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
->uint_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
->uint_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 struct lp_build_context
*uint_coord_bld
= &bld
->uint_coord_bld
;
243 LLVMValueRef half
= lp_build_const_vec(coord_bld
->type
, 0.5);
244 LLVMValueRef length_minus_one
= lp_build_sub(uint_coord_bld
, length
, uint_coord_bld
->one
);
245 LLVMValueRef coord0
, coord1
, weight
;
248 case PIPE_TEX_WRAP_REPEAT
:
249 /* mul by size and subtract 0.5 */
250 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
251 coord
= lp_build_sub(coord_bld
, coord
, half
);
252 /* convert to int, compute lerp weight */
253 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
256 coord1
= lp_build_add(uint_coord_bld
, coord0
, uint_coord_bld
->one
);
257 coord0
= LLVMBuildAnd(bld
->builder
, coord0
, length_minus_one
, "");
258 coord1
= LLVMBuildAnd(bld
->builder
, coord1
, length_minus_one
, "");
261 /* Add a bias to the texcoord to handle negative coords */
262 LLVMValueRef bias
= lp_build_mul_imm(uint_coord_bld
, length
, 1024);
264 coord0
= LLVMBuildAdd(bld
->builder
, coord0
, bias
, "");
265 coord0
= LLVMBuildURem(bld
->builder
, coord0
, length
, "");
266 mask
= lp_build_compare(bld
->builder
, int_coord_bld
->type
,
267 PIPE_FUNC_NOTEQUAL
, coord0
, length_minus_one
);
268 coord1
= LLVMBuildAnd(bld
->builder
,
269 lp_build_add(uint_coord_bld
, coord0
, uint_coord_bld
->one
),
274 case PIPE_TEX_WRAP_CLAMP
:
275 if (bld
->static_state
->normalized_coords
) {
276 /* scale coord to length */
277 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
280 /* clamp to [0, length] */
281 coord
= lp_build_clamp(coord_bld
, coord
, coord_bld
->zero
, length_f
);
283 coord
= lp_build_sub(coord_bld
, coord
, half
);
285 /* convert to int, compute lerp weight */
286 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
287 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
290 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
292 struct lp_build_context abs_coord_bld
= bld
->coord_bld
;
293 abs_coord_bld
.type
.sign
= FALSE
;
295 if (bld
->static_state
->normalized_coords
) {
296 /* mul by tex size */
297 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
);
305 /* XXX this is odd normalized ranges from 0 to length-0.5 after denorm
306 but non-normalized ranges from to 0.5 to length-0.5 after clamp.
307 Is this missing the sub 0.5? */
309 LLVMValueRef min
, max
;
310 /* clamp to [0.5, length - 0.5] */
312 max
= lp_build_sub(coord_bld
, length_f
, min
);
313 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
315 /* convert to int, compute lerp weight */
316 lp_build_ifloor_fract(&abs_coord_bld
, coord
, &coord0
, &weight
);
317 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
318 /* coord1 = min(coord1, length-1) */
319 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
323 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
326 if (bld
->static_state
->normalized_coords
) {
327 /* scale coord to length */
328 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
330 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
331 coord
= lp_build_sub(coord_bld
, coord
, half
);
332 min
= lp_build_const_vec(coord_bld
->type
, -1.0F
);
333 coord
= lp_build_clamp(coord_bld
, coord
, min
, length_f
);
334 /* convert to int, compute lerp weight */
335 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
336 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
340 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
341 /* compute mirror function */
342 coord
= lp_build_coord_mirror(bld
, coord
);
344 /* scale coord to length */
345 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
346 coord
= lp_build_sub(coord_bld
, coord
, half
);
348 /* convert to int, compute lerp weight */
349 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
350 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
352 /* coord0 = max(coord0, 0) */
353 coord0
= lp_build_max(int_coord_bld
, coord0
, int_coord_bld
->zero
);
354 /* coord1 = min(coord1, length-1) */
355 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
358 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
359 coord
= lp_build_abs(coord_bld
, coord
);
361 if (bld
->static_state
->normalized_coords
) {
362 /* scale coord to length */
363 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
366 /* clamp to [0, length] */
367 coord
= lp_build_min(coord_bld
, coord
, length_f
);
369 coord
= lp_build_sub(coord_bld
, coord
, half
);
371 /* convert to int, compute lerp weight */
372 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
373 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
376 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
378 LLVMValueRef min
