llvmpipe: handle z32s8x24 depth/stencil format
[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 */
219 static void
220 generate_fs_loop(struct gallivm_state *gallivm,
221 struct lp_fragment_shader *shader,
222 const struct lp_fragment_shader_variant_key *key,
223 LLVMBuilderRef builder,
224 struct lp_type type,
225 LLVMValueRef context_ptr,
226 LLVMValueRef num_loop,
227 struct lp_build_interp_soa_context *interp,
228 struct lp_build_sampler_soa *sampler,
229 LLVMValueRef mask_store,
230 LLVMValueRef (*out_color)[4],
231 LLVMValueRef depth_ptr,
232 LLVMValueRef depth_stride,
233 LLVMValueRef facing,
234 LLVMValueRef thread_data_ptr)
235 {
236 const struct util_format_description *zs_format_desc = NULL;
237 const struct tgsi_token *tokens = shader->base.tokens;
238 LLVMTypeRef vec_type;
239 LLVMValueRef mask_ptr, mask_val;
240 LLVMValueRef consts_ptr;
241 LLVMValueRef z;
242 LLVMValueRef z_value, s_value;
243 LLVMValueRef z_fb, s_fb;
244 LLVMValueRef stencil_refs[2];
245 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
246 struct lp_build_for_loop_state loop_state;
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
258 struct lp_bld_tgsi_system_values system_values;
259
260 memset(&system_values, 0, sizeof(system_values));
261
262 if (key->depth.enabled ||
263 key->stencil[0].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 (key->stencil[1].enabled &&
286 key->stencil[1].writemask)))))
287 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
288 }
289 else {
290 depth_mode = 0;
291 }
292
293
294 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
295 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
296
297 vec_type = lp_build_vec_type(gallivm, type);
298
299 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
300
301 lp_build_for_loop_begin(&loop_state, gallivm,
302 lp_build_const_int32(gallivm, 0),
303 LLVMIntULT,
304 num_loop,
305 lp_build_const_int32(gallivm, 1));
306
307 mask_ptr = LLVMBuildGEP(builder, mask_store,
308 &loop_state.counter, 1, "mask_ptr");
309 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
310
311 memset(outputs, 0, sizeof outputs);
312
313 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
314 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
315 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
316 lp_build_vec_type(gallivm,
317 type),
318 num_loop, "color");
319 }
320 }
321 if (dual_source_blend) {
322 assert(key->nr_cbufs <= 1);
323 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
324 out_color[1][chan] = lp_build_array_alloca(gallivm,
325 lp_build_vec_type(gallivm,
326 type),
327 num_loop, "color1");
328 }
329 }
330
331
332 /* 'mask' will control execution based on quad's pixel alive/killed state */
333 lp_build_mask_begin(&mask, gallivm, type, mask_val);
334
335 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
336 lp_build_mask_check(&mask);
337
338 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
339 z = interp->pos[2];
340
341 if (depth_mode & EARLY_DEPTH_TEST) {
342 lp_build_depth_stencil_load_swizzled(gallivm, type,
343 zs_format_desc,
344 depth_ptr, depth_stride,
345 &z_fb, &s_fb, loop_state.counter);
346 lp_build_depth_stencil_test(gallivm,
347 &key->depth,
348 key->stencil,
349 type,
350 zs_format_desc,
351 &mask,
352 stencil_refs,
353 z, z_fb, s_fb,
354 facing,
355 &z_value, &s_value,
356 !simple_shader);
357
358 if (depth_mode & EARLY_DEPTH_WRITE) {
359 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
360 NULL, NULL, NULL, loop_state.counter,
361 depth_ptr, depth_stride,
362 z_value, s_value);
363 }
364 }
365
366 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
367
368 /* Build the actual shader */
369 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
370 consts_ptr, &system_values,
371 interp->inputs,
372 outputs, sampler, &shader->info.base, NULL);
373
374 /* Alpha test */
375 if (key->alpha.enabled) {
376 int color0 = find_output_by_semantic(&shader->info.base,
377 TGSI_SEMANTIC_COLOR,
378 0);
379
380 if (color0 != -1 && outputs[color0][3]) {
381 const struct util_format_description *cbuf_format_desc;
382 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
383 LLVMValueRef alpha_ref_value;
384
385 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
386 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
387
388 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
389
390 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
391 &mask, alpha, alpha_ref_value,
392 (depth_mode & LATE_DEPTH_TEST) != 0);
393 }
394 }
395
396 /* Late Z test */
397 if (depth_mode & LATE_DEPTH_TEST) {
398 int pos0 = find_output_by_semantic(&shader->info.base,
399 TGSI_SEMANTIC_POSITION,
400 0);
401
402 if (pos0 != -1 && outputs[pos0][2]) {
403 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
404 }
405
406 lp_build_depth_stencil_load_swizzled(gallivm, type,
407 zs_format_desc,
408 depth_ptr, depth_stride,
409 &z_fb, &s_fb, loop_state.counter);
410
411 lp_build_depth_stencil_test(gallivm,
412 &key->depth,
413 key->stencil,
414 type,
415 zs_format_desc,
416 &mask,
417 stencil_refs,
418 z, z_fb, s_fb,
419 facing,
420 &z_value, &s_value,
421 !simple_shader);
422 /* Late Z write */
423 if (depth_mode & LATE_DEPTH_WRITE) {
424 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
425 NULL, NULL, NULL, loop_state.counter,
426 depth_ptr, depth_stride,
427 z_value, s_value);
428 }
429 }
430 else if ((depth_mode & EARLY_DEPTH_TEST) &&
431 (depth_mode & LATE_DEPTH_WRITE))
432 {
433 /* Need to apply a reduced mask to the depth write. Reload the
434 * depth value, update from zs_value with the new mask value and
435 * write that out.
436 */
437 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
438 &mask, z_fb, s_fb, loop_state.counter,
439 depth_ptr, depth_stride,
440 z_value, s_value);
441 }
442
443
444 /* Color write */
445 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
446 {
447 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
448 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
449 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
450 {
451 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
452 if(outputs[attrib][chan]) {
453 /* XXX: just initialize outputs to point at colors[] and
454 * skip this.
455 */
456 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
457 LLVMValueRef color_ptr;
458 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
459 &loop_state.counter, 1, "");
460 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
461 LLVMBuildStore(builder, out, color_ptr);
462 }
463 }
464 }
465 }
466
467 if (key->occlusion_count) {
468 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
469 lp_build_name(counter, "counter");
470 lp_build_occlusion_count(gallivm, type,
471 lp_build_mask_value(&mask), counter);
472 }
473
474 mask_val = lp_build_mask_end(&mask);
475 LLVMBuildStore(builder, mask_val, mask_ptr);
476 lp_build_for_loop_end(&loop_state);
477 }
478
479
480 /**
481 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
482 *
483 * Fragment Shader outputs pixels in small 2x2 blocks
484 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
485 *
486 * However in memory pixels are stored in rows
487 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
488 *
489 * @param type fragment shader type (4x or 8x float)
490 * @param num_fs number of fs_src
491 * @param dst_channels number of output channels
492 * @param fs_src output from fragment shader
493 * @param dst pointer to store result
494 * @param pad_inline is channel padding inline or at end of row
495 * @return the number of dsts
496 */
497 static int
498 generate_fs_twiddle(struct gallivm_state *gallivm,
499 struct lp_type type,
500 unsigned num_fs,
501 unsigned dst_channels,
502 LLVMValueRef fs_src[][4],
503 LLVMValueRef* dst,
504 bool pad_inline)
505 {
506 LLVMValueRef src[16];
507
508 bool swizzle_pad;
509 bool twiddle;
510 bool split;
511
512 unsigned pixels = num_fs == 4 ? 1 : 2;
513 unsigned reorder_group;
514 unsigned src_channels;
515 unsigned src_count;
516 unsigned i;
517
518 src_channels = dst_channels < 3 ? dst_channels : 4;
519 src_count = num_fs * src_channels;
520
521 assert(pixels == 2 || num_fs == 4);
522 assert(num_fs * src_channels <= Elements(src));
523
524 /*
525 * Transpose from SoA -> AoS
526 */
527 for (i = 0; i < num_fs; ++i) {
528 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
529 }
530
531 /*
532 * Pick transformation options
533 */
534 swizzle_pad = false;
535 twiddle = false;
536 split = false;
537 reorder_group = 0;
538
539 if (dst_channels == 1) {
540 twiddle = true;
541
542 if (pixels == 2) {
543 split = true;
544 }
545 } else if (dst_channels == 2) {
546 if (pixels == 1) {
547 reorder_group = 1;
548 }
549 } else if (dst_channels > 2) {
550 if (pixels == 1) {
551 reorder_group = 2;
552 } else {
553 twiddle = true;
554 }
555
556 if (!pad_inline && dst_channels == 3 && pixels > 1) {
557 swizzle_pad = true;
558 }
559 }
560
561 /*
562 * Split the src in half
563 */
564 if (split) {
565 for (i = num_fs; i > 0; --i) {
566 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
567 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
568 }
569
570 src_count *= 2;
571 type.length = 4;
572 }
573
574 /*
