1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
30 * Texture sampling -- SoA.
32 * @author Jose Fonseca <jfonseca@vmware.com>
33 * @author Brian Paul <brianp@vmware.com>
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"
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
74 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
75 unsigned sampler_unit,
82 LLVMValueRef y_stride,
83 LLVMValueRef z_stride,
84 LLVMValueRef data_ptr,
85 LLVMValueRef mipoffsets,
86 LLVMValueRef texel_out[4])
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;
94 LLVMValueRef use_border = NULL;
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)) {
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");
107 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
108 static_state->min_img_filter,
109 static_state->mag_img_filter)) {
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);
114 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
115 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
118 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
123 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
124 static_state->min_img_filter,
125 static_state->mag_img_filter)) {
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);
130 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
131 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
134 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
138 /* convert x,y,z coords to linear offset from start of texture, in bytes */
139 lp_build_sample_offset(&bld->int_coord_bld,
141 x, y, z, y_stride, z_stride,
144 offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
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.
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.
156 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
159 lp_build_fetch_rgba_soa(bld->gallivm,
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).
172 * texel = border_color;
175 * texel = sample_texture(coord);
177 * As it is now, we always sample the texture, then selectively replace
178 * the texel color results with the border color.
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);
187 for (chan = 0; chan < 4; chan++) {
188 LLVMValueRef border_chan =
189 lp_build_array_get(bld->gallivm, border_color_ptr,
190 lp_build_const_int32(bld->gallivm, chan));
191 LLVMValueRef border_chan_vec =
192 lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
194 if (!bld->texel_type.floating) {
195 border_chan_vec = LLVMBuildBitCast(builder, border_chan_vec,
196 bld->texel_bld.vec_type, "");
198 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
199 border_chan_vec, texel_out[chan]);
206 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
209 lp_build_coord_mirror(struct lp_build_sample_context *bld,
212 struct lp_build_context *coord_bld = &bld->coord_bld;
213 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
214 LLVMValueRef fract, flr, isOdd;
216 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
218 /* isOdd = flr & 1 */
219 isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
221 /* make coord positive or negative depending on isOdd */
222 coord = lp_build_set_sign(coord_bld, fract, isOdd);
224 /* convert isOdd to float */
225 isOdd = lp_build_int_to_float(coord_bld, isOdd);
227 /* add isOdd to coord */
228 coord = lp_build_add(coord_bld, coord, isOdd);
235 * Helper to compute the first coord and the weight for
236 * linear wrap repeat npot textures
239 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
240 LLVMValueRef coord_f,
241 LLVMValueRef length_i,
242 LLVMValueRef length_f,
243 LLVMValueRef *coord0_i,
244 LLVMValueRef *weight_f)
246 struct lp_build_context *coord_bld = &bld->coord_bld;
247 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
248 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
249 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
252 /* wrap with normalized floats is just fract */
253 coord_f = lp_build_fract(coord_bld, coord_f);
254 /* mul by size and subtract 0.5 */
255 coord_f = lp_build_mul(coord_bld, coord_f, length_f);
256 coord_f = lp_build_sub(coord_bld, coord_f, half);
258 * we avoided the 0.5/length division before the repeat wrap,
259 * now need to fix up edge cases with selects
261 /* convert to int, compute lerp weight */
262 lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
263 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
264 PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
265 *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
270 * Build LLVM code for texture wrap mode for linear filtering.
271 * \param x0_out returns first integer texcoord
272 * \param x1_out returns second integer texcoord
273 * \param weight_out returns linear interpolation weight
276 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
279 LLVMValueRef length_f,
283 LLVMValueRef *x0_out,
284 LLVMValueRef *x1_out,
285 LLVMValueRef *weight_out)
287 struct lp_build_context *coord_bld = &bld->coord_bld;
288 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
289 LLVMBuilderRef builder = bld->gallivm->builder;
290 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
291 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
292 LLVMValueRef coord0, coord1, weight;
295 case PIPE_TEX_WRAP_REPEAT:
297 /* mul by size and subtract 0.5 */
298 coord = lp_build_mul(coord_bld, coord, length_f);
299 coord = lp_build_sub(coord_bld, coord, half);
301 offset = lp_build_int_to_float(coord_bld, offset);
302 coord = lp_build_add(coord_bld, coord, offset);
304 /* convert to int, compute lerp weight */
305 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
306 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
308 coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
309 coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
314 offset = lp_build_int_to_float(coord_bld, offset);
315 offset = lp_build_div(coord_bld, offset, length_f);
316 coord = lp_build_add(coord_bld, coord, offset);
318 lp_build_coord_repeat_npot_linear(bld, coord,
321 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
322 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
323 coord1 = LLVMBuildAnd(builder,
324 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
329 case PIPE_TEX_WRAP_CLAMP:
330 if (bld->static_sampler_state->normalized_coords) {
331 /* scale coord to length */
332 coord = lp_build_mul(coord_bld, coord, length_f);
335 offset = lp_build_int_to_float(coord_bld, offset);
336 coord = lp_build_add(coord_bld, coord, offset);
339 /* clamp to [0, length] */
340 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
342 coord = lp_build_sub(coord_bld, coord, half);
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);
349 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
351 struct lp_build_context abs_coord_bld = bld->coord_bld;
352 abs_coord_bld.type.sign = FALSE;
354 if (bld->static_sampler_state->normalized_coords) {
355 /* mul by tex size */
356 coord = lp_build_mul(coord_bld, coord, length_f);
359 offset = lp_build_int_to_float(coord_bld, offset);
360 coord = lp_build_add(coord_bld, coord, offset);
363 /* clamp to length max */
364 coord = lp_build_min(coord_bld, coord, length_f);
366 coord = lp_build_sub(coord_bld, coord, half);
367 /* clamp to [0, length - 0.5] */
368 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
369 /* convert to int, compute lerp weight */
370 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
371 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
372 /* coord1 = min(coord1, length-1) */
