projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: split sampler and texture state
[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 normalized 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 repeatedly 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 if (src_type.norm) {
1147 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1148 blend_type.width, chans[j], src_type);
1149 }
1150
1151 /* Insert bits into correct position */
1152 chans[j] = LLVMBuildShl(builder,
1153 chans[j],
1154 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1155 "");
1156
1157 sa += src_fmt->channel[j].size;
1158
1159 if (j == 0) {
1160 res = chans[j];
1161 } else {
1162 res = LLVMBuildOr(builder, res, chans[j], "");
1163 }
1164 }
1165
1166 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1167 }
1168 }
1169
1170
1171 /**
1172 * Convert from blending format to memory format
1173 *
1174 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1175 */
1176 static void
1177 convert_from_blend_type(struct gallivm_state *gallivm,
1178 const struct util_format_description *src_fmt,
1179 struct lp_type src_type,
1180 struct lp_type dst_type,
1181 LLVMValueRef* src, // and dst
1182 unsigned num_srcs)
1183 {
1184 LLVMValueRef* dst = src;
1185 unsigned i, j, k;
1186 struct lp_type mem_type;
1187 struct lp_type blend_type;
1188 LLVMBuilderRef builder = gallivm->builder;
1189 unsigned pixels = 16 / num_srcs;
1190 bool is_arith;
1191
1192 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1193 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1194
1195 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1196
1197 /* Special case for half-floats */
1198 if (mem_type.width == 16 && mem_type.floating) {
1199 int length = dst_type.length;
1200 assert(blend_type.width == 32 && blend_type.floating);
1201
1202 dst_type.length = src_type.length;
1203
1204 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1205
1206 dst_type.length = length;
1207 is_arith = false;
1208 }
1209
1210 /* Remove any padding */
1211 if (!is_arith && (src_type.length % mem_type.length)) {
1212 src_type.length -= (src_type.length % mem_type.length);
1213
1214 for (i = 0; i < num_srcs; ++i) {
1215 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1216 }
1217 }
1218
1219 /* No bit arithmetic to do */
1220 if (!is_arith) {
1221 return;
1222 }
1223
1224 src_type.length = pixels;
1225 src_type.width = blend_type.length * blend_type.width;
1226 dst_type.length = pixels;
1227
1228 for (i = 0; i < num_srcs; ++i) {
1229 LLVMValueRef chans[4];
1230 LLVMValueRef res;
1231 unsigned sa = 0;
1232
1233 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1234
1235 for (j = 0; j < src_fmt->nr_channels; ++j) {
1236 unsigned mask = 0;
1237
1238 assert(blend_type.width > src_fmt->channel[j].size);
1239
1240 for (k = 0; k < blend_type.width; ++k) {
1241 mask |= 1 << k;
1242 }
1243
1244 /* Extract bits */
1245 chans[j] = LLVMBuildLShr(builder,
1246 dst[i],
1247 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1248 "");
1249
1250 chans[j] = LLVMBuildAnd(builder,
1251 chans[j],
1252 lp_build_const_int_vec(gallivm, src_type, mask),
1253 "");
1254
1255 /* Scale down bits */
1256 if (src_type.norm) {
1257 chans[j] = scale_bits(gallivm, blend_type.width,
1258 src_fmt->channel[j].size, chans[j], src_type);
1259 }
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 unsigned dst_alignment;
1337
1338 bool pad_inline = is_arithmetic_format(out_format_desc);
1339 bool has_alpha = false;
1340
1341 src_channels = TGSI_NUM_CHANNELS;
1342 mask_type = lp_int32_vec4_type();
1343 mask_type.length = fs_type.length;
1344
1345 /* Compute the alignment of the destination pointer in bytes */
1346 #if 0
1347 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1348 #else
1349 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1350 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1351 #endif
1352 /* Force power-of-two alignment by extracting only the least-significant-bit */
1353 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1354 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1355 dst_alignment = MIN2(dst_alignment, 16);
1356
1357 /* Do not bother executing code when mask is empty.. */
1358 if (do_branch) {
1359 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1360
1361 for (i = 0; i < num_fs; ++i) {
1362 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1363 }
1364
1365 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1366 lp_build_mask_check(&mask_ctx);
1367 }
1368
1369 partial_mask |= !variant->opaque;
1370 i32_zero = lp_build_const_int32(gallivm, 0);
1371
1372 /* Get type from output format */
1373 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1374 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1375
1376 row_type.length = fs_type.length;
1377 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1378
1379 /* Compute correct swizzle and count channels */
1380 memset(swizzle, 0xFF, TGSI_NUM_CHANNELS);
1381 dst_channels = 0;
1382
1383 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1384 /* Ensure channel is used */
1385 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1386 continue;
1387 }
1388
1389 /* Ensure not already written to (happens in case with GL_ALPHA) */
1390 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1391 continue;
1392 }
1393
1394 /* Ensure we havn't already found all channels */
1395 if (dst_channels >= out_format_desc->nr_channels) {
1396 continue;
1397 }
1398
1399 swizzle[out_format_desc->swizzle[i]] = i;
1400 ++dst_channels;
1401
1402 if (i == alpha_channel) {
1403 has_alpha = true;
1404 }
1405 }
1406
1407 /* If 3 channels then pad to include alpha for 4 element transpose */
1408 if (dst_channels == 3 && !has_alpha) {
1409 swizzle[3] = 3;
1410
1411 if (out_format_desc->nr_channels == 4) {
1412 dst_channels = 4;
1413 }
1414 }
1415
1416 /*
1417 * Load shader output
1418 */
1419 for (i = 0; i < num_fs; ++i) {
1420 /* Always load alpha for use in blending */
1421 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[alpha_channel][i], "");
1422
1423 /* Load each channel */
1424 for (j = 0; j < dst_channels; ++j) {
1425 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[swizzle[j]][i], "");
1426 }
1427
1428 /* If 3 channels then pad to include alpha for 4 element transpose */
1429 if (dst_channels == 3 && !has_alpha) {
1430 fs_src[i][3] = alpha;
1431 swizzle[3] = 3;
1432 }
1433
1434 /* We split the row_mask and row_alpha as we want 128bit interleave */
1435 if (fs_type.length == 8) {
1436 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1437 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1438
1439 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1440 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1441 } else {
1442 src_mask[i] = fs_mask[i];
1443 src_alpha[i] = alpha;
1444 }
1445 }
1446
1447 if (util_format_is_pure_integer(out_format)) {
1448 /*
1449 * In this case fs_type was really ints or uints disguised as floats,
1450 * fix that up now.
