projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: add support for nested / overlapping queries
[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 #include "lp_rast.h"
104
105
106 /** Fragment shader number (for debugging) */
107 static unsigned fs_no = 0;
108
109
110 /**
111 * Expand the relevant bits of mask_input to a n*4-dword mask for the
112 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
113 * quad mask vector to 0 or ~0.
114 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
115 * quad arguments with fs length 8.
116 *
117 * \param first_quad which quad(s) of the quad group to test, in [0,3]
118 * \param mask_input bitwise mask for the whole 4x4 stamp
119 */
120 static LLVMValueRef
121 generate_quad_mask(struct gallivm_state *gallivm,
122 struct lp_type fs_type,
123 unsigned first_quad,
124 LLVMValueRef mask_input) /* int32 */
125 {
126 LLVMBuilderRef builder = gallivm->builder;
127 struct lp_type mask_type;
128 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
129 LLVMValueRef bits[16];
130 LLVMValueRef mask;
131 int shift, i;
132
133 /*
134 * XXX: We'll need a different path for 16 x u8
135 */
136 assert(fs_type.width == 32);
137 assert(fs_type.length <= Elements(bits));
138 mask_type = lp_int_type(fs_type);
139
140 /*
141 * mask_input >>= (quad * 4)
142 */
143 switch (first_quad) {
144 case 0:
145 shift = 0;
146 break;
147 case 1:
148 assert(fs_type.length == 4);
149 shift = 2;
150 break;
151 case 2:
152 shift = 8;
153 break;
154 case 3:
155 assert(fs_type.length == 4);
156 shift = 10;
157 break;
158 default:
159 assert(0);
160 shift = 0;
161 }
162
163 mask_input = LLVMBuildLShr(builder,
164 mask_input,
165 LLVMConstInt(i32t, shift, 0),
166 "");
167
168 /*
169 * mask = { mask_input & (1 << i), for i in [0,3] }
170 */
171 mask = lp_build_broadcast(gallivm,
172 lp_build_vec_type(gallivm, mask_type),
173 mask_input);
174
175 for (i = 0; i < fs_type.length / 4; i++) {
176 unsigned j = 2 * (i % 2) + (i / 2) * 8;
177 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
178 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
179 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
180 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
181 }
182 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
183
184 /*
185 * mask = mask != 0 ? ~0 : 0
186 */
187 mask = lp_build_compare(gallivm,
188 mask_type, PIPE_FUNC_NOTEQUAL,
189 mask,
190 lp_build_const_int_vec(gallivm, mask_type, 0));
191
192 return mask;
193 }
194
195
196 #define EARLY_DEPTH_TEST 0x1
197 #define LATE_DEPTH_TEST 0x2
198 #define EARLY_DEPTH_WRITE 0x4
199 #define LATE_DEPTH_WRITE 0x8
200
201 static int
202 find_output_by_semantic( const struct tgsi_shader_info *info,
203 unsigned semantic,
204 unsigned index )
205 {
206 int i;
207
208 for (i = 0; i < info->num_outputs; i++)
209 if (info->output_semantic_name[i] == semantic &&
210 info->output_semantic_index[i] == index)
211 return i;
212
213 return -1;
214 }
215
216
217 /**
218 * Generate the fragment shader, depth/stencil test, and alpha tests.
219 */
220 static void
221 generate_fs_loop(struct gallivm_state *gallivm,
222 struct lp_fragment_shader *shader,
223 const struct lp_fragment_shader_variant_key *key,
224 LLVMBuilderRef builder,
225 struct lp_type type,
226 LLVMValueRef context_ptr,
227 LLVMValueRef num_loop,
228 struct lp_build_interp_soa_context *interp,
229 struct lp_build_sampler_soa *sampler,
230 LLVMValueRef mask_store,
231 LLVMValueRef (*out_color)[4],
232 LLVMValueRef depth_ptr,
233 LLVMValueRef depth_stride,
234 LLVMValueRef facing,
235 LLVMValueRef thread_data_ptr)
236 {
237 const struct util_format_description *zs_format_desc = NULL;
238 const struct tgsi_token *tokens = shader->base.tokens;
239 LLVMTypeRef vec_type;
240 LLVMValueRef mask_ptr, mask_val;
241 LLVMValueRef consts_ptr;
242 LLVMValueRef z;
243 LLVMValueRef z_value, s_value;
244 LLVMValueRef z_fb, s_fb;
245 LLVMValueRef stencil_refs[2];
246 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
247 struct lp_build_for_loop_state loop_state;
248 struct lp_build_mask_context mask;
249 /*
250 * TODO: figure out if simple_shader optimization is really worthwile to
251 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
252 * code since tests tend to take another codepath than real shaders.
253 */
254 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
255 shader->info.base.num_inputs < 3 &&
256 shader->info.base.num_instructions < 8) && 0;
257 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
258 util_blend_state_is_dual(&key->blend, 0);
259 unsigned attrib;
260 unsigned chan;
261 unsigned cbuf;
262 unsigned depth_mode;
263
264 struct lp_bld_tgsi_system_values system_values;
265
266 memset(&system_values, 0, sizeof(system_values));
267
268 if (key->depth.enabled ||
269 key->stencil[0].enabled) {
270
271 zs_format_desc = util_format_description(key->zsbuf_format);
272 assert(zs_format_desc);
273
274 if (!shader->info.base.writes_z) {
275 if (key->alpha.enabled || shader->info.base.uses_kill) {
276 /* With alpha test and kill, can do the depth test early
277 * and hopefully eliminate some quads. But need to do a
278 * special deferred depth write once the final mask value
279 * is known. This only works though if there's either no
280 * stencil test or the stencil value isn't written.
281 */
282 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
283 (key->stencil[1].enabled &&
284 key->stencil[1].writemask)))
285 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
286 else
287 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
288 }
289 else
290 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
291 }
292 else {
293 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
294 }
295
296 if (!(key->depth.enabled && key->depth.writemask) &&
297 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
298 (key->stencil[1].enabled &&
299 key->stencil[1].writemask))))
300 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
301 }
302 else {
303 depth_mode = 0;
304 }
305
306
307 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
308 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
309
310 vec_type = lp_build_vec_type(gallivm, type);
311
312 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
313
314 lp_build_for_loop_begin(&loop_state, gallivm,
315 lp_build_const_int32(gallivm, 0),
316 LLVMIntULT,
317 num_loop,
318 lp_build_const_int32(gallivm, 1));
319
320 mask_ptr = LLVMBuildGEP(builder, mask_store,
321 &loop_state.counter, 1, "mask_ptr");
322 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
323
324 memset(outputs, 0, sizeof outputs);
325
326 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
327 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
328 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
329 lp_build_vec_type(gallivm,
330 type),
331 num_loop, "color");
332 }
333 }
334 if (dual_source_blend) {
335 assert(key->nr_cbufs <= 1);
336 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
337 out_color[1][chan] = lp_build_array_alloca(gallivm,
338 lp_build_vec_type(gallivm,
339 type),
340 num_loop, "color1");
341 }
342 }
343
344
345 /* 'mask' will control execution based on quad's pixel alive/killed state */
346 lp_build_mask_begin(&mask, gallivm, type, mask_val);
347
348 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
349 lp_build_mask_check(&mask);
350
351 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
352 z = interp->pos[2];
353
354 if (depth_mode & EARLY_DEPTH_TEST) {
355 lp_build_depth_stencil_load_swizzled(gallivm, type,
356 zs_format_desc, key->resource_1d,
357 depth_ptr, depth_stride,
358 &z_fb, &s_fb, loop_state.counter);
359 lp_build_depth_stencil_test(gallivm,
360 &key->depth,
361 key->stencil,
362 type,
363 zs_format_desc,
364 &mask,
365 stencil_refs,
366 z, z_fb, s_fb,
367 facing,
368 &z_value, &s_value,
369 !simple_shader);
370
371 if (depth_mode & EARLY_DEPTH_WRITE) {
372 lp_build_depth_stencil_write_swizzled(gallivm, type,
373 zs_format_desc, key->resource_1d,
374 NULL, NULL, NULL, loop_state.counter,
375 depth_ptr, depth_stride,
376 z_value, s_value);
377 }
378 /*
379 * Note mask check if stencil is enabled must be after ds write not after
380 * stencil test otherwise new stencil values may not get written if all
381 * fragments got killed by depth/stencil test.
382 */
383 if (!simple_shader && key->stencil[0].enabled)
384 lp_build_mask_check(&mask);
385 }
386
387 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
388
389 /* Build the actual shader */
390 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
391 consts_ptr, &system_values,
392 interp->inputs,
393 outputs, sampler, &shader->info.base, NULL);
394
395 /* Alpha test */
396 if (key->alpha.enabled) {
397 int color0 = find_output_by_semantic(&shader->info.base,
398 TGSI_SEMANTIC_COLOR,
399 0);
400
401 if (color0 != -1 && outputs[color0][3]) {
402 const struct util_format_description *cbuf_format_desc;
403 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
404 LLVMValueRef alpha_ref_value;
405
406 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
407 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
408
409 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
410
411 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
412 &mask, alpha, alpha_ref_value,
413 (depth_mode & LATE_DEPTH_TEST) != 0);
414 }
415 }
416
417 /* Late Z test */
418 if (depth_mode & LATE_DEPTH_TEST) {
419 int pos0 = find_output_by_semantic(&shader->info.base,
420 TGSI_SEMANTIC_POSITION,
421 0);
422
423 if (pos0 != -1 && outputs[pos0][2]) {
424 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
425 }
426
427 lp_build_depth_stencil_load_swizzled(gallivm, type,
428 zs_format_desc, key->resource_1d,
429 depth_ptr, depth_stride,
430 &z_fb, &s_fb, loop_state.counter);
431
432 lp_build_depth_stencil_test(gallivm,
433 &key->depth,
434 key->stencil,
435 type,
436 zs_format_desc,
437 &mask,
438 stencil_refs,
439 z, z_fb, s_fb,
440 facing,
441 &z_value, &s_value,
442 !simple_shader);
443 /* Late Z write */
444 if (depth_mode & LATE_DEPTH_WRITE) {
445 lp_build_depth_stencil_write_swizzled(gallivm, type,
446 zs_format_desc, key->resource_1d,
447 NULL, NULL, NULL, loop_state.counter,
448 depth_ptr, depth_stride,
449 z_value, s_value);
450 }
451 }
452 else if ((depth_mode & EARLY_DEPTH_TEST) &&
453 (depth_mode & LATE_DEPTH_WRITE))
454 {
455 /* Need to apply a reduced mask to the depth write. Reload the
456 * depth value, update from zs_value with the new mask value and
457 * write that out.