373 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
377 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
378 if (bld->static_sampler_state->normalized_coords) {
379 /* scale coord to length */
380 coord = lp_build_mul(coord_bld, coord, length_f);
383 offset = lp_build_int_to_float(coord_bld, offset);
384 coord = lp_build_add(coord_bld, coord, offset);
386 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
387 /* can skip clamp (though might not work for very large coord values */
388 coord = lp_build_sub(coord_bld, coord, half);
389 /* convert to int, compute lerp weight */
390 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
391 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
394 case PIPE_TEX_WRAP_MIRROR_REPEAT:
395 /* compute mirror function */
396 coord = lp_build_coord_mirror(bld, coord);
398 /* scale coord to length */
399 coord = lp_build_mul(coord_bld, coord, length_f);
400 coord = lp_build_sub(coord_bld, coord, half);
402 offset = lp_build_int_to_float(coord_bld, offset);
403 coord = lp_build_add(coord_bld, coord, offset);
406 /* convert to int, compute lerp weight */
407 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
408 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
410 /* coord0 = max(coord0, 0) */
411 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
412 /* coord1 = min(coord1, length-1) */
413 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
416 case PIPE_TEX_WRAP_MIRROR_CLAMP:
417 if (bld->static_sampler_state->normalized_coords) {
418 /* scale coord to length */
419 coord = lp_build_mul(coord_bld, coord, length_f);
422 offset = lp_build_int_to_float(coord_bld, offset);
423 coord = lp_build_add(coord_bld, coord, offset);
425 coord = lp_build_abs(coord_bld, coord);
427 /* clamp to [0, length] */
428 coord = lp_build_min(coord_bld, coord, length_f);
430 coord = lp_build_sub(coord_bld, coord, half);
432 /* convert to int, compute lerp weight */
433 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
434 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
437 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
439 LLVMValueRef min, max;
440 struct lp_build_context abs_coord_bld = bld->coord_bld;
441 abs_coord_bld.type.sign = FALSE;
443 if (bld->static_sampler_state->normalized_coords) {
444 /* scale coord to length */
445 coord = lp_build_mul(coord_bld, coord, length_f);
448 offset = lp_build_int_to_float(coord_bld, offset);
449 coord = lp_build_add(coord_bld, coord, offset);
451 coord = lp_build_abs(coord_bld, coord);
453 /* clamp to [0.5, length - 0.5] */
455 max = lp_build_sub(coord_bld, length_f, min);
456 coord = lp_build_clamp(coord_bld, coord, min, max);
458 coord = lp_build_sub(coord_bld, coord, half);
460 /* convert to int, compute lerp weight */
461 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
462 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
466 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
468 if (bld->static_sampler_state->normalized_coords) {
469 /* scale coord to length */
470 coord = lp_build_mul(coord_bld, coord, length_f);
473 offset = lp_build_int_to_float(coord_bld, offset);
474 coord = lp_build_add(coord_bld, coord, offset);
476 coord = lp_build_abs(coord_bld, coord);
478 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
479 /* skip clamp - always positive, and other side
480 only potentially matters for very large coords */
481 coord = lp_build_sub(coord_bld, coord, half);
483 /* convert to int, compute lerp weight */
484 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
485 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
498 *weight_out = weight;
503 * Build LLVM code for texture wrap mode for nearest filtering.
504 * \param coord the incoming texcoord (nominally in [0,1])
505 * \param length the texture size along one dimension, as int vector
506 * \param length_f the texture size along one dimension, as float vector
507 * \param offset texel offset along one dimension (as int vector)
508 * \param is_pot if TRUE, length is a power of two
509 * \param wrap_mode one of PIPE_TEX_WRAP_x
512 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
515 LLVMValueRef length_f,
520 struct lp_build_context *coord_bld = &bld->coord_bld;
521 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
522 LLVMBuilderRef builder = bld->gallivm->builder;
523 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
527 case PIPE_TEX_WRAP_REPEAT:
529 coord = lp_build_mul(coord_bld, coord, length_f);
530 icoord = lp_build_ifloor(coord_bld, coord);
532 icoord = lp_build_add(int_coord_bld, icoord, offset);
534 icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
538 offset = lp_build_int_to_float(coord_bld, offset);
539 offset = lp_build_div(coord_bld, offset, length_f);
540 coord = lp_build_add(coord_bld, coord, offset);
542 /* take fraction, unnormalize */
543 coord = lp_build_fract_safe(coord_bld, coord);
544 coord = lp_build_mul(coord_bld, coord, length_f);
545 icoord = lp_build_itrunc(coord_bld, coord);
549 case PIPE_TEX_WRAP_CLAMP:
550 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
551 if (bld->static_sampler_state->normalized_coords) {
552 /* scale coord to length */
553 coord = lp_build_mul(coord_bld, coord, length_f);
557 /* use itrunc instead since we clamp to 0 anyway */
558 icoord = lp_build_itrunc(coord_bld, coord);
560 icoord = lp_build_add(int_coord_bld, icoord, offset);
563 /* clamp to [0, length - 1]. */
564 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
568 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
569 if (bld->static_sampler_state->normalized_coords) {
570 /* scale coord to length */
571 coord = lp_build_mul(coord_bld, coord, length_f);
573 /* no clamp necessary, border masking will handle this */
574 icoord = lp_build_ifloor(coord_bld, coord);
576 icoord = lp_build_add(int_coord_bld, icoord, offset);
580 case PIPE_TEX_WRAP_MIRROR_REPEAT:
582 offset = lp_build_int_to_float(coord_bld, offset);
583 offset = lp_build_div(coord_bld, offset, length_f);
584 coord = lp_build_add(coord_bld, coord, offset);
586 /* compute mirror function */
587 coord = lp_build_coord_mirror(bld, coord);
589 /* scale coord to length */
590 assert(bld->static_sampler_state->normalized_coords);
591 coord = lp_build_mul(coord_bld, coord, length_f);
593 /* itrunc == ifloor here */
594 icoord = lp_build_itrunc(coord_bld, coord);
596 /* clamp to [0, length - 1] */
597 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
600 case PIPE_TEX_WRAP_MIRROR_CLAMP:
601 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
602 if (bld->static_sampler_state->normalized_coords) {
603 /* scale coord to length */
604 coord = lp_build_mul(coord_bld, coord, length_f);
607 offset = lp_build_int_to_float(coord_bld, offset);
608 coord = lp_build_add(coord_bld, coord, offset);
610 coord = lp_build_abs(coord_bld, coord);
612 /* itrunc == ifloor here */
613 icoord = lp_build_itrunc(coord_bld, coord);
615 /* clamp to [0, length - 1] */
616 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
619 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
620 if (bld->static_sampler_state->normalized_coords) {
621 /* scale coord to length */
622 coord = lp_build_mul(coord_bld, coord, length_f);
625 offset = lp_build_int_to_float(coord_bld, offset);
626 coord = lp_build_add(coord_bld, coord, offset);
628 coord = lp_build_abs(coord_bld, coord);
630 /* itrunc == ifloor here */
631 icoord = lp_build_itrunc(coord_bld, coord);
644 * Generate code to sample a mipmap level with nearest filtering.