1451 */
1452 fs_type.floating = 0;
1453 fs_type.sign = dst_type.sign;
1454 for (i = 0; i < num_fs; ++i) {
1455 for (j = 0; j < dst_channels; ++j) {
1456 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1457 lp_build_vec_type(gallivm, fs_type), "");
1458 }
1459 if (dst_channels == 3 && !has_alpha) {
1460 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1461 lp_build_vec_type(gallivm, fs_type), "");
1462 }
1463 }
1464 }
1465
1466
1467 /*
1468 * Pixel twiddle from fragment shader order to memory order
1469 */
1470 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1471 src_channels = dst_channels < 3 ? dst_channels : 4;
1472 if (src_count != num_fs * src_channels) {
1473 unsigned ds = src_count / (num_fs * src_channels);
1474 row_type.length /= ds;
1475 fs_type.length = row_type.length;
1476 }
1477
1478 blend_type = row_type;
1479 alpha_type = fs_type;
1480 alpha_type.length = 4;
1481 mask_type.length = 4;
1482
1483 /* Convert src to row_type */
1484 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1485
1486 /* If the rows are not an SSE vector, combine them to become SSE size! */
1487 if ((row_type.width * row_type.length) % 128) {
1488 unsigned bits = row_type.width * row_type.length;
1489 unsigned combined;
1490
1491 dst_count = src_count / (vector_width / bits);
1492 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1493
1494 row_type.length *= combined;
1495 src_count /= combined;
1496
1497 bits = row_type.width * row_type.length;
1498 assert(bits == 128 || bits == 256);
1499 }
1500
1501
1502 /*
1503 * Blend Colour conversion
1504 */
1505 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1506 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1507 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1508
1509 /* Convert */
1510 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1511
1512 /* Extract alpha */
1513 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1514
1515 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1516 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1517 if (pad_inline) {
1518 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1519 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1520 } else {
1521 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1522 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1523 }
1524
1525 /*
1526 * Mask conversion
1527 */
1528 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1529
1530 if (src_count < block_height) {
1531 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1532 } else if (src_count > block_height) {
1533 for (i = src_count; i > 0; --i) {
1534 unsigned pixels = block_size / src_count;
1535 unsigned idx = i - 1;
1536
1537 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1538 }
1539 }
1540
1541 assert(mask_type.width == 32);
1542
1543 for (i = 0; i < src_count; ++i) {
1544 unsigned pixels = block_size / src_count;
1545 unsigned pixel_width = row_type.width * dst_channels;
1546
1547 if (pixel_width == 24) {
1548 mask_type.width = 8;
1549 mask_type.length = vector_width / mask_type.width;
1550 } else {
1551 mask_type.length = pixels;
1552 mask_type.width = row_type.width * dst_channels;
1553
1554 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1555
1556 mask_type.length *= dst_channels;
1557 mask_type.width /= dst_channels;
1558 }
1559
1560 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1561 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1562 }
1563
1564 /*
1565 * Alpha conversion
1566 */
1567 if (!has_alpha) {
1568 unsigned length = row_type.length;
1569 row_type.length = alpha_type.length;
1570
1571 /* Twiddle the alpha to match pixels */
1572 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1573
1574 for (i = 0; i < 4; ++i) {
1575 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1576 }
1577
1578 alpha_type = row_type;
1579 row_type.length = length;
1580
1581 /* If only one channel we can only need the single alpha value per pixel */
1582 if (src_count == 1) {
1583 assert(dst_channels == 1);
1584
1585 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1586 } else {
1587 /* If there are more srcs than rows then we need to split alpha up */
1588 if (src_count > block_height) {
1589 for (i = src_count; i > 0; --i) {
1590 unsigned pixels = block_size / src_count;
1591 unsigned idx = i - 1;
1592
1593 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1594 }
1595 }
1596
1597 /* If there is a src for each pixel broadcast the alpha across whole row */
1598 if (src_count == block_size) {
1599 for (i = 0; i < src_count; ++i) {
1600 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1601 }
1602 } else {
1603 unsigned pixels = block_size / src_count;
1604 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1605 unsigned alpha_span = 1;
1606 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1607
1608 /* Check if we need 2 src_alphas for our shuffles */
1609 if (pixels > alpha_type.length) {
1610 alpha_span = 2;
1611 }
1612
1613 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1614 for (j = 0; j < row_type.length; ++j) {
1615 if (j < pixels * channels) {
1616 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1617 } else {
1618 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1619 }
1620 }
1621
1622 for (i = 0; i < src_count; ++i) {
1623 unsigned idx1 = i, idx2 = i;
1624
1625 if (alpha_span > 1){
1626 idx1 *= alpha_span;
1627 idx2 = idx1 + 1;
1628 }
1629
1630 src_alpha[i] = LLVMBuildShuffleVector(builder,
1631 src_alpha[idx1],
1632 src_alpha[idx2],
1633 LLVMConstVector(shuffles, row_type.length),
1634 "");
1635 }
1636 }
1637 }
1638 }
1639
1640
1641 /*
1642 * Load dst from memory
1643 */
1644 if (src_count < block_height) {
1645 dst_count = block_height;
1646 } else {
1647 dst_count = src_count;
1648 }
1649
1650 dst_type.length *= 16 / dst_count;
1651
1652 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1653 dst, dst_type, dst_count, dst_alignment);
1654
1655
1656 /*
1657 * Convert from dst/output format to src/blending format.
