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 "util/u_debug.h"
39 #include "util/u_dump.h"
40 #include "util/u_memory.h"
41 #include "util/u_math.h"
42 #include "util/u_format.h"
43 #include "lp_bld_debug.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_arit.h"
48 #include "lp_bld_bitarit.h"
49 #include "lp_bld_logic.h"
50 #include "lp_bld_printf.h"
51 #include "lp_bld_swizzle.h"
52 #include "lp_bld_flow.h"
53 #include "lp_bld_gather.h"
54 #include "lp_bld_format.h"
55 #include "lp_bld_sample.h"
56 #include "lp_bld_sample_aos.h"
57 #include "lp_bld_struct.h"
58 #include "lp_bld_quad.h"
62 * Generate code to fetch a texel from a texture at int coords (x, y, z).
63 * The computation depends on whether the texture is 1D, 2D or 3D.
64 * The result, texel, will be float vectors:
65 * texel[0] = red values
66 * texel[1] = green values
67 * texel[2] = blue values
68 * texel[3] = alpha values
71 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
79 LLVMValueRef y_stride,
80 LLVMValueRef z_stride,
81 LLVMValueRef data_ptr,
82 LLVMValueRef texel_out[4])
84 const struct lp_sampler_static_state *static_state = bld->static_state;
85 const unsigned dims = bld->dims;
86 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
89 LLVMValueRef use_border = NULL;
91 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
92 if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
93 static_state->min_img_filter,
94 static_state->mag_img_filter)) {
96 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
97 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
98 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
102 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
103 static_state->min_img_filter,
104 static_state->mag_img_filter)) {
106 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
107 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
109 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
110 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
113 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
118 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
119 static_state->min_img_filter,
120 static_state->mag_img_filter)) {
122 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
123 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
125 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
126 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
129 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
133 /* convert x,y,z coords to linear offset from start of texture, in bytes */
134 lp_build_sample_offset(&bld->int_coord_bld,
136 x, y, z, y_stride, z_stride,
140 /* If we can sample the border color, it means that texcoords may
141 * lie outside the bounds of the texture image. We need to do
142 * something to prevent reading out of bounds and causing a segfault.
144 * Simply AND the texture coords with !use_border. This will cause
145 * coords which are out of bounds to become zero. Zero's guaranteed
146 * to be inside the texture image.
148 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
151 lp_build_fetch_rgba_soa(bld->builder,
159 * Note: if we find an app which frequently samples the texture border
160 * we might want to implement a true conditional here to avoid sampling
161 * the texture whenever possible (since that's quite a bit of code).
164 * texel = border_color;
167 * texel = sample_texture(coord);
169 * As it is now, we always sample the texture, then selectively replace
170 * the texel color results with the border color.
174 /* select texel color or border color depending on use_border */
175 LLVMValueRef border_color_ptr =
176 bld->dynamic_state->border_color(bld->dynamic_state,
179 for (chan = 0; chan < 4; chan++) {
180 LLVMValueRef border_chan =
181 lp_build_array_get(bld->builder, border_color_ptr,
182 lp_build_const_int32(chan));
183 LLVMValueRef border_chan_vec =
184 lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
185 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
186 border_chan_vec, texel_out[chan]);
190 apply_sampler_swizzle(bld, texel_out);
195 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
198 lp_build_coord_mirror(struct lp_build_sample_context *bld,
201 struct lp_build_context *coord_bld = &bld->coord_bld;
202 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
203 LLVMValueRef fract, flr, isOdd;
205 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
207 /* isOdd = flr & 1 */
208 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
210 /* make coord positive or negative depending on isOdd */
211 coord = lp_build_set_sign(coord_bld, fract, isOdd);
213 /* convert isOdd to float */
214 isOdd = lp_build_int_to_float(coord_bld, isOdd);
216 /* add isOdd to coord */
217 coord = lp_build_add(coord_bld, coord, isOdd);
224 * Build LLVM code for texture wrap mode for linear filtering.
