• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 -- common code.
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/format/u_format.h"
38 #include "util/u_math.h"
39 #include "util/u_cpu_detect.h"
40 #include "lp_bld_arit.h"
41 #include "lp_bld_const.h"
42 #include "lp_bld_debug.h"
43 #include "lp_bld_printf.h"
44 #include "lp_bld_flow.h"
45 #include "lp_bld_sample.h"
46 #include "lp_bld_swizzle.h"
47 #include "lp_bld_type.h"
48 #include "lp_bld_logic.h"
49 #include "lp_bld_pack.h"
50 #include "lp_bld_quad.h"
51 #include "lp_bld_bitarit.h"
52 
53 
54 /*
55  * Bri-linear factor. Should be greater than one.
56  */
57 #define BRILINEAR_FACTOR 2
58 
59 /**
60  * Does the given texture wrap mode allow sampling the texture border color?
61  * XXX maybe move this into gallium util code.
62  */
63 boolean
lp_sampler_wrap_mode_uses_border_color(unsigned mode,unsigned min_img_filter,unsigned mag_img_filter)64 lp_sampler_wrap_mode_uses_border_color(unsigned mode,
65                                        unsigned min_img_filter,
66                                        unsigned mag_img_filter)
67 {
68    switch (mode) {
69    case PIPE_TEX_WRAP_REPEAT:
70    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
71    case PIPE_TEX_WRAP_MIRROR_REPEAT:
72    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
73       return FALSE;
74    case PIPE_TEX_WRAP_CLAMP:
75    case PIPE_TEX_WRAP_MIRROR_CLAMP:
76       if (min_img_filter == PIPE_TEX_FILTER_NEAREST &&
77           mag_img_filter == PIPE_TEX_FILTER_NEAREST) {
78          return FALSE;
79       } else {
80          return TRUE;
81       }
82    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
83    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
84       return TRUE;
85    default:
86       assert(0 && "unexpected wrap mode");
87       return FALSE;
88    }
89 }
90 
91 
92 /**
93  * Initialize lp_sampler_static_texture_state object with the gallium
94  * texture/sampler_view state (this contains the parts which are
95  * considered static).
96  */
97 void
lp_sampler_static_texture_state(struct lp_static_texture_state * state,const struct pipe_sampler_view * view)98 lp_sampler_static_texture_state(struct lp_static_texture_state *state,
99                                 const struct pipe_sampler_view *view)
100 {
101    const struct pipe_resource *texture;
102 
103    memset(state, 0, sizeof *state);
104 
105    if (!view || !view->texture)
106       return;
107 
108    texture = view->texture;
109 
110    state->format            = view->format;
111    state->swizzle_r         = view->swizzle_r;
112    state->swizzle_g         = view->swizzle_g;
113    state->swizzle_b         = view->swizzle_b;
114    state->swizzle_a         = view->swizzle_a;
115 
116    state->target            = view->target;
117    state->pot_width         = util_is_power_of_two_or_zero(texture->width0);
118    state->pot_height        = util_is_power_of_two_or_zero(texture->height0);
119    state->pot_depth         = util_is_power_of_two_or_zero(texture->depth0);
120    state->level_zero_only   = !view->u.tex.last_level;
121 
122    /*
123     * the layer / element / level parameters are all either dynamic
124     * state or handled transparently wrt execution.
125     */
126 }
127 
128 /**
129  * Initialize lp_sampler_static_texture_state object with the gallium
130  * texture/sampler_view state (this contains the parts which are
131  * considered static).
132  */
133 void
lp_sampler_static_texture_state_image(struct lp_static_texture_state * state,const struct pipe_image_view * view)134 lp_sampler_static_texture_state_image(struct lp_static_texture_state *state,
135                                       const struct pipe_image_view *view)
136 {
137    const struct pipe_resource *resource;
138 
139    memset(state, 0, sizeof *state);
140 
141    if (!view || !view->resource)
142       return;
143 
144    resource = view->resource;
145 
146    state->format            = view->format;
147    state->swizzle_r         = PIPE_SWIZZLE_X;
148    state->swizzle_g         = PIPE_SWIZZLE_Y;
149    state->swizzle_b         = PIPE_SWIZZLE_Z;
150    state->swizzle_a         = PIPE_SWIZZLE_W;
151 
152    state->target            = view->resource->target;
153    state->pot_width         = util_is_power_of_two_or_zero(resource->width0);
154    state->pot_height        = util_is_power_of_two_or_zero(resource->height0);
155    state->pot_depth         = util_is_power_of_two_or_zero(resource->depth0);
156    state->level_zero_only   = 0;
157 
158    /*
159     * the layer / element / level parameters are all either dynamic
160     * state or handled transparently wrt execution.
161     */
162 }
163 
164 /**
165  * Initialize lp_sampler_static_sampler_state object with the gallium sampler
166  * state (this contains the parts which are considered static).
167  */
168 void
lp_sampler_static_sampler_state(struct lp_static_sampler_state * state,const struct pipe_sampler_state * sampler)169 lp_sampler_static_sampler_state(struct lp_static_sampler_state *state,
170                                 const struct pipe_sampler_state *sampler)
171 {
172    memset(state, 0, sizeof *state);
173 
174    if (!sampler)
175       return;
176 
177    /*
178     * We don't copy sampler state over unless it is actually enabled, to avoid
179     * spurious recompiles, as the sampler static state is part of the shader
180     * key.
181     *
182     * Ideally gallium frontends or cso_cache module would make all state
183     * canonical, but until that happens it's better to be safe than sorry here.
184     *
185     * XXX: Actually there's much more than can be done here, especially
186     * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
187     */
188 
189    state->wrap_s            = sampler->wrap_s;
190    state->wrap_t            = sampler->wrap_t;
191    state->wrap_r            = sampler->wrap_r;
192    state->min_img_filter    = sampler->min_img_filter;
193    state->mag_img_filter    = sampler->mag_img_filter;
194    state->min_mip_filter    = sampler->min_mip_filter;
195    state->seamless_cube_map = sampler->seamless_cube_map;
196 
197    if (sampler->max_lod > 0.0f) {
198       state->max_lod_pos = 1;
199    }
200 
201    if (sampler->lod_bias != 0.0f) {
202       state->lod_bias_non_zero = 1;
203    }
204 
205    if (state->min_mip_filter != PIPE_TEX_MIPFILTER_NONE ||
206        state->min_img_filter != state->mag_img_filter) {
207 
208       /* If min_lod == max_lod we can greatly simplify mipmap selection.
209        * This is a case that occurs during automatic mipmap generation.
210        */
211       if (sampler->min_lod == sampler->max_lod) {
212          state->min_max_lod_equal = 1;
213       } else {
214          if (sampler->min_lod > 0.0f) {
215             state->apply_min_lod = 1;
216          }
217 
218          /*
219           * XXX this won't do anything with the mesa state tracker which always
220           * sets max_lod to not more than actually present mip maps...
221           */
222          if (sampler->max_lod < (PIPE_MAX_TEXTURE_LEVELS - 1)) {
223             state->apply_max_lod = 1;
224          }
225       }
226    }
227 
228    state->compare_mode      = sampler->compare_mode;
229    if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
230       state->compare_func   = sampler->compare_func;
231    }
232 
233    state->normalized_coords = sampler->normalized_coords;
234 }
235 
236 
237 /**
238  * Generate code to compute coordinate gradient (rho).
239  * \param derivs  partial derivatives of (s, t, r, q) with respect to X and Y
240  *
241  * The resulting rho has bld->levelf format (per quad or per element).
242  */
243 static LLVMValueRef
lp_build_rho(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef s,LLVMValueRef t,LLVMValueRef r,LLVMValueRef cube_rho,const struct lp_derivatives * derivs)244 lp_build_rho(struct lp_build_sample_context *bld,
245              unsigned texture_unit,
246              LLVMValueRef s,
247              LLVMValueRef t,
248              LLVMValueRef r,
249              LLVMValueRef cube_rho,
250              const struct lp_derivatives *derivs)
251 {
252    struct gallivm_state *gallivm = bld->gallivm;
253    struct lp_build_context *int_size_bld = &bld->int_size_in_bld;
254    struct lp_build_context *float_size_bld = &bld->float_size_in_bld;
255    struct lp_build_context *float_bld = &bld->float_bld;
256    struct lp_build_context *coord_bld = &bld->coord_bld;
257    struct lp_build_context *rho_bld = &bld->lodf_bld;
258    const unsigned dims = bld->dims;
259    LLVMValueRef ddx_ddy[2] = {NULL};
260    LLVMBuilderRef builder = bld->gallivm->builder;
261    LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
262    LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
263    LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
264    LLVMValueRef index2 = LLVMConstInt(i32t, 2, 0);
265    LLVMValueRef rho_vec;
266    LLVMValueRef int_size, float_size;
267    LLVMValueRef rho;
268    LLVMValueRef first_level, first_level_vec;
269    unsigned length = coord_bld->type.length;
270    unsigned num_quads = length / 4;
271    boolean rho_per_quad = rho_bld->type.length != length;
272    boolean no_rho_opt = bld->no_rho_approx && (dims > 1);
273    unsigned i;
274    LLVMValueRef i32undef = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
275    LLVMValueRef rho_xvec, rho_yvec;
276 
277    /* Note that all simplified calculations will only work for isotropic filtering */
278 
279    /*
280     * rho calcs are always per quad except for explicit derivs (excluding
281     * the messy cube maps for now) when requested.
282     */
283 
284    first_level = bld->dynamic_state->first_level(bld->dynamic_state, bld->gallivm,
285                                                  bld->context_ptr, texture_unit, NULL);
286    first_level_vec = lp_build_broadcast_scalar(int_size_bld, first_level);
287    int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec, TRUE);
288    float_size = lp_build_int_to_float(float_size_bld, int_size);
289 
290    if (cube_rho) {
291       LLVMValueRef cubesize;
292       LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
293 
294       /*
295        * Cube map code did already everything except size mul and per-quad extraction.
296        * Luckily cube maps are always quadratic!
297        */
298       if (rho_per_quad) {
299          rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
300                                          rho_bld->type, cube_rho, 0);
301       }
302       else {
303          rho = lp_build_swizzle_scalar_aos(coord_bld, cube_rho, 0, 4);
304       }
305       /* Could optimize this for single quad just skip the broadcast */
306       cubesize = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
307                                             rho_bld->type, float_size, index0);
308       /* skipping sqrt hence returning rho squared */
309       cubesize = lp_build_mul(rho_bld, cubesize, cubesize);
310       rho = lp_build_mul(rho_bld, cubesize, rho);
311    }
312    else if (derivs) {
313       LLVMValueRef ddmax[3] = { NULL }, ddx[3] = { NULL }, ddy[3] = { NULL };
314       for (i = 0; i < dims; i++) {
315          LLVMValueRef floatdim;
316          LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
317 
318          floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
319                                                coord_bld->type, float_size, indexi);
320 
321          /*
322           * note that for rho_per_quad case could reduce math (at some shuffle
323           * cost), but for now use same code to per-pixel lod case.