645 * If sampling a cube texture, r = cube face in [0,5].
648 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
649 unsigned sampler_unit,
651 LLVMValueRef row_stride_vec,
652 LLVMValueRef img_stride_vec,
653 LLVMValueRef data_ptr,
654 LLVMValueRef mipoffsets,
658 const LLVMValueRef *offsets,
659 LLVMValueRef colors_out[4])
661 const unsigned dims = bld->dims;
662 LLVMValueRef width_vec;
663 LLVMValueRef height_vec;
664 LLVMValueRef depth_vec;
665 LLVMValueRef flt_size;
666 LLVMValueRef flt_width_vec;
667 LLVMValueRef flt_height_vec;
668 LLVMValueRef flt_depth_vec;
669 LLVMValueRef x, y = NULL, z = NULL;
671 lp_build_extract_image_sizes(bld,
675 &width_vec, &height_vec, &depth_vec);
677 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
679 lp_build_extract_image_sizes(bld,
680 &bld->float_size_bld,
683 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
686 * Compute integer texcoords.
688 x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, offsets[0],
689 bld->static_texture_state->pot_width,
690 bld->static_sampler_state->wrap_s);
691 lp_build_name(x, "tex.x.wrapped");
694 y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, offsets[1],
695 bld->static_texture_state->pot_height,
696 bld->static_sampler_state->wrap_t);
697 lp_build_name(y, "tex.y.wrapped");
700 z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, offsets[2],
701 bld->static_texture_state->pot_depth,
702 bld->static_sampler_state->wrap_r);
703 lp_build_name(z, "tex.z.wrapped");
706 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
707 bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
708 bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
710 lp_build_name(z, "tex.z.layer");
714 * Get texture colors.
716 lp_build_sample_texel_soa(bld, sampler_unit,
717 width_vec, height_vec, depth_vec,
719 row_stride_vec, img_stride_vec,
720 data_ptr, mipoffsets, colors_out);
725 * Generate code to sample a mipmap level with linear filtering.
726 * If sampling a cube texture, r = cube face in [0,5].
729 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
730 unsigned sampler_unit,
732 LLVMValueRef row_stride_vec,
733 LLVMValueRef img_stride_vec,
734 LLVMValueRef data_ptr,
735 LLVMValueRef mipoffsets,
739 const LLVMValueRef *offsets,
740 LLVMValueRef colors_out[4])
742 const unsigned dims = bld->dims;
743 LLVMValueRef width_vec;
744 LLVMValueRef height_vec;
745 LLVMValueRef depth_vec;
746 LLVMValueRef flt_size;
747 LLVMValueRef flt_width_vec;
748 LLVMValueRef flt_height_vec;
749 LLVMValueRef flt_depth_vec;
750 LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
751 LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
752 LLVMValueRef neighbors[2][2][4];
755 lp_build_extract_image_sizes(bld,
759 &width_vec, &height_vec, &depth_vec);
761 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
763 lp_build_extract_image_sizes(bld,
764 &bld->float_size_bld,
767 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
770 * Compute integer texcoords.
772 lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec, offsets[0],
773 bld->static_texture_state->pot_width,
774 bld->static_sampler_state->wrap_s,
776 lp_build_name(x0, "tex.x0.wrapped");
777 lp_build_name(x1, "tex.x1.wrapped");
780 lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec, offsets[1],
781 bld->static_texture_state->pot_height,
782 bld->static_sampler_state->wrap_t,
784 lp_build_name(y0, "tex.y0.wrapped");
785 lp_build_name(y1, "tex.y1.wrapped");
788 lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec, offsets[2],
789 bld->static_texture_state->pot_depth,
790 bld->static_sampler_state->wrap_r,
792 lp_build_name(z0, "tex.z0.wrapped");
793 lp_build_name(z1, "tex.z1.wrapped");
796 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
797 bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
798 bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
799 z0 = z1 = r; /* cube face or array layer */
800 lp_build_name(z0, "tex.z0.layer");
801 lp_build_name(z1, "tex.z1.layer");
806 * Get texture colors.