225 * \param x0_out returns first integer texcoord
226 * \param x1_out returns second integer texcoord
227 * \param weight_out returns linear interpolation weight
230 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
233 LLVMValueRef length_f,
236 LLVMValueRef *x0_out,
237 LLVMValueRef *x1_out,
238 LLVMValueRef *weight_out)
240 struct lp_build_context *coord_bld = &bld->coord_bld;
241 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
242 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
243 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
244 LLVMValueRef coord0, coord1, weight;
247 case PIPE_TEX_WRAP_REPEAT:
248 /* mul by size and subtract 0.5 */
249 coord = lp_build_mul(coord_bld, coord, length_f);
250 coord = lp_build_sub(coord_bld, coord, half);
251 /* convert to int, compute lerp weight */
252 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
255 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
256 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
257 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
260 /* Add a bias to the texcoord to handle negative coords */
261 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
263 coord0 = LLVMBuildAdd(bld->builder, coord0, bias, "");
264 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
265 mask = lp_build_compare(bld->builder, int_coord_bld->type,
266 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
267 coord1 = LLVMBuildAnd(bld->builder,
268 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
273 case PIPE_TEX_WRAP_CLAMP:
274 if (bld->static_state->normalized_coords) {
275 /* scale coord to length */
276 coord = lp_build_mul(coord_bld, coord, length_f);
279 /* clamp to [0, length] */
280 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
282 coord = lp_build_sub(coord_bld, coord, half);
284 /* convert to int, compute lerp weight */
285 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
286 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
289 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
291 struct lp_build_context abs_coord_bld = bld->coord_bld;
292 abs_coord_bld.type.sign = FALSE;
294 if (bld->static_state->normalized_coords) {
295 /* mul by tex size */
296 coord = lp_build_mul(coord_bld, coord, length_f);
298 /* clamp to length max */
299 coord = lp_build_min(coord_bld, coord, length_f);
301 coord = lp_build_sub(coord_bld, coord, half);
302 /* clamp to [0, length - 0.5] */
303 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
304 /* convert to int, compute lerp weight */
305 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
306 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
307 /* coord1 = min(coord1, length-1) */
308 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
312 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
315 if (bld->static_state->normalized_coords) {
316 /* scale coord to length */
317 coord = lp_build_mul(coord_bld, coord, length_f);
319 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
320 coord = lp_build_sub(coord_bld, coord, half);
321 min = lp_build_const_vec(coord_bld->type, -1.0F);
322 coord = lp_build_clamp(coord_bld, coord, min, length_f);
323 /* convert to int, compute lerp weight */
324 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
325 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
329 case PIPE_TEX_WRAP_MIRROR_REPEAT:
330 /* compute mirror function */
331 coord = lp_build_coord_mirror(bld, coord);
333 /* scale coord to length */
334 coord = lp_build_mul(coord_bld, coord, length_f);
335 coord = lp_build_sub(coord_bld, coord, half);
337 /* convert to int, compute lerp weight */
338 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
339 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
341 /* coord0 = max(coord0, 0) */
342 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
343 /* coord1 = min(coord1, length-1) */
344 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
347 case PIPE_TEX_WRAP_MIRROR_CLAMP:
348 coord = lp_build_abs(coord_bld, coord);
350 if (bld->static_state->normalized_coords) {
351 /* scale coord to length */
352 coord = lp_build_mul(coord_bld, coord, length_f);
355 /* clamp to [0, length] */
356 coord = lp_build_min(coord_bld, coord, length_f);
358 coord = lp_build_sub(coord_bld, coord, half);
360 /* convert to int, compute lerp weight */
361 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
362 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
365 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
367 LLVMValueRef min, max;
368 struct lp_build_context abs_coord_bld = bld->coord_bld;
369 abs_coord_bld.type.sign = FALSE;
370 coord = lp_build_abs(coord_bld, coord);
372 if (bld->static_state->normalized_coords) {
373 /* scale coord to length */
374 coord = lp_build_mul(coord_bld, coord, length_f);
377 /* clamp to [0.5, length - 0.5] */
379 max = lp_build_sub(coord_bld, length_f, min);
380 coord = lp_build_clamp(coord_bld, coord, min, max);
382 coord = lp_build_sub(coord_bld, coord, half);
384 /* convert to int, compute lerp weight */
385 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
386 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
390 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
392 coord = lp_build_abs(coord_bld, coord);
394 if (bld->static_state->normalized_coords) {
395 /* scale coord to length */
396 coord = lp_build_mul(coord_bld, coord, length_f);
399 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
400 /* skip -0.5 clamp (always positive), do sub first */
401 coord = lp_build_sub(coord_bld, coord, half);
402 coord = lp_build_min(coord_bld, coord, length_f);
404 /* convert to int, compute lerp weight */
405 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
406 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
419 *weight_out = weight;
424 * Build LLVM code for texture wrap mode for nearest filtering.
