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,
82 LLVMValueRef y_stride,
83 LLVMValueRef z_stride,
84 LLVMValueRef data_ptr,
85 LLVMValueRef mipoffsets,
86 LLVMValueRef texel_out[4])
88 const struct lp_sampler_static_state *static_state = bld->static_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,
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);
193 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
194 border_chan_vec, texel_out[chan]);
201 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
204 lp_build_coord_mirror(struct lp_build_sample_context *bld,
207 struct lp_build_context *coord_bld = &bld->coord_bld;
208 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
209 LLVMValueRef fract, flr, isOdd;
211 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
213 /* isOdd = flr & 1 */
214 isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
216 /* make coord positive or negative depending on isOdd */
217 coord = lp_build_set_sign(coord_bld, fract, isOdd);
219 /* convert isOdd to float */
220 isOdd = lp_build_int_to_float(coord_bld, isOdd);
222 /* add isOdd to coord */
223 coord = lp_build_add(coord_bld, coord, isOdd);
230 * Helper to compute the first coord and the weight for
231 * linear wrap repeat npot textures
234 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
235 LLVMValueRef coord_f,
236 LLVMValueRef length_i,
237 LLVMValueRef length_f,
238 LLVMValueRef *coord0_i,
239 LLVMValueRef *weight_f)
241 struct lp_build_context *coord_bld = &bld->coord_bld;
242 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
243 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
244 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
247 /* wrap with normalized floats is just fract */
248 coord_f = lp_build_fract(coord_bld, coord_f);
249 /* mul by size and subtract 0.5 */
250 coord_f = lp_build_mul(coord_bld, coord_f, length_f);
251 coord_f = lp_build_sub(coord_bld, coord_f, half);
253 * we avoided the 0.5/length division before the repeat wrap,
254 * now need to fix up edge cases with selects
256 /* convert to int, compute lerp weight */
257 lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
258 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
259 PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
260 *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
265 * Build LLVM code for texture wrap mode for linear filtering.
266 * \param x0_out returns first integer texcoord
267 * \param x1_out returns second integer texcoord
268 * \param weight_out returns linear interpolation weight
271 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
274 LLVMValueRef length_f,
277 LLVMValueRef *x0_out,
278 LLVMValueRef *x1_out,
279 LLVMValueRef *weight_out)
281 struct lp_build_context *coord_bld = &bld->coord_bld;
282 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
283 LLVMBuilderRef builder = bld->gallivm->builder;
284 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
285 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
286 LLVMValueRef coord0, coord1, weight;
289 case PIPE_TEX_WRAP_REPEAT:
291 /* mul by size and subtract 0.5 */
292 coord = lp_build_mul(coord_bld, coord, length_f);
293 coord = lp_build_sub(coord_bld, coord, half);
294 /* convert to int, compute lerp weight */
295 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
296 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
298 coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
299 coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
303 lp_build_coord_repeat_npot_linear(bld, coord,
306 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
307 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
308 coord1 = LLVMBuildAnd(builder,
309 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
314 case PIPE_TEX_WRAP_CLAMP:
315 if (bld->static_state->normalized_coords) {
316 /* scale coord to length */
317 coord = lp_build_mul(coord_bld, coord, length_f);
320 /* clamp to [0, length] */
321 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
323 coord = lp_build_sub(coord_bld, coord, half);
325 /* convert to int, compute lerp weight */
326 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
327 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
330 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
332 struct lp_build_context abs_coord_bld = bld->coord_bld;
333 abs_coord_bld.type.sign = FALSE;
335 if (bld->static_state->normalized_coords) {
336 /* mul by tex size */
337 coord = lp_build_mul(coord_bld, coord, length_f);
339 /* clamp to length max */
340 coord = lp_build_min(coord_bld, coord, length_f);
342 coord = lp_build_sub(coord_bld, coord, half);
343 /* clamp to [0, length - 0.5] */
344 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
345 /* convert to int, compute lerp weight */
346 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
347 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
348 /* coord1 = min(coord1, length-1) */
349 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
353 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
356 if (bld->static_state->normalized_coords) {
357 /* scale coord to length */
358 coord = lp_build_mul(coord_bld, coord, length_f);
360 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
361 coord = lp_build_sub(coord_bld, coord, half);
362 min = lp_build_const_vec(bld->gallivm, coord_bld->type, -1.0F);
363 coord = lp_build_clamp(coord_bld, coord, min, length_f);
364 /* convert to int, compute lerp weight */
365 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
366 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
370 case PIPE_TEX_WRAP_MIRROR_REPEAT:
371 /* compute mirror function */