808 /* get x0/x1 texels */
809 lp_build_sample_texel_soa(bld, sampler_unit,
810 width_vec, height_vec, depth_vec,
812 row_stride_vec, img_stride_vec,
813 data_ptr, mipoffsets, neighbors[0][0]);
814 lp_build_sample_texel_soa(bld, sampler_unit,
815 width_vec, height_vec, depth_vec,
817 row_stride_vec, img_stride_vec,
818 data_ptr, mipoffsets, neighbors[0][1]);
821 /* Interpolate two samples from 1D image to produce one color */
822 for (chan = 0; chan < 4; chan++) {
823 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
824 neighbors[0][0][chan],
825 neighbors[0][1][chan]);
830 LLVMValueRef colors0[4];
832 /* get x0/x1 texels at y1 */
833 lp_build_sample_texel_soa(bld, sampler_unit,
834 width_vec, height_vec, depth_vec,
836 row_stride_vec, img_stride_vec,
837 data_ptr, mipoffsets, neighbors[1][0]);
838 lp_build_sample_texel_soa(bld, sampler_unit,
839 width_vec, height_vec, depth_vec,
841 row_stride_vec, img_stride_vec,
842 data_ptr, mipoffsets, neighbors[1][1]);
844 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
845 for (chan = 0; chan < 4; chan++) {
846 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
848 neighbors[0][0][chan],
849 neighbors[0][1][chan],
850 neighbors[1][0][chan],
851 neighbors[1][1][chan]);
855 LLVMValueRef neighbors1[2][2][4];
856 LLVMValueRef colors1[4];
858 /* get x0/x1/y0/y1 texels at z1 */
859 lp_build_sample_texel_soa(bld, sampler_unit,
860 width_vec, height_vec, depth_vec,
862 row_stride_vec, img_stride_vec,
863 data_ptr, mipoffsets, neighbors1[0][0]);
864 lp_build_sample_texel_soa(bld, sampler_unit,
865 width_vec, height_vec, depth_vec,
867 row_stride_vec, img_stride_vec,
868 data_ptr, mipoffsets, neighbors1[0][1]);
869 lp_build_sample_texel_soa(bld, sampler_unit,
870 width_vec, height_vec, depth_vec,
872 row_stride_vec, img_stride_vec,
873 data_ptr, mipoffsets, neighbors1[1][0]);
874 lp_build_sample_texel_soa(bld, sampler_unit,
875 width_vec, height_vec, depth_vec,
877 row_stride_vec, img_stride_vec,
878 data_ptr, mipoffsets, neighbors1[1][1]);
880 /* Bilinear interpolate the four samples from the second Z slice */
881 for (chan = 0; chan < 4; chan++) {
882 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
884 neighbors1[0][0][chan],
885 neighbors1[0][1][chan],
886 neighbors1[1][0][chan],
887 neighbors1[1][1][chan]);
890 /* Linearly interpolate the two samples from the two 3D slices */
891 for (chan = 0; chan < 4; chan++) {
892 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
894 colors0[chan], colors1[chan]);
899 for (chan = 0; chan < 4; chan++) {
900 colors_out[chan] = colors0[chan];
908 * Sample the texture/mipmap using given image filter and mip filter.
909 * data0_ptr and data1_ptr point to the two mipmap levels to sample
910 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
911 * If we're using nearest miplevel sampling the '1' values will be null/unused.
914 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
915 unsigned sampler_unit,
921 const LLVMValueRef *offsets,
922 LLVMValueRef ilevel0,
923 LLVMValueRef ilevel1,
924 LLVMValueRef lod_fpart,
925 LLVMValueRef *colors_out)
927 LLVMBuilderRef builder = bld->gallivm->builder;
928 LLVMValueRef size0 = NULL;
929 LLVMValueRef size1 = NULL;
930 LLVMValueRef row_stride0_vec = NULL;
931 LLVMValueRef row_stride1_vec = NULL;
932 LLVMValueRef img_stride0_vec = NULL;
933 LLVMValueRef img_stride1_vec = NULL;
934 LLVMValueRef data_ptr0 = NULL;
935 LLVMValueRef data_ptr1 = NULL;
936 LLVMValueRef mipoff0 = NULL;
937 LLVMValueRef mipoff1 = NULL;
938 LLVMValueRef colors0[4], colors1[4];
941 /* sample the first mipmap level */
942 lp_build_mipmap_level_sizes(bld, ilevel0,
944 &row_stride0_vec, &img_stride0_vec);
945 if (bld->num_lods == 1) {
946 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
949 /* This path should work for num_lods 1 too but slightly less efficient */
950 data_ptr0 = bld->base_ptr;
951 mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
953 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
954 lp_build_sample_image_nearest(bld, sampler_unit,
956 row_stride0_vec, img_stride0_vec,
957 data_ptr0, mipoff0, s, t, r, offsets,
961 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
962 lp_build_sample_image_linear(bld, sampler_unit,
964 row_stride0_vec, img_stride0_vec,
965 data_ptr0, mipoff0, s, t, r, offsets,
969 /* Store the first level's colors in the output variables */
970 for (chan = 0; chan < 4; chan++) {
971 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
974 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
975 struct lp_build_if_state if_ctx;
976 LLVMValueRef need_lerp;
977 unsigned num_quads = bld->coord_bld.type.length / 4;
979 /* need_lerp = lod_fpart > 0 */
980 if (num_quads == 1) {
981 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
982 lod_fpart, bld->perquadf_bld.zero,
987 * We'll do mip filtering if any of the quads need it.
988 * It might be better to split the vectors here and only fetch/filter
989 * quads which need it.
992 * We unfortunately need to clamp lod_fpart here since we can get
993 * negative values which would screw up filtering if not all
994 * lod_fpart values have same sign.
996 lod_fpart = lp_build_max(&bld->perquadf_bld, lod_fpart,
997 bld->perquadf_bld.zero);
998 need_lerp = lp_build_compare(bld->gallivm, bld->perquadf_bld.type,
1000 lod_fpart, bld->perquadf_bld.zero);
1001 need_lerp = lp_build_any_true_range(&bld->perquadi_bld, num_quads, need_lerp);
1004 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1006 /* sample the second mipmap level */
1007 lp_build_mipmap_level_sizes(bld, ilevel1,
1009 &row_stride1_vec, &img_stride1_vec);
1010 if (bld->num_lods == 1) {
1011 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1014 data_ptr1 = bld->base_ptr;
1015 mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1017 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1018 lp_build_sample_image_nearest(bld, sampler_unit,
1020 row_stride1_vec, img_stride1_vec,
1021 data_ptr1, mipoff1, s, t, r, offsets,
1025 lp_build_sample_image_linear(bld, sampler_unit,
1027 row_stride1_vec, img_stride1_vec,
1028 data_ptr1, mipoff1, s, t, r, offsets,
1032 /* interpolate samples from the two mipmap levels */
1034 lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1035 bld->perquadf_bld.type,
1036 bld->texel_bld.type,
1039 for (chan = 0; chan < 4; chan++) {
1040 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1041 colors0[chan], colors1[chan]);
1042 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1045 lp_build_endif(&if_ctx);
1051 * Clamp layer coord to valid values.