324           */
325          if (no_rho_opt) {
326             ddx[i] = lp_build_mul(coord_bld, floatdim, derivs->ddx[i]);
327             ddy[i] = lp_build_mul(coord_bld, floatdim, derivs->ddy[i]);
328             ddx[i] = lp_build_mul(coord_bld, ddx[i], ddx[i]);
329             ddy[i] = lp_build_mul(coord_bld, ddy[i], ddy[i]);
330          }
331          else {
332             LLVMValueRef tmpx, tmpy;
333             tmpx = lp_build_abs(coord_bld, derivs->ddx[i]);
334             tmpy = lp_build_abs(coord_bld, derivs->ddy[i]);
335             ddmax[i] = lp_build_max(coord_bld, tmpx, tmpy);
336             ddmax[i] = lp_build_mul(coord_bld, floatdim, ddmax[i]);
337          }
338       }
339       if (no_rho_opt) {
340          rho_xvec = lp_build_add(coord_bld, ddx[0], ddx[1]);
341          rho_yvec = lp_build_add(coord_bld, ddy[0], ddy[1]);
342          if (dims > 2) {
343             rho_xvec = lp_build_add(coord_bld, rho_xvec, ddx[2]);
344             rho_yvec = lp_build_add(coord_bld, rho_yvec, ddy[2]);
345          }
346          rho = lp_build_max(coord_bld, rho_xvec, rho_yvec);
347          /* skipping sqrt hence returning rho squared */
348      }
349       else {
350          rho = ddmax[0];
351          if (dims > 1) {
352             rho = lp_build_max(coord_bld, rho, ddmax[1]);
353             if (dims > 2) {
354                rho = lp_build_max(coord_bld, rho, ddmax[2]);
355             }
356          }
357       }
358       if (rho_per_quad) {
359          /*
360           * rho_vec contains per-pixel rho, convert to scalar per quad.
361           */
362          rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
363                                          rho_bld->type, rho, 0);
364       }
365    }
366    else {
367       /*
368        * This looks all a bit complex, but it's not that bad
369        * (the shuffle code makes it look worse than it is).
370        * Still, might not be ideal for all cases.
371        */
372       static const unsigned char swizzle0[] = { /* no-op swizzle */
373          0, LP_BLD_SWIZZLE_DONTCARE,
374          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
375       };
376       static const unsigned char swizzle1[] = {
377          1, LP_BLD_SWIZZLE_DONTCARE,
378          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
379       };
380       static const unsigned char swizzle2[] = {
381          2, LP_BLD_SWIZZLE_DONTCARE,
382          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
383       };
384 
385       if (dims < 2) {
386          ddx_ddy[0] = lp_build_packed_ddx_ddy_onecoord(coord_bld, s);
387       }
388       else if (dims >= 2) {
389          ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
390          if (dims > 2) {
391             ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
392          }
393       }
394 
395       if (no_rho_opt) {
396          static const unsigned char swizzle01[] = { /* no-op swizzle */
397             0, 1,
398             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
399          };
400          static const unsigned char swizzle23[] = {
401             2, 3,
402             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
403          };
404          LLVMValueRef ddx_ddys, ddx_ddyt, floatdim, shuffles[LP_MAX_VECTOR_LENGTH / 4];
405 
406          for (i = 0; i < num_quads; i++) {
407             shuffles[i*4+0] = shuffles[i*4+1] = index0;
408             shuffles[i*4+2] = shuffles[i*4+3] = index1;
409          }
410          floatdim = LLVMBuildShuffleVector(builder, float_size, float_size,
411                                            LLVMConstVector(shuffles, length), "");
412          ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], floatdim);
413          ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], ddx_ddy[0]);
414          ddx_ddys = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle01);
415          ddx_ddyt = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle23);
416          rho_vec = lp_build_add(coord_bld, ddx_ddys, ddx_ddyt);
417 
418          if (dims > 2) {
419             static const unsigned char swizzle02[] = {
420                0, 2,
421                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
422             };
423             floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
424                                                   coord_bld->type, float_size, index2);
425             ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], floatdim);
426             ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], ddx_ddy[1]);
427             ddx_ddy[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
428             rho_vec = lp_build_add(coord_bld, rho_vec, ddx_ddy[1]);
429          }
430 
431          rho_xvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
432          rho_yvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
433          rho = lp_build_max(coord_bld, rho_xvec, rho_yvec);
434 
435          if (rho_per_quad) {
436             rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
437                                             rho_bld->type, rho, 0);
438          }
439          else {
440             rho = lp_build_swizzle_scalar_aos(coord_bld, rho, 0, 4);
441          }
442          /* skipping sqrt hence returning rho squared */
443       }
444       else {
445          ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
446          if (dims > 2) {
447             ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
448          }
449          else {
450             ddx_ddy[1] = NULL; /* silence compiler warning */
451          }
452 
453          if (dims < 2) {
454             rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle0);
455             rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
456          }
457          else if (dims == 2) {
458             static const unsigned char swizzle02[] = {
459                0, 2,
460                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
461             };
462             static const unsigned char swizzle13[] = {
463                1, 3,
464                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
465             };
466             rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle02);
467             rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle13);
468          }
469          else {
470             LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
471             LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
472             assert(dims == 3);
473             for (i = 0; i < num_quads; i++) {
474                shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
475                shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
476                shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
477                shuffles1[4*i + 3] = i32undef;
478                shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
479                shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
480                shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 2);
481                shuffles2[4*i + 3] = i32undef;
482             }
483             rho_xvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
484                                               LLVMConstVector(shuffles1, length), "");
485             rho_yvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
486                                               LLVMConstVector(shuffles2, length), "");
487          }
488 
489          rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
490 
491          if (bld->coord_type.length > 4) {
492             /* expand size to each quad */
493             if (dims > 1) {
494                /* could use some broadcast_vector helper for this? */
495                LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
496                for (i = 0; i < num_quads; i++) {
497                   src[i] = float_size;
498                }
499                float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
500             }
501             else {
502                float_size = lp_build_broadcast_scalar(coord_bld, float_size);
503             }
504             rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
505 
506             if (dims <= 1) {
507                rho = rho_vec;
508             }
509             else {
510                if (dims >= 2) {
511                   LLVMValueRef rho_s, rho_t, rho_r;
512 
513                   rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
514                   rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
515 
516                   rho = lp_build_max(coord_bld, rho_s, rho_t);
517 
518                   if (dims >= 3) {
519                      rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
520                      rho = lp_build_max(coord_bld, rho, rho_r);
521                   }
522                }
523             }
524             if (rho_per_quad) {
525                rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
526                                                rho_bld->type, rho, 0);
527             }
528             else {
529                rho = lp_build_swizzle_scalar_aos(coord_bld, rho, 0, 4);
530             }
531          }
532          else {
533             if (dims <= 1) {
534                rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
535             }
536             rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
537 
538             if (dims <= 1) {
539                rho = rho_vec;
540             }
541             else {
542                if (dims >= 2) {
543                   LLVMValueRef rho_s, rho_t, rho_r;
544 
545                   rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
546                   rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
547 
548                   rho = lp_build_max(float_bld, rho_s, rho_t);
549 
550                   if (dims >= 3) {
551                      rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
552                      rho = lp_build_max(float_bld, rho, rho_r);
553                   }
554                }
555             }
556             if (!rho_per_quad) {
557                rho = lp_build_broadcast_scalar(rho_bld, rho);
558             }
559          }
560       }
561    }
562 
563    return rho;
564 }
565 
566 
567 /*
568  * Bri-linear lod computation
569  *
570  * Use a piece-wise linear approximation of log2 such that:
571  * - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
572  * - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
573  *   with the steepness specified in 'factor'
574  * - exact result for 0.5, 1.5, etc.
575  *
576  *
577  *   1.0 -              /----*
578  *                     /
579  *                    /
580  *                   /
581  *   0.5 -          *
582  *                 /
583  *                /
584  *               /
585  *   0.0 - *----/
586  *
587  *         |                 |
588  *        2^0               2^1
589  *
590  * This is a technique also commonly used in hardware:
591  * - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
592  *
593  * TODO: For correctness, this should only be applied when texture is known to
594  * have regular mipmaps, i.e., mipmaps derived from the base level.
595  *
596  * TODO: This could be done in fixed point, where applicable.
597  */
598 static void
lp_build_brilinear_lod(struct lp_build_context * bld,LLVMValueRef lod,double factor,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart)599 lp_build_brilinear_lod(struct lp_build_context *bld,
600                        LLVMValueRef lod,
601                        double factor,
602                        LLVMValueRef *out_lod_ipart,
603                        LLVMValueRef *out_lod_fpart)
604 {
605    LLVMValueRef lod_fpart;
606    double pre_offset = (factor - 0.5)/factor - 0.5;
607    double post_offset = 1 - factor;
608 
609    if (0) {
610       lp_build_printf(bld->gallivm, "lod = %f\n", lod);
611    }
612 
613    lod = lp_build_add(bld, lod,
614                       lp_build_const_vec(bld->gallivm, bld->type, pre_offset));
615 
616    lp_build_ifloor_fract(bld, lod, out_lod_ipart, &lod_fpart);
617 
618    lod_fpart = lp_build_mad(bld, lod_fpart,
619                             lp_build_const_vec(bld->gallivm, bld->type, factor),
620                             lp_build_const_vec(bld->gallivm, bld->type, post_offset));
621 
622    /*
623     * It's not necessary to clamp lod_fpart since:
624     * - the above expression will never produce numbers greater than one.
625     * - the mip filtering branch is only taken if lod_fpart is positive
626     */
627 
628    *out_lod_fpart = lod_fpart;
629 
630    if (0) {
631       lp_build_printf(bld->gallivm, "lod_ipart = %i\n", *out_lod_ipart);
632       lp_build_printf(bld->gallivm, "lod_fpart = %f\n\n", *out_lod_fpart);
633    }
634 }
635 
636 
637 /*
638  * Combined log2 and brilinear lod computation.
639  *
640  * It's in all identical to calling lp_build_fast_log2() and
641  * lp_build_brilinear_lod() above, but by combining we can compute the integer
642  * and fractional part independently.
643  */
644 static void
lp_build_brilinear_rho(struct lp_build_context * bld,LLVMValueRef rho,double factor,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart)645 lp_build_brilinear_rho(struct lp_build_context *bld,
646                        LLVMValueRef rho,
647                        double factor,
648                        LLVMValueRef *out_lod_ipart,
649                        LLVMValueRef *out_lod_fpart)
650 {
651    LLVMValueRef lod_ipart;
652    LLVMValueRef lod_fpart;
653 
654    const double pre_factor = (2*factor - 0.5)/(M_SQRT2*factor);
655    const double post_offset = 1 - 2*factor;
656 
657    assert(bld->type.floating);
658 
659    assert(lp_check_value(bld->type, rho));
660 
661    /*
662     * The pre factor will make the intersections with the exact powers of two
663     * happen precisely where we want them to be, which means that the integer
664     * part will not need any post adjustments.
665     */
666    rho = lp_build_mul(bld, rho,
667                       lp_build_const_vec(bld->gallivm, bld->type, pre_factor));
668 
669    /* ipart = ifloor(log2(rho)) */
670    lod_ipart = lp_build_extract_exponent(bld, rho, 0);
671 
672    /* fpart = rho / 2**ipart */
673    lod_fpart = lp_build_extract_mantissa(bld, rho);
674 
675    lod_fpart = lp_build_mad(bld, lod_fpart,
676                             lp_build_const_vec(bld->gallivm, bld->type, factor),
677                             lp_build_const_vec(bld->gallivm, bld->type, post_offset));
678 
679    /*
680     * Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
681     * - the above expression will never produce numbers greater than one.
