projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: Make lp_build_sample_wrap_int usage more accurate.
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_sample_soa.c
1 /**************************************************************************
2 *
3 * Copyright 2009 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 above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 /**
29 * @file
30 * Texture sampling -- SoA.
31 *
32 * @author Jose Fonseca <jfonseca@vmware.com>
33 * @author Brian Paul <brianp@vmware.com>
34 */
35
36 #include "pipe/p_defines.h"
37 #include "pipe/p_state.h"
38 #include "util/u_debug.h"
39 #include "util/u_dump.h"
40 #include "util/u_memory.h"
41 #include "util/u_math.h"
42 #include "util/u_format.h"
43 #include "util/u_cpu_detect.h"
44 #include "lp_bld_debug.h"
45 #include "lp_bld_type.h"
46 #include "lp_bld_const.h"
47 #include "lp_bld_conv.h"
48 #include "lp_bld_arit.h"
49 #include "lp_bld_logic.h"
50 #include "lp_bld_swizzle.h"
51 #include "lp_bld_pack.h"
52 #include "lp_bld_flow.h"
53 #include "lp_bld_format.h"
54 #include "lp_bld_sample.h"
55
56
57 /**
58 * Keep all information for sampling code generation in a single place.
59 */
60 struct lp_build_sample_context
61 {
62 LLVMBuilderRef builder;
63
64 const struct lp_sampler_static_state *static_state;
65
66 struct lp_sampler_dynamic_state *dynamic_state;
67
68 const struct util_format_description *format_desc;
69
70 /** regular scalar float type */
71 struct lp_type float_type;
72 struct lp_build_context float_bld;
73
74 /** regular scalar float type */
75 struct lp_type int_type;
76 struct lp_build_context int_bld;
77
78 /** Incoming coordinates type and build context */
79 struct lp_type coord_type;
80 struct lp_build_context coord_bld;
81
82 /** Unsigned integer coordinates */
83 struct lp_type uint_coord_type;
84 struct lp_build_context uint_coord_bld;
85
86 /** Signed integer coordinates */
87 struct lp_type int_coord_type;
88 struct lp_build_context int_coord_bld;
89
90 /** Output texels type and build context */
91 struct lp_type texel_type;
92 struct lp_build_context texel_bld;
93 };
94
95
96 /**
97 * Does the given texture wrap mode allow sampling the texture border color?
98 * XXX maybe move this into gallium util code.
99 */
100 static boolean
101 wrap_mode_uses_border_color(unsigned mode)
102 {
103 switch (mode) {
104 case PIPE_TEX_WRAP_REPEAT:
105 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
106 case PIPE_TEX_WRAP_MIRROR_REPEAT:
107 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
108 return FALSE;
109 case PIPE_TEX_WRAP_CLAMP:
110 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
111 case PIPE_TEX_WRAP_MIRROR_CLAMP:
112 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
113 return TRUE;
114 default:
115 assert(0 && "unexpected wrap mode");
116 return FALSE;
117 }
118 }
119
120
121 static LLVMValueRef
122 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
123 LLVMValueRef data_array, LLVMValueRef level)
124 {
125 LLVMValueRef indexes[2], data_ptr;
126 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
127 indexes[1] = level;
128 data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
129 data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
130 return data_ptr;
131 }
132
133
134 static LLVMValueRef
135 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
136 LLVMValueRef data_array, int level)
137 {
138 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
139 return lp_build_get_mipmap_level(bld, data_array, lvl);
140 }
141
142
143 /**
144 * Dereference stride_array[mipmap_level] array to get a stride.
145 * Return stride as a vector.
146 */
147 static LLVMValueRef
148 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
149 LLVMValueRef stride_array, LLVMValueRef level)
150 {
151 LLVMValueRef indexes[2], stride;
152 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
153 indexes[1] = level;
154 stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
155 stride = LLVMBuildLoad(bld->builder, stride, "");
156 stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
157 return stride;
158 }
159
160
161 /** Dereference stride_array[0] array to get a stride (as vector). */
162 static LLVMValueRef
163 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
164 LLVMValueRef stride_array, int level)
165 {
166 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
167 return lp_build_get_level_stride_vec(bld, stride_array, lvl);
168 }
169
170
171 static int
172 texture_dims(enum pipe_texture_target tex)
173 {
174 switch (tex) {
175 case PIPE_TEXTURE_1D:
176 return 1;
177 case PIPE_TEXTURE_2D:
178 case PIPE_TEXTURE_CUBE:
179 return 2;
180 case PIPE_TEXTURE_3D:
181 return 3;
182 default:
183 assert(0 && "bad texture target in texture_dims()");
184 return 2;
185 }
186 }
187
188
189 static void
190 apply_sampler_swizzle(struct lp_build_sample_context *bld,
191 LLVMValueRef *texel)
192 {
193 unsigned char swizzles[4];
194
195 swizzles[0] = bld->static_state->swizzle_r;
196 swizzles[1] = bld->static_state->swizzle_g;
197 swizzles[2] = bld->static_state->swizzle_b;
198 swizzles[3] = bld->static_state->swizzle_a;
199
200 lp_build_swizzle_soa_inplace(&bld->texel_bld, texel, swizzles);
201 }
202
203
204
205 /**
206 * Generate code to fetch a texel from a texture at int coords (x, y, z).
207 * The computation depends on whether the texture is 1D, 2D or 3D.
208 * The result, texel, will be:
209 * texel[0] = red values
210 * texel[1] = green values
211 * texel[2] = blue values
212 * texel[3] = alpha values
213 */
214 static void
215 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
216 LLVMValueRef width,
217 LLVMValueRef height,
218 LLVMValueRef depth,
219 LLVMValueRef x,
220 LLVMValueRef y,
221 LLVMValueRef z,
222 LLVMValueRef y_stride,
223 LLVMValueRef z_stride,
224 LLVMValueRef data_ptr,
225 LLVMValueRef texel_out[4])
226 {
227 const int dims = texture_dims(bld->static_state->target);
228 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
229 LLVMValueRef offset;
230 LLVMValueRef i, j;
231 LLVMValueRef use_border = NULL;
232
233 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
234 if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
235 LLVMValueRef b1, b2;
236 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
237 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
238 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
239 }
240
241 if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
242 LLVMValueRef b1, b2;
243 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
244 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
245 if (use_border) {
246 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
247 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
248 }
249 else {
250 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
251 }
252 }
253
254 if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
255 LLVMValueRef b1, b2;
256 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
257 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
258 if (use_border) {
259 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
260 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
261 }
262 else {
263 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
264 }
265 }
266
267 /*
268 * Describe the coordinates in terms of pixel blocks.
269 *
270 * TODO: pixel blocks are power of two. LLVM should convert rem/div to
271 * bit arithmetic. Verify this.
272 */
273
274 if (bld->format_desc->block.width == 1) {
275 i = bld->uint_coord_bld.zero;
276 }
277 else {
278 LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width);
279 i = LLVMBuildURem(bld->builder, x, block_width, "");
280 x = LLVMBuildUDiv(bld->builder, x, block_width, "");
281 }
282
283 if (bld->format_desc->block.height == 1) {
284 j = bld->uint_coord_bld.zero;
285 }
286 else {
287 LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height);
288 j = LLVMBuildURem(bld->builder, y, block_height, "");
289 y = LLVMBuildUDiv(bld->builder, y, block_height, "");
290 }
291
292 /* convert x,y,z coords to linear offset from start of texture, in bytes */
293 offset = lp_build_sample_offset(&bld->uint_coord_bld,
294 bld->format_desc,
295 x, y, z, y_stride, z_stride);
296
297 if (use_border) {
298 /* If we can sample the border color, it means that texcoords may
299 * lie outside the bounds of the texture image. We need to do
300 * something to prevent reading out of bounds and causing a segfault.