372 coord = lp_build_coord_mirror(bld, coord);
374 /* scale coord to length */
375 coord = lp_build_mul(coord_bld, coord, length_f);
376 coord = lp_build_sub(coord_bld, coord, half);
378 /* convert to int, compute lerp weight */
379 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
380 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
382 /* coord0 = max(coord0, 0) */
383 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
384 /* coord1 = min(coord1, length-1) */
385 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
388 case PIPE_TEX_WRAP_MIRROR_CLAMP:
389 coord = lp_build_abs(coord_bld, coord);
391 if (bld->static_state->normalized_coords) {
392 /* scale coord to length */
393 coord = lp_build_mul(coord_bld, coord, length_f);
396 /* clamp to [0, length] */
397 coord = lp_build_min(coord_bld, coord, length_f);
399 coord = lp_build_sub(coord_bld, coord, half);
401 /* convert to int, compute lerp weight */
402 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
403 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
406 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
408 LLVMValueRef min, max;
409 struct lp_build_context abs_coord_bld = bld->coord_bld;
410 abs_coord_bld.type.sign = FALSE;
411 coord = lp_build_abs(coord_bld, coord);
413 if (bld->static_state->normalized_coords) {
414 /* scale coord to length */
415 coord = lp_build_mul(coord_bld, coord, length_f);
418 /* clamp to [0.5, length - 0.5] */
420 max = lp_build_sub(coord_bld, length_f, min);
421 coord = lp_build_clamp(coord_bld, coord, min, max);
423 coord = lp_build_sub(coord_bld, coord, half);
425 /* convert to int, compute lerp weight */
426 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
427 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
431 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
433 coord = lp_build_abs(coord_bld, coord);
435 if (bld->static_state->normalized_coords) {
436 /* scale coord to length */
437 coord = lp_build_mul(coord_bld, coord, length_f);
440 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
441 /* skip -0.5 clamp (always positive), do sub first */
442 coord = lp_build_sub(coord_bld, coord, half);
443 coord = lp_build_min(coord_bld, coord, length_f);
445 /* convert to int, compute lerp weight */
446 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
447 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
460 *weight_out = weight;
465 * Build LLVM code for texture wrap mode for nearest filtering.
466 * \param coord the incoming texcoord (nominally in [0,1])
467 * \param length the texture size along one dimension, as int vector
468 * \param is_pot if TRUE, length is a power of two
469 * \param wrap_mode one of PIPE_TEX_WRAP_x
472 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
475 LLVMValueRef length_f,
479 struct lp_build_context *coord_bld = &bld->coord_bld;
480 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
481 LLVMBuilderRef builder = bld->gallivm->builder;
482 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
486 case PIPE_TEX_WRAP_REPEAT:
488 coord = lp_build_mul(coord_bld, coord, length_f);
489 icoord = lp_build_ifloor(coord_bld, coord);
490 icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
493 /* take fraction, unnormalize */
494 coord = lp_build_fract_safe(coord_bld, coord);
495 coord = lp_build_mul(coord_bld, coord, length_f);
496 icoord = lp_build_itrunc(coord_bld, coord);
500 case PIPE_TEX_WRAP_CLAMP:
501 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
502 if (bld->static_state->normalized_coords) {
503 /* scale coord to length */
504 coord = lp_build_mul(coord_bld, coord, length_f);
508 /* use itrunc instead since we clamp to 0 anyway */
509 icoord = lp_build_itrunc(coord_bld, coord);
511 /* clamp to [0, length - 1]. */
512 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
516 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
517 /* Note: this is the same as CLAMP_TO_EDGE, except min = -1 */
519 LLVMValueRef min, max;
521 if (bld->static_state->normalized_coords) {
522 /* scale coord to length */
523 coord = lp_build_mul(coord_bld, coord, length_f);
526 icoord = lp_build_ifloor(coord_bld, coord);
528 /* clamp to [-1, length] */
529 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
531 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
535 case PIPE_TEX_WRAP_MIRROR_REPEAT:
536 /* compute mirror function */
537 coord = lp_build_coord_mirror(bld, coord);
539 /* scale coord to length */
540 assert(bld->static_state->normalized_coords);
541 coord = lp_build_mul(coord_bld, coord, length_f);
543 /* itrunc == ifloor here */
544 icoord = lp_build_itrunc(coord_bld, coord);
546 /* clamp to [0, length - 1] */
547 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
550 case PIPE_TEX_WRAP_MIRROR_CLAMP:
551 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
552 coord = lp_build_abs(coord_bld, coord);
554 if (bld->static_state->normalized_coords) {
555 /* scale coord to length */
556 coord = lp_build_mul(coord_bld, coord, length_f);
559 /* itrunc == ifloor here */
560 icoord = lp_build_itrunc(coord_bld, coord);
562 /* clamp to [0, length - 1] */
563 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
566 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
567 coord = lp_build_abs(coord_bld, coord);
569 if (bld->static_state->normalized_coords) {
570 /* scale coord to length */
571 coord = lp_build_mul(coord_bld, coord, length_f);
574 /* itrunc == ifloor here */
575 icoord = lp_build_itrunc(coord_bld, coord);
577 /* clamp to [0, length] */
578 icoord = lp_build_min(int_coord_bld, icoord, length);
591 * Generate code to sample a mipmap level with nearest filtering.
592 * If sampling a cube texture, r = cube face in [0,5].