682     * - the mip filtering branch is only taken if lod_fpart is positive
683     */
684 
685    *out_lod_ipart = lod_ipart;
686    *out_lod_fpart = lod_fpart;
687 }
688 
689 
690 /**
691  * Fast implementation of iround(log2(sqrt(x))), based on
692  * log2(x^n) == n*log2(x).
693  *
694  * Gives accurate results all the time.
695  * (Could be trivially extended to handle other power-of-two roots.)
696  */
697 static LLVMValueRef
lp_build_ilog2_sqrt(struct lp_build_context * bld,LLVMValueRef x)698 lp_build_ilog2_sqrt(struct lp_build_context *bld,
699                     LLVMValueRef x)
700 {
701    LLVMBuilderRef builder = bld->gallivm->builder;
702    LLVMValueRef ipart;
703    struct lp_type i_type = lp_int_type(bld->type);
704    LLVMValueRef one = lp_build_const_int_vec(bld->gallivm, i_type, 1);
705 
706    assert(bld->type.floating);
707 
708    assert(lp_check_value(bld->type, x));
709 
710    /* ipart = log2(x) + 0.5 = 0.5*(log2(x^2) + 1.0) */
711    ipart = lp_build_extract_exponent(bld, x, 1);
712    ipart = LLVMBuildAShr(builder, ipart, one, "");
713 
714    return ipart;
715 }
716 
717 
718 /**
719  * Generate code to compute texture level of detail (lambda).
720  * \param derivs  partial derivatives of (s, t, r, q) with respect to X and Y
721  * \param lod_bias  optional float vector with the shader lod bias
722  * \param explicit_lod  optional float vector with the explicit lod
723  * \param cube_rho  rho calculated by cube coord mapping (optional)
724  * \param out_lod_ipart  integer part of lod
725  * \param out_lod_fpart  float part of lod (never larger than 1 but may be negative)
726  * \param out_lod_positive  (mask) if lod is positive (i.e. texture is minified)
727  *
728  * The resulting lod can be scalar per quad or be per element.
729  */
730 void
lp_build_lod_selector(struct lp_build_sample_context * bld,boolean is_lodq,unsigned texture_unit,unsigned sampler_unit,LLVMValueRef s,LLVMValueRef t,LLVMValueRef r,LLVMValueRef cube_rho,const struct lp_derivatives * derivs,LLVMValueRef lod_bias,LLVMValueRef explicit_lod,unsigned mip_filter,LLVMValueRef * out_lod,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart,LLVMValueRef * out_lod_positive)731 lp_build_lod_selector(struct lp_build_sample_context *bld,
732                       boolean is_lodq,
733                       unsigned texture_unit,
734                       unsigned sampler_unit,
735                       LLVMValueRef s,
736                       LLVMValueRef t,
737                       LLVMValueRef r,
738                       LLVMValueRef cube_rho,
739                       const struct lp_derivatives *derivs,
740                       LLVMValueRef lod_bias, /* optional */
741                       LLVMValueRef explicit_lod, /* optional */
742                       unsigned mip_filter,
743                       LLVMValueRef *out_lod,
744                       LLVMValueRef *out_lod_ipart,
745                       LLVMValueRef *out_lod_fpart,
746                       LLVMValueRef *out_lod_positive)
747 
748 {
749    LLVMBuilderRef builder = bld->gallivm->builder;
750    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
751    struct lp_build_context *lodf_bld = &bld->lodf_bld;
752    LLVMValueRef lod;
753 
754    *out_lod_ipart = bld->lodi_bld.zero;
755    *out_lod_positive = bld->lodi_bld.zero;
756    *out_lod_fpart = lodf_bld->zero;
757 
758    /*
759     * For determining min/mag, we follow GL 4.1 spec, 3.9.12 Texture Magnification:
760     * "Implementations may either unconditionally assume c = 0 for the minification
761     * vs. magnification switch-over point, or may choose to make c depend on the
762     * combination of minification and magnification modes as follows: if the
763     * magnification filter is given by LINEAR and the minification filter is given
764     * by NEAREST_MIPMAP_NEAREST or NEAREST_MIPMAP_LINEAR, then c = 0.5. This is
765     * done to ensure that a minified texture does not appear "sharper" than a
766     * magnified texture. Otherwise c = 0."
767     * And 3.9.11 Texture Minification:
768     * "If lod is less than or equal to the constant c (see section 3.9.12) the
769     * texture is said to be magnified; if it is greater, the texture is minified."
770     * So, using 0 as switchover point always, and using magnification for lod == 0.
771     * Note that the always c = 0 behavior is new (first appearing in GL 3.1 spec),
772     * old GL versions required 0.5 for the modes listed above.
773     * I have no clue about the (undocumented) wishes of d3d9/d3d10 here!
774     */
775 
776    if (bld->static_sampler_state->min_max_lod_equal && !is_lodq) {
777       /* User is forcing sampling from a particular mipmap level.
778        * This is hit during mipmap generation.
779        */
780       LLVMValueRef min_lod =
781          dynamic_state->min_lod(dynamic_state, bld->gallivm,
782                                 bld->context_ptr, sampler_unit);
783 
784       lod = lp_build_broadcast_scalar(lodf_bld, min_lod);
785    }
786    else {
787       if (explicit_lod) {
788          if (bld->num_lods != bld->coord_type.length)
789             lod = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
790                                             lodf_bld->type, explicit_lod, 0);
791          else
792             lod = explicit_lod;
793       }
794       else {
795          LLVMValueRef rho;
796          boolean rho_squared = (bld->no_rho_approx &&
797                                 (bld->dims > 1)) || cube_rho;
798 
799          rho = lp_build_rho(bld, texture_unit, s, t, r, cube_rho, derivs);
800 
801          /*
802           * Compute lod = log2(rho)
803           */
804 
805          if (!lod_bias && !is_lodq &&
806              !bld->static_sampler_state->lod_bias_non_zero &&
807              !bld->static_sampler_state->apply_max_lod &&
808              !bld->static_sampler_state->apply_min_lod) {
809             /*
810              * Special case when there are no post-log2 adjustments, which
811              * saves instructions but keeping the integer and fractional lod
812              * computations separate from the start.
813              */
814 
815             if (mip_filter == PIPE_TEX_MIPFILTER_NONE ||
816                 mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
817                /*
818                 * Don't actually need both values all the time, lod_ipart is
819                 * needed for nearest mipfilter, lod_positive if min != mag.
820                 */
821                if (rho_squared) {
822                   *out_lod_ipart = lp_build_ilog2_sqrt(lodf_bld, rho);
823                }
824                else {
825                   *out_lod_ipart = lp_build_ilog2(lodf_bld, rho);
826                }
827                *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
828                                                 rho, lodf_bld->one);
829                return;
830             }
831             if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR &&
832                 !bld->no_brilinear && !rho_squared) {
833                /*
834                 * This can't work if rho is squared. Not sure if it could be
835                 * fixed while keeping it worthwile, could also do sqrt here
836                 * but brilinear and no_rho_opt seems like a combination not
837                 * making much sense anyway so just use ordinary path below.
838                 */
839                lp_build_brilinear_rho(lodf_bld, rho, BRILINEAR_FACTOR,
840                                       out_lod_ipart, out_lod_fpart);
841                *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
842                                                 rho, lodf_bld->one);
843                return;
844             }
845          }
846 
847          if (0) {
848             lod = lp_build_log2(lodf_bld, rho);
849          }
850          else {
851             /* get more accurate results if we just sqaure rho always */
852             if (!rho_squared)
853                rho = lp_build_mul(lodf_bld, rho, rho);
854             lod = lp_build_fast_log2(lodf_bld, rho);
855          }
856 
857          /* log2(x^2) == 0.5*log2(x) */
858          lod = lp_build_mul(lodf_bld, lod,
859                             lp_build_const_vec(bld->gallivm, lodf_bld->type, 0.5F));
860 
861          /* add shader lod bias */
862          if (lod_bias) {
863             if (bld->num_lods != bld->coord_type.length)
864                lod_bias = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
865                                                     lodf_bld->type, lod_bias, 0);
866             lod = LLVMBuildFAdd(builder, lod, lod_bias, "shader_lod_bias");
867          }
868       }
869 
870       /* add sampler lod bias */
871       if (bld->static_sampler_state->lod_bias_non_zero) {
872          LLVMValueRef sampler_lod_bias =
873             dynamic_state->lod_bias(dynamic_state, bld->gallivm,
874                                     bld->context_ptr, sampler_unit);
875          sampler_lod_bias = lp_build_broadcast_scalar(lodf_bld,
876                                                       sampler_lod_bias);
877          lod = LLVMBuildFAdd(builder, lod, sampler_lod_bias, "sampler_lod_bias");
878       }
879 
880       if (is_lodq) {
881          *out_lod = lod;
882       }
883 
884       /* clamp lod */
885       if (bld->static_sampler_state->apply_max_lod) {
886          LLVMValueRef max_lod =
887             dynamic_state->max_lod(dynamic_state, bld->gallivm,
888                                    bld->context_ptr, sampler_unit);
889          max_lod = lp_build_broadcast_scalar(lodf_bld, max_lod);
890 
891          lod = lp_build_min(lodf_bld, lod, max_lod);
892       }
893       if (bld->static_sampler_state->apply_min_lod) {
894          LLVMValueRef min_lod =
895             dynamic_state->min_lod(dynamic_state, bld->gallivm,
896                                    bld->context_ptr, sampler_unit);
897          min_lod = lp_build_broadcast_scalar(lodf_bld, min_lod);
898 
899          lod = lp_build_max(lodf_bld, lod, min_lod);
900       }
901 
902       if (is_lodq) {
903          *out_lod_fpart = lod;
904          return;
905       }
906    }
907 
908    *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
909                                     lod, lodf_bld->zero);
910 
911    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
912       if (!bld->no_brilinear) {
913          lp_build_brilinear_lod(lodf_bld, lod, BRILINEAR_FACTOR,
914                                 out_lod_ipart, out_lod_fpart);
915       }
916       else {
917          lp_build_ifloor_fract(lodf_bld, lod, out_lod_ipart, out_lod_fpart);
918       }
919 
920       lp_build_name(*out_lod_fpart, "lod_fpart");
921    }
922    else {
923       *out_lod_ipart = lp_build_iround(lodf_bld, lod);
924    }
925 
926    lp_build_name(*out_lod_ipart, "lod_ipart");
927 
928    return;
929 }
930 
931 
932 /**
933  * For PIPE_TEX_MIPFILTER_NEAREST, convert int part of lod
934  * to actual mip level.
935  * Note: this is all scalar per quad code.