458 */
459 lp_build_depth_stencil_write_swizzled(gallivm, type,
460 zs_format_desc, key->resource_1d,
461 &mask, z_fb, s_fb, loop_state.counter,
462 depth_ptr, depth_stride,
463 z_value, s_value);
464 }
465
466
467 /* Color write */
468 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
469 {
470 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
471 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
472 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
473 {
474 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
475 if(outputs[attrib][chan]) {
476 /* XXX: just initialize outputs to point at colors[] and
477 * skip this.
478 */
479 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
480 LLVMValueRef color_ptr;
481 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
482 &loop_state.counter, 1, "");
483 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
484 LLVMBuildStore(builder, out, color_ptr);
485 }
486 }
487 }
488 }
489
490 if (key->occlusion_count) {
491 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
492 lp_build_name(counter, "counter");
493 lp_build_occlusion_count(gallivm, type,
494 lp_build_mask_value(&mask), counter);
495 }
496
497 mask_val = lp_build_mask_end(&mask);
498 LLVMBuildStore(builder, mask_val, mask_ptr);
499 lp_build_for_loop_end(&loop_state);
500 }
501
502
503 /**
504 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
505 *
506 * Fragment Shader outputs pixels in small 2x2 blocks
507 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
508 *
509 * However in memory pixels are stored in rows
510 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
511 *
512 * @param type fragment shader type (4x or 8x float)
513 * @param num_fs number of fs_src
514 * @param is_1d whether we're outputting to a 1d resource
515 * @param dst_channels number of output channels
516 * @param fs_src output from fragment shader
517 * @param dst pointer to store result
518 * @param pad_inline is channel padding inline or at end of row
519 * @return the number of dsts
520 */
521 static int
522 generate_fs_twiddle(struct gallivm_state *gallivm,
523 struct lp_type type,
524 unsigned num_fs,
525 unsigned dst_channels,
526 LLVMValueRef fs_src[][4],
527 LLVMValueRef* dst,
528 bool pad_inline)
529 {
530 LLVMValueRef src[16];
531
532 bool swizzle_pad;
533 bool twiddle;
534 bool split;
535
536 unsigned pixels = type.length / 4;
537 unsigned reorder_group;
538 unsigned src_channels;
539 unsigned src_count;
540 unsigned i;
541
542 src_channels = dst_channels < 3 ? dst_channels : 4;
543 src_count = num_fs * src_channels;
544
545 assert(pixels == 2 || pixels == 1);
546 assert(num_fs * src_channels <= Elements(src));
547
548 /*
549 * Transpose from SoA -> AoS
550 */
551 for (i = 0; i < num_fs; ++i) {
552 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
553 }
554
555 /*
556 * Pick transformation options
557 */
558 swizzle_pad = false;
559 twiddle = false;
560 split = false;
561 reorder_group = 0;
562
563 if (dst_channels == 1) {
564 twiddle = true;
565
566 if (pixels == 2) {
567 split = true;
568 }
569 } else if (dst_channels == 2) {
570 if (pixels == 1) {
571 reorder_group = 1;
572 }
573 } else if (dst_channels > 2) {
574 if (pixels == 1) {
575 reorder_group = 2;
576 } else {
577 twiddle = true;
578 }
579
580 if (!pad_inline && dst_channels == 3 && pixels > 1) {
581 swizzle_pad = true;
582 }
583 }
584
585 /*
586 * Split the src in half
587 */
588 if (split) {
589 for (i = num_fs; i > 0; --i) {
590 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
591 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
592 }
593
594 src_count *= 2;
595 type.length = 4;
596 }
597
598 /*
599 * Ensure pixels are in memory order
600 */
601 if (reorder_group) {
602 /* Twiddle pixels by reordering the array, e.g.:
603 *
604 * src_count = 8 -> 0 2 1 3 4 6 5 7
605 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
606 */
607 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
608
609 for (i = 0; i < src_count; ++i) {
610 unsigned group = i / reorder_group;
611 unsigned block = (group / 4) * 4 * reorder_group;
612 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
613 dst[i] = src[j];
614 }
615 } else if (twiddle) {
616 /* Twiddle pixels across elements of array */
617 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
618 } else {
619 /* Do nothing */
620 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
621 }
622
623 /*
624 * Moves any padding between pixels to the end
625 * e.g. RGBXRGBX -> RGBRGBXX
626 */
627 if (swizzle_pad) {
628 unsigned char swizzles[16];
629 unsigned elems = pixels * dst_channels;
630
631 for (i = 0; i < type.length; ++i) {
632 if (i < elems)
633 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
634 else
635 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
636 }
637
638 for (i = 0; i < src_count; ++i) {
639 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
640 }
641 }
642
643 return src_count;
644 }
645
646
647 /**
648 * Load an unswizzled block of pixels from memory
649 */
650 static void
651 load_unswizzled_block(struct gallivm_state *gallivm,
652 LLVMValueRef base_ptr,
653 LLVMValueRef stride,
654 unsigned block_width,
655 unsigned block_height,
656 LLVMValueRef* dst,
657 struct lp_type dst_type,
658 unsigned dst_count,
659 unsigned dst_alignment)
660 {
661 LLVMBuilderRef builder = gallivm->builder;
662 unsigned row_size = dst_count / block_height;
663 unsigned i;
664
665 /* Ensure block exactly fits into dst */
666 assert((block_width * block_height) % dst_count == 0);
667
668 for (i = 0; i < dst_count; ++i) {
669 unsigned x = i % row_size;
670 unsigned y = i / row_size;
671
672 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
673 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
674
675 LLVMValueRef gep[2];
676 LLVMValueRef dst_ptr;
677
678 gep[0] = lp_build_const_int32(gallivm, 0);
679 gep[1] = LLVMBuildAdd(builder, bx, by, "");
680
681 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
682 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
683
684 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
685
686 lp_set_load_alignment(dst[i], dst_alignment);
687 }
688 }
689
690
691 /**
692 * Store an unswizzled block of pixels to memory
693 */
694 static void
695 store_unswizzled_block(struct gallivm_state *gallivm,
696 LLVMValueRef base_ptr,
697 LLVMValueRef stride,
698 unsigned block_width,
699 unsigned block_height,
700 LLVMValueRef* src,
701 struct lp_type src_type,
702 unsigned src_count,
703 unsigned src_alignment)
704 {
705 LLVMBuilderRef builder = gallivm->builder;
706 unsigned row_size = src_count / block_height;
707 unsigned i;
708
709 /* Ensure src exactly fits into block */
710 assert((block_width * block_height) % src_count == 0);
711
712 for (i = 0; i < src_count; ++i) {
713 unsigned x = i % row_size;
714 unsigned y = i / row_size;
715
716 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
717 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
718
719 LLVMValueRef gep[2];
720 LLVMValueRef src_ptr;
721
722 gep[0] = lp_build_const_int32(gallivm, 0);
723 gep[1] = LLVMBuildAdd(builder, bx, by, "");
724
725 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
726 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
727
728 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
729
730 lp_set_store_alignment(src_ptr, src_alignment);
731 }
732 }
733
734
735 /**
736 * Checks if a format description is an arithmetic format
737 *
738 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
739 */
740 static INLINE boolean
741 is_arithmetic_format(const struct util_format_description *format_desc)
742 {
743 boolean arith = false;
744 unsigned i;
745
746 for (i = 0; i < format_desc->nr_channels; ++i) {
747 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
748 arith |= (format_desc->channel[i].size % 8) != 0;
749 }
750
751 return arith;
752 }
753
754
755 /**
756 * Retrieves the type representing the memory layout for a format
757 *
758 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
759 */
760 static INLINE void
761 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
762 struct lp_type* type)
763 {
764 unsigned i;
765 unsigned chan;
766
767 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
768 /* just make this a 32bit uint */
769 type->floating = false;
770 type->fixed = false;
771 type->sign = false;
772 type->norm = false;
773 type->width = 32;
774 type->length = 1;
775 return;
776 }
777
778 for (i = 0; i < 4; i++)
779 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
780 break;
781 chan = i;
782
783 memset(type, 0, sizeof(struct lp_type));
784 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
785 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
786 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
787 type->norm = format_desc->channel[chan].normalized;
788
789 if (is_arithmetic_format(format_desc)) {
790 type->width = 0;
791 type->length = 1;
792
793 for (i = 0; i < format_desc->nr_channels; ++i) {
794 type->width += format_desc->channel[i].size;
795 }
796 } else {
797 type->width = format_desc->channel[chan].size;
798 type->length = format_desc->nr_channels;
799 }
800 }
801
802
803 /**
804 * Retrieves the type for a format which is usable in the blending code.
