_int_vec(gallivm, type, 5), "");
808 sel_mask = LLVMBuildSub(builder, sel_mask, bld.one, "");
809 tmp = lp_build_select(&bld, sel_mask, alpha_low, alpha_hi);
810 bit_pos = LLVMBuildAnd(builder, bit_pos,
811 lp_build_const_int_vec(gallivm, type, 0xffffffdf), "");
812 /* Warning: slow shift with per element count (without avx2) */
813 /*
814 * Could do pshufb here as well - just use appropriate 2 bits in bit_pos
815 * to select the right byte with pshufb. Then for the remaining one bit
816 * just do shift/select.
817 */
818 tmp = LLVMBuildLShr(builder, tmp, bit_pos, "");
819
820 /* combined expand from a4 to a8 and shift into position */
821 tmp = LLVMBuildShl(builder, tmp, lp_build_const_int_vec(gallivm, type, 28), "");
822 tmp2 = LLVMBuildLShr(builder, tmp, lp_build_const_int_vec(gallivm, type, 4), "");
823 tmp = LLVMBuildOr(builder, tmp, tmp2, "");
824
825 rgba = LLVMBuildOr(builder, tmp, rgba, "");
826
827 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
828 }
829
830 static LLVMValueRef
lp_build_lerpdxta(struct gallivm_state * gallivm,LLVMValueRef alpha0,LLVMValueRef alpha1,LLVMValueRef code,LLVMValueRef sel_mask,unsigned n)831 lp_build_lerpdxta(struct gallivm_state *gallivm,
832 LLVMValueRef alpha0,
833 LLVMValueRef alpha1,
834 LLVMValueRef code,
835 LLVMValueRef sel_mask,
836 unsigned n)
837 {
838 /*
839 * note we're doing lerp in 16bit since 32bit pmulld is only available in sse41
840 * (plus pmullw is actually faster...)
841 * we just pretend our 32bit values (which are really only 8bit) are 16bits.
842 * Note that this is obviously a disaster for the scalar case.
843 */
844 LLVMBuilderRef builder = gallivm->builder;
845 LLVMValueRef delta, ainterp;
846 LLVMValueRef weight5, weight7, weight;
847 struct lp_type type32, type16, type8;
848 struct lp_build_context bld16;
849
850 memset(&type32, 0, sizeof type32);
851 type32.width = 32;
852 type32.length = n;
853 memset(&type16, 0, sizeof type16);
854 type16.width = 16;
855 type16.length = 2*n;
856 type16.sign = TRUE;
857 memset(&type8, 0, sizeof type8);
858 type8.width = 8;
859 type8.length = 4*n;
860
861 lp_build_context_init(&bld16, gallivm, type16);
862 /* 255/7 is a bit off - increase accuracy at the expense of shift later */
863 sel_mask = LLVMBuildBitCast(builder, sel_mask, bld16.vec_type, "");
864 weight5 = lp_build_const_int_vec(gallivm, type16, 255*64/5);
865 weight7 = lp_build_const_int_vec(gallivm, type16, 255*64/7);
866 weight = lp_build_select(&bld16, sel_mask, weight7, weight5);
867
868 alpha0 = LLVMBuildBitCast(builder, alpha0, bld16.vec_type, "");
869 alpha1 = LLVMBuildBitCast(builder, alpha1, bld16.vec_type, "");
870 code = LLVMBuildBitCast(builder, code, bld16.vec_type, "");
871 /* we'll get garbage in the elements which had code 0 (or larger than 5 or 7)
872 but we don't care */
873 code = LLVMBuildSub(builder, code, bld16.one, "");
874
875 weight = LLVMBuildMul(builder, weight, code, "");
876 weight = LLVMBuildLShr(builder, weight,
877 lp_build_const_int_vec(gallivm, type16, 6), "");
878
879 delta = LLVMBuildSub(builder, alpha1, alpha0, "");
880
881 ainterp = LLVMBuildMul(builder, delta, weight, "");
882 ainterp = LLVMBuildLShr(builder, ainterp,
883 lp_build_const_int_vec(gallivm, type16, 8), "");
884
885 ainterp = LLVMBuildBitCast(builder, ainterp, lp_build_vec_type(gallivm, type8), "");
886 alpha0 = LLVMBuildBitCast(builder, alpha0, lp_build_vec_type(gallivm, type8), "");
887 ainterp = LLVMBuildAdd(builder, alpha0, ainterp, "");
888 ainterp = LLVMBuildBitCast(builder, ainterp, lp_build_vec_type(gallivm, type32), "");
889
890 return ainterp;
891 }
892
893 static LLVMValueRef
s3tc_dxt5_alpha_channel(struct gallivm_state * gallivm,bool is_signed,unsigned n,LLVMValueRef alpha_hi,LLVMValueRef alpha_lo,LLVMValueRef i,LLVMValueRef j)894 s3tc_dxt5_alpha_channel(struct gallivm_state *gallivm,
895 bool is_signed,
896 unsigned n,
897 LLVMValueRef alpha_hi, LLVMValueRef alpha_lo,
898 LLVMValueRef i, LLVMValueRef j)
899 {
900 LLVMBuilderRef builder = gallivm->builder;
901 struct lp_type type, type8;
902 LLVMValueRef tmp, alpha0, alpha1, alphac, alphac0, bit_pos, shift;
903 LLVMValueRef sel_mask, tmp_mask, alpha, alpha64, code_s;
904 LLVMValueRef mask6, mask7, ainterp;
905 LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
906 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
907 struct lp_build_context bld32;
908
909 memset(&type, 0, sizeof type);
910 type.width = 32;
911 type.length = n;
912
913 memset(&type8, 0, sizeof type8);
914 type8.width = 8;
915 type8.length = n;
916 type8.sign = is_signed;
917
918 lp_build_context_init(&bld32, gallivm, type);
919 /* this looks pretty complex for vectorization:
920 * extract a0/a1 values
921 * extract code
922 * select weights for interpolation depending on a0 > a1
923 * mul weights by code - 1
924 * lerp a0/a1/weights
925 * use selects for getting either a0, a1, interp a, interp a/0.0, interp a/1.0
926 */
927
928 alpha0 = LLVMBuildAnd(builder, alpha_lo,
929 lp_build_const_int_vec(gallivm, type, 0xff), "");
930 if (is_signed) {
931 alpha0 = LLVMBuildTrunc(builder, alpha0, lp_build_vec_type(gallivm, type8), "");
932 alpha0 = LLVMBuildSExt(builder, alpha0, lp_build_vec_type(gallivm, type), "");
933 }
934
935 alpha1 = LLVMBuildLShr(builder, alpha_lo,
936 lp_build_const_int_vec(gallivm, type, 8), "");
937 alpha1 = LLVMBuildAnd(builder, alpha1,
938 lp_build_const_int_vec(gallivm, type, 0xff), "");
939 if (is_signed) {
940 alpha1 = LLVMBuildTrunc(builder, alpha1, lp_build_vec_type(gallivm, type8), "");
941 alpha1 = LLVMBuildSExt(builder, alpha1, lp_build_vec_type(gallivm, type), "");
942 }
943
944 /* pos = 3*(4j+i) */
945 bit_pos = LLVMBuildShl(builder, j, lp_build_const_int_vec(gallivm, type, 2), "");
946 bit_pos = LLVMBuildAdd(builder, bit_pos, i, "");
947 tmp = LLVMBuildAdd(builder, bit_pos, bit_pos, "");
948 bit_pos = LLVMBuildAdd(builder, bit_pos, tmp, "");
949 /* get rid of first 2 bytes - saves shifts of alpha_lo/hi */
950 bit_pos = LLVMBuildAdd(builder, bit_pos,
951 lp_build_const_int_vec(gallivm, type, 16), "");
952
953 if (n == 1) {
954 struct lp_type type64;
955 memset(&type64, 0, sizeof type64);
956 type64.width = 64;
957 type64.length = 1;
958 /* This is pretty pointless could avoid by just directly extracting
959 64bit in the first place but makes it more complicated elsewhere */
960 alpha_lo = LLVMBuildZExt(builder, alpha_lo, i64t, "");
961 alpha_hi = LLVMBuildZExt(builder, alpha_hi, i64t, "");
962 alphac0 = LLVMBuildShl(builder, alpha_hi,
963 lp_build_const_int_vec(gallivm, type64, 32), "");
964 alphac0 = LLVMBuildOr(builder, alpha_lo, alphac0, "");
965
966 shift = LLVMBuildZExt(builder, bit_pos, i64t, "");
967 alphac0 = LLVMBuildLShr(builder, alphac0, shift, "");
968 alphac0 = LLVMBuildTrunc(builder, alphac0, i32t, "");
969 alphac = LLVMBuildAnd(builder, alphac0,
970 lp_build_const_int_vec(gallivm, type, 0x7), "");
971 }
972 else {
973 /*
974 * Using non-native vector length here (actually, with avx2 and
975 * n == 4 llvm will indeed expand to ymm regs...)
976 * At least newer llvm versions handle that ok.
977 * llvm 3.7+ will even handle the emulated 64bit shift with variable
978 * shift count without extraction (and it's actually easier to
979 * emulate than the 32bit one).
980 */
981 alpha64 = LLVMBuildShuffleVector(builder, alpha_lo, alpha_hi,
982 lp_build_const_unpackx2_shuffle(gallivm, n), "");
983
984 alpha64 = LLVMBuildBitCast(builder, alpha64, LLVMVectorType(i64t, n), "");
985 shift = LLVMBuildZExt(builder, bit_pos, LLVMVectorType(i64t, n), "");
986 alphac = LLVMBuildLShr(builder, alpha64, shift, "");
987 alphac = LLVMBuildTrunc(builder, alphac, bld32.vec_type, "");
988
989 alphac = LLVMBuildAnd(builder, alphac,
990 lp_build_const_int_vec(gallivm, type, 0x7), "");
991 }
992
993 /* signed compare is faster saves some xors */
994 type.sign = TRUE;
995 /* alpha0 > alpha1 selection */
996 sel_mask = lp_build_compare(gallivm, type, PIPE_FUNC_GREATER,
997 alpha0, alpha1);
998 ainterp = lp_build_lerpdxta(gallivm, alpha0, alpha1, alphac, sel_mask, n);
999
1000 /*
1001 * if a0 > a1 then we select a0 for case 0, a1 for case 1, interp otherwise.
