projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw,gallivm,llvmpipe: Avoid implicit casts of 32-bit shifts to 64-bits.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
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 VMWARE 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, 1ULL << (j + 0), 0);
178 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
179 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
180 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (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 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
219 */
220 static LLVMValueRef
221 lp_llvm_viewport(LLVMValueRef context_ptr,
222 struct gallivm_state *gallivm,
223 LLVMValueRef viewport_index)
224 {
225 LLVMBuilderRef builder = gallivm->builder;
226 LLVMValueRef ptr;
227 LLVMValueRef res;
228 struct lp_type viewport_type =
229 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
230
231 ptr = lp_jit_context_viewports(gallivm, context_ptr);
232 ptr = LLVMBuildPointerCast(builder, ptr,
233 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
234
235 res = lp_build_pointer_get(builder, ptr, viewport_index);
236
237 return res;
238 }
239
240
241 /**
242 * Generate the fragment shader, depth/stencil test, and alpha tests.
243 */
244 static void
245 generate_fs_loop(struct gallivm_state *gallivm,
246 struct lp_fragment_shader *shader,
247 const struct lp_fragment_shader_variant_key *key,
248 LLVMBuilderRef builder,
249 struct lp_type type,
250 LLVMValueRef context_ptr,
251 LLVMValueRef num_loop,
252 struct lp_build_interp_soa_context *interp,
253 struct lp_build_sampler_soa *sampler,
254 LLVMValueRef mask_store,
255 LLVMValueRef (*out_color)[4],
256 LLVMValueRef depth_ptr,
257 LLVMValueRef depth_stride,
258 LLVMValueRef facing,
259 LLVMValueRef thread_data_ptr)
260 {
261 const struct util_format_description *zs_format_desc = NULL;
262 const struct tgsi_token *tokens = shader->base.tokens;
263 LLVMTypeRef vec_type;
264 LLVMValueRef mask_ptr, mask_val;
265 LLVMValueRef consts_ptr, num_consts_ptr;
266 LLVMValueRef z;
267 LLVMValueRef z_value, s_value;
268 LLVMValueRef z_fb, s_fb;
269 LLVMValueRef stencil_refs[2];
270 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
271 struct lp_build_for_loop_state loop_state;
272 struct lp_build_mask_context mask;
273 /*
274 * TODO: figure out if simple_shader optimization is really worthwile to
275 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
276 * code since tests tend to take another codepath than real shaders.
277 */
278 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
279 shader->info.base.num_inputs < 3 &&
280 shader->info.base.num_instructions < 8) && 0;
281 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
282 util_blend_state_is_dual(&key->blend, 0);
283 unsigned attrib;
284 unsigned chan;
285 unsigned cbuf;
286 unsigned depth_mode;
287
288 struct lp_bld_tgsi_system_values system_values;
289
290 memset(&system_values, 0, sizeof(system_values));
291
292 if (key->depth.enabled ||
293 key->stencil[0].enabled) {
294
295 zs_format_desc = util_format_description(key->zsbuf_format);
296 assert(zs_format_desc);
297
298 if (!shader->info.base.writes_z) {
299 if (key->alpha.enabled ||
300 key->blend.alpha_to_coverage ||
301 shader->info.base.uses_kill) {
302 /* With alpha test and kill, can do the depth test early
303 * and hopefully eliminate some quads. But need to do a
304 * special deferred depth write once the final mask value
305 * is known. This only works though if there's either no
306 * stencil test or the stencil value isn't written.
307 */
308 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
309 (key->stencil[1].enabled &&
310 key->stencil[1].writemask)))
311 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
312 else
313 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
314 }
315 else
316 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
317 }
318 else {
319 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
320 }
321
322 if (!(key->depth.enabled && key->depth.writemask) &&
323 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
324 (key->stencil[1].enabled &&
325 key->stencil[1].writemask))))
326 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
327 }
328 else {
329 depth_mode = 0;
330 }
331
332
333 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
334 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
335
336 vec_type = lp_build_vec_type(gallivm, type);
337
338 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
339 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
340
341 lp_build_for_loop_begin(&loop_state, gallivm,
342 lp_build_const_int32(gallivm, 0),
343 LLVMIntULT,
344 num_loop,
345 lp_build_const_int32(gallivm, 1));
346
347 mask_ptr = LLVMBuildGEP(builder, mask_store,
348 &loop_state.counter, 1, "mask_ptr");
349 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
350
351 memset(outputs, 0, sizeof outputs);
352
353 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
354 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
355 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
356 lp_build_vec_type(gallivm,
357 type),
358 num_loop, "color");
359 }
360 }
361 if (dual_source_blend) {
362 assert(key->nr_cbufs <= 1);
363 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
364 out_color[1][chan] = lp_build_array_alloca(gallivm,
365 lp_build_vec_type(gallivm,
366 type),
367 num_loop, "color1");
368 }
369 }
370
371
372 /* 'mask' will control execution based on quad's pixel alive/killed state */
373 lp_build_mask_begin(&mask, gallivm, type, mask_val);
374
375 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
376 lp_build_mask_check(&mask);
377
378 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
379 z = interp->pos[2];
380
381 if (depth_mode & EARLY_DEPTH_TEST) {
382 lp_build_depth_stencil_load_swizzled(gallivm, type,
383 zs_format_desc, key->resource_1d,
384 depth_ptr, depth_stride,
385 &z_fb, &s_fb, loop_state.counter);
386 lp_build_depth_stencil_test(gallivm,
387 &key->depth,
388 key->stencil,
389 type,
390 zs_format_desc,
391 &mask,
392 stencil_refs,
393 z, z_fb, s_fb,
394 facing,
395 &z_value, &s_value,
396 !simple_shader);
397
398 if (depth_mode & EARLY_DEPTH_WRITE) {
399 lp_build_depth_stencil_write_swizzled(gallivm, type,
400 zs_format_desc, key->resource_1d,
401 NULL, NULL, NULL, loop_state.counter,
402 depth_ptr, depth_stride,
403 z_value, s_value);
404 }
405 /*
406 * Note mask check if stencil is enabled must be after ds write not after
407 * stencil test otherwise new stencil values may not get written if all
408 * fragments got killed by depth/stencil test.
409 */
410 if (!simple_shader && key->stencil[0].enabled)
411 lp_build_mask_check(&mask);
412 }
413
414 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
415
416 /* Build the actual shader */
417 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
418 consts_ptr, num_consts_ptr, &system_values,
419 interp->inputs,
420 outputs, sampler, &shader->info.base, NULL);
421
422 /* Alpha test */
423 if (key->alpha.enabled) {
424 int color0 = find_output_by_semantic(&shader->info.base,
425 TGSI_SEMANTIC_COLOR,
426 0);
427
428 if (color0 != -1 && outputs[color0][3]) {
429 const struct util_format_description *cbuf_format_desc;
430 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
431 LLVMValueRef alpha_ref_value;
432
433 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
434 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
435
436 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
437
438 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
439 &mask, alpha, alpha_ref_value,
440 (depth_mode & LATE_DEPTH_TEST) != 0);
441 }
442 }
443
444 /* Emulate Alpha to Coverage with Alpha test */
445 if (key->blend.alpha_to_coverage) {
446 int color0 = find_output_by_semantic(&shader->info.base,
447 TGSI_SEMANTIC_COLOR,
448 0);
449
450 if (color0 != -1 && outputs[color0][3]) {
451 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
452
453 lp_build_alpha_to_coverage(gallivm, type,
454 &mask, alpha,
455 (depth_mode & LATE_DEPTH_TEST) != 0);
456 }
457 }
458
459 /* Late Z test */
460 if (depth_mode & LATE_DEPTH_TEST) {
461 int pos0 = find_output_by_semantic(&shader->info.base,
462 TGSI_SEMANTIC_POSITION,
463 0);
464
465 if (pos0 != -1 && outputs[pos0][2]) {
466 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
467
468 /*
469 * Clamp according to ARB_depth_clamp semantics.
470 */
471 if (key->depth_clamp) {
472 LLVMValueRef viewport, min_depth, max_depth;
473 LLVMValueRef viewport_index;
474 struct lp_build_context f32_bld;
475
476 assert(type.floating);
477 lp_build_context_init(&f32_bld, gallivm, type);
478
479 /*
480 * Assumes clamping of the viewport index will occur in setup/gs. Value
481 * is passed through the rasterization stage via lp_rast_shader_inputs.
482 *
483 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
484 * semantics.
485 */
486 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
487 thread_data_ptr);
488
489 /*
490 * Load the min and max depth from the lp_jit_context.viewports
491 * array of lp_jit_viewport structures.
492 */
493 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
494
495 /* viewports[viewport_index].min_depth */
496 min_depth = LLVMBuildExtractElement(builder, viewport,
497 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH),
498 "");
499 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
500
501 /* viewports[viewport_index].max_depth */
502 max_depth = LLVMBuildExtractElement(builder, viewport,
503 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH),
504 "");
505 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
506
507 /*
508 * Clamp to the min and max depth values for the given viewport.
509 */
510 z = lp_build_clamp(&f32_bld, z, min_depth, max_depth);
511 }
512 }
513
514 lp_build_depth_stencil_load_swizzled(gallivm, type,
515 zs_format_desc, key->resource_1d,
516 depth_ptr, depth_stride,
517 &z_fb, &s_fb, loop_state.counter);
518
519 lp_build_depth_stencil_test(gallivm,
520 &key->depth,
521 key->stencil,
522 type,
523 zs_format_desc,
524 &mask,
525 stencil_refs,
526 z, z_fb, s_fb,
527 facing,
528 &z_value, &s_value,
529 !simple_shader);
530 /* Late Z write */
531 if (depth_mode & LATE_DEPTH_WRITE) {
532 lp_build_depth_stencil_write_swizzled(gallivm, type,
533 zs_format_desc, key->resource_1d,
534 NULL, NULL, NULL, loop_state.counter,
535 depth_ptr, depth_stride,
536 z_value, s_value);
537 }
538 }
539 else if ((depth_mode & EARLY_DEPTH_TEST) &&
540 (depth_mode & LATE_DEPTH_WRITE))
541 {
542 /* Need to apply a reduced mask to the depth write. Reload the
543 * depth value, update from zs_value with the new mask value and
544 * write that out.
545 */
546 lp_build_depth_stencil_write_swizzled(gallivm, type,
547 zs_format_desc, key->resource_1d,
548 &mask, z_fb, s_fb, loop_state.counter,
549 depth_ptr, depth_stride,
550 z_value, s_value);
551 }
552
553
554 /* Color write */
555 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
556 {
557 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
558 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
559 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
560 {
561 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
562 if(outputs[attrib][chan]) {
563 /* XXX: just initialize outputs to point at colors[] and
564 * skip this.
