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 "util/u_format_rgb9e5.h"
46 #include "lp_bld_debug.h"
47 #include "lp_bld_type.h"
48 #include "lp_bld_const.h"
49 #include "lp_bld_conv.h"
50 #include "lp_bld_arit.h"
51 #include "lp_bld_bitarit.h"
52 #include "lp_bld_logic.h"
53 #include "lp_bld_printf.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_flow.h"
56 #include "lp_bld_gather.h"
57 #include "lp_bld_format.h"
58 #include "lp_bld_sample.h"
59 #include "lp_bld_sample_aos.h"
60 #include "lp_bld_struct.h"
61 #include "lp_bld_quad.h"
62 #include "lp_bld_pack.h"
66 * Generate code to fetch a texel from a texture at int coords (x, y, z).
67 * The computation depends on whether the texture is 1D, 2D or 3D.
68 * The result, texel, will be float vectors:
69 * texel[0] = red values
70 * texel[1] = green values
71 * texel[2] = blue values
72 * texel[3] = alpha values
75 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_static_sampler_state
*static_state
= bld
->static_sampler_state
;
89 const unsigned dims
= bld
->dims
;
90 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
91 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
94 LLVMValueRef use_border
= NULL
;
96 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
97 if (lp_sampler_wrap_mode_uses_border_color(static_state
->wrap_s
,
98 static_state
->min_img_filter
,
99 static_state
->mag_img_filter
)) {
101 b1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, x
, int_coord_bld
->zero
);
102 b2
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, x
, width
);
103 use_border
= LLVMBuildOr(builder
, b1
, b2
, "b1_or_b2");
107 lp_sampler_wrap_mode_uses_border_color(static_state
->wrap_t
,
108 static_state
->min_img_filter
,
109 static_state
->mag_img_filter
)) {
111 b1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, y
, int_coord_bld
->zero
);
112 b2
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, y
, height
);
114 use_border
= LLVMBuildOr(builder
, use_border
, b1
, "ub_or_b1");
115 use_border
= LLVMBuildOr(builder
, use_border
, b2
, "ub_or_b2");
118 use_border
= LLVMBuildOr(builder
, b1
, b2
, "b1_or_b2");
123 lp_sampler_wrap_mode_uses_border_color(static_state
->wrap_r
,
124 static_state
->min_img_filter
,
125 static_state
->mag_img_filter
)) {
127 b1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, z
, int_coord_bld
->zero
);
128 b2
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, z
, depth
);
130 use_border
= LLVMBuildOr(builder
, use_border
, b1
, "ub_or_b1");
131 use_border
= LLVMBuildOr(builder
, use_border
, b2
, "ub_or_b2");
134 use_border
= LLVMBuildOr(builder
, b1
, b2
, "b1_or_b2");
138 /* convert x,y,z coords to linear offset from start of texture, in bytes */
139 lp_build_sample_offset(&bld
->int_coord_bld
,
141 x
, y
, z
, y_stride
, z_stride
,
144 offset
= lp_build_add(&bld
->int_coord_bld
, offset
, mipoffsets
);
148 /* If we can sample the border color, it means that texcoords may
149 * lie outside the bounds of the texture image. We need to do
150 * something to prevent reading out of bounds and causing a segfault.
152 * Simply AND the texture coords with !use_border. This will cause
153 * coords which are out of bounds to become zero. Zero's guaranteed
154 * to be inside the texture image.
156 offset
= lp_build_andnot(&bld
->int_coord_bld
, offset
, use_border
);
159 lp_build_fetch_rgba_soa(bld
->gallivm
,
167 * Note: if we find an app which frequently samples the texture border
168 * we might want to implement a true conditional here to avoid sampling
169 * the texture whenever possible (since that's quite a bit of code).
172 * texel = border_color;
175 * texel = sample_texture(coord);
177 * As it is now, we always sample the texture, then selectively replace
178 * the texel color results with the border color.
182 /* select texel color or border color depending on use_border. */
183 const struct util_format_description
*format_desc
= bld
->format_desc
;
185 struct lp_type border_type
= bld
->texel_type
;
186 border_type
.length
= 4;
188 * Only replace channels which are actually present. The others should
189 * get optimized away eventually by sampler_view swizzle anyway but it's
192 for (chan
= 0; chan
< 4; chan
++) {
194 /* reverse-map channel... */
195 for (chan_s
= 0; chan_s
< 4; chan_s
++) {
196 if (chan_s
== format_desc
->swizzle
[chan
]) {
201 /* use the already clamped color */
202 LLVMValueRef idx
= lp_build_const_int32(bld
->gallivm
, chan
);
203 LLVMValueRef border_chan
;
205 border_chan
= lp_build_extract_broadcast(bld
->gallivm
,
208 bld
->border_color_clamped
,
210 texel_out
[chan
] = lp_build_select(&bld
->texel_bld
, use_border
,
211 border_chan
, texel_out
[chan
]);
219 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
222 lp_build_coord_mirror(struct lp_build_sample_context
*bld
,
225 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
226 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
227 LLVMValueRef fract
, flr
, isOdd
;
229 lp_build_ifloor_fract(coord_bld
, coord
, &flr
, &fract
);
231 /* isOdd = flr & 1 */
232 isOdd
= LLVMBuildAnd(bld
->gallivm
->builder
, flr
, int_coord_bld
->one
, "");
234 /* make coord positive or negative depending on isOdd */
235 coord
= lp_build_set_sign(coord_bld
, fract
, isOdd
);
237 /* convert isOdd to float */
238 isOdd
= lp_build_int_to_float(coord_bld
, isOdd
);
240 /* add isOdd to coord */
241 coord
= lp_build_add(coord_bld
, coord
, isOdd
);
248 * Helper to compute the first coord and the weight for
249 * linear wrap repeat npot textures
252 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context
*bld
,
253 LLVMValueRef coord_f
,
254 LLVMValueRef length_i
,
255 LLVMValueRef length_f
,
256 LLVMValueRef
*coord0_i
,
257 LLVMValueRef
*weight_f
)
259 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
260 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
261 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, coord_bld
->type
, 0.5);
262 LLVMValueRef length_minus_one
= lp_build_sub(int_coord_bld
, length_i
,
265 /* wrap with normalized floats is just fract */
266 coord_f
= lp_build_fract(coord_bld
, coord_f
);
267 /* mul by size and subtract 0.5 */
268 coord_f
= lp_build_mul(coord_bld
, coord_f
, length_f
);
269 coord_f
= lp_build_sub(coord_bld
, coord_f
, half
);
271 * we avoided the 0.5/length division before the repeat wrap,
272 * now need to fix up edge cases with selects
274 /* convert to int, compute lerp weight */
275 lp_build_ifloor_fract(coord_bld
, coord_f
, coord0_i
, weight_f
);
276 mask
= lp_build_compare(int_coord_bld
->gallivm
, int_coord_bld
->type
,
277 PIPE_FUNC_LESS
, *coord0_i
, int_coord_bld
->zero
);
278 *coord0_i
= lp_build_select(int_coord_bld
, mask
, length_minus_one
, *coord0_i
);
283 * Build LLVM code for texture wrap mode for linear filtering.
284 * \param x0_out returns first integer texcoord
285 * \param x1_out returns second integer texcoord
286 * \param weight_out returns linear interpolation weight
289 lp_build_sample_wrap_linear(struct lp_build_sample_context
*bld
,
292 LLVMValueRef length_f
,
296 LLVMValueRef
*x0_out
,
297 LLVMValueRef
*x1_out
,
298 LLVMValueRef
*weight_out
)
300 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
301 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
302 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
303 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, coord_bld
->type
, 0.5);
304 LLVMValueRef length_minus_one
= lp_build_sub(int_coord_bld
, length
, int_coord_bld
->one
);
305 LLVMValueRef coord0
, coord1
, weight
;
308 case PIPE_TEX_WRAP_REPEAT
:
310 /* mul by size and subtract 0.5 */
311 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
312 coord
= lp_build_sub(coord_bld
, coord
, half
);
314 offset
= lp_build_int_to_float(coord_bld
, offset
);
315 coord
= lp_build_add(coord_bld
, coord
, offset
);
317 /* convert to int, compute lerp weight */
318 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
319 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
321 coord0
= LLVMBuildAnd(builder
, coord0
, length_minus_one
, "");
322 coord1
= LLVMBuildAnd(builder
, coord1
, length_minus_one
, "");
327 offset
= lp_build_int_to_float(coord_bld
, offset
);
328 offset
= lp_build_div(coord_bld
, offset
, length_f
);
329 coord
= lp_build_add(coord_bld
, coord
, offset
);
331 lp_build_coord_repeat_npot_linear(bld
, coord
,
334 mask
= lp_build_compare(int_coord_bld
->gallivm
, int_coord_bld
->type
,
335 PIPE_FUNC_NOTEQUAL
, coord0
, length_minus_one
);
336 coord1
= LLVMBuildAnd(builder
,
337 lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
),
342 case PIPE_TEX_WRAP_CLAMP
:
343 if (bld
->static_sampler_state
->normalized_coords
) {
344 /* scale coord to length */
345 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
348 offset
= lp_build_int_to_float(coord_bld
, offset
);
349 coord
= lp_build_add(coord_bld
, coord
, offset
);
352 /* clamp to [0, length] */
353 coord
= lp_build_clamp(coord_bld
, coord
, coord_bld
->zero
, length_f
);
355 coord
= lp_build_sub(coord_bld
, coord
, half
);
357 /* convert to int, compute lerp weight */
358 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
359 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
362 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
364 struct lp_build_context abs_coord_bld
= bld
->coord_bld
;
365 abs_coord_bld
.type
.sign
= FALSE
;
367 if (bld
->static_sampler_state
->normalized_coords
) {
368 /* mul by tex size */
369 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
372 offset
= lp_build_int_to_float(coord_bld
, offset
);
373 coord
= lp_build_add(coord_bld
, coord
, offset
);
376 /* clamp to length max */
377 coord
= lp_build_min(coord_bld
, coord
, length_f
);
379 coord
= lp_build_sub(coord_bld
, coord
, half
);
380 /* clamp to [0, length - 0.5] */
381 coord
= lp_build_max(coord_bld
, coord
, coord_bld
->zero
);
382 /* convert to int, compute lerp weight */
383 lp_build_ifloor_fract(&abs_coord_bld
, coord
, &coord0
, &weight
);
384 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
385 /* coord1 = min(coord1, length-1) */
386 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
390 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
391 if (bld
->static_sampler_state
->normalized_coords
) {
392 /* scale coord to length */
393 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
396 offset
= lp_build_int_to_float(coord_bld
, offset
);
397 coord
= lp_build_add(coord_bld
, coord
, offset
);
399 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
400 /* can skip clamp (though might not work for very large coord values */
401 coord
= lp_build_sub(coord_bld
, coord
, half
);
402 /* convert to int, compute lerp weight */
403 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
404 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
407 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
408 /* compute mirror function */
409 coord
= lp_build_coord_mirror(bld
, coord
);
411 /* scale coord to length */
412 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
413 coord
= lp_build_sub(coord_bld
, coord
, half
);
415 offset
= lp_build_int_to_float(coord_bld
, offset
);
416 coord
= lp_build_add(coord_bld
, coord
, offset
);
419 /* convert to int, compute lerp weight */
420 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
421 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
423 /* coord0 = max(coord0, 0) */
424 coord0
= lp_build_max(int_coord_bld
, coord0
, int_coord_bld
->zero
);
425 /* coord1 = min(coord1, length-1) */
426 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
429 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
430 if (bld
->static_sampler_state
->normalized_coords
) {
431 /* scale coord to length */
432 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
435 offset
= lp_build_int_to_float(coord_bld
, offset
);
436 coord
= lp_build_add(coord_bld
, coord
, offset
);
438 coord
= lp_build_abs(coord_bld
, coord
);
440 /* clamp to [0, length] */
441 coord
= lp_build_min(coord_bld
, coord
, length_f
);
443 coord
= lp_build_sub(coord_bld
, coord
, half
);
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
);
450 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
452 struct lp_build_context abs_coord_bld
= bld
->coord_bld
;
453 abs_coord_bld
.type
.sign
= FALSE
;
455 if (bld
->static_sampler_state
->normalized_coords
) {
456 /* scale coord to length */
457 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
460 offset
= lp_build_int_to_float(coord_bld
, offset
);
461 coord
= lp_build_add(coord_bld
, coord
, offset
);
463 coord
= lp_build_abs(coord_bld
, coord
);
465 /* clamp to length max */
466 coord
= lp_build_min(coord_bld
, coord
, length_f
);
468 coord
= lp_build_sub(coord_bld
, coord
, half
);
469 /* clamp to [0, length - 0.5] */
470 coord
= lp_build_max(coord_bld
, coord
, coord_bld
->zero
);
472 /* convert to int, compute lerp weight */
473 lp_build_ifloor_fract(&abs_coord_bld
, coord
, &coord0
, &weight
);
474 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
475 /* coord1 = min(coord1, length-1) */
476 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
480 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
482 if (bld
->static_sampler_state
->normalized_coords
) {
483 /* scale coord to length */
484 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
487 offset
= lp_build_int_to_float(coord_bld
, offset
);
488 coord
= lp_build_add(coord_bld
, coord
, offset
);
490 coord
= lp_build_abs(coord_bld
, coord
);
492 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
493 /* skip clamp - always positive, and other side
494 only potentially matters for very large coords */
495 coord
= lp_build_sub(coord_bld
, coord
, half
);
497 /* convert to int, compute lerp weight */
498 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
499 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
512 *weight_out
= weight
;
517 * Build LLVM code for texture wrap mode for nearest filtering.