575 * Ensure pixels are in memory order
576 */
577 if (reorder_group) {
578 /* Twiddle pixels by reordering the array, e.g.:
579 *
580 * src_count = 8 -> 0 2 1 3 4 6 5 7
581 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
582 */
583 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
584
585 for (i = 0; i < src_count; ++i) {
586 unsigned group = i / reorder_group;
587 unsigned block = (group / 4) * 4 * reorder_group;
588 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
589 dst[i] = src[j];
590 }
591 } else if (twiddle) {
592 /* Twiddle pixels across elements of array */
593 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
594 } else {
595 /* Do nothing */
596 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
597 }
598
599 /*
600 * Moves any padding between pixels to the end
601 * e.g. RGBXRGBX -> RGBRGBXX
602 */
603 if (swizzle_pad) {
604 unsigned char swizzles[16];
605 unsigned elems = pixels * dst_channels;
606
607 for (i = 0; i < type.length; ++i) {
608 if (i < elems)
609 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
610 else
611 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
612 }
613
614 for (i = 0; i < src_count; ++i) {
615 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
616 }
617 }
618
619 return src_count;
620 }
621
622
623 /**
624 * Load an unswizzled block of pixels from memory
625 */
626 static void
627 load_unswizzled_block(struct gallivm_state *gallivm,
628 LLVMValueRef base_ptr,
629 LLVMValueRef stride,
630 unsigned block_width,
631 unsigned block_height,
632 LLVMValueRef* dst,
633 struct lp_type dst_type,
634 unsigned dst_count,
635 unsigned dst_alignment)
636 {
637 LLVMBuilderRef builder = gallivm->builder;
638 unsigned row_size = dst_count / block_height;
639 unsigned i;
640
641 /* Ensure block exactly fits into dst */
642 assert((block_width * block_height) % dst_count == 0);
643
644 for (i = 0; i < dst_count; ++i) {
645 unsigned x = i % row_size;
646 unsigned y = i / row_size;
647
648 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
649 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
650
651 LLVMValueRef gep[2];
652 LLVMValueRef dst_ptr;
653
654 gep[0] = lp_build_const_int32(gallivm, 0);
655 gep[1] = LLVMBuildAdd(builder, bx, by, "");
656
657 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
658 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
659
660 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
661
662 lp_set_load_alignment(dst[i], dst_alignment);
663 }
664 }
665
666
667 /**
668 * Store an unswizzled block of pixels to memory
669 */
670 static void
671 store_unswizzled_block(struct gallivm_state *gallivm,
672 LLVMValueRef base_ptr,
673 LLVMValueRef stride,
674 unsigned block_width,
675 unsigned block_height,
676 LLVMValueRef* src,
677 struct lp_type src_type,
678 unsigned src_count,
679 unsigned src_alignment)
680 {
681 LLVMBuilderRef builder = gallivm->builder;
682 unsigned row_size = src_count / block_height;
683 unsigned i;
684
685 /* Ensure src exactly fits into block */
686 assert((block_width * block_height) % src_count == 0);
687
688 for (i = 0; i < src_count; ++i) {
689 unsigned x = i % row_size;
690 unsigned y = i / row_size;
691
692 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
693 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
694
695 LLVMValueRef gep[2];
696 LLVMValueRef src_ptr;
697
698 gep[0] = lp_build_const_int32(gallivm, 0);
699 gep[1] = LLVMBuildAdd(builder, bx, by, "");
700
701 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
702 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
703
704 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
705
706 lp_set_store_alignment(src_ptr, src_alignment);
707 }
708 }
709
710
711 /**
712 * Checks if a format description is an arithmetic format
713 *
714 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
715 */
716 static INLINE boolean
717 is_arithmetic_format(const struct util_format_description *format_desc)
718 {
719 boolean arith = false;
720 unsigned i;
721
722 for (i = 0; i < format_desc->nr_channels; ++i) {
723 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
724 arith |= (format_desc->channel[i].size % 8) != 0;
725 }
726
727 return arith;
728 }
729
730
731 /**
732 * Retrieves the type representing the memory layout for a format
733 *
734 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
735 */
736 static INLINE void
737 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
738 struct lp_type* type)
739 {
740 unsigned i;
741 unsigned chan;
742
743 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
744 /* just make this a 32bit uint */
745 type->floating = false;
746 type->fixed = false;
747 type->sign = false;
748 type->norm = false;
749 type->width = 32;
750 type->length = 1;
751 return;
752 }
753
754 for (i = 0; i < 4; i++)
755 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
756 break;
757 chan = i;
758
759 memset(type, 0, sizeof(struct lp_type));
760 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
761 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
762 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
763 type->norm = format_desc->channel[chan].normalized;
764
765 if (is_arithmetic_format(format_desc)) {
766 type->width = 0;
767 type->length = 1;
768
769 for (i = 0; i < format_desc->nr_channels; ++i) {
770 type->width += format_desc->channel[i].size;
771 }
772 } else {
773 type->width = format_desc->channel[chan].size;
774 type->length = format_desc->nr_channels;
775 }
776 }
777
778
779 /**
780 * Retrieves the type for a format which is usable in the blending code.
781 *
782 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
783 */
784 static INLINE void
785 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
786 struct lp_type* type)
787 {
788 unsigned i;
789 unsigned chan;
790
791 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
792 /* always use ordinary floats for blending */
793 type->floating = true;
794 type->fixed = false;
795 type->sign = true;
796 type->norm = false;
797 type->width = 32;
798 type->length = 4;
799 return;
800 }
801
802 for (i = 0; i < 4; i++)
803 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
804 break;
805 chan = i;
806
807 memset(type, 0, sizeof(struct lp_type));
808 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
809 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
810 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
811 type->norm = format_desc->channel[chan].normalized;
812 type->width = format_desc->channel[chan].size;
813 type->length = format_desc->nr_channels;
814
815 for (i = 1; i < format_desc->nr_channels; ++i) {
816 if (format_desc->channel[i].size > type->width)
817 type->width = format_desc->channel[i].size;
818 }
819
820 if (type->floating) {
821 type->width = 32;
822 } else {
823 if (type->width <= 8) {
824 type->width = 8;
825 } else if (type->width <= 16) {
826 type->width = 16;
827 } else {
828 type->width = 32;
829 }
830 }
831
832 if (is_arithmetic_format(format_desc) && type->length == 3) {
833 type->length = 4;
834 }
835 }
836
837
838 /**
839 * Scale a normalized value from src_bits to dst_bits
840 */
841 static INLINE LLVMValueRef
842 scale_bits(struct gallivm_state *gallivm,
843 int src_bits,
844 int dst_bits,
845 LLVMValueRef src,
846 struct lp_type src_type)
847 {
848 LLVMBuilderRef builder = gallivm->builder;
849 LLVMValueRef result = src;
850
851 if (dst_bits < src_bits) {
852 /* Scale down by LShr */
853 result = LLVMBuildLShr(builder,
854 src,
855 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
856 "");
857 } else if (dst_bits > src_bits) {
858 /* Scale up bits */
859 int db = dst_bits - src_bits;
860
861 /* Shift left by difference in bits */
862 result = LLVMBuildShl(builder,
863 src,
864 lp_build_const_int_vec(gallivm, src_type, db),
865 "");
866
867 if (db < src_bits) {
868 /* Enough bits in src to fill the remainder */
869 LLVMValueRef lower = LLVMBuildLShr(builder,
870 src,
871 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
872 "");
873
874 result = LLVMBuildOr(builder, result, lower, "");
875 } else if (db > src_bits) {
876 /* Need to repeatedly copy src bits to fill remainder in dst */
877 unsigned n;
878
879 for (n = src_bits; n < dst_bits; n *= 2) {
880 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
881
882 result = LLVMBuildOr(builder,
883 result,
884 LLVMBuildLShr(builder, result, shuv, ""),
885 "");
886 }
887 }
888 }
889
890 return result;
891 }
892
893
894 /**
895 * Convert from memory format to blending format
896 *
897 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
898 */
899 static void
900 convert_to_blend_type(struct gallivm_state *gallivm,
901 const struct util_format_description *src_fmt,
902 struct lp_type src_type,
903 struct lp_type dst_type,
904 LLVMValueRef* src, // and dst
905 unsigned num_srcs)
906 {
907 LLVMValueRef *dst = src;
908 LLVMBuilderRef builder = gallivm->builder;
909 struct lp_type blend_type;
910 struct lp_type mem_type;
911 unsigned i, j, k;
912 unsigned pixels = 16 / num_srcs;
913 bool is_arith;
914
915 /*
916 * full custom path for packed floats - none of the later functions would do
917 * anything useful, and given the lp_type representation they can't be fixed.