595 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
598 LLVMValueRef row_stride_vec,
599 LLVMValueRef img_stride_vec,
600 LLVMValueRef data_ptr,
601 LLVMValueRef mipoffsets,
605 LLVMValueRef colors_out[4])
607 const unsigned dims = bld->dims;
608 LLVMValueRef width_vec;
609 LLVMValueRef height_vec;
610 LLVMValueRef depth_vec;
611 LLVMValueRef flt_size;
612 LLVMValueRef flt_width_vec;
613 LLVMValueRef flt_height_vec;
614 LLVMValueRef flt_depth_vec;
615 LLVMValueRef x, y, z;
617 lp_build_extract_image_sizes(bld,
621 &width_vec, &height_vec, &depth_vec);
623 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
625 lp_build_extract_image_sizes(bld,
626 &bld->float_size_bld,
629 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
632 * Compute integer texcoords.
634 x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec,
635 bld->static_state->pot_width,
636 bld->static_state->wrap_s);
637 lp_build_name(x, "tex.x.wrapped");
640 y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec,
641 bld->static_state->pot_height,
642 bld->static_state->wrap_t);
643 lp_build_name(y, "tex.y.wrapped");
646 z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec,
647 bld->static_state->pot_depth,
648 bld->static_state->wrap_r);
649 lp_build_name(z, "tex.z.wrapped");
651 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
663 * Get texture colors.
665 lp_build_sample_texel_soa(bld, unit,
666 width_vec, height_vec, depth_vec,
668 row_stride_vec, img_stride_vec,
669 data_ptr, mipoffsets, colors_out);
674 * Generate code to sample a mipmap level with linear filtering.
675 * If sampling a cube texture, r = cube face in [0,5].
678 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
681 LLVMValueRef row_stride_vec,
682 LLVMValueRef img_stride_vec,
683 LLVMValueRef data_ptr,
684 LLVMValueRef mipoffsets,
688 LLVMValueRef colors_out[4])
690 const unsigned dims = bld->dims;
691 LLVMValueRef width_vec;
692 LLVMValueRef height_vec;
693 LLVMValueRef depth_vec;
694 LLVMValueRef flt_size;
695 LLVMValueRef flt_width_vec;
696 LLVMValueRef flt_height_vec;
697 LLVMValueRef flt_depth_vec;
698 LLVMValueRef x0, y0, z0, x1, y1, z1;
699 LLVMValueRef s_fpart, t_fpart, r_fpart;
700 LLVMValueRef neighbors[2][2][4];
703 lp_build_extract_image_sizes(bld,
707 &width_vec, &height_vec, &depth_vec);
709 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
711 lp_build_extract_image_sizes(bld,
712 &bld->float_size_bld,
715 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
718 * Compute integer texcoords.
720 lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec,
721 bld->static_state->pot_width,
722 bld->static_state->wrap_s,
724 lp_build_name(x0, "tex.x0.wrapped");
725 lp_build_name(x1, "tex.x1.wrapped");
728 lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec,
729 bld->static_state->pot_height,
730 bld->static_state->wrap_t,
732 lp_build_name(y0, "tex.y0.wrapped");
733 lp_build_name(y1, "tex.y1.wrapped");
736 lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec,
737 bld->static_state->pot_depth,
738 bld->static_state->wrap_r,
740 lp_build_name(z0, "tex.z0.wrapped");
741 lp_build_name(z1, "tex.z1.wrapped");
743 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
744 z0 = z1 = r; /* cube face */
753 y0 = y1 = t_fpart = NULL;
754 z0 = z1 = r_fpart = NULL;
758 * Get texture colors.
760 /* get x0/x1 texels */
761 lp_build_sample_texel_soa(bld, unit,
762 width_vec, height_vec, depth_vec,
764 row_stride_vec, img_stride_vec,
765 data_ptr, mipoffsets, neighbors[0][0]);
766 lp_build_sample_texel_soa(bld, unit,
767 width_vec, height_vec, depth_vec,
769 row_stride_vec, img_stride_vec,
770 data_ptr, mipoffsets, neighbors[0][1]);
773 /* Interpolate two samples from 1D image to produce one color */
774 for (chan = 0; chan < 4; chan++) {
775 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
776 neighbors[0][0][chan],
777 neighbors[0][1][chan]);
782 LLVMValueRef colors0[4];
784 /* get x0/x1 texels at y1 */
785 lp_build_sample_texel_soa(bld, unit,
786 width_vec, height_vec, depth_vec,
788 row_stride_vec, img_stride_vec,
789 data_ptr, mipoffsets, neighbors[1][0]);
790 lp_build_sample_texel_soa(bld, unit,
791 width_vec, height_vec, depth_vec,
793 row_stride_vec, img_stride_vec,
794 data_ptr, mipoffsets, neighbors[1][1]);
796 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
797 for (chan = 0; chan < 4; chan++) {
798 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
800 neighbors[0][0][chan],
801 neighbors[0][1][chan],
802 neighbors[1][0][chan],
803 neighbors[1][1][chan]);
807 LLVMValueRef neighbors1[2][2][4];
808 LLVMValueRef colors1[4];
810 /* get x0/x1/y0/y1 texels at z1 */
811 lp_build_sample_texel_soa(bld, unit,
812 width_vec, height_vec, depth_vec,
814 row_stride_vec, img_stride_vec,
815 data_ptr, mipoffsets, neighbors1[0][0]);
816 lp_build_sample_texel_soa(bld, unit,
817 width_vec, height_vec, depth_vec,
819 row_stride_vec, img_stride_vec,
820 data_ptr, mipoffsets, neighbors1[0][1]);
821 lp_build_sample_texel_soa(bld, unit,
822 width_vec, height_vec, depth_vec,
824 row_stride_vec, img_stride_vec,
825 data_ptr, mipoffsets, neighbors1[1][0]);
826 lp_build_sample_texel_soa(bld, unit,
827 width_vec, height_vec, depth_vec,
829 row_stride_vec, img_stride_vec,
830 data_ptr, mipoffsets, neighbors1[1][1]);
832 /* Bilinear interpolate the four samples from the second Z slice */
833 for (chan = 0; chan < 4; chan++) {
834 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
836 neighbors1[0][0][chan],
837 neighbors1[0][1][chan],
838 neighbors1[1][0][chan],
839 neighbors1[1][1][chan]);
842 /* Linearly interpolate the two samples from the two 3D slices */
843 for (chan = 0; chan < 4; chan++) {
844 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
846 colors0[chan], colors1[chan]);
851 for (chan = 0; chan < 4; chan++) {
852 colors_out[chan] = colors0[chan];
860 * Sample the texture/mipmap using given image filter and mip filter.
