6e48b34622fea13a50b03219ff96deb7df724cd1
[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 #include "lp_screen.h"
109 #include "compiler/nir/nir_serialize.h"
110 #include "util/mesa-sha1.h"
111 /** Fragment shader number (for debugging) */
112 static unsigned fs_no = 0;
113
114
115 /**
116 * Expand the relevant bits of mask_input to a n*4-dword mask for the
117 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
118 * quad mask vector to 0 or ~0.
119 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
120 * quad arguments with fs length 8.
121 *
122 * \param first_quad which quad(s) of the quad group to test, in [0,3]
123 * \param mask_input bitwise mask for the whole 4x4 stamp
124 */
125 static LLVMValueRef
126 generate_quad_mask(struct gallivm_state *gallivm,
127 struct lp_type fs_type,
128 unsigned first_quad,
129 unsigned sample,
130 LLVMValueRef mask_input) /* int64 */
131 {
132 LLVMBuilderRef builder = gallivm->builder;
133 struct lp_type mask_type;
134 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
135 LLVMValueRef bits[16];
136 LLVMValueRef mask, bits_vec;
137 int shift, i;
138
139 /*
140 * XXX: We'll need a different path for 16 x u8
141 */
142 assert(fs_type.width == 32);
143 assert(fs_type.length <= ARRAY_SIZE(bits));
144 mask_type = lp_int_type(fs_type);
145
146 /*
147 * mask_input >>= (quad * 4)
148 */
149 switch (first_quad) {
150 case 0:
151 shift = 0;
152 break;
153 case 1:
154 assert(fs_type.length == 4);
155 shift = 2;
156 break;
157 case 2:
158 shift = 8;
159 break;
160 case 3:
161 assert(fs_type.length == 4);
162 shift = 10;
163 break;
164 default:
165 assert(0);
166 shift = 0;
167 }
168
169 mask_input = LLVMBuildLShr(builder, mask_input, lp_build_const_int64(gallivm, 16 * sample), "");
170 mask_input = LLVMBuildTrunc(builder, mask_input,
171 i32t, "");
172 mask_input = LLVMBuildAnd(builder, mask_input, lp_build_const_int32(gallivm, 0xffff), "");
173
174 mask_input = LLVMBuildLShr(builder,
175 mask_input,
176 LLVMConstInt(i32t, shift, 0),
177 "");
178
179 /*
180 * mask = { mask_input & (1 << i), for i in [0,3] }
181 */
182 mask = lp_build_broadcast(gallivm,
183 lp_build_vec_type(gallivm, mask_type),
184 mask_input);
185
186 for (i = 0; i < fs_type.length / 4; i++) {
187 unsigned j = 2 * (i % 2) + (i / 2) * 8;
188 bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
189 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
190 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
191 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
192 }
193 bits_vec = LLVMConstVector(bits, fs_type.length);
194 mask = LLVMBuildAnd(builder, mask, bits_vec, "");
195
196 /*
197 * mask = mask == bits ? ~0 : 0
198 */
199 mask = lp_build_compare(gallivm,
200 mask_type, PIPE_FUNC_EQUAL,
201 mask, bits_vec);
202
203 return mask;
204 }
205
206
207 #define EARLY_DEPTH_TEST 0x1
208 #define LATE_DEPTH_TEST 0x2
209 #define EARLY_DEPTH_WRITE 0x4
210 #define LATE_DEPTH_WRITE 0x8
211
212 static int
213 find_output_by_semantic( const struct tgsi_shader_info *info,
214 unsigned semantic,
215 unsigned index )
216 {
217 int i;
218
219 for (i = 0; i < info->num_outputs; i++)
220 if (info->output_semantic_name[i] == semantic &&
221 info->output_semantic_index[i] == index)
222 return i;
223
224 return -1;
225 }
226
227
228 /**
229 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
230 */
231 static LLVMValueRef
232 lp_llvm_viewport(LLVMValueRef context_ptr,
233 struct gallivm_state *gallivm,
234 LLVMValueRef viewport_index)
235 {
236 LLVMBuilderRef builder = gallivm->builder;
237 LLVMValueRef ptr;
238 LLVMValueRef res;
239 struct lp_type viewport_type =
240 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
241
242 ptr = lp_jit_context_viewports(gallivm, context_ptr);
243 ptr = LLVMBuildPointerCast(builder, ptr,
244 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
245
246 res = lp_build_pointer_get(builder, ptr, viewport_index);
247
248 return res;
249 }
250
251
252 static LLVMValueRef
253 lp_build_depth_clamp(struct gallivm_state *gallivm,
254 LLVMBuilderRef builder,
255 struct lp_type type,
256 LLVMValueRef context_ptr,
257 LLVMValueRef thread_data_ptr,
258 LLVMValueRef z)
259 {
260 LLVMValueRef viewport, min_depth, max_depth;
261 LLVMValueRef viewport_index;
262 struct lp_build_context f32_bld;
263
264 assert(type.floating);
265 lp_build_context_init(&f32_bld, gallivm, type);
266
267 /*
268 * Assumes clamping of the viewport index will occur in setup/gs. Value
269 * is passed through the rasterization stage via lp_rast_shader_inputs.
270 *
271 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
272 * semantics.
273 */
274 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
275 thread_data_ptr);
276
277 /*
278 * Load the min and max depth from the lp_jit_context.viewports
279 * array of lp_jit_viewport structures.
280 */
281 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
282
283 /* viewports[viewport_index].min_depth */
284 min_depth = LLVMBuildExtractElement(builder, viewport,
285 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
286 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
287
288 /* viewports[viewport_index].max_depth */
289 max_depth = LLVMBuildExtractElement(builder, viewport,
290 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
291 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
292
293 /*
294 * Clamp to the min and max depth values for the given viewport.
295 */
296 return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
297 }
298
299 static void
300 lp_build_sample_alpha_to_coverage(struct gallivm_state *gallivm,
301 struct lp_type type,
302 unsigned coverage_samples,
303 LLVMValueRef num_loop,
304 LLVMValueRef loop_counter,
305 LLVMValueRef coverage_mask_store,
306 LLVMValueRef alpha)
307 {
308 struct lp_build_context bld;
309 LLVMBuilderRef builder = gallivm->builder;
310 float step = 1.0 / coverage_samples;
311
312 lp_build_context_init(&bld, gallivm, type);
313 for (unsigned s = 0; s < coverage_samples; s++) {
314 LLVMValueRef alpha_ref_value = lp_build_const_vec(gallivm, type, step * s);
315 LLVMValueRef test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
316
317 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, lp_build_const_int32(gallivm, s), num_loop, "");
318 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_counter, "");
319 LLVMValueRef s_mask_ptr = LLVMBuildGEP(builder, coverage_mask_store, &s_mask_idx, 1, "");
320 LLVMValueRef s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
321 s_mask = LLVMBuildAnd(builder, s_mask, test, "");
322 LLVMBuildStore(builder, s_mask, s_mask_ptr);
323 }
324 };
325
326 struct lp_build_fs_llvm_iface {
327 struct lp_build_fs_iface base;
328 struct lp_build_interp_soa_context *interp;
329 struct lp_build_for_loop_state *loop_state;
330 LLVMValueRef mask_store;
331 };
332
333 static LLVMValueRef fs_interp(const struct lp_build_fs_iface *iface,
334 struct lp_build_context *bld,
335 unsigned attrib, unsigned chan,
336 bool centroid, bool sample,
337 LLVMValueRef attrib_indir,
338 LLVMValueRef offsets[2])
339 {
340 struct lp_build_fs_llvm_iface *fs_iface = (struct lp_build_fs_llvm_iface *)iface;
341 struct lp_build_interp_soa_context *interp = fs_iface->interp;
342 unsigned loc = TGSI_INTERPOLATE_LOC_CENTER;
343 if (centroid)
344 loc = TGSI_INTERPOLATE_LOC_CENTROID;
345 if (sample)
346 loc = TGSI_INTERPOLATE_LOC_SAMPLE;
347
348 return lp_build_interp_soa(interp, bld->gallivm, fs_iface->loop_state->counter,
349 fs_iface->mask_store,
350 attrib, chan, loc, attrib_indir, offsets);
351 }
352
353 /**
354 * Generate the fragment shader, depth/stencil test, and alpha tests.
355 */
356 static void
357 generate_fs_loop(struct gallivm_state *gallivm,
358 struct lp_fragment_shader *shader,
359 const struct lp_fragment_shader_variant_key *key,
360 LLVMBuilderRef builder,
361 struct lp_type type,
362 LLVMValueRef context_ptr,
363 LLVMValueRef sample_pos_array,
364 LLVMValueRef num_loop,
365 struct lp_build_interp_soa_context *interp,
366 const struct lp_build_sampler_soa *sampler,
367 const struct lp_build_image_soa *image,
368 LLVMValueRef mask_store,
369 LLVMValueRef (*out_color)[4],
370 LLVMValueRef depth_base_ptr,
371 LLVMValueRef depth_stride,
372 LLVMValueRef depth_sample_stride,
373 LLVMValueRef facing,
374 LLVMValueRef thread_data_ptr)
375 {
376 const struct util_format_description *zs_format_desc = NULL;
377 const struct tgsi_token *tokens = shader->base.tokens;
378 struct lp_type int_type = lp_int_type(type);
379 LLVMTypeRef vec_type, int_vec_type;
380 LLVMValueRef mask_ptr = NULL, mask_val = NULL;
381 LLVMValueRef consts_ptr, num_consts_ptr;
382 LLVMValueRef ssbo_ptr, num_ssbo_ptr;
383 LLVMValueRef z;
384 LLVMValueRef z_value, s_value;
385 LLVMValueRef z_fb, s_fb;
386 LLVMValueRef depth_ptr;
387 LLVMValueRef stencil_refs[2];
388 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
389 LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
390 struct lp_build_for_loop_state loop_state, sample_loop_state;
391 struct lp_build_mask_context mask;
392 /*
393 * TODO: figure out if simple_shader optimization is really worthwile to
394 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
395 * code since tests tend to take another codepath than real shaders.
396 */
397 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
398 shader->info.base.num_inputs < 3 &&
399 shader->info.base.num_instructions < 8) && 0;
400 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
401 util_blend_state_is_dual(&key->blend, 0);
402 unsigned attrib;
403 unsigned chan;
404 unsigned cbuf;
405 unsigned depth_mode;
406
407 struct lp_bld_tgsi_system_values system_values;
408
409 memset(&system_values, 0, sizeof(system_values));
410
411 /* truncate then sign extend. */
412 system_values.front_facing = LLVMBuildTrunc(gallivm->builder, facing, LLVMInt1TypeInContext(gallivm->context), "");
413 system_values.front_facing = LLVMBuildSExt(gallivm->builder, system_values.front_facing, LLVMInt32TypeInContext(gallivm->context), "");
414
415 if (key->depth.enabled ||
416 key->stencil[0].enabled) {
417
418 zs_format_desc = util_format_description(key->zsbuf_format);
419 assert(zs_format_desc);
420
421 if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
422 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
423 else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
424 if (shader->info.base.writes_memory)
425 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
426 else if (key->alpha.enabled ||
427 key->blend.alpha_to_coverage ||
428 shader->info.base.uses_kill ||
429 shader->info.base.writes_samplemask) {
430 /* With alpha test and kill, can do the depth test early
431 * and hopefully eliminate some quads. But need to do a
432 * special deferred depth write once the final mask value
433 * is known. This only works though if there's either no
434 * stencil test or the stencil value isn't written.