518 * \param coord the incoming texcoord (nominally in [0,1])
519 * \param length the texture size along one dimension, as int vector
520 * \param length_f the texture size along one dimension, as float vector
521 * \param offset texel offset along one dimension (as int vector)
522 * \param is_pot if TRUE, length is a power of two
523 * \param wrap_mode one of PIPE_TEX_WRAP_x
526 lp_build_sample_wrap_nearest(struct lp_build_sample_context
*bld
,
529 LLVMValueRef length_f
,
534 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
535 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
536 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
537 LLVMValueRef length_minus_one
= lp_build_sub(int_coord_bld
, length
, int_coord_bld
->one
);
541 case PIPE_TEX_WRAP_REPEAT
:
543 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
544 icoord
= lp_build_ifloor(coord_bld
, coord
);
546 icoord
= lp_build_add(int_coord_bld
, icoord
, offset
);
548 icoord
= LLVMBuildAnd(builder
, icoord
, length_minus_one
, "");
552 offset
= lp_build_int_to_float(coord_bld
, offset
);
553 offset
= lp_build_div(coord_bld
, offset
, length_f
);
554 coord
= lp_build_add(coord_bld
, coord
, offset
);
556 /* take fraction, unnormalize */
557 coord
= lp_build_fract_safe(coord_bld
, coord
);
558 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
559 icoord
= lp_build_itrunc(coord_bld
, coord
);
563 case PIPE_TEX_WRAP_CLAMP
:
564 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
565 if (bld
->static_sampler_state
->normalized_coords
) {
566 /* scale coord to length */
567 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
571 /* use itrunc instead since we clamp to 0 anyway */
572 icoord
= lp_build_itrunc(coord_bld
, coord
);
574 icoord
= lp_build_add(int_coord_bld
, icoord
, offset
);
577 /* clamp to [0, length - 1]. */
578 icoord
= lp_build_clamp(int_coord_bld
, icoord
, int_coord_bld
->zero
,
582 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
583 if (bld
->static_sampler_state
->normalized_coords
) {
584 /* scale coord to length */
585 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
587 /* no clamp necessary, border masking will handle this */
588 icoord
= lp_build_ifloor(coord_bld
, coord
);
590 icoord
= lp_build_add(int_coord_bld
, icoord
, offset
);
594 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
596 offset
= lp_build_int_to_float(coord_bld
, offset
);
597 offset
= lp_build_div(coord_bld
, offset
, length_f
);
598 coord
= lp_build_add(coord_bld
, coord
, offset
);
600 /* compute mirror function */
601 coord
= lp_build_coord_mirror(bld
, coord
);
603 /* scale coord to length */
604 assert(bld
->static_sampler_state
->normalized_coords
);
605 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
607 /* itrunc == ifloor here */
608 icoord
= lp_build_itrunc(coord_bld
, coord
);
610 /* clamp to [0, length - 1] */
611 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
614 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
615 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
616 if (bld
->static_sampler_state
->normalized_coords
) {
617 /* scale coord to length */
618 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
621 offset
= lp_build_int_to_float(coord_bld
, offset
);
622 coord
= lp_build_add(coord_bld
, coord
, offset
);
624 coord
= lp_build_abs(coord_bld
, coord
);
626 /* itrunc == ifloor here */
627 icoord
= lp_build_itrunc(coord_bld
, coord
);
629 /* clamp to [0, length - 1] */
630 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
633 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
634 if (bld
->static_sampler_state
->normalized_coords
) {
635 /* scale coord to length */
636 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
639 offset
= lp_build_int_to_float(coord_bld
, offset
);
640 coord
= lp_build_add(coord_bld
, coord
, offset
);
642 coord
= lp_build_abs(coord_bld
, coord
);
644 /* itrunc == ifloor here */
645 icoord
= lp_build_itrunc(coord_bld
, coord
);
658 * Do shadow test/comparison.
659 * \param p shadow ref value
660 * \param texel the texel to compare against
663 lp_build_sample_comparefunc(struct lp_build_sample_context
*bld
,
667 struct lp_build_context
*texel_bld
= &bld
->texel_bld
;
671 //lp_build_print_value(bld->gallivm, "shadow cmp coord", p);
672 lp_build_print_value(bld
->gallivm
, "shadow cmp texel", texel
);
675 /* result = (p FUNC texel) ? 1 : 0 */
677 * honor d3d10 floating point rules here, which state that comparisons
678 * are ordered except NOT_EQUAL which is unordered.
680 if (bld
->static_sampler_state
->compare_func
!= PIPE_FUNC_NOTEQUAL
) {
681 res
= lp_build_cmp_ordered(texel_bld
, bld
->static_sampler_state
->compare_func
,
685 res
= lp_build_cmp(texel_bld
, bld
->static_sampler_state
->compare_func
,
693 * Generate code to sample a mipmap level with nearest filtering.
694 * If sampling a cube texture, r = cube face in [0,5].
697 lp_build_sample_image_nearest(struct lp_build_sample_context
*bld
,
699 LLVMValueRef row_stride_vec
,
700 LLVMValueRef img_stride_vec
,
701 LLVMValueRef data_ptr
,
702 LLVMValueRef mipoffsets
,
703 LLVMValueRef
*coords
,
704 const LLVMValueRef
*offsets
,
705 LLVMValueRef colors_out
[4])
707 const unsigned dims
= bld
->dims
;
708 LLVMValueRef width_vec
;
709 LLVMValueRef height_vec
;
710 LLVMValueRef depth_vec
;
711 LLVMValueRef flt_size
;
712 LLVMValueRef flt_width_vec
;
713 LLVMValueRef flt_height_vec
;
714 LLVMValueRef flt_depth_vec
;
715 LLVMValueRef x
, y
= NULL
, z
= NULL
;
717 lp_build_extract_image_sizes(bld
,
721 &width_vec
, &height_vec
, &depth_vec
);
723 flt_size
= lp_build_int_to_float(&bld
->float_size_bld
, size
);
725 lp_build_extract_image_sizes(bld
,
726 &bld
->float_size_bld
,
729 &flt_width_vec
, &flt_height_vec
, &flt_depth_vec
);
732 * Compute integer texcoords.
734 x
= lp_build_sample_wrap_nearest(bld
, coords
[0], width_vec
,
735 flt_width_vec
, offsets
[0],
736 bld
->static_texture_state
->pot_width
,
737 bld
->static_sampler_state
->wrap_s
);
738 lp_build_name(x
, "tex.x.wrapped");
741 y
= lp_build_sample_wrap_nearest(bld
, coords
[1], height_vec
,
742 flt_height_vec
, offsets
[1],
743 bld
->static_texture_state
->pot_height
,
744 bld
->static_sampler_state
->wrap_t
);
745 lp_build_name(y
, "tex.y.wrapped");
748 z
= lp_build_sample_wrap_nearest(bld
, coords
[2], depth_vec
,
749 flt_depth_vec
, offsets
[2],
750 bld
->static_texture_state
->pot_depth
,
751 bld
->static_sampler_state
->wrap_r
);
752 lp_build_name(z
, "tex.z.wrapped");
755 if (has_layer_coord(bld
->static_texture_state
->target
)) {
756 if (bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
757 /* add cube layer to face */
758 z
= lp_build_add(&bld
->int_coord_bld
, coords
[2], coords
[3]);
763 lp_build_name(z
, "tex.z.layer");
767 * Get texture colors.
769 lp_build_sample_texel_soa(bld
,
770 width_vec
, height_vec
, depth_vec
,
772 row_stride_vec
, img_stride_vec
,
773 data_ptr
, mipoffsets
, colors_out
);
775 if (bld
->static_sampler_state
->compare_mode
!= PIPE_TEX_COMPARE_NONE
) {
777 cmpval
= lp_build_sample_comparefunc(bld
, coords
[4], colors_out
[0]);
778 /* this is really just a AND 1.0, cmpval but llvm is clever enough */
779 colors_out
[0] = lp_build_select(&bld
->texel_bld
, cmpval
,
780 bld
->texel_bld
.one
, bld
->texel_bld
.zero
);
781 colors_out
[1] = colors_out
[2] = colors_out
[3] = colors_out
[0];
788 * Like a lerp, but inputs are 0/~0 masks, so can simplify slightly.
791 lp_build_masklerp(struct lp_build_context
*bld
,
796 struct gallivm_state
*gallivm
= bld
->gallivm
;
797 LLVMBuilderRef builder
= gallivm
->builder
;
798 LLVMValueRef weight2
;
800 weight2
= lp_build_sub(bld
, bld
->one
, weight
);
801 weight
= LLVMBuildBitCast(builder
, weight
,
802 lp_build_int_vec_type(gallivm
, bld
->type
), "");
803 weight2
= LLVMBuildBitCast(builder
, weight2
,
804 lp_build_int_vec_type(gallivm
, bld
->type
), "");
805 weight
= LLVMBuildAnd(builder
, weight
, mask1
, "");
806 weight2
= LLVMBuildAnd(builder
, weight2
, mask0
, "");
807 weight
= LLVMBuildBitCast(builder
, weight
, bld
->vec_type
, "");
808 weight2
= LLVMBuildBitCast(builder
, weight2
, bld
->vec_type
, "");
809 return lp_build_add(bld
, weight
, weight2
);
813 * Like a 2d lerp, but inputs are 0/~0 masks, so can simplify slightly.
816 lp_build_masklerp2d(struct lp_build_context
*bld
,
817 LLVMValueRef weight0
,
818 LLVMValueRef weight1
,
824 LLVMValueRef val0
= lp_build_masklerp(bld
, weight0
, mask00
, mask01
);
825 LLVMValueRef val1
= lp_build_masklerp(bld
, weight0
, mask10
, mask11
);
826 return lp_build_lerp(bld
, weight1
, val0
, val1
, 0);
830 * this is a bit excessive code for something OpenGL just recommends
831 * but does not require.
833 #define ACCURATE_CUBE_CORNERS 1
836 * Generate code to sample a mipmap level with linear filtering.
837 * If sampling a cube texture, r = cube face in [0,5].
838 * If linear_mask is present, only pixels having their mask set
839 * will receive linear filtering, the rest will use nearest.
842 lp_build_sample_image_linear(struct lp_build_sample_context
*bld
,
844 LLVMValueRef linear_mask
,
845 LLVMValueRef row_stride_vec
,
846 LLVMValueRef img_stride_vec
,
847 LLVMValueRef data_ptr
,
848 LLVMValueRef mipoffsets
,
849 LLVMValueRef
*coords
,
850 const LLVMValueRef
*offsets
,
851 LLVMValueRef colors_out
[4])
853 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
854 struct lp_build_context
*ivec_bld
= &bld
->int_coord_bld
;
855 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
856 const unsigned dims
= bld
->dims
;
857 LLVMValueRef width_vec
;
858 LLVMValueRef height_vec
;
859 LLVMValueRef depth_vec
;
860 LLVMValueRef flt_size
;
861 LLVMValueRef flt_width_vec
;
862 LLVMValueRef flt_height_vec
;
863 LLVMValueRef flt_depth_vec
;
864 LLVMValueRef fall_off
[4], have_corners
;
865 LLVMValueRef z1
= NULL
;
866 LLVMValueRef z00
= NULL
, z01
= NULL
, z10
= NULL
, z11
= NULL
;
867 LLVMValueRef x00
= NULL
, x01
= NULL
, x10
= NULL
, x11
= NULL
;
868 LLVMValueRef y00
= NULL
, y01
= NULL
, y10
= NULL
, y11
= NULL
;
869 LLVMValueRef s_fpart
, t_fpart
= NULL
, r_fpart
= NULL
;
870 LLVMValueRef xs
[4], ys
[4], zs
[4];
871 LLVMValueRef neighbors
[2][2][4];
872 int chan
, texel_index
;
873 boolean seamless_cube_filter
, accurate_cube_corners
;
875 seamless_cube_filter
= (bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE
||
876 bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) &&
877 bld
->static_sampler_state
->seamless_cube_map
;
878 accurate_cube_corners
= ACCURATE_CUBE_CORNERS
&& seamless_cube_filter
;
880 lp_build_extract_image_sizes(bld
,
884 &width_vec
, &height_vec
, &depth_vec
);
886 flt_size
= lp_build_int_to_float(&bld
->float_size_bld
, size
);
888 lp_build_extract_image_sizes(bld
,
889 &bld
->float_size_bld
,
892 &flt_width_vec
, &flt_height_vec
, &flt_depth_vec
);
895 * Compute integer texcoords.
898 if (!seamless_cube_filter
) {
899 lp_build_sample_wrap_linear(bld
, coords
[0], width_vec
,
900 flt_width_vec
, offsets
[0],
901 bld
->static_texture_state
->pot_width
,
902 bld
->static_sampler_state
->wrap_s
,
903 &x00
, &x01
, &s_fpart
);
904 lp_build_name(x00
, "tex.x0.wrapped");
905 lp_build_name(x01
, "tex.x1.wrapped");
910 lp_build_sample_wrap_linear(bld
, coords
[1], height_vec
,
911 flt_height_vec
, offsets
[1],
912 bld
->static_texture_state
->pot_height
,
913 bld
->static_sampler_state
->wrap_t
,
914 &y00
, &y10
, &t_fpart
);
915 lp_build_name(y00
, "tex.y0.wrapped");
916 lp_build_name(y10
, "tex.y1.wrapped");
921 lp_build_sample_wrap_linear(bld
, coords
[2], depth_vec
,
922 flt_depth_vec
, offsets
[2],
923 bld
->static_texture_state
->pot_depth
,
924 bld
->static_sampler_state
->wrap_r
,
925 &z00
, &z1
, &r_fpart
);
926 z01
= z10
= z11
= z00
;
927 lp_build_name(z00
, "tex.z0.wrapped");
928 lp_build_name(z1
, "tex.z1.wrapped");
931 if (has_layer_coord(bld
->static_texture_state
->target
)) {
932 if (bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
933 /* add cube layer to face */
934 z00
= z01
= z10
= z11
= z1
=
935 lp_build_add(&bld
->int_coord_bld
, coords
[2], coords
[3]);
938 z00
= z01
= z10
= z11
= z1
= coords
[2]; /* cube face or layer */
940 lp_build_name(z00
, "tex.z0.layer");
941 lp_build_name(z1
, "tex.z1.layer");
945 struct lp_build_if_state edge_if
;
947 LLVMValueRef new_faces
[4], new_xcoords
[4][2], new_ycoords
[4][2];
948 LLVMValueRef coord
, have_edge
, have_corner
;
949 LLVMValueRef fall_off_ym_notxm
, fall_off_ym_notxp
, fall_off_x
, fall_off_y
;
950 LLVMValueRef fall_off_yp_notxm
, fall_off_yp_notxp
;
951 LLVMValueRef x0
, x1
, y0
, y1
, y0_clamped
, y1_clamped
;
952 LLVMValueRef face
= coords
[2];
953 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, coord_bld
->type
, 0.5f
);
954 LLVMValueRef length_minus_one
= lp_build_sub(ivec_bld
, width_vec
, ivec_bld
->one
);
955 /* XXX drop height calcs. Could (should) do this without seamless filtering too */
956 height_vec
= width_vec
;
957 flt_height_vec
= flt_width_vec
;
959 /* XXX the overflow logic is actually sort of duplicated with trilinear,
960 * since an overflow in one mip should also have a corresponding overflow
963 /* should always have normalized coords, and offsets are undefined */
964 assert(bld
->static_sampler_state
->normalized_coords
);
965 coord
= lp_build_mul(coord_bld
, coords
[0], flt_width_vec
);
966 /* instead of clamp, build mask if overflowed */
967 coord
= lp_build_sub(coord_bld
, coord
, half
);
968 /* convert to int, compute lerp weight */
969 /* not ideal with AVX (and no AVX2) */
970 lp_build_ifloor_fract(coord_bld
, coord
, &x0
, &s_fpart
);
971 x1
= lp_build_add(ivec_bld
, x0
, ivec_bld
->one
);
972 coord
= lp_build_mul(coord_bld
, coords
[1], flt_height_vec
);
973 coord
= lp_build_sub(coord_bld
, coord
, half
);
974 lp_build_ifloor_fract(coord_bld
, coord
, &y0
, &t_fpart
);
975 y1
= lp_build_add(ivec_bld
, y0
, ivec_bld
->one
);
977 fall_off
[0] = lp_build_cmp(ivec_bld
, PIPE_FUNC_LESS
, x0
, ivec_bld
->zero
);
978 fall_off
[1] = lp_build_cmp(ivec_bld
, PIPE_FUNC_GREATER
, x1
, length_minus_one
);
979 fall_off
[2] = lp_build_cmp(ivec_bld
, PIPE_FUNC_LESS
, y0
, ivec_bld
->zero
);
980 fall_off
[3] = lp_build_cmp(ivec_bld
, PIPE_FUNC_GREATER
, y1
, length_minus_one
);
982 fall_off_x
= lp_build_or(ivec_bld
, fall_off
[0], fall_off
[1]);
983 fall_off_y
= lp_build_or(ivec_bld
, fall_off
[2], fall_off
[3]);
984 have_edge
= lp_build_or(ivec_bld
, fall_off_x
, fall_off_y
);
985 have_edge
= lp_build_any_true_range(ivec_bld
, ivec_bld
->type
.length
, have_edge
);
987 /* needed for accurate corner filtering branch later, rely on 0 init */
988 int1t
= LLVMInt1TypeInContext(bld
->gallivm
->context
);
989 have_corners
= lp_build_alloca(bld
->gallivm
, int1t
, "have_corner");
991 for (texel_index
= 0; texel_index
< 4; texel_index
++) {
992 xs
[texel_index
] = lp_build_alloca(bld
->gallivm
, ivec_bld
->vec_type
, "xs");
993 ys
[texel_index
] = lp_build_alloca(bld
->gallivm
, ivec_bld
->vec_type
, "ys");
994 zs
[texel_index
] = lp_build_alloca(bld
->gallivm
, ivec_bld
->vec_type
, "zs");
997 lp_build_if(&edge_if
, bld
->gallivm
, have_edge
);
999 have_corner
= lp_build_and(ivec_bld
, fall_off_x
, fall_off_y
);
1000 have_corner
= lp_build_any_true_range(ivec_bld
, ivec_bld
->type
.length
, have_corner
);
1001 LLVMBuildStore(builder
, have_corner
, have_corners
);
1004 * Need to feed clamped values here for cheap corner handling,
1005 * but only for y coord (as when falling off both edges we only
1006 * fall off the x one) - this should be sufficient.