936  * \param lod_ipart  int texture level of detail
937  * \param level_out  returns integer
938  * \param out_of_bounds returns per coord out_of_bounds mask if provided
939  */
940 void
lp_build_nearest_mip_level(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef lod_ipart,LLVMValueRef * level_out,LLVMValueRef * out_of_bounds)941 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
942                            unsigned texture_unit,
943                            LLVMValueRef lod_ipart,
944                            LLVMValueRef *level_out,
945                            LLVMValueRef *out_of_bounds)
946 {
947    struct lp_build_context *leveli_bld = &bld->leveli_bld;
948    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
949    LLVMValueRef first_level, last_level, level;
950 
951    first_level = dynamic_state->first_level(dynamic_state, bld->gallivm,
952                                             bld->context_ptr, texture_unit, NULL);
953    last_level = dynamic_state->last_level(dynamic_state, bld->gallivm,
954                                           bld->context_ptr, texture_unit, NULL);
955    first_level = lp_build_broadcast_scalar(leveli_bld, first_level);
956    last_level = lp_build_broadcast_scalar(leveli_bld, last_level);
957 
958    level = lp_build_add(leveli_bld, lod_ipart, first_level);
959 
960    if (out_of_bounds) {
961       LLVMValueRef out, out1;
962       out = lp_build_cmp(leveli_bld, PIPE_FUNC_LESS, level, first_level);
963       out1 = lp_build_cmp(leveli_bld, PIPE_FUNC_GREATER, level, last_level);
964       out = lp_build_or(leveli_bld, out, out1);
965       if (bld->num_mips == bld->coord_bld.type.length) {
966          *out_of_bounds = out;
967       }
968       else if (bld->num_mips == 1) {
969          *out_of_bounds = lp_build_broadcast_scalar(&bld->int_coord_bld, out);
970       }
971       else {
972          assert(bld->num_mips == bld->coord_bld.type.length / 4);
973          *out_of_bounds = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
974                                                                 leveli_bld->type,
975                                                                 bld->int_coord_bld.type,
976                                                                 out);
977       }
978       level = lp_build_andnot(&bld->int_coord_bld, level, *out_of_bounds);
979       *level_out = level;
980    }
981    else {
982       /* clamp level to legal range of levels */
983       *level_out = lp_build_clamp(leveli_bld, level, first_level, last_level);
984 
985    }
986 }
987 
988 
989 /**
990  * For PIPE_TEX_MIPFILTER_LINEAR, convert per-quad (or per element) int LOD(s)
991  * to two (per-quad) (adjacent) mipmap level indexes, and fix up float lod
992  * part accordingly.
993  * Later, we'll sample from those two mipmap levels and interpolate between them.
994  */
995 void
lp_build_linear_mip_levels(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef lod_ipart,LLVMValueRef * lod_fpart_inout,LLVMValueRef * level0_out,LLVMValueRef * level1_out)996 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
997                            unsigned texture_unit,
998                            LLVMValueRef lod_ipart,
999                            LLVMValueRef *lod_fpart_inout,
1000                            LLVMValueRef *level0_out,
1001                            LLVMValueRef *level1_out)
1002 {
1003    LLVMBuilderRef builder = bld->gallivm->builder;
1004    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
1005    struct lp_build_context *leveli_bld = &bld->leveli_bld;
1006    struct lp_build_context *levelf_bld = &bld->levelf_bld;
1007    LLVMValueRef first_level, last_level;
1008    LLVMValueRef clamp_min;
1009    LLVMValueRef clamp_max;
1010 
1011    assert(bld->num_lods == bld->num_mips);
1012 
1013    first_level = dynamic_state->first_level(dynamic_state, bld->gallivm,
1014                                             bld->context_ptr, texture_unit, NULL);
1015    last_level = dynamic_state->last_level(dynamic_state, bld->gallivm,
1016                                           bld->context_ptr, texture_unit, NULL);
1017    first_level = lp_build_broadcast_scalar(leveli_bld, first_level);
1018    last_level = lp_build_broadcast_scalar(leveli_bld, last_level);
1019 
1020    *level0_out = lp_build_add(leveli_bld, lod_ipart, first_level);
1021    *level1_out = lp_build_add(leveli_bld, *level0_out, leveli_bld->one);
1022 
1023    /*
1024     * Clamp both *level0_out and *level1_out to [first_level, last_level], with
1025     * the minimum number of comparisons, and zeroing lod_fpart in the extreme
1026     * ends in the process.
1027     */
1028 
1029    /* *level0_out < first_level */
1030    clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
1031                              *level0_out, first_level,
1032                              "clamp_lod_to_first");
1033 
1034    *level0_out = LLVMBuildSelect(builder, clamp_min,
1035                                  first_level, *level0_out, "");
1036 
1037    *level1_out = LLVMBuildSelect(builder, clamp_min,
1038                                  first_level, *level1_out, "");
1039 
1040    *lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
1041                                       levelf_bld->zero, *lod_fpart_inout, "");
1042 
1043    /* *level0_out >= last_level */
1044    clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
1045                              *level0_out, last_level,
1046                              "clamp_lod_to_last");
1047 
1048    *level0_out = LLVMBuildSelect(builder, clamp_max,
1049                                  last_level, *level0_out, "");
1050 
1051    *level1_out = LLVMBuildSelect(builder, clamp_max,
1052                                  last_level, *level1_out, "");
1053 
1054    *lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
1055                                       levelf_bld->zero, *lod_fpart_inout, "");
1056 
1057    lp_build_name(*level0_out, "texture%u_miplevel0", texture_unit);
1058    lp_build_name(*level1_out, "texture%u_miplevel1", texture_unit);
1059    lp_build_name(*lod_fpart_inout, "texture%u_mipweight", texture_unit);
1060 }
1061 
1062 
1063 /**
1064  * Return pointer to a single mipmap level.
1065  * \param level  integer mipmap level
1066  */
1067 LLVMValueRef
lp_build_get_mipmap_level(struct lp_build_sample_context * bld,LLVMValueRef level)1068 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
1069                           LLVMValueRef level)
1070 {
1071    LLVMBuilderRef builder = bld->gallivm->builder;
1072    LLVMValueRef indexes[2], data_ptr, mip_offset;
1073 
1074    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1075    indexes[1] = level;
1076    mip_offset = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1077    mip_offset = LLVMBuildLoad(builder, mip_offset, "");
1078    data_ptr = LLVMBuildGEP(builder, bld->base_ptr, &mip_offset, 1, "");
1079    return data_ptr;
1080 }
1081 
1082 /**
1083  * Return (per-pixel) offsets to mip levels.
1084  * \param level  integer mipmap level
1085  */
1086 LLVMValueRef
lp_build_get_mip_offsets(struct lp_build_sample_context * bld,LLVMValueRef level)1087 lp_build_get_mip_offsets(struct lp_build_sample_context *bld,
1088                          LLVMValueRef level)
1089 {
1090    LLVMBuilderRef builder = bld->gallivm->builder;
1091    LLVMValueRef indexes[2], offsets, offset1;
1092 
1093    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1094    if (bld->num_mips == 1) {
1095       indexes[1] = level;
1096       offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1097       offset1 = LLVMBuildLoad(builder, offset1, "");
1098       offsets = lp_build_broadcast_scalar(&bld->int_coord_bld, offset1);
1099    }
1100    else if (bld->num_mips == bld->coord_bld.type.length / 4) {
1101       unsigned i;
1102 
1103       offsets = bld->int_coord_bld.undef;
1104       for (i = 0; i < bld->num_mips; i++) {
1105          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1106          LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, 4 * i);
1107          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1108          offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1109          offset1 = LLVMBuildLoad(builder, offset1, "");
1110          offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexo, "");
1111       }
1112       offsets = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, offsets, 0, 4);
1113    }
1114    else {
1115       unsigned i;
1116 
1117       assert (bld->num_mips == bld->coord_bld.type.length);
1118 
1119       offsets = bld->int_coord_bld.undef;
1120       for (i = 0; i < bld->num_mips; i++) {
1121          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1122          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1123          offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1124          offset1 = LLVMBuildLoad(builder, offset1, "");
1125          offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexi, "");
1126       }
1127    }
1128    return offsets;
1129 }
1130 
1131 
1132 /**
1133  * Codegen equivalent for u_minify().
1134  * @param lod_scalar  if lod is a (broadcasted) scalar
1135  * Return max(1, base_size >> level);
1136  */
1137 LLVMValueRef
lp_build_minify(struct lp_build_context * bld,LLVMValueRef base_size,LLVMValueRef level,boolean lod_scalar)1138 lp_build_minify(struct lp_build_context *bld,
1139                 LLVMValueRef base_size,
1140                 LLVMValueRef level,
1141                 boolean lod_scalar)
1142 {
1143    LLVMBuilderRef builder = bld->gallivm->builder;
1144    assert(lp_check_value(bld->type, base_size));
1145    assert(lp_check_value(bld->type, level));
1146 
1147    if (level == bld->zero) {
1148       /* if we're using mipmap level zero, no minification is needed */
1149       return base_size;
1150    }
1151    else {
1152       LLVMValueRef size;
1153       assert(bld->type.sign);
1154       if (lod_scalar ||
1155          (util_cpu_caps.has_avx2 || !util_cpu_caps.has_sse)) {
1156          size = LLVMBuildLShr(builder, base_size, level, "minify");
1157          size = lp_build_max(bld, size, bld->one);
1158       }
1159       else {
1160          /*
1161           * emulate shift with float mul, since intel "forgot" shifts with
1162           * per-element shift count until avx2, which results in terrible
1163           * scalar extraction (both count and value), scalar shift,
1164           * vector reinsertion. Should not be an issue on any non-x86 cpu
1165           * with a vector instruction set.
1166           * On cpus with AMD's XOP this should also be unnecessary but I'm
1167           * not sure if llvm would emit this with current flags.
1168           */
1169          LLVMValueRef const127, const23, lf;
1170          struct lp_type ftype;
1171          struct lp_build_context fbld;
1172          ftype = lp_type_float_vec(32, bld->type.length * bld->type.width);
1173          lp_build_context_init(&fbld, bld->gallivm, ftype);
1174          const127 = lp_build_const_int_vec(bld->gallivm, bld->type, 127);
1175          const23 = lp_build_const_int_vec(bld->gallivm, bld->type, 23);
1176 
1177          /* calculate 2^(-level) float */
1178          lf = lp_build_sub(bld, const127, level);
1179          lf = lp_build_shl(bld, lf, const23);
1180          lf = LLVMBuildBitCast(builder, lf, fbld.vec_type, "");
1181 
1182          /* finish shift operation by doing float mul */
1183          base_size = lp_build_int_to_float(&fbld, base_size);
1184          size = lp_build_mul(&fbld, base_size, lf);
1185          /*
1186           * do the max also with floats because
1187           * a) non-emulated int max requires sse41
1188           *    (this is actually a lie as we could cast to 16bit values
1189           *    as 16bit is sufficient and 16bit int max is sse2)
1190           * b) with avx we can do int max 4-wide but float max 8-wide
1191           */
1192          size = lp_build_max(&fbld, size, fbld.one);
1193          size = lp_build_itrunc(&fbld, size);
1194       }
1195       return size;
1196    }
1197 }
1198 
1199 
1200 /**
1201  * Dereference stride_array[mipmap_level] array to get a stride.
1202  * Return stride as a vector.