918 */
919 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
920 LLVMValueRef tmpsrc[4];
921 /*
922 * This is pretty suboptimal for this case blending in SoA would be much
923 * better, since conversion gets us SoA values so need to convert back.
924 */
925 assert(src_type.width == 32);
926 assert(dst_type.floating);
927 assert(dst_type.width == 32);
928 assert(dst_type.length % 4 == 0);
929 for (i = 0; i < 4; i++) {
930 tmpsrc[i] = src[i];
931 }
932 for (i = 0; i < num_srcs / 4; i++) {
933 LLVMValueRef tmpsoa[4];
934 LLVMValueRef tmps = tmpsrc[i];
935 if (num_srcs == 8) {
936 LLVMValueRef shuffles[8];
937 unsigned j;
938 /* fetch was 4 values but need 8-wide output values */
939 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
940 /*
941 * for 8-wide aos transpose would give us wrong order not matching
942 * incoming converted fs values and mask. ARGH.
943 */
944 for (j = 0; j < 4; j++) {
945 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
946 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
947 }
948 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
949 LLVMConstVector(shuffles, 8), "");
950 }
951 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
952 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
953 }
954 return;
955 }
956
957 lp_mem_type_from_format_desc(src_fmt, &mem_type);
958 lp_blend_type_from_format_desc(src_fmt, &blend_type);
959
960 /* Is the format arithmetic */
961 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
962 is_arith &= !(mem_type.width == 16 && mem_type.floating);
963
964 /* Pad if necessary */
965 if (!is_arith && src_type.length < dst_type.length) {
966 for (i = 0; i < num_srcs; ++i) {
967 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
968 }
969
970 src_type.length = dst_type.length;
971 }
972
973 /* Special case for half-floats */
974 if (mem_type.width == 16 && mem_type.floating) {
975 assert(blend_type.width == 32 && blend_type.floating);
976 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
977 is_arith = false;
978 }
979
980 if (!is_arith) {
981 return;
982 }
983
984 src_type.width = blend_type.width * blend_type.length;
985 blend_type.length *= pixels;
986 src_type.length *= pixels / (src_type.length / mem_type.length);
987
988 for (i = 0; i < num_srcs; ++i) {
989 LLVMValueRef chans[4];
990 LLVMValueRef res = NULL;
991 unsigned sa = 0;
992
993 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
994
995 for (j = 0; j < src_fmt->nr_channels; ++j) {
996 unsigned mask = 0;
997
998 for (k = 0; k < src_fmt->channel[j].size; ++k) {
999 mask |= 1 << k;
1000 }
1001
1002 /* Extract bits from source */
1003 chans[j] = LLVMBuildLShr(builder,
1004 dst[i],
1005 lp_build_const_int_vec(gallivm, src_type, sa),
1006 "");
1007
1008 chans[j] = LLVMBuildAnd(builder,
1009 chans[j],
1010 lp_build_const_int_vec(gallivm, src_type, mask),
1011 "");
1012
1013 /* Scale bits */
1014 if (src_type.norm) {
1015 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1016 blend_type.width, chans[j], src_type);
1017 }
1018
1019 /* Insert bits into correct position */
1020 chans[j] = LLVMBuildShl(builder,
1021 chans[j],
1022 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1023 "");
1024
1025 sa += src_fmt->channel[j].size;
1026
1027 if (j == 0) {
1028 res = chans[j];
1029 } else {
1030 res = LLVMBuildOr(builder, res, chans[j], "");
1031 }
1032 }
1033
1034 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1035 }
1036 }
1037
1038
1039 /**
1040 * Convert from blending format to memory format
1041 *
1042 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1043 */
1044 static void
1045 convert_from_blend_type(struct gallivm_state *gallivm,
1046 const struct util_format_description *src_fmt,
1047 struct lp_type src_type,
1048 struct lp_type dst_type,
1049 LLVMValueRef* src, // and dst
1050 unsigned num_srcs)
1051 {
1052 LLVMValueRef* dst = src;
1053 unsigned i, j, k;
1054 struct lp_type mem_type;
1055 struct lp_type blend_type;
1056 LLVMBuilderRef builder = gallivm->builder;
1057 unsigned pixels = 16 / num_srcs;
1058 bool is_arith;
1059
1060 /*
1061 * full custom path for packed floats - none of the later functions would do
1062 * anything useful, and given the lp_type representation they can't be fixed.
1063 */
1064 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1065 /*
1066 * This is pretty suboptimal for this case blending in SoA would be much
1067 * better - we need to transpose the AoS values back to SoA values for
1068 * conversion/packing.
1069 */
1070 assert(src_type.floating);
1071 assert(src_type.width == 32);
1072 assert(src_type.length % 4 == 0);
1073 assert(dst_type.width == 32);
1074 for (i = 0; i < num_srcs / 4; i++) {
1075 LLVMValueRef tmpsoa[4], tmpdst;
1076 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1077 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1078 if (num_srcs == 8) {
1079 LLVMValueRef tmpaos, shuffles[8];
1080 unsigned j;
1081 /*
1082 * for 8-wide aos transpose has given us wrong order not matching
1083 * output order. HMPF. Also need to split the output values manually.
1084 */
1085 for (j = 0; j < 4; j++) {
1086 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1087 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1088 }
1089 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1090 LLVMConstVector(shuffles, 8), "");
1091 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1092 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1093 }
1094 else {
1095 src[i] = tmpdst;
1096 }
1097 }
1098 return;
1099 }
1100
1101 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1102 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1103
1104 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1105
1106 /* Special case for half-floats */
1107 if (mem_type.width == 16 && mem_type.floating) {
1108 int length = dst_type.length;
1109 assert(blend_type.width == 32 && blend_type.floating);
1110
1111 dst_type.length = src_type.length;
1112
1113 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1114
1115 dst_type.length = length;
1116 is_arith = false;
1117 }
1118
1119 /* Remove any padding */
1120 if (!is_arith && (src_type.length % mem_type.length)) {
1121 src_type.length -= (src_type.length % mem_type.length);
1122
1123 for (i = 0; i < num_srcs; ++i) {
1124 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1125 }
1126 }
1127
1128 /* No bit arithmetic to do */
1129 if (!is_arith) {
1130 return;
1131 }
1132
1133 src_type.length = pixels;
1134 src_type.width = blend_type.length * blend_type.width;
1135 dst_type.length = pixels;
1136
1137 for (i = 0; i < num_srcs; ++i) {
1138 LLVMValueRef chans[4];
1139 LLVMValueRef res = NULL;
1140 unsigned sa = 0;
1141
1142 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1143
1144 for (j = 0; j < src_fmt->nr_channels; ++j) {
1145 unsigned mask = 0;
1146
1147 assert(blend_type.width > src_fmt->channel[j].size);
1148
1149 for (k = 0; k < blend_type.width; ++k) {
1150 mask |= 1 << k;
1151 }
1152
1153 /* Extract bits */
1154 chans[j] = LLVMBuildLShr(builder,
1155 dst[i],
1156 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1157 "");
1158
1159 chans[j] = LLVMBuildAnd(builder,
1160 chans[j],
1161 lp_build_const_int_vec(gallivm, src_type, mask),
1162 "");
1163
1164 /* Scale down bits */
1165 if (src_type.norm) {
1166 chans[j] = scale_bits(gallivm, blend_type.width,
1167 src_fmt->channel[j].size, chans[j], src_type);
1168 }
1169
1170 /* Insert bits */
1171 chans[j] = LLVMBuildShl(builder,
1172 chans[j],
1173 lp_build_const_int_vec(gallivm, src_type, sa),
1174 "");
1175
1176 sa += src_fmt->channel[j].size;
1177
1178 if (j == 0) {
1179 res = chans[j];
1180 } else {
1181 res = LLVMBuildOr(builder, res, chans[j], "");
1182 }
1183 }
1184
1185 assert (dst_type.width != 24);
1186
1187 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1188 }
1189 }
1190
1191
1192 /**
1193 * Convert alpha to same blend type as src
1194 */
1195 static void
1196 convert_alpha(struct gallivm_state *gallivm,
1197 struct lp_type row_type,
1198 struct lp_type alpha_type,
1199 const unsigned block_size,
1200 const unsigned block_height,
1201 const unsigned src_count,
1202 const unsigned dst_channels,
1203 const bool pad_inline,
1204 LLVMValueRef* src_alpha)
1205 {
1206 LLVMBuilderRef builder = gallivm->builder;
1207 unsigned i, j;
1208 unsigned length = row_type.length;
1209 row_type.length = alpha_type.length;
1210
1211 /* Twiddle the alpha to match pixels */
1212 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1213
1214 for (i = 0; i < 4; ++i) {