1658 *
1659 * This is necessary as we can only read 1 row from memory at a time,
1660 * so the minimum dst_count will ever be at this point is 4.
1661 *
1662 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1663 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1664 * on all 16 pixels in that single vector at once.
1665 */
1666 if (dst_count > src_count) {
1667 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1668 }
1669
1670 /*
1671 * Blending
1672 */
1673 /* XXX this is broken for RGB8 formats -
1674 * they get expanded from 12 to 16 elements (to include alpha)
1675 * by convert_to_blend_type then reduced to 15 instead of 12
1676 * by convert_from_blend_type (a simple fix though breaks A8...).
1677 * R16G16B16 also crashes differently however something going wrong
1678 * inside llvm handling npot vector sizes seemingly.
1679 * It seems some cleanup could be done here (like skipping conversion/blend
1680 * when not needed).
1681 */
1682 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1683
1684 for (i = 0; i < src_count; ++i) {
1685 dst[i] = lp_build_blend_aos(gallivm,
1686 &variant->key.blend,
1687 out_format,
1688 row_type,
1689 rt,
1690 src[i],
1691 has_alpha ? NULL : src_alpha[i],
1692 dst[i],
1693 partial_mask ? src_mask[i] : NULL,
1694 blend_color,
1695 has_alpha ? NULL : blend_alpha,
1696 swizzle,
1697 pad_inline ? 4 : dst_channels);
1698 }
1699
1700 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1701
1702 /* Split the blend rows back to memory rows */
1703 if (dst_count > src_count) {
1704 row_type.length = dst_type.length * (dst_count / src_count);
1705
1706 if (src_count == 1) {
1707 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1708 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1709
1710 row_type.length /= 2;
1711 src_count *= 2;
1712 }
1713
1714 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1715 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1716 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1717 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1718
1719 row_type.length /= 2;
1720 src_count *= 2;
1721 }
1722
1723
1724 /*
1725 * Store blend result to memory
1726 */
1727 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1728 dst, dst_type, dst_count, dst_alignment);
1729
1730 if (do_branch) {
1731 lp_build_mask_end(&mask_ctx);
1732 }
1733 }
1734
1735
1736 /**
1737 * Generate the runtime callable function for the whole fragment pipeline.
1738 * Note that the function which we generate operates on a block of 16
1739 * pixels at at time. The block contains 2x2 quads. Each quad contains
1740 * 2x2 pixels.
1741 */
1742 static void
1743 generate_fragment(struct llvmpipe_context *lp,
1744 struct lp_fragment_shader *shader,
1745 struct lp_fragment_shader_variant *variant,
1746 unsigned partial_mask)
1747 {
1748 struct gallivm_state *gallivm = variant->gallivm;
1749 const struct lp_fragment_shader_variant_key *key = &variant->key;
1750 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1751 char func_name[256];
1752 struct lp_type fs_type;
1753 struct lp_type blend_type;
1754 LLVMTypeRef fs_elem_type;
1755 LLVMTypeRef blend_vec_type;
1756 LLVMTypeRef arg_types[12];
1757 LLVMTypeRef func_type;
1758 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1759 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1760 LLVMValueRef context_ptr;
1761 LLVMValueRef x;
1762 LLVMValueRef y;
1763 LLVMValueRef a0_ptr;
1764 LLVMValueRef dadx_ptr;
1765 LLVMValueRef dady_ptr;
1766 LLVMValueRef color_ptr_ptr;
1767 LLVMValueRef stride_ptr;
1768 LLVMValueRef depth_ptr;
1769 LLVMValueRef mask_input;
1770 LLVMValueRef counter = NULL;
1771 LLVMBasicBlockRef block;
1772 LLVMBuilderRef builder;
1773 struct lp_build_sampler_soa *sampler;
1774 struct lp_build_interp_soa_context interp;
1775 LLVMValueRef fs_mask[16 / 4];
1776 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1777 LLVMValueRef function;
1778 LLVMValueRef facing;
1779 const struct util_format_description *zs_format_desc;
1780 unsigned num_fs;
1781 unsigned i;
1782 unsigned chan;
1783 unsigned cbuf;
1784 boolean cbuf0_write_all;
1785 boolean try_loop = TRUE;
1786
1787 assert(lp_native_vector_width / 32 >= 4);
1788
1789 /* Adjust color input interpolation according to flatshade state:
1790 */
1791 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1792 for (i = 0; i < shader->info.base.num_inputs; i++) {
1793 if (inputs[i].interp == LP_INTERP_COLOR) {
1794 if (key->flatshade)
1795 inputs[i].interp = LP_INTERP_CONSTANT;
1796 else
1797 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1798 }
1799 }
1800
1801 /* check if writes to cbuf[0] are to be copied to all cbufs */
1802 cbuf0_write_all = FALSE;
1803 for (i = 0;i < shader->info.base.num_properties; i++) {
1804 if (shader->info.base.properties[i].name ==
1805 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1806 cbuf0_write_all = TRUE;
1807 break;
1808 }
1809 }
1810
1811 /* TODO: actually pick these based on the fs and color buffer
1812 * characteristics. */
1813
1814 memset(&fs_type, 0, sizeof fs_type);
1815 fs_type.floating = TRUE; /* floating point values */
1816 fs_type.sign = TRUE; /* values are signed */
1817 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1818 fs_type.width = 32; /* 32-bit float */