1054 lp_build_layer_coord(struct lp_build_sample_context *bld,
1055 unsigned texture_unit,
1058 LLVMValueRef maxlayer;
1060 maxlayer = bld->dynamic_state->depth(bld->dynamic_state,
1061 bld->gallivm, texture_unit);
1062 maxlayer = lp_build_sub(&bld->int_bld, maxlayer, bld->int_bld.one);
1063 maxlayer = lp_build_broadcast_scalar(&bld->int_coord_bld, maxlayer);
1064 return lp_build_clamp(&bld->int_coord_bld, layer,
1065 bld->int_coord_bld.zero, maxlayer);
1071 * Calculate cube face, lod, mip levels.
1074 lp_build_sample_common(struct lp_build_sample_context *bld,
1075 unsigned texture_index,
1076 unsigned sampler_index,
1080 const struct lp_derivatives *derivs, /* optional */
1081 LLVMValueRef lod_bias, /* optional */
1082 LLVMValueRef explicit_lod, /* optional */
1083 LLVMValueRef *lod_ipart,
1084 LLVMValueRef *lod_fpart,
1085 LLVMValueRef *ilevel0,
1086 LLVMValueRef *ilevel1)
1088 const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1089 const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1090 const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1091 const unsigned target = bld->static_texture_state->target;
1092 LLVMValueRef first_level, cube_rho = NULL;
1095 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1096 mip_filter, min_filter, mag_filter);
1100 * Choose cube face, recompute texcoords for the chosen face and
1101 * compute rho here too (as it requires transform of derivatives).
1103 if (target == PIPE_TEXTURE_CUBE) {
1104 LLVMValueRef face, face_s, face_t;
1105 boolean need_derivs;
1106 need_derivs = ((min_filter != mag_filter ||
1107 mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
1108 !bld->static_sampler_state->min_max_lod_equal &&
1110 lp_build_cube_lookup(bld, *s, *t, *r, derivs, &face, &face_s, &face_t,
1111 &cube_rho, need_derivs);
1112 *s = face_s; /* vec */
1113 *t = face_t; /* vec */
1114 /* use 'r' to indicate cube face */
1115 *r = face; /* vec */
1117 else if (target == PIPE_TEXTURE_1D_ARRAY) {
1118 *r = lp_build_iround(&bld->coord_bld, *t);
1119 *r = lp_build_layer_coord(bld, texture_index, *r);
1121 else if (target == PIPE_TEXTURE_2D_ARRAY) {
1122 *r = lp_build_iround(&bld->coord_bld, *r);
1123 *r = lp_build_layer_coord(bld, texture_index, *r);
1127 * Compute the level of detail (float).
1129 if (min_filter != mag_filter ||
1130 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1131 /* Need to compute lod either to choose mipmap levels or to
1132 * distinguish between minification/magnification with one mipmap level.
1134 lp_build_lod_selector(bld, texture_index, sampler_index,
1135 *s, *t, *r, cube_rho,
1136 derivs, lod_bias, explicit_lod,
1138 lod_ipart, lod_fpart);
1140 *lod_ipart = bld->perquadi_bld.zero;
1144 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1146 switch (mip_filter) {
1148 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1150 case PIPE_TEX_MIPFILTER_NONE:
1151 /* always use mip level 0 */
1152 if (target == PIPE_TEXTURE_CUBE) {
1153 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1154 * We should be able to set ilevel0 = const(0) but that causes
1155 * bad x86 code to be emitted.
1156 * XXX should probably disable that on other llvm versions.
1159 lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1162 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1163 bld->gallivm, texture_index);
1164 first_level = lp_build_broadcast_scalar(&bld->perquadi_bld, first_level);
1165 *ilevel0 = first_level;
1168 case PIPE_TEX_MIPFILTER_NEAREST:
1170 lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1172 case PIPE_TEX_MIPFILTER_LINEAR:
1175 lp_build_linear_mip_levels(bld, texture_index,
1176 *lod_ipart, lod_fpart,
1183 * General texture sampling codegen.
1184 * This function handles texture sampling for all texture targets (1D,
1185 * 2D, 3D, cube) and all filtering modes.
1188 lp_build_sample_general(struct lp_build_sample_context *bld,
1189 unsigned sampler_unit,
1193 const LLVMValueRef *offsets,
1194 LLVMValueRef lod_ipart,
1195 LLVMValueRef lod_fpart,
1196 LLVMValueRef ilevel0,
1197 LLVMValueRef ilevel1,
1198 LLVMValueRef *colors_out)
1200 struct lp_build_context *int_bld = &bld->int_bld;
1201 LLVMBuilderRef builder = bld->gallivm->builder;
1202 const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1203 const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1204 const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1205 LLVMValueRef texels[4];
1209 * Get/interpolate texture colors.
1212 for (chan = 0; chan < 4; ++chan) {
1213 texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1214 lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
1217 if (min_filter == mag_filter) {
1218 /* no need to distinguish between minification and magnification */
1219 lp_build_sample_mipmap(bld, sampler_unit,
1220 min_filter, mip_filter,
1222 ilevel0, ilevel1, lod_fpart,
1226 /* Emit conditional to choose min image filter or mag image filter
1227 * depending on the lod being > 0 or <= 0, respectively.
1229 struct lp_build_if_state if_ctx;
1230 LLVMValueRef minify;
1233 * XXX this should to all lods into account, if some are min
1234 * some max probably could hack up the coords/weights in the linear
1235 * path with selects to work for nearest.
1236 * If that's just two quads sitting next to each other it seems
1237 * quite ok to do the same filtering method on both though, at
1238 * least unless we have explicit lod (and who uses different
1239 * min/mag filter with that?)