425 * \param coord the incoming texcoord (nominally in [0,1])
426 * \param length the texture size along one dimension, as int vector
427 * \param is_pot if TRUE, length is a power of two
428 * \param wrap_mode one of PIPE_TEX_WRAP_x
431 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
434 LLVMValueRef length_f,
438 struct lp_build_context *coord_bld = &bld->coord_bld;
439 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
440 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
444 case PIPE_TEX_WRAP_REPEAT:
445 coord = lp_build_mul(coord_bld, coord, length_f);
446 icoord = lp_build_ifloor(coord_bld, coord);
448 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
450 /* Add a bias to the texcoord to handle negative coords */
451 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
452 icoord = LLVMBuildAdd(bld->builder, icoord, bias, "");
453 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
457 case PIPE_TEX_WRAP_CLAMP:
458 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
459 if (bld->static_state->normalized_coords) {
460 /* scale coord to length */
461 coord = lp_build_mul(coord_bld, coord, length_f);
465 /* use itrunc instead since we clamp to 0 anyway */
466 icoord = lp_build_itrunc(coord_bld, coord);
468 /* clamp to [0, length - 1]. */
469 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
473 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
474 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
476 LLVMValueRef min, max;
478 if (bld->static_state->normalized_coords) {
479 /* scale coord to length */
480 coord = lp_build_mul(coord_bld, coord, length_f);
483 icoord = lp_build_ifloor(coord_bld, coord);
485 /* clamp to [-1, length] */
486 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
488 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
492 case PIPE_TEX_WRAP_MIRROR_REPEAT:
493 /* compute mirror function */
494 coord = lp_build_coord_mirror(bld, coord);
496 /* scale coord to length */
497 assert(bld->static_state->normalized_coords);
498 coord = lp_build_mul(coord_bld, coord, length_f);
500 /* itrunc == ifloor here */
501 icoord = lp_build_itrunc(coord_bld, coord);
503 /* clamp to [0, length - 1] */
504 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
507 case PIPE_TEX_WRAP_MIRROR_CLAMP:
508 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
509 coord = lp_build_abs(coord_bld, coord);
511 if (bld->static_state->normalized_coords) {
512 /* scale coord to length */
513 coord = lp_build_mul(coord_bld, coord, length_f);
516 /* itrunc == ifloor here */
517 icoord = lp_build_itrunc(coord_bld, coord);
519 /* clamp to [0, length - 1] */
520 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
523 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
524 coord = lp_build_abs(coord_bld, coord);
526 if (bld->static_state->normalized_coords) {
527 /* scale coord to length */
528 coord = lp_build_mul(coord_bld, coord, length_f);
531 /* itrunc == ifloor here */
532 icoord = lp_build_itrunc(coord_bld, coord);
534 /* clamp to [0, length] */
535 icoord = lp_build_min(int_coord_bld, icoord, length);
548 * Generate code to sample a mipmap level with nearest filtering.
549 * If sampling a cube texture, r = cube face in [0,5].
552 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
555 LLVMValueRef row_stride_vec,
556 LLVMValueRef img_stride_vec,
557 LLVMValueRef data_ptr,
561 LLVMValueRef colors_out[4])
563 const unsigned dims = bld->dims;
564 LLVMValueRef width_vec;
565 LLVMValueRef height_vec;
566 LLVMValueRef depth_vec;
567 LLVMValueRef flt_size;
568 LLVMValueRef flt_width_vec;
569 LLVMValueRef flt_height_vec;
570 LLVMValueRef flt_depth_vec;
571 LLVMValueRef x, y, z;
573 lp_build_extract_image_sizes(bld,
577 &width_vec, &height_vec, &depth_vec);
579 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
581 lp_build_extract_image_sizes(bld,
582 bld->float_size_type,
585 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
588 * Compute integer texcoords.