1819 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1820 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1821
1822 memset(&blend_type, 0, sizeof blend_type);
1823 blend_type.floating = FALSE; /* values are integers */
1824 blend_type.sign = FALSE; /* values are unsigned */
1825 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1826 blend_type.width = 8; /* 8-bit ubyte values */
1827 blend_type.length = 16; /* 16 elements per vector */
1828
1829 /*
1830 * Generate the function prototype. Any change here must be reflected in
1831 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1832 */
1833
1834 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1835
1836 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1837
1838 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1839 shader->no, variant->no, partial_mask ? "partial" : "whole");
1840
1841 arg_types[0] = variant->jit_context_ptr_type; /* context */
1842 arg_types[1] = int32_type; /* x */
1843 arg_types[2] = int32_type; /* y */
1844 arg_types[3] = int32_type; /* facing */
1845 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1846 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1847 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1848 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1849 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1850 arg_types[9] = int32_type; /* mask_input */
1851 arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */
1852 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1853
1854 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1855 arg_types, Elements(arg_types), 0);
1856
1857 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1858 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1859
1860 variant->function[partial_mask] = function;
1861
1862 /* XXX: need to propagate noalias down into color param now we are
1863 * passing a pointer-to-pointer?
1864 */
1865 for(i = 0; i < Elements(arg_types); ++i)
1866 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1867 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1868
1869 context_ptr = LLVMGetParam(function, 0);
1870 x = LLVMGetParam(function, 1);
1871 y = LLVMGetParam(function, 2);
1872 facing = LLVMGetParam(function, 3);
1873 a0_ptr = LLVMGetParam(function, 4);
1874 dadx_ptr = LLVMGetParam(function, 5);
1875 dady_ptr = LLVMGetParam(function, 6);
1876 color_ptr_ptr = LLVMGetParam(function, 7);
1877 depth_ptr = LLVMGetParam(function, 8);
1878 mask_input = LLVMGetParam(function, 9);
1879 stride_ptr = LLVMGetParam(function, 11);
1880
1881 lp_build_name(context_ptr, "context");
1882 lp_build_name(x, "x");
1883 lp_build_name(y, "y");
1884 lp_build_name(a0_ptr, "a0");
1885 lp_build_name(dadx_ptr, "dadx");
1886 lp_build_name(dady_ptr, "dady");
1887 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1888 lp_build_name(depth_ptr, "depth");
1889 lp_build_name(mask_input, "mask_input");
1890 lp_build_name(stride_ptr, "stride_ptr");
1891
1892 if (key->occlusion_count) {
1893 counter = LLVMGetParam(function, 10);
1894 lp_build_name(counter, "counter");
1895 }
1896
1897 /*
1898 * Function body
1899 */
1900
1901 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1902 builder = gallivm->builder;
1903 assert(builder);
1904 LLVMPositionBuilderAtEnd(builder, block);
1905
1906 /* code generated texture sampling */
1907 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
1908
1909 zs_format_desc = util_format_description(key->zsbuf_format);
1910
1911 if (!try_loop) {
1912 /*
1913 * The shader input interpolation info is not explicitely baked in the
1914 * shader key, but everything it derives from (TGSI, and flatshade) is
1915 * already included in the shader key.
1916 */
1917 lp_build_interp_soa_init(&interp,
1918 gallivm,
1919 shader->info.base.num_inputs,
1920 inputs,
1921 builder, fs_type,
1922 FALSE,
1923 a0_ptr, dadx_ptr, dady_ptr,
1924 x, y);
1925
1926 /* loop over quads in the block */
1927 for(i = 0; i < num_fs; ++i) {
1928 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
1929 i*fs_type.length*zs_format_desc->block.bits/8,
1930 0);
1931 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1932 LLVMValueRef depth_ptr_i;
1933
1934 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
1935
1936 generate_fs(gallivm,
1937 shader, key,
1938 builder,
1939 fs_type,
1940 context_ptr,
1941 i,
1942 &interp,
1943 sampler,
1944 &fs_mask[i], /* output */
1945 out_color,
1946 depth_ptr_i,
1947 facing,
1948 partial_mask,
1949 mask_input,
1950 counter);
1951
1952 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
1953 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1954 fs_out_color[cbuf][chan][i] =
1955 out_color[cbuf * !cbuf0_write_all][chan];
1956 }
1957 }
1958 else {
1959 unsigned depth_bits = zs_format_desc->block.bits/8;
1960 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1961 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1962 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1963 num_loop, "mask_store");
1964 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1965
1966 /*
1967 * The shader input interpolation info is not explicitely baked in the
1968 * shader key, but everything it derives from (TGSI, and flatshade) is
1969 * already included in the shader key.
