projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: Eliminate color buffer swizzling.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /**
30 * @file
31 * Code generate the whole fragment pipeline.
32 *
33 * The fragment pipeline consists of the following stages:
34 * - early depth test
35 * - fragment shader
36 * - alpha test
37 * - depth/stencil test
38 * - blending
39 *
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
45 *
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
51 *
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
56 *
57 * @author Jose Fonseca <jfonseca@vmware.com>
58 */
59
60 #include <limits.h>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/u_simple_list.h"
69 #include "os/os_time.h"
70 #include "pipe/p_shader_tokens.h"
71 #include "draw/draw_context.h"
72 #include "tgsi/tgsi_dump.h"
73 #include "tgsi/tgsi_scan.h"
74 #include "tgsi/tgsi_parse.h"
75 #include "gallivm/lp_bld_type.h"
76 #include "gallivm/lp_bld_const.h"
77 #include "gallivm/lp_bld_conv.h"
78 #include "gallivm/lp_bld_init.h"
79 #include "gallivm/lp_bld_intr.h"
80 #include "gallivm/lp_bld_logic.h"
81 #include "gallivm/lp_bld_tgsi.h"
82 #include "gallivm/lp_bld_swizzle.h"
83 #include "gallivm/lp_bld_flow.h"
84 #include "gallivm/lp_bld_debug.h"
85 #include "gallivm/lp_bld_arit.h"
86 #include "gallivm/lp_bld_pack.h"
87 #include "gallivm/lp_bld_format.h"
88 #include "gallivm/lp_bld_quad.h"
89
90 #include "lp_bld_alpha.h"
91 #include "lp_bld_blend.h"
92 #include "lp_bld_depth.h"
93 #include "lp_bld_interp.h"
94 #include "lp_context.h"
95 #include "lp_debug.h"
96 #include "lp_perf.h"
97 #include "lp_setup.h"
98 #include "lp_state.h"
99 #include "lp_tex_sample.h"
100 #include "lp_flush.h"
101 #include "lp_state_fs.h"
102
103
104 /** Fragment shader number (for debugging) */
105 static unsigned fs_no = 0;
106
107
108 /**
109 * Expand the relevant bits of mask_input to a n*4-dword mask for the
110 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
111 * quad mask vector to 0 or ~0.
112 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
113 * quad arguments with fs length 8.
114 *
115 * \param first_quad which quad(s) of the quad group to test, in [0,3]
116 * \param mask_input bitwise mask for the whole 4x4 stamp
117 */
118 static LLVMValueRef
119 generate_quad_mask(struct gallivm_state *gallivm,
120 struct lp_type fs_type,
121 unsigned first_quad,
122 LLVMValueRef mask_input) /* int32 */
123 {
124 LLVMBuilderRef builder = gallivm->builder;
125 struct lp_type mask_type;
126 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
127 LLVMValueRef bits[16];
128 LLVMValueRef mask;
129 int shift, i;
130
131 /*
132 * XXX: We'll need a different path for 16 x u8
133 */
134 assert(fs_type.width == 32);
135 assert(fs_type.length <= Elements(bits));
136 mask_type = lp_int_type(fs_type);
137
138 /*
139 * mask_input >>= (quad * 4)
140 */
141 switch (first_quad) {
142 case 0:
143 shift = 0;
144 break;
145 case 1:
146 assert(fs_type.length == 4);
147 shift = 2;
148 break;
149 case 2:
150 shift = 8;
151 break;
152 case 3:
153 assert(fs_type.length == 4);
154 shift = 10;
155 break;
156 default:
157 assert(0);
158 shift = 0;
159 }
160
161 mask_input = LLVMBuildLShr(builder,
162 mask_input,
163 LLVMConstInt(i32t, shift, 0),
164 "");
165
166 /*
167 * mask = { mask_input & (1 << i), for i in [0,3] }
168 */
169 mask = lp_build_broadcast(gallivm,
170 lp_build_vec_type(gallivm, mask_type),
171 mask_input);
172
173 for (i = 0; i < fs_type.length / 4; i++) {
174 unsigned j = 2 * (i % 2) + (i / 2) * 8;
175 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
176 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
177 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
178 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
179 }
180 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
181
182 /*
183 * mask = mask != 0 ? ~0 : 0
184 */
185 mask = lp_build_compare(gallivm,
186 mask_type, PIPE_FUNC_NOTEQUAL,
187 mask,
188 lp_build_const_int_vec(gallivm, mask_type, 0));
189
190 return mask;
191 }
192
193
194 #define EARLY_DEPTH_TEST 0x1
195 #define LATE_DEPTH_TEST 0x2
196 #define EARLY_DEPTH_WRITE 0x4
197 #define LATE_DEPTH_WRITE 0x8
198
199 static int
200 find_output_by_semantic( const struct tgsi_shader_info *info,
201 unsigned semantic,
202 unsigned index )
203 {
204 int i;
205
206 for (i = 0; i < info->num_outputs; i++)
207 if (info->output_semantic_name[i] == semantic &&
208 info->output_semantic_index[i] == index)
209 return i;
210
211 return -1;
212 }
213
214
215 /**
216 * Generate the fragment shader, depth/stencil test, and alpha tests.
217 * \param i which quad in the tile, in range [0,3]
218 * \param partial_mask if 1, do mask_input testing
219 */
220 static void
221 generate_fs(struct gallivm_state *gallivm,
222 struct lp_fragment_shader *shader,
223 const struct lp_fragment_shader_variant_key *key,
224 LLVMBuilderRef builder,
225 struct lp_type type,
226 LLVMValueRef context_ptr,
227 unsigned i,
228 struct lp_build_interp_soa_context *interp,
229 struct lp_build_sampler_soa *sampler,
230 LLVMValueRef *pmask,
231 LLVMValueRef (*color)[4],
232 LLVMValueRef depth_ptr,
233 LLVMValueRef facing,
234 unsigned partial_mask,
235 LLVMValueRef mask_input,
236 LLVMValueRef counter)
237 {
238 const struct util_format_description *zs_format_desc = NULL;
239 const struct tgsi_token *tokens = shader->base.tokens;
240 LLVMTypeRef vec_type;
241 LLVMValueRef consts_ptr;
242 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
243 LLVMValueRef z;
244 LLVMValueRef zs_value = NULL;
245 LLVMValueRef stencil_refs[2];
246 struct lp_build_mask_context mask;
247 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
248 shader->info.base.num_inputs < 3 &&
249 shader->info.base.num_instructions < 8);
250 unsigned attrib;
251 unsigned chan;
252 unsigned cbuf;
253 unsigned depth_mode;
254 struct lp_bld_tgsi_system_values system_values;
255
256 memset(&system_values, 0, sizeof(system_values));
257
258 if (key->depth.enabled ||
259 key->stencil[0].enabled ||
260 key->stencil[1].enabled) {
261
262 zs_format_desc = util_format_description(key->zsbuf_format);
263 assert(zs_format_desc);
264
265 if (!shader->info.base.writes_z) {
266 if (key->alpha.enabled || shader->info.base.uses_kill)
267 /* With alpha test and kill, can do the depth test early
268 * and hopefully eliminate some quads. But need to do a
269 * special deferred depth write once the final mask value
270 * is known.
271 */
272 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
273 else
274 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
275 }
276 else {
277 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
278 }
279
280 if (!(key->depth.enabled && key->depth.writemask) &&
281 !(key->stencil[0].enabled && key->stencil[0].writemask))
282 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
283 }
284 else {
285 depth_mode = 0;
286 }
287
288 assert(i < 4);
289
290 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
291 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
292
293 vec_type = lp_build_vec_type(gallivm, type);
294
295 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
296
297 memset(outputs, 0, sizeof outputs);
298
299 /* Declare the color and z variables */
300 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
301 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
302 color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
303 }
304 }
305
306 /* do triangle edge testing */
307 if (partial_mask) {
308 *pmask = generate_quad_mask(gallivm, type,
309 i*type.length/4, mask_input);
310 }
311 else {
312 *pmask = lp_build_const_int_vec(gallivm, type, ~0);
313 }
314
315 /* 'mask' will control execution based on quad's pixel alive/killed state */
316 lp_build_mask_begin(&mask, gallivm, type, *pmask);
317
318 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
319 lp_build_mask_check(&mask);
320
321 lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
322 z = interp->pos[2];
323
324 if (depth_mode & EARLY_DEPTH_TEST) {
325 lp_build_depth_stencil_test(gallivm,
326 &key->depth,
327 key->stencil,
328 type,
329 zs_format_desc,
330 &mask,
331 stencil_refs,
332 z,
333 depth_ptr, facing,
334 &zs_value,
335 !simple_shader);
336
337 if (depth_mode & EARLY_DEPTH_WRITE) {
338 lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
339 }
340 }
341
342 lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
343
344 /* Build the actual shader */
345 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
346 consts_ptr, &system_values,
347 interp->pos, interp->inputs,
348 outputs, sampler, &shader->info.base);
349
350 /* Alpha test */
351 if (key->alpha.enabled) {
352 int color0 = find_output_by_semantic(&shader->info.base,
353 TGSI_SEMANTIC_COLOR,
354 0);
355
356 if (color0 != -1 && outputs[color0][3]) {
357 const struct util_format_description *cbuf_format_desc;
358 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
359 LLVMValueRef alpha_ref_value;
360
361 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
362 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
363
364 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
365
366 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
367 &mask, alpha, alpha_ref_value,
368 (depth_mode & LATE_DEPTH_TEST) != 0);
369 }
370 }
371
372 /* Late Z test */
373 if (depth_mode & LATE_DEPTH_TEST) {
374 int pos0 = find_output_by_semantic(&shader->info.base,
375 TGSI_SEMANTIC_POSITION,
376 0);
377
378 if (pos0 != -1 && outputs[pos0][2]) {
379 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
380 }
381
382 lp_build_depth_stencil_test(gallivm,
383 &key->depth,
384 key->stencil,
385 type,
386 zs_format_desc,
387 &mask,
388 stencil_refs,
389 z,
390 depth_ptr, facing,
391 &zs_value,
392 !simple_shader);
393 /* Late Z write */
394 if (depth_mode & LATE_DEPTH_WRITE) {
395 lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
396 }
397 }
398 else if ((depth_mode & EARLY_DEPTH_TEST) &&
399 (depth_mode & LATE_DEPTH_WRITE))
400 {
401 /* Need to apply a reduced mask to the depth write. Reload the
402 * depth value, update from zs_value with the new mask value and
403 * write that out.
