projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: Fix alignment.
[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 unsigned dst_alignment)
844 {
845 LLVMBuilderRef builder = gallivm->builder;
846 unsigned row_size = dst_count / block_height;
847 unsigned i;
848
849 /* Ensure block exactly fits into dst */
850 assert((block_width * block_height) % dst_count == 0);
851
852 for (i = 0; i < dst_count; ++i) {
853 unsigned x = i % row_size;
854 unsigned y = i / row_size;
855
856 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
857 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
858
859 LLVMValueRef gep[2];
860 LLVMValueRef dst_ptr;
861
862 gep[0] = lp_build_const_int32(gallivm, 0);
863 gep[1] = LLVMBuildAdd(builder, bx, by, "");
864
865 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
866 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
867
868 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
869
870 lp_set_load_alignment(dst[i], dst_alignment);
871 }
872 }
873
874
875 /**
876 * Store an unswizzled block of pixels to memory
877 */
878 static void
879 store_unswizzled_block(struct gallivm_state *gallivm,
880 LLVMValueRef base_ptr,
881 LLVMValueRef stride,
882 unsigned block_width,
883 unsigned block_height,
884 LLVMValueRef* src,
885 struct lp_type src_type,
886 unsigned src_count,
887 unsigned src_alignment)
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 lp_set_store_alignment(src_ptr, src_alignment);
915 }
916 }
917
918
919 /**
920 * Checks if a format description is an arithmetic format
921 *
922 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
923 */
924 static INLINE boolean
925 is_arithmetic_format(const struct util_format_description *format_desc)
926 {
927 boolean arith = false;
928 unsigned i;
929
930 for (i = 0; i < format_desc->nr_channels; ++i) {
931 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
932 arith |= (format_desc->channel[i].size % 8) != 0;
933 }
934
935 return arith;
936 }
937
938
939 /**
940 * Retrieves the type representing the memory layout for a format
941 *
942 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
943 */
944 static INLINE void
945 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
946 struct lp_type* type)
947 {
948 int i;
949
950 memset(type, 0, sizeof(struct lp_type));
951 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
952 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
953 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
954 type->norm = format_desc->channel[0].normalized;
955
956 if (is_arithmetic_format(format_desc)) {
957 type->width = 0;
958 type->length = 1;
959
960 for (i = 0; i < format_desc->nr_channels; ++i) {
961 type->width += format_desc->channel[i].size;
962 }
963 } else {
964 type->width = format_desc->channel[0].size;
965 type->length = format_desc->nr_channels;
966 }
967 }
968
969
970 /**
971 * Retrieves the type for a format which is usable in the blending code.
972 *
973 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
974 */
975 static INLINE void
976 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
977 struct lp_type* type)
978 {
979 int i;
980
981 memset(type, 0, sizeof(struct lp_type));
982 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
983 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
984 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
985 type->norm = format_desc->channel[0].normalized;
986 type->width = format_desc->channel[0].size;
987 type->length = format_desc->nr_channels;
988
989 for (i = 1; i < format_desc->nr_channels; ++i) {
990 if (format_desc->channel[i].size > type->width)
991 type->width = format_desc->channel[i].size;
992 }
993
994 if (type->floating) {
995 type->width = 32;
996 } else {
997 if (type->width <= 8) {
998 type->width = 8;
999 } else if (type->width <= 16) {
1000 type->width = 16;
1001 } else {
1002 type->width = 32;
1003 }
1004 }
1005
1006 if (is_arithmetic_format(format_desc) && type->length == 3) {
1007 type->length = 4;
1008 }
1009 }
1010
1011
1012 /**
1013 * Scale a normalised value from src_bits to dst_bits
1014 */
1015 static INLINE LLVMValueRef
1016 scale_bits(struct gallivm_state *gallivm,
1017 int src_bits,
1018 int dst_bits,
1019 LLVMValueRef src,
1020 struct lp_type src_type)
1021 {
1022 LLVMBuilderRef builder = gallivm->builder;
1023 LLVMValueRef result = src;
1024
1025 if (dst_bits < src_bits) {
1026 /* Scale down by LShr */
1027 result = LLVMBuildLShr(builder,
1028 src,
1029 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
1030 "");
1031 } else if (dst_bits > src_bits) {
1032 /* Scale up bits */
1033 int db = dst_bits - src_bits;
1034
1035 /* Shift left by difference in bits */
1036 result = LLVMBuildShl(builder,
1037 src,
1038 lp_build_const_int_vec(gallivm, src_type, db),
1039 "");
1040
1041 if (db < src_bits) {
1042 /* Enough bits in src to fill the remainder */
1043 LLVMValueRef lower = LLVMBuildLShr(builder,
1044 src,
1045 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1046 "");
1047
1048 result = LLVMBuildOr(builder, result, lower, "");
1049 } else if (db > src_bits) {
1050 /* Need to repeatedely copy src bits to fill remainder in dst */
1051 unsigned n;
1052
1053 for (n = src_bits; n < dst_bits; n *= 2) {
1054 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1055
1056 result = LLVMBuildOr(builder,
1057 result,
1058 LLVMBuildLShr(builder, result, shuv, ""),
1059 "");
1060 }
1061 }
1062 }
1063
1064 return result;
1065 }
1066
1067
1068 /**
1069 * Convert from memory format to blending format
1070 *
1071 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1072 */
1073 static void
1074 convert_to_blend_type(struct gallivm_state *gallivm,
1075 const struct util_format_description *src_fmt,
1076 struct lp_type src_type,
1077 struct lp_type dst_type,
1078 LLVMValueRef* src, // and dst
1079 unsigned num_srcs)
1080 {
1081 LLVMValueRef *dst = src;
1082 LLVMBuilderRef builder = gallivm->builder;
1083 struct lp_type blend_type;
1084 struct lp_type mem_type;
1085 unsigned i, j, k;
1086 unsigned pixels = 16 / num_srcs;
1087 bool is_arith;
1088
1089 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1090 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1091
1092 /* Is the format arithmetic */
1093 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1094 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1095
1096 /* Pad if necessary */
1097 if (!is_arith && src_type.length < dst_type.length) {
1098 for (i = 0; i < num_srcs; ++i) {
1099 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1100 }
1101
1102 src_type.length = dst_type.length;
1103 }
1104
1105 /* Special case for half-floats */
1106 if (mem_type.width == 16 && mem_type.floating) {
1107 assert(blend_type.width == 32 && blend_type.floating);
1108 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1109 is_arith = false;
1110 }
1111
1112 if (!is_arith) {
1113 return;
1114 }
1115
1116 src_type.width = blend_type.width * blend_type.length;
1117 blend_type.length *= pixels;
1118 src_type.length *= pixels / (src_type.length / mem_type.length);
1119
1120 for (i = 0; i < num_srcs; ++i) {
1121 LLVMValueRef chans[4];
1122 LLVMValueRef res;
1123 unsigned sa = 0;
1124
1125 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1126
1127 for (j = 0; j < src_fmt->nr_channels; ++j) {