805 *
806 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
807 */
808 static INLINE void
809 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
810 struct lp_type* type)
811 {
812 unsigned i;
813 unsigned chan;
814
815 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
816 /* always use ordinary floats for blending */
817 type->floating = true;
818 type->fixed = false;
819 type->sign = true;
820 type->norm = false;
821 type->width = 32;
822 type->length = 4;
823 return;
824 }
825
826 for (i = 0; i < 4; i++)
827 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
828 break;
829 chan = i;
830
831 memset(type, 0, sizeof(struct lp_type));
832 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
833 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
834 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
835 type->norm = format_desc->channel[chan].normalized;
836 type->width = format_desc->channel[chan].size;
837 type->length = format_desc->nr_channels;
838
839 for (i = 1; i < format_desc->nr_channels; ++i) {
840 if (format_desc->channel[i].size > type->width)
841 type->width = format_desc->channel[i].size;
842 }
843
844 if (type->floating) {
845 type->width = 32;
846 } else {
847 if (type->width <= 8) {
848 type->width = 8;
849 } else if (type->width <= 16) {
850 type->width = 16;
851 } else {
852 type->width = 32;
853 }
854 }
855
856 if (is_arithmetic_format(format_desc) && type->length == 3) {
857 type->length = 4;
858 }
859 }
860
861
862 /**
863 * Scale a normalized value from src_bits to dst_bits
864 */
865 static INLINE LLVMValueRef
866 scale_bits(struct gallivm_state *gallivm,
867 int src_bits,
868 int dst_bits,
869 LLVMValueRef src,
870 struct lp_type src_type)
871 {
872 LLVMBuilderRef builder = gallivm->builder;
873 LLVMValueRef result = src;
874
875 if (dst_bits < src_bits) {
876 /* Scale down by LShr */
877 result = LLVMBuildLShr(builder,
878 src,
879 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
880 "");
881 } else if (dst_bits > src_bits) {
882 /* Scale up bits */
883 int db = dst_bits - src_bits;
884
885 /* Shift left by difference in bits */
886 result = LLVMBuildShl(builder,
887 src,
888 lp_build_const_int_vec(gallivm, src_type, db),
889 "");
890
891 if (db < src_bits) {
892 /* Enough bits in src to fill the remainder */
893 LLVMValueRef lower = LLVMBuildLShr(builder,
894 src,
895 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
896 "");
897
898 result = LLVMBuildOr(builder, result, lower, "");
899 } else if (db > src_bits) {
900 /* Need to repeatedly copy src bits to fill remainder in dst */
901 unsigned n;
902
903 for (n = src_bits; n < dst_bits; n *= 2) {
904 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
905
906 result = LLVMBuildOr(builder,
907 result,
908 LLVMBuildLShr(builder, result, shuv, ""),
909 "");
910 }
911 }
912 }
913
914 return result;
915 }
916
917
918 /**
919 * Convert from memory format to blending format
920 *
921 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
922 */
923 static void
924 convert_to_blend_type(struct gallivm_state *gallivm,
925 unsigned block_size,
926 const struct util_format_description *src_fmt,
927 struct lp_type src_type,
928 struct lp_type dst_type,
929 LLVMValueRef* src, // and dst
930 unsigned num_srcs)
931 {
932 LLVMValueRef *dst = src;
933 LLVMBuilderRef builder = gallivm->builder;
934 struct lp_type blend_type;
935 struct lp_type mem_type;
936 unsigned i, j, k;
937 unsigned pixels = block_size / num_srcs;
938 bool is_arith;
939
940 /*
941 * full custom path for packed floats - none of the later functions would do
942 * anything useful, and given the lp_type representation they can't be fixed.
943 */
944 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
945 LLVMValueRef tmpsrc[4];
946 /*
947 * This is pretty suboptimal for this case blending in SoA would be much
948 * better, since conversion gets us SoA values so need to convert back.
949 */
950 assert(src_type.width == 32);
951 assert(dst_type.floating);
952 assert(dst_type.width == 32);
953 assert(dst_type.length % 4 == 0);
954 assert(num_srcs % 4 == 0);
955
956 for (i = 0; i < 4; i++) {
957 tmpsrc[i] = src[i];
958 }
959 for (i = 0; i < num_srcs / 4; i++) {
960 LLVMValueRef tmpsoa[4];
961 LLVMValueRef tmps = tmpsrc[i];
962 if (dst_type.length == 8) {
963 LLVMValueRef shuffles[8];
964 unsigned j;
965 /* fetch was 4 values but need 8-wide output values */
966 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
967 /*
968 * for 8-wide aos transpose would give us wrong order not matching
969 * incoming converted fs values and mask. ARGH.
970 */
971 for (j = 0; j < 4; j++) {
972 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
973 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
974 }
975 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
976 LLVMConstVector(shuffles, 8), "");
977 }
978 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
979 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
980 }
981 return;
982 }
983
984 lp_mem_type_from_format_desc(src_fmt, &mem_type);
985 lp_blend_type_from_format_desc(src_fmt, &blend_type);
986
987 /* Is the format arithmetic */
988 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
989 is_arith &= !(mem_type.width == 16 && mem_type.floating);
990
991 /* Pad if necessary */
992 if (!is_arith && src_type.length < dst_type.length) {
993 for (i = 0; i < num_srcs; ++i) {
994 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
995 }
996
997 src_type.length = dst_type.length;
998 }
999
1000 /* Special case for half-floats */
1001 if (mem_type.width == 16 && mem_type.floating) {
1002 assert(blend_type.width == 32 && blend_type.floating);
1003 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1004 is_arith = false;
1005 }
1006
1007 if (!is_arith) {
1008 return;
1009 }
1010
1011 src_type.width = blend_type.width * blend_type.length;
1012 blend_type.length *= pixels;
1013 src_type.length *= pixels / (src_type.length / mem_type.length);
1014
1015 for (i = 0; i < num_srcs; ++i) {
1016 LLVMValueRef chans[4];
1017 LLVMValueRef res = NULL;
1018
1019 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1020
1021 for (j = 0; j < src_fmt->nr_channels; ++j) {
1022 unsigned mask = 0;
1023 unsigned sa = src_fmt->channel[j].shift;
1024 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1025 unsigned from_lsb = j;
1026 #else
1027 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1028 #endif
1029
1030 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1031 mask |= 1 << k;
1032 }
1033
1034 /* Extract bits from source */
1035 chans[j] = LLVMBuildLShr(builder,
1036 dst[i],
1037 lp_build_const_int_vec(gallivm, src_type, sa),
1038 "");
1039
1040 chans[j] = LLVMBuildAnd(builder,
1041 chans[j],
1042 lp_build_const_int_vec(gallivm, src_type, mask),
1043 "");
1044
1045 /* Scale bits */
1046 if (src_type.norm) {
1047 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1048 blend_type.width, chans[j], src_type);
1049 }
1050
1051 /* Insert bits into correct position */
1052 chans[j] = LLVMBuildShl(builder,
1053 chans[j],
1054 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1055 "");
1056
1057 if (j == 0) {
1058 res = chans[j];
1059 } else {
1060 res = LLVMBuildOr(builder, res, chans[j], "");
1061 }
1062 }
1063
1064 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1065 }
1066 }
1067
1068
1069 /**
1070 * Convert from blending format to memory format
1071 *
1072 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1073 */
1074 static void
1075 convert_from_blend_type(struct gallivm_state *gallivm,
1076 unsigned block_size,
1077 const struct util_format_description *src_fmt,
1078 struct lp_type src_type,
1079 struct lp_type dst_type,
1080 LLVMValueRef* src, // and dst
1081 unsigned num_srcs)
1082 {
1083 LLVMValueRef* dst = src;
1084 unsigned i, j, k;
1085 struct lp_type mem_type;
1086 struct lp_type blend_type;
1087 LLVMBuilderRef builder = gallivm->builder;
1088 unsigned pixels = block_size / num_srcs;
1089 bool is_arith;
1090
1091 /*
1092 * full custom path for packed floats - none of the later functions would do
1093 * anything useful, and given the lp_type representation they can't be fixed.
1094 */
1095 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1096 /*
1097 * This is pretty suboptimal for this case blending in SoA would be much
1098 * better - we need to transpose the AoS values back to SoA values for
1099 * conversion/packing.
1100 */
1101 assert(src_type.floating);
1102 assert(src_type.width == 32);
1103 assert(src_type.length % 4 == 0);
1104 assert(dst_type.width == 32);
1105
1106 for (i = 0; i < num_srcs / 4; i++) {
1107 LLVMValueRef tmpsoa[4], tmpdst;
1108 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1109 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1110 if (src_type.length == 8) {
1111 LLVMValueRef tmpaos, shuffles[8];
1112 unsigned j;
1113 /*
1114 * for 8-wide aos transpose has given us wrong order not matching
1115 * output order. HMPF. Also need to split the output values manually.
