projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallium/llvm: implement geometry shaders in the llvm paths
[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 "util/u_dual_blend.h"
70 #include "os/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_swizzle.h"
84 #include "gallivm/lp_bld_flow.h"
85 #include "gallivm/lp_bld_debug.h"
86 #include "gallivm/lp_bld_arit.h"
87 #include "gallivm/lp_bld_pack.h"
88 #include "gallivm/lp_bld_format.h"
89 #include "gallivm/lp_bld_quad.h"
90
91 #include "lp_bld_alpha.h"
92 #include "lp_bld_blend.h"
93 #include "lp_bld_depth.h"
94 #include "lp_bld_interp.h"
95 #include "lp_context.h"
96 #include "lp_debug.h"
97 #include "lp_perf.h"
98 #include "lp_setup.h"
99 #include "lp_state.h"
100 #include "lp_tex_sample.h"
101 #include "lp_flush.h"
102 #include "lp_state_fs.h"
103
104
105 /** Fragment shader number (for debugging) */
106 static unsigned fs_no = 0;
107
108
109 /**
110 * Expand the relevant bits of mask_input to a n*4-dword mask for the
111 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
112 * quad mask vector to 0 or ~0.
113 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
114 * quad arguments with fs length 8.
115 *
116 * \param first_quad which quad(s) of the quad group to test, in [0,3]
117 * \param mask_input bitwise mask for the whole 4x4 stamp
118 */
119 static LLVMValueRef
120 generate_quad_mask(struct gallivm_state *gallivm,
121 struct lp_type fs_type,
122 unsigned first_quad,
123 LLVMValueRef mask_input) /* int32 */
124 {
125 LLVMBuilderRef builder = gallivm->builder;
126 struct lp_type mask_type;
127 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
128 LLVMValueRef bits[16];
129 LLVMValueRef mask;
130 int shift, i;
131
132 /*
133 * XXX: We'll need a different path for 16 x u8
134 */
135 assert(fs_type.width == 32);
136 assert(fs_type.length <= Elements(bits));
137 mask_type = lp_int_type(fs_type);
138
139 /*
140 * mask_input >>= (quad * 4)
141 */
142 switch (first_quad) {
143 case 0:
144 shift = 0;
145 break;
146 case 1:
147 assert(fs_type.length == 4);
148 shift = 2;
149 break;
150 case 2:
151 shift = 8;
152 break;
153 case 3:
154 assert(fs_type.length == 4);
155 shift = 10;
156 break;
157 default:
158 assert(0);
159 shift = 0;
160 }
161
162 mask_input = LLVMBuildLShr(builder,
163 mask_input,
164 LLVMConstInt(i32t, shift, 0),
165 "");
166
167 /*
168 * mask = { mask_input & (1 << i), for i in [0,3] }
169 */
170 mask = lp_build_broadcast(gallivm,
171 lp_build_vec_type(gallivm, mask_type),
172 mask_input);
173
174 for (i = 0; i < fs_type.length / 4; i++) {
175 unsigned j = 2 * (i % 2) + (i / 2) * 8;
176 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
177 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
178 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
179 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
180 }
181 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
182
183 /*
184 * mask = mask != 0 ? ~0 : 0
185 */
186 mask = lp_build_compare(gallivm,
187 mask_type, PIPE_FUNC_NOTEQUAL,
188 mask,
189 lp_build_const_int_vec(gallivm, mask_type, 0));
190
191 return mask;
192 }
193
194
195 #define EARLY_DEPTH_TEST 0x1
196 #define LATE_DEPTH_TEST 0x2
197 #define EARLY_DEPTH_WRITE 0x4
198 #define LATE_DEPTH_WRITE 0x8
199
200 static int
201 find_output_by_semantic( const struct tgsi_shader_info *info,
202 unsigned semantic,
203 unsigned index )
204 {
205 int i;
206
207 for (i = 0; i < info->num_outputs; i++)
208 if (info->output_semantic_name[i] == semantic &&
209 info->output_semantic_index[i] == index)
210 return i;
211
212 return -1;
213 }
214
215
216 /**
217 * Generate the fragment shader, depth/stencil test, and alpha tests.
218 * \param i which quad in the tile, in range [0,3]
219 * \param partial_mask if 1, do mask_input testing
220 */
221 static void
222 generate_fs(struct gallivm_state *gallivm,
223 struct lp_fragment_shader *shader,
224 const struct lp_fragment_shader_variant_key *key,
225 LLVMBuilderRef builder,
226 struct lp_type type,
227 LLVMValueRef context_ptr,
228 unsigned i,
229 struct lp_build_interp_soa_context *interp,
230 struct lp_build_sampler_soa *sampler,
231 LLVMValueRef *pmask,
232 LLVMValueRef (*color)[4],
233 LLVMValueRef depth_ptr,
234 LLVMValueRef facing,
235 unsigned partial_mask,
236 LLVMValueRef mask_input,
237 LLVMValueRef thread_data_ptr)
238 {
239 const struct util_format_description *zs_format_desc = NULL;
240 const struct tgsi_token *tokens = shader->base.tokens;
241 LLVMTypeRef vec_type;
242 LLVMValueRef consts_ptr;
243 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
244 LLVMValueRef z;
245 LLVMValueRef zs_value = NULL;
246 LLVMValueRef stencil_refs[2];
247 struct lp_build_mask_context mask;
248 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
249 shader->info.base.num_inputs < 3 &&
250 shader->info.base.num_instructions < 8);
251 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
252 util_blend_state_is_dual(&key->blend, 0);
253 unsigned attrib;
254 unsigned chan;
255 unsigned cbuf;
256 unsigned depth_mode;
257 struct lp_bld_tgsi_system_values system_values;
258
259 memset(&system_values, 0, sizeof(system_values));
260
261 if (key->depth.enabled ||
262 key->stencil[0].enabled ||
263 key->stencil[1].enabled) {
264
265 zs_format_desc = util_format_description(key->zsbuf_format);
266 assert(zs_format_desc);
267
268 if (!shader->info.base.writes_z) {
269 if (key->alpha.enabled || shader->info.base.uses_kill)
270 /* With alpha test and kill, can do the depth test early
271 * and hopefully eliminate some quads. But need to do a
272 * special deferred depth write once the final mask value
273 * is known.
274 */
275 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
276 else
277 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
278 }
279 else {
280 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
281 }
282
283 if (!(key->depth.enabled && key->depth.writemask) &&
284 !(key->stencil[0].enabled && key->stencil[0].writemask))
285 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
286 }
287 else {
288 depth_mode = 0;
289 }
290
291 assert(i < 4);
292
293 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
294 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
295
296 vec_type = lp_build_vec_type(gallivm, type);
297
298 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
299
300 memset(outputs, 0, sizeof outputs);
301
302 /* Declare the color and z variables */
303 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
304 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
305 color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
306 }
307 }
308 if (dual_source_blend) {
309 assert(key->nr_cbufs <= 1);
310 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
311 color[1][chan] = lp_build_alloca(gallivm, vec_type, "color1");
312 }
313 }
314
315 /* do triangle edge testing */
316 if (partial_mask) {
317 *pmask = generate_quad_mask(gallivm, type,
318 i*type.length/4, mask_input);
319 }
320 else {
321 *pmask = lp_build_const_int_vec(gallivm, type, ~0);
322 }
323
324 /* 'mask' will control execution based on quad's pixel alive/killed state */
325 lp_build_mask_begin(&mask, gallivm, type, *pmask);
326
327 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
328 lp_build_mask_check(&mask);
329
330 lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
331 z = interp->pos[2];
332
333 if (depth_mode & EARLY_DEPTH_TEST) {
334 lp_build_depth_stencil_test(gallivm,
335 &key->depth,
336 key->stencil,
337 type,
338 zs_format_desc,
339 &mask,
340 stencil_refs,
341 z,
342 depth_ptr, facing,
343 &zs_value,
344 !simple_shader);
345
346 if (depth_mode & EARLY_DEPTH_WRITE) {
347 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
348 }
349 }
350
351 lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
352
353 /* Build the actual shader */
354 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
355 consts_ptr, &system_values,
356 interp->pos, interp->inputs,
357 outputs, sampler, &shader->info.base, NULL);
358
359 /* Alpha test */
360 if (key->alpha.enabled) {
361 int color0 = find_output_by_semantic(&shader->info.base,
362 TGSI_SEMANTIC_COLOR,
363 0);
364
365 if (color0 != -1 && outputs[color0][3]) {
366 const struct util_format_description *cbuf_format_desc;
367 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
368 LLVMValueRef alpha_ref_value;
369
370 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
371 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
372
373 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
374
375 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
376 &mask, alpha, alpha_ref_value,
377 (depth_mode & LATE_DEPTH_TEST) != 0);
378 }
379 }
380
381 /* Late Z test */
382 if (depth_mode & LATE_DEPTH_TEST) {
383 int pos0 = find_output_by_semantic(&shader->info.base,
384 TGSI_SEMANTIC_POSITION,
385 0);
386
387 if (pos0 != -1 && outputs[pos0][2]) {
388 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
389 }
390
391 lp_build_depth_stencil_test(gallivm,
392 &key->depth,
393 key->stencil,
394 type,
395 zs_format_desc,
396 &mask,
397 stencil_refs,
398 z,
399 depth_ptr, facing,
400 &zs_value,
401 !simple_shader);
402 /* Late Z write */
403 if (depth_mode & LATE_DEPTH_WRITE) {
404 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
405 }
406 }
407 else if ((depth_mode & EARLY_DEPTH_TEST) &&
408 (depth_mode & LATE_DEPTH_WRITE))
409 {
410 /* Need to apply a reduced mask to the depth write. Reload the
411 * depth value, update from zs_value with the new mask value and
412 * write that out.
413 */
414 lp_build_deferred_depth_write(gallivm,
415 type,
416 zs_format_desc,
417 &mask,
418 depth_ptr,
419 zs_value);
420 }
421
422
423 /* Color write */
424 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
425 {
426 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
427 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
428 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
429 {
430 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
431 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
432 if(outputs[attrib][chan]) {
433 /* XXX: just initialize outputs to point at colors[] and
434 * skip this.