404 */
405 lp_build_deferred_depth_write(gallivm,
406 type,
407 zs_format_desc,
408 &mask,
409 depth_ptr,
410 zs_value);
411 }
412
413
414 /* Color write */
415 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
416 {
417 if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
418 shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
419 {
420 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
421 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
422 if(outputs[attrib][chan]) {
423 /* XXX: just initialize outputs to point at colors[] and
424 * skip this.
425 */
426 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
427 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
428 LLVMBuildStore(builder, out, color[cbuf][chan]);
429 }
430 }
431 }
432 }
433
434 if (counter)
435 lp_build_occlusion_count(gallivm, type,
436 lp_build_mask_value(&mask), counter);
437
438 *pmask = lp_build_mask_end(&mask);
439 }
440
441
442 /**
443 * Generate the fragment shader, depth/stencil test, and alpha tests.
444 */
445 static void
446 generate_fs_loop(struct gallivm_state *gallivm,
447 struct lp_fragment_shader *shader,
448 const struct lp_fragment_shader_variant_key *key,
449 LLVMBuilderRef builder,
450 struct lp_type type,
451 LLVMValueRef context_ptr,
452 LLVMValueRef num_loop,
453 struct lp_build_interp_soa_context *interp,
454 struct lp_build_sampler_soa *sampler,
455 LLVMValueRef mask_store,
456 LLVMValueRef (*out_color)[4],
457 LLVMValueRef depth_ptr,
458 unsigned depth_bits,
459 LLVMValueRef facing,
460 LLVMValueRef counter)
461 {
462 const struct util_format_description *zs_format_desc = NULL;
463 const struct tgsi_token *tokens = shader->base.tokens;
464 LLVMTypeRef vec_type;
465 LLVMValueRef mask_ptr, mask_val;
466 LLVMValueRef consts_ptr;
467 LLVMValueRef z;
468 LLVMValueRef zs_value = NULL;
469 LLVMValueRef stencil_refs[2];
470 LLVMValueRef depth_ptr_i;
471 LLVMValueRef depth_offset;
472 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
473 struct lp_build_for_loop_state loop_state;
474 struct lp_build_mask_context mask;
475 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
476 shader->info.base.num_inputs < 3 &&
477 shader->info.base.num_instructions < 8);
478 unsigned attrib;
479 unsigned chan;
480 unsigned cbuf;
481 unsigned depth_mode;
482
483 struct lp_bld_tgsi_system_values system_values;
484
485 memset(&system_values, 0, sizeof(system_values));
486
487 if (key->depth.enabled ||
488 key->stencil[0].enabled ||
489 key->stencil[1].enabled) {
490
491 zs_format_desc = util_format_description(key->zsbuf_format);
492 assert(zs_format_desc);
493
494 if (!shader->info.base.writes_z) {
495 if (key->alpha.enabled || shader->info.base.uses_kill)
496 /* With alpha test and kill, can do the depth test early
497 * and hopefully eliminate some quads. But need to do a
498 * special deferred depth write once the final mask value
499 * is known.
500 */
501 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
502 else
503 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
504 }
505 else {
506 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
507 }
508
509 if (!(key->depth.enabled && key->depth.writemask) &&
510 !(key->stencil[0].enabled && key->stencil[0].writemask))
511 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
512 }
513 else {
514 depth_mode = 0;
515 }
516
517
518 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
519 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
520
521 vec_type = lp_build_vec_type(gallivm, type);
522
523 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
524
525 lp_build_for_loop_begin(&loop_state, gallivm,
526 lp_build_const_int32(gallivm, 0),
527 LLVMIntULT,
528 num_loop,
529 lp_build_const_int32(gallivm, 1));
530
531 mask_ptr = LLVMBuildGEP(builder, mask_store,
532 &loop_state.counter, 1, "mask_ptr");
533 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
534
535 depth_offset = LLVMBuildMul(builder, loop_state.counter,
536 lp_build_const_int32(gallivm, depth_bits * type.length),
537 "");
538
539 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
540
541 memset(outputs, 0, sizeof outputs);
542
543 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
544 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
545 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
546 lp_build_vec_type(gallivm,
547 type),
548 num_loop, "color");
549 }
550 }
551
552
553
554 /* 'mask' will control execution based on quad's pixel alive/killed state */
555 lp_build_mask_begin(&mask, gallivm, type, mask_val);
556
557 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
558 lp_build_mask_check(&mask);
559
560 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
561 z = interp->pos[2];
562
563 if (depth_mode & EARLY_DEPTH_TEST) {
564 lp_build_depth_stencil_test(gallivm,
565 &key->depth,
566 key->stencil,
567 type,
568 zs_format_desc,
569 &mask,
570 stencil_refs,
571 z,
572 depth_ptr_i, facing,
573 &zs_value,
574 !simple_shader);
575
576 if (depth_mode & EARLY_DEPTH_WRITE) {
577 lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
578 }
579 }
580
581 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
582
583 /* Build the actual shader */
584 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
585 consts_ptr, &system_values,
586 interp->pos, interp->inputs,
587 outputs, sampler, &shader->info.base);
588
589 /* Alpha test */
590 if (key->alpha.enabled) {
591 int color0 = find_output_by_semantic(&shader->info.base,
592 TGSI_SEMANTIC_COLOR,
593 0);
594
595 if (color0 != -1 && outputs[color0][3]) {
596 const struct util_format_description *cbuf_format_desc;
597 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
598 LLVMValueRef alpha_ref_value;
599
600 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
601 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
602
603 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
604
605 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
606 &mask, alpha, alpha_ref_value,
607 (depth_mode & LATE_DEPTH_TEST) != 0);
608 }
609 }
610
611 /* Late Z test */
612 if (depth_mode & LATE_DEPTH_TEST) {
613 int pos0 = find_output_by_semantic(&shader->info.base,
614 TGSI_SEMANTIC_POSITION,
615 0);
616
617 if (pos0 != -1 && outputs[pos0][2]) {
618 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
619 }
620
621 lp_build_depth_stencil_test(gallivm,
622 &key->depth,
623 key->stencil,
624 type,
625 zs_format_desc,
626 &mask,
627 stencil_refs,
628 z,
629 depth_ptr_i, facing,
630 &zs_value,
631 !simple_shader);
632 /* Late Z write */
633 if (depth_mode & LATE_DEPTH_WRITE) {
634 lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
635 }
636 }
637 else if ((depth_mode & EARLY_DEPTH_TEST) &&
638 (depth_mode & LATE_DEPTH_WRITE))
639 {
640 /* Need to apply a reduced mask to the depth write. Reload the
641 * depth value, update from zs_value with the new mask value and
642 * write that out.
643 */
644 lp_build_deferred_depth_write(gallivm,
645 type,
646 zs_format_desc,
647 &mask,
648 depth_ptr_i,
649 zs_value);
650 }
651
652
653 /* Color write */
654 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
655 {
656 if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
657 shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
658 {
659 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
660 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
661 if(outputs[attrib][chan]) {
662 /* XXX: just initialize outputs to point at colors[] and
663 * skip this.
