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/simple_list.h"
69 #include "util/u_dual_blend.h"
70 #include "util/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_bitarit.h"
88 #include "gallivm/lp_bld_pack.h"
89 #include "gallivm/lp_bld_format.h"
90 #include "gallivm/lp_bld_quad.h"
91
92 #include "lp_bld_alpha.h"
93 #include "lp_bld_blend.h"
94 #include "lp_bld_depth.h"
95 #include "lp_bld_interp.h"
96 #include "lp_context.h"
97 #include "lp_debug.h"
98 #include "lp_perf.h"
99 #include "lp_setup.h"
100 #include "lp_state.h"
101 #include "lp_tex_sample.h"
102 #include "lp_flush.h"
103 #include "lp_state_fs.h"
104 #include "lp_rast.h"
105
106
107 /** Fragment shader number (for debugging) */
108 static unsigned fs_no = 0;
109
110
111 /**
112 * Expand the relevant bits of mask_input to a n*4-dword mask for the
113 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
114 * quad mask vector to 0 or ~0.
115 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
116 * quad arguments with fs length 8.
117 *
118 * \param first_quad which quad(s) of the quad group to test, in [0,3]
119 * \param mask_input bitwise mask for the whole 4x4 stamp
120 */
121 static LLVMValueRef
generate_quad_mask(struct gallivm_state * gallivm,struct lp_type fs_type,unsigned first_quad,LLVMValueRef mask_input)122 generate_quad_mask(struct gallivm_state *gallivm,
123 struct lp_type fs_type,
124 unsigned first_quad,
125 LLVMValueRef mask_input) /* int32 */
126 {
127 LLVMBuilderRef builder = gallivm->builder;
128 struct lp_type mask_type;
129 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
130 LLVMValueRef bits[16];
131 LLVMValueRef mask, bits_vec;
132 int shift, i;
133
134 /*
135 * XXX: We'll need a different path for 16 x u8
136 */
137 assert(fs_type.width == 32);
138 assert(fs_type.length <= ARRAY_SIZE(bits));
139 mask_type = lp_int_type(fs_type);
140
141 /*
142 * mask_input >>= (quad * 4)
143 */
144 switch (first_quad) {
145 case 0:
146 shift = 0;
147 break;
148 case 1:
149 assert(fs_type.length == 4);
150 shift = 2;
151 break;
152 case 2:
153 shift = 8;
154 break;
155 case 3:
156 assert(fs_type.length == 4);
157 shift = 10;
158 break;
159 default:
160 assert(0);
161 shift = 0;
162 }
163
164 mask_input = LLVMBuildLShr(builder,
165 mask_input,
166 LLVMConstInt(i32t, shift, 0),
167 "");
168
169 /*
170 * mask = { mask_input & (1 << i), for i in [0,3] }
171 */
172 mask = lp_build_broadcast(gallivm,
173 lp_build_vec_type(gallivm, mask_type),
174 mask_input);
175
176 for (i = 0; i < fs_type.length / 4; i++) {
177 unsigned j = 2 * (i % 2) + (i / 2) * 8;
178 bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
179 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
180 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
181 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
182 }
183 bits_vec = LLVMConstVector(bits, fs_type.length);
184 mask = LLVMBuildAnd(builder, mask, bits_vec, "");
185
186 /*
187 * mask = mask == bits ? ~0 : 0
188 */
189 mask = lp_build_compare(gallivm,
190 mask_type, PIPE_FUNC_EQUAL,
191 mask, bits_vec);
192
193 return mask;
194 }
195
196
197 #define EARLY_DEPTH_TEST 0x1
198 #define LATE_DEPTH_TEST 0x2
199 #define EARLY_DEPTH_WRITE 0x4
200 #define LATE_DEPTH_WRITE 0x8
201
202 static int
find_output_by_semantic(const struct tgsi_shader_info * info,unsigned semantic,unsigned index)203 find_output_by_semantic( const struct tgsi_shader_info *info,
204 unsigned semantic,
205 unsigned index )
206 {
207 int i;
208
209 for (i = 0; i < info->num_outputs; i++)
210 if (info->output_semantic_name[i] == semantic &&
211 info->output_semantic_index[i] == index)
212 return i;
213
214 return -1;
215 }
216
217
218 /**
219 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
220 */
221 static LLVMValueRef
lp_llvm_viewport(LLVMValueRef context_ptr,struct gallivm_state * gallivm,LLVMValueRef viewport_index)222 lp_llvm_viewport(LLVMValueRef context_ptr,
223 struct gallivm_state *gallivm,
224 LLVMValueRef viewport_index)
225 {
226 LLVMBuilderRef builder = gallivm->builder;
227 LLVMValueRef ptr;
228 LLVMValueRef res;
229 struct lp_type viewport_type =
230 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
231
232 ptr = lp_jit_context_viewports(gallivm, context_ptr);
233 ptr = LLVMBuildPointerCast(builder, ptr,
234 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
235
236 res = lp_build_pointer_get(builder, ptr, viewport_index);
237
238 return res;
239 }
240
241
242 static LLVMValueRef
lp_build_depth_clamp(struct gallivm_state * gallivm,LLVMBuilderRef builder,struct lp_type type,LLVMValueRef context_ptr,LLVMValueRef thread_data_ptr,LLVMValueRef z)243 lp_build_depth_clamp(struct gallivm_state *gallivm,
244 LLVMBuilderRef builder,
245 struct lp_type type,
246 LLVMValueRef context_ptr,
247 LLVMValueRef thread_data_ptr,
248 LLVMValueRef z)
249 {
250 LLVMValueRef viewport, min_depth, max_depth;
251 LLVMValueRef viewport_index;
252 struct lp_build_context f32_bld;
253
254 assert(type.floating);
255 lp_build_context_init(&f32_bld, gallivm, type);
256
257 /*
258 * Assumes clamping of the viewport index will occur in setup/gs. Value
259 * is passed through the rasterization stage via lp_rast_shader_inputs.
260 *
261 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
262 * semantics.
263 */
264 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
265 thread_data_ptr);
266
267 /*
268 * Load the min and max depth from the lp_jit_context.viewports
269 * array of lp_jit_viewport structures.
270 */
271 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
272
273 /* viewports[viewport_index].min_depth */
274 min_depth = LLVMBuildExtractElement(builder, viewport,
275 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
276 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
277
278 /* viewports[viewport_index].max_depth */
279 max_depth = LLVMBuildExtractElement(builder, viewport,
280 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
281 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
282
283 /*
284 * Clamp to the min and max depth values for the given viewport.
285 */
286 return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
287 }
288
289
290 /**
291 * Generate the fragment shader, depth/stencil test, and alpha tests.
292 */
293 static void
generate_fs_loop(struct gallivm_state * gallivm,struct lp_fragment_shader * shader,const struct lp_fragment_shader_variant_key * key,LLVMBuilderRef builder,struct lp_type type,LLVMValueRef context_ptr,LLVMValueRef num_loop,struct lp_build_interp_soa_context * interp,struct lp_build_sampler_soa * sampler,LLVMValueRef mask_store,LLVMValueRef (* out_color)[4],LLVMValueRef depth_ptr,LLVMValueRef depth_stride,LLVMValueRef facing,LLVMValueRef thread_data_ptr)294 generate_fs_loop(struct gallivm_state *gallivm,
295 struct lp_fragment_shader *shader,
296 const struct lp_fragment_shader_variant_key *key,
297 LLVMBuilderRef builder,
298 struct lp_type type,
299 LLVMValueRef context_ptr,
300 LLVMValueRef num_loop,
301 struct lp_build_interp_soa_context *interp,
302 struct lp_build_sampler_soa *sampler,
303 LLVMValueRef mask_store,
304 LLVMValueRef (*out_color)[4],
305 LLVMValueRef depth_ptr,
306 LLVMValueRef depth_stride,
307 LLVMValueRef facing,
308 LLVMValueRef thread_data_ptr)
309 {
310 const struct util_format_description *zs_format_desc = NULL;
311 const struct tgsi_token *tokens = shader->base.tokens;
312 struct lp_type int_type = lp_int_type(type);
313 LLVMTypeRef vec_type, int_vec_type;
314 LLVMValueRef mask_ptr, mask_val;
315 LLVMValueRef consts_ptr, num_consts_ptr;
316 LLVMValueRef z;
317 LLVMValueRef z_value, s_value;
318 LLVMValueRef z_fb, s_fb;
319 LLVMValueRef stencil_refs[2];
320 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
321 struct lp_build_for_loop_state loop_state;
322 struct lp_build_mask_context mask;
323 /*
324 * TODO: figure out if simple_shader optimization is really worthwile to
325 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
326 * code since tests tend to take another codepath than real shaders.
327 */
328 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
329 shader->info.base.num_inputs < 3 &&
330 shader->info.base.num_instructions < 8) && 0;
331 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
332 util_blend_state_is_dual(&key->blend, 0);
333 unsigned attrib;
334 unsigned chan;
335 unsigned cbuf;
336 unsigned depth_mode;
337
338 struct lp_bld_tgsi_system_values system_values;
339
340 memset(&system_values, 0, sizeof(system_values));
341
342 if (key->depth.enabled ||
343 key->stencil[0].enabled) {
344
345 zs_format_desc = util_format_description(key->zsbuf_format);
346 assert(zs_format_desc);
347
348 if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
349 if (key->alpha.enabled ||
350 key->blend.alpha_to_coverage ||
351 shader->info.base.uses_kill ||
352 shader->info.base.writes_samplemask) {
353 /* With alpha test and kill, can do the depth test early
354 * and hopefully eliminate some quads. But need to do a
355 * special deferred depth write once the final mask value
356 * is known. This only works though if there's either no
357 * stencil test or the stencil value isn't written.
358 */
359 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
360 (key->stencil[1].enabled &&
361 key->stencil[1].writemask)))
362 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
363 else
364 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
365 }
366 else
367 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
368 }
369 else {
370 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
371 }
372
373 if (!(key->depth.enabled && key->depth.writemask) &&
374 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
375 (key->stencil[1].enabled &&
376 key->stencil[1].writemask))))
377 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
378 }
379 else {
380 depth_mode = 0;
381 }
382
383 vec_type = lp_build_vec_type(gallivm, type);
384 int_vec_type = lp_build_vec_type(gallivm, int_type);
385
386 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
387 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
388 /* convert scalar stencil refs into vectors */
389 stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
390 stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
391
392 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
393 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
394
395 lp_build_for_loop_begin(&loop_state, gallivm,
396 lp_build_const_int32(gallivm, 0),
397 LLVMIntULT,
398 num_loop,
399 lp_build_const_int32(gallivm, 1));
400
401 mask_ptr = LLVMBuildGEP(builder, mask_store,
402 &loop_state.counter, 1, "mask_ptr");
403 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
404
405 memset(outputs, 0, sizeof outputs);
406
407 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
408 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
409 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
410 lp_build_vec_type(gallivm,
411 type),
412 num_loop, "color");
413 }
414 }
415 if (dual_source_blend) {
416 assert(key->nr_cbufs <= 1);
417 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
418 out_color[1][chan] = lp_build_array_alloca(gallivm,
419 lp_build_vec_type(gallivm,
420 type),
421 num_loop, "color1");
422 }
423 }
424
425
426 /* 'mask' will control execution based on quad's pixel alive/killed state */
427 lp_build_mask_begin(&mask, gallivm, type, mask_val);
428
429 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
430 lp_build_mask_check(&mask);
431
432 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
433 z = interp->pos[2];
434
435 if (depth_mode & EARLY_DEPTH_TEST) {
436 /*
437 * Clamp according to ARB_depth_clamp semantics.
438 */
439 if (key->depth_clamp) {
440 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
441 thread_data_ptr, z);
442 }
443 lp_build_depth_stencil_load_swizzled(gallivm, type,
444 zs_format_desc, key->resource_1d,
445 depth_ptr, depth_stride,
446 &z_fb, &s_fb, loop_state.counter);
447 lp_build_depth_stencil_test(gallivm,
448 &key->depth,
449 key->stencil,
450 type,
451 zs_format_desc,
452 &mask,
453 stencil_refs,
454 z, z_fb, s_fb,
455 facing,
456 &z_value, &s_value,
457 !simple_shader);
458
459 if (depth_mode & EARLY_DEPTH_WRITE) {
460 lp_build_depth_stencil_write_swizzled(gallivm, type,
461 zs_format_desc, key->resource_1d,
462 NULL, NULL, NULL, loop_state.counter,
463 depth_ptr, depth_stride,
464 z_value, s_value);
465 }
466 /*
467 * Note mask check if stencil is enabled must be after ds write not after
468 * stencil test otherwise new stencil values may not get written if all
469 * fragments got killed by depth/stencil test.