435 */
436 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
437 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
438 LLVMBuildStore(builder, out, color[cbuf][chan]);
439 }
440 }
441 }
442 }
443
444 if (key->occlusion_count) {
445 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
446 lp_build_name(counter, "counter");
447 lp_build_occlusion_count(gallivm, type,
448 lp_build_mask_value(&mask), counter);
449 }
450
451 *pmask = lp_build_mask_end(&mask);
452 }
453
454
455 /**
456 * Generate the fragment shader, depth/stencil test, and alpha tests.
457 */
458 static void
459 generate_fs_loop(struct gallivm_state *gallivm,
460 struct lp_fragment_shader *shader,
461 const struct lp_fragment_shader_variant_key *key,
462 LLVMBuilderRef builder,
463 struct lp_type type,
464 LLVMValueRef context_ptr,
465 LLVMValueRef num_loop,
466 struct lp_build_interp_soa_context *interp,
467 struct lp_build_sampler_soa *sampler,
468 LLVMValueRef mask_store,
469 LLVMValueRef (*out_color)[4],
470 LLVMValueRef depth_ptr,
471 unsigned depth_bits,
472 LLVMValueRef facing,
473 LLVMValueRef thread_data_ptr)
474 {
475 const struct util_format_description *zs_format_desc = NULL;
476 const struct tgsi_token *tokens = shader->base.tokens;
477 LLVMTypeRef vec_type;
478 LLVMValueRef mask_ptr, mask_val;
479 LLVMValueRef consts_ptr;
480 LLVMValueRef z;
481 LLVMValueRef zs_value = NULL;
482 LLVMValueRef stencil_refs[2];
483 LLVMValueRef depth_ptr_i;
484 LLVMValueRef depth_offset;
485 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
486 struct lp_build_for_loop_state loop_state;
487 struct lp_build_mask_context mask;
488 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
489 shader->info.base.num_inputs < 3 &&
490 shader->info.base.num_instructions < 8);
491 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
492 util_blend_state_is_dual(&key->blend, 0);
493 unsigned attrib;
494 unsigned chan;
495 unsigned cbuf;
496 unsigned depth_mode;
497
498 struct lp_bld_tgsi_system_values system_values;
499
500 memset(&system_values, 0, sizeof(system_values));
501
502 if (key->depth.enabled ||
503 key->stencil[0].enabled ||
504 key->stencil[1].enabled) {
505
506 zs_format_desc = util_format_description(key->zsbuf_format);
507 assert(zs_format_desc);
508
509 if (!shader->info.base.writes_z) {
510 if (key->alpha.enabled || shader->info.base.uses_kill)
511 /* With alpha test and kill, can do the depth test early
512 * and hopefully eliminate some quads. But need to do a
513 * special deferred depth write once the final mask value
514 * is known.
515 */
516 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
517 else
518 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
519 }
520 else {
521 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
522 }
523
524 if (!(key->depth.enabled && key->depth.writemask) &&
525 !(key->stencil[0].enabled && key->stencil[0].writemask))
526 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
527 }
528 else {
529 depth_mode = 0;
530 }
531
532
533 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
534 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
535
536 vec_type = lp_build_vec_type(gallivm, type);
537
538 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
539
540 lp_build_for_loop_begin(&loop_state, gallivm,
541 lp_build_const_int32(gallivm, 0),
542 LLVMIntULT,
543 num_loop,
544 lp_build_const_int32(gallivm, 1));
545
546 mask_ptr = LLVMBuildGEP(builder, mask_store,
547 &loop_state.counter, 1, "mask_ptr");
548 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
549
550 depth_offset = LLVMBuildMul(builder, loop_state.counter,
551 lp_build_const_int32(gallivm, depth_bits * type.length),
552 "");
553
554 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
555
556 memset(outputs, 0, sizeof outputs);
557
558 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
559 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
560 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
561 lp_build_vec_type(gallivm,
562 type),
563 num_loop, "color");
564 }
565 }
566 if (dual_source_blend) {
567 assert(key->nr_cbufs <= 1);
568 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
569 out_color[1][chan] = lp_build_array_alloca(gallivm,
570 lp_build_vec_type(gallivm,
571 type),
572 num_loop, "color1");
573 }
574 }
575
576
577 /* 'mask' will control execution based on quad's pixel alive/killed state */
578 lp_build_mask_begin(&mask, gallivm, type, mask_val);
579
580 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
581 lp_build_mask_check(&mask);
582
583 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
584 z = interp->pos[2];
585
586 if (depth_mode & EARLY_DEPTH_TEST) {
587 lp_build_depth_stencil_test(gallivm,
588 &key->depth,
589 key->stencil,
590 type,
591 zs_format_desc,
592 &mask,
593 stencil_refs,
594 z,
595 depth_ptr_i, facing,
596 &zs_value,
597 !simple_shader);
598
599 if (depth_mode & EARLY_DEPTH_WRITE) {
600 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
601 }
602 }
603
604 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
605
606 /* Build the actual shader */
607 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
608 consts_ptr, &system_values,
609 interp->pos, interp->inputs,
610 outputs, sampler, &shader->info.base, NULL);
611
612 /* Alpha test */
613 if (key->alpha.enabled) {
614 int color0 = find_output_by_semantic(&shader->info.base,
615 TGSI_SEMANTIC_COLOR,
616 0);
617
618 if (color0 != -1 && outputs[color0][3]) {
619 const struct util_format_description *cbuf_format_desc;
620 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
621 LLVMValueRef alpha_ref_value;
622
623 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
624 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
625
626 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
627
628 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
629 &mask, alpha, alpha_ref_value,
630 (depth_mode & LATE_DEPTH_TEST) != 0);
631 }
632 }
633
634 /* Late Z test */
635 if (depth_mode & LATE_DEPTH_TEST) {
636 int pos0 = find_output_by_semantic(&shader->info.base,
637 TGSI_SEMANTIC_POSITION,
638 0);
639
640 if (pos0 != -1 && outputs[pos0][2]) {
641 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
642 }
643
644 lp_build_depth_stencil_test(gallivm,
645 &key->depth,
646 key->stencil,
647 type,
648 zs_format_desc,
649 &mask,
650 stencil_refs,
651 z,
652 depth_ptr_i, facing,
653 &zs_value,
654 !simple_shader);
655 /* Late Z write */
656 if (depth_mode & LATE_DEPTH_WRITE) {
657 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
658 }
659 }
660 else if ((depth_mode & EARLY_DEPTH_TEST) &&
661 (depth_mode & LATE_DEPTH_WRITE))
662 {
663 /* Need to apply a reduced mask to the depth write. Reload the
664 * depth value, update from zs_value with the new mask value and
665 * write that out.
666 */
667 lp_build_deferred_depth_write(gallivm,
668 type,
669 zs_format_desc,
670 &mask,
671 depth_ptr_i,
672 zs_value);
673 }
674
675
676 /* Color write */
677 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
678 {
679 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
680 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
681 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
682 {
683 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
684 if(outputs[attrib][chan]) {
685 /* XXX: just initialize outputs to point at colors[] and
686 * skip this.
687 */
688 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
689 LLVMValueRef color_ptr;
690 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
691 &loop_state.counter, 1, "");
692 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
693 LLVMBuildStore(builder, out, color_ptr);
694 }
695 }
696 }
697 }
698
699 if (key->occlusion_count) {
700 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
701 lp_build_name(counter, "counter");
702 lp_build_occlusion_count(gallivm, type,
703 lp_build_mask_value(&mask), counter);
704 }
705
706 mask_val = lp_build_mask_end(&mask);
707 LLVMBuildStore(builder, mask_val, mask_ptr);
708 lp_build_for_loop_end(&loop_state);
709 }
710
711
712 /**
713 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
714 *
715 * Fragment Shader outputs pixels in small 2x2 blocks
716 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
717 *
718 * However in memory pixels are stored in rows
719 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
720 *
721 * @param type fragment shader type (4x or 8x float)
722 * @param num_fs number of fs_src
723 * @param dst_channels number of output channels
724 * @param fs_src output from fragment shader
725 * @param dst pointer to store result
726 * @param pad_inline is channel padding inline or at end of row
727 * @return the number of dsts
728 */
729 static int
730 generate_fs_twiddle(struct gallivm_state *gallivm,
731 struct lp_type type,
732 unsigned num_fs,
733 unsigned dst_channels,
734 LLVMValueRef fs_src[][4],
735 LLVMValueRef* dst,
736 bool pad_inline)
737 {
738 LLVMValueRef src[16];
739
740 bool swizzle_pad;
741 bool twiddle;
742 bool split;
743
744 unsigned pixels = num_fs == 4 ? 1 : 2;
745 unsigned reorder_group;
746 unsigned src_channels;
747 unsigned src_count;
748 unsigned i;
749
750 src_channels = dst_channels < 3 ? dst_channels : 4;
751 src_count = num_fs * src_channels;
752
753 assert(pixels == 2 || num_fs == 4);
754 assert(num_fs * src_channels <= Elements(src));
755
756 /*
757 * Transpose from SoA -> AoS
758 */
759 for (i = 0; i < num_fs; ++i) {
760 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
761 }
762
763 /*
764 * Pick transformation options
765 */
766 swizzle_pad = false;
767 twiddle = false;
768 split = false;
769 reorder_group = 0;
770
771 if (dst_channels == 1) {
772 twiddle = true;
773
774 if (pixels == 2) {
775 split = true;
776 }
777 } else if (dst_channels == 2) {
778 if (pixels == 1) {
779 reorder_group = 1;
780 }
781 } else if (dst_channels > 2) {
782 if (pixels == 1) {
783 reorder_group = 2;
784 } else {
785 twiddle = true;
786 }
787
788 if (!pad_inline && dst_channels == 3 && pixels > 1) {
789 swizzle_pad = true;
790 }
791 }
792
793 /*
794 * Split the src in half
795 */
796 if (split) {
797 for (i = num_fs; i > 0; --i) {
798 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
799 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
800 }
801
802 src_count *= 2;