590 x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec,
591 bld->static_state->pot_width,
592 bld->static_state->wrap_s);
593 lp_build_name(x, "tex.x.wrapped");
596 y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec,
597 bld->static_state->pot_height,
598 bld->static_state->wrap_t);
599 lp_build_name(y, "tex.y.wrapped");
602 z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec,
603 bld->static_state->pot_depth,
604 bld->static_state->wrap_r);
605 lp_build_name(z, "tex.z.wrapped");
607 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
619 * Get texture colors.
621 lp_build_sample_texel_soa(bld, unit,
622 width_vec, height_vec, depth_vec,
624 row_stride_vec, img_stride_vec,
625 data_ptr, colors_out);
630 * Generate code to sample a mipmap level with linear filtering.
631 * If sampling a cube texture, r = cube face in [0,5].
634 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
637 LLVMValueRef row_stride_vec,
638 LLVMValueRef img_stride_vec,
639 LLVMValueRef data_ptr,
643 LLVMValueRef colors_out[4])
645 const unsigned dims = bld->dims;
646 LLVMValueRef width_vec;
647 LLVMValueRef height_vec;
648 LLVMValueRef depth_vec;
649 LLVMValueRef flt_size;
650 LLVMValueRef flt_width_vec;
651 LLVMValueRef flt_height_vec;
652 LLVMValueRef flt_depth_vec;
653 LLVMValueRef x0, y0, z0, x1, y1, z1;
654 LLVMValueRef s_fpart, t_fpart, r_fpart;
655 LLVMValueRef neighbors[2][2][4];
658 lp_build_extract_image_sizes(bld,
662 &width_vec, &height_vec, &depth_vec);
664 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
666 lp_build_extract_image_sizes(bld,
667 bld->float_size_type,
670 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
673 * Compute integer texcoords.
675 lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec,
676 bld->static_state->pot_width,
677 bld->static_state->wrap_s,
679 lp_build_name(x0, "tex.x0.wrapped");
680 lp_build_name(x1, "tex.x1.wrapped");
683 lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec,
684 bld->static_state->pot_height,
685 bld->static_state->wrap_t,
687 lp_build_name(y0, "tex.y0.wrapped");
688 lp_build_name(y1, "tex.y1.wrapped");
691 lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec,
692 bld->static_state->pot_depth,
693 bld->static_state->wrap_r,
695 lp_build_name(z0, "tex.z0.wrapped");
696 lp_build_name(z1, "tex.z1.wrapped");
698 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
699 z0 = z1 = r; /* cube face */
708 y0 = y1 = t_fpart = NULL;
709 z0 = z1 = r_fpart = NULL;
713 * Get texture colors.
715 /* get x0/x1 texels */
716 lp_build_sample_texel_soa(bld, unit,
717 width_vec, height_vec, depth_vec,
719 row_stride_vec, img_stride_vec,
720 data_ptr, neighbors[0][0]);
721 lp_build_sample_texel_soa(bld, unit,
722 width_vec, height_vec, depth_vec,
724 row_stride_vec, img_stride_vec,
725 data_ptr, neighbors[0][1]);
728 /* Interpolate two samples from 1D image to produce one color */
729 for (chan = 0; chan < 4; chan++) {
730 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
731 neighbors[0][0][chan],
732 neighbors[0][1][chan]);
737 LLVMValueRef colors0[4];
739 /* get x0/x1 texels at y1 */
740 lp_build_sample_texel_soa(bld, unit,
741 width_vec, height_vec, depth_vec,
743 row_stride_vec, img_stride_vec,
744 data_ptr, neighbors[1][0]);
745 lp_build_sample_texel_soa(bld, unit,
746 width_vec, height_vec, depth_vec,
748 row_stride_vec, img_stride_vec,
749 data_ptr, neighbors[1][1]);
751 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
752 for (chan = 0; chan < 4; chan++) {
753 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
755 neighbors[0][0][chan],
756 neighbors[0][1][chan],
757 neighbors[1][0][chan],
758 neighbors[1][1][chan]);
762 LLVMValueRef neighbors1[2][2][4];
763 LLVMValueRef colors1[4];
765 /* get x0/x1/y0/y1 texels at z1 */
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, neighbors1[0][0]);
771 lp_build_sample_texel_soa(bld, unit,
772 width_vec, height_vec, depth_vec,
774 row_stride_vec, img_stride_vec,
775 data_ptr, neighbors1[0][1]);
776 lp_build_sample_texel_soa(bld, unit,
777 width_vec, height_vec, depth_vec,
779 row_stride_vec, img_stride_vec,
780 data_ptr, neighbors1[1][0]);
781 lp_build_sample_texel_soa(bld, unit,
782 width_vec, height_vec, depth_vec,
784 row_stride_vec, img_stride_vec,
785 data_ptr, neighbors1[1][1]);
787 /* Bilinear interpolate the four samples from the second Z slice */
788 for (chan = 0; chan < 4; chan++) {
789 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
791 neighbors1[0][0][chan],
792 neighbors1[0][1][chan],
793 neighbors1[1][0][chan],
794 neighbors1[1][1][chan]);
797 /* Linearly interpolate the two samples from the two 3D slices */
798 for (chan = 0; chan < 4; chan++) {
799 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
801 colors0[chan], colors1[chan]);
806 for (chan = 0; chan < 4; chan++) {
807 colors_out[chan] = colors0[chan];
815 * Sample the texture/mipmap using given image filter and mip filter.
