Merge remote branch 'origin/7.8'
[profile/ivi/mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_sample_soa.c
1 /**************************************************************************
2  *
3  * Copyright 2009 VMware, Inc.
4  * All Rights Reserved.
5  *
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:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
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.
25  *
26  **************************************************************************/
27
28 /**
29  * @file
30  * Texture sampling -- SoA.
31  *
32  * @author Jose Fonseca <jfonseca@vmware.com>
33  */
34
35 #include "pipe/p_defines.h"
36 #include "pipe/p_state.h"
37 #include "util/u_debug.h"
38 #include "util/u_dump.h"
39 #include "util/u_memory.h"
40 #include "util/u_math.h"
41 #include "util/u_format.h"
42 #include "util/u_cpu_detect.h"
43 #include "lp_bld_debug.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_arit.h"
48 #include "lp_bld_logic.h"
49 #include "lp_bld_swizzle.h"
50 #include "lp_bld_pack.h"
51 #include "lp_bld_flow.h"
52 #include "lp_bld_format.h"
53 #include "lp_bld_sample.h"
54
55
56 /**
57  * Keep all information for sampling code generation in a single place.
58  */
59 struct lp_build_sample_context
60 {
61    LLVMBuilderRef builder;
62
63    const struct lp_sampler_static_state *static_state;
64
65    struct lp_sampler_dynamic_state *dynamic_state;
66
67    const struct util_format_description *format_desc;
68
69    /** regular scalar float type */
70    struct lp_type float_type;
71    struct lp_build_context float_bld;
72
73    /** regular scalar float type */
74    struct lp_type int_type;
75    struct lp_build_context int_bld;
76
77    /** Incoming coordinates type and build context */
78    struct lp_type coord_type;
79    struct lp_build_context coord_bld;
80
81    /** Unsigned integer coordinates */
82    struct lp_type uint_coord_type;
83    struct lp_build_context uint_coord_bld;
84
85    /** Signed integer coordinates */
86    struct lp_type int_coord_type;
87    struct lp_build_context int_coord_bld;
88
89    /** Output texels type and build context */
90    struct lp_type texel_type;
91    struct lp_build_context texel_bld;
92 };
93
94
95 /**
96  * Does the given texture wrap mode allow sampling the texture border color?
97  * XXX maybe move this into gallium util code.
98  */
99 static boolean
100 wrap_mode_uses_border_color(unsigned mode)
101 {
102    switch (mode) {
103    case PIPE_TEX_WRAP_REPEAT:
104    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
105    case PIPE_TEX_WRAP_MIRROR_REPEAT:
106    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
107       return FALSE;
108    case PIPE_TEX_WRAP_CLAMP:
109    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
110    case PIPE_TEX_WRAP_MIRROR_CLAMP:
111    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
112       return TRUE;
113    default:
114       assert(0 && "unexpected wrap mode");
115       return FALSE;
116    }
117 }
118
119
120 static LLVMValueRef
121 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
122                           LLVMValueRef data_array, LLVMValueRef level)
123 {
124    LLVMValueRef indexes[2], data_ptr;
125    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
126    indexes[1] = level;
127    data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
128    data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
129    return data_ptr;
130 }
131
132
133 static LLVMValueRef
134 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
135                                 LLVMValueRef data_array, int level)
136 {
137    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
138    return lp_build_get_mipmap_level(bld, data_array, lvl);
139 }
140
141
142 /**
143  * Dereference stride_array[mipmap_level] array to get a stride.
144  * Return stride as a vector.
145  */
146 static LLVMValueRef
147 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
148                               LLVMValueRef stride_array, LLVMValueRef level)
149 {
150    LLVMValueRef indexes[2], stride;
151    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
152    indexes[1] = level;
153    stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
154    stride = LLVMBuildLoad(bld->builder, stride, "");
155    stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
156    return stride;
157 }
158
159
160 /** Dereference stride_array[0] array to get a stride (as vector). */
161 static LLVMValueRef
162 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
163                                     LLVMValueRef stride_array, int level)
164 {
165    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
166    return lp_build_get_level_stride_vec(bld, stride_array, lvl);
167 }
168
169
170 static int
171 texture_dims(enum pipe_texture_target tex)
172 {
173    switch (tex) {
174    case PIPE_TEXTURE_1D:
175       return 1;
176    case PIPE_TEXTURE_2D:
177    case PIPE_TEXTURE_CUBE:
178       return 2;
179    case PIPE_TEXTURE_3D:
180       return 3;
181    default:
182       assert(0 && "bad texture target in texture_dims()");
183       return 2;
184    }
185 }
186
187
188
189 /**
190  * Generate code to fetch a texel from a texture at int coords (x, y, z).
191  * The computation depends on whether the texture is 1D, 2D or 3D.
192  * The result, texel, will be:
193  *   texel[0] = red values
194  *   texel[1] = green values
195  *   texel[2] = blue values
196  *   texel[3] = alpha values
197  */
198 static void
199 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
200                           LLVMValueRef width,
201                           LLVMValueRef height,
202                           LLVMValueRef depth,
203                           LLVMValueRef x,
204                           LLVMValueRef y,
205                           LLVMValueRef z,
206                           LLVMValueRef y_stride,
207                           LLVMValueRef z_stride,
208                           LLVMValueRef data_ptr,
209                           LLVMValueRef *texel)
210 {
211    const int dims = texture_dims(bld->static_state->target);
212    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
213    LLVMValueRef offset;
214    LLVMValueRef i, j;
215    LLVMValueRef use_border = NULL;
216
217    /* use_border = x < 0 || x >= width || y < 0 || y >= height */
218    if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
219       LLVMValueRef b1, b2;
220       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
221       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
222       use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
223    }
224
225    if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
226       LLVMValueRef b1, b2;
227       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
228       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
229       if (use_border) {
230          use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
231          use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
232       }
233       else {
234          use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
235       }
236    }
237
238    if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
239       LLVMValueRef b1, b2;
240       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
241       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
242       if (use_border) {
243          use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
244          use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
245       }
246       else {
247          use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
248       }
249    }
250
251    /*
252     * Describe the coordinates in terms of pixel blocks.
253     *
254     * TODO: pixel blocks are power of two. LLVM should convert rem/div to
255     * bit arithmetic. Verify this.
256     */
257
258    if (bld->format_desc->block.width == 1) {
259       i = bld->uint_coord_bld.zero;
260    }
261    else {
262       LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width);
263       i = LLVMBuildURem(bld->builder, x, block_width, "");
264       x = LLVMBuildUDiv(bld->builder, x, block_width, "");
265    }
266
267    if (bld->format_desc->block.height == 1) {
268       j = bld->uint_coord_bld.zero;
269    }
270    else {
271       LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height);
272       j = LLVMBuildURem(bld->builder, y, block_height, "");
273       y = LLVMBuildUDiv(bld->builder, y, block_height, "");
274    }
275
276    /* convert x,y,z coords to linear offset from start of texture, in bytes */
277    offset = lp_build_sample_offset(&bld->uint_coord_bld,
278                                    bld->format_desc,
279                                    x, y, z, y_stride, z_stride);
280
281    lp_build_fetch_rgba_soa(bld->builder,
282                            bld->format_desc,
283                            bld->texel_type,
284                            data_ptr, offset,
285                            i, j,
286                            texel);
287
288    /*
289     * Note: if we find an app which frequently samples the texture border
290     * we might want to implement a true conditional here to avoid sampling
291     * the texture whenever possible (since that's quite a bit of code).
292     * Ex:
293     *   if (use_border) {
294     *      texel = border_color;
295     *   }
296     *   else {
297     *      texel = sample_texture(coord);
298     *   }
299     * As it is now, we always sample the texture, then selectively replace
300     * the texel color results with the border color.