1203  */
1204 static LLVMValueRef
lp_build_get_level_stride_vec(struct lp_build_sample_context * bld,LLVMValueRef stride_array,LLVMValueRef level)1205 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
1206                               LLVMValueRef stride_array, LLVMValueRef level)
1207 {
1208    LLVMBuilderRef builder = bld->gallivm->builder;
1209    LLVMValueRef indexes[2], stride, stride1;
1210    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1211    if (bld->num_mips == 1) {
1212       indexes[1] = level;
1213       stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1214       stride1 = LLVMBuildLoad(builder, stride1, "");
1215       stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride1);
1216    }
1217    else if (bld->num_mips == bld->coord_bld.type.length / 4) {
1218       LLVMValueRef stride1;
1219       unsigned i;
1220 
1221       stride = bld->int_coord_bld.undef;
1222       for (i = 0; i < bld->num_mips; i++) {
1223          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1224          LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, 4 * i);
1225          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1226          stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1227          stride1 = LLVMBuildLoad(builder, stride1, "");
1228          stride = LLVMBuildInsertElement(builder, stride, stride1, indexo, "");
1229       }
1230       stride = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, stride, 0, 4);
1231    }
1232    else {
1233       LLVMValueRef stride1;
1234       unsigned i;
1235 
1236       assert (bld->num_mips == bld->coord_bld.type.length);
1237 
1238       stride = bld->int_coord_bld.undef;
1239       for (i = 0; i < bld->coord_bld.type.length; i++) {
1240          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1241          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1242          stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1243          stride1 = LLVMBuildLoad(builder, stride1, "");
1244          stride = LLVMBuildInsertElement(builder, stride, stride1, indexi, "");
1245       }
1246    }
1247    return stride;
1248 }
1249 
1250 
1251 /**
1252  * When sampling a mipmap, we need to compute the width, height, depth
1253  * of the source levels from the level indexes.  This helper function
1254  * does that.
1255  */
1256 void
lp_build_mipmap_level_sizes(struct lp_build_sample_context * bld,LLVMValueRef ilevel,LLVMValueRef * out_size,LLVMValueRef * row_stride_vec,LLVMValueRef * img_stride_vec)1257 lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
1258                             LLVMValueRef ilevel,
1259                             LLVMValueRef *out_size,
1260                             LLVMValueRef *row_stride_vec,
1261                             LLVMValueRef *img_stride_vec)
1262 {
1263    const unsigned dims = bld->dims;
1264    LLVMValueRef ilevel_vec;
1265 
1266    /*
1267     * Compute width, height, depth at mipmap level 'ilevel'
1268     */
1269    if (bld->num_mips == 1) {
1270       ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
1271       *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec, TRUE);
1272    }
1273    else {
1274       LLVMValueRef int_size_vec;
1275       LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
1276       unsigned num_quads = bld->coord_bld.type.length / 4;
1277       unsigned i;
1278 
1279       if (bld->num_mips == num_quads) {
1280          /*
1281           * XXX: this should be #ifndef SANE_INSTRUCTION_SET.
1282           * intel "forgot" the variable shift count instruction until avx2.
1283           * A harmless 8x32 shift gets translated into 32 instructions
1284           * (16 extracts, 8 scalar shifts, 8 inserts), llvm is apparently
1285           * unable to recognize if there are really just 2 different shift
1286           * count values. So do the shift 4-wide before expansion.
1287           */
1288          struct lp_build_context bld4;
1289          struct lp_type type4;
1290 
1291          type4 = bld->int_coord_bld.type;
1292          type4.length = 4;
1293 
1294          lp_build_context_init(&bld4, bld->gallivm, type4);
1295 
1296          if (bld->dims == 1) {
1297             assert(bld->int_size_in_bld.type.length == 1);
1298             int_size_vec = lp_build_broadcast_scalar(&bld4,
1299                                                      bld->int_size);
1300          }
1301          else {
1302             assert(bld->int_size_in_bld.type.length == 4);
1303             int_size_vec = bld->int_size;
1304          }
1305 
1306          for (i = 0; i < num_quads; i++) {
1307             LLVMValueRef ileveli;
1308             LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1309 
1310             ileveli = lp_build_extract_broadcast(bld->gallivm,
1311                                                  bld->leveli_bld.type,
1312                                                  bld4.type,
1313                                                  ilevel,
1314                                                  indexi);
1315             tmp[i] = lp_build_minify(&bld4, int_size_vec, ileveli, TRUE);
1316          }
1317          /*
1318           * out_size is [w0, h0, d0, _, w1, h1, d1, _, ...] vector for dims > 1,
1319           * [w0, w0, w0, w0, w1, w1, w1, w1, ...] otherwise.
1320           */
1321          *out_size = lp_build_concat(bld->gallivm,
1322                                      tmp,
1323                                      bld4.type,
1324                                      num_quads);
1325       }
1326       else {
1327         /* FIXME: this is terrible and results in _huge_ vector
1328          * (for the dims > 1 case).
1329          * Should refactor this (together with extract_image_sizes) and do
1330          * something more useful. Could for instance if we have width,height
1331          * with 4-wide vector pack all elements into a 8xi16 vector
1332          * (on which we can still do useful math) instead of using a 16xi32
1333          * vector.
1334          * For dims == 1 this will create [w0, w1, w2, w3, ...] vector.
1335          * For dims > 1 this will create [w0, h0, d0, _, w1, h1, d1, _, ...] vector.
1336          */
1337          assert(bld->num_mips == bld->coord_bld.type.length);
1338          if (bld->dims == 1) {
1339             assert(bld->int_size_in_bld.type.length == 1);
1340             int_size_vec = lp_build_broadcast_scalar(&bld->int_coord_bld,
1341                                                      bld->int_size);
1342             *out_size = lp_build_minify(&bld->int_coord_bld, int_size_vec, ilevel, FALSE);
1343          }
1344          else {
1345             LLVMValueRef ilevel1;
1346             for (i = 0; i < bld->num_mips; i++) {
1347                LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1348                ilevel1 = lp_build_extract_broadcast(bld->gallivm, bld->int_coord_type,
1349                                                     bld->int_size_in_bld.type, ilevel, indexi);
1350                tmp[i] = bld->int_size;
1351                tmp[i] = lp_build_minify(&bld->int_size_in_bld, tmp[i], ilevel1, TRUE);
1352             }
1353             *out_size = lp_build_concat(bld->gallivm, tmp,
1354                                         bld->int_size_in_bld.type,
1355                                         bld->num_mips);
1356          }
1357       }
1358    }
1359 
1360    if (dims >= 2) {
1361       *row_stride_vec = lp_build_get_level_stride_vec(bld,
1362                                                       bld->row_stride_array,
1363                                                       ilevel);
1364    }
1365    if (dims == 3 || has_layer_coord(bld->static_texture_state->target)) {
1366       *img_stride_vec = lp_build_get_level_stride_vec(bld,
1367                                                       bld->img_stride_array,
1368                                                       ilevel);
1369    }
1370 }
1371 
1372 
1373 /**
1374  * Extract and broadcast texture size.
1375  *
1376  * @param size_type   type of the texture size vector (either
1377  *                    bld->int_size_type or bld->float_size_type)
1378  * @param coord_type  type of the texture size vector (either
1379  *                    bld->int_coord_type or bld->coord_type)
1380  * @param size        vector with the texture size (width, height, depth)
1381  */
1382 void
lp_build_extract_image_sizes(struct lp_build_sample_context * bld,struct lp_build_context * size_bld,struct lp_type coord_type,LLVMValueRef size,LLVMValueRef * out_width,LLVMValueRef * out_height,LLVMValueRef * out_depth)1383 lp_build_extract_image_sizes(struct lp_build_sample_context *bld,
1384                              struct lp_build_context *size_bld,
1385                              struct lp_type coord_type,
1386                              LLVMValueRef size,
1387                              LLVMValueRef *out_width,
1388                              LLVMValueRef *out_height,
1389                              LLVMValueRef *out_depth)
1390 {
1391    const unsigned dims = bld->dims;
1392    LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
1393    struct lp_type size_type = size_bld->type;
1394 
1395    if (bld->num_mips == 1) {
1396       *out_width = lp_build_extract_broadcast(bld->gallivm,
1397                                               size_type,
1398                                               coord_type,
1399                                               size,
1400                                               LLVMConstInt(i32t, 0, 0));
1401       if (dims >= 2) {
1402          *out_height = lp_build_extract_broadcast(bld->gallivm,
1403                                                   size_type,
1404                                                   coord_type,
1405                                                   size,
1406                                                   LLVMConstInt(i32t, 1, 0));
1407          if (dims == 3) {
1408             *out_depth = lp_build_extract_broadcast(bld->gallivm,
1409                                                     size_type,
1410                                                     coord_type,
1411                                                     size,
1412                                                     LLVMConstInt(i32t, 2, 0));
1413          }
1414       }
1415    }
1416    else {
1417       unsigned num_quads = bld->coord_bld.type.length / 4;
1418 
1419       if (dims == 1) {
1420          *out_width = size;
1421       }
1422       else if (bld->num_mips == num_quads) {
1423          *out_width = lp_build_swizzle_scalar_aos(size_bld, size, 0, 4);
1424          if (dims >= 2) {
1425             *out_height = lp_build_swizzle_scalar_aos(size_bld, size, 1, 4);
1426             if (dims == 3) {
1427                *out_depth = lp_build_swizzle_scalar_aos(size_bld, size, 2, 4);
1428             }
1429          }
1430       }
1431       else {
1432          assert(bld->num_mips == bld->coord_type.length);
1433          *out_width = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1434                                                 coord_type, size, 0);
1435          if (dims >= 2) {
1436             *out_height = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1437                                                     coord_type, size, 1);
1438             if (dims == 3) {
1439                *out_depth = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1440                                                       coord_type, size, 2);
1441             }
1442          }
1443       }
1444    }
1445 }
1446 
1447 
1448 /**
1449  * Unnormalize coords.
1450  *
1451  * @param flt_size  vector with the integer texture size (width, height, depth)
1452  */
1453 void
lp_build_unnormalized_coords(struct lp_build_sample_context * bld,LLVMValueRef flt_size,LLVMValueRef * s,LLVMValueRef * t,LLVMValueRef * r)1454 lp_build_unnormalized_coords(struct lp_build_sample_context *bld,
1455                              LLVMValueRef flt_size,
1456                              LLVMValueRef *s,
1457                              LLVMValueRef *t,
1458                              LLVMValueRef *r)
1459 {
1460    const unsigned dims = bld->dims;
1461    LLVMValueRef width;
1462    LLVMValueRef height = NULL;
1463    LLVMValueRef depth = NULL;
1464 
1465    lp_build_extract_image_sizes(bld,
1466                                 &bld->float_size_bld,
1467                                 bld->coord_type,
1468                                 flt_size,
1469                                 &width,
1470                                 &height,
1471                                 &depth);
1472 
1473    /* s = s * width, t = t * height */
1474    *s = lp_build_mul(&bld->coord_bld, *s, width);
1475    if (dims >= 2) {
1476       *t = lp_build_mul(&bld->coord_bld, *t, height);
1477       if (dims >= 3) {
1478          *r = lp_build_mul(&bld->coord_bld, *r, depth);
1479       }
1480    }
1481 }
1482 
1483 /**
1484  * Generate new coords and faces for cubemap texels falling off the face.
