projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm/llvmpipe: replace 'int stride' with 'int row_stride[MAX_LEVELS]'
[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 */
34
35 #include "pipe/p_defines.h"
36 #include "pipe/p_state.h"
37 #include "util/u_debug.h"
38 #include "util/u_dump.h"
39 #include "util/u_memory.h"
40 #include "util/u_math.h"
41 #include "util/u_format.h"
42 #include "util/u_cpu_detect.h"
43 #include "lp_bld_debug.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_arit.h"
48 #include "lp_bld_logic.h"
49 #include "lp_bld_swizzle.h"
50 #include "lp_bld_pack.h"
51 #include "lp_bld_format.h"
52 #include "lp_bld_sample.h"
53
54
55 /**
56 * Keep all information for sampling code generation in a single place.
57 */
58 struct lp_build_sample_context
59 {
60 LLVMBuilderRef builder;
61
62 const struct lp_sampler_static_state *static_state;
63
64 struct lp_sampler_dynamic_state *dynamic_state;
65
66 const struct util_format_description *format_desc;
67
68 /** regular scalar float type */
69 struct lp_type float_type;
70 struct lp_build_context float_bld;
71
72 /** regular scalar float type */
73 struct lp_type int_type;
74 struct lp_build_context int_bld;
75
76 /** Incoming coordinates type and build context */
77 struct lp_type coord_type;
78 struct lp_build_context coord_bld;
79
80 /** Unsigned integer coordinates */
81 struct lp_type uint_coord_type;
82 struct lp_build_context uint_coord_bld;
83
84 /** Signed integer coordinates */
85 struct lp_type int_coord_type;
86 struct lp_build_context int_coord_bld;
87
88 /** Output texels type and build context */
89 struct lp_type texel_type;
90 struct lp_build_context texel_bld;
91 };
92
93
94 /**
95 * Does the given texture wrap mode allow sampling the texture border color?
96 * XXX maybe move this into gallium util code.
97 */
98 static boolean
99 wrap_mode_uses_border_color(unsigned mode)
100 {
101 switch (mode) {
102 case PIPE_TEX_WRAP_REPEAT:
103 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
104 case PIPE_TEX_WRAP_MIRROR_REPEAT:
105 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
106 return FALSE;
107 case PIPE_TEX_WRAP_CLAMP:
108 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
109 case PIPE_TEX_WRAP_MIRROR_CLAMP:
110 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
111 return TRUE;
112 default:
113 assert(0 && "unexpected wrap mode");
114 return FALSE;
115 }
116 }
117
118
119 static LLVMValueRef
120 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
121 LLVMValueRef data_array, LLVMValueRef level)
122 {
123 LLVMValueRef indexes[2], data_ptr;
124 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
125 indexes[1] = level;
126 data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
127 data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
128 return data_ptr;
129 }
130
131
132 static LLVMValueRef
133 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
134 LLVMValueRef data_array, int level)
135 {
136 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
137 return lp_build_get_mipmap_level(bld, data_array, lvl);
138 }
139
140
141 /**
142 * Dereference stride_array[mipmap_level] array to get a stride.
143 * Return stride as a vector.
144 */
145 static LLVMValueRef
146 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
147 LLVMValueRef stride_array, LLVMValueRef level)
148 {
149 LLVMValueRef indexes[2], stride;
150 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
151 indexes[1] = level;
152 stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
153 stride = LLVMBuildLoad(bld->builder, stride, "");
154 stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
155 return stride;
156 }
157
158
159 /** Dereference stride_array[0] array to get a stride (as vector). */
160 static LLVMValueRef
161 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
162 LLVMValueRef stride_array, int level)
163 {
164 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
165 return lp_build_get_level_stride_vec(bld, stride_array, lvl);
166 }
167
168
169 static int
170 texture_dims(enum pipe_texture_target tex)
171 {
172 switch (tex) {
173 case PIPE_TEXTURE_1D:
174 return 1;
175 case PIPE_TEXTURE_2D:
176 case PIPE_TEXTURE_CUBE:
177 return 2;
178 case PIPE_TEXTURE_3D:
179 return 3;
180 default:
181 assert(0 && "bad texture target in texture_dims()");
182 return 2;
183 }
184 }
185
186
187
188 /**
189 * Generate code to fetch a texel from a texture at int coords (x, y, z).
190 * The computation depends on whether the texture is 1D, 2D or 3D.
191 * The result, texel, will be:
192 * texel[0] = red values
193 * texel[1] = green values
194 * texel[2] = blue values
195 * texel[3] = alpha values
196 */
197 static void
198 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
199 LLVMValueRef width,
200 LLVMValueRef height,
201 LLVMValueRef depth,
202 LLVMValueRef x,
203 LLVMValueRef y,
204 LLVMValueRef z,
205 LLVMValueRef y_stride,
206 LLVMValueRef z_stride,
207 LLVMValueRef data_ptr,
208 LLVMValueRef *texel)
209 {
210 const int dims = texture_dims(bld->static_state->target);
211 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
212 LLVMValueRef offset;
213 LLVMValueRef packed;
214 LLVMValueRef use_border = NULL;
215
216 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
217 if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
218 LLVMValueRef b1, b2;
219 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
220 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
221 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
222 }
223
224 if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
225 LLVMValueRef b1, b2;
226 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
227 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
228 if (use_border) {
229 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
230 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
231 }
232 else {
233 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
234 }
235 }
236
237 if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
238 LLVMValueRef b1, b2;
239 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
240 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
241 if (use_border) {
242 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
243 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
244 }
245 else {
246 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
247 }
248 }
249
250 /*
251 * Note: if we find an app which frequently samples the texture border
252 * we might want to implement a true conditional here to avoid sampling
253 * the texture whenever possible (since that's quite a bit of code).
254 * Ex:
255 * if (use_border) {
256 * texel = border_color;
257 * }
258 * else {
259 * texel = sample_texture(coord);
260 * }
261 * As it is now, we always sample the texture, then selectively replace
262 * the texel color results with the border color.