470 */
471 if (!simple_shader && key->stencil[0].enabled)
472 lp_build_mask_check(&mask);
473 }
474
475 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
476
477 /* Build the actual shader */
478 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
479 consts_ptr, num_consts_ptr, &system_values,
480 interp->inputs,
481 outputs, context_ptr, thread_data_ptr,
482 sampler, &shader->info.base, NULL);
483
484 /* Alpha test */
485 if (key->alpha.enabled) {
486 int color0 = find_output_by_semantic(&shader->info.base,
487 TGSI_SEMANTIC_COLOR,
488 0);
489
490 if (color0 != -1 && outputs[color0][3]) {
491 const struct util_format_description *cbuf_format_desc;
492 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
493 LLVMValueRef alpha_ref_value;
494
495 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
496 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
497
498 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
499
500 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
501 &mask, alpha, alpha_ref_value,
502 (depth_mode & LATE_DEPTH_TEST) != 0);
503 }
504 }
505
506 /* Emulate Alpha to Coverage with Alpha test */
507 if (key->blend.alpha_to_coverage) {
508 int color0 = find_output_by_semantic(&shader->info.base,
509 TGSI_SEMANTIC_COLOR,
510 0);
511
512 if (color0 != -1 && outputs[color0][3]) {
513 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
514
515 lp_build_alpha_to_coverage(gallivm, type,
516 &mask, alpha,
517 (depth_mode & LATE_DEPTH_TEST) != 0);
518 }
519 }
520
521 if (shader->info.base.writes_samplemask) {
522 int smaski = find_output_by_semantic(&shader->info.base,
523 TGSI_SEMANTIC_SAMPLEMASK,
524 0);
525 LLVMValueRef smask;
526 struct lp_build_context smask_bld;
527 lp_build_context_init(&smask_bld, gallivm, int_type);
528
529 assert(smaski >= 0);
530 smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
531 /*
532 * Pixel is alive according to the first sample in the mask.
533 */
534 smask = LLVMBuildBitCast(builder, smask, smask_bld.vec_type, "");
535 smask = lp_build_and(&smask_bld, smask, smask_bld.one);
536 smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, smask, smask_bld.zero);
537 lp_build_mask_update(&mask, smask);
538 }
539
540 /* Late Z test */
541 if (depth_mode & LATE_DEPTH_TEST) {
542 int pos0 = find_output_by_semantic(&shader->info.base,
543 TGSI_SEMANTIC_POSITION,
544 0);
545 int s_out = find_output_by_semantic(&shader->info.base,
546 TGSI_SEMANTIC_STENCIL,
547 0);
548 if (pos0 != -1 && outputs[pos0][2]) {
549 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
550 }
551 /*
552 * Clamp according to ARB_depth_clamp semantics.
553 */
554 if (key->depth_clamp) {
555 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
556 thread_data_ptr, z);
557 }
558
559 if (s_out != -1 && outputs[s_out][1]) {
560 /* there's only one value, and spec says to discard additional bits */
561 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
562 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
563 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
564 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
565 stencil_refs[1] = stencil_refs[0];
566 }
567
568 lp_build_depth_stencil_load_swizzled(gallivm, type,
569 zs_format_desc, key->resource_1d,
570 depth_ptr, depth_stride,
571 &z_fb, &s_fb, loop_state.counter);
572
573 lp_build_depth_stencil_test(gallivm,
574 &key->depth,
575 key->stencil,
576 type,
577 zs_format_desc,
578 &mask,
579 stencil_refs,
580 z, z_fb, s_fb,
581 facing,
582 &z_value, &s_value,
583 !simple_shader);
584 /* Late Z write */
585 if (depth_mode & LATE_DEPTH_WRITE) {
586 lp_build_depth_stencil_write_swizzled(gallivm, type,
587 zs_format_desc, key->resource_1d,
588 NULL, NULL, NULL, loop_state.counter,
589 depth_ptr, depth_stride,
590 z_value, s_value);
591 }
592 }
593 else if ((depth_mode & EARLY_DEPTH_TEST) &&
594 (depth_mode & LATE_DEPTH_WRITE))
595 {
596 /* Need to apply a reduced mask to the depth write. Reload the
597 * depth value, update from zs_value with the new mask value and
598 * write that out.
599 */
600 lp_build_depth_stencil_write_swizzled(gallivm, type,
601 zs_format_desc, key->resource_1d,
602 &mask, z_fb, s_fb, loop_state.counter,
603 depth_ptr, depth_stride,
604 z_value, s_value);
605 }
606
607
608 /* Color write */
609 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
610 {
611 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
612 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
613 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
614 {
615 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
616 if(outputs[attrib][chan]) {
617 /* XXX: just initialize outputs to point at colors[] and
618 * skip this.
619 */
620 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
621 LLVMValueRef color_ptr;
622 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
623 &loop_state.counter, 1, "");
624 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
625 LLVMBuildStore(builder, out, color_ptr);
626 }
627 }
628 }
629 }
630
631 if (key->occlusion_count) {
632 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
633 lp_build_name(counter, "counter");
634 lp_build_occlusion_count(gallivm, type,
635 lp_build_mask_value(&mask), counter);
636 }
637
638 mask_val = lp_build_mask_end(&mask);
639 LLVMBuildStore(builder, mask_val, mask_ptr);
640 lp_build_for_loop_end(&loop_state);
641 }
642
643
644 /**
645 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
646 *
647 * Fragment Shader outputs pixels in small 2x2 blocks
648 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
649 *
650 * However in memory pixels are stored in rows
651 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
652 *
653 * @param type fragment shader type (4x or 8x float)
654 * @param num_fs number of fs_src
655 * @param is_1d whether we're outputting to a 1d resource
656 * @param dst_channels number of output channels
657 * @param fs_src output from fragment shader
658 * @param dst pointer to store result
659 * @param pad_inline is channel padding inline or at end of row
660 * @return the number of dsts
661 */
662 static int
generate_fs_twiddle(struct gallivm_state * gallivm,struct lp_type type,unsigned num_fs,unsigned dst_channels,LLVMValueRef fs_src[][4],LLVMValueRef * dst,bool pad_inline)663 generate_fs_twiddle(struct gallivm_state *gallivm,
664 struct lp_type type,
665 unsigned num_fs,
666 unsigned dst_channels,
667 LLVMValueRef fs_src[][4],
668 LLVMValueRef* dst,
669 bool pad_inline)
670 {
671 LLVMValueRef src[16];
672
673 bool swizzle_pad;
674 bool twiddle;
675 bool split;
676
677 unsigned pixels = type.length / 4;
678 unsigned reorder_group;
679 unsigned src_channels;
680 unsigned src_count;
681 unsigned i;
682
683 src_channels = dst_channels < 3 ? dst_channels : 4;
684 src_count = num_fs * src_channels;
685
686 assert(pixels == 2 || pixels == 1);
687 assert(num_fs * src_channels <= ARRAY_SIZE(src));
688
689 /*
690 * Transpose from SoA -> AoS
691 */
692 for (i = 0; i < num_fs; ++i) {
693 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
694 }
695
696 /*
697 * Pick transformation options
698 */
699 swizzle_pad = false;
700 twiddle = false;
701 split = false;
702 reorder_group = 0;
703
704 if (dst_channels == 1) {
705 twiddle = true;
706
707 if (pixels == 2) {
708 split = true;
709 }
710 } else if (dst_channels == 2) {
711 if (pixels == 1) {
712 reorder_group = 1;
713 }
714 } else if (dst_channels > 2) {
715 if (pixels == 1) {
716 reorder_group = 2;
717 } else {
718 twiddle = true;
719 }
720
721 if (!pad_inline && dst_channels == 3 && pixels > 1) {
722 swizzle_pad = true;
723 }
724 }
725
726 /*
727 * Split the src in half
728 */
729 if (split) {
730 for (i = num_fs; i > 0; --i) {
731 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
732 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
733 }
734
735 src_count *= 2;
736 type.length = 4;
737 }
738
739 /*
740 * Ensure pixels are in memory order
741 */
742 if (reorder_group) {
743 /* Twiddle pixels by reordering the array, e.g.:
744 *
745 * src_count = 8 -> 0 2 1 3 4 6 5 7
746 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
747 */
748 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
749
750 for (i = 0; i < src_count; ++i) {
751 unsigned group = i / reorder_group;
752 unsigned block = (group / 4) * 4 * reorder_group;
753 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
754 dst[i] = src[j];
755 }
756 } else if (twiddle) {
757 /* Twiddle pixels across elements of array */
758 /*
759 * XXX: we should avoid this in some cases, but would need to tell
760 * lp_build_conv to reorder (or deal with it ourselves).
761 */
762 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
763 } else {
764 /* Do nothing */
765 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
766 }
767
768 /*
769 * Moves any padding between pixels to the end
770 * e.g. RGBXRGBX -> RGBRGBXX
771 */
772 if (swizzle_pad) {
773 unsigned char swizzles[16];
774 unsigned elems = pixels * dst_channels;
775
776 for (i = 0; i < type.length; ++i) {
777 if (i < elems)
778 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
779 else
780 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
781 }
782
783 for (i = 0; i < src_count; ++i) {
784 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
785 }
786 }
787
788 return src_count;
789 }
790
791
792 /*
793 * Untwiddle and transpose, much like the above.
794 * However, this is after conversion, so we get packed vectors.
795 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
796 * the vectors will look like:
797 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
798 * be swizzled here). Extending to 16bit should be trivial.
799 * Should also be extended to handle twice wide vectors with AVX2...
800 */
801 static void
fs_twiddle_transpose(struct gallivm_state * gallivm,struct lp_type type,LLVMValueRef * src,unsigned src_count,LLVMValueRef * dst)802 fs_twiddle_transpose(struct gallivm_state *gallivm,
803 struct lp_type type,
804 LLVMValueRef *src,
805 unsigned src_count,
806 LLVMValueRef *dst)
807 {
808 unsigned i, j;
809 struct lp_type type64, type16, type32;
810 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
811 LLVMBuilderRef builder = gallivm->builder;
812 LLVMValueRef tmp[4], shuf[8];
813 for (j = 0; j < 2; j++) {
814 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
815 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
816 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
817 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
818 }
819
820 assert(src_count == 4 || src_count == 2 || src_count == 1);
821 assert(type.width == 8);
822 assert(type.length == 16);
823
824 type8_t = lp_build_vec_type(gallivm, type);
825
826 type64 = type;
827 type64.length /= 8;
828 type64.width *= 8;
829 type64_t = lp_build_vec_type(gallivm, type64);
830
831 type16 = type;
832 type16.length /= 2;
833 type16.width *= 2;
834 type16_t = lp_build_vec_type(gallivm, type16);
835
836 type32 = type;
837 type32.length /= 4;
838 type32.width *= 4;
839 type32_t = lp_build_vec_type(gallivm, type32);
840
841 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
842
843 if (src_count == 1) {
844 /* transpose was no-op, just untwiddle */
845 LLVMValueRef shuf_vec;
846 shuf_vec = LLVMConstVector(shuf, 8);
847 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
848 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
849 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
850 } else if (src_count == 2) {
851 LLVMValueRef shuf_vec;
852 shuf_vec = LLVMConstVector(shuf, 4);
853
854 for (i = 0; i < 2; i++) {
855 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
856 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
857 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
858 }
859 } else {
860 for (j = 0; j < 2; j++) {
861 LLVMValueRef lo, hi, lo2, hi2;
862 /*
863 * Note that if we only really have 3 valid channels (rgb)
864 * and we don't need alpha we could substitute a undef here
865 * for the respective channel (causing llvm to drop conversion
866 * for alpha).