435 */
436 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
437 (key->stencil[1].enabled &&
438 key->stencil[1].writemask)))
439 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
440 else
441 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
442 }
443 else
444 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
445 }
446 else {
447 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
448 }
449
450 if (!(key->depth.enabled && key->depth.writemask) &&
451 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
452 (key->stencil[1].enabled &&
453 key->stencil[1].writemask))))
454 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
455 }
456 else {
457 depth_mode = 0;
458 }
459
460 vec_type = lp_build_vec_type(gallivm, type);
461 int_vec_type = lp_build_vec_type(gallivm, int_type);
462
463 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
464 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
465 /* convert scalar stencil refs into vectors */
466 stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
467 stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
468
469 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
470 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
471
472 ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
473 num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
474
475 memset(outputs, 0, sizeof outputs);
476
477 /* Allocate color storage for each fragment sample */
478 LLVMValueRef color_store_size = num_loop;
479 if (key->min_samples > 1)
480 color_store_size = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, key->min_samples), "");
481
482 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
483 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
484 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
485 lp_build_vec_type(gallivm,
486 type),
487 color_store_size, "color");
488 }
489 }
490 if (dual_source_blend) {
491 assert(key->nr_cbufs <= 1);
492 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
493 out_color[1][chan] = lp_build_array_alloca(gallivm,
494 lp_build_vec_type(gallivm,
495 type),
496 color_store_size, "color1");
497 }
498 }
499
500 lp_build_for_loop_begin(&loop_state, gallivm,
501 lp_build_const_int32(gallivm, 0),
502 LLVMIntULT,
503 num_loop,
504 lp_build_const_int32(gallivm, 1));
505
506 LLVMValueRef sample_mask_in;
507 if (key->multisample) {
508 sample_mask_in = lp_build_const_int_vec(gallivm, type, 0);
509 /* create shader execution mask by combining all sample masks. */
510 for (unsigned s = 0; s < key->coverage_samples; s++) {
511 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
512 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
513 LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
514 if (s == 0)
515 mask_val = s_mask;
516 else
517 mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
518
519 LLVMValueRef mask_in = LLVMBuildAnd(builder, s_mask, lp_build_const_int_vec(gallivm, type, (1 << s)), "");
520 sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
521 }
522 } else {
523 sample_mask_in = lp_build_const_int_vec(gallivm, type, 1);
524 mask_ptr = LLVMBuildGEP(builder, mask_store,
525 &loop_state.counter, 1, "mask_ptr");
526 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
527
528 LLVMValueRef mask_in = LLVMBuildAnd(builder, mask_val, lp_build_const_int_vec(gallivm, type, 1), "");
529 sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
530 }
531
532 /* 'mask' will control execution based on quad's pixel alive/killed state */
533 lp_build_mask_begin(&mask, gallivm, type, mask_val);
534
535 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
536 lp_build_mask_check(&mask);
537
538 /* Create storage for recombining sample masks after early Z pass. */
539 LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
540 LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
541
542 LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
543 LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
544 LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
545 LLVMTypeRef z_type = NULL, z_fb_type = NULL;
546
547 /* Run early depth once per sample */
548 if (key->multisample) {
549
550 if (zs_format_desc) {
551 struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
552 struct lp_type z_type = zs_type;
553 struct lp_type s_type = zs_type;
554 if (zs_format_desc->block.bits < type.width)
555 z_type.width = type.width;
556 else if (zs_format_desc->block.bits > 32) {
557 z_type.width = z_type.width / 2;
558 s_type.width = s_type.width / 2;
559 s_type.floating = 0;
560 }
561 z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
562 zs_samples, "z_sample_store");
563 s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
564 zs_samples, "s_sample_store");
565 z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
566 zs_samples, "z_fb_store");
567 s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
568 zs_samples, "s_fb_store");
569 }
570 lp_build_for_loop_begin(&sample_loop_state, gallivm,
571 lp_build_const_int32(gallivm, 0),
572 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
573 lp_build_const_int32(gallivm, 1));
574
575 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
576 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
577 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
578
579 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
580 s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
581 }
582
583
584 /* for multisample Z needs to be interpolated at sample points for testing. */
585 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
586 z = interp->pos[2];
587
588 depth_ptr = depth_base_ptr;
589 if (key->multisample) {
590 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
591 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
592 }
593
594 if (depth_mode & EARLY_DEPTH_TEST) {
595 /*
596 * Clamp according to ARB_depth_clamp semantics.
597 */
598 if (key->depth_clamp) {
599 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
600 thread_data_ptr, z);
601 }
602 lp_build_depth_stencil_load_swizzled(gallivm, type,
603 zs_format_desc, key->resource_1d,
604 depth_ptr, depth_stride,
605 &z_fb, &s_fb, loop_state.counter);
606 lp_build_depth_stencil_test(gallivm,
607 &key->depth,
608 key->stencil,
609 type,
610 zs_format_desc,
611 key->multisample ? NULL : &mask,
612 &s_mask,
613 stencil_refs,
614 z, z_fb, s_fb,
615 facing,
616 &z_value, &s_value,
617 !simple_shader && !key->multisample);
618
619 if (depth_mode & EARLY_DEPTH_WRITE) {
620 lp_build_depth_stencil_write_swizzled(gallivm, type,
621 zs_format_desc, key->resource_1d,
622 NULL, NULL, NULL, loop_state.counter,
623 depth_ptr, depth_stride,
624 z_value, s_value);
625 }
626 /*
627 * Note mask check if stencil is enabled must be after ds write not after
628 * stencil test otherwise new stencil values may not get written if all
629 * fragments got killed by depth/stencil test.
630 */
631 if (!simple_shader && key->stencil[0].enabled && !key->multisample)
632 lp_build_mask_check(&mask);
633
634 if (key->multisample) {
635 z_fb_type = LLVMTypeOf(z_fb);
636 z_type = LLVMTypeOf(z_value);
637 lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
638 lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
639 lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
640 lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
641 }
642 }
643
644 if (key->multisample) {
645 /*
646 * Store the post-early Z coverage mask.
647 * Recombine the resulting coverage masks post early Z into the fragment
648 * shader execution mask.
649 */
650 LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
651 tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
652 LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
653
654 LLVMBuildStore(builder, s_mask, s_mask_ptr);
655
656 lp_build_for_loop_end(&sample_loop_state);
657
658 /* recombined all the coverage masks in the shader exec mask. */
659 tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
660 lp_build_mask_update(&mask, tmp_s_mask_or);
661
662 if (key->min_samples == 1) {
663 /* for multisample Z needs to be re interpolated at pixel center */
664 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
665 lp_build_mask_update(&mask, tmp_s_mask_or);
666 }
667 }
668
669 LLVMValueRef out_sample_mask_storage = NULL;
670 if (shader->info.base.writes_samplemask) {
671 out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
672 if (key->min_samples > 1)
673 LLVMBuildStore(builder, LLVMConstNull(int_vec_type), out_sample_mask_storage);
674 }
675
676 if (key->multisample && key->min_samples > 1) {
677 lp_build_for_loop_begin(&sample_loop_state, gallivm,
678 lp_build_const_int32(gallivm, 0),
679 LLVMIntULT,
680 lp_build_const_int32(gallivm, key->min_samples),
681 lp_build_const_int32(gallivm, 1));
682
683 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
684 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
685 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
686 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
687 lp_build_mask_force(&mask, s_mask);
688 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, sample_loop_state.counter);
689 system_values.sample_id = sample_loop_state.counter;
690 } else
691 system_values.sample_id = lp_build_const_int32(gallivm, 0);
692
693 system_values.sample_mask_in = sample_mask_in;
694 system_values.sample_pos = sample_pos_array;
695
696 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, sample_loop_state.counter);
697
698 struct lp_build_fs_llvm_iface fs_iface = {
699 .base.interp_fn = fs_interp,
700 .interp = interp,
701 .loop_state = &loop_state,
702 .mask_store = mask_store,
703 };
704
705 struct lp_build_tgsi_params params;
706 memset(&params, 0, sizeof(params));
707
708 params.type = type;
709 params.mask = &mask;
710 params.fs_iface = &fs_iface.base;
711 params.consts_ptr = consts_ptr;
712 params.const_sizes_ptr = num_consts_ptr;
713 params.system_values = &system_values;
714 params.inputs = interp->inputs;
715 params.context_ptr = context_ptr;
716 params.thread_data_ptr = thread_data_ptr;
717 params.sampler = sampler;
718 params.info = &shader->info.base;
719 params.ssbo_ptr = ssbo_ptr;
720 params.ssbo_sizes_ptr = num_ssbo_ptr;
721 params.image = image;
722
723 /* Build the actual shader */
724 if (shader->base.type == PIPE_SHADER_IR_TGSI)
725 lp_build_tgsi_soa(gallivm, tokens, &params,
726 outputs);
727 else
728 lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
729 outputs);
730
731 /* Alpha test */
732 if (key->alpha.enabled) {
733 int color0 = find_output_by_semantic(&shader->info.base,
734 TGSI_SEMANTIC_COLOR,
735 0);
736
737 if (color0 != -1 && outputs[color0][3]) {
738 const struct util_format_description *cbuf_format_desc;
739 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
740 LLVMValueRef alpha_ref_value;
741
742 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
743 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
744
745 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
746
747 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
748 &mask, alpha, alpha_ref_value,
749 (depth_mode & LATE_DEPTH_TEST) != 0);
750 }
751 }
752
753 /* Emulate Alpha to Coverage with Alpha test */
754 if (key->blend.alpha_to_coverage) {
755 int color0 = find_output_by_semantic(&shader->info.base,
756 TGSI_SEMANTIC_COLOR,
757 0);
758
759 if (color0 != -1 && outputs[color0][3]) {
760 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
761
762 if (!key->multisample) {
763 lp_build_alpha_to_coverage(gallivm, type,
764 &mask, alpha,
765 (depth_mode & LATE_DEPTH_TEST) != 0);
766 } else {
767 lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
768 loop_state.counter,
769 mask_store, alpha);
770 }
771 }
772 }
773 if (key->blend.alpha_to_one && key->multisample) {
774 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {
775 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
776 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
777 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
778 if (outputs[cbuf][3]) {
779 LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
780 }
781 }
782 }
783 if (shader->info.base.writes_samplemask) {
784 LLVMValueRef output_smask = NULL;
785 int smaski = find_output_by_semantic(&shader->info.base,
786 TGSI_SEMANTIC_SAMPLEMASK,
787 0);
788 struct lp_build_context smask_bld;
789 lp_build_context_init(&smask_bld, gallivm, int_type);
790
791 assert(smaski >= 0);
792 output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
793 output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
794
795 if (key->min_samples > 1) {
796 /* only the bit corresponding to this sample is to be used. */
797 LLVMValueRef tmp_mask = LLVMBuildLoad(builder, out_sample_mask_storage, "tmp_mask");
798 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
799 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, lp_build_broadcast(gallivm, int_vec_type, out_smask_idx), "");
800 output_smask = LLVMBuildOr(builder, tmp_mask, smask_bit, "");
801 }
802
803 LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
804 }
805
806 /* Color write - per fragment sample */
807 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
808 {
809 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
810 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
811 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
812 {
813 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
814 if(outputs[attrib][chan]) {
815 /* XXX: just initialize outputs to point at colors[] and
816 * skip this.
