projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: don't allow branch to end for early Z with multisample
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /**
30 * @file
31 * Code generate the whole fragment pipeline.
32 *
33 * The fragment pipeline consists of the following stages:
34 * - early depth test
35 * - fragment shader
36 * - alpha test
37 * - depth/stencil test
38 * - blending
39 *
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
45 *
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
51 *
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
56 *
57 * @author Jose Fonseca <jfonseca@vmware.com>
58 */
59
60 #include <limits.h>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/format/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_nir.h"
84 #include "gallivm/lp_bld_swizzle.h"
85 #include "gallivm/lp_bld_flow.h"
86 #include "gallivm/lp_bld_debug.h"
87 #include "gallivm/lp_bld_arit.h"
88 #include "gallivm/lp_bld_bitarit.h"
89 #include "gallivm/lp_bld_pack.h"
90 #include "gallivm/lp_bld_format.h"
91 #include "gallivm/lp_bld_quad.h"
92
93 #include "lp_bld_alpha.h"
94 #include "lp_bld_blend.h"
95 #include "lp_bld_depth.h"
96 #include "lp_bld_interp.h"
97 #include "lp_context.h"
98 #include "lp_debug.h"
99 #include "lp_perf.h"
100 #include "lp_setup.h"
101 #include "lp_state.h"
102 #include "lp_tex_sample.h"
103 #include "lp_flush.h"
104 #include "lp_state_fs.h"
105 #include "lp_rast.h"
106 #include "nir/nir_to_tgsi_info.h"
107
108 /** Fragment shader number (for debugging) */
109 static unsigned fs_no = 0;
110
111
112 /**
113 * Expand the relevant bits of mask_input to a n*4-dword mask for the
114 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
115 * quad mask vector to 0 or ~0.
116 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
117 * quad arguments with fs length 8.
118 *
119 * \param first_quad which quad(s) of the quad group to test, in [0,3]
120 * \param mask_input bitwise mask for the whole 4x4 stamp
121 */
122 static LLVMValueRef
123 generate_quad_mask(struct gallivm_state *gallivm,
124 struct lp_type fs_type,
125 unsigned first_quad,
126 unsigned sample,
127 LLVMValueRef mask_input) /* int64 */
128 {
129 LLVMBuilderRef builder = gallivm->builder;
130 struct lp_type mask_type;
131 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
132 LLVMValueRef bits[16];
133 LLVMValueRef mask, bits_vec;
134 int shift, i;
135
136 /*
137 * XXX: We'll need a different path for 16 x u8
138 */
139 assert(fs_type.width == 32);
140 assert(fs_type.length <= ARRAY_SIZE(bits));
141 mask_type = lp_int_type(fs_type);
142
143 /*
144 * mask_input >>= (quad * 4)
145 */
146 switch (first_quad) {
147 case 0:
148 shift = 0;
149 break;
150 case 1:
151 assert(fs_type.length == 4);
152 shift = 2;
153 break;
154 case 2:
155 shift = 8;
156 break;
157 case 3:
158 assert(fs_type.length == 4);
159 shift = 10;
160 break;
161 default:
162 assert(0);
163 shift = 0;
164 }
165
166 mask_input = LLVMBuildLShr(builder, mask_input, lp_build_const_int64(gallivm, 16 * sample), "");
167 mask_input = LLVMBuildTrunc(builder, mask_input,
168 i32t, "");
169 mask_input = LLVMBuildAnd(builder, mask_input, lp_build_const_int32(gallivm, 0xffff), "");
170
171 mask_input = LLVMBuildLShr(builder,
172 mask_input,
173 LLVMConstInt(i32t, shift, 0),
174 "");
175
176 /*
177 * mask = { mask_input & (1 << i), for i in [0,3] }
178 */
179 mask = lp_build_broadcast(gallivm,
180 lp_build_vec_type(gallivm, mask_type),
181 mask_input);
182
183 for (i = 0; i < fs_type.length / 4; i++) {
184 unsigned j = 2 * (i % 2) + (i / 2) * 8;
185 bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
186 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
187 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
188 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
189 }
190 bits_vec = LLVMConstVector(bits, fs_type.length);
191 mask = LLVMBuildAnd(builder, mask, bits_vec, "");
192
193 /*
194 * mask = mask == bits ? ~0 : 0
195 */
196 mask = lp_build_compare(gallivm,
197 mask_type, PIPE_FUNC_EQUAL,
198 mask, bits_vec);
199
200 return mask;
201 }
202
203
204 #define EARLY_DEPTH_TEST 0x1
205 #define LATE_DEPTH_TEST 0x2
206 #define EARLY_DEPTH_WRITE 0x4
207 #define LATE_DEPTH_WRITE 0x8
208
209 static int
210 find_output_by_semantic( const struct tgsi_shader_info *info,
211 unsigned semantic,
212 unsigned index )
213 {
214 int i;
215
216 for (i = 0; i < info->num_outputs; i++)
217 if (info->output_semantic_name[i] == semantic &&
218 info->output_semantic_index[i] == index)
219 return i;
220
221 return -1;
222 }
223
224
225 /**
226 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
227 */
228 static LLVMValueRef
229 lp_llvm_viewport(LLVMValueRef context_ptr,
230 struct gallivm_state *gallivm,
231 LLVMValueRef viewport_index)
232 {
233 LLVMBuilderRef builder = gallivm->builder;
234 LLVMValueRef ptr;
235 LLVMValueRef res;
236 struct lp_type viewport_type =
237 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
238
239 ptr = lp_jit_context_viewports(gallivm, context_ptr);
240 ptr = LLVMBuildPointerCast(builder, ptr,
241 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
242
243 res = lp_build_pointer_get(builder, ptr, viewport_index);
244
245 return res;
246 }
247
248
249 static LLVMValueRef
250 lp_build_depth_clamp(struct gallivm_state *gallivm,
251 LLVMBuilderRef builder,
252 struct lp_type type,
253 LLVMValueRef context_ptr,
254 LLVMValueRef thread_data_ptr,
255 LLVMValueRef z)
256 {
257 LLVMValueRef viewport, min_depth, max_depth;
258 LLVMValueRef viewport_index;
259 struct lp_build_context f32_bld;
260
261 assert(type.floating);
262 lp_build_context_init(&f32_bld, gallivm, type);
263
264 /*
265 * Assumes clamping of the viewport index will occur in setup/gs. Value
266 * is passed through the rasterization stage via lp_rast_shader_inputs.
267 *
268 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
269 * semantics.
270 */
271 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
272 thread_data_ptr);
273
274 /*
275 * Load the min and max depth from the lp_jit_context.viewports
276 * array of lp_jit_viewport structures.
277 */
278 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
279
280 /* viewports[viewport_index].min_depth */
281 min_depth = LLVMBuildExtractElement(builder, viewport,
282 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
283 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
284
285 /* viewports[viewport_index].max_depth */
286 max_depth = LLVMBuildExtractElement(builder, viewport,
287 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
288 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
289
290 /*
291 * Clamp to the min and max depth values for the given viewport.
292 */
293 return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
294 }
295
296 static void
297 lp_build_sample_alpha_to_coverage(struct gallivm_state *gallivm,
298 struct lp_type type,
299 unsigned coverage_samples,
300 LLVMValueRef num_loop,
301 LLVMValueRef loop_counter,
302 LLVMValueRef coverage_mask_store,
303 LLVMValueRef alpha)
304 {
305 struct lp_build_context bld;
306 LLVMBuilderRef builder = gallivm->builder;
307 float step = 1.0 / coverage_samples;
308
309 lp_build_context_init(&bld, gallivm, type);
310 for (unsigned s = 0; s < coverage_samples; s++) {
311 LLVMValueRef alpha_ref_value = lp_build_const_vec(gallivm, type, step * s);
312 LLVMValueRef test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
313
314 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, lp_build_const_int32(gallivm, s), num_loop, "");
315 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_counter, "");
316 LLVMValueRef s_mask_ptr = LLVMBuildGEP(builder, coverage_mask_store, &s_mask_idx, 1, "");
317 LLVMValueRef s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
318 s_mask = LLVMBuildAnd(builder, s_mask, test, "");
319 LLVMBuildStore(builder, s_mask, s_mask_ptr);
320 }
321 }
322
323 /**
324 * Generate the fragment shader, depth/stencil test, and alpha tests.
325 */
326 static void
327 generate_fs_loop(struct gallivm_state *gallivm,
328 struct lp_fragment_shader *shader,
329 const struct lp_fragment_shader_variant_key *key,
330 LLVMBuilderRef builder,
331 struct lp_type type,
332 LLVMValueRef context_ptr,
333 LLVMValueRef sample_pos_array,
334 LLVMValueRef num_loop,
335 struct lp_build_interp_soa_context *interp,
336 const struct lp_build_sampler_soa *sampler,
337 const struct lp_build_image_soa *image,
338 LLVMValueRef mask_store,
339 LLVMValueRef (*out_color)[4],
340 LLVMValueRef depth_base_ptr,
341 LLVMValueRef depth_stride,
342 LLVMValueRef depth_sample_stride,
343 LLVMValueRef facing,
344 LLVMValueRef thread_data_ptr)
345 {
346 const struct util_format_description *zs_format_desc = NULL;
347 const struct tgsi_token *tokens = shader->base.tokens;
348 struct lp_type int_type = lp_int_type(type);
349 LLVMTypeRef vec_type, int_vec_type;
350 LLVMValueRef mask_ptr = NULL, mask_val = NULL;
351 LLVMValueRef consts_ptr, num_consts_ptr;
352 LLVMValueRef ssbo_ptr, num_ssbo_ptr;
353 LLVMValueRef z;
354 LLVMValueRef z_value, s_value;
355 LLVMValueRef z_fb, s_fb;
356 LLVMValueRef depth_ptr;
357 LLVMValueRef stencil_refs[2];
358 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
359 LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
360 struct lp_build_for_loop_state loop_state, sample_loop_state;
361 struct lp_build_mask_context mask;
362 /*
363 * TODO: figure out if simple_shader optimization is really worthwile to
364 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
365 * code since tests tend to take another codepath than real shaders.
366 */
367 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
368 shader->info.base.num_inputs < 3 &&
369 shader->info.base.num_instructions < 8) && 0;
370 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
371 util_blend_state_is_dual(&key->blend, 0);
372 unsigned attrib;
373 unsigned chan;
374 unsigned cbuf;
375 unsigned depth_mode;
376
377 struct lp_bld_tgsi_system_values system_values;
378
379 memset(&system_values, 0, sizeof(system_values));
380
381 /* truncate then sign extend. */
382 system_values.front_facing = LLVMBuildTrunc(gallivm->builder, facing, LLVMInt1TypeInContext(gallivm->context), "");
383 system_values.front_facing = LLVMBuildSExt(gallivm->builder, system_values.front_facing, LLVMInt32TypeInContext(gallivm->context), "");
384
385 if (key->depth.enabled ||
386 key->stencil[0].enabled) {
387
388 zs_format_desc = util_format_description(key->zsbuf_format);
389 assert(zs_format_desc);
390
391 if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
392 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
393 else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
394 if (shader->info.base.writes_memory)
395 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
396 else if (key->alpha.enabled ||
397 key->blend.alpha_to_coverage ||
398 shader->info.base.uses_kill ||
399 shader->info.base.writes_samplemask) {
400 /* With alpha test and kill, can do the depth test early
401 * and hopefully eliminate some quads. But need to do a
402 * special deferred depth write once the final mask value
403 * is known. This only works though if there's either no
404 * stencil test or the stencil value isn't written.
405 */
406 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
407 (key->stencil[1].enabled &&
408 key->stencil[1].writemask)))
409 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
410 else
411 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
412 }
413 else
414 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
415 }
416 else {
417 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
418 }
419
420 if (!(key->depth.enabled && key->depth.writemask) &&
421 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
422 (key->stencil[1].enabled &&
423 key->stencil[1].writemask))))
424 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
425 }
426 else {
427 depth_mode = 0;
428 }
429
430 vec_type = lp_build_vec_type(gallivm, type);
431 int_vec_type = lp_build_vec_type(gallivm, int_type);
432
433 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
434 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
435 /* convert scalar stencil refs into vectors */
436 stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
437 stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
438
439 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
440 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
441
442 ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
443 num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
444
445 memset(outputs, 0, sizeof outputs);
446
447 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
448 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
449 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
450 lp_build_vec_type(gallivm,
451 type),
452 num_loop, "color");
453 }
454 }
455 if (dual_source_blend) {
456 assert(key->nr_cbufs <= 1);
457 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
458 out_color[1][chan] = lp_build_array_alloca(gallivm,
459 lp_build_vec_type(gallivm,
460 type),
461 num_loop, "color1");
462 }
463 }
464
465 lp_build_for_loop_begin(&loop_state, gallivm,
466 lp_build_const_int32(gallivm, 0),
467 LLVMIntULT,
468 num_loop,
469 lp_build_const_int32(gallivm, 1));
470
471 if (key->multisample) {
472 /* create shader execution mask by combining all sample masks. */
473 for (unsigned s = 0; s < key->coverage_samples; s++) {
474 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
475 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
476 LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
477 if (s == 0)
478 mask_val = s_mask;
479 else
480 mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
481 }
482 } else {
483 mask_ptr = LLVMBuildGEP(builder, mask_store,
484 &loop_state.counter, 1, "mask_ptr");
485 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
486 }
487
488 /* 'mask' will control execution based on quad's pixel alive/killed state */
489 lp_build_mask_begin(&mask, gallivm, type, mask_val);
490
491 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
492 lp_build_mask_check(&mask);
493
494 /* Create storage for recombining sample masks after early Z pass. */
495 LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
496 LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
497
498 LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
499 LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
500 LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
501 LLVMTypeRef z_type = NULL, z_fb_type = NULL;
502
503 /* Run early depth once per sample */
504 if (key->multisample) {
505
506 if (zs_format_desc) {
507 struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
508 struct lp_type z_type = zs_type;
509 struct lp_type s_type = zs_type;
510 if (zs_format_desc->block.bits < type.width)
511 z_type.width = type.width;
512 else if (zs_format_desc->block.bits > 32) {
513 z_type.width = z_type.width / 2;
514 s_type.width = s_type.width / 2;
515 s_type.floating = 0;
516 }
517 z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
518 zs_samples, "z_sample_store");
519 s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
520 zs_samples, "s_sample_store");
521 z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
522 zs_samples, "z_fb_store");
523 s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
524 zs_samples, "s_fb_store");
525 }
526 lp_build_for_loop_begin(&sample_loop_state, gallivm,
527 lp_build_const_int32(gallivm, 0),
528 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
529 lp_build_const_int32(gallivm, 1));
530
531 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
532 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
533 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
534
535 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
536 s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
537 }
538
539
540 /* for multisample Z needs to be interpolated at sample points for testing. */
541 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
542 z = interp->pos[2];
543
544 depth_ptr = depth_base_ptr;
545 if (key->multisample) {
546 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
547 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
548 }
549
550 if (depth_mode & EARLY_DEPTH_TEST) {
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 lp_build_depth_stencil_load_swizzled(gallivm, type,
559 zs_format_desc, key->resource_1d,
560 depth_ptr, depth_stride,
561 &z_fb, &s_fb, loop_state.counter);
562 lp_build_depth_stencil_test(gallivm,
563 &key->depth,
564 key->stencil,
565 type,
566 zs_format_desc,
567 key->multisample ? NULL : &mask,
568 &s_mask,
569 stencil_refs,
570 z, z_fb, s_fb,
571 facing,
572 &z_value, &s_value,
573 !simple_shader && !key->multisample);
574
575 if (depth_mode & EARLY_DEPTH_WRITE) {
576 lp_build_depth_stencil_write_swizzled(gallivm, type,
577 zs_format_desc, key->resource_1d,
578 NULL, NULL, NULL, loop_state.counter,
579 depth_ptr, depth_stride,
580 z_value, s_value);
581 }
582 /*
583 * Note mask check if stencil is enabled must be after ds write not after
584 * stencil test otherwise new stencil values may not get written if all
585 * fragments got killed by depth/stencil test.