565 */
566 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
567 LLVMValueRef color_ptr;
568 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
569 &loop_state.counter, 1, "");
570 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
571 LLVMBuildStore(builder, out, color_ptr);
572 }
573 }
574 }
575 }
576
577 if (key->occlusion_count) {
578 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
579 lp_build_name(counter, "counter");
580 lp_build_occlusion_count(gallivm, type,
581 lp_build_mask_value(&mask), counter);
582 }
583
584 mask_val = lp_build_mask_end(&mask);
585 LLVMBuildStore(builder, mask_val, mask_ptr);
586 lp_build_for_loop_end(&loop_state);
587 }
588
589
590 /**
591 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
592 *
593 * Fragment Shader outputs pixels in small 2x2 blocks
594 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
595 *
596 * However in memory pixels are stored in rows
597 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
598 *
599 * @param type fragment shader type (4x or 8x float)
600 * @param num_fs number of fs_src
601 * @param is_1d whether we're outputting to a 1d resource
602 * @param dst_channels number of output channels
603 * @param fs_src output from fragment shader
604 * @param dst pointer to store result
605 * @param pad_inline is channel padding inline or at end of row
606 * @return the number of dsts
607 */
608 static int
609 generate_fs_twiddle(struct gallivm_state *gallivm,
610 struct lp_type type,
611 unsigned num_fs,
612 unsigned dst_channels,
613 LLVMValueRef fs_src[][4],
614 LLVMValueRef* dst,
615 bool pad_inline)
616 {
617 LLVMValueRef src[16];
618
619 bool swizzle_pad;
620 bool twiddle;
621 bool split;
622
623 unsigned pixels = type.length / 4;
624 unsigned reorder_group;
625 unsigned src_channels;
626 unsigned src_count;
627 unsigned i;
628
629 src_channels = dst_channels < 3 ? dst_channels : 4;
630 src_count = num_fs * src_channels;
631
632 assert(pixels == 2 || pixels == 1);
633 assert(num_fs * src_channels <= Elements(src));
634
635 /*
636 * Transpose from SoA -> AoS
637 */
638 for (i = 0; i < num_fs; ++i) {
639 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
640 }
641
642 /*
643 * Pick transformation options
644 */
645 swizzle_pad = false;
646 twiddle = false;
647 split = false;
648 reorder_group = 0;
649
650 if (dst_channels == 1) {
651 twiddle = true;
652
653 if (pixels == 2) {
654 split = true;
655 }
656 } else if (dst_channels == 2) {
657 if (pixels == 1) {
658 reorder_group = 1;
659 }
660 } else if (dst_channels > 2) {
661 if (pixels == 1) {
662 reorder_group = 2;
663 } else {
664 twiddle = true;
665 }
666
667 if (!pad_inline && dst_channels == 3 && pixels > 1) {
668 swizzle_pad = true;
669 }
670 }
671
672 /*
673 * Split the src in half
674 */
675 if (split) {
676 for (i = num_fs; i > 0; --i) {
677 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
678 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
679 }
680
681 src_count *= 2;
682 type.length = 4;
683 }
684
685 /*
686 * Ensure pixels are in memory order
687 */
688 if (reorder_group) {
689 /* Twiddle pixels by reordering the array, e.g.:
690 *
691 * src_count = 8 -> 0 2 1 3 4 6 5 7
692 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
693 */
694 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
695
696 for (i = 0; i < src_count; ++i) {
697 unsigned group = i / reorder_group;
698 unsigned block = (group / 4) * 4 * reorder_group;
699 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
700 dst[i] = src[j];
701 }
702 } else if (twiddle) {
703 /* Twiddle pixels across elements of array */
704 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
705 } else {
706 /* Do nothing */
707 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
708 }
709
710 /*
711 * Moves any padding between pixels to the end
712 * e.g. RGBXRGBX -> RGBRGBXX
713 */
714 if (swizzle_pad) {
715 unsigned char swizzles[16];
716 unsigned elems = pixels * dst_channels;
717
718 for (i = 0; i < type.length; ++i) {
719 if (i < elems)
720 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
721 else
722 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
723 }
724
725 for (i = 0; i < src_count; ++i) {
726 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
727 }
728 }
729
730 return src_count;
731 }
732
733
734 /**
735 * Load an unswizzled block of pixels from memory
736 */
737 static void
738 load_unswizzled_block(struct gallivm_state *gallivm,
739 LLVMValueRef base_ptr,
740 LLVMValueRef stride,
741 unsigned block_width,
742 unsigned block_height,
743 LLVMValueRef* dst,
744 struct lp_type dst_type,
745 unsigned dst_count,
746 unsigned dst_alignment)
747 {
748 LLVMBuilderRef builder = gallivm->builder;
749 unsigned row_size = dst_count / block_height;
750 unsigned i;
751
752 /* Ensure block exactly fits into dst */
753 assert((block_width * block_height) % dst_count == 0);
754
755 for (i = 0; i < dst_count; ++i) {
756 unsigned x = i % row_size;
757 unsigned y = i / row_size;
758
759 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
760 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
761
762 LLVMValueRef gep[2];
763 LLVMValueRef dst_ptr;
764
765 gep[0] = lp_build_const_int32(gallivm, 0);
766 gep[1] = LLVMBuildAdd(builder, bx, by, "");
767
768 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
769 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
770
771 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
772
773 lp_set_load_alignment(dst[i], dst_alignment);
774 }
775 }
776
777
778 /**
779 * Store an unswizzled block of pixels to memory
780 */
781 static void
782 store_unswizzled_block(struct gallivm_state *gallivm,
783 LLVMValueRef base_ptr,
784 LLVMValueRef stride,
785 unsigned block_width,
786 unsigned block_height,
787 LLVMValueRef* src,
788 struct lp_type src_type,
789 unsigned src_count,
790 unsigned src_alignment)
791 {
792 LLVMBuilderRef builder = gallivm->builder;
793 unsigned row_size = src_count / block_height;
794 unsigned i;
795
796 /* Ensure src exactly fits into block */
797 assert((block_width * block_height) % src_count == 0);
798
799 for (i = 0; i < src_count; ++i) {
800 unsigned x = i % row_size;
801 unsigned y = i / row_size;
802
803 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
804 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
805
806 LLVMValueRef gep[2];
807 LLVMValueRef src_ptr;
808
809 gep[0] = lp_build_const_int32(gallivm, 0);
810 gep[1] = LLVMBuildAdd(builder, bx, by, "");
811
812 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
813 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
814
815 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
816
817 lp_set_store_alignment(src_ptr, src_alignment);
818 }
819 }
820
821
822 /**
823 * Checks if a format description is an arithmetic format
824 *
825 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
826 */
827 static INLINE boolean
828 is_arithmetic_format(const struct util_format_description *format_desc)
829 {
830 boolean arith = false;
831 unsigned i;
832
833 for (i = 0; i < format_desc->nr_channels; ++i) {
834 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
835 arith |= (format_desc->channel[i].size % 8) != 0;
836 }
837
838 return arith;
839 }
840
841
842 /**
843 * Checks if this format requires special handling due to required expansion
844 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
845 * SoA conversion.
846 */
847 static INLINE boolean
848 format_expands_to_float_soa(const struct util_format_description *format_desc)
849 {
850 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
851 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
852 return true;
853 }
854 return false;
855 }
856
857
858 /**
859 * Retrieves the type representing the memory layout for a format
860 *
861 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
862 */
863 static INLINE void
864 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
865 struct lp_type* type)
866 {
867 unsigned i;
868 unsigned chan;
869
870 if (format_expands_to_float_soa(format_desc)) {
871 /* just make this a uint with width of block */
872 type->floating = false;
873 type->fixed = false;
874 type->sign = false;
875 type->norm = false;
876 type->width = format_desc->block.bits;
877 type->length = 1;
878 return;
879 }
880
881 for (i = 0; i < 4; i++)
882 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
883 break;
884 chan = i;
885
886 memset(type, 0, sizeof(struct lp_type));
887 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
888 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
889 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
890 type->norm = format_desc->channel[chan].normalized;
891
892 if (is_arithmetic_format(format_desc)) {
893 type->width = 0;
894 type->length = 1;
895
896 for (i = 0; i < format_desc->nr_channels; ++i) {
897 type->width += format_desc->channel[i].size;
898 }
899 } else {
900 type->width = format_desc->channel[chan].size;
901 type->length = format_desc->nr_channels;
902 }
903 }
904
905
906 /**
907 * Retrieves the type for a format which is usable in the blending code.
908 *
909 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
910 */
911 static INLINE void
912 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
913 struct lp_type* type)
914 {
915 unsigned i;
916 unsigned chan;
917
918 if (format_expands_to_float_soa(format_desc)) {
919 /* always use ordinary floats for blending */
920 type->floating = true;
921 type->fixed = false;
922 type->sign = true;
923 type->norm = false;
924 type->width = 32;
925 type->length = 4;
926 return;
927 }
928
929 for (i = 0; i < 4; i++)
930 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
931 break;
932 chan = i;
933
934 memset(type, 0, sizeof(struct lp_type));
935 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
936 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
937 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
938 type->norm = format_desc->channel[chan].normalized;
939 type->width = format_desc->channel[chan].size;
940 type->length = format_desc->nr_channels;
941
942 for (i = 1; i < format_desc->nr_channels; ++i) {
943 if (format_desc->channel[i].size > type->width)
944 type->width = format_desc->channel[i].size;
945 }
946
947 if (type->floating) {
948 type->width = 32;
949 } else {
950 if (type->width <= 8) {
951 type->width = 8;
952 } else if (type->width <= 16) {
953 type->width = 16;
954 } else {
955 type->width = 32;
956 }
957 }
958
959 if (is_arithmetic_format(format_desc) && type->length == 3) {
960 type->length = 4;
961 }
962 }
963
964
965 /**
966 * Scale a normalized value from src_bits to dst_bits.
967 *
968 * The exact calculation is
969 *
970 * dst = iround(src * dst_mask / src_mask)
971 *
972 * or with integer rounding
973 *
974 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
975 *
976 * where
977 *
978 * src_mask = (1 << src_bits) - 1
979 * dst_mask = (1 << dst_bits) - 1
980 *
981 * but we try to avoid division and multiplication through shifts.
982 */
983 static INLINE LLVMValueRef
984 scale_bits(struct gallivm_state *gallivm,
985 int src_bits,
986 int dst_bits,
987 LLVMValueRef src,
988 struct lp_type src_type)
989 {
990 LLVMBuilderRef builder = gallivm->builder;
991 LLVMValueRef result = src;
992
993 if (dst_bits < src_bits) {
994 int delta_bits = src_bits - dst_bits;
995
996 if (delta_bits <= dst_bits) {
997 /*
998 * Approximate the rescaling with a single shift.