803 type.length = 4;
804 }
805
806 /*
807 * Ensure pixels are in memory order
808 */
809 if (reorder_group) {
810 /* Twiddle pixels by reordering the array, e.g.:
811 *
812 * src_count = 8 -> 0 2 1 3 4 6 5 7
813 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
814 */
815 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
816
817 for (i = 0; i < src_count; ++i) {
818 unsigned group = i / reorder_group;
819 unsigned block = (group / 4) * 4 * reorder_group;
820 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
821 dst[i] = src[j];
822 }
823 } else if (twiddle) {
824 /* Twiddle pixels across elements of array */
825 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
826 } else {
827 /* Do nothing */
828 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
829 }
830
831 /*
832 * Moves any padding between pixels to the end
833 * e.g. RGBXRGBX -> RGBRGBXX
834 */
835 if (swizzle_pad) {
836 unsigned char swizzles[16];
837 unsigned elems = pixels * dst_channels;
838
839 for (i = 0; i < type.length; ++i) {
840 if (i < elems)
841 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
842 else
843 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
844 }
845
846 for (i = 0; i < src_count; ++i) {
847 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
848 }
849 }
850
851 return src_count;
852 }
853
854
855 /**
856 * Load an unswizzled block of pixels from memory
857 */
858 static void
859 load_unswizzled_block(struct gallivm_state *gallivm,
860 LLVMValueRef base_ptr,
861 LLVMValueRef stride,
862 unsigned block_width,
863 unsigned block_height,
864 LLVMValueRef* dst,
865 struct lp_type dst_type,
866 unsigned dst_count,
867 unsigned dst_alignment)
868 {
869 LLVMBuilderRef builder = gallivm->builder;
870 unsigned row_size = dst_count / block_height;
871 unsigned i;
872
873 /* Ensure block exactly fits into dst */
874 assert((block_width * block_height) % dst_count == 0);
875
876 for (i = 0; i < dst_count; ++i) {
877 unsigned x = i % row_size;
878 unsigned y = i / row_size;
879
880 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
881 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
882
883 LLVMValueRef gep[2];
884 LLVMValueRef dst_ptr;
885
886 gep[0] = lp_build_const_int32(gallivm, 0);
887 gep[1] = LLVMBuildAdd(builder, bx, by, "");
888
889 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
890 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
891
892 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
893
894 lp_set_load_alignment(dst[i], dst_alignment);
895 }
896 }
897
898
899 /**
900 * Store an unswizzled block of pixels to memory
901 */
902 static void
903 store_unswizzled_block(struct gallivm_state *gallivm,
904 LLVMValueRef base_ptr,
905 LLVMValueRef stride,
906 unsigned block_width,
907 unsigned block_height,
908 LLVMValueRef* src,
909 struct lp_type src_type,
910 unsigned src_count,
911 unsigned src_alignment)
912 {
913 LLVMBuilderRef builder = gallivm->builder;
914 unsigned row_size = src_count / block_height;
915 unsigned i;
916
917 /* Ensure src exactly fits into block */
918 assert((block_width * block_height) % src_count == 0);
919
920 for (i = 0; i < src_count; ++i) {
921 unsigned x = i % row_size;
922 unsigned y = i / row_size;
923
924 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
925 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
926
927 LLVMValueRef gep[2];
928 LLVMValueRef src_ptr;
929
930 gep[0] = lp_build_const_int32(gallivm, 0);
931 gep[1] = LLVMBuildAdd(builder, bx, by, "");
932
933 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
934 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
935
936 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
937
938 lp_set_store_alignment(src_ptr, src_alignment);
939 }
940 }
941
942
943 /**
944 * Checks if a format description is an arithmetic format
945 *
946 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
947 */
948 static INLINE boolean
949 is_arithmetic_format(const struct util_format_description *format_desc)
950 {
951 boolean arith = false;
952 unsigned i;
953
954 for (i = 0; i < format_desc->nr_channels; ++i) {
955 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
956 arith |= (format_desc->channel[i].size % 8) != 0;
957 }
958
959 return arith;
960 }
961
962
963 /**
964 * Retrieves the type representing the memory layout for a format
965 *
966 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
967 */
968 static INLINE void
969 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
970 struct lp_type* type)
971 {
972 unsigned i;
973 unsigned chan;
974
975 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
976 /* just make this a 32bit uint */
977 type->floating = false;
978 type->fixed = false;
979 type->sign = false;
980 type->norm = false;
981 type->width = 32;
982 type->length = 1;
983 return;
984 }
985
986 for (i = 0; i < 4; i++)
987 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
988 break;
989 chan = i;
990
991 memset(type, 0, sizeof(struct lp_type));
992 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
993 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
994 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
995 type->norm = format_desc->channel[chan].normalized;
996
997 if (is_arithmetic_format(format_desc)) {
998 type->width = 0;
999 type->length = 1;
1000
1001 for (i = 0; i < format_desc->nr_channels; ++i) {
1002 type->width += format_desc->channel[i].size;
1003 }
1004 } else {
1005 type->width = format_desc->channel[chan].size;
1006 type->length = format_desc->nr_channels;
1007 }
1008 }
1009
1010
1011 /**
1012 * Retrieves the type for a format which is usable in the blending code.
1013 *
1014 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1015 */
1016 static INLINE void
1017 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1018 struct lp_type* type)
1019 {
1020 unsigned i;
1021 unsigned chan;
1022
1023 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1024 /* always use ordinary floats for blending */
1025 type->floating = true;
1026 type->fixed = false;
1027 type->sign = true;
1028 type->norm = false;
1029 type->width = 32;
1030 type->length = 4;
1031 return;
1032 }
1033
1034 for (i = 0; i < 4; i++)
1035 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1036 break;
1037 chan = i;
1038
1039 memset(type, 0, sizeof(struct lp_type));
1040 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1041 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1042 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1043 type->norm = format_desc->channel[chan].normalized;
1044 type->width = format_desc->channel[chan].size;
1045 type->length = format_desc->nr_channels;
1046
1047 for (i = 1; i < format_desc->nr_channels; ++i) {
1048 if (format_desc->channel[i].size > type->width)
1049 type->width = format_desc->channel[i].size;
1050 }
1051
1052 if (type->floating) {
1053 type->width = 32;
1054 } else {
1055 if (type->width <= 8) {
1056 type->width = 8;
1057 } else if (type->width <= 16) {
1058 type->width = 16;
1059 } else {
1060 type->width = 32;
1061 }
1062 }
1063
1064 if (is_arithmetic_format(format_desc) && type->length == 3) {
1065 type->length = 4;
1066 }
1067 }
1068
1069
1070 /**
1071 * Scale a normalized value from src_bits to dst_bits
1072 */
1073 static INLINE LLVMValueRef
1074 scale_bits(struct gallivm_state *gallivm,
1075 int src_bits,
1076 int dst_bits,
1077 LLVMValueRef src,
1078 struct lp_type src_type)
1079 {
1080 LLVMBuilderRef builder = gallivm->builder;
1081 LLVMValueRef result = src;
1082
1083 if (dst_bits < src_bits) {
1084 /* Scale down by LShr */
1085 result = LLVMBuildLShr(builder,
1086 src,
1087 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
1088 "");
1089 } else if (dst_bits > src_bits) {
1090 /* Scale up bits */
1091 int db = dst_bits - src_bits;
1092
1093 /* Shift left by difference in bits */
1094 result = LLVMBuildShl(builder,
1095 src,
1096 lp_build_const_int_vec(gallivm, src_type, db),
1097 "");
1098
1099 if (db < src_bits) {
1100 /* Enough bits in src to fill the remainder */
1101 LLVMValueRef lower = LLVMBuildLShr(builder,
1102 src,
1103 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1104 "");
1105
1106 result = LLVMBuildOr(builder, result, lower, "");
1107 } else if (db > src_bits) {
1108 /* Need to repeatedly copy src bits to fill remainder in dst */
1109 unsigned n;
1110
1111 for (n = src_bits; n < dst_bits; n *= 2) {
1112 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1113
1114 result = LLVMBuildOr(builder,
1115 result,
1116 LLVMBuildLShr(builder, result, shuv, ""),
1117 "");
1118 }
1119 }
1120 }
1121
1122 return result;
1123 }
1124
1125
1126 /**
1127 * Convert from memory format to blending format
1128 *
1129 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1130 */
1131 static void
1132 convert_to_blend_type(struct gallivm_state *gallivm,
1133 const struct util_format_description *src_fmt,
1134 struct lp_type src_type,
1135 struct lp_type dst_type,
1136 LLVMValueRef* src, // and dst
1137 unsigned num_srcs)
1138 {
1139 LLVMValueRef *dst = src;
1140 LLVMBuilderRef builder = gallivm->builder;
1141 struct lp_type blend_type;
1142 struct lp_type mem_type;
1143 unsigned i, j, k;
1144 unsigned pixels = 16 / num_srcs;
1145 bool is_arith;
1146
1147 /*
1148 * full custom path for packed floats - none of the later functions would do
1149 * anything useful, and given the lp_type representation they can't be fixed.
1150 */
1151 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1152 LLVMValueRef tmpsrc[4];
1153 /*
1154 * This is pretty suboptimal for this case blending in SoA would be much
1155 * better, since conversion gets us SoA values so need to convert back.
1156 */
1157 assert(src_type.width == 32);
1158 assert(dst_type.floating);
1159 assert(dst_type.width == 32);
1160 assert(dst_type.length % 4 == 0);
1161 for (i = 0; i < 4; i++) {
1162 tmpsrc[i] = src[i];
1163 }
1164 for (i = 0; i < num_srcs / 4; i++) {
1165 LLVMValueRef tmpsoa[4];
1166 LLVMValueRef tmps = tmpsrc[i];
1167 if (num_srcs == 8) {
1168 LLVMValueRef shuffles[8];
1169 unsigned j;
1170 /* fetch was 4 values but need 8-wide output values */
1171 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1172 /*
1173 * for 8-wide aos transpose would give us wrong order not matching
1174 * incoming converted fs values and mask. ARGH.
