projects / mesa.git / blob
? search:


2ffe21f97d69d7657c05c6028018f984ea31bf0d
[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 "pipe/p_shader_tokens.h"
39 #include "util/u_debug.h"
40 #include "util/u_dump.h"
41 #include "util/u_memory.h"
42 #include "util/u_math.h"
43 #include "util/u_format.h"
44 #include "util/u_cpu_detect.h"
45 #include "lp_bld_debug.h"
46 #include "lp_bld_type.h"
47 #include "lp_bld_const.h"
48 #include "lp_bld_conv.h"
49 #include "lp_bld_arit.h"
50 #include "lp_bld_bitarit.h"
51 #include "lp_bld_logic.h"
52 #include "lp_bld_printf.h"
53 #include "lp_bld_swizzle.h"
54 #include "lp_bld_flow.h"
55 #include "lp_bld_gather.h"
56 #include "lp_bld_format.h"
57 #include "lp_bld_sample.h"
58 #include "lp_bld_sample_aos.h"
59 #include "lp_bld_struct.h"
60 #include "lp_bld_quad.h"
61 #include "lp_bld_pack.h"
62
63
64 /**
65 * Generate code to fetch a texel from a texture at int coords (x, y, z).
66 * The computation depends on whether the texture is 1D, 2D or 3D.
67 * The result, texel, will be float vectors:
68 * texel[0] = red values
69 * texel[1] = green values
70 * texel[2] = blue values
71 * texel[3] = alpha values
72 */
73 static void
74 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
75 unsigned sampler_unit,
76 LLVMValueRef width,
77 LLVMValueRef height,
78 LLVMValueRef depth,
79 LLVMValueRef x,
80 LLVMValueRef y,
81 LLVMValueRef z,
82 LLVMValueRef y_stride,
83 LLVMValueRef z_stride,
84 LLVMValueRef data_ptr,
85 LLVMValueRef mipoffsets,
86 LLVMValueRef texel_out[4])
87 {
88 const struct lp_static_sampler_state *static_state = bld->static_sampler_state;
89 const unsigned dims = bld->dims;
90 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
91 LLVMBuilderRef builder = bld->gallivm->builder;
92 LLVMValueRef offset;
93 LLVMValueRef i, j;
94 LLVMValueRef use_border = NULL;
95
96 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
97 if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
98 static_state->min_img_filter,
99 static_state->mag_img_filter)) {
100 LLVMValueRef b1, b2;
101 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
102 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
103 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
104 }
105
106 if (dims >= 2 &&
107 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
108 static_state->min_img_filter,
109 static_state->mag_img_filter)) {
110 LLVMValueRef b1, b2;
111 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
112 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
113 if (use_border) {
114 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
115 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
116 }
117 else {
118 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
119 }
120 }
121
122 if (dims == 3 &&
123 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
124 static_state->min_img_filter,
125 static_state->mag_img_filter)) {
126 LLVMValueRef b1, b2;
127 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
128 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
129 if (use_border) {
130 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
131 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
132 }
133 else {
134 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
135 }
136 }
137
138 /* convert x,y,z coords to linear offset from start of texture, in bytes */
139 lp_build_sample_offset(&bld->int_coord_bld,
140 bld->format_desc,
141 x, y, z, y_stride, z_stride,
142 &offset, &i, &j);
143 if (mipoffsets) {
144 offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
145 }
146
147 if (use_border) {
148 /* If we can sample the border color, it means that texcoords may
149 * lie outside the bounds of the texture image. We need to do
150 * something to prevent reading out of bounds and causing a segfault.
151 *
152 * Simply AND the texture coords with !use_border. This will cause
153 * coords which are out of bounds to become zero. Zero's guaranteed
154 * to be inside the texture image.
155 */
156 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
157 }
158
159 lp_build_fetch_rgba_soa(bld->gallivm,
160 bld->format_desc,
161 bld->texel_type,
162 data_ptr, offset,
163 i, j,
164 texel_out);
165
166 /*
167 * Note: if we find an app which frequently samples the texture border
168 * we might want to implement a true conditional here to avoid sampling
169 * the texture whenever possible (since that's quite a bit of code).
170 * Ex:
171 * if (use_border) {
172 * texel = border_color;
173 * }
174 * else {
175 * texel = sample_texture(coord);
176 * }
177 * As it is now, we always sample the texture, then selectively replace
178 * the texel color results with the border color.
179 */
180
181 if (use_border) {
182 /* select texel color or border color depending on use_border. */
183 LLVMValueRef border_color_ptr =
184 bld->dynamic_state->border_color(bld->dynamic_state,
185 bld->gallivm, sampler_unit);
186 const struct util_format_description *format_desc;
187 int chan;
188 format_desc = util_format_description(bld->static_texture_state->format);
189 /*
190 * Only replace channels which are actually present. The others should
191 * get optimized away eventually by sampler_view swizzle anyway but it's
192 * easier too as we'd need some extra logic for channels where we can't
193 * determine the format directly otherwise.
194 */
195 for (chan = 0; chan < 4; chan++) {
196 unsigned chan_s;
197 /* reverse-map channel... */
198 for (chan_s = 0; chan_s < 4; chan_s++) {
199 if (chan_s == format_desc->swizzle[chan]) {
200 break;
201 }
202 }
203 if (chan_s <= 3) {
204 LLVMValueRef border_chan =
205 lp_build_array_get(bld->gallivm, border_color_ptr,
206 lp_build_const_int32(bld->gallivm, chan));
207 LLVMValueRef border_chan_vec =
208 lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
209
210 if (!bld->texel_type.floating) {
211 border_chan_vec = LLVMBuildBitCast(builder, border_chan_vec,
212 bld->texel_bld.vec_type, "");
213 }
214 else {
215 /*
216 * For normalized format need to clamp border color (technically
217 * probably should also quantize the data). Really sucks doing this
218 * here but can't avoid at least for now since this is part of
219 * sampler state and texture format is part of sampler_view state.
220 */
221 unsigned chan_type = format_desc->channel[chan_s].type;
222 unsigned chan_norm = format_desc->channel[chan_s].normalized;
223 if (chan_type == UTIL_FORMAT_TYPE_SIGNED && chan_norm) {
224 LLVMValueRef clamp_min;
225 clamp_min = lp_build_const_vec(bld->gallivm, bld->texel_type, -1.0F);
226 border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec,
227 clamp_min,
228 bld->texel_bld.one);
229 }
230 else if (chan_type == UTIL_FORMAT_TYPE_UNSIGNED && chan_norm) {
231 border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec,
232 bld->texel_bld.zero,
233 bld->texel_bld.one);
234 }
235 /* not exactly sure about all others but I think should be ok? */
236 }
237 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
238 border_chan_vec, texel_out[chan]);
239 }
240 }
241 }
242 }
243
244
245 /**
246 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
247 */
248 static LLVMValueRef
249 lp_build_coord_mirror(struct lp_build_sample_context *bld,
250 LLVMValueRef coord)
251 {
252 struct lp_build_context *coord_bld = &bld->coord_bld;
253 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
254 LLVMValueRef fract, flr, isOdd;
255
256 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
257
258 /* isOdd = flr & 1 */
259 isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
260
261 /* make coord positive or negative depending on isOdd */
262 coord = lp_build_set_sign(coord_bld, fract, isOdd);
263
264 /* convert isOdd to float */
265 isOdd = lp_build_int_to_float(coord_bld, isOdd);
266
267 /* add isOdd to coord */
268 coord = lp_build_add(coord_bld, coord, isOdd);
269
270 return coord;
271 }
272
273
274 /**
275 * Helper to compute the first coord and the weight for
276 * linear wrap repeat npot textures
277 */
278 void
279 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
280 LLVMValueRef coord_f,
281 LLVMValueRef length_i,
282 LLVMValueRef length_f,
283 LLVMValueRef *coord0_i,
284 LLVMValueRef *weight_f)
285 {
286 struct lp_build_context *coord_bld = &bld->coord_bld;
287 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
288 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
289 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
290 int_coord_bld->one);
291 LLVMValueRef mask;
292 /* wrap with normalized floats is just fract */
293 coord_f = lp_build_fract(coord_bld, coord_f);
294 /* mul by size and subtract 0.5 */
295 coord_f = lp_build_mul(coord_bld, coord_f, length_f);
296 coord_f = lp_build_sub(coord_bld, coord_f, half);
297 /*
298 * we avoided the 0.5/length division before the repeat wrap,
299 * now need to fix up edge cases with selects
300 */
301 /* convert to int, compute lerp weight */
302 lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
303 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
304 PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
305 *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
306 }
307
308
309 /**
310 * Build LLVM code for texture wrap mode for linear filtering.