1215 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1216 }
1217
1218 alpha_type = row_type;
1219 row_type.length = length;
1220
1221 /* If only one channel we can only need the single alpha value per pixel */
1222 if (src_count == 1) {
1223 assert(dst_channels == 1);
1224
1225 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1226 } else {
1227 /* If there are more srcs than rows then we need to split alpha up */
1228 if (src_count > block_height) {
1229 for (i = src_count; i > 0; --i) {
1230 unsigned pixels = block_size / src_count;
1231 unsigned idx = i - 1;
1232
1233 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1234 }
1235 }
1236
1237 /* If there is a src for each pixel broadcast the alpha across whole row */
1238 if (src_count == block_size) {
1239 for (i = 0; i < src_count; ++i) {
1240 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1241 }
1242 } else {
1243 unsigned pixels = block_size / src_count;
1244 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1245 unsigned alpha_span = 1;
1246 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1247
1248 /* Check if we need 2 src_alphas for our shuffles */
1249 if (pixels > alpha_type.length) {
1250 alpha_span = 2;
1251 }
1252
1253 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1254 for (j = 0; j < row_type.length; ++j) {
1255 if (j < pixels * channels) {
1256 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1257 } else {
1258 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1259 }
1260 }
1261
1262 for (i = 0; i < src_count; ++i) {
1263 unsigned idx1 = i, idx2 = i;
1264
1265 if (alpha_span > 1){
1266 idx1 *= alpha_span;
1267 idx2 = idx1 + 1;
1268 }
1269
1270 src_alpha[i] = LLVMBuildShuffleVector(builder,
1271 src_alpha[idx1],
1272 src_alpha[idx2],
1273 LLVMConstVector(shuffles, row_type.length),
1274 "");
1275 }
1276 }
1277 }
1278 }
1279
1280
1281 /**
1282 * Generates the blend function for unswizzled colour buffers
1283 * Also generates the read & write from colour buffer
1284 */
1285 static void
1286 generate_unswizzled_blend(struct gallivm_state *gallivm,
1287 unsigned rt,
1288 struct lp_fragment_shader_variant *variant,
1289 enum pipe_format out_format,
1290 unsigned int num_fs,
1291 struct lp_type fs_type,
1292 LLVMValueRef* fs_mask,
1293 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1294 LLVMValueRef context_ptr,
1295 LLVMValueRef color_ptr,
1296 LLVMValueRef stride,
1297 unsigned partial_mask,
1298 boolean do_branch)
1299 {
1300 const unsigned alpha_channel = 3;
1301 const unsigned block_width = 4;
1302 const unsigned block_height = 4;
1303 const unsigned block_size = block_width * block_height;
1304 const unsigned lp_integer_vector_width = 128;
1305
1306 LLVMBuilderRef builder = gallivm->builder;
1307 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1308 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1309 LLVMValueRef src_alpha[4 * 4];
1310 LLVMValueRef src1_alpha[4 * 4];
1311 LLVMValueRef src_mask[4 * 4];
1312 LLVMValueRef src[4 * 4];
1313 LLVMValueRef src1[4 * 4];
1314 LLVMValueRef dst[4 * 4];
1315 LLVMValueRef blend_color;
1316 LLVMValueRef blend_alpha;
1317 LLVMValueRef i32_zero;
1318 LLVMValueRef check_mask;
1319
1320 struct lp_build_mask_context mask_ctx;
1321 struct lp_type mask_type;
1322 struct lp_type blend_type;
1323 struct lp_type row_type;
1324 struct lp_type dst_type;
1325
1326 unsigned char swizzle[TGSI_NUM_CHANNELS];
1327 unsigned vector_width;
1328 unsigned src_channels = TGSI_NUM_CHANNELS;
1329 unsigned dst_channels;
1330 unsigned dst_count;
1331 unsigned src_count;
1332 unsigned i, j;
1333
1334 const struct util_format_description* out_format_desc = util_format_description(out_format);
1335
1336 unsigned dst_alignment;
1337
1338 bool pad_inline = is_arithmetic_format(out_format_desc);
1339 bool has_alpha = false;
1340 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1341 util_blend_state_is_dual(&variant->key.blend, 0);
1342
1343 mask_type = lp_int32_vec4_type();
1344 mask_type.length = fs_type.length;
1345
1346 /* Compute the alignment of the destination pointer in bytes */
1347 #if 0
1348 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1349 #else
1350 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1351 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1352 #endif
1353 /* Force power-of-two alignment by extracting only the least-significant-bit */
1354 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1355 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1356 dst_alignment = MIN2(dst_alignment, 16);
1357
1358 /* Do not bother executing code when mask is empty.. */
1359 if (do_branch) {
1360 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1361
1362 for (i = 0; i < num_fs; ++i) {
1363 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1364 }
1365
1366 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1367 lp_build_mask_check(&mask_ctx);
1368 }
1369
1370 partial_mask |= !variant->opaque;
1371 i32_zero = lp_build_const_int32(gallivm, 0);
1372
1373 /* Get type from output format */
1374 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1375 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1376
1377 row_type.length = fs_type.length;
1378 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1379
1380 /* Compute correct swizzle and count channels */
1381 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1382 dst_channels = 0;
1383
1384 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1385 /* Ensure channel is used */
1386 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1387 continue;
1388 }
1389
1390 /* Ensure not already written to (happens in case with GL_ALPHA) */
1391 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1392 continue;
1393 }
1394
1395 /* Ensure we havn't already found all channels */
1396 if (dst_channels >= out_format_desc->nr_channels) {
1397 continue;
1398 }
1399
1400 swizzle[out_format_desc->swizzle[i]] = i;
1401 ++dst_channels;
1402
1403 if (i == alpha_channel) {
1404 has_alpha = true;
1405 }
1406 }
1407
1408 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1409 /* the code above can't work for layout_other */
1410 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1411 has_alpha = true;
1412 swizzle[0] = 0;
1413 swizzle[1] = 1;
1414 swizzle[2] = 2;
1415 swizzle[3] = 3;
1416 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1417 }
1418
1419 /* If 3 channels then pad to include alpha for 4 element transpose */
1420 if (dst_channels == 3 && !has_alpha) {
1421 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1422 if (swizzle[i] > TGSI_NUM_CHANNELS)
1423 swizzle[i] = 3;
1424 }
1425 if (out_format_desc->nr_channels == 4) {
1426 dst_channels = 4;
1427 }
1428 }
1429
1430 /*
1431 * Load shader output
1432 */
1433 for (i = 0; i < num_fs; ++i) {
1434 /* Always load alpha for use in blending */
1435 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1436
1437 /* Load each channel */
1438 for (j = 0; j < dst_channels; ++j) {
1439 assert(swizzle[j] < 4);
1440 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1441 }
1442
1443 /* If 3 channels then pad to include alpha for 4 element transpose */
1444 /*
1445 * XXX If we include that here maybe could actually use it instead of
1446 * separate alpha for blending?
1447 */
1448 if (dst_channels == 3 && !has_alpha) {
1449 fs_src[i][3] = alpha;
1450 }
1451
1452 /* We split the row_mask and row_alpha as we want 128bit interleave */
1453 if (fs_type.length == 8) {
1454 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1455 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1456
1457 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1458 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1459 } else {
1460 src_mask[i] = fs_mask[i];
1461 src_alpha[i] = alpha;
1462 }
1463 }
1464 if (dual_source_blend) {
1465 /* same as above except different src/dst, skip masks and comments... */
1466 for (i = 0; i < num_fs; ++i) {
1467 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1468
1469 for (j = 0; j < dst_channels; ++j) {
1470 assert(swizzle[j] < 4);
1471 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1472 }
1473 if (dst_channels == 3 && !has_alpha) {
1474 fs_src1[i][3] = alpha;
1475 }
1476 if (fs_type.length == 8) {
1477 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1478 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1479 } else {
1480 src1_alpha[i] = alpha;
1481 }
1482 }
1483 }
1484
1485 if (util_format_is_pure_integer(out_format)) {
1486 /*
1487 * In this case fs_type was really ints or uints disguised as floats,
1488 * fix that up now.