816 * data0_ptr and data1_ptr point to the two mipmap levels to sample
817 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
818 * If we're using nearest miplevel sampling the '1' values will be null/unused.
821 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
828 LLVMValueRef ilevel0,
829 LLVMValueRef ilevel1,
830 LLVMValueRef lod_fpart,
831 LLVMValueRef *colors_out)
833 LLVMBuilderRef builder = bld->builder;
836 LLVMValueRef row_stride0_vec;
837 LLVMValueRef row_stride1_vec;
838 LLVMValueRef img_stride0_vec;
839 LLVMValueRef img_stride1_vec;
840 LLVMValueRef data_ptr0;
841 LLVMValueRef data_ptr1;
842 LLVMValueRef colors0[4], colors1[4];
845 /* sample the first mipmap level */
846 lp_build_mipmap_level_sizes(bld, ilevel0,
848 &row_stride0_vec, &img_stride0_vec);
849 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
850 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
851 lp_build_sample_image_nearest(bld, unit,
853 row_stride0_vec, img_stride0_vec,
858 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
859 lp_build_sample_image_linear(bld, unit,
861 row_stride0_vec, img_stride0_vec,
866 /* Store the first level's colors in the output variables */
867 for (chan = 0; chan < 4; chan++) {
868 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
871 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
872 struct lp_build_if_state if_ctx;
873 LLVMValueRef need_lerp;
875 /* need_lerp = lod_fpart > 0 */
876 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
881 lp_build_if(&if_ctx, builder, need_lerp);
883 /* sample the second mipmap level */
884 lp_build_mipmap_level_sizes(bld, ilevel1,
886 &row_stride1_vec, &img_stride1_vec);
887 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
888 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
889 lp_build_sample_image_nearest(bld, unit,
891 row_stride1_vec, img_stride1_vec,
896 lp_build_sample_image_linear(bld, unit,
898 row_stride1_vec, img_stride1_vec,
903 /* interpolate samples from the two mipmap levels */
905 lod_fpart = lp_build_broadcast_scalar(&bld->texel_bld, lod_fpart);
907 for (chan = 0; chan < 4; chan++) {
908 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
909 colors0[chan], colors1[chan]);
910 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
913 lp_build_endif(&if_ctx);
920 * General texture sampling codegen.
921 * This function handles texture sampling for all texture targets (1D,
922 * 2D, 3D, cube) and all filtering modes.
925 lp_build_sample_general(struct lp_build_sample_context *bld,
930 const LLVMValueRef *ddx,
931 const LLVMValueRef *ddy,
932 LLVMValueRef lod_bias, /* optional */
933 LLVMValueRef explicit_lod, /* optional */
934 LLVMValueRef *colors_out)
936 struct lp_build_context *int_bld = &bld->int_bld;
937 LLVMBuilderRef builder = bld->builder;
938 const unsigned mip_filter = bld->static_state->min_mip_filter;
939 const unsigned min_filter = bld->static_state->min_img_filter;
940 const unsigned mag_filter = bld->static_state->mag_img_filter;
941 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
942 LLVMValueRef ilevel0, ilevel1 = NULL;
943 LLVMValueRef face_ddx[4], face_ddy[4];
944 LLVMValueRef texels[4];
945 LLVMTypeRef i32t = LLVMInt32Type();
946 LLVMValueRef i32t_zero = LLVMConstInt(i32t, 0, 0);
950 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
951 mip_filter, min_filter, mag_filter);
955 * Choose cube face, recompute texcoords and derivatives for the chosen face.