263 */
264
265 /* convert x,y,z coords to linear offset from start of texture, in bytes */
266 offset = lp_build_sample_offset(&bld->uint_coord_bld,
267 bld->format_desc,
268 x, y, z, y_stride, z_stride);
269
270 assert(bld->format_desc->block.width == 1);
271 assert(bld->format_desc->block.height == 1);
272 assert(bld->format_desc->block.bits <= bld->texel_type.width);
273
274 /* gather the texels from the texture */
275 packed = lp_build_gather(bld->builder,
276 bld->texel_type.length,
277 bld->format_desc->block.bits,
278 bld->texel_type.width,
279 data_ptr, offset);
280
281 texel[0] = texel[1] = texel[2] = texel[3] = NULL;
282
283 /* convert texels to float rgba */
284 lp_build_unpack_rgba_soa(bld->builder,
285 bld->format_desc,
286 bld->texel_type,
287 packed, texel);
288
289 if (use_border) {
290 /* select texel color or border color depending on use_border */
291 int chan;
292 for (chan = 0; chan < 4; chan++) {
293 LLVMValueRef border_chan =
294 lp_build_const_scalar(bld->texel_type,
295 bld->static_state->border_color[chan]);
296 texel[chan] = lp_build_select(&bld->texel_bld, use_border,
297 border_chan, texel[chan]);
298 }
299 }
300 }
301
302
303 static LLVMValueRef
304 lp_build_sample_packed(struct lp_build_sample_context *bld,
305 LLVMValueRef x,
306 LLVMValueRef y,
307 LLVMValueRef y_stride,
308 LLVMValueRef data_array)
309 {
310 LLVMValueRef offset;
311 LLVMValueRef data_ptr;
312
313 offset = lp_build_sample_offset(&bld->uint_coord_bld,
314 bld->format_desc,
315 x, y, NULL, y_stride, NULL);
316
317 assert(bld->format_desc->block.width == 1);
318 assert(bld->format_desc->block.height == 1);
319 assert(bld->format_desc->block.bits <= bld->texel_type.width);
320
321 /* get pointer to mipmap level 0 data */
322 data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
323
324 return lp_build_gather(bld->builder,
325 bld->texel_type.length,
326 bld->format_desc->block.bits,
327 bld->texel_type.width,
328 data_ptr, offset);
329 }
330
331
332 /**
333 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
334 */
335 static LLVMValueRef
336 lp_build_coord_mirror(struct lp_build_sample_context *bld,
337 LLVMValueRef coord)
338 {
339 struct lp_build_context *coord_bld = &bld->coord_bld;
340 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
341 LLVMValueRef fract, flr, isOdd;
342
343 /* fract = coord - floor(coord) */
344 fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
345
346 /* flr = ifloor(coord); */
347 flr = lp_build_ifloor(coord_bld, coord);
348
349 /* isOdd = flr & 1 */
350 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
351
352 /* make coord positive or negative depending on isOdd */
353 coord = lp_build_set_sign(coord_bld, fract, isOdd);
354
355 /* convert isOdd to float */
356 isOdd = lp_build_int_to_float(coord_bld, isOdd);
357
358 /* add isOdd to coord */
359 coord = lp_build_add(coord_bld, coord, isOdd);
360
361 return coord;
362 }
363
364
365 /**
366 * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
367 * Return whether the given mode is supported by that function.
368 */
369 static boolean
370 is_simple_wrap_mode(unsigned mode)
371 {
372 switch (mode) {
373 case PIPE_TEX_WRAP_REPEAT:
374 case PIPE_TEX_WRAP_CLAMP:
375 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
376 return TRUE;
377 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
378 default:
379 return FALSE;
380 }
381 }
382
383
384 /**
385 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
386 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
387 * \param length the texture size along one dimension
388 * \param is_pot if TRUE, length is a power of two
389 * \param wrap_mode one of PIPE_TEX_WRAP_x
390 */
391 static LLVMValueRef
392 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
393 LLVMValueRef coord,
394 LLVMValueRef length,
395 boolean is_pot,
396 unsigned wrap_mode)
397 {
398 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
399 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
400 LLVMValueRef length_minus_one;
401
402 length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
403
404 switch(wrap_mode) {
405 case PIPE_TEX_WRAP_REPEAT:
406 if(is_pot)
407 coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
408 else
409 /* Signed remainder won't give the right results for negative
410 * dividends but unsigned remainder does.*/
411 coord = LLVMBuildURem(bld->builder, coord, length, "");
412 break;
413
414 case PIPE_TEX_WRAP_CLAMP:
415 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
416 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
417 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
418 coord = lp_build_min(int_coord_bld, coord, length_minus_one);
419 break;
420
421 case PIPE_TEX_WRAP_MIRROR_REPEAT:
422 case PIPE_TEX_WRAP_MIRROR_CLAMP:
423 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
424 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
425 /* FIXME */
426 _debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
427 util_dump_tex_wrap(wrap_mode, TRUE));
428 coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
429 coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
430 break;
431
432 default:
433 assert(0);
434 }
435
436 return coord;
437 }
438
439
440 /**
441 * Build LLVM code for texture wrap mode for linear filtering.