1008 y0_clamped
= lp_build_max(ivec_bld
, y0
, ivec_bld
->zero
);
1009 y1_clamped
= lp_build_min(ivec_bld
, y1
, length_minus_one
);
1012 * Get all possible new coords.
1014 lp_build_cube_new_coords(ivec_bld
, face
,
1015 x0
, x1
, y0_clamped
, y1_clamped
,
1017 new_faces
, new_xcoords
, new_ycoords
);
1019 /* handle fall off x-, x+ direction */
1020 /* determine new coords, face (not both fall_off vars can be true at same time) */
1021 x00
= lp_build_select(ivec_bld
, fall_off
[0], new_xcoords
[0][0], x0
);
1022 y00
= lp_build_select(ivec_bld
, fall_off
[0], new_ycoords
[0][0], y0_clamped
);
1023 x10
= lp_build_select(ivec_bld
, fall_off
[0], new_xcoords
[0][1], x0
);
1024 y10
= lp_build_select(ivec_bld
, fall_off
[0], new_ycoords
[0][1], y1_clamped
);
1025 x01
= lp_build_select(ivec_bld
, fall_off
[1], new_xcoords
[1][0], x1
);
1026 y01
= lp_build_select(ivec_bld
, fall_off
[1], new_ycoords
[1][0], y0_clamped
);
1027 x11
= lp_build_select(ivec_bld
, fall_off
[1], new_xcoords
[1][1], x1
);
1028 y11
= lp_build_select(ivec_bld
, fall_off
[1], new_ycoords
[1][1], y1_clamped
);
1030 z00
= z10
= lp_build_select(ivec_bld
, fall_off
[0], new_faces
[0], face
);
1031 z01
= z11
= lp_build_select(ivec_bld
, fall_off
[1], new_faces
[1], face
);
1033 /* handle fall off y-, y+ direction */
1035 * Cheap corner logic: just hack up things so a texel doesn't fall
1036 * off both sides (which means filter weights will be wrong but we'll only
1037 * use valid texels in the filter).
1038 * This means however (y) coords must additionally be clamped (see above).
1039 * This corner handling should be fully OpenGL (but not d3d10) compliant.
1041 fall_off_ym_notxm
= lp_build_andnot(ivec_bld
, fall_off
[2], fall_off
[0]);
1042 fall_off_ym_notxp
= lp_build_andnot(ivec_bld
, fall_off
[2], fall_off
[1]);
1043 fall_off_yp_notxm
= lp_build_andnot(ivec_bld
, fall_off
[3], fall_off
[0]);
1044 fall_off_yp_notxp
= lp_build_andnot(ivec_bld
, fall_off
[3], fall_off
[1]);
1046 x00
= lp_build_select(ivec_bld
, fall_off_ym_notxm
, new_xcoords
[2][0], x00
);
1047 y00
= lp_build_select(ivec_bld
, fall_off_ym_notxm
, new_ycoords
[2][0], y00
);
1048 x01
= lp_build_select(ivec_bld
, fall_off_ym_notxp
, new_xcoords
[2][1], x01
);
1049 y01
= lp_build_select(ivec_bld
, fall_off_ym_notxp
, new_ycoords
[2][1], y01
);
1050 x10
= lp_build_select(ivec_bld
, fall_off_yp_notxm
, new_xcoords
[3][0], x10
);
1051 y10
= lp_build_select(ivec_bld
, fall_off_yp_notxm
, new_ycoords
[3][0], y10
);
1052 x11
= lp_build_select(ivec_bld
, fall_off_yp_notxp
, new_xcoords
[3][1], x11
);
1053 y11
= lp_build_select(ivec_bld
, fall_off_yp_notxp
, new_ycoords
[3][1], y11
);
1055 z00
= lp_build_select(ivec_bld
, fall_off_ym_notxm
, new_faces
[2], z00
);
1056 z01
= lp_build_select(ivec_bld
, fall_off_ym_notxp
, new_faces
[2], z01
);
1057 z10
= lp_build_select(ivec_bld
, fall_off_yp_notxm
, new_faces
[3], z10
);
1058 z11
= lp_build_select(ivec_bld
, fall_off_yp_notxp
, new_faces
[3], z11
);
1060 if (bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
1061 /* now can add cube layer to face (per sample) */
1062 z00
= lp_build_add(ivec_bld
, z00
, coords
[3]);
1063 z01
= lp_build_add(ivec_bld
, z01
, coords
[3]);
1064 z10
= lp_build_add(ivec_bld
, z10
, coords
[3]);
1065 z11
= lp_build_add(ivec_bld
, z11
, coords
[3]);
1068 LLVMBuildStore(builder
, x00
, xs
[0]);
1069 LLVMBuildStore(builder
, x01
, xs
[1]);
1070 LLVMBuildStore(builder
, x10
, xs
[2]);
1071 LLVMBuildStore(builder
, x11
, xs
[3]);
1072 LLVMBuildStore(builder
, y00
, ys
[0]);
1073 LLVMBuildStore(builder
, y01
, ys
[1]);
1074 LLVMBuildStore(builder
, y10
, ys
[2]);
1075 LLVMBuildStore(builder
, y11
, ys
[3]);
1076 LLVMBuildStore(builder
, z00
, zs
[0]);
1077 LLVMBuildStore(builder
, z01
, zs
[1]);
1078 LLVMBuildStore(builder
, z10
, zs
[2]);
1079 LLVMBuildStore(builder
, z11
, zs
[3]);
1081 lp_build_else(&edge_if
);
1083 LLVMBuildStore(builder
, x0
, xs
[0]);
1084 LLVMBuildStore(builder
, x1
, xs
[1]);
1085 LLVMBuildStore(builder
, x0
, xs
[2]);
1086 LLVMBuildStore(builder
, x1
, xs
[3]);
1087 LLVMBuildStore(builder
, y0
, ys
[0]);
1088 LLVMBuildStore(builder
, y0
, ys
[1]);
1089 LLVMBuildStore(builder
, y1
, ys
[2]);
1090 LLVMBuildStore(builder
, y1
, ys
[3]);
1091 if (bld
->static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
1092 LLVMValueRef cube_layer
= lp_build_add(ivec_bld
, face
, coords
[3]);
1093 LLVMBuildStore(builder
, cube_layer
, zs
[0]);
1094 LLVMBuildStore(builder
, cube_layer
, zs
[1]);
1095 LLVMBuildStore(builder
, cube_layer
, zs
[2]);
1096 LLVMBuildStore(builder
, cube_layer
, zs
[3]);
1099 LLVMBuildStore(builder
, face
, zs
[0]);
1100 LLVMBuildStore(builder
, face
, zs
[1]);
1101 LLVMBuildStore(builder
, face
, zs
[2]);
1102 LLVMBuildStore(builder
, face
, zs
[3]);
1105 lp_build_endif(&edge_if
);
1107 x00
= LLVMBuildLoad(builder
, xs
[0], "");
1108 x01
= LLVMBuildLoad(builder
, xs
[1], "");
1109 x10
= LLVMBuildLoad(builder
, xs
[2], "");
1110 x11
= LLVMBuildLoad(builder
, xs
[3], "");
1111 y00
= LLVMBuildLoad(builder
, ys
[0], "");
1112 y01
= LLVMBuildLoad(builder
, ys
[1], "");
1113 y10
= LLVMBuildLoad(builder
, ys
[2], "");
1114 y11
= LLVMBuildLoad(builder
, ys
[3], "");
1115 z00
= LLVMBuildLoad(builder
, zs
[0], "");
1116 z01
= LLVMBuildLoad(builder
, zs
[1], "");
1117 z10
= LLVMBuildLoad(builder
, zs
[2], "");
1118 z11
= LLVMBuildLoad(builder
, zs
[3], "");
1123 * Whack filter weights into place. Whatever texel had more weight is
1124 * the one which should have been selected by nearest filtering hence
1125 * just use 100% weight for it.
1127 struct lp_build_context
*c_bld
= &bld
->coord_bld
;
1128 LLVMValueRef w1_mask
, w1_weight
;
1129 LLVMValueRef half
= lp_build_const_vec(bld
->gallivm
, c_bld
->type
, 0.5f
);
1131 w1_mask
= lp_build_cmp(c_bld
, PIPE_FUNC_GREATER
, s_fpart
, half
);
1132 /* this select is really just a "and" */
1133 w1_weight
= lp_build_select(c_bld
, w1_mask
, c_bld
->one
, c_bld
->zero
);
1134 s_fpart
= lp_build_select(c_bld
, linear_mask
, s_fpart
, w1_weight
);
1136 w1_mask
= lp_build_cmp(c_bld
, PIPE_FUNC_GREATER
, t_fpart
, half
);
1137 w1_weight
= lp_build_select(c_bld
, w1_mask
, c_bld
->one
, c_bld
->zero
);
1138 t_fpart
= lp_build_select(c_bld
, linear_mask
, t_fpart
, w1_weight
);
1140 w1_mask
= lp_build_cmp(c_bld
, PIPE_FUNC_GREATER
, r_fpart
, half
);
1141 w1_weight
= lp_build_select(c_bld
, w1_mask
, c_bld
->one
, c_bld
->zero
);
1142 r_fpart
= lp_build_select(c_bld
, linear_mask
, r_fpart
, w1_weight
);
1148 * Get texture colors.
1150 /* get x0/x1 texels */
1151 lp_build_sample_texel_soa(bld
,
1152 width_vec
, height_vec
, depth_vec
,
1154 row_stride_vec
, img_stride_vec
,
1155 data_ptr
, mipoffsets
, neighbors
[0][0]);
1156 lp_build_sample_texel_soa(bld
,
1157 width_vec
, height_vec
, depth_vec
,
1159 row_stride_vec
, img_stride_vec
,
1160 data_ptr
, mipoffsets
, neighbors
[0][1]);
1163 if (bld
->static_sampler_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
) {
1164 /* Interpolate two samples from 1D image to produce one color */
1165 for (chan
= 0; chan
< 4; chan
++) {
1166 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
, s_fpart
,
1167 neighbors
[0][0][chan
],
1168 neighbors
[0][1][chan
],
1173 LLVMValueRef cmpval0
, cmpval1
;
1174 cmpval0
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][0][0]);
1175 cmpval1
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][1][0]);
1176 /* simplified lerp, AND mask with weight and add */
1177 colors_out
[0] = lp_build_masklerp(&bld
->texel_bld
, s_fpart
,
1179 colors_out
[1] = colors_out
[2] = colors_out
[3] = colors_out
[0];
1184 struct lp_build_if_state corner_if
;
1185 LLVMValueRef colors0
[4], colorss
[4];
1187 /* get x0/x1 texels at y1 */
1188 lp_build_sample_texel_soa(bld
,
1189 width_vec
, height_vec
, depth_vec
,
1191 row_stride_vec
, img_stride_vec
,
1192 data_ptr
, mipoffsets
, neighbors
[1][0]);
1193 lp_build_sample_texel_soa(bld
,
1194 width_vec
, height_vec
, depth_vec
,
1196 row_stride_vec
, img_stride_vec
,
1197 data_ptr
, mipoffsets
, neighbors
[1][1]);
1200 * To avoid having to duplicate linear_mask / fetch code use
1201 * another branch (with corner condition though edge would work
1204 if (accurate_cube_corners
) {
1205 LLVMValueRef w00
, w01
, w10
, w11
, wx0
, wy0
;
1206 LLVMValueRef c_weight
, c00
, c01
, c10
, c11
;
1207 LLVMValueRef have_corner
, one_third
, tmp
;
1209 colorss
[0] = lp_build_alloca(bld
->gallivm
, coord_bld
->vec_type
, "cs");
1210 colorss
[1] = lp_build_alloca(bld
->gallivm
, coord_bld
->vec_type
, "cs");
1211 colorss
[2] = lp_build_alloca(bld
->gallivm
, coord_bld
->vec_type
, "cs");
1212 colorss
[3] = lp_build_alloca(bld
->gallivm
, coord_bld
->vec_type
, "cs");
1214 have_corner
= LLVMBuildLoad(builder
, have_corners
, "");
1216 lp_build_if(&corner_if
, bld
->gallivm
, have_corner
);
1219 * we can't use standard 2d lerp as we need per-element weight
1220 * in case of corners, so just calculate bilinear result as
1221 * w00*s00 + w01*s01 + w10*s10 + w11*s11.
1222 * (This is actually less work than using 2d lerp, 7 vs. 9 instructions,
1223 * however calculating the weights needs another 6, so actually probably
1224 * not slower than 2d lerp only for 4 channels as weights only need
1225 * to be calculated once - of course fixing the weights has additional cost.)