301 *
302 * Simply AND the texture coords with !use_border. This will cause
303 * coords which are out of bounds to become zero. Zero's guaranteed
304 * to be inside the texture image.
305 */
306 offset = lp_build_andc(&bld->uint_coord_bld, offset, use_border);
307 }
308
309 lp_build_fetch_rgba_soa(bld->builder,
310 bld->format_desc,
311 bld->texel_type,
312 data_ptr, offset,
313 i, j,
314 texel_out);
315
316 apply_sampler_swizzle(bld, texel_out);
317
318 /*
319 * Note: if we find an app which frequently samples the texture border
320 * we might want to implement a true conditional here to avoid sampling
321 * the texture whenever possible (since that's quite a bit of code).
322 * Ex:
323 * if (use_border) {
324 * texel = border_color;
325 * }
326 * else {
327 * texel = sample_texture(coord);
328 * }
329 * As it is now, we always sample the texture, then selectively replace
330 * the texel color results with the border color.
331 */
332
333 if (use_border) {
334 /* select texel color or border color depending on use_border */
335 int chan;
336 for (chan = 0; chan < 4; chan++) {
337 LLVMValueRef border_chan =
338 lp_build_const_vec(bld->texel_type,
339 bld->static_state->border_color[chan]);
340 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
341 border_chan, texel_out[chan]);
342 }
343 }
344 }
345
346
347 static LLVMValueRef
348 lp_build_sample_packed(struct lp_build_sample_context *bld,
349 LLVMValueRef x,
350 LLVMValueRef y,
351 LLVMValueRef y_stride,
352 LLVMValueRef data_array)
353 {
354 LLVMValueRef offset;
355 LLVMValueRef data_ptr;
356
357 offset = lp_build_sample_offset(&bld->uint_coord_bld,
358 bld->format_desc,
359 x, y, NULL, y_stride, NULL);
360
361 assert(bld->format_desc->block.width == 1);
362 assert(bld->format_desc->block.height == 1);
363 assert(bld->format_desc->block.bits <= bld->texel_type.width);
364
365 /* get pointer to mipmap level 0 data */
366 data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
367
368 return lp_build_gather(bld->builder,
369 bld->texel_type.length,
370 bld->format_desc->block.bits,
371 bld->texel_type.width,
372 data_ptr, offset);
373 }
374
375
376 /**
377 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
378 */
379 static LLVMValueRef
380 lp_build_coord_mirror(struct lp_build_sample_context *bld,
381 LLVMValueRef coord)
382 {
383 struct lp_build_context *coord_bld = &bld->coord_bld;
384 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
385 LLVMValueRef fract, flr, isOdd;
386
387 /* fract = coord - floor(coord) */
388 fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
389
390 /* flr = ifloor(coord); */
391 flr = lp_build_ifloor(coord_bld, coord);
392
393 /* isOdd = flr & 1 */
394 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
395
396 /* make coord positive or negative depending on isOdd */
397 coord = lp_build_set_sign(coord_bld, fract, isOdd);
398
399 /* convert isOdd to float */
400 isOdd = lp_build_int_to_float(coord_bld, isOdd);
401
402 /* add isOdd to coord */
403 coord = lp_build_add(coord_bld, coord, isOdd);
404
405 return coord;
406 }
407
408
409 /**
410 * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
411 * Return whether the given mode is supported by that function.
412 */
413 static boolean
414 is_simple_wrap_mode(unsigned mode)
415 {
416 switch (mode) {
417 case PIPE_TEX_WRAP_REPEAT:
418 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
419 return TRUE;
420 default:
421 return FALSE;
422 }
423 }
424
425
426 /**
427 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
428 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
429 * \param length the texture size along one dimension
430 * \param is_pot if TRUE, length is a power of two
431 * \param wrap_mode one of PIPE_TEX_WRAP_x
432 */
433 static LLVMValueRef
434 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
435 LLVMValueRef coord,
436 LLVMValueRef length,
437 boolean is_pot,
438 unsigned wrap_mode)
439 {
440 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
441 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
442 LLVMValueRef length_minus_one;
443
444 length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
445
446 switch(wrap_mode) {
447 case PIPE_TEX_WRAP_REPEAT:
448 if(is_pot)
449 coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
450 else
451 /* Signed remainder won't give the right results for negative
452 * dividends but unsigned remainder does.*/
453 coord = LLVMBuildURem(bld->builder, coord, length, "");
454 break;
455
456 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
457 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
458 coord = lp_build_min(int_coord_bld, coord, length_minus_one);
459 break;
460
461 case PIPE_TEX_WRAP_CLAMP:
462 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
463 case PIPE_TEX_WRAP_MIRROR_REPEAT:
464 case PIPE_TEX_WRAP_MIRROR_CLAMP:
465 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
466 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
467 default:
468 assert(0);
469 }
470
471 return coord;
472 }
473
474
475 /**
476 * Build LLVM code for texture wrap mode for linear filtering.
477 * \param x0_out returns first integer texcoord
478 * \param x1_out returns second integer texcoord
479 * \param weight_out returns linear interpolation weight
480 */
481 static void
482 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
483 LLVMValueRef coord,
484 LLVMValueRef length,
485 boolean is_pot,
486 unsigned wrap_mode,
487 LLVMValueRef *x0_out,
488 LLVMValueRef *x1_out,
489 LLVMValueRef *weight_out)
490 {
491 struct lp_build_context *coord_bld = &bld->coord_bld;
492 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
493 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
494 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
495 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
496 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
497 LLVMValueRef coord0, coord1, weight;
498
499 switch(wrap_mode) {
500 case PIPE_TEX_WRAP_REPEAT:
501 /* mul by size and subtract 0.5 */
502 coord = lp_build_mul(coord_bld, coord, length_f);
503 coord = lp_build_sub(coord_bld, coord, half);
504 /* convert to int */
505 coord0 = lp_build_ifloor(coord_bld, coord);
506 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
507 /* compute lerp weight */
508 weight = lp_build_fract(coord_bld, coord);
509 /* repeat wrap */
510 if (is_pot) {
511 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
512 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
513 }
514 else {
515 /* Signed remainder won't give the right results for negative
516 * dividends but unsigned remainder does.*/
517 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
518 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
519 }
520 break;
521
522 case PIPE_TEX_WRAP_CLAMP:
523 if (bld->static_state->normalized_coords) {
524 /* scale coord to length */
525 coord = lp_build_mul(coord_bld, coord, length_f);
526 }
527
528 /* clamp to [0, length] */
529 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
530
531 coord = lp_build_sub(coord_bld, coord, half);
532
533 weight = lp_build_fract(coord_bld, coord);
534 coord0 = lp_build_ifloor(coord_bld, coord);
535 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
536 break;
537
538 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
539 if (bld->static_state->normalized_coords) {
540 /* clamp to [0,1] */
541 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
542 /* mul by tex size and subtract 0.5 */
543 coord = lp_build_mul(coord_bld, coord, length_f);
544 coord = lp_build_sub(coord_bld, coord, half);
545 }
546 else {
547 LLVMValueRef min, max;
548 /* clamp to [0.5, length - 0.5] */
549 min = half;
550 max = lp_build_sub(coord_bld, length_f, min);
551 coord = lp_build_clamp(coord_bld, coord, min, max);
552 }
553 /* compute lerp weight */
554 weight = lp_build_fract(coord_bld, coord);
555 /* coord0 = floor(coord); */
556 coord0 = lp_build_ifloor(coord_bld, coord);
557 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
558 /* coord0 = max(coord0, 0) */
559 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
560 /* coord1 = min(coord1, length-1) */