867 */
868 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
869 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
870 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
871 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
872 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
873 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
874 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
875 }
876 }
877 }
878
879
880 /**
881 * Load an unswizzled block of pixels from memory
882 */
883 static void
load_unswizzled_block(struct gallivm_state * gallivm,LLVMValueRef base_ptr,LLVMValueRef stride,unsigned block_width,unsigned block_height,LLVMValueRef * dst,struct lp_type dst_type,unsigned dst_count,unsigned dst_alignment)884 load_unswizzled_block(struct gallivm_state *gallivm,
885 LLVMValueRef base_ptr,
886 LLVMValueRef stride,
887 unsigned block_width,
888 unsigned block_height,
889 LLVMValueRef* dst,
890 struct lp_type dst_type,
891 unsigned dst_count,
892 unsigned dst_alignment)
893 {
894 LLVMBuilderRef builder = gallivm->builder;
895 unsigned row_size = dst_count / block_height;
896 unsigned i;
897
898 /* Ensure block exactly fits into dst */
899 assert((block_width * block_height) % dst_count == 0);
900
901 for (i = 0; i < dst_count; ++i) {
902 unsigned x = i % row_size;
903 unsigned y = i / row_size;
904
905 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
906 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
907
908 LLVMValueRef gep[2];
909 LLVMValueRef dst_ptr;
910
911 gep[0] = lp_build_const_int32(gallivm, 0);
912 gep[1] = LLVMBuildAdd(builder, bx, by, "");
913
914 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
915 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
916 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
917
918 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
919
920 LLVMSetAlignment(dst[i], dst_alignment);
921 }
922 }
923
924
925 /**
926 * Store an unswizzled block of pixels to memory
927 */
928 static void
store_unswizzled_block(struct gallivm_state * gallivm,LLVMValueRef base_ptr,LLVMValueRef stride,unsigned block_width,unsigned block_height,LLVMValueRef * src,struct lp_type src_type,unsigned src_count,unsigned src_alignment)929 store_unswizzled_block(struct gallivm_state *gallivm,
930 LLVMValueRef base_ptr,
931 LLVMValueRef stride,
932 unsigned block_width,
933 unsigned block_height,
934 LLVMValueRef* src,
935 struct lp_type src_type,
936 unsigned src_count,
937 unsigned src_alignment)
938 {
939 LLVMBuilderRef builder = gallivm->builder;
940 unsigned row_size = src_count / block_height;
941 unsigned i;
942
943 /* Ensure src exactly fits into block */
944 assert((block_width * block_height) % src_count == 0);
945
946 for (i = 0; i < src_count; ++i) {
947 unsigned x = i % row_size;
948 unsigned y = i / row_size;
949
950 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
951 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
952
953 LLVMValueRef gep[2];
954 LLVMValueRef src_ptr;
955
956 gep[0] = lp_build_const_int32(gallivm, 0);
957 gep[1] = LLVMBuildAdd(builder, bx, by, "");
958
959 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
960 src_ptr = LLVMBuildBitCast(builder, src_ptr,
961 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
962
963 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
964
965 LLVMSetAlignment(src_ptr, src_alignment);
966 }
967 }
968
969
970 /**
971 * Checks if a format description is an arithmetic format
972 *
973 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
974 */
975 static inline boolean
is_arithmetic_format(const struct util_format_description * format_desc)976 is_arithmetic_format(const struct util_format_description *format_desc)
977 {
978 boolean arith = false;
979 unsigned i;
980
981 for (i = 0; i < format_desc->nr_channels; ++i) {
982 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
983 arith |= (format_desc->channel[i].size % 8) != 0;
984 }
985
986 return arith;
987 }
988
989
990 /**
991 * Checks if this format requires special handling due to required expansion
992 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
993 * SoA conversion.
994 */
995 static inline boolean
format_expands_to_float_soa(const struct util_format_description * format_desc)996 format_expands_to_float_soa(const struct util_format_description *format_desc)
997 {
998 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
999 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1000 return true;
1001 }
1002 return false;
1003 }
1004
1005
1006 /**
1007 * Retrieves the type representing the memory layout for a format
1008 *
1009 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1010 */
1011 static inline void
lp_mem_type_from_format_desc(const struct util_format_description * format_desc,struct lp_type * type)1012 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1013 struct lp_type* type)
1014 {
1015 unsigned i;
1016 unsigned chan;
1017
1018 if (format_expands_to_float_soa(format_desc)) {
1019 /* just make this a uint with width of block */
1020 type->floating = false;
1021 type->fixed = false;
1022 type->sign = false;
1023 type->norm = false;
1024 type->width = format_desc->block.bits;
1025 type->length = 1;
1026 return;
1027 }
1028
1029 for (i = 0; i < 4; i++)
1030 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1031 break;
1032 chan = i;
1033
1034 memset(type, 0, sizeof(struct lp_type));
1035 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1036 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1037 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1038 type->norm = format_desc->channel[chan].normalized;
1039
1040 if (is_arithmetic_format(format_desc)) {
1041 type->width = 0;
1042 type->length = 1;
1043
1044 for (i = 0; i < format_desc->nr_channels; ++i) {
1045 type->width += format_desc->channel[i].size;
1046 }
1047 } else {
1048 type->width = format_desc->channel[chan].size;
1049 type->length = format_desc->nr_channels;
1050 }
1051 }
1052
1053
1054 /**
1055 * Retrieves the type for a format which is usable in the blending code.
1056 *
1057 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1058 */
1059 static inline void
lp_blend_type_from_format_desc(const struct util_format_description * format_desc,struct lp_type * type)1060 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1061 struct lp_type* type)
1062 {
1063 unsigned i;
1064 unsigned chan;
1065
1066 if (format_expands_to_float_soa(format_desc)) {
1067 /* always use ordinary floats for blending */
1068 type->floating = true;
1069 type->fixed = false;
1070 type->sign = true;
1071 type->norm = false;
1072 type->width = 32;
1073 type->length = 4;
1074 return;
1075 }
1076
1077 for (i = 0; i < 4; i++)
1078 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1079 break;
1080 chan = i;
1081
1082 memset(type, 0, sizeof(struct lp_type));
1083 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1084 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1085 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1086 type->norm = format_desc->channel[chan].normalized;
1087 type->width = format_desc->channel[chan].size;
1088 type->length = format_desc->nr_channels;
1089
1090 for (i = 1; i < format_desc->nr_channels; ++i) {
1091 if (format_desc->channel[i].size > type->width)
1092 type->width = format_desc->channel[i].size;
1093 }
1094
1095 if (type->floating) {
1096 type->width = 32;
1097 } else {
1098 if (type->width <= 8) {
1099 type->width = 8;
1100 } else if (type->width <= 16) {
1101 type->width = 16;
1102 } else {
1103 type->width = 32;
1104 }
1105 }
1106
1107 if (is_arithmetic_format(format_desc) && type->length == 3) {
1108 type->length = 4;
1109 }
1110 }
1111
1112
1113 /**
1114 * Scale a normalized value from src_bits to dst_bits.
1115 *
1116 * The exact calculation is
1117 *
1118 * dst = iround(src * dst_mask / src_mask)
1119 *
1120 * or with integer rounding
1121 *
1122 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1123 *
1124 * where
1125 *
1126 * src_mask = (1 << src_bits) - 1
1127 * dst_mask = (1 << dst_bits) - 1
1128 *
1129 * but we try to avoid division and multiplication through shifts.
1130 */
1131 static inline LLVMValueRef
scale_bits(struct gallivm_state * gallivm,int src_bits,int dst_bits,LLVMValueRef src,struct lp_type src_type)1132 scale_bits(struct gallivm_state *gallivm,
1133 int src_bits,
1134 int dst_bits,
1135 LLVMValueRef src,
1136 struct lp_type src_type)
1137 {
1138 LLVMBuilderRef builder = gallivm->builder;
1139 LLVMValueRef result = src;
1140
1141 if (dst_bits < src_bits) {
1142 int delta_bits = src_bits - dst_bits;
1143
1144 if (delta_bits <= dst_bits) {
1145 /*
1146 * Approximate the rescaling with a single shift.
1147 *
1148 * This gives the wrong rounding.
1149 */
1150
1151 result = LLVMBuildLShr(builder,
1152 src,
1153 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1154 "");
1155
1156 } else {
1157 /*
1158 * Try more accurate rescaling.
1159 */
1160
1161 /*
1162 * Drop the least significant bits to make space for the multiplication.
1163 *
1164 * XXX: A better approach would be to use a wider integer type as intermediate. But
1165 * this is enough to convert alpha from 16bits -> 2 when rendering to
1166 * PIPE_FORMAT_R10G10B10A2_UNORM.
1167 */
1168 result = LLVMBuildLShr(builder,
1169 src,
1170 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1171 "");
1172
1173
1174 result = LLVMBuildMul(builder,
1175 result,
1176 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1177 "");
1178
1179 /*
1180 * Add a rounding term before the division.
1181 *
1182 * TODO: Handle signed integers too.
1183 */
1184 if (!src_type.sign) {
1185 result = LLVMBuildAdd(builder,
1186 result,
1187 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1188 "");
1189 }
1190
1191 /*
1192 * Approximate the division by src_mask with a src_bits shift.
1193 *
1194 * Given the src has already been shifted by dst_bits, all we need
1195 * to do is to shift by the difference.
1196 */
1197
1198 result = LLVMBuildLShr(builder,
1199 result,
1200 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1201 "");
1202 }
1203
1204 } else if (dst_bits > src_bits) {
1205 /* Scale up bits */
1206 int db = dst_bits - src_bits;
1207
1208 /* Shift left by difference in bits */
1209 result = LLVMBuildShl(builder,
1210 src,
1211 lp_build_const_int_vec(gallivm, src_type, db),
1212 "");
1213
1214 if (db <= src_bits) {
1215 /* Enough bits in src to fill the remainder */
1216 LLVMValueRef lower = LLVMBuildLShr(builder,
1217 src,
1218 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1219 "");
1220
1221 result = LLVMBuildOr(builder, result, lower, "");
1222 } else if (db > src_bits) {
1223 /* Need to repeatedly copy src bits to fill remainder in dst */
1224 unsigned n;
1225
1226 for (n = src_bits; n < dst_bits; n *= 2) {
1227 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1228
1229 result = LLVMBuildOr(builder,
1230 result,
1231 LLVMBuildLShr(builder, result, shuv, ""),
1232 "");
1233 }
1234 }
1235 }
1236
1237 return result;
1238 }
1239
1240 /**
1241 * If RT is a smallfloat (needing denorms) format
1242 */
1243 static inline int
have_smallfloat_format(struct lp_type dst_type,enum pipe_format format)1244 have_smallfloat_format(struct lp_type dst_type,
1245 enum pipe_format format)
1246 {
1247 return ((dst_type.floating && dst_type.width != 32) ||
1248 /* due to format handling hacks this format doesn't have floating set
1249 * here (and actually has width set to 32 too) so special case this. */
1250 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1251 }
1252
1253
1254 /**
1255 * Convert from memory format to blending format
1256 *
1257 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1258 */
1259 static void
convert_to_blend_type(struct gallivm_state * gallivm,unsigned block_size,const struct util_format_description * src_fmt,struct lp_type src_type,struct lp_type dst_type,LLVMValueRef * src,unsigned num_srcs)1260 convert_to_blend_type(struct gallivm_state *gallivm,
1261 unsigned block_size,
1262 const struct util_format_description *src_fmt,
1263 struct lp_type src_type,
1264 struct lp_type dst_type,
1265 LLVMValueRef* src, // and dst
1266 unsigned num_srcs)
1267 {
1268 LLVMValueRef *dst = src;
1269 LLVMBuilderRef builder = gallivm->builder;
1270 struct lp_type blend_type;
1271 struct lp_type mem_type;
1272 unsigned i, j;
1273 unsigned pixels = block_size / num_srcs;
1274 bool is_arith;
1275
1276 /*
1277 * full custom path for packed floats and srgb formats - none of the later
1278 * functions would do anything useful, and given the lp_type representation they
1279 * can't be fixed. Should really have some SoA blend path for these kind of
1280 * formats rather than hacking them in here.
1281 */
1282 if (format_expands_to_float_soa(src_fmt)) {
1283 LLVMValueRef tmpsrc[4];
1284 /*
1285 * This is pretty suboptimal for this case blending in SoA would be much
1286 * better, since conversion gets us SoA values so need to convert back.
1287 */
1288 assert(src_type.width == 32 || src_type.width == 16);
1289 assert(dst_type.floating);
1290 assert(dst_type.width == 32);
1291 assert(dst_type.length % 4 == 0);
1292 assert(num_srcs % 4 == 0);
1293
1294 if (src_type.width == 16) {
1295 /* expand 4x16bit values to 4x32bit */
1296 struct lp_type type32x4 = src_type;
1297 LLVMTypeRef ltype32x4;
1298 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1299 type32x4.width = 32;
1300 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1301 for (i = 0; i < num_fetch; i++) {
1302 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1303 }
1304 src_type.width = 32;
1305 }
1306 for (i = 0; i < 4; i++) {
1307 tmpsrc[i] = src[i];
1308 }
1309 for (i = 0; i < num_srcs / 4; i++) {
1310 LLVMValueRef tmpsoa[4];
1311 LLVMValueRef tmps = tmpsrc[i];
1312 if (dst_type.length == 8) {
1313 LLVMValueRef shuffles[8];
1314 unsigned j;
1315 /* fetch was 4 values but need 8-wide output values */
1316 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1317 /*
1318 * for 8-wide aos transpose would give us wrong order not matching
1319 * incoming converted fs values and mask. ARGH.
