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