561 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
562 break;
563
564 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
565 {
566 LLVMValueRef min, max;
567 if (bld->static_state->normalized_coords) {
568 /* scale coord to length */
569 coord = lp_build_mul(coord_bld, coord, length_f);
570 }
571 /* clamp to [-0.5, length + 0.5] */
572 min = lp_build_const_vec(coord_bld->type, -0.5F);
573 max = lp_build_sub(coord_bld, length_f, min);
574 coord = lp_build_clamp(coord_bld, coord, min, max);
575 coord = lp_build_sub(coord_bld, coord, half);
576 /* compute lerp weight */
577 weight = lp_build_fract(coord_bld, coord);
578 /* convert to int */
579 coord0 = lp_build_ifloor(coord_bld, coord);
580 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
581 }
582 break;
583
584 case PIPE_TEX_WRAP_MIRROR_REPEAT:
585 /* compute mirror function */
586 coord = lp_build_coord_mirror(bld, coord);
587
588 /* scale coord to length */
589 coord = lp_build_mul(coord_bld, coord, length_f);
590 coord = lp_build_sub(coord_bld, coord, half);
591
592 /* compute lerp weight */
593 weight = lp_build_fract(coord_bld, coord);
594
595 /* convert to int coords */
596 coord0 = lp_build_ifloor(coord_bld, coord);
597 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
598
599 /* coord0 = max(coord0, 0) */
600 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
601 /* coord1 = min(coord1, length-1) */
602 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
603 break;
604
605 case PIPE_TEX_WRAP_MIRROR_CLAMP:
606 coord = lp_build_abs(coord_bld, coord);
607
608 if (bld->static_state->normalized_coords) {
609 /* scale coord to length */
610 coord = lp_build_mul(coord_bld, coord, length_f);
611 }
612
613 /* clamp to [0, length] */
614 coord = lp_build_min(coord_bld, coord, length_f);
615
616 coord = lp_build_sub(coord_bld, coord, half);
617
618 weight = lp_build_fract(coord_bld, coord);
619 coord0 = lp_build_ifloor(coord_bld, coord);
620 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
621 break;
622
623 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
624 {
625 LLVMValueRef min, max;
626
627 coord = lp_build_abs(coord_bld, coord);
628
629 if (bld->static_state->normalized_coords) {
630 /* scale coord to length */
631 coord = lp_build_mul(coord_bld, coord, length_f);
632 }
633
634 /* clamp to [0.5, length - 0.5] */
635 min = half;
636 max = lp_build_sub(coord_bld, length_f, min);
637 coord = lp_build_clamp(coord_bld, coord, min, max);
638
639 coord = lp_build_sub(coord_bld, coord, half);
640
641 weight = lp_build_fract(coord_bld, coord);
642 coord0 = lp_build_ifloor(coord_bld, coord);
643 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
644 }
645 break;
646
647 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
648 {
649 LLVMValueRef min, max;
650
651 coord = lp_build_abs(coord_bld, coord);
652
653 if (bld->static_state->normalized_coords) {
654 /* scale coord to length */
655 coord = lp_build_mul(coord_bld, coord, length_f);
656 }
657
658 /* clamp to [-0.5, length + 0.5] */
659 min = lp_build_negate(coord_bld, half);
660 max = lp_build_sub(coord_bld, length_f, min);
661 coord = lp_build_clamp(coord_bld, coord, min, max);
662
663 coord = lp_build_sub(coord_bld, coord, half);
664
665 weight = lp_build_fract(coord_bld, coord);
666 coord0 = lp_build_ifloor(coord_bld, coord);
667 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
668 }
669 break;
670
671 default:
672 assert(0);
673 coord0 = NULL;
674 coord1 = NULL;
675 weight = NULL;
676 }
677
678 *x0_out = coord0;
679 *x1_out = coord1;
680 *weight_out = weight;
681 }
682
683
684 /**
685 * Build LLVM code for texture wrap mode for nearest filtering.
686 * \param coord the incoming texcoord (nominally in [0,1])
687 * \param length the texture size along one dimension, as int
688 * \param is_pot if TRUE, length is a power of two
689 * \param wrap_mode one of PIPE_TEX_WRAP_x
690 */
691 static LLVMValueRef
692 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
693 LLVMValueRef coord,
694 LLVMValueRef length,
695 boolean is_pot,
696 unsigned wrap_mode)
697 {
698 struct lp_build_context *coord_bld = &bld->coord_bld;
699 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
700 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
701 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
702 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
703 LLVMValueRef icoord;
704
705 switch(wrap_mode) {
706 case PIPE_TEX_WRAP_REPEAT:
707 coord = lp_build_mul(coord_bld, coord, length_f);
708 icoord = lp_build_ifloor(coord_bld, coord);
709 if (is_pot)
710 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
711 else
712 /* Signed remainder won't give the right results for negative
713 * dividends but unsigned remainder does.*/
714 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
715 break;
716
717 case PIPE_TEX_WRAP_CLAMP:
718 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
719 if (bld->static_state->normalized_coords) {
720 /* scale coord to length */
721 coord = lp_build_mul(coord_bld, coord, length_f);
722 }
723
724 /* floor */
725 icoord = lp_build_ifloor(coord_bld, coord);
726
727 /* clamp to [0, length - 1]. */
728 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
729 length_minus_one);
730 break;
731
732 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
733 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
734 {
735 LLVMValueRef min, max;
736
737 if (bld->static_state->normalized_coords) {
738 /* scale coord to length */
739 coord = lp_build_mul(coord_bld, coord, length_f);
740 }
741
742 icoord = lp_build_ifloor(coord_bld, coord);
743
744 /* clamp to [-1, length] */
745 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
746 max = length;
747 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
748 }
749 break;
750
751 case PIPE_TEX_WRAP_MIRROR_REPEAT:
752 /* compute mirror function */
753 coord = lp_build_coord_mirror(bld, coord);
754
755 /* scale coord to length */
756 assert(bld->static_state->normalized_coords);
757 coord = lp_build_mul(coord_bld, coord, length_f);
758
759 icoord = lp_build_ifloor(coord_bld, coord);
760
761 /* clamp to [0, length - 1] */
762 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
763 break;
764
765 case PIPE_TEX_WRAP_MIRROR_CLAMP:
766 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
767 coord = lp_build_abs(coord_bld, coord);
768
769 if (bld->static_state->normalized_coords) {
770 /* scale coord to length */
771 coord = lp_build_mul(coord_bld, coord, length_f);
772 }
773
774 icoord = lp_build_ifloor(coord_bld, coord);
775
776 /* clamp to [0, length - 1] */
777 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
778 break;
779
780 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
781 coord = lp_build_abs(coord_bld, coord);
782
783 if (bld->static_state->normalized_coords) {
784 /* scale coord to length */
785 coord = lp_build_mul(coord_bld, coord, length_f);
786 }
787
788 icoord = lp_build_ifloor(coord_bld, coord);
789
790 /* clamp to [0, length] */
791 icoord = lp_build_min(int_coord_bld, icoord, length);
792 break;
793
794 default:
795 assert(0);
796 icoord = NULL;
797 }
798
799 return icoord;
800 }
801
802
803 /**
804 * Codegen equivalent for u_minify().
805 * Return max(1, base_size >> level);
806 */
807 static LLVMValueRef
808 lp_build_minify(struct lp_build_sample_context *bld,
809 LLVMValueRef base_size,
810 LLVMValueRef level)
811 {
812 LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
813 size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
814 return size;
815 }
816
817
818 /**
819 * Generate code to compute texture level of detail (lambda).
820 * \param s vector of texcoord s values
821 * \param t vector of texcoord t values
822 * \param r vector of texcoord r values
823 * \param lod_bias optional float vector with the shader lod bias
824 * \param explicit_lod optional float vector with the explicit lod
825 * \param width scalar int texture width
826 * \param height scalar int texture height
827 * \param depth scalar int texture depth
828 *
829 * XXX: The resulting lod is scalar, so ignore all but the first element of
830 * derivatives, lod_bias, etc that are passed by the shader.