311 * \param x0_out returns first integer texcoord
312 * \param x1_out returns second integer texcoord
313 * \param weight_out returns linear interpolation weight
314 */
315 static void
316 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
317 LLVMValueRef coord,
318 LLVMValueRef length,
319 LLVMValueRef length_f,
320 LLVMValueRef offset,
321 boolean is_pot,
322 unsigned wrap_mode,
323 LLVMValueRef *x0_out,
324 LLVMValueRef *x1_out,
325 LLVMValueRef *weight_out)
326 {
327 struct lp_build_context *coord_bld = &bld->coord_bld;
328 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
329 LLVMBuilderRef builder = bld->gallivm->builder;
330 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
331 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
332 LLVMValueRef coord0, coord1, weight;
333
334 switch(wrap_mode) {
335 case PIPE_TEX_WRAP_REPEAT:
336 if (is_pot) {
337 /* mul by size and subtract 0.5 */
338 coord = lp_build_mul(coord_bld, coord, length_f);
339 coord = lp_build_sub(coord_bld, coord, half);
340 if (offset) {
341 offset = lp_build_int_to_float(coord_bld, offset);
342 coord = lp_build_add(coord_bld, coord, offset);
343 }
344 /* convert to int, compute lerp weight */
345 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
346 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
347 /* repeat wrap */
348 coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
349 coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
350 }
351 else {
352 LLVMValueRef mask;
353 if (offset) {
354 offset = lp_build_int_to_float(coord_bld, offset);
355 offset = lp_build_div(coord_bld, offset, length_f);
356 coord = lp_build_add(coord_bld, coord, offset);
357 }
358 lp_build_coord_repeat_npot_linear(bld, coord,
359 length, length_f,
360 &coord0, &weight);
361 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
362 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
363 coord1 = LLVMBuildAnd(builder,
364 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
365 mask, "");
366 }
367 break;
368
369 case PIPE_TEX_WRAP_CLAMP:
370 if (bld->static_sampler_state->normalized_coords) {
371 /* scale coord to length */
372 coord = lp_build_mul(coord_bld, coord, length_f);
373 }
374 if (offset) {
375 offset = lp_build_int_to_float(coord_bld, offset);
376 coord = lp_build_add(coord_bld, coord, offset);
377 }
378
379 /* clamp to [0, length] */
380 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
381
382 coord = lp_build_sub(coord_bld, coord, half);
383
384 /* convert to int, compute lerp weight */
385 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
386 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
387 break;
388
389 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
390 {
391 struct lp_build_context abs_coord_bld = bld->coord_bld;
392 abs_coord_bld.type.sign = FALSE;
393
394 if (bld->static_sampler_state->normalized_coords) {
395 /* mul by tex size */
396 coord = lp_build_mul(coord_bld, coord, length_f);
397 }
398 if (offset) {
399 offset = lp_build_int_to_float(coord_bld, offset);
400 coord = lp_build_add(coord_bld, coord, offset);
401 }
402
403 /* clamp to length max */
404 coord = lp_build_min(coord_bld, coord, length_f);
405 /* subtract 0.5 */
406 coord = lp_build_sub(coord_bld, coord, half);
407 /* clamp to [0, length - 0.5] */
408 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
409 /* convert to int, compute lerp weight */
410 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
411 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
412 /* coord1 = min(coord1, length-1) */
413 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
414 break;
415 }
416
417 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
418 if (bld->static_sampler_state->normalized_coords) {
419 /* scale coord to length */
420 coord = lp_build_mul(coord_bld, coord, length_f);
421 }
422 if (offset) {
423 offset = lp_build_int_to_float(coord_bld, offset);
424 coord = lp_build_add(coord_bld, coord, offset);
425 }
426 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
427 /* can skip clamp (though might not work for very large coord values */
428 coord = lp_build_sub(coord_bld, coord, half);
429 /* convert to int, compute lerp weight */
430 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
431 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
432 break;
433
434 case PIPE_TEX_WRAP_MIRROR_REPEAT:
435 /* compute mirror function */
436 coord = lp_build_coord_mirror(bld, coord);
437
438 /* scale coord to length */
439 coord = lp_build_mul(coord_bld, coord, length_f);
440 coord = lp_build_sub(coord_bld, coord, half);
441 if (offset) {
442 offset = lp_build_int_to_float(coord_bld, offset);
443 coord = lp_build_add(coord_bld, coord, offset);
444 }
445
446 /* convert to int, compute lerp weight */
447 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
448 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
449
450 /* coord0 = max(coord0, 0) */
451 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
452 /* coord1 = min(coord1, length-1) */
453 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
454 break;
455
456 case PIPE_TEX_WRAP_MIRROR_CLAMP:
457 if (bld->static_sampler_state->normalized_coords) {
458 /* scale coord to length */
459 coord = lp_build_mul(coord_bld, coord, length_f);
460 }
461 if (offset) {
462 offset = lp_build_int_to_float(coord_bld, offset);
463 coord = lp_build_add(coord_bld, coord, offset);
464 }
465 coord = lp_build_abs(coord_bld, coord);
466
467 /* clamp to [0, length] */
468 coord = lp_build_min(coord_bld, coord, length_f);
469
470 coord = lp_build_sub(coord_bld, coord, half);
471
472 /* convert to int, compute lerp weight */
473 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
474 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
475 break;
476
477 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
478 {
479 struct lp_build_context abs_coord_bld = bld->coord_bld;
480 abs_coord_bld.type.sign = FALSE;
481
482 if (bld->static_sampler_state->normalized_coords) {
483 /* scale coord to length */
484 coord = lp_build_mul(coord_bld, coord, length_f);
485 }
486 if (offset) {
487 offset = lp_build_int_to_float(coord_bld, offset);
488 coord = lp_build_add(coord_bld, coord, offset);
489 }
490 coord = lp_build_abs(coord_bld, coord);
491
492 /* clamp to length max */
493 coord = lp_build_min(coord_bld, coord, length_f);
494 /* subtract 0.5 */
495 coord = lp_build_sub(coord_bld, coord, half);
496 /* clamp to [0, length - 0.5] */
497 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
498
499 /* convert to int, compute lerp weight */
500 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
501 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
502 /* coord1 = min(coord1, length-1) */
503 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
504 }
505 break;
506
507 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
508 {
509 if (bld->static_sampler_state->normalized_coords) {
510 /* scale coord to length */
511 coord = lp_build_mul(coord_bld, coord, length_f);
512 }
513 if (offset) {
514 offset = lp_build_int_to_float(coord_bld, offset);
515 coord = lp_build_add(coord_bld, coord, offset);
516 }
517 coord = lp_build_abs(coord_bld, coord);
518
519 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
520 /* skip clamp - always positive, and other side
521 only potentially matters for very large coords */
522 coord = lp_build_sub(coord_bld, coord, half);
523
524 /* convert to int, compute lerp weight */
525 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
526 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
527 }
528 break;
529
530 default:
531 assert(0);
532 coord0 = NULL;
533 coord1 = NULL;
534 weight = NULL;
535 }
536
537 *x0_out = coord0;
538 *x1_out = coord1;
539 *weight_out = weight;
540 }
541
542
543 /**
544 * Build LLVM code for texture wrap mode for nearest filtering.
545 * \param coord the incoming texcoord (nominally in [0,1])
546 * \param length the texture size along one dimension, as int vector
547 * \param length_f the texture size along one dimension, as float vector
548 * \param offset texel offset along one dimension (as int vector)
549 * \param is_pot if TRUE, length is a power of two
550 * \param wrap_mode one of PIPE_TEX_WRAP_x
551 */
552 static LLVMValueRef
553 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
554 LLVMValueRef coord,
555 LLVMValueRef length,
556 LLVMValueRef length_f,
557 LLVMValueRef offset,
558 boolean is_pot,
559 unsigned wrap_mode)
560 {
561 struct lp_build_context *coord_bld = &bld->coord_bld;
562 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
563 LLVMBuilderRef builder = bld->gallivm->builder;
564 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
565 LLVMValueRef icoord;
566
567 switch(wrap_mode) {
568 case PIPE_TEX_WRAP_REPEAT:
569 if (is_pot) {
570 coord = lp_build_mul(coord_bld, coord, length_f);
571 icoord = lp_build_ifloor(coord_bld, coord);
572 if (offset) {
573 icoord = lp_build_add(int_coord_bld, icoord, offset);
574 }
575 icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
576 }
577 else {
578 if (offset) {
579 offset = lp_build_int_to_float(coord_bld, offset);
580 offset = lp_build_div(coord_bld, offset, length_f);
581 coord = lp_build_add(coord_bld, coord, offset);
582 }
583 /* take fraction, unnormalize */
584 coord = lp_build_fract_safe(coord_bld, coord);
585 coord = lp_build_mul(coord_bld, coord, length_f);
586 icoord = lp_build_itrunc(coord_bld, coord);
587 }
588 break;
589
590 case PIPE_TEX_WRAP_CLAMP:
591 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
592 if (bld->static_sampler_state->normalized_coords) {
593 /* scale coord to length */
594 coord = lp_build_mul(coord_bld, coord, length_f);
595 }
596
597 /* floor */
598 /* use itrunc instead since we clamp to 0 anyway */
599 icoord = lp_build_itrunc(coord_bld, coord);
600 if (offset) {
601 icoord = lp_build_add(int_coord_bld, icoord, offset);
602 }
603
604 /* clamp to [0, length - 1]. */
605 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
606 length_minus_one);
607 break;
608
609 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
610 if (bld->static_sampler_state->normalized_coords) {
611 /* scale coord to length */
612 coord = lp_build_mul(coord_bld, coord, length_f);
613 }
614 /* no clamp necessary, border masking will handle this */
615 icoord = lp_build_ifloor(coord_bld, coord);
616 if (offset) {
617 icoord = lp_build_add(int_coord_bld, icoord, offset);
618 }
619 break;
620
621 case PIPE_TEX_WRAP_MIRROR_REPEAT:
622 if (offset) {
623 offset = lp_build_int_to_float(coord_bld, offset);
624 offset = lp_build_div(coord_bld, offset, length_f);
625 coord = lp_build_add(coord_bld, coord, offset);
626 }
627 /* compute mirror function */
628 coord = lp_build_coord_mirror(bld, coord);
629
630 /* scale coord to length */
631 assert(bld->static_sampler_state->normalized_coords);
632 coord = lp_build_mul(coord_bld, coord, length_f);
633
634 /* itrunc == ifloor here */
635 icoord = lp_build_itrunc(coord_bld, coord);
636
637 /* clamp to [0, length - 1] */
638 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
639 break;
640
641 case PIPE_TEX_WRAP_MIRROR_CLAMP:
642 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
643 if (bld->static_sampler_state->normalized_coords) {
644 /* scale coord to length */
645 coord = lp_build_mul(coord_bld, coord, length_f);
646 }
647 if (offset) {
648 offset = lp_build_int_to_float(coord_bld, offset);
649 coord = lp_build_add(coord_bld, coord, offset);
650 }
651 coord = lp_build_abs(coord_bld, coord);
652
653 /* itrunc == ifloor here */
654 icoord = lp_build_itrunc(coord_bld, coord);
655
656 /* clamp to [0, length - 1] */
657 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
658 break;
659
660 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
661 if (bld->static_sampler_state->normalized_coords) {
662 /* scale coord to length */
663 coord = lp_build_mul(coord_bld, coord, length_f);
664 }
665 if (offset) {
666 offset = lp_build_int_to_float(coord_bld, offset);
667 coord = lp_build_add(coord_bld, coord, offset);
668 }
669 coord = lp_build_abs(coord_bld, coord);
670
671 /* itrunc == ifloor here */
672 icoord = lp_build_itrunc(coord_bld, coord);
673 break;
674
675 default:
676 assert(0);
677 icoord = NULL;
678 }
679
680 return icoord;
681 }
682
683
684 /**
685 * Do shadow test/comparison.