999 *
1000 * This gives the wrong rounding.
1001 */
1002
1003 result = LLVMBuildLShr(builder,
1004 src,
1005 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1006 "");
1007
1008 } else {
1009 /*
1010 * Try more accurate rescaling.
1011 */
1012
1013 /*
1014 * Drop the least significant bits to make space for the multiplication.
1015 *
1016 * XXX: A better approach would be to use a wider integer type as intermediate. But
1017 * this is enough to convert alpha from 16bits -> 2 when rendering to
1018 * PIPE_FORMAT_R10G10B10A2_UNORM.
1019 */
1020 result = LLVMBuildLShr(builder,
1021 src,
1022 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1023 "");
1024
1025
1026 result = LLVMBuildMul(builder,
1027 result,
1028 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1029 "");
1030
1031 /*
1032 * Add a rounding term before the division.
1033 *
1034 * TODO: Handle signed integers too.
1035 */
1036 if (!src_type.sign) {
1037 result = LLVMBuildAdd(builder,
1038 result,
1039 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1040 "");
1041 }
1042
1043 /*
1044 * Approximate the division by src_mask with a src_bits shift.
1045 *
1046 * Given the src has already been shifted by dst_bits, all we need
1047 * to do is to shift by the difference.
1048 */
1049
1050 result = LLVMBuildLShr(builder,
1051 result,
1052 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1053 "");
1054 }
1055
1056 } else if (dst_bits > src_bits) {
1057 /* Scale up bits */
1058 int db = dst_bits - src_bits;
1059
1060 /* Shift left by difference in bits */
1061 result = LLVMBuildShl(builder,
1062 src,
1063 lp_build_const_int_vec(gallivm, src_type, db),
1064 "");
1065
1066 if (db < src_bits) {
1067 /* Enough bits in src to fill the remainder */
1068 LLVMValueRef lower = LLVMBuildLShr(builder,
1069 src,
1070 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1071 "");
1072
1073 result = LLVMBuildOr(builder, result, lower, "");
1074 } else if (db > src_bits) {
1075 /* Need to repeatedly copy src bits to fill remainder in dst */
1076 unsigned n;
1077
1078 for (n = src_bits; n < dst_bits; n *= 2) {
1079 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1080
1081 result = LLVMBuildOr(builder,
1082 result,
1083 LLVMBuildLShr(builder, result, shuv, ""),
1084 "");
1085 }
1086 }
1087 }
1088
1089 return result;
1090 }
1091
1092 /**
1093 * If RT is a smallfloat (needing denorms) format
1094 */
1095 static INLINE int
1096 have_smallfloat_format(struct lp_type dst_type,
1097 enum pipe_format format)
1098 {
1099 return ((dst_type.floating && dst_type.width != 32) ||
1100 /* due to format handling hacks this format doesn't have floating set
1101 * here (and actually has width set to 32 too) so special case this. */
1102 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1103 }
1104
1105
1106 /**
1107 * Convert from memory format to blending format
1108 *
1109 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1110 */
1111 static void
1112 convert_to_blend_type(struct gallivm_state *gallivm,
1113 unsigned block_size,
1114 const struct util_format_description *src_fmt,
1115 struct lp_type src_type,
1116 struct lp_type dst_type,
1117 LLVMValueRef* src, // and dst
1118 unsigned num_srcs)
1119 {
1120 LLVMValueRef *dst = src;
1121 LLVMBuilderRef builder = gallivm->builder;
1122 struct lp_type blend_type;
1123 struct lp_type mem_type;
1124 unsigned i, j, k;
1125 unsigned pixels = block_size / num_srcs;
1126 bool is_arith;
1127
1128 /*
1129 * full custom path for packed floats and srgb formats - none of the later
1130 * functions would do anything useful, and given the lp_type representation they
1131 * can't be fixed. Should really have some SoA blend path for these kind of
1132 * formats rather than hacking them in here.
1133 */
1134 if (format_expands_to_float_soa(src_fmt)) {
1135 LLVMValueRef tmpsrc[4];
1136 /*
1137 * This is pretty suboptimal for this case blending in SoA would be much
1138 * better, since conversion gets us SoA values so need to convert back.
1139 */
1140 assert(src_type.width == 32 || src_type.width == 16);
1141 assert(dst_type.floating);
1142 assert(dst_type.width == 32);
1143 assert(dst_type.length % 4 == 0);
1144 assert(num_srcs % 4 == 0);
1145
1146 if (src_type.width == 16) {
1147 /* expand 4x16bit values to 4x32bit */
1148 struct lp_type type32x4 = src_type;
1149 LLVMTypeRef ltype32x4;
1150 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1151 type32x4.width = 32;
1152 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1153 for (i = 0; i < num_fetch; i++) {
1154 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1155 }
1156 src_type.width = 32;
1157 }
1158 for (i = 0; i < 4; i++) {
1159 tmpsrc[i] = src[i];
1160 }
1161 for (i = 0; i < num_srcs / 4; i++) {
1162 LLVMValueRef tmpsoa[4];
1163 LLVMValueRef tmps = tmpsrc[i];
1164 if (dst_type.length == 8) {
1165 LLVMValueRef shuffles[8];
1166 unsigned j;
1167 /* fetch was 4 values but need 8-wide output values */
1168 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1169 /*
1170 * for 8-wide aos transpose would give us wrong order not matching
1171 * incoming converted fs values and mask. ARGH.
1172 */
1173 for (j = 0; j < 4; j++) {
1174 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1175 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1176 }
1177 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1178 LLVMConstVector(shuffles, 8), "");
1179 }
1180 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1181 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1182 }
1183 else {
1184 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1185 }
1186 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1187 }
1188 return;
1189 }
1190
1191 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1192 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1193
1194 /* Is the format arithmetic */
1195 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1196 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1197
1198 /* Pad if necessary */
1199 if (!is_arith && src_type.length < dst_type.length) {
1200 for (i = 0; i < num_srcs; ++i) {
1201 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1202 }
1203
1204 src_type.length = dst_type.length;
1205 }
1206
1207 /* Special case for half-floats */
1208 if (mem_type.width == 16 && mem_type.floating) {
1209 assert(blend_type.width == 32 && blend_type.floating);
1210 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1211 is_arith = false;
1212 }
1213
1214 if (!is_arith) {
1215 return;
1216 }
1217
1218 src_type.width = blend_type.width * blend_type.length;
1219 blend_type.length *= pixels;
1220 src_type.length *= pixels / (src_type.length / mem_type.length);
1221
1222 for (i = 0; i < num_srcs; ++i) {
1223 LLVMValueRef chans[4];
1224 LLVMValueRef res = NULL;
1225
1226 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1227
1228 for (j = 0; j < src_fmt->nr_channels; ++j) {
1229 unsigned mask = 0;
1230 unsigned sa = src_fmt->channel[j].shift;
1231 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1232 unsigned from_lsb = j;
1233 #else
1234 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1235 #endif
1236
1237 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1238 mask |= 1 << k;
1239 }
1240
1241 /* Extract bits from source */
1242 chans[j] = LLVMBuildLShr(builder,
1243 dst[i],
1244 lp_build_const_int_vec(gallivm, src_type, sa),
1245 "");
1246
1247 chans[j] = LLVMBuildAnd(builder,
1248 chans[j],
1249 lp_build_const_int_vec(gallivm, src_type, mask),
1250 "");
1251
1252 /* Scale bits */
1253 if (src_type.norm) {
1254 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1255 blend_type.width, chans[j], src_type);
1256 }
1257
1258 /* Insert bits into correct position */
1259 chans[j] = LLVMBuildShl(builder,
1260 chans[j],
1261 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1262 "");
1263
1264 if (j == 0) {
1265 res = chans[j];
1266 } else {
1267 res = LLVMBuildOr(builder, res, chans[j], "");
1268 }
1269 }
1270
1271 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1272 }
1273 }
1274
1275
1276 /**
1277 * Convert from blending format to memory format
1278 *
1279 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1280 */
1281 static void
1282 convert_from_blend_type(struct gallivm_state *gallivm,
1283 unsigned block_size,
1284 const struct util_format_description *src_fmt,
1285 struct lp_type src_type,
1286 struct lp_type dst_type,
1287 LLVMValueRef* src, // and dst
1288 unsigned num_srcs)
1289 {
1290 LLVMValueRef* dst = src;
1291 unsigned i, j, k;
1292 struct lp_type mem_type;
1293 struct lp_type blend_type;
1294 LLVMBuilderRef builder = gallivm->builder;
1295 unsigned pixels = block_size / num_srcs;
1296 bool is_arith;
1297
1298 /*
1299 * full custom path for packed floats and srgb formats - none of the later
1300 * functions would do anything useful, and given the lp_type representation they
1301 * can't be fixed. Should really have some SoA blend path for these kind of
1302 * formats rather than hacking them in here.
1303 */
1304 if (format_expands_to_float_soa(src_fmt)) {
1305 /*
1306 * This is pretty suboptimal for this case blending in SoA would be much
1307 * better - we need to transpose the AoS values back to SoA values for
1308 * conversion/packing.
1309 */
1310 assert(src_type.floating);
1311 assert(src_type.width == 32);
1312 assert(src_type.length % 4 == 0);
1313 assert(dst_type.width == 32 || dst_type.width == 16);
1314
1315 for (i = 0; i < num_srcs / 4; i++) {
1316 LLVMValueRef tmpsoa[4], tmpdst;
1317 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1318 /* really really need SoA here */
1319
1320 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1321 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1322 }
1323 else {
1324 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1325 src_type, tmpsoa);
1326 }
1327
1328 if (src_type.length == 8) {
1329 LLVMValueRef tmpaos, shuffles[8];
1330 unsigned j;
1331 /*
1332 * for 8-wide aos transpose has given us wrong order not matching
1333 * output order. HMPF. Also need to split the output values manually.