442 * \param x0_out returns first integer texcoord
443 * \param x1_out returns second integer texcoord
444 * \param weight_out returns linear interpolation weight
445 */
446 static void
447 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
448 LLVMValueRef coord,
449 LLVMValueRef length,
450 boolean is_pot,
451 unsigned wrap_mode,
452 LLVMValueRef *x0_out,
453 LLVMValueRef *x1_out,
454 LLVMValueRef *weight_out)
455 {
456 struct lp_build_context *coord_bld = &bld->coord_bld;
457 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
458 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
459 LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
460 LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5);
461 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
462 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
463 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
464 LLVMValueRef coord0, coord1, weight;
465
466 switch(wrap_mode) {
467 case PIPE_TEX_WRAP_REPEAT:
468 /* mul by size and subtract 0.5 */
469 coord = lp_build_mul(coord_bld, coord, length_f);
470 coord = lp_build_sub(coord_bld, coord, half);
471 /* convert to int */
472 coord0 = lp_build_ifloor(coord_bld, coord);
473 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
474 /* compute lerp weight */
475 weight = lp_build_fract(coord_bld, coord);
476 /* repeat wrap */
477 if (is_pot) {
478 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
479 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
480 }
481 else {
482 /* Signed remainder won't give the right results for negative
483 * dividends but unsigned remainder does.*/
484 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
485 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
486 }
487 break;
488
489 case PIPE_TEX_WRAP_CLAMP:
490 if (bld->static_state->normalized_coords) {
491 coord = lp_build_mul(coord_bld, coord, length_f);
492 }
493 weight = lp_build_fract(coord_bld, coord);
494 coord0 = lp_build_clamp(coord_bld, coord, coord_bld->zero,
495 length_f_minus_one);
496 coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
497 coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
498 length_f_minus_one);
499 coord0 = lp_build_ifloor(coord_bld, coord0);
500 coord1 = lp_build_ifloor(coord_bld, coord1);
501 break;
502
503 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
504 if (bld->static_state->normalized_coords) {
505 /* clamp to [0,1] */
506 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
507 /* mul by tex size and subtract 0.5 */
508 coord = lp_build_mul(coord_bld, coord, length_f);
509 coord = lp_build_sub(coord_bld, coord, half);
510 }
511 else {
512 LLVMValueRef min, max;
513 /* clamp to [0.5, length - 0.5] */
514 min = lp_build_const_scalar(coord_bld->type, 0.5F);
515 max = lp_build_sub(coord_bld, length_f, min);
516 coord = lp_build_clamp(coord_bld, coord, min, max);
517 }
518 /* compute lerp weight */
519 weight = lp_build_fract(coord_bld, coord);
520 /* coord0 = floor(coord); */
521 coord0 = lp_build_ifloor(coord_bld, coord);
522 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
523 /* coord0 = max(coord0, 0) */
524 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
525 /* coord1 = min(coord1, length-1) */
526 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
527 break;
528
529 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
530 {
531 LLVMValueRef min, max;
532 if (bld->static_state->normalized_coords) {
533 /* min = -1.0 / (2 * length) = -0.5 / length */
534 min = lp_build_mul(coord_bld,
535 lp_build_const_scalar(coord_bld->type, -0.5F),
536 lp_build_rcp(coord_bld, length_f));
537 /* max = 1.0 - min */
538 max = lp_build_sub(coord_bld, coord_bld->one, min);
539 /* coord = clamp(coord, min, max) */
540 coord = lp_build_clamp(coord_bld, coord, min, max);
541 /* scale coord to length (and sub 0.5?) */
542 coord = lp_build_mul(coord_bld, coord, length_f);
543 coord = lp_build_sub(coord_bld, coord, half);
544 }
545 else {
546 /* clamp to [-0.5, length + 0.5] */
547 min = lp_build_const_scalar(coord_bld->type, -0.5F);
548 max = lp_build_sub(coord_bld, length_f, min);
549 coord = lp_build_clamp(coord_bld, coord, min, max);
550 coord = lp_build_sub(coord_bld, coord, half);
551 }
552 /* compute lerp weight */
553 weight = lp_build_fract(coord_bld, coord);
554 /* convert to int */
555 coord0 = lp_build_ifloor(coord_bld, coord);
556 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
557 }
558 break;
559
560 case PIPE_TEX_WRAP_MIRROR_REPEAT:
561 /* compute mirror function */
562 coord = lp_build_coord_mirror(bld, coord);
563
564 /* scale coord to length */
565 coord = lp_build_mul(coord_bld, coord, length_f);
566 coord = lp_build_sub(coord_bld, coord, half);
567
568 /* compute lerp weight */
569 weight = lp_build_fract(coord_bld, coord);
570
571 /* convert to int coords */
572 coord0 = lp_build_ifloor(coord_bld, coord);
573 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
574
575 /* coord0 = max(coord0, 0) */
576 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
577 /* coord1 = min(coord1, length-1) */
578 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
579 break;
580
581 case PIPE_TEX_WRAP_MIRROR_CLAMP:
582 {
583 LLVMValueRef min, max;
584 /* min = 1.0 / (2 * length) */
585 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
586 /* max = 1.0 - min */
587 max = lp_build_sub(coord_bld, coord_bld->one, min);
588
589 coord = lp_build_abs(coord_bld, coord);
590 coord = lp_build_clamp(coord_bld, coord, min, max);
591 coord = lp_build_mul(coord_bld, coord, length_f);
592 if(0)coord = lp_build_sub(coord_bld, coord, half);
593 weight = lp_build_fract(coord_bld, coord);
594 coord0 = lp_build_ifloor(coord_bld, coord);
595 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
596 }
597 break;
598
599 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
600 {
601 LLVMValueRef min, max;
602 /* min = 1.0 / (2 * length) */
603 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
604 /* max = 1.0 - min */
605 max = lp_build_sub(coord_bld, coord_bld->one, min);
606
607 coord = lp_build_abs(coord_bld, coord);
608 coord = lp_build_clamp(coord_bld, coord, min, max);
609 coord = lp_build_mul(coord_bld, coord, length_f);
610 coord = lp_build_sub(coord_bld, coord, half);
611 weight = lp_build_fract(coord_bld, coord);
612 coord0 = lp_build_ifloor(coord_bld, coord);
613 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
614 }
615 break;
616
617 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
618 {
619 LLVMValueRef min, max;
620 /* min = -1.0 / (2 * length) = -0.5 / length */
621 min = lp_build_mul(coord_bld,
622 lp_build_const_scalar(coord_bld->type, -0.5F),
623 lp_build_rcp(coord_bld, length_f));
624 /* max = 1.0 - min */
625 max = lp_build_sub(coord_bld, coord_bld->one, min);
626
627 coord = lp_build_abs(coord_bld, coord);
628 coord = lp_build_clamp(coord_bld, coord, min, max);
629 coord = lp_build_mul(coord_bld, coord, length_f);
630 coord = lp_build_sub(coord_bld, coord, half);
631 weight = lp_build_fract(coord_bld, coord);
632 coord0 = lp_build_ifloor(coord_bld, coord);
633 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
634 }
635 break;
636
637 default:
638 assert(0);
639 coord0 = NULL;
640 coord1 = NULL;
641 weight = NULL;
642 }
643
644 *x0_out = coord0;
645 *x1_out = coord1;
646 *weight_out = weight;
647 }
648
649
650 /**
651 * Build LLVM code for texture wrap mode for nearest filtering.