686 * \param p shadow ref value
687 * \param texel the texel to compare against
688 */
689 static LLVMValueRef
690 lp_build_sample_comparefunc(struct lp_build_sample_context *bld,
691 LLVMValueRef p,
692 LLVMValueRef texel)
693 {
694 struct lp_build_context *texel_bld = &bld->texel_bld;
695 LLVMValueRef res;
696
697 if (0) {
698 //lp_build_print_value(bld->gallivm, "shadow cmp coord", p);
699 lp_build_print_value(bld->gallivm, "shadow cmp texel", texel);
700 }
701
702 /* result = (p FUNC texel) ? 1 : 0 */
703 /*
704 * honor d3d10 floating point rules here, which state that comparisons
705 * are ordered except NOT_EQUAL which is unordered.
706 */
707 if (bld->static_sampler_state->compare_func != PIPE_FUNC_NOTEQUAL) {
708 res = lp_build_cmp_ordered(texel_bld, bld->static_sampler_state->compare_func,
709 p, texel);
710 }
711 else {
712 res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
713 p, texel);
714 }
715 return res;
716 }
717
718
719 /**
720 * Generate code to sample a mipmap level with nearest filtering.
721 * If sampling a cube texture, r = cube face in [0,5].
722 */
723 static void
724 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
725 unsigned sampler_unit,
726 LLVMValueRef size,
727 LLVMValueRef row_stride_vec,
728 LLVMValueRef img_stride_vec,
729 LLVMValueRef data_ptr,
730 LLVMValueRef mipoffsets,
731 LLVMValueRef *coords,
732 const LLVMValueRef *offsets,
733 LLVMValueRef colors_out[4])
734 {
735 const unsigned dims = bld->dims;
736 LLVMValueRef width_vec;
737 LLVMValueRef height_vec;
738 LLVMValueRef depth_vec;
739 LLVMValueRef flt_size;
740 LLVMValueRef flt_width_vec;
741 LLVMValueRef flt_height_vec;
742 LLVMValueRef flt_depth_vec;
743 LLVMValueRef x, y = NULL, z = NULL;
744
745 lp_build_extract_image_sizes(bld,
746 &bld->int_size_bld,
747 bld->int_coord_type,
748 size,
749 &width_vec, &height_vec, &depth_vec);
750
751 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
752
753 lp_build_extract_image_sizes(bld,
754 &bld->float_size_bld,
755 bld->coord_type,
756 flt_size,
757 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
758
759 /*
760 * Compute integer texcoords.
761 */
762 x = lp_build_sample_wrap_nearest(bld, coords[0], width_vec,
763 flt_width_vec, offsets[0],
764 bld->static_texture_state->pot_width,
765 bld->static_sampler_state->wrap_s);
766 lp_build_name(x, "tex.x.wrapped");
767
768 if (dims >= 2) {
769 y = lp_build_sample_wrap_nearest(bld, coords[1], height_vec,
770 flt_height_vec, offsets[1],
771 bld->static_texture_state->pot_height,
772 bld->static_sampler_state->wrap_t);
773 lp_build_name(y, "tex.y.wrapped");
774
775 if (dims == 3) {
776 z = lp_build_sample_wrap_nearest(bld, coords[2], depth_vec,
777 flt_depth_vec, offsets[2],
778 bld->static_texture_state->pot_depth,
779 bld->static_sampler_state->wrap_r);
780 lp_build_name(z, "tex.z.wrapped");
781 }
782 }
783 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
784 bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
785 bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
786 z = coords[2];
787 lp_build_name(z, "tex.z.layer");
788 }
789
790 /*
791 * Get texture colors.
792 */
793 lp_build_sample_texel_soa(bld, sampler_unit,
794 width_vec, height_vec, depth_vec,
795 x, y, z,
796 row_stride_vec, img_stride_vec,
797 data_ptr, mipoffsets, colors_out);
798
799 if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
800 LLVMValueRef cmpval;
801 cmpval = lp_build_sample_comparefunc(bld, coords[4], colors_out[0]);
802 /* this is really just a AND 1.0, cmpval but llvm is clever enough */
803 colors_out[0] = lp_build_select(&bld->texel_bld, cmpval,
804 bld->texel_bld.one, bld->texel_bld.zero);
805 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
806 }
807
808 }
809
810
811 /**
812 * Like a lerp, but inputs are 0/~0 masks, so can simplify slightly.
813 */
814 static LLVMValueRef
815 lp_build_masklerp(struct lp_build_context *bld,
816 LLVMValueRef weight,
817 LLVMValueRef mask0,
818 LLVMValueRef mask1)
819 {
820 struct gallivm_state *gallivm = bld->gallivm;
821 LLVMBuilderRef builder = gallivm->builder;
822 LLVMValueRef weight2;
823
824 weight2 = lp_build_sub(bld, bld->one, weight);
825 weight = LLVMBuildBitCast(builder, weight,
826 lp_build_int_vec_type(gallivm, bld->type), "");
827 weight2 = LLVMBuildBitCast(builder, weight2,
828 lp_build_int_vec_type(gallivm, bld->type), "");
829 weight = LLVMBuildAnd(builder, weight, mask1, "");
830 weight2 = LLVMBuildAnd(builder, weight2, mask0, "");
831 weight = LLVMBuildBitCast(builder, weight, bld->vec_type, "");
832 weight2 = LLVMBuildBitCast(builder, weight2, bld->vec_type, "");
833 return lp_build_add(bld, weight, weight2);
834 }
835
836 /**
837 * Like a 2d lerp, but inputs are 0/~0 masks, so can simplify slightly.
838 */
839 static LLVMValueRef
840 lp_build_masklerp2d(struct lp_build_context *bld,
841 LLVMValueRef weight0,
842 LLVMValueRef weight1,
843 LLVMValueRef mask00,
844 LLVMValueRef mask01,
845 LLVMValueRef mask10,
846 LLVMValueRef mask11)
847 {
848 LLVMValueRef val0 = lp_build_masklerp(bld, weight0, mask00, mask01);
849 LLVMValueRef val1 = lp_build_masklerp(bld, weight0, mask10, mask11);
850 return lp_build_lerp(bld, weight1, val0, val1, 0);
851 }
852
853 /**
854 * Generate code to sample a mipmap level with linear filtering.
855 * If sampling a cube texture, r = cube face in [0,5].
856 */
857 static void
858 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
859 unsigned sampler_unit,
860 LLVMValueRef size,
861 LLVMValueRef row_stride_vec,
862 LLVMValueRef img_stride_vec,
863 LLVMValueRef data_ptr,
864 LLVMValueRef mipoffsets,
865 LLVMValueRef *coords,
866 const LLVMValueRef *offsets,
867 LLVMValueRef colors_out[4])
868 {
869 const unsigned dims = bld->dims;
870 LLVMValueRef width_vec;
871 LLVMValueRef height_vec;
872 LLVMValueRef depth_vec;
873 LLVMValueRef flt_size;
874 LLVMValueRef flt_width_vec;
875 LLVMValueRef flt_height_vec;
876 LLVMValueRef flt_depth_vec;
877 LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
878 LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
879 LLVMValueRef neighbors[2][2][4];
880 int chan;
881
882 lp_build_extract_image_sizes(bld,
883 &bld->int_size_bld,
884 bld->int_coord_type,
885 size,
886 &width_vec, &height_vec, &depth_vec);
887
888 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
889
890 lp_build_extract_image_sizes(bld,
891 &bld->float_size_bld,
892 bld->coord_type,
893 flt_size,
894 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
895
896 /*
897 * Compute integer texcoords.