664 */
665 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
666 LLVMValueRef color_ptr;
667 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
668 &loop_state.counter, 1, "");
669 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
670 LLVMBuildStore(builder, out, color_ptr);
671 }
672 }
673 }
674 }
675
676 if (key->occlusion_count) {
677 lp_build_name(counter, "counter");
678 lp_build_occlusion_count(gallivm, type,
679 lp_build_mask_value(&mask), counter);
680 }
681
682 mask_val = lp_build_mask_end(&mask);
683 LLVMBuildStore(builder, mask_val, mask_ptr);
684 lp_build_for_loop_end(&loop_state);
685 }
686
687
688 /**
689 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
690 *
691 * Fragment Shader outputs pixels in small 2x2 blocks
692 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
693 *
694 * However in memory pixels are stored in rows
695 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
696 *
697 * @param type fragment shader type (4x or 8x float)
698 * @param num_fs number of fs_src
699 * @param dst_channels number of output channels
700 * @param fs_src output from fragment shader
701 * @param dst pointer to store result
702 * @param pad_inline is channel padding inline or at end of row
703 * @return the number of dsts
704 */
705 static int
706 generate_fs_twiddle(struct gallivm_state *gallivm,
707 struct lp_type type,
708 unsigned num_fs,
709 unsigned dst_channels,
710 LLVMValueRef fs_src[][4],
711 LLVMValueRef* dst,
712 bool pad_inline)
713 {
714 LLVMValueRef src[16];
715
716 bool swizzle_pad;
717 bool twiddle;
718 bool split;
719
720 unsigned pixels = num_fs == 4 ? 1 : 2;
721 unsigned reorder_group;
722 unsigned src_channels;
723 unsigned src_count;
724 unsigned i;
725
726 src_channels = dst_channels < 3 ? dst_channels : 4;
727 src_count = num_fs * src_channels;
728
729 assert(pixels == 2 || num_fs == 4);
730 assert(num_fs * src_channels <= Elements(src));
731
732 /*
733 * Transpose from SoA -> AoS
734 */
735 for (i = 0; i < num_fs; ++i) {
736 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
737 }
738
739 /*
740 * Pick transformation options
741 */
742 swizzle_pad = false;
743 twiddle = false;
744 split = false;
745 reorder_group = 0;
746
747 if (dst_channels == 1) {
748 twiddle = true;
749
750 if (pixels == 2) {
751 split = true;
752 }
753 } else if (dst_channels == 2) {
754 if (pixels == 1) {
755 reorder_group = 1;
756 }
757 } else if (dst_channels > 2) {
758 if (pixels == 1) {
759 reorder_group = 2;
760 } else {
761 twiddle = true;
762 }
763
764 if (!pad_inline && dst_channels == 3 && pixels > 1) {
765 swizzle_pad = true;
766 }
767 }
768
769 /*
770 * Split the src in half
771 */
772 if (split) {
773 for (i = num_fs; i > 0; --i) {
774 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
775 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
776 }
777
778 src_count *= 2;
779 type.length = 4;
780 }
781
782 /*
783 * Ensure pixels are in memory order
784 */
785 if (reorder_group) {
786 /* Twiddle pixels by reordering the array, e.g.:
787 *
788 * src_count = 8 -> 0 2 1 3 4 6 5 7
789 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
790 */
791 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
792
793 for (i = 0; i < src_count; ++i) {
794 unsigned group = i / reorder_group;
795 unsigned block = (group / 4) * 4 * reorder_group;
796 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
797 dst[i] = src[j];
798 }
799 } else if (twiddle) {
800 /* Twiddle pixels across elements of array */
801 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
802 } else {
803 /* Do nothing */
804 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
805 }
806
807 /*
808 * Moves any padding between pixels to the end
809 * e.g. RGBXRGBX -> RGBRGBXX
810 */
811 if (swizzle_pad) {
812 unsigned char swizzles[16];
813 unsigned elems = pixels * dst_channels;
814
815 for (i = 0; i < type.length; ++i) {
816 if (i < elems)
817 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
818 else
819 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
820 }
821
822 for (i = 0; i < src_count; ++i) {
823 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
824 }
825 }
826
827 return src_count;
828 }
829
830
831 /**
832 * Load an unswizzled block of pixels from memory
833 */
834 static void
835 load_unswizzled_block(struct gallivm_state *gallivm,
836 LLVMValueRef base_ptr,
837 LLVMValueRef stride,
838 unsigned block_width,
839 unsigned block_height,
840 LLVMValueRef* dst,
841 struct lp_type dst_type,
842 unsigned dst_count)
843 {
844 LLVMBuilderRef builder = gallivm->builder;
845 unsigned row_size = dst_count / block_height;
846 unsigned i;
847
848 /* Ensure block exactly fits into dst */
849 assert((block_width * block_height) % dst_count == 0);
850
851 for (i = 0; i < dst_count; ++i) {
852 unsigned x = i % row_size;
853 unsigned y = i / row_size;
854
855 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
856 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
857
858 LLVMValueRef gep[2];
859 LLVMValueRef dst_ptr;
860
861 gep[0] = lp_build_const_int32(gallivm, 0);
862 gep[1] = LLVMBuildAdd(builder, bx, by, "");
863
864 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
865 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
866
867 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
868
869 if ((dst_type.length % 3) == 0) {
870 lp_set_load_alignment(dst[i], dst_type.width / 8);
871 }
872 }
873 }
874
875
876 /**
877 * Store an unswizzled block of pixels to memory
878 */
879 static void
880 store_unswizzled_block(struct gallivm_state *gallivm,
881 LLVMValueRef base_ptr,
882 LLVMValueRef stride,
883 unsigned block_width,
884 unsigned block_height,
885 LLVMValueRef* src,
886 struct lp_type src_type,
887 unsigned src_count)
888 {
889 LLVMBuilderRef builder = gallivm->builder;
890 unsigned row_size = src_count / block_height;
891 unsigned i;
892
893 /* Ensure src exactly fits into block */
894 assert((block_width * block_height) % src_count == 0);
895
896 for (i = 0; i < src_count; ++i) {
897 unsigned x = i % row_size;
898 unsigned y = i / row_size;
899
900 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
901 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
902
903 LLVMValueRef gep[2];
904 LLVMValueRef src_ptr;
905
906 gep[0] = lp_build_const_int32(gallivm, 0);
907 gep[1] = LLVMBuildAdd(builder, bx, by, "");
908
909 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
910 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
911
912 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
913
914 if ((src_type.length % 3) == 0) {
915 lp_set_store_alignment(src_ptr, src_type.width / 8);
916 }
917 }
918 }
919
920
921 /**
922 * Checks if a format description is an arithmetic format
923 *
924 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
925 */
926 static INLINE boolean
927 is_arithmetic_format(const struct util_format_description *format_desc)
928 {
929 boolean arith = false;
930 unsigned i;
931
932 for (i = 0; i < format_desc->nr_channels; ++i) {
933 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
934 arith |= (format_desc->channel[i].size % 8) != 0;
935 }
936
937 return arith;
938 }
939
940
941 /**
942 * Retrieves the type representing the memory layout for a format
943 *
944 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
945 */
946 static INLINE void
947 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
948 struct lp_type* type)
949 {
950 int i;
951
952 memset(type, 0, sizeof(struct lp_type));
953 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
954 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
955 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
956 type->norm = format_desc->channel[0].normalized;
957
958 if (is_arithmetic_format(format_desc)) {
959 type->width = 0;
960 type->length = 1;
961
962 for (i = 0; i < format_desc->nr_channels; ++i) {
963 type->width += format_desc->channel[i].size;
964 }
965 } else {
966 type->width = format_desc->channel[0].size;
967 type->length = format_desc->nr_channels;
968 }
969 }
970
971
972 /**
973 * Retrieves the type for a format which is usable in the blending code.