652 * \param coord the incoming texcoord (nominally in [0,1])
653 * \param length the texture size along one dimension, as int
654 * \param is_pot if TRUE, length is a power of two
655 * \param wrap_mode one of PIPE_TEX_WRAP_x
656 */
657 static LLVMValueRef
658 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
659 LLVMValueRef coord,
660 LLVMValueRef length,
661 boolean is_pot,
662 unsigned wrap_mode)
663 {
664 struct lp_build_context *coord_bld = &bld->coord_bld;
665 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
666 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
667 LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
668 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
669 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
670 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
671 LLVMValueRef icoord;
672
673 switch(wrap_mode) {
674 case PIPE_TEX_WRAP_REPEAT:
675 coord = lp_build_mul(coord_bld, coord, length_f);
676 icoord = lp_build_ifloor(coord_bld, coord);
677 if (is_pot)
678 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
679 else
680 /* Signed remainder won't give the right results for negative
681 * dividends but unsigned remainder does.*/
682 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
683 break;
684
685 case PIPE_TEX_WRAP_CLAMP:
686 /* mul by size */
687 if (bld->static_state->normalized_coords) {
688 coord = lp_build_mul(coord_bld, coord, length_f);
689 }
690 /* floor */
691 icoord = lp_build_ifloor(coord_bld, coord);
692 /* clamp to [0, size-1]. Note: int coord builder type */
693 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
694 length_minus_one);
695 break;
696
697 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
698 {
699 LLVMValueRef min, max;
700 if (bld->static_state->normalized_coords) {
701 /* min = 1.0 / (2 * length) */
702 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
703 /* max = length - min */
704 max = lp_build_sub(coord_bld, length_f, min);
705 /* scale coord to length */
706 coord = lp_build_mul(coord_bld, coord, length_f);
707 }
708 else {
709 /* clamp to [0.5, length - 0.5] */
710 min = lp_build_const_scalar(coord_bld->type, 0.5F);
711 max = lp_build_sub(coord_bld, length_f, min);
712 }
713 /* coord = clamp(coord, min, max) */
714 coord = lp_build_clamp(coord_bld, coord, min, max);
715 icoord = lp_build_ifloor(coord_bld, coord);
716 }
717 break;
718
719 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
720 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
721 {
722 LLVMValueRef min, max;
723 if (bld->static_state->normalized_coords) {
724 /* min = -1.0 / (2 * length) = -0.5 / length */
725 min = lp_build_mul(coord_bld,
726 lp_build_const_scalar(coord_bld->type, -0.5F),
727 lp_build_rcp(coord_bld, length_f));
728 /* max = length - min */
729 max = lp_build_sub(coord_bld, length_f, min);
730 /* scale coord to length */
731 coord = lp_build_mul(coord_bld, coord, length_f);
732 }
733 else {
734 /* clamp to [-0.5, length + 0.5] */
735 min = lp_build_const_scalar(coord_bld->type, -0.5F);
736 max = lp_build_sub(coord_bld, length_f, min);
737 }
738 /* coord = clamp(coord, min, max) */
739 coord = lp_build_clamp(coord_bld, coord, min, max);
740 icoord = lp_build_ifloor(coord_bld, coord);
741 }
742 break;
743
744 case PIPE_TEX_WRAP_MIRROR_REPEAT:
745 {
746 LLVMValueRef min, max;
747 /* min = 1.0 / (2 * length) */
748 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
749 /* max = length - min */
750 max = lp_build_sub(coord_bld, length_f, min);
751
752 /* compute mirror function */
753 coord = lp_build_coord_mirror(bld, coord);
754
755 /* scale coord to length */
756 coord = lp_build_mul(coord_bld, coord, length_f);
757
758 /* coord = clamp(coord, min, max) */
759 coord = lp_build_clamp(coord_bld, coord, min, max);
760 icoord = lp_build_ifloor(coord_bld, coord);
761 }
762 break;
763
764 case PIPE_TEX_WRAP_MIRROR_CLAMP:
765 coord = lp_build_abs(coord_bld, coord);
766 coord = lp_build_mul(coord_bld, coord, length_f);
767 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
768 icoord = lp_build_ifloor(coord_bld, coord);
769 break;
770
771 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
772 {
773 LLVMValueRef min, max;
774 /* min = 1.0 / (2 * length) */
775 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
776 /* max = length - min */
777 max = lp_build_sub(coord_bld, length_f, min);
778
779 coord = lp_build_abs(coord_bld, coord);
780 coord = lp_build_mul(coord_bld, coord, length_f);
781 coord = lp_build_clamp(coord_bld, coord, min, max);
782 icoord = lp_build_ifloor(coord_bld, coord);
783 }
784 break;
785
786 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
787 {
788 LLVMValueRef min, max;
789 /* min = 1.0 / (2 * length) */
790 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
791 min = lp_build_negate(coord_bld, min);
792 /* max = length - min */
793 max = lp_build_sub(coord_bld, length_f, min);
794
795 coord = lp_build_abs(coord_bld, coord);
796 coord = lp_build_mul(coord_bld, coord, length_f);
797 coord = lp_build_clamp(coord_bld, coord, min, max);
798 icoord = lp_build_ifloor(coord_bld, coord);
799 }
800 break;
801
802 default:
803 assert(0);
804 icoord = NULL;
805 }
806
807 return icoord;
808 }
809
810
811 /**
812 * Codegen equivalent for u_minify().
813 * Return max(1, base_size >> level);
814 */
815 static LLVMValueRef
816 lp_build_minify(struct lp_build_sample_context *bld,
817 LLVMValueRef base_size,
818 LLVMValueRef level)
819 {
820 LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
821 size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
822 return size;
823 }
824
825
826 /**
827 * Generate code to compute texture level of detail (lambda).
828 * \param s vector of texcoord s values
829 * \param t vector of texcoord t values
830 * \param r vector of texcoord r values
831 * \param width scalar int texture width
832 * \param height scalar int texture height
833 * \param depth scalar int texture depth
834 */
835 static LLVMValueRef
836 lp_build_lod_selector(struct lp_build_sample_context *bld,
837 LLVMValueRef s,
838 LLVMValueRef t,
839 LLVMValueRef r,
840 LLVMValueRef width,
841 LLVMValueRef height,
842 LLVMValueRef depth)
843
844 {
845 const int dims = texture_dims(bld->static_state->target);
846 struct lp_build_context *float_bld = &bld->float_bld;
847 LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(), bld->static_state->lod_bias);
848 LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
849 LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(), bld->static_state->max_lod);
850
851 LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
852 LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
853 LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
854
855 LLVMValueRef s0, s1, s2;
856 LLVMValueRef t0, t1, t2;
857 LLVMValueRef r0, r1, r2;
858 LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
859 LLVMValueRef rho, lod;
860
861 /*
862 * dsdx = abs(s[1] - s[0]);
863 * dsdy = abs(s[2] - s[0]);
864 * dtdx = abs(t[1] - t[0]);
865 * dtdy = abs(t[2] - t[0]);
866 * drdx = abs(r[1] - r[0]);
867 * drdy = abs(r[2] - r[0]);
868 * XXX we're assuming a four-element quad in 2x2 layout here.