1002 * else we select a0 for case 0, a1 for case 1,
1003 * interp for case 2-5, 00 for 6 and 0xff(ffffff) for 7
1004 * a = (c == 0) ? a0 : a1
1005 * a = (c > 1) ? ainterp : a
1006 * Finally handle case 6/7 for !(a0 > a1)
1007 * a = (!(a0 > a1) && c == 6) ? 0 : a (andnot with mask)
1008 * a = (!(a0 > a1) && c == 7) ? 0xffffffff : a (or with mask)
1009 */
1010 tmp_mask = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL,
1011 alphac, bld32.zero);
1012 alpha = lp_build_select(&bld32, tmp_mask, alpha0, alpha1);
1013 tmp_mask = lp_build_compare(gallivm, type, PIPE_FUNC_GREATER,
1014 alphac, bld32.one);
1015 alpha = lp_build_select(&bld32, tmp_mask, ainterp, alpha);
1016
1017 code_s = LLVMBuildAnd(builder, alphac,
1018 LLVMBuildNot(builder, sel_mask, ""), "");
1019 mask6 = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL,
1020 code_s, lp_build_const_int_vec(gallivm, type, 6));
1021 mask7 = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL,
1022 code_s, lp_build_const_int_vec(gallivm, type, 7));
1023 if (is_signed) {
1024 alpha = lp_build_select(&bld32, mask6, lp_build_const_int_vec(gallivm, type, -127), alpha);
1025 alpha = lp_build_select(&bld32, mask7, lp_build_const_int_vec(gallivm, type, 127), alpha);
1026 } else {
1027 alpha = LLVMBuildAnd(builder, alpha, LLVMBuildNot(builder, mask6, ""), "");
1028 alpha = LLVMBuildOr(builder, alpha, mask7, "");
1029 }
1030 /* There can be garbage in upper bits, mask them off for rgtc formats */
1031 alpha = LLVMBuildAnd(builder, alpha, lp_build_const_int_vec(gallivm, type, 0xff), "");
1032
1033 return alpha;
1034 }
1035
1036 /**
1037 * Convert from <n x i128> s3tc dxt5 to <4n x i8> RGBA AoS
1038 * @param colors is a <n x i32> vector with n x 2x16bit colors
1039 * @param codewords is a <n x i32> vector containing the codewords
1040 * @param alphas is a <n x i64> vector containing the alpha values
1041 * @param i is a <n x i32> vector with the x pixel coordinate (0 to 3)
1042 * @param j is a <n x i32> vector with the y pixel coordinate (0 to 3)
1043 */
1044 static LLVMValueRef
s3tc_dxt5_full_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef colors,LLVMValueRef codewords,LLVMValueRef alpha_lo,LLVMValueRef alpha_hi,LLVMValueRef i,LLVMValueRef j)1045 s3tc_dxt5_full_to_rgba_aos(struct gallivm_state *gallivm,
1046 unsigned n,
1047 enum pipe_format format,
1048 LLVMValueRef colors,
1049 LLVMValueRef codewords,
1050 LLVMValueRef alpha_lo,
1051 LLVMValueRef alpha_hi,
1052 LLVMValueRef i,
1053 LLVMValueRef j)
1054 {
1055 LLVMBuilderRef builder = gallivm->builder;
1056 LLVMValueRef rgba, alpha;
1057 struct lp_type type, type8;
1058 struct lp_build_context bld32;
1059
1060 memset(&type, 0, sizeof type);
1061 type.width = 32;
1062 type.length = n;
1063
1064 memset(&type8, 0, sizeof type8);
1065 type8.width = 8;
1066 type8.length = n*4;
1067
1068 assert(lp_check_value(type, i));
1069 assert(lp_check_value(type, j));
1070
1071 lp_build_context_init(&bld32, gallivm, type);
1072
1073 assert(lp_check_value(type, i));
1074 assert(lp_check_value(type, j));
1075
1076 rgba = s3tc_dxt1_to_rgba_aos(gallivm, n, format,
1077 colors, codewords, i, j);
1078
1079 rgba = LLVMBuildBitCast(builder, rgba, bld32.vec_type, "");
1080
1081 alpha = s3tc_dxt5_alpha_channel(gallivm, false, n, alpha_hi, alpha_lo, i, j);
1082 alpha = LLVMBuildShl(builder, alpha, lp_build_const_int_vec(gallivm, type, 24), "");
1083 rgba = LLVMBuildOr(builder, alpha, rgba, "");
1084
1085 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
1086 }
1087
1088
1089 static void
lp_build_gather_s3tc_simple_scalar(struct gallivm_state * gallivm,const struct util_format_description * format_desc,LLVMValueRef * dxt_block,LLVMValueRef ptr)1090 lp_build_gather_s3tc_simple_scalar(struct gallivm_state *gallivm,
1091 const struct util_format_description *format_desc,
1092 LLVMValueRef *dxt_block,
1093 LLVMValueRef ptr)
1094 {
1095 LLVMBuilderRef builder = gallivm->builder;
1096 unsigned block_bits = format_desc->block.bits;
1097 LLVMValueRef elem, shuf;
1098 LLVMTypeRef type32 = LLVMIntTypeInContext(gallivm->context, 32);
1099 LLVMTypeRef src_type = LLVMIntTypeInContext(gallivm->context, block_bits);
1100 LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
1101 LLVMTypeRef type32_4 = LLVMVectorType(type32, 4);
1102
1103 assert(block_bits == 64 || block_bits == 128);
1104
1105 ptr = LLVMBuildBitCast(builder, ptr, src_ptr_type, "");
1106 elem = LLVMBuildLoad(builder, ptr, "");
1107
1108 if (block_bits == 128) {
1109 /* just return block as is */
1110 *dxt_block = LLVMBuildBitCast(builder, elem, type32_4, "");
1111 }
1112 else {
1113 LLVMTypeRef type32_2 = LLVMVectorType(type32, 2);
1114 shuf = lp_build_const_extend_shuffle(gallivm, 2, 4);
1115 elem = LLVMBuildBitCast(builder, elem, type32_2, "");
1116 *dxt_block = LLVMBuildShuffleVector(builder, elem,
1117 LLVMGetUndef(type32_2), shuf, "");
1118 }
1119 }
1120
1121
1122 static void
s3tc_store_cached_block(struct gallivm_state * gallivm,LLVMValueRef * col,LLVMValueRef tag_value,LLVMValueRef hash_index,LLVMValueRef cache)1123 s3tc_store_cached_block(struct gallivm_state *gallivm,
1124 LLVMValueRef *col,
1125 LLVMValueRef tag_value,
1126 LLVMValueRef hash_index,
1127 LLVMValueRef cache)
1128 {
1129 LLVMBuilderRef builder = gallivm->builder;
1130 LLVMValueRef ptr, indices[3];
1131 LLVMTypeRef type_ptr4x32;
1132 unsigned count;
1133
1134 type_ptr4x32 = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
1135 indices[0] = lp_build_const_int32(gallivm, 0);
1136 indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS);
1137 indices[2] = hash_index;
1138 ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
1139 LLVMBuildStore(builder, tag_value, ptr);
1140
1141 indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_DATA);
1142 hash_index = LLVMBuildMul(builder, hash_index,
1143 lp_build_const_int32(gallivm, 16), "");
1144 for (count = 0; count < 4; count++) {
1145 indices[2] = hash_index;
1146 ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
1147 ptr = LLVMBuildBitCast(builder, ptr, type_ptr4x32, "");
1148 LLVMBuildStore(builder, col[count], ptr);
1149 hash_index = LLVMBuildAdd(builder, hash_index,
1150 lp_build_const_int32(gallivm, 4), "");
1151 }
1152 }
1153
1154 static LLVMValueRef
s3tc_lookup_cached_pixel(struct gallivm_state * gallivm,LLVMValueRef ptr,LLVMValueRef index)1155 s3tc_lookup_cached_pixel(struct gallivm_state *gallivm,
1156 LLVMValueRef ptr,
1157 LLVMValueRef index)
1158 {
1159 LLVMBuilderRef builder = gallivm->builder;
1160 LLVMValueRef member_ptr, indices[3];
1161
1162 indices[0] = lp_build_const_int32(gallivm, 0);
1163 indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_DATA);
1164 indices[2] = index;
1165 member_ptr = LLVMBuildGEP(builder, ptr, indices, ARRAY_SIZE(indices), "");
1166 return LLVMBuildLoad(builder, member_ptr, "cache_data");
1167 }
1168
1169 static LLVMValueRef
s3tc_lookup_tag_data(struct gallivm_state * gallivm,LLVMValueRef ptr,LLVMValueRef index)1170 s3tc_lookup_tag_data(struct gallivm_state *gallivm,
1171 LLVMValueRef ptr,
1172 LLVMValueRef index)
1173 {
1174 LLVMBuilderRef builder = gallivm->builder;
1175 LLVMValueRef member_ptr, indices[3];
1176
1177 indices[0] = lp_build_const_int32(gallivm, 0);
1178 indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS);
1179 indices[2] = index;
1180 member_ptr = LLVMBuildGEP(builder, ptr, indices, ARRAY_SIZE(indices), "");
1181 return LLVMBuildLoad(builder, member_ptr, "tag_data");
1182 }
1183
1184 #if LP_BUILD_FORMAT_CACHE_DEBUG
1185 static void
s3tc_update_cache_access(struct gallivm_state * gallivm,LLVMValueRef ptr,unsigned count,unsigned index)1186 s3tc_update_cache_access(struct gallivm_state *gallivm,
1187 LLVMValueRef ptr,
1188 unsigned count,
1189 unsigned index)
1190 {
1191 LLVMBuilderRef builder = gallivm->builder;
1192 LLVMValueRef member_ptr, cache_access;
1193
1194 assert(index == LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL ||
1195 index == LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
1196
1197 member_ptr = lp_build_struct_get_ptr(gallivm, ptr, index, "");
1198 cache_access = LLVMBuildLoad(builder, member_ptr, "cache_access");
1199 cache_access = LLVMBuildAdd(builder, cache_access,
1200 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
1201 count, 0), "");
1202 LLVMBuildStore(builder, cache_access, member_ptr);
1203 }
1204 #endif
1205
1206 /**
1207 * Calculate 1/3(v1-v0) + v0 and 2*1/3(v1-v0) + v0.
1208 * The lerp is performed between the first 2 32bit colors
1209 * in the source vector, both results are returned packed in result vector.
1210 */
1211 static LLVMValueRef
lp_build_lerp23_single(struct lp_build_context * bld,LLVMValueRef v01)1212 lp_build_lerp23_single(struct lp_build_context *bld,
1213 LLVMValueRef v01)
1214 {
1215 struct gallivm_state *gallivm = bld->gallivm;
1216 LLVMValueRef x, mul, delta, res, v0, v1, elems[8];
1217 const struct lp_type type = bld->type;
1218 LLVMBuilderRef builder = bld->gallivm->builder;
1219 struct lp_type i16_type = lp_wider_type(type);
1220 struct lp_type i32_type = lp_wider_type(i16_type);
1221 struct lp_build_context bld2;
1222
1223 assert(!type.floating && !type.fixed && !type.norm && type.width == 8);
1224
1225 lp_build_context_init(&bld2, gallivm, i16_type);
1226 bld2.type.sign = TRUE;
1227
1228 /* weights 256/3, 256*2/3, with correct rounding */
1229 elems[0] = elems[1] = elems[2] = elems[3] =
1230 lp_build_const_elem(gallivm, i16_type, 255*1/3);
1231 elems[4] = elems[5] = elems[6] = elems[7] =
1232 lp_build_const_elem(gallivm, i16_type, 171);
1233 x = LLVMConstVector(elems, 8);
1234
1235 /*
1236 * v01 has col0 in 32bit elem 0, col1 in elem 1.
1237 * Interleave/unpack will give us separate v0/v1 vectors.
1238 */
1239 v01 = lp_build_interleave2(gallivm, i32_type, v01, v01, 0);
1240 v01 = LLVMBuildBitCast(builder, v01, bld->vec_type, "");
1241
1242 lp_build_unpack2(gallivm, type, i16_type, v01, &v0, &v1);
1243 delta = lp_build_sub(&bld2, v1, v0);
1244
1245 mul = LLVMBuildMul(builder, x, delta, "");
1246
1247 mul = LLVMBuildLShr(builder, mul, lp_build_const_int_vec(gallivm, i16_type, 8), "");
1248 /* lerp optimization: pack now, do add afterwards */
1249 res = lp_build_pack2(gallivm, i16_type, type, mul, bld2.undef);
1250 /* only lower 2 elems are valid - for these v0 is really v0 */
1251 return lp_build_add(bld, res, v01);
1252 }
1253
1254 /*
1255 * decode one dxt1 block.
1256 */
1257 static void
s3tc_decode_block_dxt1(struct gallivm_state * gallivm,enum pipe_format format,LLVMValueRef dxt_block,LLVMValueRef * col)1258 s3tc_decode_block_dxt1(struct gallivm_state *gallivm,
1259 enum pipe_format format,
1260 LLVMValueRef dxt_block,
1261 LLVMValueRef *col)
1262 {
1263 LLVMBuilderRef builder = gallivm->builder;
1264 LLVMValueRef color01, color23, color01_16, color0123;
1265 LLVMValueRef rgba, tmp, a, sel_mask, indices, code, const2;
1266 struct lp_type type8, type32, type16, type64;
1267 struct lp_build_context bld8, bld32, bld16, bld64;
1268 unsigned i;
1269 boolean is_dxt1_variant = format_dxt1_variant(format);
1270
1271 memset(&type32, 0, sizeof type32);
1272 type32.width = 32;
1273 type32.length = 4;
1274 type32.sign = TRUE;
1275
1276 memset(&type8, 0, sizeof type8);
1277 type8.width = 8;
1278 type8.length = 16;
1279
1280 memset(&type16, 0, sizeof type16);
1281 type16.width = 16;
1282 type16.length = 8;
1283
1284 memset(&type64, 0, sizeof type64);
1285 type64.width = 64;
1286 type64.length = 2;
1287
1288 a = lp_build_const_int_vec(gallivm, type32, 0xff000000);
1289 const2 = lp_build_const_int_vec(gallivm, type32, 2);
1290
1291 lp_build_context_init(&bld32, gallivm, type32);
1292 lp_build_context_init(&bld16, gallivm, type16);
1293 lp_build_context_init(&bld8, gallivm, type8);
1294 lp_build_context_init(&bld64, gallivm, type64);
1295
1296 if (is_dxt1_variant) {
1297 color01 = lp_build_shuffle1undef(gallivm, dxt_block, 0, 4);
1298 code = lp_build_shuffle1undef(gallivm, dxt_block, 1, 4);
1299 } else {
1300 color01 = lp_build_shuffle1undef(gallivm, dxt_block, 2, 4);
1301 code = lp_build_shuffle1undef(gallivm, dxt_block, 3, 4);
1302 }
1303 code = LLVMBuildBitCast(builder, code, bld8.vec_type, "");
1304 /* expand bytes to dwords */
1305 code = lp_build_interleave2(gallivm, type8, code, code, 0);
1306 code = lp_build_interleave2(gallivm, type8, code, code, 0);
1307
1308
1309 /*
1310 * works as follows:
1311 * - expand color0/color1 to rgba8888
1312 * - calculate color2/3 (interpolation) according to color0 < color1 rules
1313 * - calculate color2/3 according to color0 >= color1 rules
1314 * - do selection of color2/3 according to comparison of color0/1
1315 * - extract indices.