974 *
975 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
976 */
977 static INLINE void
978 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
979 struct lp_type* type)
980 {
981 int i;
982
983 memset(type, 0, sizeof(struct lp_type));
984 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
985 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
986 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
987 type->norm = format_desc->channel[0].normalized;
988 type->width = format_desc->channel[0].size;
989 type->length = format_desc->nr_channels;
990
991 for (i = 1; i < format_desc->nr_channels; ++i) {
992 if (format_desc->channel[i].size > type->width)
993 type->width = format_desc->channel[i].size;
994 }
995
996 if (type->floating) {
997 type->width = 32;
998 } else {
999 if (type->width <= 8) {
1000 type->width = 8;
1001 } else if (type->width <= 16) {
1002 type->width = 16;
1003 } else {
1004 type->width = 32;
1005 }
1006 }
1007
1008 if (is_arithmetic_format(format_desc) && type->length == 3) {
1009 type->length = 4;
1010 }
1011 }
1012
1013
1014 /**
1015 * Scale a normalised value from src_bits to dst_bits
1016 */
1017 static INLINE LLVMValueRef
1018 scale_bits(struct gallivm_state *gallivm,
1019 int src_bits,
1020 int dst_bits,
1021 LLVMValueRef src,
1022 struct lp_type src_type)
1023 {
1024 LLVMBuilderRef builder = gallivm->builder;
1025 LLVMValueRef result = src;
1026
1027 if (dst_bits < src_bits) {
1028 /* Scale down by LShr */
1029 result = LLVMBuildLShr(builder,
1030 src,
1031 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
1032 "");
1033 } else if (dst_bits > src_bits) {
1034 /* Scale up bits */
1035 int db = dst_bits - src_bits;
1036
1037 /* Shift left by difference in bits */
1038 result = LLVMBuildShl(builder,
1039 src,
1040 lp_build_const_int_vec(gallivm, src_type, db),
1041 "");
1042
1043 if (db < src_bits) {
1044 /* Enough bits in src to fill the remainder */
1045 LLVMValueRef lower = LLVMBuildLShr(builder,
1046 src,
1047 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1048 "");
1049
1050 result = LLVMBuildOr(builder, result, lower, "");
1051 } else if (db > src_bits) {
1052 /* Need to repeatedely copy src bits to fill remainder in dst */
1053 unsigned n;
1054
1055 for (n = src_bits; n < dst_bits; n *= 2) {
1056 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1057
1058 result = LLVMBuildOr(builder,
1059 result,
1060 LLVMBuildLShr(builder, result, shuv, ""),
1061 "");
1062 }
1063 }
1064 }
1065
1066 return result;
1067 }
1068
1069
1070 /**
1071 * Convert from memory format to blending format
1072 *
1073 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1074 */
1075 static void
1076 convert_to_blend_type(struct gallivm_state *gallivm,
1077 const struct util_format_description *src_fmt,
1078 struct lp_type src_type,
1079 struct lp_type dst_type,
1080 LLVMValueRef* src,
1081 unsigned num_srcs,
1082 LLVMValueRef* dst)
1083 {
1084 LLVMBuilderRef builder = gallivm->builder;
1085 struct lp_type blend_type;
1086 struct lp_type mem_type;
1087 unsigned i, j, k;
1088 unsigned pixels = 16 / num_srcs;
1089 bool is_arith;
1090
1091 memcpy(dst, src, sizeof(LLVMValueRef*) * num_srcs);
1092
1093 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1094 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1095
1096 /* Is the format arithmetic */
1097 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1098 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1099
1100 /* Pad if necessary */
1101 if (!is_arith && src_type.length < dst_type.length) {
1102 for (i = 0; i < num_srcs; ++i) {
1103 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1104 }
1105
1106 src_type.length = dst_type.length;
1107 }
1108
1109 /* Special case for half-floats */
1110 if (mem_type.width == 16 && mem_type.floating) {
1111 assert(blend_type.width == 32 && blend_type.floating);
1112 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1113 is_arith = false;
1114 }
1115
1116 if (!is_arith) {
1117 return;
1118 }
1119
1120 src_type.width = blend_type.width * blend_type.length;
1121 blend_type.length *= pixels;
1122 src_type.length *= pixels / (src_type.length / mem_type.length);
1123
1124 for (i = 0; i < num_srcs; ++i) {
1125 LLVMValueRef chans[4];
1126 LLVMValueRef res;
1127 unsigned sa = 0;
1128
1129 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1130
1131 for (j = 0; j < src_fmt->nr_channels; ++j) {
1132 unsigned mask = 0;
1133
1134 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1135 mask |= 1 << k;
1136 }
1137
1138 /* Extract bits from source */
1139 chans[j] = LLVMBuildLShr(builder,
1140 dst[i],
1141 lp_build_const_int_vec(gallivm, src_type, sa),
1142 "");
1143
1144 chans[j] = LLVMBuildAnd(builder,
1145 chans[j],
1146 lp_build_const_int_vec(gallivm, src_type, mask),
1147 "");
1148
1149 /* Scale bits */
1150 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size, blend_type.width, chans[j], src_type);
1151
1152 /* Insert bits into correct position */
1153 chans[j] = LLVMBuildShl(builder,
1154 chans[j],
1155 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1156 "");
1157
1158 sa += src_fmt->channel[j].size;
1159
1160 if (j == 0) {
1161 res = chans[j];
1162 } else {
1163 res = LLVMBuildOr(builder, res, chans[j], "");
1164 }
1165 }
1166
1167 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1168 }
1169 }
1170
1171
1172 /**
1173 * Convert from blending format to memory format
1174 *
1175 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1176 */
1177 static void
1178 convert_from_blend_type(struct gallivm_state *gallivm,
1179 const struct util_format_description *src_fmt,
1180 struct lp_type src_type,
1181 struct lp_type dst_type,
1182 LLVMValueRef* src,
1183 unsigned num_srcs,
1184 LLVMValueRef* dst)
1185 {
1186 unsigned i, j, k;
1187 struct lp_type mem_type;
1188 struct lp_type blend_type;
1189 LLVMBuilderRef builder = gallivm->builder;
1190 unsigned pixels = 16 / num_srcs;
1191 bool is_arith;
1192
1193 memcpy(dst, src, sizeof(LLVMValueRef*) * num_srcs);
1194
1195 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1196 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1197
1198 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1199
1200 /* Special case for half-floats */
1201 if (mem_type.width == 16 && mem_type.floating) {
1202 int length = dst_type.length;
1203 assert(blend_type.width == 32 && blend_type.floating);
1204
1205 dst_type.length = src_type.length;
1206
1207 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1208
1209 dst_type.length = length;
1210 is_arith = false;
1211 }
1212
1213 /* Remove any padding */
1214 if (!is_arith && (src_type.length % mem_type.length)) {
1215 src_type.length -= (src_type.length % mem_type.length);
1216
1217 for (i = 0; i < num_srcs; ++i) {
1218 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1219 }
1220 }
1221
1222 /* No bit arithmitic to do */
1223 if (!is_arith) {
1224 return;
1225 }
1226
1227 src_type.length = pixels;
1228 src_type.width = blend_type.length * blend_type.width;
1229 dst_type.length = pixels;
1230
1231 for (i = 0; i < num_srcs; ++i) {
1232 LLVMValueRef chans[4];
1233 LLVMValueRef res;
1234 unsigned sa = 0;
1235
1236 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1237
1238 for (j = 0; j < src_fmt->nr_channels; ++j) {
1239 unsigned mask = 0;
1240
1241 assert(blend_type.width > src_fmt->channel[j].size);
1242
1243 for (k = 0; k < blend_type.width; ++k) {
1244 mask |= 1 << k;
1245 }
1246
1247 /* Extract bits */
1248 chans[j] = LLVMBuildLShr(builder,
1249 dst[i],
1250 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1251 "");
1252
1253 chans[j] = LLVMBuildAnd(builder,
1254 chans[j],
1255 lp_build_const_int_vec(gallivm, src_type, mask),
1256 "");
1257
1258 /* Scale down bits */
1259 chans[j] = scale_bits(gallivm, blend_type.width, src_fmt->channel[j].size, chans[j], src_type);
1260
1261 /* Insert bits */
1262 chans[j] = LLVMBuildShl(builder,
1263 chans[j],
1264 lp_build_const_int_vec(gallivm, src_type, sa),
1265 "");
1266
1267 sa += src_fmt->channel[j].size;
1268
1269 if (j == 0) {
1270 res = chans[j];
1271 } else {
1272 res = LLVMBuildOr(builder, res, chans[j], "");
1273 }
1274 }
1275
1276 assert (dst_type.width != 24);
1277
1278 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1279 }
1280 }
1281
1282
1283 /**
1284 * Generates the blend function for unswizzled colour buffers
1285 * Also generates the read & write from colour buffer
1286 */
1287 static void
1288 generate_unswizzled_blend(struct gallivm_state *gallivm,
1289 unsigned rt,
1290 struct lp_fragment_shader_variant *variant,
1291 enum pipe_format out_format,
1292 unsigned int num_fs,
1293 struct lp_type fs_type,
1294 LLVMValueRef* fs_mask,
1295 LLVMValueRef fs_out_color[TGSI_NUM_CHANNELS][4],
1296 LLVMValueRef context_ptr,
1297 LLVMValueRef color_ptr,
1298 LLVMValueRef stride,
1299 unsigned partial_mask,
1300 boolean do_branch)
1301 {
1302 const unsigned alpha_channel = 3;
1303 const unsigned block_width = 4;
1304 const unsigned block_height = 4;
1305 const unsigned block_size = block_width * block_height;
1306 const unsigned lp_integer_vector_width = 128;
1307
1308 LLVMBuilderRef builder = gallivm->builder;
1309 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1310 LLVMValueRef src_alpha[4 * 4];
1311 LLVMValueRef src_mask[4 * 4];
1312 LLVMValueRef src[4 * 4];
1313 LLVMValueRef dst[4 * 4];
1314 LLVMValueRef blend_color;
1315 LLVMValueRef blend_alpha;
1316 LLVMValueRef i32_zero;
1317 LLVMValueRef check_mask;
1318
1319 struct lp_build_mask_context mask_ctx;
1320 struct lp_type mask_type;
1321 struct lp_type blend_type;
1322 struct lp_type alpha_type;
1323 struct lp_type row_type;
1324 struct lp_type dst_type;
1325
1326 unsigned char swizzle[TGSI_NUM_CHANNELS];
1327 unsigned vector_width;
1328 unsigned dst_channels;
1329 unsigned src_channels;
1330 unsigned dst_count;
1331 unsigned src_count;
1332 unsigned i, j;
1333
1334 const struct util_format_description* out_format_desc = util_format_description(out_format);
1335
1336 bool pad_inline = is_arithmetic_format(out_format_desc);
1337 bool has_alpha = false;
1338
1339 src_channels = TGSI_NUM_CHANNELS;
1340 mask_type = lp_int32_vec4_type();
1341 mask_type.length = fs_type.length;
1342
1343 /* Do not bother executing code when mask is empty.. */
1344 if (do_branch) {