831 */
832 static LLVMValueRef
833 lp_build_lod_selector(struct lp_build_sample_context *bld,
834 LLVMValueRef s,
835 LLVMValueRef t,
836 LLVMValueRef r,
837 const LLVMValueRef *ddx,
838 const LLVMValueRef *ddy,
839 LLVMValueRef lod_bias, /* optional */
840 LLVMValueRef explicit_lod, /* optional */
841 LLVMValueRef width,
842 LLVMValueRef height,
843 LLVMValueRef depth)
844
845 {
846 if (bld->static_state->min_lod == bld->static_state->max_lod) {
847 /* User is forcing sampling from a particular mipmap level.
848 * This is hit during mipmap generation.
849 */
850 return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
851 }
852 else {
853 struct lp_build_context *float_bld = &bld->float_bld;
854 LLVMValueRef sampler_lod_bias = LLVMConstReal(LLVMFloatType(),
855 bld->static_state->lod_bias);
856 LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
857 bld->static_state->min_lod);
858 LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
859 bld->static_state->max_lod);
860 LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
861 LLVMValueRef lod;
862
863 if (explicit_lod) {
864 lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
865 index0, "");
866 }
867 else {
868 const int dims = texture_dims(bld->static_state->target);
869 LLVMValueRef dsdx, dsdy;
870 LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
871 LLVMValueRef rho;
872
873 /*
874 * dsdx = abs(s[1] - s[0]);
875 * dsdy = abs(s[2] - s[0]);
876 * dtdx = abs(t[1] - t[0]);
877 * dtdy = abs(t[2] - t[0]);
878 * drdx = abs(r[1] - r[0]);
879 * drdy = abs(r[2] - r[0]);
880 */
881 dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
882 dsdx = lp_build_abs(float_bld, dsdx);
883 dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
884 dsdy = lp_build_abs(float_bld, dsdy);
885 if (dims > 1) {
886 dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
887 dtdx = lp_build_abs(float_bld, dtdx);
888 dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
889 dtdy = lp_build_abs(float_bld, dtdy);
890 if (dims > 2) {
891 drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
892 drdx = lp_build_abs(float_bld, drdx);
893 drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
894 drdy = lp_build_abs(float_bld, drdy);
895 }
896 }
897
898 /* Compute rho = max of all partial derivatives scaled by texture size.
899 * XXX this could be vectorized somewhat
900 */
901 rho = LLVMBuildMul(bld->builder,
902 lp_build_max(float_bld, dsdx, dsdy),
903 lp_build_int_to_float(float_bld, width), "");
904 if (dims > 1) {
905 LLVMValueRef max;
906 max = LLVMBuildMul(bld->builder,
907 lp_build_max(float_bld, dtdx, dtdy),
908 lp_build_int_to_float(float_bld, height), "");
909 rho = lp_build_max(float_bld, rho, max);
910 if (dims > 2) {
911 max = LLVMBuildMul(bld->builder,
912 lp_build_max(float_bld, drdx, drdy),
913 lp_build_int_to_float(float_bld, depth), "");
914 rho = lp_build_max(float_bld, rho, max);
915 }
916 }
917
918 /* compute lod = log2(rho) */
919 lod = lp_build_log2(float_bld, rho);
920
921 /* add shader lod bias */
922 if (lod_bias) {
923 lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
924 index0, "");
925 lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
926 }
927 }
928
929 /* add sampler lod bias */
930 lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
931
932 /* clamp lod */
933 lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
934
935 return lod;
936 }
937 }
938
939
940 /**
941 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
942 * mipmap level index.
943 * Note: this is all scalar code.
944 * \param lod scalar float texture level of detail
945 * \param level_out returns integer
946 */
947 static void
948 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
949 unsigned unit,
950 LLVMValueRef lod,
951 LLVMValueRef *level_out)
952 {
953 struct lp_build_context *float_bld = &bld->float_bld;
954 struct lp_build_context *int_bld = &bld->int_bld;
955 LLVMValueRef last_level, level;
956
957 LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
958
959 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
960 bld->builder, unit);
961
962 /* convert float lod to integer */
963 level = lp_build_iround(float_bld, lod);
964
965 /* clamp level to legal range of levels */
966 *level_out = lp_build_clamp(int_bld, level, zero, last_level);
967 }
968
969
970 /**
971 * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
972 * two (adjacent) mipmap level indexes. Later, we'll sample from those
973 * two mipmap levels and interpolate between them.
974 */
975 static void
976 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
977 unsigned unit,
978 LLVMValueRef lod,
979 LLVMValueRef *level0_out,
980 LLVMValueRef *level1_out,
981 LLVMValueRef *weight_out)
982 {
983 struct lp_build_context *float_bld = &bld->float_bld;
984 struct lp_build_context *int_bld = &bld->int_bld;
985 LLVMValueRef last_level, level;
986
987 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
988 bld->builder, unit);
989
990 /* convert float lod to integer */
991 level = lp_build_ifloor(float_bld, lod);
992
993 /* compute level 0 and clamp to legal range of levels */
994 *level0_out = lp_build_clamp(int_bld, level,
995 int_bld->zero,
996 last_level);
997 /* compute level 1 and clamp to legal range of levels */
998 level = lp_build_add(int_bld, level, int_bld->one);
999 *level1_out = lp_build_clamp(int_bld, level,
1000 int_bld->zero,
1001 last_level);
1002
1003 *weight_out = lp_build_fract(float_bld, lod);
1004 }
1005
1006
1007 /**
1008 * Generate code to sample a mipmap level with nearest filtering.
1009 * If sampling a cube texture, r = cube face in [0,5].
1010 */
1011 static void
1012 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1013 LLVMValueRef width_vec,
1014 LLVMValueRef height_vec,
1015 LLVMValueRef depth_vec,
1016 LLVMValueRef row_stride_vec,
1017 LLVMValueRef img_stride_vec,
1018 LLVMValueRef data_ptr,
1019 LLVMValueRef s,
1020 LLVMValueRef t,
1021 LLVMValueRef r,
1022 LLVMValueRef colors_out[4])
1023 {
1024 const int dims = texture_dims(bld->static_state->target);
1025 LLVMValueRef x, y, z;
1026
1027 /*
1028 * Compute integer texcoords.
1029 */
1030 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1031 bld->static_state->pot_width,
1032 bld->static_state->wrap_s);
1033 lp_build_name(x, "tex.x.wrapped");
1034
1035 if (dims >= 2) {
1036 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1037 bld->static_state->pot_height,
1038 bld->static_state->wrap_t);
1039 lp_build_name(y, "tex.y.wrapped");
1040
1041 if (dims == 3) {
1042 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1043 bld->static_state->pot_height,
1044 bld->static_state->wrap_r);
1045 lp_build_name(z, "tex.z.wrapped");
1046 }
1047 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1048 z = r;
1049 }
1050 else {
1051 z = NULL;
1052 }
1053 }
1054 else {
1055 y = z = NULL;
1056 }
1057
1058 /*
1059 * Get texture colors.
1060 */
1061 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1062 x, y, z,
1063 row_stride_vec, img_stride_vec,
1064 data_ptr, colors_out);
1065 }
1066
1067
1068 /**
1069 * Generate code to sample a mipmap level with linear filtering.
1070 * If sampling a cube texture, r = cube face in [0,5].
1071 */
1072 static void
1073 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1074 LLVMValueRef width_vec,
1075 LLVMValueRef height_vec,
1076 LLVMValueRef depth_vec,
1077 LLVMValueRef row_stride_vec,
1078 LLVMValueRef img_stride_vec,
1079 LLVMValueRef data_ptr,
1080 LLVMValueRef s,
1081 LLVMValueRef t,
1082 LLVMValueRef r,
1083 LLVMValueRef colors_out[4])
1084 {
1085 const int dims = texture_dims(bld->static_state->target);
1086 LLVMValueRef x0, y0, z0, x1, y1, z1;
1087 LLVMValueRef s_fpart, t_fpart, r_fpart;
1088 LLVMValueRef neighbors[2][2][4];
1089 int chan;
1090
1091 /*
1092 * Compute integer texcoords.