898 */
899 lp_build_sample_wrap_linear(bld, coords[0], width_vec,
900 flt_width_vec, offsets[0],
901 bld->static_texture_state->pot_width,
902 bld->static_sampler_state->wrap_s,
903 &x0, &x1, &s_fpart);
904 lp_build_name(x0, "tex.x0.wrapped");
905 lp_build_name(x1, "tex.x1.wrapped");
906
907 if (dims >= 2) {
908 lp_build_sample_wrap_linear(bld, coords[1], height_vec,
909 flt_height_vec, offsets[1],
910 bld->static_texture_state->pot_height,
911 bld->static_sampler_state->wrap_t,
912 &y0, &y1, &t_fpart);
913 lp_build_name(y0, "tex.y0.wrapped");
914 lp_build_name(y1, "tex.y1.wrapped");
915
916 if (dims == 3) {
917 lp_build_sample_wrap_linear(bld, coords[2], depth_vec,
918 flt_depth_vec, offsets[2],
919 bld->static_texture_state->pot_depth,
920 bld->static_sampler_state->wrap_r,
921 &z0, &z1, &r_fpart);
922 lp_build_name(z0, "tex.z0.wrapped");
923 lp_build_name(z1, "tex.z1.wrapped");
924 }
925 }
926 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
927 bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
928 bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
929 z0 = z1 = coords[2]; /* cube face or layer */
930 lp_build_name(z0, "tex.z0.layer");
931 lp_build_name(z1, "tex.z1.layer");
932 }
933
934
935 /*
936 * Get texture colors.
937 */
938 /* get x0/x1 texels */
939 lp_build_sample_texel_soa(bld, sampler_unit,
940 width_vec, height_vec, depth_vec,
941 x0, y0, z0,
942 row_stride_vec, img_stride_vec,
943 data_ptr, mipoffsets, neighbors[0][0]);
944 lp_build_sample_texel_soa(bld, sampler_unit,
945 width_vec, height_vec, depth_vec,
946 x1, y0, z0,
947 row_stride_vec, img_stride_vec,
948 data_ptr, mipoffsets, neighbors[0][1]);
949
950 if (dims == 1) {
951 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
952 /* Interpolate two samples from 1D image to produce one color */
953 for (chan = 0; chan < 4; chan++) {
954 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
955 neighbors[0][0][chan],
956 neighbors[0][1][chan],
957 0);
958 }
959 }
960 else {
961 LLVMValueRef cmpval0, cmpval1;
962 cmpval0 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
963 cmpval1 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
964 /* simplified lerp, AND mask with weight and add */
965 colors_out[0] = lp_build_masklerp(&bld->texel_bld, s_fpart,
966 cmpval0, cmpval1);
967 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
968 }
969 }
970 else {
971 /* 2D/3D texture */
972 LLVMValueRef colors0[4];
973
974 /* get x0/x1 texels at y1 */
975 lp_build_sample_texel_soa(bld, sampler_unit,
976 width_vec, height_vec, depth_vec,
977 x0, y1, z0,
978 row_stride_vec, img_stride_vec,
979 data_ptr, mipoffsets, neighbors[1][0]);
980 lp_build_sample_texel_soa(bld, sampler_unit,
981 width_vec, height_vec, depth_vec,
982 x1, y1, z0,
983 row_stride_vec, img_stride_vec,
984 data_ptr, mipoffsets, neighbors[1][1]);
985
986 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
987 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
988 for (chan = 0; chan < 4; chan++) {
989 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
990 s_fpart, t_fpart,
991 neighbors[0][0][chan],
992 neighbors[0][1][chan],
993 neighbors[1][0][chan],
994 neighbors[1][1][chan],
995 0);
996 }
997 }
998 else {
999 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1000 cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1001 cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1002 cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1003 cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1004 colors0[0] = lp_build_masklerp2d(&bld->texel_bld, s_fpart, t_fpart,
1005 cmpval00, cmpval01, cmpval10, cmpval11);
1006 colors0[1] = colors0[2] = colors0[3] = colors0[0];
1007 }
1008
1009 if (dims == 3) {
1010 LLVMValueRef neighbors1[2][2][4];
1011 LLVMValueRef colors1[4];
1012
1013 /* get x0/x1/y0/y1 texels at z1 */
1014 lp_build_sample_texel_soa(bld, sampler_unit,
1015 width_vec, height_vec, depth_vec,
1016 x0, y0, z1,
1017 row_stride_vec, img_stride_vec,
1018 data_ptr, mipoffsets, neighbors1[0][0]);
1019 lp_build_sample_texel_soa(bld, sampler_unit,
1020 width_vec, height_vec, depth_vec,
1021 x1, y0, z1,
1022 row_stride_vec, img_stride_vec,
1023 data_ptr, mipoffsets, neighbors1[0][1]);
1024 lp_build_sample_texel_soa(bld, sampler_unit,
1025 width_vec, height_vec, depth_vec,
1026 x0, y1, z1,
1027 row_stride_vec, img_stride_vec,
1028 data_ptr, mipoffsets, neighbors1[1][0]);
1029 lp_build_sample_texel_soa(bld, sampler_unit,
1030 width_vec, height_vec, depth_vec,
1031 x1, y1, z1,
1032 row_stride_vec, img_stride_vec,
1033 data_ptr, mipoffsets, neighbors1[1][1]);
1034
1035 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1036 /* Bilinear interpolate the four samples from the second Z slice */
1037 for (chan = 0; chan < 4; chan++) {
1038 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1039 s_fpart, t_fpart,
1040 neighbors1[0][0][chan],
1041 neighbors1[0][1][chan],
1042 neighbors1[1][0][chan],
1043 neighbors1[1][1][chan],
1044 0);
1045 }
1046 /* Linearly interpolate the two samples from the two 3D slices */
1047 for (chan = 0; chan < 4; chan++) {
1048 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1049 r_fpart,
1050 colors0[chan], colors1[chan],
1051 0);
1052 }
1053 }
1054 else {
1055 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1056 cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1057 cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1058 cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1059 cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1060 colors1[0] = lp_build_masklerp2d(&bld->texel_bld, s_fpart, t_fpart,
1061 cmpval00, cmpval01, cmpval10, cmpval11);
1062 /* Linearly interpolate the two samples from the two 3D slices */
1063 colors_out[0] = lp_build_lerp(&bld->texel_bld,
1064 r_fpart,
1065 colors0[0], colors1[0],
1066 0);
1067 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1068 }
1069 }
1070 else {
1071 /* 2D tex */
1072 for (chan = 0; chan < 4; chan++) {
1073 colors_out[chan] = colors0[chan];
1074 }
1075 }
1076 }
1077 }
1078
1079
1080 /**
1081 * Sample the texture/mipmap using given image filter and mip filter.
1082 * data0_ptr and data1_ptr point to the two mipmap levels to sample
1083 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
1084 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1085 */
1086 static void
1087 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1088 unsigned sampler_unit,
1089 unsigned img_filter,
1090 unsigned mip_filter,
1091 LLVMValueRef *coords,
1092 const LLVMValueRef *offsets,
1093 LLVMValueRef ilevel0,
1094 LLVMValueRef ilevel1,
1095 LLVMValueRef lod_fpart,
1096 LLVMValueRef *colors_out)
1097 {
1098 LLVMBuilderRef builder = bld->gallivm->builder;
1099 LLVMValueRef size0 = NULL;
1100 LLVMValueRef size1 = NULL;
1101 LLVMValueRef row_stride0_vec = NULL;
1102 LLVMValueRef row_stride1_vec = NULL;
1103 LLVMValueRef img_stride0_vec = NULL;
1104 LLVMValueRef img_stride1_vec = NULL;
1105 LLVMValueRef data_ptr0 = NULL;
1106 LLVMValueRef data_ptr1 = NULL;
1107 LLVMValueRef mipoff0 = NULL;
1108 LLVMValueRef mipoff1 = NULL;
1109 LLVMValueRef colors0[4], colors1[4];
1110 unsigned chan;
1111
1112 /* sample the first mipmap level */
1113 lp_build_mipmap_level_sizes(bld, ilevel0,
1114 &size0,
1115 &row_stride0_vec, &img_stride0_vec);
1116 if (bld->num_lods == 1) {
1117 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
1118 }
1119 else {
1120 /* This path should work for num_lods 1 too but slightly less efficient */
1121 data_ptr0 = bld->base_ptr;
1122 mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
1123 }
1124 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1125 lp_build_sample_image_nearest(bld, sampler_unit,
1126 size0,
1127 row_stride0_vec, img_stride0_vec,
1128 data_ptr0, mipoff0, coords, offsets,
1129 colors0);
1130 }
1131 else {
1132 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1133 lp_build_sample_image_linear(bld, sampler_unit,
1134 size0,
1135 row_stride0_vec, img_stride0_vec,
1136 data_ptr0, mipoff0, coords, offsets,
1137 colors0);
1138 }
1139
1140 /* Store the first level's colors in the output variables */
1141 for (chan = 0; chan < 4; chan++) {
1142 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1143 }
1144
1145 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1146 struct lp_build_if_state if_ctx;
1147 LLVMValueRef need_lerp;
1148
1149 /* need_lerp = lod_fpart > 0 */
1150 if (bld->num_lods == 1) {
1151 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
1152 lod_fpart, bld->levelf_bld.zero,
1153 "need_lerp");
1154 }
1155 else {
1156 /*
1157 * We'll do mip filtering if any of the quads (or individual
1158 * pixel in case of per-pixel lod) need it.