1128 unsigned mask = 0;
1129
1130 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1131 mask |= 1 << k;
1132 }
1133
1134 /* Extract bits from source */
1135 chans[j] = LLVMBuildLShr(builder,
1136 dst[i],
1137 lp_build_const_int_vec(gallivm, src_type, sa),
1138 "");
1139
1140 chans[j] = LLVMBuildAnd(builder,
1141 chans[j],
1142 lp_build_const_int_vec(gallivm, src_type, mask),
1143 "");
1144
1145 /* Scale bits */
1146 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size, blend_type.width, chans[j], src_type);
1147
1148 /* Insert bits into correct position */
1149 chans[j] = LLVMBuildShl(builder,
1150 chans[j],
1151 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1152 "");
1153
1154 sa += src_fmt->channel[j].size;
1155
1156 if (j == 0) {
1157 res = chans[j];
1158 } else {
1159 res = LLVMBuildOr(builder, res, chans[j], "");
1160 }
1161 }
1162
1163 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1164 }
1165 }
1166
1167
1168 /**
1169 * Convert from blending format to memory format
1170 *
1171 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1172 */
1173 static void
1174 convert_from_blend_type(struct gallivm_state *gallivm,
1175 const struct util_format_description *src_fmt,
1176 struct lp_type src_type,
1177 struct lp_type dst_type,
1178 LLVMValueRef* src, // and dst
1179 unsigned num_srcs)
1180 {
1181 LLVMValueRef* dst = src;
1182 unsigned i, j, k;
1183 struct lp_type mem_type;
1184 struct lp_type blend_type;
1185 LLVMBuilderRef builder = gallivm->builder;
1186 unsigned pixels = 16 / num_srcs;
1187 bool is_arith;
1188
1189 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1190 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1191
1192 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1193
1194 /* Special case for half-floats */
1195 if (mem_type.width == 16 && mem_type.floating) {
1196 int length = dst_type.length;
1197 assert(blend_type.width == 32 && blend_type.floating);
1198
1199 dst_type.length = src_type.length;
1200
1201 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1202
1203 dst_type.length = length;
1204 is_arith = false;
1205 }
1206
1207 /* Remove any padding */
1208 if (!is_arith && (src_type.length % mem_type.length)) {
1209 src_type.length -= (src_type.length % mem_type.length);
1210
1211 for (i = 0; i < num_srcs; ++i) {
1212 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1213 }
1214 }
1215
1216 /* No bit arithmitic to do */
1217 if (!is_arith) {
1218 return;
1219 }
1220
1221 src_type.length = pixels;
1222 src_type.width = blend_type.length * blend_type.width;
1223 dst_type.length = pixels;
1224
1225 for (i = 0; i < num_srcs; ++i) {
1226 LLVMValueRef chans[4];
1227 LLVMValueRef res;
1228 unsigned sa = 0;
1229
1230 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1231
1232 for (j = 0; j < src_fmt->nr_channels; ++j) {
1233 unsigned mask = 0;
1234
1235 assert(blend_type.width > src_fmt->channel[j].size);
1236
1237 for (k = 0; k < blend_type.width; ++k) {
1238 mask |= 1 << k;
1239 }
1240
1241 /* Extract bits */
1242 chans[j] = LLVMBuildLShr(builder,
1243 dst[i],
1244 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1245 "");
1246
1247 chans[j] = LLVMBuildAnd(builder,
1248 chans[j],
1249 lp_build_const_int_vec(gallivm, src_type, mask),
1250 "");
1251
1252 /* Scale down bits */
1253 chans[j] = scale_bits(gallivm, blend_type.width, src_fmt->channel[j].size, chans[j], src_type);
1254
1255 /* Insert bits */
1256 chans[j] = LLVMBuildShl(builder,
1257 chans[j],
1258 lp_build_const_int_vec(gallivm, src_type, sa),
1259 "");
1260
1261 sa += src_fmt->channel[j].size;
1262
1263 if (j == 0) {
1264 res = chans[j];
1265 } else {
1266 res = LLVMBuildOr(builder, res, chans[j], "");
1267 }
1268 }
1269
1270 assert (dst_type.width != 24);
1271
1272 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1273 }
1274 }
1275
1276
1277 /**
1278 * Generates the blend function for unswizzled colour buffers
1279 * Also generates the read & write from colour buffer
1280 */
1281 static void
1282 generate_unswizzled_blend(struct gallivm_state *gallivm,
1283 unsigned rt,
1284 struct lp_fragment_shader_variant *variant,
1285 enum pipe_format out_format,
1286 unsigned int num_fs,
1287 struct lp_type fs_type,
1288 LLVMValueRef* fs_mask,
1289 LLVMValueRef fs_out_color[TGSI_NUM_CHANNELS][4],
1290 LLVMValueRef context_ptr,
1291 LLVMValueRef color_ptr,
1292 LLVMValueRef stride,
1293 unsigned partial_mask,
1294 boolean do_branch)
1295 {
1296 const unsigned alpha_channel = 3;
1297 const unsigned block_width = 4;
1298 const unsigned block_height = 4;
1299 const unsigned block_size = block_width * block_height;
1300 const unsigned lp_integer_vector_width = 128;
1301
1302 LLVMBuilderRef builder = gallivm->builder;
1303 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1304 LLVMValueRef src_alpha[4 * 4];
1305 LLVMValueRef src_mask[4 * 4];
1306 LLVMValueRef src[4 * 4];
1307 LLVMValueRef dst[4 * 4];
1308 LLVMValueRef blend_color;
1309 LLVMValueRef blend_alpha;
1310 LLVMValueRef i32_zero;
1311 LLVMValueRef check_mask;
1312
1313 struct lp_build_mask_context mask_ctx;
1314 struct lp_type mask_type;
1315 struct lp_type blend_type;
1316 struct lp_type alpha_type;
1317 struct lp_type row_type;
1318 struct lp_type dst_type;
1319
1320 unsigned char swizzle[TGSI_NUM_CHANNELS];
1321 unsigned vector_width;
1322 unsigned dst_channels;
1323 unsigned src_channels;
1324 unsigned dst_count;
1325 unsigned src_count;
1326 unsigned i, j;
1327
1328 const struct util_format_description* out_format_desc = util_format_description(out_format);
1329
1330 unsigned dst_alignment;
1331
1332 bool pad_inline = is_arithmetic_format(out_format_desc);
1333 bool has_alpha = false;
1334
1335 src_channels = TGSI_NUM_CHANNELS;
1336 mask_type = lp_int32_vec4_type();
1337 mask_type.length = fs_type.length;
1338
1339 /* Compute the alignment of the destination pointer in bytes */
1340 #if 0
1341 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1342 #else
1343 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1344 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1345 #endif
1346 /* Force power-of-two alignment by extracting only the least-significant-bit */
1347 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1348 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1349 dst_alignment = MIN2(dst_alignment, 16);
1350
1351 /* Do not bother executing code when mask is empty.. */
1352 if (do_branch) {
1353 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1354
1355 for (i = 0; i < num_fs; ++i) {
1356 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1357 }
1358
1359 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1360 lp_build_mask_check(&mask_ctx);
1361 }
1362
1363 partial_mask |= !variant->opaque;
1364 i32_zero = lp_build_const_int32(gallivm, 0);
1365
1366 /* Get type from output format */
1367 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1368 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1369
1370 row_type.length = fs_type.length;
1371 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1372
1373 /* Compute correct swizzle and count channels */
1374 memset(swizzle, 0xFF, TGSI_NUM_CHANNELS);
1375 dst_channels = 0;
1376
1377 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1378 /* Ensure channel is used */