, max
;
379 struct lp_build_context abs_coord_bld
= bld
->coord_bld
;
380 abs_coord_bld
.type
.sign
= FALSE
;
381 coord
= lp_build_abs(coord_bld
, coord
);
383 if (bld
->static_state
->normalized_coords
) {
384 /* scale coord to length */
385 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
388 /* clamp to [0.5, length - 0.5] */
390 max
= lp_build_sub(coord_bld
, length_f
, min
);
391 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
393 coord
= lp_build_sub(coord_bld
, coord
, half
);
395 /* convert to int, compute lerp weight */
396 lp_build_ifloor_fract(&abs_coord_bld
, coord
, &coord0
, &weight
);
397 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
401 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
403 coord
= lp_build_abs(coord_bld
, coord
);
405 if (bld
->static_state
->normalized_coords
) {
406 /* scale coord to length */
407 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
410 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
411 /* skip -0.5 clamp (always positive), do sub first */
412 coord
= lp_build_sub(coord_bld
, coord
, half
);
413 coord
= lp_build_min(coord_bld
, coord
, length_f
);
415 /* convert to int, compute lerp weight */
416 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
417 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
430 *weight_out
= weight
;
435 * Build LLVM code for texture wrap mode for nearest filtering.
436 * \param coord the incoming texcoord (nominally in [0,1])
437 * \param length the texture size along one dimension, as int vector
438 * \param is_pot if TRUE, length is a power of two
439 * \param wrap_mode one of PIPE_TEX_WRAP_x
442 lp_build_sample_wrap_nearest(struct lp_build_sample_context
*bld
,
445 LLVMValueRef length_f
,
449 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
450 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
451 struct lp_build_context
*uint_coord_bld
= &bld
->uint_coord_bld
;
452 LLVMValueRef length_minus_one
= lp_build_sub(uint_coord_bld
, length
, uint_coord_bld
->one
);
456 case PIPE_TEX_WRAP_REPEAT
:
457 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
458 icoord
= lp_build_ifloor(coord_bld
, coord
);
460 icoord
= LLVMBuildAnd(bld
->builder
, icoord
, length_minus_one
, "");
462 /* Add a bias to the texcoord to handle negative coords */
463 LLVMValueRef bias
= lp_build_mul_imm(uint_coord_bld
, length
, 1024);
464 icoord
= LLVMBuildAdd(bld
->builder
, icoord
, bias
, "");
465 icoord
= LLVMBuildURem(bld
->builder
, icoord
, length
, "");
469 case PIPE_TEX_WRAP_CLAMP
:
470 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
471 if (bld
->static_state
->normalized_coords
) {
472 /* scale coord to length */
473 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
477 /* use itrunc instead since we clamp to 0 anyway */
478 icoord
= lp_build_itrunc(coord_bld
, coord
);
480 /* clamp to [0, length - 1]. */
481 icoord
= lp_build_clamp(int_coord_bld
, icoord
, int_coord_bld
->zero
,
485 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
486 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
488 LLVMValueRef min
, max
;
490 if (bld
->static_state
->normalized_coords
) {
491 /* scale coord to length */
492 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
495 icoord
= lp_build_ifloor(coord_bld
, coord
);
497 /* clamp to [-1, length] */
498 min
= lp_build_negate(int_coord_bld
, int_coord_bld
->one
);
500 icoord
= lp_build_clamp(int_coord_bld
, icoord
, min
, max
);
504 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
505 /* compute mirror function */
506 coord
= lp_build_coord_mirror(bld
, coord
);
508 /* scale coord to length */
509 assert(bld
->static_state
->normalized_coords
);
510 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
512 /* itrunc == ifloor here */
513 icoord
= lp_build_itrunc(coord_bld
, coord
);
515 /* clamp to [0, length - 1] */
516 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
519 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
520 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
521 coord
= lp_build_abs(coord_bld
, coord
);
523 if (bld
->static_state
->normalized_coords
) {
524 /* scale coord to length */
525 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
528 /* itrunc == ifloor here */
529 icoord
= lp_build_itrunc(coord_bld
, coord
);
531 /* clamp to [0, length - 1] */
532 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
535 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
536 coord
= lp_build_abs(coord_bld
, coord
);
538 if (bld
->static_state
->normalized_coords
) {
539 /* scale coord to length */
540 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] */
547 icoord
= lp_build_min(int_coord_bld
, icoord
, length
);
560 * Generate code to sample a mipmap level with nearest filtering.