1489 */
1490 fs_type.floating = 0;
1491 fs_type.sign = dst_type.sign;
1492 for (i = 0; i < num_fs; ++i) {
1493 for (j = 0; j < dst_channels; ++j) {
1494 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1495 lp_build_vec_type(gallivm, fs_type), "");
1496 }
1497 if (dst_channels == 3 && !has_alpha) {
1498 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1499 lp_build_vec_type(gallivm, fs_type), "");
1500 }
1501 }
1502 }
1503
1504 /*
1505 * Pixel twiddle from fragment shader order to memory order
1506 */
1507 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1508 if (dual_source_blend) {
1509 generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
1510 }
1511
1512 src_channels = dst_channels < 3 ? dst_channels : 4;
1513 if (src_count != num_fs * src_channels) {
1514 unsigned ds = src_count / (num_fs * src_channels);
1515 row_type.length /= ds;
1516 fs_type.length = row_type.length;
1517 }
1518
1519 blend_type = row_type;
1520 mask_type.length = 4;
1521
1522 /* Convert src to row_type */
1523 if (dual_source_blend) {
1524 struct lp_type old_row_type = row_type;
1525 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1526 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1527 }
1528 else {
1529 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1530 }
1531
1532 /* If the rows are not an SSE vector, combine them to become SSE size! */
1533 if ((row_type.width * row_type.length) % 128) {
1534 unsigned bits = row_type.width * row_type.length;
1535 unsigned combined;
1536
1537 dst_count = src_count / (vector_width / bits);
1538 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1539 if (dual_source_blend) {
1540 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1541 }
1542
1543 row_type.length *= combined;
1544 src_count /= combined;
1545
1546 bits = row_type.width * row_type.length;
1547 assert(bits == 128 || bits == 256);
1548 }
1549
1550
1551 /*
1552 * Blend Colour conversion
1553 */
1554 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1555 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1556 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1557
1558 /* Convert */
1559 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1560
1561 /* Extract alpha */
1562 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1563
1564 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1565 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1566 if (pad_inline) {
1567 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1568 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1569 } else {
1570 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1571 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1572 }
1573
1574 /*
1575 * Mask conversion
1576 */
1577 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1578
1579 if (src_count < block_height) {
1580 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1581 } else if (src_count > block_height) {
1582 for (i = src_count; i > 0; --i) {
1583 unsigned pixels = block_size / src_count;
1584 unsigned idx = i - 1;
1585
1586 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1587 }
1588 }
1589
1590 assert(mask_type.width == 32);
1591
1592 for (i = 0; i < src_count; ++i) {
1593 unsigned pixels = block_size / src_count;
1594 unsigned pixel_width = row_type.width * dst_channels;
1595
1596 if (pixel_width == 24) {
1597 mask_type.width = 8;
1598 mask_type.length = vector_width / mask_type.width;
1599 } else {
1600 mask_type.length = pixels;
1601 mask_type.width = row_type.width * dst_channels;
1602
1603 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1604
1605 mask_type.length *= dst_channels;
1606 mask_type.width /= dst_channels;
1607 }
1608
1609 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1610 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1611 }
1612
1613 /*
1614 * Alpha conversion
1615 */
1616 if (!has_alpha) {
1617 struct lp_type alpha_type = fs_type;
1618 alpha_type.length = 4;
1619 convert_alpha(gallivm, row_type, alpha_type,
1620 block_size, block_height,
1621 src_count, dst_channels,
1622 pad_inline, src_alpha);
1623 if (dual_source_blend) {
1624 convert_alpha(gallivm, row_type, alpha_type,
1625 block_size, block_height,
1626 src_count, dst_channels,
1627 pad_inline, src1_alpha);
1628 }
1629 }
1630
1631
1632 /*
1633 * Load dst from memory
1634 */
1635 if (src_count < block_height) {
1636 dst_count = block_height;
1637 } else {
1638 dst_count = src_count;
1639 }
1640
1641 dst_type.length *= 16 / dst_count;
1642
1643 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1644 /*
1645 * we need multiple values at once for the conversion, so can as well
1646 * load them vectorized here too instead of concatenating later.
1647 * (Still need concatenation later for 8-wide vectors).
1648 */
1649 dst_count = block_height;
1650 dst_type.length = block_width;
1651 }
1652
1653 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1654 dst, dst_type, dst_count, dst_alignment);
1655
1656
1657 /*
1658 * Convert from dst/output format to src/blending format.
1659 *
1660 * This is necessary as we can only read 1 row from memory at a time,
1661 * so the minimum dst_count will ever be at this point is 4.
1662 *
1663 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1664 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1665 * on all 16 pixels in that single vector at once.
1666 */
1667 if (dst_count > src_count) {
1668 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1669 }
1670
1671 /*
1672 * Blending
1673 */
1674 /* XXX this is broken for RGB8 formats -
1675 * they get expanded from 12 to 16 elements (to include alpha)
1676 * by convert_to_blend_type then reduced to 15 instead of 12
1677 * by convert_from_blend_type (a simple fix though breaks A8...).
1678 * R16G16B16 also crashes differently however something going wrong
1679 * inside llvm handling npot vector sizes seemingly.
1680 * It seems some cleanup could be done here (like skipping conversion/blend
1681 * when not needed).
1682 */
1683 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1684
1685 for (i = 0; i < src_count; ++i) {
1686 dst[i] = lp_build_blend_aos(gallivm,
1687 &variant->key.blend,
1688 out_format,
1689 row_type,
1690 rt,
1691 src[i],
1692 has_alpha ? NULL : src_alpha[i],
1693 src1[i],
1694 has_alpha ? NULL : src1_alpha[i],
1695 dst[i],
1696 partial_mask ? src_mask[i] : NULL,
1697 blend_color,
1698 has_alpha ? NULL : blend_alpha,
1699 swizzle,
1700 pad_inline ? 4 : dst_channels);
1701 }
1702
1703 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1704
1705 /* Split the blend rows back to memory rows */
1706 if (dst_count > src_count) {
1707 row_type.length = dst_type.length * (dst_count / src_count);
1708
1709 if (src_count == 1) {
1710 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1711 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1712
1713 row_type.length /= 2;
1714 src_count *= 2;
1715 }
1716
1717 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1718 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1719 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1720 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1721
1722 row_type.length /= 2;
1723 src_count *= 2;
1724 }
1725
1726
1727 /*
1728 * Store blend result to memory
1729 */
1730 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1731 dst, dst_type, dst_count, dst_alignment);
1732
1733 if (do_branch) {
1734 lp_build_mask_end(&mask_ctx);
1735 }
1736 }
1737
1738
1739 /**
1740 * Generate the runtime callable function for the whole fragment pipeline.
1741 * Note that the function which we generate operates on a block of 16
1742 * pixels at at time. The block contains 2x2 quads. Each quad contains
1743 * 2x2 pixels.
1744 */
1745 static void
1746 generate_fragment(struct llvmpipe_context *lp,
1747 struct lp_fragment_shader *shader,
1748 struct lp_fragment_shader_variant *variant,
1749 unsigned partial_mask)
1750 {
1751 struct gallivm_state *gallivm = variant->gallivm;
1752 const struct lp_fragment_shader_variant_key *key = &variant->key;
1753 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1754 char func_name[256];
1755 struct lp_type fs_type;
1756 struct lp_type blend_type;
1757 LLVMTypeRef fs_elem_type;
1758 LLVMTypeRef blend_vec_type;
1759 LLVMTypeRef arg_types[13];
1760 LLVMTypeRef func_type;
1761 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1762 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1763 LLVMValueRef context_ptr;
1764 LLVMValueRef x;
1765 LLVMValueRef y;
1766 LLVMValueRef a0_ptr;
1767 LLVMValueRef dadx_ptr;
1768 LLVMValueRef dady_ptr;
1769 LLVMValueRef color_ptr_ptr;
1770 LLVMValueRef stride_ptr;
1771 LLVMValueRef depth_ptr;
1772 LLVMValueRef depth_stride;
1773 LLVMValueRef mask_input;
1774 LLVMValueRef thread_data_ptr;
1775 LLVMBasicBlockRef block;
1776 LLVMBuilderRef builder;
1777 struct lp_build_sampler_soa *sampler;
1778 struct lp_build_interp_soa_context interp;
1779 LLVMValueRef fs_mask[16 / 4];
1780 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1781 LLVMValueRef function;
1782 LLVMValueRef facing;
1783 unsigned num_fs;
1784 unsigned i;
1785 unsigned chan;
1786 unsigned cbuf;
1787 boolean cbuf0_write_all;
1788 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1789 util_blend_state_is_dual(&key->blend, 0);
1790
1791 assert(lp_native_vector_width / 32 >= 4);
1792
1793 /* Adjust color input interpolation according to flatshade state:
1794 */
1795 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1796 for (i = 0; i < shader->info.base.num_inputs; i++) {
1797 if (inputs[i].interp == LP_INTERP_COLOR) {
1798 if (key->flatshade)
1799 inputs[i].interp = LP_INTERP_CONSTANT;
1800 else
1801 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1802 }
1803 }
1804
1805 /* check if writes to cbuf[0] are to be copied to all cbufs */
1806 cbuf0_write_all = FALSE;
1807 for (i = 0;i < shader->info.base.num_properties; i++) {
1808 if (shader->info.base.properties[i].name ==
1809 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1810 cbuf0_write_all = TRUE;
1811 break;
1812 }
1813 }
1814
1815 /* TODO: actually pick these based on the fs and color buffer
1816 * characteristics. */
1817
1818 memset(&fs_type, 0, sizeof fs_type);
1819 fs_type.floating = TRUE; /* floating point values */
1820 fs_type.sign = TRUE; /* values are signed */
1821 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1822 fs_type.width = 32; /* 32-bit float */
1823 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1824 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1825
1826 memset(&blend_type, 0, sizeof blend_type);
1827 blend_type.floating = FALSE; /* values are integers */
1828 blend_type.sign = FALSE; /* values are unsigned */
1829 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1830 blend_type.width = 8; /* 8-bit ubyte values */
1831 blend_type.length = 16; /* 16 elements per vector */
1832
1833 /*
1834 * Generate the function prototype. Any change here must be reflected in
1835 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1836 */
1837
1838 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1839
1840 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1841
1842 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1843 shader->no, variant->no, partial_mask ? "partial" : "whole");
1844
1845 arg_types[0] = variant->jit_context_ptr_type; /* context */
1846 arg_types[1] = int32_type; /* x */
1847 arg_types[2] = int32_type; /* y */
1848 arg_types[3] = int32_type; /* facing */
1849 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1850 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1851 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1852 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1853 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1854 arg_types[9] = int32_type; /* mask_input */
1855 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
1856 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1857 arg_types[12] = int32_type; /* depth_stride */
1858
1859 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1860 arg_types, Elements(arg_types), 0);
1861
1862 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1863 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1864
1865 variant->function[partial_mask] = function;
1866
1867 /* XXX: need to propagate noalias down into color param now we are
1868 * passing a pointer-to-pointer?