1241 if (bld->num_lods > 1)
1242 lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
1243 lp_build_const_int32(bld->gallivm, 0), "");
1245 /* minify = lod >= 0.0 */
1246 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1247 lod_ipart, int_bld->zero, "");
1249 lp_build_if(&if_ctx, bld->gallivm, minify);
1251 /* Use the minification filter */
1252 lp_build_sample_mipmap(bld, sampler_unit,
1253 min_filter, mip_filter,
1255 ilevel0, ilevel1, lod_fpart,
1258 lp_build_else(&if_ctx);
1260 /* Use the magnification filter */
1261 lp_build_sample_mipmap(bld, sampler_unit,
1262 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1264 ilevel0, NULL, NULL,
1267 lp_build_endif(&if_ctx);
1270 for (chan = 0; chan < 4; ++chan) {
1271 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1272 lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
1278 * Texel fetch function.
1279 * In contrast to general sampling there is no filtering, no coord minification,
1280 * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
1281 * directly to be applied to the selected mip level (after adding texel offsets).
1282 * This function handles texel fetch for all targets where texel fetch is supported
1283 * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
1286 lp_build_fetch_texel(struct lp_build_sample_context *bld,
1287 unsigned texture_unit,
1288 const LLVMValueRef *coords,
1289 LLVMValueRef explicit_lod,
1290 const LLVMValueRef *offsets,
1291 LLVMValueRef *colors_out)
1293 struct lp_build_context *perquadi_bld = &bld->perquadi_bld;
1294 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1295 unsigned dims = bld->dims, chan;
1296 unsigned target = bld->static_texture_state->target;
1297 LLVMValueRef size, ilevel;
1298 LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
1299 LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
1300 LLVMValueRef width, height, depth, i, j;
1301 LLVMValueRef offset, out_of_bounds, out1;
1303 /* XXX just like ordinary sampling, we don't handle per-pixel lod (yet). */
1304 if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
1305 ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
1306 perquadi_bld->type, explicit_lod, 0);
1307 lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel);
1311 ilevel = lp_build_const_int32(bld->gallivm, 0);
1313 lp_build_mipmap_level_sizes(bld, ilevel,
1315 &row_stride_vec, &img_stride_vec);
1316 lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
1317 size, &width, &height, &depth);
1319 if (target == PIPE_TEXTURE_1D_ARRAY ||
1320 target == PIPE_TEXTURE_2D_ARRAY) {
1321 if (target == PIPE_TEXTURE_1D_ARRAY) {
1322 z = lp_build_layer_coord(bld, texture_unit, y);
1325 z = lp_build_layer_coord(bld, texture_unit, z);
1329 /* This is a lot like border sampling */
1331 /* XXX coords are really unsigned, offsets are signed */
1332 x = lp_build_add(int_coord_bld, x, offsets[0]);
1334 out_of_bounds = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
1335 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
1336 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1340 y = lp_build_add(int_coord_bld, y, offsets[1]);
1342 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
1343 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1344 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
1345 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1349 z = lp_build_add(int_coord_bld, z, offsets[2]);
1351 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
1352 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1353 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
1354 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1358 lp_build_sample_offset(int_coord_bld,
1360 x, y, z, row_stride_vec, img_stride_vec,
1363 if (bld->static_texture_state->target != PIPE_BUFFER) {
1364 offset = lp_build_add(int_coord_bld, offset,
1365 lp_build_get_mip_offsets(bld, ilevel));
1368 offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
1370 lp_build_fetch_rgba_soa(bld->gallivm,
1373 bld->base_ptr, offset,
1379 * Not needed except for ARB_robust_buffer_access_behavior.
1380 * Could use min/max above instead of out-of-bounds comparisons
1381 * (in fact cast to unsigned and min only is sufficient)
1382 * if we don't care about the result returned for out-of-bounds.
1384 for (chan = 0; chan < 4; chan++) {
1385 colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
1386 bld->texel_bld.zero, colors_out[chan]);
1393 * Do shadow test/comparison.
1394 * \param coords incoming texcoords
1395 * \param texel the texel to compare against (use the X channel)
1396 * Ideally this should really be done per-sample.
1399 lp_build_sample_compare(struct lp_build_sample_context *bld,
1400 const LLVMValueRef *coords,
1401 LLVMValueRef texel[4])
1403 struct lp_build_context *texel_bld = &bld->texel_bld;
1404 LLVMBuilderRef builder = bld->gallivm->builder;
1405 LLVMValueRef res, p;
1406 const unsigned chan = 0;
1408 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1411 if (bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY ||
1412 bld->static_texture_state->target == PIPE_TEXTURE_CUBE) {
1421 LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
1422 LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
1423 LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
1424 lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
1428 /* Clamp p coords to [0,1] */
1429 p = lp_build_clamp(&bld->coord_bld, p,
1430 bld->coord_bld.zero,
1431 bld->coord_bld.one);
1433 /* result = (p FUNC texel) ? 1 : 0 */
1434 res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
1436 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1438 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1442 texel[3] = texel_bld->one;
1447 * Just set texels to white instead of actually sampling the texture.
1451 lp_build_sample_nop(struct gallivm_state *gallivm,
1452 struct lp_type type,
1453 const LLVMValueRef *coords,
1454 LLVMValueRef texel_out[4])
1456 LLVMValueRef one = lp_build_one(gallivm, type);
1459 for (chan = 0; chan < 4; chan++) {
1460 texel_out[chan] = one;
1466 * Build texture sampling code.
1467 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1469 * \param type vector float type to use for coords, etc.
1470 * \param is_fetch if this is a texel fetch instruction.