561 * If sampling a cube texture, r = cube face in [0,5].
564 lp_build_sample_image_nearest(struct lp_build_sample_context
*bld
,
567 LLVMValueRef row_stride_vec
,
568 LLVMValueRef img_stride_vec
,
569 LLVMValueRef data_ptr
,
573 LLVMValueRef colors_out
[4])
575 const unsigned dims
= bld
->dims
;
576 LLVMValueRef width_vec
;
577 LLVMValueRef height_vec
;
578 LLVMValueRef depth_vec
;
579 LLVMValueRef flt_size
;
580 LLVMValueRef flt_width_vec
;
581 LLVMValueRef flt_height_vec
;
582 LLVMValueRef flt_depth_vec
;
583 LLVMValueRef x
, y
, z
;
585 lp_build_extract_image_sizes(bld
,
589 &width_vec
, &height_vec
, &depth_vec
);
591 flt_size
= lp_build_int_to_float(&bld
->float_size_bld
, size
);
593 lp_build_extract_image_sizes(bld
,
594 bld
->float_size_type
,
597 &flt_width_vec
, &flt_height_vec
, &flt_depth_vec
);
600 * Compute integer texcoords.
602 x
= lp_build_sample_wrap_nearest(bld
, s
, width_vec
, flt_width_vec
,
603 bld
->static_state
->pot_width
,
604 bld
->static_state
->wrap_s
);
605 lp_build_name(x
, "tex.x.wrapped");
608 y
= lp_build_sample_wrap_nearest(bld
, t
, height_vec
, flt_height_vec
,
609 bld
->static_state
->pot_height
,
610 bld
->static_state
->wrap_t
);
611 lp_build_name(y
, "tex.y.wrapped");
614 z
= lp_build_sample_wrap_nearest(bld
, r
, depth_vec
, flt_depth_vec
,
615 bld
->static_state
->pot_depth
,
616 bld
->static_state
->wrap_r
);
617 lp_build_name(z
, "tex.z.wrapped");
619 else if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
631 * Get texture colors.
633 lp_build_sample_texel_soa(bld
, unit
,
634 width_vec
, height_vec
, depth_vec
,
636 row_stride_vec
, img_stride_vec
,
637 data_ptr
, colors_out
);
642 * Generate code to sample a mipmap level with linear filtering.
643 * If sampling a cube texture, r = cube face in [0,5].
646 lp_build_sample_image_linear(struct lp_build_sample_context
*bld
,
649 LLVMValueRef row_stride_vec
,
650 LLVMValueRef img_stride_vec
,
651 LLVMValueRef data_ptr
,
655 LLVMValueRef colors_out
[4])
657 const unsigned dims
= bld
->dims
;
658 LLVMValueRef width_vec
;
659 LLVMValueRef height_vec
;
660 LLVMValueRef depth_vec
;
661 LLVMValueRef flt_size
;
662 LLVMValueRef flt_width_vec
;
663 LLVMValueRef flt_height_vec
;
664 LLVMValueRef flt_depth_vec
;
665 LLVMValueRef x0
, y0
, z0
, x1
, y1
, z1
;
666 LLVMValueRef s_fpart
, t_fpart
, r_fpart
;
667 LLVMValueRef neighbors
[2][2][4];
670 lp_build_extract_image_sizes(bld
,
674 &width_vec
, &height_vec
, &depth_vec
);
676 flt_size
= lp_build_int_to_float(&bld
->float_size_bld
, size
);
678 lp_build_extract_image_sizes(bld
,
679 bld
->float_size_type
,
682 &flt_width_vec
, &flt_height_vec
, &flt_depth_vec
);
685 * Compute integer texcoords.