1334 */
1335 for (j = 0; j < 4; j++) {
1336 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1337 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1338 }
1339 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1340 LLVMConstVector(shuffles, 8), "");
1341 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1342 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1343 }
1344 else {
1345 src[i] = tmpdst;
1346 }
1347 }
1348 if (dst_type.width == 16) {
1349 struct lp_type type16x8 = dst_type;
1350 struct lp_type type32x4 = dst_type;
1351 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1352 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1353 type16x8.length = 8;
1354 type32x4.width = 32;
1355 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1356 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1357 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1358 /* We could do vector truncation but it doesn't generate very good code */
1359 for (i = 0; i < num_fetch; i++) {
1360 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1361 src[i], lp_build_zero(gallivm, type32x4));
1362 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1363 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1364 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1365 }
1366 }
1367 return;
1368 }
1369
1370 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1371 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1372
1373 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1374
1375 /* Special case for half-floats */
1376 if (mem_type.width == 16 && mem_type.floating) {
1377 int length = dst_type.length;
1378 assert(blend_type.width == 32 && blend_type.floating);
1379
1380 dst_type.length = src_type.length;
1381
1382 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1383
1384 dst_type.length = length;
1385 is_arith = false;
1386 }
1387
1388 /* Remove any padding */
1389 if (!is_arith && (src_type.length % mem_type.length)) {
1390 src_type.length -= (src_type.length % mem_type.length);
1391
1392 for (i = 0; i < num_srcs; ++i) {
1393 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1394 }
1395 }
1396
1397 /* No bit arithmetic to do */
1398 if (!is_arith) {
1399 return;
1400 }
1401
1402 src_type.length = pixels;
1403 src_type.width = blend_type.length * blend_type.width;
1404 dst_type.length = pixels;
1405
1406 for (i = 0; i < num_srcs; ++i) {
1407 LLVMValueRef chans[4];
1408 LLVMValueRef res = NULL;
1409
1410 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1411
1412 for (j = 0; j < src_fmt->nr_channels; ++j) {
1413 unsigned mask = 0;
1414 unsigned sa = src_fmt->channel[j].shift;
1415 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1416 unsigned from_lsb = j;
1417 #else
1418 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1419 #endif
1420
1421 assert(blend_type.width > src_fmt->channel[j].size);
1422
1423 for (k = 0; k < blend_type.width; ++k) {
1424 mask |= 1 << k;
1425 }
1426
1427 /* Extract bits */
1428 chans[j] = LLVMBuildLShr(builder,
1429 dst[i],
1430 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1431 "");
1432
1433 chans[j] = LLVMBuildAnd(builder,
1434 chans[j],
1435 lp_build_const_int_vec(gallivm, src_type, mask),
1436 "");
1437
1438 /* Scale down bits */
1439 if (src_type.norm) {
1440 chans[j] = scale_bits(gallivm, blend_type.width,
1441 src_fmt->channel[j].size, chans[j], src_type);
1442 }
1443
1444 /* Insert bits */
1445 chans[j] = LLVMBuildShl(builder,
1446 chans[j],
1447 lp_build_const_int_vec(gallivm, src_type, sa),
1448 "");
1449
1450 sa += src_fmt->channel[j].size;
1451
1452 if (j == 0) {
1453 res = chans[j];
1454 } else {
1455 res = LLVMBuildOr(builder, res, chans[j], "");
1456 }
1457 }
1458
1459 assert (dst_type.width != 24);
1460
1461 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1462 }
1463 }
1464
1465
1466 /**
1467 * Convert alpha to same blend type as src
1468 */
1469 static void
1470 convert_alpha(struct gallivm_state *gallivm,
1471 struct lp_type row_type,
1472 struct lp_type alpha_type,
1473 const unsigned block_size,
1474 const unsigned block_height,
1475 const unsigned src_count,
1476 const unsigned dst_channels,
1477 const bool pad_inline,
1478 LLVMValueRef* src_alpha)
1479 {
1480 LLVMBuilderRef builder = gallivm->builder;
1481 unsigned i, j;
1482 unsigned length = row_type.length;
1483 row_type.length = alpha_type.length;
1484
1485 /* Twiddle the alpha to match pixels */
1486 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1487
1488 /*
1489 * TODO this should use single lp_build_conv call for
1490 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1491 */
1492 for (i = 0; i < block_height; ++i) {
1493 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1494 }
1495
1496 alpha_type = row_type;
1497 row_type.length = length;
1498
1499 /* If only one channel we can only need the single alpha value per pixel */
1500 if (src_count == 1 && dst_channels == 1) {
1501
1502 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1503 } else {
1504 /* If there are more srcs than rows then we need to split alpha up */
1505 if (src_count > block_height) {
1506 for (i = src_count; i > 0; --i) {
1507 unsigned pixels = block_size / src_count;
1508 unsigned idx = i - 1;
1509
1510 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1511 (idx * pixels) % 4, pixels);
1512 }
1513 }
1514
1515 /* If there is a src for each pixel broadcast the alpha across whole row */
1516 if (src_count == block_size) {
1517 for (i = 0; i < src_count; ++i) {
1518 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1519 }
1520 } else {
1521 unsigned pixels = block_size / src_count;
1522 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1523 unsigned alpha_span = 1;
1524 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1525
1526 /* Check if we need 2 src_alphas for our shuffles */
1527 if (pixels > alpha_type.length) {
1528 alpha_span = 2;
1529 }
1530
1531 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1532 for (j = 0; j < row_type.length; ++j) {
1533 if (j < pixels * channels) {
1534 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1535 } else {
1536 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1537 }
1538 }
1539
1540 for (i = 0; i < src_count; ++i) {
1541 unsigned idx1 = i, idx2 = i;
1542
1543 if (alpha_span > 1){
1544 idx1 *= alpha_span;
1545 idx2 = idx1 + 1;
1546 }
1547
1548 src_alpha[i] = LLVMBuildShuffleVector(builder,
1549 src_alpha[idx1],
1550 src_alpha[idx2],
1551 LLVMConstVector(shuffles, row_type.length),
1552 "");
1553 }
1554 }
1555 }
1556 }
1557
1558
1559 /**
1560 * Generates the blend function for unswizzled colour buffers
1561 * Also generates the read & write from colour buffer
1562 */
1563 static void
1564 generate_unswizzled_blend(struct gallivm_state *gallivm,
1565 unsigned rt,
1566 struct lp_fragment_shader_variant *variant,
1567 enum pipe_format out_format,
1568 unsigned int num_fs,
1569 struct lp_type fs_type,
1570 LLVMValueRef* fs_mask,
1571 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1572 LLVMValueRef context_ptr,
1573 LLVMValueRef color_ptr,
1574 LLVMValueRef stride,
1575 unsigned partial_mask,
1576 boolean do_branch)
1577 {
1578 const unsigned alpha_channel = 3;
1579 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1580 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1581 const unsigned block_size = block_width * block_height;
1582 const unsigned lp_integer_vector_width = 128;
1583
1584 LLVMBuilderRef builder = gallivm->builder;
1585 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1586 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1587 LLVMValueRef src_alpha[4 * 4];
1588 LLVMValueRef src1_alpha[4 * 4];
1589 LLVMValueRef src_mask[4 * 4];
1590 LLVMValueRef src[4 * 4];
1591 LLVMValueRef src1[4 * 4];
1592 LLVMValueRef dst[4 * 4];
1593 LLVMValueRef blend_color;
1594 LLVMValueRef blend_alpha;
1595 LLVMValueRef i32_zero;
1596 LLVMValueRef check_mask;
1597 LLVMValueRef undef_src_val;
1598
1599 struct lp_build_mask_context mask_ctx;
1600 struct lp_type mask_type;
1601 struct lp_type blend_type;
1602 struct lp_type row_type;
1603 struct lp_type dst_type;
1604
1605 unsigned char swizzle[TGSI_NUM_CHANNELS];
1606 unsigned vector_width;
1607 unsigned src_channels = TGSI_NUM_CHANNELS;
1608 unsigned dst_channels;
1609 unsigned dst_count;
1610 unsigned src_count;
1611 unsigned i, j;
1612
1613 const struct util_format_description* out_format_desc = util_format_description(out_format);
1614
1615 unsigned dst_alignment;
1616
1617 bool pad_inline = is_arithmetic_format(out_format_desc);
1618 bool has_alpha = false;
1619 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1620 util_blend_state_is_dual(&variant->key.blend, 0);
1621
1622 const boolean is_1d = variant->key.resource_1d;
1623 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1624 LLVMValueRef fpstate = 0;
1625
1626 /* Get type from output format */
1627 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1628 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1629
1630 /*
1631 * Technically this code should go into lp_build_smallfloat_to_float
1632 * and lp_build_float_to_smallfloat but due to the
1633 * http://llvm.org/bugs/show_bug.cgi?id=6393
1634 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1635 * So the ordering is important here and there shouldn't be any
1636 * llvm ir instrunctions in this function before
1637 * this, otherwise half-float format conversions won't work
1638 * (again due to llvm bug #6393).
1639 */
1640 if (have_smallfloat_format(dst_type, out_format)) {
1641 /* We need to make sure that denorms are ok for half float
1642 conversions */
1643 fpstate = lp_build_fpstate_get(gallivm);
1644 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
1645 }
1646
1647 mask_type = lp_int32_vec4_type();
1648 mask_type.length = fs_type.length;
1649
1650 for (i = num_fs; i < num_fullblock_fs; i++) {
1651 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1652 }
1653
1654 /* Do not bother executing code when mask is empty.. */
1655 if (do_branch) {
1656 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1657
1658 for (i = 0; i < num_fullblock_fs; ++i) {
1659 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1660 }
1661
1662 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1663 lp_build_mask_check(&mask_ctx);
1664 }
1665
1666 partial_mask |= !variant->opaque;
1667 i32_zero = lp_build_const_int32(gallivm, 0);
1668
1669 undef_src_val = lp_build_undef(gallivm, fs_type);
1670
1671 row_type.length = fs_type.length;
1672 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1673
1674 /* Compute correct swizzle and count channels */
1675 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1676 dst_channels = 0;
1677
1678 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1679 /* Ensure channel is used */
1680 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1681 continue;
1682 }
1683
1684 /* Ensure not already written to (happens in case with GL_ALPHA) */
1685 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1686 continue;
1687 }
1688
1689 /* Ensure we havn't already found all channels */
1690 if (dst_channels >= out_format_desc->nr_channels) {
1691 continue;
1692 }
1693
1694 swizzle[out_format_desc->swizzle[i]] = i;
1695 ++dst_channels;
1696
1697 if (i == alpha_channel) {
1698 has_alpha = true;
1699 }
1700 }
1701
1702 if (format_expands_to_float_soa(out_format_desc)) {
1703 /*
1704 * the code above can't work for layout_other
1705 * for srgb it would sort of work but we short-circuit swizzles, etc.
1706 * as that is done as part of unpack / pack.