1175 */
1176 for (j = 0; j < 4; j++) {
1177 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1178 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1179 }
1180 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1181 LLVMConstVector(shuffles, 8), "");
1182 }
1183 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1184 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1185 }
1186 return;
1187 }
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 the format arithmetic */
1193 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1194 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1195
1196 /* Pad if necessary */
1197 if (!is_arith && src_type.length < dst_type.length) {
1198 for (i = 0; i < num_srcs; ++i) {
1199 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1200 }
1201
1202 src_type.length = dst_type.length;
1203 }
1204
1205 /* Special case for half-floats */
1206 if (mem_type.width == 16 && mem_type.floating) {
1207 assert(blend_type.width == 32 && blend_type.floating);
1208 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1209 is_arith = false;
1210 }
1211
1212 if (!is_arith) {
1213 return;
1214 }
1215
1216 src_type.width = blend_type.width * blend_type.length;
1217 blend_type.length *= pixels;
1218 src_type.length *= pixels / (src_type.length / mem_type.length);
1219
1220 for (i = 0; i < num_srcs; ++i) {
1221 LLVMValueRef chans[4];
1222 LLVMValueRef res = NULL;
1223 unsigned sa = 0;
1224
1225 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1226
1227 for (j = 0; j < src_fmt->nr_channels; ++j) {
1228 unsigned mask = 0;
1229
1230 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1231 mask |= 1 << k;
1232 }
1233
1234 /* Extract bits from source */
1235 chans[j] = LLVMBuildLShr(builder,
1236 dst[i],
1237 lp_build_const_int_vec(gallivm, src_type, sa),
1238 "");
1239
1240 chans[j] = LLVMBuildAnd(builder,
1241 chans[j],
1242 lp_build_const_int_vec(gallivm, src_type, mask),
1243 "");
1244
1245 /* Scale bits */
1246 if (src_type.norm) {
1247 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1248 blend_type.width, chans[j], src_type);
1249 }
1250
1251 /* Insert bits into correct position */
1252 chans[j] = LLVMBuildShl(builder,
1253 chans[j],
1254 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1255 "");
1256
1257 sa += src_fmt->channel[j].size;
1258
1259 if (j == 0) {
1260 res = chans[j];
1261 } else {
1262 res = LLVMBuildOr(builder, res, chans[j], "");
1263 }
1264 }
1265
1266 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1267 }
1268 }
1269
1270
1271 /**
1272 * Convert from blending format to memory format
1273 *
1274 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1275 */
1276 static void
1277 convert_from_blend_type(struct gallivm_state *gallivm,
1278 const struct util_format_description *src_fmt,
1279 struct lp_type src_type,
1280 struct lp_type dst_type,
1281 LLVMValueRef* src, // and dst
1282 unsigned num_srcs)
1283 {
1284 LLVMValueRef* dst = src;
1285 unsigned i, j, k;
1286 struct lp_type mem_type;
1287 struct lp_type blend_type;
1288 LLVMBuilderRef builder = gallivm->builder;
1289 unsigned pixels = 16 / num_srcs;
1290 bool is_arith;
1291
1292 /*
1293 * full custom path for packed floats - none of the later functions would do
1294 * anything useful, and given the lp_type representation they can't be fixed.
1295 */
1296 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1297 /*
1298 * This is pretty suboptimal for this case blending in SoA would be much
1299 * better - we need to transpose the AoS values back to SoA values for
1300 * conversion/packing.
1301 */
1302 assert(src_type.floating);
1303 assert(src_type.width == 32);
1304 assert(src_type.length % 4 == 0);
1305 assert(dst_type.width == 32);
1306 for (i = 0; i < num_srcs / 4; i++) {
1307 LLVMValueRef tmpsoa[4], tmpdst;
1308 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1309 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1310 if (num_srcs == 8) {
1311 LLVMValueRef tmpaos, shuffles[8];
1312 unsigned j;
1313 /*
1314 * for 8-wide aos transpose has given us wrong order not matching
1315 * output order. HMPF. Also need to split the output values manually.
1316 */
1317 for (j = 0; j < 4; j++) {
1318 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1319 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1320 }
1321 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1322 LLVMConstVector(shuffles, 8), "");
1323 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1324 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1325 }
1326 else {
1327 src[i] = tmpdst;
1328 }
1329 }
1330 return;
1331 }
1332
1333 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1334 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1335
1336 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1337
1338 /* Special case for half-floats */
1339 if (mem_type.width == 16 && mem_type.floating) {
1340 int length = dst_type.length;
1341 assert(blend_type.width == 32 && blend_type.floating);
1342
1343 dst_type.length = src_type.length;
1344
1345 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1346
1347 dst_type.length = length;
1348 is_arith = false;
1349 }
1350
1351 /* Remove any padding */
1352 if (!is_arith && (src_type.length % mem_type.length)) {
1353 src_type.length -= (src_type.length % mem_type.length);
1354
1355 for (i = 0; i < num_srcs; ++i) {
1356 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1357 }
1358 }
1359
1360 /* No bit arithmetic to do */
1361 if (!is_arith) {
1362 return;
1363 }
1364
1365 src_type.length = pixels;
1366 src_type.width = blend_type.length * blend_type.width;
1367 dst_type.length = pixels;
1368
1369 for (i = 0; i < num_srcs; ++i) {
1370 LLVMValueRef chans[4];
1371 LLVMValueRef res = NULL;
1372 unsigned sa = 0;
1373
1374 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1375
1376 for (j = 0; j < src_fmt->nr_channels; ++j) {
1377 unsigned mask = 0;
1378
1379 assert(blend_type.width > src_fmt->channel[j].size);
1380
1381 for (k = 0; k < blend_type.width; ++k) {
1382 mask |= 1 << k;
1383 }
1384
1385 /* Extract bits */
1386 chans[j] = LLVMBuildLShr(builder,
1387 dst[i],
1388 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1389 "");
1390
1391 chans[j] = LLVMBuildAnd(builder,
1392 chans[j],
1393 lp_build_const_int_vec(gallivm, src_type, mask),
1394 "");
1395
1396 /* Scale down bits */
1397 if (src_type.norm) {
1398 chans[j] = scale_bits(gallivm, blend_type.width,
1399 src_fmt->channel[j].size, chans[j], src_type);
1400 }
1401
1402 /* Insert bits */
1403 chans[j] = LLVMBuildShl(builder,
1404 chans[j],
1405 lp_build_const_int_vec(gallivm, src_type, sa),
1406 "");
1407
1408 sa += src_fmt->channel[j].size;
1409
1410 if (j == 0) {
1411 res = chans[j];
1412 } else {
1413 res = LLVMBuildOr(builder, res, chans[j], "");
1414 }
1415 }
1416
1417 assert (dst_type.width != 24);
1418
1419 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1420 }
1421 }
1422
1423
1424 /**
1425 * Convert alpha to same blend type as src
1426 */
1427 static void
1428 convert_alpha(struct gallivm_state *gallivm,
1429 struct lp_type row_type,
1430 struct lp_type alpha_type,
1431 const unsigned block_size,
1432 const unsigned block_height,
1433 const unsigned src_count,
1434 const unsigned dst_channels,
1435 const bool pad_inline,
1436 LLVMValueRef* src_alpha)
1437 {
1438 LLVMBuilderRef builder = gallivm->builder;
1439 unsigned i, j;
1440 unsigned length = row_type.length;
1441 row_type.length = alpha_type.length;
1442
1443 /* Twiddle the alpha to match pixels */
1444 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1445
1446 for (i = 0; i < 4; ++i) {
1447 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1448 }
1449
1450 alpha_type = row_type;
1451 row_type.length = length;
1452
1453 /* If only one channel we can only need the single alpha value per pixel */
1454 if (src_count == 1) {
1455 assert(dst_channels == 1);
1456
1457 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1458 } else {
1459 /* If there are more srcs than rows then we need to split alpha up */
1460 if (src_count > block_height) {
1461 for (i = src_count; i > 0; --i) {
1462 unsigned pixels = block_size / src_count;
1463 unsigned idx = i - 1;
1464
1465 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1466 }
1467 }
1468
1469 /* If there is a src for each pixel broadcast the alpha across whole row */
1470 if (src_count == block_size) {
1471 for (i = 0; i < src_count; ++i) {
1472 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1473 }
1474 } else {
1475 unsigned pixels = block_size / src_count;
1476 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1477 unsigned alpha_span = 1;
1478 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1479
1480 /* Check if we need 2 src_alphas for our shuffles */
1481 if (pixels > alpha_type.length) {
1482 alpha_span = 2;
1483 }
1484
1485 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1486 for (j = 0; j < row_type.length; ++j) {
1487 if (j < pixels * channels) {
1488 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1489 } else {
1490 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1491 }
1492 }
1493
1494 for (i = 0; i < src_count; ++i) {
1495 unsigned idx1 = i, idx2 = i;
1496
1497 if (alpha_span > 1){
1498 idx1 *= alpha_span;
1499 idx2 = idx1 + 1;
1500 }
1501
1502 src_alpha[i] = LLVMBuildShuffleVector(builder,
1503 src_alpha[idx1],
1504 src_alpha[idx2],
1505 LLVMConstVector(shuffles, row_type.length),
1506 "");
1507 }
1508 }
1509 }
1510 }
1511
1512
1513 /**
1514 * Generates the blend function for unswizzled colour buffers
1515 * Also generates the read & write from colour buffer
1516 */
1517 static void
1518 generate_unswizzled_blend(struct gallivm_state *gallivm,
1519 unsigned rt,
1520 struct lp_fragment_shader_variant *variant,
1521 enum pipe_format out_format,
1522 unsigned int num_fs,
1523 struct lp_type fs_type,
1524 LLVMValueRef* fs_mask,
1525 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1526 LLVMValueRef context_ptr,
1527 LLVMValueRef color_ptr,
1528 LLVMValueRef stride,
1529 unsigned partial_mask,
1530 boolean do_branch)
1531 {
1532 const unsigned alpha_channel = 3;
1533 const unsigned block_width = 4;
1534 const unsigned block_height = 4;
1535 const unsigned block_size = block_width * block_height;
1536 const unsigned lp_integer_vector_width = 128;
1537
1538 LLVMBuilderRef builder = gallivm->builder;
1539 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1540 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1541 LLVMValueRef src_alpha[4 * 4];
1542 LLVMValueRef src1_alpha[4 * 4];
1543 LLVMValueRef src_mask[4 * 4];
1544 LLVMValueRef src[4 * 4];
1545 LLVMValueRef src1[4 * 4];
1546 LLVMValueRef dst[4 * 4];
1547 LLVMValueRef blend_color;
1548 LLVMValueRef blend_alpha;
1549 LLVMValueRef i32_zero;
1550 LLVMValueRef check_mask;
1551
1552 struct lp_build_mask_context mask_ctx;
1553 struct lp_type mask_type;
1554 struct lp_type blend_type;
1555 struct lp_type row_type;
1556 struct lp_type dst_type;
1557
1558 unsigned char swizzle[TGSI_NUM_CHANNELS];
1559 unsigned vector_width;
1560 unsigned src_channels = TGSI_NUM_CHANNELS;