1116 */
1117 for (j = 0; j < 4; j++) {
1118 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1119 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1120 }
1121 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1122 LLVMConstVector(shuffles, 8), "");
1123 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1124 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1125 }
1126 else {
1127 src[i] = tmpdst;
1128 }
1129 }
1130 return;
1131 }
1132
1133 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1134 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1135
1136 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1137
1138 /* Special case for half-floats */
1139 if (mem_type.width == 16 && mem_type.floating) {
1140 int length = dst_type.length;
1141 assert(blend_type.width == 32 && blend_type.floating);
1142
1143 dst_type.length = src_type.length;
1144
1145 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1146
1147 dst_type.length = length;
1148 is_arith = false;
1149 }
1150
1151 /* Remove any padding */
1152 if (!is_arith && (src_type.length % mem_type.length)) {
1153 src_type.length -= (src_type.length % mem_type.length);
1154
1155 for (i = 0; i < num_srcs; ++i) {
1156 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1157 }
1158 }
1159
1160 /* No bit arithmetic to do */
1161 if (!is_arith) {
1162 return;
1163 }
1164
1165 src_type.length = pixels;
1166 src_type.width = blend_type.length * blend_type.width;
1167 dst_type.length = pixels;
1168
1169 for (i = 0; i < num_srcs; ++i) {
1170 LLVMValueRef chans[4];
1171 LLVMValueRef res = NULL;
1172
1173 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1174
1175 for (j = 0; j < src_fmt->nr_channels; ++j) {
1176 unsigned mask = 0;
1177 unsigned sa = src_fmt->channel[j].shift;
1178 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1179 unsigned from_lsb = j;
1180 #else
1181 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1182 #endif
1183
1184 assert(blend_type.width > src_fmt->channel[j].size);
1185
1186 for (k = 0; k < blend_type.width; ++k) {
1187 mask |= 1 << k;
1188 }
1189
1190 /* Extract bits */
1191 chans[j] = LLVMBuildLShr(builder,
1192 dst[i],
1193 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1194 "");
1195
1196 chans[j] = LLVMBuildAnd(builder,
1197 chans[j],
1198 lp_build_const_int_vec(gallivm, src_type, mask),
1199 "");
1200
1201 /* Scale down bits */
1202 if (src_type.norm) {
1203 chans[j] = scale_bits(gallivm, blend_type.width,
1204 src_fmt->channel[j].size, chans[j], src_type);
1205 }
1206
1207 /* Insert bits */
1208 chans[j] = LLVMBuildShl(builder,
1209 chans[j],
1210 lp_build_const_int_vec(gallivm, src_type, sa),
1211 "");
1212
1213 sa += src_fmt->channel[j].size;
1214
1215 if (j == 0) {
1216 res = chans[j];
1217 } else {
1218 res = LLVMBuildOr(builder, res, chans[j], "");
1219 }
1220 }
1221
1222 assert (dst_type.width != 24);
1223
1224 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1225 }
1226 }
1227
1228
1229 /**
1230 * Convert alpha to same blend type as src
1231 */
1232 static void
1233 convert_alpha(struct gallivm_state *gallivm,
1234 struct lp_type row_type,
1235 struct lp_type alpha_type,
1236 const unsigned block_size,
1237 const unsigned block_height,
1238 const unsigned src_count,
1239 const unsigned dst_channels,
1240 const bool pad_inline,
1241 LLVMValueRef* src_alpha)
1242 {
1243 LLVMBuilderRef builder = gallivm->builder;
1244 unsigned i, j;
1245 unsigned length = row_type.length;
1246 row_type.length = alpha_type.length;
1247
1248 /* Twiddle the alpha to match pixels */
1249 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1250
1251 /*
1252 * TODO this should use single lp_build_conv call for
1253 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1254 */
1255 for (i = 0; i < block_height; ++i) {
1256 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1257 }
1258
1259 alpha_type = row_type;
1260 row_type.length = length;
1261
1262 /* If only one channel we can only need the single alpha value per pixel */
1263 if (src_count == 1 && dst_channels == 1) {
1264
1265 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1266 } else {
1267 /* If there are more srcs than rows then we need to split alpha up */
1268 if (src_count > block_height) {
1269 for (i = src_count; i > 0; --i) {
1270 unsigned pixels = block_size / src_count;
1271 unsigned idx = i - 1;
1272
1273 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1274 (idx * pixels) % 4, pixels);
1275 }
1276 }
1277
1278 /* If there is a src for each pixel broadcast the alpha across whole row */
1279 if (src_count == block_size) {
1280 for (i = 0; i < src_count; ++i) {
1281 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1282 }
1283 } else {
1284 unsigned pixels = block_size / src_count;
1285 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1286 unsigned alpha_span = 1;
1287 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1288
1289 /* Check if we need 2 src_alphas for our shuffles */
1290 if (pixels > alpha_type.length) {
1291 alpha_span = 2;
1292 }
1293
1294 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1295 for (j = 0; j < row_type.length; ++j) {
1296 if (j < pixels * channels) {
1297 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1298 } else {
1299 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1300 }
1301 }
1302
1303 for (i = 0; i < src_count; ++i) {
1304 unsigned idx1 = i, idx2 = i;
1305
1306 if (alpha_span > 1){
1307 idx1 *= alpha_span;
1308 idx2 = idx1 + 1;
1309 }
1310
1311 src_alpha[i] = LLVMBuildShuffleVector(builder,
1312 src_alpha[idx1],
1313 src_alpha[idx2],
1314 LLVMConstVector(shuffles, row_type.length),
1315 "");
1316 }
1317 }
1318 }
1319 }
1320
1321
1322 /**
1323 * Generates the blend function for unswizzled colour buffers
1324 * Also generates the read & write from colour buffer
1325 */
1326 static void
1327 generate_unswizzled_blend(struct gallivm_state *gallivm,
1328 unsigned rt,
1329 struct lp_fragment_shader_variant *variant,
1330 enum pipe_format out_format,
1331 unsigned int num_fs,
1332 struct lp_type fs_type,
1333 LLVMValueRef* fs_mask,
1334 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1335 LLVMValueRef context_ptr,
1336 LLVMValueRef color_ptr,
1337 LLVMValueRef stride,
1338 unsigned partial_mask,
1339 boolean do_branch)
1340 {
1341 const unsigned alpha_channel = 3;
1342 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1343 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1344 const unsigned block_size = block_width * block_height;
1345 const unsigned lp_integer_vector_width = 128;
1346
1347 LLVMBuilderRef builder = gallivm->builder;
1348 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1349 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1350 LLVMValueRef src_alpha[4 * 4];
1351 LLVMValueRef src1_alpha[4 * 4];
1352 LLVMValueRef src_mask[4 * 4];
1353 LLVMValueRef src[4 * 4];
1354 LLVMValueRef src1[4 * 4];
1355 LLVMValueRef dst[4 * 4];
1356 LLVMValueRef blend_color;
1357 LLVMValueRef blend_alpha;
1358 LLVMValueRef i32_zero;
1359 LLVMValueRef check_mask;
1360 LLVMValueRef undef_src_val;
1361
1362 struct lp_build_mask_context mask_ctx;
1363 struct lp_type mask_type;
1364 struct lp_type blend_type;
1365 struct lp_type row_type;
1366 struct lp_type dst_type;
1367
1368 unsigned char swizzle[TGSI_NUM_CHANNELS];
1369 unsigned vector_width;
1370 unsigned src_channels = TGSI_NUM_CHANNELS;
1371 unsigned dst_channels;
1372 unsigned dst_count;
1373 unsigned src_count;
1374 unsigned i, j;
1375
1376 const struct util_format_description* out_format_desc = util_format_description(out_format);
1377
1378 unsigned dst_alignment;
1379
1380 bool pad_inline = is_arithmetic_format(out_format_desc);
1381 bool has_alpha = false;
1382 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1383 util_blend_state_is_dual(&variant->key.blend, 0);
1384
1385 const boolean is_1d = variant->key.resource_1d;
1386 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1387
1388 mask_type = lp_int32_vec4_type();
1389 mask_type.length = fs_type.length;
1390
1391 for (i = num_fs; i < num_fullblock_fs; i++) {
1392 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1393 }
1394
1395 /* Do not bother executing code when mask is empty.. */
1396 if (do_branch) {
1397 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1398
1399 for (i = 0; i < num_fullblock_fs; ++i) {
1400 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1401 }
1402
1403 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1404 lp_build_mask_check(&mask_ctx);
1405 }
1406
1407 partial_mask |= !variant->opaque;
1408 i32_zero = lp_build_const_int32(gallivm, 0);
1409
1410 #if HAVE_LLVM < 0x0302
1411 /*
1412 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1413 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1414 */
1415 undef_src_val = lp_build_zero(gallivm, fs_type);
1416 #else
1417 undef_src_val = lp_build_undef(gallivm, fs_type);
1418 #endif
1419
1420
1421 /* Get type from output format */
1422 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1423 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1424
1425 row_type.length = fs_type.length;
1426 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1427
1428 /* Compute correct swizzle and count channels */
1429 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1430 dst_channels = 0;
1431
1432 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1433 /* Ensure channel is used */
1434 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1435 continue;
1436 }
1437
1438 /* Ensure not already written to (happens in case with GL_ALPHA) */
1439 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1440 continue;
1441 }
1442
1443 /* Ensure we havn't already found all channels */
1444 if (dst_channels >= out_format_desc->nr_channels) {
1445 continue;
1446 }
1447
1448 swizzle[out_format_desc->swizzle[i]] = i;
1449 ++dst_channels;
1450
1451 if (i == alpha_channel) {
1452 has_alpha = true;
1453 }
1454 }
1455
1456 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1457 /* the code above can't work for layout_other */
1458 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1459 has_alpha = true;
1460 swizzle[0] = 0;
1461 swizzle[1] = 1;
1462 swizzle[2] = 2;
1463 swizzle[3] = 3;
1464 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1465 }
1466
1467 /* If 3 channels then pad to include alpha for 4 element transpose */
1468 if (dst_channels == 3 && !has_alpha) {
1469 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1470 if (swizzle[i] > TGSI_NUM_CHANNELS)
1471 swizzle[i] = 3;
1472 }
1473 if (out_format_desc->nr_channels == 4) {
1474 dst_channels = 4;
1475 }
1476 }
1477
1478 /*
1479 * Load shader output
1480 */
1481 for (i = 0; i < num_fullblock_fs; ++i) {
1482 /* Always load alpha for use in blending */
1483 LLVMValueRef alpha;
1484 if (i < num_fs) {
1485 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1486 }
1487 else {
1488 alpha = undef_src_val;
1489 }
1490
1491 /* Load each channel */
1492 for (j = 0; j < dst_channels; ++j) {
1493 assert(swizzle[j] < 4);
1494 if (i < num_fs) {
1495 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1496 }
1497 else {
1498 fs_src[i][j] = undef_src_val;
1499 }
1500 }
1501
1502 /* If 3 channels then pad to include alpha for 4 element transpose */
1503 /*
1504 * XXX If we include that here maybe could actually use it instead of
1505 * separate alpha for blending?