1345 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1346
1347 for (i = 0; i < num_fs; ++i) {
1348 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1349 }
1350
1351 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1352 lp_build_mask_check(&mask_ctx);
1353 }
1354
1355 partial_mask |= !variant->opaque;
1356 i32_zero = lp_build_const_int32(gallivm, 0);
1357
1358 /* Get type from output format */
1359 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1360 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1361
1362 row_type.length = fs_type.length;
1363 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1364
1365 /* Compute correct swizzle and count channels */
1366 memset(swizzle, 0xFF, TGSI_NUM_CHANNELS);
1367 dst_channels = 0;
1368
1369 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1370 /* Ensure channel is used */
1371 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1372 continue;
1373 }
1374
1375 /* Ensure not already written to (happens in case with GL_ALPHA) */
1376 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1377 continue;
1378 }
1379
1380 /* Ensure we havn't already found all channels */
1381 if (dst_channels >= out_format_desc->nr_channels) {
1382 continue;
1383 }
1384
1385 swizzle[out_format_desc->swizzle[i]] = i;
1386 ++dst_channels;
1387
1388 if (i == alpha_channel) {
1389 has_alpha = true;
1390 }
1391 }
1392
1393 /* If 3 channels then pad to include alpha for 4 element transpose */
1394 if (dst_channels == 3 && !has_alpha) {
1395 swizzle[3] = 3;
1396
1397 if (out_format_desc->nr_channels == 4) {
1398 dst_channels = 4;
1399 }
1400 }
1401
1402 /*
1403 * Load shader output
1404 */
1405 for (i = 0; i < num_fs; ++i) {
1406 /* Always load alpha for use in blending */
1407 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[alpha_channel][i], "");
1408
1409 /* Load each channel */
1410 for (j = 0; j < dst_channels; ++j) {
1411 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[swizzle[j]][i], "");
1412 }
1413
1414 /* If 3 channels then pad to include alpha for 4 element transpose */
1415 if (dst_channels == 3 && !has_alpha) {
1416 fs_src[i][3] = alpha;
1417 swizzle[3] = 3;
1418 }
1419
1420 /* We split the row_mask and row_alpha as we want 128bit interleave */
1421 if (fs_type.length == 8) {
1422 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1423 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1424
1425 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1426 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1427 } else {
1428 src_mask[i] = fs_mask[i];
1429 src_alpha[i] = alpha;
1430 }
1431 }
1432
1433
1434 /*
1435 * Pixel twiddle from fragment shader order to memory order
1436 */
1437 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1438 src_channels = dst_channels < 3 ? dst_channels : 4;
1439 if (src_count != num_fs * src_channels) {
1440 unsigned ds = src_count / (num_fs * src_channels);
1441 row_type.length /= ds;
1442 fs_type.length = row_type.length;
1443 }
1444
1445 blend_type = row_type;
1446 alpha_type = fs_type;
1447 alpha_type.length = 4;
1448 mask_type.length = 4;
1449
1450 /* Convert src to row_type */
1451 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1452
1453 /* If the rows are not an SSE vector, combine them to become SSE size! */
1454 if ((row_type.width * row_type.length) % 128) {
1455 unsigned bits = row_type.width * row_type.length;
1456 unsigned combined;
1457
1458 dst_count = src_count / (vector_width / bits);
1459 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1460
1461 row_type.length *= combined;
1462 src_count /= combined;
1463
1464 bits = row_type.width * row_type.length;
1465 assert(bits == 128 || bits == 256);
1466 }
1467
1468
1469 /*
1470 * Blend Colour conversion
1471 */
1472 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1473 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1474 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1475
1476 /* Convert */
1477 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1478
1479 /* Extract alpha */
1480 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1481
1482 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1483 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1484 if (pad_inline) {
1485 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1486 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1487 } else {
1488 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1489 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1490 }
1491
1492 /*
1493 * Mask conversion
1494 */
1495 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1496
1497 if (src_count < block_height) {
1498 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1499 } else if (src_count > block_height) {
1500 for (i = src_count; i > 0; --i) {
1501 unsigned pixels = block_size / src_count;
1502 unsigned idx = i - 1;
1503
1504 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1505 }
1506 }
1507
1508 assert(mask_type.width == 32);
1509
1510 for (i = 0; i < src_count; ++i) {
1511 unsigned pixels = block_size / src_count;
1512 unsigned pixel_width = row_type.width * dst_channels;
1513
1514 if (pixel_width == 24) {
1515 mask_type.width = 8;
1516 mask_type.length = vector_width / mask_type.width;
1517 } else {
1518 mask_type.length = pixels;
1519 mask_type.width = row_type.width * dst_channels;
1520
1521 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1522
1523 mask_type.length *= dst_channels;
1524 mask_type.width /= dst_channels;
1525 }
1526
1527 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1528 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1529 }
1530
1531 /*
1532 * Alpha conversion
1533 */
1534 if (!has_alpha) {
1535 unsigned length = row_type.length;
1536 row_type.length = alpha_type.length;
1537
1538 /* Twiddle the alpha to match pixels */
1539 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1540
1541 for (i = 0; i < 4; ++i) {
1542 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1543 }
1544
1545 alpha_type = row_type;
1546 row_type.length = length;
1547
1548 /* If only one channel we can only need the single alpha value per pixel */
1549 if (src_count == 1) {
1550 assert(dst_channels == 1);
1551
1552 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1553 } else {
1554 /* If there are more srcs than rows then we need to split alpha up */
1555 if (src_count > block_height) {
1556 for (i = src_count; i > 0; --i) {
1557 unsigned pixels = block_size / src_count;
1558 unsigned idx = i - 1;
1559
1560 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1561 }
1562 }
1563
1564 /* If there is a src for each pixel broadcast the alpha across whole row */
1565 if (src_count == block_size) {
1566 for (i = 0; i < src_count; ++i) {
1567 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1568 }
1569 } else {
1570 unsigned pixels = block_size / src_count;
1571 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1572 unsigned alpha_span = 1;
1573
1574 /* Check if we need 2 src_alphas for our shuffles */
1575 if (pixels > alpha_type.length) {
1576 alpha_span = 2;
1577 }
1578
1579 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1580 for (i = 0; i < src_count; ++i) {
1581 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1582 unsigned idx1 = i, idx2 = i;
1583
1584 if (alpha_span > 1){
1585 idx1 *= alpha_span;
1586 idx2 = idx1 + 1;
1587 }
1588
1589 for (j = 0; j < row_type.length; ++j) {
1590 if (j < pixels * channels) {
1591 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1592 } else {
1593 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1594 }
1595 }
1596
1597 src_alpha[i] = LLVMBuildShuffleVector(builder,
1598 src_alpha[idx1],
1599 src_alpha[idx2],
1600 LLVMConstVector(shuffles, row_type.length),
1601 "");
1602 }
1603 }
1604 }
1605 }
1606
1607
1608 /*
1609 * Load dst from memory
1610 */
1611 if (src_count < block_height) {
1612 dst_count = block_height;
1613 } else {
1614 dst_count = src_count;
1615 }
1616
1617 dst_type.length *= 16 / dst_count;
1618
1619 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height, dst, dst_type, dst_count);
1620
1621
1622 /*
1623 * Convert from dst/output format to src/blending format.
1624 *
1625 * This is necessary as we can only read 1 row from memory at a time,
1626 * so the minimum dst_count will ever be at this point is 4.
1627 *
1628 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1629 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1630 * on all 16 pixels in that single vector at once.
1631 */
1632 if (dst_count > src_count) {
1633 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1634 }
1635
1636 /*
1637 * Blending
1638 */
1639 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count, dst);
1640
1641 for (i = 0; i < src_count; ++i) {
1642 dst[i] = lp_build_blend_aos(gallivm,
1643 &variant->key.blend,
1644 variant->key.cbuf_format,
1645 row_type,
1646 rt,
1647 src[i],
1648 has_alpha ? NULL : src_alpha[i],
1649 dst[i],
1650 partial_mask ? src_mask[i] : NULL,
1651 blend_color,
1652 has_alpha ? NULL : blend_alpha,
1653 swizzle,
1654 pad_inline ? 4 : dst_channels);
1655 }
1656
1657 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count, dst);
1658
1659 /* Split the blend rows back to memory rows */
1660 if (dst_count > src_count) {
1661 row_type.length = dst_type.length * (dst_count / src_count);
1662
1663 if (src_count == 1) {
1664 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1665 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1666
1667 row_type.length /= 2;
1668 src_count *= 2;
1669 }
1670
1671 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1672 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1673 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1674 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1675
1676 row_type.length /= 2;
1677 src_count *= 2;
1678 }
1679
1680
1681 /*
1682 * Store blend result to memory
1683 */
1684 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height, dst, dst_type, dst_count);
1685
1686 if (do_branch) {
1687 lp_build_mask_end(&mask_ctx);
1688 }
1689 }
1690
1691
1692 /**
1693 * Generate the runtime callable function for the whole fragment pipeline.
1694 * Note that the function which we generate operates on a block of 16
1695 * pixels at at time. The block contains 2x2 quads. Each quad contains
1696 * 2x2 pixels.