817 */
818 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
819 LLVMValueRef color_ptr;
820 LLVMValueRef color_idx = loop_state.counter;
821 if (key->min_samples > 1)
822 color_idx = LLVMBuildAdd(builder, color_idx,
823 LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
824 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
825 &color_idx, 1, "");
826 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
827 LLVMBuildStore(builder, out, color_ptr);
828 }
829 }
830 }
831 }
832
833 if (key->multisample && key->min_samples > 1) {
834 LLVMBuildStore(builder, lp_build_mask_value(&mask), s_mask_ptr);
835 lp_build_for_loop_end(&sample_loop_state);
836 }
837
838 if (key->multisample) {
839 /* execute depth test for each sample */
840 lp_build_for_loop_begin(&sample_loop_state, gallivm,
841 lp_build_const_int32(gallivm, 0),
842 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
843 lp_build_const_int32(gallivm, 1));
844
845 /* load the per-sample coverage mask */
846 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
847 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
848 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
849
850 /* combine the execution mask post fragment shader with the coverage mask. */
851 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
852 if (key->min_samples == 1)
853 s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
854
855 /* if the shader writes sample mask use that */
856 if (shader->info.base.writes_samplemask) {
857 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
858 out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
859 LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
860 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
861 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
862 smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
863
864 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
865 }
866 }
867
868 depth_ptr = depth_base_ptr;
869 if (key->multisample) {
870 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
871 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
872 }
873
874 /* Late Z test */
875 if (depth_mode & LATE_DEPTH_TEST) {
876 int pos0 = find_output_by_semantic(&shader->info.base,
877 TGSI_SEMANTIC_POSITION,
878 0);
879 int s_out = find_output_by_semantic(&shader->info.base,
880 TGSI_SEMANTIC_STENCIL,
881 0);
882 if (pos0 != -1 && outputs[pos0][2]) {
883 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
884 }
885 /*
886 * Clamp according to ARB_depth_clamp semantics.
887 */
888 if (key->depth_clamp) {
889 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
890 thread_data_ptr, z);
891 }
892
893 if (s_out != -1 && outputs[s_out][1]) {
894 /* there's only one value, and spec says to discard additional bits */
895 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
896 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
897 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
898 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
899 stencil_refs[1] = stencil_refs[0];
900 }
901
902 lp_build_depth_stencil_load_swizzled(gallivm, type,
903 zs_format_desc, key->resource_1d,
904 depth_ptr, depth_stride,
905 &z_fb, &s_fb, loop_state.counter);
906
907 lp_build_depth_stencil_test(gallivm,
908 &key->depth,
909 key->stencil,
910 type,
911 zs_format_desc,
912 key->multisample ? NULL : &mask,
913 &s_mask,
914 stencil_refs,
915 z, z_fb, s_fb,
916 facing,
917 &z_value, &s_value,
918 !simple_shader);
919 /* Late Z write */
920 if (depth_mode & LATE_DEPTH_WRITE) {
921 lp_build_depth_stencil_write_swizzled(gallivm, type,
922 zs_format_desc, key->resource_1d,
923 NULL, NULL, NULL, loop_state.counter,
924 depth_ptr, depth_stride,
925 z_value, s_value);
926 }
927 }
928 else if ((depth_mode & EARLY_DEPTH_TEST) &&
929 (depth_mode & LATE_DEPTH_WRITE))
930 {
931 /* Need to apply a reduced mask to the depth write. Reload the
932 * depth value, update from zs_value with the new mask value and
933 * write that out.
934 */
935 if (key->multisample) {
936 z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
937 s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
938 z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
939 s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
940 }
941 lp_build_depth_stencil_write_swizzled(gallivm, type,
942 zs_format_desc, key->resource_1d,
943 key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
944 depth_ptr, depth_stride,
945 z_value, s_value);
946 }
947
948 if (key->occlusion_count) {
949 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
950 lp_build_name(counter, "counter");
951
952 lp_build_occlusion_count(gallivm, type,
953 key->multisample ? s_mask : lp_build_mask_value(&mask), counter);
954 }
955
956 if (key->multisample) {
957 /* store the sample mask for this loop */
958 LLVMBuildStore(builder, s_mask, s_mask_ptr);
959 lp_build_for_loop_end(&sample_loop_state);
960 }
961
962 mask_val = lp_build_mask_end(&mask);
963 if (!key->multisample)
964 LLVMBuildStore(builder, mask_val, mask_ptr);
965 lp_build_for_loop_end(&loop_state);
966 }
967
968
969 /**
970 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
971 *
972 * Fragment Shader outputs pixels in small 2x2 blocks
973 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
974 *
975 * However in memory pixels are stored in rows
976 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
977 *
978 * @param type fragment shader type (4x or 8x float)
979 * @param num_fs number of fs_src
980 * @param is_1d whether we're outputting to a 1d resource
981 * @param dst_channels number of output channels
982 * @param fs_src output from fragment shader
983 * @param dst pointer to store result
984 * @param pad_inline is channel padding inline or at end of row
985 * @return the number of dsts
986 */
987 static int
988 generate_fs_twiddle(struct gallivm_state *gallivm,
989 struct lp_type type,
990 unsigned num_fs,
991 unsigned dst_channels,
992 LLVMValueRef fs_src[][4],
993 LLVMValueRef* dst,
994 bool pad_inline)
995 {
996 LLVMValueRef src[16];
997
998 bool swizzle_pad;
999 bool twiddle;
1000 bool split;
1001
1002 unsigned pixels = type.length / 4;
1003 unsigned reorder_group;
1004 unsigned src_channels;
1005 unsigned src_count;
1006 unsigned i;
1007
1008 src_channels = dst_channels < 3 ? dst_channels : 4;
1009 src_count = num_fs * src_channels;
1010
1011 assert(pixels == 2 || pixels == 1);
1012 assert(num_fs * src_channels <= ARRAY_SIZE(src));
1013
1014 /*
1015 * Transpose from SoA -> AoS
1016 */
1017 for (i = 0; i < num_fs; ++i) {
1018 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
1019 }
1020
1021 /*
1022 * Pick transformation options
1023 */
1024 swizzle_pad = false;
1025 twiddle = false;
1026 split = false;
1027 reorder_group = 0;
1028
1029 if (dst_channels == 1) {
1030 twiddle = true;
1031
1032 if (pixels == 2) {
1033 split = true;
1034 }
1035 } else if (dst_channels == 2) {
1036 if (pixels == 1) {
1037 reorder_group = 1;
1038 }
1039 } else if (dst_channels > 2) {
1040 if (pixels == 1) {
1041 reorder_group = 2;
1042 } else {
1043 twiddle = true;
1044 }
1045
1046 if (!pad_inline && dst_channels == 3 && pixels > 1) {
1047 swizzle_pad = true;
1048 }
1049 }
1050
1051 /*
1052 * Split the src in half
1053 */
1054 if (split) {
1055 for (i = num_fs; i > 0; --i) {
1056 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
1057 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
1058 }
1059
1060 src_count *= 2;
1061 type.length = 4;
1062 }
1063
1064 /*
1065 * Ensure pixels are in memory order
1066 */
1067 if (reorder_group) {
1068 /* Twiddle pixels by reordering the array, e.g.:
1069 *
1070 * src_count = 8 -> 0 2 1 3 4 6 5 7
1071 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1072 */
1073 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
1074
1075 for (i = 0; i < src_count; ++i) {
1076 unsigned group = i / reorder_group;
1077 unsigned block = (group / 4) * 4 * reorder_group;
1078 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
1079 dst[i] = src[j];
1080 }
1081 } else if (twiddle) {
1082 /* Twiddle pixels across elements of array */
1083 /*
1084 * XXX: we should avoid this in some cases, but would need to tell
1085 * lp_build_conv to reorder (or deal with it ourselves).
1086 */
1087 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1088 } else {
1089 /* Do nothing */
1090 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1091 }
1092
1093 /*
1094 * Moves any padding between pixels to the end
1095 * e.g. RGBXRGBX -> RGBRGBXX
1096 */
1097 if (swizzle_pad) {
1098 unsigned char swizzles[16];
1099 unsigned elems = pixels * dst_channels;
1100
1101 for (i = 0; i < type.length; ++i) {
1102 if (i < elems)
1103 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1104 else
1105 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1106 }
1107
1108 for (i = 0; i < src_count; ++i) {
1109 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1110 }
1111 }
1112
1113 return src_count;
1114 }
1115
1116
1117 /*
1118 * Untwiddle and transpose, much like the above.
1119 * However, this is after conversion, so we get packed vectors.
1120 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1121 * the vectors will look like:
1122 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1123 * be swizzled here). Extending to 16bit should be trivial.
1124 * Should also be extended to handle twice wide vectors with AVX2...
1125 */
1126 static void
1127 fs_twiddle_transpose(struct gallivm_state *gallivm,
1128 struct lp_type type,
1129 LLVMValueRef *src,
1130 unsigned src_count,
1131 LLVMValueRef *dst)
1132 {
1133 unsigned i, j;
1134 struct lp_type type64, type16, type32;
1135 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1136 LLVMBuilderRef builder = gallivm->builder;
1137 LLVMValueRef tmp[4], shuf[8];
1138 for (j = 0; j < 2; j++) {
1139 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1140 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1141 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1142 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1143 }
1144
1145 assert(src_count == 4 || src_count == 2 || src_count == 1);
1146 assert(type.width == 8);
1147 assert(type.length == 16);
1148
1149 type8_t = lp_build_vec_type(gallivm, type);
1150
1151 type64 = type;
1152 type64.length /= 8;
1153 type64.width *= 8;
1154 type64_t = lp_build_vec_type(gallivm, type64);
1155
1156 type16 = type;
1157 type16.length /= 2;
1158 type16.width *= 2;
1159 type16_t = lp_build_vec_type(gallivm, type16);
1160
1161 type32 = type;
1162 type32.length /= 4;
1163 type32.width *= 4;
1164 type32_t = lp_build_vec_type(gallivm, type32);
1165
1166 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1167
1168 if (src_count == 1) {
1169 /* transpose was no-op, just untwiddle */
1170 LLVMValueRef shuf_vec;
1171 shuf_vec = LLVMConstVector(shuf, 8);
1172 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1173 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1174 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1175 } else if (src_count == 2) {
1176 LLVMValueRef shuf_vec;
1177 shuf_vec = LLVMConstVector(shuf, 4);
1178
1179 for (i = 0; i < 2; i++) {
1180 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1181 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1182 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1183 }
1184 } else {
1185 for (j = 0; j < 2; j++) {
1186 LLVMValueRef lo, hi, lo2, hi2;
1187 /*
1188 * Note that if we only really have 3 valid channels (rgb)
1189 * and we don't need alpha we could substitute a undef here
1190 * for the respective channel (causing llvm to drop conversion
1191 * for alpha).