869 */
870 s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0");
871 s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1");
872 s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2");
873 dsdx = LLVMBuildSub(bld->builder, s1, s0, "");
874 dsdx = lp_build_abs(float_bld, dsdx);
875 dsdy = LLVMBuildSub(bld->builder, s2, s0, "");
876 dsdy = lp_build_abs(float_bld, dsdy);
877 if (dims > 1) {
878 t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0");
879 t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1");
880 t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2");
881 dtdx = LLVMBuildSub(bld->builder, t1, t0, "");
882 dtdx = lp_build_abs(float_bld, dtdx);
883 dtdy = LLVMBuildSub(bld->builder, t2, t0, "");
884 dtdy = lp_build_abs(float_bld, dtdy);
885 if (dims > 2) {
886 r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0");
887 r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1");
888 r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2");
889 drdx = LLVMBuildSub(bld->builder, r1, r0, "");
890 drdx = lp_build_abs(float_bld, drdx);
891 drdy = LLVMBuildSub(bld->builder, r2, r0, "");
892 drdy = lp_build_abs(float_bld, drdy);
893 }
894 }
895
896 /* Compute rho = max of all partial derivatives scaled by texture size.
897 * XXX this could be vectorized somewhat
898 */
899 rho = LLVMBuildMul(bld->builder,
900 lp_build_max(float_bld, dsdx, dsdy),
901 lp_build_int_to_float(float_bld, width), "");
902 if (dims > 1) {
903 LLVMValueRef max;
904 max = LLVMBuildMul(bld->builder,
905 lp_build_max(float_bld, dtdx, dtdy),
906 lp_build_int_to_float(float_bld, height), "");
907 rho = lp_build_max(float_bld, rho, max);
908 if (dims > 2) {
909 max = LLVMBuildMul(bld->builder,
910 lp_build_max(float_bld, drdx, drdy),
911 lp_build_int_to_float(float_bld, depth), "");
912 rho = lp_build_max(float_bld, rho, max);
913 }
914 }
915
916 /* compute lod = log2(rho) */
917 lod = lp_build_log2(float_bld, rho);
918
919 /* add lod bias */
920 lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "LOD bias");
921
922 /* clamp lod */
923 lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
924
925 return lod;
926 }
927
928
929 /**
930 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
931 * mipmap level index.
932 * Note: this is all scalar code.
933 * \param lod scalar float texture level of detail
934 * \param level_out returns integer
935 */
936 static void
937 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
938 unsigned unit,
939 LLVMValueRef lod,
940 LLVMValueRef *level_out)
941 {
942 struct lp_build_context *float_bld = &bld->float_bld;
943 struct lp_build_context *int_bld = &bld->int_bld;
944 LLVMValueRef last_level, level;
945
946 LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
947
948 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
949 bld->builder, unit);
950
951 /* convert float lod to integer */
952 level = lp_build_iround(float_bld, lod);
953
954 /* clamp level to legal range of levels */
955 *level_out = lp_build_clamp(int_bld, level, zero, last_level);
956 }
957
958
959 /**
960 * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
961 * two (adjacent) mipmap level indexes. Later, we'll sample from those
962 * two mipmap levels and interpolate between them.
963 */
964 static void
965 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
966 unsigned unit,
967 LLVMValueRef lod,
968 LLVMValueRef *level0_out,
969 LLVMValueRef *level1_out,
970 LLVMValueRef *weight_out)
971 {
972 struct lp_build_context *float_bld = &bld->float_bld;
973 struct lp_build_context *int_bld = &bld->int_bld;
974 LLVMValueRef last_level, level;
975
976 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
977 bld->builder, unit);
978
979 /* convert float lod to integer */
980 level = lp_build_ifloor(float_bld, lod);
981
982 /* compute level 0 and clamp to legal range of levels */
983 *level0_out = lp_build_clamp(int_bld, level,
984 int_bld->zero,
985 last_level);
986 /* compute level 1 and clamp to legal range of levels */
987 *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
988 *level1_out = lp_build_min(int_bld, *level1_out, int_bld->zero);
989
990 *weight_out = lp_build_fract(float_bld, lod);
991 }
992
993
994 /**
995 * Generate code to sample a mipmap level with nearest filtering.
996 */
997 static void
998 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
999 LLVMValueRef width_vec,
1000 LLVMValueRef height_vec,
1001 LLVMValueRef depth_vec,
1002 LLVMValueRef row_stride_vec,
1003 LLVMValueRef img_stride_vec,
1004 LLVMValueRef data_ptr,
1005 LLVMValueRef s,
1006 LLVMValueRef t,
1007 LLVMValueRef r,
1008 LLVMValueRef colors_out[4])
1009 {
1010 const int dims = texture_dims(bld->static_state->target);
1011 LLVMValueRef x, y, z;
1012
1013 /*
1014 * Compute integer texcoords.
1015 */
1016 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1017 bld->static_state->pot_width,
1018 bld->static_state->wrap_s);
1019 lp_build_name(x, "tex.x.wrapped");
1020
1021 if (dims >= 2) {
1022 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1023 bld->static_state->pot_height,
1024 bld->static_state->wrap_t);
1025 lp_build_name(y, "tex.y.wrapped");
1026
1027 if (dims == 3) {
1028 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1029 bld->static_state->pot_height,
1030 bld->static_state->wrap_r);
1031 lp_build_name(z, "tex.z.wrapped");
1032 }
1033 else {
1034 z = NULL;
1035 }
1036 }
1037 else {
1038 y = z = NULL;
1039 }
1040
1041 /*
1042 * Get texture colors.
1043 */
1044 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1045 x, y, z,
1046 row_stride_vec, img_stride_vec,
1047 data_ptr, colors_out);
1048 }
1049
1050
1051 /**
1052 * Generate code to sample a mipmap level with linear filtering.
1053 * 1D, 2D and 3D images are suppored.
1054 */
1055 static void
1056 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1057 LLVMValueRef width_vec,
1058 LLVMValueRef height_vec,
1059 LLVMValueRef depth_vec,
1060 LLVMValueRef row_stride_vec,
1061 LLVMValueRef img_stride_vec,
1062 LLVMValueRef data_ptr,
1063 LLVMValueRef s,
1064 LLVMValueRef t,
1065 LLVMValueRef r,
1066 LLVMValueRef colors_out[4])
1067 {
1068 const int dims = texture_dims(bld->static_state->target);
1069 LLVMValueRef x0, y0, z0, x1, y1, z1;
1070 LLVMValueRef s_fpart, t_fpart, r_fpart;
1071 LLVMValueRef neighbors[2][2][4];
1072 int chan;
1073
1074 /*
1075 * Compute integer texcoords.