1093 */
1094 lp_build_sample_wrap_linear(bld, s, width_vec,
1095 bld->static_state->pot_width,
1096 bld->static_state->wrap_s,
1097 &x0, &x1, &s_fpart);
1098 lp_build_name(x0, "tex.x0.wrapped");
1099 lp_build_name(x1, "tex.x1.wrapped");
1100
1101 if (dims >= 2) {
1102 lp_build_sample_wrap_linear(bld, t, height_vec,
1103 bld->static_state->pot_height,
1104 bld->static_state->wrap_t,
1105 &y0, &y1, &t_fpart);
1106 lp_build_name(y0, "tex.y0.wrapped");
1107 lp_build_name(y1, "tex.y1.wrapped");
1108
1109 if (dims == 3) {
1110 lp_build_sample_wrap_linear(bld, r, depth_vec,
1111 bld->static_state->pot_depth,
1112 bld->static_state->wrap_r,
1113 &z0, &z1, &r_fpart);
1114 lp_build_name(z0, "tex.z0.wrapped");
1115 lp_build_name(z1, "tex.z1.wrapped");
1116 }
1117 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1118 z0 = z1 = r; /* cube face */
1119 r_fpart = NULL;
1120 }
1121 else {
1122 z0 = z1 = NULL;
1123 r_fpart = NULL;
1124 }
1125 }
1126 else {
1127 y0 = y1 = t_fpart = NULL;
1128 z0 = z1 = r_fpart = NULL;
1129 }
1130
1131 /*
1132 * Get texture colors.
1133 */
1134 /* get x0/x1 texels */
1135 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1136 x0, y0, z0,
1137 row_stride_vec, img_stride_vec,
1138 data_ptr, neighbors[0][0]);
1139 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1140 x1, y0, z0,
1141 row_stride_vec, img_stride_vec,
1142 data_ptr, neighbors[0][1]);
1143
1144 if (dims == 1) {
1145 /* Interpolate two samples from 1D image to produce one color */
1146 for (chan = 0; chan < 4; chan++) {
1147 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1148 neighbors[0][0][chan],
1149 neighbors[0][1][chan]);
1150 }
1151 }
1152 else {
1153 /* 2D/3D texture */
1154 LLVMValueRef colors0[4];
1155
1156 /* get x0/x1 texels at y1 */
1157 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1158 x0, y1, z0,
1159 row_stride_vec, img_stride_vec,
1160 data_ptr, neighbors[1][0]);
1161 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1162 x1, y1, z0,
1163 row_stride_vec, img_stride_vec,
1164 data_ptr, neighbors[1][1]);
1165
1166 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1167 for (chan = 0; chan < 4; chan++) {
1168 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1169 s_fpart, t_fpart,
1170 neighbors[0][0][chan],
1171 neighbors[0][1][chan],
1172 neighbors[1][0][chan],
1173 neighbors[1][1][chan]);
1174 }
1175
1176 if (dims == 3) {
1177 LLVMValueRef neighbors1[2][2][4];
1178 LLVMValueRef colors1[4];
1179
1180 /* get x0/x1/y0/y1 texels at z1 */
1181 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1182 x0, y0, z1,
1183 row_stride_vec, img_stride_vec,
1184 data_ptr, neighbors1[0][0]);
1185 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1186 x1, y0, z1,
1187 row_stride_vec, img_stride_vec,
1188 data_ptr, neighbors1[0][1]);
1189 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1190 x0, y1, z1,
1191 row_stride_vec, img_stride_vec,
1192 data_ptr, neighbors1[1][0]);
1193 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1194 x1, y1, z1,
1195 row_stride_vec, img_stride_vec,
1196 data_ptr, neighbors1[1][1]);
1197
1198 /* Bilinear interpolate the four samples from the second Z slice */
1199 for (chan = 0; chan < 4; chan++) {
1200 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1201 s_fpart, t_fpart,
1202 neighbors1[0][0][chan],
1203 neighbors1[0][1][chan],
1204 neighbors1[1][0][chan],
1205 neighbors1[1][1][chan]);
1206 }
1207
1208 /* Linearly interpolate the two samples from the two 3D slices */
1209 for (chan = 0; chan < 4; chan++) {
1210 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1211 r_fpart,
1212 colors0[chan], colors1[chan]);
1213 }
1214 }
1215 else {
1216 /* 2D tex */
1217 for (chan = 0; chan < 4; chan++) {
1218 colors_out[chan] = colors0[chan];
1219 }
1220 }
1221 }
1222 }
1223
1224
1225 /** Helper used by lp_build_cube_lookup() */
1226 static LLVMValueRef
1227 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1228 {
1229 /* ima = -0.5 / abs(coord); */
1230 LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
1231 LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1232 LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
1233 lp_build_rcp(coord_bld, absCoord));
1234 return ima;
1235 }
1236
1237
1238 /**
1239 * Helper used by lp_build_cube_lookup()
1240 * \param sign scalar +1 or -1
1241 * \param coord float vector
1242 * \param ima float vector
1243 */
1244 static LLVMValueRef
1245 lp_build_cube_coord(struct lp_build_context *coord_bld,
1246 LLVMValueRef sign, int negate_coord,
1247 LLVMValueRef coord, LLVMValueRef ima)
1248 {
1249 /* return negate(coord) * ima * sign + 0.5; */
1250 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
1251 LLVMValueRef res;
1252
1253 assert(negate_coord == +1 || negate_coord == -1);
1254
1255 if (negate_coord == -1) {
1256 coord = lp_build_negate(coord_bld, coord);
1257 }
1258
1259 res = lp_build_mul(coord_bld, coord, ima);
1260 if (sign) {
1261 sign = lp_build_broadcast_scalar(coord_bld, sign);
1262 res = lp_build_mul(coord_bld, res, sign);
1263 }
1264 res = lp_build_add(coord_bld, res, half);
1265
1266 return res;
1267 }
1268
1269
1270 /** Helper used by lp_build_cube_lookup()
1271 * Return (major_coord >= 0) ? pos_face : neg_face;
1272 */
1273 static LLVMValueRef
1274 lp_build_cube_face(struct lp_build_sample_context *bld,
1275 LLVMValueRef major_coord,
1276 unsigned pos_face, unsigned neg_face)
1277 {
1278 LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1279 major_coord,
1280 bld->float_bld.zero, "");
1281 LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1282 LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1283 LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1284 return res;
1285 }
1286
1287
1288
1289 /**
1290 * Generate code to do cube face selection and per-face texcoords.
1291 */
1292 static void
1293 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1294 LLVMValueRef s,
1295 LLVMValueRef t,
1296 LLVMValueRef r,
1297 LLVMValueRef *face,
1298 LLVMValueRef *face_s,
1299 LLVMValueRef *face_t)
1300 {
1301 struct lp_build_context *float_bld = &bld->float_bld;
1302 struct lp_build_context *coord_bld = &bld->coord_bld;
1303 LLVMValueRef rx, ry, rz;
1304 LLVMValueRef arx, ary, arz;
1305 LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1306 LLVMValueRef arx_ge_ary, arx_ge_arz;
1307 LLVMValueRef ary_ge_arx, ary_ge_arz;
1308 LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1309 LLVMValueRef rx_pos, ry_pos, rz_pos;
1310
1311 assert(bld->coord_bld.type.length == 4);
1312
1313 /*
1314 * Use the average of the four pixel's texcoords to choose the face.