1970 */
1971 lp_build_interp_soa_init(&interp,
1972 gallivm,
1973 shader->info.base.num_inputs,
1974 inputs,
1975 builder, fs_type,
1976 TRUE,
1977 a0_ptr, dadx_ptr, dady_ptr,
1978 x, y);
1979
1980 for (i = 0; i < num_fs; i++) {
1981 LLVMValueRef mask;
1982 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1983 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1984 &indexi, 1, "mask_ptr");
1985
1986 if (partial_mask) {
1987 mask = generate_quad_mask(gallivm, fs_type,
1988 i*fs_type.length/4, mask_input);
1989 }
1990 else {
1991 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1992 }
1993 LLVMBuildStore(builder, mask, mask_ptr);
1994 }
1995
1996 generate_fs_loop(gallivm,
1997 shader, key,
1998 builder,
1999 fs_type,
2000 context_ptr,
2001 num_loop,
2002 &interp,
2003 sampler,
2004 mask_store, /* output */
2005 color_store,
2006 depth_ptr,
2007 depth_bits,
2008 facing,
2009 counter);
2010
2011 for (i = 0; i < num_fs; i++) {
2012 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2013 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2014 &indexi, 1, "");
2015 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2016 /* This is fucked up need to reorganize things */
2017 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2018 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2019 ptr = LLVMBuildGEP(builder,
2020 color_store[cbuf * !cbuf0_write_all][chan],
2021 &indexi, 1, "");
2022 fs_out_color[cbuf][chan][i] = ptr;
2023 }
2024 }
2025 }
2026 }
2027
2028 sampler->destroy(sampler);
2029
2030 /* Loop over color outputs / color buffers to do blending.
2031 */
2032 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2033 LLVMValueRef color_ptr;
2034 LLVMValueRef stride;
2035 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2036
2037 boolean do_branch = ((key->depth.enabled
2038 || key->stencil[0].enabled
2039 || key->alpha.enabled)
2040 && !shader->info.base.uses_kill);
2041
2042 color_ptr = LLVMBuildLoad(builder,
2043 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2044 "");
2045
2046 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2047
2048 stride = LLVMBuildLoad(builder,
2049 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2050 "");
2051
2052 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2053 num_fs, fs_type, fs_mask, fs_out_color[cbuf],
2054 context_ptr, color_ptr, stride, partial_mask, do_branch);
2055 }
2056
2057 LLVMBuildRetVoid(builder);
2058
2059 gallivm_verify_function(gallivm, function);
2060
2061 variant->nr_instrs += lp_build_count_instructions(function);
2062 }
2063
2064
2065 static void
2066 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2067 {
2068 unsigned i;
2069
2070 debug_printf("fs variant %p:\n", (void *) key);
2071
2072 if (key->flatshade) {
2073 debug_printf("flatshade = 1\n");
2074 }
2075 for (i = 0; i < key->nr_cbufs; ++i) {
2076 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2077 }
2078 if (key->depth.enabled) {
2079 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2080 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2081 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2082 }
2083
2084 for (i = 0; i < 2; ++i) {
2085 if (key->stencil[i].enabled) {
2086 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2087 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2088 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2089 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2090 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2091 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2092 }
2093 }
2094
2095 if (key->alpha.enabled) {
2096 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2097 }
2098
2099 if (key->occlusion_count) {
2100 debug_printf("occlusion_count = 1\n");
2101 }
2102
2103 if (key->blend.logicop_enable) {
2104 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2105 }
2106 else if (key->blend.rt[0].blend_enable) {
2107 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2108 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2109 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2110 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2111 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2112 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2113 }
2114 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2115 for (i = 0; i < key->nr_samplers; ++i) {
2116 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2117 debug_printf("sampler[%u] = \n", i);
2118 debug_printf(" .wrap = %s %s %s\n",
2119 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2120 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2121 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2122 debug_printf(" .min_img_filter = %s\n",
2123 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2124 debug_printf(" .min_mip_filter = %s\n",
2125 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2126 debug_printf(" .mag_img_filter = %s\n",
2127 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2128 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2129 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2130 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2131 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2132 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2133 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2134 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2135 }
2136 for (i = 0; i < key->nr_sampler_views; ++i) {
2137 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2138 debug_printf("texture[%u] = \n", i);
2139 debug_printf(" .format = %s\n",
2140 util_format_name(texture->format));
2141 debug_printf(" .target = %s\n",
2142 util_dump_tex_target(texture->target, TRUE));
2143 debug_printf(" .level_zero_only = %u\n",
2144 texture->level_zero_only);
2145 debug_printf(" .pot = %u %u %u\n",
2146 texture->pot_width,
2147 texture->pot_height,
2148 texture->pot_depth);
2149 }
2150 }
2151
2152
2153 void
2154 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2155 {
2156 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2157 variant->shader->no, variant->no);
2158 tgsi_dump(variant->shader->base.tokens, 0);
2159 dump_fs_variant_key(&variant->key);
2160 debug_printf("variant->opaque = %u\n", variant->opaque);
2161 debug_printf("\n");
2162 }
2163
2164
2165 /**
2166 * Generate a new fragment shader variant from the shader code and
2167 * other state indicated by the key.