586 */
587 if (!simple_shader && key->stencil[0].enabled && !key->multisample)
588 lp_build_mask_check(&mask);
589
590 if (key->multisample) {
591 z_fb_type = LLVMTypeOf(z_fb);
592 z_type = LLVMTypeOf(z_value);
593 lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
594 lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
595 lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
596 lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
597 }
598 }
599
600 if (key->multisample) {
601 /*
602 * Store the post-early Z coverage mask.
603 * Recombine the resulting coverage masks post early Z into the fragment
604 * shader execution mask.
605 */
606 LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
607 tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
608 LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
609
610 LLVMBuildStore(builder, s_mask, s_mask_ptr);
611
612 lp_build_for_loop_end(&sample_loop_state);
613
614 /* recombined all the coverage masks in the shader exec mask. */
615 tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
616 lp_build_mask_update(&mask, tmp_s_mask_or);
617
618 /* for multisample Z needs to be re interpolated at pixel center */
619 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
620 }
621
622 LLVMValueRef out_sample_mask_storage = NULL;
623 if (shader->info.base.writes_samplemask) {
624 out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
625 }
626 system_values.sample_pos = sample_pos_array;
627
628 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, NULL, NULL);
629
630 struct lp_build_tgsi_params params;
631 memset(&params, 0, sizeof(params));
632
633 params.type = type;
634 params.mask = &mask;
635 params.consts_ptr = consts_ptr;
636 params.const_sizes_ptr = num_consts_ptr;
637 params.system_values = &system_values;
638 params.inputs = interp->inputs;
639 params.context_ptr = context_ptr;
640 params.thread_data_ptr = thread_data_ptr;
641 params.sampler = sampler;
642 params.info = &shader->info.base;
643 params.ssbo_ptr = ssbo_ptr;
644 params.ssbo_sizes_ptr = num_ssbo_ptr;
645 params.image = image;
646
647 /* Build the actual shader */
648 if (shader->base.type == PIPE_SHADER_IR_TGSI)
649 lp_build_tgsi_soa(gallivm, tokens, &params,
650 outputs);
651 else
652 lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
653 outputs);
654
655 /* Alpha test */
656 if (key->alpha.enabled) {
657 int color0 = find_output_by_semantic(&shader->info.base,
658 TGSI_SEMANTIC_COLOR,
659 0);
660
661 if (color0 != -1 && outputs[color0][3]) {
662 const struct util_format_description *cbuf_format_desc;
663 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
664 LLVMValueRef alpha_ref_value;
665
666 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
667 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
668
669 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
670
671 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
672 &mask, alpha, alpha_ref_value,
673 (depth_mode & LATE_DEPTH_TEST) != 0);
674 }
675 }
676
677 /* Emulate Alpha to Coverage with Alpha test */
678 if (key->blend.alpha_to_coverage) {
679 int color0 = find_output_by_semantic(&shader->info.base,
680 TGSI_SEMANTIC_COLOR,
681 0);
682
683 if (color0 != -1 && outputs[color0][3]) {
684 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
685
686 if (!key->multisample) {
687 lp_build_alpha_to_coverage(gallivm, type,
688 &mask, alpha,
689 (depth_mode & LATE_DEPTH_TEST) != 0);
690 } else {
691 lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
692 loop_state.counter,
693 mask_store, alpha);
694 }
695 }
696 }
697 if (key->blend.alpha_to_one && key->multisample) {
698 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {
699 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
700 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
701 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
702 if (outputs[cbuf][3]) {
703 LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
704 }
705 }
706 }
707 if (shader->info.base.writes_samplemask) {
708 LLVMValueRef output_smask = NULL;
709 int smaski = find_output_by_semantic(&shader->info.base,
710 TGSI_SEMANTIC_SAMPLEMASK,
711 0);
712 struct lp_build_context smask_bld;
713 lp_build_context_init(&smask_bld, gallivm, int_type);
714
715 assert(smaski >= 0);
716 output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
717 /*
718 * Pixel is alive according to the first sample in the mask.
719 */
720 output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
721 if (!key->multisample) {
722 output_smask = lp_build_and(&smask_bld, output_smask, smask_bld.one);
723 output_smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, output_smask, smask_bld.zero);
724 lp_build_mask_update(&mask, output_smask);
725 }
726 LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
727 }
728
729 if (key->multisample) {
730 /* execute depth test for each sample */
731 lp_build_for_loop_begin(&sample_loop_state, gallivm,
732 lp_build_const_int32(gallivm, 0),
733 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
734 lp_build_const_int32(gallivm, 1));
735
736 /* load the per-sample coverage mask */
737 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
738 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
739 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
740
741 /* combine the execution mask post fragment shader with the coverage mask. */
742 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
743 s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
744
745 /* if the shader writes sample mask use that */
746 if (shader->info.base.writes_samplemask) {
747 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
748 out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
749 LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
750 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
751 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
752 smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
753
754 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
755 }
756 }
757
758 depth_ptr = depth_base_ptr;
759 if (key->multisample) {
760 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
761 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
762 }
763
764 /* Late Z test */
765 if (depth_mode & LATE_DEPTH_TEST) {
766 int pos0 = find_output_by_semantic(&shader->info.base,
767 TGSI_SEMANTIC_POSITION,
768 0);
769 int s_out = find_output_by_semantic(&shader->info.base,
770 TGSI_SEMANTIC_STENCIL,
771 0);
772 if (pos0 != -1 && outputs[pos0][2]) {
773 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
774 }
775 /*
776 * Clamp according to ARB_depth_clamp semantics.
777 */
778 if (key->depth_clamp) {
779 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
780 thread_data_ptr, z);
781 }
782
783 if (s_out != -1 && outputs[s_out][1]) {
784 /* there's only one value, and spec says to discard additional bits */
785 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
786 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
787 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
788 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
789 stencil_refs[1] = stencil_refs[0];
790 }
791
792 lp_build_depth_stencil_load_swizzled(gallivm, type,
793 zs_format_desc, key->resource_1d,
794 depth_ptr, depth_stride,
795 &z_fb, &s_fb, loop_state.counter);
796
797 lp_build_depth_stencil_test(gallivm,
798 &key->depth,
799 key->stencil,
800 type,
801 zs_format_desc,
802 key->multisample ? NULL : &mask,
803 &s_mask,
804 stencil_refs,
805 z, z_fb, s_fb,
806 facing,
807 &z_value, &s_value,
808 !simple_shader);
809 /* Late Z write */
810 if (depth_mode & LATE_DEPTH_WRITE) {
811 lp_build_depth_stencil_write_swizzled(gallivm, type,
812 zs_format_desc, key->resource_1d,
813 NULL, NULL, NULL, loop_state.counter,
814 depth_ptr, depth_stride,
815 z_value, s_value);
816 }
817 }
818 else if ((depth_mode & EARLY_DEPTH_TEST) &&
819 (depth_mode & LATE_DEPTH_WRITE))
820 {
821 /* Need to apply a reduced mask to the depth write. Reload the
822 * depth value, update from zs_value with the new mask value and
823 * write that out.
824 */
825 if (key->multisample) {
826 z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
827 s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
828 z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
829 s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
830 }
831 lp_build_depth_stencil_write_swizzled(gallivm, type,
832 zs_format_desc, key->resource_1d,
833 key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
834 depth_ptr, depth_stride,
835 z_value, s_value);
836 }
837
838 if (key->multisample) {
839 /* store the sample mask for this loop */
840 LLVMBuildStore(builder, s_mask, s_mask_ptr);
841 lp_build_for_loop_end(&sample_loop_state);
842 }
843
844 /* Color write */
845 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
846 {
847 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
848 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
849 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
850 {
851 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
852 if(outputs[attrib][chan]) {
853 /* XXX: just initialize outputs to point at colors[] and
854 * skip this.
855 */
856 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
857 LLVMValueRef color_ptr;
858 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
859 &loop_state.counter, 1, "");
860 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
861 LLVMBuildStore(builder, out, color_ptr);
862 }
863 }
864 }
865 }
866
867 if (key->occlusion_count) {
868 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
869 lp_build_name(counter, "counter");
870 lp_build_occlusion_count(gallivm, type,
871 lp_build_mask_value(&mask), counter);
872 }
873
874 mask_val = lp_build_mask_end(&mask);
875 if (!key->multisample)
876 LLVMBuildStore(builder, mask_val, mask_ptr);
877 lp_build_for_loop_end(&loop_state);
878 }
879
880
881 /**
882 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
883 *
884 * Fragment Shader outputs pixels in small 2x2 blocks
885 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
886 *
887 * However in memory pixels are stored in rows
888 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
889 *
890 * @param type fragment shader type (4x or 8x float)
891 * @param num_fs number of fs_src
892 * @param is_1d whether we're outputting to a 1d resource
893 * @param dst_channels number of output channels
894 * @param fs_src output from fragment shader
895 * @param dst pointer to store result
896 * @param pad_inline is channel padding inline or at end of row
897 * @return the number of dsts
898 */
899 static int
900 generate_fs_twiddle(struct gallivm_state *gallivm,
901 struct lp_type type,
902 unsigned num_fs,
903 unsigned dst_channels,
904 LLVMValueRef fs_src[][4],
905 LLVMValueRef* dst,
906 bool pad_inline)
907 {
908 LLVMValueRef src[16];
909
910 bool swizzle_pad;
911 bool twiddle;
912 bool split;
913
914 unsigned pixels = type.length / 4;
915 unsigned reorder_group;
916 unsigned src_channels;
917 unsigned src_count;
918 unsigned i;
919
920 src_channels = dst_channels < 3 ? dst_channels : 4;
921 src_count = num_fs * src_channels;
922
923 assert(pixels == 2 || pixels == 1);
924 assert(num_fs * src_channels <= ARRAY_SIZE(src));
925
926 /*
927 * Transpose from SoA -> AoS
928 */
929 for (i = 0; i < num_fs; ++i) {
930 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
931 }
932
933 /*
934 * Pick transformation options
935 */
936 swizzle_pad = false;
937 twiddle = false;
938 split = false;
939 reorder_group = 0;
940
941 if (dst_channels == 1) {
942 twiddle = true;
943
944 if (pixels == 2) {
945 split = true;
946 }
947 } else if (dst_channels == 2) {
948 if (pixels == 1) {
949 reorder_group = 1;
950 }
951 } else if (dst_channels > 2) {
952 if (pixels == 1) {
953 reorder_group = 2;
954 } else {
955 twiddle = true;
956 }
957
958 if (!pad_inline && dst_channels == 3 && pixels > 1) {
959 swizzle_pad = true;
960 }
961 }
962
963 /*
964 * Split the src in half
965 */
966 if (split) {
967 for (i = num_fs; i > 0; --i) {
968 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
969 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
970 }
971
972 src_count *= 2;
973 type.length = 4;
974 }
975
976 /*
977 * Ensure pixels are in memory order
978 */
979 if (reorder_group) {
980 /* Twiddle pixels by reordering the array, e.g.:
981 *
982 * src_count = 8 -> 0 2 1 3 4 6 5 7
983 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
984 */
985 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
986
987 for (i = 0; i < src_count; ++i) {
988 unsigned group = i / reorder_group;
989 unsigned block = (group / 4) * 4 * reorder_group;
990 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
991 dst[i] = src[j];
992 }
993 } else if (twiddle) {
994 /* Twiddle pixels across elements of array */
995 /*
996 * XXX: we should avoid this in some cases, but would need to tell
997 * lp_build_conv to reorder (or deal with it ourselves).
998 */
999 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1000 } else {
1001 /* Do nothing */
1002 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1003 }
1004
1005 /*
1006 * Moves any padding between pixels to the end
1007 * e.g. RGBXRGBX -> RGBRGBXX
1008 */
1009 if (swizzle_pad) {
1010 unsigned char swizzles[16];
1011 unsigned elems = pixels * dst_channels;
1012
1013 for (i = 0; i < type.length; ++i) {
1014 if (i < elems)
1015 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1016 else
1017 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1018 }
1019
1020 for (i = 0; i < src_count; ++i) {
1021 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1022 }
1023 }
1024
1025 return src_count;
1026 }
1027
1028
1029 /*
1030 * Untwiddle and transpose, much like the above.
1031 * However, this is after conversion, so we get packed vectors.
1032 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1033 * the vectors will look like:
1034 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1035 * be swizzled here). Extending to 16bit should be trivial.
1036 * Should also be extended to handle twice wide vectors with AVX2...