1485  *
1486  * @param face   face (center) of the pixel
1487  * @param x0     lower x coord
1488  * @param x1     higher x coord (must be x0 + 1)
1489  * @param y0     lower y coord
1490  * @param y1     higher y coord (must be x0 + 1)
1491  * @param max_coord     texture cube (level) size - 1
1492  * @param next_faces    new face values when falling off
1493  * @param next_xcoords  new x coord values when falling off
1494  * @param next_ycoords  new y coord values when falling off
1495  *
1496  * The arrays hold the new values when under/overflow of
1497  * lower x, higher x, lower y, higher y coord would occur (in this order).
1498  * next_xcoords/next_ycoords have two entries each (for both new lower and
1499  * higher coord).
1500  */
1501 void
lp_build_cube_new_coords(struct lp_build_context * ivec_bld,LLVMValueRef face,LLVMValueRef x0,LLVMValueRef x1,LLVMValueRef y0,LLVMValueRef y1,LLVMValueRef max_coord,LLVMValueRef next_faces[4],LLVMValueRef next_xcoords[4][2],LLVMValueRef next_ycoords[4][2])1502 lp_build_cube_new_coords(struct lp_build_context *ivec_bld,
1503                         LLVMValueRef face,
1504                         LLVMValueRef x0,
1505                         LLVMValueRef x1,
1506                         LLVMValueRef y0,
1507                         LLVMValueRef y1,
1508                         LLVMValueRef max_coord,
1509                         LLVMValueRef next_faces[4],
1510                         LLVMValueRef next_xcoords[4][2],
1511                         LLVMValueRef next_ycoords[4][2])
1512 {
1513    /*
1514     * Lookup tables aren't nice for simd code hence try some logic here.
1515     * (Note that while it would not be necessary to do per-sample (4) lookups
1516     * when using a LUT as it's impossible that texels fall off of positive
1517     * and negative edges simultaneously, it would however be necessary to
1518     * do 2 lookups for corner handling as in this case texels both fall off
1519     * of x and y axes.)
1520     */
1521    /*
1522     * Next faces (for face 012345):
1523     * x < 0.0  : 451110
1524     * x >= 1.0 : 540001
1525     * y < 0.0  : 225422
1526     * y >= 1.0 : 334533
1527     * Hence nfx+ (and nfy+) == nfx- (nfy-) xor 1
1528     * nfx-: face > 1 ? (face == 5 ? 0 : 1) : (4 + face & 1)
1529     * nfy+: face & ~4 > 1 ? face + 2 : 3;
1530     * This could also use pshufb instead, but would need (manually coded)
1531     * ssse3 intrinsic (llvm won't do non-constant shuffles).
1532     */
1533    struct gallivm_state *gallivm = ivec_bld->gallivm;
1534    LLVMValueRef sel, sel_f2345, sel_f23, sel_f2, tmpsel, tmp;
1535    LLVMValueRef faceand1, sel_fand1, maxmx0, maxmx1, maxmy0, maxmy1;
1536    LLVMValueRef c2 = lp_build_const_int_vec(gallivm, ivec_bld->type, 2);
1537    LLVMValueRef c3 = lp_build_const_int_vec(gallivm, ivec_bld->type, 3);
1538    LLVMValueRef c4 = lp_build_const_int_vec(gallivm, ivec_bld->type, 4);
1539    LLVMValueRef c5 = lp_build_const_int_vec(gallivm, ivec_bld->type, 5);
1540 
1541    sel = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, face, c5);
1542    tmpsel = lp_build_select(ivec_bld, sel, ivec_bld->zero, ivec_bld->one);
1543    sel_f2345 = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, face, ivec_bld->one);
1544    faceand1 = lp_build_and(ivec_bld, face, ivec_bld->one);
1545    tmp = lp_build_add(ivec_bld, faceand1, c4);
1546    next_faces[0] = lp_build_select(ivec_bld, sel_f2345, tmpsel, tmp);
1547    next_faces[1] = lp_build_xor(ivec_bld, next_faces[0], ivec_bld->one);
1548 
1549    tmp = lp_build_andnot(ivec_bld, face, c4);
1550    sel_f23 = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, tmp, ivec_bld->one);
1551    tmp = lp_build_add(ivec_bld, face, c2);
1552    next_faces[3] = lp_build_select(ivec_bld, sel_f23, tmp, c3);
1553    next_faces[2] = lp_build_xor(ivec_bld, next_faces[3], ivec_bld->one);
1554 
1555    /*
1556     * new xcoords (for face 012345):
1557     * x < 0.0  : max   max   t     max-t max  max
1558     * x >= 1.0 : 0     0     max-t t     0    0
1559     * y < 0.0  : max   0     max-s s     s    max-s
1560     * y >= 1.0 : max   0     s     max-s s    max-s
1561     *
1562     * ncx[1] = face & ~4 > 1 ? (face == 2 ? max-t : t) : 0
1563     * ncx[0] = max - ncx[1]
1564     * ncx[3] = face > 1 ? (face & 1 ? max-s : s) : (face & 1) ? 0 : max
1565     * ncx[2] = face & ~4 > 1 ? max - ncx[3] : ncx[3]
1566     */
1567    sel_f2 = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, face, c2);
1568    maxmy0 = lp_build_sub(ivec_bld, max_coord, y0);
1569    tmp = lp_build_select(ivec_bld, sel_f2, maxmy0, y0);
1570    next_xcoords[1][0] = lp_build_select(ivec_bld, sel_f23, tmp, ivec_bld->zero);
1571    next_xcoords[0][0] = lp_build_sub(ivec_bld, max_coord, next_xcoords[1][0]);
1572    maxmy1 = lp_build_sub(ivec_bld, max_coord, y1);
1573    tmp = lp_build_select(ivec_bld, sel_f2, maxmy1, y1);
1574    next_xcoords[1][1] = lp_build_select(ivec_bld, sel_f23, tmp, ivec_bld->zero);
1575    next_xcoords[0][1] = lp_build_sub(ivec_bld, max_coord, next_xcoords[1][1]);
1576 
1577    sel_fand1 = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, faceand1, ivec_bld->one);
1578 
1579    tmpsel = lp_build_select(ivec_bld, sel_fand1, ivec_bld->zero, max_coord);
1580    maxmx0 = lp_build_sub(ivec_bld, max_coord, x0);
1581    tmp = lp_build_select(ivec_bld, sel_fand1, maxmx0, x0);
1582    next_xcoords[3][0] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1583    tmp = lp_build_sub(ivec_bld, max_coord, next_xcoords[3][0]);
1584    next_xcoords[2][0] = lp_build_select(ivec_bld, sel_f23, tmp, next_xcoords[3][0]);
1585    maxmx1 = lp_build_sub(ivec_bld, max_coord, x1);
1586    tmp = lp_build_select(ivec_bld, sel_fand1, maxmx1, x1);
1587    next_xcoords[3][1] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1588    tmp = lp_build_sub(ivec_bld, max_coord, next_xcoords[3][1]);
1589    next_xcoords[2][1] = lp_build_select(ivec_bld, sel_f23, tmp, next_xcoords[3][1]);
1590 
1591    /*
1592     * new ycoords (for face 012345):
1593     * x < 0.0  : t     t     0     max   t    t
1594     * x >= 1.0 : t     t     0     max   t    t
1595     * y < 0.0  : max-s s     0     max   max  0
1596     * y >= 1.0 : s     max-s 0     max   0    max
1597     *
1598     * ncy[0] = face & ~4 > 1 ? (face == 2 ? 0 : max) : t
1599     * ncy[1] = ncy[0]
1600     * ncy[3] = face > 1 ? (face & 1 ? max : 0) : (face & 1) ? max-s : max
1601     * ncx[2] = face & ~4 > 1 ? max - ncx[3] : ncx[3]
1602     */
1603    tmp = lp_build_select(ivec_bld, sel_f2, ivec_bld->zero, max_coord);
1604    next_ycoords[0][0] = lp_build_select(ivec_bld, sel_f23, tmp, y0);
1605    next_ycoords[1][0] = next_ycoords[0][0];
1606    next_ycoords[0][1] = lp_build_select(ivec_bld, sel_f23, tmp, y1);
1607    next_ycoords[1][1] = next_ycoords[0][1];
1608 
1609    tmpsel = lp_build_select(ivec_bld, sel_fand1, maxmx0, x0);
1610    tmp = lp_build_select(ivec_bld, sel_fand1, max_coord, ivec_bld->zero);
1611    next_ycoords[3][0] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1612    tmp = lp_build_sub(ivec_bld, max_coord, next_ycoords[3][0]);
1613    next_ycoords[2][0] = lp_build_select(ivec_bld, sel_f23, next_ycoords[3][0], tmp);
1614    tmpsel = lp_build_select(ivec_bld, sel_fand1, maxmx1, x1);
1615    tmp = lp_build_select(ivec_bld, sel_fand1, max_coord, ivec_bld->zero);
1616    next_ycoords[3][1] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1617    tmp = lp_build_sub(ivec_bld, max_coord, next_ycoords[3][1]);
1618    next_ycoords[2][1] = lp_build_select(ivec_bld, sel_f23, next_ycoords[3][1], tmp);
1619 }
1620 
1621 
1622 /** Helper used by lp_build_cube_lookup() */
1623 static LLVMValueRef
lp_build_cube_imapos(struct lp_build_context * coord_bld,LLVMValueRef coord)1624 lp_build_cube_imapos(struct lp_build_context *coord_bld, LLVMValueRef coord)
1625 {
1626    /* ima = +0.5 / abs(coord); */
1627    LLVMValueRef posHalf = lp_build_const_vec(coord_bld->gallivm, coord_bld->type, 0.5);
1628    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1629    LLVMValueRef ima = lp_build_div(coord_bld, posHalf, absCoord);
1630    return ima;
1631 }
1632 
1633 
1634 /** Helper for doing 3-wise selection.
1635  * Returns sel1 ? val2 : (sel0 ? val0 : val1).
1636  */
1637 static LLVMValueRef
lp_build_select3(struct lp_build_context * sel_bld,LLVMValueRef sel0,LLVMValueRef sel1,LLVMValueRef val0,LLVMValueRef val1,LLVMValueRef val2)1638 lp_build_select3(struct lp_build_context *sel_bld,
1639                  LLVMValueRef sel0,
1640                  LLVMValueRef sel1,
1641                  LLVMValueRef val0,
1642                  LLVMValueRef val1,
1643                  LLVMValueRef val2)
1644 {
1645    LLVMValueRef tmp;
1646    tmp = lp_build_select(sel_bld, sel0, val0, val1);
1647    return lp_build_select(sel_bld, sel1, val2, tmp);
1648 }
1649 
1650 
1651 /**
1652  * Generate code to do cube face selection and compute per-face texcoords.