2168 */
2169 static struct lp_fragment_shader_variant *
2170 generate_variant(struct llvmpipe_context *lp,
2171 struct lp_fragment_shader *shader,
2172 const struct lp_fragment_shader_variant_key *key)
2173 {
2174 struct lp_fragment_shader_variant *variant;
2175 const struct util_format_description *cbuf0_format_desc;
2176 boolean fullcolormask;
2177
2178 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2179 if(!variant)
2180 return NULL;
2181
2182 variant->gallivm = gallivm_create();
2183 if (!variant->gallivm) {
2184 FREE(variant);
2185 return NULL;
2186 }
2187
2188 variant->shader = shader;
2189 variant->list_item_global.base = variant;
2190 variant->list_item_local.base = variant;
2191 variant->no = shader->variants_created++;
2192
2193 memcpy(&variant->key, key, shader->variant_key_size);
2194
2195 /*
2196 * Determine whether we are touching all channels in the color buffer.
2197 */
2198 fullcolormask = FALSE;
2199 if (key->nr_cbufs == 1) {
2200 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2201 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2202 }
2203
2204 variant->opaque =
2205 !key->blend.logicop_enable &&
2206 !key->blend.rt[0].blend_enable &&
2207 fullcolormask &&
2208 !key->stencil[0].enabled &&
2209 !key->alpha.enabled &&
2210 !key->depth.enabled &&
2211 !shader->info.base.uses_kill
2212 ? TRUE : FALSE;
2213
2214 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2215 lp_debug_fs_variant(variant);
2216 }
2217
2218 lp_jit_init_types(variant);
2219
2220 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2221 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2222
2223 if (variant->jit_function[RAST_WHOLE] == NULL) {
2224 if (variant->opaque) {
2225 /* Specialized shader, which doesn't need to read the color buffer. */
2226 generate_fragment(lp, shader, variant, RAST_WHOLE);
2227 }
2228 }
2229
2230 /*
2231 * Compile everything
2232 */
2233
2234 gallivm_compile_module(variant->gallivm);
2235
2236 if (variant->function[RAST_EDGE_TEST]) {
2237 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2238 gallivm_jit_function(variant->gallivm,
2239 variant->function[RAST_EDGE_TEST]);
2240 }
2241
2242 if (variant->function[RAST_WHOLE]) {
2243 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2244 gallivm_jit_function(variant->gallivm,
2245 variant->function[RAST_WHOLE]);
2246 } else if (!variant->jit_function[RAST_WHOLE]) {
2247 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2248 }
2249
2250 return variant;
2251 }
2252
2253
2254 static void *
2255 llvmpipe_create_fs_state(struct pipe_context *pipe,
2256 const struct pipe_shader_state *templ)
2257 {
2258 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2259 struct lp_fragment_shader *shader;
2260 int nr_samplers;
2261 int nr_sampler_views;
2262 int i;
2263
2264 shader = CALLOC_STRUCT(lp_fragment_shader);
2265 if (!shader)
2266 return NULL;
2267
2268 shader->no = fs_no++;
2269 make_empty_list(&shader->variants);
2270
2271 /* get/save the summary info for this shader */
2272 lp_build_tgsi_info(templ->tokens, &shader->info);
2273
2274 /* we need to keep a local copy of the tokens */
2275 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2276
2277 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2278 if (shader->draw_data == NULL) {
2279 FREE((void *) shader->base.tokens);
2280 FREE(shader);
2281 return NULL;
2282 }
2283
2284 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2285 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2286
2287 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2288 state[MAX2(nr_samplers, nr_sampler_views)]);
2289
2290 for (i = 0; i < shader->info.base.num_inputs; i++) {
2291 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2292 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2293
2294 switch (shader->info.base.input_interpolate[i]) {
2295 case TGSI_INTERPOLATE_CONSTANT:
2296 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2297 break;
2298 case TGSI_INTERPOLATE_LINEAR:
2299 shader->inputs[i].interp = LP_INTERP_LINEAR;
2300 break;
2301 case TGSI_INTERPOLATE_PERSPECTIVE:
2302 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2303 break;
2304 case TGSI_INTERPOLATE_COLOR:
2305 shader->inputs[i].interp = LP_INTERP_COLOR;
2306 break;
2307 default:
2308 assert(0);
2309 break;
2310 }
2311
2312 switch (shader->info.base.input_semantic_name[i]) {
2313 case TGSI_SEMANTIC_FACE:
2314 shader->inputs[i].interp = LP_INTERP_FACING;
2315 break;
2316 case TGSI_SEMANTIC_POSITION:
2317 /* Position was already emitted above
2318 */
2319 shader->inputs[i].interp = LP_INTERP_POSITION;
2320 shader->inputs[i].src_index = 0;
2321 continue;
2322 }
2323
2324 shader->inputs[i].src_index = i+1;
2325 }
2326
2327 if (LP_DEBUG & DEBUG_TGSI) {
2328 unsigned attrib;
2329 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2330 shader->no, (void *) shader);
2331 tgsi_dump(templ->tokens, 0);
2332 debug_printf("usage masks:\n");
2333 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2334 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2335 debug_printf(" IN[%u].%s%s%s%s\n",
2336 attrib,
2337 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2338 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2339 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2340 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2341 }
2342 debug_printf("\n");
2343 }
2344
2345 return shader;
2346 }
2347
2348
2349 static void
2350 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2351 {
2352 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2353
2354 if (llvmpipe->fs == fs)
2355 return;
2356
2357 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2358
2359 draw_bind_fragment_shader(llvmpipe->draw,
2360 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2361
2362 llvmpipe->dirty |= LP_NEW_FS;
2363 }
2364
2365
2366 /**
2367 * Remove shader variant from two lists: the shader's variant list
2368 * and the context's variant list.