1037 */
1038 static void
1039 fs_twiddle_transpose(struct gallivm_state *gallivm,
1040 struct lp_type type,
1041 LLVMValueRef *src,
1042 unsigned src_count,
1043 LLVMValueRef *dst)
1044 {
1045 unsigned i, j;
1046 struct lp_type type64, type16, type32;
1047 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1048 LLVMBuilderRef builder = gallivm->builder;
1049 LLVMValueRef tmp[4], shuf[8];
1050 for (j = 0; j < 2; j++) {
1051 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1052 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1053 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1054 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1055 }
1056
1057 assert(src_count == 4 || src_count == 2 || src_count == 1);
1058 assert(type.width == 8);
1059 assert(type.length == 16);
1060
1061 type8_t = lp_build_vec_type(gallivm, type);
1062
1063 type64 = type;
1064 type64.length /= 8;
1065 type64.width *= 8;
1066 type64_t = lp_build_vec_type(gallivm, type64);
1067
1068 type16 = type;
1069 type16.length /= 2;
1070 type16.width *= 2;
1071 type16_t = lp_build_vec_type(gallivm, type16);
1072
1073 type32 = type;
1074 type32.length /= 4;
1075 type32.width *= 4;
1076 type32_t = lp_build_vec_type(gallivm, type32);
1077
1078 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1079
1080 if (src_count == 1) {
1081 /* transpose was no-op, just untwiddle */
1082 LLVMValueRef shuf_vec;
1083 shuf_vec = LLVMConstVector(shuf, 8);
1084 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1085 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1086 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1087 } else if (src_count == 2) {
1088 LLVMValueRef shuf_vec;
1089 shuf_vec = LLVMConstVector(shuf, 4);
1090
1091 for (i = 0; i < 2; i++) {
1092 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1093 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1094 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1095 }
1096 } else {
1097 for (j = 0; j < 2; j++) {
1098 LLVMValueRef lo, hi, lo2, hi2;
1099 /*
1100 * Note that if we only really have 3 valid channels (rgb)
1101 * and we don't need alpha we could substitute a undef here
1102 * for the respective channel (causing llvm to drop conversion
1103 * for alpha).
1104 */
1105 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1106 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1107 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1108 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1109 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1110 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1111 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1112 }
1113 }
1114 }
1115
1116
1117 /**
1118 * Load an unswizzled block of pixels from memory
1119 */
1120 static void
1121 load_unswizzled_block(struct gallivm_state *gallivm,
1122 LLVMValueRef base_ptr,
1123 LLVMValueRef stride,
1124 unsigned block_width,
1125 unsigned block_height,
1126 LLVMValueRef* dst,
1127 struct lp_type dst_type,
1128 unsigned dst_count,
1129 unsigned dst_alignment)
1130 {
1131 LLVMBuilderRef builder = gallivm->builder;
1132 unsigned row_size = dst_count / block_height;
1133 unsigned i;
1134
1135 /* Ensure block exactly fits into dst */
1136 assert((block_width * block_height) % dst_count == 0);
1137
1138 for (i = 0; i < dst_count; ++i) {
1139 unsigned x = i % row_size;
1140 unsigned y = i / row_size;
1141
1142 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1143 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1144
1145 LLVMValueRef gep[2];
1146 LLVMValueRef dst_ptr;
1147
1148 gep[0] = lp_build_const_int32(gallivm, 0);
1149 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1150
1151 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1152 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1153 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1154
1155 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1156
1157 LLVMSetAlignment(dst[i], dst_alignment);
1158 }
1159 }
1160
1161
1162 /**
1163 * Store an unswizzled block of pixels to memory
1164 */
1165 static void
1166 store_unswizzled_block(struct gallivm_state *gallivm,
1167 LLVMValueRef base_ptr,
1168 LLVMValueRef stride,
1169 unsigned block_width,
1170 unsigned block_height,
1171 LLVMValueRef* src,
1172 struct lp_type src_type,
1173 unsigned src_count,
1174 unsigned src_alignment)
1175 {
1176 LLVMBuilderRef builder = gallivm->builder;
1177 unsigned row_size = src_count / block_height;
1178 unsigned i;
1179
1180 /* Ensure src exactly fits into block */
1181 assert((block_width * block_height) % src_count == 0);
1182
1183 for (i = 0; i < src_count; ++i) {
1184 unsigned x = i % row_size;
1185 unsigned y = i / row_size;
1186
1187 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1188 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1189
1190 LLVMValueRef gep[2];
1191 LLVMValueRef src_ptr;
1192
1193 gep[0] = lp_build_const_int32(gallivm, 0);
1194 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1195
1196 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1197 src_ptr = LLVMBuildBitCast(builder, src_ptr,
1198 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1199
1200 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1201
1202 LLVMSetAlignment(src_ptr, src_alignment);
1203 }
1204 }
1205
1206
1207 /**
1208 * Checks if a format description is an arithmetic format
1209 *
1210 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1211 */
1212 static inline boolean
1213 is_arithmetic_format(const struct util_format_description *format_desc)
1214 {
1215 boolean arith = false;
1216 unsigned i;
1217
1218 for (i = 0; i < format_desc->nr_channels; ++i) {
1219 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1220 arith |= (format_desc->channel[i].size % 8) != 0;
1221 }
1222
1223 return arith;
1224 }
1225
1226
1227 /**
1228 * Checks if this format requires special handling due to required expansion
1229 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1230 * SoA conversion.
1231 */
1232 static inline boolean
1233 format_expands_to_float_soa(const struct util_format_description *format_desc)
1234 {
1235 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
1236 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1237 return true;
1238 }
1239 return false;
1240 }
1241
1242
1243 /**
1244 * Retrieves the type representing the memory layout for a format
1245 *
1246 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1247 */
1248 static inline void
1249 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1250 struct lp_type* type)
1251 {
1252 unsigned i;
1253 unsigned chan;
1254
1255 if (format_expands_to_float_soa(format_desc)) {
1256 /* just make this a uint with width of block */
1257 type->floating = false;
1258 type->fixed = false;
1259 type->sign = false;
1260 type->norm = false;
1261 type->width = format_desc->block.bits;
1262 type->length = 1;
1263 return;
1264 }
1265
1266 for (i = 0; i < 4; i++)
1267 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1268 break;
1269 chan = i;
1270
1271 memset(type, 0, sizeof(struct lp_type));
1272 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1273 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1274 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1275 type->norm = format_desc->channel[chan].normalized;
1276
1277 if (is_arithmetic_format(format_desc)) {
1278 type->width = 0;
1279 type->length = 1;
1280
1281 for (i = 0; i < format_desc->nr_channels; ++i) {
1282 type->width += format_desc->channel[i].size;
1283 }
1284 } else {
1285 type->width = format_desc->channel[chan].size;
1286 type->length = format_desc->nr_channels;
1287 }
1288 }
1289
1290
1291 /**
1292 * Retrieves the type for a format which is usable in the blending code.
1293 *
1294 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1295 */
1296 static inline void
1297 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1298 struct lp_type* type)
1299 {
1300 unsigned i;
1301 unsigned chan;
1302
1303 if (format_expands_to_float_soa(format_desc)) {
1304 /* always use ordinary floats for blending */
1305 type->floating = true;
1306 type->fixed = false;
1307 type->sign = true;
1308 type->norm = false;
1309 type->width = 32;
1310 type->length = 4;
1311 return;
1312 }
1313
1314 for (i = 0; i < 4; i++)
1315 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1316 break;
1317 chan = i;
1318
1319 memset(type, 0, sizeof(struct lp_type));
1320 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1321 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1322 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1323 type->norm = format_desc->channel[chan].normalized;
1324 type->width = format_desc->channel[chan].size;
1325 type->length = format_desc->nr_channels;
1326
1327 for (i = 1; i < format_desc->nr_channels; ++i) {
1328 if (format_desc->channel[i].size > type->width)
1329 type->width = format_desc->channel[i].size;
1330 }
1331
1332 if (type->floating) {
1333 type->width = 32;
1334 } else {
1335 if (type->width <= 8) {
1336 type->width = 8;
1337 } else if (type->width <= 16) {
1338 type->width = 16;
1339 } else {
1340 type->width = 32;
1341 }
1342 }
1343
1344 if (is_arithmetic_format(format_desc) && type->length == 3) {
1345 type->length = 4;
1346 }
1347 }
1348
1349
1350 /**
1351 * Scale a normalized value from src_bits to dst_bits.
1352 *
1353 * The exact calculation is
1354 *
1355 * dst = iround(src * dst_mask / src_mask)
1356 *
1357 * or with integer rounding
1358 *
1359 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1360 *
1361 * where
1362 *
1363 * src_mask = (1 << src_bits) - 1
1364 * dst_mask = (1 << dst_bits) - 1
1365 *
1366 * but we try to avoid division and multiplication through shifts.
1367 */
1368 static inline LLVMValueRef
1369 scale_bits(struct gallivm_state *gallivm,
1370 int src_bits,
1371 int dst_bits,
1372 LLVMValueRef src,
1373 struct lp_type src_type)
1374 {
1375 LLVMBuilderRef builder = gallivm->builder;
1376 LLVMValueRef result = src;
1377
1378 if (dst_bits < src_bits) {
1379 int delta_bits = src_bits - dst_bits;
1380
1381 if (delta_bits <= dst_bits) {
1382 /*
1383 * Approximate the rescaling with a single shift.
1384 *
1385 * This gives the wrong rounding.
1386 */
1387
1388 result = LLVMBuildLShr(builder,
1389 src,
1390 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1391 "");
1392
1393 } else {
1394 /*
1395 * Try more accurate rescaling.
1396 */
1397
1398 /*
1399 * Drop the least significant bits to make space for the multiplication.
1400 *
1401 * XXX: A better approach would be to use a wider integer type as intermediate. But
1402 * this is enough to convert alpha from 16bits -> 2 when rendering to
1403 * PIPE_FORMAT_R10G10B10A2_UNORM.
1404 */
1405 result = LLVMBuildLShr(builder,
1406 src,
1407 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1408 "");
1409
1410
1411 result = LLVMBuildMul(builder,
1412 result,
1413 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1414 "");
1415
1416 /*
1417 * Add a rounding term before the division.
1418 *
1419 * TODO: Handle signed integers too.
1420 */
1421 if (!src_type.sign) {
1422 result = LLVMBuildAdd(builder,
1423 result,
1424 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1425 "");
1426 }
1427
1428 /*
1429 * Approximate the division by src_mask with a src_bits shift.
1430 *
1431 * Given the src has already been shifted by dst_bits, all we need
1432 * to do is to shift by the difference.
1433 */
1434
1435 result = LLVMBuildLShr(builder,
1436 result,
1437 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1438 "");
1439 }
1440
1441 } else if (dst_bits > src_bits) {
1442 /* Scale up bits */
1443 int db = dst_bits - src_bits;
1444
1445 /* Shift left by difference in bits */
1446 result = LLVMBuildShl(builder,
1447 src,
1448 lp_build_const_int_vec(gallivm, src_type, db),
1449 "");
1450
1451 if (db <= src_bits) {
1452 /* Enough bits in src to fill the remainder */
1453 LLVMValueRef lower = LLVMBuildLShr(builder,
1454 src,
1455 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1456 "");
1457
1458 result = LLVMBuildOr(builder, result, lower, "");
1459 } else if (db > src_bits) {
1460 /* Need to repeatedly copy src bits to fill remainder in dst */
1461 unsigned n;
1462
1463 for (n = src_bits; n < dst_bits; n *= 2) {
1464 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1465
1466 result = LLVMBuildOr(builder,
1467 result,
1468 LLVMBuildLShr(builder, result, shuv, ""),
1469 "");
1470 }
1471 }
1472 }
1473
1474 return result;
1475 }
1476
1477 /**
1478 * If RT is a smallfloat (needing denorms) format
1479 */
1480 static inline int
1481 have_smallfloat_format(struct lp_type dst_type,
1482 enum pipe_format format)
1483 {
1484 return ((dst_type.floating && dst_type.width != 32) ||
1485 /* due to format handling hacks this format doesn't have floating set
1486 * here (and actually has width set to 32 too) so special case this. */
1487 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1488 }
1489
1490
1491 /**
1492 * Convert from memory format to blending format
1493 *
1494 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1495 */
1496 static void
1497 convert_to_blend_type(struct gallivm_state *gallivm,
1498 unsigned block_size,
1499 const struct util_format_description *src_fmt,
1500 struct lp_type src_type,
1501 struct lp_type dst_type,
1502 LLVMValueRef* src, // and dst
1503 unsigned num_srcs)
1504 {
1505 LLVMValueRef *dst = src;
1506 LLVMBuilderRef builder = gallivm->builder;
1507 struct lp_type blend_type;
1508 struct lp_type mem_type;
1509 unsigned i, j;
1510 unsigned pixels = block_size / num_srcs;
1511 bool is_arith;
1512
1513 /*
1514 * full custom path for packed floats and srgb formats - none of the later
1515 * functions would do anything useful, and given the lp_type representation they
1516 * can't be fixed. Should really have some SoA blend path for these kind of
1517 * formats rather than hacking them in here.
1518 */
1519 if (format_expands_to_float_soa(src_fmt)) {
1520 LLVMValueRef tmpsrc[4];
1521 /*
1522 * This is pretty suboptimal for this case blending in SoA would be much
1523 * better, since conversion gets us SoA values so need to convert back.
1524 */
1525 assert(src_type.width == 32 || src_type.width == 16);
1526 assert(dst_type.floating);
1527 assert(dst_type.width == 32);
1528 assert(dst_type.length % 4 == 0);
1529 assert(num_srcs % 4 == 0);
1530
1531 if (src_type.width == 16) {
1532 /* expand 4x16bit values to 4x32bit */
1533 struct lp_type type32x4 = src_type;
1534 LLVMTypeRef ltype32x4;
1535 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1536 type32x4.width = 32;
1537 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1538 for (i = 0; i < num_fetch; i++) {
1539 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1540 }
1541 src_type.width = 32;
1542 }
1543 for (i = 0; i < 4; i++) {
1544 tmpsrc[i] = src[i];
1545 }
1546 for (i = 0; i < num_srcs / 4; i++) {
1547 LLVMValueRef tmpsoa[4];
1548 LLVMValueRef tmps = tmpsrc[i];
1549 if (dst_type.length == 8) {
1550 LLVMValueRef shuffles[8];
1551 unsigned j;
1552 /* fetch was 4 values but need 8-wide output values */
1553 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1554 /*
1555 * for 8-wide aos transpose would give us wrong order not matching
1556 * incoming converted fs values and mask. ARGH.