1653  */
1654 void
lp_build_cube_lookup(struct lp_build_sample_context * bld,LLVMValueRef * coords,const struct lp_derivatives * derivs_in,LLVMValueRef * rho,struct lp_derivatives * derivs_out,boolean need_derivs)1655 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1656                      LLVMValueRef *coords,
1657                      const struct lp_derivatives *derivs_in, /* optional */
1658                      LLVMValueRef *rho,
1659                      struct lp_derivatives *derivs_out, /* optional */
1660                      boolean need_derivs)
1661 {
1662    struct lp_build_context *coord_bld = &bld->coord_bld;
1663    LLVMBuilderRef builder = bld->gallivm->builder;
1664    struct gallivm_state *gallivm = bld->gallivm;
1665    LLVMValueRef si, ti, ri;
1666 
1667    /*
1668     * Do per-pixel face selection. We cannot however (as we used to do)
1669     * simply calculate the derivs afterwards (which is very bogus for
1670     * explicit derivs btw) because the values would be "random" when
1671     * not all pixels lie on the same face. So what we do here is just
1672     * calculate the derivatives after scaling the coords by the absolute
1673     * value of the inverse major axis, and essentially do rho calculation
1674     * steps as if it were a 3d texture. This is perfect if all pixels hit
1675     * the same face, but not so great at edges, I believe the max error
1676     * should be sqrt(2) with no_rho_approx or 2 otherwise (essentially measuring
1677     * the 3d distance between 2 points on the cube instead of measuring up/down
1678     * the edge). Still this is possibly a win over just selecting the same face
1679     * for all pixels. Unfortunately, something like that doesn't work for
1680     * explicit derivatives.
1681     */
1682    struct lp_build_context *cint_bld = &bld->int_coord_bld;
1683    struct lp_type intctype = cint_bld->type;
1684    LLVMTypeRef coord_vec_type = coord_bld->vec_type;
1685    LLVMTypeRef cint_vec_type = cint_bld->vec_type;
1686    LLVMValueRef as, at, ar, face, face_s, face_t;
1687    LLVMValueRef as_ge_at, maxasat, ar_ge_as_at;
1688    LLVMValueRef snewx, tnewx, snewy, tnewy, snewz, tnewz;
1689    LLVMValueRef tnegi, rnegi;
1690    LLVMValueRef ma, mai, signma, signmabit, imahalfpos;
1691    LLVMValueRef posHalf = lp_build_const_vec(gallivm, coord_bld->type, 0.5);
1692    LLVMValueRef signmask = lp_build_const_int_vec(gallivm, intctype,
1693                                                   1LL << (intctype.width - 1));
1694    LLVMValueRef signshift = lp_build_const_int_vec(gallivm, intctype,
1695                                                    intctype.width -1);
1696    LLVMValueRef facex = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_X);
1697    LLVMValueRef facey = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Y);
1698    LLVMValueRef facez = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Z);
1699    LLVMValueRef s = coords[0];
1700    LLVMValueRef t = coords[1];
1701    LLVMValueRef r = coords[2];
1702 
1703    assert(PIPE_TEX_FACE_NEG_X == PIPE_TEX_FACE_POS_X + 1);
1704    assert(PIPE_TEX_FACE_NEG_Y == PIPE_TEX_FACE_POS_Y + 1);
1705    assert(PIPE_TEX_FACE_NEG_Z == PIPE_TEX_FACE_POS_Z + 1);
1706 
1707    /*
1708     * get absolute value (for x/y/z face selection) and sign bit
1709     * (for mirroring minor coords and pos/neg face selection)
1710     * of the original coords.
1711     */
1712    as = lp_build_abs(&bld->coord_bld, s);
1713    at = lp_build_abs(&bld->coord_bld, t);
1714    ar = lp_build_abs(&bld->coord_bld, r);
1715 
1716    /*
1717     * major face determination: select x if x > y else select y
1718     * select z if z >= max(x,y) else select previous result
1719     * if some axis are the same we chose z over y, y over x - the
1720     * dx10 spec seems to ask for it while OpenGL doesn't care (if we
1721     * wouldn't care could save a select or two if using different
1722     * compares and doing at_g_as_ar last since tnewx and tnewz are the
1723     * same).
1724     */
1725    as_ge_at = lp_build_cmp(coord_bld, PIPE_FUNC_GREATER, as, at);
1726    maxasat = lp_build_max(coord_bld, as, at);
1727    ar_ge_as_at = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, ar, maxasat);
1728 
1729    if (need_derivs && (derivs_in || (bld->no_quad_lod && bld->no_rho_approx))) {
1730       /*
1731        * XXX: This is really really complex.
1732        * It is a bit overkill to use this for implicit derivatives as well,
1733        * no way this is worth the cost in practice, but seems to be the
1734        * only way for getting accurate and per-pixel lod values.
1735        */
1736       LLVMValueRef ima, imahalf, tmp, ddx[3], ddy[3];
1737       LLVMValueRef madx, mady, madxdivma, madydivma;
1738       LLVMValueRef sdxi, tdxi, rdxi, sdyi, tdyi, rdyi;
1739       LLVMValueRef tdxnegi, rdxnegi, tdynegi, rdynegi;
1740       LLVMValueRef sdxnewx, sdxnewy, sdxnewz, tdxnewx, tdxnewy, tdxnewz;
1741       LLVMValueRef sdynewx, sdynewy, sdynewz, tdynewx, tdynewy, tdynewz;
1742       LLVMValueRef face_sdx, face_tdx, face_sdy, face_tdy;
1743       /*
1744        * s = 1/2 * ( sc / ma + 1)
1745        * t = 1/2 * ( tc / ma + 1)
1746        *
1747        * s' = 1/2 * (sc' * ma - sc * ma') / ma^2
1748        * t' = 1/2 * (tc' * ma - tc * ma') / ma^2
1749        *
1750        * dx.s = 0.5 * (dx.sc - sc * dx.ma / ma) / ma
1751        * dx.t = 0.5 * (dx.tc - tc * dx.ma / ma) / ma
1752        * dy.s = 0.5 * (dy.sc - sc * dy.ma / ma) / ma
1753        * dy.t = 0.5 * (dy.tc - tc * dy.ma / ma) / ma
1754        */
1755 
1756       /* select ma, calculate ima */
1757       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1758       mai = LLVMBuildBitCast(builder, ma, cint_vec_type, "");
1759       signmabit = LLVMBuildAnd(builder, mai, signmask, "");
1760       ima = lp_build_div(coord_bld, coord_bld->one, ma);
1761       imahalf = lp_build_mul(coord_bld, posHalf, ima);
1762       imahalfpos = lp_build_abs(coord_bld, imahalf);
1763 
1764       if (!derivs_in) {
1765          ddx[0] = lp_build_ddx(coord_bld, s);
1766          ddx[1] = lp_build_ddx(coord_bld, t);
1767          ddx[2] = lp_build_ddx(coord_bld, r);
1768          ddy[0] = lp_build_ddy(coord_bld, s);
1769          ddy[1] = lp_build_ddy(coord_bld, t);
1770          ddy[2] = lp_build_ddy(coord_bld, r);
1771       }
1772       else {
1773          ddx[0] = derivs_in->ddx[0];
1774          ddx[1] = derivs_in->ddx[1];
1775          ddx[2] = derivs_in->ddx[2];
1776          ddy[0] = derivs_in->ddy[0];
1777          ddy[1] = derivs_in->ddy[1];
1778          ddy[2] = derivs_in->ddy[2];
1779       }
1780 
1781       /* select major derivatives */
1782       madx = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, ddx[0], ddx[1], ddx[2]);
1783       mady = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, ddy[0], ddy[1], ddy[2]);
1784 
1785       si = LLVMBuildBitCast(builder, s, cint_vec_type, "");
1786       ti = LLVMBuildBitCast(builder, t, cint_vec_type, "");
1787       ri = LLVMBuildBitCast(builder, r, cint_vec_type, "");
1788 
1789       sdxi = LLVMBuildBitCast(builder, ddx[0], cint_vec_type, "");
1790       tdxi = LLVMBuildBitCast(builder, ddx[1], cint_vec_type, "");
1791       rdxi = LLVMBuildBitCast(builder, ddx[2], cint_vec_type, "");
1792 
1793       sdyi = LLVMBuildBitCast(builder, ddy[0], cint_vec_type, "");
1794       tdyi = LLVMBuildBitCast(builder, ddy[1], cint_vec_type, "");
1795       rdyi = LLVMBuildBitCast(builder, ddy[2], cint_vec_type, "");
1796 
1797       /*
1798        * compute all possible new s/t coords, which does the mirroring,
1799        * and do the same for derivs minor axes.
1800        * snewx = signma * -r;
1801        * tnewx = -t;
1802        * snewy = s;
1803        * tnewy = signma * r;
1804        * snewz = signma * s;
1805        * tnewz = -t;
1806        */
1807       tnegi = LLVMBuildXor(builder, ti, signmask, "");
1808       rnegi = LLVMBuildXor(builder, ri, signmask, "");
1809       tdxnegi = LLVMBuildXor(builder, tdxi, signmask, "");
1810       rdxnegi = LLVMBuildXor(builder, rdxi, signmask, "");
1811       tdynegi = LLVMBuildXor(builder, tdyi, signmask, "");
1812       rdynegi = LLVMBuildXor(builder, rdyi, signmask, "");
1813 
1814       snewx = LLVMBuildXor(builder, signmabit, rnegi, "");
1815       tnewx = tnegi;
1816       sdxnewx = LLVMBuildXor(builder, signmabit, rdxnegi, "");
1817       tdxnewx = tdxnegi;
1818       sdynewx = LLVMBuildXor(builder, signmabit, rdynegi, "");
1819       tdynewx = tdynegi;
1820 
1821       snewy = si;
1822       tnewy = LLVMBuildXor(builder, signmabit, ri, "");
1823       sdxnewy = sdxi;
1824       tdxnewy = LLVMBuildXor(builder, signmabit, rdxi, "");
1825       sdynewy = sdyi;
1826       tdynewy = LLVMBuildXor(builder, signmabit, rdyi, "");
1827 
1828       snewz = LLVMBuildXor(builder, signmabit, si, "");
1829       tnewz = tnegi;
1830       sdxnewz = LLVMBuildXor(builder, signmabit, sdxi, "");
1831       tdxnewz = tdxnegi;
1832       sdynewz = LLVMBuildXor(builder, signmabit, sdyi, "");
1833       tdynewz = tdynegi;
1834 
1835       /* select the mirrored values */
1836       face = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, facex, facey, facez);
1837       face_s = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, snewx, snewy, snewz);
1838       face_t = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tnewx, tnewy, tnewz);
1839       face_sdx = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, sdxnewx, sdxnewy, sdxnewz);
1840       face_tdx = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tdxnewx, tdxnewy, tdxnewz);
1841       face_sdy = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, sdynewx, sdynewy, sdynewz);
1842       face_tdy = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tdynewx, tdynewy, tdynewz);
1843 
1844       face_s = LLVMBuildBitCast(builder, face_s, coord_vec_type, "");
1845       face_t = LLVMBuildBitCast(builder, face_t, coord_vec_type, "");
1846       face_sdx = LLVMBuildBitCast(builder, face_sdx, coord_vec_type, "");
1847       face_tdx = LLVMBuildBitCast(builder, face_tdx, coord_vec_type, "");
1848       face_sdy = LLVMBuildBitCast(builder, face_sdy, coord_vec_type, "");
1849       face_tdy = LLVMBuildBitCast(builder, face_tdy, coord_vec_type, "");
1850 
1851       /* deriv math, dx.s = 0.5 * (dx.sc - sc * dx.ma / ma) / ma */
1852       madxdivma = lp_build_mul(coord_bld, madx, ima);
1853       tmp = lp_build_mul(coord_bld, madxdivma, face_s);
1854       tmp = lp_build_sub(coord_bld, face_sdx, tmp);
1855       derivs_out->ddx[0] = lp_build_mul(coord_bld, tmp, imahalf);
1856 
1857       /* dx.t = 0.5 * (dx.tc - tc * dx.ma / ma) / ma */
1858       tmp = lp_build_mul(coord_bld, madxdivma, face_t);
1859       tmp = lp_build_sub(coord_bld, face_tdx, tmp);
1860       derivs_out->ddx[1] = lp_build_mul(coord_bld, tmp, imahalf);
1861 
1862       /* dy.s = 0.5 * (dy.sc - sc * dy.ma / ma) / ma */
1863       madydivma = lp_build_mul(coord_bld, mady, ima);
1864       tmp = lp_build_mul(coord_bld, madydivma, face_s);
1865       tmp = lp_build_sub(coord_bld, face_sdy, tmp);
1866       derivs_out->ddy[0] = lp_build_mul(coord_bld, tmp, imahalf);
1867 
1868       /* dy.t = 0.5 * (dy.tc - tc * dy.ma / ma) / ma */
1869       tmp = lp_build_mul(coord_bld, madydivma, face_t);
1870       tmp = lp_build_sub(coord_bld, face_tdy, tmp);
1871       derivs_out->ddy[1] = lp_build_mul(coord_bld, tmp, imahalf);
1872 
1873       signma = LLVMBuildLShr(builder, mai, signshift, "");
1874       coords[2] = LLVMBuildOr(builder, face, signma, "face");
1875 
1876       /* project coords */
1877       face_s = lp_build_mul(coord_bld, face_s, imahalfpos);
1878       face_t = lp_build_mul(coord_bld, face_t, imahalfpos);
1879 
1880       coords[0] = lp_build_add(coord_bld, face_s, posHalf);
1881       coords[1] = lp_build_add(coord_bld, face_t, posHalf);
1882 
1883       return;
1884    }
1885 
1886    else if (need_derivs) {
1887       LLVMValueRef ddx_ddy[2], tmp[3], rho_vec;
1888       static const unsigned char swizzle0[] = { /* no-op swizzle */
1889          0, LP_BLD_SWIZZLE_DONTCARE,
1890          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1891       };
1892       static const unsigned char swizzle1[] = {
1893          1, LP_BLD_SWIZZLE_DONTCARE,
1894          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1895       };
1896       static const unsigned char swizzle01[] = { /* no-op swizzle */
1897          0, 1,
1898          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1899       };
1900       static const unsigned char swizzle23[] = {
1901          2, 3,
1902          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1903       };
1904       static const unsigned char swizzle02[] = {
1905          0, 2,
1906          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1907       };
1908 
1909       /*
1910        * scale the s/t/r coords pre-select/mirror so we can calculate
1911        * "reasonable" derivs.