1869 */
1870 for(i = 0; i < Elements(arg_types); ++i)
1871 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1872 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1873
1874 context_ptr = LLVMGetParam(function, 0);
1875 x = LLVMGetParam(function, 1);
1876 y = LLVMGetParam(function, 2);
1877 facing = LLVMGetParam(function, 3);
1878 a0_ptr = LLVMGetParam(function, 4);
1879 dadx_ptr = LLVMGetParam(function, 5);
1880 dady_ptr = LLVMGetParam(function, 6);
1881 color_ptr_ptr = LLVMGetParam(function, 7);
1882 depth_ptr = LLVMGetParam(function, 8);
1883 mask_input = LLVMGetParam(function, 9);
1884 thread_data_ptr = LLVMGetParam(function, 10);
1885 stride_ptr = LLVMGetParam(function, 11);
1886 depth_stride = LLVMGetParam(function, 12);
1887
1888 lp_build_name(context_ptr, "context");
1889 lp_build_name(x, "x");
1890 lp_build_name(y, "y");
1891 lp_build_name(a0_ptr, "a0");
1892 lp_build_name(dadx_ptr, "dadx");
1893 lp_build_name(dady_ptr, "dady");
1894 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1895 lp_build_name(depth_ptr, "depth");
1896 lp_build_name(thread_data_ptr, "thread_data");
1897 lp_build_name(mask_input, "mask_input");
1898 lp_build_name(stride_ptr, "stride_ptr");
1899 lp_build_name(depth_stride, "depth_stride");
1900
1901 /*
1902 * Function body
1903 */
1904
1905 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1906 builder = gallivm->builder;
1907 assert(builder);
1908 LLVMPositionBuilderAtEnd(builder, block);
1909
1910 /* code generated texture sampling */
1911 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
1912
1913 {
1914 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1915 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1916 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1917 num_loop, "mask_store");
1918 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1919
1920 /*
1921 * The shader input interpolation info is not explicitely baked in the
1922 * shader key, but everything it derives from (TGSI, and flatshade) is
1923 * already included in the shader key.
1924 */
1925 lp_build_interp_soa_init(&interp,
1926 gallivm,
1927 shader->info.base.num_inputs,
1928 inputs,
1929 shader->info.base.pixel_center_integer,
1930 builder, fs_type,
1931 a0_ptr, dadx_ptr, dady_ptr,
1932 x, y);
1933
1934 for (i = 0; i < num_fs; i++) {
1935 LLVMValueRef mask;
1936 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1937 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1938 &indexi, 1, "mask_ptr");
1939
1940 if (partial_mask) {
1941 mask = generate_quad_mask(gallivm, fs_type,
1942 i*fs_type.length/4, mask_input);
1943 }
1944 else {
1945 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1946 }
1947 LLVMBuildStore(builder, mask, mask_ptr);
1948 }
1949
1950 generate_fs_loop(gallivm,
1951 shader, key,
1952 builder,
1953 fs_type,
1954 context_ptr,
1955 num_loop,
1956 &interp,
1957 sampler,
1958 mask_store, /* output */
1959 color_store,
1960 depth_ptr,
1961 depth_stride,
1962 facing,
1963 thread_data_ptr);
1964
1965 for (i = 0; i < num_fs; i++) {
1966 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1967 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1968 &indexi, 1, "");
1969 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1970 /* This is fucked up need to reorganize things */
1971 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1972 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1973 ptr = LLVMBuildGEP(builder,
1974 color_store[cbuf * !cbuf0_write_all][chan],
1975 &indexi, 1, "");
1976 fs_out_color[cbuf][chan][i] = ptr;
1977 }
1978 }
1979 if (dual_source_blend) {
1980 /* only support one dual source blend target hence always use output 1 */
1981 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1982 ptr = LLVMBuildGEP(builder,
1983 color_store[1][chan],
1984 &indexi, 1, "");
1985 fs_out_color[1][chan][i] = ptr;
1986 }
1987 }
1988 }
1989 }
1990
1991 sampler->destroy(sampler);
1992
1993 /* Loop over color outputs / color buffers to do blending.
1994 */
1995 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1996 LLVMValueRef color_ptr;
1997 LLVMValueRef stride;
1998 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
1999
2000 boolean do_branch = ((key->depth.enabled
2001 || key->stencil[0].enabled
2002 || key->alpha.enabled)
2003 && !shader->info.base.uses_kill);
2004
2005 color_ptr = LLVMBuildLoad(builder,
2006 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2007 "");
2008
2009 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2010
2011 stride = LLVMBuildLoad(builder,
2012 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2013 "");
2014
2015 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2016 num_fs, fs_type, fs_mask, fs_out_color,
2017 context_ptr, color_ptr, stride, partial_mask, do_branch);
2018 }
2019
2020 LLVMBuildRetVoid(builder);
2021
2022 gallivm_verify_function(gallivm, function);
2023
2024 variant->nr_instrs += lp_build_count_instructions(function);
2025 }
2026
2027
2028 static void
2029 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2030 {
2031 unsigned i;
2032
2033 debug_printf("fs variant %p:\n", (void *) key);
2034
2035 if (key->flatshade) {
2036 debug_printf("flatshade = 1\n");
2037 }
2038 for (i = 0; i < key->nr_cbufs; ++i) {
2039 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2040 }
2041 if (key->depth.enabled) {
2042 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2043 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2044 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2045 }
2046
2047 for (i = 0; i < 2; ++i) {
2048 if (key->stencil[i].enabled) {
2049 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2050 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2051 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2052 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2053 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2054 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2055 }
2056 }
2057
2058 if (key->alpha.enabled) {
2059 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2060 }
2061
2062 if (key->occlusion_count) {
2063 debug_printf("occlusion_count = 1\n");
2064 }
2065
2066 if (key->blend.logicop_enable) {
2067 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2068 }
2069 else if (key->blend.rt[0].blend_enable) {
2070 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2071 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2072 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2073 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2074 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2075 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2076 }
2077 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2078 for (i = 0; i < key->nr_samplers; ++i) {
2079 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2080 debug_printf("sampler[%u] = \n", i);
2081 debug_printf(" .wrap = %s %s %s\n",
2082 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2083 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2084 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2085 debug_printf(" .min_img_filter = %s\n",
2086 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2087 debug_printf(" .min_mip_filter = %s\n",
2088 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2089 debug_printf(" .mag_img_filter = %s\n",
2090 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2091 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2092 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2093 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2094 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2095 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2096 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2097 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2098 }
2099 for (i = 0; i < key->nr_sampler_views; ++i) {
2100 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2101 debug_printf("texture[%u] = \n", i);
2102 debug_printf(" .format = %s\n",
2103 util_format_name(texture->format));
2104 debug_printf(" .target = %s\n",
2105 util_dump_tex_target(texture->target, TRUE));
2106 debug_printf(" .level_zero_only = %u\n",
2107 texture->level_zero_only);
2108 debug_printf(" .pot = %u %u %u\n",
2109 texture->pot_width,
2110 texture->pot_height,
2111 texture->pot_depth);
2112 }
2113 }
2114
2115
2116 void
2117 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2118 {
2119 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2120 variant->shader->no, variant->no);
2121 tgsi_dump(variant->shader->base.tokens, 0);
2122 dump_fs_variant_key(&variant->key);
2123 debug_printf("variant->opaque = %u\n", variant->opaque);
2124 debug_printf("\n");
2125 }
2126
2127
2128 /**
2129 * Generate a new fragment shader variant from the shader code and
2130 * other state indicated by the key.