1316 * - use compare/select to select the correct color. Since we have 2bit
1317 * indices (and 4 colors), needs at least three compare/selects.
1318 */
1319
1320 /*
1321 * expand the two colors
1322 */
1323 color01 = LLVMBuildBitCast(builder, color01, bld16.vec_type, "");
1324 color01 = lp_build_interleave2(gallivm, type16, color01,
1325 bld16.zero, 0);
1326 color01_16 = LLVMBuildBitCast(builder, color01, bld32.vec_type, "");
1327 color01 = color_expand_565_to_8888(gallivm, 4, color01_16);
1328
1329 /*
1330 * interpolate colors
1331 * color2_1 is 2/3 color0 + 1/3 color1
1332 * color3_1 is 1/3 color0 + 2/3 color1
1333 * color2_2 is 1/2 color0 + 1/2 color1
1334 * color3_2 is 0
1335 */
1336
1337 /* TODO: since this is now always scalar, should
1338 * probably just use control flow here instead of calculating
1339 * both cases and then selection
1340 */
1341 if (format == PIPE_FORMAT_DXT1_RGBA ||
1342 format == PIPE_FORMAT_DXT1_SRGBA) {
1343 color01 = LLVMBuildOr(builder, color01, a, "");
1344 }
1345 /* can combine 2 lerps into one mostly */
1346 color23 = lp_build_lerp23_single(&bld8, color01);
1347 color23 = LLVMBuildBitCast(builder, color23, bld32.vec_type, "");
1348
1349 /* dxt3/5 always use 4-color encoding */
1350 if (is_dxt1_variant) {
1351 LLVMValueRef color23_2, color2_2;
1352
1353 if (util_get_cpu_caps()->has_sse2) {
1354 LLVMValueRef intrargs[2];
1355 intrargs[0] = LLVMBuildBitCast(builder, color01, bld8.vec_type, "");
1356 /* same interleave as for lerp23 - correct result in 2nd element */
1357 intrargs[1] = lp_build_interleave2(gallivm, type32, color01, color01, 0);
1358 intrargs[1] = LLVMBuildBitCast(builder, intrargs[1], bld8.vec_type, "");
1359 color2_2 = lp_build_pavgb(&bld8, intrargs[0], intrargs[1]);
1360 }
1361 else {
1362 LLVMValueRef v01, v0, v1, vhalf;
1363 /*
1364 * This isn't as expensive as it looks (the unpack is the same as
1365 * for lerp23, which is the reason why we do the pointless
1366 * interleave2 too), with correct rounding (the two lower elements
1367 * will be the same).
1368 */
1369 v01 = lp_build_interleave2(gallivm, type32, color01, color01, 0);
1370 v01 = LLVMBuildBitCast(builder, v01, bld8.vec_type, "");
1371 lp_build_unpack2(gallivm, type8, type16, v01, &v0, &v1);
1372 vhalf = lp_build_add(&bld16, v0, v1);
1373 vhalf = LLVMBuildLShr(builder, vhalf, bld16.one, "");
1374 color2_2 = lp_build_pack2(gallivm, type16, type8, vhalf, bld16.undef);
1375 }
1376 /* shuffle in color 3 as elem 2 zero, color 2 elem 1 */
1377 color23_2 = LLVMBuildBitCast(builder, color2_2, bld64.vec_type, "");
1378 color23_2 = LLVMBuildLShr(builder, color23_2,
1379 lp_build_const_int_vec(gallivm, type64, 32), "");
1380 color23_2 = LLVMBuildBitCast(builder, color23_2, bld32.vec_type, "");
1381
1382 tmp = LLVMBuildBitCast(builder, color01_16, bld64.vec_type, "");
1383 tmp = LLVMBuildLShr(builder, tmp,
1384 lp_build_const_int_vec(gallivm, type64, 32), "");
1385 tmp = LLVMBuildBitCast(builder, tmp, bld32.vec_type, "");
1386 sel_mask = lp_build_compare(gallivm, type32, PIPE_FUNC_GREATER,
1387 color01_16, tmp);
1388 sel_mask = lp_build_interleave2(gallivm, type32, sel_mask, sel_mask, 0);
1389 color23 = lp_build_select(&bld32, sel_mask, color23, color23_2);
1390 }
1391
1392 if (util_get_cpu_caps()->has_ssse3) {
1393 /*
1394 * Use pshufb as mini-lut. (Only doable with intrinsics as the
1395 * final shuffles are non-constant. pshufb is awesome!)
1396 */
1397 LLVMValueRef shuf[16], low2mask;
1398 LLVMValueRef intrargs[2], lut_ind, lut_adj;
1399
1400 color01 = LLVMBuildBitCast(builder, color01, bld64.vec_type, "");
1401 color23 = LLVMBuildBitCast(builder, color23, bld64.vec_type, "");
1402 color0123 = lp_build_interleave2(gallivm, type64, color01, color23, 0);
1403 color0123 = LLVMBuildBitCast(builder, color0123, bld32.vec_type, "");
1404
1405 if (format == PIPE_FORMAT_DXT1_RGB ||
1406 format == PIPE_FORMAT_DXT1_SRGB) {
1407 color0123 = LLVMBuildOr(builder, color0123, a, "");
1408 }
1409
1410 /* shuffle as r0r1r2r3g0g1... */
1411 for (i = 0; i < 4; i++) {
1412 shuf[4*i] = lp_build_const_int32(gallivm, 0 + i);
1413 shuf[4*i+1] = lp_build_const_int32(gallivm, 4 + i);
1414 shuf[4*i+2] = lp_build_const_int32(gallivm, 8 + i);
1415 shuf[4*i+3] = lp_build_const_int32(gallivm, 12 + i);
1416 }
1417 color0123 = LLVMBuildBitCast(builder, color0123, bld8.vec_type, "");
1418 color0123 = LLVMBuildShuffleVector(builder, color0123, bld8.undef,
1419 LLVMConstVector(shuf, 16), "");
1420
1421 /* lowest 2 bits of each 8 bit value contain index into "LUT" */
1422 low2mask = lp_build_const_int_vec(gallivm, type8, 3);
1423 /* add 0/4/8/12 for r/g/b/a */
1424 lut_adj = lp_build_const_int_vec(gallivm, type32, 0x0c080400);
1425 lut_adj = LLVMBuildBitCast(builder, lut_adj, bld8.vec_type, "");
1426 intrargs[0] = color0123;
1427 for (i = 0; i < 4; i++) {
1428 lut_ind = LLVMBuildAnd(builder, code, low2mask, "");
1429 lut_ind = LLVMBuildOr(builder, lut_ind, lut_adj, "");
1430 intrargs[1] = lut_ind;
1431 col[i] = lp_build_intrinsic(builder, "llvm.x86.ssse3.pshuf.b.128",
1432 bld8.vec_type, intrargs, 2, 0);
1433 col[i] = LLVMBuildBitCast(builder, col[i], bld32.vec_type, "");
1434 code = LLVMBuildBitCast(builder, code, bld32.vec_type, "");
1435 code = LLVMBuildLShr(builder, code, const2, "");
1436 code = LLVMBuildBitCast(builder, code, bld8.vec_type, "");
1437 }
1438 }
1439 else {
1440 /* Thanks to vectorization can do 4 texels in parallel */
1441 LLVMValueRef color0, color1, color2, color3;
1442 if (format == PIPE_FORMAT_DXT1_RGB ||
1443 format == PIPE_FORMAT_DXT1_SRGB) {
1444 color01 = LLVMBuildOr(builder, color01, a, "");
1445 color23 = LLVMBuildOr(builder, color23, a, "");
1446 }
1447 color0 = LLVMBuildShuffleVector(builder, color01, bld32.undef,
1448 lp_build_const_shuffle1(gallivm, 0, 4), "");
1449 color1 = LLVMBuildShuffleVector(builder, color01, bld32.undef,
1450 lp_build_const_shuffle1(gallivm, 1, 4), "");
1451 color2 = LLVMBuildShuffleVector(builder, color23, bld32.undef,
1452 lp_build_const_shuffle1(gallivm, 0, 4), "");
1453 color3 = LLVMBuildShuffleVector(builder, color23, bld32.undef,
1454 lp_build_const_shuffle1(gallivm, 1, 4), "");
1455 code = LLVMBuildBitCast(builder, code, bld32.vec_type, "");
1456
1457 for (i = 0; i < 4; i++) {
1458 /* select the colors */
1459 LLVMValueRef selmasklo, rgba01, rgba23, bitlo;
1460 bitlo = bld32.one;
1461 indices = LLVMBuildAnd(builder, code, bitlo, "");
1462 selmasklo = lp_build_compare(gallivm, type32, PIPE_FUNC_EQUAL,
1463 indices, bitlo);
1464 rgba01 = lp_build_select(&bld32, selmasklo, color1, color0);
1465
1466 LLVMValueRef selmaskhi;
1467 indices = LLVMBuildAnd(builder, code, const2, "");
1468 selmaskhi = lp_build_compare(gallivm, type32, PIPE_FUNC_EQUAL,
1469 indices, const2);
1470 rgba23 = lp_build_select(&bld32, selmasklo, color3, color2);
1471 rgba = lp_build_select(&bld32, selmaskhi, rgba23, rgba01);
1472
1473 /*
1474 * Note that this will give "wrong" order.
1475 * col0 will be rgba0, rgba4, rgba8, rgba12, col1 rgba1, rgba5, ...
1476 * This would be easily fixable by using different shuffle, bitlo/hi
1477 * vectors above (and different shift), but seems slightly easier to
1478 * deal with for dxt3/dxt5 alpha too. So instead change lookup.
1479 */
1480 col[i] = rgba;
1481 code = LLVMBuildLShr(builder, code, const2, "");
1482 }
1483 }
1484 }
1485
1486 /*
1487 * decode one dxt3 block.
1488 */
1489 static void
s3tc_decode_block_dxt3(struct gallivm_state * gallivm,enum pipe_format format,LLVMValueRef dxt_block,LLVMValueRef * col)1490 s3tc_decode_block_dxt3(struct gallivm_state *gallivm,
1491 enum pipe_format format,
1492 LLVMValueRef dxt_block,
1493 LLVMValueRef *col)
1494 {
1495 LLVMBuilderRef builder = gallivm->builder;
1496 LLVMValueRef alpha, alphas0, alphas1, shift4_16, a[4], mask8hi;
1497 struct lp_type type32, type8, type16;
1498 unsigned i;
1499
1500 memset(&type32, 0, sizeof type32);
1501 type32.width = 32;
1502 type32.length = 4;
1503
1504 memset(&type8, 0, sizeof type8);
1505 type8.width = 8;
1506 type8.length = 16;
1507
1508 memset(&type16, 0, sizeof type16);
1509 type16.width = 16;
1510 type16.length = 8;
1511
1512 s3tc_decode_block_dxt1(gallivm, format, dxt_block, col);
1513
1514 shift4_16 = lp_build_const_int_vec(gallivm, type16, 4);
1515 mask8hi = lp_build_const_int_vec(gallivm, type32, 0xff000000);
1516
1517 alpha = LLVMBuildBitCast(builder, dxt_block,
1518 lp_build_vec_type(gallivm, type8), "");
1519 alpha = lp_build_interleave2(gallivm, type8, alpha, alpha, 0);
1520 alpha = LLVMBuildBitCast(builder, alpha,
1521 lp_build_vec_type(gallivm, type16), "");
1522 alpha = LLVMBuildAnd(builder, alpha,
1523 lp_build_const_int_vec(gallivm, type16, 0xf00f), "");
1524 alphas0 = LLVMBuildLShr(builder, alpha, shift4_16, "");
1525 alphas1 = LLVMBuildShl(builder, alpha, shift4_16, "");
1526 alpha = LLVMBuildOr(builder, alphas0, alpha, "");
1527 alpha = LLVMBuildOr(builder, alphas1, alpha, "");
1528 alpha = LLVMBuildBitCast(builder, alpha,
1529 lp_build_vec_type(gallivm, type32), "");
1530 /*
1531 * alpha now contains elems 0,1,2,3,... (ubytes)
1532 * we need 0,4,8,12, 1,5,9,13 etc. in dwords to match color (which
1533 * is just as easy as "natural" order - 3 shift/and instead of 6 unpack).
