projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: move color storing earlier in frag shader
[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, mask_store, 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 /* Color write */
730 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
731 {
732 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
733 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
734 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
735 {
736 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
737 if(outputs[attrib][chan]) {
738 /* XXX: just initialize outputs to point at colors[] and
739 * skip this.
740 */
741 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
742 LLVMValueRef color_ptr;
743 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
744 &loop_state.counter, 1, "");
745 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
746 LLVMBuildStore(builder, out, color_ptr);
747 }
748 }
749 }
750 }
751
752
753 if (key->multisample) {
754 /* execute depth test for each sample */
755 lp_build_for_loop_begin(&sample_loop_state, gallivm,
756 lp_build_const_int32(gallivm, 0),
757 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
758 lp_build_const_int32(gallivm, 1));
759
760 /* load the per-sample coverage mask */
761 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
762 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
763 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
764
765 /* combine the execution mask post fragment shader with the coverage mask. */
766 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
767 s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
768
769 /* if the shader writes sample mask use that */
770 if (shader->info.base.writes_samplemask) {
771 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
772 out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
773 LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
774 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
775 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
776 smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
777
778 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
779 }
780 }
781
782 depth_ptr = depth_base_ptr;
783 if (key->multisample) {
784 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
785 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
786 }
787
788 /* Late Z test */
789 if (depth_mode & LATE_DEPTH_TEST) {
790 int pos0 = find_output_by_semantic(&shader->info.base,
791 TGSI_SEMANTIC_POSITION,
792 0);
793 int s_out = find_output_by_semantic(&shader->info.base,
794 TGSI_SEMANTIC_STENCIL,
795 0);
796 if (pos0 != -1 && outputs[pos0][2]) {
797 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
798 }
799 /*
800 * Clamp according to ARB_depth_clamp semantics.
801 */
802 if (key->depth_clamp) {
803 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
804 thread_data_ptr, z);
805 }
806
807 if (s_out != -1 && outputs[s_out][1]) {
808 /* there's only one value, and spec says to discard additional bits */
809 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
810 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
811 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
812 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
813 stencil_refs[1] = stencil_refs[0];
814 }
815
816 lp_build_depth_stencil_load_swizzled(gallivm, type,
817 zs_format_desc, key->resource_1d,
818 depth_ptr, depth_stride,
819 &z_fb, &s_fb, loop_state.counter);
820
821 lp_build_depth_stencil_test(gallivm,
822 &key->depth,
823 key->stencil,
824 type,
825 zs_format_desc,
826 key->multisample ? NULL : &mask,
827 &s_mask,
828 stencil_refs,
829 z, z_fb, s_fb,
830 facing,
831 &z_value, &s_value,
832 !simple_shader);
833 /* Late Z write */
834 if (depth_mode & LATE_DEPTH_WRITE) {
835 lp_build_depth_stencil_write_swizzled(gallivm, type,
836 zs_format_desc, key->resource_1d,
837 NULL, NULL, NULL, loop_state.counter,
838 depth_ptr, depth_stride,
839 z_value, s_value);
840 }
841 }
842 else if ((depth_mode & EARLY_DEPTH_TEST) &&
843 (depth_mode & LATE_DEPTH_WRITE))
844 {
845 /* Need to apply a reduced mask to the depth write. Reload the
846 * depth value, update from zs_value with the new mask value and
847 * write that out.
848 */
849 if (key->multisample) {
850 z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
851 s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
852 z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
853 s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
854 }
855 lp_build_depth_stencil_write_swizzled(gallivm, type,
856 zs_format_desc, key->resource_1d,
857 key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
858 depth_ptr, depth_stride,
859 z_value, s_value);
860 }
861
862 if (key->multisample) {
863 /* store the sample mask for this loop */
864 LLVMBuildStore(builder, s_mask, s_mask_ptr);
865 lp_build_for_loop_end(&sample_loop_state);
866 }
867
868 if (key->occlusion_count) {
869 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
870 lp_build_name(counter, "counter");
871 lp_build_occlusion_count(gallivm, type,
872 lp_build_mask_value(&mask), counter);
873 }
874
875 mask_val = lp_build_mask_end(&mask);
876 if (!key->multisample)
877 LLVMBuildStore(builder, mask_val, mask_ptr);
878 lp_build_for_loop_end(&loop_state);
879 }
880
881
882 /**
883 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
884 *
885 * Fragment Shader outputs pixels in small 2x2 blocks
886 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
887 *
888 * However in memory pixels are stored in rows
889 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
890 *
891 * @param type fragment shader type (4x or 8x float)
892 * @param num_fs number of fs_src
893 * @param is_1d whether we're outputting to a 1d resource
894 * @param dst_channels number of output channels
895 * @param fs_src output from fragment shader
896 * @param dst pointer to store result
897 * @param pad_inline is channel padding inline or at end of row
898 * @return the number of dsts
899 */
900 static int
901 generate_fs_twiddle(struct gallivm_state *gallivm,
902 struct lp_type type,
903 unsigned num_fs,
904 unsigned dst_channels,
905 LLVMValueRef fs_src[][4],
906 LLVMValueRef* dst,
907 bool pad_inline)
908 {
909 LLVMValueRef src[16];
910
911 bool swizzle_pad;
912 bool twiddle;
913 bool split;
914
915 unsigned pixels = type.length / 4;
916 unsigned reorder_group;
917 unsigned src_channels;
918 unsigned src_count;
919 unsigned i;
920
921 src_channels = dst_channels < 3 ? dst_channels : 4;
922 src_count = num_fs * src_channels;
923
924 assert(pixels == 2 || pixels == 1);
925 assert(num_fs * src_channels <= ARRAY_SIZE(src));
926
927 /*
928 * Transpose from SoA -> AoS
929 */
930 for (i = 0; i < num_fs; ++i) {
931 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
932 }
933
934 /*
935 * Pick transformation options
936 */
937 swizzle_pad = false;
938 twiddle = false;
939 split = false;
940 reorder_group = 0;
941
942 if (dst_channels == 1) {
943 twiddle = true;
944
945 if (pixels == 2) {
946 split = true;
947 }
948 } else if (dst_channels == 2) {
949 if (pixels == 1) {
950 reorder_group = 1;
951 }
952 } else if (dst_channels > 2) {
953 if (pixels == 1) {
954 reorder_group = 2;
955 } else {
956 twiddle = true;
957 }
958
959 if (!pad_inline && dst_channels == 3 && pixels > 1) {
960 swizzle_pad = true;
961 }
962 }
963
964 /*
965 * Split the src in half
966 */
967 if (split) {
968 for (i = num_fs; i > 0; --i) {
969 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
970 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
971 }
972
973 src_count *= 2;
974 type.length = 4;
975 }
976
977 /*
978 * Ensure pixels are in memory order
979 */
980 if (reorder_group) {
981 /* Twiddle pixels by reordering the array, e.g.:
982 *
983 * src_count = 8 -> 0 2 1 3 4 6 5 7
984 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
985 */
986 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
987
988 for (i = 0; i < src_count; ++i) {
989 unsigned group = i / reorder_group;
990 unsigned block = (group / 4) * 4 * reorder_group;
991 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
992 dst[i] = src[j];
993 }
994 } else if (twiddle) {
995 /* Twiddle pixels across elements of array */
996 /*
997 * XXX: we should avoid this in some cases, but would need to tell
998 * lp_build_conv to reorder (or deal with it ourselves).
999 */
1000 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1001 } else {
1002 /* Do nothing */
1003 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1004 }
1005
1006 /*
1007 * Moves any padding between pixels to the end
1008 * e.g. RGBXRGBX -> RGBRGBXX
1009 */
1010 if (swizzle_pad) {
1011 unsigned char swizzles[16];
1012 unsigned elems = pixels * dst_channels;
1013
1014 for (i = 0; i < type.length; ++i) {
1015 if (i < elems)
1016 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1017 else
1018 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1019 }
1020
1021 for (i = 0; i < src_count; ++i) {
1022 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1023 }
1024 }
1025
1026 return src_count;
1027 }
1028
1029
1030 /*
1031 * Untwiddle and transpose, much like the above.
1032 * However, this is after conversion, so we get packed vectors.
1033 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1034 * the vectors will look like:
1035 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1036 * be swizzled here). Extending to 16bit should be trivial.
1037 * Should also be extended to handle twice wide vectors with AVX2...
1038 */
1039 static void
1040 fs_twiddle_transpose(struct gallivm_state *gallivm,
1041 struct lp_type type,
1042 LLVMValueRef *src,
1043 unsigned src_count,
1044 LLVMValueRef *dst)
1045 {
1046 unsigned i, j;
1047 struct lp_type type64, type16, type32;
1048 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1049 LLVMBuilderRef builder = gallivm->builder;
1050 LLVMValueRef tmp[4], shuf[8];
1051 for (j = 0; j < 2; j++) {
1052 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1053 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1054 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1055 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1056 }
1057
1058 assert(src_count == 4 || src_count == 2 || src_count == 1);
1059 assert(type.width == 8);
1060 assert(type.length == 16);
1061
1062 type8_t = lp_build_vec_type(gallivm, type);
1063
1064 type64 = type;
1065 type64.length /= 8;
1066 type64.width *= 8;
1067 type64_t = lp_build_vec_type(gallivm, type64);
1068
1069 type16 = type;
1070 type16.length /= 2;
1071 type16.width *= 2;
1072 type16_t = lp_build_vec_type(gallivm, type16);
1073
1074 type32 = type;
1075 type32.length /= 4;
1076 type32.width *= 4;
1077 type32_t = lp_build_vec_type(gallivm, type32);
1078
1079 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1080
1081 if (src_count == 1) {
1082 /* transpose was no-op, just untwiddle */
1083 LLVMValueRef shuf_vec;
1084 shuf_vec = LLVMConstVector(shuf, 8);
1085 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1086 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1087 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1088 } else if (src_count == 2) {
1089 LLVMValueRef shuf_vec;
1090 shuf_vec = LLVMConstVector(shuf, 4);
1091
1092 for (i = 0; i < 2; i++) {
1093 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1094 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1095 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1096 }
1097 } else {
1098 for (j = 0; j < 2; j++) {
1099 LLVMValueRef lo, hi, lo2, hi2;
1100 /*
1101 * Note that if we only really have 3 valid channels (rgb)
1102 * and we don't need alpha we could substitute a undef here
1103 * for the respective channel (causing llvm to drop conversion
1104 * for alpha).