1159 * It might be better to split the vectors here and only fetch/filter
1160 * quads which need it.
1161 */
1162 /*
1163 * We unfortunately need to clamp lod_fpart here since we can get
1164 * negative values which would screw up filtering if not all
1165 * lod_fpart values have same sign.
1166 */
1167 lod_fpart = lp_build_max(&bld->levelf_bld, lod_fpart,
1168 bld->levelf_bld.zero);
1169 need_lerp = lp_build_compare(bld->gallivm, bld->levelf_bld.type,
1170 PIPE_FUNC_GREATER,
1171 lod_fpart, bld->levelf_bld.zero);
1172 need_lerp = lp_build_any_true_range(&bld->leveli_bld, bld->num_lods, need_lerp);
1173 }
1174
1175 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1176 {
1177 /* sample the second mipmap level */
1178 lp_build_mipmap_level_sizes(bld, ilevel1,
1179 &size1,
1180 &row_stride1_vec, &img_stride1_vec);
1181 if (bld->num_lods == 1) {
1182 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1183 }
1184 else {
1185 data_ptr1 = bld->base_ptr;
1186 mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1187 }
1188 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1189 lp_build_sample_image_nearest(bld, sampler_unit,
1190 size1,
1191 row_stride1_vec, img_stride1_vec,
1192 data_ptr1, mipoff1, coords, offsets,
1193 colors1);
1194 }
1195 else {
1196 lp_build_sample_image_linear(bld, sampler_unit,
1197 size1,
1198 row_stride1_vec, img_stride1_vec,
1199 data_ptr1, mipoff1, coords, offsets,
1200 colors1);
1201 }
1202
1203 /* interpolate samples from the two mipmap levels */
1204
1205 if (bld->num_lods != bld->coord_type.length)
1206 lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1207 bld->levelf_bld.type,
1208 bld->texel_bld.type,
1209 lod_fpart);
1210
1211 for (chan = 0; chan < 4; chan++) {
1212 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1213 colors0[chan], colors1[chan],
1214 0);
1215 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1216 }
1217 }
1218 lp_build_endif(&if_ctx);
1219 }
1220 }
1221
1222
1223 /**
1224 * Build (per-coord) layer value.
1225 * Either clamp layer to valid values or fill in optional out_of_bounds
1226 * value and just return value unclamped.
1227 */
1228 static LLVMValueRef
1229 lp_build_layer_coord(struct lp_build_sample_context *bld,
1230 unsigned texture_unit,
1231 LLVMValueRef layer,
1232 LLVMValueRef *out_of_bounds)
1233 {
1234 LLVMValueRef num_layers;
1235 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1236
1237 num_layers = bld->dynamic_state->depth(bld->dynamic_state,
1238 bld->gallivm, texture_unit);
1239
1240 if (out_of_bounds) {
1241 LLVMValueRef out1, out;
1242 num_layers = lp_build_broadcast_scalar(int_coord_bld, num_layers);
1243 out = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, layer, int_coord_bld->zero);
1244 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, layer, num_layers);
1245 *out_of_bounds = lp_build_or(int_coord_bld, out, out1);
1246 return layer;
1247 }
1248 else {
1249 LLVMValueRef maxlayer;
1250 maxlayer = lp_build_sub(&bld->int_bld, num_layers, bld->int_bld.one);
1251 maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer);
1252 return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer);
1253 }
1254 }
1255
1256
1257 /**
1258 * Calculate cube face, lod, mip levels.
1259 */
1260 static void
1261 lp_build_sample_common(struct lp_build_sample_context *bld,
1262 unsigned texture_index,
1263 unsigned sampler_index,
1264 LLVMValueRef *coords,
1265 const struct lp_derivatives *derivs, /* optional */
1266 LLVMValueRef lod_bias, /* optional */
1267 LLVMValueRef explicit_lod, /* optional */
1268 LLVMValueRef *lod_ipart,
1269 LLVMValueRef *lod_fpart,
1270 LLVMValueRef *ilevel0,
1271 LLVMValueRef *ilevel1)
1272 {
1273 const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1274 const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1275 const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1276 const unsigned target = bld->static_texture_state->target;
1277 LLVMValueRef first_level, cube_rho = NULL;
1278
1279 /*
1280 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1281 mip_filter, min_filter, mag_filter);
1282 */
1283
1284 /*
1285 * Choose cube face, recompute texcoords for the chosen face and
1286 * compute rho here too (as it requires transform of derivatives).
1287 */
1288 if (target == PIPE_TEXTURE_CUBE) {
1289 boolean need_derivs;
1290 need_derivs = ((min_filter != mag_filter ||
1291 mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
1292 !bld->static_sampler_state->min_max_lod_equal &&
1293 !explicit_lod);
1294 lp_build_cube_lookup(bld, coords, derivs, &cube_rho, need_derivs);
1295 }
1296 else if (target == PIPE_TEXTURE_1D_ARRAY ||
1297 target == PIPE_TEXTURE_2D_ARRAY) {
1298 coords[2] = lp_build_iround(&bld->coord_bld, coords[2]);
1299 coords[2] = lp_build_layer_coord(bld, texture_index, coords[2], NULL);
1300 }
1301
1302 if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
1303 /*
1304 * Clamp p coords to [0,1] for fixed function depth texture format here.
1305 * Technically this is not entirely correct for unorm depth as the ref value
1306 * should be converted to the depth format (quantization!) and comparison
1307 * then done in texture format. This would actually help performance (since
1308 * only need to do it once and could save the per-sample conversion of texels
1309 * to floats instead), but it would need more messy code (would need to push
1310 * at least some bits down to actual fetch so conversion could be skipped,
1311 * and would have ugly interaction with border color, would need to convert
1312 * border color to that format too or do some other tricks to make it work).
1313 */
1314 const struct util_format_description *format_desc;
1315 unsigned chan_type;
1316 format_desc = util_format_description(bld->static_texture_state->format);
1317 /* not entirely sure we couldn't end up with non-valid swizzle here */
1318 chan_type = format_desc->swizzle[0] <= UTIL_FORMAT_SWIZZLE_W ?
1319 format_desc->channel[format_desc->swizzle[0]].type :
1320 UTIL_FORMAT_TYPE_FLOAT;
1321 if (chan_type != UTIL_FORMAT_TYPE_FLOAT) {
1322 coords[4] = lp_build_clamp(&bld->coord_bld, coords[4],
1323 bld->coord_bld.zero, bld->coord_bld.one);
1324 }
1325 }
1326
1327 /*
1328 * Compute the level of detail (float).
1329 */
1330 if (min_filter != mag_filter ||
1331 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1332 /* Need to compute lod either to choose mipmap levels or to
1333 * distinguish between minification/magnification with one mipmap level.
1334 */
1335 lp_build_lod_selector(bld, texture_index, sampler_index,
1336 coords[0], coords[1], coords[2], cube_rho,
1337 derivs, lod_bias, explicit_lod,
1338 mip_filter,
1339 lod_ipart, lod_fpart);
1340 } else {
1341 *lod_ipart = bld->leveli_bld.zero;
1342 }
1343
1344 /*
1345 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1346 */
1347 switch (mip_filter) {
1348 default:
1349 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1350 /* fall-through */
1351 case PIPE_TEX_MIPFILTER_NONE:
1352 /* always use mip level 0 */
1353 if (HAVE_LLVM == 0x0207 && target == PIPE_TEXTURE_CUBE) {
1354 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1355 * We should be able to set ilevel0 = const(0) but that causes
1356 * bad x86 code to be emitted.
1357 */
1358 assert(*lod_ipart);
1359 lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL);
1360 }
1361 else {
1362 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1363 bld->gallivm, texture_index);
1364 first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level);
1365 *ilevel0 = first_level;
1366 }
1367 break;
1368 case PIPE_TEX_MIPFILTER_NEAREST:
1369 assert(*lod_ipart);
1370 lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL);
1371 break;
1372 case PIPE_TEX_MIPFILTER_LINEAR:
1373 assert(*lod_ipart);
1374 assert(*lod_fpart);
1375 lp_build_linear_mip_levels(bld, texture_index,
1376 *lod_ipart, lod_fpart,
1377 ilevel0, ilevel1);
1378 break;
1379 }
1380 }
1381
1382 /**
1383 * General texture sampling codegen.
1384 * This function handles texture sampling for all texture targets (1D,
1385 * 2D, 3D, cube) and all filtering modes.
1386 */
1387 static void
1388 lp_build_sample_general(struct lp_build_sample_context *bld,
1389 unsigned sampler_unit,
1390 LLVMValueRef *coords,
1391 const LLVMValueRef *offsets,
1392 LLVMValueRef lod_ipart,
1393 LLVMValueRef lod_fpart,
1394 LLVMValueRef ilevel0,
1395 LLVMValueRef ilevel1,
1396 LLVMValueRef *colors_out)
1397 {
1398 struct lp_build_context *int_bld = &bld->int_bld;
1399 LLVMBuilderRef builder = bld->gallivm->builder;
1400 const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1401 const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1402 const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1403 LLVMValueRef texels[4];
1404 unsigned chan;
1405
1406 /*
1407 * Get/interpolate texture colors.