687 lp_build_sample_wrap_linear(bld
, s
, width_vec
, flt_width_vec
,
688 bld
->static_state
->pot_width
,
689 bld
->static_state
->wrap_s
,
691 lp_build_name(x0
, "tex.x0.wrapped");
692 lp_build_name(x1
, "tex.x1.wrapped");
695 lp_build_sample_wrap_linear(bld
, t
, height_vec
, flt_height_vec
,
696 bld
->static_state
->pot_height
,
697 bld
->static_state
->wrap_t
,
699 lp_build_name(y0
, "tex.y0.wrapped");
700 lp_build_name(y1
, "tex.y1.wrapped");
703 lp_build_sample_wrap_linear(bld
, r
, depth_vec
, flt_depth_vec
,
704 bld
->static_state
->pot_depth
,
705 bld
->static_state
->wrap_r
,
707 lp_build_name(z0
, "tex.z0.wrapped");
708 lp_build_name(z1
, "tex.z1.wrapped");
710 else if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
711 z0
= z1
= r
; /* cube face */
720 y0
= y1
= t_fpart
= NULL
;
721 z0
= z1
= r_fpart
= NULL
;
725 * Get texture colors.
727 /* get x0/x1 texels */
728 lp_build_sample_texel_soa(bld
, unit
,
729 width_vec
, height_vec
, depth_vec
,
731 row_stride_vec
, img_stride_vec
,
732 data_ptr
, neighbors
[0][0]);
733 lp_build_sample_texel_soa(bld
, unit
,
734 width_vec
, height_vec
, depth_vec
,
736 row_stride_vec
, img_stride_vec
,
737 data_ptr
, neighbors
[0][1]);
740 /* Interpolate two samples from 1D image to produce one color */
741 for (chan
= 0; chan
< 4; chan
++) {
742 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
, s_fpart
,
743 neighbors
[0][0][chan
],
744 neighbors
[0][1][chan
]);
749 LLVMValueRef colors0
[4];
751 /* get x0/x1 texels at y1 */
752 lp_build_sample_texel_soa(bld
, unit
,
753 width_vec
, height_vec
, depth_vec
,
755 row_stride_vec
, img_stride_vec
,
756 data_ptr
, neighbors
[1][0]);
757 lp_build_sample_texel_soa(bld
, unit
,
758 width_vec
, height_vec
, depth_vec
,
760 row_stride_vec
, img_stride_vec
,
761 data_ptr
, neighbors
[1][1]);
763 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
764 for (chan
= 0; chan
< 4; chan
++) {
765 colors0
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
767 neighbors
[0][0][chan
],
768 neighbors
[0][1][chan
],
769 neighbors
[1][0][chan
],
770 neighbors
[1][1][chan
]);
774 LLVMValueRef neighbors1
[2][2][4];
775 LLVMValueRef colors1
[4];
777 /* get x0/x1/y0/y1 texels at z1 */
778 lp_build_sample_texel_soa(bld
, unit
,
779 width_vec
, height_vec
, depth_vec
,
781 row_stride_vec
, img_stride_vec
,
782 data_ptr
, neighbors1
[0][0]);
783 lp_build_sample_texel_soa(bld
, unit
,
784 width_vec
, height_vec
, depth_vec
,
786 row_stride_vec
, img_stride_vec
,
787 data_ptr
, neighbors1
[0][1]);
788 lp_build_sample_texel_soa(bld
, unit
,
789 width_vec
, height_vec
, depth_vec
,
791 row_stride_vec
, img_stride_vec
,
792 data_ptr
, neighbors1
[1][0]);
793 lp_build_sample_texel_soa(bld
, unit
,
794 width_vec
, height_vec
, depth_vec
,
796 row_stride_vec
, img_stride_vec
,
797 data_ptr
, neighbors1
[1][1]);
799 /* Bilinear interpolate the four samples from the second Z slice */
800 for (chan
= 0; chan
< 4; chan
++) {
801 colors1
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
803 neighbors1
[0][0][chan
],
804 neighbors1
[0][1][chan
],
805 neighbors1
[1][0][chan
],
806 neighbors1
[1][1][chan
]);
809 /* Linearly interpolate the two samples from the two 3D slices */
810 for (chan
= 0; chan
< 4; chan
++) {
811 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
,
813 colors0
[chan
], colors1
[chan
]);
818 for (chan
= 0; chan
< 4; chan
++) {
819 colors_out
[chan
] = colors0
[chan
];
827 * Sample the texture/mipmap using given image filter and mip filter.