1315 */
1316 rx = lp_build_mul(float_bld, c25,
1317 lp_build_sum_vector(&bld->coord_bld, s));
1318 ry = lp_build_mul(float_bld, c25,
1319 lp_build_sum_vector(&bld->coord_bld, t));
1320 rz = lp_build_mul(float_bld, c25,
1321 lp_build_sum_vector(&bld->coord_bld, r));
1322
1323 arx = lp_build_abs(float_bld, rx);
1324 ary = lp_build_abs(float_bld, ry);
1325 arz = lp_build_abs(float_bld, rz);
1326
1327 /*
1328 * Compare sign/magnitude of rx,ry,rz to determine face
1329 */
1330 arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1331 arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1332 ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1333 ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1334
1335 arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1336 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1337
1338 rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1339 ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1340 rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1341
1342 {
1343 struct lp_build_flow_context *flow_ctx;
1344 struct lp_build_if_state if_ctx;
1345
1346 flow_ctx = lp_build_flow_create(bld->builder);
1347 lp_build_flow_scope_begin(flow_ctx);
1348
1349 *face_s = bld->coord_bld.undef;
1350 *face_t = bld->coord_bld.undef;
1351 *face = bld->int_bld.undef;
1352
1353 lp_build_name(*face_s, "face_s");
1354 lp_build_name(*face_t, "face_t");
1355 lp_build_name(*face, "face");
1356
1357 lp_build_flow_scope_declare(flow_ctx, face_s);
1358 lp_build_flow_scope_declare(flow_ctx, face_t);
1359 lp_build_flow_scope_declare(flow_ctx, face);
1360
1361 lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1362 {
1363 /* +/- X face */
1364 LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1365 LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1366 *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1367 *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1368 *face = lp_build_cube_face(bld, rx,
1369 PIPE_TEX_FACE_POS_X,
1370 PIPE_TEX_FACE_NEG_X);
1371 }
1372 lp_build_else(&if_ctx);
1373 {
1374 struct lp_build_flow_context *flow_ctx2;
1375 struct lp_build_if_state if_ctx2;
1376
1377 LLVMValueRef face_s2 = bld->coord_bld.undef;
1378 LLVMValueRef face_t2 = bld->coord_bld.undef;
1379 LLVMValueRef face2 = bld->int_bld.undef;
1380
1381 flow_ctx2 = lp_build_flow_create(bld->builder);
1382 lp_build_flow_scope_begin(flow_ctx2);
1383 lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1384 lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1385 lp_build_flow_scope_declare(flow_ctx2, &face2);
1386
1387 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1388
1389 lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1390 {
1391 /* +/- Y face */
1392 LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1393 LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1394 face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1395 face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1396 face2 = lp_build_cube_face(bld, ry,
1397 PIPE_TEX_FACE_POS_Y,
1398 PIPE_TEX_FACE_NEG_Y);
1399 }
1400 lp_build_else(&if_ctx2);
1401 {
1402 /* +/- Z face */
1403 LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1404 LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1405 face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1406 face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1407 face2 = lp_build_cube_face(bld, rz,
1408 PIPE_TEX_FACE_POS_Z,
1409 PIPE_TEX_FACE_NEG_Z);
1410 }
1411 lp_build_endif(&if_ctx2);
1412 lp_build_flow_scope_end(flow_ctx2);
1413 lp_build_flow_destroy(flow_ctx2);
1414
1415 *face_s = face_s2;
1416 *face_t = face_t2;
1417 *face = face2;
1418 }
1419
1420 lp_build_endif(&if_ctx);
1421 lp_build_flow_scope_end(flow_ctx);
1422 lp_build_flow_destroy(flow_ctx);
1423 }
1424 }
1425
1426
1427
1428 /**
1429 * Sample the texture/mipmap using given image filter and mip filter.
1430 * data0_ptr and data1_ptr point to the two mipmap levels to sample
1431 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
1432 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1433 */
1434 static void
1435 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1436 unsigned img_filter,
1437 unsigned mip_filter,
1438 LLVMValueRef s,
1439 LLVMValueRef t,
1440 LLVMValueRef r,
1441 LLVMValueRef lod_fpart,
1442 LLVMValueRef width0_vec,
1443 LLVMValueRef width1_vec,
1444 LLVMValueRef height0_vec,
1445 LLVMValueRef height1_vec,
1446 LLVMValueRef depth0_vec,
1447 LLVMValueRef depth1_vec,
1448 LLVMValueRef row_stride0_vec,
1449 LLVMValueRef row_stride1_vec,
1450 LLVMValueRef img_stride0_vec,
1451 LLVMValueRef img_stride1_vec,
1452 LLVMValueRef data_ptr0,
1453 LLVMValueRef data_ptr1,
1454 LLVMValueRef *colors_out)
1455 {
1456 LLVMValueRef colors0[4], colors1[4];
1457 int chan;
1458
1459 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1460 lp_build_sample_image_nearest(bld,
1461 width0_vec, height0_vec, depth0_vec,
1462 row_stride0_vec, img_stride0_vec,
1463 data_ptr0, s, t, r, colors0);
1464
1465 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1466 /* sample the second mipmap level, and interp */
1467 lp_build_sample_image_nearest(bld,
1468 width1_vec, height1_vec, depth1_vec,
1469 row_stride1_vec, img_stride1_vec,
1470 data_ptr1, s, t, r, colors1);
1471 }
1472 }
1473 else {
1474 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1475
1476 lp_build_sample_image_linear(bld,
1477 width0_vec, height0_vec, depth0_vec,
1478 row_stride0_vec, img_stride0_vec,
1479 data_ptr0, s, t, r, colors0);
1480
1481 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1482 /* sample the second mipmap level, and interp */
1483 lp_build_sample_image_linear(bld,
1484 width1_vec, height1_vec, depth1_vec,
1485 row_stride1_vec, img_stride1_vec,
1486 data_ptr1, s, t, r, colors1);
1487 }
1488 }
1489
1490 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1491 /* interpolate samples from the two mipmap levels */
1492 for (chan = 0; chan < 4; chan++) {
1493 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1494 colors0[chan], colors1[chan]);
1495 }
1496 }
1497 else {
1498 /* use first/only level's colors */
1499 for (chan = 0; chan < 4; chan++) {
1500 colors_out[chan] = colors0[chan];
1501 }
1502 }
1503 }
1504
1505
1506
1507 /**
1508 * General texture sampling codegen.
1509 * This function handles texture sampling for all texture targets (1D,
1510 * 2D, 3D, cube) and all filtering modes.
1511 */
1512 static void
1513 lp_build_sample_general(struct lp_build_sample_context *bld,
1514 unsigned unit,
1515 LLVMValueRef s,
1516 LLVMValueRef t,
1517 LLVMValueRef r,
1518 const LLVMValueRef *ddx,
1519 const LLVMValueRef *ddy,
1520 LLVMValueRef lod_bias, /* optional */
1521 LLVMValueRef explicit_lod, /* optional */
1522 LLVMValueRef width,
1523 LLVMValueRef height,
1524 LLVMValueRef depth,
1525 LLVMValueRef width_vec,
1526 LLVMValueRef height_vec,
1527 LLVMValueRef depth_vec,
1528 LLVMValueRef row_stride_array,
1529 LLVMValueRef img_stride_array,
1530 LLVMValueRef data_array,
1531 LLVMValueRef *colors_out)
1532 {
1533 struct lp_build_context *float_bld = &bld->float_bld;
1534 const unsigned mip_filter = bld->static_state->min_mip_filter;
1535 const unsigned min_filter = bld->static_state->min_img_filter;
1536 const unsigned mag_filter = bld->static_state->mag_img_filter;
1537 const int dims = texture_dims(bld->static_state->target);
1538 LLVMValueRef lod = NULL, lod_fpart = NULL;
1539 LLVMValueRef ilevel0, ilevel1 = NULL, ilevel0_vec, ilevel1_vec = NULL;
1540 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1541 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1542 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1543 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1544 LLVMValueRef data_ptr0, data_ptr1 = NULL;
1545
1546 /*
1547 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1548 mip_filter, min_filter, mag_filter);
1549 */
1550
1551 /*
1552 * Compute the level of detail (float).
1553 */
1554 if (min_filter != mag_filter ||
1555 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1556 /* Need to compute lod either to choose mipmap levels or to
1557 * distinguish between minification/magnification with one mipmap level.