1471 * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
1474 lp_build_sample_soa(struct gallivm_state *gallivm,
1475 const struct lp_static_texture_state *static_texture_state,
1476 const struct lp_static_sampler_state *static_sampler_state,
1477 struct lp_sampler_dynamic_state *dynamic_state,
1478 struct lp_type type,
1480 unsigned texture_index,
1481 unsigned sampler_index,
1482 const LLVMValueRef *coords,
1483 const LLVMValueRef *offsets,
1484 const struct lp_derivatives *derivs, /* optional */
1485 LLVMValueRef lod_bias, /* optional */
1486 LLVMValueRef explicit_lod, /* optional */
1487 LLVMValueRef texel_out[4])
1489 unsigned dims = texture_dims(static_texture_state->target);
1490 unsigned num_quads = type.length / 4;
1491 unsigned mip_filter;
1492 struct lp_build_sample_context bld;
1493 struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
1494 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1495 LLVMBuilderRef builder = gallivm->builder;
1496 LLVMValueRef tex_width;
1502 enum pipe_format fmt = static_texture_state->format;
1503 debug_printf("Sample from %s\n", util_format_name(fmt));
1506 assert(type.floating);
1508 /* Setup our build context */
1509 memset(&bld, 0, sizeof bld);
1510 bld.gallivm = gallivm;
1511 bld.static_sampler_state = &derived_sampler_state;
1512 bld.static_texture_state = static_texture_state;
1513 bld.dynamic_state = dynamic_state;
1514 bld.format_desc = util_format_description(static_texture_state->format);
1517 bld.vector_width = lp_type_width(type);
1519 bld.float_type = lp_type_float(32);
1520 bld.int_type = lp_type_int(32);
1521 bld.coord_type = type;
1522 bld.int_coord_type = lp_int_type(type);
1523 bld.float_size_in_type = lp_type_float(32);
1524 bld.float_size_in_type.length = dims > 1 ? 4 : 1;
1525 bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
1526 bld.texel_type = type;
1527 bld.perquadf_type = type;
1528 /* we want native vector size to be able to use our intrinsics */
1529 bld.perquadf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
1530 bld.perquadi_type = lp_int_type(bld.perquadf_type);
1532 /* always using the first channel hopefully should be safe,
1533 * if not things WILL break in other places anyway.
1535 if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1536 bld.format_desc->channel[0].pure_integer) {
1537 if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1538 bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1540 else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1541 bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
1545 if (!static_texture_state->level_zero_only) {
1546 derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
1548 derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
1550 mip_filter = derived_sampler_state.min_mip_filter;
1553 debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
1557 * There are other situations where at least the multiple int lods could be
1558 * avoided like min and max lod being equal.
1560 if ((is_fetch && explicit_lod && bld.static_texture_state->target != PIPE_BUFFER) ||
1561 (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
1562 bld.num_lods = num_quads;
1568 bld.float_size_type = bld.float_size_in_type;
1569 bld.float_size_type.length = bld.num_lods > 1 ? type.length :
1570 bld.float_size_in_type.length;
1571 bld.int_size_type = lp_int_type(bld.float_size_type);
1573 lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1574 lp_build_context_init(&bld.float_vec_bld, gallivm, type);
1575 lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1576 lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1577 lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1578 lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
1579 lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
1580 lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1581 lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1582 lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1583 lp_build_context_init(&bld.perquadf_bld, gallivm, bld.perquadf_type);
1584 lp_build_context_init(&bld.perquadi_bld, gallivm, bld.perquadi_type);
1586 /* Get the dynamic state */
1587 tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index);
1588 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index);
1589 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index);
1590 bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index);
1591 bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index);
1592 /* Note that mip_offsets is an array[level] of offsets to texture images */
1598 /* width, height, depth as single int vector */
1600 bld.int_size = tex_width;
1603 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
1604 tex_width, LLVMConstInt(i32t, 0, 0), "");
1606 LLVMValueRef tex_height =
1607 dynamic_state->height(dynamic_state, gallivm, texture_index);
1608 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1609 tex_height, LLVMConstInt(i32t, 1, 0), "");
1611 LLVMValueRef tex_depth =
1612 dynamic_state->depth(dynamic_state, gallivm, texture_index);
1613 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1614 tex_depth, LLVMConstInt(i32t, 2, 0), "");
1620 /* For debug: no-op texture sampling */
1621 lp_build_sample_nop(gallivm,
1627 else if (is_fetch) {
1628 lp_build_fetch_texel(&bld, texture_index, coords,
1629 explicit_lod, offsets,
1634 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
1635 LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
1636 boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
1637 lp_is_simple_wrap_mode(static_sampler_state->wrap_s) &&
1638 lp_is_simple_wrap_mode(static_sampler_state->wrap_t);
1640 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1641 !use_aos && util_format_fits_8unorm(bld.format_desc)) {
1642 debug_printf("%s: using floating point linear filtering for %s\n",
1643 __FUNCTION__, bld.format_desc->short_name);
1644 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1645 static_sampler_state->min_img_filter,
1646 static_sampler_state->mag_img_filter,
1647 static_sampler_state->min_mip_filter,
1648 static_sampler_state->wrap_s,
1649 static_sampler_state->wrap_t);
1652 lp_build_sample_common(&bld, texture_index, sampler_index,
1654 derivs, lod_bias, explicit_lod,
1655 &lod_ipart, &lod_fpart,
1656 &ilevel0, &ilevel1);
1659 * we only try 8-wide sampling with soa as it appears to
1660 * be a loss with aos with AVX (but it should work).
1661 * (It should be faster if we'd support avx2)
1663 if (num_quads == 1 || !use_aos) {
1665 if (num_quads > 1) {
1666 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1667 LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
1669 * These parameters are the same for all quads,
1670 * could probably simplify.