957 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
958 LLVMValueRef face, face_s, face_t;
959 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
960 s = face_s; /* vec */
961 t = face_t; /* vec */
962 /* use 'r' to indicate cube face */
963 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
965 /* recompute ddx, ddy using the new (s,t) face texcoords */
966 face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
967 face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
970 face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
971 face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
979 * Compute the level of detail (float).
981 if (min_filter != mag_filter ||
982 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
983 /* Need to compute lod either to choose mipmap levels or to
984 * distinguish between minification/magnification with one mipmap level.
986 lp_build_lod_selector(bld, unit, ddx, ddy,
987 lod_bias, explicit_lod,
989 &lod_ipart, &lod_fpart);
991 lod_ipart = i32t_zero;
995 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
997 switch (mip_filter) {
999 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1001 case PIPE_TEX_MIPFILTER_NONE:
1002 /* always use mip level 0 */
1003 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1004 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1005 * We should be able to set ilevel0 = const(0) but that causes
1006 * bad x86 code to be emitted.
1009 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1012 ilevel0 = i32t_zero;
1015 case PIPE_TEX_MIPFILTER_NEAREST:
1017 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1019 case PIPE_TEX_MIPFILTER_LINEAR:
1022 lp_build_linear_mip_levels(bld, unit,
1023 lod_ipart, &lod_fpart,
1024 &ilevel0, &ilevel1);
1029 * Get/interpolate texture colors.
1032 for (chan = 0; chan < 4; ++chan) {
1033 texels[chan] = lp_build_alloca(builder, bld->texel_bld.vec_type, "");
1034 lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]);
1037 if (min_filter == mag_filter) {
1038 /* no need to distinquish between minification and magnification */
1039 lp_build_sample_mipmap(bld, unit,
1040 min_filter, mip_filter,
1042 ilevel0, ilevel1, lod_fpart,
1046 /* Emit conditional to choose min image filter or mag image filter
1047 * depending on the lod being > 0 or <= 0, respectively.
1049 struct lp_build_if_state if_ctx;
1050 LLVMValueRef minify;
1052 /* minify = lod >= 0.0 */
1053 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1054 lod_ipart, int_bld->zero, "");
1056 lp_build_if(&if_ctx, builder, minify);
1058 /* Use the minification filter */
1059 lp_build_sample_mipmap(bld, unit,
1060 min_filter, mip_filter,
1062 ilevel0, ilevel1, lod_fpart,
1065 lp_build_else(&if_ctx);
1067 /* Use the magnification filter */
1068 lp_build_sample_mipmap(bld, unit,
1069 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1071 i32t_zero, NULL, NULL,
1074 lp_build_endif(&if_ctx);
1077 for (chan = 0; chan < 4; ++chan) {
1078 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1079 lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]);
1085 * Do shadow test/comparison.
1086 * \param p the texcoord Z (aka R, aka P) component
1087 * \param texel the texel to compare against (use the X channel)
1090 lp_build_sample_compare(struct lp_build_sample_context *bld,
1092 LLVMValueRef texel[4])
1094 struct lp_build_context *texel_bld = &bld->texel_bld;
1096 const unsigned chan = 0;
1098 if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1103 LLVMValueRef indx = lp_build_const_int32(0);
1104 LLVMValueRef coord = LLVMBuildExtractElement(bld->builder, p, indx, "");
1105 LLVMValueRef tex = LLVMBuildExtractElement(bld->builder,
1106 texel[chan], indx, "");
1107 lp_build_printf(bld->builder, "shadow compare coord %f to texture %f\n",
1111 /* result = (p FUNC texel) ? 1 : 0 */
1112 res = lp_build_cmp(texel_bld, bld->static_state->compare_func,
1114 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1116 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1120 texel[3] = texel_bld->one;
1125 * Just set texels to white instead of actually sampling the texture.