1558 */
1559 lod = lp_build_lod_selector(bld, s, t, r, ddx, ddy,
1560 lod_bias, explicit_lod,
1561 width, height, depth);
1562 }
1563
1564 /*
1565 * Compute integer mipmap level(s) to fetch texels from.
1566 */
1567 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1568 /* always use mip level 0 */
1569 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1570 }
1571 else {
1572 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1573 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1574 }
1575 else {
1576 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1577 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1578 &lod_fpart);
1579 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1580 }
1581 }
1582
1583 /*
1584 * Convert scalar integer mipmap levels into vectors.
1585 */
1586 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1587 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1588 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1589
1590 /*
1591 * Compute width, height at mipmap level 'ilevel0'
1592 */
1593 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1594 if (dims >= 2) {
1595 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1596 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1597 ilevel0);
1598 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1599 img_stride0_vec = lp_build_get_level_stride_vec(bld,
1600 img_stride_array,
1601 ilevel0);
1602 if (dims == 3) {
1603 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1604 }
1605 }
1606 }
1607 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1608 /* compute width, height, depth for second mipmap level at 'ilevel1' */
1609 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1610 if (dims >= 2) {
1611 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1612 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1613 ilevel1);
1614 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1615 img_stride1_vec = lp_build_get_level_stride_vec(bld,
1616 img_stride_array,
1617 ilevel1);
1618 if (dims ==3) {
1619 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1620 }
1621 }
1622 }
1623 }
1624
1625 /*
1626 * Choose cube face, recompute per-face texcoords.
1627 */
1628 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1629 LLVMValueRef face, face_s, face_t;
1630 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1631 s = face_s; /* vec */
1632 t = face_t; /* vec */
1633 /* use 'r' to indicate cube face */
1634 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1635 }
1636
1637 /*
1638 * Get pointer(s) to image data for mipmap level(s).
1639 */
1640 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1641 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1642 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1643 }
1644
1645 /*
1646 * Get/interpolate texture colors.
1647 */
1648 if (min_filter == mag_filter) {
1649 /* no need to distinquish between minification and magnification */
1650 lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart,
1651 width0_vec, width1_vec,
1652 height0_vec, height1_vec,
1653 depth0_vec, depth1_vec,
1654 row_stride0_vec, row_stride1_vec,
1655 img_stride0_vec, img_stride1_vec,
1656 data_ptr0, data_ptr1,
1657 colors_out);
1658 }
1659 else {
1660 /* Emit conditional to choose min image filter or mag image filter
1661 * depending on the lod being >0 or <= 0, respectively.
1662 */
1663 struct lp_build_flow_context *flow_ctx;
1664 struct lp_build_if_state if_ctx;
1665 LLVMValueRef minify;
1666
1667 flow_ctx = lp_build_flow_create(bld->builder);
1668 lp_build_flow_scope_begin(flow_ctx);
1669
1670 lp_build_flow_scope_declare(flow_ctx, &colors_out[0]);
1671 lp_build_flow_scope_declare(flow_ctx, &colors_out[1]);
1672 lp_build_flow_scope_declare(flow_ctx, &colors_out[2]);
1673 lp_build_flow_scope_declare(flow_ctx, &colors_out[3]);
1674
1675 /* minify = lod > 0.0 */
1676 minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1677 lod, float_bld->zero, "");
1678
1679 lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
1680 {
1681 /* Use the minification filter */
1682 lp_build_sample_mipmap(bld, min_filter, mip_filter,
1683 s, t, r, lod_fpart,
1684 width0_vec, width1_vec,
1685 height0_vec, height1_vec,
1686 depth0_vec, depth1_vec,
1687 row_stride0_vec, row_stride1_vec,
1688 img_stride0_vec, img_stride1_vec,
1689 data_ptr0, data_ptr1,
1690 colors_out);
1691 }
1692 lp_build_else(&if_ctx);
1693 {
1694 /* Use the magnification filter */
1695 lp_build_sample_mipmap(bld, mag_filter, mip_filter,
1696 s, t, r, lod_fpart,
1697 width0_vec, width1_vec,
1698 height0_vec, height1_vec,
1699 depth0_vec, depth1_vec,
1700 row_stride0_vec, row_stride1_vec,
1701 img_stride0_vec, img_stride1_vec,
1702 data_ptr0, data_ptr1,
1703 colors_out);
1704 }
1705 lp_build_endif(&if_ctx);
1706
1707 lp_build_flow_scope_end(flow_ctx);
1708 lp_build_flow_destroy(flow_ctx);
1709 }
1710 }
1711
1712
1713
1714 static void
1715 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1716 struct lp_type dst_type,
1717 LLVMValueRef packed,
1718 LLVMValueRef *rgba)
1719 {
1720 LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
1721 unsigned chan;
1722
1723 /* Decode the input vector components */
1724 for (chan = 0; chan < 4; ++chan) {
1725 unsigned start = chan*8;
1726 unsigned stop = start + 8;
1727 LLVMValueRef input;
1728
1729 input = packed;
1730
1731 if(start)
1732 input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(dst_type, start), "");
1733
1734 if(stop < 32)
1735 input = LLVMBuildAnd(builder, input, mask, "");
1736
1737 input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1738
1739 rgba[chan] = input;
1740 }
1741 }
1742
1743
1744 static void
1745 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1746 LLVMValueRef s,
1747 LLVMValueRef t,
1748 LLVMValueRef width,
1749 LLVMValueRef height,
1750 LLVMValueRef stride_array,
1751 LLVMValueRef data_array,
1752 LLVMValueRef texel_out[4])
1753 {
1754 LLVMBuilderRef builder = bld->builder;
1755 struct lp_build_context i32, h16, u8n;
1756 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1757 LLVMValueRef i32_c8, i32_c128, i32_c255;
1758 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1759 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1760 LLVMValueRef x0, x1;
1761 LLVMValueRef y0, y1;
1762 LLVMValueRef neighbors[2][2];
1763 LLVMValueRef neighbors_lo[2][2];
1764 LLVMValueRef neighbors_hi[2][2];
1765 LLVMValueRef packed, packed_lo, packed_hi;
1766 LLVMValueRef unswizzled[4];
1767 LLVMValueRef stride;
1768
1769 assert(bld->static_state->target == PIPE_TEXTURE_2D);
1770 assert(bld->static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR);
1771 assert(bld->static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR);
1772 assert(bld->static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE);
1773
1774 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1775 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1776 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1777
1778 i32_vec_type = lp_build_vec_type(i32.type);
1779 h16_vec_type = lp_build_vec_type(h16.type);
1780 u8n_vec_type = lp_build_vec_type(u8n.type);
1781
1782 if (bld->static_state->normalized_coords) {
1783 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1784 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1785 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1786 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1787 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1788 }
1789
1790 /* scale coords by 256 (8 fractional bits) */
1791 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1792 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1793
1794 /* convert float to int */
1795 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1796 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1797
1798 /* subtract 0.5 (add -128) */
1799 i32_c128 = lp_build_const_int_vec(i32.type, -128);
1800 s = LLVMBuildAdd(builder, s, i32_c128, "");
1801 t = LLVMBuildAdd(builder, t, i32_c128, "");
1802
1803 /* compute floor (shift right 8) */
1804 i32_c8 = lp_build_const_int_vec(i32.type, 8);
1805 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1806 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1807
1808 /* compute fractional part (AND with 0xff) */
1809 i32_c255 = lp_build_const_int_vec(i32.type, 255);
1810 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1811 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1812
1813 x0 = s_ipart;
1814 y0 = t_ipart;
1815
1816 x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1817 y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1818
1819 x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
1820 bld->static_state->wrap_s);
1821 y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1822 bld->static_state->wrap_t);
1823
1824 x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
1825 bld->static_state->wrap_s);
1826 y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1827 bld->static_state->wrap_t);
1828
1829 /*
1830 * Transform 4 x i32 in
1831 *
1832 * s_fpart = {s0, s1, s2, s3}
1833 *
1834 * into 8 x i16
1835 *
1836 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1837 *
1838 * into two 8 x i16
1839 *
1840 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1841 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1842 *
1843 * and likewise for t_fpart. There is no risk of loosing precision here
1844 * since the fractional parts only use the lower 8bits.