861 * data0_ptr and data1_ptr point to the two mipmap levels to sample
862 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
863 * If we're using nearest miplevel sampling the '1' values will be null/unused.
866 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
873 LLVMValueRef ilevel0,
874 LLVMValueRef ilevel1,
875 LLVMValueRef lod_fpart,
876 LLVMValueRef *colors_out)
878 LLVMBuilderRef builder = bld->gallivm->builder;
879 LLVMValueRef size0 = NULL;
880 LLVMValueRef size1 = NULL;
881 LLVMValueRef row_stride0_vec = NULL;
882 LLVMValueRef row_stride1_vec = NULL;
883 LLVMValueRef img_stride0_vec = NULL;
884 LLVMValueRef img_stride1_vec = NULL;
885 LLVMValueRef data_ptr0 = NULL;
886 LLVMValueRef data_ptr1 = NULL;
887 LLVMValueRef mipoff0 = NULL;
888 LLVMValueRef mipoff1 = NULL;
889 LLVMValueRef colors0[4], colors1[4];
892 /* sample the first mipmap level */
893 lp_build_mipmap_level_sizes(bld, ilevel0,
895 &row_stride0_vec, &img_stride0_vec);
896 if (bld->num_lods == 1) {
897 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
900 /* This path should work for num_lods 1 too but slightly less efficient */
901 data_ptr0 = bld->base_ptr;
902 mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
904 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
905 lp_build_sample_image_nearest(bld, unit,
907 row_stride0_vec, img_stride0_vec,
908 data_ptr0, mipoff0, s, t, r,
912 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
913 lp_build_sample_image_linear(bld, unit,
915 row_stride0_vec, img_stride0_vec,
916 data_ptr0, mipoff0, s, t, r,
920 /* Store the first level's colors in the output variables */
921 for (chan = 0; chan < 4; chan++) {
922 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
925 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
926 struct lp_build_if_state if_ctx;
927 LLVMValueRef need_lerp;
928 unsigned num_quads = bld->coord_bld.type.length / 4;
930 /* need_lerp = lod_fpart > 0 */
931 if (num_quads == 1) {
932 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
933 lod_fpart, bld->perquadf_bld.zero,
938 * We'll do mip filtering if any of the quads need it.
939 * It might be better to split the vectors here and only fetch/filter
940 * quads which need it.
943 * We unfortunately need to clamp lod_fpart here since we can get
944 * negative values which would screw up filtering if not all
945 * lod_fpart values have same sign.
947 lod_fpart = lp_build_max(&bld->perquadf_bld, lod_fpart,
948 bld->perquadf_bld.zero);
949 need_lerp = lp_build_compare(bld->gallivm, bld->perquadf_bld.type,
951 lod_fpart, bld->perquadf_bld.zero);
952 need_lerp = lp_build_any_true_range(&bld->perquadi_bld, num_quads, need_lerp);
955 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
957 /* sample the second mipmap level */
958 lp_build_mipmap_level_sizes(bld, ilevel1,
960 &row_stride1_vec, &img_stride1_vec);
961 if (bld->num_lods == 1) {
962 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
965 data_ptr1 = bld->base_ptr;
966 mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
968 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
969 lp_build_sample_image_nearest(bld, unit,
971 row_stride1_vec, img_stride1_vec,
972 data_ptr1, mipoff1, s, t, r,
976 lp_build_sample_image_linear(bld, unit,
978 row_stride1_vec, img_stride1_vec,
979 data_ptr1, mipoff1, s, t, r,
983 /* interpolate samples from the two mipmap levels */
985 lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
986 bld->perquadf_bld.type,
990 for (chan = 0; chan < 4; chan++) {
991 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
992 colors0[chan], colors1[chan]);
993 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
996 lp_build_endif(&if_ctx);
1001 * Calculate cube face, lod, mip levels.