1707 */
1708 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1709 has_alpha = true;
1710 swizzle[0] = 0;
1711 swizzle[1] = 1;
1712 swizzle[2] = 2;
1713 swizzle[3] = 3;
1714 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1715 }
1716
1717 /* If 3 channels then pad to include alpha for 4 element transpose */
1718 if (dst_channels == 3 && !has_alpha) {
1719 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1720 if (swizzle[i] > TGSI_NUM_CHANNELS)
1721 swizzle[i] = 3;
1722 }
1723 if (out_format_desc->nr_channels == 4) {
1724 dst_channels = 4;
1725 }
1726 }
1727
1728 /*
1729 * Load shader output
1730 */
1731 for (i = 0; i < num_fullblock_fs; ++i) {
1732 /* Always load alpha for use in blending */
1733 LLVMValueRef alpha;
1734 if (i < num_fs) {
1735 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1736 }
1737 else {
1738 alpha = undef_src_val;
1739 }
1740
1741 /* Load each channel */
1742 for (j = 0; j < dst_channels; ++j) {
1743 assert(swizzle[j] < 4);
1744 if (i < num_fs) {
1745 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1746 }
1747 else {
1748 fs_src[i][j] = undef_src_val;
1749 }
1750 }
1751
1752 /* If 3 channels then pad to include alpha for 4 element transpose */
1753 /*
1754 * XXX If we include that here maybe could actually use it instead of
1755 * separate alpha for blending?
1756 */
1757 if (dst_channels == 3 && !has_alpha) {
1758 fs_src[i][3] = alpha;
1759 }
1760
1761 /* We split the row_mask and row_alpha as we want 128bit interleave */
1762 if (fs_type.length == 8) {
1763 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1764 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1765
1766 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1767 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1768 } else {
1769 src_mask[i] = fs_mask[i];
1770 src_alpha[i] = alpha;
1771 }
1772 }
1773 if (dual_source_blend) {
1774 /* same as above except different src/dst, skip masks and comments... */
1775 for (i = 0; i < num_fullblock_fs; ++i) {
1776 LLVMValueRef alpha;
1777 if (i < num_fs) {
1778 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1779 }
1780 else {
1781 alpha = undef_src_val;
1782 }
1783
1784 for (j = 0; j < dst_channels; ++j) {
1785 assert(swizzle[j] < 4);
1786 if (i < num_fs) {
1787 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1788 }
1789 else {
1790 fs_src1[i][j] = undef_src_val;
1791 }
1792 }
1793 if (dst_channels == 3 && !has_alpha) {
1794 fs_src1[i][3] = alpha;
1795 }
1796 if (fs_type.length == 8) {
1797 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1798 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1799 } else {
1800 src1_alpha[i] = alpha;
1801 }
1802 }
1803 }
1804
1805 if (util_format_is_pure_integer(out_format)) {
1806 /*
1807 * In this case fs_type was really ints or uints disguised as floats,
1808 * fix that up now.
1809 */
1810 fs_type.floating = 0;
1811 fs_type.sign = dst_type.sign;
1812 for (i = 0; i < num_fullblock_fs; ++i) {
1813 for (j = 0; j < dst_channels; ++j) {
1814 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1815 lp_build_vec_type(gallivm, fs_type), "");
1816 }
1817 if (dst_channels == 3 && !has_alpha) {
1818 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1819 lp_build_vec_type(gallivm, fs_type), "");
1820 }
1821 }
1822 }
1823
1824 /*
1825 * Pixel twiddle from fragment shader order to memory order
1826 */
1827 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
1828 dst_channels, fs_src, src, pad_inline);
1829 if (dual_source_blend) {
1830 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
1831 fs_src1, src1, pad_inline);
1832 }
1833
1834 src_channels = dst_channels < 3 ? dst_channels : 4;
1835 if (src_count != num_fullblock_fs * src_channels) {
1836 unsigned ds = src_count / (num_fullblock_fs * src_channels);
1837 row_type.length /= ds;
1838 fs_type.length = row_type.length;
1839 }
1840
1841 blend_type = row_type;
1842 mask_type.length = 4;
1843
1844 /* Convert src to row_type */
1845 if (dual_source_blend) {
1846 struct lp_type old_row_type = row_type;
1847 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1848 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1849 }
1850 else {
1851 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1852 }
1853
1854 /* If the rows are not an SSE vector, combine them to become SSE size! */
1855 if ((row_type.width * row_type.length) % 128) {
1856 unsigned bits = row_type.width * row_type.length;
1857 unsigned combined;
1858
1859 assert(src_count >= (vector_width / bits));
1860
1861 dst_count = src_count / (vector_width / bits);
1862
1863 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1864 if (dual_source_blend) {
1865 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1866 }
1867
1868 row_type.length *= combined;
1869 src_count /= combined;
1870
1871 bits = row_type.width * row_type.length;
1872 assert(bits == 128 || bits == 256);
1873 }
1874
1875
1876 /*
1877 * Blend Colour conversion
1878 */
1879 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1880 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1881 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1882
1883 /* Convert */
1884 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1885
1886 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1887 /*
1888 * since blending is done with floats, there was no conversion.
1889 * However, the rules according to fixed point renderbuffers still
1890 * apply, that is we must clamp inputs to 0.0/1.0.
1891 * (This would apply to separate alpha conversion too but we currently
1892 * force has_alpha to be true.)
1893 * TODO: should skip this with "fake" blend, since post-blend conversion
1894 * will clamp anyway.
1895 * TODO: could also skip this if fragment color clamping is enabled. We
1896 * don't support it natively so it gets baked into the shader however, so
1897 * can't really tell here.
1898 */
1899 struct lp_build_context f32_bld;
1900 assert(row_type.floating);
1901 lp_build_context_init(&f32_bld, gallivm, row_type);
1902 for (i = 0; i < src_count; i++) {
1903 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
1904 }
1905 if (dual_source_blend) {
1906 for (i = 0; i < src_count; i++) {
1907 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
1908 }
1909 }
1910 /* probably can't be different than row_type but better safe than sorry... */
1911 lp_build_context_init(&f32_bld, gallivm, blend_type);
1912 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
1913 }
1914
1915 /* Extract alpha */
1916 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1917
1918 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1919 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1920 if (pad_inline) {
1921 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1922 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1923 } else {
1924 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1925 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1926 }
1927
1928 /*
1929 * Mask conversion
1930 */
1931 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
1932
1933 if (src_count < block_height) {
1934 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1935 } else if (src_count > block_height) {
1936 for (i = src_count; i > 0; --i) {
1937 unsigned pixels = block_size / src_count;
1938 unsigned idx = i - 1;
1939
1940 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
1941 (idx * pixels) % 4, pixels);
1942 }
1943 }
1944
1945 assert(mask_type.width == 32);
1946
1947 for (i = 0; i < src_count; ++i) {
1948 unsigned pixels = block_size / src_count;
1949 unsigned pixel_width = row_type.width * dst_channels;
1950
1951 if (pixel_width == 24) {
1952 mask_type.width = 8;
1953 mask_type.length = vector_width / mask_type.width;
1954 } else {
1955 mask_type.length = pixels;
1956 mask_type.width = row_type.width * dst_channels;
1957
1958 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1959
1960 mask_type.length *= dst_channels;
1961 mask_type.width /= dst_channels;
1962 }
1963
1964 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1965 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1966 }
1967
1968 /*
1969 * Alpha conversion
1970 */
1971 if (!has_alpha) {
1972 struct lp_type alpha_type = fs_type;
1973 alpha_type.length = 4;
1974 convert_alpha(gallivm, row_type, alpha_type,
1975 block_size, block_height,
1976 src_count, dst_channels,
1977 pad_inline, src_alpha);
1978 if (dual_source_blend) {
1979 convert_alpha(gallivm, row_type, alpha_type,
1980 block_size, block_height,
1981 src_count, dst_channels,
1982 pad_inline, src1_alpha);
1983 }
1984 }
1985
1986
1987 /*
1988 * Load dst from memory
1989 */
1990 if (src_count < block_height) {
1991 dst_count = block_height;
1992 } else {
1993 dst_count = src_count;
1994 }
1995
1996 dst_type.length *= block_size / dst_count;
1997
1998 if (format_expands_to_float_soa(out_format_desc)) {
1999 /*
2000 * we need multiple values at once for the conversion, so can as well
2001 * load them vectorized here too instead of concatenating later.
2002 * (Still need concatenation later for 8-wide vectors).
2003 */
2004 dst_count = block_height;
2005 dst_type.length = block_width;
2006 }
2007
2008 /*
2009 * Compute the alignment of the destination pointer in bytes
2010 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2011 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2012 * 1d tex but can't distinguish here) so need to stick with per-pixel
2013 * alignment in this case.
2014 */
2015 if (is_1d) {
2016 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2017 }
2018 else {
2019 dst_alignment = dst_type.length * dst_type.width / 8;
2020 }
2021 /* Force power-of-two alignment by extracting only the least-significant-bit */
2022 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2023 /*
2024 * Resource base and stride pointers are aligned to 16 bytes, so that's
2025 * the maximum alignment we can guarantee
2026 */
2027 dst_alignment = MIN2(16, dst_alignment);
2028
2029 if (is_1d) {
2030 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2031 dst, dst_type, dst_count / 4, dst_alignment);
2032 for (i = dst_count / 4; i < dst_count; i++) {
2033 dst[i] = lp_build_undef(gallivm, dst_type);
2034 }
2035
2036 }
2037 else {
2038 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2039 dst, dst_type, dst_count, dst_alignment);
2040 }
2041
2042
2043 /*
2044 * Convert from dst/output format to src/blending format.
2045 *
2046 * This is necessary as we can only read 1 row from memory at a time,
2047 * so the minimum dst_count will ever be at this point is 4.
2048 *
2049 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2050 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2051 * on all 16 pixels in that single vector at once.
2052 */
2053 if (dst_count > src_count) {
2054 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
2055 }
2056
2057 /*
2058 * Blending
2059 */
2060 /* XXX this is broken for RGB8 formats -
2061 * they get expanded from 12 to 16 elements (to include alpha)
2062 * by convert_to_blend_type then reduced to 15 instead of 12
2063 * by convert_from_blend_type (a simple fix though breaks A8...).
2064 * R16G16B16 also crashes differently however something going wrong
2065 * inside llvm handling npot vector sizes seemingly.
2066 * It seems some cleanup could be done here (like skipping conversion/blend
2067 * when not needed).
2068 */
2069 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
2070
2071 /*
2072 * FIXME: Really should get logic ops / masks out of generic blend / row
2073 * format. Logic ops will definitely not work on the blend float format
2074 * used for SRGB here and I think OpenGL expects this to work as expected
2075 * (that is incoming values converted to srgb then logic op applied).