301     */
302
303    if (use_border) {
304       /* select texel color or border color depending on use_border */
305       int chan;
306       for (chan = 0; chan < 4; chan++) {
307          LLVMValueRef border_chan =
308             lp_build_const_vec(bld->texel_type,
309                                   bld->static_state->border_color[chan]);
310          texel[chan] = lp_build_select(&bld->texel_bld, use_border,
311                                        border_chan, texel[chan]);
312       }
313    }
314 }
315
316
317 static LLVMValueRef
318 lp_build_sample_packed(struct lp_build_sample_context *bld,
319                        LLVMValueRef x,
320                        LLVMValueRef y,
321                        LLVMValueRef y_stride,
322                        LLVMValueRef data_array)
323 {
324    LLVMValueRef offset;
325    LLVMValueRef data_ptr;
326
327    offset = lp_build_sample_offset(&bld->uint_coord_bld,
328                                    bld->format_desc,
329                                    x, y, NULL, y_stride, NULL);
330
331    assert(bld->format_desc->block.width == 1);
332    assert(bld->format_desc->block.height == 1);
333    assert(bld->format_desc->block.bits <= bld->texel_type.width);
334
335    /* get pointer to mipmap level 0 data */
336    data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
337
338    return lp_build_gather(bld->builder,
339                           bld->texel_type.length,
340                           bld->format_desc->block.bits,
341                           bld->texel_type.width,
342                           data_ptr, offset);
343 }
344
345
346 /**
347  * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
348  */
349 static LLVMValueRef
350 lp_build_coord_mirror(struct lp_build_sample_context *bld,
351                       LLVMValueRef coord)
352 {
353    struct lp_build_context *coord_bld = &bld->coord_bld;
354    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
355    LLVMValueRef fract, flr, isOdd;
356
357    /* fract = coord - floor(coord) */
358    fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
359
360    /* flr = ifloor(coord); */
361    flr = lp_build_ifloor(coord_bld, coord);
362
363    /* isOdd = flr & 1 */
364    isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
365
366    /* make coord positive or negative depending on isOdd */
367    coord = lp_build_set_sign(coord_bld, fract, isOdd);
368
369    /* convert isOdd to float */
370    isOdd = lp_build_int_to_float(coord_bld, isOdd);
371
372    /* add isOdd to coord */
373    coord = lp_build_add(coord_bld, coord, isOdd);
374
375    return coord;
376 }
377
378
379 /**
380  * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
381  * Return whether the given mode is supported by that function.
382  */
383 static boolean
384 is_simple_wrap_mode(unsigned mode)
385 {
386    switch (mode) {
387    case PIPE_TEX_WRAP_REPEAT:
388    case PIPE_TEX_WRAP_CLAMP:
389    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
390       return TRUE;
391    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
392    default:
393       return FALSE;
394    }
395 }
396
397
398 /**
399  * Build LLVM code for texture wrap mode, for scaled integer texcoords.
400  * \param coord  the incoming texcoord (s,t,r or q) scaled to the texture size
401  * \param length  the texture size along one dimension
402  * \param is_pot  if TRUE, length is a power of two
403  * \param wrap_mode  one of PIPE_TEX_WRAP_x
404  */
405 static LLVMValueRef
406 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
407                          LLVMValueRef coord,
408                          LLVMValueRef length,
409                          boolean is_pot,
410                          unsigned wrap_mode)
411 {
412    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
413    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
414    LLVMValueRef length_minus_one;
415
416    length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
417
418    switch(wrap_mode) {
419    case PIPE_TEX_WRAP_REPEAT:
420       if(is_pot)
421          coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
422       else
423          /* Signed remainder won't give the right results for negative
424           * dividends but unsigned remainder does.*/
425          coord = LLVMBuildURem(bld->builder, coord, length, "");
426       break;
427
428    case PIPE_TEX_WRAP_CLAMP:
429    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
430    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
431       coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
432       coord = lp_build_min(int_coord_bld, coord, length_minus_one);
433       break;
434
435    case PIPE_TEX_WRAP_MIRROR_REPEAT:
436    case PIPE_TEX_WRAP_MIRROR_CLAMP:
437    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
438    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
439       /* FIXME */
440       _debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
441                     util_dump_tex_wrap(wrap_mode, TRUE));
442       coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
443       coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
444       break;
445
446    default:
447       assert(0);
448    }
449
450    return coord;
451 }
452
453
454 /**
455  * Build LLVM code for texture wrap mode for linear filtering.
456  * \param x0_out  returns first integer texcoord
457  * \param x1_out  returns second integer texcoord
458  * \param weight_out  returns linear interpolation weight
459  */
460 static void
461 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
462                             LLVMValueRef coord,
463                             LLVMValueRef length,
464                             boolean is_pot,
465                             unsigned wrap_mode,
466                             LLVMValueRef *x0_out,
467                             LLVMValueRef *x1_out,
468                             LLVMValueRef *weight_out)
469 {
470    struct lp_build_context *coord_bld = &bld->coord_bld;
471    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
472    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
473    LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
474    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
475    LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
476    LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
477    LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
478    LLVMValueRef coord0, coord1, weight;
479
480    switch(wrap_mode) {
481    case PIPE_TEX_WRAP_REPEAT:
482       /* mul by size and subtract 0.5 */
483       coord = lp_build_mul(coord_bld, coord, length_f);
484       coord = lp_build_sub(coord_bld, coord, half);
485       /* convert to int */
486       coord0 = lp_build_ifloor(coord_bld, coord);
487       coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
488       /* compute lerp weight */
489       weight = lp_build_fract(coord_bld, coord);
490       /* repeat wrap */
491       if (is_pot) {
492          coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
493          coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
494       }
495       else {
496          /* Signed remainder won't give the right results for negative
497           * dividends but unsigned remainder does.*/
498          coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
499          coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
500       }
501       break;
502
503    case PIPE_TEX_WRAP_CLAMP:
504       if (bld->static_state->normalized_coords) {
505          coord = lp_build_mul(coord_bld, coord, length_f);
506       }
507       weight = lp_build_fract(coord_bld, coord);
508       coord0 = lp_build_clamp(coord_bld, coord, coord_bld->zero,
509                               length_f_minus_one);
510       coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
511       coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
512                               length_f_minus_one);
513       coord0 = lp_build_ifloor(coord_bld, coord0);
514       coord1 = lp_build_ifloor(coord_bld, coord1);
515       break;
516
517    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
518       if (bld->static_state->normalized_coords) {
519          /* clamp to [0,1] */
520          coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
521          /* mul by tex size and subtract 0.5 */
522          coord = lp_build_mul(coord_bld, coord, length_f);
523          coord = lp_build_sub(coord_bld, coord, half);
524       }
525       else {
526          LLVMValueRef min, max;
527          /* clamp to [0.5, length - 0.5] */
528          min = lp_build_const_vec(coord_bld->type, 0.5F);
529          max = lp_build_sub(coord_bld, length_f, min);
530          coord = lp_build_clamp(coord_bld, coord, min, max);
531       }
532       /* compute lerp weight */
533       weight = lp_build_fract(coord_bld, coord);
534       /* coord0 = floor(coord); */
535       coord0 = lp_build_ifloor(coord_bld, coord);
536       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
537       /* coord0 = max(coord0, 0) */
538       coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
539       /* coord1 = min(coord1, length-1) */
540       coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
541       break;
542
543    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
544       {
545          LLVMValueRef min, max;
546          if (bld->static_state->normalized_coords) {
547             /* min = -1.0 / (2 * length) = -0.5 / length */
548             min = lp_build_mul(coord_bld,
549                                lp_build_const_vec(coord_bld->type, -0.5F),
550                                lp_build_rcp(coord_bld, length_f));
551             /* max = 1.0 - min */
552             max = lp_build_sub(coord_bld, coord_bld->one, min);
553             /* coord = clamp(coord, min, max) */
554             coord = lp_build_clamp(coord_bld, coord, min, max);
555             /* scale coord to length (and sub 0.5?) */
556             coord = lp_build_mul(coord_bld, coord, length_f);
557             coord = lp_build_sub(coord_bld, coord, half);
558          }
559          else {
560             /* clamp to [-0.5, length + 0.5] */
561             min = lp_build_const_vec(coord_bld->type, -0.5F);
562             max = lp_build_sub(coord_bld, length_f, min);