1076 */
1077 lp_build_sample_wrap_linear(bld, s, width_vec,
1078 bld->static_state->pot_width,
1079 bld->static_state->wrap_s,
1080 &x0, &x1, &s_fpart);
1081 lp_build_name(x0, "tex.x0.wrapped");
1082 lp_build_name(x1, "tex.x1.wrapped");
1083
1084 if (dims >= 2) {
1085 lp_build_sample_wrap_linear(bld, t, height_vec,
1086 bld->static_state->pot_height,
1087 bld->static_state->wrap_t,
1088 &y0, &y1, &t_fpart);
1089 lp_build_name(y0, "tex.y0.wrapped");
1090 lp_build_name(y1, "tex.y1.wrapped");
1091
1092 if (dims == 3) {
1093 lp_build_sample_wrap_linear(bld, r, depth_vec,
1094 bld->static_state->pot_depth,
1095 bld->static_state->wrap_r,
1096 &z0, &z1, &r_fpart);
1097 lp_build_name(z0, "tex.z0.wrapped");
1098 lp_build_name(z1, "tex.z1.wrapped");
1099 }
1100 else {
1101 z0 = z1 = r_fpart = NULL;
1102 }
1103 }
1104 else {
1105 y0 = y1 = t_fpart = NULL;
1106 z0 = z1 = r_fpart = NULL;
1107 }
1108
1109 /*
1110 * Get texture colors.
1111 */
1112 /* get x0/x1 texels */
1113 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1114 x0, y0, z0,
1115 row_stride_vec, img_stride_vec,
1116 data_ptr, neighbors[0][0]);
1117 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1118 x1, y0, z0,
1119 row_stride_vec, img_stride_vec,
1120 data_ptr, neighbors[0][1]);
1121
1122 if (dims == 1) {
1123 /* Interpolate two samples from 1D image to produce one color */
1124 for (chan = 0; chan < 4; chan++) {
1125 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1126 neighbors[0][0][chan],
1127 neighbors[0][1][chan]);
1128 }
1129 }
1130 else {
1131 /* 2D/3D texture */
1132 LLVMValueRef colors0[4];
1133
1134 /* get x0/x1 texels at y1 */
1135 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1136 x0, y1, z0,
1137 row_stride_vec, img_stride_vec,
1138 data_ptr, neighbors[1][0]);
1139 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1140 x1, y1, z0,
1141 row_stride_vec, img_stride_vec,
1142 data_ptr, neighbors[1][1]);
1143
1144 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1145 for (chan = 0; chan < 4; chan++) {
1146 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1147 s_fpart, t_fpart,
1148 neighbors[0][0][chan],
1149 neighbors[0][1][chan],
1150 neighbors[1][0][chan],
1151 neighbors[1][1][chan]);
1152 }
1153
1154 if (dims == 3) {
1155 LLVMValueRef neighbors1[2][2][4];
1156 LLVMValueRef colors1[4];
1157
1158 /* get x0/x1/y0/y1 texels at z1 */
1159 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1160 x0, y0, z1,
1161 row_stride_vec, img_stride_vec,
1162 data_ptr, neighbors1[0][0]);
1163 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1164 x1, y0, z1,
1165 row_stride_vec, img_stride_vec,
1166 data_ptr, neighbors1[0][1]);
1167 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1168 x0, y1, z1,
1169 row_stride_vec, img_stride_vec,
1170 data_ptr, neighbors1[1][0]);
1171 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1172 x1, y1, z1,
1173 row_stride_vec, img_stride_vec,
1174 data_ptr, neighbors1[1][1]);
1175
1176 /* Bilinear interpolate the four samples from the second Z slice */
1177 for (chan = 0; chan < 4; chan++) {
1178 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1179 s_fpart, t_fpart,
1180 neighbors1[0][0][chan],
1181 neighbors1[0][1][chan],
1182 neighbors1[1][0][chan],
1183 neighbors1[1][1][chan]);
1184 }
1185
1186 /* Linearly interpolate the two samples from the two 3D slices */
1187 for (chan = 0; chan < 4; chan++) {
1188 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1189 r_fpart,
1190 colors0[chan], colors1[chan]);
1191 }
1192 }
1193 else {
1194 /* 2D tex */
1195 for (chan = 0; chan < 4; chan++) {
1196 colors_out[chan] = colors0[chan];
1197 }
1198 }
1199 }
1200 }
1201
1202
1203
1204 /**
1205 * General texture sampling codegen.
1206 * This function handles texture sampling for all texture targets (1D,
1207 * 2D, 3D, cube) and all filtering modes.
1208 */
1209 static void
1210 lp_build_sample_general(struct lp_build_sample_context *bld,
1211 unsigned unit,
1212 LLVMValueRef s,
1213 LLVMValueRef t,
1214 LLVMValueRef r,
1215 LLVMValueRef width,
1216 LLVMValueRef height,
1217 LLVMValueRef depth,
1218 LLVMValueRef width_vec,
1219 LLVMValueRef height_vec,
1220 LLVMValueRef depth_vec,
1221 LLVMValueRef row_stride_array,
1222 LLVMValueRef img_stride_vec,
1223 LLVMValueRef data_array,
1224 LLVMValueRef *colors_out)
1225 {
1226 const unsigned mip_filter = bld->static_state->min_mip_filter;
1227 const unsigned min_filter = bld->static_state->min_img_filter;
1228 const unsigned mag_filter = bld->static_state->mag_img_filter;
1229 const int dims = texture_dims(bld->static_state->target);
1230 LLVMValueRef lod, lod_fpart;
1231 LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
1232 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1233 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1234 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1235 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1236 LLVMValueRef data_ptr0, data_ptr1;
1237 int chan;
1238
1239 /*
1240 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1241 mip_filter, min_filter, mag_filter);
1242 */
1243
1244 /*
1245 * Compute the level of detail (mipmap level index(es)).
1246 */
1247 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1248 /* always use mip level 0 */
1249 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1250 }
1251 else {
1252 /* compute float LOD */
1253 lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
1254
1255 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1256 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1257 }
1258 else {
1259 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1260 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1261 &lod_fpart);
1262 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1263 }
1264 }
1265
1266 /*
1267 * Convert scalar integer mipmap levels into vectors.