1320 */
1321 for (j = 0; j < 4; j++) {
1322 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1323 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1324 }
1325 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1326 LLVMConstVector(shuffles, 8), "");
1327 }
1328 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1329 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1330 }
1331 else {
1332 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1333 }
1334 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1335 }
1336 return;
1337 }
1338
1339 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1340 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1341
1342 /* Is the format arithmetic */
1343 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1344 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1345
1346 /* Pad if necessary */
1347 if (!is_arith && src_type.length < dst_type.length) {
1348 for (i = 0; i < num_srcs; ++i) {
1349 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1350 }
1351
1352 src_type.length = dst_type.length;
1353 }
1354
1355 /* Special case for half-floats */
1356 if (mem_type.width == 16 && mem_type.floating) {
1357 assert(blend_type.width == 32 && blend_type.floating);
1358 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1359 is_arith = false;
1360 }
1361
1362 if (!is_arith) {
1363 return;
1364 }
1365
1366 src_type.width = blend_type.width * blend_type.length;
1367 blend_type.length *= pixels;
1368 src_type.length *= pixels / (src_type.length / mem_type.length);
1369
1370 for (i = 0; i < num_srcs; ++i) {
1371 LLVMValueRef chans[4];
1372 LLVMValueRef res = NULL;
1373
1374 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1375
1376 for (j = 0; j < src_fmt->nr_channels; ++j) {
1377 unsigned mask = 0;
1378 unsigned sa = src_fmt->channel[j].shift;
1379 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1380 unsigned from_lsb = j;
1381 #else
1382 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1383 #endif
1384
1385 mask = (1 << src_fmt->channel[j].size) - 1;
1386
1387 /* Extract bits from source */
1388 chans[j] = LLVMBuildLShr(builder,
1389 dst[i],
1390 lp_build_const_int_vec(gallivm, src_type, sa),
1391 "");
1392
1393 chans[j] = LLVMBuildAnd(builder,
1394 chans[j],
1395 lp_build_const_int_vec(gallivm, src_type, mask),
1396 "");
1397
1398 /* Scale bits */
1399 if (src_type.norm) {
1400 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1401 blend_type.width, chans[j], src_type);
1402 }
1403
1404 /* Insert bits into correct position */
1405 chans[j] = LLVMBuildShl(builder,
1406 chans[j],
1407 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1408 "");
1409
1410 if (j == 0) {
1411 res = chans[j];
1412 } else {
1413 res = LLVMBuildOr(builder, res, chans[j], "");
1414 }
1415 }
1416
1417 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1418 }
1419 }
1420
1421
1422 /**
1423 * Convert from blending format to memory format
1424 *
1425 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1426 */
1427 static void
convert_from_blend_type(struct gallivm_state * gallivm,unsigned block_size,const struct util_format_description * src_fmt,struct lp_type src_type,struct lp_type dst_type,LLVMValueRef * src,unsigned num_srcs)1428 convert_from_blend_type(struct gallivm_state *gallivm,
1429 unsigned block_size,
1430 const struct util_format_description *src_fmt,
1431 struct lp_type src_type,
1432 struct lp_type dst_type,
1433 LLVMValueRef* src, // and dst
1434 unsigned num_srcs)
1435 {
1436 LLVMValueRef* dst = src;
1437 unsigned i, j, k;
1438 struct lp_type mem_type;
1439 struct lp_type blend_type;
1440 LLVMBuilderRef builder = gallivm->builder;
1441 unsigned pixels = block_size / num_srcs;
1442 bool is_arith;
1443
1444 /*
1445 * full custom path for packed floats and srgb formats - none of the later
1446 * functions would do anything useful, and given the lp_type representation they
1447 * can't be fixed. Should really have some SoA blend path for these kind of
1448 * formats rather than hacking them in here.
1449 */
1450 if (format_expands_to_float_soa(src_fmt)) {
1451 /*
1452 * This is pretty suboptimal for this case blending in SoA would be much
1453 * better - we need to transpose the AoS values back to SoA values for
1454 * conversion/packing.
1455 */
1456 assert(src_type.floating);
1457 assert(src_type.width == 32);
1458 assert(src_type.length % 4 == 0);
1459 assert(dst_type.width == 32 || dst_type.width == 16);
1460
1461 for (i = 0; i < num_srcs / 4; i++) {
1462 LLVMValueRef tmpsoa[4], tmpdst;
1463 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1464 /* really really need SoA here */
1465
1466 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1467 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1468 }
1469 else {
1470 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1471 src_type, tmpsoa);
1472 }
1473
1474 if (src_type.length == 8) {
1475 LLVMValueRef tmpaos, shuffles[8];
1476 unsigned j;
1477 /*
1478 * for 8-wide aos transpose has given us wrong order not matching
1479 * output order. HMPF. Also need to split the output values manually.
1480 */
1481 for (j = 0; j < 4; j++) {
1482 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1483 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1484 }
1485 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1486 LLVMConstVector(shuffles, 8), "");
1487 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1488 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1489 }
1490 else {
1491 src[i] = tmpdst;
1492 }
1493 }
1494 if (dst_type.width == 16) {
1495 struct lp_type type16x8 = dst_type;
1496 struct lp_type type32x4 = dst_type;
1497 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1498 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1499 type16x8.length = 8;
1500 type32x4.width = 32;
1501 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1502 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1503 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1504 /* We could do vector truncation but it doesn't generate very good code */
1505 for (i = 0; i < num_fetch; i++) {
1506 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1507 src[i], lp_build_zero(gallivm, type32x4));
1508 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1509 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1510 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1511 }
1512 }
1513 return;
1514 }
1515
1516 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1517 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1518
1519 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1520
1521 /* Special case for half-floats */
1522 if (mem_type.width == 16 && mem_type.floating) {
1523 int length = dst_type.length;
1524 assert(blend_type.width == 32 && blend_type.floating);
1525
1526 dst_type.length = src_type.length;
1527
1528 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1529
1530 dst_type.length = length;
1531 is_arith = false;
1532 }
1533
1534 /* Remove any padding */
1535 if (!is_arith && (src_type.length % mem_type.length)) {
1536 src_type.length -= (src_type.length % mem_type.length);
1537
1538 for (i = 0; i < num_srcs; ++i) {
1539 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1540 }
1541 }
1542
1543 /* No bit arithmetic to do */
1544 if (!is_arith) {
1545 return;
1546 }
1547
1548 src_type.length = pixels;
1549 src_type.width = blend_type.length * blend_type.width;
1550 dst_type.length = pixels;
1551
1552 for (i = 0; i < num_srcs; ++i) {
1553 LLVMValueRef chans[4];
1554 LLVMValueRef res = NULL;
1555
1556 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1557
1558 for (j = 0; j < src_fmt->nr_channels; ++j) {
1559 unsigned mask = 0;
1560 unsigned sa = src_fmt->channel[j].shift;
1561 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1562 unsigned from_lsb = j;
1563 #else
1564 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1565 #endif
1566
1567 assert(blend_type.width > src_fmt->channel[j].size);
1568
1569 for (k = 0; k < blend_type.width; ++k) {
1570 mask |= 1 << k;
1571 }
1572
1573 /* Extract bits */
1574 chans[j] = LLVMBuildLShr(builder,
1575 dst[i],
1576 lp_build_const_int_vec(gallivm, src_type,
1577 from_lsb * blend_type.width),
1578 "");
1579
1580 chans[j] = LLVMBuildAnd(builder,
1581 chans[j],
1582 lp_build_const_int_vec(gallivm, src_type, mask),
1583 "");
1584
1585 /* Scale down bits */
1586 if (src_type.norm) {
1587 chans[j] = scale_bits(gallivm, blend_type.width,
1588 src_fmt->channel[j].size, chans[j], src_type);
1589 }
1590
1591 /* Insert bits */
1592 chans[j] = LLVMBuildShl(builder,
1593 chans[j],
1594 lp_build_const_int_vec(gallivm, src_type, sa),
1595 "");
1596
1597 sa += src_fmt->channel[j].size;
1598
1599 if (j == 0) {
1600 res = chans[j];
1601 } else {
1602 res = LLVMBuildOr(builder, res, chans[j], "");
1603 }
1604 }
1605
1606 assert (dst_type.width != 24);
1607
1608 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1609 }
1610 }
1611
1612
1613 /**
1614 * Convert alpha to same blend type as src
1615 */
1616 static void
convert_alpha(struct gallivm_state * gallivm,struct lp_type row_type,struct lp_type alpha_type,const unsigned block_size,const unsigned block_height,const unsigned src_count,const unsigned dst_channels,const bool pad_inline,LLVMValueRef * src_alpha)1617 convert_alpha(struct gallivm_state *gallivm,
1618 struct lp_type row_type,
1619 struct lp_type alpha_type,
1620 const unsigned block_size,
1621 const unsigned block_height,
1622 const unsigned src_count,
1623 const unsigned dst_channels,
1624 const bool pad_inline,
1625 LLVMValueRef* src_alpha)
1626 {
1627 LLVMBuilderRef builder = gallivm->builder;
1628 unsigned i, j;
1629 unsigned length = row_type.length;
1630 row_type.length = alpha_type.length;
1631
1632 /* Twiddle the alpha to match pixels */
1633 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1634
1635 /*
1636 * TODO this should use single lp_build_conv call for
1637 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1638 */
1639 for (i = 0; i < block_height; ++i) {
1640 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1641 }
1642
1643 alpha_type = row_type;
1644 row_type.length = length;
1645
1646 /* If only one channel we can only need the single alpha value per pixel */
1647 if (src_count == 1 && dst_channels == 1) {
1648
1649 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1650 } else {
1651 /* If there are more srcs than rows then we need to split alpha up */
1652 if (src_count > block_height) {
1653 for (i = src_count; i > 0; --i) {
1654 unsigned pixels = block_size / src_count;
1655 unsigned idx = i - 1;
1656
1657 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1658 (idx * pixels) % 4, pixels);
1659 }
1660 }
1661
1662 /* If there is a src for each pixel broadcast the alpha across whole row */
1663 if (src_count == block_size) {
1664 for (i = 0; i < src_count; ++i) {
1665 src_alpha[i] = lp_build_broadcast(gallivm,
1666 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1667 }
1668 } else {
1669 unsigned pixels = block_size / src_count;
1670 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1671 unsigned alpha_span = 1;
1672 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1673
1674 /* Check if we need 2 src_alphas for our shuffles */
1675 if (pixels > alpha_type.length) {
1676 alpha_span = 2;
1677 }
1678
1679 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1680 for (j = 0; j < row_type.length; ++j) {
1681 if (j < pixels * channels) {
1682 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1683 } else {
1684 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1685 }
1686 }
1687
1688 for (i = 0; i < src_count; ++i) {
1689 unsigned idx1 = i, idx2 = i;
1690
1691 if (alpha_span > 1){
1692 idx1 *= alpha_span;
1693 idx2 = idx1 + 1;
1694 }
1695
1696 src_alpha[i] = LLVMBuildShuffleVector(builder,
1697 src_alpha[idx1],
1698 src_alpha[idx2],
1699 LLVMConstVector(shuffles, row_type.length),
1700 "");
1701 }
1702 }
1703 }
1704 }
1705
1706
1707 /**
1708 * Generates the blend function for unswizzled colour buffers
1709 * Also generates the read & write from colour buffer
1710 */
1711 static void
generate_unswizzled_blend(struct gallivm_state * gallivm,unsigned rt,struct lp_fragment_shader_variant * variant,enum pipe_format out_format,unsigned int num_fs,struct lp_type fs_type,LLVMValueRef * fs_mask,LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],LLVMValueRef context_ptr,LLVMValueRef color_ptr,LLVMValueRef stride,unsigned partial_mask,boolean do_branch)1712 generate_unswizzled_blend(struct gallivm_state *gallivm,
1713 unsigned rt,
1714 struct lp_fragment_shader_variant *variant,
1715 enum pipe_format out_format,
1716 unsigned int num_fs,
1717 struct lp_type fs_type,
1718 LLVMValueRef* fs_mask,
1719 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1720 LLVMValueRef context_ptr,
1721 LLVMValueRef color_ptr,
1722 LLVMValueRef stride,
1723 unsigned partial_mask,
1724 boolean do_branch)
1725 {
1726 const unsigned alpha_channel = 3;
1727 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1728 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1729 const unsigned block_size = block_width * block_height;
1730 const unsigned lp_integer_vector_width = 128;
1731
1732 LLVMBuilderRef builder = gallivm->builder;
1733 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1734 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1735 LLVMValueRef src_alpha[4 * 4];
1736 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1737 LLVMValueRef src_mask[4 * 4];
1738 LLVMValueRef src[4 * 4];
1739 LLVMValueRef src1[4 * 4];
1740 LLVMValueRef dst[4 * 4];
1741 LLVMValueRef blend_color;
1742 LLVMValueRef blend_alpha;
1743 LLVMValueRef i32_zero;
1744 LLVMValueRef check_mask;
1745 LLVMValueRef undef_src_val;
1746
1747 struct lp_build_mask_context mask_ctx;
1748 struct lp_type mask_type;
1749 struct lp_type blend_type;
1750 struct lp_type row_type;
1751 struct lp_type dst_type;
1752 struct lp_type ls_type;
1753
1754 unsigned char swizzle[TGSI_NUM_CHANNELS];
1755 unsigned vector_width;
1756 unsigned src_channels = TGSI_NUM_CHANNELS;
1757 unsigned dst_channels;
1758 unsigned dst_count;
1759 unsigned src_count;
1760 unsigned i, j;
1761
1762 const struct util_format_description* out_format_desc = util_format_description(out_format);
1763
1764 unsigned dst_alignment;
1765
1766 bool pad_inline = is_arithmetic_format(out_format_desc);
1767 bool has_alpha = false;
1768 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1769 util_blend_state_is_dual(&variant->key.blend, 0);
1770
1771 const boolean is_1d = variant->key.resource_1d;
1772 boolean twiddle_after_convert = FALSE;
1773 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1774 LLVMValueRef fpstate = 0;
1775
1776 /* Get type from output format */
1777 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1778 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1779
1780 /*
1781 * Technically this code should go into lp_build_smallfloat_to_float
1782 * and lp_build_float_to_smallfloat but due to the
1783 * http://llvm.org/bugs/show_bug.cgi?id=6393
1784 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1785 * So the ordering is important here and there shouldn't be any
1786 * llvm ir instrunctions in this function before
1787 * this, otherwise half-float format conversions won't work
1788 * (again due to llvm bug #6393).