1534 */
1535 a[0] = LLVMBuildShl(builder, alpha,
1536 lp_build_const_int_vec(gallivm, type32, 24), "");
1537 a[1] = LLVMBuildShl(builder, alpha,
1538 lp_build_const_int_vec(gallivm, type32, 16), "");
1539 a[1] = LLVMBuildAnd(builder, a[1], mask8hi, "");
1540 a[2] = LLVMBuildShl(builder, alpha,
1541 lp_build_const_int_vec(gallivm, type32, 8), "");
1542 a[2] = LLVMBuildAnd(builder, a[2], mask8hi, "");
1543 a[3] = LLVMBuildAnd(builder, alpha, mask8hi, "");
1544
1545 for (i = 0; i < 4; i++) {
1546 col[i] = LLVMBuildOr(builder, col[i], a[i], "");
1547 }
1548 }
1549
1550
1551 static LLVMValueRef
lp_build_lerpdxta_block(struct gallivm_state * gallivm,LLVMValueRef alpha0,LLVMValueRef alpha1,LLVMValueRef code,LLVMValueRef sel_mask)1552 lp_build_lerpdxta_block(struct gallivm_state *gallivm,
1553 LLVMValueRef alpha0,
1554 LLVMValueRef alpha1,
1555 LLVMValueRef code,
1556 LLVMValueRef sel_mask)
1557 {
1558 LLVMBuilderRef builder = gallivm->builder;
1559 LLVMValueRef delta, ainterp;
1560 LLVMValueRef weight5, weight7, weight;
1561 struct lp_type type16;
1562 struct lp_build_context bld;
1563
1564 memset(&type16, 0, sizeof type16);
1565 type16.width = 16;
1566 type16.length = 8;
1567 type16.sign = TRUE;
1568
1569 lp_build_context_init(&bld, gallivm, type16);
1570 /*
1571 * 256/7 is only 36.57 so we'd lose quite some precision. Since it would
1572 * actually be desirable to do this here with even higher accuracy than
1573 * even 8 bit (more or less required for rgtc, albeit that's not handled
1574 * here right now), shift the weights after multiplication by code.
1575 */
1576 weight5 = lp_build_const_int_vec(gallivm, type16, 256*64/5);
1577 weight7 = lp_build_const_int_vec(gallivm, type16, 256*64/7);
1578 weight = lp_build_select(&bld, sel_mask, weight7, weight5);
1579
1580 /*
1581 * we'll get garbage in the elements which had code 0 (or larger than
1582 * 5 or 7) but we don't care (or rather, need to fix up anyway).
1583 */
1584 code = LLVMBuildSub(builder, code, bld.one, "");
1585
1586 weight = LLVMBuildMul(builder, weight, code, "");
1587 weight = LLVMBuildLShr(builder, weight,
1588 lp_build_const_int_vec(gallivm, type16, 6), "");
1589
1590 delta = LLVMBuildSub(builder, alpha1, alpha0, "");
1591
1592 ainterp = LLVMBuildMul(builder, delta, weight, "");
1593 ainterp = LLVMBuildLShr(builder, ainterp,
1594 lp_build_const_int_vec(gallivm, type16, 8), "");
1595
1596 /* lerp is done later (with packed values) */
1597
1598 return ainterp;
1599 }
1600
1601
1602 /*
1603 * decode one dxt5 block.
1604 */
1605 static void
s3tc_decode_block_dxt5(struct gallivm_state * gallivm,enum pipe_format format,LLVMValueRef dxt_block,LLVMValueRef * col)1606 s3tc_decode_block_dxt5(struct gallivm_state *gallivm,
1607 enum pipe_format format,
1608 LLVMValueRef dxt_block,
1609 LLVMValueRef *col)
1610 {
1611 LLVMBuilderRef builder = gallivm->builder;
1612 LLVMValueRef alpha, alpha0, alpha1, ares;
1613 LLVMValueRef ainterp, ainterp0, ainterp1, shuffle1, sel_mask, sel_mask2;
1614 LLVMValueRef a[4], acode, tmp0, tmp1;
1615 LLVMTypeRef i64t, i32t;
1616 struct lp_type type32, type64, type8, type16;
1617 struct lp_build_context bld16, bld8;
1618 unsigned i;
1619
1620 memset(&type32, 0, sizeof type32);
1621 type32.width = 32;
1622 type32.length = 4;
1623
1624 memset(&type64, 0, sizeof type64);
1625 type64.width = 64;
1626 type64.length = 2;
1627
1628 memset(&type8, 0, sizeof type8);
1629 type8.width = 8;
1630 type8.length = 16;
1631
1632 memset(&type16, 0, sizeof type16);
1633 type16.width = 16;
1634 type16.length = 8;
1635
1636 lp_build_context_init(&bld16, gallivm, type16);
1637 lp_build_context_init(&bld8, gallivm, type8);
1638
1639 i64t = lp_build_vec_type(gallivm, type64);
1640 i32t = lp_build_vec_type(gallivm, type32);
1641
1642 s3tc_decode_block_dxt1(gallivm, format, dxt_block, col);
1643
1644 /*
1645 * three possible strategies for vectorizing alpha:
1646 * 1) compute all 8 values then use scalar extraction
1647 * (i.e. have all 8 alpha values packed in one 64bit scalar
1648 * and do something like ax = vals >> (codex * 8) followed
1649 * by inserting these values back into color)
1650 * 2) same as 8 but just use pshufb as a mini-LUT for selection.
1651 * (without pshufb would need boatloads of cmp/selects trying to
1652 * keep things vectorized for essentially scalar selection).
1653 * 3) do something similar to the uncached case
1654 * needs more calculations (need to calc 16 values instead of 8 though
1655 * that's only an issue for the lerp which we need to do twice otherwise
1656 * everything still fits into 128bit) but keeps things vectorized mostly.
1657 * Trying 3) here though not sure it's really faster...
1658 * With pshufb, we try 2) (cheaper and more accurate)
1659 */
1660
1661 /*
1662 * Ideally, we'd use 2 variable 16bit shifts here (byte shifts wouldn't
1663 * help since code crosses 8bit boundaries). But variable shifts are
1664 * AVX2 only, and even then only dword/quadword (intel _really_ hates
1665 * shifts!). Instead, emulate by 16bit muls.
1666 * Also, the required byte shuffles are essentially non-emulatable, so
1667 * require ssse3 (albeit other archs might do them fine).
1668 * This is not directly tied to ssse3 - just need sane byte shuffles.
1669 * But ordering is going to be different below so use same condition.
1670 */
1671
1672
1673 /* vectorize alpha */
1674 alpha = LLVMBuildBitCast(builder, dxt_block, i64t, "");
1675 alpha0 = LLVMBuildAnd(builder, alpha,
1676 lp_build_const_int_vec(gallivm, type64, 0xff), "");
1677 alpha0 = LLVMBuildBitCast(builder, alpha0, bld16.vec_type, "");
1678 alpha = LLVMBuildBitCast(builder, alpha, bld16.vec_type, "");
1679 alpha1 = LLVMBuildLShr(builder, alpha,
1680 lp_build_const_int_vec(gallivm, type16, 8), "");
1681 alpha = LLVMBuildBitCast(builder, alpha, i64t, "");
1682 shuffle1 = lp_build_const_shuffle1(gallivm, 0, 8);
1683 alpha0 = LLVMBuildShuffleVector(builder, alpha0, alpha0, shuffle1, "");
1684 alpha1 = LLVMBuildShuffleVector(builder, alpha1, alpha1, shuffle1, "");
1685
1686 type16.sign = TRUE;
1687 sel_mask = lp_build_compare(gallivm, type16, PIPE_FUNC_GREATER,
1688 alpha0, alpha1);
1689 type16.sign = FALSE;
1690 sel_mask = LLVMBuildBitCast(builder, sel_mask, bld8.vec_type, "");
1691
1692 if (!util_get_cpu_caps()->has_ssse3) {
1693 LLVMValueRef acodeg, mask1, acode0, acode1;
1694
1695 /* extraction of the 3 bit values into something more useful is HARD */
1696 /* first steps are actually scalar */
1697 acode = LLVMBuildLShr(builder, alpha,
1698 lp_build_const_int_vec(gallivm, type64, 16), "");
1699 tmp0 = LLVMBuildAnd(builder, acode,
1700 lp_build_const_int_vec(gallivm, type64, 0xffffff), "");
1701 tmp1 = LLVMBuildLShr(builder, acode,
1702 lp_build_const_int_vec(gallivm, type64, 24), "");
1703 tmp0 = LLVMBuildBitCast(builder, tmp0, i32t, "");
1704 tmp1 = LLVMBuildBitCast(builder, tmp1, i32t, "");
1705 acode = lp_build_interleave2(gallivm, type32, tmp0, tmp1, 0);
1706 /* now have 2x24bit in 4x32bit, order 01234567, 89..., undef, undef */
1707 tmp0 = LLVMBuildAnd(builder, acode,
1708 lp_build_const_int_vec(gallivm, type32, 0xfff), "");
1709 tmp1 = LLVMBuildLShr(builder, acode,
1710 lp_build_const_int_vec(gallivm, type32, 12), "");
1711 acode = lp_build_interleave2(gallivm, type32, tmp0, tmp1, 0);
1712 /* now have 4x12bit in 4x32bit, order 0123, 4567, ,,, */
1713 tmp0 = LLVMBuildAnd(builder, acode,
1714 lp_build_const_int_vec(gallivm, type32, 0x3f), "");
1715 tmp1 = LLVMBuildLShr(builder, acode,
1716 lp_build_const_int_vec(gallivm, type32, 6), "");
1717 /* use signed pack doesn't matter and otherwise need sse41 */
1718 type32.sign = type16.sign = TRUE;
1719 acode = lp_build_pack2(gallivm, type32, type16, tmp0, tmp1);
1720 type32.sign = type16.sign = FALSE;
1721 /* now have 8x6bit in 8x16bit, 01, 45, 89, ..., 23, 67, ... */
1722 acode0 = LLVMBuildAnd(builder, acode,
1723 lp_build_const_int_vec(gallivm, type16, 0x7), "");
1724 acode1 = LLVMBuildLShr(builder, acode,
1725 lp_build_const_int_vec(gallivm, type16, 3), "");
1726 acode = lp_build_pack2(gallivm, type16, type8, acode0, acode1);
1727 /* acode0 contains elems 0,4,8,12,2,6,10,14, acode1 1,5,9,... */
1728
1729 acodeg = LLVMBuildAnd(builder, acode,
1730 LLVMBuildNot(builder, sel_mask, ""), "");
1731 mask1 = lp_build_compare(gallivm, type8, PIPE_FUNC_EQUAL,
1732 acode, bld8.one);
1733
1734 sel_mask = LLVMBuildBitCast(builder, sel_mask, bld16.vec_type, "");
1735 ainterp0 = lp_build_lerpdxta_block(gallivm, alpha0, alpha1, acode0, sel_mask);
1736 ainterp1 = lp_build_lerpdxta_block(gallivm, alpha0, alpha1, acode1, sel_mask);
1737 sel_mask = LLVMBuildBitCast(builder, sel_mask, bld8.vec_type, "");
1738 ainterp = lp_build_pack2(gallivm, type16, type8, ainterp0, ainterp1);
1739 alpha0 = lp_build_pack2(gallivm, type16, type8, alpha0, alpha0);
1740 alpha1 = lp_build_pack2(gallivm, type16, type8, alpha1, alpha1);
1741 ainterp = LLVMBuildAdd(builder, ainterp, alpha0, "");
1742 /* Fix up val01 */
1743 sel_mask2 = lp_build_compare(gallivm, type8, PIPE_FUNC_EQUAL,
1744 acode, bld8.zero);
1745 ainterp = lp_build_select(&bld8, sel_mask2, alpha0, ainterp);
1746 ainterp = lp_build_select(&bld8, mask1, alpha1, ainterp);
1747
1748 /* fix up val67 if a0 <= a1 */
1749 sel_mask2 = lp_build_compare(gallivm, type8, PIPE_FUNC_EQUAL,
1750 acodeg, lp_build_const_int_vec(gallivm, type8, 6));
1751 ares = LLVMBuildAnd(builder, ainterp, LLVMBuildNot(builder, sel_mask2, ""), "");
1752 sel_mask2 = lp_build_compare(gallivm, type8, PIPE_FUNC_EQUAL,
1753 acodeg, lp_build_const_int_vec(gallivm, type8, 7));
1754 ares = LLVMBuildOr(builder, ares, sel_mask2, "");
1755
1756 /* unpack in right order (0,4,8,12,1,5,..) */
1757 /* this gives us zero, a0, zero, a4, zero, a8, ... for tmp0 */
1758 tmp0 = lp_build_interleave2(gallivm, type8, bld8.zero, ares, 0);
1759 tmp1 = lp_build_interleave2(gallivm, type8, bld8.zero, ares, 1);
1760 tmp0 = LLVMBuildBitCast(builder, tmp0, bld16.vec_type, "");
1761 tmp1 = LLVMBuildBitCast(builder, tmp1, bld16.vec_type, "");
1762
1763 a[0] = lp_build_interleave2(gallivm, type16, bld16.zero, tmp0, 0);
1764 a[1] = lp_build_interleave2(gallivm, type16, bld16.zero, tmp1, 0);
1765 a[2] = lp_build_interleave2(gallivm, type16, bld16.zero, tmp0, 1);
1766 a[3] = lp_build_interleave2(gallivm, type16, bld16.zero, tmp1, 1);
1767 }
1768 else {
1769 LLVMValueRef elems[16], intrargs[2], shufa, mulclo, mulchi, mask8hi;
1770 LLVMTypeRef type16s = LLVMInt16TypeInContext(gallivm->context);
1771 LLVMTypeRef type8s = LLVMInt8TypeInContext(gallivm->context);
1772 unsigned i, j;
1773 /*
1774 * Ideally, we'd use 2 variable 16bit shifts here (byte shifts wouldn't
1775 * help since code crosses 8bit boundaries). But variable shifts are
1776 * AVX2 only, and even then only dword/quadword (intel _really_ hates
1777 * shifts!). Instead, emulate by 16bit muls.