1227 wx0
= lp_build_sub(coord_bld
, coord_bld
->one
, s_fpart
);
1228 wy0
= lp_build_sub(coord_bld
, coord_bld
->one
, t_fpart
);
1229 w00
= lp_build_mul(coord_bld
, wx0
, wy0
);
1230 w01
= lp_build_mul(coord_bld
, s_fpart
, wy0
);
1231 w10
= lp_build_mul(coord_bld
, wx0
, t_fpart
);
1232 w11
= lp_build_mul(coord_bld
, s_fpart
, t_fpart
);
1234 /* find corner weight */
1235 c00
= lp_build_and(ivec_bld
, fall_off
[0], fall_off
[2]);
1236 c_weight
= lp_build_select(coord_bld
, c00
, w00
, coord_bld
->zero
);
1237 c01
= lp_build_and(ivec_bld
, fall_off
[1], fall_off
[2]);
1238 c_weight
= lp_build_select(coord_bld
, c01
, w01
, c_weight
);
1239 c10
= lp_build_and(ivec_bld
, fall_off
[0], fall_off
[3]);
1240 c_weight
= lp_build_select(coord_bld
, c10
, w10
, c_weight
);
1241 c11
= lp_build_and(ivec_bld
, fall_off
[1], fall_off
[3]);
1242 c_weight
= lp_build_select(coord_bld
, c11
, w11
, c_weight
);
1245 * add 1/3 of the corner weight to each of the 3 other samples
1246 * and null out corner weight
1248 one_third
= lp_build_const_vec(bld
->gallivm
, coord_bld
->type
, 1.0f
/3.0f
);
1249 c_weight
= lp_build_mul(coord_bld
, c_weight
, one_third
);
1250 w00
= lp_build_add(coord_bld
, w00
, c_weight
);
1251 c00
= LLVMBuildBitCast(builder
, c00
, coord_bld
->vec_type
, "");
1252 w00
= lp_build_andnot(coord_bld
, w00
, c00
);
1253 w01
= lp_build_add(coord_bld
, w01
, c_weight
);
1254 c01
= LLVMBuildBitCast(builder
, c01
, coord_bld
->vec_type
, "");
1255 w01
= lp_build_andnot(coord_bld
, w01
, c01
);
1256 w10
= lp_build_add(coord_bld
, w10
, c_weight
);
1257 c10
= LLVMBuildBitCast(builder
, c10
, coord_bld
->vec_type
, "");
1258 w10
= lp_build_andnot(coord_bld
, w10
, c10
);
1259 w11
= lp_build_add(coord_bld
, w11
, c_weight
);
1260 c11
= LLVMBuildBitCast(builder
, c11
, coord_bld
->vec_type
, "");
1261 w11
= lp_build_andnot(coord_bld
, w11
, c11
);
1263 if (bld
->static_sampler_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
) {
1264 for (chan
= 0; chan
< 4; chan
++) {
1265 colors0
[chan
] = lp_build_mul(coord_bld
, w00
, neighbors
[0][0][chan
]);
1266 tmp
= lp_build_mul(coord_bld
, w01
, neighbors
[0][1][chan
]);
1267 colors0
[chan
] = lp_build_add(coord_bld
, tmp
, colors0
[chan
]);
1268 tmp
= lp_build_mul(coord_bld
, w10
, neighbors
[1][0][chan
]);
1269 colors0
[chan
] = lp_build_add(coord_bld
, tmp
, colors0
[chan
]);
1270 tmp
= lp_build_mul(coord_bld
, w11
, neighbors
[1][1][chan
]);
1271 colors0
[chan
] = lp_build_add(coord_bld
, tmp
, colors0
[chan
]);
1275 LLVMValueRef cmpval00
, cmpval01
, cmpval10
, cmpval11
;
1276 cmpval00
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][0][0]);
1277 cmpval01
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][1][0]);
1278 cmpval10
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][0][0]);
1279 cmpval11
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][1][0]);
1280 /* inputs to interpolation are just masks so just add masked weights together */
1281 cmpval00
= LLVMBuildBitCast(builder
, cmpval00
, coord_bld
->vec_type
, "");
1282 cmpval01
= LLVMBuildBitCast(builder
, cmpval01
, coord_bld
->vec_type
, "");
1283 cmpval10
= LLVMBuildBitCast(builder
, cmpval10
, coord_bld
->vec_type
, "");
1284 cmpval11
= LLVMBuildBitCast(builder
, cmpval11
, coord_bld
->vec_type
, "");
1285 colors0
[0] = lp_build_and(coord_bld
, w00
, cmpval00
);
1286 tmp
= lp_build_and(coord_bld
, w01
, cmpval01
);
1287 colors0
[0] = lp_build_add(coord_bld
, tmp
, colors0
[0]);
1288 tmp
= lp_build_and(coord_bld
, w10
, cmpval10
);
1289 colors0
[0] = lp_build_add(coord_bld
, tmp
, colors0
[0]);
1290 tmp
= lp_build_and(coord_bld
, w11
, cmpval11
);
1291 colors0
[0] = lp_build_add(coord_bld
, tmp
, colors0
[0]);
1292 colors0
[1] = colors0
[2] = colors0
[3] = colors0
[0];
1295 LLVMBuildStore(builder
, colors0
[0], colorss
[0]);
1296 LLVMBuildStore(builder
, colors0
[1], colorss
[1]);
1297 LLVMBuildStore(builder
, colors0
[2], colorss
[2]);
1298 LLVMBuildStore(builder
, colors0
[3], colorss
[3]);
1300 lp_build_else(&corner_if
);
1303 if (bld
->static_sampler_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
) {
1304 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1305 for (chan
= 0; chan
< 4; chan
++) {
1306 colors0
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
1308 neighbors
[0][0][chan
],
1309 neighbors
[0][1][chan
],
1310 neighbors
[1][0][chan
],
1311 neighbors
[1][1][chan
],
1316 LLVMValueRef cmpval00
, cmpval01
, cmpval10
, cmpval11
;
1317 cmpval00
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][0][0]);
1318 cmpval01
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][1][0]);
1319 cmpval10
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][0][0]);
1320 cmpval11
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][1][0]);
1321 colors0
[0] = lp_build_masklerp2d(&bld
->texel_bld
, s_fpart
, t_fpart
,
1322 cmpval00
, cmpval01
, cmpval10
, cmpval11
);
1323 colors0
[1] = colors0
[2] = colors0
[3] = colors0
[0];
1326 if (accurate_cube_corners
) {
1327 LLVMBuildStore(builder
, colors0
[0], colorss
[0]);
1328 LLVMBuildStore(builder
, colors0
[1], colorss
[1]);
1329 LLVMBuildStore(builder
, colors0
[2], colorss
[2]);
1330 LLVMBuildStore(builder
, colors0
[3], colorss
[3]);
1332 lp_build_endif(&corner_if
);
1334 colors0
[0] = LLVMBuildLoad(builder
, colorss
[0], "");
1335 colors0
[1] = LLVMBuildLoad(builder
, colorss
[1], "");
1336 colors0
[2] = LLVMBuildLoad(builder
, colorss
[2], "");
1337 colors0
[3] = LLVMBuildLoad(builder
, colorss
[3], "");
1341 LLVMValueRef neighbors1
[2][2][4];
1342 LLVMValueRef colors1
[4];
1344 /* get x0/x1/y0/y1 texels at z1 */
1345 lp_build_sample_texel_soa(bld
,
1346 width_vec
, height_vec
, depth_vec
,
1348 row_stride_vec
, img_stride_vec
,
1349 data_ptr
, mipoffsets
, neighbors1
[0][0]);
1350 lp_build_sample_texel_soa(bld
,
1351 width_vec
, height_vec
, depth_vec
,
1353 row_stride_vec
, img_stride_vec
,
1354 data_ptr
, mipoffsets
, neighbors1
[0][1]);
1355 lp_build_sample_texel_soa(bld
,
1356 width_vec
, height_vec
, depth_vec
,
1358 row_stride_vec
, img_stride_vec
,
1359 data_ptr
, mipoffsets
, neighbors1
[1][0]);
1360 lp_build_sample_texel_soa(bld
,
1361 width_vec
, height_vec
, depth_vec
,
1363 row_stride_vec
, img_stride_vec
,
1364 data_ptr
, mipoffsets
, neighbors1
[1][1]);
1366 if (bld
->static_sampler_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
) {
1367 /* Bilinear interpolate the four samples from the second Z slice */
1368 for (chan
= 0; chan
< 4; chan
++) {
1369 colors1
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
1371 neighbors1
[0][0][chan
],
1372 neighbors1
[0][1][chan
],
1373 neighbors1
[1][0][chan
],
1374 neighbors1
[1][1][chan
],
1377 /* Linearly interpolate the two samples from the two 3D slices */
1378 for (chan
= 0; chan
< 4; chan
++) {
1379 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
,
1381 colors0
[chan
], colors1
[chan
],
1386 LLVMValueRef cmpval00
, cmpval01
, cmpval10
, cmpval11
;
1387 cmpval00
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][0][0]);
1388 cmpval01
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[0][1][0]);
1389 cmpval10
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][0][0]);
1390 cmpval11
= lp_build_sample_comparefunc(bld
, coords
[4], neighbors
[1][1][0]);
1391 colors1
[0] = lp_build_masklerp2d(&bld
->texel_bld
, s_fpart
, t_fpart
,
1392 cmpval00
, cmpval01
, cmpval10
, cmpval11
);
1393 /* Linearly interpolate the two samples from the two 3D slices */
1394 colors_out
[0] = lp_build_lerp(&bld
->texel_bld
,
1396 colors0
[0], colors1
[0],
1398 colors_out
[1] = colors_out
[2] = colors_out
[3] = colors_out
[0];
1403 for (chan
= 0; chan
< 4; chan
++) {
1404 colors_out
[chan
] = colors0
[chan
];
1412 * Sample the texture/mipmap using given image filter and mip filter.
1413 * ilevel0 and ilevel1 indicate the two mipmap levels to sample
1414 * from (vectors or scalars).
1415 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1418 lp_build_sample_mipmap(struct lp_build_sample_context
*bld
,
1419 unsigned img_filter
,
1420 unsigned mip_filter
,
1421 LLVMValueRef
*coords
,
1422 const LLVMValueRef
*offsets
,
1423 LLVMValueRef ilevel0
,
1424 LLVMValueRef ilevel1
,
1425 LLVMValueRef lod_fpart
,
1426 LLVMValueRef
*colors_out
)
1428 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1429 LLVMValueRef size0
= NULL
;
1430 LLVMValueRef size1
= NULL
;
1431 LLVMValueRef row_stride0_vec
= NULL
;
1432 LLVMValueRef row_stride1_vec
= NULL
;
1433 LLVMValueRef img_stride0_vec
= NULL
;
1434 LLVMValueRef img_stride1_vec
= NULL
;
1435 LLVMValueRef data_ptr0
= NULL
;
1436 LLVMValueRef data_ptr1
= NULL
;
1437 LLVMValueRef mipoff0
= NULL
;
1438 LLVMValueRef mipoff1
= NULL
;
1439 LLVMValueRef colors0
[4], colors1
[4];
1442 /* sample the first mipmap level */
1443 lp_build_mipmap_level_sizes(bld
, ilevel0
,
1445 &row_stride0_vec
, &img_stride0_vec
);
1446 if (bld
->num_mips
== 1) {
1447 data_ptr0
= lp_build_get_mipmap_level(bld
, ilevel0
);
1450 /* This path should work for num_lods 1 too but slightly less efficient */
1451 data_ptr0
= bld
->base_ptr
;
1452 mipoff0
= lp_build_get_mip_offsets(bld
, ilevel0
);
1454 if (img_filter
== PIPE_TEX_FILTER_NEAREST
) {
1455 lp_build_sample_image_nearest(bld
, size0
,
1456 row_stride0_vec
, img_stride0_vec
,
1457 data_ptr0
, mipoff0
, coords
, offsets
,
1461 assert(img_filter
== PIPE_TEX_FILTER_LINEAR
);
1462 lp_build_sample_image_linear(bld
, size0
, NULL
,
1463 row_stride0_vec
, img_stride0_vec
,
1464 data_ptr0
, mipoff0
, coords
, offsets
,
1468 /* Store the first level's colors in the output variables */
1469 for (chan
= 0; chan
< 4; chan
++) {
1470 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
1473 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
1474 struct lp_build_if_state if_ctx
;
1475 LLVMValueRef need_lerp
;
1477 /* need_lerp = lod_fpart > 0 */
1478 if (bld
->num_lods
== 1) {
1479 need_lerp
= LLVMBuildFCmp(builder
, LLVMRealUGT
,
1480 lod_fpart
, bld
->lodf_bld
.zero
,
1485 * We'll do mip filtering if any of the quads (or individual
1486 * pixel in case of per-pixel lod) need it.
1487 * It might be better to split the vectors here and only fetch/filter
1488 * quads which need it (if there's one lod per quad).
1490 need_lerp
= lp_build_compare(bld
->gallivm
, bld
->lodf_bld
.type
,
1492 lod_fpart
, bld
->lodf_bld
.zero
);
1493 need_lerp
= lp_build_any_true_range(&bld
->lodi_bld
, bld
->num_lods
, need_lerp
);
1496 lp_build_if(&if_ctx
, bld
->gallivm
, need_lerp
);
1499 * We unfortunately need to clamp lod_fpart here since we can get
1500 * negative values which would screw up filtering if not all
1501 * lod_fpart values have same sign.
1503 lod_fpart
= lp_build_max(&bld
->lodf_bld
, lod_fpart
,
1504 bld
->lodf_bld
.zero
);
1505 /* sample the second mipmap level */
1506 lp_build_mipmap_level_sizes(bld
, ilevel1
,
1508 &row_stride1_vec
, &img_stride1_vec
);
1509 if (bld
->num_mips
== 1) {
1510 data_ptr1
= lp_build_get_mipmap_level(bld
, ilevel1
);
1513 data_ptr1
= bld
->base_ptr
;
1514 mipoff1
= lp_build_get_mip_offsets(bld
, ilevel1
);
1516 if (img_filter
== PIPE_TEX_FILTER_NEAREST
) {
1517 lp_build_sample_image_nearest(bld
, size1
,
1518 row_stride1_vec
, img_stride1_vec
,
1519 data_ptr1
, mipoff1
, coords
, offsets
,
1523 lp_build_sample_image_linear(bld
, size1
, NULL
,
1524 row_stride1_vec
, img_stride1_vec
,
1525 data_ptr1
, mipoff1
, coords
, offsets
,
1529 /* interpolate samples from the two mipmap levels */
1531 if (bld
->num_lods
!= bld
->coord_type
.length
)
1532 lod_fpart
= lp_build_unpack_broadcast_aos_scalars(bld
->gallivm
,
1534 bld
->texel_bld
.type
,
1537 for (chan
= 0; chan
< 4; chan
++) {
1538 colors0
[chan
] = lp_build_lerp(&bld
->texel_bld
, lod_fpart
,
1539 colors0
[chan
], colors1
[chan
],
1541 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
1544 lp_build_endif(&if_ctx
);
1550 * Sample the texture/mipmap using given mip filter, and using
1551 * both nearest and linear filtering at the same time depending
1553 * lod can be per quad but linear_mask is always per pixel.
1554 * ilevel0 and ilevel1 indicate the two mipmap levels to sample
1555 * from (vectors or scalars).
1556 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1559 lp_build_sample_mipmap_both(struct lp_build_sample_context
*bld
,
1560 LLVMValueRef linear_mask
,
1561 unsigned mip_filter
,
1562 LLVMValueRef
*coords
,
1563 const LLVMValueRef
*offsets
,
1564 LLVMValueRef ilevel0
,
1565 LLVMValueRef ilevel1
,
1566 LLVMValueRef lod_fpart
,
1567 LLVMValueRef lod_positive
,
1568 LLVMValueRef
*colors_out
)
1570 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1571 LLVMValueRef size0
= NULL
;
1572 LLVMValueRef size1
= NULL
;
1573 LLVMValueRef row_stride0_vec
= NULL
;
1574 LLVMValueRef row_stride1_vec
= NULL
;
1575 LLVMValueRef img_stride0_vec
= NULL
;
1576 LLVMValueRef img_stride1_vec
= NULL
;
1577 LLVMValueRef data_ptr0
= NULL
;
1578 LLVMValueRef data_ptr1
= NULL
;
1579 LLVMValueRef mipoff0
= NULL
;
1580 LLVMValueRef mipoff1
= NULL
;
1581 LLVMValueRef colors0
[4], colors1
[4];
1584 /* sample the first mipmap level */
1585 lp_build_mipmap_level_sizes(bld
, ilevel0
,
1587 &row_stride0_vec
, &img_stride0_vec
);
1588 if (bld
->num_mips
== 1) {
1589 data_ptr0
= lp_build_get_mipmap_level(bld
, ilevel0
);
1592 /* This path should work for num_lods 1 too but slightly less efficient */
1593 data_ptr0
= bld
->base_ptr
;
1594 mipoff0
= lp_build_get_mip_offsets(bld
, ilevel0
);
1597 lp_build_sample_image_linear(bld
, size0
, linear_mask
,
1598 row_stride0_vec
, img_stride0_vec
,
1599 data_ptr0
, mipoff0
, coords
, offsets
,
1602 /* Store the first level's colors in the output variables */
1603 for (chan
= 0; chan
< 4; chan
++) {
1604 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
1607 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
1608 struct lp_build_if_state if_ctx
;
1609 LLVMValueRef need_lerp
;
1612 * We'll do mip filtering if any of the quads (or individual
1613 * pixel in case of per-pixel lod) need it.