1557 */
1558 for (j = 0; j < 4; j++) {
1559 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1560 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1561 }
1562 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1563 LLVMConstVector(shuffles, 8), "");
1564 }
1565 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1566 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1567 }
1568 else {
1569 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1570 }
1571 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1572 }
1573 return;
1574 }
1575
1576 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1577 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1578
1579 /* Is the format arithmetic */
1580 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1581 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1582
1583 /* Pad if necessary */
1584 if (!is_arith && src_type.length < dst_type.length) {
1585 for (i = 0; i < num_srcs; ++i) {
1586 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1587 }
1588
1589 src_type.length = dst_type.length;
1590 }
1591
1592 /* Special case for half-floats */
1593 if (mem_type.width == 16 && mem_type.floating) {
1594 assert(blend_type.width == 32 && blend_type.floating);
1595 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1596 is_arith = false;
1597 }
1598
1599 if (!is_arith) {
1600 return;
1601 }
1602
1603 src_type.width = blend_type.width * blend_type.length;
1604 blend_type.length *= pixels;
1605 src_type.length *= pixels / (src_type.length / mem_type.length);
1606
1607 for (i = 0; i < num_srcs; ++i) {
1608 LLVMValueRef chans[4];
1609 LLVMValueRef res = NULL;
1610
1611 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1612
1613 for (j = 0; j < src_fmt->nr_channels; ++j) {
1614 unsigned mask = 0;
1615 unsigned sa = src_fmt->channel[j].shift;
1616 #if UTIL_ARCH_LITTLE_ENDIAN
1617 unsigned from_lsb = j;
1618 #else
1619 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1620 #endif
1621
1622 mask = (1 << src_fmt->channel[j].size) - 1;
1623
1624 /* Extract bits from source */
1625 chans[j] = LLVMBuildLShr(builder,
1626 dst[i],
1627 lp_build_const_int_vec(gallivm, src_type, sa),
1628 "");
1629
1630 chans[j] = LLVMBuildAnd(builder,
1631 chans[j],
1632 lp_build_const_int_vec(gallivm, src_type, mask),
1633 "");
1634
1635 /* Scale bits */
1636 if (src_type.norm) {
1637 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1638 blend_type.width, chans[j], src_type);
1639 }
1640
1641 /* Insert bits into correct position */
1642 chans[j] = LLVMBuildShl(builder,
1643 chans[j],
1644 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1645 "");
1646
1647 if (j == 0) {
1648 res = chans[j];
1649 } else {
1650 res = LLVMBuildOr(builder, res, chans[j], "");
1651 }
1652 }
1653
1654 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1655 }
1656 }
1657
1658
1659 /**
1660 * Convert from blending format to memory format
1661 *
1662 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1663 */
1664 static void
1665 convert_from_blend_type(struct gallivm_state *gallivm,
1666 unsigned block_size,
1667 const struct util_format_description *src_fmt,
1668 struct lp_type src_type,
1669 struct lp_type dst_type,
1670 LLVMValueRef* src, // and dst
1671 unsigned num_srcs)
1672 {
1673 LLVMValueRef* dst = src;
1674 unsigned i, j, k;
1675 struct lp_type mem_type;
1676 struct lp_type blend_type;
1677 LLVMBuilderRef builder = gallivm->builder;
1678 unsigned pixels = block_size / num_srcs;
1679 bool is_arith;
1680
1681 /*
1682 * full custom path for packed floats and srgb formats - none of the later
1683 * functions would do anything useful, and given the lp_type representation they
1684 * can't be fixed. Should really have some SoA blend path for these kind of
1685 * formats rather than hacking them in here.
1686 */
1687 if (format_expands_to_float_soa(src_fmt)) {
1688 /*
1689 * This is pretty suboptimal for this case blending in SoA would be much
1690 * better - we need to transpose the AoS values back to SoA values for
1691 * conversion/packing.
1692 */
1693 assert(src_type.floating);
1694 assert(src_type.width == 32);
1695 assert(src_type.length % 4 == 0);
1696 assert(dst_type.width == 32 || dst_type.width == 16);
1697
1698 for (i = 0; i < num_srcs / 4; i++) {
1699 LLVMValueRef tmpsoa[4], tmpdst;
1700 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1701 /* really really need SoA here */
1702
1703 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1704 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1705 }
1706 else {
1707 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1708 src_type, tmpsoa);
1709 }
1710
1711 if (src_type.length == 8) {
1712 LLVMValueRef tmpaos, shuffles[8];
1713 unsigned j;
1714 /*
1715 * for 8-wide aos transpose has given us wrong order not matching
1716 * output order. HMPF. Also need to split the output values manually.
1717 */
1718 for (j = 0; j < 4; j++) {
1719 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1720 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1721 }
1722 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1723 LLVMConstVector(shuffles, 8), "");
1724 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1725 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1726 }
1727 else {
1728 src[i] = tmpdst;
1729 }
1730 }
1731 if (dst_type.width == 16) {
1732 struct lp_type type16x8 = dst_type;
1733 struct lp_type type32x4 = dst_type;
1734 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1735 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1736 type16x8.length = 8;
1737 type32x4.width = 32;
1738 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1739 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1740 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1741 /* We could do vector truncation but it doesn't generate very good code */
1742 for (i = 0; i < num_fetch; i++) {
1743 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1744 src[i], lp_build_zero(gallivm, type32x4));
1745 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1746 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1747 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1748 }
1749 }
1750 return;
1751 }
1752
1753 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1754 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1755
1756 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1757
1758 /* Special case for half-floats */
1759 if (mem_type.width == 16 && mem_type.floating) {
1760 int length = dst_type.length;
1761 assert(blend_type.width == 32 && blend_type.floating);
1762
1763 dst_type.length = src_type.length;
1764
1765 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1766
1767 dst_type.length = length;
1768 is_arith = false;
1769 }
1770
1771 /* Remove any padding */
1772 if (!is_arith && (src_type.length % mem_type.length)) {
1773 src_type.length -= (src_type.length % mem_type.length);
1774
1775 for (i = 0; i < num_srcs; ++i) {
1776 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1777 }
1778 }
1779
1780 /* No bit arithmetic to do */
1781 if (!is_arith) {
1782 return;
1783 }
1784
1785 src_type.length = pixels;
1786 src_type.width = blend_type.length * blend_type.width;
1787 dst_type.length = pixels;
1788
1789 for (i = 0; i < num_srcs; ++i) {
1790 LLVMValueRef chans[4];
1791 LLVMValueRef res = NULL;
1792
1793 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1794
1795 for (j = 0; j < src_fmt->nr_channels; ++j) {
1796 unsigned mask = 0;
1797 unsigned sa = src_fmt->channel[j].shift;
1798 unsigned sz_a = src_fmt->channel[j].size;
1799 #if UTIL_ARCH_LITTLE_ENDIAN
1800 unsigned from_lsb = j;
1801 #else
1802 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1803 #endif
1804
1805 assert(blend_type.width > src_fmt->channel[j].size);
1806
1807 for (k = 0; k < blend_type.width; ++k) {
1808 mask |= 1 << k;
1809 }
1810
1811 /* Extract bits */
1812 chans[j] = LLVMBuildLShr(builder,
1813 dst[i],
1814 lp_build_const_int_vec(gallivm, src_type,
1815 from_lsb * blend_type.width),
1816 "");
1817
1818 chans[j] = LLVMBuildAnd(builder,
1819 chans[j],
1820 lp_build_const_int_vec(gallivm, src_type, mask),
1821 "");
1822
1823 /* Scale down bits */
1824 if (src_type.norm) {
1825 chans[j] = scale_bits(gallivm, blend_type.width,
1826 src_fmt->channel[j].size, chans[j], src_type);
1827 } else if (!src_type.floating && sz_a < blend_type.width) {
1828 LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
1829 LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
1830 chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
1831 }
1832
1833 /* Insert bits */
1834 chans[j] = LLVMBuildShl(builder,
1835 chans[j],
1836 lp_build_const_int_vec(gallivm, src_type, sa),
1837 "");
1838
1839 sa += src_fmt->channel[j].size;
1840
1841 if (j == 0) {
1842 res = chans[j];
1843 } else {
1844 res = LLVMBuildOr(builder, res, chans[j], "");
1845 }
1846 }
1847
1848 assert (dst_type.width != 24);
1849
1850 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1851 }
1852 }
1853
1854
1855 /**
1856 * Convert alpha to same blend type as src
1857 */
1858 static void
1859 convert_alpha(struct gallivm_state *gallivm,
1860 struct lp_type row_type,
1861 struct lp_type alpha_type,
1862 const unsigned block_size,
1863 const unsigned block_height,
1864 const unsigned src_count,
1865 const unsigned dst_channels,
1866 const bool pad_inline,
1867 LLVMValueRef* src_alpha)
1868 {
1869 LLVMBuilderRef builder = gallivm->builder;
1870 unsigned i, j;
1871 unsigned length = row_type.length;
1872 row_type.length = alpha_type.length;
1873
1874 /* Twiddle the alpha to match pixels */
1875 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1876
1877 /*
1878 * TODO this should use single lp_build_conv call for
1879 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1880 */
1881 for (i = 0; i < block_height; ++i) {
1882 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1883 }
1884
1885 alpha_type = row_type;
1886 row_type.length = length;
1887
1888 /* If only one channel we can only need the single alpha value per pixel */
1889 if (src_count == 1 && dst_channels == 1) {
1890
1891 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1892 } else {
1893 /* If there are more srcs than rows then we need to split alpha up */
1894 if (src_count > block_height) {
1895 for (i = src_count; i > 0; --i) {
1896 unsigned pixels = block_size / src_count;
1897 unsigned idx = i - 1;
1898
1899 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1900 (idx * pixels) % 4, pixels);
1901 }
1902 }
1903
1904 /* If there is a src for each pixel broadcast the alpha across whole row */
1905 if (src_count == block_size) {
1906 for (i = 0; i < src_count; ++i) {
1907 src_alpha[i] = lp_build_broadcast(gallivm,
1908 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1909 }
1910 } else {
1911 unsigned pixels = block_size / src_count;
1912 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1913 unsigned alpha_span = 1;
1914 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1915
1916 /* Check if we need 2 src_alphas for our shuffles */
1917 if (pixels > alpha_type.length) {
1918 alpha_span = 2;
1919 }
1920
1921 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1922 for (j = 0; j < row_type.length; ++j) {
1923 if (j < pixels * channels) {
1924 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1925 } else {
1926 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1927 }
1928 }
1929
1930 for (i = 0; i < src_count; ++i) {
1931 unsigned idx1 = i, idx2 = i;
1932
1933 if (alpha_span > 1){
1934 idx1 *= alpha_span;
1935 idx2 = idx1 + 1;
1936 }
1937
1938 src_alpha[i] = LLVMBuildShuffleVector(builder,
1939 src_alpha[idx1],
1940 src_alpha[idx2],
1941 LLVMConstVector(shuffles, row_type.length),
1942 "");
1943 }
1944 }
1945 }
1946 }
1947
1948
1949 /**
1950 * Generates the blend function for unswizzled colour buffers
1951 * Also generates the read & write from colour buffer
1952 */
1953 static void
1954 generate_unswizzled_blend(struct gallivm_state *gallivm,
1955 unsigned rt,
1956 struct lp_fragment_shader_variant *variant,
1957 enum pipe_format out_format,
1958 unsigned int num_fs,
1959 struct lp_type fs_type,
1960 LLVMValueRef* fs_mask,
1961 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1962 LLVMValueRef context_ptr,
1963 LLVMValueRef color_ptr,
1964 LLVMValueRef stride,
1965 unsigned partial_mask,
1966 boolean do_branch)
1967 {
1968 const unsigned alpha_channel = 3;
1969 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1970 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1971 const unsigned block_size = block_width * block_height;
1972 const unsigned lp_integer_vector_width = 128;
1973
1974 LLVMBuilderRef builder = gallivm->builder;
1975 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1976 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1977 LLVMValueRef src_alpha[4 * 4];
1978 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1979 LLVMValueRef src_mask[4 * 4];
1980 LLVMValueRef src[4 * 4];
1981 LLVMValueRef src1[4 * 4];
1982 LLVMValueRef dst[4 * 4];
1983 LLVMValueRef blend_color;
1984 LLVMValueRef blend_alpha;
1985 LLVMValueRef i32_zero;
1986 LLVMValueRef check_mask;
1987 LLVMValueRef undef_src_val;
1988
1989 struct lp_build_mask_context mask_ctx;
1990 struct lp_type mask_type;
1991 struct lp_type blend_type;
1992 struct lp_type row_type;
1993 struct lp_type dst_type;
1994 struct lp_type ls_type;
1995
1996 unsigned char swizzle[TGSI_NUM_CHANNELS];
1997 unsigned vector_width;
1998 unsigned src_channels = TGSI_NUM_CHANNELS;
1999 unsigned dst_channels;
2000 unsigned dst_count;
2001 unsigned src_count;
2002 unsigned i, j;
2003
2004 const struct util_format_description* out_format_desc = util_format_description(out_format);
2005
2006 unsigned dst_alignment;
2007
2008 bool pad_inline = is_arithmetic_format(out_format_desc);
2009 bool has_alpha = false;
2010 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2011 util_blend_state_is_dual(&variant->key.blend, 0);
2012
2013 const boolean is_1d = variant->key.resource_1d;
2014 boolean twiddle_after_convert = FALSE;
2015 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2016 LLVMValueRef fpstate = 0;
2017
2018 /* Get type from output format */
2019 lp_blend_type_from_format_desc(out_format_desc, &row_type);
2020 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2021
2022 /*
2023 * Technically this code should go into lp_build_smallfloat_to_float
2024 * and lp_build_float_to_smallfloat but due to the
2025 * http://llvm.org/bugs/show_bug.cgi?id=6393
2026 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2027 * So the ordering is important here and there shouldn't be any
2028 * llvm ir instrunctions in this function before
2029 * this, otherwise half-float format conversions won't work
2030 * (again due to llvm bug #6393).