1912        */
1913       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1914       imahalfpos = lp_build_cube_imapos(coord_bld, ma);
1915       s = lp_build_mul(coord_bld, s, imahalfpos);
1916       t = lp_build_mul(coord_bld, t, imahalfpos);
1917       r = lp_build_mul(coord_bld, r, imahalfpos);
1918 
1919       /*
1920        * This isn't quite the same as the "ordinary" (3d deriv) path since we
1921        * know the texture is square which simplifies things (we can omit the
1922        * size mul which happens very early completely here and do it at the
1923        * very end).
1924        * Also always do calculations according to GALLIVM_DEBUG_NO_RHO_APPROX
1925        * since the error can get quite big otherwise at edges.
1926        * (With no_rho_approx max error is sqrt(2) at edges, same as it is
1927        * without no_rho_approx for 2d textures, otherwise it would be factor 2.)
1928        */
1929       ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
1930       ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
1931 
1932       ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], ddx_ddy[0]);
1933       ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], ddx_ddy[1]);
1934 
1935       tmp[0] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle01);
1936       tmp[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle23);
1937       tmp[2] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
1938 
1939       rho_vec = lp_build_add(coord_bld, tmp[0], tmp[1]);
1940       rho_vec = lp_build_add(coord_bld, rho_vec, tmp[2]);
1941 
1942       tmp[0] = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
1943       tmp[1] = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
1944       *rho = lp_build_max(coord_bld, tmp[0], tmp[1]);
1945    }
1946 
1947    if (!need_derivs) {
1948       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1949    }
1950    mai = LLVMBuildBitCast(builder, ma, cint_vec_type, "");
1951    signmabit = LLVMBuildAnd(builder, mai, signmask, "");
1952 
1953    si = LLVMBuildBitCast(builder, s, cint_vec_type, "");
1954    ti = LLVMBuildBitCast(builder, t, cint_vec_type, "");
1955    ri = LLVMBuildBitCast(builder, r, cint_vec_type, "");
1956 
1957    /*
1958     * compute all possible new s/t coords, which does the mirroring
1959     * snewx = signma * -r;
1960     * tnewx = -t;
1961     * snewy = s;
1962     * tnewy = signma * r;
1963     * snewz = signma * s;
1964     * tnewz = -t;
1965     */
1966    tnegi = LLVMBuildXor(builder, ti, signmask, "");
1967    rnegi = LLVMBuildXor(builder, ri, signmask, "");
1968 
1969    snewx = LLVMBuildXor(builder, signmabit, rnegi, "");
1970    tnewx = tnegi;
1971 
1972    snewy = si;
1973    tnewy = LLVMBuildXor(builder, signmabit, ri, "");
1974 
1975    snewz = LLVMBuildXor(builder, signmabit, si, "");
1976    tnewz = tnegi;
1977 
1978    /* select the mirrored values */
1979    face_s = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, snewx, snewy, snewz);
1980    face_t = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tnewx, tnewy, tnewz);
1981    face = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, facex, facey, facez);
1982 
1983    face_s = LLVMBuildBitCast(builder, face_s, coord_vec_type, "");
1984    face_t = LLVMBuildBitCast(builder, face_t, coord_vec_type, "");
1985 
1986    /* add +1 for neg face */
1987    /* XXX with AVX probably want to use another select here -
1988     * as long as we ensure vblendvps gets used we can actually
1989     * skip the comparison and just use sign as a "mask" directly.
1990     */
1991    signma = LLVMBuildLShr(builder, mai, signshift, "");
1992    coords[2] = LLVMBuildOr(builder, face, signma, "face");
1993 
1994    /* project coords */
1995    if (!need_derivs) {
1996       imahalfpos = lp_build_cube_imapos(coord_bld, ma);
1997       face_s = lp_build_mul(coord_bld, face_s, imahalfpos);
1998       face_t = lp_build_mul(coord_bld, face_t, imahalfpos);
1999    }
2000 
2001    coords[0] = lp_build_add(coord_bld, face_s, posHalf);
2002    coords[1] = lp_build_add(coord_bld, face_t, posHalf);
2003 }
2004 
2005 
2006 /**
2007  * Compute the partial offset of a pixel block along an arbitrary axis.
2008  *
2009  * @param coord   coordinate in pixels
2010  * @param stride  number of bytes between rows of successive pixel blocks
2011  * @param block_length  number of pixels in a pixels block along the coordinate
2012  *                      axis
2013  * @param out_offset    resulting relative offset of the pixel block in bytes
2014  * @param out_subcoord  resulting sub-block pixel coordinate
2015  */
2016 void
lp_build_sample_partial_offset(struct lp_build_context * bld,unsigned block_length,LLVMValueRef coord,LLVMValueRef stride,LLVMValueRef * out_offset,LLVMValueRef * out_subcoord)2017 lp_build_sample_partial_offset(struct lp_build_context *bld,
2018                                unsigned block_length,
2019                                LLVMValueRef coord,
2020                                LLVMValueRef stride,
2021                                LLVMValueRef *out_offset,
2022                                LLVMValueRef *out_subcoord)
2023 {
2024    LLVMBuilderRef builder = bld->gallivm->builder;
2025    LLVMValueRef offset;
2026    LLVMValueRef subcoord;
2027 
2028    if (block_length == 1) {
2029       subcoord = bld->zero;
2030    }
2031    else {
2032       /*
2033        * Pixel blocks have power of two dimensions. LLVM should convert the
2034        * rem/div to bit arithmetic.
2035        * TODO: Verify this.
2036        * It does indeed BUT it does transform it to scalar (and back) when doing so
2037        * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
2038        * The generated code looks seriously unfunny and is quite expensive.
2039        */
2040 #if 0
2041       LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
2042       subcoord = LLVMBuildURem(builder, coord, block_width, "");
2043       coord    = LLVMBuildUDiv(builder, coord, block_width, "");
2044 #else
2045       unsigned logbase2 = util_logbase2(block_length);
2046       LLVMValueRef block_shift = lp_build_const_int_vec(bld->gallivm, bld->type, logbase2);
2047       LLVMValueRef block_mask = lp_build_const_int_vec(bld->gallivm, bld->type, block_length - 1);
2048       subcoord = LLVMBuildAnd(builder, coord, block_mask, "");
2049       coord = LLVMBuildLShr(builder, coord, block_shift, "");
2050 #endif
2051    }
2052 
2053    offset = lp_build_mul(bld, coord, stride);
2054 
2055    assert(out_offset);
2056    assert(out_subcoord);
2057 
2058    *out_offset = offset;
2059    *out_subcoord = subcoord;
2060 }
2061 
2062 
2063 /**
2064  * Compute the offset of a pixel block.
2065  *
2066  * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
2067  *
2068  * Returns the relative offset and i,j sub-block coordinates
2069  */
2070 void
lp_build_sample_offset(struct lp_build_context * bld,const struct util_format_description * format_desc,LLVMValueRef x,LLVMValueRef y,LLVMValueRef z,LLVMValueRef y_stride,LLVMValueRef z_stride,LLVMValueRef * out_offset,LLVMValueRef * out_i,LLVMValueRef * out_j)2071 lp_build_sample_offset(struct lp_build_context *bld,
2072                        const struct util_format_description *format_desc,
2073                        LLVMValueRef x,
2074                        LLVMValueRef y,
2075                        LLVMValueRef z,
2076                        LLVMValueRef y_stride,
2077                        LLVMValueRef z_stride,
2078                        LLVMValueRef *out_offset,
2079                        LLVMValueRef *out_i,
2080                        LLVMValueRef *out_j)
2081 {
2082    LLVMValueRef x_stride;
2083    LLVMValueRef offset;
2084 
2085    x_stride = lp_build_const_vec(bld->gallivm, bld->type,
2086                                  format_desc->block.bits/8);
2087 
2088    lp_build_sample_partial_offset(bld,
2089                                   format_desc->block.width,
2090                                   x, x_stride,
2091                                   &offset, out_i);
2092 
2093    if (y && y_stride) {
2094       LLVMValueRef y_offset;
2095       lp_build_sample_partial_offset(bld,
2096                                      format_desc->block.height,
2097                                      y, y_stride,
2098                                      &y_offset, out_j);
2099       offset = lp_build_add(bld, offset, y_offset);
2100    }
2101    else {
2102       *out_j = bld->zero;
2103    }
2104 
2105    if (z && z_stride) {
2106       LLVMValueRef z_offset;
2107       LLVMValueRef k;
2108       lp_build_sample_partial_offset(bld,
2109                                      1, /* pixel blocks are always 2D */
2110                                      z, z_stride,
2111                                      &z_offset, &k);
2112       offset = lp_build_add(bld, offset, z_offset);
2113    }
2114 
2115    *out_offset = offset;
2116 }
2117