1004 lp_build_sample_common(struct lp_build_sample_context *bld,
1009 const struct lp_derivatives *derivs,
1010 LLVMValueRef lod_bias, /* optional */
1011 LLVMValueRef explicit_lod, /* optional */
1012 LLVMValueRef *lod_ipart,
1013 LLVMValueRef *lod_fpart,
1014 LLVMValueRef *ilevel0,
1015 LLVMValueRef *ilevel1)
1017 const unsigned mip_filter = bld->static_state->min_mip_filter;
1018 const unsigned min_filter = bld->static_state->min_img_filter;
1019 const unsigned mag_filter = bld->static_state->mag_img_filter;
1020 LLVMValueRef first_level;
1021 struct lp_derivatives face_derivs;
1024 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1025 mip_filter, min_filter, mag_filter);
1029 * Choose cube face, recompute texcoords and derivatives for the chosen face.
1031 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1032 LLVMValueRef face, face_s, face_t;
1033 lp_build_cube_lookup(bld, *s, *t, *r, &face, &face_s, &face_t);
1034 *s = face_s; /* vec */
1035 *t = face_t; /* vec */
1036 /* use 'r' to indicate cube face */
1037 *r = face; /* vec */
1039 /* recompute ddx, ddy using the new (s,t) face texcoords */
1040 face_derivs.ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(&bld->coord_bld, *s, *t);
1041 face_derivs.ddx_ddy[1] = NULL;
1042 derivs = &face_derivs;
1046 * Compute the level of detail (float).
1048 if (min_filter != mag_filter ||
1049 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1050 /* Need to compute lod either to choose mipmap levels or to
1051 * distinguish between minification/magnification with one mipmap level.
1053 lp_build_lod_selector(bld, unit, derivs,
1054 lod_bias, explicit_lod,
1056 lod_ipart, lod_fpart);
1058 *lod_ipart = bld->perquadi_bld.zero;
1062 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1064 switch (mip_filter) {
1066 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1068 case PIPE_TEX_MIPFILTER_NONE:
1069 /* always use mip level 0 */
1070 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1071 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1072 * We should be able to set ilevel0 = const(0) but that causes
1073 * bad x86 code to be emitted.
1074 * XXX should probably disable that on other llvm versions.
1077 lp_build_nearest_mip_level(bld, unit, *lod_ipart, ilevel0);
1080 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1081 bld->gallivm, unit);
1082 first_level = lp_build_broadcast_scalar(&bld->perquadi_bld, first_level);
1083 *ilevel0 = first_level;
1086 case PIPE_TEX_MIPFILTER_NEAREST:
1088 lp_build_nearest_mip_level(bld, unit, *lod_ipart, ilevel0);
1090 case PIPE_TEX_MIPFILTER_LINEAR:
1093 lp_build_linear_mip_levels(bld, unit,
1094 *lod_ipart, lod_fpart,
1101 * General texture sampling codegen.
1102 * This function handles texture sampling for all texture targets (1D,
1103 * 2D, 3D, cube) and all filtering modes.
1106 lp_build_sample_general(struct lp_build_sample_context *bld,
1111 LLVMValueRef lod_ipart,
1112 LLVMValueRef lod_fpart,
1113 LLVMValueRef ilevel0,
1114 LLVMValueRef ilevel1,
1115 LLVMValueRef *colors_out)
1117 struct lp_build_context *int_bld = &bld->int_bld;
1118 LLVMBuilderRef builder = bld->gallivm->builder;
1119 const unsigned mip_filter = bld->static_state->min_mip_filter;
1120 const unsigned min_filter = bld->static_state->min_img_filter;
1121 const unsigned mag_filter = bld->static_state->mag_img_filter;
1122 LLVMValueRef texels[4];
1126 * Get/interpolate texture colors.
1129 for (chan = 0; chan < 4; ++chan) {
1130 texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1131 lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]);
1134 if (min_filter == mag_filter) {
1135 /* no need to distinguish between minification and magnification */
1136 lp_build_sample_mipmap(bld, unit,
1137 min_filter, mip_filter,
1139 ilevel0, ilevel1, lod_fpart,
1143 /* Emit conditional to choose min image filter or mag image filter
1144 * depending on the lod being > 0 or <= 0, respectively.
1146 struct lp_build_if_state if_ctx;
1147 LLVMValueRef minify;
1150 * XXX this should to all lods into account, if some are min
1151 * some max probably could hack up the coords/weights in the linear
1152 * path with selects to work for nearest.
1153 * If that's just two quads sitting next to each other it seems
1154 * quite ok to do the same filtering method on both though, at
1155 * least unless we have explicit lod (and who uses different
1156 * min/mag filter with that?)
1159 /* minify = lod >= 0.0 */
1160 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1161 lod_ipart, int_bld->zero, "");
1163 lp_build_if(&if_ctx, bld->gallivm, minify);
1165 /* Use the minification filter */
1166 lp_build_sample_mipmap(bld, unit,
1167 min_filter, mip_filter,
1169 ilevel0, ilevel1, lod_fpart,
1172 lp_build_else(&if_ctx);
1174 /* Use the magnification filter */
1175 lp_build_sample_mipmap(bld, unit,
1176 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1178 ilevel0, NULL, NULL,
1181 lp_build_endif(&if_ctx);
1184 for (chan = 0; chan < 4; ++chan) {
1185 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1186 lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]);
1192 * Texel fetch function.