828 * data0_ptr and data1_ptr point to the two mipmap levels to sample
829 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
830 * If we're using nearest miplevel sampling the '1' values will be null/unused.
833 lp_build_sample_mipmap(struct lp_build_sample_context
*bld
,
840 LLVMValueRef ilevel0
,
841 LLVMValueRef ilevel1
,
842 LLVMValueRef lod_fpart
,
843 LLVMValueRef
*colors_out
)
845 LLVMBuilderRef builder
= bld
->builder
;
848 LLVMValueRef row_stride0_vec
;
849 LLVMValueRef row_stride1_vec
;
850 LLVMValueRef img_stride0_vec
;
851 LLVMValueRef img_stride1_vec
;
852 LLVMValueRef data_ptr0
;
853 LLVMValueRef data_ptr1
;
854 LLVMValueRef colors0
[4], colors1
[4];
857 /* sample the first mipmap level */
858 lp_build_mipmap_level_sizes(bld
, ilevel0
,
860 &row_stride0_vec
, &img_stride0_vec
);
861 data_ptr0
= lp_build_get_mipmap_level(bld
, ilevel0
);
862 if (img_filter
== PIPE_TEX_FILTER_NEAREST
) {
863 lp_build_sample_image_nearest(bld
, unit
,
865 row_stride0_vec
, img_stride0_vec
,
870 assert(img_filter
== PIPE_TEX_FILTER_LINEAR
);
871 lp_build_sample_image_linear(bld
, unit
,
873 row_stride0_vec
, img_stride0_vec
,
878 /* Store the first level's colors in the output variables */
879 for (chan
= 0; chan
< 4; chan
++) {
880 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
883 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
884 struct lp_build_flow_context
*flow_ctx
;
885 struct lp_build_if_state if_ctx
;
886 LLVMValueRef need_lerp
;
888 flow_ctx
= lp_build_flow_create(builder
);
890 /* need_lerp = lod_fpart > 0 */
891 need_lerp
= LLVMBuildFCmp(builder
, LLVMRealUGT
,
896 lp_build_if(&if_ctx
, flow_ctx
, builder
, need_lerp
);
898 /* sample the second mipmap level */
899 lp_build_mipmap_level_sizes(bld
, ilevel1
,
901 &row_stride1_vec
, &img_stride1_vec
);
902 data_ptr1
= lp_build_get_mipmap_level(bld
, ilevel1
);
903 if (img_filter
== PIPE_TEX_FILTER_NEAREST
) {
904 lp_build_sample_image_nearest(bld
, unit
,
906 row_stride1_vec
, img_stride1_vec
,
911 lp_build_sample_image_linear(bld
, unit
,
913 row_stride1_vec
, img_stride1_vec
,
918 /* interpolate samples from the two mipmap levels */
920 lod_fpart
= lp_build_broadcast_scalar(&bld
->texel_bld
, lod_fpart
);
922 for (chan
= 0; chan
< 4; chan
++) {
923 colors0
[chan
] = lp_build_lerp(&bld
->texel_bld
, lod_fpart
,
924 colors0
[chan
], colors1
[chan
]);
925 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
928 lp_build_endif(&if_ctx
);
930 lp_build_flow_destroy(flow_ctx
);
937 * General texture sampling codegen.
938 * This function handles texture sampling for all texture targets (1D,
939 * 2D, 3D, cube) and all filtering modes.