1408 */
1409
1410 for (chan = 0; chan < 4; ++chan) {
1411 texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1412 lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
1413 }
1414
1415 if (min_filter == mag_filter) {
1416 /* no need to distinguish between minification and magnification */
1417 lp_build_sample_mipmap(bld, sampler_unit,
1418 min_filter, mip_filter,
1419 coords, offsets,
1420 ilevel0, ilevel1, lod_fpart,
1421 texels);
1422 }
1423 else {
1424 /* Emit conditional to choose min image filter or mag image filter
1425 * depending on the lod being > 0 or <= 0, respectively.
1426 */
1427 struct lp_build_if_state if_ctx;
1428 LLVMValueRef minify;
1429
1430 /*
1431 * XXX this should to all lods into account, if some are min
1432 * some max probably could hack up the coords/weights in the linear
1433 * path with selects to work for nearest.
1434 * If that's just two quads sitting next to each other it seems
1435 * quite ok to do the same filtering method on both though, at
1436 * least unless we have explicit lod (and who uses different
1437 * min/mag filter with that?)
1438 */
1439 if (bld->num_lods > 1)
1440 lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
1441 lp_build_const_int32(bld->gallivm, 0), "");
1442
1443 /* minify = lod >= 0.0 */
1444 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1445 lod_ipart, int_bld->zero, "");
1446
1447 lp_build_if(&if_ctx, bld->gallivm, minify);
1448 {
1449 /* Use the minification filter */
1450 lp_build_sample_mipmap(bld, sampler_unit,
1451 min_filter, mip_filter,
1452 coords, offsets,
1453 ilevel0, ilevel1, lod_fpart,
1454 texels);
1455 }
1456 lp_build_else(&if_ctx);
1457 {
1458 /* Use the magnification filter */
1459 lp_build_sample_mipmap(bld, sampler_unit,
1460 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1461 coords, offsets,
1462 ilevel0, NULL, NULL,
1463 texels);
1464 }
1465 lp_build_endif(&if_ctx);
1466 }
1467
1468 for (chan = 0; chan < 4; ++chan) {
1469 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1470 lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
1471 }
1472 }
1473
1474
1475 /**
1476 * Texel fetch function.
1477 * In contrast to general sampling there is no filtering, no coord minification,
1478 * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
1479 * directly to be applied to the selected mip level (after adding texel offsets).
1480 * This function handles texel fetch for all targets where texel fetch is supported
1481 * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
1482 */
1483 static void
1484 lp_build_fetch_texel(struct lp_build_sample_context *bld,
1485 unsigned texture_unit,
1486 const LLVMValueRef *coords,
1487 LLVMValueRef explicit_lod,
1488 const LLVMValueRef *offsets,
1489 LLVMValueRef *colors_out)
1490 {
1491 struct lp_build_context *perquadi_bld = &bld->leveli_bld;
1492 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1493 unsigned dims = bld->dims, chan;
1494 unsigned target = bld->static_texture_state->target;
1495 boolean out_of_bound_ret_zero = TRUE;
1496 LLVMValueRef size, ilevel;
1497 LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
1498 LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
1499 LLVMValueRef width, height, depth, i, j;
1500 LLVMValueRef offset, out_of_bounds, out1;
1501
1502 out_of_bounds = int_coord_bld->zero;
1503
1504 if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
1505 if (bld->num_lods != int_coord_bld->type.length) {
1506 ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
1507 perquadi_bld->type, explicit_lod, 0);
1508 }
1509 else {
1510 ilevel = explicit_lod;
1511 }
1512 lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel,
1513 out_of_bound_ret_zero ? &out_of_bounds : NULL);
1514 }
1515 else {
1516 assert(bld->num_lods == 1);
1517 if (bld->static_texture_state->target != PIPE_BUFFER) {
1518 ilevel = bld->dynamic_state->first_level(bld->dynamic_state,
1519 bld->gallivm, texture_unit);
1520 }
1521 else {
1522 ilevel = lp_build_const_int32(bld->gallivm, 0);
1523 }
1524 }
1525 lp_build_mipmap_level_sizes(bld, ilevel,
1526 &size,
1527 &row_stride_vec, &img_stride_vec);
1528 lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
1529 size, &width, &height, &depth);
1530
1531 if (target == PIPE_TEXTURE_1D_ARRAY ||
1532 target == PIPE_TEXTURE_2D_ARRAY) {
1533 if (out_of_bound_ret_zero) {
1534 z = lp_build_layer_coord(bld, texture_unit, z, &out1);
1535 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1536 }
1537 else {
1538 z = lp_build_layer_coord(bld, texture_unit, z, NULL);
1539 }
1540 }
1541
1542 /* This is a lot like border sampling */
1543 if (offsets[0]) {
1544 /*
1545 * coords are really unsigned, offsets are signed, but I don't think
1546 * exceeding 31 bits is possible
1547 */
1548 x = lp_build_add(int_coord_bld, x, offsets[0]);
1549 }
1550 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
1551 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1552 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
1553 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1554
1555 if (dims >= 2) {
1556 if (offsets[1]) {
1557 y = lp_build_add(int_coord_bld, y, offsets[1]);
1558 }
1559 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
1560 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1561 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
1562 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1563
1564 if (dims >= 3) {
1565 if (offsets[2]) {
1566 z = lp_build_add(int_coord_bld, z, offsets[2]);
1567 }
1568 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
1569 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1570 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
1571 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1572 }
1573 }
1574
1575 lp_build_sample_offset(int_coord_bld,
1576 bld->format_desc,
1577 x, y, z, row_stride_vec, img_stride_vec,
1578 &offset, &i, &j);
1579
1580 if (bld->static_texture_state->target != PIPE_BUFFER) {
1581 offset = lp_build_add(int_coord_bld, offset,
1582 lp_build_get_mip_offsets(bld, ilevel));
1583 }
1584
1585 offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
1586
1587 lp_build_fetch_rgba_soa(bld->gallivm,
1588 bld->format_desc,
1589 bld->texel_type,
1590 bld->base_ptr, offset,
1591 i, j,
1592 colors_out);
1593
1594 if (out_of_bound_ret_zero) {
1595 /*
1596 * Only needed for ARB_robust_buffer_access_behavior and d3d10.
1597 * Could use min/max above instead of out-of-bounds comparisons
1598 * if we don't care about the result returned for out-of-bounds.
1599 */
1600 for (chan = 0; chan < 4; chan++) {
1601 colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
1602 bld->texel_bld.zero, colors_out[chan]);
1603 }
1604 }
1605 }
1606
1607
1608 /**
1609 * Just set texels to white instead of actually sampling the texture.
1610 * For debugging.
1611 */
1612 void
1613 lp_build_sample_nop(struct gallivm_state *gallivm,
1614 struct lp_type type,
1615 const LLVMValueRef *coords,
1616 LLVMValueRef texel_out[4])
1617 {
1618 LLVMValueRef one = lp_build_one(gallivm, type);
1619 unsigned chan;
1620
1621 for (chan = 0; chan < 4; chan++) {
1622 texel_out[chan] = one;
1623 }
1624 }
1625
1626
1627 /**
1628 * Build texture sampling code.
1629 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1630 * R, G, B, A.
1631 * \param type vector float type to use for coords, etc.
1632 * \param is_fetch if this is a texel fetch instruction.
1633 * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
1634 */
1635 void
1636 lp_build_sample_soa(struct gallivm_state *gallivm,
1637 const struct lp_static_texture_state *static_texture_state,
1638 const struct lp_static_sampler_state *static_sampler_state,
1639 struct lp_sampler_dynamic_state *dynamic_state,
1640 struct lp_type type,
1641 boolean is_fetch,
1642 unsigned texture_index,
1643 unsigned sampler_index,
1644 const LLVMValueRef *coords,
1645 const LLVMValueRef *offsets,
1646 const struct lp_derivatives *derivs, /* optional */
1647 LLVMValueRef lod_bias, /* optional */
1648 LLVMValueRef explicit_lod, /* optional */
1649 enum lp_sampler_lod_property lod_property,
1650 LLVMValueRef texel_out[4])
1651 {
1652 unsigned target = static_texture_state->target;
1653 unsigned dims = texture_dims(target);
1654 unsigned num_quads = type.length / 4;
1655 unsigned mip_filter, i;
1656 struct lp_build_sample_context bld;
1657 struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
1658 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1659 LLVMBuilderRef builder = gallivm->builder;
1660 LLVMValueRef tex_width, newcoords[5];
1661
1662 if (0) {
1663 enum pipe_format fmt = static_texture_state->format;
1664 debug_printf("Sample from %s\n", util_format_name(fmt));
1665 }
1666
1667 assert(type.floating);
1668
1669 /* Setup our build context */
1670 memset(&bld, 0, sizeof bld);
1671 bld.gallivm = gallivm;
1672 bld.static_sampler_state = &derived_sampler_state;
1673 bld.static_texture_state = static_texture_state;
1674 bld.dynamic_state = dynamic_state;
1675 bld.format_desc = util_format_description(static_texture_state->format);
1676 bld.dims = dims;
1677
1678 bld.vector_width = lp_type_width(type);
1679
1680 bld.float_type = lp_type_float(32);
1681 bld.int_type = lp_type_int(32);
1682 bld.coord_type = type;
1683 bld.int_coord_type = lp_int_type(type);
1684 bld.float_size_in_type = lp_type_float(32);
1685 bld.float_size_in_type.length = dims > 1 ? 4 : 1;
1686 bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
1687 bld.texel_type = type;
1688
1689 /* always using the first channel hopefully should be safe,
1690 * if not things WILL break in other places anyway.