2131 */
2132 static struct lp_fragment_shader_variant *
2133 generate_variant(struct llvmpipe_context *lp,
2134 struct lp_fragment_shader *shader,
2135 const struct lp_fragment_shader_variant_key *key)
2136 {
2137 struct lp_fragment_shader_variant *variant;
2138 const struct util_format_description *cbuf0_format_desc;
2139 boolean fullcolormask;
2140
2141 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2142 if(!variant)
2143 return NULL;
2144
2145 variant->gallivm = gallivm_create();
2146 if (!variant->gallivm) {
2147 FREE(variant);
2148 return NULL;
2149 }
2150
2151 variant->shader = shader;
2152 variant->list_item_global.base = variant;
2153 variant->list_item_local.base = variant;
2154 variant->no = shader->variants_created++;
2155
2156 memcpy(&variant->key, key, shader->variant_key_size);
2157
2158 /*
2159 * Determine whether we are touching all channels in the color buffer.
2160 */
2161 fullcolormask = FALSE;
2162 if (key->nr_cbufs == 1) {
2163 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2164 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2165 }
2166
2167 variant->opaque =
2168 !key->blend.logicop_enable &&
2169 !key->blend.rt[0].blend_enable &&
2170 fullcolormask &&
2171 !key->stencil[0].enabled &&
2172 !key->alpha.enabled &&
2173 !key->depth.enabled &&
2174 !shader->info.base.uses_kill
2175 ? TRUE : FALSE;
2176
2177 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2178 lp_debug_fs_variant(variant);
2179 }
2180
2181 lp_jit_init_types(variant);
2182
2183 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2184 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2185
2186 if (variant->jit_function[RAST_WHOLE] == NULL) {
2187 if (variant->opaque) {
2188 /* Specialized shader, which doesn't need to read the color buffer. */
2189 generate_fragment(lp, shader, variant, RAST_WHOLE);
2190 }
2191 }
2192
2193 /*
2194 * Compile everything
2195 */
2196
2197 gallivm_compile_module(variant->gallivm);
2198
2199 if (variant->function[RAST_EDGE_TEST]) {
2200 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2201 gallivm_jit_function(variant->gallivm,
2202 variant->function[RAST_EDGE_TEST]);
2203 }
2204
2205 if (variant->function[RAST_WHOLE]) {
2206 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2207 gallivm_jit_function(variant->gallivm,
2208 variant->function[RAST_WHOLE]);
2209 } else if (!variant->jit_function[RAST_WHOLE]) {
2210 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2211 }
2212
2213 return variant;
2214 }
2215
2216
2217 static void *
2218 llvmpipe_create_fs_state(struct pipe_context *pipe,
2219 const struct pipe_shader_state *templ)
2220 {
2221 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2222 struct lp_fragment_shader *shader;
2223 int nr_samplers;
2224 int nr_sampler_views;
2225 int i;
2226
2227 shader = CALLOC_STRUCT(lp_fragment_shader);
2228 if (!shader)
2229 return NULL;
2230
2231 shader->no = fs_no++;
2232 make_empty_list(&shader->variants);
2233
2234 /* get/save the summary info for this shader */
2235 lp_build_tgsi_info(templ->tokens, &shader->info);
2236
2237 /* we need to keep a local copy of the tokens */
2238 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2239
2240 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2241 if (shader->draw_data == NULL) {
2242 FREE((void *) shader->base.tokens);
2243 FREE(shader);
2244 return NULL;
2245 }
2246
2247 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2248 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2249
2250 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2251 state[MAX2(nr_samplers, nr_sampler_views)]);
2252
2253 for (i = 0; i < shader->info.base.num_inputs; i++) {
2254 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2255 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2256
2257 switch (shader->info.base.input_interpolate[i]) {
2258 case TGSI_INTERPOLATE_CONSTANT:
2259 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2260 break;
2261 case TGSI_INTERPOLATE_LINEAR:
2262 shader->inputs[i].interp = LP_INTERP_LINEAR;
2263 break;
2264 case TGSI_INTERPOLATE_PERSPECTIVE:
2265 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2266 break;
2267 case TGSI_INTERPOLATE_COLOR:
2268 shader->inputs[i].interp = LP_INTERP_COLOR;
2269 break;
2270 default:
2271 assert(0);
2272 break;
2273 }
2274
2275 switch (shader->info.base.input_semantic_name[i]) {
2276 case TGSI_SEMANTIC_FACE:
2277 shader->inputs[i].interp = LP_INTERP_FACING;
2278 break;
2279 case TGSI_SEMANTIC_POSITION:
2280 /* Position was already emitted above
2281 */
2282 shader->inputs[i].interp = LP_INTERP_POSITION;
2283 shader->inputs[i].src_index = 0;
2284 continue;
2285 }
2286
2287 shader->inputs[i].src_index = i+1;
2288 }
2289
2290 if (LP_DEBUG & DEBUG_TGSI) {
2291 unsigned attrib;
2292 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2293 shader->no, (void *) shader);
2294 tgsi_dump(templ->tokens, 0);
2295 debug_printf("usage masks:\n");
2296 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2297 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2298 debug_printf(" IN[%u].%s%s%s%s\n",
2299 attrib,
2300 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2301 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2302 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2303 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2304 }
2305 debug_printf("\n");
2306 }
2307
2308 return shader;
2309 }
2310
2311
2312 static void
2313 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2314 {
2315 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2316
2317 if (llvmpipe->fs == fs)
2318 return;
2319
2320 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2321
2322 draw_bind_fragment_shader(llvmpipe->draw,
2323 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2324
2325 llvmpipe->dirty |= LP_NEW_FS;
2326 }
2327
2328
2329 /**
2330 * Remove shader variant from two lists: the shader's variant list
2331 * and the context's variant list.
2332 */
2333 void
2334 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2335 struct lp_fragment_shader_variant *variant)
2336 {
2337 unsigned i;
2338
2339 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2340 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2341 " #%u v total cached #%u\n",
2342 variant->shader->no,
2343 variant->no,
2344 variant->shader->variants_created,
2345 variant->shader->variants_cached,
2346 lp->nr_fs_variants);
2347 }
2348
2349 /* free all the variant's JIT'd functions */
2350 for (i = 0; i < Elements(variant->function); i++) {
2351 if (variant->function[i]) {
2352 gallivm_free_function(variant->gallivm,
2353 variant->function[i],
2354 variant->jit_function[i]);
2355 }
2356 }
2357
2358 gallivm_destroy(variant->gallivm);
2359
2360 /* remove from shader's list */
2361 remove_from_list(&variant->list_item_local);
2362 variant->shader->variants_cached--;
2363
2364 /* remove from context's list */
2365 remove_from_list(&variant->list_item_global);
2366 lp->nr_fs_variants--;
2367 lp->nr_fs_instrs -= variant->nr_instrs;
2368
2369 FREE(variant);
2370 }
2371
2372
2373 static void
2374 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2375 {
2376 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2377 struct lp_fragment_shader *shader = fs;
2378 struct lp_fs_variant_list_item *li;
2379
2380 assert(fs != llvmpipe->fs);
2381
2382 /*
2383 * XXX: we need to flush the context until we have some sort of reference
2384 * counting in fragment shaders as they may still be binned
2385 * Flushing alone might not sufficient we need to wait on it too.
2386 */
2387 llvmpipe_finish(pipe, __FUNCTION__);
2388
2389 /* Delete all the variants */
2390 li = first_elem(&shader->variants);
2391 while(!at_end(&shader->variants, li)) {
2392 struct lp_fs_variant_list_item *next = next_elem(li);
2393 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2394 li = next;
2395 }
2396
2397 /* Delete draw module's data */
2398 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2399
2400 assert(shader->variants_cached == 0);
2401 FREE((void *) shader->base.tokens);
2402 FREE(shader);
2403 }
2404
2405
2406
2407 static void
2408 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2409 uint shader, uint index,
2410 struct pipe_constant_buffer *cb)
2411 {
2412 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2413 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2414
2415 assert(shader < PIPE_SHADER_TYPES);
2416 assert(index < Elements(llvmpipe->constants[shader]));
2417
2418 /* note: reference counting */
2419 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2420
2421 if (shader == PIPE_SHADER_VERTEX ||
2422 shader == PIPE_SHADER_GEOMETRY) {
2423 /* Pass the constants to the 'draw' module */
2424 const unsigned size = cb ? cb->buffer_size : 0;
2425 const ubyte *data;
2426
2427 if (constants) {
2428 data = (ubyte *) llvmpipe_resource_data(constants);
2429 }
2430 else if (cb && cb->user_buffer) {
2431 data = (ubyte *) cb->user_buffer;
2432 }
2433 else {
2434 data = NULL;
2435 }
2436
2437 if (data)
2438 data += cb->buffer_offset;
2439
2440 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2441 index, data, size);
2442 }
2443
2444 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2445
2446 if (cb && cb->user_buffer) {
2447 pipe_resource_reference(&constants, NULL);
2448 }
2449 }
2450
2451
2452 /**
2453 * Return the blend factor equivalent to a destination alpha of one.