1778 * Also, the required byte shuffles are essentially non-emulatable, so
1779 * require ssse3 (albeit other archs might do them fine, but the
1780 * complete path is ssse3 only for now).
1781 */
1782 for (i = 0, j = 0; i < 16; i += 8, j += 3) {
1783 elems[i+0] = elems[i+1] = elems[i+2] = lp_build_const_int32(gallivm, j+2);
1784 elems[i+3] = elems[i+4] = lp_build_const_int32(gallivm, j+3);
1785 elems[i+5] = elems[i+6] = elems[i+7] = lp_build_const_int32(gallivm, j+4);
1786 }
1787 shufa = LLVMConstVector(elems, 16);
1788 alpha = LLVMBuildBitCast(builder, alpha, bld8.vec_type, "");
1789 acode = LLVMBuildShuffleVector(builder, alpha, bld8.undef, shufa, "");
1790 acode = LLVMBuildBitCast(builder, acode, bld16.vec_type, "");
1791 /*
1792 * Put 0/2/4/6 into high 3 bits of 16 bits (save AND mask)
1793 * Do the same for 1/3/5/7 (albeit still need mask there - ideally
1794 * we'd place them into bits 4-7 so could save shift but impossible.)
1795 */
1796 for (i = 0; i < 8; i += 4) {
1797 elems[i+0] = LLVMConstInt(type16s, 1 << (13-0), 0);
1798 elems[i+1] = LLVMConstInt(type16s, 1 << (13-6), 0);
1799 elems[i+2] = LLVMConstInt(type16s, 1 << (13-4), 0);
1800 elems[i+3] = LLVMConstInt(type16s, 1 << (13-2), 0);
1801 }
1802 mulclo = LLVMConstVector(elems, 8);
1803 for (i = 0; i < 8; i += 4) {
1804 elems[i+0] = LLVMConstInt(type16s, 1 << (13-3), 0);
1805 elems[i+1] = LLVMConstInt(type16s, 1 << (13-9), 0);
1806 elems[i+2] = LLVMConstInt(type16s, 1 << (13-7), 0);
1807 elems[i+3] = LLVMConstInt(type16s, 1 << (13-5), 0);
1808 }
1809 mulchi = LLVMConstVector(elems, 8);
1810
1811 tmp0 = LLVMBuildMul(builder, acode, mulclo, "");
1812 tmp1 = LLVMBuildMul(builder, acode, mulchi, "");
1813 tmp0 = LLVMBuildLShr(builder, tmp0,
1814 lp_build_const_int_vec(gallivm, type16, 13), "");
1815 tmp1 = LLVMBuildLShr(builder, tmp1,
1816 lp_build_const_int_vec(gallivm, type16, 5), "");
1817 tmp1 = LLVMBuildAnd(builder, tmp1,
1818 lp_build_const_int_vec(gallivm, type16, 0x700), "");
1819 acode = LLVMBuildOr(builder, tmp0, tmp1, "");
1820 acode = LLVMBuildBitCast(builder, acode, bld8.vec_type, "");
1821
1822 /*
1823 * Note that ordering is different here to non-ssse3 path:
1824 * 0/1/2/3/4/5...
1825 */
1826
1827 LLVMValueRef weight0, weight1, weight, delta;
1828 LLVMValueRef constff_elem7, const0_elem6;
1829 /* weights, correctly rounded (round(256*x/7)) */
1830 elems[0] = LLVMConstInt(type16s, 256, 0);
1831 elems[1] = LLVMConstInt(type16s, 0, 0);
1832 elems[2] = LLVMConstInt(type16s, 219, 0);
1833 elems[3] = LLVMConstInt(type16s, 183, 0);
1834 elems[4] = LLVMConstInt(type16s, 146, 0);
1835 elems[5] = LLVMConstInt(type16s, 110, 0);
1836 elems[6] = LLVMConstInt(type16s, 73, 0);
1837 elems[7] = LLVMConstInt(type16s, 37, 0);
1838 weight0 = LLVMConstVector(elems, 8);
1839
1840 elems[0] = LLVMConstInt(type16s, 256, 0);
1841 elems[1] = LLVMConstInt(type16s, 0, 0);
1842 elems[2] = LLVMConstInt(type16s, 205, 0);
1843 elems[3] = LLVMConstInt(type16s, 154, 0);
1844 elems[4] = LLVMConstInt(type16s, 102, 0);
1845 elems[5] = LLVMConstInt(type16s, 51, 0);
1846 elems[6] = LLVMConstInt(type16s, 0, 0);
1847 elems[7] = LLVMConstInt(type16s, 0, 0);
1848 weight1 = LLVMConstVector(elems, 8);
1849
1850 weight0 = LLVMBuildBitCast(builder, weight0, bld8.vec_type, "");
1851 weight1 = LLVMBuildBitCast(builder, weight1, bld8.vec_type, "");
1852 weight = lp_build_select(&bld8, sel_mask, weight0, weight1);
1853 weight = LLVMBuildBitCast(builder, weight, bld16.vec_type, "");
1854
1855 for (i = 0; i < 16; i++) {
1856 elems[i] = LLVMConstNull(type8s);
1857 }
1858 elems[7] = LLVMConstInt(type8s, 255, 0);
1859 constff_elem7 = LLVMConstVector(elems, 16);
1860
1861 for (i = 0; i < 16; i++) {
1862 elems[i] = LLVMConstInt(type8s, 255, 0);
1863 }
1864 elems[6] = LLVMConstInt(type8s, 0, 0);
1865 const0_elem6 = LLVMConstVector(elems, 16);
1866
1867 /* standard simple lerp - but the version we need isn't available */
1868 delta = LLVMBuildSub(builder, alpha0, alpha1, "");
1869 ainterp = LLVMBuildMul(builder, delta, weight, "");
1870 ainterp = LLVMBuildLShr(builder, ainterp,
1871 lp_build_const_int_vec(gallivm, type16, 8), "");
1872 ainterp = LLVMBuildBitCast(builder, ainterp, bld8.vec_type, "");
1873 alpha1 = LLVMBuildBitCast(builder, alpha1, bld8.vec_type, "");
1874 ainterp = LLVMBuildAdd(builder, ainterp, alpha1, "");
1875 ainterp = LLVMBuildBitCast(builder, ainterp, bld16.vec_type, "");
1876 ainterp = lp_build_pack2(gallivm, type16, type8, ainterp, bld16.undef);
1877
1878 /* fixing 0/0xff case is slightly more complex */
1879 constff_elem7 = LLVMBuildAnd(builder, constff_elem7,
1880 LLVMBuildNot(builder, sel_mask, ""), "");
1881 const0_elem6 = LLVMBuildOr(builder, const0_elem6, sel_mask, "");
1882 ainterp = LLVMBuildOr(builder, ainterp, constff_elem7, "");
1883 ainterp = LLVMBuildAnd(builder, ainterp, const0_elem6, "");
1884
1885 /* now pick all 16 elements at once! */
1886 intrargs[0] = ainterp;
1887 intrargs[1] = acode;
1888 ares = lp_build_intrinsic(builder, "llvm.x86.ssse3.pshuf.b.128",
1889 bld8.vec_type, intrargs, 2, 0);
1890
1891 ares = LLVMBuildBitCast(builder, ares, i32t, "");
1892 mask8hi = lp_build_const_int_vec(gallivm, type32, 0xff000000);
1893 a[0] = LLVMBuildShl(builder, ares,
1894 lp_build_const_int_vec(gallivm, type32, 24), "");
1895 a[1] = LLVMBuildShl(builder, ares,
1896 lp_build_const_int_vec(gallivm, type32, 16), "");
1897 a[1] = LLVMBuildAnd(builder, a[1], mask8hi, "");
1898 a[2] = LLVMBuildShl(builder, ares,
1899 lp_build_const_int_vec(gallivm, type32, 8), "");
1900 a[2] = LLVMBuildAnd(builder, a[2], mask8hi, "");
1901 a[3] = LLVMBuildAnd(builder, ares, mask8hi, "");
1902 }
1903
1904 for (i = 0; i < 4; i++) {
1905 a[i] = LLVMBuildBitCast(builder, a[i], i32t, "");
1906 col[i] = LLVMBuildOr(builder, col[i], a[i], "");
1907 }
1908 }
1909
1910
1911 static void
generate_update_cache_one_block(struct gallivm_state * gallivm,LLVMValueRef function,const struct util_format_description * format_desc)1912 generate_update_cache_one_block(struct gallivm_state *gallivm,
1913 LLVMValueRef function,
1914 const struct util_format_description *format_desc)
1915 {
1916 LLVMBasicBlockRef block;
1917 LLVMBuilderRef old_builder;
1918 LLVMValueRef ptr_addr;
1919 LLVMValueRef hash_index;
1920 LLVMValueRef cache;
1921 LLVMValueRef dxt_block, tag_value;
1922 LLVMValueRef col[LP_MAX_VECTOR_LENGTH];
1923
1924 ptr_addr = LLVMGetParam(function, 0);
1925 hash_index = LLVMGetParam(function, 1);
1926 cache = LLVMGetParam(function, 2);
1927
1928 lp_build_name(ptr_addr, "ptr_addr" );
1929 lp_build_name(hash_index, "hash_index");
1930 lp_build_name(cache, "cache_addr");
1931
1932 /*
1933 * Function body
1934 */
1935
1936 old_builder = gallivm->builder;
1937 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1938 gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
1939 LLVMPositionBuilderAtEnd(gallivm->builder, block);
1940
1941 lp_build_gather_s3tc_simple_scalar(gallivm, format_desc, &dxt_block,
1942 ptr_addr);
1943
1944 switch (format_desc->format) {
1945 case PIPE_FORMAT_DXT1_RGB:
1946 case PIPE_FORMAT_DXT1_RGBA:
1947 case PIPE_FORMAT_DXT1_SRGB:
1948 case PIPE_FORMAT_DXT1_SRGBA:
1949 s3tc_decode_block_dxt1(gallivm, format_desc->format, dxt_block, col);
1950 break;
1951 case PIPE_FORMAT_DXT3_RGBA:
1952 case PIPE_FORMAT_DXT3_SRGBA:
1953 s3tc_decode_block_dxt3(gallivm, format_desc->format, dxt_block, col);
1954 break;
1955 case PIPE_FORMAT_DXT5_RGBA:
1956 case PIPE_FORMAT_DXT5_SRGBA:
1957 s3tc_decode_block_dxt5(gallivm, format_desc->format, dxt_block, col);
1958 break;
1959 default:
1960 assert(0);
1961 s3tc_decode_block_dxt1(gallivm, format_desc->format, dxt_block, col);
1962 break;
1963 }
1964
1965 tag_value = LLVMBuildPtrToInt(gallivm->builder, ptr_addr,
1966 LLVMInt64TypeInContext(gallivm->context), "");
1967 s3tc_store_cached_block(gallivm, col, tag_value, hash_index, cache);
1968
1969 LLVMBuildRetVoid(gallivm->builder);
1970
1971 LLVMDisposeBuilder(gallivm->builder);
1972 gallivm->builder = old_builder;
1973
1974 gallivm_verify_function(gallivm, function);
1975 }
1976
1977
1978 static void
update_cached_block(struct gallivm_state * gallivm,const struct util_format_description * format_desc,LLVMValueRef ptr_addr,LLVMValueRef hash_index,LLVMValueRef cache)1979 update_cached_block(struct gallivm_state *gallivm,
1980 const struct util_format_description *format_desc,
1981 LLVMValueRef ptr_addr,
1982 LLVMValueRef hash_index,
1983 LLVMValueRef cache)
1984
1985 {
1986 LLVMBuilderRef builder = gallivm->builder;
1987 LLVMModuleRef module = gallivm->module;
1988 char name[256];
1989 LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
1990 LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
1991 LLVMValueRef function, inst;
1992 LLVMBasicBlockRef bb;
1993 LLVMValueRef args[3];
1994
1995 snprintf(name, sizeof name, "%s_update_cache_one_block",
1996 format_desc->short_name);
1997 function = LLVMGetNamedFunction(module, name);
1998
1999 if (!function) {
2000 LLVMTypeRef ret_type;
2001 LLVMTypeRef arg_types[3];
2002 LLVMTypeRef function_type;
2003 unsigned arg;
2004
2005 /*
2006 * Generate the function prototype.