1129 lp_build_sample_nop(struct lp_type type,
1130 LLVMValueRef texel_out[4])
1132 LLVMValueRef one = lp_build_one(type);
1135 for (chan = 0; chan < 4; chan++) {
1136 texel_out[chan] = one;
1142 * Build texture sampling code.
1143 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1145 * \param type vector float type to use for coords, etc.
1146 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1147 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1150 lp_build_sample_soa(LLVMBuilderRef builder,
1151 const struct lp_sampler_static_state *static_state,
1152 struct lp_sampler_dynamic_state *dynamic_state,
1153 struct lp_type type,
1155 unsigned num_coords,
1156 const LLVMValueRef *coords,
1157 const LLVMValueRef ddx[4],
1158 const LLVMValueRef ddy[4],
1159 LLVMValueRef lod_bias, /* optional */
1160 LLVMValueRef explicit_lod, /* optional */
1161 LLVMValueRef texel_out[4])
1163 unsigned dims = texture_dims(static_state->target);
1164 struct lp_build_sample_context bld;
1165 LLVMTypeRef i32t = LLVMInt32Type();
1170 struct lp_type float_vec_type;
1173 enum pipe_format fmt = static_state->format;
1174 debug_printf("Sample from %s\n", util_format_name(fmt));
1177 assert(type.floating);
1179 /* Setup our build context */
1180 memset(&bld, 0, sizeof bld);
1181 bld.builder = builder;
1182 bld.static_state = static_state;
1183 bld.dynamic_state = dynamic_state;
1184 bld.format_desc = util_format_description(static_state->format);
1187 bld.float_type = lp_type_float(32);
1188 bld.int_type = lp_type_int(32);
1189 bld.coord_type = type;
1190 bld.int_coord_type = lp_int_type(type);
1191 bld.float_size_type = lp_type_float(32);
1192 bld.float_size_type.length = dims > 1 ? 4 : 1;
1193 bld.int_size_type = lp_int_type(bld.float_size_type);
1194 bld.texel_type = type;
1196 float_vec_type = lp_type_float_vec(32);
1198 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1199 lp_build_context_init(&bld.float_vec_bld, builder, float_vec_type);
1200 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1201 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1202 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1203 lp_build_context_init(&bld.int_size_bld, builder, bld.int_size_type);
1204 lp_build_context_init(&bld.float_size_bld, builder, bld.float_size_type);
1205 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1207 /* Get the dynamic state */
1208 bld.width = dynamic_state->width(dynamic_state, builder, unit);
1209 bld.height = dynamic_state->height(dynamic_state, builder, unit);
1210 bld.depth = dynamic_state->depth(dynamic_state, builder, unit);
1211 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1212 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
1213 bld.data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1214 /* Note that data_array is an array[level] of pointers to texture images */
1220 /* width, height, depth as single int vector */
1222 bld.int_size = bld.width;
1225 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_bld.undef,
1226 bld.width, LLVMConstInt(i32t, 0, 0), "");
1228 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1229 bld.height, LLVMConstInt(i32t, 1, 0), "");
1231 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1232 bld.depth, LLVMConstInt(i32t, 2, 0), "");
1238 /* For debug: no-op texture sampling */
1239 lp_build_sample_nop(bld.texel_type, texel_out);
1241 else if (util_format_fits_8unorm(bld.format_desc) &&
1242 lp_is_simple_wrap_mode(static_state->wrap_s) &&
1243 lp_is_simple_wrap_mode(static_state->wrap_t)) {
1244 /* do sampling/filtering with fixed pt arithmetic */
1245 lp_build_sample_aos(&bld, unit, s, t, r, ddx, ddy,
1246 lod_bias, explicit_lod,
1251 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1252 util_format_fits_8unorm(bld.format_desc)) {
1253 debug_printf("%s: using floating point linear filtering for %s\n",
1254 __FUNCTION__, bld.format_desc->short_name);
1255 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1256 static_state->min_img_filter,
1257 static_state->mag_img_filter,
1258 static_state->min_mip_filter,
1259 static_state->wrap_s,
1260 static_state->wrap_t);
1263 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
1264 lod_bias, explicit_lod,
1268 lp_build_sample_compare(&bld, r, texel_out);