942 lp_build_sample_general(struct lp_build_sample_context
*bld
,
947 const LLVMValueRef
*ddx
,
948 const LLVMValueRef
*ddy
,
949 LLVMValueRef lod_bias
, /* optional */
950 LLVMValueRef explicit_lod
, /* optional */
951 LLVMValueRef
*colors_out
)
953 struct lp_build_context
*int_bld
= &bld
->int_bld
;
954 LLVMBuilderRef builder
= bld
->builder
;
955 const unsigned mip_filter
= bld
->static_state
->min_mip_filter
;
956 const unsigned min_filter
= bld
->static_state
->min_img_filter
;
957 const unsigned mag_filter
= bld
->static_state
->mag_img_filter
;
958 LLVMValueRef lod_ipart
= NULL
, lod_fpart
= NULL
;
959 LLVMValueRef ilevel0
, ilevel1
= NULL
;
960 LLVMValueRef face_ddx
[4], face_ddy
[4];
961 LLVMValueRef texels
[4];
962 LLVMTypeRef i32t
= LLVMInt32Type();
963 LLVMValueRef i32t_zero
= LLVMConstInt(i32t
, 0, 0);
967 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
968 mip_filter, min_filter, mag_filter);
972 * Choose cube face, recompute texcoords and derivatives for the chosen face.
974 if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
975 LLVMValueRef face
, face_s
, face_t
;
976 lp_build_cube_lookup(bld
, s
, t
, r
, &face
, &face_s
, &face_t
);
977 s
= face_s
; /* vec */
978 t
= face_t
; /* vec */
979 /* use 'r' to indicate cube face */
980 r
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, face
); /* vec */
982 /* recompute ddx, ddy using the new (s,t) face texcoords */
983 face_ddx
[0] = lp_build_ddx(&bld
->coord_bld
, s
);
984 face_ddx
[1] = lp_build_ddx(&bld
->coord_bld
, t
);
987 face_ddy
[0] = lp_build_ddy(&bld
->coord_bld
, s
);
988 face_ddy
[1] = lp_build_ddy(&bld
->coord_bld
, t
);
996 * Compute the level of detail (float).
998 if (min_filter
!= mag_filter
||
999 mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) {
1000 /* Need to compute lod either to choose mipmap levels or to
1001 * distinguish between minification/magnification with one mipmap level.
1003 lp_build_lod_selector(bld
, unit
, ddx
, ddy
,
1004 lod_bias
, explicit_lod
,
1006 &lod_ipart
, &lod_fpart
);
1008 lod_ipart
= i32t_zero
;
1012 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1014 switch (mip_filter
) {
1016 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1018 case PIPE_TEX_MIPFILTER_NONE
:
1019 /* always use mip level 0 */
1020 if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
1021 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1022 * We should be able to set ilevel0 = const(0) but that causes
1023 * bad x86 code to be emitted.
1026 lp_build_nearest_mip_level(bld
, unit
, lod_ipart
, &ilevel0
);
1029 ilevel0
= i32t_zero
;
1032 case PIPE_TEX_MIPFILTER_NEAREST
:
1034 lp_build_nearest_mip_level(bld
, unit
, lod_ipart
, &ilevel0
);
1036 case PIPE_TEX_MIPFILTER_LINEAR
:
1039 lp_build_linear_mip_levels(bld
, unit
,
1040 lod_ipart
, &lod_fpart
,
1041 &ilevel0
, &ilevel1
);
1046 * Get/interpolate texture colors.
1049 for (chan
= 0; chan
< 4; ++chan
) {
1050 texels
[chan
] = lp_build_alloca(builder
, bld
->texel_bld
.vec_type
, "");
1051 lp_build_name(texels
[chan
], "sampler%u_texel_%c_var", unit
, "xyzw"[chan
]);
1054 if (min_filter
== mag_filter
) {
1055 /* no need to distinquish between minification and magnification */
1056 lp_build_sample_mipmap(bld
, unit
,
1057 min_filter
, mip_filter
,
1059 ilevel0
, ilevel1
, lod_fpart
,
1063 /* Emit conditional to choose min image filter or mag image filter
1064 * depending on the lod being > 0 or <= 0, respectively.