2454 */
2455 static INLINE unsigned
2456 force_dst_alpha_one(unsigned factor)
2457 {
2458 switch(factor) {
2459 case PIPE_BLENDFACTOR_DST_ALPHA:
2460 return PIPE_BLENDFACTOR_ONE;
2461 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2462 return PIPE_BLENDFACTOR_ZERO;
2463 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2464 return PIPE_BLENDFACTOR_ZERO;
2465 }
2466
2467 return factor;
2468 }
2469
2470
2471 /**
2472 * We need to generate several variants of the fragment pipeline to match
2473 * all the combinations of the contributing state atoms.
2474 *
2475 * TODO: there is actually no reason to tie this to context state -- the
2476 * generated code could be cached globally in the screen.
2477 */
2478 static void
2479 make_variant_key(struct llvmpipe_context *lp,
2480 struct lp_fragment_shader *shader,
2481 struct lp_fragment_shader_variant_key *key)
2482 {
2483 unsigned i;
2484
2485 memset(key, 0, shader->variant_key_size);
2486
2487 if (lp->framebuffer.zsbuf) {
2488 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2489 const struct util_format_description *zsbuf_desc =
2490 util_format_description(zsbuf_format);
2491
2492 if (lp->depth_stencil->depth.enabled &&
2493 util_format_has_depth(zsbuf_desc)) {
2494 key->zsbuf_format = zsbuf_format;
2495 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2496 }
2497 if (lp->depth_stencil->stencil[0].enabled &&
2498 util_format_has_stencil(zsbuf_desc)) {
2499 key->zsbuf_format = zsbuf_format;
2500 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2501 }
2502 }
2503
2504 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2505 if (!lp->framebuffer.nr_cbufs ||
2506 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2507 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2508 }
2509 if(key->alpha.enabled)
2510 key->alpha.func = lp->depth_stencil->alpha.func;
2511 /* alpha.ref_value is passed in jit_context */
2512
2513 key->flatshade = lp->rasterizer->flatshade;
2514 if (lp->active_occlusion_query) {
2515 key->occlusion_count = TRUE;
2516 }
2517
2518 if (lp->framebuffer.nr_cbufs) {
2519 memcpy(&key->blend, lp->blend, sizeof key->blend);
2520 }
2521
2522 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2523
2524 if (!key->blend.independent_blend_enable) {
2525 /* we always need independent blend otherwise the fixups below won't work */
2526 for (i = 1; i < key->nr_cbufs; i++) {
2527 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2528 }
2529 key->blend.independent_blend_enable = 1;
2530 }
2531
2532 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2533 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2534 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2535 const struct util_format_description *format_desc;
2536
2537 key->cbuf_format[i] = format;
2538
2539 format_desc = util_format_description(format);
2540 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2541 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2542
2543 /*
2544 * Mask out color channels not present in the color buffer.
2545 */
2546 blend_rt->colormask &= util_format_colormask(format_desc);
2547
2548 /*
2549 * Disable blend for integer formats.
2550 */
2551 if (util_format_is_pure_integer(format)) {
2552 blend_rt->blend_enable = 0;
2553 }
2554
2555 /*
2556 * Our swizzled render tiles always have an alpha channel, but the linear
2557 * render target format often does not, so force here the dst alpha to be
2558 * one.
2559 *
2560 * This is not a mere optimization. Wrong results will be produced if the
2561 * dst alpha is used, the dst format does not have alpha, and the previous
2562 * rendering was not flushed from the swizzled to linear buffer. For
2563 * example, NonPowTwo DCT.
2564 *
2565 * TODO: This should be generalized to all channels for better
2566 * performance, but only alpha causes correctness issues.
2567 *
2568 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2569 */
2570 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2571 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2572 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2573 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2574 blend_rt->alpha_func = blend_rt->rgb_func;
2575 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2576 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2577 }
2578 }
2579
2580 /* This value will be the same for all the variants of a given shader:
2581 */
2582 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2583
2584 for(i = 0; i < key->nr_samplers; ++i) {
2585 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2586 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2587 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2588 }
2589 }
2590
2591 /*
2592 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2593 * are dx10-style? Can't really have mixed opcodes, at least not
2594 * if we want to skip the holes here (without rescanning tgsi).
2595 */
2596 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2597 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2598 for(i = 0; i < key->nr_sampler_views; ++i) {
2599 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2600 lp_sampler_static_texture_state(&key->state[i].texture_state,
2601 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2602 }
2603 }
2604 }
2605 else {
2606 key->nr_sampler_views = key->nr_samplers;
2607 for(i = 0; i < key->nr_sampler_views; ++i) {
2608 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2609 lp_sampler_static_texture_state(&key->state[i].texture_state,
2610 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2611 }
2612 }
2613 }
2614 }
2615
2616
2617
2618 /**
2619 * Update fragment shader state. This is called just prior to drawing
2620 * something when some fragment-related state has changed.
2621 */
2622 void
2623 llvmpipe_update_fs(struct llvmpipe_context *lp)
2624 {
2625 struct lp_fragment_shader *shader = lp->fs;
2626 struct lp_fragment_shader_variant_key key;
2627 struct lp_fragment_shader_variant *variant = NULL;
2628 struct lp_fs_variant_list_item *li;
2629
2630 make_variant_key(lp, shader, &key);
2631
2632 /* Search the variants for one which matches the key */
2633 li = first_elem(&shader->variants);
2634 while(!at_end(&shader->variants, li)) {
2635 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2636 variant = li->base;
2637 break;
2638 }
2639 li = next_elem(li);
2640 }
2641
2642 if (variant) {
2643 /* Move this variant to the head of the list to implement LRU
2644 * deletion of shader's when we have too many.
2645 */
2646 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2647 }
2648 else {
2649 /* variant not found, create it now */
2650 int64_t t0, t1, dt;
2651 unsigned i;
2652 unsigned variants_to_cull;
2653
2654 if (0) {
2655 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2656 lp->nr_fs_variants,
2657 lp->nr_fs_instrs,
2658 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2659 }
2660
2661 /* First, check if we've exceeded the max number of shader variants.
2662 * If so, free 25% of them (the least recently used ones).
2663 */
2664 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2665
2666 if (variants_to_cull ||
2667 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2668 struct pipe_context *pipe = &lp->pipe;
2669
2670 /*
2671 * XXX: we need to flush the context until we have some sort of
2672 * reference counting in fragment shaders as they may still be binned
2673 * Flushing alone might not be sufficient we need to wait on it too.
2674 */
2675 llvmpipe_finish(pipe, __FUNCTION__);
2676
2677 /*
2678 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2679 * number of shader variants (potentially all of them) could be
2680 * pending for destruction on flush.
2681 */
2682
2683 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2684 struct lp_fs_variant_list_item *item;
2685 if (is_empty_list(&lp->fs_variants_list)) {
2686 break;
2687 }
2688 item = last_elem(&lp->fs_variants_list);
2689 assert(item);
2690 assert(item->base);
2691 llvmpipe_remove_shader_variant(lp, item->base);
2692 }
2693 }
2694
2695 /*
2696 * Generate the new variant.
2697 */
2698 t0 = os_time_get();
2699 variant = generate_variant(lp, shader, &key);
2700 t1 = os_time_get();
2701 dt = t1 - t0;
2702 LP_COUNT_ADD(llvm_compile_time, dt);
2703 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2704
2705 llvmpipe_variant_count++;
2706
2707 /* Put the new variant into the list */
2708 if (variant) {
2709 insert_at_head(&shader->variants, &variant->list_item_local);
2710 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2711 lp->nr_fs_variants++;
2712 lp->nr_fs_instrs += variant->nr_instrs;
2713 shader->variants_cached++;
2714 }
2715 }
2716
2717 /* Bind this variant */
2718 lp_setup_set_fs_variant(lp->setup, variant);
2719 }
2720
2721
2722
2723
2724
2725 void
2726 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2727 {
2728 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2729 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2730 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2731
2732 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2733 }