1697 */
1698 static void
1699 generate_fragment(struct llvmpipe_context *lp,
1700 struct lp_fragment_shader *shader,
1701 struct lp_fragment_shader_variant *variant,
1702 unsigned partial_mask)
1703 {
1704 struct gallivm_state *gallivm = variant->gallivm;
1705 const struct lp_fragment_shader_variant_key *key = &variant->key;
1706 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1707 char func_name[256];
1708 struct lp_type fs_type;
1709 struct lp_type blend_type;
1710 LLVMTypeRef fs_elem_type;
1711 LLVMTypeRef blend_vec_type;
1712 LLVMTypeRef arg_types[12];
1713 LLVMTypeRef func_type;
1714 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1715 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1716 LLVMValueRef context_ptr;
1717 LLVMValueRef x;
1718 LLVMValueRef y;
1719 LLVMValueRef a0_ptr;
1720 LLVMValueRef dadx_ptr;
1721 LLVMValueRef dady_ptr;
1722 LLVMValueRef color_ptr_ptr;
1723 LLVMValueRef stride_ptr;
1724 LLVMValueRef depth_ptr;
1725 LLVMValueRef mask_input;
1726 LLVMValueRef counter = NULL;
1727 LLVMBasicBlockRef block;
1728 LLVMBuilderRef builder;
1729 struct lp_build_sampler_soa *sampler;
1730 struct lp_build_interp_soa_context interp;
1731 LLVMValueRef fs_mask[16 / 4];
1732 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1733 LLVMValueRef function;
1734 LLVMValueRef facing;
1735 const struct util_format_description *zs_format_desc;
1736 unsigned num_fs;
1737 unsigned i;
1738 unsigned chan;
1739 unsigned cbuf;
1740 boolean cbuf0_write_all;
1741 boolean try_loop = TRUE;
1742
1743 assert(lp_native_vector_width / 32 >= 4);
1744
1745 /* Adjust color input interpolation according to flatshade state:
1746 */
1747 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1748 for (i = 0; i < shader->info.base.num_inputs; i++) {
1749 if (inputs[i].interp == LP_INTERP_COLOR) {
1750 if (key->flatshade)
1751 inputs[i].interp = LP_INTERP_CONSTANT;
1752 else
1753 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1754 }
1755 }
1756
1757 /* check if writes to cbuf[0] are to be copied to all cbufs */
1758 cbuf0_write_all = FALSE;
1759 for (i = 0;i < shader->info.base.num_properties; i++) {
1760 if (shader->info.base.properties[i].name ==
1761 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1762 cbuf0_write_all = TRUE;
1763 break;
1764 }
1765 }
1766
1767 /* TODO: actually pick these based on the fs and color buffer
1768 * characteristics. */
1769
1770 memset(&fs_type, 0, sizeof fs_type);
1771 fs_type.floating = TRUE; /* floating point values */
1772 fs_type.sign = TRUE; /* values are signed */
1773 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1774 fs_type.width = 32; /* 32-bit float */
1775 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1776 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1777
1778 memset(&blend_type, 0, sizeof blend_type);
1779 blend_type.floating = FALSE; /* values are integers */
1780 blend_type.sign = FALSE; /* values are unsigned */
1781 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1782 blend_type.width = 8; /* 8-bit ubyte values */
1783 blend_type.length = 16; /* 16 elements per vector */
1784
1785 /*
1786 * Generate the function prototype. Any change here must be reflected in
1787 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1788 */
1789
1790 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1791
1792 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1793
1794 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1795 shader->no, variant->no, partial_mask ? "partial" : "whole");
1796
1797 arg_types[0] = variant->jit_context_ptr_type; /* context */
1798 arg_types[1] = int32_type; /* x */
1799 arg_types[2] = int32_type; /* y */
1800 arg_types[3] = int32_type; /* facing */
1801 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1802 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1803 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1804 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1805 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1806 arg_types[9] = int32_type; /* mask_input */
1807 arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */
1808 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1809
1810 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1811 arg_types, Elements(arg_types), 0);
1812
1813 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1814 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1815
1816 variant->function[partial_mask] = function;
1817
1818 /* XXX: need to propagate noalias down into color param now we are
1819 * passing a pointer-to-pointer?
1820 */
1821 for(i = 0; i < Elements(arg_types); ++i)
1822 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1823 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1824
1825 context_ptr = LLVMGetParam(function, 0);
1826 x = LLVMGetParam(function, 1);
1827 y = LLVMGetParam(function, 2);
1828 facing = LLVMGetParam(function, 3);
1829 a0_ptr = LLVMGetParam(function, 4);
1830 dadx_ptr = LLVMGetParam(function, 5);
1831 dady_ptr = LLVMGetParam(function, 6);
1832 color_ptr_ptr = LLVMGetParam(function, 7);
1833 depth_ptr = LLVMGetParam(function, 8);
1834 mask_input = LLVMGetParam(function, 9);
1835 stride_ptr = LLVMGetParam(function, 11);
1836
1837 lp_build_name(context_ptr, "context");
1838 lp_build_name(x, "x");
1839 lp_build_name(y, "y");
1840 lp_build_name(a0_ptr, "a0");
1841 lp_build_name(dadx_ptr, "dadx");
1842 lp_build_name(dady_ptr, "dady");
1843 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1844 lp_build_name(depth_ptr, "depth");
1845 lp_build_name(mask_input, "mask_input");
1846 lp_build_name(stride_ptr, "stride_ptr");
1847
1848 if (key->occlusion_count) {
1849 counter = LLVMGetParam(function, 10);
1850 lp_build_name(counter, "counter");
1851 }
1852
1853 /*
1854 * Function body
1855 */
1856
1857 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1858 builder = gallivm->builder;
1859 assert(builder);
1860 LLVMPositionBuilderAtEnd(builder, block);
1861
1862 /* code generated texture sampling */
1863 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
1864
1865 zs_format_desc = util_format_description(key->zsbuf_format);
1866
1867 if (!try_loop) {
1868 /*
1869 * The shader input interpolation info is not explicitely baked in the
1870 * shader key, but everything it derives from (TGSI, and flatshade) is
1871 * already included in the shader key.
1872 */
1873 lp_build_interp_soa_init(&interp,
1874 gallivm,
1875 shader->info.base.num_inputs,
1876 inputs,
1877 builder, fs_type,
1878 FALSE,
1879 a0_ptr, dadx_ptr, dady_ptr,
1880 x, y);
1881
1882 /* loop over quads in the block */
1883 for(i = 0; i < num_fs; ++i) {
1884 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
1885 i*fs_type.length*zs_format_desc->block.bits/8,
1886 0);
1887 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1888 LLVMValueRef depth_ptr_i;
1889
1890 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
1891
1892 generate_fs(gallivm,
1893 shader, key,
1894 builder,
1895 fs_type,
1896 context_ptr,
1897 i,
1898 &interp,
1899 sampler,
1900 &fs_mask[i], /* output */
1901 out_color,
1902 depth_ptr_i,
1903 facing,
1904 partial_mask,
1905 mask_input,
1906 counter);
1907
1908 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
1909 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1910 fs_out_color[cbuf][chan][i] =
1911 out_color[cbuf * !cbuf0_write_all][chan];
1912 }
1913 }
1914 else {
1915 unsigned depth_bits = zs_format_desc->block.bits/8;
1916 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1917 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1918 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1919 num_loop, "mask_store");
1920 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1921
1922 /*
1923 * The shader input interpolation info is not explicitely baked in the
1924 * shader key, but everything it derives from (TGSI, and flatshade) is
1925 * already included in the shader key.
1926 */
1927 lp_build_interp_soa_init(&interp,
1928 gallivm,
1929 shader->info.base.num_inputs,
1930 inputs,
1931 builder, fs_type,
1932 TRUE,
1933 a0_ptr, dadx_ptr, dady_ptr,
1934 x, y);
1935
1936 for (i = 0; i < num_fs; i++) {
1937 LLVMValueRef mask;
1938 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1939 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1940 &indexi, 1, "mask_ptr");
1941
1942 if (partial_mask) {
1943 mask = generate_quad_mask(gallivm, fs_type,
1944 i*fs_type.length/4, mask_input);
1945 }
1946 else {
1947 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1948 }
1949 LLVMBuildStore(builder, mask, mask_ptr);
1950 }
1951
1952 generate_fs_loop(gallivm,
1953 shader, key,
1954 builder,
1955 fs_type,
1956 context_ptr,
1957 num_loop,
1958 &interp,
1959 sampler,
1960 mask_store, /* output */
1961 color_store,
1962 depth_ptr,
1963 depth_bits,
1964 facing,
1965 counter);
1966
1967 for (i = 0; i < num_fs; i++) {
1968 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1969 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1970 &indexi, 1, "");
1971 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1972 /* This is fucked up need to reorganize things */
1973 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1974 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1975 ptr = LLVMBuildGEP(builder,
1976 color_store[cbuf * !cbuf0_write_all][chan],
1977 &indexi, 1, "");
1978 fs_out_color[cbuf][chan][i] = ptr;
1979 }
1980 }
1981 }
1982 }
1983
1984 sampler->destroy(sampler);
1985
1986 /* Loop over color outputs / color buffers to do blending.