1561 unsigned dst_channels;
1562 unsigned dst_count;
1563 unsigned src_count;
1564 unsigned i, j;
1565
1566 const struct util_format_description* out_format_desc = util_format_description(out_format);
1567
1568 unsigned dst_alignment;
1569
1570 bool pad_inline = is_arithmetic_format(out_format_desc);
1571 bool has_alpha = false;
1572 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1573 util_blend_state_is_dual(&variant->key.blend, 0);
1574
1575 mask_type = lp_int32_vec4_type();
1576 mask_type.length = fs_type.length;
1577
1578 /* Compute the alignment of the destination pointer in bytes */
1579 #if 0
1580 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1581 #else
1582 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1583 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1584 #endif
1585 /* Force power-of-two alignment by extracting only the least-significant-bit */
1586 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1587 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1588 dst_alignment = MIN2(dst_alignment, 16);
1589
1590 /* Do not bother executing code when mask is empty.. */
1591 if (do_branch) {
1592 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1593
1594 for (i = 0; i < num_fs; ++i) {
1595 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1596 }
1597
1598 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1599 lp_build_mask_check(&mask_ctx);
1600 }
1601
1602 partial_mask |= !variant->opaque;
1603 i32_zero = lp_build_const_int32(gallivm, 0);
1604
1605 /* Get type from output format */
1606 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1607 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1608
1609 row_type.length = fs_type.length;
1610 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1611
1612 /* Compute correct swizzle and count channels */
1613 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1614 dst_channels = 0;
1615
1616 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1617 /* Ensure channel is used */
1618 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1619 continue;
1620 }
1621
1622 /* Ensure not already written to (happens in case with GL_ALPHA) */
1623 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1624 continue;
1625 }
1626
1627 /* Ensure we havn't already found all channels */
1628 if (dst_channels >= out_format_desc->nr_channels) {
1629 continue;
1630 }
1631
1632 swizzle[out_format_desc->swizzle[i]] = i;
1633 ++dst_channels;
1634
1635 if (i == alpha_channel) {
1636 has_alpha = true;
1637 }
1638 }
1639
1640 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1641 /* the code above can't work for layout_other */
1642 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1643 has_alpha = true;
1644 swizzle[0] = 0;
1645 swizzle[1] = 1;
1646 swizzle[2] = 2;
1647 swizzle[3] = 3;
1648 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1649 }
1650
1651 /* If 3 channels then pad to include alpha for 4 element transpose */
1652 if (dst_channels == 3 && !has_alpha) {
1653 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1654 if (swizzle[i] > TGSI_NUM_CHANNELS)
1655 swizzle[i] = 3;
1656 }
1657 if (out_format_desc->nr_channels == 4) {
1658 dst_channels = 4;
1659 }
1660 }
1661
1662 /*
1663 * Load shader output
1664 */
1665 for (i = 0; i < num_fs; ++i) {
1666 /* Always load alpha for use in blending */
1667 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1668
1669 /* Load each channel */
1670 for (j = 0; j < dst_channels; ++j) {
1671 assert(swizzle[j] < 4);
1672 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1673 }
1674
1675 /* If 3 channels then pad to include alpha for 4 element transpose */
1676 /*
1677 * XXX If we include that here maybe could actually use it instead of
1678 * separate alpha for blending?
1679 */
1680 if (dst_channels == 3 && !has_alpha) {
1681 fs_src[i][3] = alpha;
1682 }
1683
1684 /* We split the row_mask and row_alpha as we want 128bit interleave */
1685 if (fs_type.length == 8) {
1686 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1687 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1688
1689 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1690 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1691 } else {
1692 src_mask[i] = fs_mask[i];
1693 src_alpha[i] = alpha;
1694 }
1695 }
1696 if (dual_source_blend) {
1697 /* same as above except different src/dst, skip masks and comments... */
1698 for (i = 0; i < num_fs; ++i) {
1699 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1700
1701 for (j = 0; j < dst_channels; ++j) {
1702 assert(swizzle[j] < 4);
1703 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1704 }
1705 if (dst_channels == 3 && !has_alpha) {
1706 fs_src1[i][3] = alpha;
1707 }
1708 if (fs_type.length == 8) {
1709 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1710 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1711 } else {
1712 src1_alpha[i] = alpha;
1713 }
1714 }
1715 }
1716
1717 if (util_format_is_pure_integer(out_format)) {
1718 /*
1719 * In this case fs_type was really ints or uints disguised as floats,
1720 * fix that up now.
1721 */
1722 fs_type.floating = 0;
1723 fs_type.sign = dst_type.sign;
1724 for (i = 0; i < num_fs; ++i) {
1725 for (j = 0; j < dst_channels; ++j) {
1726 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1727 lp_build_vec_type(gallivm, fs_type), "");
1728 }
1729 if (dst_channels == 3 && !has_alpha) {
1730 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1731 lp_build_vec_type(gallivm, fs_type), "");
1732 }
1733 }
1734 }
1735
1736 /*
1737 * Pixel twiddle from fragment shader order to memory order
1738 */
1739 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1740 if (dual_source_blend) {
1741 generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
1742 }
1743
1744 src_channels = dst_channels < 3 ? dst_channels : 4;
1745 if (src_count != num_fs * src_channels) {
1746 unsigned ds = src_count / (num_fs * src_channels);
1747 row_type.length /= ds;
1748 fs_type.length = row_type.length;
1749 }
1750
1751 blend_type = row_type;
1752 mask_type.length = 4;
1753
1754 /* Convert src to row_type */
1755 if (dual_source_blend) {
1756 struct lp_type old_row_type = row_type;
1757 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1758 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1759 }
1760 else {
1761 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1762 }
1763
1764 /* If the rows are not an SSE vector, combine them to become SSE size! */
1765 if ((row_type.width * row_type.length) % 128) {
1766 unsigned bits = row_type.width * row_type.length;
1767 unsigned combined;
1768
1769 dst_count = src_count / (vector_width / bits);
1770 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1771 if (dual_source_blend) {
1772 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1773 }
1774
1775 row_type.length *= combined;
1776 src_count /= combined;
1777
1778 bits = row_type.width * row_type.length;
1779 assert(bits == 128 || bits == 256);
1780 }
1781
1782
1783 /*
1784 * Blend Colour conversion
1785 */
1786 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1787 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1788 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1789
1790 /* Convert */
1791 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1792
1793 /* Extract alpha */
1794 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1795
1796 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1797 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1798 if (pad_inline) {
1799 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1800 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1801 } else {
1802 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1803 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1804 }
1805
1806 /*
1807 * Mask conversion
1808 */
1809 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1810
1811 if (src_count < block_height) {
1812 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1813 } else if (src_count > block_height) {
1814 for (i = src_count; i > 0; --i) {
1815 unsigned pixels = block_size / src_count;
1816 unsigned idx = i - 1;
1817
1818 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1819 }
1820 }
1821
1822 assert(mask_type.width == 32);
1823
1824 for (i = 0; i < src_count; ++i) {
1825 unsigned pixels = block_size / src_count;
1826 unsigned pixel_width = row_type.width * dst_channels;
1827
1828 if (pixel_width == 24) {
1829 mask_type.width = 8;
1830 mask_type.length = vector_width / mask_type.width;
1831 } else {
1832 mask_type.length = pixels;
1833 mask_type.width = row_type.width * dst_channels;
1834
1835 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1836
1837 mask_type.length *= dst_channels;
1838 mask_type.width /= dst_channels;
1839 }
1840
1841 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1842 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1843 }
1844
1845 /*
1846 * Alpha conversion
1847 */
1848 if (!has_alpha) {
1849 struct lp_type alpha_type = fs_type;
1850 alpha_type.length = 4;
1851 convert_alpha(gallivm, row_type, alpha_type,
1852 block_size, block_height,
1853 src_count, dst_channels,
1854 pad_inline, src_alpha);
1855 if (dual_source_blend) {
1856 convert_alpha(gallivm, row_type, alpha_type,
1857 block_size, block_height,
1858 src_count, dst_channels,
1859 pad_inline, src1_alpha);
1860 }
1861 }
1862
1863
1864 /*
1865 * Load dst from memory
1866 */
1867 if (src_count < block_height) {
1868 dst_count = block_height;
1869 } else {
1870 dst_count = src_count;
1871 }
1872
1873 dst_type.length *= 16 / dst_count;
1874
1875 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1876 /*
1877 * we need multiple values at once for the conversion, so can as well
1878 * load them vectorized here too instead of concatenating later.
1879 * (Still need concatenation later for 8-wide vectors).
1880 */
1881 dst_count = block_height;
1882 dst_type.length = block_width;
1883 }
1884
1885 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1886 dst, dst_type, dst_count, dst_alignment);
1887
1888
1889 /*
1890 * Convert from dst/output format to src/blending format.
1891 *
1892 * This is necessary as we can only read 1 row from memory at a time,
1893 * so the minimum dst_count will ever be at this point is 4.
1894 *
1895 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1896 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1897 * on all 16 pixels in that single vector at once.
1898 */
1899 if (dst_count > src_count) {
1900 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1901 }
1902
1903 /*
1904 * Blending
1905 */
1906 /* XXX this is broken for RGB8 formats -
1907 * they get expanded from 12 to 16 elements (to include alpha)
1908 * by convert_to_blend_type then reduced to 15 instead of 12
1909 * by convert_from_blend_type (a simple fix though breaks A8...).
1910 * R16G16B16 also crashes differently however something going wrong
1911 * inside llvm handling npot vector sizes seemingly.
1912 * It seems some cleanup could be done here (like skipping conversion/blend
1913 * when not needed).