563             coord = lp_build_clamp(coord_bld, coord, min, max);
564             coord = lp_build_sub(coord_bld, coord, half);
565          }
566          /* compute lerp weight */
567          weight = lp_build_fract(coord_bld, coord);
568          /* convert to int */
569          coord0 = lp_build_ifloor(coord_bld, coord);
570          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
571       }
572       break;
573
574    case PIPE_TEX_WRAP_MIRROR_REPEAT:
575       /* compute mirror function */
576       coord = lp_build_coord_mirror(bld, coord);
577
578       /* scale coord to length */
579       coord = lp_build_mul(coord_bld, coord, length_f);
580       coord = lp_build_sub(coord_bld, coord, half);
581
582       /* compute lerp weight */
583       weight = lp_build_fract(coord_bld, coord);
584
585       /* convert to int coords */
586       coord0 = lp_build_ifloor(coord_bld, coord);
587       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
588
589       /* coord0 = max(coord0, 0) */
590       coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
591       /* coord1 = min(coord1, length-1) */
592       coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
593       break;
594
595    case PIPE_TEX_WRAP_MIRROR_CLAMP:
596       {
597          LLVMValueRef min, max;
598          /* min = 1.0 / (2 * length) */
599          min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
600          /* max = 1.0 - min */
601          max = lp_build_sub(coord_bld, coord_bld->one, min);
602
603          coord = lp_build_abs(coord_bld, coord);
604          coord = lp_build_clamp(coord_bld, coord, min, max);
605          coord = lp_build_mul(coord_bld, coord, length_f);
606          if(0)coord = lp_build_sub(coord_bld, coord, half);
607          weight = lp_build_fract(coord_bld, coord);
608          coord0 = lp_build_ifloor(coord_bld, coord);
609          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
610       }
611       break;
612
613    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
614       {
615          LLVMValueRef min, max;
616          /* min = 1.0 / (2 * length) */
617          min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
618          /* max = 1.0 - min */
619          max = lp_build_sub(coord_bld, coord_bld->one, min);
620
621          coord = lp_build_abs(coord_bld, coord);
622          coord = lp_build_clamp(coord_bld, coord, min, max);
623          coord = lp_build_mul(coord_bld, coord, length_f);
624          coord = lp_build_sub(coord_bld, coord, half);
625          weight = lp_build_fract(coord_bld, coord);
626          coord0 = lp_build_ifloor(coord_bld, coord);
627          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
628       }
629       break;
630
631    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
632       {
633          LLVMValueRef min, max;
634          /* min = -1.0 / (2 * length) = -0.5 / length */
635          min = lp_build_mul(coord_bld,
636                             lp_build_const_vec(coord_bld->type, -0.5F),
637                             lp_build_rcp(coord_bld, length_f));
638          /* max = 1.0 - min */
639          max = lp_build_sub(coord_bld, coord_bld->one, min);
640
641          coord = lp_build_abs(coord_bld, coord);
642          coord = lp_build_clamp(coord_bld, coord, min, max);
643          coord = lp_build_mul(coord_bld, coord, length_f);
644          coord = lp_build_sub(coord_bld, coord, half);
645          weight = lp_build_fract(coord_bld, coord);
646          coord0 = lp_build_ifloor(coord_bld, coord);
647          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
648       }
649       break;
650
651    default:
652       assert(0);
653       coord0 = NULL;
654       coord1 = NULL;
655       weight = NULL;
656    }
657
658    *x0_out = coord0;
659    *x1_out = coord1;
660    *weight_out = weight;
661 }
662
663
664 /**
665  * Build LLVM code for texture wrap mode for nearest filtering.
666  * \param coord  the incoming texcoord (nominally in [0,1])
667  * \param length  the texture size along one dimension, as int
668  * \param is_pot  if TRUE, length is a power of two
669  * \param wrap_mode  one of PIPE_TEX_WRAP_x
670  */
671 static LLVMValueRef
672 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
673                              LLVMValueRef coord,
674                              LLVMValueRef length,
675                              boolean is_pot,
676                              unsigned wrap_mode)
677 {
678    struct lp_build_context *coord_bld = &bld->coord_bld;
679    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
680    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
681    LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
682    LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
683    LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
684    LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
685    LLVMValueRef icoord;
686    
687    switch(wrap_mode) {
688    case PIPE_TEX_WRAP_REPEAT:
689       coord = lp_build_mul(coord_bld, coord, length_f);
690       icoord = lp_build_ifloor(coord_bld, coord);
691       if (is_pot)
692          icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
693       else
694          /* Signed remainder won't give the right results for negative
695           * dividends but unsigned remainder does.*/
696          icoord = LLVMBuildURem(bld->builder, icoord, length, "");
697       break;
698
699    case PIPE_TEX_WRAP_CLAMP:
700       /* mul by size */
701       if (bld->static_state->normalized_coords) {
702          coord = lp_build_mul(coord_bld, coord, length_f);
703       }
704       /* floor */
705       icoord = lp_build_ifloor(coord_bld, coord);
706       /* clamp to [0, size-1].  Note: int coord builder type */
707       icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
708                               length_minus_one);
709       break;
710
711    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
712       {
713          LLVMValueRef min, max;
714          if (bld->static_state->normalized_coords) {
715             /* min = 1.0 / (2 * length) */
716             min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
717             /* max = length - min */
718             max = lp_build_sub(coord_bld, length_f, min);
719             /* scale coord to length */
720             coord = lp_build_mul(coord_bld, coord, length_f);
721          }
722          else {
723             /* clamp to [0.5, length - 0.5] */
724             min = lp_build_const_vec(coord_bld->type, 0.5F);
725             max = lp_build_sub(coord_bld, length_f, min);
726          }
727          /* coord = clamp(coord, min, max) */
728          coord = lp_build_clamp(coord_bld, coord, min, max);
729          icoord = lp_build_ifloor(coord_bld, coord);
730       }
731       break;
732
733    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
734       /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
735       {
736          LLVMValueRef min, max;
737          if (bld->static_state->normalized_coords) {
738             /* min = -1.0 / (2 * length) = -0.5 / length */
739             min = lp_build_mul(coord_bld,
740                                lp_build_const_vec(coord_bld->type, -0.5F),
741                                lp_build_rcp(coord_bld, length_f));
742             /* max = length - min */
743             max = lp_build_sub(coord_bld, length_f, min);
744             /* scale coord to length */
745             coord = lp_build_mul(coord_bld, coord, length_f);
746          }
747          else {
748             /* clamp to [-0.5, length + 0.5] */
749             min = lp_build_const_vec(coord_bld->type, -0.5F);
750             max = lp_build_sub(coord_bld, length_f, min);
751          }
752          /* coord = clamp(coord, min, max) */
753          coord = lp_build_clamp(coord_bld, coord, min, max);
754          icoord = lp_build_ifloor(coord_bld, coord);
755       }
756       break;
757
758    case PIPE_TEX_WRAP_MIRROR_REPEAT:
759       {
760          LLVMValueRef min, max;
761          /* min = 1.0 / (2 * length) */
762          min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
763          /* max = length - min */
764          max = lp_build_sub(coord_bld, length_f, min);
765
766          /* compute mirror function */
767          coord = lp_build_coord_mirror(bld, coord);
768
769          /* scale coord to length */
770          coord = lp_build_mul(coord_bld, coord, length_f);
771
772          /* coord = clamp(coord, min, max) */
773          coord = lp_build_clamp(coord_bld, coord, min, max);
774          icoord = lp_build_ifloor(coord_bld, coord);
775       }
776       break;
777
778    case PIPE_TEX_WRAP_MIRROR_CLAMP:
779       coord = lp_build_abs(coord_bld, coord);
780       coord = lp_build_mul(coord_bld, coord, length_f);
781       coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
782       icoord = lp_build_ifloor(coord_bld, coord);
783       break;
784
785    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
786       {
787          LLVMValueRef min, max;
788          /* min = 1.0 / (2 * length) */
789          min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
790          /* max = length - min */
791          max = lp_build_sub(coord_bld, length_f, min);
792
793          coord = lp_build_abs(coord_bld, coord);
794          coord = lp_build_mul(coord_bld, coord, length_f);
795          coord = lp_build_clamp(coord_bld, coord, min, max);
796          icoord = lp_build_ifloor(coord_bld, coord);
797       }
798       break;
799
800    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
801       {
802          LLVMValueRef min, max;
803          /* min = 1.0 / (2 * length) */
804          min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
805          min = lp_build_negate(coord_bld, min);
806          /* max = length - min */
807          max = lp_build_sub(coord_bld, length_f, min);
808
809          coord = lp_build_abs(coord_bld, coord);
810          coord = lp_build_mul(coord_bld, coord, length_f);
811          coord = lp_build_clamp(coord_bld, coord, min, max);
812          icoord = lp_build_ifloor(coord_bld, coord);
813       }
814       break;
815
816    default:
817       assert(0);
818       icoord = NULL;
819    }
820
821    return icoord;
822 }
823
824
825 /**
826  * Codegen equivalent for u_minify().