1691 */
1692 if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1693 bld.format_desc->channel[0].pure_integer) {
1694 if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1695 bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1696 }
1697 else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1698 bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
1699 }
1700 }
1701 else if (util_format_has_stencil(bld.format_desc) &&
1702 !util_format_has_depth(bld.format_desc)) {
1703 /* for stencil only formats, sample stencil (uint) */
1704 bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1705 }
1706
1707 if (!static_texture_state->level_zero_only) {
1708 derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
1709 } else {
1710 derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
1711 }
1712 mip_filter = derived_sampler_state.min_mip_filter;
1713
1714 if (0) {
1715 debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
1716 }
1717
1718 /*
1719 * This is all a bit complicated different paths are chosen for performance
1720 * reasons.
1721 * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
1722 * everything (the last two options are equivalent for 4-wide case).
1723 * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
1724 * lod is calculated then the lod value extracted afterwards so making this
1725 * case basically the same as far as lod handling is concerned for the
1726 * further sample/filter code as the 1 lod for everything case.
1727 * Different lod handling mostly shows up when building mipmap sizes
1728 * (lp_build_mipmap_level_sizes() and friends) and also in filtering
1729 * (getting the fractional part of the lod to the right texels).
1730 */
1731
1732 /*
1733 * There are other situations where at least the multiple int lods could be
1734 * avoided like min and max lod being equal.
1735 */
1736 if (explicit_lod && lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
1737 ((is_fetch && target != PIPE_BUFFER) ||
1738 (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
1739 bld.num_lods = type.length;
1740 /* TODO: for true scalar_lod should only use 1 lod value */
1741 else if ((is_fetch && explicit_lod && target != PIPE_BUFFER ) ||
1742 (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
1743 bld.num_lods = num_quads;
1744 }
1745 else {
1746 bld.num_lods = 1;
1747 }
1748
1749 bld.levelf_type = type;
1750 /* we want native vector size to be able to use our intrinsics */
1751 if (bld.num_lods != type.length) {
1752 bld.levelf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
1753 }
1754 bld.leveli_type = lp_int_type(bld.levelf_type);
1755 bld.float_size_type = bld.float_size_in_type;
1756 /* Note: size vectors may not be native. They contain minified w/h/d/_ values,
1757 * with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */
1758 if (bld.num_lods > 1) {
1759 bld.float_size_type.length = bld.num_lods == type.length ?
1760 bld.num_lods * bld.float_size_in_type.length :
1761 type.length;
1762 }
1763 bld.int_size_type = lp_int_type(bld.float_size_type);
1764
1765 lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1766 lp_build_context_init(&bld.float_vec_bld, gallivm, type);
1767 lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1768 lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1769 lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1770 lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
1771 lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
1772 lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1773 lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1774 lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1775 lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
1776 lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
1777
1778 /* Get the dynamic state */
1779 tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index);
1780 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index);
1781 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index);
1782 bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index);
1783 bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index);
1784 /* Note that mip_offsets is an array[level] of offsets to texture images */
1785
1786 /* width, height, depth as single int vector */
1787 if (dims <= 1) {
1788 bld.int_size = tex_width;
1789 }
1790 else {
1791 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
1792 tex_width, LLVMConstInt(i32t, 0, 0), "");
1793 if (dims >= 2) {
1794 LLVMValueRef tex_height =
1795 dynamic_state->height(dynamic_state, gallivm, texture_index);
1796 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1797 tex_height, LLVMConstInt(i32t, 1, 0), "");
1798 if (dims >= 3) {
1799 LLVMValueRef tex_depth =
1800 dynamic_state->depth(dynamic_state, gallivm, texture_index);
1801 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1802 tex_depth, LLVMConstInt(i32t, 2, 0), "");
1803 }
1804 }
1805 }
1806
1807 for (i = 0; i < 5; i++) {
1808 newcoords[i] = coords[i];
1809 }
1810
1811 if (0) {
1812 /* For debug: no-op texture sampling */
1813 lp_build_sample_nop(gallivm,
1814 bld.texel_type,
1815 newcoords,
1816 texel_out);
1817 }
1818
1819 else if (is_fetch) {
1820 lp_build_fetch_texel(&bld, texture_index, newcoords,
1821 explicit_lod, offsets,
1822 texel_out);
1823 }
1824
1825 else {
1826 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
1827 LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
1828 boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
1829 lp_is_simple_wrap_mode(static_sampler_state->wrap_s) &&
1830 lp_is_simple_wrap_mode(static_sampler_state->wrap_t) &&
1831 /* not sure this is strictly needed or simply impossible */
1832 static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE;
1833
1834 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1835 !use_aos && util_format_fits_8unorm(bld.format_desc)) {
1836 debug_printf("%s: using floating point linear filtering for %s\n",
1837 __FUNCTION__, bld.format_desc->short_name);
1838 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1839 static_sampler_state->min_img_filter,
1840 static_sampler_state->mag_img_filter,
1841 static_sampler_state->min_mip_filter,
1842 static_sampler_state->wrap_s,
1843 static_sampler_state->wrap_t);
1844 }
1845
1846 lp_build_sample_common(&bld, texture_index, sampler_index,
1847 newcoords,
1848 derivs, lod_bias, explicit_lod,
1849 &lod_ipart, &lod_fpart,
1850 &ilevel0, &ilevel1);
1851
1852 /*
1853 * we only try 8-wide sampling with soa as it appears to
1854 * be a loss with aos with AVX (but it should work).
1855 * (It should be faster if we'd support avx2)
1856 */
1857 if (num_quads == 1 || !use_aos) {
1858
1859 if (num_quads > 1) {
1860 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1861 LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
1862 /*
1863 * These parameters are the same for all quads,
1864 * could probably simplify.
1865 */
1866 lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
1867 ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
1868 }
1869 }
1870 if (use_aos) {
1871 /* do sampling/filtering with fixed pt arithmetic */
1872 lp_build_sample_aos(&bld, sampler_index,
1873 newcoords[0], newcoords[1],
1874 newcoords[2],
1875 offsets, lod_ipart, lod_fpart,
1876 ilevel0, ilevel1,
1877 texel_out);
1878 }
1879
1880 else {
1881 lp_build_sample_general(&bld, sampler_index,
1882 newcoords, offsets,
1883 lod_ipart, lod_fpart,
1884 ilevel0, ilevel1,
1885 texel_out);
1886 }
1887 }
1888 else {
1889 unsigned j;
1890 struct lp_build_sample_context bld4;
1891 struct lp_type type4 = type;
1892 unsigned i;
1893 LLVMValueRef texelout4[4];
1894 LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
1895
1896 type4.length = 4;
1897
1898 /* Setup our build context */
1899 memset(&bld4, 0, sizeof bld4);
1900 bld4.gallivm = bld.gallivm;
1901 bld4.static_texture_state = bld.static_texture_state;
1902 bld4.static_sampler_state = bld.static_sampler_state;
1903 bld4.dynamic_state = bld.dynamic_state;
1904 bld4.format_desc = bld.format_desc;
1905 bld4.dims = bld.dims;
1906 bld4.row_stride_array = bld.row_stride_array;
1907 bld4.img_stride_array = bld.img_stride_array;
1908 bld4.base_ptr = bld.base_ptr;
1909 bld4.mip_offsets = bld.mip_offsets;
1910 bld4.int_size = bld.int_size;
1911
1912 bld4.vector_width = lp_type_width(type4);
1913
1914 bld4.float_type = lp_type_float(32);
1915 bld4.int_type = lp_type_int(32);
1916 bld4.coord_type = type4;
1917 bld4.int_coord_type = lp_int_type(type4);
1918 bld4.float_size_in_type = lp_type_float(32);
1919 bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
1920 bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
1921 bld4.texel_type = bld.texel_type;
1922 bld4.texel_type.length = 4;
1923 bld4.levelf_type = type4;
1924 /* we want native vector size to be able to use our intrinsics */
1925 bld4.levelf_type.length = 1;
1926 bld4.leveli_type = lp_int_type(bld4.levelf_type);
1927
1928 if (explicit_lod && lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
1929 ((is_fetch && target != PIPE_BUFFER) ||
1930 (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
1931 bld4.num_lods = type4.length;
1932 else
1933 bld4.num_lods = 1;
1934
1935 bld4.levelf_type = type4;
1936 /* we want native vector size to be able to use our intrinsics */
1937 if (bld4.num_lods != type4.length) {
1938 bld4.levelf_type.length = 1;
1939 }
1940 bld4.leveli_type = lp_int_type(bld4.levelf_type);
1941 bld4.float_size_type = bld4.float_size_in_type;
1942 if (bld4.num_lods > 1) {
1943 bld4.float_size_type.length = bld4.num_lods == type4.length ?