1268 */
1269 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1270 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1271 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1272
1273 /*
1274 * Compute width, height at mipmap level 'ilevel0'
1275 */
1276 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1277 if (dims >= 2) {
1278 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1279 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1280 ilevel0);
1281 if (dims == 3) {
1282 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1283 }
1284 }
1285 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1286 /* compute width, height, depth for second mipmap level at ilevel1 */
1287 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1288 if (dims >= 2) {
1289 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1290 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1291 ilevel1);
1292 if (dims == 3) {
1293 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1294 }
1295 }
1296 }
1297
1298 /*
1299 * Choose cube face, recompute texcoords.
1300 */
1301 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1302
1303 }
1304
1305 /*
1306 * Get pointer(s) to image data for mipmap level(s).
1307 */
1308 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1309 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1310 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1311 }
1312
1313 /*
1314 * Get/interpolate texture colors.
1315 */
1316 /* XXX temporarily force this path: */
1317 if (1 /*min_filter == mag_filter*/) {
1318 /* same filter for minification or magnification */
1319 LLVMValueRef colors0[4], colors1[4];
1320
1321 if (min_filter == PIPE_TEX_FILTER_NEAREST) {
1322 lp_build_sample_image_nearest(bld,
1323 width0_vec, height0_vec, depth0_vec,
1324 row_stride0_vec, img_stride0_vec,
1325 data_ptr0, s, t, r, colors0);
1326
1327 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1328 /* sample the second mipmap level, and interp */
1329 lp_build_sample_image_nearest(bld,
1330 width1_vec, height1_vec, depth1_vec,
1331 row_stride1_vec, img_stride1_vec,
1332 data_ptr1, s, t, r, colors1);
1333 }
1334 }
1335 else {
1336 assert(min_filter == PIPE_TEX_FILTER_LINEAR);
1337
1338 lp_build_sample_image_linear(bld,
1339 width0_vec, height0_vec, depth0_vec,
1340 row_stride0_vec, img_stride0_vec,
1341 data_ptr0, s, t, r, colors0);
1342
1343
1344 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1345 /* sample the second mipmap level, and interp */
1346 lp_build_sample_image_linear(bld,
1347 width1_vec, height1_vec, depth1_vec,
1348 row_stride1_vec, img_stride1_vec,
1349 data_ptr1, s, t, r, colors1);
1350 }
1351 }
1352
1353 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1354 /* interpolate samples from the two mipmap levels */
1355 for (chan = 0; chan < 4; chan++) {
1356 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1357 colors0[chan], colors1[chan]);
1358 }
1359 }
1360 else {
1361 /* use first/only level's colors */
1362 for (chan = 0; chan < 4; chan++) {
1363 colors_out[chan] = colors0[chan];
1364 }
1365 }
1366 }
1367 else {
1368 /* emit conditional to choose min image filter or mag image filter
1369 * depending on the lod being >0 or <= 0, respectively.
1370 */
1371 abort();
1372 }
1373 }
1374
1375
1376
1377 static void
1378 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1379 struct lp_type dst_type,
1380 LLVMValueRef packed,
1381 LLVMValueRef *rgba)
1382 {
1383 LLVMValueRef mask = lp_build_int_const_scalar(dst_type, 0xff);
1384 unsigned chan;
1385
1386 /* Decode the input vector components */
1387 for (chan = 0; chan < 4; ++chan) {
1388 unsigned start = chan*8;
1389 unsigned stop = start + 8;
1390 LLVMValueRef input;
1391
1392 input = packed;
1393
1394 if(start)
1395 input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(dst_type, start), "");
1396
1397 if(stop < 32)
1398 input = LLVMBuildAnd(builder, input, mask, "");
1399
1400 input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1401
1402 rgba[chan] = input;
1403 }
1404 }
1405
1406
1407 static void
1408 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1409 LLVMValueRef s,
1410 LLVMValueRef t,
1411 LLVMValueRef width,
1412 LLVMValueRef height,
1413 LLVMValueRef stride_array,
1414 LLVMValueRef data_array,
1415 LLVMValueRef *texel)
1416 {
1417 LLVMBuilderRef builder = bld->builder;
1418 struct lp_build_context i32, h16, u8n;
1419 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1420 LLVMValueRef i32_c8, i32_c128, i32_c255;
1421 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1422 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1423 LLVMValueRef x0, x1;
1424 LLVMValueRef y0, y1;
1425 LLVMValueRef neighbors[2][2];
1426 LLVMValueRef neighbors_lo[2][2];
1427 LLVMValueRef neighbors_hi[2][2];
1428 LLVMValueRef packed, packed_lo, packed_hi;
1429 LLVMValueRef unswizzled[4];
1430 LLVMValueRef stride;
1431
1432 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1433 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1434 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1435
1436 i32_vec_type = lp_build_vec_type(i32.type);
1437 h16_vec_type = lp_build_vec_type(h16.type);
1438 u8n_vec_type = lp_build_vec_type(u8n.type);
1439
1440 if (bld->static_state->normalized_coords) {
1441 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1442 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1443 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1444 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1445 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1446 }
1447
1448 /* scale coords by 256 (8 fractional bits) */
1449 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1450 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1451
1452 /* convert float to int */
1453 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1454 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1455
1456 /* subtract 0.5 (add -128) */
1457 i32_c128 = lp_build_int_const_scalar(i32.type, -128);
1458 s = LLVMBuildAdd(builder, s, i32_c128, "");
1459 t = LLVMBuildAdd(builder, t, i32_c128, "");
1460
1461 /* compute floor (shift right 8) */
1462 i32_c8 = lp_build_int_const_scalar(i32.type, 8);
1463 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1464 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1465
1466 /* compute fractional part (AND with 0xff) */
1467 i32_c255 = lp_build_int_const_scalar(i32.type, 255);
1468 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1469 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1470
1471 x0 = s_ipart;
1472 y0 = t_ipart;
1473
1474 x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1475 y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1476
1477 x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
1478 bld->static_state->wrap_s);
1479 y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1480 bld->static_state->wrap_t);
1481
1482 x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
1483 bld->static_state->wrap_s);
1484 y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1485 bld->static_state->wrap_t);
1486
1487 /*
1488 * Transform 4 x i32 in
1489 *
1490 * s_fpart = {s0, s1, s2, s3}
1491 *
1492 * into 8 x i16
1493 *
1494 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1495 *
1496 * into two 8 x i16
1497 *
1498 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1499 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1500 *
1501 * and likewise for t_fpart. There is no risk of loosing precision here
1502 * since the fractional parts only use the lower 8bits.