827  * Return max(1, base_size >> level);
828  */
829 static LLVMValueRef
830 lp_build_minify(struct lp_build_sample_context *bld,
831                 LLVMValueRef base_size,
832                 LLVMValueRef level)
833 {
834    LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
835    size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
836    return size;
837 }
838
839
840 /**
841  * Generate code to compute texture level of detail (lambda).
842  * \param s  vector of texcoord s values
843  * \param t  vector of texcoord t values
844  * \param r  vector of texcoord r values
845  * \param width  scalar int texture width
846  * \param height  scalar int texture height
847  * \param depth  scalar int texture depth
848  */
849 static LLVMValueRef
850 lp_build_lod_selector(struct lp_build_sample_context *bld,
851                       LLVMValueRef s,
852                       LLVMValueRef t,
853                       LLVMValueRef r,
854                       LLVMValueRef width,
855                       LLVMValueRef height,
856                       LLVMValueRef depth)
857
858 {
859    if (bld->static_state->min_lod == bld->static_state->max_lod) {
860       /* User is forcing sampling from a particular mipmap level.
861        * This is hit during mipmap generation.
862        */
863       return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
864    }
865    else {
866       const int dims = texture_dims(bld->static_state->target);
867       struct lp_build_context *float_bld = &bld->float_bld;
868       LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(),
869                                             bld->static_state->lod_bias);
870       LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
871                                            bld->static_state->min_lod);
872       LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
873                                            bld->static_state->max_lod);
874
875       LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
876       LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
877       LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
878
879       LLVMValueRef s0, s1, s2;
880       LLVMValueRef t0, t1, t2;
881       LLVMValueRef r0, r1, r2;
882       LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
883       LLVMValueRef rho, lod;
884
885       /*
886        * dsdx = abs(s[1] - s[0]);
887        * dsdy = abs(s[2] - s[0]);
888        * dtdx = abs(t[1] - t[0]);
889        * dtdy = abs(t[2] - t[0]);
890        * drdx = abs(r[1] - r[0]);
891        * drdy = abs(r[2] - r[0]);
892        * XXX we're assuming a four-element quad in 2x2 layout here.
893        */
894       s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0");
895       s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1");
896       s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2");
897       dsdx = LLVMBuildSub(bld->builder, s1, s0, "");
898       dsdx = lp_build_abs(float_bld, dsdx);
899       dsdy = LLVMBuildSub(bld->builder, s2, s0, "");
900       dsdy = lp_build_abs(float_bld, dsdy);
901       if (dims > 1) {
902          t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0");
903          t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1");
904          t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2");
905          dtdx = LLVMBuildSub(bld->builder, t1, t0, "");
906          dtdx = lp_build_abs(float_bld, dtdx);
907          dtdy = LLVMBuildSub(bld->builder, t2, t0, "");
908          dtdy = lp_build_abs(float_bld, dtdy);
909          if (dims > 2) {
910             r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0");
911             r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1");
912             r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2");
913             drdx = LLVMBuildSub(bld->builder, r1, r0, "");
914             drdx = lp_build_abs(float_bld, drdx);
915             drdy = LLVMBuildSub(bld->builder, r2, r0, "");
916             drdy = lp_build_abs(float_bld, drdy);
917          }
918       }
919
920       /* Compute rho = max of all partial derivatives scaled by texture size.
921        * XXX this could be vectorized somewhat
922        */
923       rho = LLVMBuildMul(bld->builder,
924                          lp_build_max(float_bld, dsdx, dsdy),
925                          lp_build_int_to_float(float_bld, width), "");
926       if (dims > 1) {
927          LLVMValueRef max;
928          max = LLVMBuildMul(bld->builder,
929                             lp_build_max(float_bld, dtdx, dtdy),
930                             lp_build_int_to_float(float_bld, height), "");
931          rho = lp_build_max(float_bld, rho, max);
932          if (dims > 2) {
933             max = LLVMBuildMul(bld->builder,
934                                lp_build_max(float_bld, drdx, drdy),
935                                lp_build_int_to_float(float_bld, depth), "");
936             rho = lp_build_max(float_bld, rho, max);
937          }
938       }
939
940       /* compute lod = log2(rho) */
941       lod = lp_build_log2(float_bld, rho);
942
943       /* add lod bias */
944       lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "LOD bias");
945
946       /* clamp lod */
947       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
948
949       return lod;
950    }
951 }
952
953
954 /**
955  * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
956  * mipmap level index.
957  * Note: this is all scalar code.
958  * \param lod  scalar float texture level of detail
959  * \param level_out  returns integer 
960  */
961 static void
962 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
963                            unsigned unit,
964                            LLVMValueRef lod,
965                            LLVMValueRef *level_out)
966 {
967    struct lp_build_context *float_bld = &bld->float_bld;
968    struct lp_build_context *int_bld = &bld->int_bld;
969    LLVMValueRef last_level, level;
970
971    LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
972
973    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
974                                                bld->builder, unit);
975
976    /* convert float lod to integer */
977    level = lp_build_iround(float_bld, lod);
978
979    /* clamp level to legal range of levels */
980    *level_out = lp_build_clamp(int_bld, level, zero, last_level);
981 }
982
983
984 /**
985  * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
986  * two (adjacent) mipmap level indexes.  Later, we'll sample from those
987  * two mipmap levels and interpolate between them.
988  */
989 static void
990 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
991                            unsigned unit,
992                            LLVMValueRef lod,
993                            LLVMValueRef *level0_out,
994                            LLVMValueRef *level1_out,
995                            LLVMValueRef *weight_out)
996 {
997    struct lp_build_context *float_bld = &bld->float_bld;
998    struct lp_build_context *int_bld = &bld->int_bld;
999    LLVMValueRef last_level, level;
1000
1001    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
1002                                                bld->builder, unit);
1003
1004    /* convert float lod to integer */
1005    level = lp_build_ifloor(float_bld, lod);
1006
1007    /* compute level 0 and clamp to legal range of levels */
1008    *level0_out = lp_build_clamp(int_bld, level,
1009                                 int_bld->zero,
1010                                 last_level);
1011    /* compute level 1 and clamp to legal range of levels */
1012    *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
1013    *level1_out = lp_build_min(int_bld, *level1_out, last_level);
1014
1015    *weight_out = lp_build_fract(float_bld, lod);
1016 }
1017
1018
1019 /**
1020  * Generate code to sample a mipmap level with nearest filtering.
1021  * If sampling a cube texture, r = cube face in [0,5].
1022  */
1023 static void
1024 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1025                               LLVMValueRef width_vec,
1026                               LLVMValueRef height_vec,
1027                               LLVMValueRef depth_vec,
1028                               LLVMValueRef row_stride_vec,
1029                               LLVMValueRef img_stride_vec,
1030                               LLVMValueRef data_ptr,
1031                               LLVMValueRef s,
1032                               LLVMValueRef t,
1033                               LLVMValueRef r,
1034                               LLVMValueRef colors_out[4])
1035 {
1036    const int dims = texture_dims(bld->static_state->target);
1037    LLVMValueRef x, y, z;
1038
1039    /*
1040     * Compute integer texcoords.
1041     */
1042    x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1043                                     bld->static_state->pot_width,
1044                                     bld->static_state->wrap_s);
1045    lp_build_name(x, "tex.x.wrapped");
1046
1047    if (dims >= 2) {
1048       y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1049                                        bld->static_state->pot_height,
1050                                        bld->static_state->wrap_t);
1051       lp_build_name(y, "tex.y.wrapped");
1052
1053       if (dims == 3) {
1054          z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1055                                           bld->static_state->pot_height,
1056                                           bld->static_state->wrap_r);
1057          lp_build_name(z, "tex.z.wrapped");
1058       }
1059       else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1060          z = r;
1061       }
1062       else {
1063          z = NULL;
1064       }
1065    }
1066    else {
1067       y = z = NULL;
1068    }
1069
1070    /*
1071     * Get texture colors.