1914 */
1915 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1916
1917 for (i = 0; i < src_count; ++i) {
1918 dst[i] = lp_build_blend_aos(gallivm,
1919 &variant->key.blend,
1920 out_format,
1921 row_type,
1922 rt,
1923 src[i],
1924 has_alpha ? NULL : src_alpha[i],
1925 src1[i],
1926 has_alpha ? NULL : src1_alpha[i],
1927 dst[i],
1928 partial_mask ? src_mask[i] : NULL,
1929 blend_color,
1930 has_alpha ? NULL : blend_alpha,
1931 swizzle,
1932 pad_inline ? 4 : dst_channels);
1933 }
1934
1935 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1936
1937 /* Split the blend rows back to memory rows */
1938 if (dst_count > src_count) {
1939 row_type.length = dst_type.length * (dst_count / src_count);
1940
1941 if (src_count == 1) {
1942 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1943 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1944
1945 row_type.length /= 2;
1946 src_count *= 2;
1947 }
1948
1949 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1950 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1951 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1952 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1953
1954 row_type.length /= 2;
1955 src_count *= 2;
1956 }
1957
1958
1959 /*
1960 * Store blend result to memory
1961 */
1962 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1963 dst, dst_type, dst_count, dst_alignment);
1964
1965 if (do_branch) {
1966 lp_build_mask_end(&mask_ctx);
1967 }
1968 }
1969
1970
1971 /**
1972 * Generate the runtime callable function for the whole fragment pipeline.
1973 * Note that the function which we generate operates on a block of 16
1974 * pixels at at time. The block contains 2x2 quads. Each quad contains
1975 * 2x2 pixels.
1976 */
1977 static void
1978 generate_fragment(struct llvmpipe_context *lp,
1979 struct lp_fragment_shader *shader,
1980 struct lp_fragment_shader_variant *variant,
1981 unsigned partial_mask)
1982 {
1983 struct gallivm_state *gallivm = variant->gallivm;
1984 const struct lp_fragment_shader_variant_key *key = &variant->key;
1985 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1986 char func_name[256];
1987 struct lp_type fs_type;
1988 struct lp_type blend_type;
1989 LLVMTypeRef fs_elem_type;
1990 LLVMTypeRef blend_vec_type;
1991 LLVMTypeRef arg_types[12];
1992 LLVMTypeRef func_type;
1993 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1994 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1995 LLVMValueRef context_ptr;
1996 LLVMValueRef x;
1997 LLVMValueRef y;
1998 LLVMValueRef a0_ptr;
1999 LLVMValueRef dadx_ptr;
2000 LLVMValueRef dady_ptr;
2001 LLVMValueRef color_ptr_ptr;
2002 LLVMValueRef stride_ptr;
2003 LLVMValueRef depth_ptr;
2004 LLVMValueRef mask_input;
2005 LLVMValueRef thread_data_ptr;
2006 LLVMBasicBlockRef block;
2007 LLVMBuilderRef builder;
2008 struct lp_build_sampler_soa *sampler;
2009 struct lp_build_interp_soa_context interp;
2010 LLVMValueRef fs_mask[16 / 4];
2011 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2012 LLVMValueRef function;
2013 LLVMValueRef facing;
2014 const struct util_format_description *zs_format_desc;
2015 unsigned num_fs;
2016 unsigned i;
2017 unsigned chan;
2018 unsigned cbuf;
2019 boolean cbuf0_write_all;
2020 boolean try_loop = TRUE;
2021 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2022 util_blend_state_is_dual(&key->blend, 0);
2023
2024 assert(lp_native_vector_width / 32 >= 4);
2025
2026 /* Adjust color input interpolation according to flatshade state:
2027 */
2028 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2029 for (i = 0; i < shader->info.base.num_inputs; i++) {
2030 if (inputs[i].interp == LP_INTERP_COLOR) {
2031 if (key->flatshade)
2032 inputs[i].interp = LP_INTERP_CONSTANT;
2033 else
2034 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2035 }
2036 }
2037
2038 /* check if writes to cbuf[0] are to be copied to all cbufs */
2039 cbuf0_write_all = FALSE;
2040 for (i = 0;i < shader->info.base.num_properties; i++) {
2041 if (shader->info.base.properties[i].name ==
2042 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
2043 cbuf0_write_all = TRUE;
2044 break;
2045 }
2046 }
2047
2048 /* TODO: actually pick these based on the fs and color buffer
2049 * characteristics. */
2050
2051 memset(&fs_type, 0, sizeof fs_type);
2052 fs_type.floating = TRUE; /* floating point values */
2053 fs_type.sign = TRUE; /* values are signed */
2054 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2055 fs_type.width = 32; /* 32-bit float */
2056 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2057 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2058
2059 memset(&blend_type, 0, sizeof blend_type);
2060 blend_type.floating = FALSE; /* values are integers */
2061 blend_type.sign = FALSE; /* values are unsigned */
2062 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2063 blend_type.width = 8; /* 8-bit ubyte values */
2064 blend_type.length = 16; /* 16 elements per vector */
2065
2066 /*
2067 * Generate the function prototype. Any change here must be reflected in
2068 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2069 */
2070
2071 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2072
2073 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2074
2075 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2076 shader->no, variant->no, partial_mask ? "partial" : "whole");
2077
2078 arg_types[0] = variant->jit_context_ptr_type; /* context */
2079 arg_types[1] = int32_type; /* x */
2080 arg_types[2] = int32_type; /* y */
2081 arg_types[3] = int32_type; /* facing */
2082 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2083 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2084 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2085 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2086 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2087 arg_types[9] = int32_type; /* mask_input */
2088 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2089 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2090
2091 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2092 arg_types, Elements(arg_types), 0);
2093
2094 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2095 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2096
2097 variant->function[partial_mask] = function;
2098
2099 /* XXX: need to propagate noalias down into color param now we are
2100 * passing a pointer-to-pointer?
2101 */
2102 for(i = 0; i < Elements(arg_types); ++i)
2103 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2104 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
2105
2106 context_ptr = LLVMGetParam(function, 0);
2107 x = LLVMGetParam(function, 1);
2108 y = LLVMGetParam(function, 2);
2109 facing = LLVMGetParam(function, 3);
2110 a0_ptr = LLVMGetParam(function, 4);
2111 dadx_ptr = LLVMGetParam(function, 5);
2112 dady_ptr = LLVMGetParam(function, 6);
2113 color_ptr_ptr = LLVMGetParam(function, 7);
2114 depth_ptr = LLVMGetParam(function, 8);
2115 mask_input = LLVMGetParam(function, 9);
2116 thread_data_ptr = LLVMGetParam(function, 10);
2117 stride_ptr = LLVMGetParam(function, 11);
2118
2119 lp_build_name(context_ptr, "context");
2120 lp_build_name(x, "x");
2121 lp_build_name(y, "y");
2122 lp_build_name(a0_ptr, "a0");
2123 lp_build_name(dadx_ptr, "dadx");
2124 lp_build_name(dady_ptr, "dady");
2125 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2126 lp_build_name(depth_ptr, "depth");
2127 lp_build_name(thread_data_ptr, "thread_data");
2128 lp_build_name(mask_input, "mask_input");
2129 lp_build_name(stride_ptr, "stride_ptr");
2130
2131 /*
2132 * Function body
2133 */
2134
2135 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2136 builder = gallivm->builder;
2137 assert(builder);
2138 LLVMPositionBuilderAtEnd(builder, block);
2139
2140 /* code generated texture sampling */
2141 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
2142
2143 zs_format_desc = util_format_description(key->zsbuf_format);
2144
2145 if (!try_loop) {
2146 /*
2147 * The shader input interpolation info is not explicitely baked in the
2148 * shader key, but everything it derives from (TGSI, and flatshade) is
2149 * already included in the shader key.
2150 */
2151 lp_build_interp_soa_init(&interp,
2152 gallivm,
2153 shader->info.base.num_inputs,
2154 inputs,
2155 builder, fs_type,
2156 FALSE,
2157 a0_ptr, dadx_ptr, dady_ptr,
2158 x, y);
2159
2160 /* loop over quads in the block */
2161 for(i = 0; i < num_fs; ++i) {
2162 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
2163 i*fs_type.length*zs_format_desc->block.bits/8,
2164 0);
2165 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2166 LLVMValueRef depth_ptr_i;
2167
2168 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
2169
2170 generate_fs(gallivm,
2171 shader, key,
2172 builder,
2173 fs_type,
2174 context_ptr,
2175 i,
2176 &interp,
2177 sampler,
2178 &fs_mask[i], /* output */
2179 out_color,
2180 depth_ptr_i,
2181 facing,
2182 partial_mask,
2183 mask_input,
2184 thread_data_ptr);
2185
2186 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2187 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
2188 fs_out_color[cbuf][chan][i] =
2189 out_color[cbuf * !cbuf0_write_all][chan];
2190 }
2191 if (dual_source_blend) {
2192 /* only support one dual source blend target hence always use output 1 */
2193 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
2194 fs_out_color[1][chan][i] =
2195 out_color[1][chan];
2196 }
2197 }
2198 }
2199 else {
2200 unsigned depth_bits = zs_format_desc->block.bits/8;
2201 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2202 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2203 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2204 num_loop, "mask_store");
2205 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2206
2207 /*
2208 * The shader input interpolation info is not explicitely baked in the
2209 * shader key, but everything it derives from (TGSI, and flatshade) is
2210 * already included in the shader key.
2211 */
2212 lp_build_interp_soa_init(&interp,
2213 gallivm,
2214 shader->info.base.num_inputs,
2215 inputs,
2216 builder, fs_type,
2217 TRUE,
2218 a0_ptr, dadx_ptr, dady_ptr,
2219 x, y);
2220
2221 for (i = 0; i < num_fs; i++) {
2222 LLVMValueRef mask;
2223 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2224 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2225 &indexi, 1, "mask_ptr");
2226
2227 if (partial_mask) {
2228 mask = generate_quad_mask(gallivm, fs_type,
2229 i*fs_type.length/4, mask_input);
2230 }
2231 else {
2232 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2233 }
2234 LLVMBuildStore(builder, mask, mask_ptr);
2235 }
2236
2237 generate_fs_loop(gallivm,
2238 shader, key,
2239 builder,
2240 fs_type,
2241 context_ptr,
2242 num_loop,
2243 &interp,
2244 sampler,
2245 mask_store, /* output */
2246 color_store,
2247 depth_ptr,
2248 depth_bits,
2249 facing,
2250 thread_data_ptr);
2251
2252 for (i = 0; i < num_fs; i++) {
2253 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2254 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2255 &indexi, 1, "");
2256 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2257 /* This is fucked up need to reorganize things */
2258 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2259 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2260 ptr = LLVMBuildGEP(builder,
2261 color_store[cbuf * !cbuf0_write_all][chan],
2262 &indexi, 1, "");
2263 fs_out_color[cbuf][chan][i] = ptr;
2264 }
2265 }
2266 if (dual_source_blend) {
2267 /* only support one dual source blend target hence always use output 1 */
2268 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2269 ptr = LLVMBuildGEP(builder,
2270 color_store[1][chan],
2271 &indexi, 1, "");
2272 fs_out_color[1][chan][i] = ptr;
2273 }
2274 }
2275 }
2276 }
2277
2278 sampler->destroy(sampler);
2279
2280 /* Loop over color outputs / color buffers to do blending.