1192 */
1193 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1194 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1195 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1196 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1197 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1198 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1199 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1200 }
1201 }
1202 }
1203
1204
1205 /**
1206 * Load an unswizzled block of pixels from memory
1207 */
1208 static void
1209 load_unswizzled_block(struct gallivm_state *gallivm,
1210 LLVMValueRef base_ptr,
1211 LLVMValueRef stride,
1212 unsigned block_width,
1213 unsigned block_height,
1214 LLVMValueRef* dst,
1215 struct lp_type dst_type,
1216 unsigned dst_count,
1217 unsigned dst_alignment)
1218 {
1219 LLVMBuilderRef builder = gallivm->builder;
1220 unsigned row_size = dst_count / block_height;
1221 unsigned i;
1222
1223 /* Ensure block exactly fits into dst */
1224 assert((block_width * block_height) % dst_count == 0);
1225
1226 for (i = 0; i < dst_count; ++i) {
1227 unsigned x = i % row_size;
1228 unsigned y = i / row_size;
1229
1230 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1231 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1232
1233 LLVMValueRef gep[2];
1234 LLVMValueRef dst_ptr;
1235
1236 gep[0] = lp_build_const_int32(gallivm, 0);
1237 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1238
1239 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1240 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1241 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1242
1243 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1244
1245 LLVMSetAlignment(dst[i], dst_alignment);
1246 }
1247 }
1248
1249
1250 /**
1251 * Store an unswizzled block of pixels to memory
1252 */
1253 static void
1254 store_unswizzled_block(struct gallivm_state *gallivm,
1255 LLVMValueRef base_ptr,
1256 LLVMValueRef stride,
1257 unsigned block_width,
1258 unsigned block_height,
1259 LLVMValueRef* src,
1260 struct lp_type src_type,
1261 unsigned src_count,
1262 unsigned src_alignment)
1263 {
1264 LLVMBuilderRef builder = gallivm->builder;
1265 unsigned row_size = src_count / block_height;
1266 unsigned i;
1267
1268 /* Ensure src exactly fits into block */
1269 assert((block_width * block_height) % src_count == 0);
1270
1271 for (i = 0; i < src_count; ++i) {
1272 unsigned x = i % row_size;
1273 unsigned y = i / row_size;
1274
1275 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1276 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1277
1278 LLVMValueRef gep[2];
1279 LLVMValueRef src_ptr;
1280
1281 gep[0] = lp_build_const_int32(gallivm, 0);
1282 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1283
1284 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1285 src_ptr = LLVMBuildBitCast(builder, src_ptr,
1286 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1287
1288 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1289
1290 LLVMSetAlignment(src_ptr, src_alignment);
1291 }
1292 }
1293
1294
1295 /**
1296 * Checks if a format description is an arithmetic format
1297 *
1298 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1299 */
1300 static inline boolean
1301 is_arithmetic_format(const struct util_format_description *format_desc)
1302 {
1303 boolean arith = false;
1304 unsigned i;
1305
1306 for (i = 0; i < format_desc->nr_channels; ++i) {
1307 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1308 arith |= (format_desc->channel[i].size % 8) != 0;
1309 }
1310
1311 return arith;
1312 }
1313
1314
1315 /**
1316 * Checks if this format requires special handling due to required expansion
1317 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1318 * SoA conversion.
1319 */
1320 static inline boolean
1321 format_expands_to_float_soa(const struct util_format_description *format_desc)
1322 {
1323 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
1324 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1325 return true;
1326 }
1327 return false;
1328 }
1329
1330
1331 /**
1332 * Retrieves the type representing the memory layout for a format
1333 *
1334 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1335 */
1336 static inline void
1337 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1338 struct lp_type* type)
1339 {
1340 unsigned i;
1341 unsigned chan;
1342
1343 if (format_expands_to_float_soa(format_desc)) {
1344 /* just make this a uint with width of block */
1345 type->floating = false;
1346 type->fixed = false;
1347 type->sign = false;
1348 type->norm = false;
1349 type->width = format_desc->block.bits;
1350 type->length = 1;
1351 return;
1352 }
1353
1354 for (i = 0; i < 4; i++)
1355 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1356 break;
1357 chan = i;
1358
1359 memset(type, 0, sizeof(struct lp_type));
1360 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1361 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1362 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1363 type->norm = format_desc->channel[chan].normalized;
1364
1365 if (is_arithmetic_format(format_desc)) {
1366 type->width = 0;
1367 type->length = 1;
1368
1369 for (i = 0; i < format_desc->nr_channels; ++i) {
1370 type->width += format_desc->channel[i].size;
1371 }
1372 } else {
1373 type->width = format_desc->channel[chan].size;
1374 type->length = format_desc->nr_channels;
1375 }
1376 }
1377
1378
1379 /**
1380 * Retrieves the type for a format which is usable in the blending code.
1381 *
1382 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1383 */
1384 static inline void
1385 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1386 struct lp_type* type)
1387 {
1388 unsigned i;
1389 unsigned chan;
1390
1391 if (format_expands_to_float_soa(format_desc)) {
1392 /* always use ordinary floats for blending */
1393 type->floating = true;
1394 type->fixed = false;
1395 type->sign = true;
1396 type->norm = false;
1397 type->width = 32;
1398 type->length = 4;
1399 return;
1400 }
1401
1402 for (i = 0; i < 4; i++)
1403 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1404 break;
1405 chan = i;
1406
1407 memset(type, 0, sizeof(struct lp_type));
1408 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1409 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1410 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1411 type->norm = format_desc->channel[chan].normalized;
1412 type->width = format_desc->channel[chan].size;
1413 type->length = format_desc->nr_channels;
1414
1415 for (i = 1; i < format_desc->nr_channels; ++i) {
1416 if (format_desc->channel[i].size > type->width)
1417 type->width = format_desc->channel[i].size;
1418 }
1419
1420 if (type->floating) {
1421 type->width = 32;
1422 } else {
1423 if (type->width <= 8) {
1424 type->width = 8;
1425 } else if (type->width <= 16) {
1426 type->width = 16;
1427 } else {
1428 type->width = 32;
1429 }
1430 }
1431
1432 if (is_arithmetic_format(format_desc) && type->length == 3) {
1433 type->length = 4;
1434 }
1435 }
1436
1437
1438 /**
1439 * Scale a normalized value from src_bits to dst_bits.
1440 *
1441 * The exact calculation is
1442 *
1443 * dst = iround(src * dst_mask / src_mask)
1444 *
1445 * or with integer rounding
1446 *
1447 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1448 *
1449 * where
1450 *
1451 * src_mask = (1 << src_bits) - 1
1452 * dst_mask = (1 << dst_bits) - 1
1453 *
1454 * but we try to avoid division and multiplication through shifts.
1455 */
1456 static inline LLVMValueRef
1457 scale_bits(struct gallivm_state *gallivm,
1458 int src_bits,
1459 int dst_bits,
1460 LLVMValueRef src,
1461 struct lp_type src_type)
1462 {
1463 LLVMBuilderRef builder = gallivm->builder;
1464 LLVMValueRef result = src;
1465
1466 if (dst_bits < src_bits) {
1467 int delta_bits = src_bits - dst_bits;
1468
1469 if (delta_bits <= dst_bits) {
1470 /*
1471 * Approximate the rescaling with a single shift.
1472 *
1473 * This gives the wrong rounding.
1474 */
1475
1476 result = LLVMBuildLShr(builder,
1477 src,
1478 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1479 "");
1480
1481 } else {
1482 /*
1483 * Try more accurate rescaling.
1484 */
1485
1486 /*
1487 * Drop the least significant bits to make space for the multiplication.
1488 *
1489 * XXX: A better approach would be to use a wider integer type as intermediate. But
1490 * this is enough to convert alpha from 16bits -> 2 when rendering to
1491 * PIPE_FORMAT_R10G10B10A2_UNORM.
1492 */
1493 result = LLVMBuildLShr(builder,
1494 src,
1495 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1496 "");
1497
1498
1499 result = LLVMBuildMul(builder,
1500 result,
1501 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1502 "");
1503
1504 /*
1505 * Add a rounding term before the division.
1506 *
1507 * TODO: Handle signed integers too.
1508 */
1509 if (!src_type.sign) {
1510 result = LLVMBuildAdd(builder,
1511 result,
1512 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1513 "");
1514 }
1515
1516 /*
1517 * Approximate the division by src_mask with a src_bits shift.
1518 *
1519 * Given the src has already been shifted by dst_bits, all we need
1520 * to do is to shift by the difference.
1521 */
1522
1523 result = LLVMBuildLShr(builder,
1524 result,
1525 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1526 "");
1527 }
1528
1529 } else if (dst_bits > src_bits) {
1530 /* Scale up bits */
1531 int db = dst_bits - src_bits;
1532
1533 /* Shift left by difference in bits */
1534 result = LLVMBuildShl(builder,
1535 src,
1536 lp_build_const_int_vec(gallivm, src_type, db),
1537 "");
1538
1539 if (db <= src_bits) {
1540 /* Enough bits in src to fill the remainder */
1541 LLVMValueRef lower = LLVMBuildLShr(builder,
1542 src,
1543 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1544 "");
1545
1546 result = LLVMBuildOr(builder, result, lower, "");
1547 } else if (db > src_bits) {
1548 /* Need to repeatedly copy src bits to fill remainder in dst */
1549 unsigned n;
1550
1551 for (n = src_bits; n < dst_bits; n *= 2) {
1552 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1553
1554 result = LLVMBuildOr(builder,
1555 result,
1556 LLVMBuildLShr(builder, result, shuv, ""),
1557 "");
1558 }
1559 }
1560 }
1561
1562 return result;
1563 }
1564
1565 /**
1566 * If RT is a smallfloat (needing denorms) format
1567 */
1568 static inline int
1569 have_smallfloat_format(struct lp_type dst_type,
1570 enum pipe_format format)
1571 {
1572 return ((dst_type.floating && dst_type.width != 32) ||
1573 /* due to format handling hacks this format doesn't have floating set
1574 * here (and actually has width set to 32 too) so special case this. */
1575 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1576 }
1577
1578
1579 /**
1580 * Convert from memory format to blending format
1581 *
1582 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1583 */
1584 static void
1585 convert_to_blend_type(struct gallivm_state *gallivm,
1586 unsigned block_size,
1587 const struct util_format_description *src_fmt,
1588 struct lp_type src_type,
1589 struct lp_type dst_type,
1590 LLVMValueRef* src, // and dst
1591 unsigned num_srcs)
1592 {
1593 LLVMValueRef *dst = src;
1594 LLVMBuilderRef builder = gallivm->builder;
1595 struct lp_type blend_type;
1596 struct lp_type mem_type;
1597 unsigned i, j;
1598 unsigned pixels = block_size / num_srcs;
1599 bool is_arith;
1600
1601 /*
1602 * full custom path for packed floats and srgb formats - none of the later
1603 * functions would do anything useful, and given the lp_type representation they
1604 * can't be fixed. Should really have some SoA blend path for these kind of
1605 * formats rather than hacking them in here.