1072     */
1073    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1074                              x, y, z,
1075                              row_stride_vec, img_stride_vec,
1076                              data_ptr, colors_out);
1077 }
1078
1079
1080 /**
1081  * Generate code to sample a mipmap level with linear filtering.
1082  * If sampling a cube texture, r = cube face in [0,5].
1083  */
1084 static void
1085 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1086                              LLVMValueRef width_vec,
1087                              LLVMValueRef height_vec,
1088                              LLVMValueRef depth_vec,
1089                              LLVMValueRef row_stride_vec,
1090                              LLVMValueRef img_stride_vec,
1091                              LLVMValueRef data_ptr,
1092                              LLVMValueRef s,
1093                              LLVMValueRef t,
1094                              LLVMValueRef r,
1095                              LLVMValueRef colors_out[4])
1096 {
1097    const int dims = texture_dims(bld->static_state->target);
1098    LLVMValueRef x0, y0, z0, x1, y1, z1;
1099    LLVMValueRef s_fpart, t_fpart, r_fpart;
1100    LLVMValueRef neighbors[2][2][4];
1101    int chan;
1102
1103    /*
1104     * Compute integer texcoords.
1105     */
1106    lp_build_sample_wrap_linear(bld, s, width_vec,
1107                                bld->static_state->pot_width,
1108                                bld->static_state->wrap_s,
1109                                &x0, &x1, &s_fpart);
1110    lp_build_name(x0, "tex.x0.wrapped");
1111    lp_build_name(x1, "tex.x1.wrapped");
1112
1113    if (dims >= 2) {
1114       lp_build_sample_wrap_linear(bld, t, height_vec,
1115                                   bld->static_state->pot_height,
1116                                   bld->static_state->wrap_t,
1117                                   &y0, &y1, &t_fpart);
1118       lp_build_name(y0, "tex.y0.wrapped");
1119       lp_build_name(y1, "tex.y1.wrapped");
1120
1121       if (dims == 3) {
1122          lp_build_sample_wrap_linear(bld, r, depth_vec,
1123                                      bld->static_state->pot_depth,
1124                                      bld->static_state->wrap_r,
1125                                      &z0, &z1, &r_fpart);
1126          lp_build_name(z0, "tex.z0.wrapped");
1127          lp_build_name(z1, "tex.z1.wrapped");
1128       }
1129       else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1130          z0 = z1 = r;  /* cube face */
1131          r_fpart = NULL;
1132       }
1133       else {
1134          z0 = z1 = NULL;
1135          r_fpart = NULL;
1136       }
1137    }
1138    else {
1139       y0 = y1 = t_fpart = NULL;
1140       z0 = z1 = r_fpart = NULL;
1141    }
1142
1143    /*
1144     * Get texture colors.
1145     */
1146    /* get x0/x1 texels */
1147    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1148                              x0, y0, z0,
1149                              row_stride_vec, img_stride_vec,
1150                              data_ptr, neighbors[0][0]);
1151    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1152                              x1, y0, z0,
1153                              row_stride_vec, img_stride_vec,
1154                              data_ptr, neighbors[0][1]);
1155
1156    if (dims == 1) {
1157       /* Interpolate two samples from 1D image to produce one color */
1158       for (chan = 0; chan < 4; chan++) {
1159          colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1160                                           neighbors[0][0][chan],
1161                                           neighbors[0][1][chan]);
1162       }
1163    }
1164    else {
1165       /* 2D/3D texture */
1166       LLVMValueRef colors0[4];
1167
1168       /* get x0/x1 texels at y1 */
1169       lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1170                                 x0, y1, z0,
1171                                 row_stride_vec, img_stride_vec,
1172                                 data_ptr, neighbors[1][0]);
1173       lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1174                                 x1, y1, z0,
1175                                 row_stride_vec, img_stride_vec,
1176                                 data_ptr, neighbors[1][1]);
1177
1178       /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1179       for (chan = 0; chan < 4; chan++) {
1180          colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1181                                           s_fpart, t_fpart,
1182                                           neighbors[0][0][chan],
1183                                           neighbors[0][1][chan],
1184                                           neighbors[1][0][chan],
1185                                           neighbors[1][1][chan]);
1186       }
1187
1188       if (dims == 3) {
1189          LLVMValueRef neighbors1[2][2][4];
1190          LLVMValueRef colors1[4];
1191
1192          /* get x0/x1/y0/y1 texels at z1 */
1193          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1194                                    x0, y0, z1,
1195                                    row_stride_vec, img_stride_vec,
1196                                    data_ptr, neighbors1[0][0]);
1197          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1198                                    x1, y0, z1,
1199                                    row_stride_vec, img_stride_vec,
1200                                    data_ptr, neighbors1[0][1]);
1201          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1202                                    x0, y1, z1,
1203                                    row_stride_vec, img_stride_vec,
1204                                    data_ptr, neighbors1[1][0]);
1205          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1206                                    x1, y1, z1,
1207                                    row_stride_vec, img_stride_vec,
1208                                    data_ptr, neighbors1[1][1]);
1209
1210          /* Bilinear interpolate the four samples from the second Z slice */
1211          for (chan = 0; chan < 4; chan++) {
1212             colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1213                                              s_fpart, t_fpart,
1214                                              neighbors1[0][0][chan],
1215                                              neighbors1[0][1][chan],
1216                                              neighbors1[1][0][chan],
1217                                              neighbors1[1][1][chan]);
1218          }
1219
1220          /* Linearly interpolate the two samples from the two 3D slices */
1221          for (chan = 0; chan < 4; chan++) {
1222             colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1223                                              r_fpart,
1224                                              colors0[chan], colors1[chan]);
1225          }
1226       }
1227       else {
1228          /* 2D tex */
1229          for (chan = 0; chan < 4; chan++) {
1230             colors_out[chan] = colors0[chan];
1231          }
1232       }
1233    }
1234 }
1235
1236
1237 /** Helper used by lp_build_cube_lookup() */
1238 static LLVMValueRef
1239 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1240 {
1241    /* ima = -0.5 / abs(coord); */
1242    LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
1243    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1244    LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
1245                                    lp_build_rcp(coord_bld, absCoord));
1246    return ima;
1247 }
1248
1249
1250 /**
1251  * Helper used by lp_build_cube_lookup()
1252  * \param sign  scalar +1 or -1
1253  * \param coord  float vector
1254  * \param ima  float vector
1255  */
1256 static LLVMValueRef
1257 lp_build_cube_coord(struct lp_build_context *coord_bld,
1258                     LLVMValueRef sign, int negate_coord,
1259                     LLVMValueRef coord, LLVMValueRef ima)
1260 {
1261    /* return negate(coord) * ima * sign + 0.5; */
1262    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
1263    LLVMValueRef res;
1264
1265    assert(negate_coord == +1 || negate_coord == -1);
1266
1267    if (negate_coord == -1) {
1268       coord = lp_build_negate(coord_bld, coord);
1269    }
1270
1271    res = lp_build_mul(coord_bld, coord, ima);
1272    if (sign) {
1273       sign = lp_build_broadcast_scalar(coord_bld, sign);
1274       res = lp_build_mul(coord_bld, res, sign);
1275    }
1276    res = lp_build_add(coord_bld, res, half);
1277
1278    return res;
1279 }
1280
1281
1282 /** Helper used by lp_build_cube_lookup()
1283  * Return (major_coord >= 0) ? pos_face : neg_face;
1284  */
1285 static LLVMValueRef
1286 lp_build_cube_face(struct lp_build_sample_context *bld,
1287                    LLVMValueRef major_coord,
1288                    unsigned pos_face, unsigned neg_face)
1289 {
1290    LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1291                                     major_coord,
1292                                     bld->float_bld.zero, "");
1293    LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1294    LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1295    LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1296    return res;
1297 }
1298
1299
1300
1301 /**
1302  * Generate code to do cube face selection and per-face texcoords.