1503 */
1504
1505 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1506 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1507
1508 {
1509 LLVMTypeRef elem_type = LLVMInt32Type();
1510 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1511 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1512 LLVMValueRef shuffle_lo;
1513 LLVMValueRef shuffle_hi;
1514 unsigned i, j;
1515
1516 for(j = 0; j < h16.type.length; j += 4) {
1517 unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1518 LLVMValueRef index;
1519
1520 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1521 for(i = 0; i < 4; ++i)
1522 shuffles_lo[j + i] = index;
1523
1524 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1525 for(i = 0; i < 4; ++i)
1526 shuffles_hi[j + i] = index;
1527 }
1528
1529 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1530 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1531
1532 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1533 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1534 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1535 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1536 }
1537
1538 stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1539
1540 /*
1541 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1542 *
1543 * rgba0 rgba1 rgba2 rgba3
1544 *
1545 * bit cast them into 16 x u8
1546 *
1547 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1548 *
1549 * unpack them into two 8 x i16:
1550 *
1551 * r0 g0 b0 a0 r1 g1 b1 a1
1552 * r2 g2 b2 a2 r3 g3 b3 a3
1553 *
1554 * The higher 8 bits of the resulting elements will be zero.
1555 */
1556
1557 neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1558 neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1559 neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1560 neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1561
1562 neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1563 neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1564 neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1565 neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1566
1567 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1568 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1569 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1570 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1571
1572 /*
1573 * Linear interpolate with 8.8 fixed point.
1574 */
1575
1576 packed_lo = lp_build_lerp_2d(&h16,
1577 s_fpart_lo, t_fpart_lo,
1578 neighbors_lo[0][0],
1579 neighbors_lo[0][1],
1580 neighbors_lo[1][0],
1581 neighbors_lo[1][1]);
1582
1583 packed_hi = lp_build_lerp_2d(&h16,
1584 s_fpart_hi, t_fpart_hi,
1585 neighbors_hi[0][0],
1586 neighbors_hi[0][1],
1587 neighbors_hi[1][0],
1588 neighbors_hi[1][1]);
1589
1590 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1591
1592 /*
1593 * Convert to SoA and swizzle.
1594 */
1595
1596 packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
1597
1598 lp_build_rgba8_to_f32_soa(bld->builder,
1599 bld->texel_type,
1600 packed, unswizzled);
1601
1602 lp_build_format_swizzle_soa(bld->format_desc,
1603 bld->texel_type, unswizzled,
1604 texel);
1605 }
1606
1607
1608 static void
1609 lp_build_sample_compare(struct lp_build_sample_context *bld,
1610 LLVMValueRef p,
1611 LLVMValueRef *texel)
1612 {
1613 struct lp_build_context *texel_bld = &bld->texel_bld;
1614 LLVMValueRef res;
1615 unsigned chan;
1616
1617 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1618 return;
1619
1620 /* TODO: Compare before swizzling, to avoid redundant computations */
1621 res = NULL;
1622 for(chan = 0; chan < 4; ++chan) {
1623 LLVMValueRef cmp;
1624 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1625 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1626
1627 if(res)
1628 res = lp_build_add(texel_bld, res, cmp);
1629 else
1630 res = cmp;
1631 }
1632
1633 assert(res);
1634 res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25));
1635
1636 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1637 for(chan = 0; chan < 3; ++chan)
1638 texel[chan] = res;
1639 texel[3] = texel_bld->one;
1640 }
1641
1642
1643 /**
1644 * Build texture sampling code.
1645 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1646 * R, G, B, A.
1647 * \param type vector float type to use for coords, etc.
1648 */
1649 void
1650 lp_build_sample_soa(LLVMBuilderRef builder,
1651 const struct lp_sampler_static_state *static_state,
1652 struct lp_sampler_dynamic_state *dynamic_state,
1653 struct lp_type type,
1654 unsigned unit,
1655 unsigned num_coords,
1656 const LLVMValueRef *coords,
1657 LLVMValueRef lodbias,
1658 LLVMValueRef *texel)
1659 {
1660 struct lp_build_sample_context bld;
1661 LLVMValueRef width, width_vec;
1662 LLVMValueRef height, height_vec;
1663 LLVMValueRef depth, depth_vec;
1664 LLVMValueRef stride_array;
1665 LLVMValueRef data_array;
1666 LLVMValueRef s;
1667 LLVMValueRef t;
1668 LLVMValueRef r;
1669
1670 (void) lp_build_lod_selector; /* temporary to silence warning */
1671 (void) lp_build_nearest_mip_level;
1672 (void) lp_build_linear_mip_levels;
1673 (void) lp_build_minify;
1674
1675 /* Setup our build context */
1676 memset(&bld, 0, sizeof bld);
1677 bld.builder = builder;
1678 bld.static_state = static_state;
1679 bld.dynamic_state = dynamic_state;
1680 bld.format_desc = util_format_description(static_state->format);
1681
1682 bld.float_type = lp_type_float(32);
1683 bld.int_type = lp_type_int(32);
1684 bld.coord_type = type;
1685 bld.uint_coord_type = lp_uint_type(type);
1686 bld.int_coord_type = lp_int_type(type);
1687 bld.texel_type = type;
1688
1689 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1690 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1691 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1692 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
1693 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1694 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1695
1696 /* Get the dynamic state */
1697 width = dynamic_state->width(dynamic_state, builder, unit);
1698 height = dynamic_state->height(dynamic_state, builder, unit);
1699 depth = dynamic_state->depth(dynamic_state, builder, unit);
1700 stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1701 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1702 /* Note that data_array is an array[level] of pointers to texture images */
1703
1704 s = coords[0];
1705 t = coords[1];
1706 r = coords[2];
1707
1708 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
1709 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
1710 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
1711
1712 if (lp_format_is_rgba8(bld.format_desc) &&
1713 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
1714 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
1715 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
1716 is_simple_wrap_mode(static_state->wrap_s) &&
1717 is_simple_wrap_mode(static_state->wrap_t)) {
1718 /* special case */
1719 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
1720 stride_array, data_array, texel);
1721 }
1722 else {
1723 lp_build_sample_general(&bld, unit, s, t, r,
1724 width, height, depth,
1725 width_vec, height_vec, depth_vec,
1726 stride_array, NULL, data_array,
1727 texel);
1728 }
1729
1730 lp_build_sample_compare(&bld, r, texel);
1731 }