1845 */
1846
1847 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1848 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1849
1850 {
1851 LLVMTypeRef elem_type = LLVMInt32Type();
1852 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1853 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1854 LLVMValueRef shuffle_lo;
1855 LLVMValueRef shuffle_hi;
1856 unsigned i, j;
1857
1858 for(j = 0; j < h16.type.length; j += 4) {
1859 unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1860 LLVMValueRef index;
1861
1862 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1863 for(i = 0; i < 4; ++i)
1864 shuffles_lo[j + i] = index;
1865
1866 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1867 for(i = 0; i < 4; ++i)
1868 shuffles_hi[j + i] = index;
1869 }
1870
1871 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1872 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1873
1874 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1875 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1876 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1877 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1878 }
1879
1880 stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1881
1882 /*
1883 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1884 *
1885 * rgba0 rgba1 rgba2 rgba3
1886 *
1887 * bit cast them into 16 x u8
1888 *
1889 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1890 *
1891 * unpack them into two 8 x i16:
1892 *
1893 * r0 g0 b0 a0 r1 g1 b1 a1
1894 * r2 g2 b2 a2 r3 g3 b3 a3
1895 *
1896 * The higher 8 bits of the resulting elements will be zero.
1897 */
1898
1899 neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1900 neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1901 neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1902 neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1903
1904 neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1905 neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1906 neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1907 neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1908
1909 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1910 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1911 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1912 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1913
1914 /*
1915 * Linear interpolate with 8.8 fixed point.
1916 */
1917
1918 packed_lo = lp_build_lerp_2d(&h16,
1919 s_fpart_lo, t_fpart_lo,
1920 neighbors_lo[0][0],
1921 neighbors_lo[0][1],
1922 neighbors_lo[1][0],
1923 neighbors_lo[1][1]);
1924
1925 packed_hi = lp_build_lerp_2d(&h16,
1926 s_fpart_hi, t_fpart_hi,
1927 neighbors_hi[0][0],
1928 neighbors_hi[0][1],
1929 neighbors_hi[1][0],
1930 neighbors_hi[1][1]);
1931
1932 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1933
1934 /*
1935 * Convert to SoA and swizzle.
1936 */
1937
1938 packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
1939
1940 lp_build_rgba8_to_f32_soa(bld->builder,
1941 bld->texel_type,
1942 packed, unswizzled);
1943
1944 lp_build_format_swizzle_soa(bld->format_desc,
1945 &bld->texel_bld,
1946 unswizzled, texel_out);
1947
1948 apply_sampler_swizzle(bld, texel_out);
1949 }
1950
1951
1952 static void
1953 lp_build_sample_compare(struct lp_build_sample_context *bld,
1954 LLVMValueRef p,
1955 LLVMValueRef texel[4])
1956 {
1957 struct lp_build_context *texel_bld = &bld->texel_bld;
1958 LLVMValueRef res;
1959 unsigned chan;
1960
1961 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1962 return;
1963
1964 /* TODO: Compare before swizzling, to avoid redundant computations */
1965 res = NULL;
1966 for(chan = 0; chan < 4; ++chan) {
1967 LLVMValueRef cmp;
1968 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1969 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1970
1971 if(res)
1972 res = lp_build_add(texel_bld, res, cmp);
1973 else
1974 res = cmp;
1975 }
1976
1977 assert(res);
1978 res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
1979
1980 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1981 for(chan = 0; chan < 3; ++chan)
1982 texel[chan] = res;
1983 texel[3] = texel_bld->one;
1984 }
1985
1986
1987 /**
1988 * Just set texels to white instead of actually sampling the texture.
1989 * For debugging.
1990 */
1991 static void
1992 lp_build_sample_nop(struct lp_build_sample_context *bld,
1993 LLVMValueRef texel_out[4])
1994 {
1995 struct lp_build_context *texel_bld = &bld->texel_bld;
1996 unsigned chan;
1997
1998 for (chan = 0; chan < 4; chan++) {
1999 /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
2000 texel_out[chan] = texel_bld->one;
2001 }
2002 }
2003
2004
2005 /**
2006 * Build texture sampling code.
2007 * 'texel' will return a vector of four LLVMValueRefs corresponding to
2008 * R, G, B, A.
2009 * \param type vector float type to use for coords, etc.
2010 */
2011 void
2012 lp_build_sample_soa(LLVMBuilderRef builder,
2013 const struct lp_sampler_static_state *static_state,
2014 struct lp_sampler_dynamic_state *dynamic_state,
2015 struct lp_type type,
2016 unsigned unit,
2017 unsigned num_coords,
2018 const LLVMValueRef *coords,
2019 const LLVMValueRef *ddx,
2020 const LLVMValueRef *ddy,
2021 LLVMValueRef lod_bias, /* optional */
2022 LLVMValueRef explicit_lod, /* optional */
2023 LLVMValueRef texel_out[4])
2024 {
2025 struct lp_build_sample_context bld;
2026 LLVMValueRef width, width_vec;
2027 LLVMValueRef height, height_vec;
2028 LLVMValueRef depth, depth_vec;
2029 LLVMValueRef row_stride_array, img_stride_array;
2030 LLVMValueRef data_array;
2031 LLVMValueRef s;
2032 LLVMValueRef t;
2033 LLVMValueRef r;
2034
2035 if (0) {
2036 enum pipe_format fmt = static_state->format;
2037 debug_printf("Sample from %s\n", util_format_name(fmt));
2038 }
2039
2040 /* Setup our build context */
2041 memset(&bld, 0, sizeof bld);
2042 bld.builder = builder;
2043 bld.static_state = static_state;
2044 bld.dynamic_state = dynamic_state;
2045 bld.format_desc = util_format_description(static_state->format);
2046
2047 bld.float_type = lp_type_float(32);
2048 bld.int_type = lp_type_int(32);
2049 bld.coord_type = type;
2050 bld.uint_coord_type = lp_uint_type(type);
2051 bld.int_coord_type = lp_int_type(type);
2052 bld.texel_type = type;
2053
2054 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
2055 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
2056 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
2057 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
2058 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
2059 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
2060
2061 /* Get the dynamic state */
2062 width = dynamic_state->width(dynamic_state, builder, unit);
2063 height = dynamic_state->height(dynamic_state, builder, unit);
2064 depth = dynamic_state->depth(dynamic_state, builder, unit);
2065 row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
2066 img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
2067 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
2068 /* Note that data_array is an array[level] of pointers to texture images */
2069
2070 s = coords[0];
2071 t = coords[1];
2072 r = coords[2];
2073
2074 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
2075 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
2076 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
2077
2078 if (0) {
2079 /* For debug: no-op texture sampling */
2080 lp_build_sample_nop(&bld, texel_out);
2081 }
2082 else if (util_format_is_rgba8_variant(bld.format_desc) &&
2083 static_state->target == PIPE_TEXTURE_2D &&
2084 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
2085 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
2086 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
2087 is_simple_wrap_mode(static_state->wrap_s) &&
2088 is_simple_wrap_mode(static_state->wrap_t)) {
2089 /* special case */
2090 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
2091 row_stride_array, data_array, texel_out);
2092 }
2093 else {
2094 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
2095 lod_bias, explicit_lod,
2096 width, height, depth,
2097 width_vec, height_vec, depth_vec,
2098 row_stride_array, img_stride_array,
2099 data_array,
2100 texel_out);
2101 }
2102
2103 lp_build_sample_compare(&bld, r, texel_out);
2104 }