2369 */
2370 void
2371 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2372 struct lp_fragment_shader_variant *variant)
2373 {
2374 unsigned i;
2375
2376 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2377 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2378 " #%u v total cached #%u\n",
2379 variant->shader->no,
2380 variant->no,
2381 variant->shader->variants_created,
2382 variant->shader->variants_cached,
2383 lp->nr_fs_variants);
2384 }
2385
2386 /* free all the variant's JIT'd functions */
2387 for (i = 0; i < Elements(variant->function); i++) {
2388 if (variant->function[i]) {
2389 gallivm_free_function(variant->gallivm,
2390 variant->function[i],
2391 variant->jit_function[i]);
2392 }
2393 }
2394
2395 gallivm_destroy(variant->gallivm);
2396
2397 /* remove from shader's list */
2398 remove_from_list(&variant->list_item_local);
2399 variant->shader->variants_cached--;
2400
2401 /* remove from context's list */
2402 remove_from_list(&variant->list_item_global);
2403 lp->nr_fs_variants--;
2404 lp->nr_fs_instrs -= variant->nr_instrs;
2405
2406 FREE(variant);
2407 }
2408
2409
2410 static void
2411 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2412 {
2413 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2414 struct lp_fragment_shader *shader = fs;
2415 struct lp_fs_variant_list_item *li;
2416
2417 assert(fs != llvmpipe->fs);
2418
2419 /*
2420 * XXX: we need to flush the context until we have some sort of reference
2421 * counting in fragment shaders as they may still be binned
2422 * Flushing alone might not sufficient we need to wait on it too.
2423 */
2424 llvmpipe_finish(pipe, __FUNCTION__);
2425
2426 /* Delete all the variants */
2427 li = first_elem(&shader->variants);
2428 while(!at_end(&shader->variants, li)) {
2429 struct lp_fs_variant_list_item *next = next_elem(li);
2430 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2431 li = next;
2432 }
2433
2434 /* Delete draw module's data */
2435 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2436
2437 assert(shader->variants_cached == 0);
2438 FREE((void *) shader->base.tokens);
2439 FREE(shader);
2440 }
2441
2442
2443
2444 static void
2445 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2446 uint shader, uint index,
2447 struct pipe_constant_buffer *cb)
2448 {
2449 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2450 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2451
2452 assert(shader < PIPE_SHADER_TYPES);
2453 assert(index < Elements(llvmpipe->constants[shader]));
2454
2455 /* note: reference counting */
2456 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2457
2458 if (shader == PIPE_SHADER_VERTEX ||
2459 shader == PIPE_SHADER_GEOMETRY) {
2460 /* Pass the constants to the 'draw' module */
2461 const unsigned size = cb ? cb->buffer_size : 0;
2462 const ubyte *data;
2463
2464 if (constants) {
2465 data = (ubyte *) llvmpipe_resource_data(constants);
2466 }
2467 else if (cb && cb->user_buffer) {
2468 data = (ubyte *) cb->user_buffer;
2469 }
2470 else {
2471 data = NULL;
2472 }
2473
2474 if (data)
2475 data += cb->buffer_offset;
2476
2477 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2478 index, data, size);
2479 }
2480
2481 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2482
2483 if (cb && cb->user_buffer) {
2484 pipe_resource_reference(&constants, NULL);
2485 }
2486 }
2487
2488
2489 /**
2490 * Return the blend factor equivalent to a destination alpha of one.
2491 */
2492 static INLINE unsigned
2493 force_dst_alpha_one(unsigned factor)
2494 {
2495 switch(factor) {
2496 case PIPE_BLENDFACTOR_DST_ALPHA:
2497 return PIPE_BLENDFACTOR_ONE;
2498 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2499 return PIPE_BLENDFACTOR_ZERO;
2500 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2501 return PIPE_BLENDFACTOR_ZERO;
2502 }
2503
2504 return factor;
2505 }
2506
2507
2508 /**
2509 * We need to generate several variants of the fragment pipeline to match
2510 * all the combinations of the contributing state atoms.
2511 *
2512 * TODO: there is actually no reason to tie this to context state -- the
2513 * generated code could be cached globally in the screen.
2514 */
2515 static void
2516 make_variant_key(struct llvmpipe_context *lp,
2517 struct lp_fragment_shader *shader,
2518 struct lp_fragment_shader_variant_key *key)
2519 {
2520 unsigned i;
2521
2522 memset(key, 0, shader->variant_key_size);
2523
2524 if (lp->framebuffer.zsbuf) {
2525 if (lp->depth_stencil->depth.enabled) {
2526 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2527 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2528 }
2529 if (lp->depth_stencil->stencil[0].enabled) {
2530 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2531 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2532 }
2533 }
2534
2535 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2536 if (!lp->framebuffer.nr_cbufs ||
2537 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2538 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2539 }
2540 if(key->alpha.enabled)
2541 key->alpha.func = lp->depth_stencil->alpha.func;
2542 /* alpha.ref_value is passed in jit_context */
2543
2544 key->flatshade = lp->rasterizer->flatshade;
2545 if (lp->active_occlusion_query) {
2546 key->occlusion_count = TRUE;
2547 }
2548
2549 if (lp->framebuffer.nr_cbufs) {
2550 memcpy(&key->blend, lp->blend, sizeof key->blend);
2551 }
2552
2553 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2554
2555 if (!key->blend.independent_blend_enable) {
2556 /* we always need independent blend otherwise the fixups below won't work */
2557 for (i = 1; i < key->nr_cbufs; i++) {
2558 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2559 }
2560 key->blend.independent_blend_enable = 1;
2561 }
2562
2563 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2564 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2565 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2566 const struct util_format_description *format_desc;
2567
2568 key->cbuf_format[i] = format;
2569
2570 format_desc = util_format_description(format);
2571 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2572 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2573
2574 /*
2575 * Mask out color channels not present in the color buffer.