1379 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1380 continue;
1381 }
1382
1383 /* Ensure not already written to (happens in case with GL_ALPHA) */
1384 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1385 continue;
1386 }
1387
1388 /* Ensure we havn't already found all channels */
1389 if (dst_channels >= out_format_desc->nr_channels) {
1390 continue;
1391 }
1392
1393 swizzle[out_format_desc->swizzle[i]] = i;
1394 ++dst_channels;
1395
1396 if (i == alpha_channel) {
1397 has_alpha = true;
1398 }
1399 }
1400
1401 /* If 3 channels then pad to include alpha for 4 element transpose */
1402 if (dst_channels == 3 && !has_alpha) {
1403 swizzle[3] = 3;
1404
1405 if (out_format_desc->nr_channels == 4) {
1406 dst_channels = 4;
1407 }
1408 }
1409
1410 /*
1411 * Load shader output
1412 */
1413 for (i = 0; i < num_fs; ++i) {
1414 /* Always load alpha for use in blending */
1415 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[alpha_channel][i], "");
1416
1417 /* Load each channel */
1418 for (j = 0; j < dst_channels; ++j) {
1419 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[swizzle[j]][i], "");
1420 }
1421
1422 /* If 3 channels then pad to include alpha for 4 element transpose */
1423 if (dst_channels == 3 && !has_alpha) {
1424 fs_src[i][3] = alpha;
1425 swizzle[3] = 3;
1426 }
1427
1428 /* We split the row_mask and row_alpha as we want 128bit interleave */
1429 if (fs_type.length == 8) {
1430 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1431 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1432
1433 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1434 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1435 } else {
1436 src_mask[i] = fs_mask[i];
1437 src_alpha[i] = alpha;
1438 }
1439 }
1440
1441
1442 /*
1443 * Pixel twiddle from fragment shader order to memory order
1444 */
1445 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1446 src_channels = dst_channels < 3 ? dst_channels : 4;
1447 if (src_count != num_fs * src_channels) {
1448 unsigned ds = src_count / (num_fs * src_channels);
1449 row_type.length /= ds;
1450 fs_type.length = row_type.length;
1451 }
1452
1453 blend_type = row_type;
1454 alpha_type = fs_type;
1455 alpha_type.length = 4;
1456 mask_type.length = 4;
1457
1458 /* Convert src to row_type */
1459 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1460
1461 /* If the rows are not an SSE vector, combine them to become SSE size! */
1462 if ((row_type.width * row_type.length) % 128) {
1463 unsigned bits = row_type.width * row_type.length;
1464 unsigned combined;
1465
1466 dst_count = src_count / (vector_width / bits);
1467 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1468
1469 row_type.length *= combined;
1470 src_count /= combined;
1471
1472 bits = row_type.width * row_type.length;
1473 assert(bits == 128 || bits == 256);
1474 }
1475
1476
1477 /*
1478 * Blend Colour conversion
1479 */
1480 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1481 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1482 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1483
1484 /* Convert */
1485 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1486
1487 /* Extract alpha */
1488 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1489
1490 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1491 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1492 if (pad_inline) {
1493 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1494 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1495 } else {
1496 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1497 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1498 }
1499
1500 /*
1501 * Mask conversion
1502 */
1503 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1504
1505 if (src_count < block_height) {
1506 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1507 } else if (src_count > block_height) {
1508 for (i = src_count; i > 0; --i) {
1509 unsigned pixels = block_size / src_count;
1510 unsigned idx = i - 1;
1511
1512 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1513 }
1514 }
1515
1516 assert(mask_type.width == 32);
1517
1518 for (i = 0; i < src_count; ++i) {
1519 unsigned pixels = block_size / src_count;
1520 unsigned pixel_width = row_type.width * dst_channels;
1521
1522 if (pixel_width == 24) {
1523 mask_type.width = 8;
1524 mask_type.length = vector_width / mask_type.width;
1525 } else {
1526 mask_type.length = pixels;
1527 mask_type.width = row_type.width * dst_channels;
1528
1529 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1530
1531 mask_type.length *= dst_channels;
1532 mask_type.width /= dst_channels;
1533 }
1534
1535 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1536 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1537 }
1538
1539 /*
1540 * Alpha conversion
1541 */
1542 if (!has_alpha) {
1543 unsigned length = row_type.length;
1544 row_type.length = alpha_type.length;
1545
1546 /* Twiddle the alpha to match pixels */
1547 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1548
1549 for (i = 0; i < 4; ++i) {
1550 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1551 }
1552
1553 alpha_type = row_type;
1554 row_type.length = length;
1555
1556 /* If only one channel we can only need the single alpha value per pixel */
1557 if (src_count == 1) {
1558 assert(dst_channels == 1);
1559
1560 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1561 } else {
1562 /* If there are more srcs than rows then we need to split alpha up */
1563 if (src_count > block_height) {
1564 for (i = src_count; i > 0; --i) {
1565 unsigned pixels = block_size / src_count;
1566 unsigned idx = i - 1;
1567
1568 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1569 }
1570 }
1571
1572 /* If there is a src for each pixel broadcast the alpha across whole row */
1573 if (src_count == block_size) {
1574 for (i = 0; i < src_count; ++i) {
1575 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1576 }
1577 } else {
1578 unsigned pixels = block_size / src_count;
1579 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1580 unsigned alpha_span = 1;
1581
1582 /* Check if we need 2 src_alphas for our shuffles */
1583 if (pixels > alpha_type.length) {
1584 alpha_span = 2;
1585 }
1586
1587 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1588 for (i = 0; i < src_count; ++i) {
1589 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1590 unsigned idx1 = i, idx2 = i;
1591
1592 if (alpha_span > 1){
1593 idx1 *= alpha_span;
1594 idx2 = idx1 + 1;
1595 }
1596
1597 for (j = 0; j < row_type.length; ++j) {
1598 if (j < pixels * channels) {
1599 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1600 } else {
1601 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1602 }
1603 }
1604
1605 src_alpha[i] = LLVMBuildShuffleVector(builder,
1606 src_alpha[idx1],
1607 src_alpha[idx2],
1608 LLVMConstVector(shuffles, row_type.length),
1609 "");
1610 }
1611 }
1612 }
1613 }
1614
1615
1616 /*
1617 * Load dst from memory
1618 */
1619 if (src_count < block_height) {
1620 dst_count = block_height;
1621 } else {
1622 dst_count = src_count;
1623 }
1624
1625 dst_type.length *= 16 / dst_count;
1626
1627 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1628 dst, dst_type, dst_count, dst_alignment);
1629
1630
1631 /*
1632 * Convert from dst/output format to src/blending format.