2031 */
2032 if (have_smallfloat_format(dst_type, out_format)) {
2033 /* We need to make sure that denorms are ok for half float
2034 conversions */
2035 fpstate = lp_build_fpstate_get(gallivm);
2036 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2037 }
2038
2039 mask_type = lp_int32_vec4_type();
2040 mask_type.length = fs_type.length;
2041
2042 for (i = num_fs; i < num_fullblock_fs; i++) {
2043 fs_mask[i] = lp_build_zero(gallivm, mask_type);
2044 }
2045
2046 /* Do not bother executing code when mask is empty.. */
2047 if (do_branch) {
2048 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2049
2050 for (i = 0; i < num_fullblock_fs; ++i) {
2051 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2052 }
2053
2054 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2055 lp_build_mask_check(&mask_ctx);
2056 }
2057
2058 partial_mask |= !variant->opaque;
2059 i32_zero = lp_build_const_int32(gallivm, 0);
2060
2061 undef_src_val = lp_build_undef(gallivm, fs_type);
2062
2063 row_type.length = fs_type.length;
2064 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2065
2066 /* Compute correct swizzle and count channels */
2067 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2068 dst_channels = 0;
2069
2070 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2071 /* Ensure channel is used */
2072 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2073 continue;
2074 }
2075
2076 /* Ensure not already written to (happens in case with GL_ALPHA) */
2077 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2078 continue;
2079 }
2080
2081 /* Ensure we havn't already found all channels */
2082 if (dst_channels >= out_format_desc->nr_channels) {
2083 continue;
2084 }
2085
2086 swizzle[out_format_desc->swizzle[i]] = i;
2087 ++dst_channels;
2088
2089 if (i == alpha_channel) {
2090 has_alpha = true;
2091 }
2092 }
2093
2094 if (format_expands_to_float_soa(out_format_desc)) {
2095 /*
2096 * the code above can't work for layout_other
2097 * for srgb it would sort of work but we short-circuit swizzles, etc.
2098 * as that is done as part of unpack / pack.
2099 */
2100 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2101 has_alpha = true;
2102 swizzle[0] = 0;
2103 swizzle[1] = 1;
2104 swizzle[2] = 2;
2105 swizzle[3] = 3;
2106 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2107 }
2108
2109 /* If 3 channels then pad to include alpha for 4 element transpose */
2110 if (dst_channels == 3) {
2111 assert (!has_alpha);
2112 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2113 if (swizzle[i] > TGSI_NUM_CHANNELS)
2114 swizzle[i] = 3;
2115 }
2116 if (out_format_desc->nr_channels == 4) {
2117 dst_channels = 4;
2118 /*
2119 * We use alpha from the color conversion, not separate one.
2120 * We had to include it for transpose, hence it will get converted
2121 * too (albeit when doing transpose after conversion, that would
2122 * no longer be the case necessarily).
2123 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2124 * otherwise we really have padding, not alpha, included.)
2125 */
2126 has_alpha = true;
2127 }
2128 }
2129
2130 /*
2131 * Load shader output
2132 */
2133 for (i = 0; i < num_fullblock_fs; ++i) {
2134 /* Always load alpha for use in blending */
2135 LLVMValueRef alpha;
2136 if (i < num_fs) {
2137 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2138 }
2139 else {
2140 alpha = undef_src_val;
2141 }
2142
2143 /* Load each channel */
2144 for (j = 0; j < dst_channels; ++j) {
2145 assert(swizzle[j] < 4);
2146 if (i < num_fs) {
2147 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2148 }
2149 else {
2150 fs_src[i][j] = undef_src_val;
2151 }
2152 }
2153
2154 /* If 3 channels then pad to include alpha for 4 element transpose */
2155 /*
2156 * XXX If we include that here maybe could actually use it instead of
2157 * separate alpha for blending?
2158 * (Difficult though we actually convert pad channels, not alpha.)
2159 */
2160 if (dst_channels == 3 && !has_alpha) {
2161 fs_src[i][3] = alpha;
2162 }
2163
2164 /* We split the row_mask and row_alpha as we want 128bit interleave */
2165 if (fs_type.length == 8) {
2166 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
2167 0, src_channels);
2168 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
2169 src_channels, src_channels);
2170
2171 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2172 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2173 src_channels, src_channels);
2174 } else {
2175 src_mask[i] = fs_mask[i];
2176 src_alpha[i] = alpha;
2177 }
2178 }
2179 if (dual_source_blend) {
2180 /* same as above except different src/dst, skip masks and comments... */
2181 for (i = 0; i < num_fullblock_fs; ++i) {
2182 LLVMValueRef alpha;
2183 if (i < num_fs) {
2184 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2185 }
2186 else {
2187 alpha = undef_src_val;
2188 }
2189
2190 for (j = 0; j < dst_channels; ++j) {
2191 assert(swizzle[j] < 4);
2192 if (i < num_fs) {
2193 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2194 }
2195 else {
2196 fs_src1[i][j] = undef_src_val;
2197 }
2198 }
2199 if (dst_channels == 3 && !has_alpha) {
2200 fs_src1[i][3] = alpha;
2201 }
2202 if (fs_type.length == 8) {
2203 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2204 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2205 src_channels, src_channels);
2206 } else {
2207 src1_alpha[i] = alpha;
2208 }
2209 }
2210 }
2211
2212 if (util_format_is_pure_integer(out_format)) {
2213 /*
2214 * In this case fs_type was really ints or uints disguised as floats,
2215 * fix that up now.
2216 */
2217 fs_type.floating = 0;
2218 fs_type.sign = dst_type.sign;
2219 for (i = 0; i < num_fullblock_fs; ++i) {
2220 for (j = 0; j < dst_channels; ++j) {
2221 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2222 lp_build_vec_type(gallivm, fs_type), "");
2223 }
2224 if (dst_channels == 3 && !has_alpha) {
2225 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2226 lp_build_vec_type(gallivm, fs_type), "");
2227 }
2228 }
2229 }
2230
2231 /*
2232 * We actually should generally do conversion first (for non-1d cases)
2233 * when the blend format is 8 or 16 bits. The reason is obvious,
2234 * there's 2 or 4 times less vectors to deal with for the interleave...
2235 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2236 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2237 * unpack only with 128bit vectors).
2238 * Note: for 16bit sizes really need matching pack conversion code
2239 */
2240 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2241 twiddle_after_convert = TRUE;
2242 }
2243
2244 /*
2245 * Pixel twiddle from fragment shader order to memory order
2246 */
2247 if (!twiddle_after_convert) {
2248 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2249 dst_channels, fs_src, src, pad_inline);
2250 if (dual_source_blend) {
2251 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2252 fs_src1, src1, pad_inline);
2253 }
2254 } else {
2255 src_count = num_fullblock_fs * dst_channels;
2256 /*
2257 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2258 * (AVX) turn out the same later when untwiddling/transpose (albeit
2259 * for true AVX2 path untwiddle needs to be different).
2260 * For now just order by colors first (so we can use unpack later).
2261 */
2262 for (j = 0; j < num_fullblock_fs; j++) {
2263 for (i = 0; i < dst_channels; i++) {
2264 src[i*num_fullblock_fs + j] = fs_src[j][i];
2265 if (dual_source_blend) {
2266 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2267 }
2268 }
2269 }
2270 }
2271
2272 src_channels = dst_channels < 3 ? dst_channels : 4;
2273 if (src_count != num_fullblock_fs * src_channels) {
2274 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2275 row_type.length /= ds;
2276 fs_type.length = row_type.length;
2277 }
2278
2279 blend_type = row_type;
2280 mask_type.length = 4;
2281
2282 /* Convert src to row_type */
2283 if (dual_source_blend) {
2284 struct lp_type old_row_type = row_type;
2285 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2286 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2287 }
2288 else {
2289 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2290 }
2291
2292 /* If the rows are not an SSE vector, combine them to become SSE size! */
2293 if ((row_type.width * row_type.length) % 128) {
2294 unsigned bits = row_type.width * row_type.length;
2295 unsigned combined;
2296
2297 assert(src_count >= (vector_width / bits));
2298
2299 dst_count = src_count / (vector_width / bits);
2300
2301 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2302 if (dual_source_blend) {
2303 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2304 }
2305
2306 row_type.length *= combined;
2307 src_count /= combined;
2308
2309 bits = row_type.width * row_type.length;
2310 assert(bits == 128 || bits == 256);
2311 }
2312
2313 if (twiddle_after_convert) {
2314 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2315 if (dual_source_blend) {
2316 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2317 }
2318 }
2319
2320 /*
2321 * Blend Colour conversion
2322 */
2323 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2324 blend_color = LLVMBuildPointerCast(builder, blend_color,
2325 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2326 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2327 &i32_zero, 1, ""), "");
2328
2329 /* Convert */
2330 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2331
2332 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2333 /*
2334 * since blending is done with floats, there was no conversion.
2335 * However, the rules according to fixed point renderbuffers still
2336 * apply, that is we must clamp inputs to 0.0/1.0.
2337 * (This would apply to separate alpha conversion too but we currently
2338 * force has_alpha to be true.)
2339 * TODO: should skip this with "fake" blend, since post-blend conversion
2340 * will clamp anyway.
2341 * TODO: could also skip this if fragment color clamping is enabled. We
2342 * don't support it natively so it gets baked into the shader however, so
2343 * can't really tell here.
2344 */
2345 struct lp_build_context f32_bld;
2346 assert(row_type.floating);
2347 lp_build_context_init(&f32_bld, gallivm, row_type);
2348 for (i = 0; i < src_count; i++) {
2349 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2350 }
2351 if (dual_source_blend) {
2352 for (i = 0; i < src_count; i++) {
2353 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2354 }
2355 }
2356 /* probably can't be different than row_type but better safe than sorry... */
2357 lp_build_context_init(&f32_bld, gallivm, blend_type);
2358 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2359 }
2360
2361 /* Extract alpha */
2362 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2363
2364 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2365 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2366 if (pad_inline) {
2367 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2368 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2369 } else {
2370 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2371 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2372 }
2373
2374 /*
2375 * Mask conversion
2376 */
2377 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2378
2379 if (src_count < block_height) {
2380 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2381 } else if (src_count > block_height) {
2382 for (i = src_count; i > 0; --i) {
2383 unsigned pixels = block_size / src_count;
2384 unsigned idx = i - 1;
2385
2386 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2387 (idx * pixels) % 4, pixels);
2388 }
2389 }
2390
2391 assert(mask_type.width == 32);
2392
2393 for (i = 0; i < src_count; ++i) {
2394 unsigned pixels = block_size / src_count;
2395 unsigned pixel_width = row_type.width * dst_channels;
2396
2397 if (pixel_width == 24) {
2398 mask_type.width = 8;
2399 mask_type.length = vector_width / mask_type.width;
2400 } else {
2401 mask_type.length = pixels;
2402 mask_type.width = row_type.width * dst_channels;
2403
2404 /*
2405 * If mask_type width is smaller than 32bit, this doesn't quite
2406 * generate the most efficient code (could use some pack).
2407 */
2408 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2409 lp_build_int_vec_type(gallivm, mask_type), "");
2410
2411 mask_type.length *= dst_channels;
2412 mask_type.width /= dst_channels;
2413 }
2414
2415 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2416 lp_build_int_vec_type(gallivm, mask_type), "");
2417 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2418 }
2419
2420 /*
2421 * Alpha conversion
2422 */
2423 if (!has_alpha) {
2424 struct lp_type alpha_type = fs_type;
2425 alpha_type.length = 4;
2426 convert_alpha(gallivm, row_type, alpha_type,
2427 block_size, block_height,
2428 src_count, dst_channels,
2429 pad_inline, src_alpha);
2430 if (dual_source_blend) {
2431 convert_alpha(gallivm, row_type, alpha_type,
2432 block_size, block_height,
2433 src_count, dst_channels,
2434 pad_inline, src1_alpha);
2435 }
2436 }
2437
2438
2439 /*
2440 * Load dst from memory
2441 */
2442 if (src_count < block_height) {
2443 dst_count = block_height;
2444 } else {
2445 dst_count = src_count;
2446 }
2447
2448 dst_type.length *= block_size / dst_count;
2449
2450 if (format_expands_to_float_soa(out_format_desc)) {
2451 /*
2452 * we need multiple values at once for the conversion, so can as well
2453 * load them vectorized here too instead of concatenating later.
2454 * (Still need concatenation later for 8-wide vectors).
2455 */
2456 dst_count = block_height;
2457 dst_type.length = block_width;
2458 }
2459
2460 /*
2461 * Compute the alignment of the destination pointer in bytes
2462 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2463 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2464 * 1d tex but can't distinguish here) so need to stick with per-pixel
2465 * alignment in this case.
2466 */
2467 if (is_1d) {
2468 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2469 }
2470 else {
2471 dst_alignment = dst_type.length * dst_type.width / 8;
2472 }
2473 /* Force power-of-two alignment by extracting only the least-significant-bit */
2474 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2475 /*
2476 * Resource base and stride pointers are aligned to 16 bytes, so that's
2477 * the maximum alignment we can guarantee
2478 */
2479 dst_alignment = MIN2(16, dst_alignment);
2480
2481 ls_type = dst_type;
2482
2483 if (dst_count > src_count) {
2484 if ((dst_type.width == 8 || dst_type.width == 16) &&
2485 util_is_power_of_two_or_zero(dst_type.length) &&
2486 dst_type.length * dst_type.width < 128) {
2487 /*
2488 * Never try to load values as 4xi8 which we will then
2489 * concatenate to larger vectors. This gives llvm a real
2490 * headache (the problem is the type legalizer (?) will
2491 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2492 * then the shuffles to concatenate are more or less impossible
2493 * - llvm is easily capable of generating a sequence of 32
2494 * pextrb/pinsrb instructions for that. Albeit it appears to
2495 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2496 * width to avoid the trouble (16bit seems not as bad, llvm
2497 * probably recognizes the load+shuffle as only one shuffle
2498 * is necessary, but we can do just the same anyway).
2499 */
2500 ls_type.length = dst_type.length * dst_type.width / 32;
2501 ls_type.width = 32;
2502 }
2503 }
2504
2505 if (is_1d) {
2506 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2507 dst, ls_type, dst_count / 4, dst_alignment);
2508 for (i = dst_count / 4; i < dst_count; i++) {
2509 dst[i] = lp_build_undef(gallivm, ls_type);
2510 }
2511
2512 }
2513 else {
2514 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2515 dst, ls_type, dst_count, dst_alignment);
2516 }
2517
2518
2519 /*
2520 * Convert from dst/output format to src/blending format.
2521 *
2522 * This is necessary as we can only read 1 row from memory at a time,
2523 * so the minimum dst_count will ever be at this point is 4.
2524 *
2525 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2526 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2527 * on all 16 pixels in that single vector at once.