2007 */
2008
2009 ret_type = LLVMVoidTypeInContext(gallivm->context);
2010 arg_types[0] = pi8t;
2011 arg_types[1] = LLVMInt32TypeInContext(gallivm->context);
2012 arg_types[2] = LLVMTypeOf(cache); // XXX: put right type here
2013 function_type = LLVMFunctionType(ret_type, arg_types, ARRAY_SIZE(arg_types), 0);
2014 function = LLVMAddFunction(module, name, function_type);
2015
2016 for (arg = 0; arg < ARRAY_SIZE(arg_types); ++arg)
2017 if (LLVMGetTypeKind(arg_types[arg]) == LLVMPointerTypeKind)
2018 lp_add_function_attr(function, arg + 1, LP_FUNC_ATTR_NOALIAS);
2019
2020 LLVMSetFunctionCallConv(function, LLVMFastCallConv);
2021 LLVMSetVisibility(function, LLVMHiddenVisibility);
2022 generate_update_cache_one_block(gallivm, function, format_desc);
2023 }
2024
2025 args[0] = ptr_addr;
2026 args[1] = hash_index;
2027 args[2] = cache;
2028
2029 LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");
2030 bb = LLVMGetInsertBlock(builder);
2031 inst = LLVMGetLastInstruction(bb);
2032 LLVMSetInstructionCallConv(inst, LLVMFastCallConv);
2033 }
2034
2035 /*
2036 * cached lookup
2037 */
2038 static LLVMValueRef
compressed_fetch_cached(struct gallivm_state * gallivm,const struct util_format_description * format_desc,unsigned n,LLVMValueRef base_ptr,LLVMValueRef offset,LLVMValueRef i,LLVMValueRef j,LLVMValueRef cache)2039 compressed_fetch_cached(struct gallivm_state *gallivm,
2040 const struct util_format_description *format_desc,
2041 unsigned n,
2042 LLVMValueRef base_ptr,
2043 LLVMValueRef offset,
2044 LLVMValueRef i,
2045 LLVMValueRef j,
2046 LLVMValueRef cache)
2047
2048 {
2049 LLVMBuilderRef builder = gallivm->builder;
2050 unsigned count, low_bit, log2size;
2051 LLVMValueRef color, offset_stored, addr, ptr_addrtrunc, tmp;
2052 LLVMValueRef ij_index, hash_index, hash_mask, block_index;
2053 LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
2054 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
2055 LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
2056 struct lp_type type;
2057 struct lp_build_context bld32;
2058 memset(&type, 0, sizeof type);
2059 type.width = 32;
2060 type.length = n;
2061
2062 lp_build_context_init(&bld32, gallivm, type);
2063
2064 /*
2065 * compute hash - we use direct mapped cache, the hash function could
2066 * be better but it needs to be simple
2067 * per-element:
2068 * compare offset with offset stored at tag (hash)
2069 * if not equal extract block, store block, update tag
2070 * extract color from cache
2071 * assemble colors
2072 */
2073
2074 low_bit = util_logbase2(format_desc->block.bits / 8);
2075 log2size = util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE);
2076 addr = LLVMBuildPtrToInt(builder, base_ptr, i64t, "");
2077 ptr_addrtrunc = LLVMBuildPtrToInt(builder, base_ptr, i32t, "");
2078 ptr_addrtrunc = lp_build_broadcast_scalar(&bld32, ptr_addrtrunc);
2079 /* For the hash function, first mask off the unused lowest bits. Then just
2080 do some xor with address bits - only use lower 32bits */
2081 ptr_addrtrunc = LLVMBuildAdd(builder, offset, ptr_addrtrunc, "");
2082 ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
2083 lp_build_const_int_vec(gallivm, type, low_bit), "");
2084 /* This only really makes sense for size 64,128,256 */
2085 hash_index = ptr_addrtrunc;
2086 ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
2087 lp_build_const_int_vec(gallivm, type, 2*log2size), "");
2088 hash_index = LLVMBuildXor(builder, ptr_addrtrunc, hash_index, "");
2089 tmp = LLVMBuildLShr(builder, hash_index,
2090 lp_build_const_int_vec(gallivm, type, log2size), "");
2091 hash_index = LLVMBuildXor(builder, hash_index, tmp, "");
2092
2093 hash_mask = lp_build_const_int_vec(gallivm, type, LP_BUILD_FORMAT_CACHE_SIZE - 1);
2094 hash_index = LLVMBuildAnd(builder, hash_index, hash_mask, "");
2095 ij_index = LLVMBuildShl(builder, i, lp_build_const_int_vec(gallivm, type, 2), "");
2096 ij_index = LLVMBuildAdd(builder, ij_index, j, "");
2097 block_index = LLVMBuildShl(builder, hash_index,
2098 lp_build_const_int_vec(gallivm, type, 4), "");
2099 block_index = LLVMBuildAdd(builder, ij_index, block_index, "");
2100
2101 if (n > 1) {
2102 color = bld32.undef;
2103 for (count = 0; count < n; count++) {
2104 LLVMValueRef index, cond, colorx;
2105 LLVMValueRef block_indexx, hash_indexx, addrx, offsetx, ptr_addrx;
2106 struct lp_build_if_state if_ctx;
2107
2108 index = lp_build_const_int32(gallivm, count);
2109 offsetx = LLVMBuildExtractElement(builder, offset, index, "");
2110 addrx = LLVMBuildZExt(builder, offsetx, i64t, "");
2111 addrx = LLVMBuildAdd(builder, addrx, addr, "");
2112 block_indexx = LLVMBuildExtractElement(builder, block_index, index, "");
2113 hash_indexx = LLVMBuildLShr(builder, block_indexx,
2114 lp_build_const_int32(gallivm, 4), "");
2115 offset_stored = s3tc_lookup_tag_data(gallivm, cache, hash_indexx);
2116 cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addrx, "");
2117
2118 lp_build_if(&if_ctx, gallivm, cond);
2119 {
2120 ptr_addrx = LLVMBuildIntToPtr(builder, addrx,
2121 LLVMPointerType(i8t, 0), "");
2122 update_cached_block(gallivm, format_desc, ptr_addrx, hash_indexx, cache);
2123 #if LP_BUILD_FORMAT_CACHE_DEBUG
2124 s3tc_update_cache_access(gallivm, cache, 1,
2125 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
2126 #endif
2127 }
2128 lp_build_endif(&if_ctx);
2129
2130 colorx = s3tc_lookup_cached_pixel(gallivm, cache, block_indexx);
2131
2132 color = LLVMBuildInsertElement(builder, color, colorx,
2133 lp_build_const_int32(gallivm, count), "");
2134 }
2135 }
2136 else {
2137 LLVMValueRef cond;
2138 struct lp_build_if_state if_ctx;
2139
2140 tmp = LLVMBuildZExt(builder, offset, i64t, "");
2141 addr = LLVMBuildAdd(builder, tmp, addr, "");
2142 offset_stored = s3tc_lookup_tag_data(gallivm, cache, hash_index);
2143 cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addr, "");
2144
2145 lp_build_if(&if_ctx, gallivm, cond);
2146 {
2147 tmp = LLVMBuildIntToPtr(builder, addr, LLVMPointerType(i8t, 0), "");
2148 update_cached_block(gallivm, format_desc, tmp, hash_index, cache);
2149 #if LP_BUILD_FORMAT_CACHE_DEBUG
2150 s3tc_update_cache_access(gallivm, cache, 1,
2151 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
2152 #endif
2153 }
2154 lp_build_endif(&if_ctx);
2155
2156 color = s3tc_lookup_cached_pixel(gallivm, cache, block_index);
2157 }
2158 #if LP_BUILD_FORMAT_CACHE_DEBUG
2159 s3tc_update_cache_access(gallivm, cache, n,
2160 LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL);
2161 #endif
2162 return LLVMBuildBitCast(builder, color, LLVMVectorType(i8t, n * 4), "");
2163 }
2164
2165
2166 static LLVMValueRef
s3tc_dxt5_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef colors,LLVMValueRef codewords,LLVMValueRef alpha_lo,LLVMValueRef alpha_hi,LLVMValueRef i,LLVMValueRef j)2167 s3tc_dxt5_to_rgba_aos(struct gallivm_state *gallivm,
2168 unsigned n,
2169 enum pipe_format format,
2170 LLVMValueRef colors,
2171 LLVMValueRef codewords,
2172 LLVMValueRef alpha_lo,
2173 LLVMValueRef alpha_hi,
2174 LLVMValueRef i,
2175 LLVMValueRef j)
2176 {
2177 return s3tc_dxt5_full_to_rgba_aos(gallivm, n, format, colors,
2178 codewords, alpha_lo, alpha_hi, i, j);
2179 }
2180
2181
2182 /**
2183 * @param n number of pixels processed (usually n=4, but it should also work with n=1
2184 * and multiples of 4)
2185 * @param base_ptr base pointer (32bit or 64bit pointer depending on the architecture)
2186 * @param offset <n x i32> vector with the relative offsets of the S3TC blocks
2187 * @param i is a <n x i32> vector with the x subpixel coordinate (0..3)
2188 * @param j is a <n x i32> vector with the y subpixel coordinate (0..3)
2189 * @return a <4*n x i8> vector with the pixel RGBA values in AoS
2190 */
2191 LLVMValueRef
lp_build_fetch_s3tc_rgba_aos(struct gallivm_state * gallivm,const struct util_format_description * format_desc,unsigned n,LLVMValueRef base_ptr,LLVMValueRef offset,LLVMValueRef i,LLVMValueRef j,LLVMValueRef cache)2192 lp_build_fetch_s3tc_rgba_aos(struct gallivm_state *gallivm,
2193 const struct util_format_description *format_desc,
2194 unsigned n,
2195 LLVMValueRef base_ptr,
2196 LLVMValueRef offset,
2197 LLVMValueRef i,
2198 LLVMValueRef j,
2199 LLVMValueRef cache)
2200 {
2201 LLVMValueRef rgba;
2202 LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
2203 LLVMBuilderRef builder = gallivm->builder;
2204
2205 assert(format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC);
2206 assert(format_desc->block.width == 4);
2207 assert(format_desc->block.height == 4);
2208
2209 assert((n == 1) || (n % 4 == 0));
2210
2211 /* debug_printf("format = %d\n", format_desc->format);*/
2212 if (cache) {
2213 rgba = compressed_fetch_cached(gallivm, format_desc, n,
2214 base_ptr, offset, i, j, cache);
2215 return rgba;
2216 }
2217
2218 /*
2219 * Could use n > 8 here with avx2, but doesn't seem faster.