1506 */
1507 if (dst_channels == 3 && !has_alpha) {
1508 fs_src[i][3] = alpha;
1509 }
1510
1511 /* We split the row_mask and row_alpha as we want 128bit interleave */
1512 if (fs_type.length == 8) {
1513 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1514 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1515
1516 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1517 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1518 } else {
1519 src_mask[i] = fs_mask[i];
1520 src_alpha[i] = alpha;
1521 }
1522 }
1523 if (dual_source_blend) {
1524 /* same as above except different src/dst, skip masks and comments... */
1525 for (i = 0; i < num_fullblock_fs; ++i) {
1526 LLVMValueRef alpha;
1527 if (i < num_fs) {
1528 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1529 }
1530 else {
1531 alpha = undef_src_val;
1532 }
1533
1534 for (j = 0; j < dst_channels; ++j) {
1535 assert(swizzle[j] < 4);
1536 if (i < num_fs) {
1537 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1538 }
1539 else {
1540 fs_src1[i][j] = undef_src_val;
1541 }
1542 }
1543 if (dst_channels == 3 && !has_alpha) {
1544 fs_src1[i][3] = alpha;
1545 }
1546 if (fs_type.length == 8) {
1547 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1548 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1549 } else {
1550 src1_alpha[i] = alpha;
1551 }
1552 }
1553 }
1554
1555 if (util_format_is_pure_integer(out_format)) {
1556 /*
1557 * In this case fs_type was really ints or uints disguised as floats,
1558 * fix that up now.
1559 */
1560 fs_type.floating = 0;
1561 fs_type.sign = dst_type.sign;
1562 for (i = 0; i < num_fullblock_fs; ++i) {
1563 for (j = 0; j < dst_channels; ++j) {
1564 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1565 lp_build_vec_type(gallivm, fs_type), "");
1566 }
1567 if (dst_channels == 3 && !has_alpha) {
1568 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1569 lp_build_vec_type(gallivm, fs_type), "");
1570 }
1571 }
1572 }
1573
1574 /*
1575 * Pixel twiddle from fragment shader order to memory order
1576 */
1577 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
1578 dst_channels, fs_src, src, pad_inline);
1579 if (dual_source_blend) {
1580 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
1581 fs_src1, src1, pad_inline);
1582 }
1583
1584 src_channels = dst_channels < 3 ? dst_channels : 4;
1585 if (src_count != num_fullblock_fs * src_channels) {
1586 unsigned ds = src_count / (num_fullblock_fs * src_channels);
1587 row_type.length /= ds;
1588 fs_type.length = row_type.length;
1589 }
1590
1591 blend_type = row_type;
1592 mask_type.length = 4;
1593
1594 /* Convert src to row_type */
1595 if (dual_source_blend) {
1596 struct lp_type old_row_type = row_type;
1597 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1598 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1599 }
1600 else {
1601 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1602 }
1603
1604 /* If the rows are not an SSE vector, combine them to become SSE size! */
1605 if ((row_type.width * row_type.length) % 128) {
1606 unsigned bits = row_type.width * row_type.length;
1607 unsigned combined;
1608
1609 assert(src_count >= (vector_width / bits));
1610
1611 dst_count = src_count / (vector_width / bits);
1612
1613 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1614 if (dual_source_blend) {
1615 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1616 }
1617
1618 row_type.length *= combined;
1619 src_count /= combined;
1620
1621 bits = row_type.width * row_type.length;
1622 assert(bits == 128 || bits == 256);
1623 }
1624
1625
1626 /*
1627 * Blend Colour conversion
1628 */
1629 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1630 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1631 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1632
1633 /* Convert */
1634 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1635
1636 /* Extract alpha */
1637 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1638
1639 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1640 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1641 if (pad_inline) {
1642 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1643 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1644 } else {
1645 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1646 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1647 }
1648
1649 /*
1650 * Mask conversion
1651 */
1652 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
1653
1654 if (src_count < block_height) {
1655 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1656 } else if (src_count > block_height) {
1657 for (i = src_count; i > 0; --i) {
1658 unsigned pixels = block_size / src_count;
1659 unsigned idx = i - 1;
1660
1661 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
1662 (idx * pixels) % 4, pixels);
1663 }
1664 }
1665
1666 assert(mask_type.width == 32);
1667
1668 for (i = 0; i < src_count; ++i) {
1669 unsigned pixels = block_size / src_count;
1670 unsigned pixel_width = row_type.width * dst_channels;
1671
1672 if (pixel_width == 24) {
1673 mask_type.width = 8;
1674 mask_type.length = vector_width / mask_type.width;
1675 } else {
1676 mask_type.length = pixels;
1677 mask_type.width = row_type.width * dst_channels;
1678
1679 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1680
1681 mask_type.length *= dst_channels;
1682 mask_type.width /= dst_channels;
1683 }
1684
1685 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1686 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1687 }
1688
1689 /*
1690 * Alpha conversion
1691 */
1692 if (!has_alpha) {
1693 struct lp_type alpha_type = fs_type;
1694 alpha_type.length = 4;
1695 convert_alpha(gallivm, row_type, alpha_type,
1696 block_size, block_height,
1697 src_count, dst_channels,
1698 pad_inline, src_alpha);
1699 if (dual_source_blend) {
1700 convert_alpha(gallivm, row_type, alpha_type,
1701 block_size, block_height,
1702 src_count, dst_channels,
1703 pad_inline, src1_alpha);
1704 }
1705 }
1706
1707
1708 /*
1709 * Load dst from memory
1710 */
1711 if (src_count < block_height) {
1712 dst_count = block_height;
1713 } else {
1714 dst_count = src_count;
1715 }
1716
1717 dst_type.length *= block_size / dst_count;
1718
1719 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1720 /*
1721 * we need multiple values at once for the conversion, so can as well
1722 * load them vectorized here too instead of concatenating later.
1723 * (Still need concatenation later for 8-wide vectors).
1724 */
1725 dst_count = block_height;
1726 dst_type.length = block_width;
1727 }
1728
1729 /*
1730 * Compute the alignment of the destination pointer in bytes
1731 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1732 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1733 * 1d tex but can't distinguish here) so need to stick with per-pixel
1734 * alignment in this case.
1735 */
1736 if (is_1d) {
1737 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1738 }
1739 else {
1740 dst_alignment = dst_type.length * dst_type.width / 8;
1741 }
1742 /* Force power-of-two alignment by extracting only the least-significant-bit */
1743 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1744 /*
1745 * Resource base and stride pointers are aligned to 16 bytes, so that's
1746 * the maximum alignment we can guarantee
1747 */
1748 dst_alignment = MIN2(16, dst_alignment);
1749
1750 if (is_1d) {
1751 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
1752 dst, dst_type, dst_count / 4, dst_alignment);
1753 for (i = dst_count / 4; i < dst_count; i++) {
1754 dst[i] = lp_build_undef(gallivm, dst_type);
1755 }
1756
1757 }
1758 else {
1759 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1760 dst, dst_type, dst_count, dst_alignment);
1761 }
1762
1763
1764 /*
1765 * Convert from dst/output format to src/blending format.
1766 *
1767 * This is necessary as we can only read 1 row from memory at a time,
1768 * so the minimum dst_count will ever be at this point is 4.
1769 *
1770 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1771 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1772 * on all 16 pixels in that single vector at once.
1773 */
1774 if (dst_count > src_count) {
1775 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1776 }
1777
1778 /*
1779 * Blending
1780 */
1781 /* XXX this is broken for RGB8 formats -
1782 * they get expanded from 12 to 16 elements (to include alpha)
1783 * by convert_to_blend_type then reduced to 15 instead of 12
1784 * by convert_from_blend_type (a simple fix though breaks A8...).
1785 * R16G16B16 also crashes differently however something going wrong
1786 * inside llvm handling npot vector sizes seemingly.
1787 * It seems some cleanup could be done here (like skipping conversion/blend
1788 * when not needed).
1789 */
1790 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
1791
1792 for (i = 0; i < src_count; ++i) {
1793 dst[i] = lp_build_blend_aos(gallivm,
1794 &variant->key.blend,
1795 out_format,
1796 row_type,
1797 rt,
1798 src[i],
1799 has_alpha ? NULL : src_alpha[i],
1800 src1[i],
1801 has_alpha ? NULL : src1_alpha[i],
1802 dst[i],
1803 partial_mask ? src_mask[i] : NULL,
1804 blend_color,
1805 has_alpha ? NULL : blend_alpha,
1806 swizzle,
1807 pad_inline ? 4 : dst_channels);
1808 }
1809
1810 convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
1811
1812 /* Split the blend rows back to memory rows */
1813 if (dst_count > src_count) {
1814 row_type.length = dst_type.length * (dst_count / src_count);
1815
1816 if (src_count == 1) {
1817 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1818 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1819
1820 row_type.length /= 2;
1821 src_count *= 2;
1822 }
1823
1824 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1825 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1826 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1827 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1828
1829 row_type.length /= 2;
1830 src_count *= 2;
1831 }
1832
1833 /*
1834 * Store blend result to memory
1835 */
1836 if (is_1d) {
1837 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
1838 dst, dst_type, dst_count / 4, dst_alignment);
1839 }
1840 else {
1841 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1842 dst, dst_type, dst_count, dst_alignment);
1843 }
1844
1845 if (do_branch) {
1846 lp_build_mask_end(&mask_ctx);
1847 }
1848 }
1849
1850
1851 /**
1852 * Generate the runtime callable function for the whole fragment pipeline.