2281 */
2282 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2283 LLVMValueRef color_ptr;
2284 LLVMValueRef stride;
2285 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2286
2287 boolean do_branch = ((key->depth.enabled
2288 || key->stencil[0].enabled
2289 || key->alpha.enabled)
2290 && !shader->info.base.uses_kill);
2291
2292 color_ptr = LLVMBuildLoad(builder,
2293 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2294 "");
2295
2296 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2297
2298 stride = LLVMBuildLoad(builder,
2299 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2300 "");
2301
2302 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2303 num_fs, fs_type, fs_mask, fs_out_color,
2304 context_ptr, color_ptr, stride, partial_mask, do_branch);
2305 }
2306
2307 LLVMBuildRetVoid(builder);
2308
2309 gallivm_verify_function(gallivm, function);
2310
2311 variant->nr_instrs += lp_build_count_instructions(function);
2312 }
2313
2314
2315 static void
2316 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2317 {
2318 unsigned i;
2319
2320 debug_printf("fs variant %p:\n", (void *) key);
2321
2322 if (key->flatshade) {
2323 debug_printf("flatshade = 1\n");
2324 }
2325 for (i = 0; i < key->nr_cbufs; ++i) {
2326 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2327 }
2328 if (key->depth.enabled) {
2329 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2330 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2331 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2332 }
2333
2334 for (i = 0; i < 2; ++i) {
2335 if (key->stencil[i].enabled) {
2336 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2337 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2338 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2339 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2340 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2341 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2342 }
2343 }
2344
2345 if (key->alpha.enabled) {
2346 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2347 }
2348
2349 if (key->occlusion_count) {
2350 debug_printf("occlusion_count = 1\n");
2351 }
2352
2353 if (key->blend.logicop_enable) {
2354 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2355 }
2356 else if (key->blend.rt[0].blend_enable) {
2357 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2358 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2359 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2360 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2361 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2362 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2363 }
2364 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2365 for (i = 0; i < key->nr_samplers; ++i) {
2366 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2367 debug_printf("sampler[%u] = \n", i);
2368 debug_printf(" .wrap = %s %s %s\n",
2369 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2370 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2371 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2372 debug_printf(" .min_img_filter = %s\n",
2373 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2374 debug_printf(" .min_mip_filter = %s\n",
2375 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2376 debug_printf(" .mag_img_filter = %s\n",
2377 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2378 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2379 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2380 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2381 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2382 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2383 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2384 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2385 }
2386 for (i = 0; i < key->nr_sampler_views; ++i) {
2387 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2388 debug_printf("texture[%u] = \n", i);
2389 debug_printf(" .format = %s\n",
2390 util_format_name(texture->format));
2391 debug_printf(" .target = %s\n",
2392 util_dump_tex_target(texture->target, TRUE));
2393 debug_printf(" .level_zero_only = %u\n",
2394 texture->level_zero_only);
2395 debug_printf(" .pot = %u %u %u\n",
2396 texture->pot_width,
2397 texture->pot_height,
2398 texture->pot_depth);
2399 }
2400 }
2401
2402
2403 void
2404 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2405 {
2406 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2407 variant->shader->no, variant->no);
2408 tgsi_dump(variant->shader->base.tokens, 0);
2409 dump_fs_variant_key(&variant->key);
2410 debug_printf("variant->opaque = %u\n", variant->opaque);
2411 debug_printf("\n");
2412 }
2413
2414
2415 /**
2416 * Generate a new fragment shader variant from the shader code and
2417 * other state indicated by the key.
2418 */
2419 static struct lp_fragment_shader_variant *
2420 generate_variant(struct llvmpipe_context *lp,
2421 struct lp_fragment_shader *shader,
2422 const struct lp_fragment_shader_variant_key *key)
2423 {
2424 struct lp_fragment_shader_variant *variant;
2425 const struct util_format_description *cbuf0_format_desc;
2426 boolean fullcolormask;
2427
2428 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2429 if(!variant)
2430 return NULL;
2431
2432 variant->gallivm = gallivm_create();
2433 if (!variant->gallivm) {
2434 FREE(variant);
2435 return NULL;
2436 }
2437
2438 variant->shader = shader;
2439 variant->list_item_global.base = variant;
2440 variant->list_item_local.base = variant;
2441 variant->no = shader->variants_created++;
2442
2443 memcpy(&variant->key, key, shader->variant_key_size);
2444
2445 /*
2446 * Determine whether we are touching all channels in the color buffer.
2447 */
2448 fullcolormask = FALSE;
2449 if (key->nr_cbufs == 1) {
2450 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2451 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2452 }
2453
2454 variant->opaque =
2455 !key->blend.logicop_enable &&
2456 !key->blend.rt[0].blend_enable &&
2457 fullcolormask &&
2458 !key->stencil[0].enabled &&
2459 !key->alpha.enabled &&
2460 !key->depth.enabled &&
2461 !shader->info.base.uses_kill
2462 ? TRUE : FALSE;
2463
2464 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2465 lp_debug_fs_variant(variant);
2466 }
2467
2468 lp_jit_init_types(variant);
2469
2470 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2471 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2472
2473 if (variant->jit_function[RAST_WHOLE] == NULL) {
2474 if (variant->opaque) {
2475 /* Specialized shader, which doesn't need to read the color buffer. */
2476 generate_fragment(lp, shader, variant, RAST_WHOLE);
2477 }
2478 }
2479
2480 /*
2481 * Compile everything
2482 */
2483
2484 gallivm_compile_module(variant->gallivm);
2485
2486 if (variant->function[RAST_EDGE_TEST]) {
2487 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2488 gallivm_jit_function(variant->gallivm,
2489 variant->function[RAST_EDGE_TEST]);
2490 }
2491
2492 if (variant->function[RAST_WHOLE]) {
2493 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2494 gallivm_jit_function(variant->gallivm,
2495 variant->function[RAST_WHOLE]);
2496 } else if (!variant->jit_function[RAST_WHOLE]) {
2497 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2498 }
2499
2500 return variant;
2501 }
2502
2503
2504 static void *
2505 llvmpipe_create_fs_state(struct pipe_context *pipe,
2506 const struct pipe_shader_state *templ)
2507 {
2508 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2509 struct lp_fragment_shader *shader;
2510 int nr_samplers;
2511 int nr_sampler_views;
2512 int i;
2513
2514 shader = CALLOC_STRUCT(lp_fragment_shader);
2515 if (!shader)
2516 return NULL;
2517
2518 shader->no = fs_no++;
2519 make_empty_list(&shader->variants);
2520
2521 /* get/save the summary info for this shader */
2522 lp_build_tgsi_info(templ->tokens, &shader->info);
2523
2524 /* we need to keep a local copy of the tokens */
2525 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2526
2527 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2528 if (shader->draw_data == NULL) {
2529 FREE((void *) shader->base.tokens);
2530 FREE(shader);
2531 return NULL;
2532 }
2533
2534 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2535 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2536
2537 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2538 state[MAX2(nr_samplers, nr_sampler_views)]);
2539
2540 for (i = 0; i < shader->info.base.num_inputs; i++) {
2541 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2542 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2543
2544 switch (shader->info.base.input_interpolate[i]) {
2545 case TGSI_INTERPOLATE_CONSTANT:
2546 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2547 break;
2548 case TGSI_INTERPOLATE_LINEAR:
2549 shader->inputs[i].interp = LP_INTERP_LINEAR;
2550 break;
2551 case TGSI_INTERPOLATE_PERSPECTIVE:
2552 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2553 break;
2554 case TGSI_INTERPOLATE_COLOR:
2555 shader->inputs[i].interp = LP_INTERP_COLOR;
2556 break;
2557 default:
2558 assert(0);
2559 break;
2560 }
2561
2562 switch (shader->info.base.input_semantic_name[i]) {
2563 case TGSI_SEMANTIC_FACE:
2564 shader->inputs[i].interp = LP_INTERP_FACING;
2565 break;
2566 case TGSI_SEMANTIC_POSITION:
2567 /* Position was already emitted above
2568 */
2569 shader->inputs[i].interp = LP_INTERP_POSITION;
2570 shader->inputs[i].src_index = 0;
2571 continue;
2572 }
2573
2574 shader->inputs[i].src_index = i+1;
2575 }
2576
2577 if (LP_DEBUG & DEBUG_TGSI) {
2578 unsigned attrib;
2579 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2580 shader->no, (void *) shader);
2581 tgsi_dump(templ->tokens, 0);
2582 debug_printf("usage masks:\n");
2583 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2584 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2585 debug_printf(" IN[%u].%s%s%s%s\n",
2586 attrib,
2587 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2588 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2589 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2590 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2591 }
2592 debug_printf("\n");
2593 }
2594
2595 return shader;
2596 }
2597
2598
2599 static void
2600 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2601 {
2602 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2603
2604 if (llvmpipe->fs == fs)
2605 return;
2606
2607 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2608
2609 draw_bind_fragment_shader(llvmpipe->draw,
2610 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2611
2612 llvmpipe->dirty |= LP_NEW_FS;
2613 }
2614
2615
2616 /**
2617 * Remove shader variant from two lists: the shader's variant list
2618 * and the context's variant list.