1987 */
1988 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1989 LLVMValueRef color_ptr;
1990 LLVMValueRef stride;
1991 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
1992 unsigned rt = key->blend.independent_blend_enable ? cbuf : 0;
1993
1994 boolean do_branch = ((key->depth.enabled
1995 || key->stencil[0].enabled
1996 || key->alpha.enabled)
1997 && !shader->info.base.uses_kill);
1998
1999 color_ptr = LLVMBuildLoad(builder,
2000 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2001 "");
2002
2003 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2004
2005 stride = LLVMBuildLoad(builder,
2006 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2007 "");
2008
2009 generate_unswizzled_blend(gallivm, rt, variant, key->cbuf_format[cbuf],
2010 num_fs, fs_type, fs_mask, fs_out_color[cbuf],
2011 context_ptr, color_ptr, stride, partial_mask, do_branch);
2012 }
2013
2014 LLVMBuildRetVoid(builder);
2015
2016 gallivm_verify_function(gallivm, function);
2017
2018 variant->nr_instrs += lp_build_count_instructions(function);
2019 }
2020
2021
2022 static void
2023 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2024 {
2025 unsigned i;
2026
2027 debug_printf("fs variant %p:\n", (void *) key);
2028
2029 if (key->flatshade) {
2030 debug_printf("flatshade = 1\n");
2031 }
2032 for (i = 0; i < key->nr_cbufs; ++i) {
2033 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2034 }
2035 if (key->depth.enabled) {
2036 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2037 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2038 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2039 }
2040
2041 for (i = 0; i < 2; ++i) {
2042 if (key->stencil[i].enabled) {
2043 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2044 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2045 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2046 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2047 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2048 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2049 }
2050 }
2051
2052 if (key->alpha.enabled) {
2053 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2054 }
2055
2056 if (key->occlusion_count) {
2057 debug_printf("occlusion_count = 1\n");
2058 }
2059
2060 if (key->blend.logicop_enable) {
2061 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2062 }
2063 else if (key->blend.rt[0].blend_enable) {
2064 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2065 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2066 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2067 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2068 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2069 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2070 }
2071 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2072 for (i = 0; i < key->nr_samplers; ++i) {
2073 debug_printf("sampler[%u] = \n", i);
2074 debug_printf(" .format = %s\n",
2075 util_format_name(key->sampler[i].format));
2076 debug_printf(" .target = %s\n",
2077 util_dump_tex_target(key->sampler[i].target, TRUE));
2078 debug_printf(" .pot = %u %u %u\n",
2079 key->sampler[i].pot_width,
2080 key->sampler[i].pot_height,
2081 key->sampler[i].pot_depth);
2082 debug_printf(" .wrap = %s %s %s\n",
2083 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
2084 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
2085 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
2086 debug_printf(" .min_img_filter = %s\n",
2087 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
2088 debug_printf(" .min_mip_filter = %s\n",
2089 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
2090 debug_printf(" .mag_img_filter = %s\n",
2091 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
2092 if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
2093 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
2094 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
2095 debug_printf(" .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
2096 debug_printf(" .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
2097 debug_printf(" .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
2098 debug_printf(" .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
2099 }
2100 }
2101
2102
2103 void
2104 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2105 {
2106 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2107 variant->shader->no, variant->no);
2108 tgsi_dump(variant->shader->base.tokens, 0);
2109 dump_fs_variant_key(&variant->key);
2110 debug_printf("variant->opaque = %u\n", variant->opaque);
2111 debug_printf("\n");
2112 }
2113
2114
2115 /**
2116 * Generate a new fragment shader variant from the shader code and
2117 * other state indicated by the key.
2118 */
2119 static struct lp_fragment_shader_variant *
2120 generate_variant(struct llvmpipe_context *lp,
2121 struct lp_fragment_shader *shader,
2122 const struct lp_fragment_shader_variant_key *key)
2123 {
2124 struct lp_fragment_shader_variant *variant;
2125 const struct util_format_description *cbuf0_format_desc;
2126 boolean fullcolormask;
2127
2128 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2129 if(!variant)
2130 return NULL;
2131
2132 variant->gallivm = gallivm_create();
2133 if (!variant->gallivm) {
2134 FREE(variant);
2135 return NULL;
2136 }
2137
2138 variant->shader = shader;
2139 variant->list_item_global.base = variant;
2140 variant->list_item_local.base = variant;
2141 variant->no = shader->variants_created++;
2142
2143 memcpy(&variant->key, key, shader->variant_key_size);
2144
2145 /*
2146 * Determine whether we are touching all channels in the color buffer.
2147 */
2148 fullcolormask = FALSE;
2149 if (key->nr_cbufs == 1) {
2150 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2151 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2152 }
2153
2154 variant->opaque =
2155 !key->blend.logicop_enable &&
2156 !key->blend.rt[0].blend_enable &&
2157 fullcolormask &&
2158 !key->stencil[0].enabled &&
2159 !key->alpha.enabled &&
2160 !key->depth.enabled &&
2161 !shader->info.base.uses_kill
2162 ? TRUE : FALSE;
2163
2164 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2165 lp_debug_fs_variant(variant);
2166 }
2167
2168 lp_jit_init_types(variant);
2169
2170 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2171 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2172
2173 if (variant->jit_function[RAST_WHOLE] == NULL) {
2174 if (variant->opaque) {
2175 /* Specialized shader, which doesn't need to read the color buffer. */
2176 generate_fragment(lp, shader, variant, RAST_WHOLE);
2177 }
2178 }
2179
2180 /*
2181 * Compile everything
2182 */
2183
2184 gallivm_compile_module(variant->gallivm);
2185
2186 if (variant->function[RAST_EDGE_TEST]) {
2187 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2188 gallivm_jit_function(variant->gallivm,
2189 variant->function[RAST_EDGE_TEST]);
2190 }
2191
2192 if (variant->function[RAST_WHOLE]) {
2193 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2194 gallivm_jit_function(variant->gallivm,
2195 variant->function[RAST_WHOLE]);
2196 } else if (!variant->jit_function[RAST_WHOLE]) {
2197 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2198 }
2199
2200 return variant;
2201 }
2202
2203
2204 static void *
2205 llvmpipe_create_fs_state(struct pipe_context *pipe,
2206 const struct pipe_shader_state *templ)
2207 {
2208 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2209 struct lp_fragment_shader *shader;
2210 int nr_samplers;
2211 int i;
2212
2213 shader = CALLOC_STRUCT(lp_fragment_shader);
2214 if (!shader)
2215 return NULL;
2216
2217 shader->no = fs_no++;
2218 make_empty_list(&shader->variants);
2219
2220 /* get/save the summary info for this shader */
2221 lp_build_tgsi_info(templ->tokens, &shader->info);
2222
2223 /* we need to keep a local copy of the tokens */
2224 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2225
2226 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2227 if (shader->draw_data == NULL) {
2228 FREE((void *) shader->base.tokens);
2229 FREE(shader);
2230 return NULL;
2231 }
2232
2233 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2234
2235 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2236 sampler[nr_samplers]);
2237
2238 for (i = 0; i < shader->info.base.num_inputs; i++) {
2239 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2240 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2241
2242 switch (shader->info.base.input_interpolate[i]) {
2243 case TGSI_INTERPOLATE_CONSTANT:
2244 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2245 break;
2246 case TGSI_INTERPOLATE_LINEAR:
2247 shader->inputs[i].interp = LP_INTERP_LINEAR;
2248 break;
2249 case TGSI_INTERPOLATE_PERSPECTIVE:
2250 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2251 break;
2252 case TGSI_INTERPOLATE_COLOR:
2253 shader->inputs[i].interp = LP_INTERP_COLOR;
2254 break;
2255 default:
2256 assert(0);
2257 break;
2258 }
2259
2260 switch (shader->info.base.input_semantic_name[i]) {
2261 case TGSI_SEMANTIC_FACE:
2262 shader->inputs[i].interp = LP_INTERP_FACING;
2263 break;
2264 case TGSI_SEMANTIC_POSITION:
2265 /* Position was already emitted above
2266 */
2267 shader->inputs[i].interp = LP_INTERP_POSITION;
2268 shader->inputs[i].src_index = 0;
2269 continue;
2270 }
2271
2272 shader->inputs[i].src_index = i+1;
2273 }
2274
2275 if (LP_DEBUG & DEBUG_TGSI) {
2276 unsigned attrib;
2277 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2278 shader->no, (void *) shader);
2279 tgsi_dump(templ->tokens, 0);
2280 debug_printf("usage masks:\n");
2281 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2282 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2283 debug_printf(" IN[%u].%s%s%s%s\n",
2284 attrib,
2285 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2286 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2287 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2288 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2289 }
2290 debug_printf("\n");
2291 }
2292
2293 return shader;
2294 }
2295
2296
2297 static void
2298 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2299 {
2300 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2301
2302 if (llvmpipe->fs == fs)
2303 return;
2304
2305 draw_flush(llvmpipe->draw);
2306
2307 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2308
2309 draw_bind_fragment_shader(llvmpipe->draw,
2310 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2311
2312 llvmpipe->dirty |= LP_NEW_FS;
2313 }
2314
2315
2316 /**
2317 * Remove shader variant from two lists: the shader's variant list
2318 * and the context's variant list.