1606 */
1607 if (format_expands_to_float_soa(src_fmt)) {
1608 LLVMValueRef tmpsrc[4];
1609 /*
1610 * This is pretty suboptimal for this case blending in SoA would be much
1611 * better, since conversion gets us SoA values so need to convert back.
1612 */
1613 assert(src_type.width == 32 || src_type.width == 16);
1614 assert(dst_type.floating);
1615 assert(dst_type.width == 32);
1616 assert(dst_type.length % 4 == 0);
1617 assert(num_srcs % 4 == 0);
1618
1619 if (src_type.width == 16) {
1620 /* expand 4x16bit values to 4x32bit */
1621 struct lp_type type32x4 = src_type;
1622 LLVMTypeRef ltype32x4;
1623 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1624 type32x4.width = 32;
1625 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1626 for (i = 0; i < num_fetch; i++) {
1627 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1628 }
1629 src_type.width = 32;
1630 }
1631 for (i = 0; i < 4; i++) {
1632 tmpsrc[i] = src[i];
1633 }
1634 for (i = 0; i < num_srcs / 4; i++) {
1635 LLVMValueRef tmpsoa[4];
1636 LLVMValueRef tmps = tmpsrc[i];
1637 if (dst_type.length == 8) {
1638 LLVMValueRef shuffles[8];
1639 unsigned j;
1640 /* fetch was 4 values but need 8-wide output values */
1641 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1642 /*
1643 * for 8-wide aos transpose would give us wrong order not matching
1644 * incoming converted fs values and mask. ARGH.
1645 */
1646 for (j = 0; j < 4; j++) {
1647 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1648 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1649 }
1650 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1651 LLVMConstVector(shuffles, 8), "");
1652 }
1653 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1654 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1655 }
1656 else {
1657 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1658 }
1659 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1660 }
1661 return;
1662 }
1663
1664 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1665 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1666
1667 /* Is the format arithmetic */
1668 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1669 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1670
1671 /* Pad if necessary */
1672 if (!is_arith && src_type.length < dst_type.length) {
1673 for (i = 0; i < num_srcs; ++i) {
1674 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1675 }
1676
1677 src_type.length = dst_type.length;
1678 }
1679
1680 /* Special case for half-floats */
1681 if (mem_type.width == 16 && mem_type.floating) {
1682 assert(blend_type.width == 32 && blend_type.floating);
1683 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1684 is_arith = false;
1685 }
1686
1687 if (!is_arith) {
1688 return;
1689 }
1690
1691 src_type.width = blend_type.width * blend_type.length;
1692 blend_type.length *= pixels;
1693 src_type.length *= pixels / (src_type.length / mem_type.length);
1694
1695 for (i = 0; i < num_srcs; ++i) {
1696 LLVMValueRef chans[4];
1697 LLVMValueRef res = NULL;
1698
1699 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1700
1701 for (j = 0; j < src_fmt->nr_channels; ++j) {
1702 unsigned mask = 0;
1703 unsigned sa = src_fmt->channel[j].shift;
1704 #if UTIL_ARCH_LITTLE_ENDIAN
1705 unsigned from_lsb = j;
1706 #else
1707 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1708 #endif
1709
1710 mask = (1 << src_fmt->channel[j].size) - 1;
1711
1712 /* Extract bits from source */
1713 chans[j] = LLVMBuildLShr(builder,
1714 dst[i],
1715 lp_build_const_int_vec(gallivm, src_type, sa),
1716 "");
1717
1718 chans[j] = LLVMBuildAnd(builder,
1719 chans[j],
1720 lp_build_const_int_vec(gallivm, src_type, mask),
1721 "");
1722
1723 /* Scale bits */
1724 if (src_type.norm) {
1725 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1726 blend_type.width, chans[j], src_type);
1727 }
1728
1729 /* Insert bits into correct position */
1730 chans[j] = LLVMBuildShl(builder,
1731 chans[j],
1732 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1733 "");
1734
1735 if (j == 0) {
1736 res = chans[j];
1737 } else {
1738 res = LLVMBuildOr(builder, res, chans[j], "");
1739 }
1740 }
1741
1742 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1743 }
1744 }
1745
1746
1747 /**
1748 * Convert from blending format to memory format
1749 *
1750 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1751 */
1752 static void
1753 convert_from_blend_type(struct gallivm_state *gallivm,
1754 unsigned block_size,
1755 const struct util_format_description *src_fmt,
1756 struct lp_type src_type,
1757 struct lp_type dst_type,
1758 LLVMValueRef* src, // and dst
1759 unsigned num_srcs)
1760 {
1761 LLVMValueRef* dst = src;
1762 unsigned i, j, k;
1763 struct lp_type mem_type;
1764 struct lp_type blend_type;
1765 LLVMBuilderRef builder = gallivm->builder;
1766 unsigned pixels = block_size / num_srcs;
1767 bool is_arith;
1768
1769 /*
1770 * full custom path for packed floats and srgb formats - none of the later
1771 * functions would do anything useful, and given the lp_type representation they
1772 * can't be fixed. Should really have some SoA blend path for these kind of
1773 * formats rather than hacking them in here.
1774 */
1775 if (format_expands_to_float_soa(src_fmt)) {
1776 /*
1777 * This is pretty suboptimal for this case blending in SoA would be much
1778 * better - we need to transpose the AoS values back to SoA values for
1779 * conversion/packing.
1780 */
1781 assert(src_type.floating);
1782 assert(src_type.width == 32);
1783 assert(src_type.length % 4 == 0);
1784 assert(dst_type.width == 32 || dst_type.width == 16);
1785
1786 for (i = 0; i < num_srcs / 4; i++) {
1787 LLVMValueRef tmpsoa[4], tmpdst;
1788 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1789 /* really really need SoA here */
1790
1791 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1792 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1793 }
1794 else {
1795 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1796 src_type, tmpsoa);
1797 }
1798
1799 if (src_type.length == 8) {
1800 LLVMValueRef tmpaos, shuffles[8];
1801 unsigned j;
1802 /*
1803 * for 8-wide aos transpose has given us wrong order not matching
1804 * output order. HMPF. Also need to split the output values manually.
1805 */
1806 for (j = 0; j < 4; j++) {
1807 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1808 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1809 }
1810 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1811 LLVMConstVector(shuffles, 8), "");
1812 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1813 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1814 }
1815 else {
1816 src[i] = tmpdst;
1817 }
1818 }
1819 if (dst_type.width == 16) {
1820 struct lp_type type16x8 = dst_type;
1821 struct lp_type type32x4 = dst_type;
1822 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1823 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1824 type16x8.length = 8;
1825 type32x4.width = 32;
1826 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1827 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1828 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1829 /* We could do vector truncation but it doesn't generate very good code */
1830 for (i = 0; i < num_fetch; i++) {
1831 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1832 src[i], lp_build_zero(gallivm, type32x4));
1833 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1834 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1835 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1836 }
1837 }
1838 return;
1839 }
1840
1841 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1842 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1843
1844 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1845
1846 /* Special case for half-floats */
1847 if (mem_type.width == 16 && mem_type.floating) {
1848 int length = dst_type.length;
1849 assert(blend_type.width == 32 && blend_type.floating);
1850
1851 dst_type.length = src_type.length;
1852
1853 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1854
1855 dst_type.length = length;
1856 is_arith = false;
1857 }
1858
1859 /* Remove any padding */
1860 if (!is_arith && (src_type.length % mem_type.length)) {
1861 src_type.length -= (src_type.length % mem_type.length);
1862
1863 for (i = 0; i < num_srcs; ++i) {
1864 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1865 }
1866 }
1867
1868 /* No bit arithmetic to do */
1869 if (!is_arith) {
1870 return;
1871 }
1872
1873 src_type.length = pixels;
1874 src_type.width = blend_type.length * blend_type.width;
1875 dst_type.length = pixels;
1876
1877 for (i = 0; i < num_srcs; ++i) {
1878 LLVMValueRef chans[4];
1879 LLVMValueRef res = NULL;
1880
1881 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1882
1883 for (j = 0; j < src_fmt->nr_channels; ++j) {
1884 unsigned mask = 0;
1885 unsigned sa = src_fmt->channel[j].shift;
1886 unsigned sz_a = src_fmt->channel[j].size;
1887 #if UTIL_ARCH_LITTLE_ENDIAN
1888 unsigned from_lsb = j;
1889 #else
1890 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1891 #endif
1892
1893 assert(blend_type.width > src_fmt->channel[j].size);
1894
1895 for (k = 0; k < blend_type.width; ++k) {
1896 mask |= 1 << k;
1897 }
1898
1899 /* Extract bits */
1900 chans[j] = LLVMBuildLShr(builder,
1901 dst[i],
1902 lp_build_const_int_vec(gallivm, src_type,
1903 from_lsb * blend_type.width),
1904 "");
1905
1906 chans[j] = LLVMBuildAnd(builder,
1907 chans[j],
1908 lp_build_const_int_vec(gallivm, src_type, mask),
1909 "");
1910
1911 /* Scale down bits */
1912 if (src_type.norm) {
1913 chans[j] = scale_bits(gallivm, blend_type.width,
1914 src_fmt->channel[j].size, chans[j], src_type);
1915 } else if (!src_type.floating && sz_a < blend_type.width) {
1916 LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
1917 LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
1918 chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
1919 }
1920
1921 /* Insert bits */
1922 chans[j] = LLVMBuildShl(builder,
1923 chans[j],
1924 lp_build_const_int_vec(gallivm, src_type, sa),
1925 "");
1926
1927 sa += src_fmt->channel[j].size;
1928
1929 if (j == 0) {
1930 res = chans[j];
1931 } else {
1932 res = LLVMBuildOr(builder, res, chans[j], "");
1933 }
1934 }
1935
1936 assert (dst_type.width != 24);
1937
1938 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1939 }
1940 }
1941
1942
1943 /**
1944 * Convert alpha to same blend type as src
1945 */
1946 static void
1947 convert_alpha(struct gallivm_state *gallivm,
1948 struct lp_type row_type,
1949 struct lp_type alpha_type,
1950 const unsigned block_size,
1951 const unsigned block_height,
1952 const unsigned src_count,
1953 const unsigned dst_channels,
1954 const bool pad_inline,
1955 LLVMValueRef* src_alpha)
1956 {
1957 LLVMBuilderRef builder = gallivm->builder;
1958 unsigned i, j;
1959 unsigned length = row_type.length;
1960 row_type.length = alpha_type.length;
1961
1962 /* Twiddle the alpha to match pixels */
1963 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1964
1965 /*
1966 * TODO this should use single lp_build_conv call for
1967 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1968 */
1969 for (i = 0; i < block_height; ++i) {
1970 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1971 }