1672 lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
1673 ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
1677 /* do sampling/filtering with fixed pt arithmetic */
1678 lp_build_sample_aos(&bld, sampler_index,
1680 lod_ipart, lod_fpart,
1686 lp_build_sample_general(&bld, sampler_index,
1688 lod_ipart, lod_fpart,
1695 struct lp_build_sample_context bld4;
1696 struct lp_type type4 = type;
1698 LLVMValueRef texelout4[4];
1699 LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
1703 /* Setup our build context */
1704 memset(&bld4, 0, sizeof bld4);
1705 bld4.gallivm = bld.gallivm;
1706 bld4.static_texture_state = bld.static_texture_state;
1707 bld4.static_sampler_state = bld.static_sampler_state;
1708 bld4.dynamic_state = bld.dynamic_state;
1709 bld4.format_desc = bld.format_desc;
1710 bld4.dims = bld.dims;
1711 bld4.row_stride_array = bld.row_stride_array;
1712 bld4.img_stride_array = bld.img_stride_array;
1713 bld4.base_ptr = bld.base_ptr;
1714 bld4.mip_offsets = bld.mip_offsets;
1715 bld4.int_size = bld.int_size;
1717 bld4.vector_width = lp_type_width(type4);
1719 bld4.float_type = lp_type_float(32);
1720 bld4.int_type = lp_type_int(32);
1721 bld4.coord_type = type4;
1722 bld4.int_coord_type = lp_int_type(type4);
1723 bld4.float_size_in_type = lp_type_float(32);
1724 bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
1725 bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
1726 bld4.texel_type = bld.texel_type;
1727 bld4.texel_type.length = 4;
1728 bld4.perquadf_type = type4;
1729 /* we want native vector size to be able to use our intrinsics */
1730 bld4.perquadf_type.length = 1;
1731 bld4.perquadi_type = lp_int_type(bld4.perquadf_type);
1734 bld4.int_size_type = bld4.int_size_in_type;
1735 bld4.float_size_type = bld4.float_size_in_type;
1737 lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
1738 lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
1739 lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
1740 lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
1741 lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
1742 lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
1743 lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
1744 lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
1745 lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
1746 lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
1747 lp_build_context_init(&bld4.perquadf_bld, gallivm, bld4.perquadf_type);
1748 lp_build_context_init(&bld4.perquadi_bld, gallivm, bld4.perquadi_type);
1750 for (i = 0; i < num_quads; i++) {
1751 LLVMValueRef s4, t4, r4;
1752 LLVMValueRef lod_iparts, lod_fparts = NULL;
1753 LLVMValueRef ilevel0s, ilevel1s = NULL;
1754 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1755 LLVMValueRef offsets4[4] = { NULL };
1757 s4 = lp_build_extract_range(gallivm, s, 4*i, 4);
1758 t4 = lp_build_extract_range(gallivm, t, 4*i, 4);
1759 r4 = lp_build_extract_range(gallivm, r, 4*i, 4);
1762 offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
1764 offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
1766 offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
1770 lod_iparts = LLVMBuildExtractElement(builder, lod_ipart, indexi, "");
1771 ilevel0s = LLVMBuildExtractElement(builder, ilevel0, indexi, "");
1772 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1773 ilevel1s = LLVMBuildExtractElement(builder, ilevel1, indexi, "");
1774 lod_fparts = LLVMBuildExtractElement(builder, lod_fpart, indexi, "");
1778 /* do sampling/filtering with fixed pt arithmetic */
1779 lp_build_sample_aos(&bld4, sampler_index,
1780 s4, t4, r4, offsets4,
1781 lod_iparts, lod_fparts,
1787 lp_build_sample_general(&bld4, sampler_index,
1788 s4, t4, r4, offsets4,
1789 lod_iparts, lod_fparts,
1793 for (j = 0; j < 4; j++) {
1794 texelouttmp[j][i] = texelout4[j];
1798 for (j = 0; j < 4; j++) {
1799 texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
1803 lp_build_sample_compare(&bld, coords, texel_out);
1806 if (static_texture_state->target != PIPE_BUFFER) {
1807 apply_sampler_swizzle(&bld, texel_out);
1811 * texel type can be a (32bit) int/uint (for pure int formats only),
1812 * however we are expected to always return floats (storage is untyped).
1814 if (!bld.texel_type.floating) {
1816 for (chan = 0; chan < 4; chan++) {
1817 texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
1818 lp_build_vec_type(gallivm, type), "");
1824 lp_build_size_query_soa(struct gallivm_state *gallivm,
1825 const struct lp_static_texture_state *static_state,
1826 struct lp_sampler_dynamic_state *dynamic_state,
1827 struct lp_type int_type,
1828 unsigned texture_unit,
1829 boolean need_nr_mips,
1830 LLVMValueRef explicit_lod,
1831 LLVMValueRef *sizes_out)
1835 LLVMValueRef first_level = NULL;
1838 struct lp_build_context bld_int_vec;
1840 dims = texture_dims(static_state->target);
1842 switch (static_state->target) {
1843 case PIPE_TEXTURE_1D_ARRAY:
1844 case PIPE_TEXTURE_2D_ARRAY:
1852 assert(!int_type.floating);
1854 lp_build_context_init(&bld_int_vec, gallivm, lp_type_int_vec(32, 128));
1857 lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), "");
1858 first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit);
1859 lod = lp_build_broadcast_scalar(&bld_int_vec,
1860 LLVMBuildAdd(gallivm->builder, lod, first_level, "lod"));
1863 lod = bld_int_vec.zero;
1867 size = bld_int_vec.zero;
1870 size = bld_int_vec.undef;
1873 size = LLVMBuildInsertElement(gallivm->builder, size,
1874 dynamic_state->width(dynamic_state, gallivm, texture_unit),
1875 lp_build_const_int32(gallivm, 0), "");
1878 size = LLVMBuildInsertElement(gallivm->builder, size,
1879 dynamic_state->height(dynamic_state, gallivm, texture_unit),
1880 lp_build_const_int32(gallivm, 1), "");
1884 size = LLVMBuildInsertElement(gallivm->builder, size,
1885 dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1886 lp_build_const_int32(gallivm, 2), "");
1889 size = lp_build_minify(&bld_int_vec, size, lod);
1892 size = LLVMBuildInsertElement(gallivm->builder, size,
1893 dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1894 lp_build_const_int32(gallivm, dims), "");
1897 * XXX for out-of-bounds lod, should set size to zero vector here
1898 * (for dx10-style only, i.e. need_nr_mips)
1901 for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
1902 sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec.type, int_type,
1904 lp_build_const_int32(gallivm, i));
1908 * if there's no explicit_lod (buffers, rects) queries requiring nr of
1909 * mips would be illegal.
1911 if (need_nr_mips && explicit_lod) {
1912 struct lp_build_context bld_int_scalar;
1913 LLVMValueRef num_levels;
1914 lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
1916 if (static_state->level_zero_only) {
1917 num_levels = bld_int_scalar.one;
1920 LLVMValueRef last_level;
1922 last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit);
1923 num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
1924 num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one);
1926 sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type),