1633 *
1634 * This is necessary as we can only read 1 row from memory at a time,
1635 * so the minimum dst_count will ever be at this point is 4.
1636 *
1637 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1638 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1639 * on all 16 pixels in that single vector at once.
1640 */
1641 if (dst_count > src_count) {
1642 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1643 }
1644
1645 /*
1646 * Blending
1647 */
1648 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1649
1650 for (i = 0; i < src_count; ++i) {
1651 dst[i] = lp_build_blend_aos(gallivm,
1652 &variant->key.blend,
1653 variant->key.cbuf_format,
1654 row_type,
1655 rt,
1656 src[i],
1657 has_alpha ? NULL : src_alpha[i],
1658 dst[i],
1659 partial_mask ? src_mask[i] : NULL,
1660 blend_color,
1661 has_alpha ? NULL : blend_alpha,
1662 swizzle,
1663 pad_inline ? 4 : dst_channels);
1664 }
1665
1666 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1667
1668 /* Split the blend rows back to memory rows */
1669 if (dst_count > src_count) {
1670 row_type.length = dst_type.length * (dst_count / src_count);
1671
1672 if (src_count == 1) {
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 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1681 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1682 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1683 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1684
1685 row_type.length /= 2;
1686 src_count *= 2;
1687 }
1688
1689
1690 /*
1691 * Store blend result to memory
1692 */
1693 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1694 dst, dst_type, dst_count, dst_alignment);
1695
1696 if (do_branch) {
1697 lp_build_mask_end(&mask_ctx);
1698 }
1699 }
1700
1701
1702 /**
1703 * Generate the runtime callable function for the whole fragment pipeline.
1704 * Note that the function which we generate operates on a block of 16
1705 * pixels at at time. The block contains 2x2 quads. Each quad contains
1706 * 2x2 pixels.
1707 */
1708 static void
1709 generate_fragment(struct llvmpipe_context *lp,
1710 struct lp_fragment_shader *shader,
1711 struct lp_fragment_shader_variant *variant,
1712 unsigned partial_mask)
1713 {
1714 struct gallivm_state *gallivm = variant->gallivm;
1715 const struct lp_fragment_shader_variant_key *key = &variant->key;
1716 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1717 char func_name[256];
1718 struct lp_type fs_type;
1719 struct lp_type blend_type;
1720 LLVMTypeRef fs_elem_type;
1721 LLVMTypeRef blend_vec_type;
1722 LLVMTypeRef arg_types[12];
1723 LLVMTypeRef func_type;
1724 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1725 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1726 LLVMValueRef context_ptr;
1727 LLVMValueRef x;
1728 LLVMValueRef y;
1729 LLVMValueRef a0_ptr;
1730 LLVMValueRef dadx_ptr;
1731 LLVMValueRef dady_ptr;
1732 LLVMValueRef color_ptr_ptr;
1733 LLVMValueRef stride_ptr;
1734 LLVMValueRef depth_ptr;
1735 LLVMValueRef mask_input;
1736 LLVMValueRef counter = NULL;
1737 LLVMBasicBlockRef block;
1738 LLVMBuilderRef builder;
1739 struct lp_build_sampler_soa *sampler;
1740 struct lp_build_interp_soa_context interp;
1741 LLVMValueRef fs_mask[16 / 4];
1742 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1743 LLVMValueRef function;
1744 LLVMValueRef facing;
1745 const struct util_format_description *zs_format_desc;
1746 unsigned num_fs;
1747 unsigned i;
1748 unsigned chan;
1749 unsigned cbuf;
1750 boolean cbuf0_write_all;
1751 boolean try_loop = TRUE;
1752
1753 assert(lp_native_vector_width / 32 >= 4);
1754
1755 /* Adjust color input interpolation according to flatshade state:
1756 */
1757 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1758 for (i = 0; i < shader->info.base.num_inputs; i++) {
1759 if (inputs[i].interp == LP_INTERP_COLOR) {
1760 if (key->flatshade)
1761 inputs[i].interp = LP_INTERP_CONSTANT;
1762 else
1763 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1764 }
1765 }
1766
1767 /* check if writes to cbuf[0] are to be copied to all cbufs */
1768 cbuf0_write_all = FALSE;
1769 for (i = 0;i < shader->info.base.num_properties; i++) {
1770 if (shader->info.base.properties[i].name ==
1771 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1772 cbuf0_write_all = TRUE;
1773 break;
1774 }
1775 }
1776
1777 /* TODO: actually pick these based on the fs and color buffer
1778 * characteristics. */
1779
1780 memset(&fs_type, 0, sizeof fs_type);
1781 fs_type.floating = TRUE; /* floating point values */
1782 fs_type.sign = TRUE; /* values are signed */
1783 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1784 fs_type.width = 32; /* 32-bit float */
1785 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1786 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1787
1788 memset(&blend_type, 0, sizeof blend_type);
1789 blend_type.floating = FALSE; /* values are integers */
1790 blend_type.sign = FALSE; /* values are unsigned */
1791 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1792 blend_type.width = 8; /* 8-bit ubyte values */
1793 blend_type.length = 16; /* 16 elements per vector */
1794
1795 /*
1796 * Generate the function prototype. Any change here must be reflected in
1797 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1798 */
1799
1800 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1801
1802 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1803
1804 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1805 shader->no, variant->no, partial_mask ? "partial" : "whole");
1806
1807 arg_types[0] = variant->jit_context_ptr_type; /* context */
1808 arg_types[1] = int32_type; /* x */
1809 arg_types[2] = int32_type; /* y */
1810 arg_types[3] = int32_type; /* facing */
1811 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1812 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1813 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1814 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1815 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1816 arg_types[9] = int32_type; /* mask_input */
1817 arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */
1818 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1819
1820 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1821 arg_types, Elements(arg_types), 0);
1822
1823 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1824 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1825
1826 variant->function[partial_mask] = function;
1827
1828 /* XXX: need to propagate noalias down into color param now we are
1829 * passing a pointer-to-pointer?