1066 struct lp_build_flow_context
*flow_ctx
;
1067 struct lp_build_if_state if_ctx
;
1068 LLVMValueRef minify
;
1070 flow_ctx
= lp_build_flow_create(builder
);
1072 /* minify = lod >= 0.0 */
1073 minify
= LLVMBuildICmp(builder
, LLVMIntSGE
,
1074 lod_ipart
, int_bld
->zero
, "");
1076 lp_build_if(&if_ctx
, flow_ctx
, builder
, minify
);
1078 /* Use the minification filter */
1079 lp_build_sample_mipmap(bld
, unit
,
1080 min_filter
, mip_filter
,
1082 ilevel0
, ilevel1
, lod_fpart
,
1085 lp_build_else(&if_ctx
);
1087 /* Use the magnification filter */
1088 lp_build_sample_mipmap(bld
, unit
,
1089 mag_filter
, PIPE_TEX_MIPFILTER_NONE
,
1091 i32t_zero
, NULL
, NULL
,
1094 lp_build_endif(&if_ctx
);
1096 lp_build_flow_destroy(flow_ctx
);
1099 for (chan
= 0; chan
< 4; ++chan
) {
1100 colors_out
[chan
] = LLVMBuildLoad(builder
, texels
[chan
], "");
1101 lp_build_name(colors_out
[chan
], "sampler%u_texel_%c", unit
, "xyzw"[chan
]);
1107 * Do shadow test/comparison.
1108 * \param p the texcoord Z (aka R, aka P) component
1109 * \param texel the texel to compare against (use the X channel)
1112 lp_build_sample_compare(struct lp_build_sample_context
*bld
,
1114 LLVMValueRef texel
[4])
1116 struct lp_build_context
*texel_bld
= &bld
->texel_bld
;
1118 const unsigned chan
= 0;
1120 if (bld
->static_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
)
1125 LLVMValueRef indx
= lp_build_const_int32(0);
1126 LLVMValueRef coord
= LLVMBuildExtractElement(bld
->builder
, p
, indx
, "");
1127 LLVMValueRef tex
= LLVMBuildExtractElement(bld
->builder
,
1128 texel
[chan
], indx
, "");
1129 lp_build_printf(bld
->builder
, "shadow compare coord %f to texture %f\n",
1133 /* result = (p FUNC texel) ? 1 : 0 */
1134 res
= lp_build_cmp(texel_bld
, bld
->static_state
->compare_func
,
1136 res
= lp_build_select(texel_bld
, res
, texel_bld
->one
, texel_bld
->zero
);
1138 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1142 texel
[3] = texel_bld
->one
;
1147 * Just set texels to white instead of actually sampling the texture.
1151 lp_build_sample_nop(struct lp_type type
,
1152 LLVMValueRef texel_out
[4])
1154 LLVMValueRef one
= lp_build_one(type
);
1157 for (chan
= 0; chan
< 4; chan
++) {
1158 texel_out
[chan
] = one
;
1164 * Build texture sampling code.
1165 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1167 * \param type vector float type to use for coords, etc.