1303  */
1304 static void
1305 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1306                      LLVMValueRef s,
1307                      LLVMValueRef t,
1308                      LLVMValueRef r,
1309                      LLVMValueRef *face,
1310                      LLVMValueRef *face_s,
1311                      LLVMValueRef *face_t)
1312 {
1313    struct lp_build_context *float_bld = &bld->float_bld;
1314    struct lp_build_context *coord_bld = &bld->coord_bld;
1315    LLVMValueRef rx, ry, rz;
1316    LLVMValueRef arx, ary, arz;
1317    LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1318    LLVMValueRef arx_ge_ary, arx_ge_arz;
1319    LLVMValueRef ary_ge_arx, ary_ge_arz;
1320    LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1321    LLVMValueRef rx_pos, ry_pos, rz_pos;
1322
1323    assert(bld->coord_bld.type.length == 4);
1324
1325    /*
1326     * Use the average of the four pixel's texcoords to choose the face.
1327     */
1328    rx = lp_build_mul(float_bld, c25,
1329                      lp_build_sum_vector(&bld->coord_bld, s));
1330    ry = lp_build_mul(float_bld, c25,
1331                      lp_build_sum_vector(&bld->coord_bld, t));
1332    rz = lp_build_mul(float_bld, c25,
1333                      lp_build_sum_vector(&bld->coord_bld, r));
1334
1335    arx = lp_build_abs(float_bld, rx);
1336    ary = lp_build_abs(float_bld, ry);
1337    arz = lp_build_abs(float_bld, rz);
1338
1339    /*
1340     * Compare sign/magnitude of rx,ry,rz to determine face
1341     */
1342    arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1343    arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1344    ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1345    ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1346
1347    arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1348    ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1349
1350    rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1351    ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1352    rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1353
1354    {
1355       struct lp_build_flow_context *flow_ctx;
1356       struct lp_build_if_state if_ctx;
1357
1358       flow_ctx = lp_build_flow_create(bld->builder);
1359       lp_build_flow_scope_begin(flow_ctx);
1360
1361       *face_s = bld->coord_bld.undef;
1362       *face_t = bld->coord_bld.undef;
1363       *face = bld->int_bld.undef;
1364
1365       lp_build_name(*face_s, "face_s");
1366       lp_build_name(*face_t, "face_t");
1367       lp_build_name(*face, "face");
1368
1369       lp_build_flow_scope_declare(flow_ctx, face_s);
1370       lp_build_flow_scope_declare(flow_ctx, face_t);
1371       lp_build_flow_scope_declare(flow_ctx, face);
1372
1373       lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1374       {
1375          /* +/- X face */
1376          LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1377          LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1378          *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1379          *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1380          *face = lp_build_cube_face(bld, rx,
1381                                     PIPE_TEX_FACE_POS_X,
1382                                     PIPE_TEX_FACE_NEG_X);
1383       }
1384       lp_build_else(&if_ctx);
1385       {
1386          struct lp_build_flow_context *flow_ctx2;
1387          struct lp_build_if_state if_ctx2;
1388
1389          LLVMValueRef face_s2 = bld->coord_bld.undef;
1390          LLVMValueRef face_t2 = bld->coord_bld.undef;
1391          LLVMValueRef face2 = bld->int_bld.undef;
1392
1393          flow_ctx2 = lp_build_flow_create(bld->builder);
1394          lp_build_flow_scope_begin(flow_ctx2);
1395          lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1396          lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1397          lp_build_flow_scope_declare(flow_ctx2, &face2);
1398
1399          ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1400
1401          lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1402          {
1403             /* +/- Y face */
1404             LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1405             LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1406             face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1407             face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1408             face2 = lp_build_cube_face(bld, ry,
1409                                        PIPE_TEX_FACE_POS_Y,
1410                                        PIPE_TEX_FACE_NEG_Y);
1411          }
1412          lp_build_else(&if_ctx2);
1413          {
1414             /* +/- Z face */
1415             LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1416             LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1417             face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1418             face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1419             face2 = lp_build_cube_face(bld, rz,
1420                                        PIPE_TEX_FACE_POS_Z,
1421                                        PIPE_TEX_FACE_NEG_Z);
1422          }
1423          lp_build_endif(&if_ctx2);
1424          lp_build_flow_scope_end(flow_ctx2);
1425          lp_build_flow_destroy(flow_ctx2);
1426
1427          *face_s = face_s2;
1428          *face_t = face_t2;
1429          *face = face2;
1430       }
1431
1432       lp_build_endif(&if_ctx);
1433       lp_build_flow_scope_end(flow_ctx);
1434       lp_build_flow_destroy(flow_ctx);
1435    }
1436 }
1437
1438
1439
1440 /**
1441  * Sample the texture/mipmap using given image filter and mip filter.
1442  * data0_ptr and data1_ptr point to the two mipmap levels to sample
1443  * from.  width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
1444  * If we're using nearest miplevel sampling the '1' values will be null/unused.
1445  */
1446 static void
1447 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1448                        unsigned img_filter,
1449                        unsigned mip_filter,
1450                        LLVMValueRef s,
1451                        LLVMValueRef t,
1452                        LLVMValueRef r,
1453                        LLVMValueRef lod_fpart,
1454                        LLVMValueRef width0_vec,
1455                        LLVMValueRef width1_vec,
1456                        LLVMValueRef height0_vec,
1457                        LLVMValueRef height1_vec,
1458                        LLVMValueRef depth0_vec,
1459                        LLVMValueRef depth1_vec,
1460                        LLVMValueRef row_stride0_vec,
1461                        LLVMValueRef row_stride1_vec,
1462                        LLVMValueRef img_stride0_vec,
1463                        LLVMValueRef img_stride1_vec,
1464                        LLVMValueRef data_ptr0,
1465                        LLVMValueRef data_ptr1,
1466                        LLVMValueRef *colors_out)
1467 {
1468    LLVMValueRef colors0[4], colors1[4];
1469    int chan;
1470
1471    if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1472       lp_build_sample_image_nearest(bld,
1473                                     width0_vec, height0_vec, depth0_vec,
1474                                     row_stride0_vec, img_stride0_vec,
1475                                     data_ptr0, s, t, r, colors0);
1476
1477       if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1478          /* sample the second mipmap level, and interp */
1479          lp_build_sample_image_nearest(bld,
1480                                        width1_vec, height1_vec, depth1_vec,
1481                                        row_stride1_vec, img_stride1_vec,
1482                                        data_ptr1, s, t, r, colors1);
1483       }
1484    }
1485    else {
1486       assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1487
1488       lp_build_sample_image_linear(bld,
1489                                    width0_vec, height0_vec, depth0_vec,
1490                                    row_stride0_vec, img_stride0_vec,
1491                                    data_ptr0, s, t, r, colors0);
1492
1493       if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1494          /* sample the second mipmap level, and interp */
1495          lp_build_sample_image_linear(bld,
1496                                       width1_vec, height1_vec, depth1_vec,
1497                                       row_stride1_vec, img_stride1_vec,
1498                                       data_ptr1, s, t, r, colors1);
1499       }
1500    }
1501
1502    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1503       /* interpolate samples from the two mipmap levels */
1504       for (chan = 0; chan < 4; chan++) {
1505          colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1506                                           colors0[chan], colors1[chan]);
1507       }
1508    }
1509    else {
1510       /* use first/only level's colors */
1511       for (chan = 0; chan < 4; chan++) {
1512          colors_out[chan] = colors0[chan];
1513       }
1514    }
1515 }
1516
1517
1518
1519 /**
1520  * General texture sampling codegen.
1521  * This function handles texture sampling for all texture targets (1D,
1522  * 2D, 3D, cube) and all filtering modes.