2528 */
2529 if (dst_count > src_count) {
2530 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2531 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2532 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2533 for (i = 0; i < dst_count; i++) {
2534 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2535 }
2536 }
2537
2538 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2539
2540 if (ls_type.length != dst_type.length) {
2541 struct lp_type tmp_type = dst_type;
2542 tmp_type.length = dst_type.length * 4 / src_count;
2543 for (i = 0; i < src_count; i++) {
2544 dst[i] = LLVMBuildBitCast(builder, dst[i],
2545 lp_build_vec_type(gallivm, tmp_type), "");
2546 }
2547 }
2548 }
2549
2550 /*
2551 * Blending
2552 */
2553 /* XXX this is broken for RGB8 formats -
2554 * they get expanded from 12 to 16 elements (to include alpha)
2555 * by convert_to_blend_type then reduced to 15 instead of 12
2556 * by convert_from_blend_type (a simple fix though breaks A8...).
2557 * R16G16B16 also crashes differently however something going wrong
2558 * inside llvm handling npot vector sizes seemingly.
2559 * It seems some cleanup could be done here (like skipping conversion/blend
2560 * when not needed).
2561 */
2562 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2563 row_type, dst, src_count);
2564
2565 /*
2566 * FIXME: Really should get logic ops / masks out of generic blend / row
2567 * format. Logic ops will definitely not work on the blend float format
2568 * used for SRGB here and I think OpenGL expects this to work as expected
2569 * (that is incoming values converted to srgb then logic op applied).
2570 */
2571 for (i = 0; i < src_count; ++i) {
2572 dst[i] = lp_build_blend_aos(gallivm,
2573 &variant->key.blend,
2574 out_format,
2575 row_type,
2576 rt,
2577 src[i],
2578 has_alpha ? NULL : src_alpha[i],
2579 src1[i],
2580 has_alpha ? NULL : src1_alpha[i],
2581 dst[i],
2582 partial_mask ? src_mask[i] : NULL,
2583 blend_color,
2584 has_alpha ? NULL : blend_alpha,
2585 swizzle,
2586 pad_inline ? 4 : dst_channels);
2587 }
2588
2589 convert_from_blend_type(gallivm, block_size, out_format_desc,
2590 row_type, dst_type, dst, src_count);
2591
2592 /* Split the blend rows back to memory rows */
2593 if (dst_count > src_count) {
2594 row_type.length = dst_type.length * (dst_count / src_count);
2595
2596 if (src_count == 1) {
2597 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2598 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2599
2600 row_type.length /= 2;
2601 src_count *= 2;
2602 }
2603
2604 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2605 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2606 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2607 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2608
2609 row_type.length /= 2;
2610 src_count *= 2;
2611 }
2612
2613 /*
2614 * Store blend result to memory
2615 */
2616 if (is_1d) {
2617 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2618 dst, dst_type, dst_count / 4, dst_alignment);
2619 }
2620 else {
2621 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2622 dst, dst_type, dst_count, dst_alignment);
2623 }
2624
2625 if (have_smallfloat_format(dst_type, out_format)) {
2626 lp_build_fpstate_set(gallivm, fpstate);
2627 }
2628
2629 if (do_branch) {
2630 lp_build_mask_end(&mask_ctx);
2631 }
2632 }
2633
2634
2635 /**
2636 * Generate the runtime callable function for the whole fragment pipeline.
2637 * Note that the function which we generate operates on a block of 16
2638 * pixels at at time. The block contains 2x2 quads. Each quad contains
2639 * 2x2 pixels.
2640 */
2641 static void
2642 generate_fragment(struct llvmpipe_context *lp,
2643 struct lp_fragment_shader *shader,
2644 struct lp_fragment_shader_variant *variant,
2645 unsigned partial_mask)
2646 {
2647 struct gallivm_state *gallivm = variant->gallivm;
2648 struct lp_fragment_shader_variant_key *key = &variant->key;
2649 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2650 char func_name[64];
2651 struct lp_type fs_type;
2652 struct lp_type blend_type;
2653 LLVMTypeRef fs_elem_type;
2654 LLVMTypeRef blend_vec_type;
2655 LLVMTypeRef arg_types[15];
2656 LLVMTypeRef func_type;
2657 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2658 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2659 LLVMValueRef context_ptr;
2660 LLVMValueRef x;
2661 LLVMValueRef y;
2662 LLVMValueRef a0_ptr;
2663 LLVMValueRef dadx_ptr;
2664 LLVMValueRef dady_ptr;
2665 LLVMValueRef color_ptr_ptr;
2666 LLVMValueRef stride_ptr;
2667 LLVMValueRef color_sample_stride_ptr;
2668 LLVMValueRef depth_ptr;
2669 LLVMValueRef depth_stride;
2670 LLVMValueRef depth_sample_stride;
2671 LLVMValueRef mask_input;
2672 LLVMValueRef thread_data_ptr;
2673 LLVMBasicBlockRef block;
2674 LLVMBuilderRef builder;
2675 struct lp_build_sampler_soa *sampler;
2676 struct lp_build_image_soa *image;
2677 struct lp_build_interp_soa_context interp;
2678 LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
2679 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2680 LLVMValueRef function;
2681 LLVMValueRef facing;
2682 unsigned num_fs;
2683 unsigned i;
2684 unsigned chan;
2685 unsigned cbuf;
2686 boolean cbuf0_write_all;
2687 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2688 util_blend_state_is_dual(&key->blend, 0);
2689
2690 assert(lp_native_vector_width / 32 >= 4);
2691
2692 /* Adjust color input interpolation according to flatshade state:
2693 */
2694 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2695 for (i = 0; i < shader->info.base.num_inputs; i++) {
2696 if (inputs[i].interp == LP_INTERP_COLOR) {
2697 if (key->flatshade)
2698 inputs[i].interp = LP_INTERP_CONSTANT;
2699 else
2700 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2701 }
2702 }
2703
2704 /* check if writes to cbuf[0] are to be copied to all cbufs */
2705 cbuf0_write_all =
2706 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2707
2708 /* TODO: actually pick these based on the fs and color buffer
2709 * characteristics. */
2710
2711 memset(&fs_type, 0, sizeof fs_type);
2712 fs_type.floating = TRUE; /* floating point values */
2713 fs_type.sign = TRUE; /* values are signed */
2714 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2715 fs_type.width = 32; /* 32-bit float */
2716 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2717
2718 memset(&blend_type, 0, sizeof blend_type);
2719 blend_type.floating = FALSE; /* values are integers */
2720 blend_type.sign = FALSE; /* values are unsigned */
2721 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2722 blend_type.width = 8; /* 8-bit ubyte values */
2723 blend_type.length = 16; /* 16 elements per vector */
2724
2725 /*
2726 * Generate the function prototype. Any change here must be reflected in
2727 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2728 */
2729
2730 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2731
2732 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2733
2734 snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2735 shader->no, variant->no, partial_mask ? "partial" : "whole");
2736
2737 arg_types[0] = variant->jit_context_ptr_type; /* context */
2738 arg_types[1] = int32_type; /* x */
2739 arg_types[2] = int32_type; /* y */
2740 arg_types[3] = int32_type; /* facing */
2741 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2742 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2743 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2744 arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
2745 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2746 arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
2747 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2748 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2749 arg_types[12] = int32_type; /* depth_stride */
2750 arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
2751 arg_types[14] = int32_type; /* depth sample stride */
2752
2753 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2754 arg_types, ARRAY_SIZE(arg_types), 0);
2755
2756 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2757 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2758
2759 variant->function[partial_mask] = function;
2760
2761 /* XXX: need to propagate noalias down into color param now we are
2762 * passing a pointer-to-pointer?
2763 */
2764 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2765 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2766 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2767
2768 context_ptr = LLVMGetParam(function, 0);
2769 x = LLVMGetParam(function, 1);
2770 y = LLVMGetParam(function, 2);
2771 facing = LLVMGetParam(function, 3);
2772 a0_ptr = LLVMGetParam(function, 4);
2773 dadx_ptr = LLVMGetParam(function, 5);
2774 dady_ptr = LLVMGetParam(function, 6);
2775 color_ptr_ptr = LLVMGetParam(function, 7);
2776 depth_ptr = LLVMGetParam(function, 8);
2777 mask_input = LLVMGetParam(function, 9);
2778 thread_data_ptr = LLVMGetParam(function, 10);
2779 stride_ptr = LLVMGetParam(function, 11);
2780 depth_stride = LLVMGetParam(function, 12);
2781 color_sample_stride_ptr = LLVMGetParam(function, 13);
2782 depth_sample_stride = LLVMGetParam(function, 14);
2783
2784 lp_build_name(context_ptr, "context");
2785 lp_build_name(x, "x");
2786 lp_build_name(y, "y");
2787 lp_build_name(a0_ptr, "a0");
2788 lp_build_name(dadx_ptr, "dadx");
2789 lp_build_name(dady_ptr, "dady");
2790 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2791 lp_build_name(depth_ptr, "depth");
2792 lp_build_name(mask_input, "mask_input");
2793 lp_build_name(thread_data_ptr, "thread_data");
2794 lp_build_name(stride_ptr, "stride_ptr");
2795 lp_build_name(depth_stride, "depth_stride");
2796 lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
2797 lp_build_name(depth_sample_stride, "depth_sample_stride");
2798
2799 /*
2800 * Function body
2801 */
2802
2803 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2804 builder = gallivm->builder;
2805 assert(builder);
2806 LLVMPositionBuilderAtEnd(builder, block);
2807
2808 /*
2809 * Must not count ps invocations if there's a null shader.
2810 * (It would be ok to count with null shader if there's d/s tests,
2811 * but only if there's d/s buffers too, which is different
2812 * to implicit rasterization disable which must not depend
2813 * on the d/s buffers.)
2814 * Could use popcount on mask, but pixel accuracy is not required.
2815 * Could disable if there's no stats query, but maybe not worth it.
2816 */
2817 if (shader->info.base.num_instructions > 1) {
2818 LLVMValueRef invocs, val;
2819 invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
2820 val = LLVMBuildLoad(builder, invocs, "");
2821 val = LLVMBuildAdd(builder, val,
2822 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
2823 "invoc_count");
2824 LLVMBuildStore(builder, val, invocs);
2825 }
2826
2827 /* code generated texture sampling */
2828 sampler = lp_llvm_sampler_soa_create(key->samplers);
2829 image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
2830
2831 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2832 /* for 1d resources only run "upper half" of stamp */
2833 if (key->resource_1d)
2834 num_fs /= 2;
2835
2836 {
2837 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2838 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2839 LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
2840 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2841 num_loop_samp, "mask_store");
2842
2843 LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
2844 LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, flt_type, "");
2845 LLVMValueRef sample_pos_array;
2846
2847 if (key->multisample && key->coverage_samples == 4) {
2848 LLVMValueRef sample_pos_arr[8];
2849 for (unsigned i = 0; i < 4; i++) {
2850 sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
2851 sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
2852 }
2853 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
2854 } else {
2855 LLVMValueRef sample_pos_arr[2];
2856 sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
2857 sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
2858 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
2859 }
2860 LLVMSetInitializer(glob_sample_pos, sample_pos_array);
2861
2862 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2863 boolean pixel_center_integer =
2864 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2865
2866 /*
2867 * The shader input interpolation info is not explicitely baked in the
2868 * shader key, but everything it derives from (TGSI, and flatshade) is
2869 * already included in the shader key.
2870 */
2871 lp_build_interp_soa_init(&interp,
2872 gallivm,
2873 shader->info.base.num_inputs,
2874 inputs,
2875 pixel_center_integer,
2876 key->coverage_samples, glob_sample_pos,
2877 num_loop,
2878 key->depth_clamp,
2879 builder, fs_type,
2880 a0_ptr, dadx_ptr, dady_ptr,
2881 x, y);
2882
2883 for (i = 0; i < num_fs; i++) {
2884 if (key->multisample) {
2885 LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
2886
2887 /*
2888 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2889 * store to the per sample mask storage. Or all of them together to generate
2890 * the fragment shader mask. (sample shading TODO).
2891 * Take the incoming state coverage mask into account.
2892 */
2893 for (unsigned s = 0; s < key->coverage_samples; s++) {
2894 LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
2895 LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
2896 &sindexi, 1, "sample_mask_ptr");
2897 LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
2898 i*fs_type.length/4, s, mask_input);
2899
2900 LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
2901 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
2902 smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
2903 smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
2904
2905 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
2906 LLVMBuildStore(builder, s_mask, sample_mask_ptr);
2907 }
2908 } else {
2909 LLVMValueRef mask;
2910 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2911 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2912 &indexi, 1, "mask_ptr");
2913
2914 if (partial_mask) {
2915 mask = generate_quad_mask(gallivm, fs_type,
2916 i*fs_type.length/4, 0, mask_input);
2917 }
2918 else {
2919 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2920 }
2921 LLVMBuildStore(builder, mask, mask_ptr);
2922 }
2923 }
2924
2925 generate_fs_loop(gallivm,
2926 shader, key,
2927 builder,
2928 fs_type,
2929 context_ptr,
2930 glob_sample_pos,
2931 num_loop,
2932 &interp,
2933 sampler,
2934 image,
2935 mask_store, /* output */
2936 color_store,
2937 depth_ptr,
2938 depth_stride,
2939 depth_sample_stride,
2940 facing,
2941 thread_data_ptr);
2942
2943 for (i = 0; i < num_fs; i++) {
2944 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2945 LLVMValueRef ptr;
2946 for (unsigned s = 0; s < key->coverage_samples; s++) {
2947 int idx = (i + (s * num_fs));
2948 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
2949 ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
2950
2951 fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
2952 }
2953
2954 /* This is fucked up need to reorganize things */
2955 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2956 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2957 ptr = LLVMBuildGEP(builder,
2958 color_store[cbuf * !cbuf0_write_all][chan],
2959 &indexi, 1, "");
2960 fs_out_color[cbuf][chan][i] = ptr;
2961 }
2962 }
2963 if (dual_source_blend) {
2964 /* only support one dual source blend target hence always use output 1 */
2965 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2966 ptr = LLVMBuildGEP(builder,
2967 color_store[1][chan],
2968 &indexi, 1, "");
2969 fs_out_color[1][chan][i] = ptr;
2970 }
2971 }
2972 }
2973 }
2974
2975 sampler->destroy(sampler);
2976 image->destroy(image);
2977 /* Loop over color outputs / color buffers to do blending.