1105 */
1106 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1107 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1108 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1109 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1110 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1111 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1112 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1113 }
1114 }
1115 }
1116
1117
1118 /**
1119 * Load an unswizzled block of pixels from memory
1120 */
1121 static void
1122 load_unswizzled_block(struct gallivm_state *gallivm,
1123 LLVMValueRef base_ptr,
1124 LLVMValueRef stride,
1125 unsigned block_width,
1126 unsigned block_height,
1127 LLVMValueRef* dst,
1128 struct lp_type dst_type,
1129 unsigned dst_count,
1130 unsigned dst_alignment)
1131 {
1132 LLVMBuilderRef builder = gallivm->builder;
1133 unsigned row_size = dst_count / block_height;
1134 unsigned i;
1135
1136 /* Ensure block exactly fits into dst */
1137 assert((block_width * block_height) % dst_count == 0);
1138
1139 for (i = 0; i < dst_count; ++i) {
1140 unsigned x = i % row_size;
1141 unsigned y = i / row_size;
1142
1143 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1144 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1145
1146 LLVMValueRef gep[2];
1147 LLVMValueRef dst_ptr;
1148
1149 gep[0] = lp_build_const_int32(gallivm, 0);
1150 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1151
1152 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1153 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1154 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1155
1156 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1157
1158 LLVMSetAlignment(dst[i], dst_alignment);
1159 }
1160 }
1161
1162
1163 /**
1164 * Store an unswizzled block of pixels to memory
1165 */
1166 static void
1167 store_unswizzled_block(struct gallivm_state *gallivm,
1168 LLVMValueRef base_ptr,
1169 LLVMValueRef stride,
1170 unsigned block_width,
1171 unsigned block_height,
1172 LLVMValueRef* src,
1173 struct lp_type src_type,
1174 unsigned src_count,
1175 unsigned src_alignment)
1176 {
1177 LLVMBuilderRef builder = gallivm->builder;
1178 unsigned row_size = src_count / block_height;
1179 unsigned i;
1180
1181 /* Ensure src exactly fits into block */
1182 assert((block_width * block_height) % src_count == 0);
1183
1184 for (i = 0; i < src_count; ++i) {
1185 unsigned x = i % row_size;
1186 unsigned y = i / row_size;
1187
1188 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1189 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1190
1191 LLVMValueRef gep[2];
1192 LLVMValueRef src_ptr;
1193
1194 gep[0] = lp_build_const_int32(gallivm, 0);
1195 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1196
1197 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1198 src_ptr = LLVMBuildBitCast(builder, src_ptr,
1199 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1200
1201 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1202
1203 LLVMSetAlignment(src_ptr, src_alignment);
1204 }
1205 }
1206
1207
1208 /**
1209 * Checks if a format description is an arithmetic format
1210 *
1211 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1212 */
1213 static inline boolean
1214 is_arithmetic_format(const struct util_format_description *format_desc)
1215 {
1216 boolean arith = false;
1217 unsigned i;
1218
1219 for (i = 0; i < format_desc->nr_channels; ++i) {
1220 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1221 arith |= (format_desc->channel[i].size % 8) != 0;
1222 }
1223
1224 return arith;
1225 }
1226
1227
1228 /**
1229 * Checks if this format requires special handling due to required expansion
1230 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1231 * SoA conversion.
1232 */
1233 static inline boolean
1234 format_expands_to_float_soa(const struct util_format_description *format_desc)
1235 {
1236 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
1237 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1238 return true;
1239 }
1240 return false;
1241 }
1242
1243
1244 /**
1245 * Retrieves the type representing the memory layout for a format
1246 *
1247 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1248 */
1249 static inline void
1250 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1251 struct lp_type* type)
1252 {
1253 unsigned i;
1254 unsigned chan;
1255
1256 if (format_expands_to_float_soa(format_desc)) {
1257 /* just make this a uint with width of block */
1258 type->floating = false;
1259 type->fixed = false;
1260 type->sign = false;
1261 type->norm = false;
1262 type->width = format_desc->block.bits;
1263 type->length = 1;
1264 return;
1265 }
1266
1267 for (i = 0; i < 4; i++)
1268 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1269 break;
1270 chan = i;
1271
1272 memset(type, 0, sizeof(struct lp_type));
1273 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1274 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1275 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1276 type->norm = format_desc->channel[chan].normalized;
1277
1278 if (is_arithmetic_format(format_desc)) {
1279 type->width = 0;
1280 type->length = 1;
1281
1282 for (i = 0; i < format_desc->nr_channels; ++i) {
1283 type->width += format_desc->channel[i].size;
1284 }
1285 } else {
1286 type->width = format_desc->channel[chan].size;
1287 type->length = format_desc->nr_channels;
1288 }
1289 }
1290
1291
1292 /**
1293 * Retrieves the type for a format which is usable in the blending code.
1294 *
1295 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1296 */
1297 static inline void
1298 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1299 struct lp_type* type)
1300 {
1301 unsigned i;
1302 unsigned chan;
1303
1304 if (format_expands_to_float_soa(format_desc)) {
1305 /* always use ordinary floats for blending */
1306 type->floating = true;
1307 type->fixed = false;
1308 type->sign = true;
1309 type->norm = false;
1310 type->width = 32;
1311 type->length = 4;
1312 return;
1313 }
1314
1315 for (i = 0; i < 4; i++)
1316 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1317 break;
1318 chan = i;
1319
1320 memset(type, 0, sizeof(struct lp_type));
1321 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1322 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1323 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1324 type->norm = format_desc->channel[chan].normalized;
1325 type->width = format_desc->channel[chan].size;
1326 type->length = format_desc->nr_channels;
1327
1328 for (i = 1; i < format_desc->nr_channels; ++i) {
1329 if (format_desc->channel[i].size > type->width)
1330 type->width = format_desc->channel[i].size;
1331 }
1332
1333 if (type->floating) {
1334 type->width = 32;
1335 } else {
1336 if (type->width <= 8) {
1337 type->width = 8;
1338 } else if (type->width <= 16) {
1339 type->width = 16;
1340 } else {
1341 type->width = 32;
1342 }
1343 }
1344
1345 if (is_arithmetic_format(format_desc) && type->length == 3) {
1346 type->length = 4;
1347 }
1348 }
1349
1350
1351 /**
1352 * Scale a normalized value from src_bits to dst_bits.
1353 *
1354 * The exact calculation is
1355 *
1356 * dst = iround(src * dst_mask / src_mask)
1357 *
1358 * or with integer rounding
1359 *
1360 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1361 *
1362 * where
1363 *
1364 * src_mask = (1 << src_bits) - 1
1365 * dst_mask = (1 << dst_bits) - 1
1366 *
1367 * but we try to avoid division and multiplication through shifts.
1368 */
1369 static inline LLVMValueRef
1370 scale_bits(struct gallivm_state *gallivm,
1371 int src_bits,
1372 int dst_bits,
1373 LLVMValueRef src,
1374 struct lp_type src_type)
1375 {
1376 LLVMBuilderRef builder = gallivm->builder;
1377 LLVMValueRef result = src;
1378
1379 if (dst_bits < src_bits) {
1380 int delta_bits = src_bits - dst_bits;
1381
1382 if (delta_bits <= dst_bits) {
1383 /*
1384 * Approximate the rescaling with a single shift.
1385 *
1386 * This gives the wrong rounding.
1387 */
1388
1389 result = LLVMBuildLShr(builder,
1390 src,
1391 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1392 "");
1393
1394 } else {
1395 /*
1396 * Try more accurate rescaling.
1397 */
1398
1399 /*
1400 * Drop the least significant bits to make space for the multiplication.
1401 *
1402 * XXX: A better approach would be to use a wider integer type as intermediate. But
1403 * this is enough to convert alpha from 16bits -> 2 when rendering to
1404 * PIPE_FORMAT_R10G10B10A2_UNORM.
1405 */
1406 result = LLVMBuildLShr(builder,
1407 src,
1408 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1409 "");
1410
1411
1412 result = LLVMBuildMul(builder,
1413 result,
1414 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1415 "");
1416
1417 /*
1418 * Add a rounding term before the division.
1419 *
1420 * TODO: Handle signed integers too.
1421 */
1422 if (!src_type.sign) {
1423 result = LLVMBuildAdd(builder,
1424 result,
1425 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1426 "");
1427 }
1428
1429 /*
1430 * Approximate the division by src_mask with a src_bits shift.
1431 *
1432 * Given the src has already been shifted by dst_bits, all we need
1433 * to do is to shift by the difference.
1434 */
1435
1436 result = LLVMBuildLShr(builder,
1437 result,
1438 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1439 "");
1440 }
1441
1442 } else if (dst_bits > src_bits) {
1443 /* Scale up bits */
1444 int db = dst_bits - src_bits;
1445
1446 /* Shift left by difference in bits */
1447 result = LLVMBuildShl(builder,
1448 src,
1449 lp_build_const_int_vec(gallivm, src_type, db),
1450 "");
1451
1452 if (db <= src_bits) {
1453 /* Enough bits in src to fill the remainder */
1454 LLVMValueRef lower = LLVMBuildLShr(builder,
1455 src,
1456 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1457 "");
1458
1459 result = LLVMBuildOr(builder, result, lower, "");
1460 } else if (db > src_bits) {
1461 /* Need to repeatedly copy src bits to fill remainder in dst */
1462 unsigned n;
1463
1464 for (n = src_bits; n < dst_bits; n *= 2) {
1465 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1466
1467 result = LLVMBuildOr(builder,
1468 result,
1469 LLVMBuildLShr(builder, result, shuv, ""),
1470 "");
1471 }
1472 }
1473 }
1474
1475 return result;
1476 }
1477
1478 /**
1479 * If RT is a smallfloat (needing denorms) format
1480 */
1481 static inline int
1482 have_smallfloat_format(struct lp_type dst_type,
1483 enum pipe_format format)
1484 {
1485 return ((dst_type.floating && dst_type.width != 32) ||
1486 /* due to format handling hacks this format doesn't have floating set
1487 * here (and actually has width set to 32 too) so special case this. */
1488 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1489 }
1490
1491
1492 /**
1493 * Convert from memory format to blending format
1494 *
1495 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1496 */
1497 static void
1498 convert_to_blend_type(struct gallivm_state *gallivm,
1499 unsigned block_size,
1500 const struct util_format_description *src_fmt,
1501 struct lp_type src_type,
1502 struct lp_type dst_type,
1503 LLVMValueRef* src, // and dst
1504 unsigned num_srcs)
1505 {
1506 LLVMValueRef *dst = src;
1507 LLVMBuilderRef builder = gallivm->builder;
1508 struct lp_type blend_type;
1509 struct lp_type mem_type;
1510 unsigned i, j;
1511 unsigned pixels = block_size / num_srcs;
1512 bool is_arith;
1513
1514 /*
1515 * full custom path for packed floats and srgb formats - none of the later
1516 * functions would do anything useful, and given the lp_type representation they
1517 * can't be fixed. Should really have some SoA blend path for these kind of
1518 * formats rather than hacking them in here.
1519 */
1520 if (format_expands_to_float_soa(src_fmt)) {
1521 LLVMValueRef tmpsrc[4];
1522 /*
1523 * This is pretty suboptimal for this case blending in SoA would be much
1524 * better, since conversion gets us SoA values so need to convert back.
1525 */
1526 assert(src_type.width == 32 || src_type.width == 16);
1527 assert(dst_type.floating);
1528 assert(dst_type.width == 32);
1529 assert(dst_type.length % 4 == 0);
1530 assert(num_srcs % 4 == 0);
1531
1532 if (src_type.width == 16) {
1533 /* expand 4x16bit values to 4x32bit */
1534 struct lp_type type32x4 = src_type;
1535 LLVMTypeRef ltype32x4;
1536 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1537 type32x4.width = 32;
1538 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1539 for (i = 0; i < num_fetch; i++) {
1540 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1541 }
1542 src_type.width = 32;
1543 }
1544 for (i = 0; i < 4; i++) {
1545 tmpsrc[i] = src[i];
1546 }
1547 for (i = 0; i < num_srcs / 4; i++) {
1548 LLVMValueRef tmpsoa[4];
1549 LLVMValueRef tmps = tmpsrc[i];
1550 if (dst_type.length == 8) {
1551 LLVMValueRef shuffles[8];
1552 unsigned j;
1553 /* fetch was 4 values but need 8-wide output values */
1554 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1555 /*
1556 * for 8-wide aos transpose would give us wrong order not matching
1557 * incoming converted fs values and mask. ARGH.