1789 */
1790 if (have_smallfloat_format(dst_type, out_format)) {
1791 /* We need to make sure that denorms are ok for half float
1792 conversions */
1793 fpstate = lp_build_fpstate_get(gallivm);
1794 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
1795 }
1796
1797 mask_type = lp_int32_vec4_type();
1798 mask_type.length = fs_type.length;
1799
1800 for (i = num_fs; i < num_fullblock_fs; i++) {
1801 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1802 }
1803
1804 /* Do not bother executing code when mask is empty.. */
1805 if (do_branch) {
1806 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1807
1808 for (i = 0; i < num_fullblock_fs; ++i) {
1809 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1810 }
1811
1812 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1813 lp_build_mask_check(&mask_ctx);
1814 }
1815
1816 partial_mask |= !variant->opaque;
1817 i32_zero = lp_build_const_int32(gallivm, 0);
1818
1819 undef_src_val = lp_build_undef(gallivm, fs_type);
1820
1821 row_type.length = fs_type.length;
1822 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1823
1824 /* Compute correct swizzle and count channels */
1825 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1826 dst_channels = 0;
1827
1828 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1829 /* Ensure channel is used */
1830 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1831 continue;
1832 }
1833
1834 /* Ensure not already written to (happens in case with GL_ALPHA) */
1835 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1836 continue;
1837 }
1838
1839 /* Ensure we havn't already found all channels */
1840 if (dst_channels >= out_format_desc->nr_channels) {
1841 continue;
1842 }
1843
1844 swizzle[out_format_desc->swizzle[i]] = i;
1845 ++dst_channels;
1846
1847 if (i == alpha_channel) {
1848 has_alpha = true;
1849 }
1850 }
1851
1852 if (format_expands_to_float_soa(out_format_desc)) {
1853 /*
1854 * the code above can't work for layout_other
1855 * for srgb it would sort of work but we short-circuit swizzles, etc.
1856 * as that is done as part of unpack / pack.
1857 */
1858 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1859 has_alpha = true;
1860 swizzle[0] = 0;
1861 swizzle[1] = 1;
1862 swizzle[2] = 2;
1863 swizzle[3] = 3;
1864 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1865 }
1866
1867 /* If 3 channels then pad to include alpha for 4 element transpose */
1868 if (dst_channels == 3) {
1869 assert (!has_alpha);
1870 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1871 if (swizzle[i] > TGSI_NUM_CHANNELS)
1872 swizzle[i] = 3;
1873 }
1874 if (out_format_desc->nr_channels == 4) {
1875 dst_channels = 4;
1876 /*
1877 * We use alpha from the color conversion, not separate one.
1878 * We had to include it for transpose, hence it will get converted
1879 * too (albeit when doing transpose after conversion, that would
1880 * no longer be the case necessarily).
1881 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1882 * otherwise we really have padding, not alpha, included.)
1883 */
1884 has_alpha = true;
1885 }
1886 }
1887
1888 /*
1889 * Load shader output
1890 */
1891 for (i = 0; i < num_fullblock_fs; ++i) {
1892 /* Always load alpha for use in blending */
1893 LLVMValueRef alpha;
1894 if (i < num_fs) {
1895 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1896 }
1897 else {
1898 alpha = undef_src_val;
1899 }
1900
1901 /* Load each channel */
1902 for (j = 0; j < dst_channels; ++j) {
1903 assert(swizzle[j] < 4);
1904 if (i < num_fs) {
1905 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1906 }
1907 else {
1908 fs_src[i][j] = undef_src_val;
1909 }
1910 }
1911
1912 /* If 3 channels then pad to include alpha for 4 element transpose */
1913 /*
1914 * XXX If we include that here maybe could actually use it instead of
1915 * separate alpha for blending?
1916 * (Difficult though we actually convert pad channels, not alpha.)
1917 */
1918 if (dst_channels == 3 && !has_alpha) {
1919 fs_src[i][3] = alpha;
1920 }
1921
1922 /* We split the row_mask and row_alpha as we want 128bit interleave */
1923 if (fs_type.length == 8) {
1924 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
1925 0, src_channels);
1926 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
1927 src_channels, src_channels);
1928
1929 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1930 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
1931 src_channels, src_channels);
1932 } else {
1933 src_mask[i] = fs_mask[i];
1934 src_alpha[i] = alpha;
1935 }
1936 }
1937 if (dual_source_blend) {
1938 /* same as above except different src/dst, skip masks and comments... */
1939 for (i = 0; i < num_fullblock_fs; ++i) {
1940 LLVMValueRef alpha;
1941 if (i < num_fs) {
1942 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1943 }
1944 else {
1945 alpha = undef_src_val;
1946 }
1947
1948 for (j = 0; j < dst_channels; ++j) {
1949 assert(swizzle[j] < 4);
1950 if (i < num_fs) {
1951 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1952 }
1953 else {
1954 fs_src1[i][j] = undef_src_val;
1955 }
1956 }
1957 if (dst_channels == 3 && !has_alpha) {
1958 fs_src1[i][3] = alpha;
1959 }
1960 if (fs_type.length == 8) {
1961 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1962 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
1963 src_channels, src_channels);
1964 } else {
1965 src1_alpha[i] = alpha;
1966 }
1967 }
1968 }
1969
1970 if (util_format_is_pure_integer(out_format)) {
1971 /*
1972 * In this case fs_type was really ints or uints disguised as floats,
1973 * fix that up now.
1974 */
1975 fs_type.floating = 0;
1976 fs_type.sign = dst_type.sign;
1977 for (i = 0; i < num_fullblock_fs; ++i) {
1978 for (j = 0; j < dst_channels; ++j) {
1979 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1980 lp_build_vec_type(gallivm, fs_type), "");
1981 }
1982 if (dst_channels == 3 && !has_alpha) {
1983 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1984 lp_build_vec_type(gallivm, fs_type), "");
1985 }
1986 }
1987 }
1988
1989 /*
1990 * We actually should generally do conversion first (for non-1d cases)
1991 * when the blend format is 8 or 16 bits. The reason is obvious,
1992 * there's 2 or 4 times less vectors to deal with for the interleave...
1993 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
1994 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
1995 * unpack only with 128bit vectors).
1996 * Note: for 16bit sizes really need matching pack conversion code
1997 */
1998 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
1999 twiddle_after_convert = TRUE;
2000 }
2001
2002 /*
2003 * Pixel twiddle from fragment shader order to memory order
2004 */
2005 if (!twiddle_after_convert) {
2006 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2007 dst_channels, fs_src, src, pad_inline);
2008 if (dual_source_blend) {
2009 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2010 fs_src1, src1, pad_inline);
2011 }
2012 } else {
2013 src_count = num_fullblock_fs * dst_channels;
2014 /*
2015 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2016 * (AVX) turn out the same later when untwiddling/transpose (albeit
2017 * for true AVX2 path untwiddle needs to be different).
2018 * For now just order by colors first (so we can use unpack later).
2019 */
2020 for (j = 0; j < num_fullblock_fs; j++) {
2021 for (i = 0; i < dst_channels; i++) {
2022 src[i*num_fullblock_fs + j] = fs_src[j][i];
2023 if (dual_source_blend) {
2024 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2025 }
2026 }
2027 }
2028 }
2029
2030 src_channels = dst_channels < 3 ? dst_channels : 4;
2031 if (src_count != num_fullblock_fs * src_channels) {
2032 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2033 row_type.length /= ds;
2034 fs_type.length = row_type.length;
2035 }
2036
2037 blend_type = row_type;
2038 mask_type.length = 4;
2039
2040 /* Convert src to row_type */
2041 if (dual_source_blend) {
2042 struct lp_type old_row_type = row_type;
2043 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2044 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2045 }
2046 else {
2047 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2048 }
2049
2050 /* If the rows are not an SSE vector, combine them to become SSE size! */
2051 if ((row_type.width * row_type.length) % 128) {
2052 unsigned bits = row_type.width * row_type.length;
2053 unsigned combined;
2054
2055 assert(src_count >= (vector_width / bits));
2056
2057 dst_count = src_count / (vector_width / bits);
2058
2059 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2060 if (dual_source_blend) {
2061 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2062 }
2063
2064 row_type.length *= combined;
2065 src_count /= combined;
2066
2067 bits = row_type.width * row_type.length;
2068 assert(bits == 128 || bits == 256);
2069 }
2070
2071 if (twiddle_after_convert) {
2072 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2073 if (dual_source_blend) {
2074 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2075 }
2076 }
2077
2078 /*
2079 * Blend Colour conversion
2080 */
2081 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2082 blend_color = LLVMBuildPointerCast(builder, blend_color,
2083 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2084 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2085 &i32_zero, 1, ""), "");
2086
2087 /* Convert */
2088 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2089
2090 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2091 /*
2092 * since blending is done with floats, there was no conversion.
2093 * However, the rules according to fixed point renderbuffers still
2094 * apply, that is we must clamp inputs to 0.0/1.0.
2095 * (This would apply to separate alpha conversion too but we currently
2096 * force has_alpha to be true.)
2097 * TODO: should skip this with "fake" blend, since post-blend conversion
2098 * will clamp anyway.
2099 * TODO: could also skip this if fragment color clamping is enabled. We
2100 * don't support it natively so it gets baked into the shader however, so
2101 * can't really tell here.
2102 */
2103 struct lp_build_context f32_bld;
2104 assert(row_type.floating);
2105 lp_build_context_init(&f32_bld, gallivm, row_type);
2106 for (i = 0; i < src_count; i++) {
2107 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2108 }
2109 if (dual_source_blend) {
2110 for (i = 0; i < src_count; i++) {
2111 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2112 }
2113 }
2114 /* probably can't be different than row_type but better safe than sorry... */
2115 lp_build_context_init(&f32_bld, gallivm, blend_type);
2116 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2117 }
2118
2119 /* Extract alpha */
2120 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2121
2122 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2123 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2124 if (pad_inline) {
2125 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2126 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2127 } else {
2128 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2129 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2130 }
2131
2132 /*
2133 * Mask conversion
2134 */
2135 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2136
2137 if (src_count < block_height) {
2138 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2139 } else if (src_count > block_height) {
2140 for (i = src_count; i > 0; --i) {
2141 unsigned pixels = block_size / src_count;
2142 unsigned idx = i - 1;
2143
2144 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2145 (idx * pixels) % 4, pixels);
2146 }
2147 }
2148
2149 assert(mask_type.width == 32);
2150
2151 for (i = 0; i < src_count; ++i) {
2152 unsigned pixels = block_size / src_count;
2153 unsigned pixel_width = row_type.width * dst_channels;
2154
2155 if (pixel_width == 24) {
2156 mask_type.width = 8;
2157 mask_type.length = vector_width / mask_type.width;
2158 } else {
2159 mask_type.length = pixels;
2160 mask_type.width = row_type.width * dst_channels;
2161
2162 /*
2163 * If mask_type width is smaller than 32bit, this doesn't quite
2164 * generate the most efficient code (could use some pack).
2165 */
2166 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2167 lp_build_int_vec_type(gallivm, mask_type), "");
2168
2169 mask_type.length *= dst_channels;
2170 mask_type.width /= dst_channels;
2171 }
2172
2173 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2174 lp_build_int_vec_type(gallivm, mask_type), "");
2175 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2176 }
2177
2178 /*
2179 * Alpha conversion
2180 */
2181 if (!has_alpha) {
2182 struct lp_type alpha_type = fs_type;
2183 alpha_type.length = 4;
2184 convert_alpha(gallivm, row_type, alpha_type,
2185 block_size, block_height,
2186 src_count, dst_channels,
2187 pad_inline, src_alpha);
2188 if (dual_source_blend) {
2189 convert_alpha(gallivm, row_type, alpha_type,
2190 block_size, block_height,
2191 src_count, dst_channels,
2192 pad_inline, src1_alpha);
2193 }
2194 }
2195
2196
2197 /*
2198 * Load dst from memory
2199 */
2200 if (src_count < block_height) {
2201 dst_count = block_height;
2202 } else {
2203 dst_count = src_count;
2204 }
2205
2206 dst_type.length *= block_size / dst_count;
2207
2208 if (format_expands_to_float_soa(out_format_desc)) {
2209 /*
2210 * we need multiple values at once for the conversion, so can as well
2211 * load them vectorized here too instead of concatenating later.
2212 * (Still need concatenation later for 8-wide vectors).