1830 */
1831 for(i = 0; i < Elements(arg_types); ++i)
1832 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1833 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1834
1835 context_ptr = LLVMGetParam(function, 0);
1836 x = LLVMGetParam(function, 1);
1837 y = LLVMGetParam(function, 2);
1838 facing = LLVMGetParam(function, 3);
1839 a0_ptr = LLVMGetParam(function, 4);
1840 dadx_ptr = LLVMGetParam(function, 5);
1841 dady_ptr = LLVMGetParam(function, 6);
1842 color_ptr_ptr = LLVMGetParam(function, 7);
1843 depth_ptr = LLVMGetParam(function, 8);
1844 mask_input = LLVMGetParam(function, 9);
1845 stride_ptr = LLVMGetParam(function, 11);
1846
1847 lp_build_name(context_ptr, "context");
1848 lp_build_name(x, "x");
1849 lp_build_name(y, "y");
1850 lp_build_name(a0_ptr, "a0");
1851 lp_build_name(dadx_ptr, "dadx");
1852 lp_build_name(dady_ptr, "dady");
1853 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1854 lp_build_name(depth_ptr, "depth");
1855 lp_build_name(mask_input, "mask_input");
1856 lp_build_name(stride_ptr, "stride_ptr");
1857
1858 if (key->occlusion_count) {
1859 counter = LLVMGetParam(function, 10);
1860 lp_build_name(counter, "counter");
1861 }
1862
1863 /*
1864 * Function body
1865 */
1866
1867 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1868 builder = gallivm->builder;
1869 assert(builder);
1870 LLVMPositionBuilderAtEnd(builder, block);
1871
1872 /* code generated texture sampling */
1873 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
1874
1875 zs_format_desc = util_format_description(key->zsbuf_format);
1876
1877 if (!try_loop) {
1878 /*
1879 * The shader input interpolation info is not explicitely baked in the
1880 * shader key, but everything it derives from (TGSI, and flatshade) is
1881 * already included in the shader key.
1882 */
1883 lp_build_interp_soa_init(&interp,
1884 gallivm,
1885 shader->info.base.num_inputs,
1886 inputs,
1887 builder, fs_type,
1888 FALSE,
1889 a0_ptr, dadx_ptr, dady_ptr,
1890 x, y);
1891
1892 /* loop over quads in the block */
1893 for(i = 0; i < num_fs; ++i) {
1894 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
1895 i*fs_type.length*zs_format_desc->block.bits/8,
1896 0);
1897 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1898 LLVMValueRef depth_ptr_i;
1899
1900 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
1901
1902 generate_fs(gallivm,
1903 shader, key,
1904 builder,
1905 fs_type,
1906 context_ptr,
1907 i,
1908 &interp,
1909 sampler,
1910 &fs_mask[i], /* output */
1911 out_color,
1912 depth_ptr_i,
1913 facing,
1914 partial_mask,
1915 mask_input,
1916 counter);
1917
1918 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
1919 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1920 fs_out_color[cbuf][chan][i] =
1921 out_color[cbuf * !cbuf0_write_all][chan];
1922 }
1923 }
1924 else {
1925 unsigned depth_bits = zs_format_desc->block.bits/8;
1926 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1927 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1928 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1929 num_loop, "mask_store");
1930 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1931
1932 /*
1933 * The shader input interpolation info is not explicitely baked in the
1934 * shader key, but everything it derives from (TGSI, and flatshade) is
1935 * already included in the shader key.
1936 */
1937 lp_build_interp_soa_init(&interp,
1938 gallivm,
1939 shader->info.base.num_inputs,
1940 inputs,
1941 builder, fs_type,
1942 TRUE,
1943 a0_ptr, dadx_ptr, dady_ptr,
1944 x, y);
1945
1946 for (i = 0; i < num_fs; i++) {
1947 LLVMValueRef mask;
1948 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1949 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1950 &indexi, 1, "mask_ptr");
1951
1952 if (partial_mask) {
1953 mask = generate_quad_mask(gallivm, fs_type,
1954 i*fs_type.length/4, mask_input);
1955 }
1956 else {
1957 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1958 }
1959 LLVMBuildStore(builder, mask, mask_ptr);
1960 }
1961
1962 generate_fs_loop(gallivm,
1963 shader, key,
1964 builder,
1965 fs_type,
1966 context_ptr,
1967 num_loop,
1968 &interp,
1969 sampler,
1970 mask_store, /* output */
1971 color_store,
1972 depth_ptr,
1973 depth_bits,
1974 facing,
1975 counter);
1976
1977 for (i = 0; i < num_fs; i++) {
1978 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1979 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1980 &indexi, 1, "");
1981 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1982 /* This is fucked up need to reorganize things */
1983 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1984 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1985 ptr = LLVMBuildGEP(builder,
1986 color_store[cbuf * !cbuf0_write_all][chan],
1987 &indexi, 1, "");
1988 fs_out_color[cbuf][chan][i] = ptr;
1989 }
1990 }
1991 }
1992 }
1993
1994 sampler->destroy(sampler);
1995
1996 /* Loop over color outputs / color buffers to do blending.