2978 */
2979 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2980 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2981 LLVMValueRef color_ptr;
2982 LLVMValueRef stride;
2983 LLVMValueRef sample_stride = NULL;
2984 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2985
2986 boolean do_branch = ((key->depth.enabled
2987 || key->stencil[0].enabled
2988 || key->alpha.enabled)
2989 && !shader->info.base.uses_kill);
2990
2991 color_ptr = LLVMBuildLoad(builder,
2992 LLVMBuildGEP(builder, color_ptr_ptr,
2993 &index, 1, ""),
2994 "");
2995
2996 stride = LLVMBuildLoad(builder,
2997 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2998 "");
2999
3000 if (key->multisample)
3001 sample_stride = LLVMBuildLoad(builder,
3002 LLVMBuildGEP(builder, color_sample_stride_ptr,
3003 &index, 1, ""), "");
3004
3005 for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3006 unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3007 LLVMValueRef out_ptr = color_ptr;;
3008
3009 if (key->multisample) {
3010 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3011 out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3012 }
3013 out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3014
3015 lp_build_name(out_ptr, "color_ptr%d", cbuf);
3016
3017 generate_unswizzled_blend(gallivm, cbuf, variant,
3018 key->cbuf_format[cbuf],
3019 num_fs, fs_type, &fs_mask[mask_idx], fs_out_color,
3020 context_ptr, out_ptr, stride,
3021 partial_mask, do_branch);
3022 }
3023 }
3024 }
3025
3026 LLVMBuildRetVoid(builder);
3027
3028 gallivm_verify_function(gallivm, function);
3029 }
3030
3031
3032 static void
3033 dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3034 {
3035 unsigned i;
3036
3037 debug_printf("fs variant %p:\n", (void *) key);
3038
3039 if (key->flatshade) {
3040 debug_printf("flatshade = 1\n");
3041 }
3042 if (key->multisample) {
3043 debug_printf("multisample = 1\n");
3044 debug_printf("coverage samples = %d\n", key->coverage_samples);
3045 }
3046 for (i = 0; i < key->nr_cbufs; ++i) {
3047 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3048 debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3049 }
3050 if (key->depth.enabled || key->stencil[0].enabled) {
3051 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3052 debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3053 }
3054 if (key->depth.enabled) {
3055 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3056 debug_printf("depth.writemask = %u\n", key->depth.writemask);
3057 }
3058
3059 for (i = 0; i < 2; ++i) {
3060 if (key->stencil[i].enabled) {
3061 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3062 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3063 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3064 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3065 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3066 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3067 }
3068 }
3069
3070 if (key->alpha.enabled) {
3071 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3072 }
3073
3074 if (key->occlusion_count) {
3075 debug_printf("occlusion_count = 1\n");
3076 }
3077
3078 if (key->blend.logicop_enable) {
3079 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3080 }
3081 else if (key->blend.rt[0].blend_enable) {
3082 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
3083 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3084 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3085 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
3086 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3087 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3088 }
3089 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3090 if (key->blend.alpha_to_coverage) {
3091 debug_printf("blend.alpha_to_coverage is enabled\n");
3092 }
3093 for (i = 0; i < key->nr_samplers; ++i) {
3094 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
3095 debug_printf("sampler[%u] = \n", i);
3096 debug_printf(" .wrap = %s %s %s\n",
3097 util_str_tex_wrap(sampler->wrap_s, TRUE),
3098 util_str_tex_wrap(sampler->wrap_t, TRUE),
3099 util_str_tex_wrap(sampler->wrap_r, TRUE));
3100 debug_printf(" .min_img_filter = %s\n",
3101 util_str_tex_filter(sampler->min_img_filter, TRUE));
3102 debug_printf(" .min_mip_filter = %s\n",
3103 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3104 debug_printf(" .mag_img_filter = %s\n",
3105 util_str_tex_filter(sampler->mag_img_filter, TRUE));
3106 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3107 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3108 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
3109 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3110 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3111 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
3112 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
3113 }
3114 for (i = 0; i < key->nr_sampler_views; ++i) {
3115 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
3116 debug_printf("texture[%u] = \n", i);
3117 debug_printf(" .format = %s\n",
3118 util_format_name(texture->format));
3119 debug_printf(" .target = %s\n",
3120 util_str_tex_target(texture->target, TRUE));
3121 debug_printf(" .level_zero_only = %u\n",
3122 texture->level_zero_only);
3123 debug_printf(" .pot = %u %u %u\n",
3124 texture->pot_width,
3125 texture->pot_height,
3126 texture->pot_depth);
3127 }
3128 struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3129 for (i = 0; i < key->nr_images; ++i) {
3130 const struct lp_static_texture_state *image = &images[i].image_state;
3131 debug_printf("image[%u] = \n", i);
3132 debug_printf(" .format = %s\n",
3133 util_format_name(image->format));
3134 debug_printf(" .target = %s\n",
3135 util_str_tex_target(image->target, TRUE));
3136 debug_printf(" .level_zero_only = %u\n",
3137 image->level_zero_only);
3138 debug_printf(" .pot = %u %u %u\n",
3139 image->pot_width,
3140 image->pot_height,
3141 image->pot_depth);
3142 }
3143 }
3144
3145
3146 void
3147 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3148 {
3149 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3150 variant->shader->no, variant->no);
3151 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3152 tgsi_dump(variant->shader->base.tokens, 0);
3153 else
3154 nir_print_shader(variant->shader->base.ir.nir, stderr);
3155 dump_fs_variant_key(&variant->key);
3156 debug_printf("variant->opaque = %u\n", variant->opaque);
3157 debug_printf("\n");
3158 }
3159
3160
3161 /**
3162 * Generate a new fragment shader variant from the shader code and
3163 * other state indicated by the key.
3164 */
3165 static struct lp_fragment_shader_variant *
3166 generate_variant(struct llvmpipe_context *lp,
3167 struct lp_fragment_shader *shader,
3168 const struct lp_fragment_shader_variant_key *key)
3169 {
3170 struct lp_fragment_shader_variant *variant;
3171 const struct util_format_description *cbuf0_format_desc = NULL;
3172 boolean fullcolormask;
3173 char module_name[64];
3174
3175 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3176 if (!variant)
3177 return NULL;
3178
3179 memset(variant, 0, sizeof(*variant));
3180 snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3181 shader->no, shader->variants_created);
3182
3183 variant->gallivm = gallivm_create(module_name, lp->context);
3184 if (!variant->gallivm) {
3185 FREE(variant);
3186 return NULL;
3187 }
3188
3189 variant->shader = shader;
3190 variant->list_item_global.base = variant;
3191 variant->list_item_local.base = variant;
3192 variant->no = shader->variants_created++;
3193
3194 memcpy(&variant->key, key, shader->variant_key_size);
3195
3196 /*
3197 * Determine whether we are touching all channels in the color buffer.
3198 */
3199 fullcolormask = FALSE;
3200 if (key->nr_cbufs == 1) {
3201 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3202 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3203 }
3204
3205 variant->opaque =
3206 !key->blend.logicop_enable &&
3207 !key->blend.rt[0].blend_enable &&
3208 fullcolormask &&
3209 !key->stencil[0].enabled &&
3210 !key->alpha.enabled &&
3211 !key->blend.alpha_to_coverage &&
3212 !key->depth.enabled &&
3213 !shader->info.base.uses_kill &&
3214 !shader->info.base.writes_samplemask
3215 ? TRUE : FALSE;
3216
3217 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3218 lp_debug_fs_variant(variant);
3219 }
3220
3221 lp_jit_init_types(variant);
3222
3223 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
3224 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
3225
3226 if (variant->jit_function[RAST_WHOLE] == NULL) {
3227 if (variant->opaque) {
3228 /* Specialized shader, which doesn't need to read the color buffer. */
3229 generate_fragment(lp, shader, variant, RAST_WHOLE);
3230 }
3231 }
3232
3233 /*
3234 * Compile everything
3235 */
3236
3237 gallivm_compile_module(variant->gallivm);
3238
3239 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
3240
3241 if (variant->function[RAST_EDGE_TEST]) {
3242 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
3243 gallivm_jit_function(variant->gallivm,
3244 variant->function[RAST_EDGE_TEST]);
3245 }
3246
3247 if (variant->function[RAST_WHOLE]) {
3248 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
3249 gallivm_jit_function(variant->gallivm,
3250 variant->function[RAST_WHOLE]);
3251 } else if (!variant->jit_function[RAST_WHOLE]) {
3252 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
3253 }
3254
3255 gallivm_free_ir(variant->gallivm);
3256
3257 return variant;
3258 }
3259
3260
3261 static void *
3262 llvmpipe_create_fs_state(struct pipe_context *pipe,
3263 const struct pipe_shader_state *templ)
3264 {
3265 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3266 struct lp_fragment_shader *shader;
3267 int nr_samplers;
3268 int nr_sampler_views;
3269 int nr_images;
3270 int i;
3271
3272 shader = CALLOC_STRUCT(lp_fragment_shader);
3273 if (!shader)
3274 return NULL;
3275
3276 shader->no = fs_no++;
3277 make_empty_list(&shader->variants);
3278
3279 shader->base.type = templ->type;
3280 if (templ->type == PIPE_SHADER_IR_TGSI) {
3281 /* get/save the summary info for this shader */
3282 lp_build_tgsi_info(templ->tokens, &shader->info);
3283
3284 /* we need to keep a local copy of the tokens */
3285 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
3286 } else {
3287 shader->base.ir.nir = templ->ir.nir;
3288 nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
3289 }
3290
3291 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
3292 if (shader->draw_data == NULL) {
3293 FREE((void *) shader->base.tokens);
3294 FREE(shader);
3295 return NULL;
3296 }
3297
3298 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3299 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3300 nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3301 shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
3302
3303 for (i = 0; i < shader->info.base.num_inputs; i++) {
3304 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
3305 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
3306 shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
3307
3308 switch (shader->info.base.input_interpolate[i]) {
3309 case TGSI_INTERPOLATE_CONSTANT:
3310 shader->inputs[i].interp = LP_INTERP_CONSTANT;
3311 break;
3312 case TGSI_INTERPOLATE_LINEAR:
3313 shader->inputs[i].interp = LP_INTERP_LINEAR;
3314 break;
3315 case TGSI_INTERPOLATE_PERSPECTIVE:
3316 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
3317 break;
3318 case TGSI_INTERPOLATE_COLOR:
3319 shader->inputs[i].interp = LP_INTERP_COLOR;
3320 break;
3321 default:
3322 assert(0);
3323 break;
3324 }
3325
3326 switch (shader->info.base.input_semantic_name[i]) {
3327 case TGSI_SEMANTIC_FACE:
3328 shader->inputs[i].interp = LP_INTERP_FACING;
3329 break;
3330 case TGSI_SEMANTIC_POSITION:
3331 /* Position was already emitted above
3332 */
3333 shader->inputs[i].interp = LP_INTERP_POSITION;
3334 shader->inputs[i].src_index = 0;
3335 continue;
3336 }
3337
3338 /* XXX this is a completely pointless index map... */
3339 shader->inputs[i].src_index = i+1;
3340 }
3341
3342 if (LP_DEBUG & DEBUG_TGSI) {
3343 unsigned attrib;
3344 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3345 shader->no, (void *) shader);
3346 tgsi_dump(templ->tokens, 0);
3347 debug_printf("usage masks:\n");
3348 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
3349 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
3350 debug_printf(" IN[%u].%s%s%s%s\n",
3351 attrib,
3352 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
3353 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
3354 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
3355 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
3356 }
3357 debug_printf("\n");
3358 }
3359
3360 return shader;
3361 }
3362
3363
3364 static void
3365 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
3366 {
3367 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3368 struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
3369 if (llvmpipe->fs == lp_fs)
3370 return;
3371
3372 draw_bind_fragment_shader(llvmpipe->draw,
3373 (lp_fs ? lp_fs->draw_data : NULL));
3374
3375 llvmpipe->fs = lp_fs;
3376
3377 llvmpipe->dirty |= LP_NEW_FS;
3378 }
3379
3380
3381 /**
3382 * Remove shader variant from two lists: the shader's variant list
3383 * and the context's variant list.
3384 */
3385 static void
3386 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3387 struct lp_fragment_shader_variant *variant)
3388 {
3389 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3390 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3391 "v total cached %u inst %u total inst %u\n",
3392 variant->shader->no, variant->no,
3393 variant->shader->variants_created,
3394 variant->shader->variants_cached,
3395 lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3396 }
3397
3398 gallivm_destroy(variant->gallivm);
3399
3400 /* remove from shader's list */
3401 remove_from_list(&variant->list_item_local);
3402 variant->shader->variants_cached--;
3403
3404 /* remove from context's list */
3405 remove_from_list(&variant->list_item_global);
3406 lp->nr_fs_variants--;
3407 lp->nr_fs_instrs -= variant->nr_instrs;
3408
3409 FREE(variant);
3410 }
3411
3412
3413 static void
3414 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3415 {
3416 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3417 struct lp_fragment_shader *shader = fs;
3418 struct lp_fs_variant_list_item *li;
3419
3420 assert(fs != llvmpipe->fs);
3421
3422 /*
3423 * XXX: we need to flush the context until we have some sort of reference
3424 * counting in fragment shaders as they may still be binned
3425 * Flushing alone might not sufficient we need to wait on it too.