2213 */
2214 dst_count = block_height;
2215 dst_type.length = block_width;
2216 }
2217
2218 /*
2219 * Compute the alignment of the destination pointer in bytes
2220 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2221 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2222 * 1d tex but can't distinguish here) so need to stick with per-pixel
2223 * alignment in this case.
2224 */
2225 if (is_1d) {
2226 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2227 }
2228 else {
2229 dst_alignment = dst_type.length * dst_type.width / 8;
2230 }
2231 /* Force power-of-two alignment by extracting only the least-significant-bit */
2232 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2233 /*
2234 * Resource base and stride pointers are aligned to 16 bytes, so that's
2235 * the maximum alignment we can guarantee
2236 */
2237 dst_alignment = MIN2(16, dst_alignment);
2238
2239 ls_type = dst_type;
2240
2241 if (dst_count > src_count) {
2242 if ((dst_type.width == 8 || dst_type.width == 16) &&
2243 util_is_power_of_two(dst_type.length) &&
2244 dst_type.length * dst_type.width < 128) {
2245 /*
2246 * Never try to load values as 4xi8 which we will then
2247 * concatenate to larger vectors. This gives llvm a real
2248 * headache (the problem is the type legalizer (?) will
2249 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2250 * then the shuffles to concatenate are more or less impossible
2251 * - llvm is easily capable of generating a sequence of 32
2252 * pextrb/pinsrb instructions for that. Albeit it appears to
2253 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2254 * width to avoid the trouble (16bit seems not as bad, llvm
2255 * probably recognizes the load+shuffle as only one shuffle
2256 * is necessary, but we can do just the same anyway).
2257 */
2258 ls_type.length = dst_type.length * dst_type.width / 32;
2259 ls_type.width = 32;
2260 }
2261 }
2262
2263 if (is_1d) {
2264 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2265 dst, ls_type, dst_count / 4, dst_alignment);
2266 for (i = dst_count / 4; i < dst_count; i++) {
2267 dst[i] = lp_build_undef(gallivm, ls_type);
2268 }
2269
2270 }
2271 else {
2272 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2273 dst, ls_type, dst_count, dst_alignment);
2274 }
2275
2276
2277 /*
2278 * Convert from dst/output format to src/blending format.
2279 *
2280 * This is necessary as we can only read 1 row from memory at a time,
2281 * so the minimum dst_count will ever be at this point is 4.
2282 *
2283 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2284 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2285 * on all 16 pixels in that single vector at once.
2286 */
2287 if (dst_count > src_count) {
2288 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2289 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2290 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2291 for (i = 0; i < dst_count; i++) {
2292 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2293 }
2294 }
2295
2296 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2297
2298 if (ls_type.length != dst_type.length) {
2299 struct lp_type tmp_type = dst_type;
2300 tmp_type.length = dst_type.length * 4 / src_count;
2301 for (i = 0; i < src_count; i++) {
2302 dst[i] = LLVMBuildBitCast(builder, dst[i],
2303 lp_build_vec_type(gallivm, tmp_type), "");
2304 }
2305 }
2306 }
2307
2308 /*
2309 * Blending
2310 */
2311 /* XXX this is broken for RGB8 formats -
2312 * they get expanded from 12 to 16 elements (to include alpha)
2313 * by convert_to_blend_type then reduced to 15 instead of 12
2314 * by convert_from_blend_type (a simple fix though breaks A8...).
2315 * R16G16B16 also crashes differently however something going wrong
2316 * inside llvm handling npot vector sizes seemingly.
2317 * It seems some cleanup could be done here (like skipping conversion/blend
2318 * when not needed).
2319 */
2320 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2321 row_type, dst, src_count);
2322
2323 /*
2324 * FIXME: Really should get logic ops / masks out of generic blend / row
2325 * format. Logic ops will definitely not work on the blend float format
2326 * used for SRGB here and I think OpenGL expects this to work as expected
2327 * (that is incoming values converted to srgb then logic op applied).
2328 */
2329 for (i = 0; i < src_count; ++i) {
2330 dst[i] = lp_build_blend_aos(gallivm,
2331 &variant->key.blend,
2332 out_format,
2333 row_type,
2334 rt,
2335 src[i],
2336 has_alpha ? NULL : src_alpha[i],
2337 src1[i],
2338 has_alpha ? NULL : src1_alpha[i],
2339 dst[i],
2340 partial_mask ? src_mask[i] : NULL,
2341 blend_color,
2342 has_alpha ? NULL : blend_alpha,
2343 swizzle,
2344 pad_inline ? 4 : dst_channels);
2345 }
2346
2347 convert_from_blend_type(gallivm, block_size, out_format_desc,
2348 row_type, dst_type, dst, src_count);
2349
2350 /* Split the blend rows back to memory rows */
2351 if (dst_count > src_count) {
2352 row_type.length = dst_type.length * (dst_count / src_count);
2353
2354 if (src_count == 1) {
2355 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2356 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2357
2358 row_type.length /= 2;
2359 src_count *= 2;
2360 }
2361
2362 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2363 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2364 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2365 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2366
2367 row_type.length /= 2;
2368 src_count *= 2;
2369 }
2370
2371 /*
2372 * Store blend result to memory
2373 */
2374 if (is_1d) {
2375 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2376 dst, dst_type, dst_count / 4, dst_alignment);
2377 }
2378 else {
2379 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2380 dst, dst_type, dst_count, dst_alignment);
2381 }
2382
2383 if (have_smallfloat_format(dst_type, out_format)) {
2384 lp_build_fpstate_set(gallivm, fpstate);
2385 }
2386
2387 if (do_branch) {
2388 lp_build_mask_end(&mask_ctx);
2389 }
2390 }
2391
2392
2393 /**
2394 * Generate the runtime callable function for the whole fragment pipeline.
2395 * Note that the function which we generate operates on a block of 16
2396 * pixels at at time. The block contains 2x2 quads. Each quad contains
2397 * 2x2 pixels.
2398 */
2399 static void
generate_fragment(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant * variant,unsigned partial_mask)2400 generate_fragment(struct llvmpipe_context *lp,
2401 struct lp_fragment_shader *shader,
2402 struct lp_fragment_shader_variant *variant,
2403 unsigned partial_mask)
2404 {
2405 struct gallivm_state *gallivm = variant->gallivm;
2406 const struct lp_fragment_shader_variant_key *key = &variant->key;
2407 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2408 char func_name[64];
2409 struct lp_type fs_type;
2410 struct lp_type blend_type;
2411 LLVMTypeRef fs_elem_type;
2412 LLVMTypeRef blend_vec_type;
2413 LLVMTypeRef arg_types[13];
2414 LLVMTypeRef func_type;
2415 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2416 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2417 LLVMValueRef context_ptr;
2418 LLVMValueRef x;
2419 LLVMValueRef y;
2420 LLVMValueRef a0_ptr;
2421 LLVMValueRef dadx_ptr;
2422 LLVMValueRef dady_ptr;
2423 LLVMValueRef color_ptr_ptr;
2424 LLVMValueRef stride_ptr;
2425 LLVMValueRef depth_ptr;
2426 LLVMValueRef depth_stride;
2427 LLVMValueRef mask_input;
2428 LLVMValueRef thread_data_ptr;
2429 LLVMBasicBlockRef block;
2430 LLVMBuilderRef builder;
2431 struct lp_build_sampler_soa *sampler;
2432 struct lp_build_interp_soa_context interp;
2433 LLVMValueRef fs_mask[16 / 4];
2434 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2435 LLVMValueRef function;
2436 LLVMValueRef facing;
2437 unsigned num_fs;
2438 unsigned i;
2439 unsigned chan;
2440 unsigned cbuf;
2441 boolean cbuf0_write_all;
2442 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2443 util_blend_state_is_dual(&key->blend, 0);
2444
2445 assert(lp_native_vector_width / 32 >= 4);
2446
2447 /* Adjust color input interpolation according to flatshade state:
2448 */
2449 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2450 for (i = 0; i < shader->info.base.num_inputs; i++) {
2451 if (inputs[i].interp == LP_INTERP_COLOR) {
2452 if (key->flatshade)
2453 inputs[i].interp = LP_INTERP_CONSTANT;
2454 else
2455 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2456 }
2457 }
2458
2459 /* check if writes to cbuf[0] are to be copied to all cbufs */
2460 cbuf0_write_all =
2461 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2462
2463 /* TODO: actually pick these based on the fs and color buffer
2464 * characteristics. */
2465
2466 memset(&fs_type, 0, sizeof fs_type);
2467 fs_type.floating = TRUE; /* floating point values */
2468 fs_type.sign = TRUE; /* values are signed */
2469 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2470 fs_type.width = 32; /* 32-bit float */
2471 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2472
2473 memset(&blend_type, 0, sizeof blend_type);
2474 blend_type.floating = FALSE; /* values are integers */
2475 blend_type.sign = FALSE; /* values are unsigned */
2476 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2477 blend_type.width = 8; /* 8-bit ubyte values */
2478 blend_type.length = 16; /* 16 elements per vector */
2479
2480 /*
2481 * Generate the function prototype. Any change here must be reflected in
2482 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2483 */
2484
2485 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2486
2487 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2488
2489 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2490 shader->no, variant->no, partial_mask ? "partial" : "whole");
2491
2492 arg_types[0] = variant->jit_context_ptr_type; /* context */
2493 arg_types[1] = int32_type; /* x */
2494 arg_types[2] = int32_type; /* y */
2495 arg_types[3] = int32_type; /* facing */
2496 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2497 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2498 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2499 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2500 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2501 arg_types[9] = int32_type; /* mask_input */
2502 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2503 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2504 arg_types[12] = int32_type; /* depth_stride */
2505
2506 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2507 arg_types, ARRAY_SIZE(arg_types), 0);
2508
2509 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2510 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2511
2512 variant->function[partial_mask] = function;
2513
2514 /* XXX: need to propagate noalias down into color param now we are
2515 * passing a pointer-to-pointer?
2516 */
2517 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2518 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2519 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2520
2521 context_ptr = LLVMGetParam(function, 0);
2522 x = LLVMGetParam(function, 1);
2523 y = LLVMGetParam(function, 2);
2524 facing = LLVMGetParam(function, 3);
2525 a0_ptr = LLVMGetParam(function, 4);
2526 dadx_ptr = LLVMGetParam(function, 5);
2527 dady_ptr = LLVMGetParam(function, 6);
2528 color_ptr_ptr = LLVMGetParam(function, 7);
2529 depth_ptr = LLVMGetParam(function, 8);
2530 mask_input = LLVMGetParam(function, 9);
2531 thread_data_ptr = LLVMGetParam(function, 10);
2532 stride_ptr = LLVMGetParam(function, 11);
2533 depth_stride = LLVMGetParam(function, 12);
2534
2535 lp_build_name(context_ptr, "context");
2536 lp_build_name(x, "x");
2537 lp_build_name(y, "y");
2538 lp_build_name(a0_ptr, "a0");
2539 lp_build_name(dadx_ptr, "dadx");
2540 lp_build_name(dady_ptr, "dady");
2541 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2542 lp_build_name(depth_ptr, "depth");
2543 lp_build_name(mask_input, "mask_input");
2544 lp_build_name(thread_data_ptr, "thread_data");
2545 lp_build_name(stride_ptr, "stride_ptr");
2546 lp_build_name(depth_stride, "depth_stride");
2547
2548 /*
2549 * Function body
2550 */
2551
2552 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2553 builder = gallivm->builder;
2554 assert(builder);
2555 LLVMPositionBuilderAtEnd(builder, block);
2556
2557 /* code generated texture sampling */
2558 sampler = lp_llvm_sampler_soa_create(key->state);
2559
2560 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2561 /* for 1d resources only run "upper half" of stamp */
2562 if (key->resource_1d)
2563 num_fs /= 2;
2564
2565 {
2566 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2567 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2568 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2569 num_loop, "mask_store");
2570 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2571 boolean pixel_center_integer =
2572 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2573
2574 /*
2575 * The shader input interpolation info is not explicitely baked in the
2576 * shader key, but everything it derives from (TGSI, and flatshade) is
2577 * already included in the shader key.