1558 */
1559 for (j = 0; j < 4; j++) {
1560 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1561 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1562 }
1563 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1564 LLVMConstVector(shuffles, 8), "");
1565 }
1566 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1567 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1568 }
1569 else {
1570 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1571 }
1572 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1573 }
1574 return;
1575 }
1576
1577 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1578 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1579
1580 /* Is the format arithmetic */
1581 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1582 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1583
1584 /* Pad if necessary */
1585 if (!is_arith && src_type.length < dst_type.length) {
1586 for (i = 0; i < num_srcs; ++i) {
1587 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1588 }
1589
1590 src_type.length = dst_type.length;
1591 }
1592
1593 /* Special case for half-floats */
1594 if (mem_type.width == 16 && mem_type.floating) {
1595 assert(blend_type.width == 32 && blend_type.floating);
1596 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1597 is_arith = false;
1598 }
1599
1600 if (!is_arith) {
1601 return;
1602 }
1603
1604 src_type.width = blend_type.width * blend_type.length;
1605 blend_type.length *= pixels;
1606 src_type.length *= pixels / (src_type.length / mem_type.length);
1607
1608 for (i = 0; i < num_srcs; ++i) {
1609 LLVMValueRef chans[4];
1610 LLVMValueRef res = NULL;
1611
1612 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1613
1614 for (j = 0; j < src_fmt->nr_channels; ++j) {
1615 unsigned mask = 0;
1616 unsigned sa = src_fmt->channel[j].shift;
1617 #if UTIL_ARCH_LITTLE_ENDIAN
1618 unsigned from_lsb = j;
1619 #else
1620 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1621 #endif
1622
1623 mask = (1 << src_fmt->channel[j].size) - 1;
1624
1625 /* Extract bits from source */
1626 chans[j] = LLVMBuildLShr(builder,
1627 dst[i],
1628 lp_build_const_int_vec(gallivm, src_type, sa),
1629 "");
1630
1631 chans[j] = LLVMBuildAnd(builder,
1632 chans[j],
1633 lp_build_const_int_vec(gallivm, src_type, mask),
1634 "");
1635
1636 /* Scale bits */
1637 if (src_type.norm) {
1638 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1639 blend_type.width, chans[j], src_type);
1640 }
1641
1642 /* Insert bits into correct position */
1643 chans[j] = LLVMBuildShl(builder,
1644 chans[j],
1645 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1646 "");
1647
1648 if (j == 0) {
1649 res = chans[j];
1650 } else {
1651 res = LLVMBuildOr(builder, res, chans[j], "");
1652 }
1653 }
1654
1655 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1656 }
1657 }
1658
1659
1660 /**
1661 * Convert from blending format to memory format
1662 *
1663 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1664 */
1665 static void
1666 convert_from_blend_type(struct gallivm_state *gallivm,
1667 unsigned block_size,
1668 const struct util_format_description *src_fmt,
1669 struct lp_type src_type,
1670 struct lp_type dst_type,
1671 LLVMValueRef* src, // and dst
1672 unsigned num_srcs)
1673 {
1674 LLVMValueRef* dst = src;
1675 unsigned i, j, k;
1676 struct lp_type mem_type;
1677 struct lp_type blend_type;
1678 LLVMBuilderRef builder = gallivm->builder;
1679 unsigned pixels = block_size / num_srcs;
1680 bool is_arith;
1681
1682 /*
1683 * full custom path for packed floats and srgb formats - none of the later
1684 * functions would do anything useful, and given the lp_type representation they
1685 * can't be fixed. Should really have some SoA blend path for these kind of
1686 * formats rather than hacking them in here.
1687 */
1688 if (format_expands_to_float_soa(src_fmt)) {
1689 /*
1690 * This is pretty suboptimal for this case blending in SoA would be much
1691 * better - we need to transpose the AoS values back to SoA values for
1692 * conversion/packing.
1693 */
1694 assert(src_type.floating);
1695 assert(src_type.width == 32);
1696 assert(src_type.length % 4 == 0);
1697 assert(dst_type.width == 32 || dst_type.width == 16);
1698
1699 for (i = 0; i < num_srcs / 4; i++) {
1700 LLVMValueRef tmpsoa[4], tmpdst;
1701 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1702 /* really really need SoA here */
1703
1704 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1705 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1706 }
1707 else {
1708 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1709 src_type, tmpsoa);
1710 }
1711
1712 if (src_type.length == 8) {
1713 LLVMValueRef tmpaos, shuffles[8];
1714 unsigned j;
1715 /*
1716 * for 8-wide aos transpose has given us wrong order not matching
1717 * output order. HMPF. Also need to split the output values manually.
1718 */
1719 for (j = 0; j < 4; j++) {
1720 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1721 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1722 }
1723 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1724 LLVMConstVector(shuffles, 8), "");
1725 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1726 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1727 }
1728 else {
1729 src[i] = tmpdst;
1730 }
1731 }
1732 if (dst_type.width == 16) {
1733 struct lp_type type16x8 = dst_type;
1734 struct lp_type type32x4 = dst_type;
1735 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1736 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1737 type16x8.length = 8;
1738 type32x4.width = 32;
1739 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1740 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1741 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1742 /* We could do vector truncation but it doesn't generate very good code */
1743 for (i = 0; i < num_fetch; i++) {
1744 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1745 src[i], lp_build_zero(gallivm, type32x4));
1746 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1747 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1748 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1749 }
1750 }
1751 return;
1752 }
1753
1754 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1755 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1756
1757 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1758
1759 /* Special case for half-floats */
1760 if (mem_type.width == 16 && mem_type.floating) {
1761 int length = dst_type.length;
1762 assert(blend_type.width == 32 && blend_type.floating);
1763
1764 dst_type.length = src_type.length;
1765
1766 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1767
1768 dst_type.length = length;
1769 is_arith = false;
1770 }
1771
1772 /* Remove any padding */
1773 if (!is_arith && (src_type.length % mem_type.length)) {
1774 src_type.length -= (src_type.length % mem_type.length);
1775
1776 for (i = 0; i < num_srcs; ++i) {
1777 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1778 }
1779 }
1780
1781 /* No bit arithmetic to do */
1782 if (!is_arith) {
1783 return;
1784 }
1785
1786 src_type.length = pixels;
1787 src_type.width = blend_type.length * blend_type.width;
1788 dst_type.length = pixels;
1789
1790 for (i = 0; i < num_srcs; ++i) {
1791 LLVMValueRef chans[4];
1792 LLVMValueRef res = NULL;
1793
1794 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1795
1796 for (j = 0; j < src_fmt->nr_channels; ++j) {
1797 unsigned mask = 0;
1798 unsigned sa = src_fmt->channel[j].shift;
1799 unsigned sz_a = src_fmt->channel[j].size;
1800 #if UTIL_ARCH_LITTLE_ENDIAN
1801 unsigned from_lsb = j;
1802 #else
1803 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1804 #endif
1805
1806 assert(blend_type.width > src_fmt->channel[j].size);
1807
1808 for (k = 0; k < blend_type.width; ++k) {
1809 mask |= 1 << k;
1810 }
1811
1812 /* Extract bits */
1813 chans[j] = LLVMBuildLShr(builder,
1814 dst[i],
1815 lp_build_const_int_vec(gallivm, src_type,
1816 from_lsb * blend_type.width),
1817 "");
1818
1819 chans[j] = LLVMBuildAnd(builder,
1820 chans[j],
1821 lp_build_const_int_vec(gallivm, src_type, mask),
1822 "");
1823
1824 /* Scale down bits */
1825 if (src_type.norm) {
1826 chans[j] = scale_bits(gallivm, blend_type.width,
1827 src_fmt->channel[j].size, chans[j], src_type);
1828 } else if (!src_type.floating && sz_a < blend_type.width) {
1829 LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
1830 LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
1831 chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
1832 }
1833
1834 /* Insert bits */
1835 chans[j] = LLVMBuildShl(builder,
1836 chans[j],
1837 lp_build_const_int_vec(gallivm, src_type, sa),
1838 "");
1839
1840 sa += src_fmt->channel[j].size;
1841
1842 if (j == 0) {
1843 res = chans[j];
1844 } else {
1845 res = LLVMBuildOr(builder, res, chans[j], "");
1846 }
1847 }
1848
1849 assert (dst_type.width != 24);
1850
1851 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1852 }
1853 }
1854
1855
1856 /**
1857 * Convert alpha to same blend type as src
1858 */
1859 static void
1860 convert_alpha(struct gallivm_state *gallivm,
1861 struct lp_type row_type,
1862 struct lp_type alpha_type,
1863 const unsigned block_size,
1864 const unsigned block_height,
1865 const unsigned src_count,
1866 const unsigned dst_channels,
1867 const bool pad_inline,
1868 LLVMValueRef* src_alpha)
1869 {
1870 LLVMBuilderRef builder = gallivm->builder;
1871 unsigned i, j;
1872 unsigned length = row_type.length;
1873 row_type.length = alpha_type.length;
1874
1875 /* Twiddle the alpha to match pixels */
1876 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1877
1878 /*
1879 * TODO this should use single lp_build_conv call for
1880 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1881 */
1882 for (i = 0; i < block_height; ++i) {
1883 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1884 }
1885
1886 alpha_type = row_type;
1887 row_type.length = length;
1888
1889 /* If only one channel we can only need the single alpha value per pixel */
1890 if (src_count == 1 && dst_channels == 1) {
1891
1892 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1893 } else {
1894 /* If there are more srcs than rows then we need to split alpha up */
1895 if (src_count > block_height) {
1896 for (i = src_count; i > 0; --i) {
1897 unsigned pixels = block_size / src_count;
1898 unsigned idx = i - 1;
1899
1900 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1901 (idx * pixels) % 4, pixels);
1902 }
1903 }
1904
1905 /* If there is a src for each pixel broadcast the alpha across whole row */
1906 if (src_count == block_size) {
1907 for (i = 0; i < src_count; ++i) {
1908 src_alpha[i] = lp_build_broadcast(gallivm,
1909 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1910 }
1911 } else {
1912 unsigned pixels = block_size / src_count;
1913 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1914 unsigned alpha_span = 1;
1915 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1916
1917 /* Check if we need 2 src_alphas for our shuffles */
1918 if (pixels > alpha_type.length) {
1919 alpha_span = 2;
1920 }
1921
1922 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1923 for (j = 0; j < row_type.length; ++j) {
1924 if (j < pixels * channels) {
1925 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1926 } else {
1927 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1928 }
1929 }
1930
1931 for (i = 0; i < src_count; ++i) {
1932 unsigned idx1 = i, idx2 = i;
1933
1934 if (alpha_span > 1){
1935 idx1 *= alpha_span;
1936 idx2 = idx1 + 1;
1937 }
1938
1939 src_alpha[i] = LLVMBuildShuffleVector(builder,
1940 src_alpha[idx1],
1941 src_alpha[idx2],
1942 LLVMConstVector(shuffles, row_type.length),
1943 "");
1944 }
1945 }
1946 }
1947 }
1948
1949
1950 /**
1951 * Generates the blend function for unswizzled colour buffers
1952 * Also generates the read & write from colour buffer
1953 */
1954 static void
1955 generate_unswizzled_blend(struct gallivm_state *gallivm,
1956 unsigned rt,
1957 struct lp_fragment_shader_variant *variant,
1958 enum pipe_format out_format,
1959 unsigned int num_fs,
1960 struct lp_type fs_type,
1961 LLVMValueRef* fs_mask,
1962 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1963 LLVMValueRef context_ptr,
1964 LLVMValueRef color_ptr,
1965 LLVMValueRef stride,
1966 unsigned partial_mask,
1967 boolean do_branch)
1968 {
1969 const unsigned alpha_channel = 3;
1970 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1971 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1972 const unsigned block_size = block_width * block_height;
1973 const unsigned lp_integer_vector_width = 128;
1974
1975 LLVMBuilderRef builder = gallivm->builder;
1976 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1977 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1978 LLVMValueRef src_alpha[4 * 4];
1979 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1980 LLVMValueRef src_mask[4 * 4];
1981 LLVMValueRef src[4 * 4];
1982 LLVMValueRef src1[4 * 4];
1983 LLVMValueRef dst[4 * 4];
1984 LLVMValueRef blend_color;
1985 LLVMValueRef blend_alpha;
1986 LLVMValueRef i32_zero;
1987 LLVMValueRef check_mask;
1988 LLVMValueRef undef_src_val;
1989
1990 struct lp_build_mask_context mask_ctx;
1991 struct lp_type mask_type;
1992 struct lp_type blend_type;
1993 struct lp_type row_type;
1994 struct lp_type dst_type;
1995 struct lp_type ls_type;
1996
1997 unsigned char swizzle[TGSI_NUM_CHANNELS];
1998 unsigned vector_width;
1999 unsigned src_channels = TGSI_NUM_CHANNELS;
2000 unsigned dst_channels;
2001 unsigned dst_count;
2002 unsigned src_count;
2003 unsigned i, j;
2004
2005 const struct util_format_description* out_format_desc = util_format_description(out_format);
2006
2007 unsigned dst_alignment;
2008
2009 bool pad_inline = is_arithmetic_format(out_format_desc);
2010 bool has_alpha = false;
2011 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2012 util_blend_state_is_dual(&variant->key.blend, 0);
2013
2014 const boolean is_1d = variant->key.resource_1d;
2015 boolean twiddle_after_convert = FALSE;
2016 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2017 LLVMValueRef fpstate = 0;
2018
2019 /* Get type from output format */
2020 lp_blend_type_from_format_desc(out_format_desc, &row_type);
2021 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2022
2023 /*
2024 * Technically this code should go into lp_build_smallfloat_to_float
2025 * and lp_build_float_to_smallfloat but due to the
2026 * http://llvm.org/bugs/show_bug.cgi?id=6393
2027 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2028 * So the ordering is important here and there shouldn't be any
2029 * llvm ir instrunctions in this function before
2030 * this, otherwise half-float format conversions won't work
2031 * (again due to llvm bug #6393).