1853 * Note that the function which we generate operates on a block of 16
1854 * pixels at at time. The block contains 2x2 quads. Each quad contains
1855 * 2x2 pixels.
1856 */
1857 static void
1858 generate_fragment(struct llvmpipe_context *lp,
1859 struct lp_fragment_shader *shader,
1860 struct lp_fragment_shader_variant *variant,
1861 unsigned partial_mask)
1862 {
1863 struct gallivm_state *gallivm = variant->gallivm;
1864 const struct lp_fragment_shader_variant_key *key = &variant->key;
1865 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1866 char func_name[256];
1867 struct lp_type fs_type;
1868 struct lp_type blend_type;
1869 LLVMTypeRef fs_elem_type;
1870 LLVMTypeRef blend_vec_type;
1871 LLVMTypeRef arg_types[13];
1872 LLVMTypeRef func_type;
1873 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1874 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1875 LLVMValueRef context_ptr;
1876 LLVMValueRef x;
1877 LLVMValueRef y;
1878 LLVMValueRef a0_ptr;
1879 LLVMValueRef dadx_ptr;
1880 LLVMValueRef dady_ptr;
1881 LLVMValueRef color_ptr_ptr;
1882 LLVMValueRef stride_ptr;
1883 LLVMValueRef depth_ptr;
1884 LLVMValueRef depth_stride;
1885 LLVMValueRef mask_input;
1886 LLVMValueRef thread_data_ptr;
1887 LLVMBasicBlockRef block;
1888 LLVMBuilderRef builder;
1889 struct lp_build_sampler_soa *sampler;
1890 struct lp_build_interp_soa_context interp;
1891 LLVMValueRef fs_mask[16 / 4];
1892 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1893 LLVMValueRef function;
1894 LLVMValueRef facing;
1895 unsigned num_fs;
1896 unsigned i;
1897 unsigned chan;
1898 unsigned cbuf;
1899 boolean cbuf0_write_all;
1900 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1901 util_blend_state_is_dual(&key->blend, 0);
1902
1903 assert(lp_native_vector_width / 32 >= 4);
1904
1905 /* Adjust color input interpolation according to flatshade state:
1906 */
1907 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1908 for (i = 0; i < shader->info.base.num_inputs; i++) {
1909 if (inputs[i].interp == LP_INTERP_COLOR) {
1910 if (key->flatshade)
1911 inputs[i].interp = LP_INTERP_CONSTANT;
1912 else
1913 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1914 }
1915 }
1916
1917 /* check if writes to cbuf[0] are to be copied to all cbufs */
1918 cbuf0_write_all = FALSE;
1919 for (i = 0;i < shader->info.base.num_properties; i++) {
1920 if (shader->info.base.properties[i].name ==
1921 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1922 cbuf0_write_all = TRUE;
1923 break;
1924 }
1925 }
1926
1927 /* TODO: actually pick these based on the fs and color buffer
1928 * characteristics. */
1929
1930 memset(&fs_type, 0, sizeof fs_type);
1931 fs_type.floating = TRUE; /* floating point values */
1932 fs_type.sign = TRUE; /* values are signed */
1933 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1934 fs_type.width = 32; /* 32-bit float */
1935 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1936
1937 memset(&blend_type, 0, sizeof blend_type);
1938 blend_type.floating = FALSE; /* values are integers */
1939 blend_type.sign = FALSE; /* values are unsigned */
1940 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1941 blend_type.width = 8; /* 8-bit ubyte values */
1942 blend_type.length = 16; /* 16 elements per vector */
1943
1944 /*
1945 * Generate the function prototype. Any change here must be reflected in
1946 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1947 */
1948
1949 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1950
1951 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1952
1953 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1954 shader->no, variant->no, partial_mask ? "partial" : "whole");
1955
1956 arg_types[0] = variant->jit_context_ptr_type; /* context */
1957 arg_types[1] = int32_type; /* x */
1958 arg_types[2] = int32_type; /* y */
1959 arg_types[3] = int32_type; /* facing */
1960 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1961 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1962 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1963 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1964 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1965 arg_types[9] = int32_type; /* mask_input */
1966 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
1967 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1968 arg_types[12] = int32_type; /* depth_stride */
1969
1970 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1971 arg_types, Elements(arg_types), 0);
1972
1973 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1974 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1975
1976 variant->function[partial_mask] = function;
1977
1978 /* XXX: need to propagate noalias down into color param now we are
1979 * passing a pointer-to-pointer?
1980 */
1981 for(i = 0; i < Elements(arg_types); ++i)
1982 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1983 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1984
1985 context_ptr = LLVMGetParam(function, 0);
1986 x = LLVMGetParam(function, 1);
1987 y = LLVMGetParam(function, 2);
1988 facing = LLVMGetParam(function, 3);
1989 a0_ptr = LLVMGetParam(function, 4);
1990 dadx_ptr = LLVMGetParam(function, 5);
1991 dady_ptr = LLVMGetParam(function, 6);
1992 color_ptr_ptr = LLVMGetParam(function, 7);
1993 depth_ptr = LLVMGetParam(function, 8);
1994 mask_input = LLVMGetParam(function, 9);
1995 thread_data_ptr = LLVMGetParam(function, 10);
1996 stride_ptr = LLVMGetParam(function, 11);
1997 depth_stride = LLVMGetParam(function, 12);
1998
1999 lp_build_name(context_ptr, "context");
2000 lp_build_name(x, "x");
2001 lp_build_name(y, "y");
2002 lp_build_name(a0_ptr, "a0");
2003 lp_build_name(dadx_ptr, "dadx");
2004 lp_build_name(dady_ptr, "dady");
2005 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2006 lp_build_name(depth_ptr, "depth");
2007 lp_build_name(thread_data_ptr, "thread_data");
2008 lp_build_name(mask_input, "mask_input");
2009 lp_build_name(stride_ptr, "stride_ptr");
2010 lp_build_name(depth_stride, "depth_stride");
2011
2012 /*
2013 * Function body
2014 */
2015
2016 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2017 builder = gallivm->builder;
2018 assert(builder);
2019 LLVMPositionBuilderAtEnd(builder, block);
2020
2021 /* code generated texture sampling */
2022 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
2023
2024 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2025 /* for 1d resources only run "upper half" of stamp */
2026 if (key->resource_1d)
2027 num_fs /= 2;
2028
2029 {
2030 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2031 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2032 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2033 num_loop, "mask_store");
2034 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2035
2036 /*
2037 * The shader input interpolation info is not explicitely baked in the
2038 * shader key, but everything it derives from (TGSI, and flatshade) is
2039 * already included in the shader key.
2040 */
2041 lp_build_interp_soa_init(&interp,
2042 gallivm,
2043 shader->info.base.num_inputs,
2044 inputs,
2045 shader->info.base.pixel_center_integer,
2046 builder, fs_type,
2047 a0_ptr, dadx_ptr, dady_ptr,
2048 x, y);
2049
2050 for (i = 0; i < num_fs; i++) {
2051 LLVMValueRef mask;
2052 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2053 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2054 &indexi, 1, "mask_ptr");
2055
2056 if (partial_mask) {
2057 mask = generate_quad_mask(gallivm, fs_type,
2058 i*fs_type.length/4, mask_input);
2059 }
2060 else {
2061 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2062 }
2063 LLVMBuildStore(builder, mask, mask_ptr);
2064 }
2065
2066 generate_fs_loop(gallivm,
2067 shader, key,
2068 builder,
2069 fs_type,
2070 context_ptr,
2071 num_loop,
2072 &interp,
2073 sampler,
2074 mask_store, /* output */
2075 color_store,
2076 depth_ptr,
2077 depth_stride,
2078 facing,
2079 thread_data_ptr);
2080
2081 for (i = 0; i < num_fs; i++) {
2082 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2083 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2084 &indexi, 1, "");
2085 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2086 /* This is fucked up need to reorganize things */
2087 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2088 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2089 ptr = LLVMBuildGEP(builder,
2090 color_store[cbuf * !cbuf0_write_all][chan],
2091 &indexi, 1, "");
2092 fs_out_color[cbuf][chan][i] = ptr;
2093 }
2094 }
2095 if (dual_source_blend) {
2096 /* only support one dual source blend target hence always use output 1 */
2097 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2098 ptr = LLVMBuildGEP(builder,
2099 color_store[1][chan],
2100 &indexi, 1, "");
2101 fs_out_color[1][chan][i] = ptr;
2102 }
2103 }
2104 }
2105 }
2106
2107 sampler->destroy(sampler);
2108
2109 /* Loop over color outputs / color buffers to do blending.