1972
1973 alpha_type = row_type;
1974 row_type.length = length;
1975
1976 /* If only one channel we can only need the single alpha value per pixel */
1977 if (src_count == 1 && dst_channels == 1) {
1978
1979 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1980 } else {
1981 /* If there are more srcs than rows then we need to split alpha up */
1982 if (src_count > block_height) {
1983 for (i = src_count; i > 0; --i) {
1984 unsigned pixels = block_size / src_count;
1985 unsigned idx = i - 1;
1986
1987 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1988 (idx * pixels) % 4, pixels);
1989 }
1990 }
1991
1992 /* If there is a src for each pixel broadcast the alpha across whole row */
1993 if (src_count == block_size) {
1994 for (i = 0; i < src_count; ++i) {
1995 src_alpha[i] = lp_build_broadcast(gallivm,
1996 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1997 }
1998 } else {
1999 unsigned pixels = block_size / src_count;
2000 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
2001 unsigned alpha_span = 1;
2002 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
2003
2004 /* Check if we need 2 src_alphas for our shuffles */
2005 if (pixels > alpha_type.length) {
2006 alpha_span = 2;
2007 }
2008
2009 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
2010 for (j = 0; j < row_type.length; ++j) {
2011 if (j < pixels * channels) {
2012 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
2013 } else {
2014 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
2015 }
2016 }
2017
2018 for (i = 0; i < src_count; ++i) {
2019 unsigned idx1 = i, idx2 = i;
2020
2021 if (alpha_span > 1){
2022 idx1 *= alpha_span;
2023 idx2 = idx1 + 1;
2024 }
2025
2026 src_alpha[i] = LLVMBuildShuffleVector(builder,
2027 src_alpha[idx1],
2028 src_alpha[idx2],
2029 LLVMConstVector(shuffles, row_type.length),
2030 "");
2031 }
2032 }
2033 }
2034 }
2035
2036
2037 /**
2038 * Generates the blend function for unswizzled colour buffers
2039 * Also generates the read & write from colour buffer
2040 */
2041 static void
2042 generate_unswizzled_blend(struct gallivm_state *gallivm,
2043 unsigned rt,
2044 struct lp_fragment_shader_variant *variant,
2045 enum pipe_format out_format,
2046 unsigned int num_fs,
2047 struct lp_type fs_type,
2048 LLVMValueRef* fs_mask,
2049 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
2050 LLVMValueRef context_ptr,
2051 LLVMValueRef color_ptr,
2052 LLVMValueRef stride,
2053 unsigned partial_mask,
2054 boolean do_branch)
2055 {
2056 const unsigned alpha_channel = 3;
2057 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
2058 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
2059 const unsigned block_size = block_width * block_height;
2060 const unsigned lp_integer_vector_width = 128;
2061
2062 LLVMBuilderRef builder = gallivm->builder;
2063 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
2064 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
2065 LLVMValueRef src_alpha[4 * 4];
2066 LLVMValueRef src1_alpha[4 * 4] = { NULL };
2067 LLVMValueRef src_mask[4 * 4];
2068 LLVMValueRef src[4 * 4];
2069 LLVMValueRef src1[4 * 4];
2070 LLVMValueRef dst[4 * 4];
2071 LLVMValueRef blend_color;
2072 LLVMValueRef blend_alpha;
2073 LLVMValueRef i32_zero;
2074 LLVMValueRef check_mask;
2075 LLVMValueRef undef_src_val;
2076
2077 struct lp_build_mask_context mask_ctx;
2078 struct lp_type mask_type;
2079 struct lp_type blend_type;
2080 struct lp_type row_type;
2081 struct lp_type dst_type;
2082 struct lp_type ls_type;
2083
2084 unsigned char swizzle[TGSI_NUM_CHANNELS];
2085 unsigned vector_width;
2086 unsigned src_channels = TGSI_NUM_CHANNELS;
2087 unsigned dst_channels;
2088 unsigned dst_count;
2089 unsigned src_count;
2090 unsigned i, j;
2091
2092 const struct util_format_description* out_format_desc = util_format_description(out_format);
2093
2094 unsigned dst_alignment;
2095
2096 bool pad_inline = is_arithmetic_format(out_format_desc);
2097 bool has_alpha = false;
2098 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2099 util_blend_state_is_dual(&variant->key.blend, 0);
2100
2101 const boolean is_1d = variant->key.resource_1d;
2102 boolean twiddle_after_convert = FALSE;
2103 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2104 LLVMValueRef fpstate = 0;
2105
2106 /* Get type from output format */
2107 lp_blend_type_from_format_desc(out_format_desc, &row_type);
2108 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2109
2110 /*
2111 * Technically this code should go into lp_build_smallfloat_to_float
2112 * and lp_build_float_to_smallfloat but due to the
2113 * http://llvm.org/bugs/show_bug.cgi?id=6393
2114 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2115 * So the ordering is important here and there shouldn't be any
2116 * llvm ir instrunctions in this function before
2117 * this, otherwise half-float format conversions won't work
2118 * (again due to llvm bug #6393).
2119 */
2120 if (have_smallfloat_format(dst_type, out_format)) {
2121 /* We need to make sure that denorms are ok for half float
2122 conversions */
2123 fpstate = lp_build_fpstate_get(gallivm);
2124 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2125 }
2126
2127 mask_type = lp_int32_vec4_type();
2128 mask_type.length = fs_type.length;
2129
2130 for (i = num_fs; i < num_fullblock_fs; i++) {
2131 fs_mask[i] = lp_build_zero(gallivm, mask_type);
2132 }
2133
2134 /* Do not bother executing code when mask is empty.. */
2135 if (do_branch) {
2136 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2137
2138 for (i = 0; i < num_fullblock_fs; ++i) {
2139 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2140 }
2141
2142 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2143 lp_build_mask_check(&mask_ctx);
2144 }
2145
2146 partial_mask |= !variant->opaque;
2147 i32_zero = lp_build_const_int32(gallivm, 0);
2148
2149 undef_src_val = lp_build_undef(gallivm, fs_type);
2150
2151 row_type.length = fs_type.length;
2152 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2153
2154 /* Compute correct swizzle and count channels */
2155 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2156 dst_channels = 0;
2157
2158 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2159 /* Ensure channel is used */
2160 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2161 continue;
2162 }
2163
2164 /* Ensure not already written to (happens in case with GL_ALPHA) */
2165 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2166 continue;
2167 }
2168
2169 /* Ensure we havn't already found all channels */
2170 if (dst_channels >= out_format_desc->nr_channels) {
2171 continue;
2172 }
2173
2174 swizzle[out_format_desc->swizzle[i]] = i;
2175 ++dst_channels;
2176
2177 if (i == alpha_channel) {
2178 has_alpha = true;
2179 }
2180 }
2181
2182 if (format_expands_to_float_soa(out_format_desc)) {
2183 /*
2184 * the code above can't work for layout_other
2185 * for srgb it would sort of work but we short-circuit swizzles, etc.
2186 * as that is done as part of unpack / pack.
2187 */
2188 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2189 has_alpha = true;
2190 swizzle[0] = 0;
2191 swizzle[1] = 1;
2192 swizzle[2] = 2;
2193 swizzle[3] = 3;
2194 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2195 }
2196
2197 /* If 3 channels then pad to include alpha for 4 element transpose */
2198 if (dst_channels == 3) {
2199 assert (!has_alpha);
2200 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2201 if (swizzle[i] > TGSI_NUM_CHANNELS)
2202 swizzle[i] = 3;
2203 }
2204 if (out_format_desc->nr_channels == 4) {
2205 dst_channels = 4;
2206 /*
2207 * We use alpha from the color conversion, not separate one.
2208 * We had to include it for transpose, hence it will get converted
2209 * too (albeit when doing transpose after conversion, that would
2210 * no longer be the case necessarily).
2211 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2212 * otherwise we really have padding, not alpha, included.)
2213 */
2214 has_alpha = true;
2215 }
2216 }
2217
2218 /*
2219 * Load shader output
2220 */
2221 for (i = 0; i < num_fullblock_fs; ++i) {
2222 /* Always load alpha for use in blending */
2223 LLVMValueRef alpha;
2224 if (i < num_fs) {
2225 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2226 }
2227 else {
2228 alpha = undef_src_val;
2229 }
2230
2231 /* Load each channel */
2232 for (j = 0; j < dst_channels; ++j) {
2233 assert(swizzle[j] < 4);
2234 if (i < num_fs) {
2235 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2236 }
2237 else {
2238 fs_src[i][j] = undef_src_val;
2239 }
2240 }
2241
2242 /* If 3 channels then pad to include alpha for 4 element transpose */
2243 /*
2244 * XXX If we include that here maybe could actually use it instead of
2245 * separate alpha for blending?
2246 * (Difficult though we actually convert pad channels, not alpha.)
2247 */
2248 if (dst_channels == 3 && !has_alpha) {
2249 fs_src[i][3] = alpha;
2250 }
2251
2252 /* We split the row_mask and row_alpha as we want 128bit interleave */
2253 if (fs_type.length == 8) {
2254 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
2255 0, src_channels);
2256 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
2257 src_channels, src_channels);
2258
2259 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2260 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2261 src_channels, src_channels);
2262 } else {
2263 src_mask[i] = fs_mask[i];
2264 src_alpha[i] = alpha;
2265 }
2266 }
2267 if (dual_source_blend) {
2268 /* same as above except different src/dst, skip masks and comments... */
2269 for (i = 0; i < num_fullblock_fs; ++i) {
2270 LLVMValueRef alpha;
2271 if (i < num_fs) {
2272 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2273 }
2274 else {
2275 alpha = undef_src_val;
2276 }
2277
2278 for (j = 0; j < dst_channels; ++j) {
2279 assert(swizzle[j] < 4);
2280 if (i < num_fs) {
2281 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2282 }
2283 else {
2284 fs_src1[i][j] = undef_src_val;
2285 }
2286 }
2287 if (dst_channels == 3 && !has_alpha) {
2288 fs_src1[i][3] = alpha;
2289 }
2290 if (fs_type.length == 8) {
2291 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2292 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2293 src_channels, src_channels);
2294 } else {
2295 src1_alpha[i] = alpha;
2296 }
2297 }
2298 }
2299
2300 if (util_format_is_pure_integer(out_format)) {
2301 /*
2302 * In this case fs_type was really ints or uints disguised as floats,
2303 * fix that up now.
2304 */
2305 fs_type.floating = 0;
2306 fs_type.sign = dst_type.sign;
2307 for (i = 0; i < num_fullblock_fs; ++i) {
2308 for (j = 0; j < dst_channels; ++j) {
2309 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2310 lp_build_vec_type(gallivm, fs_type), "");
2311 }
2312 if (dst_channels == 3 && !has_alpha) {
2313 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2314 lp_build_vec_type(gallivm, fs_type), "");
2315 }
2316 }
2317 }
2318
2319 /*
2320 * We actually should generally do conversion first (for non-1d cases)
2321 * when the blend format is 8 or 16 bits. The reason is obvious,
2322 * there's 2 or 4 times less vectors to deal with for the interleave...
2323 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2324 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2325 * unpack only with 128bit vectors).