2076 */
2077 for (i = 0; i < src_count; ++i) {
2078 dst[i] = lp_build_blend_aos(gallivm,
2079 &variant->key.blend,
2080 out_format,
2081 row_type,
2082 rt,
2083 src[i],
2084 has_alpha ? NULL : src_alpha[i],
2085 src1[i],
2086 has_alpha ? NULL : src1_alpha[i],
2087 dst[i],
2088 partial_mask ? src_mask[i] : NULL,
2089 blend_color,
2090 has_alpha ? NULL : blend_alpha,
2091 swizzle,
2092 pad_inline ? 4 : dst_channels);
2093 }
2094
2095 convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
2096
2097 /* Split the blend rows back to memory rows */
2098 if (dst_count > src_count) {
2099 row_type.length = dst_type.length * (dst_count / src_count);
2100
2101 if (src_count == 1) {
2102 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2103 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2104
2105 row_type.length /= 2;
2106 src_count *= 2;
2107 }
2108
2109 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2110 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2111 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2112 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2113
2114 row_type.length /= 2;
2115 src_count *= 2;
2116 }
2117
2118 /*
2119 * Store blend result to memory
2120 */
2121 if (is_1d) {
2122 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2123 dst, dst_type, dst_count / 4, dst_alignment);
2124 }
2125 else {
2126 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2127 dst, dst_type, dst_count, dst_alignment);
2128 }
2129
2130 if (have_smallfloat_format(dst_type, out_format)) {
2131 lp_build_fpstate_set(gallivm, fpstate);
2132 }
2133
2134 if (do_branch) {
2135 lp_build_mask_end(&mask_ctx);
2136 }
2137 }
2138
2139
2140 /**
2141 * Generate the runtime callable function for the whole fragment pipeline.
2142 * Note that the function which we generate operates on a block of 16
2143 * pixels at at time. The block contains 2x2 quads. Each quad contains
2144 * 2x2 pixels.
2145 */
2146 static void
2147 generate_fragment(struct llvmpipe_context *lp,
2148 struct lp_fragment_shader *shader,
2149 struct lp_fragment_shader_variant *variant,
2150 unsigned partial_mask)
2151 {
2152 struct gallivm_state *gallivm = variant->gallivm;
2153 const struct lp_fragment_shader_variant_key *key = &variant->key;
2154 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2155 char func_name[64];
2156 struct lp_type fs_type;
2157 struct lp_type blend_type;
2158 LLVMTypeRef fs_elem_type;
2159 LLVMTypeRef blend_vec_type;
2160 LLVMTypeRef arg_types[13];
2161 LLVMTypeRef func_type;
2162 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2163 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2164 LLVMValueRef context_ptr;
2165 LLVMValueRef x;
2166 LLVMValueRef y;
2167 LLVMValueRef a0_ptr;
2168 LLVMValueRef dadx_ptr;
2169 LLVMValueRef dady_ptr;
2170 LLVMValueRef color_ptr_ptr;
2171 LLVMValueRef stride_ptr;
2172 LLVMValueRef depth_ptr;
2173 LLVMValueRef depth_stride;
2174 LLVMValueRef mask_input;
2175 LLVMValueRef thread_data_ptr;
2176 LLVMBasicBlockRef block;
2177 LLVMBuilderRef builder;
2178 struct lp_build_sampler_soa *sampler;
2179 struct lp_build_interp_soa_context interp;
2180 LLVMValueRef fs_mask[16 / 4];
2181 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2182 LLVMValueRef function;
2183 LLVMValueRef facing;
2184 unsigned num_fs;
2185 unsigned i;
2186 unsigned chan;
2187 unsigned cbuf;
2188 boolean cbuf0_write_all;
2189 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2190 util_blend_state_is_dual(&key->blend, 0);
2191
2192 assert(lp_native_vector_width / 32 >= 4);
2193
2194 /* Adjust color input interpolation according to flatshade state:
2195 */
2196 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2197 for (i = 0; i < shader->info.base.num_inputs; i++) {
2198 if (inputs[i].interp == LP_INTERP_COLOR) {
2199 if (key->flatshade)
2200 inputs[i].interp = LP_INTERP_CONSTANT;
2201 else
2202 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2203 }
2204 }
2205
2206 /* check if writes to cbuf[0] are to be copied to all cbufs */
2207 cbuf0_write_all =
2208 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2209
2210 /* TODO: actually pick these based on the fs and color buffer
2211 * characteristics. */
2212
2213 memset(&fs_type, 0, sizeof fs_type);
2214 fs_type.floating = TRUE; /* floating point values */
2215 fs_type.sign = TRUE; /* values are signed */
2216 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2217 fs_type.width = 32; /* 32-bit float */
2218 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2219
2220 memset(&blend_type, 0, sizeof blend_type);
2221 blend_type.floating = FALSE; /* values are integers */
2222 blend_type.sign = FALSE; /* values are unsigned */
2223 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2224 blend_type.width = 8; /* 8-bit ubyte values */
2225 blend_type.length = 16; /* 16 elements per vector */
2226
2227 /*
2228 * Generate the function prototype. Any change here must be reflected in
2229 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2230 */
2231
2232 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2233
2234 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2235
2236 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2237 shader->no, variant->no, partial_mask ? "partial" : "whole");
2238
2239 arg_types[0] = variant->jit_context_ptr_type; /* context */
2240 arg_types[1] = int32_type; /* x */
2241 arg_types[2] = int32_type; /* y */
2242 arg_types[3] = int32_type; /* facing */
2243 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2244 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2245 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2246 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2247 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2248 arg_types[9] = int32_type; /* mask_input */
2249 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2250 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2251 arg_types[12] = int32_type; /* depth_stride */
2252
2253 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2254 arg_types, Elements(arg_types), 0);
2255
2256 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2257 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2258
2259 variant->function[partial_mask] = function;
2260
2261 /* XXX: need to propagate noalias down into color param now we are
2262 * passing a pointer-to-pointer?
2263 */
2264 for(i = 0; i < Elements(arg_types); ++i)
2265 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2266 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
2267
2268 context_ptr = LLVMGetParam(function, 0);
2269 x = LLVMGetParam(function, 1);
2270 y = LLVMGetParam(function, 2);
2271 facing = LLVMGetParam(function, 3);
2272 a0_ptr = LLVMGetParam(function, 4);
2273 dadx_ptr = LLVMGetParam(function, 5);
2274 dady_ptr = LLVMGetParam(function, 6);
2275 color_ptr_ptr = LLVMGetParam(function, 7);
2276 depth_ptr = LLVMGetParam(function, 8);
2277 mask_input = LLVMGetParam(function, 9);
2278 thread_data_ptr = LLVMGetParam(function, 10);
2279 stride_ptr = LLVMGetParam(function, 11);
2280 depth_stride = LLVMGetParam(function, 12);
2281
2282 lp_build_name(context_ptr, "context");
2283 lp_build_name(x, "x");
2284 lp_build_name(y, "y");
2285 lp_build_name(a0_ptr, "a0");
2286 lp_build_name(dadx_ptr, "dadx");
2287 lp_build_name(dady_ptr, "dady");
2288 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2289 lp_build_name(depth_ptr, "depth");
2290 lp_build_name(thread_data_ptr, "thread_data");
2291 lp_build_name(mask_input, "mask_input");
2292 lp_build_name(stride_ptr, "stride_ptr");
2293 lp_build_name(depth_stride, "depth_stride");
2294
2295 /*
2296 * Function body
2297 */
2298
2299 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2300 builder = gallivm->builder;
2301 assert(builder);
2302 LLVMPositionBuilderAtEnd(builder, block);
2303
2304 /* code generated texture sampling */
2305 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
2306
2307 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2308 /* for 1d resources only run "upper half" of stamp */
2309 if (key->resource_1d)
2310 num_fs /= 2;
2311
2312 {
2313 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2314 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2315 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2316 num_loop, "mask_store");
2317 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2318 boolean pixel_center_integer =
2319 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2320
2321 /*
2322 * The shader input interpolation info is not explicitely baked in the
2323 * shader key, but everything it derives from (TGSI, and flatshade) is
2324 * already included in the shader key.
2325 */
2326 lp_build_interp_soa_init(&interp,
2327 gallivm,
2328 shader->info.base.num_inputs,
2329 inputs,
2330 pixel_center_integer,
2331 builder, fs_type,
2332 a0_ptr, dadx_ptr, dady_ptr,
2333 x, y);
2334
2335 for (i = 0; i < num_fs; i++) {
2336 LLVMValueRef mask;
2337 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2338 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2339 &indexi, 1, "mask_ptr");
2340
2341 if (partial_mask) {
2342 mask = generate_quad_mask(gallivm, fs_type,
2343 i*fs_type.length/4, mask_input);
2344 }
2345 else {
2346 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2347 }
2348 LLVMBuildStore(builder, mask, mask_ptr);
2349 }
2350
2351 generate_fs_loop(gallivm,
2352 shader, key,
2353 builder,
2354 fs_type,
2355 context_ptr,
2356 num_loop,
2357 &interp,
2358 sampler,
2359 mask_store, /* output */
2360 color_store,
2361 depth_ptr,
2362 depth_stride,
2363 facing,
2364 thread_data_ptr);
2365
2366 for (i = 0; i < num_fs; i++) {
2367 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2368 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2369 &indexi, 1, "");
2370 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2371 /* This is fucked up need to reorganize things */
2372 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2373 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2374 ptr = LLVMBuildGEP(builder,
2375 color_store[cbuf * !cbuf0_write_all][chan],
2376 &indexi, 1, "");
2377 fs_out_color[cbuf][chan][i] = ptr;
2378 }
2379 }
2380 if (dual_source_blend) {
2381 /* only support one dual source blend target hence always use output 1 */
2382 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2383 ptr = LLVMBuildGEP(builder,
2384 color_store[1][chan],
2385 &indexi, 1, "");
2386 fs_out_color[1][chan][i] = ptr;
2387 }
2388 }
2389 }
2390 }
2391
2392 sampler->destroy(sampler);
2393
2394 /* Loop over color outputs / color buffers to do blending.