2110 */
2111 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2112 LLVMValueRef color_ptr;
2113 LLVMValueRef stride;
2114 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2115
2116 boolean do_branch = ((key->depth.enabled
2117 || key->stencil[0].enabled
2118 || key->alpha.enabled)
2119 && !shader->info.base.uses_kill);
2120
2121 color_ptr = LLVMBuildLoad(builder,
2122 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2123 "");
2124
2125 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2126
2127 stride = LLVMBuildLoad(builder,
2128 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2129 "");
2130
2131 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2132 num_fs, fs_type, fs_mask, fs_out_color,
2133 context_ptr, color_ptr, stride, partial_mask, do_branch);
2134 }
2135
2136 LLVMBuildRetVoid(builder);
2137
2138 gallivm_verify_function(gallivm, function);
2139
2140 variant->nr_instrs += lp_build_count_instructions(function);
2141 }
2142
2143
2144 static void
2145 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2146 {
2147 unsigned i;
2148
2149 debug_printf("fs variant %p:\n", (void *) key);
2150
2151 if (key->flatshade) {
2152 debug_printf("flatshade = 1\n");
2153 }
2154 for (i = 0; i < key->nr_cbufs; ++i) {
2155 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2156 }
2157 if (key->depth.enabled) {
2158 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2159 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2160 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2161 }
2162
2163 for (i = 0; i < 2; ++i) {
2164 if (key->stencil[i].enabled) {
2165 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2166 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2167 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2168 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2169 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2170 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2171 }
2172 }
2173
2174 if (key->alpha.enabled) {
2175 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2176 }
2177
2178 if (key->occlusion_count) {
2179 debug_printf("occlusion_count = 1\n");
2180 }
2181
2182 if (key->blend.logicop_enable) {
2183 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2184 }
2185 else if (key->blend.rt[0].blend_enable) {
2186 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2187 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2188 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2189 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2190 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2191 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2192 }
2193 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2194 for (i = 0; i < key->nr_samplers; ++i) {
2195 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2196 debug_printf("sampler[%u] = \n", i);
2197 debug_printf(" .wrap = %s %s %s\n",
2198 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2199 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2200 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2201 debug_printf(" .min_img_filter = %s\n",
2202 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2203 debug_printf(" .min_mip_filter = %s\n",
2204 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2205 debug_printf(" .mag_img_filter = %s\n",
2206 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2207 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2208 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2209 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2210 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2211 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2212 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2213 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2214 }
2215 for (i = 0; i < key->nr_sampler_views; ++i) {
2216 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2217 debug_printf("texture[%u] = \n", i);
2218 debug_printf(" .format = %s\n",
2219 util_format_name(texture->format));
2220 debug_printf(" .target = %s\n",
2221 util_dump_tex_target(texture->target, TRUE));
2222 debug_printf(" .level_zero_only = %u\n",
2223 texture->level_zero_only);
2224 debug_printf(" .pot = %u %u %u\n",
2225 texture->pot_width,
2226 texture->pot_height,
2227 texture->pot_depth);
2228 }
2229 }
2230
2231
2232 void
2233 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2234 {
2235 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2236 variant->shader->no, variant->no);
2237 tgsi_dump(variant->shader->base.tokens, 0);
2238 dump_fs_variant_key(&variant->key);
2239 debug_printf("variant->opaque = %u\n", variant->opaque);
2240 debug_printf("\n");
2241 }
2242
2243
2244 /**
2245 * Generate a new fragment shader variant from the shader code and
2246 * other state indicated by the key.
2247 */
2248 static struct lp_fragment_shader_variant *
2249 generate_variant(struct llvmpipe_context *lp,
2250 struct lp_fragment_shader *shader,
2251 const struct lp_fragment_shader_variant_key *key)
2252 {
2253 struct lp_fragment_shader_variant *variant;
2254 const struct util_format_description *cbuf0_format_desc;
2255 boolean fullcolormask;
2256
2257 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2258 if(!variant)
2259 return NULL;
2260
2261 variant->gallivm = gallivm_create();
2262 if (!variant->gallivm) {
2263 FREE(variant);
2264 return NULL;
2265 }
2266
2267 variant->shader = shader;
2268 variant->list_item_global.base = variant;
2269 variant->list_item_local.base = variant;
2270 variant->no = shader->variants_created++;
2271
2272 memcpy(&variant->key, key, shader->variant_key_size);
2273
2274 /*
2275 * Determine whether we are touching all channels in the color buffer.
2276 */
2277 fullcolormask = FALSE;
2278 if (key->nr_cbufs == 1) {
2279 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2280 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2281 }
2282
2283 variant->opaque =
2284 !key->blend.logicop_enable &&
2285 !key->blend.rt[0].blend_enable &&
2286 fullcolormask &&
2287 !key->stencil[0].enabled &&
2288 !key->alpha.enabled &&
2289 !key->depth.enabled &&
2290 !shader->info.base.uses_kill
2291 ? TRUE : FALSE;
2292
2293 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2294 lp_debug_fs_variant(variant);
2295 }
2296
2297 lp_jit_init_types(variant);
2298
2299 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2300 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2301
2302 if (variant->jit_function[RAST_WHOLE] == NULL) {
2303 if (variant->opaque) {
2304 /* Specialized shader, which doesn't need to read the color buffer. */
2305 generate_fragment(lp, shader, variant, RAST_WHOLE);
2306 }
2307 }
2308
2309 /*
2310 * Compile everything
2311 */
2312
2313 gallivm_compile_module(variant->gallivm);
2314
2315 if (variant->function[RAST_EDGE_TEST]) {
2316 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2317 gallivm_jit_function(variant->gallivm,
2318 variant->function[RAST_EDGE_TEST]);
2319 }
2320
2321 if (variant->function[RAST_WHOLE]) {
2322 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2323 gallivm_jit_function(variant->gallivm,
2324 variant->function[RAST_WHOLE]);
2325 } else if (!variant->jit_function[RAST_WHOLE]) {
2326 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2327 }
2328
2329 return variant;
2330 }
2331
2332
2333 static void *
2334 llvmpipe_create_fs_state(struct pipe_context *pipe,
2335 const struct pipe_shader_state *templ)
2336 {
2337 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2338 struct lp_fragment_shader *shader;
2339 int nr_samplers;
2340 int nr_sampler_views;
2341 int i;
2342
2343 shader = CALLOC_STRUCT(lp_fragment_shader);
2344 if (!shader)
2345 return NULL;
2346
2347 shader->no = fs_no++;
2348 make_empty_list(&shader->variants);
2349
2350 /* get/save the summary info for this shader */
2351 lp_build_tgsi_info(templ->tokens, &shader->info);
2352
2353 /* we need to keep a local copy of the tokens */
2354 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2355
2356 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2357 if (shader->draw_data == NULL) {
2358 FREE((void *) shader->base.tokens);
2359 FREE(shader);
2360 return NULL;
2361 }
2362
2363 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2364 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2365
2366 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2367 state[MAX2(nr_samplers, nr_sampler_views)]);
2368
2369 for (i = 0; i < shader->info.base.num_inputs; i++) {
2370 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2371 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2372
2373 switch (shader->info.base.input_interpolate[i]) {
2374 case TGSI_INTERPOLATE_CONSTANT:
2375 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2376 break;
2377 case TGSI_INTERPOLATE_LINEAR:
2378 shader->inputs[i].interp = LP_INTERP_LINEAR;
2379 break;
2380 case TGSI_INTERPOLATE_PERSPECTIVE:
2381 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2382 break;
2383 case TGSI_INTERPOLATE_COLOR:
2384 shader->inputs[i].interp = LP_INTERP_COLOR;
2385 break;
2386 default:
2387 assert(0);
2388 break;
2389 }
2390
2391 switch (shader->info.base.input_semantic_name[i]) {
2392 case TGSI_SEMANTIC_FACE:
2393 shader->inputs[i].interp = LP_INTERP_FACING;
2394 break;
2395 case TGSI_SEMANTIC_POSITION:
2396 /* Position was already emitted above
2397 */
2398 shader->inputs[i].interp = LP_INTERP_POSITION;
2399 shader->inputs[i].src_index = 0;
2400 continue;
2401 }
2402
2403 shader->inputs[i].src_index = i+1;
2404 }
2405
2406 if (LP_DEBUG & DEBUG_TGSI) {
2407 unsigned attrib;
2408 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2409 shader->no, (void *) shader);
2410 tgsi_dump(templ->tokens, 0);
2411 debug_printf("usage masks:\n");
2412 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2413 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2414 debug_printf(" IN[%u].%s%s%s%s\n",
2415 attrib,
2416 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2417 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2418 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2419 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2420 }
2421 debug_printf("\n");
2422 }
2423
2424 return shader;
2425 }
2426
2427
2428 static void
2429 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2430 {
2431 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2432
2433 if (llvmpipe->fs == fs)
2434 return;
2435
2436 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2437
2438 draw_bind_fragment_shader(llvmpipe->draw,
2439 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2440
2441 llvmpipe->dirty |= LP_NEW_FS;
2442 }
2443
2444
2445 /**
2446 * Remove shader variant from two lists: the shader's variant list
2447 * and the context's variant list.
2448 */
2449 void
2450 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2451 struct lp_fragment_shader_variant *variant)
2452 {
2453 unsigned i;
2454
2455 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2456 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2457 " #%u v total cached #%u\n",
2458 variant->shader->no,
2459 variant->no,
2460 variant->shader->variants_created,
2461 variant->shader->variants_cached,
2462 lp->nr_fs_variants);
2463 }
2464
2465 /* free all the variant's JIT'd functions */
2466 for (i = 0; i < Elements(variant->function); i++) {
2467 if (variant->function[i]) {
2468 gallivm_free_function(variant->gallivm,
2469 variant->function[i],
2470 variant->jit_function[i]);
2471 }
2472 }
2473
2474 gallivm_destroy(variant->gallivm);
2475
2476 /* remove from shader's list */
2477 remove_from_list(&variant->list_item_local);
2478 variant->shader->variants_cached--;
2479
2480 /* remove from context's list */
2481 remove_from_list(&variant->list_item_global);
2482 lp->nr_fs_variants--;
2483 lp->nr_fs_instrs -= variant->nr_instrs;
2484
2485 FREE(variant);
2486 }
2487
2488
2489 static void
2490 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2491 {
2492 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2493 struct lp_fragment_shader *shader = fs;
2494 struct lp_fs_variant_list_item *li;
2495
2496 assert(fs != llvmpipe->fs);
2497
2498 /*
2499 * XXX: we need to flush the context until we have some sort of reference
2500 * counting in fragment shaders as they may still be binned
2501 * Flushing alone might not sufficient we need to wait on it too.