2326 * Note: for 16bit sizes really need matching pack conversion code
2327 */
2328 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2329 twiddle_after_convert = TRUE;
2330 }
2331
2332 /*
2333 * Pixel twiddle from fragment shader order to memory order
2334 */
2335 if (!twiddle_after_convert) {
2336 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2337 dst_channels, fs_src, src, pad_inline);
2338 if (dual_source_blend) {
2339 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2340 fs_src1, src1, pad_inline);
2341 }
2342 } else {
2343 src_count = num_fullblock_fs * dst_channels;
2344 /*
2345 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2346 * (AVX) turn out the same later when untwiddling/transpose (albeit
2347 * for true AVX2 path untwiddle needs to be different).
2348 * For now just order by colors first (so we can use unpack later).
2349 */
2350 for (j = 0; j < num_fullblock_fs; j++) {
2351 for (i = 0; i < dst_channels; i++) {
2352 src[i*num_fullblock_fs + j] = fs_src[j][i];
2353 if (dual_source_blend) {
2354 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2355 }
2356 }
2357 }
2358 }
2359
2360 src_channels = dst_channels < 3 ? dst_channels : 4;
2361 if (src_count != num_fullblock_fs * src_channels) {
2362 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2363 row_type.length /= ds;
2364 fs_type.length = row_type.length;
2365 }
2366
2367 blend_type = row_type;
2368 mask_type.length = 4;
2369
2370 /* Convert src to row_type */
2371 if (dual_source_blend) {
2372 struct lp_type old_row_type = row_type;
2373 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2374 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2375 }
2376 else {
2377 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2378 }
2379
2380 /* If the rows are not an SSE vector, combine them to become SSE size! */
2381 if ((row_type.width * row_type.length) % 128) {
2382 unsigned bits = row_type.width * row_type.length;
2383 unsigned combined;
2384
2385 assert(src_count >= (vector_width / bits));
2386
2387 dst_count = src_count / (vector_width / bits);
2388
2389 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2390 if (dual_source_blend) {
2391 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2392 }
2393
2394 row_type.length *= combined;
2395 src_count /= combined;
2396
2397 bits = row_type.width * row_type.length;
2398 assert(bits == 128 || bits == 256);
2399 }
2400
2401 if (twiddle_after_convert) {
2402 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2403 if (dual_source_blend) {
2404 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2405 }
2406 }
2407
2408 /*
2409 * Blend Colour conversion
2410 */
2411 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2412 blend_color = LLVMBuildPointerCast(builder, blend_color,
2413 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2414 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2415 &i32_zero, 1, ""), "");
2416
2417 /* Convert */
2418 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2419
2420 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2421 /*
2422 * since blending is done with floats, there was no conversion.
2423 * However, the rules according to fixed point renderbuffers still
2424 * apply, that is we must clamp inputs to 0.0/1.0.
2425 * (This would apply to separate alpha conversion too but we currently
2426 * force has_alpha to be true.)
2427 * TODO: should skip this with "fake" blend, since post-blend conversion
2428 * will clamp anyway.
2429 * TODO: could also skip this if fragment color clamping is enabled. We
2430 * don't support it natively so it gets baked into the shader however, so
2431 * can't really tell here.
2432 */
2433 struct lp_build_context f32_bld;
2434 assert(row_type.floating);
2435 lp_build_context_init(&f32_bld, gallivm, row_type);
2436 for (i = 0; i < src_count; i++) {
2437 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2438 }
2439 if (dual_source_blend) {
2440 for (i = 0; i < src_count; i++) {
2441 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2442 }
2443 }
2444 /* probably can't be different than row_type but better safe than sorry... */
2445 lp_build_context_init(&f32_bld, gallivm, blend_type);
2446 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2447 }
2448
2449 /* Extract alpha */
2450 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2451
2452 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2453 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2454 if (pad_inline) {
2455 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2456 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2457 } else {
2458 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2459 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2460 }
2461
2462 /*
2463 * Mask conversion
2464 */
2465 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2466
2467 if (src_count < block_height) {
2468 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2469 } else if (src_count > block_height) {
2470 for (i = src_count; i > 0; --i) {
2471 unsigned pixels = block_size / src_count;
2472 unsigned idx = i - 1;
2473
2474 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2475 (idx * pixels) % 4, pixels);
2476 }
2477 }
2478
2479 assert(mask_type.width == 32);
2480
2481 for (i = 0; i < src_count; ++i) {
2482 unsigned pixels = block_size / src_count;
2483 unsigned pixel_width = row_type.width * dst_channels;
2484
2485 if (pixel_width == 24) {
2486 mask_type.width = 8;
2487 mask_type.length = vector_width / mask_type.width;
2488 } else {
2489 mask_type.length = pixels;
2490 mask_type.width = row_type.width * dst_channels;
2491
2492 /*
2493 * If mask_type width is smaller than 32bit, this doesn't quite
2494 * generate the most efficient code (could use some pack).
2495 */
2496 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2497 lp_build_int_vec_type(gallivm, mask_type), "");
2498
2499 mask_type.length *= dst_channels;
2500 mask_type.width /= dst_channels;
2501 }
2502
2503 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2504 lp_build_int_vec_type(gallivm, mask_type), "");
2505 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2506 }
2507
2508 /*
2509 * Alpha conversion
2510 */
2511 if (!has_alpha) {
2512 struct lp_type alpha_type = fs_type;
2513 alpha_type.length = 4;
2514 convert_alpha(gallivm, row_type, alpha_type,
2515 block_size, block_height,
2516 src_count, dst_channels,
2517 pad_inline, src_alpha);
2518 if (dual_source_blend) {
2519 convert_alpha(gallivm, row_type, alpha_type,
2520 block_size, block_height,
2521 src_count, dst_channels,
2522 pad_inline, src1_alpha);
2523 }
2524 }
2525
2526
2527 /*
2528 * Load dst from memory
2529 */
2530 if (src_count < block_height) {
2531 dst_count = block_height;
2532 } else {
2533 dst_count = src_count;
2534 }
2535
2536 dst_type.length *= block_size / dst_count;
2537
2538 if (format_expands_to_float_soa(out_format_desc)) {
2539 /*
2540 * we need multiple values at once for the conversion, so can as well
2541 * load them vectorized here too instead of concatenating later.
2542 * (Still need concatenation later for 8-wide vectors).
2543 */
2544 dst_count = block_height;
2545 dst_type.length = block_width;
2546 }
2547
2548 /*
2549 * Compute the alignment of the destination pointer in bytes
2550 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2551 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2552 * 1d tex but can't distinguish here) so need to stick with per-pixel
2553 * alignment in this case.
2554 */
2555 if (is_1d) {
2556 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2557 }
2558 else {
2559 dst_alignment = dst_type.length * dst_type.width / 8;
2560 }
2561 /* Force power-of-two alignment by extracting only the least-significant-bit */
2562 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2563 /*
2564 * Resource base and stride pointers are aligned to 16 bytes, so that's
2565 * the maximum alignment we can guarantee
2566 */
2567 dst_alignment = MIN2(16, dst_alignment);
2568
2569 ls_type = dst_type;
2570
2571 if (dst_count > src_count) {
2572 if ((dst_type.width == 8 || dst_type.width == 16) &&
2573 util_is_power_of_two_or_zero(dst_type.length) &&
2574 dst_type.length * dst_type.width < 128) {
2575 /*
2576 * Never try to load values as 4xi8 which we will then
2577 * concatenate to larger vectors. This gives llvm a real
2578 * headache (the problem is the type legalizer (?) will
2579 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2580 * then the shuffles to concatenate are more or less impossible
2581 * - llvm is easily capable of generating a sequence of 32
2582 * pextrb/pinsrb instructions for that. Albeit it appears to
2583 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2584 * width to avoid the trouble (16bit seems not as bad, llvm
2585 * probably recognizes the load+shuffle as only one shuffle
2586 * is necessary, but we can do just the same anyway).
2587 */
2588 ls_type.length = dst_type.length * dst_type.width / 32;
2589 ls_type.width = 32;
2590 }
2591 }
2592
2593 if (is_1d) {
2594 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2595 dst, ls_type, dst_count / 4, dst_alignment);
2596 for (i = dst_count / 4; i < dst_count; i++) {
2597 dst[i] = lp_build_undef(gallivm, ls_type);
2598 }
2599
2600 }
2601 else {
2602 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2603 dst, ls_type, dst_count, dst_alignment);
2604 }
2605
2606
2607 /*
2608 * Convert from dst/output format to src/blending format.
2609 *
2610 * This is necessary as we can only read 1 row from memory at a time,
2611 * so the minimum dst_count will ever be at this point is 4.
2612 *
2613 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2614 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2615 * on all 16 pixels in that single vector at once.
2616 */
2617 if (dst_count > src_count) {
2618 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2619 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2620 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2621 for (i = 0; i < dst_count; i++) {
2622 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2623 }
2624 }
2625
2626 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2627
2628 if (ls_type.length != dst_type.length) {
2629 struct lp_type tmp_type = dst_type;
2630 tmp_type.length = dst_type.length * 4 / src_count;
2631 for (i = 0; i < src_count; i++) {
2632 dst[i] = LLVMBuildBitCast(builder, dst[i],
2633 lp_build_vec_type(gallivm, tmp_type), "");
2634 }
2635 }
2636 }
2637
2638 /*
2639 * Blending
2640 */
2641 /* XXX this is broken for RGB8 formats -
2642 * they get expanded from 12 to 16 elements (to include alpha)
2643 * by convert_to_blend_type then reduced to 15 instead of 12
2644 * by convert_from_blend_type (a simple fix though breaks A8...).
2645 * R16G16B16 also crashes differently however something going wrong
2646 * inside llvm handling npot vector sizes seemingly.
2647 * It seems some cleanup could be done here (like skipping conversion/blend
2648 * when not needed).
2649 */
2650 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2651 row_type, dst, src_count);
2652
2653 /*
2654 * FIXME: Really should get logic ops / masks out of generic blend / row
2655 * format. Logic ops will definitely not work on the blend float format
2656 * used for SRGB here and I think OpenGL expects this to work as expected
2657 * (that is incoming values converted to srgb then logic op applied).
2658 */
2659 for (i = 0; i < src_count; ++i) {
2660 dst[i] = lp_build_blend_aos(gallivm,
2661 &variant->key.blend,
2662 out_format,
2663 row_type,
2664 rt,
2665 src[i],
2666 has_alpha ? NULL : src_alpha[i],
2667 src1[i],
2668 has_alpha ? NULL : src1_alpha[i],
2669 dst[i],
2670 partial_mask ? src_mask[i] : NULL,
2671 blend_color,
2672 has_alpha ? NULL : blend_alpha,
2673 swizzle,
2674 pad_inline ? 4 : dst_channels);
2675 }
2676
2677 convert_from_blend_type(gallivm, block_size, out_format_desc,
2678 row_type, dst_type, dst, src_count);
2679
2680 /* Split the blend rows back to memory rows */
2681 if (dst_count > src_count) {
2682 row_type.length = dst_type.length * (dst_count / src_count);
2683
2684 if (src_count == 1) {
2685 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2686 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2687
2688 row_type.length /= 2;
2689 src_count *= 2;
2690 }
2691
2692 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2693 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2694 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2695 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2696
2697 row_type.length /= 2;
2698 src_count *= 2;
2699 }
2700
2701 /*
2702 * Store blend result to memory
2703 */
2704 if (is_1d) {
2705 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2706 dst, dst_type, dst_count / 4, dst_alignment);
2707 }
2708 else {
2709 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2710 dst, dst_type, dst_count, dst_alignment);
2711 }
2712
2713 if (have_smallfloat_format(dst_type, out_format)) {
2714 lp_build_fpstate_set(gallivm, fpstate);
2715 }
2716
2717 if (do_branch) {
2718 lp_build_mask_end(&mask_ctx);
2719 }
2720 }
2721
2722
2723 /**
2724 * Generate the runtime callable function for the whole fragment pipeline.