2319 */
2320 void
2321 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2322 struct lp_fragment_shader_variant *variant)
2323 {
2324 unsigned i;
2325
2326 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2327 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2328 " #%u v total cached #%u\n",
2329 variant->shader->no,
2330 variant->no,
2331 variant->shader->variants_created,
2332 variant->shader->variants_cached,
2333 lp->nr_fs_variants);
2334 }
2335
2336 /* free all the variant's JIT'd functions */
2337 for (i = 0; i < Elements(variant->function); i++) {
2338 if (variant->function[i]) {
2339 gallivm_free_function(variant->gallivm,
2340 variant->function[i],
2341 variant->jit_function[i]);
2342 }
2343 }
2344
2345 gallivm_destroy(variant->gallivm);
2346
2347 /* remove from shader's list */
2348 remove_from_list(&variant->list_item_local);
2349 variant->shader->variants_cached--;
2350
2351 /* remove from context's list */
2352 remove_from_list(&variant->list_item_global);
2353 lp->nr_fs_variants--;
2354 lp->nr_fs_instrs -= variant->nr_instrs;
2355
2356 FREE(variant);
2357 }
2358
2359
2360 static void
2361 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2362 {
2363 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2364 struct lp_fragment_shader *shader = fs;
2365 struct lp_fs_variant_list_item *li;
2366
2367 assert(fs != llvmpipe->fs);
2368
2369 /*
2370 * XXX: we need to flush the context until we have some sort of reference
2371 * counting in fragment shaders as they may still be binned
2372 * Flushing alone might not sufficient we need to wait on it too.
2373 */
2374 llvmpipe_finish(pipe, __FUNCTION__);
2375
2376 /* Delete all the variants */
2377 li = first_elem(&shader->variants);
2378 while(!at_end(&shader->variants, li)) {
2379 struct lp_fs_variant_list_item *next = next_elem(li);
2380 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2381 li = next;
2382 }
2383
2384 /* Delete draw module's data */
2385 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2386
2387 assert(shader->variants_cached == 0);
2388 FREE((void *) shader->base.tokens);
2389 FREE(shader);
2390 }
2391
2392
2393
2394 static void
2395 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2396 uint shader, uint index,
2397 struct pipe_constant_buffer *cb)
2398 {
2399 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2400 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2401 unsigned size;
2402 const void *data;
2403
2404 if (cb && cb->user_buffer) {
2405 constants = llvmpipe_user_buffer_create(pipe->screen,
2406 (void *) cb->user_buffer,
2407 cb->buffer_size,
2408 PIPE_BIND_CONSTANT_BUFFER);
2409 }
2410
2411 size = constants ? constants->width0 : 0;
2412 data = constants ? llvmpipe_resource_data(constants) : NULL;
2413
2414 assert(shader < PIPE_SHADER_TYPES);
2415 assert(index < PIPE_MAX_CONSTANT_BUFFERS);
2416
2417 if(llvmpipe->constants[shader][index] == constants)
2418 return;
2419
2420 draw_flush(llvmpipe->draw);
2421
2422 /* note: reference counting */
2423 pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
2424
2425 if(shader == PIPE_SHADER_VERTEX ||
2426 shader == PIPE_SHADER_GEOMETRY) {
2427 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2428 index, data, size);
2429 }
2430
2431 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2432
2433 if (cb && cb->user_buffer) {
2434 pipe_resource_reference(&constants, NULL);
2435 }
2436 }
2437
2438
2439 /**
2440 * Return the blend factor equivalent to a destination alpha of one.
2441 */
2442 static INLINE unsigned
2443 force_dst_alpha_one(unsigned factor)
2444 {
2445 switch(factor) {
2446 case PIPE_BLENDFACTOR_DST_ALPHA:
2447 return PIPE_BLENDFACTOR_ONE;
2448 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2449 return PIPE_BLENDFACTOR_ZERO;
2450 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2451 return PIPE_BLENDFACTOR_ZERO;
2452 }
2453
2454 return factor;
2455 }
2456
2457
2458 /**
2459 * We need to generate several variants of the fragment pipeline to match
2460 * all the combinations of the contributing state atoms.
2461 *
2462 * TODO: there is actually no reason to tie this to context state -- the
2463 * generated code could be cached globally in the screen.
2464 */
2465 static void
2466 make_variant_key(struct llvmpipe_context *lp,
2467 struct lp_fragment_shader *shader,
2468 struct lp_fragment_shader_variant_key *key)
2469 {
2470 unsigned i;
2471
2472 memset(key, 0, shader->variant_key_size);
2473
2474 if (lp->framebuffer.zsbuf) {
2475 if (lp->depth_stencil->depth.enabled) {
2476 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2477 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2478 }
2479 if (lp->depth_stencil->stencil[0].enabled) {
2480 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2481 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2482 }
2483 }
2484
2485 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2486 if(key->alpha.enabled)
2487 key->alpha.func = lp->depth_stencil->alpha.func;
2488 /* alpha.ref_value is passed in jit_context */
2489
2490 key->flatshade = lp->rasterizer->flatshade;
2491 if (lp->active_query_count) {
2492 key->occlusion_count = TRUE;
2493 }
2494
2495 if (lp->framebuffer.nr_cbufs) {
2496 memcpy(&key->blend, lp->blend, sizeof key->blend);
2497 }
2498
2499 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2500 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2501 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2502 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2503 const struct util_format_description *format_desc;
2504
2505 key->cbuf_format[i] = format;
2506
2507 format_desc = util_format_description(format);
2508 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2509 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2510
2511 blend_rt->colormask = lp->blend->rt[i].colormask;
2512
2513 /*
2514 * Mask out color channels not present in the color buffer.
2515 */
2516 blend_rt->colormask &= util_format_colormask(format_desc);
2517
2518 /*
2519 * Our swizzled render tiles always have an alpha channel, but the linear
2520 * render target format often does not, so force here the dst alpha to be
2521 * one.
2522 *
2523 * This is not a mere optimization. Wrong results will be produced if the
2524 * dst alpha is used, the dst format does not have alpha, and the previous
2525 * rendering was not flushed from the swizzled to linear buffer. For
2526 * example, NonPowTwo DCT.
2527 *
2528 * TODO: This should be generalized to all channels for better
2529 * performance, but only alpha causes correctness issues.
2530 *
2531 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2532 */
2533 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2534 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2535 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2536 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2537 blend_rt->alpha_func = blend_rt->rgb_func;
2538 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2539 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2540 }
2541 }
2542
2543 /* This value will be the same for all the variants of a given shader:
2544 */
2545 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2546
2547 for(i = 0; i < key->nr_samplers; ++i) {
2548 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2549 lp_sampler_static_state(&key->sampler[i],
2550 lp->sampler_views[PIPE_SHADER_FRAGMENT][i],
2551 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2552 }
2553 }
2554 }
2555
2556
2557
2558 /**
2559 * Update fragment shader state. This is called just prior to drawing
2560 * something when some fragment-related state has changed.
2561 */
2562 void
2563 llvmpipe_update_fs(struct llvmpipe_context *lp)
2564 {
2565 struct lp_fragment_shader *shader = lp->fs;
2566 struct lp_fragment_shader_variant_key key;
2567 struct lp_fragment_shader_variant *variant = NULL;
2568 struct lp_fs_variant_list_item *li;
2569
2570 make_variant_key(lp, shader, &key);
2571
2572 /* Search the variants for one which matches the key */
2573 li = first_elem(&shader->variants);
2574 while(!at_end(&shader->variants, li)) {
2575 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2576 variant = li->base;
2577 break;
2578 }
2579 li = next_elem(li);
2580 }
2581
2582 if (variant) {
2583 /* Move this variant to the head of the list to implement LRU
2584 * deletion of shader's when we have too many.
2585 */
2586 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2587 }
2588 else {
2589 /* variant not found, create it now */
2590 int64_t t0, t1, dt;
2591 unsigned i;
2592 unsigned variants_to_cull;
2593
2594 if (0) {
2595 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2596 lp->nr_fs_variants,
2597 lp->nr_fs_instrs,
2598 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2599 }
2600
2601 /* First, check if we've exceeded the max number of shader variants.
2602 * If so, free 25% of them (the least recently used ones).
2603 */
2604 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2605
2606 if (variants_to_cull ||
2607 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2608 struct pipe_context *pipe = &lp->pipe;
2609
2610 /*
2611 * XXX: we need to flush the context until we have some sort of
2612 * reference counting in fragment shaders as they may still be binned
2613 * Flushing alone might not be sufficient we need to wait on it too.
2614 */
2615 llvmpipe_finish(pipe, __FUNCTION__);
2616
2617 /*
2618 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2619 * number of shader variants (potentially all of them) could be
2620 * pending for destruction on flush.
2621 */
2622
2623 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2624 struct lp_fs_variant_list_item *item;
2625 if (is_empty_list(&lp->fs_variants_list)) {
2626 break;
2627 }
2628 item = last_elem(&lp->fs_variants_list);
2629 assert(item);
2630 assert(item->base);
2631 llvmpipe_remove_shader_variant(lp, item->base);
2632 }
2633 }
2634
2635 /*
2636 * Generate the new variant.
2637 */
2638 t0 = os_time_get();
2639 variant = generate_variant(lp, shader, &key);
2640 t1 = os_time_get();
2641 dt = t1 - t0;
2642 LP_COUNT_ADD(llvm_compile_time, dt);
2643 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2644
2645 llvmpipe_variant_count++;
2646
2647 /* Put the new variant into the list */
2648 if (variant) {
2649 insert_at_head(&shader->variants, &variant->list_item_local);
2650 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2651 lp->nr_fs_variants++;
2652 lp->nr_fs_instrs += variant->nr_instrs;
2653 shader->variants_cached++;
2654 }
2655 }
2656
2657 /* Bind this variant */
2658 lp_setup_set_fs_variant(lp->setup, variant);
2659 }
2660
2661
2662
2663
2664
2665
2666
2667 void
2668 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2669 {
2670 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2671 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2672 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2673
2674 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2675 }