2576 */
2577 blend_rt->colormask &= util_format_colormask(format_desc);
2578
2579 /*
2580 * Disable blend for integer formats.
2581 */
2582 if (util_format_is_pure_integer(format)) {
2583 blend_rt->blend_enable = 0;
2584 }
2585
2586 /*
2587 * Our swizzled render tiles always have an alpha channel, but the linear
2588 * render target format often does not, so force here the dst alpha to be
2589 * one.
2590 *
2591 * This is not a mere optimization. Wrong results will be produced if the
2592 * dst alpha is used, the dst format does not have alpha, and the previous
2593 * rendering was not flushed from the swizzled to linear buffer. For
2594 * example, NonPowTwo DCT.
2595 *
2596 * TODO: This should be generalized to all channels for better
2597 * performance, but only alpha causes correctness issues.
2598 *
2599 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2600 */
2601 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2602 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2603 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2604 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2605 blend_rt->alpha_func = blend_rt->rgb_func;
2606 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2607 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2608 }
2609 }
2610
2611 /* This value will be the same for all the variants of a given shader:
2612 */
2613 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2614
2615 for(i = 0; i < key->nr_samplers; ++i) {
2616 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2617 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2618 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2619 }
2620 }
2621
2622 /*
2623 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2624 * are dx10-style? Can't really have mixed opcodes, at least not
2625 * if we want to skip the holes here (without rescanning tgsi).
2626 */
2627 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2628 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2629 for(i = 0; i < key->nr_sampler_views; ++i) {
2630 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2631 lp_sampler_static_texture_state(&key->state[i].texture_state,
2632 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2633 }
2634 }
2635 }
2636 else {
2637 key->nr_sampler_views = key->nr_samplers;
2638 for(i = 0; i < key->nr_sampler_views; ++i) {
2639 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2640 lp_sampler_static_texture_state(&key->state[i].texture_state,
2641 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2642 }
2643 }
2644 }
2645 }
2646
2647
2648
2649 /**
2650 * Update fragment shader state. This is called just prior to drawing
2651 * something when some fragment-related state has changed.
2652 */
2653 void
2654 llvmpipe_update_fs(struct llvmpipe_context *lp)
2655 {
2656 struct lp_fragment_shader *shader = lp->fs;
2657 struct lp_fragment_shader_variant_key key;
2658 struct lp_fragment_shader_variant *variant = NULL;
2659 struct lp_fs_variant_list_item *li;
2660
2661 make_variant_key(lp, shader, &key);
2662
2663 /* Search the variants for one which matches the key */
2664 li = first_elem(&shader->variants);
2665 while(!at_end(&shader->variants, li)) {
2666 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2667 variant = li->base;
2668 break;
2669 }
2670 li = next_elem(li);
2671 }
2672
2673 if (variant) {
2674 /* Move this variant to the head of the list to implement LRU
2675 * deletion of shader's when we have too many.
2676 */
2677 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2678 }
2679 else {
2680 /* variant not found, create it now */
2681 int64_t t0, t1, dt;
2682 unsigned i;
2683 unsigned variants_to_cull;
2684
2685 if (0) {
2686 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2687 lp->nr_fs_variants,
2688 lp->nr_fs_instrs,
2689 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2690 }
2691
2692 /* First, check if we've exceeded the max number of shader variants.
2693 * If so, free 25% of them (the least recently used ones).
2694 */
2695 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2696
2697 if (variants_to_cull ||
2698 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2699 struct pipe_context *pipe = &lp->pipe;
2700
2701 /*
2702 * XXX: we need to flush the context until we have some sort of
2703 * reference counting in fragment shaders as they may still be binned
2704 * Flushing alone might not be sufficient we need to wait on it too.
2705 */
2706 llvmpipe_finish(pipe, __FUNCTION__);
2707
2708 /*
2709 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2710 * number of shader variants (potentially all of them) could be
2711 * pending for destruction on flush.
2712 */
2713
2714 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2715 struct lp_fs_variant_list_item *item;
2716 if (is_empty_list(&lp->fs_variants_list)) {
2717 break;
2718 }
2719 item = last_elem(&lp->fs_variants_list);
2720 assert(item);
2721 assert(item->base);
2722 llvmpipe_remove_shader_variant(lp, item->base);
2723 }
2724 }
2725
2726 /*
2727 * Generate the new variant.
2728 */
2729 t0 = os_time_get();
2730 variant = generate_variant(lp, shader, &key);
2731 t1 = os_time_get();
2732 dt = t1 - t0;
2733 LP_COUNT_ADD(llvm_compile_time, dt);
2734 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2735
2736 llvmpipe_variant_count++;
2737
2738 /* Put the new variant into the list */
2739 if (variant) {
2740 insert_at_head(&shader->variants, &variant->list_item_local);
2741 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2742 lp->nr_fs_variants++;
2743 lp->nr_fs_instrs += variant->nr_instrs;
2744 shader->variants_cached++;
2745 }
2746 }
2747
2748 /* Bind this variant */
2749 lp_setup_set_fs_variant(lp->setup, variant);
2750 }
2751
2752
2753
2754
2755
2756 void
2757 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2758 {
2759 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2760 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2761 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2762
2763 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2764 }