1193 * In contrast to general sampling there is no filtering, no coord minification,
1194 * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
1195 * directly to be applied to the selected mip level (after adding texel offsets).
1196 * This function handles texel fetch for all targets where texel fetch is supported
1197 * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
1200 lp_build_fetch_texel(struct lp_build_sample_context *bld,
1202 const LLVMValueRef *coords,
1203 LLVMValueRef explicit_lod,
1204 const LLVMValueRef *offsets,
1205 LLVMValueRef *colors_out)
1207 struct lp_build_context *perquadi_bld = &bld->perquadi_bld;
1208 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1209 unsigned dims = bld->dims, chan;
1210 LLVMValueRef size, ilevel;
1211 LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
1212 LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
1213 LLVMValueRef width, height, depth, i, j;
1214 LLVMValueRef offset, out_of_bounds, out1;
1216 /* XXX just like ordinary sampling, we don't handle per-pixel lod (yet). */
1217 if (explicit_lod && bld->static_state->target != PIPE_BUFFER) {
1218 ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
1219 perquadi_bld->type, explicit_lod, 0);
1220 lp_build_nearest_mip_level(bld, unit, ilevel, &ilevel);
1224 ilevel = lp_build_const_int32(bld->gallivm, 0);
1226 lp_build_mipmap_level_sizes(bld, ilevel,
1228 &row_stride_vec, &img_stride_vec);
1229 lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
1230 size, &width, &height, &depth);
1232 /* This is a lot like border sampling */
1234 /* XXX coords are really unsigned, offsets are signed */
1235 x = lp_build_add(int_coord_bld, x, offsets[0]);
1237 out_of_bounds = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
1238 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
1239 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1243 y = lp_build_add(int_coord_bld, y, offsets[1]);
1245 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
1246 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1247 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
1248 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1252 z = lp_build_add(int_coord_bld, z, offsets[2]);
1254 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
1255 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1256 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
1257 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1261 lp_build_sample_offset(int_coord_bld,
1263 x, y, z, row_stride_vec, img_stride_vec,
1266 if (bld->static_state->target != PIPE_BUFFER) {
1267 offset = lp_build_add(int_coord_bld, offset,
1268 lp_build_get_mip_offsets(bld, ilevel));
1271 offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
1273 lp_build_fetch_rgba_soa(bld->gallivm,
1276 bld->base_ptr, offset,
1282 * Not needed except for ARB_robust_buffer_access_behavior.
1283 * Could use min/max above instead of out-of-bounds comparisons
1284 * (in fact cast to unsigned and min only is sufficient)
1285 * if we don't care about the result returned for out-of-bounds.
1287 for (chan = 0; chan < 4; chan++) {
1288 colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
1289 bld->texel_bld.zero, colors_out[chan]);
1296 * Do shadow test/comparison.
1297 * \param p the texcoord Z (aka R, aka P) component
1298 * \param texel the texel to compare against (use the X channel)
1299 * Ideally this should really be done per-sample.
1302 lp_build_sample_compare(struct lp_build_sample_context *bld,
1304 LLVMValueRef texel[4])
1306 struct lp_build_context *texel_bld = &bld->texel_bld;
1307 LLVMBuilderRef builder = bld->gallivm->builder;
1309 const unsigned chan = 0;
1311 if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1316 LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
1317 LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
1318 LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
1319 lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
1323 /* Clamp p coords to [0,1] */
1324 p = lp_build_clamp(&bld->coord_bld, p,
1325 bld->coord_bld.zero,
1326 bld->coord_bld.one);
1328 /* result = (p FUNC texel) ? 1 : 0 */
1329 res = lp_build_cmp(texel_bld, bld->static_state->compare_func,
1331 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1333 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1337 texel[3] = texel_bld->one;
1342 * Just set texels to white instead of actually sampling the texture.
1346 lp_build_sample_nop(struct gallivm_state *gallivm,
1347 struct lp_type type,
1348 const LLVMValueRef *coords,
1349 LLVMValueRef texel_out[4])
1351 LLVMValueRef one = lp_build_one(gallivm, type);
1354 for (chan = 0; chan < 4; chan++) {
1355 texel_out[chan] = one;
1361 * Build texture sampling code.
1362 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1364 * \param type vector float type to use for coords, etc.
1365 * \param is_fetch if this is a texel fetch instruction.