1997 */
1998 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1999 LLVMValueRef color_ptr;
2000 LLVMValueRef stride;
2001 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2002 unsigned rt = key->blend.independent_blend_enable ? cbuf : 0;
2003
2004 boolean do_branch = ((key->depth.enabled
2005 || key->stencil[0].enabled
2006 || key->alpha.enabled)
2007 && !shader->info.base.uses_kill);
2008
2009 color_ptr = LLVMBuildLoad(builder,
2010 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2011 "");
2012
2013 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2014
2015 stride = LLVMBuildLoad(builder,
2016 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2017 "");
2018
2019 generate_unswizzled_blend(gallivm, rt, variant, key->cbuf_format[cbuf],
2020 num_fs, fs_type, fs_mask, fs_out_color[cbuf],
2021 context_ptr, color_ptr, stride, partial_mask, do_branch);
2022 }
2023
2024 LLVMBuildRetVoid(builder);
2025
2026 gallivm_verify_function(gallivm, function);
2027
2028 variant->nr_instrs += lp_build_count_instructions(function);
2029 }
2030
2031
2032 static void
2033 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2034 {
2035 unsigned i;
2036
2037 debug_printf("fs variant %p:\n", (void *) key);
2038
2039 if (key->flatshade) {
2040 debug_printf("flatshade = 1\n");
2041 }
2042 for (i = 0; i < key->nr_cbufs; ++i) {
2043 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2044 }
2045 if (key->depth.enabled) {
2046 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2047 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2048 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2049 }
2050
2051 for (i = 0; i < 2; ++i) {
2052 if (key->stencil[i].enabled) {
2053 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2054 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2055 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2056 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2057 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2058 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2059 }
2060 }
2061
2062 if (key->alpha.enabled) {
2063 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2064 }
2065
2066 if (key->occlusion_count) {
2067 debug_printf("occlusion_count = 1\n");
2068 }
2069
2070 if (key->blend.logicop_enable) {
2071 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2072 }
2073 else if (key->blend.rt[0].blend_enable) {
2074 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2075 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2076 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2077 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2078 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2079 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2080 }
2081 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2082 for (i = 0; i < key->nr_samplers; ++i) {
2083 debug_printf("sampler[%u] = \n", i);
2084 debug_printf(" .format = %s\n",
2085 util_format_name(key->sampler[i].format));
2086 debug_printf(" .target = %s\n",
2087 util_dump_tex_target(key->sampler[i].target, TRUE));
2088 debug_printf(" .pot = %u %u %u\n",
2089 key->sampler[i].pot_width,
2090 key->sampler[i].pot_height,
2091 key->sampler[i].pot_depth);
2092 debug_printf(" .wrap = %s %s %s\n",
2093 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
2094 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
2095 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
2096 debug_printf(" .min_img_filter = %s\n",
2097 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
2098 debug_printf(" .min_mip_filter = %s\n",
2099 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
2100 debug_printf(" .mag_img_filter = %s\n",
2101 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
2102 if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
2103 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
2104 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
2105 debug_printf(" .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
2106 debug_printf(" .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
2107 debug_printf(" .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
2108 debug_printf(" .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
2109 }
2110 }
2111
2112
2113 void
2114 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2115 {
2116 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2117 variant->shader->no, variant->no);
2118 tgsi_dump(variant->shader->base.tokens, 0);
2119 dump_fs_variant_key(&variant->key);
2120 debug_printf("variant->opaque = %u\n", variant->opaque);
2121 debug_printf("\n");
2122 }
2123
2124
2125 /**
2126 * Generate a new fragment shader variant from the shader code and
2127 * other state indicated by the key.
2128 */
2129 static struct lp_fragment_shader_variant *
2130 generate_variant(struct llvmpipe_context *lp,
2131 struct lp_fragment_shader *shader,
2132 const struct lp_fragment_shader_variant_key *key)
2133 {
2134 struct lp_fragment_shader_variant *variant;
2135 const struct util_format_description *cbuf0_format_desc;
2136 boolean fullcolormask;
2137
2138 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2139 if(!variant)
2140 return NULL;
2141
2142 variant->gallivm = gallivm_create();
2143 if (!variant->gallivm) {
2144 FREE(variant);
2145 return NULL;
2146 }
2147
2148 variant->shader = shader;
2149 variant->list_item_global.base = variant;
2150 variant->list_item_local.base = variant;
2151 variant->no = shader->variants_created++;
2152
2153 memcpy(&variant->key, key, shader->variant_key_size);
2154
2155 /*
2156 * Determine whether we are touching all channels in the color buffer.
2157 */
2158 fullcolormask = FALSE;
2159 if (key->nr_cbufs == 1) {
2160 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2161 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2162 }
2163
2164 variant->opaque =
2165 !key->blend.logicop_enable &&
2166 !key->blend.rt[0].blend_enable &&
2167 fullcolormask &&
2168 !key->stencil[0].enabled &&
2169 !key->alpha.enabled &&
2170 !key->depth.enabled &&
2171 !shader->info.base.uses_kill
2172 ? TRUE : FALSE;
2173
2174 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2175 lp_debug_fs_variant(variant);
2176 }
2177
2178 lp_jit_init_types(variant);
2179
2180 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2181 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2182
2183 if (variant->jit_function[RAST_WHOLE] == NULL) {
2184 if (variant->opaque) {
2185 /* Specialized shader, which doesn't need to read the color buffer. */
2186 generate_fragment(lp, shader, variant, RAST_WHOLE);
2187 }
2188 }
2189
2190 /*
2191 * Compile everything
2192 */
2193
2194 gallivm_compile_module(variant->gallivm);
2195
2196 if (variant->function[RAST_EDGE_TEST]) {
2197 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2198 gallivm_jit_function(variant->gallivm,
2199 variant->function[RAST_EDGE_TEST]);
2200 }
2201
2202 if (variant->function[RAST_WHOLE]) {
2203 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2204 gallivm_jit_function(variant->gallivm,