2395 */
2396 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2397 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2398 LLVMValueRef color_ptr;
2399 LLVMValueRef stride;
2400 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2401
2402 boolean do_branch = ((key->depth.enabled
2403 || key->stencil[0].enabled
2404 || key->alpha.enabled)
2405 && !shader->info.base.uses_kill);
2406
2407 color_ptr = LLVMBuildLoad(builder,
2408 LLVMBuildGEP(builder, color_ptr_ptr,
2409 &index, 1, ""),
2410 "");
2411
2412 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2413
2414 stride = LLVMBuildLoad(builder,
2415 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2416 "");
2417
2418 generate_unswizzled_blend(gallivm, cbuf, variant,
2419 key->cbuf_format[cbuf],
2420 num_fs, fs_type, fs_mask, fs_out_color,
2421 context_ptr, color_ptr, stride,
2422 partial_mask, do_branch);
2423 }
2424 }
2425
2426 LLVMBuildRetVoid(builder);
2427
2428 gallivm_verify_function(gallivm, function);
2429 }
2430
2431
2432 static void
2433 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2434 {
2435 unsigned i;
2436
2437 debug_printf("fs variant %p:\n", (void *) key);
2438
2439 if (key->flatshade) {
2440 debug_printf("flatshade = 1\n");
2441 }
2442 for (i = 0; i < key->nr_cbufs; ++i) {
2443 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2444 }
2445 if (key->depth.enabled) {
2446 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2447 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2448 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2449 }
2450
2451 for (i = 0; i < 2; ++i) {
2452 if (key->stencil[i].enabled) {
2453 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2454 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2455 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2456 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2457 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2458 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2459 }
2460 }
2461
2462 if (key->alpha.enabled) {
2463 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2464 }
2465
2466 if (key->occlusion_count) {
2467 debug_printf("occlusion_count = 1\n");
2468 }
2469
2470 if (key->blend.logicop_enable) {
2471 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2472 }
2473 else if (key->blend.rt[0].blend_enable) {
2474 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2475 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2476 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2477 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2478 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2479 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2480 }
2481 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2482 if (key->blend.alpha_to_coverage) {
2483 debug_printf("blend.alpha_to_coverage is enabled\n");
2484 }
2485 for (i = 0; i < key->nr_samplers; ++i) {
2486 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2487 debug_printf("sampler[%u] = \n", i);
2488 debug_printf(" .wrap = %s %s %s\n",
2489 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2490 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2491 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2492 debug_printf(" .min_img_filter = %s\n",
2493 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2494 debug_printf(" .min_mip_filter = %s\n",
2495 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2496 debug_printf(" .mag_img_filter = %s\n",
2497 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2498 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2499 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2500 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2501 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2502 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2503 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2504 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2505 }
2506 for (i = 0; i < key->nr_sampler_views; ++i) {
2507 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2508 debug_printf("texture[%u] = \n", i);
2509 debug_printf(" .format = %s\n",
2510 util_format_name(texture->format));
2511 debug_printf(" .target = %s\n",
2512 util_dump_tex_target(texture->target, TRUE));
2513 debug_printf(" .level_zero_only = %u\n",
2514 texture->level_zero_only);
2515 debug_printf(" .pot = %u %u %u\n",
2516 texture->pot_width,
2517 texture->pot_height,
2518 texture->pot_depth);
2519 }
2520 }
2521
2522
2523 void
2524 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2525 {
2526 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2527 variant->shader->no, variant->no);
2528 tgsi_dump(variant->shader->base.tokens, 0);
2529 dump_fs_variant_key(&variant->key);
2530 debug_printf("variant->opaque = %u\n", variant->opaque);
2531 debug_printf("\n");
2532 }
2533
2534
2535 /**
2536 * Generate a new fragment shader variant from the shader code and
2537 * other state indicated by the key.
2538 */
2539 static struct lp_fragment_shader_variant *
2540 generate_variant(struct llvmpipe_context *lp,
2541 struct lp_fragment_shader *shader,
2542 const struct lp_fragment_shader_variant_key *key)
2543 {
2544 struct lp_fragment_shader_variant *variant;
2545 const struct util_format_description *cbuf0_format_desc;
2546 boolean fullcolormask;
2547 char module_name[64];
2548
2549 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2550 if(!variant)
2551 return NULL;
2552
2553 util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
2554 shader->no, shader->variants_created);
2555
2556 variant->gallivm = gallivm_create(module_name, lp->context);
2557 if (!variant->gallivm) {
2558 FREE(variant);
2559 return NULL;
2560 }
2561
2562 variant->shader = shader;
2563 variant->list_item_global.base = variant;
2564 variant->list_item_local.base = variant;
2565 variant->no = shader->variants_created++;
2566
2567 memcpy(&variant->key, key, shader->variant_key_size);
2568
2569 /*
2570 * Determine whether we are touching all channels in the color buffer.
2571 */
2572 fullcolormask = FALSE;
2573 if (key->nr_cbufs == 1) {
2574 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2575 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2576 }
2577
2578 variant->opaque =
2579 !key->blend.logicop_enable &&
2580 !key->blend.rt[0].blend_enable &&
2581 fullcolormask &&
2582 !key->stencil[0].enabled &&
2583 !key->alpha.enabled &&
2584 !key->blend.alpha_to_coverage &&
2585 !key->depth.enabled &&
2586 !shader->info.base.uses_kill
2587 ? TRUE : FALSE;
2588
2589 if ((shader->info.base.num_tokens <= 1) &&
2590 !key->depth.enabled && !key->stencil[0].enabled) {
2591 variant->ps_inv_multiplier = 0;
2592 } else {
2593 variant->ps_inv_multiplier = 1;
2594 }
2595
2596 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2597 lp_debug_fs_variant(variant);
2598 }
2599
2600 lp_jit_init_types(variant);
2601
2602 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2603 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2604
2605 if (variant->jit_function[RAST_WHOLE] == NULL) {
2606 if (variant->opaque) {
2607 /* Specialized shader, which doesn't need to read the color buffer. */
2608 generate_fragment(lp, shader, variant, RAST_WHOLE);
2609 }
2610 }
2611
2612 /*
2613 * Compile everything
2614 */
2615
2616 gallivm_compile_module(variant->gallivm);
2617
2618 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
2619
2620 if (variant->function[RAST_EDGE_TEST]) {
2621 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2622 gallivm_jit_function(variant->gallivm,
2623 variant->function[RAST_EDGE_TEST]);
2624 }
2625
2626 if (variant->function[RAST_WHOLE]) {
2627 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2628 gallivm_jit_function(variant->gallivm,
2629 variant->function[RAST_WHOLE]);
2630 } else if (!variant->jit_function[RAST_WHOLE]) {
2631 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2632 }
2633
2634 gallivm_free_ir(variant->gallivm);
2635
2636 return variant;
2637 }
2638
2639
2640 static void *
2641 llvmpipe_create_fs_state(struct pipe_context *pipe,
2642 const struct pipe_shader_state *templ)
2643 {
2644 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2645 struct lp_fragment_shader *shader;
2646 int nr_samplers;
2647 int nr_sampler_views;
2648 int i;
2649
2650 shader = CALLOC_STRUCT(lp_fragment_shader);
2651 if (!shader)
2652 return NULL;
2653
2654 shader->no = fs_no++;
2655 make_empty_list(&shader->variants);
2656
2657 /* get/save the summary info for this shader */
2658 lp_build_tgsi_info(templ->tokens, &shader->info);
2659
2660 /* we need to keep a local copy of the tokens */
2661 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2662
2663 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2664 if (shader->draw_data == NULL) {
2665 FREE((void *) shader->base.tokens);
2666 FREE(shader);
2667 return NULL;
2668 }
2669
2670 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2671 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2672
2673 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2674 state[MAX2(nr_samplers, nr_sampler_views)]);
2675
2676 for (i = 0; i < shader->info.base.num_inputs; i++) {
2677 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2678 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2679
2680 switch (shader->info.base.input_interpolate[i]) {
2681 case TGSI_INTERPOLATE_CONSTANT:
2682 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2683 break;
2684 case TGSI_INTERPOLATE_LINEAR:
2685 shader->inputs[i].interp = LP_INTERP_LINEAR;
2686 break;
2687 case TGSI_INTERPOLATE_PERSPECTIVE:
2688 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2689 break;
2690 case TGSI_INTERPOLATE_COLOR:
2691 shader->inputs[i].interp = LP_INTERP_COLOR;
2692 break;
2693 default:
2694 assert(0);
2695 break;
2696 }
2697
2698 switch (shader->info.base.input_semantic_name[i]) {
2699 case TGSI_SEMANTIC_FACE:
2700 shader->inputs[i].interp = LP_INTERP_FACING;
2701 break;
2702 case TGSI_SEMANTIC_POSITION:
2703 /* Position was already emitted above
2704 */
2705 shader->inputs[i].interp = LP_INTERP_POSITION;
2706 shader->inputs[i].src_index = 0;
2707 continue;
2708 }
2709
2710 shader->inputs[i].src_index = i+1;
2711 }
2712
2713 if (LP_DEBUG & DEBUG_TGSI) {
2714 unsigned attrib;
2715 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2716 shader->no, (void *) shader);
2717 tgsi_dump(templ->tokens, 0);
2718 debug_printf("usage masks:\n");
2719 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2720 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2721 debug_printf(" IN[%u].%s%s%s%s\n",
2722 attrib,
2723 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2724 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2725 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2726 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2727 }
2728 debug_printf("\n");
2729 }
2730
2731 return shader;
2732 }
2733
2734
2735 static void
2736 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2737 {
2738 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2739
2740 if (llvmpipe->fs == fs)
2741 return;
2742
2743 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2744
2745 draw_bind_fragment_shader(llvmpipe->draw,
2746 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2747
2748 llvmpipe->dirty |= LP_NEW_FS;
2749 }
2750
2751
2752 /**
2753 * Remove shader variant from two lists: the shader's variant list
2754 * and the context's variant list.
2755 */
2756 void
2757 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2758 struct lp_fragment_shader_variant *variant)
2759 {
2760 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2761 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2762 " #%u v total cached #%u\n",
2763 variant->shader->no,
2764 variant->no,
2765 variant->shader->variants_created,
2766 variant->shader->variants_cached,
2767 lp->nr_fs_variants);
2768 }
2769
2770 gallivm_destroy(variant->gallivm);
2771
2772 /* remove from shader's list */
2773 remove_from_list(&variant->list_item_local);
2774 variant->shader->variants_cached--;
2775
2776 /* remove from context's list */
2777 remove_from_list(&variant->list_item_global);
2778 lp->nr_fs_variants--;
2779 lp->nr_fs_instrs -= variant->nr_instrs;
2780
2781 FREE(variant);
2782 }
2783
2784
2785 static void
2786 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2787 {
2788 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2789 struct lp_fragment_shader *shader = fs;
2790 struct lp_fs_variant_list_item *li;
2791
2792 assert(fs != llvmpipe->fs);
2793
2794 /*
2795 * XXX: we need to flush the context until we have some sort of reference
2796 * counting in fragment shaders as they may still be binned
2797 * Flushing alone might not sufficient we need to wait on it too.