1614 * Note using lod_positive here not lod_fpart since it may be the same
1615 * condition as that used in the outer "if" in the caller hence llvm
1616 * should be able to merge the branches in this case.
1618 need_lerp
= lp_build_any_true_range(&bld
->lodi_bld
, bld
->num_lods
, lod_positive
);
1620 lp_build_if(&if_ctx
, bld
->gallivm
, need_lerp
);
1623 * We unfortunately need to clamp lod_fpart here since we can get
1624 * negative values which would screw up filtering if not all
1625 * lod_fpart values have same sign.
1627 lod_fpart
= lp_build_max(&bld
->lodf_bld
, lod_fpart
,
1628 bld
->lodf_bld
.zero
);
1629 /* sample the second mipmap level */
1630 lp_build_mipmap_level_sizes(bld
, ilevel1
,
1632 &row_stride1_vec
, &img_stride1_vec
);
1633 if (bld
->num_mips
== 1) {
1634 data_ptr1
= lp_build_get_mipmap_level(bld
, ilevel1
);
1637 data_ptr1
= bld
->base_ptr
;
1638 mipoff1
= lp_build_get_mip_offsets(bld
, ilevel1
);
1641 lp_build_sample_image_linear(bld
, size1
, linear_mask
,
1642 row_stride1_vec
, img_stride1_vec
,
1643 data_ptr1
, mipoff1
, coords
, offsets
,
1646 /* interpolate samples from the two mipmap levels */
1648 if (bld
->num_lods
!= bld
->coord_type
.length
)
1649 lod_fpart
= lp_build_unpack_broadcast_aos_scalars(bld
->gallivm
,
1651 bld
->texel_bld
.type
,
1654 for (chan
= 0; chan
< 4; chan
++) {
1655 colors0
[chan
] = lp_build_lerp(&bld
->texel_bld
, lod_fpart
,
1656 colors0
[chan
], colors1
[chan
],
1658 LLVMBuildStore(builder
, colors0
[chan
], colors_out
[chan
]);
1661 lp_build_endif(&if_ctx
);
1667 * Build (per-coord) layer value.
1668 * Either clamp layer to valid values or fill in optional out_of_bounds
1669 * value and just return value unclamped.
1672 lp_build_layer_coord(struct lp_build_sample_context
*bld
,
1673 unsigned texture_unit
,
1674 boolean is_cube_array
,
1676 LLVMValueRef
*out_of_bounds
)
1678 LLVMValueRef num_layers
;
1679 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
1681 num_layers
= bld
->dynamic_state
->depth(bld
->dynamic_state
,
1682 bld
->gallivm
, texture_unit
);
1684 if (out_of_bounds
) {
1685 LLVMValueRef out1
, out
;
1686 assert(!is_cube_array
);
1687 num_layers
= lp_build_broadcast_scalar(int_coord_bld
, num_layers
);
1688 out
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, layer
, int_coord_bld
->zero
);
1689 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, layer
, num_layers
);
1690 *out_of_bounds
= lp_build_or(int_coord_bld
, out
, out1
);
1694 LLVMValueRef maxlayer
;
1695 LLVMValueRef s
= is_cube_array
? lp_build_const_int32(bld
->gallivm
, 6) :
1697 maxlayer
= lp_build_sub(&bld
->int_bld
, num_layers
, s
);
1698 maxlayer
= lp_build_broadcast_scalar(int_coord_bld
, maxlayer
);
1699 return lp_build_clamp(int_coord_bld
, layer
, int_coord_bld
->zero
, maxlayer
);
1705 * Calculate cube face, lod, mip levels.
1708 lp_build_sample_common(struct lp_build_sample_context
*bld
,
1709 unsigned texture_index
,
1710 unsigned sampler_index
,
1711 LLVMValueRef
*coords
,
1712 const struct lp_derivatives
*derivs
, /* optional */
1713 LLVMValueRef lod_bias
, /* optional */
1714 LLVMValueRef explicit_lod
, /* optional */
1715 LLVMValueRef
*lod_pos_or_zero
,
1716 LLVMValueRef
*lod_fpart
,
1717 LLVMValueRef
*ilevel0
,
1718 LLVMValueRef
*ilevel1
)
1720 const unsigned mip_filter
= bld
->static_sampler_state
->min_mip_filter
;
1721 const unsigned min_filter
= bld
->static_sampler_state
->min_img_filter
;
1722 const unsigned mag_filter
= bld
->static_sampler_state
->mag_img_filter
;
1723 const unsigned target
= bld
->static_texture_state
->target
;
1724 LLVMValueRef first_level
, cube_rho
= NULL
;
1725 LLVMValueRef lod_ipart
= NULL
;
1726 struct lp_derivatives cube_derivs
;
1729 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1730 mip_filter, min_filter, mag_filter);
1734 * Choose cube face, recompute texcoords for the chosen face and
1735 * compute rho here too (as it requires transform of derivatives).
1737 if (target
== PIPE_TEXTURE_CUBE
|| target
== PIPE_TEXTURE_CUBE_ARRAY
) {
1738 boolean need_derivs
;
1739 need_derivs
= ((min_filter
!= mag_filter
||
1740 mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) &&
1741 !bld
->static_sampler_state
->min_max_lod_equal
&&
1743 lp_build_cube_lookup(bld
, coords
, derivs
, &cube_rho
, &cube_derivs
, need_derivs
);
1744 derivs
= &cube_derivs
;
1745 if (target
== PIPE_TEXTURE_CUBE_ARRAY
) {
1746 /* calculate cube layer coord now */
1747 LLVMValueRef layer
= lp_build_iround(&bld
->coord_bld
, coords
[3]);
1748 LLVMValueRef six
= lp_build_const_int_vec(bld
->gallivm
, bld
->int_coord_type
, 6);
1749 layer
= lp_build_mul(&bld
->int_coord_bld
, layer
, six
);
1750 coords
[3] = lp_build_layer_coord(bld
, texture_index
, TRUE
, layer
, NULL
);
1751 /* because of seamless filtering can't add it to face (coords[2]) here. */
1754 else if (target
== PIPE_TEXTURE_1D_ARRAY
||
1755 target
== PIPE_TEXTURE_2D_ARRAY
) {
1756 coords
[2] = lp_build_iround(&bld
->coord_bld
, coords
[2]);
1757 coords
[2] = lp_build_layer_coord(bld
, texture_index
, FALSE
, coords
[2], NULL
);
1760 if (bld
->static_sampler_state
->compare_mode
!= PIPE_TEX_COMPARE_NONE
) {
1762 * Clamp p coords to [0,1] for fixed function depth texture format here.
1763 * Technically this is not entirely correct for unorm depth as the ref value
1764 * should be converted to the depth format (quantization!) and comparison
1765 * then done in texture format. This would actually help performance (since
1766 * only need to do it once and could save the per-sample conversion of texels
1767 * to floats instead), but it would need more messy code (would need to push
1768 * at least some bits down to actual fetch so conversion could be skipped,
1769 * and would have ugly interaction with border color, would need to convert
1770 * border color to that format too or do some other tricks to make it work).
1772 const struct util_format_description
*format_desc
= bld
->format_desc
;
1774 /* not entirely sure we couldn't end up with non-valid swizzle here */
1775 chan_type
= format_desc
->swizzle
[0] <= UTIL_FORMAT_SWIZZLE_W
?
1776 format_desc
->channel
[format_desc
->swizzle
[0]].type
:
1777 UTIL_FORMAT_TYPE_FLOAT
;
1778 if (chan_type
!= UTIL_FORMAT_TYPE_FLOAT
) {
1779 coords
[4] = lp_build_clamp(&bld
->coord_bld
, coords
[4],
1780 bld
->coord_bld
.zero
, bld
->coord_bld
.one
);
1785 * Compute the level of detail (float).
1787 if (min_filter
!= mag_filter
||
1788 mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) {
1789 /* Need to compute lod either to choose mipmap levels or to
1790 * distinguish between minification/magnification with one mipmap level.
1792 lp_build_lod_selector(bld
, texture_index
, sampler_index
,
1793 coords
[0], coords
[1], coords
[2], cube_rho
,
1794 derivs
, lod_bias
, explicit_lod
,
1796 &lod_ipart
, lod_fpart
, lod_pos_or_zero
);
1798 lod_ipart
= bld
->lodi_bld
.zero
;
1799 *lod_pos_or_zero
= bld
->lodi_bld
.zero
;
1802 if (bld
->num_lods
!= bld
->num_mips
) {
1803 /* only makes sense if there's just a single mip level */
1804 assert(bld
->num_mips
== 1);
1805 lod_ipart
= lp_build_extract_range(bld
->gallivm
, lod_ipart
, 0, 1);
1809 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1811 switch (mip_filter
) {
1813 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1815 case PIPE_TEX_MIPFILTER_NONE
:
1816 /* always use mip level 0 */
1817 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
1818 bld
->gallivm
, texture_index
);
1819 first_level
= lp_build_broadcast_scalar(&bld
->leveli_bld
, first_level
);
1820 *ilevel0
= first_level
;
1822 case PIPE_TEX_MIPFILTER_NEAREST
:
1824 lp_build_nearest_mip_level(bld
, texture_index
, lod_ipart
, ilevel0
, NULL
);
1826 case PIPE_TEX_MIPFILTER_LINEAR
:
1829 lp_build_linear_mip_levels(bld
, texture_index
,
1830 lod_ipart
, lod_fpart
,
1837 lp_build_clamp_border_color(struct lp_build_sample_context
*bld
,
1838 unsigned sampler_unit
)
1840 struct gallivm_state
*gallivm
= bld
->gallivm
;
1841 LLVMBuilderRef builder
= gallivm
->builder
;
1842 LLVMValueRef border_color_ptr
=
1843 bld
->dynamic_state
->border_color(bld
->dynamic_state
,
1844 gallivm
, sampler_unit
);
1845 LLVMValueRef border_color
;
1846 const struct util_format_description
*format_desc
= bld
->format_desc
;
1847 struct lp_type vec4_type
= bld
->texel_type
;
1848 struct lp_build_context vec4_bld
;
1849 LLVMValueRef min_clamp
= NULL
;
1850 LLVMValueRef max_clamp
= NULL
;
1853 * For normalized format need to clamp border color (technically
1854 * probably should also quantize the data). Really sucks doing this
1855 * here but can't avoid at least for now since this is part of
1856 * sampler state and texture format is part of sampler_view state.
1857 * GL expects also expects clamping for uint/sint formats too so
1858 * do that as well (d3d10 can't end up here with uint/sint since it
1859 * only supports them with ld).
1861 vec4_type
.length
= 4;
1862 lp_build_context_init(&vec4_bld
, gallivm
, vec4_type
);
1865 * Vectorized clamping of border color. Loading is a bit of a hack since
1866 * we just cast the pointer to float array to pointer to vec4
1869 border_color_ptr
= lp_build_array_get_ptr(gallivm
, border_color_ptr
,
1870 lp_build_const_int32(gallivm
, 0));
1871 border_color_ptr
= LLVMBuildBitCast(builder
, border_color_ptr
,
1872 LLVMPointerType(vec4_bld
.vec_type
, 0), "");
1873 border_color
= LLVMBuildLoad(builder
, border_color_ptr
, "");
1874 /* we don't have aligned type in the dynamic state unfortunately */
1875 lp_set_load_alignment(border_color
, 4);
1878 * Instead of having some incredibly complex logic which will try to figure out
1879 * clamping necessary for each channel, simply use the first channel, and treat
1880 * mixed signed/unsigned normalized formats specially.
1881 * (Mixed non-normalized, which wouldn't work at all here, do not exist for a
1884 if (format_desc
->layout
== UTIL_FORMAT_LAYOUT_PLAIN
) {
1886 /* d/s needs special handling because both present means just sampling depth */
1887 if (util_format_is_depth_and_stencil(format_desc
->format
)) {
1888 chan
= format_desc
->swizzle
[0];
1891 chan
= util_format_get_first_non_void_channel(format_desc
->format
);
1893 if (chan
>= 0 && chan
<= UTIL_FORMAT_SWIZZLE_W
) {
1894 unsigned chan_type
= format_desc
->channel
[chan
].type
;
1895 unsigned chan_norm
= format_desc
->channel
[chan
].normalized
;
1896 unsigned chan_pure
= format_desc
->channel
[chan
].pure_integer
;
1897 if (chan_type
== UTIL_FORMAT_TYPE_SIGNED
) {
1899 min_clamp
= lp_build_const_vec(gallivm
, vec4_type
, -1.0F
);
1900 max_clamp
= vec4_bld
.one
;
1902 else if (chan_pure
) {
1904 * Border color was stored as int, hence need min/max clamp
1905 * only if chan has less than 32 bits..
1907 unsigned chan_size
= format_desc
->channel
[chan
].size
;
1908 if (chan_size
< 32) {
1909 min_clamp
= lp_build_const_int_vec(gallivm
, vec4_type
,
1910 0 - (1 << (chan_size
- 1)));
1911 max_clamp
= lp_build_const_int_vec(gallivm
, vec4_type
,
1912 (1 << (chan_size
- 1)) - 1);
1915 /* TODO: no idea about non-pure, non-normalized! */
1917 else if (chan_type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
1919 min_clamp
= vec4_bld
.zero
;
1920 max_clamp
= vec4_bld
.one
;
1923 * Need a ugly hack here, because we don't have Z32_FLOAT_X8X24
1924 * we use Z32_FLOAT_S8X24 to imply sampling depth component
1925 * and ignoring stencil, which will blow up here if we try to
1926 * do a uint clamp in a float texel build...
1927 * And even if we had that format, mesa st also thinks using z24s8
1928 * means depth sampling ignoring stencil.
1930 else if (chan_pure
) {
1932 * Border color was stored as uint, hence never need min
1933 * clamp, and only need max clamp if chan has less than 32 bits.