1366 * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
1369 lp_build_sample_soa(struct gallivm_state *gallivm,
1370 const struct lp_sampler_static_state *static_state,
1371 struct lp_sampler_dynamic_state *dynamic_state,
1372 struct lp_type type,
1375 const LLVMValueRef *coords,
1376 const LLVMValueRef *offsets,
1377 const struct lp_derivatives *derivs,
1378 LLVMValueRef lod_bias, /* optional */
1379 LLVMValueRef explicit_lod, /* optional */
1380 LLVMValueRef texel_out[4])
1382 unsigned dims = texture_dims(static_state->target);
1383 unsigned num_quads = type.length / 4;
1384 unsigned mip_filter = static_state->min_mip_filter;
1385 struct lp_build_sample_context bld;
1386 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1387 LLVMBuilderRef builder = gallivm->builder;
1388 LLVMValueRef tex_width, tex_height, tex_depth;
1394 enum pipe_format fmt = static_state->format;
1395 debug_printf("Sample from %s\n", util_format_name(fmt));
1398 assert(type.floating);
1400 /* Setup our build context */
1401 memset(&bld, 0, sizeof bld);
1402 bld.gallivm = gallivm;
1403 bld.static_state = static_state;
1404 bld.dynamic_state = dynamic_state;
1405 bld.format_desc = util_format_description(static_state->format);
1408 bld.vector_width = lp_type_width(type);
1410 bld.float_type = lp_type_float(32);
1411 bld.int_type = lp_type_int(32);
1412 bld.coord_type = type;
1413 bld.int_coord_type = lp_int_type(type);
1414 bld.float_size_in_type = lp_type_float(32);
1415 bld.float_size_in_type.length = dims > 1 ? 4 : 1;
1416 bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
1417 bld.texel_type = type;
1418 bld.perquadf_type = type;
1419 /* we want native vector size to be able to use our intrinsics */
1420 bld.perquadf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
1421 bld.perquadi_type = lp_int_type(bld.perquadf_type);
1424 * There are other situations where at least the multiple int lods could be
1425 * avoided like min and max lod being equal.
1427 if ((is_fetch && explicit_lod && bld.static_state->target != PIPE_BUFFER) ||
1428 (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
1429 bld.num_lods = num_quads;
1435 bld.float_size_type = bld.float_size_in_type;
1436 bld.float_size_type.length = bld.num_lods > 1 ? type.length :
1437 bld.float_size_in_type.length;
1438 bld.int_size_type = lp_int_type(bld.float_size_type);
1440 lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1441 lp_build_context_init(&bld.float_vec_bld, gallivm, type);
1442 lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1443 lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1444 lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1445 lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
1446 lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
1447 lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1448 lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1449 lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1450 lp_build_context_init(&bld.perquadf_bld, gallivm, bld.perquadf_type);
1451 lp_build_context_init(&bld.perquadi_bld, gallivm, bld.perquadi_type);
1453 /* Get the dynamic state */
1454 tex_width = dynamic_state->width(dynamic_state, gallivm, unit);
1455 tex_height = dynamic_state->height(dynamic_state, gallivm, unit);
1456 tex_depth = dynamic_state->depth(dynamic_state, gallivm, unit);
1457 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, unit);
1458 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, unit);
1459 bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, unit);
1460 bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, unit);
1461 /* Note that mip_offsets is an array[level] of offsets to texture images */
1467 /* width, height, depth as single int vector */
1469 bld.int_size = tex_width;
1472 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
1473 tex_width, LLVMConstInt(i32t, 0, 0), "");
1475 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1476 tex_height, LLVMConstInt(i32t, 1, 0), "");
1478 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1479 tex_depth, LLVMConstInt(i32t, 2, 0), "");
1485 /* For debug: no-op texture sampling */
1486 lp_build_sample_nop(gallivm,
1492 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
1493 LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
1494 boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
1495 lp_is_simple_wrap_mode(static_state->wrap_s) &&
1496 lp_is_simple_wrap_mode(static_state->wrap_t);
1498 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1499 !use_aos && util_format_fits_8unorm(bld.format_desc)) {
1500 debug_printf("%s: using floating point linear filtering for %s\n",
1501 __FUNCTION__, bld.format_desc->short_name);
1502 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1503 static_state->min_img_filter,
1504 static_state->mag_img_filter,
1505 static_state->min_mip_filter,
1506 static_state->wrap_s,
1507 static_state->wrap_t);
1511 lp_build_fetch_texel(&bld, unit, coords,
1512 explicit_lod, offsets,
1515 if (static_state->target != PIPE_BUFFER) {
1516 apply_sampler_swizzle(&bld, texel_out);
1522 lp_build_sample_common(&bld, unit,
1524 derivs, lod_bias, explicit_lod,
1525 &lod_ipart, &lod_fpart,
1526 &ilevel0, &ilevel1);
1529 * we only try 8-wide sampling with soa as it appears to
1530 * be a loss with aos with AVX (but it should work).
1531 * (It should be faster if we'd support avx2)
1533 if (num_quads == 1 || !use_aos) {
1535 if (num_quads > 1) {
1536 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1537 LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
1539 * These parameters are the same for all quads,
1540 * could probably simplify.