2032 */
2033 if (have_smallfloat_format(dst_type, out_format)) {
2034 /* We need to make sure that denorms are ok for half float
2035 conversions */
2036 fpstate = lp_build_fpstate_get(gallivm);
2037 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2038 }
2039
2040 mask_type = lp_int32_vec4_type();
2041 mask_type.length = fs_type.length;
2042
2043 for (i = num_fs; i < num_fullblock_fs; i++) {
2044 fs_mask[i] = lp_build_zero(gallivm, mask_type);
2045 }
2046
2047 /* Do not bother executing code when mask is empty.. */
2048 if (do_branch) {
2049 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2050
2051 for (i = 0; i < num_fullblock_fs; ++i) {
2052 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2053 }
2054
2055 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2056 lp_build_mask_check(&mask_ctx);
2057 }
2058
2059 partial_mask |= !variant->opaque;
2060 i32_zero = lp_build_const_int32(gallivm, 0);
2061
2062 undef_src_val = lp_build_undef(gallivm, fs_type);
2063
2064 row_type.length = fs_type.length;
2065 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2066
2067 /* Compute correct swizzle and count channels */
2068 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2069 dst_channels = 0;
2070
2071 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2072 /* Ensure channel is used */
2073 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2074 continue;
2075 }
2076
2077 /* Ensure not already written to (happens in case with GL_ALPHA) */
2078 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2079 continue;
2080 }
2081
2082 /* Ensure we havn't already found all channels */
2083 if (dst_channels >= out_format_desc->nr_channels) {
2084 continue;
2085 }
2086
2087 swizzle[out_format_desc->swizzle[i]] = i;
2088 ++dst_channels;
2089
2090 if (i == alpha_channel) {
2091 has_alpha = true;
2092 }
2093 }
2094
2095 if (format_expands_to_float_soa(out_format_desc)) {
2096 /*
2097 * the code above can't work for layout_other
2098 * for srgb it would sort of work but we short-circuit swizzles, etc.
2099 * as that is done as part of unpack / pack.
2100 */
2101 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2102 has_alpha = true;
2103 swizzle[0] = 0;
2104 swizzle[1] = 1;
2105 swizzle[2] = 2;
2106 swizzle[3] = 3;
2107 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2108 }
2109
2110 /* If 3 channels then pad to include alpha for 4 element transpose */
2111 if (dst_channels == 3) {
2112 assert (!has_alpha);
2113 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2114 if (swizzle[i] > TGSI_NUM_CHANNELS)
2115 swizzle[i] = 3;
2116 }
2117 if (out_format_desc->nr_channels == 4) {
2118 dst_channels = 4;
2119 /*
2120 * We use alpha from the color conversion, not separate one.
2121 * We had to include it for transpose, hence it will get converted
2122 * too (albeit when doing transpose after conversion, that would
2123 * no longer be the case necessarily).
2124 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2125 * otherwise we really have padding, not alpha, included.)
2126 */
2127 has_alpha = true;
2128 }
2129 }
2130
2131 /*
2132 * Load shader output
2133 */
2134 for (i = 0; i < num_fullblock_fs; ++i) {
2135 /* Always load alpha for use in blending */
2136 LLVMValueRef alpha;
2137 if (i < num_fs) {
2138 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2139 }
2140 else {
2141 alpha = undef_src_val;
2142 }
2143
2144 /* Load each channel */
2145 for (j = 0; j < dst_channels; ++j) {
2146 assert(swizzle[j] < 4);
2147 if (i < num_fs) {
2148 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2149 }
2150 else {
2151 fs_src[i][j] = undef_src_val;
2152 }
2153 }
2154
2155 /* If 3 channels then pad to include alpha for 4 element transpose */
2156 /*
2157 * XXX If we include that here maybe could actually use it instead of
2158 * separate alpha for blending?
2159 * (Difficult though we actually convert pad channels, not alpha.)
2160 */
2161 if (dst_channels == 3 && !has_alpha) {
2162 fs_src[i][3] = alpha;
2163 }
2164
2165 /* We split the row_mask and row_alpha as we want 128bit interleave */
2166 if (fs_type.length == 8) {
2167 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
2168 0, src_channels);
2169 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
2170 src_channels, src_channels);
2171
2172 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2173 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2174 src_channels, src_channels);
2175 } else {
2176 src_mask[i] = fs_mask[i];
2177 src_alpha[i] = alpha;
2178 }
2179 }
2180 if (dual_source_blend) {
2181 /* same as above except different src/dst, skip masks and comments... */
2182 for (i = 0; i < num_fullblock_fs; ++i) {
2183 LLVMValueRef alpha;
2184 if (i < num_fs) {
2185 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2186 }
2187 else {
2188 alpha = undef_src_val;
2189 }
2190
2191 for (j = 0; j < dst_channels; ++j) {
2192 assert(swizzle[j] < 4);
2193 if (i < num_fs) {
2194 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2195 }
2196 else {
2197 fs_src1[i][j] = undef_src_val;
2198 }
2199 }
2200 if (dst_channels == 3 && !has_alpha) {
2201 fs_src1[i][3] = alpha;
2202 }
2203 if (fs_type.length == 8) {
2204 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2205 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2206 src_channels, src_channels);
2207 } else {
2208 src1_alpha[i] = alpha;
2209 }
2210 }
2211 }
2212
2213 if (util_format_is_pure_integer(out_format)) {
2214 /*
2215 * In this case fs_type was really ints or uints disguised as floats,
2216 * fix that up now.
2217 */
2218 fs_type.floating = 0;
2219 fs_type.sign = dst_type.sign;
2220 for (i = 0; i < num_fullblock_fs; ++i) {
2221 for (j = 0; j < dst_channels; ++j) {
2222 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2223 lp_build_vec_type(gallivm, fs_type), "");
2224 }
2225 if (dst_channels == 3 && !has_alpha) {
2226 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2227 lp_build_vec_type(gallivm, fs_type), "");
2228 }
2229 }
2230 }
2231
2232 /*
2233 * We actually should generally do conversion first (for non-1d cases)
2234 * when the blend format is 8 or 16 bits. The reason is obvious,
2235 * there's 2 or 4 times less vectors to deal with for the interleave...
2236 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2237 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2238 * unpack only with 128bit vectors).
2239 * Note: for 16bit sizes really need matching pack conversion code
2240 */
2241 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2242 twiddle_after_convert = TRUE;
2243 }
2244
2245 /*
2246 * Pixel twiddle from fragment shader order to memory order
2247 */
2248 if (!twiddle_after_convert) {
2249 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2250 dst_channels, fs_src, src, pad_inline);
2251 if (dual_source_blend) {
2252 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2253 fs_src1, src1, pad_inline);
2254 }
2255 } else {
2256 src_count = num_fullblock_fs * dst_channels;
2257 /*
2258 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2259 * (AVX) turn out the same later when untwiddling/transpose (albeit
2260 * for true AVX2 path untwiddle needs to be different).
2261 * For now just order by colors first (so we can use unpack later).
2262 */
2263 for (j = 0; j < num_fullblock_fs; j++) {
2264 for (i = 0; i < dst_channels; i++) {
2265 src[i*num_fullblock_fs + j] = fs_src[j][i];
2266 if (dual_source_blend) {
2267 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2268 }
2269 }
2270 }
2271 }
2272
2273 src_channels = dst_channels < 3 ? dst_channels : 4;
2274 if (src_count != num_fullblock_fs * src_channels) {
2275 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2276 row_type.length /= ds;
2277 fs_type.length = row_type.length;
2278 }
2279
2280 blend_type = row_type;
2281 mask_type.length = 4;
2282
2283 /* Convert src to row_type */
2284 if (dual_source_blend) {
2285 struct lp_type old_row_type = row_type;
2286 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2287 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2288 }
2289 else {
2290 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2291 }
2292
2293 /* If the rows are not an SSE vector, combine them to become SSE size! */
2294 if ((row_type.width * row_type.length) % 128) {
2295 unsigned bits = row_type.width * row_type.length;
2296 unsigned combined;
2297
2298 assert(src_count >= (vector_width / bits));
2299
2300 dst_count = src_count / (vector_width / bits);
2301
2302 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2303 if (dual_source_blend) {
2304 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2305 }
2306
2307 row_type.length *= combined;
2308 src_count /= combined;
2309
2310 bits = row_type.width * row_type.length;
2311 assert(bits == 128 || bits == 256);
2312 }
2313
2314 if (twiddle_after_convert) {
2315 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2316 if (dual_source_blend) {
2317 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2318 }
2319 }
2320
2321 /*
2322 * Blend Colour conversion
2323 */
2324 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2325 blend_color = LLVMBuildPointerCast(builder, blend_color,
2326 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2327 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2328 &i32_zero, 1, ""), "");
2329
2330 /* Convert */
2331 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2332
2333 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2334 /*
2335 * since blending is done with floats, there was no conversion.
2336 * However, the rules according to fixed point renderbuffers still
2337 * apply, that is we must clamp inputs to 0.0/1.0.
2338 * (This would apply to separate alpha conversion too but we currently
2339 * force has_alpha to be true.)
2340 * TODO: should skip this with "fake" blend, since post-blend conversion
2341 * will clamp anyway.
2342 * TODO: could also skip this if fragment color clamping is enabled. We
2343 * don't support it natively so it gets baked into the shader however, so
2344 * can't really tell here.
2345 */
2346 struct lp_build_context f32_bld;
2347 assert(row_type.floating);
2348 lp_build_context_init(&f32_bld, gallivm, row_type);
2349 for (i = 0; i < src_count; i++) {
2350 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2351 }
2352 if (dual_source_blend) {
2353 for (i = 0; i < src_count; i++) {
2354 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2355 }
2356 }
2357 /* probably can't be different than row_type but better safe than sorry... */
2358 lp_build_context_init(&f32_bld, gallivm, blend_type);
2359 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2360 }
2361
2362 /* Extract alpha */
2363 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2364
2365 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2366 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2367 if (pad_inline) {
2368 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2369 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2370 } else {
2371 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2372 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2373 }
2374
2375 /*
2376 * Mask conversion
2377 */
2378 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2379
2380 if (src_count < block_height) {
2381 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2382 } else if (src_count > block_height) {
2383 for (i = src_count; i > 0; --i) {
2384 unsigned pixels = block_size / src_count;
2385 unsigned idx = i - 1;
2386
2387 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2388 (idx * pixels) % 4, pixels);
2389 }
2390 }
2391
2392 assert(mask_type.width == 32);
2393
2394 for (i = 0; i < src_count; ++i) {
2395 unsigned pixels = block_size / src_count;
2396 unsigned pixel_width = row_type.width * dst_channels;
2397
2398 if (pixel_width == 24) {
2399 mask_type.width = 8;
2400 mask_type.length = vector_width / mask_type.width;
2401 } else {
2402 mask_type.length = pixels;
2403 mask_type.width = row_type.width * dst_channels;
2404
2405 /*
2406 * If mask_type width is smaller than 32bit, this doesn't quite
2407 * generate the most efficient code (could use some pack).