2205 variant->function[RAST_WHOLE]);
2206 } else if (!variant->jit_function[RAST_WHOLE]) {
2207 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2208 }
2209
2210 return variant;
2211 }
2212
2213
2214 static void *
2215 llvmpipe_create_fs_state(struct pipe_context *pipe,
2216 const struct pipe_shader_state *templ)
2217 {
2218 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2219 struct lp_fragment_shader *shader;
2220 int nr_samplers;
2221 int i;
2222
2223 shader = CALLOC_STRUCT(lp_fragment_shader);
2224 if (!shader)
2225 return NULL;
2226
2227 shader->no = fs_no++;
2228 make_empty_list(&shader->variants);
2229
2230 /* get/save the summary info for this shader */
2231 lp_build_tgsi_info(templ->tokens, &shader->info);
2232
2233 /* we need to keep a local copy of the tokens */
2234 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2235
2236 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2237 if (shader->draw_data == NULL) {
2238 FREE((void *) shader->base.tokens);
2239 FREE(shader);
2240 return NULL;
2241 }
2242
2243 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2244
2245 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2246 sampler[nr_samplers]);
2247
2248 for (i = 0; i < shader->info.base.num_inputs; i++) {
2249 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2250 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2251
2252 switch (shader->info.base.input_interpolate[i]) {
2253 case TGSI_INTERPOLATE_CONSTANT:
2254 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2255 break;
2256 case TGSI_INTERPOLATE_LINEAR:
2257 shader->inputs[i].interp = LP_INTERP_LINEAR;
2258 break;
2259 case TGSI_INTERPOLATE_PERSPECTIVE:
2260 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2261 break;
2262 case TGSI_INTERPOLATE_COLOR:
2263 shader->inputs[i].interp = LP_INTERP_COLOR;
2264 break;
2265 default:
2266 assert(0);
2267 break;
2268 }
2269
2270 switch (shader->info.base.input_semantic_name[i]) {
2271 case TGSI_SEMANTIC_FACE:
2272 shader->inputs[i].interp = LP_INTERP_FACING;
2273 break;
2274 case TGSI_SEMANTIC_POSITION:
2275 /* Position was already emitted above
2276 */
2277 shader->inputs[i].interp = LP_INTERP_POSITION;
2278 shader->inputs[i].src_index = 0;
2279 continue;
2280 }
2281
2282 shader->inputs[i].src_index = i+1;
2283 }
2284
2285 if (LP_DEBUG & DEBUG_TGSI) {
2286 unsigned attrib;
2287 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2288 shader->no, (void *) shader);
2289 tgsi_dump(templ->tokens, 0);
2290 debug_printf("usage masks:\n");
2291 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2292 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2293 debug_printf(" IN[%u].%s%s%s%s\n",
2294 attrib,
2295 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2296 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2297 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2298 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2299 }
2300 debug_printf("\n");
2301 }
2302
2303 return shader;
2304 }
2305
2306
2307 static void
2308 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2309 {
2310 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2311
2312 if (llvmpipe->fs == fs)
2313 return;
2314
2315 draw_flush(llvmpipe->draw);
2316
2317 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2318
2319 draw_bind_fragment_shader(llvmpipe->draw,
2320 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2321
2322 llvmpipe->dirty |= LP_NEW_FS;
2323 }
2324
2325
2326 /**
2327 * Remove shader variant from two lists: the shader's variant list
2328 * and the context's variant list.
2329 */
2330 void
2331 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2332 struct lp_fragment_shader_variant *variant)
2333 {
2334 unsigned i;
2335
2336 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2337 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2338 " #%u v total cached #%u\n",
2339 variant->shader->no,
2340 variant->no,
2341 variant->shader->variants_created,
2342 variant->shader->variants_cached,
2343 lp->nr_fs_variants);
2344 }
2345
2346 /* free all the variant's JIT'd functions */
2347 for (i = 0; i < Elements(variant->function); i++) {
2348 if (variant->function[i]) {
2349 gallivm_free_function(variant->gallivm,
2350 variant->function[i],
2351 variant->jit_function[i]);
2352 }
2353 }
2354
2355 gallivm_destroy(variant->gallivm);
2356
2357 /* remove from shader's list */
2358 remove_from_list(&variant->list_item_local);
2359 variant->shader->variants_cached--;
2360
2361 /* remove from context's list */
2362 remove_from_list(&variant->list_item_global);
2363 lp->nr_fs_variants--;
2364 lp->nr_fs_instrs -= variant->nr_instrs;
2365
2366 FREE(variant);
2367 }
2368
2369
2370 static void
2371 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2372 {
2373 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2374 struct lp_fragment_shader *shader = fs;
2375 struct lp_fs_variant_list_item *li;
2376
2377 assert(fs != llvmpipe->fs);
2378
2379 /*
2380 * XXX: we need to flush the context until we have some sort of reference
2381 * counting in fragment shaders as they may still be binned
2382 * Flushing alone might not sufficient we need to wait on it too.
2383 */
2384 llvmpipe_finish(pipe, __FUNCTION__);
2385
2386 /* Delete all the variants */
2387 li = first_elem(&shader->variants);
2388 while(!at_end(&shader->variants, li)) {
2389 struct lp_fs_variant_list_item *next = next_elem(li);
2390 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2391 li = next;
2392 }
2393
2394 /* Delete draw module's data */
2395 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2396
2397 assert(shader->variants_cached == 0);
2398 FREE((void *) shader->base.tokens);
2399 FREE(shader);
2400 }
2401
2402
2403
2404 static void
2405 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2406 uint shader, uint index,
2407 struct pipe_constant_buffer *cb)
2408 {
2409 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2410 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2411 unsigned size;
2412 const void *data;
2413
2414 if (cb && cb->user_buffer) {
2415 constants = llvmpipe_user_buffer_create(pipe->screen,
2416 (void *) cb->user_buffer,
2417 cb->buffer_size,
2418 PIPE_BIND_CONSTANT_BUFFER);
2419 }
2420
2421 size = constants ? constants->width0 : 0;
2422 data = constants ? llvmpipe_resource_data(constants) : NULL;
2423
2424 assert(shader < PIPE_SHADER_TYPES);
2425 assert(index < PIPE_MAX_CONSTANT_BUFFERS);
2426
2427 if(llvmpipe->constants[shader][index] == constants)
2428 return;
2429
2430 draw_flush(llvmpipe->draw);
2431
2432 /* note: reference counting */
2433 pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
2434
2435 if(shader == PIPE_SHADER_VERTEX ||
2436 shader == PIPE_SHADER_GEOMETRY) {
2437 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2438 index, data, size);
2439 }
2440
2441 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2442
2443 if (cb && cb->user_buffer) {
2444 pipe_resource_reference(&constants, NULL);
2445 }
2446 }
2447
2448
2449 /**
2450 * Return the blend factor equivalent to a destination alpha of one.