2619 */
2620 void
2621 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2622 struct lp_fragment_shader_variant *variant)
2623 {
2624 unsigned i;
2625
2626 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2627 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2628 " #%u v total cached #%u\n",
2629 variant->shader->no,
2630 variant->no,
2631 variant->shader->variants_created,
2632 variant->shader->variants_cached,
2633 lp->nr_fs_variants);
2634 }
2635
2636 /* free all the variant's JIT'd functions */
2637 for (i = 0; i < Elements(variant->function); i++) {
2638 if (variant->function[i]) {
2639 gallivm_free_function(variant->gallivm,
2640 variant->function[i],
2641 variant->jit_function[i]);
2642 }
2643 }
2644
2645 gallivm_destroy(variant->gallivm);
2646
2647 /* remove from shader's list */
2648 remove_from_list(&variant->list_item_local);
2649 variant->shader->variants_cached--;
2650
2651 /* remove from context's list */
2652 remove_from_list(&variant->list_item_global);
2653 lp->nr_fs_variants--;
2654 lp->nr_fs_instrs -= variant->nr_instrs;
2655
2656 FREE(variant);
2657 }
2658
2659
2660 static void
2661 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2662 {
2663 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2664 struct lp_fragment_shader *shader = fs;
2665 struct lp_fs_variant_list_item *li;
2666
2667 assert(fs != llvmpipe->fs);
2668
2669 /*
2670 * XXX: we need to flush the context until we have some sort of reference
2671 * counting in fragment shaders as they may still be binned
2672 * Flushing alone might not sufficient we need to wait on it too.
2673 */
2674 llvmpipe_finish(pipe, __FUNCTION__);
2675
2676 /* Delete all the variants */
2677 li = first_elem(&shader->variants);
2678 while(!at_end(&shader->variants, li)) {
2679 struct lp_fs_variant_list_item *next = next_elem(li);
2680 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2681 li = next;
2682 }
2683
2684 /* Delete draw module's data */
2685 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2686
2687 assert(shader->variants_cached == 0);
2688 FREE((void *) shader->base.tokens);
2689 FREE(shader);
2690 }
2691
2692
2693
2694 static void
2695 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2696 uint shader, uint index,
2697 struct pipe_constant_buffer *cb)
2698 {
2699 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2700 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2701
2702 assert(shader < PIPE_SHADER_TYPES);
2703 assert(index < Elements(llvmpipe->constants[shader]));
2704
2705 /* note: reference counting */
2706 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2707
2708 if (shader == PIPE_SHADER_VERTEX ||
2709 shader == PIPE_SHADER_GEOMETRY) {
2710 /* Pass the constants to the 'draw' module */
2711 const unsigned size = cb ? cb->buffer_size : 0;
2712 const ubyte *data;
2713
2714 if (constants) {
2715 data = (ubyte *) llvmpipe_resource_data(constants);
2716 }
2717 else if (cb && cb->user_buffer) {
2718 data = (ubyte *) cb->user_buffer;
2719 }
2720 else {
2721 data = NULL;
2722 }
2723
2724 if (data)
2725 data += cb->buffer_offset;
2726
2727 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2728 index, data, size);
2729 }
2730
2731 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2732
2733 if (cb && cb->user_buffer) {
2734 pipe_resource_reference(&constants, NULL);
2735 }
2736 }
2737
2738
2739 /**
2740 * Return the blend factor equivalent to a destination alpha of one.
2741 */
2742 static INLINE unsigned
2743 force_dst_alpha_one(unsigned factor)
2744 {
2745 switch(factor) {
2746 case PIPE_BLENDFACTOR_DST_ALPHA:
2747 return PIPE_BLENDFACTOR_ONE;
2748 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2749 return PIPE_BLENDFACTOR_ZERO;
2750 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2751 return PIPE_BLENDFACTOR_ZERO;
2752 }
2753
2754 return factor;
2755 }
2756
2757
2758 /**
2759 * We need to generate several variants of the fragment pipeline to match
2760 * all the combinations of the contributing state atoms.
2761 *
2762 * TODO: there is actually no reason to tie this to context state -- the
2763 * generated code could be cached globally in the screen.
2764 */
2765 static void
2766 make_variant_key(struct llvmpipe_context *lp,
2767 struct lp_fragment_shader *shader,
2768 struct lp_fragment_shader_variant_key *key)
2769 {
2770 unsigned i;
2771
2772 memset(key, 0, shader->variant_key_size);
2773
2774 if (lp->framebuffer.zsbuf) {
2775 if (lp->depth_stencil->depth.enabled) {
2776 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2777 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2778 }
2779 if (lp->depth_stencil->stencil[0].enabled) {
2780 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2781 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2782 }
2783 }
2784
2785 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2786 if (!lp->framebuffer.nr_cbufs ||
2787 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2788 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2789 }
2790 if(key->alpha.enabled)
2791 key->alpha.func = lp->depth_stencil->alpha.func;
2792 /* alpha.ref_value is passed in jit_context */
2793
2794 key->flatshade = lp->rasterizer->flatshade;
2795 if (lp->active_occlusion_query) {
2796 key->occlusion_count = TRUE;
2797 }
2798
2799 if (lp->framebuffer.nr_cbufs) {
2800 memcpy(&key->blend, lp->blend, sizeof key->blend);
2801 }
2802
2803 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2804
2805 if (!key->blend.independent_blend_enable) {
2806 /* we always need independent blend otherwise the fixups below won't work */
2807 for (i = 1; i < key->nr_cbufs; i++) {
2808 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2809 }
2810 key->blend.independent_blend_enable = 1;
2811 }
2812
2813 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2814 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2815 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2816 const struct util_format_description *format_desc;
2817
2818 key->cbuf_format[i] = format;
2819
2820 format_desc = util_format_description(format);
2821 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2822 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2823
2824 /*
2825 * Mask out color channels not present in the color buffer.
2826 */
2827 blend_rt->colormask &= util_format_colormask(format_desc);
2828
2829 /*
2830 * Disable blend for integer formats.
2831 */
2832 if (util_format_is_pure_integer(format)) {
2833 blend_rt->blend_enable = 0;
2834 }
2835
2836 /*
2837 * Our swizzled render tiles always have an alpha channel, but the linear
2838 * render target format often does not, so force here the dst alpha to be
2839 * one.
2840 *
2841 * This is not a mere optimization. Wrong results will be produced if the
2842 * dst alpha is used, the dst format does not have alpha, and the previous
2843 * rendering was not flushed from the swizzled to linear buffer. For
2844 * example, NonPowTwo DCT.
2845 *
2846 * TODO: This should be generalized to all channels for better
2847 * performance, but only alpha causes correctness issues.
2848 *
2849 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2850 */
2851 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2852 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2853 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2854 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2855 blend_rt->alpha_func = blend_rt->rgb_func;
2856 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2857 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2858 }
2859 }
2860
2861 /* This value will be the same for all the variants of a given shader:
2862 */
2863 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2864
2865 for(i = 0; i < key->nr_samplers; ++i) {
2866 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2867 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2868 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2869 }
2870 }
2871
2872 /*
2873 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2874 * are dx10-style? Can't really have mixed opcodes, at least not
2875 * if we want to skip the holes here (without rescanning tgsi).
2876 */
2877 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2878 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2879 for(i = 0; i < key->nr_sampler_views; ++i) {
2880 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2881 lp_sampler_static_texture_state(&key->state[i].texture_state,
2882 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2883 }
2884 }
2885 }
2886 else {
2887 key->nr_sampler_views = key->nr_samplers;
2888 for(i = 0; i < key->nr_sampler_views; ++i) {
2889 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2890 lp_sampler_static_texture_state(&key->state[i].texture_state,
2891 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2892 }
2893 }
2894 }
2895 }
2896
2897
2898
2899 /**
2900 * Update fragment shader state. This is called just prior to drawing
2901 * something when some fragment-related state has changed.
2902 */
2903 void
2904 llvmpipe_update_fs(struct llvmpipe_context *lp)
2905 {
2906 struct lp_fragment_shader *shader = lp->fs;
2907 struct lp_fragment_shader_variant_key key;
2908 struct lp_fragment_shader_variant *variant = NULL;
2909 struct lp_fs_variant_list_item *li;
2910
2911 make_variant_key(lp, shader, &key);
2912
2913 /* Search the variants for one which matches the key */
2914 li = first_elem(&shader->variants);
2915 while(!at_end(&shader->variants, li)) {
2916 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2917 variant = li->base;
2918 break;
2919 }
2920 li = next_elem(li);
2921 }
2922
2923 if (variant) {
2924 /* Move this variant to the head of the list to implement LRU
2925 * deletion of shader's when we have too many.
2926 */
2927 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2928 }
2929 else {
2930 /* variant not found, create it now */
2931 int64_t t0, t1, dt;
2932 unsigned i;
2933 unsigned variants_to_cull;
2934
2935 if (0) {
2936 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2937 lp->nr_fs_variants,
2938 lp->nr_fs_instrs,
2939 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2940 }
2941
2942 /* First, check if we've exceeded the max number of shader variants.
2943 * If so, free 25% of them (the least recently used ones).
2944 */
2945 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2946
2947 if (variants_to_cull ||
2948 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2949 struct pipe_context *pipe = &lp->pipe;
2950
2951 /*
2952 * XXX: we need to flush the context until we have some sort of
2953 * reference counting in fragment shaders as they may still be binned
2954 * Flushing alone might not be sufficient we need to wait on it too.
2955 */
2956 llvmpipe_finish(pipe, __FUNCTION__);
2957
2958 /*
2959 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2960 * number of shader variants (potentially all of them) could be
2961 * pending for destruction on flush.
2962 */
2963
2964 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2965 struct lp_fs_variant_list_item *item;
2966 if (is_empty_list(&lp->fs_variants_list)) {
2967 break;
2968 }
2969 item = last_elem(&lp->fs_variants_list);
2970 assert(item);
2971 assert(item->base);
2972 llvmpipe_remove_shader_variant(lp, item->base);
2973 }
2974 }
2975
2976 /*
2977 * Generate the new variant.
2978 */
2979 t0 = os_time_get();
2980 variant = generate_variant(lp, shader, &key);
2981 t1 = os_time_get();
2982 dt = t1 - t0;
2983 LP_COUNT_ADD(llvm_compile_time, dt);
2984 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2985
2986 llvmpipe_variant_count++;
2987
2988 /* Put the new variant into the list */
2989 if (variant) {
2990 insert_at_head(&shader->variants, &variant->list_item_local);
2991 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2992 lp->nr_fs_variants++;
2993 lp->nr_fs_instrs += variant->nr_instrs;
2994 shader->variants_cached++;
2995 }
2996 }
2997
2998 /* Bind this variant */
2999 lp_setup_set_fs_variant(lp->setup, variant);
3000 }
3001
3002
3003
3004
3005
3006 void
3007 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3008 {
3009 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3010 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3011 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3012
3013 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3014 }