1944 bld4.num_lods * bld4.float_size_in_type.length :
1945 type4.length;
1946 }
1947 bld4.int_size_type = lp_int_type(bld4.float_size_type);
1948
1949 lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
1950 lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
1951 lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
1952 lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
1953 lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
1954 lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
1955 lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
1956 lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
1957 lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
1958 lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
1959 lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
1960 lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
1961
1962 for (i = 0; i < num_quads; i++) {
1963 LLVMValueRef s4, t4, r4;
1964 LLVMValueRef lod_ipart4, lod_fpart4 = NULL;
1965 LLVMValueRef ilevel04, ilevel14 = NULL;
1966 LLVMValueRef offsets4[4] = { NULL };
1967 unsigned num_lods = bld4.num_lods;
1968
1969 s4 = lp_build_extract_range(gallivm, newcoords[0], 4*i, 4);
1970 t4 = lp_build_extract_range(gallivm, newcoords[1], 4*i, 4);
1971 r4 = lp_build_extract_range(gallivm, newcoords[2], 4*i, 4);
1972
1973 if (offsets[0]) {
1974 offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
1975 if (dims > 1) {
1976 offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
1977 if (dims > 2) {
1978 offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
1979 }
1980 }
1981 }
1982 lod_ipart4 = lp_build_extract_range(gallivm, lod_ipart, num_lods * i, num_lods);
1983 ilevel04 = lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
1984 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1985 ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods);
1986 lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods);
1987 }
1988
1989 if (use_aos) {
1990 /* do sampling/filtering with fixed pt arithmetic */
1991 lp_build_sample_aos(&bld4, sampler_index,
1992 s4, t4, r4, offsets4,
1993 lod_ipart4, lod_fpart4,
1994 ilevel04, ilevel14,
1995 texelout4);
1996 }
1997
1998 else {
1999 /* this path is currently unreachable and hence might break easily... */
2000 LLVMValueRef newcoords4[5];
2001 newcoords4[0] = s4;
2002 newcoords4[1] = t4;
2003 newcoords4[2] = r4;
2004 newcoords4[3] = lp_build_extract_range(gallivm, newcoords[3], 4*i, 4);
2005 newcoords4[4] = lp_build_extract_range(gallivm, newcoords[4], 4*i, 4);
2006
2007 lp_build_sample_general(&bld4, sampler_index,
2008 newcoords4, offsets4,
2009 lod_ipart4, lod_fpart4,
2010 ilevel04, ilevel14,
2011 texelout4);
2012 }
2013 for (j = 0; j < 4; j++) {
2014 texelouttmp[j][i] = texelout4[j];
2015 }
2016 }
2017
2018 for (j = 0; j < 4; j++) {
2019 texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
2020 }
2021 }
2022 }
2023
2024 if (target != PIPE_BUFFER) {
2025 apply_sampler_swizzle(&bld, texel_out);
2026 }
2027
2028 /*
2029 * texel type can be a (32bit) int/uint (for pure int formats only),
2030 * however we are expected to always return floats (storage is untyped).
2031 */
2032 if (!bld.texel_type.floating) {
2033 unsigned chan;
2034 for (chan = 0; chan < 4; chan++) {
2035 texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
2036 lp_build_vec_type(gallivm, type), "");
2037 }
2038 }
2039 }
2040
2041 void
2042 lp_build_size_query_soa(struct gallivm_state *gallivm,
2043 const struct lp_static_texture_state *static_state,
2044 struct lp_sampler_dynamic_state *dynamic_state,
2045 struct lp_type int_type,
2046 unsigned texture_unit,
2047 unsigned target,
2048 boolean is_sviewinfo,
2049 enum lp_sampler_lod_property lod_property,
2050 LLVMValueRef explicit_lod,
2051 LLVMValueRef *sizes_out)
2052 {
2053 LLVMValueRef lod, level, size;
2054 LLVMValueRef first_level = NULL;
2055 int dims, i;
2056 boolean has_array;
2057 unsigned num_lods = 1;
2058 struct lp_build_context bld_int_vec4;
2059
2060 /*
2061 * Do some sanity verification about bound texture and shader dcl target.
2062 * Not entirely sure what's possible but assume array/non-array
2063 * always compatible (probably not ok for OpenGL but d3d10 has no
2064 * distinction of arrays at the resource level).
2065 * Everything else looks bogus (though not entirely sure about rect/2d).
2066 * Currently disabled because it causes assertion failures if there's
2067 * nothing bound (or rather a dummy texture, not that this case would
2068 * return the right values).
2069 */
2070 if (0 && static_state->target != target) {
2071 if (static_state->target == PIPE_TEXTURE_1D)
2072 assert(target == PIPE_TEXTURE_1D_ARRAY);
2073 else if (static_state->target == PIPE_TEXTURE_1D_ARRAY)
2074 assert(target == PIPE_TEXTURE_1D);
2075 else if (static_state->target == PIPE_TEXTURE_2D)
2076 assert(target == PIPE_TEXTURE_2D_ARRAY);
2077 else if (static_state->target == PIPE_TEXTURE_2D_ARRAY)
2078 assert(target == PIPE_TEXTURE_2D);
2079 else if (static_state->target == PIPE_TEXTURE_CUBE)
2080 assert(target == PIPE_TEXTURE_CUBE_ARRAY);
2081 else if (static_state->target == PIPE_TEXTURE_CUBE_ARRAY)
2082 assert(target == PIPE_TEXTURE_CUBE);
2083 else
2084 assert(0);
2085 }
2086
2087 dims = texture_dims(target);
2088
2089 switch (target) {
2090 case PIPE_TEXTURE_1D_ARRAY:
2091 case PIPE_TEXTURE_2D_ARRAY:
2092 has_array = TRUE;
2093 break;
2094 default:
2095 has_array = FALSE;
2096 break;
2097 }
2098
2099 assert(!int_type.floating);
2100
2101 lp_build_context_init(&bld_int_vec4, gallivm, lp_type_int_vec(32, 128));
2102
2103 if (explicit_lod) {
2104 /* FIXME: this needs to honor per-element lod */
2105 lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), "");
2106 first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit);
2107 level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level");
2108 lod = lp_build_broadcast_scalar(&bld_int_vec4, level);
2109 } else {
2110 lod = bld_int_vec4.zero;
2111 }
2112
2113 size = bld_int_vec4.undef;
2114
2115 size = LLVMBuildInsertElement(gallivm->builder, size,
2116 dynamic_state->width(dynamic_state, gallivm, texture_unit),
2117 lp_build_const_int32(gallivm, 0), "");
2118
2119 if (dims >= 2) {
2120 size = LLVMBuildInsertElement(gallivm->builder, size,
2121 dynamic_state->height(dynamic_state, gallivm, texture_unit),
2122 lp_build_const_int32(gallivm, 1), "");
2123 }
2124
2125 if (dims >= 3) {
2126 size = LLVMBuildInsertElement(gallivm->builder, size,
2127 dynamic_state->depth(dynamic_state, gallivm, texture_unit),
2128 lp_build_const_int32(gallivm, 2), "");
2129 }
2130
2131 size = lp_build_minify(&bld_int_vec4, size, lod);
2132
2133 if (has_array)
2134 size = LLVMBuildInsertElement(gallivm->builder, size,
2135 dynamic_state->depth(dynamic_state, gallivm, texture_unit),
2136 lp_build_const_int32(gallivm, dims), "");
2137
2138 /*
2139 * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels)
2140 * if level is out of bounds (note this can't cover unbound texture
2141 * here, which also requires returning zero).
2142 */
2143 if (explicit_lod && is_sviewinfo) {
2144 LLVMValueRef last_level, out, out1;
2145 struct lp_build_context leveli_bld;
2146
2147 /* everything is scalar for now */
2148 lp_build_context_init(&leveli_bld, gallivm, lp_type_int_vec(32, 32));
2149 last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit);
2150
2151 out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level);
2152 out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level);
2153 out = lp_build_or(&leveli_bld, out, out1);
2154 if (num_lods == 1) {
2155 out = lp_build_broadcast_scalar(&bld_int_vec4, out);
2156 }
2157 else {
2158 /* TODO */
2159 assert(0);
2160 }
2161 size = lp_build_andnot(&bld_int_vec4, size, out);
2162 }
2163 for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
2164 sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec4.type, int_type,
2165 size,
2166 lp_build_const_int32(gallivm, i));
2167 }
2168 if (is_sviewinfo) {
2169 for (; i < 4; i++) {
2170 sizes_out[i] = lp_build_const_vec(gallivm, int_type, 0.0);
2171 }
2172 }
2173
2174 /*
2175 * if there's no explicit_lod (buffers, rects) queries requiring nr of
2176 * mips would be illegal.
2177 */
2178 if (is_sviewinfo && explicit_lod) {
2179 struct lp_build_context bld_int_scalar;
2180 LLVMValueRef num_levels;
2181 lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
2182
2183 if (static_state->level_zero_only) {
2184 num_levels = bld_int_scalar.one;
2185 }
2186 else {
2187 LLVMValueRef last_level;
2188
2189 last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit);
2190 num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
2191 num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one);
2192 }
2193 sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type),
2194 num_levels);
2195 }
2196 }