2451 */
2452 static INLINE unsigned
2453 force_dst_alpha_one(unsigned factor)
2454 {
2455 switch(factor) {
2456 case PIPE_BLENDFACTOR_DST_ALPHA:
2457 return PIPE_BLENDFACTOR_ONE;
2458 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2459 return PIPE_BLENDFACTOR_ZERO;
2460 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2461 return PIPE_BLENDFACTOR_ZERO;
2462 }
2463
2464 return factor;
2465 }
2466
2467
2468 /**
2469 * We need to generate several variants of the fragment pipeline to match
2470 * all the combinations of the contributing state atoms.
2471 *
2472 * TODO: there is actually no reason to tie this to context state -- the
2473 * generated code could be cached globally in the screen.
2474 */
2475 static void
2476 make_variant_key(struct llvmpipe_context *lp,
2477 struct lp_fragment_shader *shader,
2478 struct lp_fragment_shader_variant_key *key)
2479 {
2480 unsigned i;
2481
2482 memset(key, 0, shader->variant_key_size);
2483
2484 if (lp->framebuffer.zsbuf) {
2485 if (lp->depth_stencil->depth.enabled) {
2486 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2487 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2488 }
2489 if (lp->depth_stencil->stencil[0].enabled) {
2490 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2491 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2492 }
2493 }
2494
2495 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2496 if(key->alpha.enabled)
2497 key->alpha.func = lp->depth_stencil->alpha.func;
2498 /* alpha.ref_value is passed in jit_context */
2499
2500 key->flatshade = lp->rasterizer->flatshade;
2501 if (lp->active_occlusion_query) {
2502 key->occlusion_count = TRUE;
2503 }
2504
2505 if (lp->framebuffer.nr_cbufs) {
2506 memcpy(&key->blend, lp->blend, sizeof key->blend);
2507 }
2508
2509 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2510 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2511 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2512 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2513 const struct util_format_description *format_desc;
2514
2515 key->cbuf_format[i] = format;
2516
2517 format_desc = util_format_description(format);
2518 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2519 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2520
2521 blend_rt->colormask = lp->blend->rt[i].colormask;
2522
2523 /*
2524 * Mask out color channels not present in the color buffer.
2525 */
2526 blend_rt->colormask &= util_format_colormask(format_desc);
2527
2528 /*
2529 * Our swizzled render tiles always have an alpha channel, but the linear
2530 * render target format often does not, so force here the dst alpha to be
2531 * one.
2532 *
2533 * This is not a mere optimization. Wrong results will be produced if the
2534 * dst alpha is used, the dst format does not have alpha, and the previous
2535 * rendering was not flushed from the swizzled to linear buffer. For
2536 * example, NonPowTwo DCT.
2537 *
2538 * TODO: This should be generalized to all channels for better
2539 * performance, but only alpha causes correctness issues.
2540 *
2541 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2542 */
2543 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2544 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2545 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2546 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2547 blend_rt->alpha_func = blend_rt->rgb_func;
2548 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2549 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2550 }
2551 }
2552
2553 /* This value will be the same for all the variants of a given shader:
2554 */
2555 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2556
2557 for(i = 0; i < key->nr_samplers; ++i) {
2558 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2559 lp_sampler_static_state(&key->sampler[i],
2560 lp->sampler_views[PIPE_SHADER_FRAGMENT][i],
2561 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2562 }
2563 }
2564 }
2565
2566
2567
2568 /**
2569 * Update fragment shader state. This is called just prior to drawing
2570 * something when some fragment-related state has changed.
2571 */
2572 void
2573 llvmpipe_update_fs(struct llvmpipe_context *lp)
2574 {
2575 struct lp_fragment_shader *shader = lp->fs;
2576 struct lp_fragment_shader_variant_key key;
2577 struct lp_fragment_shader_variant *variant = NULL;
2578 struct lp_fs_variant_list_item *li;
2579
2580 make_variant_key(lp, shader, &key);
2581
2582 /* Search the variants for one which matches the key */
2583 li = first_elem(&shader->variants);
2584 while(!at_end(&shader->variants, li)) {
2585 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2586 variant = li->base;
2587 break;
2588 }
2589 li = next_elem(li);
2590 }
2591
2592 if (variant) {
2593 /* Move this variant to the head of the list to implement LRU
2594 * deletion of shader's when we have too many.
2595 */
2596 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2597 }
2598 else {
2599 /* variant not found, create it now */
2600 int64_t t0, t1, dt;
2601 unsigned i;
2602 unsigned variants_to_cull;
2603
2604 if (0) {
2605 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2606 lp->nr_fs_variants,
2607 lp->nr_fs_instrs,
2608 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2609 }
2610
2611 /* First, check if we've exceeded the max number of shader variants.
2612 * If so, free 25% of them (the least recently used ones).
2613 */
2614 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2615
2616 if (variants_to_cull ||
2617 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2618 struct pipe_context *pipe = &lp->pipe;
2619
2620 /*
2621 * XXX: we need to flush the context until we have some sort of
2622 * reference counting in fragment shaders as they may still be binned
2623 * Flushing alone might not be sufficient we need to wait on it too.
2624 */
2625 llvmpipe_finish(pipe, __FUNCTION__);
2626
2627 /*
2628 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2629 * number of shader variants (potentially all of them) could be
2630 * pending for destruction on flush.
2631 */
2632
2633 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2634 struct lp_fs_variant_list_item *item;
2635 if (is_empty_list(&lp->fs_variants_list)) {
2636 break;
2637 }
2638 item = last_elem(&lp->fs_variants_list);
2639 assert(item);
2640 assert(item->base);
2641 llvmpipe_remove_shader_variant(lp, item->base);
2642 }
2643 }
2644
2645 /*
2646 * Generate the new variant.
2647 */
2648 t0 = os_time_get();
2649 variant = generate_variant(lp, shader, &key);
2650 t1 = os_time_get();
2651 dt = t1 - t0;
2652 LP_COUNT_ADD(llvm_compile_time, dt);
2653 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2654
2655 llvmpipe_variant_count++;
2656
2657 /* Put the new variant into the list */
2658 if (variant) {
2659 insert_at_head(&shader->variants, &variant->list_item_local);
2660 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2661 lp->nr_fs_variants++;
2662 lp->nr_fs_instrs += variant->nr_instrs;
2663 shader->variants_cached++;
2664 }
2665 }
2666
2667 /* Bind this variant */
2668 lp_setup_set_fs_variant(lp->setup, variant);
2669 }
2670
2671
2672
2673
2674
2675
2676
2677 void
2678 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2679 {
2680 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2681 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2682 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2683
2684 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2685 }