1168 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1169 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1172 lp_build_sample_soa(LLVMBuilderRef builder
,
1173 const struct lp_sampler_static_state
*static_state
,
1174 struct lp_sampler_dynamic_state
*dynamic_state
,
1175 struct lp_type type
,
1177 unsigned num_coords
,
1178 const LLVMValueRef
*coords
,
1179 const LLVMValueRef ddx
[4],
1180 const LLVMValueRef ddy
[4],
1181 LLVMValueRef lod_bias
, /* optional */
1182 LLVMValueRef explicit_lod
, /* optional */
1183 LLVMValueRef texel_out
[4])
1185 unsigned dims
= texture_dims(static_state
->target
);
1186 struct lp_build_sample_context bld
;
1187 LLVMTypeRef i32t
= LLVMInt32Type();
1192 struct lp_type float_vec_type
;
1195 enum pipe_format fmt
= static_state
->format
;
1196 debug_printf("Sample from %s\n", util_format_name(fmt
));
1199 assert(type
.floating
);
1201 /* Setup our build context */
1202 memset(&bld
, 0, sizeof bld
);
1203 bld
.builder
= builder
;
1204 bld
.static_state
= static_state
;
1205 bld
.dynamic_state
= dynamic_state
;
1206 bld
.format_desc
= util_format_description(static_state
->format
);
1209 bld
.float_type
= lp_type_float(32);
1210 bld
.int_type
= lp_type_int(32);
1211 bld
.coord_type
= type
;
1212 bld
.uint_coord_type
= lp_uint_type(type
);
1213 bld
.int_coord_type
= lp_int_type(type
);
1214 bld
.float_size_type
= lp_type_float(32);
1215 bld
.float_size_type
.length
= dims
> 1 ? 4 : 1;
1216 bld
.int_size_type
= lp_int_type(bld
.float_size_type
);
1217 bld
.texel_type
= type
;
1219 float_vec_type
= lp_type_float_vec(32);
1221 lp_build_context_init(&bld
.float_bld
, builder
, bld
.float_type
);
1222 lp_build_context_init(&bld
.float_vec_bld
, builder
, float_vec_type
);
1223 lp_build_context_init(&bld
.int_bld
, builder
, bld
.int_type
);
1224 lp_build_context_init(&bld
.coord_bld
, builder
, bld
.coord_type
);
1225 lp_build_context_init(&bld
.uint_coord_bld
, builder
, bld
.uint_coord_type
);
1226 lp_build_context_init(&bld
.int_coord_bld
, builder
, bld
.int_coord_type
);
1227 lp_build_context_init(&bld
.int_size_bld
, builder
, bld
.int_size_type
);
1228 lp_build_context_init(&bld
.float_size_bld
, builder
, bld
.float_size_type
);
1229 lp_build_context_init(&bld
.texel_bld
, builder
, bld
.texel_type
);
1231 /* Get the dynamic state */
1232 bld
.width
= dynamic_state
->width(dynamic_state
, builder
, unit
);
1233 bld
.height
= dynamic_state
->height(dynamic_state
, builder
, unit
);
1234 bld
.depth
= dynamic_state
->depth(dynamic_state
, builder
, unit
);
1235 bld
.row_stride_array
= dynamic_state
->row_stride(dynamic_state
, builder
, unit
);
1236 bld
.img_stride_array
= dynamic_state
->img_stride(dynamic_state
, builder
, unit
);
1237 bld
.data_array
= dynamic_state
->data_ptr(dynamic_state
, builder
, unit
);
1238 /* Note that data_array is an array[level] of pointers to texture images */
1244 /* width, height, depth as single int vector */
1246 bld
.int_size
= bld
.width
;
1249 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size_bld
.undef
,
1250 bld
.width
, LLVMConstInt(i32t
, 0, 0), "");
1252 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size
,
1253 bld
.height
, LLVMConstInt(i32t
, 1, 0), "");
1255 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size
,
1256 bld
.depth
, LLVMConstInt(i32t
, 2, 0), "");
1262 /* For debug: no-op texture sampling */
1263 lp_build_sample_nop(bld
.texel_type
, texel_out
);
1265 else if (util_format_fits_8unorm(bld
.format_desc
) &&
1266 lp_is_simple_wrap_mode(static_state
->wrap_s
) &&
1267 lp_is_simple_wrap_mode(static_state
->wrap_t
)) {
1268 /* do sampling/filtering with fixed pt arithmetic */
1269 lp_build_sample_aos(&bld
, unit
, s
, t
, r
, ddx
, ddy
,
1270 lod_bias
, explicit_lod
,
1275 if ((gallivm_debug
& GALLIVM_DEBUG_PERF
) &&
1276 util_format_fits_8unorm(bld
.format_desc
)) {
1277 debug_printf("%s: using floating point linear filtering for %s\n",
1278 __FUNCTION__
, bld
.format_desc
->short_name
);
1279 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1280 static_state
->min_img_filter
,
1281 static_state
->mag_img_filter
,
1282 static_state
->min_mip_filter
,
1283 static_state
->wrap_s
,
1284 static_state
->wrap_t
);
1287 lp_build_sample_general(&bld
, unit
, s
, t
, r
, ddx
, ddy
,
1288 lod_bias
, explicit_lod
,
1292 lp_build_sample_compare(&bld
, r
, texel_out
);