2578 */
2579 lp_build_interp_soa_init(&interp,
2580 gallivm,
2581 shader->info.base.num_inputs,
2582 inputs,
2583 pixel_center_integer,
2584 key->depth_clamp,
2585 builder, fs_type,
2586 a0_ptr, dadx_ptr, dady_ptr,
2587 x, y);
2588
2589 for (i = 0; i < num_fs; i++) {
2590 LLVMValueRef mask;
2591 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2592 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2593 &indexi, 1, "mask_ptr");
2594
2595 if (partial_mask) {
2596 mask = generate_quad_mask(gallivm, fs_type,
2597 i*fs_type.length/4, mask_input);
2598 }
2599 else {
2600 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2601 }
2602 LLVMBuildStore(builder, mask, mask_ptr);
2603 }
2604
2605 generate_fs_loop(gallivm,
2606 shader, key,
2607 builder,
2608 fs_type,
2609 context_ptr,
2610 num_loop,
2611 &interp,
2612 sampler,
2613 mask_store, /* output */
2614 color_store,
2615 depth_ptr,
2616 depth_stride,
2617 facing,
2618 thread_data_ptr);
2619
2620 for (i = 0; i < num_fs; i++) {
2621 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2622 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2623 &indexi, 1, "");
2624 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2625 /* This is fucked up need to reorganize things */
2626 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2627 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2628 ptr = LLVMBuildGEP(builder,
2629 color_store[cbuf * !cbuf0_write_all][chan],
2630 &indexi, 1, "");
2631 fs_out_color[cbuf][chan][i] = ptr;
2632 }
2633 }
2634 if (dual_source_blend) {
2635 /* only support one dual source blend target hence always use output 1 */
2636 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2637 ptr = LLVMBuildGEP(builder,
2638 color_store[1][chan],
2639 &indexi, 1, "");
2640 fs_out_color[1][chan][i] = ptr;
2641 }
2642 }
2643 }
2644 }
2645
2646 sampler->destroy(sampler);
2647
2648 /* Loop over color outputs / color buffers to do blending.
2649 */
2650 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2651 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2652 LLVMValueRef color_ptr;
2653 LLVMValueRef stride;
2654 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2655
2656 boolean do_branch = ((key->depth.enabled
2657 || key->stencil[0].enabled
2658 || key->alpha.enabled)
2659 && !shader->info.base.uses_kill);
2660
2661 color_ptr = LLVMBuildLoad(builder,
2662 LLVMBuildGEP(builder, color_ptr_ptr,
2663 &index, 1, ""),
2664 "");
2665
2666 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2667
2668 stride = LLVMBuildLoad(builder,
2669 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2670 "");
2671
2672 generate_unswizzled_blend(gallivm, cbuf, variant,
2673 key->cbuf_format[cbuf],
2674 num_fs, fs_type, fs_mask, fs_out_color,
2675 context_ptr, color_ptr, stride,
2676 partial_mask, do_branch);
2677 }
2678 }
2679
2680 LLVMBuildRetVoid(builder);
2681
2682 gallivm_verify_function(gallivm, function);
2683 }
2684
2685
2686 static void
dump_fs_variant_key(const struct lp_fragment_shader_variant_key * key)2687 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2688 {
2689 unsigned i;
2690
2691 debug_printf("fs variant %p:\n", (void *) key);
2692
2693 if (key->flatshade) {
2694 debug_printf("flatshade = 1\n");
2695 }
2696 for (i = 0; i < key->nr_cbufs; ++i) {
2697 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2698 }
2699 if (key->depth.enabled || key->stencil[0].enabled) {
2700 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2701 }
2702 if (key->depth.enabled) {
2703 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
2704 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2705 }
2706
2707 for (i = 0; i < 2; ++i) {
2708 if (key->stencil[i].enabled) {
2709 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
2710 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
2711 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
2712 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
2713 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2714 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2715 }
2716 }
2717
2718 if (key->alpha.enabled) {
2719 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
2720 }
2721
2722 if (key->occlusion_count) {
2723 debug_printf("occlusion_count = 1\n");
2724 }
2725
2726 if (key->blend.logicop_enable) {
2727 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
2728 }
2729 else if (key->blend.rt[0].blend_enable) {
2730 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
2731 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2732 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2733 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
2734 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2735 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2736 }
2737 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2738 if (key->blend.alpha_to_coverage) {
2739 debug_printf("blend.alpha_to_coverage is enabled\n");
2740 }
2741 for (i = 0; i < key->nr_samplers; ++i) {
2742 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2743 debug_printf("sampler[%u] = \n", i);
2744 debug_printf(" .wrap = %s %s %s\n",
2745 util_str_tex_wrap(sampler->wrap_s, TRUE),
2746 util_str_tex_wrap(sampler->wrap_t, TRUE),
2747 util_str_tex_wrap(sampler->wrap_r, TRUE));
2748 debug_printf(" .min_img_filter = %s\n",
2749 util_str_tex_filter(sampler->min_img_filter, TRUE));
2750 debug_printf(" .min_mip_filter = %s\n",
2751 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
2752 debug_printf(" .mag_img_filter = %s\n",
2753 util_str_tex_filter(sampler->mag_img_filter, TRUE));
2754 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2755 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
2756 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2757 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2758 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2759 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2760 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2761 }
2762 for (i = 0; i < key->nr_sampler_views; ++i) {
2763 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2764 debug_printf("texture[%u] = \n", i);
2765 debug_printf(" .format = %s\n",
2766 util_format_name(texture->format));
2767 debug_printf(" .target = %s\n",
2768 util_str_tex_target(texture->target, TRUE));
2769 debug_printf(" .level_zero_only = %u\n",
2770 texture->level_zero_only);
2771 debug_printf(" .pot = %u %u %u\n",
2772 texture->pot_width,
2773 texture->pot_height,
2774 texture->pot_depth);
2775 }
2776 }
2777
2778
2779 void
lp_debug_fs_variant(const struct lp_fragment_shader_variant * variant)2780 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2781 {
2782 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2783 variant->shader->no, variant->no);
2784 tgsi_dump(variant->shader->base.tokens, 0);
2785 dump_fs_variant_key(&variant->key);
2786 debug_printf("variant->opaque = %u\n", variant->opaque);
2787 debug_printf("\n");
2788 }
2789
2790
2791 /**
2792 * Generate a new fragment shader variant from the shader code and
2793 * other state indicated by the key.
2794 */
2795 static struct lp_fragment_shader_variant *
generate_variant(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,const struct lp_fragment_shader_variant_key * key)2796 generate_variant(struct llvmpipe_context *lp,
2797 struct lp_fragment_shader *shader,
2798 const struct lp_fragment_shader_variant_key *key)
2799 {
2800 struct lp_fragment_shader_variant *variant;
2801 const struct util_format_description *cbuf0_format_desc = NULL;
2802 boolean fullcolormask;
2803 char module_name[64];
2804
2805 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2806 if (!variant)
2807 return NULL;
2808
2809 util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
2810 shader->no, shader->variants_created);
2811
2812 variant->gallivm = gallivm_create(module_name, lp->context);
2813 if (!variant->gallivm) {
2814 FREE(variant);
2815 return NULL;
2816 }
2817
2818 variant->shader = shader;
2819 variant->list_item_global.base = variant;
2820 variant->list_item_local.base = variant;
2821 variant->no = shader->variants_created++;
2822
2823 memcpy(&variant->key, key, shader->variant_key_size);
2824
2825 /*
2826 * Determine whether we are touching all channels in the color buffer.
2827 */
2828 fullcolormask = FALSE;
2829 if (key->nr_cbufs == 1) {
2830 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2831 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2832 }
2833
2834 variant->opaque =
2835 !key->blend.logicop_enable &&
2836 !key->blend.rt[0].blend_enable &&
2837 fullcolormask &&
2838 !key->stencil[0].enabled &&
2839 !key->alpha.enabled &&
2840 !key->blend.alpha_to_coverage &&
2841 !key->depth.enabled &&
2842 !shader->info.base.uses_kill &&
2843 !shader->info.base.writes_samplemask
2844 ? TRUE : FALSE;
2845
2846 if ((shader->info.base.num_tokens <= 1) &&
2847 !key->depth.enabled && !key->stencil[0].enabled) {
2848 variant->ps_inv_multiplier = 0;
2849 } else {
2850 variant->ps_inv_multiplier = 1;
2851 }
2852
2853 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2854 lp_debug_fs_variant(variant);
2855 }
2856
2857 lp_jit_init_types(variant);
2858
2859 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2860 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2861
2862 if (variant->jit_function[RAST_WHOLE] == NULL) {
2863 if (variant->opaque) {
2864 /* Specialized shader, which doesn't need to read the color buffer. */
2865 generate_fragment(lp, shader, variant, RAST_WHOLE);
2866 }
2867 }
2868
2869 /*
2870 * Compile everything
2871 */
2872
2873 gallivm_compile_module(variant->gallivm);
2874
2875 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
2876
2877 if (variant->function[RAST_EDGE_TEST]) {
2878 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2879 gallivm_jit_function(variant->gallivm,
2880 variant->function[RAST_EDGE_TEST]);
2881 }
2882
2883 if (variant->function[RAST_WHOLE]) {
2884 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2885 gallivm_jit_function(variant->gallivm,
2886 variant->function[RAST_WHOLE]);
2887 } else if (!variant->jit_function[RAST_WHOLE]) {
2888 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2889 }
2890
2891 gallivm_free_ir(variant->gallivm);
2892
2893 return variant;
2894 }
2895
2896
2897 static void *
llvmpipe_create_fs_state(struct pipe_context * pipe,const struct pipe_shader_state * templ)2898 llvmpipe_create_fs_state(struct pipe_context *pipe,
2899 const struct pipe_shader_state *templ)
2900 {
2901 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2902 struct lp_fragment_shader *shader;
2903 int nr_samplers;
2904 int nr_sampler_views;
2905 int i;
2906
2907 shader = CALLOC_STRUCT(lp_fragment_shader);
2908 if (!shader)
2909 return NULL;
2910
2911 shader->no = fs_no++;
2912 make_empty_list(&shader->variants);
2913
2914 /* get/save the summary info for this shader */
2915 lp_build_tgsi_info(templ->tokens, &shader->info);
2916
2917 /* we need to keep a local copy of the tokens */
2918 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2919
2920 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2921 if (shader->draw_data == NULL) {
2922 FREE((void *) shader->base.tokens);
2923 FREE(shader);
2924 return NULL;
2925 }
2926
2927 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2928 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2929
2930 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2931 state[MAX2(nr_samplers, nr_sampler_views)]);
2932
2933 for (i = 0; i < shader->info.base.num_inputs; i++) {
2934 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2935 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2936
2937 switch (shader->info.base.input_interpolate[i]) {
2938 case TGSI_INTERPOLATE_CONSTANT:
2939 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2940 break;
2941 case TGSI_INTERPOLATE_LINEAR:
2942 shader->inputs[i].interp = LP_INTERP_LINEAR;
2943 break;
2944 case TGSI_INTERPOLATE_PERSPECTIVE:
2945 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2946 break;
2947 case TGSI_INTERPOLATE_COLOR:
2948 shader->inputs[i].interp = LP_INTERP_COLOR;
2949 break;
2950 default:
2951 assert(0);
2952 break;
2953 }
2954
2955 switch (shader->info.base.input_semantic_name[i]) {
2956 case TGSI_SEMANTIC_FACE:
2957 shader->inputs[i].interp = LP_INTERP_FACING;
2958 break;
2959 case TGSI_SEMANTIC_POSITION:
2960 /* Position was already emitted above
2961 */
2962 shader->inputs[i].interp = LP_INTERP_POSITION;
2963 shader->inputs[i].src_index = 0;
2964 continue;
2965 }
2966
2967 /* XXX this is a completely pointless index map... */
2968 shader->inputs[i].src_index = i+1;
2969 }
2970
2971 if (LP_DEBUG & DEBUG_TGSI) {
2972 unsigned attrib;
2973 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2974 shader->no, (void *) shader);
2975 tgsi_dump(templ->tokens, 0);
2976 debug_printf("usage masks:\n");
2977 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2978 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2979 debug_printf(" IN[%u].%s%s%s%s\n",
2980 attrib,
2981 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2982 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2983 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2984 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2985 }
2986 debug_printf("\n");
2987 }
2988
2989 return shader;
2990 }
2991
2992
2993 static void
llvmpipe_bind_fs_state(struct pipe_context * pipe,void * fs)2994 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2995 {
2996 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2997
2998 if (llvmpipe->fs == fs)
2999 return;
3000
3001 llvmpipe->fs = (struct lp_fragment_shader *) fs;
3002
3003 draw_bind_fragment_shader(llvmpipe->draw,
3004 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
3005
3006 llvmpipe->dirty |= LP_NEW_FS;
3007 }
3008
3009
3010 /**
3011 * Remove shader variant from two lists: the shader's variant list
3012 * and the context's variant list.
3013 */
3014 void
llvmpipe_remove_shader_variant(struct llvmpipe_context * lp,struct lp_fragment_shader_variant * variant)3015 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3016 struct lp_fragment_shader_variant *variant)
3017 {
3018 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3019 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3020 "v total cached %u inst %u total inst %u\n",
3021 variant->shader->no, variant->no,
3022 variant->shader->variants_created,
3023 variant->shader->variants_cached,
3024 lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3025 }
3026
3027 gallivm_destroy(variant->gallivm);
3028
3029 /* remove from shader's list */
3030 remove_from_list(&variant->list_item_local);
3031 variant->shader->variants_cached--;
3032
3033 /* remove from context's list */
3034 remove_from_list(&variant->list_item_global);
3035 lp->nr_fs_variants--;
3036 lp->nr_fs_instrs -= variant->nr_instrs;
3037
3038 FREE(variant);
3039 }
3040
3041
3042 static void
llvmpipe_delete_fs_state(struct pipe_context * pipe,void * fs)3043 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3044 {
3045 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3046 struct lp_fragment_shader *shader = fs;
3047 struct lp_fs_variant_list_item *li;
3048
3049 assert(fs != llvmpipe->fs);
3050
3051 /*
3052 * XXX: we need to flush the context until we have some sort of reference
3053 * counting in fragment shaders as they may still be binned
3054 * Flushing alone might not sufficient we need to wait on it too.