1523  */
1524 static void
1525 lp_build_sample_general(struct lp_build_sample_context *bld,
1526                         unsigned unit,
1527                         LLVMValueRef s,
1528                         LLVMValueRef t,
1529                         LLVMValueRef r,
1530                         LLVMValueRef width,
1531                         LLVMValueRef height,
1532                         LLVMValueRef depth,
1533                         LLVMValueRef width_vec,
1534                         LLVMValueRef height_vec,
1535                         LLVMValueRef depth_vec,
1536                         LLVMValueRef row_stride_array,
1537                         LLVMValueRef img_stride_vec,
1538                         LLVMValueRef data_array,
1539                         LLVMValueRef *colors_out)
1540 {
1541    struct lp_build_context *float_bld = &bld->float_bld;
1542    const unsigned mip_filter = bld->static_state->min_mip_filter;
1543    const unsigned min_filter = bld->static_state->min_img_filter;
1544    const unsigned mag_filter = bld->static_state->mag_img_filter;
1545    const int dims = texture_dims(bld->static_state->target);
1546    LLVMValueRef lod, lod_fpart;
1547    LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
1548    LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1549    LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1550    LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1551    LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1552    LLVMValueRef data_ptr0, data_ptr1;
1553
1554    /*
1555    printf("%s mip %d  min %d  mag %d\n", __FUNCTION__,
1556           mip_filter, min_filter, mag_filter);
1557    */
1558
1559    /*
1560     * Compute the level of detail (float).
1561     */
1562    if (min_filter != mag_filter ||
1563        mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1564       /* Need to compute lod either to choose mipmap levels or to
1565        * distinguish between minification/magnification with one mipmap level.
1566        */
1567       lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
1568    }
1569
1570    /*
1571     * Compute integer mipmap level(s) to fetch texels from.
1572     */
1573    if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1574       /* always use mip level 0 */
1575       ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1576    }
1577    else {
1578       if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1579          lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1580       }
1581       else {
1582          assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1583          lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1584                                     &lod_fpart);
1585          lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1586       }
1587    }
1588
1589    /*
1590     * Convert scalar integer mipmap levels into vectors.
1591     */
1592    ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1593    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1594       ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1595
1596    /*
1597     * Compute width, height at mipmap level 'ilevel0'
1598     */
1599    width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1600    if (dims >= 2) {
1601       height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1602       row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1603                                                       ilevel0);
1604       if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1605          img_stride0_vec = lp_build_mul(&bld->int_coord_bld,
1606                                         row_stride0_vec, height0_vec);
1607          if (dims == 3) {
1608             depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1609          }
1610       }
1611    }
1612    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1613       /* compute width, height, depth for second mipmap level at 'ilevel1' */
1614       width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1615       if (dims >= 2) {
1616          height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1617          row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1618                                                          ilevel1);
1619          if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1620             img_stride1_vec = lp_build_mul(&bld->int_coord_bld,
1621                                            row_stride1_vec, height1_vec);
1622             if (dims ==3) {
1623                depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1624             }
1625          }
1626       }
1627    }
1628
1629    /*
1630     * Choose cube face, recompute per-face texcoords.
1631     */
1632    if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1633       LLVMValueRef face, face_s, face_t;
1634       lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1635       s = face_s; /* vec */
1636       t = face_t; /* vec */
1637       /* use 'r' to indicate cube face */
1638       r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1639    }
1640
1641    /*
1642     * Get pointer(s) to image data for mipmap level(s).
1643     */
1644    data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1645    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1646       data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1647    }
1648
1649    /*
1650     * Get/interpolate texture colors.
1651     */
1652    if (min_filter == mag_filter) {
1653       /* no need to distinquish between minification and magnification */
1654       lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart,
1655                              width0_vec, width1_vec,
1656                              height0_vec, height1_vec,
1657                              depth0_vec, depth1_vec,
1658                              row_stride0_vec, row_stride1_vec,
1659                              img_stride0_vec, img_stride1_vec,
1660                              data_ptr0, data_ptr1,
1661                              colors_out);
1662    }
1663    else {
1664       /* Emit conditional to choose min image filter or mag image filter
1665        * depending on the lod being >0 or <= 0, respectively.
1666        */
1667       struct lp_build_flow_context *flow_ctx;
1668       struct lp_build_if_state if_ctx;
1669       LLVMValueRef minify;
1670
1671       flow_ctx = lp_build_flow_create(bld->builder);
1672       lp_build_flow_scope_begin(flow_ctx);
1673
1674       lp_build_flow_scope_declare(flow_ctx, &colors_out[0]);
1675       lp_build_flow_scope_declare(flow_ctx, &colors_out[1]);
1676       lp_build_flow_scope_declare(flow_ctx, &colors_out[2]);
1677       lp_build_flow_scope_declare(flow_ctx, &colors_out[3]);
1678
1679       /* minify = lod > 0.0 */
1680       minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1681                              lod, float_bld->zero, "");
1682
1683       lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
1684       {
1685          /* Use the minification filter */
1686          lp_build_sample_mipmap(bld, min_filter, mip_filter,
1687                                 s, t, r, lod_fpart,
1688                                 width0_vec, width1_vec,
1689                                 height0_vec, height1_vec,
1690                                 depth0_vec, depth1_vec,
1691                                 row_stride0_vec, row_stride1_vec,
1692                                 img_stride0_vec, img_stride1_vec,
1693                                 data_ptr0, data_ptr1,
1694                                 colors_out);
1695       }
1696       lp_build_else(&if_ctx);
1697       {
1698          /* Use the magnification filter */
1699          lp_build_sample_mipmap(bld, mag_filter, mip_filter,
1700                                 s, t, r, lod_fpart,
1701                                 width0_vec, width1_vec,
1702                                 height0_vec, height1_vec,
1703                                 depth0_vec, depth1_vec,
1704                                 row_stride0_vec, row_stride1_vec,
1705                                 img_stride0_vec, img_stride1_vec,
1706                                 data_ptr0, data_ptr1,
1707                                 colors_out);
1708       }
1709       lp_build_endif(&if_ctx);
1710
1711       lp_build_flow_scope_end(flow_ctx);
1712       lp_build_flow_destroy(flow_ctx);
1713    }
1714 }
1715
1716
1717
1718 static void
1719 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1720                           struct lp_type dst_type,
1721                           LLVMValueRef packed,
1722                           LLVMValueRef *rgba)
1723 {
1724    LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
1725    unsigned chan;
1726
1727    /* Decode the input vector components */
1728    for (chan = 0; chan < 4; ++chan) {
1729       unsigned start = chan*8;
1730       unsigned stop = start + 8;
1731       LLVMValueRef input;
1732
1733       input = packed;
1734
1735       if(start)
1736          input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(dst_type, start), "");
1737
1738       if(stop < 32)
1739          input = LLVMBuildAnd(builder, input, mask, "");
1740
1741       input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1742
1743       rgba[chan] = input;
1744    }
1745 }
1746
1747
1748 static void
1749 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1750                               LLVMValueRef s,
1751                               LLVMValueRef t,
1752                               LLVMValueRef width,
1753                               LLVMValueRef height,
1754                               LLVMValueRef stride_array,
1755                               LLVMValueRef data_array,
1756                               LLVMValueRef *texel)
1757 {
1758    LLVMBuilderRef builder = bld->builder;
1759    struct lp_build_context i32, h16, u8n;
1760    LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1761    LLVMValueRef i32_c8, i32_c128, i32_c255;
1762    LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1763    LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1764    LLVMValueRef x0, x1;
1765    LLVMValueRef y0, y1;
1766    LLVMValueRef neighbors[2][2];
1767    LLVMValueRef neighbors_lo[2][2];
1768    LLVMValueRef neighbors_hi[2][2];
1769    LLVMValueRef packed, packed_lo, packed_hi;
1770    LLVMValueRef unswizzled[4];
1771    LLVMValueRef stride;
1772
1773    lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1774    lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1775    lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1776
1777    i32_vec_type = lp_build_vec_type(i32.type);
1778    h16_vec_type = lp_build_vec_type(h16.type);
1779    u8n_vec_type = lp_build_vec_type(u8n.type);
1780
1781    if (bld->static_state->normalized_coords) {
1782       LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1783       LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1784       LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1785       s = lp_build_mul(&bld->coord_bld, s, fp_width);
1786       t = lp_build_mul(&bld->coord_bld, t, fp_height);
1787    }
1788
1789    /* scale coords by 256 (8 fractional bits) */
1790    s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1791    t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1792
1793    /* convert float to int */
1794    s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1795    t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1796
1797    /* subtract 0.5 (add -128) */
1798    i32_c128 = lp_build_const_int_vec(i32.type, -128);
1799    s = LLVMBuildAdd(builder, s, i32_c128, "");
1800    t = LLVMBuildAdd(builder, t, i32_c128, "");
1801
1802    /* compute floor (shift right 8) */
1803    i32_c8 = lp_build_const_int_vec(i32.type, 8);
1804    s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1805    t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1806
1807    /* compute fractional part (AND with 0xff) */
1808    i32_c255 = lp_build_const_int_vec(i32.type, 255);
1809    s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1810    t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1811
1812    x0 = s_ipart;
1813    y0 = t_ipart;
1814
1815    x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1816    y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1817
1818    x0 = lp_build_sample_wrap_int(bld, x0, width,  bld->static_state->pot_width,
1819                                  bld->static_state->wrap_s);
1820    y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1821                                  bld->static_state->wrap_t);
1822
1823    x1 = lp_build_sample_wrap_int(bld, x1, width,  bld->static_state->pot_width,
1824                                  bld->static_state->wrap_s);
1825    y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1826                                  bld->static_state->wrap_t);
1827
1828    /*
1829     * Transform 4 x i32 in
1830     *
1831     *   s_fpart = {s0, s1, s2, s3}
1832     *
1833     * into 8 x i16
1834     *
1835     *   s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1836     *
1837     * into two 8 x i16
1838     *
1839     *   s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1840     *   s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1841     *
1842     * and likewise for t_fpart. There is no risk of loosing precision here
1843     * since the fractional parts only use the lower 8bits.