2408 */
2409 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2410 lp_build_int_vec_type(gallivm, mask_type), "");
2411
2412 mask_type.length *= dst_channels;
2413 mask_type.width /= dst_channels;
2414 }
2415
2416 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2417 lp_build_int_vec_type(gallivm, mask_type), "");
2418 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2419 }
2420
2421 /*
2422 * Alpha conversion
2423 */
2424 if (!has_alpha) {
2425 struct lp_type alpha_type = fs_type;
2426 alpha_type.length = 4;
2427 convert_alpha(gallivm, row_type, alpha_type,
2428 block_size, block_height,
2429 src_count, dst_channels,
2430 pad_inline, src_alpha);
2431 if (dual_source_blend) {
2432 convert_alpha(gallivm, row_type, alpha_type,
2433 block_size, block_height,
2434 src_count, dst_channels,
2435 pad_inline, src1_alpha);
2436 }
2437 }
2438
2439
2440 /*
2441 * Load dst from memory
2442 */
2443 if (src_count < block_height) {
2444 dst_count = block_height;
2445 } else {
2446 dst_count = src_count;
2447 }
2448
2449 dst_type.length *= block_size / dst_count;
2450
2451 if (format_expands_to_float_soa(out_format_desc)) {
2452 /*
2453 * we need multiple values at once for the conversion, so can as well
2454 * load them vectorized here too instead of concatenating later.
2455 * (Still need concatenation later for 8-wide vectors).
2456 */
2457 dst_count = block_height;
2458 dst_type.length = block_width;
2459 }
2460
2461 /*
2462 * Compute the alignment of the destination pointer in bytes
2463 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2464 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2465 * 1d tex but can't distinguish here) so need to stick with per-pixel
2466 * alignment in this case.
2467 */
2468 if (is_1d) {
2469 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2470 }
2471 else {
2472 dst_alignment = dst_type.length * dst_type.width / 8;
2473 }
2474 /* Force power-of-two alignment by extracting only the least-significant-bit */
2475 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2476 /*
2477 * Resource base and stride pointers are aligned to 16 bytes, so that's
2478 * the maximum alignment we can guarantee
2479 */
2480 dst_alignment = MIN2(16, dst_alignment);
2481
2482 ls_type = dst_type;
2483
2484 if (dst_count > src_count) {
2485 if ((dst_type.width == 8 || dst_type.width == 16) &&
2486 util_is_power_of_two_or_zero(dst_type.length) &&
2487 dst_type.length * dst_type.width < 128) {
2488 /*
2489 * Never try to load values as 4xi8 which we will then
2490 * concatenate to larger vectors. This gives llvm a real
2491 * headache (the problem is the type legalizer (?) will
2492 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2493 * then the shuffles to concatenate are more or less impossible
2494 * - llvm is easily capable of generating a sequence of 32
2495 * pextrb/pinsrb instructions for that. Albeit it appears to
2496 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2497 * width to avoid the trouble (16bit seems not as bad, llvm
2498 * probably recognizes the load+shuffle as only one shuffle
2499 * is necessary, but we can do just the same anyway).
2500 */
2501 ls_type.length = dst_type.length * dst_type.width / 32;
2502 ls_type.width = 32;
2503 }
2504 }
2505
2506 if (is_1d) {
2507 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2508 dst, ls_type, dst_count / 4, dst_alignment);
2509 for (i = dst_count / 4; i < dst_count; i++) {
2510 dst[i] = lp_build_undef(gallivm, ls_type);
2511 }
2512
2513 }
2514 else {
2515 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2516 dst, ls_type, dst_count, dst_alignment);
2517 }
2518
2519
2520 /*
2521 * Convert from dst/output format to src/blending format.
2522 *
2523 * This is necessary as we can only read 1 row from memory at a time,
2524 * so the minimum dst_count will ever be at this point is 4.
2525 *
2526 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2527 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2528 * on all 16 pixels in that single vector at once.
2529 */
2530 if (dst_count > src_count) {
2531 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2532 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2533 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2534 for (i = 0; i < dst_count; i++) {
2535 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2536 }
2537 }
2538
2539 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2540
2541 if (ls_type.length != dst_type.length) {
2542 struct lp_type tmp_type = dst_type;
2543 tmp_type.length = dst_type.length * 4 / src_count;
2544 for (i = 0; i < src_count; i++) {
2545 dst[i] = LLVMBuildBitCast(builder, dst[i],
2546 lp_build_vec_type(gallivm, tmp_type), "");
2547 }
2548 }
2549 }
2550
2551 /*
2552 * Blending
2553 */
2554 /* XXX this is broken for RGB8 formats -
2555 * they get expanded from 12 to 16 elements (to include alpha)
2556 * by convert_to_blend_type then reduced to 15 instead of 12
2557 * by convert_from_blend_type (a simple fix though breaks A8...).
2558 * R16G16B16 also crashes differently however something going wrong
2559 * inside llvm handling npot vector sizes seemingly.
2560 * It seems some cleanup could be done here (like skipping conversion/blend
2561 * when not needed).
2562 */
2563 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2564 row_type, dst, src_count);
2565
2566 /*
2567 * FIXME: Really should get logic ops / masks out of generic blend / row
2568 * format. Logic ops will definitely not work on the blend float format
2569 * used for SRGB here and I think OpenGL expects this to work as expected
2570 * (that is incoming values converted to srgb then logic op applied).
2571 */
2572 for (i = 0; i < src_count; ++i) {
2573 dst[i] = lp_build_blend_aos(gallivm,
2574 &variant->key.blend,
2575 out_format,
2576 row_type,
2577 rt,
2578 src[i],
2579 has_alpha ? NULL : src_alpha[i],
2580 src1[i],
2581 has_alpha ? NULL : src1_alpha[i],
2582 dst[i],
2583 partial_mask ? src_mask[i] : NULL,
2584 blend_color,
2585 has_alpha ? NULL : blend_alpha,
2586 swizzle,
2587 pad_inline ? 4 : dst_channels);
2588 }
2589
2590 convert_from_blend_type(gallivm, block_size, out_format_desc,
2591 row_type, dst_type, dst, src_count);
2592
2593 /* Split the blend rows back to memory rows */
2594 if (dst_count > src_count) {
2595 row_type.length = dst_type.length * (dst_count / src_count);
2596
2597 if (src_count == 1) {
2598 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2599 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2600
2601 row_type.length /= 2;
2602 src_count *= 2;
2603 }
2604
2605 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2606 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2607 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2608 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2609
2610 row_type.length /= 2;
2611 src_count *= 2;
2612 }
2613
2614 /*
2615 * Store blend result to memory
2616 */
2617 if (is_1d) {
2618 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2619 dst, dst_type, dst_count / 4, dst_alignment);
2620 }
2621 else {
2622 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2623 dst, dst_type, dst_count, dst_alignment);
2624 }
2625
2626 if (have_smallfloat_format(dst_type, out_format)) {
2627 lp_build_fpstate_set(gallivm, fpstate);
2628 }
2629
2630 if (do_branch) {
2631 lp_build_mask_end(&mask_ctx);
2632 }
2633 }
2634
2635
2636 /**
2637 * Generate the runtime callable function for the whole fragment pipeline.
2638 * Note that the function which we generate operates on a block of 16
2639 * pixels at at time. The block contains 2x2 quads. Each quad contains
2640 * 2x2 pixels.
2641 */
2642 static void
2643 generate_fragment(struct llvmpipe_context *lp,
2644 struct lp_fragment_shader *shader,
2645 struct lp_fragment_shader_variant *variant,
2646 unsigned partial_mask)
2647 {
2648 struct gallivm_state *gallivm = variant->gallivm;
2649 struct lp_fragment_shader_variant_key *key = &variant->key;
2650 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2651 char func_name[64];
2652 struct lp_type fs_type;
2653 struct lp_type blend_type;
2654 LLVMTypeRef fs_elem_type;
2655 LLVMTypeRef blend_vec_type;
2656 LLVMTypeRef arg_types[15];
2657 LLVMTypeRef func_type;
2658 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2659 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2660 LLVMValueRef context_ptr;
2661 LLVMValueRef x;
2662 LLVMValueRef y;
2663 LLVMValueRef a0_ptr;
2664 LLVMValueRef dadx_ptr;
2665 LLVMValueRef dady_ptr;
2666 LLVMValueRef color_ptr_ptr;
2667 LLVMValueRef stride_ptr;
2668 LLVMValueRef color_sample_stride_ptr;
2669 LLVMValueRef depth_ptr;
2670 LLVMValueRef depth_stride;
2671 LLVMValueRef depth_sample_stride;
2672 LLVMValueRef mask_input;
2673 LLVMValueRef thread_data_ptr;
2674 LLVMBasicBlockRef block;
2675 LLVMBuilderRef builder;
2676 struct lp_build_sampler_soa *sampler;
2677 struct lp_build_image_soa *image;
2678 struct lp_build_interp_soa_context interp;
2679 LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
2680 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2681 LLVMValueRef function;
2682 LLVMValueRef facing;
2683 unsigned num_fs;
2684 unsigned i;
2685 unsigned chan;
2686 unsigned cbuf;
2687 boolean cbuf0_write_all;
2688 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2689 util_blend_state_is_dual(&key->blend, 0);
2690
2691 assert(lp_native_vector_width / 32 >= 4);
2692
2693 /* Adjust color input interpolation according to flatshade state:
2694 */
2695 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2696 for (i = 0; i < shader->info.base.num_inputs; i++) {
2697 if (inputs[i].interp == LP_INTERP_COLOR) {
2698 if (key->flatshade)
2699 inputs[i].interp = LP_INTERP_CONSTANT;
2700 else
2701 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2702 }
2703 }
2704
2705 /* check if writes to cbuf[0] are to be copied to all cbufs */
2706 cbuf0_write_all =
2707 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2708
2709 /* TODO: actually pick these based on the fs and color buffer
2710 * characteristics. */
2711
2712 memset(&fs_type, 0, sizeof fs_type);
2713 fs_type.floating = TRUE; /* floating point values */
2714 fs_type.sign = TRUE; /* values are signed */
2715 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2716 fs_type.width = 32; /* 32-bit float */
2717 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2718
2719 memset(&blend_type, 0, sizeof blend_type);
2720 blend_type.floating = FALSE; /* values are integers */
2721 blend_type.sign = FALSE; /* values are unsigned */
2722 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2723 blend_type.width = 8; /* 8-bit ubyte values */
2724 blend_type.length = 16; /* 16 elements per vector */
2725
2726 /*
2727 * Generate the function prototype. Any change here must be reflected in
2728 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2729 */
2730
2731 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2732
2733 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2734
2735 snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2736 shader->no, variant->no, partial_mask ? "partial" : "whole");
2737
2738 arg_types[0] = variant->jit_context_ptr_type; /* context */
2739 arg_types[1] = int32_type; /* x */
2740 arg_types[2] = int32_type; /* y */
2741 arg_types[3] = int32_type; /* facing */
2742 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2743 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2744 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2745 arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
2746 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2747 arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
2748 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2749 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2750 arg_types[12] = int32_type; /* depth_stride */
2751 arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
2752 arg_types[14] = int32_type; /* depth sample stride */
2753
2754 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2755 arg_types, ARRAY_SIZE(arg_types), 0);
2756
2757 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2758 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2759
2760 variant->function[partial_mask] = function;
2761
2762 /* XXX: need to propagate noalias down into color param now we are
2763 * passing a pointer-to-pointer?