3055 */
3056 llvmpipe_finish(pipe, __FUNCTION__);
3057
3058 /* Delete all the variants */
3059 li = first_elem(&shader->variants);
3060 while(!at_end(&shader->variants, li)) {
3061 struct lp_fs_variant_list_item *next = next_elem(li);
3062 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3063 li = next;
3064 }
3065
3066 /* Delete draw module's data */
3067 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3068
3069 assert(shader->variants_cached == 0);
3070 FREE((void *) shader->base.tokens);
3071 FREE(shader);
3072 }
3073
3074
3075
3076 static void
llvmpipe_set_constant_buffer(struct pipe_context * pipe,enum pipe_shader_type shader,uint index,const struct pipe_constant_buffer * cb)3077 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3078 enum pipe_shader_type shader, uint index,
3079 const struct pipe_constant_buffer *cb)
3080 {
3081 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3082 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3083
3084 assert(shader < PIPE_SHADER_TYPES);
3085 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3086
3087 /* note: reference counting */
3088 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3089
3090 if (constants) {
3091 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3092 debug_printf("Illegal set constant without bind flag\n");
3093 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3094 }
3095 }
3096
3097 if (shader == PIPE_SHADER_VERTEX ||
3098 shader == PIPE_SHADER_GEOMETRY) {
3099 /* Pass the constants to the 'draw' module */
3100 const unsigned size = cb ? cb->buffer_size : 0;
3101 const ubyte *data;
3102
3103 if (constants) {
3104 data = (ubyte *) llvmpipe_resource_data(constants);
3105 }
3106 else if (cb && cb->user_buffer) {
3107 data = (ubyte *) cb->user_buffer;
3108 }
3109 else {
3110 data = NULL;
3111 }
3112
3113 if (data)
3114 data += cb->buffer_offset;
3115
3116 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3117 index, data, size);
3118 }
3119 else {
3120 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3121 }
3122
3123 if (cb && cb->user_buffer) {
3124 pipe_resource_reference(&constants, NULL);
3125 }
3126 }
3127
3128
3129 /**
3130 * Return the blend factor equivalent to a destination alpha of one.
3131 */
3132 static inline unsigned
force_dst_alpha_one(unsigned factor,boolean clamped_zero)3133 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3134 {
3135 switch(factor) {
3136 case PIPE_BLENDFACTOR_DST_ALPHA:
3137 return PIPE_BLENDFACTOR_ONE;
3138 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3139 return PIPE_BLENDFACTOR_ZERO;
3140 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3141 if (clamped_zero)
3142 return PIPE_BLENDFACTOR_ZERO;
3143 else
3144 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3145 }
3146
3147 return factor;
3148 }
3149
3150
3151 /**
3152 * We need to generate several variants of the fragment pipeline to match
3153 * all the combinations of the contributing state atoms.
3154 *
3155 * TODO: there is actually no reason to tie this to context state -- the
3156 * generated code could be cached globally in the screen.
3157 */
3158 static void
make_variant_key(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant_key * key)3159 make_variant_key(struct llvmpipe_context *lp,
3160 struct lp_fragment_shader *shader,
3161 struct lp_fragment_shader_variant_key *key)
3162 {
3163 unsigned i;
3164
3165 memset(key, 0, shader->variant_key_size);
3166
3167 if (lp->framebuffer.zsbuf) {
3168 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3169 const struct util_format_description *zsbuf_desc =
3170 util_format_description(zsbuf_format);
3171
3172 if (lp->depth_stencil->depth.enabled &&
3173 util_format_has_depth(zsbuf_desc)) {
3174 key->zsbuf_format = zsbuf_format;
3175 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3176 }
3177 if (lp->depth_stencil->stencil[0].enabled &&
3178 util_format_has_stencil(zsbuf_desc)) {
3179 key->zsbuf_format = zsbuf_format;
3180 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3181 }
3182 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3183 key->resource_1d = TRUE;
3184 }
3185 }
3186
3187 /*
3188 * Propagate the depth clamp setting from the rasterizer state.
3189 * depth_clip == 0 implies depth clamping is enabled.
3190 *
3191 * When clip_halfz is enabled, then always clamp the depth values.
3192 *
3193 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3194 * clamp is always active in d3d10, regardless if depth clip is
3195 * enabled or not).
3196 * (GL has an always-on [0,1] clamp on fs depth output instead
3197 * to ensure the depth values stay in range. Doesn't look like
3198 * we do that, though...)
3199 */
3200 if (lp->rasterizer->clip_halfz) {
3201 key->depth_clamp = 1;
3202 } else {
3203 key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
3204 }
3205
3206 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3207 if (!lp->framebuffer.nr_cbufs ||
3208 !lp->framebuffer.cbufs[0] ||
3209 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3210 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3211 }
3212 if(key->alpha.enabled)
3213 key->alpha.func = lp->depth_stencil->alpha.func;
3214 /* alpha.ref_value is passed in jit_context */
3215
3216 key->flatshade = lp->rasterizer->flatshade;
3217 if (lp->active_occlusion_queries) {
3218 key->occlusion_count = TRUE;
3219 }
3220
3221 if (lp->framebuffer.nr_cbufs) {
3222 memcpy(&key->blend, lp->blend, sizeof key->blend);
3223 }
3224
3225 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3226
3227 if (!key->blend.independent_blend_enable) {
3228 /* we always need independent blend otherwise the fixups below won't work */
3229 for (i = 1; i < key->nr_cbufs; i++) {
3230 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3231 }
3232 key->blend.independent_blend_enable = 1;
3233 }
3234
3235 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3236 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3237
3238 if (lp->framebuffer.cbufs[i]) {
3239 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3240 const struct util_format_description *format_desc;
3241
3242 key->cbuf_format[i] = format;
3243
3244 /*
3245 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3246 * mixing of 2d textures with height 1 and 1d textures, so make sure
3247 * we pick 1d if any cbuf or zsbuf is 1d.
3248 */
3249 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3250 key->resource_1d = TRUE;
3251 }
3252
3253 format_desc = util_format_description(format);
3254 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3255 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3256
3257 /*
3258 * Mask out color channels not present in the color buffer.
3259 */
3260 blend_rt->colormask &= util_format_colormask(format_desc);
3261
3262 /*
3263 * Disable blend for integer formats.
3264 */
3265 if (util_format_is_pure_integer(format)) {
3266 blend_rt->blend_enable = 0;
3267 }
3268
3269 /*
3270 * Our swizzled render tiles always have an alpha channel, but the
3271 * linear render target format often does not, so force here the dst
3272 * alpha to be one.
3273 *
3274 * This is not a mere optimization. Wrong results will be produced if
3275 * the dst alpha is used, the dst format does not have alpha, and the
3276 * previous rendering was not flushed from the swizzled to linear
3277 * buffer. For example, NonPowTwo DCT.
3278 *
3279 * TODO: This should be generalized to all channels for better
3280 * performance, but only alpha causes correctness issues.
3281 *
3282 * Also, force rgb/alpha func/factors match, to make AoS blending
3283 * easier.
3284 */
3285 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3286 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3287 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3288 boolean clamped_zero = !util_format_is_float(format) &&
3289 !util_format_is_snorm(format);
3290 blend_rt->rgb_src_factor =
3291 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3292 blend_rt->rgb_dst_factor =
3293 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3294 blend_rt->alpha_func = blend_rt->rgb_func;
3295 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3296 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3297 }
3298 }
3299 else {
3300 /* no color buffer for this fragment output */
3301 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3302 blend_rt->colormask = 0x0;
3303 blend_rt->blend_enable = 0;
3304 }
3305 }
3306
3307 /* This value will be the same for all the variants of a given shader:
3308 */
3309 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3310
3311 for(i = 0; i < key->nr_samplers; ++i) {
3312 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3313 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
3314 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3315 }
3316 }
3317
3318 /*
3319 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3320 * are dx10-style? Can't really have mixed opcodes, at least not
3321 * if we want to skip the holes here (without rescanning tgsi).
3322 */
3323 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3324 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3325 for(i = 0; i < key->nr_sampler_views; ++i) {
3326 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
3327 lp_sampler_static_texture_state(&key->state[i].texture_state,
3328 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3329 }
3330 }
3331 }
3332 else {
3333 key->nr_sampler_views = key->nr_samplers;
3334 for(i = 0; i < key->nr_sampler_views; ++i) {
3335 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3336 lp_sampler_static_texture_state(&key->state[i].texture_state,
3337 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3338 }
3339 }
3340 }
3341 }
3342
3343
3344
3345 /**
3346 * Update fragment shader state. This is called just prior to drawing
3347 * something when some fragment-related state has changed.
3348 */
3349 void
llvmpipe_update_fs(struct llvmpipe_context * lp)3350 llvmpipe_update_fs(struct llvmpipe_context *lp)
3351 {
3352 struct lp_fragment_shader *shader = lp->fs;
3353 struct lp_fragment_shader_variant_key key;
3354 struct lp_fragment_shader_variant *variant = NULL;
3355 struct lp_fs_variant_list_item *li;
3356
3357 make_variant_key(lp, shader, &key);
3358
3359 /* Search the variants for one which matches the key */
3360 li = first_elem(&shader->variants);
3361 while(!at_end(&shader->variants, li)) {
3362 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
3363 variant = li->base;
3364 break;
3365 }
3366 li = next_elem(li);
3367 }
3368
3369 if (variant) {
3370 /* Move this variant to the head of the list to implement LRU
3371 * deletion of shader's when we have too many.
3372 */
3373 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3374 }
3375 else {
3376 /* variant not found, create it now */
3377 int64_t t0, t1, dt;
3378 unsigned i;
3379 unsigned variants_to_cull;
3380
3381 if (LP_DEBUG & DEBUG_FS) {
3382 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3383 lp->nr_fs_variants,
3384 lp->nr_fs_instrs,
3385 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3386 }
3387
3388 /* First, check if we've exceeded the max number of shader variants.
3389 * If so, free 6.25% of them (the least recently used ones).
3390 */
3391 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
3392
3393 if (variants_to_cull ||
3394 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3395 struct pipe_context *pipe = &lp->pipe;
3396
3397 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
3398 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3399 "\t%u instrs,\t%u instrs/variant\n",
3400 shader->variants_cached,
3401 lp->nr_fs_variants, lp->nr_fs_instrs,
3402 lp->nr_fs_instrs / lp->nr_fs_variants);
3403 }
3404
3405 /*
3406 * XXX: we need to flush the context until we have some sort of
3407 * reference counting in fragment shaders as they may still be binned
3408 * Flushing alone might not be sufficient we need to wait on it too.
3409 */
3410 llvmpipe_finish(pipe, __FUNCTION__);
3411
3412 /*
3413 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3414 * number of shader variants (potentially all of them) could be
3415 * pending for destruction on flush.
3416 */
3417
3418 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3419 struct lp_fs_variant_list_item *item;
3420 if (is_empty_list(&lp->fs_variants_list)) {
3421 break;
3422 }
3423 item = last_elem(&lp->fs_variants_list);
3424 assert(item);
3425 assert(item->base);
3426 llvmpipe_remove_shader_variant(lp, item->base);
3427 }
3428 }
3429
3430 /*
3431 * Generate the new variant.
3432 */
3433 t0 = os_time_get();
3434 variant = generate_variant(lp, shader, &key);
3435 t1 = os_time_get();
3436 dt = t1 - t0;
3437 LP_COUNT_ADD(llvm_compile_time, dt);
3438 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3439
3440 /* Put the new variant into the list */
3441 if (variant) {
3442 insert_at_head(&shader->variants, &variant->list_item_local);
3443 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3444 lp->nr_fs_variants++;
3445 lp->nr_fs_instrs += variant->nr_instrs;
3446 shader->variants_cached++;
3447 }
3448 }
3449
3450 /* Bind this variant */
3451 lp_setup_set_fs_variant(lp->setup, variant);
3452 }
3453
3454
3455
3456
3457
3458 void
llvmpipe_init_fs_funcs(struct llvmpipe_context * llvmpipe)3459 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3460 {
3461 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3462 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3463 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3464
3465 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3466 }
3467
3468 /*
3469 * Rasterization is disabled if there is no pixel shader and
3470 * both depth and stencil testing are disabled:
3471 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3472 */
3473 boolean
llvmpipe_rasterization_disabled(struct llvmpipe_context * lp)3474 llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
3475 {
3476 boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
3477
3478 return (null_fs &&
3479 !lp->depth_stencil->depth.enabled &&
3480 !lp->depth_stencil->stencil[0].enabled);
3481 }
3482