3426 */
3427 llvmpipe_finish(pipe, __FUNCTION__);
3428
3429 /* Delete all the variants */
3430 li = first_elem(&shader->variants);
3431 while(!at_end(&shader->variants, li)) {
3432 struct lp_fs_variant_list_item *next = next_elem(li);
3433 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3434 li = next;
3435 }
3436
3437 /* Delete draw module's data */
3438 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3439
3440 if (shader->base.ir.nir)
3441 ralloc_free(shader->base.ir.nir);
3442 assert(shader->variants_cached == 0);
3443 FREE((void *) shader->base.tokens);
3444 FREE(shader);
3445 }
3446
3447
3448
3449 static void
3450 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3451 enum pipe_shader_type shader, uint index,
3452 const struct pipe_constant_buffer *cb)
3453 {
3454 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3455 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3456
3457 assert(shader < PIPE_SHADER_TYPES);
3458 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3459
3460 /* note: reference counting */
3461 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3462
3463 if (constants) {
3464 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3465 debug_printf("Illegal set constant without bind flag\n");
3466 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3467 }
3468 }
3469
3470 if (shader == PIPE_SHADER_VERTEX ||
3471 shader == PIPE_SHADER_GEOMETRY ||
3472 shader == PIPE_SHADER_TESS_CTRL ||
3473 shader == PIPE_SHADER_TESS_EVAL) {
3474 /* Pass the constants to the 'draw' module */
3475 const unsigned size = cb ? cb->buffer_size : 0;
3476 const ubyte *data;
3477
3478 if (constants) {
3479 data = (ubyte *) llvmpipe_resource_data(constants);
3480 }
3481 else if (cb && cb->user_buffer) {
3482 data = (ubyte *) cb->user_buffer;
3483 }
3484 else {
3485 data = NULL;
3486 }
3487
3488 if (data)
3489 data += cb->buffer_offset;
3490
3491 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3492 index, data, size);
3493 }
3494 else if (shader == PIPE_SHADER_COMPUTE)
3495 llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
3496 else
3497 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3498
3499 if (cb && cb->user_buffer) {
3500 pipe_resource_reference(&constants, NULL);
3501 }
3502 }
3503
3504 static void
3505 llvmpipe_set_shader_buffers(struct pipe_context *pipe,
3506 enum pipe_shader_type shader, unsigned start_slot,
3507 unsigned count, const struct pipe_shader_buffer *buffers,
3508 unsigned writable_bitmask)
3509 {
3510 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3511 unsigned i, idx;
3512 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3513 const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
3514
3515 util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
3516
3517 if (shader == PIPE_SHADER_VERTEX ||
3518 shader == PIPE_SHADER_GEOMETRY ||
3519 shader == PIPE_SHADER_TESS_CTRL ||
3520 shader == PIPE_SHADER_TESS_EVAL) {
3521 const unsigned size = buffer ? buffer->buffer_size : 0;
3522 const ubyte *data = NULL;
3523 if (buffer && buffer->buffer)
3524 data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
3525 if (data)
3526 data += buffer->buffer_offset;
3527 draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
3528 i, data, size);
3529 } else if (shader == PIPE_SHADER_COMPUTE) {
3530 llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
3531 } else if (shader == PIPE_SHADER_FRAGMENT) {
3532 llvmpipe->dirty |= LP_NEW_FS_SSBOS;
3533 }
3534 }
3535 }
3536
3537 static void
3538 llvmpipe_set_shader_images(struct pipe_context *pipe,
3539 enum pipe_shader_type shader, unsigned start_slot,
3540 unsigned count, const struct pipe_image_view *images)
3541 {
3542 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3543 unsigned i, idx;
3544
3545 draw_flush(llvmpipe->draw);
3546 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3547 const struct pipe_image_view *image = images ? &images[idx] : NULL;
3548
3549 util_copy_image_view(&llvmpipe->images[shader][i], image);
3550 }
3551
3552 llvmpipe->num_images[shader] = start_slot + count;
3553 if (shader == PIPE_SHADER_VERTEX ||
3554 shader == PIPE_SHADER_GEOMETRY ||
3555 shader == PIPE_SHADER_TESS_CTRL ||
3556 shader == PIPE_SHADER_TESS_EVAL) {
3557 draw_set_images(llvmpipe->draw,
3558 shader,
3559 llvmpipe->images[shader],
3560 start_slot + count);
3561 } else if (shader == PIPE_SHADER_COMPUTE)
3562 llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
3563 else
3564 llvmpipe->dirty |= LP_NEW_FS_IMAGES;
3565 }
3566
3567 /**
3568 * Return the blend factor equivalent to a destination alpha of one.
3569 */
3570 static inline unsigned
3571 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3572 {
3573 switch(factor) {
3574 case PIPE_BLENDFACTOR_DST_ALPHA:
3575 return PIPE_BLENDFACTOR_ONE;
3576 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3577 return PIPE_BLENDFACTOR_ZERO;
3578 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3579 if (clamped_zero)
3580 return PIPE_BLENDFACTOR_ZERO;
3581 else
3582 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3583 }
3584
3585 return factor;
3586 }
3587
3588
3589 /**
3590 * We need to generate several variants of the fragment pipeline to match
3591 * all the combinations of the contributing state atoms.
3592 *
3593 * TODO: there is actually no reason to tie this to context state -- the
3594 * generated code could be cached globally in the screen.
3595 */
3596 static struct lp_fragment_shader_variant_key *
3597 make_variant_key(struct llvmpipe_context *lp,
3598 struct lp_fragment_shader *shader,
3599 char *store)
3600 {
3601 unsigned i;
3602 struct lp_fragment_shader_variant_key *key;
3603
3604 key = (struct lp_fragment_shader_variant_key *)store;
3605
3606 memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
3607
3608 if (lp->framebuffer.zsbuf) {
3609 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3610 const struct util_format_description *zsbuf_desc =
3611 util_format_description(zsbuf_format);
3612
3613 if (lp->depth_stencil->depth.enabled &&
3614 util_format_has_depth(zsbuf_desc)) {
3615 key->zsbuf_format = zsbuf_format;
3616 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3617 }
3618 if (lp->depth_stencil->stencil[0].enabled &&
3619 util_format_has_stencil(zsbuf_desc)) {
3620 key->zsbuf_format = zsbuf_format;
3621 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3622 }
3623 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3624 key->resource_1d = TRUE;
3625 }
3626 key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
3627 }
3628
3629 /*
3630 * Propagate the depth clamp setting from the rasterizer state.
3631 * depth_clip == 0 implies depth clamping is enabled.
3632 *
3633 * When clip_halfz is enabled, then always clamp the depth values.
3634 *
3635 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3636 * clamp is always active in d3d10, regardless if depth clip is
3637 * enabled or not).
3638 * (GL has an always-on [0,1] clamp on fs depth output instead
3639 * to ensure the depth values stay in range. Doesn't look like
3640 * we do that, though...)
3641 */
3642 if (lp->rasterizer->clip_halfz) {
3643 key->depth_clamp = 1;
3644 } else {
3645 key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
3646 }
3647
3648 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3649 if (!lp->framebuffer.nr_cbufs ||
3650 !lp->framebuffer.cbufs[0] ||
3651 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3652 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3653 }
3654 if(key->alpha.enabled)
3655 key->alpha.func = lp->depth_stencil->alpha.func;
3656 /* alpha.ref_value is passed in jit_context */
3657
3658 key->flatshade = lp->rasterizer->flatshade;
3659 key->multisample = lp->rasterizer->multisample;
3660 if (lp->active_occlusion_queries && !lp->queries_disabled) {
3661 key->occlusion_count = TRUE;
3662 }
3663
3664 if (lp->framebuffer.nr_cbufs) {
3665 memcpy(&key->blend, lp->blend, sizeof key->blend);
3666 }
3667
3668 key->coverage_samples = 1;
3669 if (key->multisample)
3670 key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
3671 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3672
3673 if (!key->blend.independent_blend_enable) {
3674 /* we always need independent blend otherwise the fixups below won't work */
3675 for (i = 1; i < key->nr_cbufs; i++) {
3676 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3677 }
3678 key->blend.independent_blend_enable = 1;
3679 }
3680
3681 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3682 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3683
3684 if (lp->framebuffer.cbufs[i]) {
3685 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3686 const struct util_format_description *format_desc;
3687
3688 key->cbuf_format[i] = format;
3689 key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
3690
3691 /*
3692 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3693 * mixing of 2d textures with height 1 and 1d textures, so make sure
3694 * we pick 1d if any cbuf or zsbuf is 1d.
3695 */
3696 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3697 key->resource_1d = TRUE;
3698 }
3699
3700 format_desc = util_format_description(format);
3701 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3702 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3703
3704 /*
3705 * Mask out color channels not present in the color buffer.
3706 */
3707 blend_rt->colormask &= util_format_colormask(format_desc);
3708
3709 /*
3710 * Disable blend for integer formats.
3711 */
3712 if (util_format_is_pure_integer(format)) {
3713 blend_rt->blend_enable = 0;
3714 }
3715
3716 /*
3717 * Our swizzled render tiles always have an alpha channel, but the
3718 * linear render target format often does not, so force here the dst
3719 * alpha to be one.
3720 *
3721 * This is not a mere optimization. Wrong results will be produced if
3722 * the dst alpha is used, the dst format does not have alpha, and the
3723 * previous rendering was not flushed from the swizzled to linear
3724 * buffer. For example, NonPowTwo DCT.
3725 *
3726 * TODO: This should be generalized to all channels for better
3727 * performance, but only alpha causes correctness issues.
3728 *
3729 * Also, force rgb/alpha func/factors match, to make AoS blending
3730 * easier.
3731 */
3732 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3733 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3734 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3735 boolean clamped_zero = !util_format_is_float(format) &&
3736 !util_format_is_snorm(format);
3737 blend_rt->rgb_src_factor =
3738 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3739 blend_rt->rgb_dst_factor =
3740 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3741 blend_rt->alpha_func = blend_rt->rgb_func;
3742 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3743 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3744 }
3745 }
3746 else {
3747 /* no color buffer for this fragment output */
3748 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3749 key->cbuf_nr_samples[i] = 0;
3750 blend_rt->colormask = 0x0;
3751 blend_rt->blend_enable = 0;
3752 }
3753 }
3754
3755 /* This value will be the same for all the variants of a given shader:
3756 */
3757 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3758
3759 struct lp_sampler_static_state *fs_sampler;
3760
3761 fs_sampler = key->samplers;
3762
3763 memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
3764
3765 for(i = 0; i < key->nr_samplers; ++i) {
3766 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3767 lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
3768 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3769 }
3770 }
3771
3772 /*
3773 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3774 * are dx10-style? Can't really have mixed opcodes, at least not
3775 * if we want to skip the holes here (without rescanning tgsi).
3776 */
3777 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3778 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3779 for(i = 0; i < key->nr_sampler_views; ++i) {
3780 /*
3781 * Note sview may exceed what's representable by file_mask.
3782 * This will still work, the only downside is that not actually
3783 * used views may be included in the shader key.
3784 */
3785 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
3786 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3787 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3788 }
3789 }
3790 }
3791 else {
3792 key->nr_sampler_views = key->nr_samplers;
3793 for(i = 0; i < key->nr_sampler_views; ++i) {
3794 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3795 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3796 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3797 }
3798 }
3799 }
3800
3801 struct lp_image_static_state *lp_image;
3802 lp_image = lp_fs_variant_key_images(key);
3803 key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3804 for (i = 0; i < key->nr_images; ++i) {
3805 if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
3806 lp_sampler_static_texture_state_image(&lp_image[i].image_state,
3807 &lp->images[PIPE_SHADER_FRAGMENT][i]);
3808 }
3809 }
3810 return key;
3811 }
3812
3813
3814
3815 /**
3816 * Update fragment shader state. This is called just prior to drawing
3817 * something when some fragment-related state has changed.
3818 */
3819 void
3820 llvmpipe_update_fs(struct llvmpipe_context *lp)
3821 {
3822 struct lp_fragment_shader *shader = lp->fs;
3823 struct lp_fragment_shader_variant_key *key;
3824 struct lp_fragment_shader_variant *variant = NULL;
3825 struct lp_fs_variant_list_item *li;
3826 char store[LP_FS_MAX_VARIANT_KEY_SIZE];
3827
3828 key = make_variant_key(lp, shader, store);
3829
3830 /* Search the variants for one which matches the key */
3831 li = first_elem(&shader->variants);
3832 while(!at_end(&shader->variants, li)) {
3833 if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
3834 variant = li->base;
3835 break;
3836 }
3837 li = next_elem(li);
3838 }
3839
3840 if (variant) {
3841 /* Move this variant to the head of the list to implement LRU
3842 * deletion of shader's when we have too many.
3843 */
3844 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3845 }
3846 else {
3847 /* variant not found, create it now */
3848 int64_t t0, t1, dt;
3849 unsigned i;
3850 unsigned variants_to_cull;
3851
3852 if (LP_DEBUG & DEBUG_FS) {
3853 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3854 lp->nr_fs_variants,
3855 lp->nr_fs_instrs,
3856 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3857 }
3858
3859 /* First, check if we've exceeded the max number of shader variants.
3860 * If so, free 6.25% of them (the least recently used ones).
3861 */
3862 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
3863
3864 if (variants_to_cull ||
3865 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3866 struct pipe_context *pipe = &lp->pipe;
3867
3868 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
3869 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3870 "\t%u instrs,\t%u instrs/variant\n",
3871 shader->variants_cached,
3872 lp->nr_fs_variants, lp->nr_fs_instrs,
3873 lp->nr_fs_instrs / lp->nr_fs_variants);
3874 }
3875
3876 /*
3877 * XXX: we need to flush the context until we have some sort of
3878 * reference counting in fragment shaders as they may still be binned
3879 * Flushing alone might not be sufficient we need to wait on it too.
3880 */
3881 llvmpipe_finish(pipe, __FUNCTION__);
3882
3883 /*
3884 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3885 * number of shader variants (potentially all of them) could be
3886 * pending for destruction on flush.
3887 */
3888
3889 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3890 struct lp_fs_variant_list_item *item;
3891 if (is_empty_list(&lp->fs_variants_list)) {
3892 break;
3893 }
3894 item = last_elem(&lp->fs_variants_list);
3895 assert(item);
3896 assert(item->base);
3897 llvmpipe_remove_shader_variant(lp, item->base);
3898 }
3899 }
3900
3901 /*
3902 * Generate the new variant.
3903 */
3904 t0 = os_time_get();
3905 variant = generate_variant(lp, shader, key);
3906 t1 = os_time_get();
3907 dt = t1 - t0;
3908 LP_COUNT_ADD(llvm_compile_time, dt);
3909 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3910
3911 /* Put the new variant into the list */
3912 if (variant) {
3913 insert_at_head(&shader->variants, &variant->list_item_local);
3914 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3915 lp->nr_fs_variants++;
3916 lp->nr_fs_instrs += variant->nr_instrs;
3917 shader->variants_cached++;
3918 }
3919 }
3920
3921 /* Bind this variant */
3922 lp_setup_set_fs_variant(lp->setup, variant);
3923 }
3924
3925
3926
3927
3928
3929 void
3930 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3931 {
3932 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3933 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3934 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3935
3936 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3937
3938 llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
3939 llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
3940 }
3941
3942