2764 */
2765 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2766 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2767 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2768
2769 context_ptr = LLVMGetParam(function, 0);
2770 x = LLVMGetParam(function, 1);
2771 y = LLVMGetParam(function, 2);
2772 facing = LLVMGetParam(function, 3);
2773 a0_ptr = LLVMGetParam(function, 4);
2774 dadx_ptr = LLVMGetParam(function, 5);
2775 dady_ptr = LLVMGetParam(function, 6);
2776 color_ptr_ptr = LLVMGetParam(function, 7);
2777 depth_ptr = LLVMGetParam(function, 8);
2778 mask_input = LLVMGetParam(function, 9);
2779 thread_data_ptr = LLVMGetParam(function, 10);
2780 stride_ptr = LLVMGetParam(function, 11);
2781 depth_stride = LLVMGetParam(function, 12);
2782 color_sample_stride_ptr = LLVMGetParam(function, 13);
2783 depth_sample_stride = LLVMGetParam(function, 14);
2784
2785 lp_build_name(context_ptr, "context");
2786 lp_build_name(x, "x");
2787 lp_build_name(y, "y");
2788 lp_build_name(a0_ptr, "a0");
2789 lp_build_name(dadx_ptr, "dadx");
2790 lp_build_name(dady_ptr, "dady");
2791 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2792 lp_build_name(depth_ptr, "depth");
2793 lp_build_name(mask_input, "mask_input");
2794 lp_build_name(thread_data_ptr, "thread_data");
2795 lp_build_name(stride_ptr, "stride_ptr");
2796 lp_build_name(depth_stride, "depth_stride");
2797 lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
2798 lp_build_name(depth_sample_stride, "depth_sample_stride");
2799
2800 /*
2801 * Function body
2802 */
2803
2804 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2805 builder = gallivm->builder;
2806 assert(builder);
2807 LLVMPositionBuilderAtEnd(builder, block);
2808
2809 /*
2810 * Must not count ps invocations if there's a null shader.
2811 * (It would be ok to count with null shader if there's d/s tests,
2812 * but only if there's d/s buffers too, which is different
2813 * to implicit rasterization disable which must not depend
2814 * on the d/s buffers.)
2815 * Could use popcount on mask, but pixel accuracy is not required.
2816 * Could disable if there's no stats query, but maybe not worth it.
2817 */
2818 if (shader->info.base.num_instructions > 1) {
2819 LLVMValueRef invocs, val;
2820 invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
2821 val = LLVMBuildLoad(builder, invocs, "");
2822 val = LLVMBuildAdd(builder, val,
2823 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
2824 "invoc_count");
2825 LLVMBuildStore(builder, val, invocs);
2826 }
2827
2828 /* code generated texture sampling */
2829 sampler = lp_llvm_sampler_soa_create(key->samplers);
2830 image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
2831
2832 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2833 /* for 1d resources only run "upper half" of stamp */
2834 if (key->resource_1d)
2835 num_fs /= 2;
2836
2837 {
2838 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2839 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2840 LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
2841 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2842 num_loop_samp, "mask_store");
2843
2844 LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
2845 LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, flt_type, "");
2846 LLVMValueRef sample_pos_array;
2847
2848 if (key->multisample && key->coverage_samples == 4) {
2849 LLVMValueRef sample_pos_arr[8];
2850 for (unsigned i = 0; i < 4; i++) {
2851 sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
2852 sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
2853 }
2854 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
2855 } else {
2856 LLVMValueRef sample_pos_arr[2];
2857 sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
2858 sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
2859 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
2860 }
2861 LLVMSetInitializer(glob_sample_pos, sample_pos_array);
2862
2863 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2864 boolean pixel_center_integer =
2865 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2866
2867 /*
2868 * The shader input interpolation info is not explicitely baked in the
2869 * shader key, but everything it derives from (TGSI, and flatshade) is
2870 * already included in the shader key.
2871 */
2872 lp_build_interp_soa_init(&interp,
2873 gallivm,
2874 shader->info.base.num_inputs,
2875 inputs,
2876 pixel_center_integer,
2877 key->coverage_samples, glob_sample_pos,
2878 num_loop,
2879 key->depth_clamp,
2880 builder, fs_type,
2881 a0_ptr, dadx_ptr, dady_ptr,
2882 x, y);
2883
2884 for (i = 0; i < num_fs; i++) {
2885 if (key->multisample) {
2886 LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
2887
2888 /*
2889 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2890 * store to the per sample mask storage. Or all of them together to generate
2891 * the fragment shader mask. (sample shading TODO).
2892 * Take the incoming state coverage mask into account.
2893 */
2894 for (unsigned s = 0; s < key->coverage_samples; s++) {
2895 LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
2896 LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
2897 &sindexi, 1, "sample_mask_ptr");
2898 LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
2899 i*fs_type.length/4, s, mask_input);
2900
2901 LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
2902 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
2903 smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
2904 smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
2905
2906 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
2907 LLVMBuildStore(builder, s_mask, sample_mask_ptr);
2908 }
2909 } else {
2910 LLVMValueRef mask;
2911 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2912 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2913 &indexi, 1, "mask_ptr");
2914
2915 if (partial_mask) {
2916 mask = generate_quad_mask(gallivm, fs_type,
2917 i*fs_type.length/4, 0, mask_input);
2918 }
2919 else {
2920 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2921 }
2922 LLVMBuildStore(builder, mask, mask_ptr);
2923 }
2924 }
2925
2926 generate_fs_loop(gallivm,
2927 shader, key,
2928 builder,
2929 fs_type,
2930 context_ptr,
2931 glob_sample_pos,
2932 num_loop,
2933 &interp,
2934 sampler,
2935 image,
2936 mask_store, /* output */
2937 color_store,
2938 depth_ptr,
2939 depth_stride,
2940 depth_sample_stride,
2941 facing,
2942 thread_data_ptr);
2943
2944 for (i = 0; i < num_fs; i++) {
2945 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2946 LLVMValueRef ptr;
2947 for (unsigned s = 0; s < key->coverage_samples; s++) {
2948 int idx = (i + (s * num_fs));
2949 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
2950 ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
2951
2952 fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
2953 }
2954
2955 /* This is fucked up need to reorganize things */
2956 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2957 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2958 ptr = LLVMBuildGEP(builder,
2959 color_store[cbuf * !cbuf0_write_all][chan],
2960 &indexi, 1, "");
2961 fs_out_color[cbuf][chan][i] = ptr;
2962 }
2963 }
2964 if (dual_source_blend) {
2965 /* only support one dual source blend target hence always use output 1 */
2966 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2967 ptr = LLVMBuildGEP(builder,
2968 color_store[1][chan],
2969 &indexi, 1, "");
2970 fs_out_color[1][chan][i] = ptr;
2971 }
2972 }
2973 }
2974 }
2975
2976 sampler->destroy(sampler);
2977 image->destroy(image);
2978 /* Loop over color outputs / color buffers to do blending.
2979 */
2980 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2981 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2982 LLVMValueRef color_ptr;
2983 LLVMValueRef stride;
2984 LLVMValueRef sample_stride = NULL;
2985 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2986
2987 boolean do_branch = ((key->depth.enabled
2988 || key->stencil[0].enabled
2989 || key->alpha.enabled)
2990 && !shader->info.base.uses_kill);
2991
2992 color_ptr = LLVMBuildLoad(builder,
2993 LLVMBuildGEP(builder, color_ptr_ptr,
2994 &index, 1, ""),
2995 "");
2996
2997 stride = LLVMBuildLoad(builder,
2998 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2999 "");
3000
3001 if (key->multisample)
3002 sample_stride = LLVMBuildLoad(builder,
3003 LLVMBuildGEP(builder, color_sample_stride_ptr,
3004 &index, 1, ""), "");
3005
3006 for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3007 unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3008 LLVMValueRef out_ptr = color_ptr;;
3009
3010 if (key->multisample) {
3011 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3012 out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3013 }
3014 out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3015
3016 lp_build_name(out_ptr, "color_ptr%d", cbuf);
3017
3018 generate_unswizzled_blend(gallivm, cbuf, variant,
3019 key->cbuf_format[cbuf],
3020 num_fs, fs_type, &fs_mask[mask_idx], fs_out_color,
3021 context_ptr, out_ptr, stride,
3022 partial_mask, do_branch);
3023 }
3024 }
3025 }
3026
3027 LLVMBuildRetVoid(builder);
3028
3029 gallivm_verify_function(gallivm, function);
3030 }
3031
3032
3033 static void
3034 dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3035 {
3036 unsigned i;
3037
3038 debug_printf("fs variant %p:\n", (void *) key);
3039
3040 if (key->flatshade) {
3041 debug_printf("flatshade = 1\n");
3042 }
3043 if (key->multisample) {
3044 debug_printf("multisample = 1\n");
3045 debug_printf("coverage samples = %d\n", key->coverage_samples);
3046 }
3047 for (i = 0; i < key->nr_cbufs; ++i) {
3048 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3049 debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3050 }
3051 if (key->depth.enabled || key->stencil[0].enabled) {
3052 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3053 debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3054 }
3055 if (key->depth.enabled) {
3056 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3057 debug_printf("depth.writemask = %u\n", key->depth.writemask);
3058 }
3059
3060 for (i = 0; i < 2; ++i) {
3061 if (key->stencil[i].enabled) {
3062 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3063 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3064 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3065 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3066 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3067 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3068 }
3069 }
3070
3071 if (key->alpha.enabled) {
3072 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3073 }
3074
3075 if (key->occlusion_count) {
3076 debug_printf("occlusion_count = 1\n");
3077 }
3078
3079 if (key->blend.logicop_enable) {
3080 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3081 }
3082 else if (key->blend.rt[0].blend_enable) {
3083 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
3084 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3085 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3086 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
3087 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3088 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3089 }
3090 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3091 if (key->blend.alpha_to_coverage) {
3092 debug_printf("blend.alpha_to_coverage is enabled\n");
3093 }
3094 for (i = 0; i < key->nr_samplers; ++i) {
3095 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
3096 debug_printf("sampler[%u] = \n", i);
3097 debug_printf(" .wrap = %s %s %s\n",
3098 util_str_tex_wrap(sampler->wrap_s, TRUE),
3099 util_str_tex_wrap(sampler->wrap_t, TRUE),
3100 util_str_tex_wrap(sampler->wrap_r, TRUE));
3101 debug_printf(" .min_img_filter = %s\n",
3102 util_str_tex_filter(sampler->min_img_filter, TRUE));
3103 debug_printf(" .min_mip_filter = %s\n",
3104 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3105 debug_printf(" .mag_img_filter = %s\n",
3106 util_str_tex_filter(sampler->mag_img_filter, TRUE));
3107 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3108 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3109 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
3110 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3111 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3112 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
3113 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
3114 }
3115 for (i = 0; i < key->nr_sampler_views; ++i) {
3116 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
3117 debug_printf("texture[%u] = \n", i);
3118 debug_printf(" .format = %s\n",
3119 util_format_name(texture->format));
3120 debug_printf(" .target = %s\n",
3121 util_str_tex_target(texture->target, TRUE));
3122 debug_printf(" .level_zero_only = %u\n",
3123 texture->level_zero_only);
3124 debug_printf(" .pot = %u %u %u\n",
3125 texture->pot_width,
3126 texture->pot_height,
3127 texture->pot_depth);
3128 }
3129 struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3130 for (i = 0; i < key->nr_images; ++i) {
3131 const struct lp_static_texture_state *image = &images[i].image_state;
3132 debug_printf("image[%u] = \n", i);
3133 debug_printf(" .format = %s\n",
3134 util_format_name(image->format));
3135 debug_printf(" .target = %s\n",
3136 util_str_tex_target(image->target, TRUE));
3137 debug_printf(" .level_zero_only = %u\n",
3138 image->level_zero_only);
3139 debug_printf(" .pot = %u %u %u\n",
3140 image->pot_width,
3141 image->pot_height,
3142 image->pot_depth);
3143 }
3144 }
3145
3146
3147 void
3148 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3149 {
3150 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3151 variant->shader->no, variant->no);
3152 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3153 tgsi_dump(variant->shader->base.tokens, 0);
3154 else
3155 nir_print_shader(variant->shader->base.ir.nir, stderr);
3156 dump_fs_variant_key(&variant->key);
3157 debug_printf("variant->opaque = %u\n", variant->opaque);
3158 debug_printf("\n");
3159 }
3160
3161
3162 /**
3163 * Generate a new fragment shader variant from the shader code and
3164 * other state indicated by the key.