1935 unsigned chan_size
= format_desc
->channel
[chan
].size
;
1936 if (chan_size
< 32) {
1937 max_clamp
= lp_build_const_int_vec(gallivm
, vec4_type
,
1938 (1 << chan_size
) - 1);
1940 /* TODO: no idea about non-pure, non-normalized! */
1943 else if (chan_type
== UTIL_FORMAT_TYPE_FIXED
) {
1944 /* TODO: I have no idea what clamp this would need if any! */
1947 /* mixed plain formats (or different pure size) */
1948 switch (format_desc
->format
) {
1949 case PIPE_FORMAT_B10G10R10A2_UINT
:
1950 case PIPE_FORMAT_R10G10B10A2_UINT
:
1952 unsigned max10
= (1 << 10) - 1;
1953 max_clamp
= lp_build_const_aos(gallivm
, vec4_type
, max10
, max10
,
1954 max10
, (1 << 2) - 1, NULL
);
1957 case PIPE_FORMAT_R10SG10SB10SA2U_NORM
:
1958 min_clamp
= lp_build_const_aos(gallivm
, vec4_type
, -1.0F
, -1.0F
,
1960 max_clamp
= vec4_bld
.one
;
1962 case PIPE_FORMAT_R8SG8SB8UX8U_NORM
:
1963 case PIPE_FORMAT_R5SG5SB6U_NORM
:
1964 min_clamp
= lp_build_const_aos(gallivm
, vec4_type
, -1.0F
, -1.0F
,
1966 max_clamp
= vec4_bld
.one
;
1973 /* cannot figure this out from format description */
1974 if (format_desc
->layout
== UTIL_FORMAT_LAYOUT_S3TC
) {
1975 /* s3tc formats are always unorm */
1976 min_clamp
= vec4_bld
.zero
;
1977 max_clamp
= vec4_bld
.one
;
1979 else if (format_desc
->layout
== UTIL_FORMAT_LAYOUT_RGTC
||
1980 format_desc
->layout
== UTIL_FORMAT_LAYOUT_ETC
) {
1981 switch (format_desc
->format
) {
1982 case PIPE_FORMAT_RGTC1_UNORM
:
1983 case PIPE_FORMAT_RGTC2_UNORM
:
1984 case PIPE_FORMAT_LATC1_UNORM
:
1985 case PIPE_FORMAT_LATC2_UNORM
:
1986 case PIPE_FORMAT_ETC1_RGB8
:
1987 min_clamp
= vec4_bld
.zero
;
1988 max_clamp
= vec4_bld
.one
;
1990 case PIPE_FORMAT_RGTC1_SNORM
:
1991 case PIPE_FORMAT_RGTC2_SNORM
:
1992 case PIPE_FORMAT_LATC1_SNORM
:
1993 case PIPE_FORMAT_LATC2_SNORM
:
1994 min_clamp
= lp_build_const_vec(gallivm
, vec4_type
, -1.0F
);
1995 max_clamp
= vec4_bld
.one
;
2003 * all others from subsampled/other group, though we don't care
2004 * about yuv (and should not have any from zs here)
2006 else if (format_desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_YUV
){
2007 switch (format_desc
->format
) {
2008 case PIPE_FORMAT_R8G8_B8G8_UNORM
:
2009 case PIPE_FORMAT_G8R8_G8B8_UNORM
:
2010 case PIPE_FORMAT_G8R8_B8R8_UNORM
:
2011 case PIPE_FORMAT_R8G8_R8B8_UNORM
:
2012 case PIPE_FORMAT_R1_UNORM
: /* doesn't make sense but ah well */
2013 min_clamp
= vec4_bld
.zero
;
2014 max_clamp
= vec4_bld
.one
;
2016 case PIPE_FORMAT_R8G8Bx_SNORM
:
2017 min_clamp
= lp_build_const_vec(gallivm
, vec4_type
, -1.0F
);
2018 max_clamp
= vec4_bld
.one
;
2021 * Note smallfloat formats usually don't need clamping
2022 * (they still have infinite range) however this is not
2023 * true for r11g11b10 and r9g9b9e5, which can't represent
2024 * negative numbers (and additionally r9g9b9e5 can't represent
2025 * very large numbers). d3d10 seems happy without clamping in
2026 * this case, but gl spec is pretty clear: "for floating
2027 * point and integer formats, border values are clamped to
2028 * the representable range of the format" so do that here.
2030 case PIPE_FORMAT_R11G11B10_FLOAT
:
2031 min_clamp
= vec4_bld
.zero
;
2033 case PIPE_FORMAT_R9G9B9E5_FLOAT
:
2034 min_clamp
= vec4_bld
.zero
;
2035 max_clamp
= lp_build_const_vec(gallivm
, vec4_type
, MAX_RGB9E5
);
2045 border_color
= lp_build_max(&vec4_bld
, border_color
, min_clamp
);
2048 border_color
= lp_build_min(&vec4_bld
, border_color
, max_clamp
);
2051 bld
->border_color_clamped
= border_color
;
2056 * General texture sampling codegen.
2057 * This function handles texture sampling for all texture targets (1D,
2058 * 2D, 3D, cube) and all filtering modes.
2061 lp_build_sample_general(struct lp_build_sample_context
*bld
,
2062 unsigned sampler_unit
,
2063 LLVMValueRef
*coords
,
2064 const LLVMValueRef
*offsets
,
2065 LLVMValueRef lod_positive
,
2066 LLVMValueRef lod_fpart
,
2067 LLVMValueRef ilevel0
,
2068 LLVMValueRef ilevel1
,
2069 LLVMValueRef
*colors_out
)
2071 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
2072 const struct lp_static_sampler_state
*sampler_state
= bld
->static_sampler_state
;
2073 const unsigned mip_filter
= sampler_state
->min_mip_filter
;
2074 const unsigned min_filter
= sampler_state
->min_img_filter
;
2075 const unsigned mag_filter
= sampler_state
->mag_img_filter
;
2076 LLVMValueRef texels
[4];
2079 /* if we need border color, (potentially) clamp it now */
2080 if (lp_sampler_wrap_mode_uses_border_color(sampler_state
->wrap_s
,
2084 lp_sampler_wrap_mode_uses_border_color(sampler_state
->wrap_t
,
2088 lp_sampler_wrap_mode_uses_border_color(sampler_state
->wrap_r
,
2091 lp_build_clamp_border_color(bld
, sampler_unit
);
2096 * Get/interpolate texture colors.
2099 for (chan
= 0; chan
< 4; ++chan
) {
2100 texels
[chan
] = lp_build_alloca(bld
->gallivm
, bld
->texel_bld
.vec_type
, "");
2101 lp_build_name(texels
[chan
], "sampler%u_texel_%c_var", sampler_unit
, "xyzw"[chan
]);
2104 if (min_filter
== mag_filter
) {
2105 /* no need to distinguish between minification and magnification */
2106 lp_build_sample_mipmap(bld
, min_filter
, mip_filter
,
2108 ilevel0
, ilevel1
, lod_fpart
,
2113 * Could also get rid of the if-logic and always use mipmap_both, both
2114 * for the single lod and multi-lod case if nothing really uses this.
2116 if (bld
->num_lods
== 1) {
2117 /* Emit conditional to choose min image filter or mag image filter
2118 * depending on the lod being > 0 or <= 0, respectively.
2120 struct lp_build_if_state if_ctx
;
2122 lod_positive
= LLVMBuildTrunc(builder
, lod_positive
,
2123 LLVMInt1TypeInContext(bld
->gallivm
->context
), "");
2125 lp_build_if(&if_ctx
, bld
->gallivm
, lod_positive
);
2127 /* Use the minification filter */
2128 lp_build_sample_mipmap(bld
, min_filter
, mip_filter
,
2130 ilevel0
, ilevel1
, lod_fpart
,
2133 lp_build_else(&if_ctx
);
2135 /* Use the magnification filter */
2136 lp_build_sample_mipmap(bld
, mag_filter
, PIPE_TEX_MIPFILTER_NONE
,
2138 ilevel0
, NULL
, NULL
,
2141 lp_build_endif(&if_ctx
);
2144 LLVMValueRef need_linear
, linear_mask
;
2145 unsigned mip_filter_for_nearest
;
2146 struct lp_build_if_state if_ctx
;
2148 if (min_filter
== PIPE_TEX_FILTER_LINEAR
) {
2149 linear_mask
= lod_positive
;
2150 mip_filter_for_nearest
= PIPE_TEX_MIPFILTER_NONE
;
2153 linear_mask
= lp_build_not(&bld
->lodi_bld
, lod_positive
);
2154 mip_filter_for_nearest
= mip_filter
;
2156 need_linear
= lp_build_any_true_range(&bld
->lodi_bld
, bld
->num_lods
,
2159 if (bld
->num_lods
!= bld
->coord_type
.length
) {
2160 linear_mask
= lp_build_unpack_broadcast_aos_scalars(bld
->gallivm
,
2162 bld
->int_coord_type
,
2166 lp_build_if(&if_ctx
, bld
->gallivm
, need_linear
);
2169 * Do sampling with both filters simultaneously. This means using
2170 * a linear filter and doing some tricks (with weights) for the pixels
2171 * which need nearest filter.
2172 * Note that it's probably rare some pixels need nearest and some
2173 * linear filter but the fixups required for the nearest pixels
2174 * aren't all that complicated so just always run a combined path
2175 * if at least some pixels require linear.
2177 lp_build_sample_mipmap_both(bld
, linear_mask
, mip_filter
,
2180 lod_fpart
, lod_positive
,
2183 lp_build_else(&if_ctx
);
2186 * All pixels require just nearest filtering, which is way
2187 * cheaper than linear, hence do a separate path for that.
2189 lp_build_sample_mipmap(bld
, PIPE_TEX_FILTER_NEAREST
,
2190 mip_filter_for_nearest
,
2192 ilevel0
, ilevel1
, lod_fpart
,
2195 lp_build_endif(&if_ctx
);
2199 for (chan
= 0; chan
< 4; ++chan
) {
2200 colors_out
[chan
] = LLVMBuildLoad(builder
, texels
[chan
], "");
2201 lp_build_name(colors_out
[chan
], "sampler%u_texel_%c", sampler_unit
, "xyzw"[chan
]);
2207 * Texel fetch function.
2208 * In contrast to general sampling there is no filtering, no coord minification,
2209 * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
2210 * directly to be applied to the selected mip level (after adding texel offsets).
2211 * This function handles texel fetch for all targets where texel fetch is supported
2212 * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
2215 lp_build_fetch_texel(struct lp_build_sample_context
*bld
,
2216 unsigned texture_unit
,
2217 const LLVMValueRef
*coords
,
2218 LLVMValueRef explicit_lod
,
2219 const LLVMValueRef
*offsets
,
2220 LLVMValueRef
*colors_out
)
2222 struct lp_build_context
*perquadi_bld
= &bld
->lodi_bld
;
2223 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
2224 unsigned dims
= bld
->dims
, chan
;
2225 unsigned target
= bld
->static_texture_state
->target
;
2226 boolean out_of_bound_ret_zero
= TRUE
;
2227 LLVMValueRef size
, ilevel
;
2228 LLVMValueRef row_stride_vec
= NULL
, img_stride_vec
= NULL
;
2229 LLVMValueRef x
= coords
[0], y
= coords
[1], z
= coords
[2];
2230 LLVMValueRef width
, height
, depth
, i
, j
;
2231 LLVMValueRef offset
, out_of_bounds
, out1
;
2233 out_of_bounds
= int_coord_bld
->zero
;
2235 if (explicit_lod
&& bld
->static_texture_state
->target
!= PIPE_BUFFER
) {
2236 if (bld
->num_mips
!= int_coord_bld
->type
.length
) {
2237 ilevel
= lp_build_pack_aos_scalars(bld
->gallivm
, int_coord_bld
->type
,
2238 perquadi_bld
->type
, explicit_lod
, 0);
2241 ilevel
= explicit_lod
;
2243 lp_build_nearest_mip_level(bld
, texture_unit
, ilevel
, &ilevel
,
2244 out_of_bound_ret_zero
? &out_of_bounds
: NULL
);
2247 assert(bld
->num_mips
== 1);
2248 if (bld
->static_texture_state
->target
!= PIPE_BUFFER
) {
2249 ilevel
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
2250 bld
->gallivm
, texture_unit
);
2253 ilevel
= lp_build_const_int32(bld
->gallivm
, 0);
2256 lp_build_mipmap_level_sizes(bld
, ilevel
,
2258 &row_stride_vec
, &img_stride_vec
);
2259 lp_build_extract_image_sizes(bld
, &bld
->int_size_bld
, int_coord_bld
->type
,
2260 size
, &width
, &height
, &depth
);
2262 if (target
== PIPE_TEXTURE_1D_ARRAY
||
2263 target
== PIPE_TEXTURE_2D_ARRAY
) {
2264 if (out_of_bound_ret_zero
) {
2265 z
= lp_build_layer_coord(bld
, texture_unit
, FALSE
, z
, &out1
);
2266 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2269 z
= lp_build_layer_coord(bld
, texture_unit
, FALSE
, z
, NULL
);
2273 /* This is a lot like border sampling */
2276 * coords are really unsigned, offsets are signed, but I don't think
2277 * exceeding 31 bits is possible
2279 x
= lp_build_add(int_coord_bld
, x
, offsets
[0]);
2281 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, x
, int_coord_bld
->zero
);
2282 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2283 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, x
, width
);
2284 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2288 y
= lp_build_add(int_coord_bld
, y
, offsets
[1]);
2290 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, y
, int_coord_bld
->zero
);
2291 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2292 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, y
, height
);
2293 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2297 z
= lp_build_add(int_coord_bld
, z
, offsets
[2]);
2299 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_LESS
, z
, int_coord_bld
->zero
);
2300 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2301 out1
= lp_build_cmp(int_coord_bld
, PIPE_FUNC_GEQUAL
, z
, depth
);
2302 out_of_bounds
= lp_build_or(int_coord_bld
, out_of_bounds
, out1
);
2306 lp_build_sample_offset(int_coord_bld
,
2308 x
, y
, z
, row_stride_vec
, img_stride_vec
,
2311 if (bld
->static_texture_state
->target
!= PIPE_BUFFER
) {
2312 offset
= lp_build_add(int_coord_bld
, offset
,
2313 lp_build_get_mip_offsets(bld
, ilevel
));
2316 offset
= lp_build_andnot(int_coord_bld
, offset
, out_of_bounds
);
2318 lp_build_fetch_rgba_soa(bld
->gallivm
,
2321 bld
->base_ptr
, offset
,
2325 if (out_of_bound_ret_zero
) {
2327 * Only needed for ARB_robust_buffer_access_behavior and d3d10.
2328 * Could use min/max above instead of out-of-bounds comparisons
2329 * if we don't care about the result returned for out-of-bounds.
2331 for (chan
= 0; chan
< 4; chan
++) {
2332 colors_out
[chan
] = lp_build_select(&bld
->texel_bld
, out_of_bounds
,
2333 bld
->texel_bld
.zero
, colors_out
[chan
]);
2340 * Just set texels to white instead of actually sampling the texture.
2344 lp_build_sample_nop(struct gallivm_state
*gallivm
,
2345 struct lp_type type
,
2346 const LLVMValueRef
*coords
,
2347 LLVMValueRef texel_out
[4])
2349 LLVMValueRef one
= lp_build_one(gallivm
, type
);
2352 for (chan
= 0; chan
< 4; chan
++) {
2353 texel_out
[chan
] = one
;
2359 * Build texture sampling code.
2360 * 'texel' will return a vector of four LLVMValueRefs corresponding to
2362 * \param type vector float type to use for coords, etc.
2363 * \param is_fetch if this is a texel fetch instruction.