1844     */
1845
1846    s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1847    t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1848
1849    {
1850       LLVMTypeRef elem_type = LLVMInt32Type();
1851       LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1852       LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1853       LLVMValueRef shuffle_lo;
1854       LLVMValueRef shuffle_hi;
1855       unsigned i, j;
1856
1857       for(j = 0; j < h16.type.length; j += 4) {
1858          unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1859          LLVMValueRef index;
1860
1861          index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1862          for(i = 0; i < 4; ++i)
1863             shuffles_lo[j + i] = index;
1864
1865          index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1866          for(i = 0; i < 4; ++i)
1867             shuffles_hi[j + i] = index;
1868       }
1869
1870       shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1871       shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1872
1873       s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1874       t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1875       s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1876       t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1877    }
1878
1879    stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1880
1881    /*
1882     * Fetch the pixels as 4 x 32bit (rgba order might differ):
1883     *
1884     *   rgba0 rgba1 rgba2 rgba3
1885     *
1886     * bit cast them into 16 x u8
1887     *
1888     *   r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1889     *
1890     * unpack them into two 8 x i16:
1891     *
1892     *   r0 g0 b0 a0 r1 g1 b1 a1
1893     *   r2 g2 b2 a2 r3 g3 b3 a3
1894     *
1895     * The higher 8 bits of the resulting elements will be zero.
1896     */
1897
1898    neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1899    neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1900    neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1901    neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1902
1903    neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1904    neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1905    neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1906    neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1907
1908    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1909    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1910    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1911    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1912
1913    /*
1914     * Linear interpolate with 8.8 fixed point.
1915     */
1916
1917    packed_lo = lp_build_lerp_2d(&h16,
1918                                 s_fpart_lo, t_fpart_lo,
1919                                 neighbors_lo[0][0],
1920                                 neighbors_lo[0][1],
1921                                 neighbors_lo[1][0],
1922                                 neighbors_lo[1][1]);
1923
1924    packed_hi = lp_build_lerp_2d(&h16,
1925                                 s_fpart_hi, t_fpart_hi,
1926                                 neighbors_hi[0][0],
1927                                 neighbors_hi[0][1],
1928                                 neighbors_hi[1][0],
1929                                 neighbors_hi[1][1]);
1930
1931    packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1932
1933    /*
1934     * Convert to SoA and swizzle.
1935     */
1936
1937    packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
1938
1939    lp_build_rgba8_to_f32_soa(bld->builder,
1940                              bld->texel_type,
1941                              packed, unswizzled);
1942
1943    lp_build_format_swizzle_soa(bld->format_desc,
1944                                bld->texel_type, unswizzled,
1945                                texel);
1946 }
1947
1948
1949 static void
1950 lp_build_sample_compare(struct lp_build_sample_context *bld,
1951                         LLVMValueRef p,
1952                         LLVMValueRef *texel)
1953 {
1954    struct lp_build_context *texel_bld = &bld->texel_bld;
1955    LLVMValueRef res;
1956    unsigned chan;
1957
1958    if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1959       return;
1960
1961    /* TODO: Compare before swizzling, to avoid redundant computations */
1962    res = NULL;
1963    for(chan = 0; chan < 4; ++chan) {
1964       LLVMValueRef cmp;
1965       cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1966       cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1967
1968       if(res)
1969          res = lp_build_add(texel_bld, res, cmp);
1970       else
1971          res = cmp;
1972    }
1973
1974    assert(res);
1975    res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
1976
1977    /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1978    for(chan = 0; chan < 3; ++chan)
1979       texel[chan] = res;
1980    texel[3] = texel_bld->one;
1981 }
1982
1983
1984 /**
1985  * Build texture sampling code.
1986  * 'texel' will return a vector of four LLVMValueRefs corresponding to
1987  * R, G, B, A.
1988  * \param type  vector float type to use for coords, etc.
1989  */
1990 void
1991 lp_build_sample_soa(LLVMBuilderRef builder,
1992                     const struct lp_sampler_static_state *static_state,
1993                     struct lp_sampler_dynamic_state *dynamic_state,
1994                     struct lp_type type,
1995                     unsigned unit,
1996                     unsigned num_coords,
1997                     const LLVMValueRef *coords,
1998                     LLVMValueRef lodbias,
1999                     LLVMValueRef *texel)
2000 {
2001    struct lp_build_sample_context bld;
2002    LLVMValueRef width, width_vec;
2003    LLVMValueRef height, height_vec;
2004    LLVMValueRef depth, depth_vec;
2005    LLVMValueRef stride_array;
2006    LLVMValueRef data_array;
2007    LLVMValueRef s;
2008    LLVMValueRef t;
2009    LLVMValueRef r;
2010
2011    (void) lp_build_lod_selector;   /* temporary to silence warning */
2012    (void) lp_build_nearest_mip_level;
2013    (void) lp_build_linear_mip_levels;
2014    (void) lp_build_minify;
2015
2016    /* Setup our build context */
2017    memset(&bld, 0, sizeof bld);
2018    bld.builder = builder;
2019    bld.static_state = static_state;
2020    bld.dynamic_state = dynamic_state;
2021    bld.format_desc = util_format_description(static_state->format);
2022
2023    bld.float_type = lp_type_float(32);
2024    bld.int_type = lp_type_int(32);
2025    bld.coord_type = type;
2026    bld.uint_coord_type = lp_uint_type(type);
2027    bld.int_coord_type = lp_int_type(type);
2028    bld.texel_type = type;
2029
2030    lp_build_context_init(&bld.float_bld, builder, bld.float_type);
2031    lp_build_context_init(&bld.int_bld, builder, bld.int_type);
2032    lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
2033    lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
2034    lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
2035    lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
2036
2037    /* Get the dynamic state */
2038    width = dynamic_state->width(dynamic_state, builder, unit);
2039    height = dynamic_state->height(dynamic_state, builder, unit);
2040    depth = dynamic_state->depth(dynamic_state, builder, unit);
2041    stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
2042    data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
2043    /* Note that data_array is an array[level] of pointers to texture images */
2044
2045    s = coords[0];
2046    t = coords[1];
2047    r = coords[2];
2048
2049    width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
2050    height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
2051    depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
2052
2053    if (util_format_is_rgba8_variant(bld.format_desc) &&
2054        static_state->target == PIPE_TEXTURE_2D &&
2055        static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
2056        static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
2057        static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
2058        is_simple_wrap_mode(static_state->wrap_s) &&
2059        is_simple_wrap_mode(static_state->wrap_t)) {
2060       /* special case */
2061       lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
2062                                     stride_array, data_array, texel);
2063    }
2064    else {
2065       lp_build_sample_general(&bld, unit, s, t, r,
2066                               width, height, depth,
2067                               width_vec, height_vec, depth_vec,
2068                               stride_array, NULL, data_array,
2069                               texel);
2070    }
2071
2072    lp_build_sample_compare(&bld, r, texel);
2073 }