2502 */
2503 llvmpipe_finish(pipe, __FUNCTION__);
2504
2505 /* Delete all the variants */
2506 li = first_elem(&shader->variants);
2507 while(!at_end(&shader->variants, li)) {
2508 struct lp_fs_variant_list_item *next = next_elem(li);
2509 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2510 li = next;
2511 }
2512
2513 /* Delete draw module's data */
2514 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2515
2516 assert(shader->variants_cached == 0);
2517 FREE((void *) shader->base.tokens);
2518 FREE(shader);
2519 }
2520
2521
2522
2523 static void
2524 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2525 uint shader, uint index,
2526 struct pipe_constant_buffer *cb)
2527 {
2528 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2529 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2530
2531 assert(shader < PIPE_SHADER_TYPES);
2532 assert(index < Elements(llvmpipe->constants[shader]));
2533
2534 /* note: reference counting */
2535 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2536
2537 if (shader == PIPE_SHADER_VERTEX ||
2538 shader == PIPE_SHADER_GEOMETRY) {
2539 /* Pass the constants to the 'draw' module */
2540 const unsigned size = cb ? cb->buffer_size : 0;
2541 const ubyte *data;
2542
2543 if (constants) {
2544 data = (ubyte *) llvmpipe_resource_data(constants);
2545 }
2546 else if (cb && cb->user_buffer) {
2547 data = (ubyte *) cb->user_buffer;
2548 }
2549 else {
2550 data = NULL;
2551 }
2552
2553 if (data)
2554 data += cb->buffer_offset;
2555
2556 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2557 index, data, size);
2558 }
2559
2560 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2561
2562 if (cb && cb->user_buffer) {
2563 pipe_resource_reference(&constants, NULL);
2564 }
2565 }
2566
2567
2568 /**
2569 * Return the blend factor equivalent to a destination alpha of one.
2570 */
2571 static INLINE unsigned
2572 force_dst_alpha_one(unsigned factor)
2573 {
2574 switch(factor) {
2575 case PIPE_BLENDFACTOR_DST_ALPHA:
2576 return PIPE_BLENDFACTOR_ONE;
2577 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2578 return PIPE_BLENDFACTOR_ZERO;
2579 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2580 return PIPE_BLENDFACTOR_ZERO;
2581 }
2582
2583 return factor;
2584 }
2585
2586
2587 /**
2588 * We need to generate several variants of the fragment pipeline to match
2589 * all the combinations of the contributing state atoms.
2590 *
2591 * TODO: there is actually no reason to tie this to context state -- the
2592 * generated code could be cached globally in the screen.
2593 */
2594 static void
2595 make_variant_key(struct llvmpipe_context *lp,
2596 struct lp_fragment_shader *shader,
2597 struct lp_fragment_shader_variant_key *key)
2598 {
2599 unsigned i;
2600
2601 memset(key, 0, shader->variant_key_size);
2602
2603 if (lp->framebuffer.zsbuf) {
2604 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2605 const struct util_format_description *zsbuf_desc =
2606 util_format_description(zsbuf_format);
2607
2608 if (lp->depth_stencil->depth.enabled &&
2609 util_format_has_depth(zsbuf_desc)) {
2610 key->zsbuf_format = zsbuf_format;
2611 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2612 }
2613 if (lp->depth_stencil->stencil[0].enabled &&
2614 util_format_has_stencil(zsbuf_desc)) {
2615 key->zsbuf_format = zsbuf_format;
2616 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2617 }
2618 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
2619 key->resource_1d = TRUE;
2620 }
2621 }
2622
2623 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2624 if (!lp->framebuffer.nr_cbufs ||
2625 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2626 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2627 }
2628 if(key->alpha.enabled)
2629 key->alpha.func = lp->depth_stencil->alpha.func;
2630 /* alpha.ref_value is passed in jit_context */
2631
2632 key->flatshade = lp->rasterizer->flatshade;
2633 if (lp->active_occlusion_queries) {
2634 key->occlusion_count = TRUE;
2635 }
2636
2637 if (lp->framebuffer.nr_cbufs) {
2638 memcpy(&key->blend, lp->blend, sizeof key->blend);
2639 }
2640
2641 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2642
2643 if (!key->blend.independent_blend_enable) {
2644 /* we always need independent blend otherwise the fixups below won't work */
2645 for (i = 1; i < key->nr_cbufs; i++) {
2646 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2647 }
2648 key->blend.independent_blend_enable = 1;
2649 }
2650
2651 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2652 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2653 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2654 const struct util_format_description *format_desc;
2655
2656 key->cbuf_format[i] = format;
2657
2658 /*
2659 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2660 * mixing of 2d textures with height 1 and 1d textures, so make sure
2661 * we pick 1d if any cbuf or zsbuf is 1d.
2662 */
2663 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[0]->texture)) {
2664 key->resource_1d = TRUE;
2665 }
2666
2667 format_desc = util_format_description(format);
2668 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2669 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2670
2671 /*
2672 * Mask out color channels not present in the color buffer.
2673 */
2674 blend_rt->colormask &= util_format_colormask(format_desc);
2675
2676 /*
2677 * Disable blend for integer formats.
2678 */
2679 if (util_format_is_pure_integer(format)) {
2680 blend_rt->blend_enable = 0;
2681 }
2682
2683 /*
2684 * Our swizzled render tiles always have an alpha channel, but the linear
2685 * render target format often does not, so force here the dst alpha to be
2686 * one.
2687 *
2688 * This is not a mere optimization. Wrong results will be produced if the
2689 * dst alpha is used, the dst format does not have alpha, and the previous
2690 * rendering was not flushed from the swizzled to linear buffer. For
2691 * example, NonPowTwo DCT.
2692 *
2693 * TODO: This should be generalized to all channels for better
2694 * performance, but only alpha causes correctness issues.
2695 *
2696 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2697 */
2698 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2699 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2700 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2701 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2702 blend_rt->alpha_func = blend_rt->rgb_func;
2703 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2704 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2705 }
2706 }
2707
2708 /* This value will be the same for all the variants of a given shader:
2709 */
2710 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2711
2712 for(i = 0; i < key->nr_samplers; ++i) {
2713 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2714 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2715 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2716 }
2717 }
2718
2719 /*
2720 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2721 * are dx10-style? Can't really have mixed opcodes, at least not
2722 * if we want to skip the holes here (without rescanning tgsi).
2723 */
2724 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2725 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2726 for(i = 0; i < key->nr_sampler_views; ++i) {
2727 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2728 lp_sampler_static_texture_state(&key->state[i].texture_state,
2729 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2730 }
2731 }
2732 }
2733 else {
2734 key->nr_sampler_views = key->nr_samplers;
2735 for(i = 0; i < key->nr_sampler_views; ++i) {
2736 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2737 lp_sampler_static_texture_state(&key->state[i].texture_state,
2738 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2739 }
2740 }
2741 }
2742 }
2743
2744
2745
2746 /**
2747 * Update fragment shader state. This is called just prior to drawing
2748 * something when some fragment-related state has changed.
2749 */
2750 void
2751 llvmpipe_update_fs(struct llvmpipe_context *lp)
2752 {
2753 struct lp_fragment_shader *shader = lp->fs;
2754 struct lp_fragment_shader_variant_key key;
2755 struct lp_fragment_shader_variant *variant = NULL;
2756 struct lp_fs_variant_list_item *li;
2757
2758 make_variant_key(lp, shader, &key);
2759
2760 /* Search the variants for one which matches the key */
2761 li = first_elem(&shader->variants);
2762 while(!at_end(&shader->variants, li)) {
2763 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2764 variant = li->base;
2765 break;
2766 }
2767 li = next_elem(li);
2768 }
2769
2770 if (variant) {
2771 /* Move this variant to the head of the list to implement LRU
2772 * deletion of shader's when we have too many.
2773 */
2774 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2775 }
2776 else {
2777 /* variant not found, create it now */
2778 int64_t t0, t1, dt;
2779 unsigned i;
2780 unsigned variants_to_cull;
2781
2782 if (0) {
2783 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2784 lp->nr_fs_variants,
2785 lp->nr_fs_instrs,
2786 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2787 }
2788
2789 /* First, check if we've exceeded the max number of shader variants.
2790 * If so, free 25% of them (the least recently used ones).
2791 */
2792 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2793
2794 if (variants_to_cull ||
2795 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2796 struct pipe_context *pipe = &lp->pipe;
2797
2798 /*
2799 * XXX: we need to flush the context until we have some sort of
2800 * reference counting in fragment shaders as they may still be binned
2801 * Flushing alone might not be sufficient we need to wait on it too.
2802 */
2803 llvmpipe_finish(pipe, __FUNCTION__);
2804
2805 /*
2806 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2807 * number of shader variants (potentially all of them) could be
2808 * pending for destruction on flush.
2809 */
2810
2811 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2812 struct lp_fs_variant_list_item *item;
2813 if (is_empty_list(&lp->fs_variants_list)) {
2814 break;
2815 }
2816 item = last_elem(&lp->fs_variants_list);
2817 assert(item);
2818 assert(item->base);
2819 llvmpipe_remove_shader_variant(lp, item->base);
2820 }
2821 }
2822
2823 /*
2824 * Generate the new variant.
2825 */
2826 t0 = os_time_get();
2827 variant = generate_variant(lp, shader, &key);
2828 t1 = os_time_get();
2829 dt = t1 - t0;
2830 LP_COUNT_ADD(llvm_compile_time, dt);
2831 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2832
2833 llvmpipe_variant_count++;
2834
2835 /* Put the new variant into the list */
2836 if (variant) {
2837 insert_at_head(&shader->variants, &variant->list_item_local);
2838 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2839 lp->nr_fs_variants++;
2840 lp->nr_fs_instrs += variant->nr_instrs;
2841 shader->variants_cached++;
2842 }
2843 }
2844
2845 /* Bind this variant */
2846 lp_setup_set_fs_variant(lp->setup, variant);
2847 }
2848
2849
2850
2851
2852
2853 void
2854 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2855 {
2856 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2857 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2858 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2859
2860 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2861 }