2364 * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
2367 lp_build_sample_soa(struct gallivm_state
*gallivm
,
2368 const struct lp_static_texture_state
*static_texture_state
,
2369 const struct lp_static_sampler_state
*static_sampler_state
,
2370 struct lp_sampler_dynamic_state
*dynamic_state
,
2371 struct lp_type type
,
2373 unsigned texture_index
,
2374 unsigned sampler_index
,
2375 const LLVMValueRef
*coords
,
2376 const LLVMValueRef
*offsets
,
2377 const struct lp_derivatives
*derivs
, /* optional */
2378 LLVMValueRef lod_bias
, /* optional */
2379 LLVMValueRef explicit_lod
, /* optional */
2380 enum lp_sampler_lod_property lod_property
,
2381 LLVMValueRef texel_out
[4])
2383 unsigned target
= static_texture_state
->target
;
2384 unsigned dims
= texture_dims(target
);
2385 unsigned num_quads
= type
.length
/ 4;
2386 unsigned mip_filter
, min_img_filter
, mag_img_filter
, i
;
2387 struct lp_build_sample_context bld
;
2388 struct lp_static_sampler_state derived_sampler_state
= *static_sampler_state
;
2389 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
2390 LLVMBuilderRef builder
= gallivm
->builder
;
2391 LLVMValueRef tex_width
, newcoords
[5];
2394 enum pipe_format fmt
= static_texture_state
->format
;
2395 debug_printf("Sample from %s\n", util_format_name(fmt
));
2398 if (static_texture_state
->format
== PIPE_FORMAT_NONE
) {
2400 * If there's nothing bound, format is NONE, and we must return
2401 * all zero as mandated by d3d10 in this case.
2404 LLVMValueRef zero
= lp_build_const_vec(gallivm
, type
, 0.0F
);
2405 for (chan
= 0; chan
< 4; chan
++) {
2406 texel_out
[chan
] = zero
;
2411 assert(type
.floating
);
2413 /* Setup our build context */
2414 memset(&bld
, 0, sizeof bld
);
2415 bld
.gallivm
= gallivm
;
2416 bld
.static_sampler_state
= &derived_sampler_state
;
2417 bld
.static_texture_state
= static_texture_state
;
2418 bld
.dynamic_state
= dynamic_state
;
2419 bld
.format_desc
= util_format_description(static_texture_state
->format
);
2422 bld
.vector_width
= lp_type_width(type
);
2424 bld
.float_type
= lp_type_float(32);
2425 bld
.int_type
= lp_type_int(32);
2426 bld
.coord_type
= type
;
2427 bld
.int_coord_type
= lp_int_type(type
);
2428 bld
.float_size_in_type
= lp_type_float(32);
2429 bld
.float_size_in_type
.length
= dims
> 1 ? 4 : 1;
2430 bld
.int_size_in_type
= lp_int_type(bld
.float_size_in_type
);
2431 bld
.texel_type
= type
;
2433 /* always using the first channel hopefully should be safe,
2434 * if not things WILL break in other places anyway.
2436 if (bld
.format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
&&
2437 bld
.format_desc
->channel
[0].pure_integer
) {
2438 if (bld
.format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_SIGNED
) {
2439 bld
.texel_type
= lp_type_int_vec(type
.width
, type
.width
* type
.length
);
2441 else if (bld
.format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
2442 bld
.texel_type
= lp_type_uint_vec(type
.width
, type
.width
* type
.length
);
2445 else if (util_format_has_stencil(bld
.format_desc
) &&
2446 !util_format_has_depth(bld
.format_desc
)) {
2447 /* for stencil only formats, sample stencil (uint) */
2448 bld
.texel_type
= lp_type_int_vec(type
.width
, type
.width
* type
.length
);
2451 if (!static_texture_state
->level_zero_only
) {
2452 derived_sampler_state
.min_mip_filter
= static_sampler_state
->min_mip_filter
;
2454 derived_sampler_state
.min_mip_filter
= PIPE_TEX_MIPFILTER_NONE
;
2456 mip_filter
= derived_sampler_state
.min_mip_filter
;
2459 debug_printf(" .min_mip_filter = %u\n", derived_sampler_state
.min_mip_filter
);
2462 if (static_texture_state
->target
== PIPE_TEXTURE_CUBE
||
2463 static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
)
2466 * Seamless filtering ignores wrap modes.
2467 * Setting to CLAMP_TO_EDGE is correct for nearest filtering, for
2468 * bilinear it's not correct but way better than using for instance repeat.
2469 * Note we even set this for non-seamless. Technically GL allows any wrap
2470 * mode, which made sense when supporting true borders (can get seamless
2471 * effect with border and CLAMP_TO_BORDER), but gallium doesn't support
2472 * borders and d3d9 requires wrap modes to be ignored and it's a pain to fix
2473 * up the sampler state (as it makes it texture dependent).
2475 derived_sampler_state
.wrap_s
= PIPE_TEX_WRAP_CLAMP_TO_EDGE
;
2476 derived_sampler_state
.wrap_t
= PIPE_TEX_WRAP_CLAMP_TO_EDGE
;
2479 min_img_filter
= derived_sampler_state
.min_img_filter
;
2480 mag_img_filter
= derived_sampler_state
.mag_img_filter
;
2484 * This is all a bit complicated different paths are chosen for performance
2486 * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
2487 * everything (the last two options are equivalent for 4-wide case).
2488 * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
2489 * lod is calculated then the lod value extracted afterwards so making this
2490 * case basically the same as far as lod handling is concerned for the
2491 * further sample/filter code as the 1 lod for everything case.
2492 * Different lod handling mostly shows up when building mipmap sizes
2493 * (lp_build_mipmap_level_sizes() and friends) and also in filtering
2494 * (getting the fractional part of the lod to the right texels).
2498 * There are other situations where at least the multiple int lods could be
2499 * avoided like min and max lod being equal.
2501 bld
.num_mips
= bld
.num_lods
= 1;
2503 if ((gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) &&
2504 (gallivm_debug
& GALLIVM_DEBUG_NO_RHO_APPROX
) &&
2505 (static_texture_state
->target
== PIPE_TEXTURE_CUBE
||
2506 static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) &&
2507 (!is_fetch
&& mip_filter
!= PIPE_TEX_MIPFILTER_NONE
)) {
2509 * special case for using per-pixel lod even for implicit lod,
2510 * which is generally never required (ok by APIs) except to please
2511 * some (somewhat broken imho) tests (because per-pixel face selection
2512 * can cause derivatives to be different for pixels outside the primitive
2513 * due to the major axis division even if pre-project derivatives are
2516 bld
.num_mips
= type
.length
;
2517 bld
.num_lods
= type
.length
;
2519 else if (lod_property
== LP_SAMPLER_LOD_PER_ELEMENT
||
2520 (explicit_lod
|| lod_bias
|| derivs
)) {
2521 if ((is_fetch
&& target
!= PIPE_BUFFER
) ||
2522 (!is_fetch
&& mip_filter
!= PIPE_TEX_MIPFILTER_NONE
)) {
2523 bld
.num_mips
= type
.length
;
2524 bld
.num_lods
= type
.length
;
2526 else if (!is_fetch
&& min_img_filter
!= mag_img_filter
) {
2528 bld
.num_lods
= type
.length
;
2531 /* TODO: for true scalar_lod should only use 1 lod value */
2532 else if ((is_fetch
&& explicit_lod
&& target
!= PIPE_BUFFER
) ||
2533 (!is_fetch
&& mip_filter
!= PIPE_TEX_MIPFILTER_NONE
)) {
2534 bld
.num_mips
= num_quads
;
2535 bld
.num_lods
= num_quads
;
2537 else if (!is_fetch
&& min_img_filter
!= mag_img_filter
) {
2539 bld
.num_lods
= num_quads
;
2543 bld
.lodf_type
= type
;
2544 /* we want native vector size to be able to use our intrinsics */
2545 if (bld
.num_lods
!= type
.length
) {
2546 /* TODO: this currently always has to be per-quad or per-element */
2547 bld
.lodf_type
.length
= type
.length
> 4 ? ((type
.length
+ 15) / 16) * 4 : 1;
2549 bld
.lodi_type
= lp_int_type(bld
.lodf_type
);
2550 bld
.levelf_type
= bld
.lodf_type
;
2551 if (bld
.num_mips
== 1) {
2552 bld
.levelf_type
.length
= 1;
2554 bld
.leveli_type
= lp_int_type(bld
.levelf_type
);
2555 bld
.float_size_type
= bld
.float_size_in_type
;
2556 /* Note: size vectors may not be native. They contain minified w/h/d/_ values,
2557 * with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */
2558 if (bld
.num_mips
> 1) {
2559 bld
.float_size_type
.length
= bld
.num_mips
== type
.length
?
2560 bld
.num_mips
* bld
.float_size_in_type
.length
:
2563 bld
.int_size_type
= lp_int_type(bld
.float_size_type
);
2565 lp_build_context_init(&bld
.float_bld
, gallivm
, bld
.float_type
);
2566 lp_build_context_init(&bld
.float_vec_bld
, gallivm
, type
);
2567 lp_build_context_init(&bld
.int_bld
, gallivm
, bld
.int_type
);
2568 lp_build_context_init(&bld
.coord_bld
, gallivm
, bld
.coord_type
);
2569 lp_build_context_init(&bld
.int_coord_bld
, gallivm
, bld
.int_coord_type
);
2570 lp_build_context_init(&bld
.int_size_in_bld
, gallivm
, bld
.int_size_in_type
);
2571 lp_build_context_init(&bld
.float_size_in_bld
, gallivm
, bld
.float_size_in_type
);
2572 lp_build_context_init(&bld
.int_size_bld
, gallivm
, bld
.int_size_type
);
2573 lp_build_context_init(&bld
.float_size_bld
, gallivm
, bld
.float_size_type
);
2574 lp_build_context_init(&bld
.texel_bld
, gallivm
, bld
.texel_type
);
2575 lp_build_context_init(&bld
.levelf_bld
, gallivm
, bld
.levelf_type
);
2576 lp_build_context_init(&bld
.leveli_bld
, gallivm
, bld
.leveli_type
);
2577 lp_build_context_init(&bld
.lodf_bld
, gallivm
, bld
.lodf_type
);
2578 lp_build_context_init(&bld
.lodi_bld
, gallivm
, bld
.lodi_type
);
2580 /* Get the dynamic state */
2581 tex_width
= dynamic_state
->width(dynamic_state
, gallivm
, texture_index
);
2582 bld
.row_stride_array
= dynamic_state
->row_stride(dynamic_state
, gallivm
, texture_index
);
2583 bld
.img_stride_array
= dynamic_state
->img_stride(dynamic_state
, gallivm
, texture_index
);
2584 bld
.base_ptr
= dynamic_state
->base_ptr(dynamic_state
, gallivm
, texture_index
);
2585 bld
.mip_offsets
= dynamic_state
->mip_offsets(dynamic_state
, gallivm
, texture_index
);
2586 /* Note that mip_offsets is an array[level] of offsets to texture images */
2588 /* width, height, depth as single int vector */
2590 bld
.int_size
= tex_width
;
2593 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size_in_bld
.undef
,
2594 tex_width
, LLVMConstInt(i32t
, 0, 0), "");
2596 LLVMValueRef tex_height
=
2597 dynamic_state
->height(dynamic_state
, gallivm
, texture_index
);
2598 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size
,
2599 tex_height
, LLVMConstInt(i32t
, 1, 0), "");
2601 LLVMValueRef tex_depth
=
2602 dynamic_state
->depth(dynamic_state
, gallivm
, texture_index
);
2603 bld
.int_size
= LLVMBuildInsertElement(builder
, bld
.int_size
,
2604 tex_depth
, LLVMConstInt(i32t
, 2, 0), "");
2609 for (i
= 0; i
< 5; i
++) {
2610 newcoords
[i
] = coords
[i
];
2614 /* For debug: no-op texture sampling */
2615 lp_build_sample_nop(gallivm
,
2621 else if (is_fetch
) {
2622 lp_build_fetch_texel(&bld
, texture_index
, newcoords
,
2623 explicit_lod
, offsets
,
2628 LLVMValueRef lod_fpart
= NULL
, lod_positive
= NULL
;
2629 LLVMValueRef ilevel0
= NULL
, ilevel1
= NULL
;
2630 boolean use_aos
= util_format_fits_8unorm(bld
.format_desc
) &&
2631 /* not sure this is strictly needed or simply impossible */
2632 derived_sampler_state
.compare_mode
== PIPE_TEX_COMPARE_NONE
&&
2633 lp_is_simple_wrap_mode(derived_sampler_state
.wrap_s
);
2635 use_aos
&= bld
.num_lods
<= num_quads
||
2636 derived_sampler_state
.min_img_filter
==
2637 derived_sampler_state
.mag_img_filter
;
2639 use_aos
&= lp_is_simple_wrap_mode(derived_sampler_state
.wrap_t
);
2641 use_aos
&= lp_is_simple_wrap_mode(derived_sampler_state
.wrap_r
);
2644 if ((static_texture_state
->target
== PIPE_TEXTURE_CUBE
||
2645 static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) &&
2646 derived_sampler_state
.seamless_cube_map
&&
2647 (derived_sampler_state
.min_img_filter
== PIPE_TEX_FILTER_LINEAR
||
2648 derived_sampler_state
.mag_img_filter
== PIPE_TEX_FILTER_LINEAR
)) {
2649 /* theoretically possible with AoS filtering but not implemented (complex!) */
2653 if ((gallivm_debug
& GALLIVM_DEBUG_PERF
) &&
2654 !use_aos
&& util_format_fits_8unorm(bld
.format_desc
)) {
2655 debug_printf("%s: using floating point linear filtering for %s\n",
2656 __FUNCTION__
, bld
.format_desc
->short_name
);
2657 debug_printf(" min_img %d mag_img %d mip %d target %d seamless %d"
2658 " wraps %d wrapt %d wrapr %d\n",
2659 derived_sampler_state
.min_img_filter
,
2660 derived_sampler_state
.mag_img_filter
,
2661 derived_sampler_state
.min_mip_filter
,
2662 static_texture_state
->target
,
2663 derived_sampler_state
.seamless_cube_map
,
2664 derived_sampler_state
.wrap_s
,
2665 derived_sampler_state
.wrap_t
,
2666 derived_sampler_state
.wrap_r
);
2669 lp_build_sample_common(&bld
, texture_index
, sampler_index
,
2671 derivs
, lod_bias
, explicit_lod
,
2672 &lod_positive
, &lod_fpart
,
2673 &ilevel0
, &ilevel1
);
2675 if (use_aos
&& static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
2676 /* The aos path doesn't do seamless filtering so simply add cube layer
2679 newcoords
[2] = lp_build_add(&bld
.int_coord_bld
, newcoords
[2], newcoords
[3]);
2683 * we only try 8-wide sampling with soa as it appears to
2684 * be a loss with aos with AVX (but it should work, except
2685 * for conformance if min_filter != mag_filter if num_lods > 1).