2220 */
2221 if (n > 4) {
2222 unsigned count;
2223 LLVMTypeRef i8_vectype = LLVMVectorType(i8t, 4 * n);
2224 LLVMTypeRef i128_type = LLVMIntTypeInContext(gallivm->context, 128);
2225 LLVMTypeRef i128_vectype = LLVMVectorType(i128_type, n / 4);
2226 LLVMTypeRef i324_vectype = LLVMVectorType(LLVMInt32TypeInContext(
2227 gallivm->context), 4);
2228 LLVMValueRef offset4, i4, j4, rgba4[LP_MAX_VECTOR_LENGTH/16];
2229 struct lp_type lp_324_vectype = lp_type_uint_vec(32, 128);
2230
2231 assert(n / 4 <= ARRAY_SIZE(rgba4));
2232
2233 rgba = LLVMGetUndef(i128_vectype);
2234
2235 for (count = 0; count < n / 4; count++) {
2236 LLVMValueRef colors, codewords, alpha_lo = NULL, alpha_hi = NULL;
2237
2238 i4 = lp_build_extract_range(gallivm, i, count * 4, 4);
2239 j4 = lp_build_extract_range(gallivm, j, count * 4, 4);
2240 offset4 = lp_build_extract_range(gallivm, offset, count * 4, 4);
2241
2242 lp_build_gather_s3tc(gallivm, 4, format_desc, &colors, &codewords,
2243 &alpha_lo, &alpha_hi, base_ptr, offset4);
2244
2245 switch (format_desc->format) {
2246 case PIPE_FORMAT_DXT1_RGB:
2247 case PIPE_FORMAT_DXT1_RGBA:
2248 case PIPE_FORMAT_DXT1_SRGB:
2249 case PIPE_FORMAT_DXT1_SRGBA:
2250 rgba4[count] = s3tc_dxt1_to_rgba_aos(gallivm, 4, format_desc->format,
2251 colors, codewords, i4, j4);
2252 break;
2253 case PIPE_FORMAT_DXT3_RGBA:
2254 case PIPE_FORMAT_DXT3_SRGBA:
2255 rgba4[count] = s3tc_dxt3_to_rgba_aos(gallivm, 4, format_desc->format, colors,
2256 codewords, alpha_lo, alpha_hi, i4, j4);
2257 break;
2258 case PIPE_FORMAT_DXT5_RGBA:
2259 case PIPE_FORMAT_DXT5_SRGBA:
2260 rgba4[count] = s3tc_dxt5_to_rgba_aos(gallivm, 4, format_desc->format, colors,
2261 codewords, alpha_lo, alpha_hi, i4, j4);
2262 break;
2263 default:
2264 assert(0);
2265 rgba4[count] = LLVMGetUndef(LLVMVectorType(i8t, 4));
2266 break;
2267 }
2268 /* shuffles typically give best results with dword elements...*/
2269 rgba4[count] = LLVMBuildBitCast(builder, rgba4[count], i324_vectype, "");
2270 }
2271 rgba = lp_build_concat(gallivm, rgba4, lp_324_vectype, n / 4);
2272 rgba = LLVMBuildBitCast(builder, rgba, i8_vectype, "");
2273 }
2274 else {
2275 LLVMValueRef colors, codewords, alpha_lo = NULL, alpha_hi = NULL;
2276
2277 lp_build_gather_s3tc(gallivm, n, format_desc, &colors, &codewords,
2278 &alpha_lo, &alpha_hi, base_ptr, offset);
2279
2280 switch (format_desc->format) {
2281 case PIPE_FORMAT_DXT1_RGB:
2282 case PIPE_FORMAT_DXT1_RGBA:
2283 case PIPE_FORMAT_DXT1_SRGB:
2284 case PIPE_FORMAT_DXT1_SRGBA:
2285 rgba = s3tc_dxt1_to_rgba_aos(gallivm, n, format_desc->format,
2286 colors, codewords, i, j);
2287 break;
2288 case PIPE_FORMAT_DXT3_RGBA:
2289 case PIPE_FORMAT_DXT3_SRGBA:
2290 rgba = s3tc_dxt3_to_rgba_aos(gallivm, n, format_desc->format, colors,
2291 codewords, alpha_lo, alpha_hi, i, j);
2292 break;
2293 case PIPE_FORMAT_DXT5_RGBA:
2294 case PIPE_FORMAT_DXT5_SRGBA:
2295 rgba = s3tc_dxt5_to_rgba_aos(gallivm, n, format_desc->format, colors,
2296 codewords, alpha_lo, alpha_hi, i, j);
2297 break;
2298 default:
2299 assert(0);
2300 rgba = LLVMGetUndef(LLVMVectorType(i8t, 4*n));
2301 break;
2302 }
2303 }
2304
2305 /* always return just decompressed values - srgb conversion is done later */
2306
2307 return rgba;
2308 }
2309
2310 /**
2311 * Gather elements from scatter positions in memory into vectors.
2312 * This is customised for fetching texels from s3tc textures.
2313 * For SSE, typical value is length=4.
2314 *
2315 * @param length length of the offsets
2316 * @param colors the stored colors of the blocks will be extracted into this.
2317 * @param codewords the codewords of the blocks will be extracted into this.
2318 * @param alpha_lo used for storing lower 32bit of alpha components for dxt3/5
2319 * @param alpha_hi used for storing higher 32bit of alpha components for dxt3/5
2320 * @param base_ptr base pointer, should be a i8 pointer type.
2321 * @param offsets vector with offsets
2322 */
2323 static void
lp_build_gather_rgtc(struct gallivm_state * gallivm,unsigned length,const struct util_format_description * format_desc,LLVMValueRef * red_lo,LLVMValueRef * red_hi,LLVMValueRef * green_lo,LLVMValueRef * green_hi,LLVMValueRef base_ptr,LLVMValueRef offsets)2324 lp_build_gather_rgtc(struct gallivm_state *gallivm,
2325 unsigned length,
2326 const struct util_format_description *format_desc,
2327 LLVMValueRef *red_lo, LLVMValueRef *red_hi,
2328 LLVMValueRef *green_lo, LLVMValueRef *green_hi,
2329 LLVMValueRef base_ptr,
2330 LLVMValueRef offsets)
2331 {
2332 LLVMBuilderRef builder = gallivm->builder;
2333 unsigned block_bits = format_desc->block.bits;
2334 unsigned i;
2335 LLVMValueRef elems[8];
2336 LLVMTypeRef type32 = LLVMInt32TypeInContext(gallivm->context);
2337 LLVMTypeRef type64 = LLVMInt64TypeInContext(gallivm->context);
2338 LLVMTypeRef type32dxt;
2339 struct lp_type lp_type32dxt;
2340
2341 memset(&lp_type32dxt, 0, sizeof lp_type32dxt);
2342 lp_type32dxt.width = 32;
2343 lp_type32dxt.length = block_bits / 32;
2344 type32dxt = lp_build_vec_type(gallivm, lp_type32dxt);
2345
2346 assert(block_bits == 64 || block_bits == 128);
2347 assert(length == 1 || length == 4 || length == 8);
2348
2349 for (i = 0; i < length; ++i) {
2350 elems[i] = lp_build_gather_elem(gallivm, length,
2351 block_bits, block_bits, TRUE,
2352 base_ptr, offsets, i, FALSE);
2353 elems[i] = LLVMBuildBitCast(builder, elems[i], type32dxt, "");
2354 }
2355 if (length == 1) {
2356 LLVMValueRef elem = elems[0];
2357
2358 *red_lo = LLVMBuildExtractElement(builder, elem,
2359 lp_build_const_int32(gallivm, 0), "");
2360 *red_hi = LLVMBuildExtractElement(builder, elem,
2361 lp_build_const_int32(gallivm, 1), "");
2362
2363 if (block_bits == 128) {
2364 *green_lo = LLVMBuildExtractElement(builder, elem,
2365 lp_build_const_int32(gallivm, 2), "");
2366 *green_hi = LLVMBuildExtractElement(builder, elem,
2367 lp_build_const_int32(gallivm, 3), "");
2368 } else {
2369 *green_lo = NULL;
2370 *green_hi = NULL;
2371 }
2372 } else {
2373 LLVMValueRef tmp[4];
2374 struct lp_type lp_type32, lp_type64;
2375 memset(&lp_type32, 0, sizeof lp_type32);
2376 lp_type32.width = 32;
2377 lp_type32.length = length;
2378 lp_type32.sign = lp_type32dxt.sign;
2379 memset(&lp_type64, 0, sizeof lp_type64);
2380 lp_type64.width = 64;
2381 lp_type64.length = length/2;
2382 if (block_bits == 128) {
2383 if (length == 8) {
2384 for (i = 0; i < 4; ++i) {
2385 tmp[0] = elems[i];
2386 tmp[1] = elems[i+4];
2387 elems[i] = lp_build_concat(gallivm, tmp, lp_type32dxt, 2);
2388 }
2389 }
2390 lp_build_transpose_aos(gallivm, lp_type32, elems, tmp);
2391 *green_lo = tmp[2];
2392 *green_hi = tmp[3];
2393 *red_lo = tmp[0];
2394 *red_hi = tmp[1];
2395 } else {
2396 LLVMValueRef red01, red23;
2397 LLVMTypeRef type64_vec = LLVMVectorType(type64, length/2);
2398 LLVMTypeRef type32_vec = LLVMVectorType(type32, length);
2399
2400 for (i = 0; i < length; ++i) {
2401 /* no-op shuffle */
2402 elems[i] = LLVMBuildShuffleVector(builder, elems[i],
2403 LLVMGetUndef(type32dxt),
2404 lp_build_const_extend_shuffle(gallivm, 2, 4), "");
2405 }
2406 if (length == 8) {
2407 struct lp_type lp_type32_4 = {0};
2408 lp_type32_4.width = 32;
2409 lp_type32_4.length = 4;
2410 for (i = 0; i < 4; ++i) {
2411 tmp[0] = elems[i];
2412 tmp[1] = elems[i+4];
2413 elems[i] = lp_build_concat(gallivm, tmp, lp_type32_4, 2);
2414 }
2415 }
2416 red01 = lp_build_interleave2_half(gallivm, lp_type32, elems[0], elems[1], 0);
2417 red23 = lp_build_interleave2_half(gallivm, lp_type32, elems[2], elems[3], 0);
2418 red01 = LLVMBuildBitCast(builder, red01, type64_vec, "");
2419 red23 = LLVMBuildBitCast(builder, red23, type64_vec, "");
2420 *red_lo = lp_build_interleave2_half(gallivm, lp_type64, red01, red23, 0);
2421 *red_hi = lp_build_interleave2_half(gallivm, lp_type64, red01, red23, 1);
2422 *red_lo = LLVMBuildBitCast(builder, *red_lo, type32_vec, "");
2423 *red_hi = LLVMBuildBitCast(builder, *red_hi, type32_vec, "");
2424 *green_lo = NULL;
2425 *green_hi = NULL;
2426 }
2427 }
2428 }
2429
2430 static LLVMValueRef
rgtc1_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef red_lo,LLVMValueRef red_hi,LLVMValueRef i,LLVMValueRef j)2431 rgtc1_to_rgba_aos(struct gallivm_state *gallivm,
2432 unsigned n,
2433 enum pipe_format format,
2434 LLVMValueRef red_lo,
2435 LLVMValueRef red_hi,
2436 LLVMValueRef i,
2437 LLVMValueRef j)
2438 {
2439 LLVMBuilderRef builder = gallivm->builder;
2440 bool is_signed = (format == PIPE_FORMAT_RGTC1_SNORM);
2441 LLVMValueRef red = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, red_hi, red_lo, i, j);
2442 LLVMValueRef rgba;
2443 struct lp_type type, type8;
2444 memset(&type, 0, sizeof type);
2445 type.width = 32;
2446 type.length = n;
2447 memset(&type8, 0, sizeof type8);
2448 type8.width = 8;
2449 type8.length = n*4;
2450 rgba = lp_build_const_int_vec(gallivm, type, is_signed ? (0x7f << 24) : (0xff << 24));