2798 */
2799 llvmpipe_finish(pipe, __FUNCTION__);
2800
2801 /* Delete all the variants */
2802 li = first_elem(&shader->variants);
2803 while(!at_end(&shader->variants, li)) {
2804 struct lp_fs_variant_list_item *next = next_elem(li);
2805 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2806 li = next;
2807 }
2808
2809 /* Delete draw module's data */
2810 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2811
2812 assert(shader->variants_cached == 0);
2813 FREE((void *) shader->base.tokens);
2814 FREE(shader);
2815 }
2816
2817
2818
2819 static void
2820 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2821 uint shader, uint index,
2822 struct pipe_constant_buffer *cb)
2823 {
2824 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2825 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2826
2827 assert(shader < PIPE_SHADER_TYPES);
2828 assert(index < Elements(llvmpipe->constants[shader]));
2829
2830 /* note: reference counting */
2831 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2832
2833 if (shader == PIPE_SHADER_VERTEX ||
2834 shader == PIPE_SHADER_GEOMETRY) {
2835 /* Pass the constants to the 'draw' module */
2836 const unsigned size = cb ? cb->buffer_size : 0;
2837 const ubyte *data;
2838
2839 if (constants) {
2840 data = (ubyte *) llvmpipe_resource_data(constants);
2841 }
2842 else if (cb && cb->user_buffer) {
2843 data = (ubyte *) cb->user_buffer;
2844 }
2845 else {
2846 data = NULL;
2847 }
2848
2849 if (data)
2850 data += cb->buffer_offset;
2851
2852 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2853 index, data, size);
2854 }
2855
2856 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2857
2858 if (cb && cb->user_buffer) {
2859 pipe_resource_reference(&constants, NULL);
2860 }
2861 }
2862
2863
2864 /**
2865 * Return the blend factor equivalent to a destination alpha of one.
2866 */
2867 static INLINE unsigned
2868 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
2869 {
2870 switch(factor) {
2871 case PIPE_BLENDFACTOR_DST_ALPHA:
2872 return PIPE_BLENDFACTOR_ONE;
2873 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2874 return PIPE_BLENDFACTOR_ZERO;
2875 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2876 if (clamped_zero)
2877 return PIPE_BLENDFACTOR_ZERO;
2878 else
2879 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
2880 }
2881
2882 return factor;
2883 }
2884
2885
2886 /**
2887 * We need to generate several variants of the fragment pipeline to match
2888 * all the combinations of the contributing state atoms.
2889 *
2890 * TODO: there is actually no reason to tie this to context state -- the
2891 * generated code could be cached globally in the screen.
2892 */
2893 static void
2894 make_variant_key(struct llvmpipe_context *lp,
2895 struct lp_fragment_shader *shader,
2896 struct lp_fragment_shader_variant_key *key)
2897 {
2898 unsigned i;
2899
2900 memset(key, 0, shader->variant_key_size);
2901
2902 if (lp->framebuffer.zsbuf) {
2903 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2904 const struct util_format_description *zsbuf_desc =
2905 util_format_description(zsbuf_format);
2906
2907 if (lp->depth_stencil->depth.enabled &&
2908 util_format_has_depth(zsbuf_desc)) {
2909 key->zsbuf_format = zsbuf_format;
2910 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2911 }
2912 if (lp->depth_stencil->stencil[0].enabled &&
2913 util_format_has_stencil(zsbuf_desc)) {
2914 key->zsbuf_format = zsbuf_format;
2915 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2916 }
2917 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
2918 key->resource_1d = TRUE;
2919 }
2920 }
2921
2922 /*
2923 * Propagate the depth clamp setting from the rasterizer state.
2924 * depth_clip == 0 implies depth clamping is enabled.
2925 *
2926 * When clip_halfz is enabled, then always clamp the depth values.
2927 */
2928 if (lp->rasterizer->clip_halfz) {
2929 key->depth_clamp = 1;
2930 } else {
2931 key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
2932 }
2933
2934 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2935 if (!lp->framebuffer.nr_cbufs ||
2936 !lp->framebuffer.cbufs[0] ||
2937 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2938 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2939 }
2940 if(key->alpha.enabled)
2941 key->alpha.func = lp->depth_stencil->alpha.func;
2942 /* alpha.ref_value is passed in jit_context */
2943
2944 key->flatshade = lp->rasterizer->flatshade;
2945 if (lp->active_occlusion_queries) {
2946 key->occlusion_count = TRUE;
2947 }
2948
2949 if (lp->framebuffer.nr_cbufs) {
2950 memcpy(&key->blend, lp->blend, sizeof key->blend);
2951 }
2952
2953 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2954
2955 if (!key->blend.independent_blend_enable) {
2956 /* we always need independent blend otherwise the fixups below won't work */
2957 for (i = 1; i < key->nr_cbufs; i++) {
2958 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2959 }
2960 key->blend.independent_blend_enable = 1;
2961 }
2962
2963 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2964 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2965
2966 if (lp->framebuffer.cbufs[i]) {
2967 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2968 const struct util_format_description *format_desc;
2969
2970 key->cbuf_format[i] = format;
2971
2972 /*
2973 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2974 * mixing of 2d textures with height 1 and 1d textures, so make sure
2975 * we pick 1d if any cbuf or zsbuf is 1d.
2976 */
2977 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
2978 key->resource_1d = TRUE;
2979 }
2980
2981 format_desc = util_format_description(format);
2982 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2983 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2984
2985 /*
2986 * Mask out color channels not present in the color buffer.
2987 */
2988 blend_rt->colormask &= util_format_colormask(format_desc);
2989
2990 /*
2991 * Disable blend for integer formats.
2992 */
2993 if (util_format_is_pure_integer(format)) {
2994 blend_rt->blend_enable = 0;
2995 }
2996
2997 /*
2998 * Our swizzled render tiles always have an alpha channel, but the
2999 * linear render target format often does not, so force here the dst
3000 * alpha to be one.
3001 *
3002 * This is not a mere optimization. Wrong results will be produced if
3003 * the dst alpha is used, the dst format does not have alpha, and the
3004 * previous rendering was not flushed from the swizzled to linear
3005 * buffer. For example, NonPowTwo DCT.
3006 *
3007 * TODO: This should be generalized to all channels for better
3008 * performance, but only alpha causes correctness issues.
3009 *
3010 * Also, force rgb/alpha func/factors match, to make AoS blending
3011 * easier.
3012 */
3013 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
3014 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3015 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3016 boolean clamped_zero = !util_format_is_float(format) &&
3017 !util_format_is_snorm(format);
3018 blend_rt->rgb_src_factor =
3019 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3020 blend_rt->rgb_dst_factor =
3021 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3022 blend_rt->alpha_func = blend_rt->rgb_func;
3023 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3024 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3025 }
3026 }
3027 else {
3028 /* no color buffer for this fragment output */
3029 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3030 blend_rt->colormask = 0x0;
3031 blend_rt->blend_enable = 0;
3032 }
3033 }
3034
3035 /* This value will be the same for all the variants of a given shader:
3036 */
3037 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3038
3039 for(i = 0; i < key->nr_samplers; ++i) {
3040 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3041 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
3042 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3043 }
3044 }
3045
3046 /*
3047 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3048 * are dx10-style? Can't really have mixed opcodes, at least not
3049 * if we want to skip the holes here (without rescanning tgsi).
3050 */
3051 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3052 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3053 for(i = 0; i < key->nr_sampler_views; ++i) {
3054 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
3055 lp_sampler_static_texture_state(&key->state[i].texture_state,
3056 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3057 }
3058 }
3059 }
3060 else {
3061 key->nr_sampler_views = key->nr_samplers;
3062 for(i = 0; i < key->nr_sampler_views; ++i) {
3063 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3064 lp_sampler_static_texture_state(&key->state[i].texture_state,
3065 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3066 }
3067 }
3068 }
3069 }
3070
3071
3072
3073 /**
3074 * Update fragment shader state. This is called just prior to drawing
3075 * something when some fragment-related state has changed.
3076 */
3077 void
3078 llvmpipe_update_fs(struct llvmpipe_context *lp)
3079 {
3080 struct lp_fragment_shader *shader = lp->fs;
3081 struct lp_fragment_shader_variant_key key;
3082 struct lp_fragment_shader_variant *variant = NULL;
3083 struct lp_fs_variant_list_item *li;
3084
3085 make_variant_key(lp, shader, &key);
3086
3087 /* Search the variants for one which matches the key */
3088 li = first_elem(&shader->variants);
3089 while(!at_end(&shader->variants, li)) {
3090 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
3091 variant = li->base;
3092 break;
3093 }
3094 li = next_elem(li);
3095 }
3096
3097 if (variant) {
3098 /* Move this variant to the head of the list to implement LRU
3099 * deletion of shader's when we have too many.
3100 */
3101 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3102 }
3103 else {
3104 /* variant not found, create it now */
3105 int64_t t0, t1, dt;
3106 unsigned i;
3107 unsigned variants_to_cull;
3108
3109 if (0) {
3110 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3111 lp->nr_fs_variants,
3112 lp->nr_fs_instrs,
3113 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3114 }
3115
3116 /* First, check if we've exceeded the max number of shader variants.
3117 * If so, free 25% of them (the least recently used ones).
3118 */
3119 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
3120
3121 if (variants_to_cull ||
3122 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3123 struct pipe_context *pipe = &lp->pipe;
3124
3125 /*
3126 * XXX: we need to flush the context until we have some sort of
3127 * reference counting in fragment shaders as they may still be binned
3128 * Flushing alone might not be sufficient we need to wait on it too.
3129 */
3130 llvmpipe_finish(pipe, __FUNCTION__);
3131
3132 /*
3133 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3134 * number of shader variants (potentially all of them) could be
3135 * pending for destruction on flush.
3136 */
3137
3138 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3139 struct lp_fs_variant_list_item *item;
3140 if (is_empty_list(&lp->fs_variants_list)) {
3141 break;
3142 }
3143 item = last_elem(&lp->fs_variants_list);
3144 assert(item);
3145 assert(item->base);
3146 llvmpipe_remove_shader_variant(lp, item->base);
3147 }
3148 }
3149
3150 /*
3151 * Generate the new variant.
3152 */
3153 t0 = os_time_get();
3154 variant = generate_variant(lp, shader, &key);
3155 t1 = os_time_get();
3156 dt = t1 - t0;
3157 LP_COUNT_ADD(llvm_compile_time, dt);
3158 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3159
3160 /* Put the new variant into the list */
3161 if (variant) {
3162 insert_at_head(&shader->variants, &variant->list_item_local);
3163 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3164 lp->nr_fs_variants++;
3165 lp->nr_fs_instrs += variant->nr_instrs;
3166 shader->variants_cached++;
3167 }
3168 }
3169
3170 /* Bind this variant */
3171 lp_setup_set_fs_variant(lp->setup, variant);
3172 }
3173
3174
3175
3176
3177
3178 void
3179 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3180 {
3181 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3182 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3183 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3184
3185 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3186 }
3187
3188 /*
3189 * Rasterization is disabled if there is no pixel shader and
3190 * both depth and stencil testing are disabled:
3191 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3192 */
3193 boolean
3194 llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
3195 {
3196 boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
3197
3198 return (null_fs &&
3199 !lp->depth_stencil->depth.enabled &&
3200 !lp->depth_stencil->stencil[0].enabled);
3201 }