2686 * (It should be faster if we'd support avx2)
2688 if (num_quads
== 1 || !use_aos
) {
2690 /* do sampling/filtering with fixed pt arithmetic */
2691 lp_build_sample_aos(&bld
, sampler_index
,
2692 newcoords
[0], newcoords
[1],
2694 offsets
, lod_positive
, lod_fpart
,
2700 lp_build_sample_general(&bld
, sampler_index
,
2702 lod_positive
, lod_fpart
,
2709 struct lp_build_sample_context bld4
;
2710 struct lp_type type4
= type
;
2712 LLVMValueRef texelout4
[4];
2713 LLVMValueRef texelouttmp
[4][LP_MAX_VECTOR_LENGTH
/16];
2717 /* Setup our build context */
2718 memset(&bld4
, 0, sizeof bld4
);
2719 bld4
.gallivm
= bld
.gallivm
;
2720 bld4
.static_texture_state
= bld
.static_texture_state
;
2721 bld4
.static_sampler_state
= bld
.static_sampler_state
;
2722 bld4
.dynamic_state
= bld
.dynamic_state
;
2723 bld4
.format_desc
= bld
.format_desc
;
2724 bld4
.dims
= bld
.dims
;
2725 bld4
.row_stride_array
= bld
.row_stride_array
;
2726 bld4
.img_stride_array
= bld
.img_stride_array
;
2727 bld4
.base_ptr
= bld
.base_ptr
;
2728 bld4
.mip_offsets
= bld
.mip_offsets
;
2729 bld4
.int_size
= bld
.int_size
;
2731 bld4
.vector_width
= lp_type_width(type4
);
2733 bld4
.float_type
= lp_type_float(32);
2734 bld4
.int_type
= lp_type_int(32);
2735 bld4
.coord_type
= type4
;
2736 bld4
.int_coord_type
= lp_int_type(type4
);
2737 bld4
.float_size_in_type
= lp_type_float(32);
2738 bld4
.float_size_in_type
.length
= dims
> 1 ? 4 : 1;
2739 bld4
.int_size_in_type
= lp_int_type(bld4
.float_size_in_type
);
2740 bld4
.texel_type
= bld
.texel_type
;
2741 bld4
.texel_type
.length
= 4;
2743 bld4
.num_mips
= bld4
.num_lods
= 1;
2744 if ((gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) &&
2745 (gallivm_debug
& GALLIVM_DEBUG_NO_RHO_APPROX
) &&
2746 (static_texture_state
->target
== PIPE_TEXTURE_CUBE
||
2747 static_texture_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
) &&
2748 (!is_fetch
&& mip_filter
!= PIPE_TEX_MIPFILTER_NONE
)) {
2749 bld4
.num_mips
= type4
.length
;
2750 bld4
.num_lods
= type4
.length
;
2752 if (lod_property
== LP_SAMPLER_LOD_PER_ELEMENT
&&
2753 (explicit_lod
|| lod_bias
|| derivs
)) {
2754 if ((is_fetch
&& target
!= PIPE_BUFFER
) ||
2755 (!is_fetch
&& mip_filter
!= PIPE_TEX_MIPFILTER_NONE
)) {
2756 bld4
.num_mips
= type4
.length
;
2757 bld4
.num_lods
= type4
.length
;
2759 else if (!is_fetch
&& min_img_filter
!= mag_img_filter
) {
2761 bld4
.num_lods
= type4
.length
;
2765 /* we want native vector size to be able to use our intrinsics */
2766 bld4
.lodf_type
= type4
;
2767 if (bld4
.num_lods
!= type4
.length
) {
2768 bld4
.lodf_type
.length
= 1;
2770 bld4
.lodi_type
= lp_int_type(bld4
.lodf_type
);
2771 bld4
.levelf_type
= type4
;
2772 if (bld4
.num_mips
!= type4
.length
) {
2773 bld4
.levelf_type
.length
= 1;
2775 bld4
.leveli_type
= lp_int_type(bld4
.levelf_type
);
2776 bld4
.float_size_type
= bld4
.float_size_in_type
;
2777 if (bld4
.num_mips
> 1) {
2778 bld4
.float_size_type
.length
= bld4
.num_mips
== type4
.length
?
2779 bld4
.num_mips
* bld4
.float_size_in_type
.length
:
2782 bld4
.int_size_type
= lp_int_type(bld4
.float_size_type
);
2784 lp_build_context_init(&bld4
.float_bld
, gallivm
, bld4
.float_type
);
2785 lp_build_context_init(&bld4
.float_vec_bld
, gallivm
, type4
);
2786 lp_build_context_init(&bld4
.int_bld
, gallivm
, bld4
.int_type
);
2787 lp_build_context_init(&bld4
.coord_bld
, gallivm
, bld4
.coord_type
);
2788 lp_build_context_init(&bld4
.int_coord_bld
, gallivm
, bld4
.int_coord_type
);
2789 lp_build_context_init(&bld4
.int_size_in_bld
, gallivm
, bld4
.int_size_in_type
);
2790 lp_build_context_init(&bld4
.float_size_in_bld
, gallivm
, bld4
.float_size_in_type
);
2791 lp_build_context_init(&bld4
.int_size_bld
, gallivm
, bld4
.int_size_type
);
2792 lp_build_context_init(&bld4
.float_size_bld
, gallivm
, bld4
.float_size_type
);
2793 lp_build_context_init(&bld4
.texel_bld
, gallivm
, bld4
.texel_type
);
2794 lp_build_context_init(&bld4
.levelf_bld
, gallivm
, bld4
.levelf_type
);
2795 lp_build_context_init(&bld4
.leveli_bld
, gallivm
, bld4
.leveli_type
);
2796 lp_build_context_init(&bld4
.lodf_bld
, gallivm
, bld4
.lodf_type
);
2797 lp_build_context_init(&bld4
.lodi_bld
, gallivm
, bld4
.lodi_type
);
2799 for (i
= 0; i
< num_quads
; i
++) {
2800 LLVMValueRef s4
, t4
, r4
;
2801 LLVMValueRef lod_positive4
, lod_fpart4
= NULL
;
2802 LLVMValueRef ilevel04
, ilevel14
= NULL
;
2803 LLVMValueRef offsets4
[4] = { NULL
};
2804 unsigned num_lods
= bld4
.num_lods
;
2806 s4
= lp_build_extract_range(gallivm
, newcoords
[0], 4*i
, 4);
2807 t4
= lp_build_extract_range(gallivm
, newcoords
[1], 4*i
, 4);
2808 r4
= lp_build_extract_range(gallivm
, newcoords
[2], 4*i
, 4);
2811 offsets4
[0] = lp_build_extract_range(gallivm
, offsets
[0], 4*i
, 4);
2813 offsets4
[1] = lp_build_extract_range(gallivm
, offsets
[1], 4*i
, 4);
2815 offsets4
[2] = lp_build_extract_range(gallivm
, offsets
[2], 4*i
, 4);
2819 lod_positive4
= lp_build_extract_range(gallivm
, lod_positive
, num_lods
* i
, num_lods
);
2820 ilevel04
= bld
.num_mips
== 1 ? ilevel0
:
2821 lp_build_extract_range(gallivm
, ilevel0
, num_lods
* i
, num_lods
);
2822 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
2823 ilevel14
= lp_build_extract_range(gallivm
, ilevel1
, num_lods
* i
, num_lods
);
2824 lod_fpart4
= lp_build_extract_range(gallivm
, lod_fpart
, num_lods
* i
, num_lods
);
2828 /* do sampling/filtering with fixed pt arithmetic */
2829 lp_build_sample_aos(&bld4
, sampler_index
,
2830 s4
, t4
, r4
, offsets4
,
2831 lod_positive4
, lod_fpart4
,
2837 /* this path is currently unreachable and hence might break easily... */
2838 LLVMValueRef newcoords4
[5];
2842 newcoords4
[3] = lp_build_extract_range(gallivm
, newcoords
[3], 4*i
, 4);
2843 newcoords4
[4] = lp_build_extract_range(gallivm
, newcoords
[4], 4*i
, 4);
2845 lp_build_sample_general(&bld4
, sampler_index
,
2846 newcoords4
, offsets4
,
2847 lod_positive4
, lod_fpart4
,
2851 for (j
= 0; j
< 4; j
++) {
2852 texelouttmp
[j
][i
] = texelout4
[j
];
2856 for (j
= 0; j
< 4; j
++) {
2857 texel_out
[j
] = lp_build_concat(gallivm
, texelouttmp
[j
], type4
, num_quads
);
2862 if (target
!= PIPE_BUFFER
) {
2863 apply_sampler_swizzle(&bld
, texel_out
);
2867 * texel type can be a (32bit) int/uint (for pure int formats only),
2868 * however we are expected to always return floats (storage is untyped).
2870 if (!bld
.texel_type
.floating
) {
2872 for (chan
= 0; chan
< 4; chan
++) {
2873 texel_out
[chan
] = LLVMBuildBitCast(builder
, texel_out
[chan
],
2874 lp_build_vec_type(gallivm
, type
), "");
2880 lp_build_size_query_soa(struct gallivm_state
*gallivm
,
2881 const struct lp_static_texture_state
*static_state
,
2882 struct lp_sampler_dynamic_state
*dynamic_state
,
2883 struct lp_type int_type
,
2884 unsigned texture_unit
,
2886 boolean is_sviewinfo
,
2887 enum lp_sampler_lod_property lod_property
,
2888 LLVMValueRef explicit_lod
,
2889 LLVMValueRef
*sizes_out
)
2891 LLVMValueRef lod
, level
, size
;
2892 LLVMValueRef first_level
= NULL
;
2895 unsigned num_lods
= 1;
2896 struct lp_build_context bld_int_vec4
;
2898 if (static_state
->format
== PIPE_FORMAT_NONE
) {
2900 * If there's nothing bound, format is NONE, and we must return
2901 * all zero as mandated by d3d10 in this case.
2904 LLVMValueRef zero
= lp_build_const_vec(gallivm
, int_type
, 0.0F
);
2905 for (chan
= 0; chan
< 4; chan
++) {
2906 sizes_out
[chan
] = zero
;
2912 * Do some sanity verification about bound texture and shader dcl target.
2913 * Not entirely sure what's possible but assume array/non-array
2914 * always compatible (probably not ok for OpenGL but d3d10 has no
2915 * distinction of arrays at the resource level).
2916 * Everything else looks bogus (though not entirely sure about rect/2d).
2917 * Currently disabled because it causes assertion failures if there's
2918 * nothing bound (or rather a dummy texture, not that this case would
2919 * return the right values).
2921 if (0 && static_state
->target
!= target
) {
2922 if (static_state
->target
== PIPE_TEXTURE_1D
)
2923 assert(target
== PIPE_TEXTURE_1D_ARRAY
);
2924 else if (static_state
->target
== PIPE_TEXTURE_1D_ARRAY
)
2925 assert(target
== PIPE_TEXTURE_1D
);
2926 else if (static_state
->target
== PIPE_TEXTURE_2D
)
2927 assert(target
== PIPE_TEXTURE_2D_ARRAY
);
2928 else if (static_state
->target
== PIPE_TEXTURE_2D_ARRAY
)
2929 assert(target
== PIPE_TEXTURE_2D
);
2930 else if (static_state
->target
== PIPE_TEXTURE_CUBE
)
2931 assert(target
== PIPE_TEXTURE_CUBE_ARRAY
);
2932 else if (static_state
->target
== PIPE_TEXTURE_CUBE_ARRAY
)
2933 assert(target
== PIPE_TEXTURE_CUBE
);
2938 dims
= texture_dims(target
);
2941 case PIPE_TEXTURE_1D_ARRAY
:
2942 case PIPE_TEXTURE_2D_ARRAY
:
2943 case PIPE_TEXTURE_CUBE_ARRAY
:
2951 assert(!int_type
.floating
);
2953 lp_build_context_init(&bld_int_vec4
, gallivm
, lp_type_int_vec(32, 128));
2956 /* FIXME: this needs to honor per-element lod */
2957 lod
= LLVMBuildExtractElement(gallivm
->builder
, explicit_lod
, lp_build_const_int32(gallivm
, 0), "");
2958 first_level
= dynamic_state
->first_level(dynamic_state
, gallivm
, texture_unit
);
2959 level
= LLVMBuildAdd(gallivm
->builder
, lod
, first_level
, "level");
2960 lod
= lp_build_broadcast_scalar(&bld_int_vec4
, level
);
2962 lod
= bld_int_vec4
.zero
;
2965 size
= bld_int_vec4
.undef
;
2967 size
= LLVMBuildInsertElement(gallivm
->builder
, size
,
2968 dynamic_state
->width(dynamic_state
, gallivm
, texture_unit
),
2969 lp_build_const_int32(gallivm
, 0), "");
2972 size
= LLVMBuildInsertElement(gallivm
->builder
, size
,
2973 dynamic_state
->height(dynamic_state
, gallivm
, texture_unit
),
2974 lp_build_const_int32(gallivm
, 1), "");
2978 size
= LLVMBuildInsertElement(gallivm
->builder
, size
,
2979 dynamic_state
->depth(dynamic_state
, gallivm
, texture_unit
),
2980 lp_build_const_int32(gallivm
, 2), "");
2983 size
= lp_build_minify(&bld_int_vec4
, size
, lod
, TRUE
);
2986 LLVMValueRef layers
= dynamic_state
->depth(dynamic_state
, gallivm
, texture_unit
);
2987 if (target
== PIPE_TEXTURE_CUBE_ARRAY
) {
2989 * It looks like GL wants number of cubes, d3d10.1 has it undefined?
2990 * Could avoid this by passing in number of cubes instead of total
2991 * number of layers (might make things easier elsewhere too).
2993 LLVMValueRef six
= lp_build_const_int32(gallivm
, 6);
2994 layers
= LLVMBuildSDiv(gallivm
->builder
, layers
, six
, "");
2996 size
= LLVMBuildInsertElement(gallivm
->builder
, size
, layers
,
2997 lp_build_const_int32(gallivm
, dims
), "");
3001 * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels)
3002 * if level is out of bounds (note this can't cover unbound texture
3003 * here, which also requires returning zero).
3005 if (explicit_lod
&& is_sviewinfo
) {
3006 LLVMValueRef last_level
, out
, out1
;
3007 struct lp_build_context leveli_bld
;
3009 /* everything is scalar for now */
3010 lp_build_context_init(&leveli_bld
, gallivm
, lp_type_int_vec(32, 32));
3011 last_level
= dynamic_state
->last_level(dynamic_state
, gallivm
, texture_unit
);
3013 out
= lp_build_cmp(&leveli_bld
, PIPE_FUNC_LESS
, level
, first_level
);
3014 out1
= lp_build_cmp(&leveli_bld
, PIPE_FUNC_GREATER
, level
, last_level
);
3015 out
= lp_build_or(&leveli_bld
, out
, out1
);
3016 if (num_lods
== 1) {
3017 out
= lp_build_broadcast_scalar(&bld_int_vec4
, out
);
3023 size
= lp_build_andnot(&bld_int_vec4
, size
, out
);
3025 for (i
= 0; i
< dims
+ (has_array
? 1 : 0); i
++) {
3026 sizes_out
[i
] = lp_build_extract_broadcast(gallivm
, bld_int_vec4
.type
, int_type
,
3028 lp_build_const_int32(gallivm
, i
));
3031 for (; i
< 4; i
++) {
3032 sizes_out
[i
] = lp_build_const_vec(gallivm
, int_type
, 0.0);
3037 * if there's no explicit_lod (buffers, rects) queries requiring nr of
3038 * mips would be illegal.
3040 if (is_sviewinfo
&& explicit_lod
) {
3041 struct lp_build_context bld_int_scalar
;
3042 LLVMValueRef num_levels
;
3043 lp_build_context_init(&bld_int_scalar
, gallivm
, lp_type_int(32));
3045 if (static_state
->level_zero_only
) {
3046 num_levels
= bld_int_scalar
.one
;
3049 LLVMValueRef last_level
;
3051 last_level
= dynamic_state
->last_level(dynamic_state
, gallivm
, texture_unit
);
3052 num_levels
= lp_build_sub(&bld_int_scalar
, last_level
, first_level
);
3053 num_levels
= lp_build_add(&bld_int_scalar
, num_levels
, bld_int_scalar
.one
);
3055 sizes_out
[3] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, int_type
),