3165 */
3166 static struct lp_fragment_shader_variant *
3167 generate_variant(struct llvmpipe_context *lp,
3168 struct lp_fragment_shader *shader,
3169 const struct lp_fragment_shader_variant_key *key)
3170 {
3171 struct lp_fragment_shader_variant *variant;
3172 const struct util_format_description *cbuf0_format_desc = NULL;
3173 boolean fullcolormask;
3174 char module_name[64];
3175
3176 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3177 if (!variant)
3178 return NULL;
3179
3180 memset(variant, 0, sizeof(*variant));
3181 snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3182 shader->no, shader->variants_created);
3183
3184 variant->gallivm = gallivm_create(module_name, lp->context);
3185 if (!variant->gallivm) {
3186 FREE(variant);
3187 return NULL;
3188 }
3189
3190 variant->shader = shader;
3191 variant->list_item_global.base = variant;
3192 variant->list_item_local.base = variant;
3193 variant->no = shader->variants_created++;
3194
3195 memcpy(&variant->key, key, shader->variant_key_size);
3196
3197 /*
3198 * Determine whether we are touching all channels in the color buffer.
3199 */
3200 fullcolormask = FALSE;
3201 if (key->nr_cbufs == 1) {
3202 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3203 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3204 }
3205
3206 variant->opaque =
3207 !key->blend.logicop_enable &&
3208 !key->blend.rt[0].blend_enable &&
3209 fullcolormask &&
3210 !key->stencil[0].enabled &&
3211 !key->alpha.enabled &&
3212 !key->blend.alpha_to_coverage &&
3213 !key->depth.enabled &&
3214 !shader->info.base.uses_kill &&
3215 !shader->info.base.writes_samplemask
3216 ? TRUE : FALSE;
3217
3218 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3219 lp_debug_fs_variant(variant);
3220 }
3221
3222 lp_jit_init_types(variant);
3223
3224 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
3225 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
3226
3227 if (variant->jit_function[RAST_WHOLE] == NULL) {
3228 if (variant->opaque) {
3229 /* Specialized shader, which doesn't need to read the color buffer. */
3230 generate_fragment(lp, shader, variant, RAST_WHOLE);
3231 }
3232 }
3233
3234 /*
3235 * Compile everything
3236 */
3237
3238 gallivm_compile_module(variant->gallivm);
3239
3240 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
3241
3242 if (variant->function[RAST_EDGE_TEST]) {
3243 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
3244 gallivm_jit_function(variant->gallivm,
3245 variant->function[RAST_EDGE_TEST]);
3246 }
3247
3248 if (variant->function[RAST_WHOLE]) {
3249 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
3250 gallivm_jit_function(variant->gallivm,
3251 variant->function[RAST_WHOLE]);
3252 } else if (!variant->jit_function[RAST_WHOLE]) {
3253 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
3254 }
3255
3256 gallivm_free_ir(variant->gallivm);
3257
3258 return variant;
3259 }
3260
3261
3262 static void *
3263 llvmpipe_create_fs_state(struct pipe_context *pipe,
3264 const struct pipe_shader_state *templ)
3265 {
3266 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3267 struct lp_fragment_shader *shader;
3268 int nr_samplers;
3269 int nr_sampler_views;
3270 int nr_images;
3271 int i;
3272
3273 shader = CALLOC_STRUCT(lp_fragment_shader);
3274 if (!shader)
3275 return NULL;
3276
3277 shader->no = fs_no++;
3278 make_empty_list(&shader->variants);
3279
3280 shader->base.type = templ->type;
3281 if (templ->type == PIPE_SHADER_IR_TGSI) {
3282 /* get/save the summary info for this shader */
3283 lp_build_tgsi_info(templ->tokens, &shader->info);
3284
3285 /* we need to keep a local copy of the tokens */
3286 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
3287 } else {
3288 shader->base.ir.nir = templ->ir.nir;
3289 nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
3290 }
3291
3292 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
3293 if (shader->draw_data == NULL) {
3294 FREE((void *) shader->base.tokens);
3295 FREE(shader);
3296 return NULL;
3297 }
3298
3299 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3300 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3301 nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3302 shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
3303
3304 for (i = 0; i < shader->info.base.num_inputs; i++) {
3305 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
3306 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
3307 shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
3308
3309 switch (shader->info.base.input_interpolate[i]) {
3310 case TGSI_INTERPOLATE_CONSTANT:
3311 shader->inputs[i].interp = LP_INTERP_CONSTANT;
3312 break;
3313 case TGSI_INTERPOLATE_LINEAR:
3314 shader->inputs[i].interp = LP_INTERP_LINEAR;
3315 break;
3316 case TGSI_INTERPOLATE_PERSPECTIVE:
3317 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
3318 break;
3319 case TGSI_INTERPOLATE_COLOR:
3320 shader->inputs[i].interp = LP_INTERP_COLOR;
3321 break;
3322 default:
3323 assert(0);
3324 break;
3325 }
3326
3327 switch (shader->info.base.input_semantic_name[i]) {
3328 case TGSI_SEMANTIC_FACE:
3329 shader->inputs[i].interp = LP_INTERP_FACING;
3330 break;
3331 case TGSI_SEMANTIC_POSITION:
3332 /* Position was already emitted above
3333 */
3334 shader->inputs[i].interp = LP_INTERP_POSITION;
3335 shader->inputs[i].src_index = 0;
3336 continue;
3337 }
3338
3339 /* XXX this is a completely pointless index map... */
3340 shader->inputs[i].src_index = i+1;
3341 }
3342
3343 if (LP_DEBUG & DEBUG_TGSI) {
3344 unsigned attrib;
3345 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3346 shader->no, (void *) shader);
3347 tgsi_dump(templ->tokens, 0);
3348 debug_printf("usage masks:\n");
3349 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
3350 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
3351 debug_printf(" IN[%u].%s%s%s%s\n",
3352 attrib,
3353 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
3354 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
3355 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
3356 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
3357 }
3358 debug_printf("\n");
3359 }
3360
3361 return shader;
3362 }
3363
3364
3365 static void
3366 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
3367 {
3368 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3369 struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
3370 if (llvmpipe->fs == lp_fs)
3371 return;
3372
3373 draw_bind_fragment_shader(llvmpipe->draw,
3374 (lp_fs ? lp_fs->draw_data : NULL));
3375
3376 llvmpipe->fs = lp_fs;
3377
3378 llvmpipe->dirty |= LP_NEW_FS;
3379 }
3380
3381
3382 /**
3383 * Remove shader variant from two lists: the shader's variant list
3384 * and the context's variant list.
3385 */
3386 static void
3387 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3388 struct lp_fragment_shader_variant *variant)
3389 {
3390 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3391 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3392 "v total cached %u inst %u total inst %u\n",
3393 variant->shader->no, variant->no,
3394 variant->shader->variants_created,
3395 variant->shader->variants_cached,
3396 lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3397 }
3398
3399 gallivm_destroy(variant->gallivm);
3400
3401 /* remove from shader's list */
3402 remove_from_list(&variant->list_item_local);
3403 variant->shader->variants_cached--;
3404
3405 /* remove from context's list */
3406 remove_from_list(&variant->list_item_global);
3407 lp->nr_fs_variants--;
3408 lp->nr_fs_instrs -= variant->nr_instrs;
3409
3410 FREE(variant);
3411 }
3412
3413
3414 static void
3415 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3416 {
3417 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3418 struct lp_fragment_shader *shader = fs;
3419 struct lp_fs_variant_list_item *li;
3420
3421 assert(fs != llvmpipe->fs);
3422
3423 /*
3424 * XXX: we need to flush the context until we have some sort of reference
3425 * counting in fragment shaders as they may still be binned
3426 * Flushing alone might not sufficient we need to wait on it too.
3427 */
3428 llvmpipe_finish(pipe, __FUNCTION__);
3429
3430 /* Delete all the variants */
3431 li = first_elem(&shader->variants);
3432 while(!at_end(&shader->variants, li)) {
3433 struct lp_fs_variant_list_item *next = next_elem(li);
3434 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3435 li = next;
3436 }
3437
3438 /* Delete draw module's data */
3439 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3440
3441 if (shader->base.ir.nir)
3442 ralloc_free(shader->base.ir.nir);
3443 assert(shader->variants_cached == 0);
3444 FREE((void *) shader->base.tokens);
3445 FREE(shader);
3446 }
3447
3448
3449
3450 static void
3451 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3452 enum pipe_shader_type shader, uint index,
3453 const struct pipe_constant_buffer *cb)
3454 {
3455 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3456 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3457
3458 assert(shader < PIPE_SHADER_TYPES);
3459 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3460
3461 /* note: reference counting */
3462 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3463
3464 if (constants) {
3465 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3466 debug_printf("Illegal set constant without bind flag\n");
3467 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3468 }
3469 }
3470
3471 if (shader == PIPE_SHADER_VERTEX ||
3472 shader == PIPE_SHADER_GEOMETRY ||
3473 shader == PIPE_SHADER_TESS_CTRL ||
3474 shader == PIPE_SHADER_TESS_EVAL) {
3475 /* Pass the constants to the 'draw' module */
3476 const unsigned size = cb ? cb->buffer_size : 0;
3477 const ubyte *data;
3478
3479 if (constants) {
3480 data = (ubyte *) llvmpipe_resource_data(constants);
3481 }
3482 else if (cb && cb->user_buffer) {
3483 data = (ubyte *) cb->user_buffer;
3484 }
3485 else {
3486 data = NULL;
3487 }
3488
3489 if (data)
3490 data += cb->buffer_offset;
3491
3492 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3493 index, data, size);
3494 }
3495 else if (shader == PIPE_SHADER_COMPUTE)
3496 llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
3497 else
3498 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3499
3500 if (cb && cb->user_buffer) {
3501 pipe_resource_reference(&constants, NULL);
3502 }
3503 }
3504
3505 static void
3506 llvmpipe_set_shader_buffers(struct pipe_context *pipe,
3507 enum pipe_shader_type shader, unsigned start_slot,
3508 unsigned count, const struct pipe_shader_buffer *buffers,
3509 unsigned writable_bitmask)
3510 {
3511 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3512 unsigned i, idx;
3513 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3514 const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
3515
3516 util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
3517
3518 if (shader == PIPE_SHADER_VERTEX ||
3519 shader == PIPE_SHADER_GEOMETRY ||
3520 shader == PIPE_SHADER_TESS_CTRL ||
3521 shader == PIPE_SHADER_TESS_EVAL) {
3522 const unsigned size = buffer ? buffer->buffer_size : 0;
3523 const ubyte *data = NULL;
3524 if (buffer && buffer->buffer)
3525 data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
3526 if (data)
3527 data += buffer->buffer_offset;
3528 draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
3529 i, data, size);
3530 } else if (shader == PIPE_SHADER_COMPUTE) {
3531 llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
3532 } else if (shader == PIPE_SHADER_FRAGMENT) {
3533 llvmpipe->dirty |= LP_NEW_FS_SSBOS;
3534 }
3535 }
3536 }
3537
3538 static void
3539 llvmpipe_set_shader_images(struct pipe_context *pipe,
3540 enum pipe_shader_type shader, unsigned start_slot,
3541 unsigned count, const struct pipe_image_view *images)
3542 {
3543 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3544 unsigned i, idx;
3545
3546 draw_flush(llvmpipe->draw);
3547 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3548 const struct pipe_image_view *image = images ? &images[idx] : NULL;
3549
3550 util_copy_image_view(&llvmpipe->images[shader][i], image);
3551 }
3552
3553 llvmpipe->num_images[shader] = start_slot + count;
3554 if (shader == PIPE_SHADER_VERTEX ||
3555 shader == PIPE_SHADER_GEOMETRY ||
3556 shader == PIPE_SHADER_TESS_CTRL ||
3557 shader == PIPE_SHADER_TESS_EVAL) {
3558 draw_set_images(llvmpipe->draw,
3559 shader,
3560 llvmpipe->images[shader],
3561 start_slot + count);
3562 } else if (shader == PIPE_SHADER_COMPUTE)
3563 llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
3564 else
3565 llvmpipe->dirty |= LP_NEW_FS_IMAGES;
3566 }
3567
3568 /**
3569 * Return the blend factor equivalent to a destination alpha of one.