1542 lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
1543 ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
1547 /* do sampling/filtering with fixed pt arithmetic */
1548 lp_build_sample_aos(&bld, unit,
1550 lod_ipart, lod_fpart,
1556 lp_build_sample_general(&bld, unit,
1558 lod_ipart, lod_fpart,
1565 struct lp_build_sample_context bld4;
1566 struct lp_type type4 = type;
1568 LLVMValueRef texelout4[4];
1569 LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
1573 /* Setup our build context */
1574 memset(&bld4, 0, sizeof bld4);
1575 bld4.gallivm = bld.gallivm;
1576 bld4.static_state = bld.static_state;
1577 bld4.dynamic_state = bld.dynamic_state;
1578 bld4.format_desc = bld.format_desc;
1579 bld4.dims = bld.dims;
1580 bld4.row_stride_array = bld.row_stride_array;
1581 bld4.img_stride_array = bld.img_stride_array;
1582 bld4.base_ptr = bld.base_ptr;
1583 bld4.mip_offsets = bld.mip_offsets;
1584 bld4.int_size = bld.int_size;
1586 bld4.vector_width = lp_type_width(type4);
1588 bld4.float_type = lp_type_float(32);
1589 bld4.int_type = lp_type_int(32);
1590 bld4.coord_type = type4;
1591 bld4.int_coord_type = lp_int_type(type4);
1592 bld4.float_size_in_type = lp_type_float(32);
1593 bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
1594 bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
1595 bld4.texel_type = type4;
1596 bld4.perquadf_type = type4;
1597 /* we want native vector size to be able to use our intrinsics */
1598 bld4.perquadf_type.length = 1;
1599 bld4.perquadi_type = lp_int_type(bld4.perquadf_type);
1602 bld4.int_size_type = bld4.int_size_in_type;
1603 bld4.float_size_type = bld4.float_size_in_type;
1605 lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
1606 lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
1607 lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
1608 lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
1609 lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
1610 lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
1611 lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
1612 lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
1613 lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
1614 lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
1615 lp_build_context_init(&bld4.perquadf_bld, gallivm, bld4.perquadf_type);
1616 lp_build_context_init(&bld4.perquadi_bld, gallivm, bld4.perquadi_type);
1618 for (i = 0; i < num_quads; i++) {
1619 LLVMValueRef s4, t4, r4;
1620 LLVMValueRef lod_iparts, lod_fparts = NULL;
1621 LLVMValueRef ilevel0s, ilevel1s = NULL;
1622 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1624 s4 = lp_build_extract_range(gallivm, s, 4*i, 4);
1625 t4 = lp_build_extract_range(gallivm, t, 4*i, 4);
1626 r4 = lp_build_extract_range(gallivm, r, 4*i, 4);
1627 lod_iparts = LLVMBuildExtractElement(builder, lod_ipart, indexi, "");
1628 ilevel0s = LLVMBuildExtractElement(builder, ilevel0, indexi, "");
1629 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1630 ilevel1s = LLVMBuildExtractElement(builder, ilevel1, indexi, "");
1631 lod_fparts = LLVMBuildExtractElement(builder, lod_fpart, indexi, "");
1635 /* do sampling/filtering with fixed pt arithmetic */
1636 lp_build_sample_aos(&bld4, unit,
1638 lod_iparts, lod_fparts,
1644 lp_build_sample_general(&bld4, unit,
1646 lod_iparts, lod_fparts,
1650 for (j = 0; j < 4; j++) {
1651 texelouttmp[j][i] = texelout4[j];
1655 for (j = 0; j < 4; j++) {
1656 texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
1661 lp_build_sample_compare(&bld, r, texel_out);
1663 apply_sampler_swizzle(&bld, texel_out);
1667 lp_build_size_query_soa(struct gallivm_state *gallivm,
1668 const struct lp_sampler_static_state *static_state,
1669 struct lp_sampler_dynamic_state *dynamic_state,
1670 struct lp_type int_type,
1672 LLVMValueRef explicit_lod,
1673 LLVMValueRef *sizes_out)
1678 struct lp_build_context bld_int_vec;
1680 switch (static_state->target) {
1681 case PIPE_TEXTURE_1D:
1685 case PIPE_TEXTURE_2D:
1686 case PIPE_TEXTURE_CUBE:
1687 case PIPE_TEXTURE_RECT:
1690 case PIPE_TEXTURE_3D:
1699 assert(!int_type.floating);
1701 lp_build_context_init(&bld_int_vec, gallivm, lp_type_int_vec(32, 128));
1704 LLVMValueRef first_level;
1705 lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), "");
1706 first_level = dynamic_state->first_level(dynamic_state, gallivm, unit);
1707 lod = lp_build_broadcast_scalar(&bld_int_vec,
1708 LLVMBuildAdd(gallivm->builder, lod, first_level, "lod"));
1711 lod = bld_int_vec.zero;
1714 size = bld_int_vec.undef;
1716 size = LLVMBuildInsertElement(gallivm->builder, size,
1717 dynamic_state->width(dynamic_state, gallivm, unit),
1718 lp_build_const_int32(gallivm, 0), "");
1721 size = LLVMBuildInsertElement(gallivm->builder, size,
1722 dynamic_state->height(dynamic_state, gallivm, unit),
1723 lp_build_const_int32(gallivm, 1), "");
1727 size = LLVMBuildInsertElement(gallivm->builder, size,
1728 dynamic_state->depth(dynamic_state, gallivm, unit),
1729 lp_build_const_int32(gallivm, 2), "");
1732 size = lp_build_minify(&bld_int_vec, size, lod);
1734 for (i=0; i < dims; i++) {
1735 sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec.type, int_type,
1737 lp_build_const_int32(gallivm, i));