2451 rgba = LLVMBuildOr(builder, rgba, red, "");
2452 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
2453 }
2454
2455 static LLVMValueRef
rgtc2_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef red_lo,LLVMValueRef red_hi,LLVMValueRef green_lo,LLVMValueRef green_hi,LLVMValueRef i,LLVMValueRef j)2456 rgtc2_to_rgba_aos(struct gallivm_state *gallivm,
2457 unsigned n,
2458 enum pipe_format format,
2459 LLVMValueRef red_lo,
2460 LLVMValueRef red_hi,
2461 LLVMValueRef green_lo,
2462 LLVMValueRef green_hi,
2463 LLVMValueRef i,
2464 LLVMValueRef j)
2465 {
2466 LLVMBuilderRef builder = gallivm->builder;
2467 bool is_signed = (format == PIPE_FORMAT_RGTC2_SNORM);
2468 LLVMValueRef red = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, red_hi, red_lo, i, j);
2469 LLVMValueRef green = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, green_hi, green_lo, i, j);
2470 LLVMValueRef rgba;
2471 struct lp_type type, type8;
2472 memset(&type, 0, sizeof type);
2473 type.width = 32;
2474 type.length = n;
2475 memset(&type8, 0, sizeof type8);
2476 type8.width = 8;
2477 type8.length = n*4;
2478 rgba = lp_build_const_int_vec(gallivm, type, is_signed ? (0x7f << 24) : (0xff << 24));
2479 rgba = LLVMBuildOr(builder, rgba, red, "");
2480 green = LLVMBuildShl(builder, green, lp_build_const_int_vec(gallivm, type, 8), "");
2481 rgba = LLVMBuildOr(builder, rgba, green, "");
2482 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
2483 }
2484
2485 static LLVMValueRef
latc1_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef red_lo,LLVMValueRef red_hi,LLVMValueRef i,LLVMValueRef j)2486 latc1_to_rgba_aos(struct gallivm_state *gallivm,
2487 unsigned n,
2488 enum pipe_format format,
2489 LLVMValueRef red_lo,
2490 LLVMValueRef red_hi,
2491 LLVMValueRef i,
2492 LLVMValueRef j)
2493 {
2494 LLVMBuilderRef builder = gallivm->builder;
2495 bool is_signed = (format == PIPE_FORMAT_LATC1_SNORM);
2496 LLVMValueRef red = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, red_hi, red_lo, i, j);
2497 LLVMValueRef rgba, temp;
2498 struct lp_type type, type8;
2499 memset(&type, 0, sizeof type);
2500 type.width = 32;
2501 type.length = n;
2502 memset(&type8, 0, sizeof type8);
2503 type8.width = 8;
2504 type8.length = n*4;
2505 rgba = lp_build_const_int_vec(gallivm, type, is_signed ? (0x7f << 24) : (0xff << 24));
2506 rgba = LLVMBuildOr(builder, rgba, red, "");
2507 temp = LLVMBuildShl(builder, red, lp_build_const_int_vec(gallivm, type, 8), "");
2508 rgba = LLVMBuildOr(builder, rgba, temp, "");
2509 temp = LLVMBuildShl(builder, red, lp_build_const_int_vec(gallivm, type, 16), "");
2510 rgba = LLVMBuildOr(builder, rgba, temp, "");
2511 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
2512 }
2513
2514 static LLVMValueRef
latc2_to_rgba_aos(struct gallivm_state * gallivm,unsigned n,enum pipe_format format,LLVMValueRef red_lo,LLVMValueRef red_hi,LLVMValueRef green_lo,LLVMValueRef green_hi,LLVMValueRef i,LLVMValueRef j)2515 latc2_to_rgba_aos(struct gallivm_state *gallivm,
2516 unsigned n,
2517 enum pipe_format format,
2518 LLVMValueRef red_lo,
2519 LLVMValueRef red_hi,
2520 LLVMValueRef green_lo,
2521 LLVMValueRef green_hi,
2522 LLVMValueRef i,
2523 LLVMValueRef j)
2524 {
2525 LLVMBuilderRef builder = gallivm->builder;
2526 bool is_signed = (format == PIPE_FORMAT_LATC2_SNORM);
2527 LLVMValueRef red = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, red_hi, red_lo, i, j);
2528 LLVMValueRef green = s3tc_dxt5_alpha_channel(gallivm, is_signed, n, green_hi, green_lo, i, j);
2529 LLVMValueRef rgba, temp;
2530 struct lp_type type, type8;
2531 memset(&type, 0, sizeof type);
2532 type.width = 32;
2533 type.length = n;
2534 memset(&type8, 0, sizeof type8);
2535 type8.width = 8;
2536 type8.length = n*4;
2537
2538 temp = LLVMBuildShl(builder, red, lp_build_const_int_vec(gallivm, type, 8), "");
2539 rgba = LLVMBuildOr(builder, red, temp, "");
2540 temp = LLVMBuildShl(builder, red, lp_build_const_int_vec(gallivm, type, 16), "");
2541 rgba = LLVMBuildOr(builder, rgba, temp, "");
2542 temp = LLVMBuildShl(builder, green, lp_build_const_int_vec(gallivm, type, 24), "");
2543 rgba = LLVMBuildOr(builder, rgba, temp, "");
2544 return LLVMBuildBitCast(builder, rgba, lp_build_vec_type(gallivm, type8), "");
2545 }
2546
2547 /**
2548 * @param n number of pixels processed (usually n=4, but it should also work with n=1
2549 * and multiples of 4)
2550 * @param base_ptr base pointer (32bit or 64bit pointer depending on the architecture)
2551 * @param offset <n x i32> vector with the relative offsets of the S3TC blocks
2552 * @param i is a <n x i32> vector with the x subpixel coordinate (0..3)
2553 * @param j is a <n x i32> vector with the y subpixel coordinate (0..3)
2554 * @return a <4*n x i8> vector with the pixel RGBA values in AoS
2555 */
2556 LLVMValueRef
lp_build_fetch_rgtc_rgba_aos(struct gallivm_state * gallivm,const struct util_format_description * format_desc,unsigned n,LLVMValueRef base_ptr,LLVMValueRef offset,LLVMValueRef i,LLVMValueRef j,LLVMValueRef cache)2557 lp_build_fetch_rgtc_rgba_aos(struct gallivm_state *gallivm,
2558 const struct util_format_description *format_desc,
2559 unsigned n,
2560 LLVMValueRef base_ptr,
2561 LLVMValueRef offset,
2562 LLVMValueRef i,
2563 LLVMValueRef j,
2564 LLVMValueRef cache)
2565 {
2566 LLVMValueRef rgba;
2567 LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
2568 LLVMBuilderRef builder = gallivm->builder;
2569 LLVMValueRef red_lo, red_hi, green_lo, green_hi;
2570 assert(format_desc->layout == UTIL_FORMAT_LAYOUT_RGTC);
2571 assert(format_desc->block.width == 4);
2572 assert(format_desc->block.height == 4);
2573
2574 assert((n == 1) || (n % 4 == 0));
2575
2576 if (n > 4) {
2577 unsigned count;
2578 LLVMTypeRef i128_type = LLVMIntTypeInContext(gallivm->context, 128);
2579 LLVMTypeRef i128_vectype = LLVMVectorType(i128_type, n / 4);
2580 LLVMTypeRef i8_vectype = LLVMVectorType(i8t, 4 * n);
2581 LLVMTypeRef i324_vectype = LLVMVectorType(LLVMInt32TypeInContext(
2582 gallivm->context), 4);
2583 LLVMValueRef offset4, i4, j4, rgba4[LP_MAX_VECTOR_LENGTH/16];
2584 struct lp_type lp_324_vectype = lp_type_uint_vec(32, 128);
2585
2586 rgba = LLVMGetUndef(i128_vectype);
2587
2588 for (count = 0; count < n / 4; count++) {
2589
2590 i4 = lp_build_extract_range(gallivm, i, count * 4, 4);
2591 j4 = lp_build_extract_range(gallivm, j, count * 4, 4);
2592 offset4 = lp_build_extract_range(gallivm, offset, count * 4, 4);
2593
2594 lp_build_gather_rgtc(gallivm, 4, format_desc, &red_lo, &red_hi,
2595 &green_lo, &green_hi, base_ptr, offset4);
2596
2597 switch (format_desc->format) {
2598 case PIPE_FORMAT_RGTC1_UNORM:
2599 case PIPE_FORMAT_RGTC1_SNORM:
2600 rgba4[count] = rgtc1_to_rgba_aos(gallivm, 4, format_desc->format,
2601 red_lo, red_hi, i4, j4);
2602 break;
2603 case PIPE_FORMAT_RGTC2_UNORM:
2604 case PIPE_FORMAT_RGTC2_SNORM:
2605 rgba4[count] = rgtc2_to_rgba_aos(gallivm, 4, format_desc->format,
2606 red_lo, red_hi, green_lo, green_hi, i4, j4);
2607 break;
2608 case PIPE_FORMAT_LATC1_UNORM:
2609 case PIPE_FORMAT_LATC1_SNORM:
2610 rgba4[count] = latc1_to_rgba_aos(gallivm, 4, format_desc->format,
2611 red_lo, red_hi, i4, j4);
2612 break;
2613 case PIPE_FORMAT_LATC2_UNORM:
2614 case PIPE_FORMAT_LATC2_SNORM:
2615 rgba4[count] = latc2_to_rgba_aos(gallivm, 4, format_desc->format,
2616 red_lo, red_hi, green_lo, green_hi, i4, j4);
2617 break;
2618 default:
2619 assert(0);
2620 rgba4[count] = LLVMGetUndef(LLVMVectorType(i8t, 4));
2621 break;
2622 }
2623 /* shuffles typically give best results with dword elements...*/
2624 rgba4[count] = LLVMBuildBitCast(builder, rgba4[count], i324_vectype, "");
2625 }
2626 rgba = lp_build_concat(gallivm, rgba4, lp_324_vectype, n / 4);
2627 rgba = LLVMBuildBitCast(builder, rgba, i8_vectype, "");
2628 } else {
2629 LLVMValueRef red_lo, red_hi, green_lo, green_hi;
2630
2631 lp_build_gather_rgtc(gallivm, n, format_desc, &red_lo, &red_hi,
2632 &green_lo, &green_hi, base_ptr, offset);
2633
2634 switch (format_desc->format) {
2635 case PIPE_FORMAT_RGTC1_UNORM:
2636 case PIPE_FORMAT_RGTC1_SNORM:
2637 rgba = rgtc1_to_rgba_aos(gallivm, n, format_desc->format,
2638 red_lo, red_hi, i, j);
2639 break;
2640 case PIPE_FORMAT_RGTC2_UNORM:
2641 case PIPE_FORMAT_RGTC2_SNORM:
2642 rgba = rgtc2_to_rgba_aos(gallivm, n, format_desc->format,
2643 red_lo, red_hi, green_lo, green_hi, i, j);
2644 break;
2645 case PIPE_FORMAT_LATC1_UNORM:
2646 case PIPE_FORMAT_LATC1_SNORM:
2647 rgba = latc1_to_rgba_aos(gallivm, n, format_desc->format,
2648 red_lo, red_hi, i, j);
2649 break;
2650 case PIPE_FORMAT_LATC2_UNORM:
2651 case PIPE_FORMAT_LATC2_SNORM:
2652 rgba = latc2_to_rgba_aos(gallivm, n, format_desc->format,
2653 red_lo, red_hi, green_lo, green_hi, i, j);
2654 break;
2655 default:
2656 assert(0);
2657 rgba = LLVMGetUndef(LLVMVectorType(i8t, 4*n));
2658 break;
2659 }
2660 }
2661 return rgba;
2662 }
2663