3570 */
3571 static inline unsigned
3572 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3573 {
3574 switch(factor) {
3575 case PIPE_BLENDFACTOR_DST_ALPHA:
3576 return PIPE_BLENDFACTOR_ONE;
3577 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3578 return PIPE_BLENDFACTOR_ZERO;
3579 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3580 if (clamped_zero)
3581 return PIPE_BLENDFACTOR_ZERO;
3582 else
3583 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3584 }
3585
3586 return factor;
3587 }
3588
3589
3590 /**
3591 * We need to generate several variants of the fragment pipeline to match
3592 * all the combinations of the contributing state atoms.
3593 *
3594 * TODO: there is actually no reason to tie this to context state -- the
3595 * generated code could be cached globally in the screen.
3596 */
3597 static struct lp_fragment_shader_variant_key *
3598 make_variant_key(struct llvmpipe_context *lp,
3599 struct lp_fragment_shader *shader,
3600 char *store)
3601 {
3602 unsigned i;
3603 struct lp_fragment_shader_variant_key *key;
3604
3605 key = (struct lp_fragment_shader_variant_key *)store;
3606
3607 memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
3608
3609 if (lp->framebuffer.zsbuf) {
3610 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3611 const struct util_format_description *zsbuf_desc =
3612 util_format_description(zsbuf_format);
3613
3614 if (lp->depth_stencil->depth.enabled &&
3615 util_format_has_depth(zsbuf_desc)) {
3616 key->zsbuf_format = zsbuf_format;
3617 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3618 }
3619 if (lp->depth_stencil->stencil[0].enabled &&
3620 util_format_has_stencil(zsbuf_desc)) {
3621 key->zsbuf_format = zsbuf_format;
3622 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3623 }
3624 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3625 key->resource_1d = TRUE;
3626 }
3627 key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
3628 }
3629
3630 /*
3631 * Propagate the depth clamp setting from the rasterizer state.
3632 * depth_clip == 0 implies depth clamping is enabled.
3633 *
3634 * When clip_halfz is enabled, then always clamp the depth values.
3635 *
3636 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3637 * clamp is always active in d3d10, regardless if depth clip is
3638 * enabled or not).
3639 * (GL has an always-on [0,1] clamp on fs depth output instead
3640 * to ensure the depth values stay in range. Doesn't look like
3641 * we do that, though...)
3642 */
3643 if (lp->rasterizer->clip_halfz) {
3644 key->depth_clamp = 1;
3645 } else {
3646 key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
3647 }
3648
3649 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3650 if (!lp->framebuffer.nr_cbufs ||
3651 !lp->framebuffer.cbufs[0] ||
3652 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3653 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3654 }
3655 if(key->alpha.enabled)
3656 key->alpha.func = lp->depth_stencil->alpha.func;
3657 /* alpha.ref_value is passed in jit_context */
3658
3659 key->flatshade = lp->rasterizer->flatshade;
3660 key->multisample = lp->rasterizer->multisample;
3661 if (lp->active_occlusion_queries && !lp->queries_disabled) {
3662 key->occlusion_count = TRUE;
3663 }
3664
3665 if (lp->framebuffer.nr_cbufs) {
3666 memcpy(&key->blend, lp->blend, sizeof key->blend);
3667 }
3668
3669 key->coverage_samples = 1;
3670 if (key->multisample)
3671 key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
3672 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3673
3674 if (!key->blend.independent_blend_enable) {
3675 /* we always need independent blend otherwise the fixups below won't work */
3676 for (i = 1; i < key->nr_cbufs; i++) {
3677 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3678 }
3679 key->blend.independent_blend_enable = 1;
3680 }
3681
3682 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3683 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3684
3685 if (lp->framebuffer.cbufs[i]) {
3686 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3687 const struct util_format_description *format_desc;
3688
3689 key->cbuf_format[i] = format;
3690 key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
3691
3692 /*
3693 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3694 * mixing of 2d textures with height 1 and 1d textures, so make sure
3695 * we pick 1d if any cbuf or zsbuf is 1d.
3696 */
3697 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3698 key->resource_1d = TRUE;
3699 }
3700
3701 format_desc = util_format_description(format);
3702 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3703 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3704
3705 /*
3706 * Mask out color channels not present in the color buffer.
3707 */
3708 blend_rt->colormask &= util_format_colormask(format_desc);
3709
3710 /*
3711 * Disable blend for integer formats.
3712 */
3713 if (util_format_is_pure_integer(format)) {
3714 blend_rt->blend_enable = 0;
3715 }
3716
3717 /*
3718 * Our swizzled render tiles always have an alpha channel, but the
3719 * linear render target format often does not, so force here the dst
3720 * alpha to be one.
3721 *
3722 * This is not a mere optimization. Wrong results will be produced if
3723 * the dst alpha is used, the dst format does not have alpha, and the
3724 * previous rendering was not flushed from the swizzled to linear
3725 * buffer. For example, NonPowTwo DCT.
3726 *
3727 * TODO: This should be generalized to all channels for better
3728 * performance, but only alpha causes correctness issues.
3729 *
3730 * Also, force rgb/alpha func/factors match, to make AoS blending
3731 * easier.
3732 */
3733 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3734 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3735 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3736 boolean clamped_zero = !util_format_is_float(format) &&
3737 !util_format_is_snorm(format);
3738 blend_rt->rgb_src_factor =
3739 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3740 blend_rt->rgb_dst_factor =
3741 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3742 blend_rt->alpha_func = blend_rt->rgb_func;
3743 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3744 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3745 }
3746 }
3747 else {
3748 /* no color buffer for this fragment output */
3749 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3750 key->cbuf_nr_samples[i] = 0;
3751 blend_rt->colormask = 0x0;
3752 blend_rt->blend_enable = 0;
3753 }
3754 }
3755
3756 /* This value will be the same for all the variants of a given shader:
3757 */
3758 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3759
3760 struct lp_sampler_static_state *fs_sampler;
3761
3762 fs_sampler = key->samplers;
3763
3764 memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
3765
3766 for(i = 0; i < key->nr_samplers; ++i) {
3767 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3768 lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
3769 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3770 }
3771 }
3772
3773 /*
3774 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3775 * are dx10-style? Can't really have mixed opcodes, at least not
3776 * if we want to skip the holes here (without rescanning tgsi).
3777 */
3778 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3779 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3780 for(i = 0; i < key->nr_sampler_views; ++i) {
3781 /*
3782 * Note sview may exceed what's representable by file_mask.
3783 * This will still work, the only downside is that not actually
3784 * used views may be included in the shader key.
3785 */
3786 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
3787 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3788 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3789 }
3790 }
3791 }
3792 else {
3793 key->nr_sampler_views = key->nr_samplers;
3794 for(i = 0; i < key->nr_sampler_views; ++i) {
3795 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3796 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3797 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3798 }
3799 }
3800 }
3801
3802 struct lp_image_static_state *lp_image;
3803 lp_image = lp_fs_variant_key_images(key);
3804 key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3805 for (i = 0; i < key->nr_images; ++i) {
3806 if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
3807 lp_sampler_static_texture_state_image(&lp_image[i].image_state,
3808 &lp->images[PIPE_SHADER_FRAGMENT][i]);
3809 }
3810 }
3811 return key;
3812 }
3813
3814
3815
3816 /**
3817 * Update fragment shader state. This is called just prior to drawing
3818 * something when some fragment-related state has changed.
3819 */
3820 void
3821 llvmpipe_update_fs(struct llvmpipe_context *lp)
3822 {
3823 struct lp_fragment_shader *shader = lp->fs;
3824 struct lp_fragment_shader_variant_key *key;
3825 struct lp_fragment_shader_variant *variant = NULL;
3826 struct lp_fs_variant_list_item *li;
3827 char store[LP_FS_MAX_VARIANT_KEY_SIZE];
3828
3829 key = make_variant_key(lp, shader, store);
3830
3831 /* Search the variants for one which matches the key */
3832 li = first_elem(&shader->variants);
3833 while(!at_end(&shader->variants, li)) {
3834 if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
3835 variant = li->base;
3836 break;
3837 }
3838 li = next_elem(li);
3839 }
3840
3841 if (variant) {
3842 /* Move this variant to the head of the list to implement LRU
3843 * deletion of shader's when we have too many.
3844 */
3845 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3846 }
3847 else {
3848 /* variant not found, create it now */
3849 int64_t t0, t1, dt;
3850 unsigned i;
3851 unsigned variants_to_cull;
3852
3853 if (LP_DEBUG & DEBUG_FS) {
3854 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3855 lp->nr_fs_variants,
3856 lp->nr_fs_instrs,
3857 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3858 }
3859
3860 /* First, check if we've exceeded the max number of shader variants.
3861 * If so, free 6.25% of them (the least recently used ones).
3862 */
3863 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
3864
3865 if (variants_to_cull ||
3866 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3867 struct pipe_context *pipe = &lp->pipe;
3868
3869 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
3870 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3871 "\t%u instrs,\t%u instrs/variant\n",
3872 shader->variants_cached,
3873 lp->nr_fs_variants, lp->nr_fs_instrs,
3874 lp->nr_fs_instrs / lp->nr_fs_variants);
3875 }
3876
3877 /*
3878 * XXX: we need to flush the context until we have some sort of
3879 * reference counting in fragment shaders as they may still be binned
3880 * Flushing alone might not be sufficient we need to wait on it too.
3881 */
3882 llvmpipe_finish(pipe, __FUNCTION__);
3883
3884 /*
3885 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3886 * number of shader variants (potentially all of them) could be
3887 * pending for destruction on flush.
3888 */
3889
3890 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3891 struct lp_fs_variant_list_item *item;
3892 if (is_empty_list(&lp->fs_variants_list)) {
3893 break;
3894 }
3895 item = last_elem(&lp->fs_variants_list);
3896 assert(item);
3897 assert(item->base);
3898 llvmpipe_remove_shader_variant(lp, item->base);
3899 }
3900 }
3901
3902 /*
3903 * Generate the new variant.
3904 */
3905 t0 = os_time_get();
3906 variant = generate_variant(lp, shader, key);
3907 t1 = os_time_get();
3908 dt = t1 - t0;
3909 LP_COUNT_ADD(llvm_compile_time, dt);
3910 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3911
3912 /* Put the new variant into the list */
3913 if (variant) {
3914 insert_at_head(&shader->variants, &variant->list_item_local);
3915 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3916 lp->nr_fs_variants++;
3917 lp->nr_fs_instrs += variant->nr_instrs;
3918 shader->variants_cached++;
3919 }
3920 }
3921
3922 /* Bind this variant */
3923 lp_setup_set_fs_variant(lp->setup, variant);
3924 }
3925
3926
3927
3928
3929
3930 void
3931 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3932 {
3933 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3934 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3935 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3936
3937 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3938
3939 llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
3940 llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
3941 }
3942
3943