2725 * Note that the function which we generate operates on a block of 16
2726 * pixels at at time. The block contains 2x2 quads. Each quad contains
2727 * 2x2 pixels.
2728 */
2729 static void
2730 generate_fragment(struct llvmpipe_context *lp,
2731 struct lp_fragment_shader *shader,
2732 struct lp_fragment_shader_variant *variant,
2733 unsigned partial_mask)
2734 {
2735 struct gallivm_state *gallivm = variant->gallivm;
2736 struct lp_fragment_shader_variant_key *key = &variant->key;
2737 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2738 char func_name[64];
2739 struct lp_type fs_type;
2740 struct lp_type blend_type;
2741 LLVMTypeRef fs_elem_type;
2742 LLVMTypeRef blend_vec_type;
2743 LLVMTypeRef arg_types[15];
2744 LLVMTypeRef func_type;
2745 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2746 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2747 LLVMValueRef context_ptr;
2748 LLVMValueRef x;
2749 LLVMValueRef y;
2750 LLVMValueRef a0_ptr;
2751 LLVMValueRef dadx_ptr;
2752 LLVMValueRef dady_ptr;
2753 LLVMValueRef color_ptr_ptr;
2754 LLVMValueRef stride_ptr;
2755 LLVMValueRef color_sample_stride_ptr;
2756 LLVMValueRef depth_ptr;
2757 LLVMValueRef depth_stride;
2758 LLVMValueRef depth_sample_stride;
2759 LLVMValueRef mask_input;
2760 LLVMValueRef thread_data_ptr;
2761 LLVMBasicBlockRef block;
2762 LLVMBuilderRef builder;
2763 struct lp_build_sampler_soa *sampler;
2764 struct lp_build_image_soa *image;
2765 struct lp_build_interp_soa_context interp;
2766 LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
2767 LLVMValueRef fs_out_color[LP_MAX_SAMPLES][PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2768 LLVMValueRef function;
2769 LLVMValueRef facing;
2770 unsigned num_fs;
2771 unsigned i;
2772 unsigned chan;
2773 unsigned cbuf;
2774 boolean cbuf0_write_all;
2775 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2776 util_blend_state_is_dual(&key->blend, 0);
2777
2778 assert(lp_native_vector_width / 32 >= 4);
2779
2780 /* Adjust color input interpolation according to flatshade state:
2781 */
2782 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2783 for (i = 0; i < shader->info.base.num_inputs; i++) {
2784 if (inputs[i].interp == LP_INTERP_COLOR) {
2785 if (key->flatshade)
2786 inputs[i].interp = LP_INTERP_CONSTANT;
2787 else
2788 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2789 }
2790 }
2791
2792 /* check if writes to cbuf[0] are to be copied to all cbufs */
2793 cbuf0_write_all =
2794 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2795
2796 /* TODO: actually pick these based on the fs and color buffer
2797 * characteristics. */
2798
2799 memset(&fs_type, 0, sizeof fs_type);
2800 fs_type.floating = TRUE; /* floating point values */
2801 fs_type.sign = TRUE; /* values are signed */
2802 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2803 fs_type.width = 32; /* 32-bit float */
2804 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2805
2806 memset(&blend_type, 0, sizeof blend_type);
2807 blend_type.floating = FALSE; /* values are integers */
2808 blend_type.sign = FALSE; /* values are unsigned */
2809 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2810 blend_type.width = 8; /* 8-bit ubyte values */
2811 blend_type.length = 16; /* 16 elements per vector */
2812
2813 /*
2814 * Generate the function prototype. Any change here must be reflected in
2815 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2816 */
2817
2818 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2819
2820 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2821
2822 snprintf(func_name, sizeof(func_name), "fs_variant_%s",
2823 partial_mask ? "partial" : "whole");
2824
2825 arg_types[0] = variant->jit_context_ptr_type; /* context */
2826 arg_types[1] = int32_type; /* x */
2827 arg_types[2] = int32_type; /* y */
2828 arg_types[3] = int32_type; /* facing */
2829 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2830 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2831 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2832 arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
2833 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2834 arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
2835 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2836 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2837 arg_types[12] = int32_type; /* depth_stride */
2838 arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
2839 arg_types[14] = int32_type; /* depth sample stride */
2840
2841 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2842 arg_types, ARRAY_SIZE(arg_types), 0);
2843
2844 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2845 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2846
2847 variant->function[partial_mask] = function;
2848
2849 /* XXX: need to propagate noalias down into color param now we are
2850 * passing a pointer-to-pointer?
2851 */
2852 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2853 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2854 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2855
2856 if (variant->gallivm->cache->data_size)
2857 return;
2858
2859 context_ptr = LLVMGetParam(function, 0);
2860 x = LLVMGetParam(function, 1);
2861 y = LLVMGetParam(function, 2);
2862 facing = LLVMGetParam(function, 3);
2863 a0_ptr = LLVMGetParam(function, 4);
2864 dadx_ptr = LLVMGetParam(function, 5);
2865 dady_ptr = LLVMGetParam(function, 6);
2866 color_ptr_ptr = LLVMGetParam(function, 7);
2867 depth_ptr = LLVMGetParam(function, 8);
2868 mask_input = LLVMGetParam(function, 9);
2869 thread_data_ptr = LLVMGetParam(function, 10);
2870 stride_ptr = LLVMGetParam(function, 11);
2871 depth_stride = LLVMGetParam(function, 12);
2872 color_sample_stride_ptr = LLVMGetParam(function, 13);
2873 depth_sample_stride = LLVMGetParam(function, 14);
2874
2875 lp_build_name(context_ptr, "context");
2876 lp_build_name(x, "x");
2877 lp_build_name(y, "y");
2878 lp_build_name(a0_ptr, "a0");
2879 lp_build_name(dadx_ptr, "dadx");
2880 lp_build_name(dady_ptr, "dady");
2881 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2882 lp_build_name(depth_ptr, "depth");
2883 lp_build_name(mask_input, "mask_input");
2884 lp_build_name(thread_data_ptr, "thread_data");
2885 lp_build_name(stride_ptr, "stride_ptr");
2886 lp_build_name(depth_stride, "depth_stride");
2887 lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
2888 lp_build_name(depth_sample_stride, "depth_sample_stride");
2889
2890 /*
2891 * Function body
2892 */
2893
2894 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2895 builder = gallivm->builder;
2896 assert(builder);
2897 LLVMPositionBuilderAtEnd(builder, block);
2898
2899 /*
2900 * Must not count ps invocations if there's a null shader.
2901 * (It would be ok to count with null shader if there's d/s tests,
2902 * but only if there's d/s buffers too, which is different
2903 * to implicit rasterization disable which must not depend
2904 * on the d/s buffers.)
2905 * Could use popcount on mask, but pixel accuracy is not required.
2906 * Could disable if there's no stats query, but maybe not worth it.
2907 */
2908 if (shader->info.base.num_instructions > 1) {
2909 LLVMValueRef invocs, val;
2910 invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
2911 val = LLVMBuildLoad(builder, invocs, "");
2912 val = LLVMBuildAdd(builder, val,
2913 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
2914 "invoc_count");
2915 LLVMBuildStore(builder, val, invocs);
2916 }
2917
2918 /* code generated texture sampling */
2919 sampler = lp_llvm_sampler_soa_create(key->samplers);
2920 image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
2921
2922 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2923 /* for 1d resources only run "upper half" of stamp */
2924 if (key->resource_1d)
2925 num_fs /= 2;
2926
2927 {
2928 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2929 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2930 LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
2931 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2932 num_loop_samp, "mask_store");
2933
2934 LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
2935 LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, LLVMArrayType(flt_type, key->coverage_samples * 2), "");
2936 LLVMValueRef sample_pos_array;
2937
2938 if (key->multisample && key->coverage_samples == 4) {
2939 LLVMValueRef sample_pos_arr[8];
2940 for (unsigned i = 0; i < 4; i++) {
2941 sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
2942 sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
2943 }
2944 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
2945 } else {
2946 LLVMValueRef sample_pos_arr[2];
2947 sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
2948 sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
2949 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
2950 }
2951 LLVMSetInitializer(glob_sample_pos, sample_pos_array);
2952
2953 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2954 boolean pixel_center_integer =
2955 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2956
2957 /*
2958 * The shader input interpolation info is not explicitely baked in the
2959 * shader key, but everything it derives from (TGSI, and flatshade) is
2960 * already included in the shader key.
2961 */
2962 lp_build_interp_soa_init(&interp,
2963 gallivm,
2964 shader->info.base.num_inputs,
2965 inputs,
2966 pixel_center_integer,
2967 key->coverage_samples, glob_sample_pos,
2968 num_loop,
2969 key->depth_clamp,
2970 builder, fs_type,
2971 a0_ptr, dadx_ptr, dady_ptr,
2972 x, y);
2973
2974 for (i = 0; i < num_fs; i++) {
2975 if (key->multisample) {
2976 LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
2977
2978 /*
2979 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2980 * store to the per sample mask storage. Or all of them together to generate
2981 * the fragment shader mask. (sample shading TODO).
2982 * Take the incoming state coverage mask into account.
2983 */
2984 for (unsigned s = 0; s < key->coverage_samples; s++) {
2985 LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
2986 LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
2987 &sindexi, 1, "sample_mask_ptr");
2988 LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
2989 i*fs_type.length/4, s, mask_input);
2990
2991 LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
2992 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
2993 smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
2994 smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
2995
2996 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
2997 LLVMBuildStore(builder, s_mask, sample_mask_ptr);
2998 }
2999 } else {
3000 LLVMValueRef mask;
3001 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
3002 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
3003 &indexi, 1, "mask_ptr");
3004
3005 if (partial_mask) {
3006 mask = generate_quad_mask(gallivm, fs_type,
3007 i*fs_type.length/4, 0, mask_input);
3008 }
3009 else {
3010 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
3011 }
3012 LLVMBuildStore(builder, mask, mask_ptr);
3013 }
3014 }
3015
3016 generate_fs_loop(gallivm,
3017 shader, key,
3018 builder,
3019 fs_type,
3020 context_ptr,
3021 glob_sample_pos,
3022 num_loop,
3023 &interp,
3024 sampler,
3025 image,
3026 mask_store, /* output */
3027 color_store,
3028 depth_ptr,
3029 depth_stride,
3030 depth_sample_stride,
3031 facing,
3032 thread_data_ptr);
3033
3034 for (i = 0; i < num_fs; i++) {
3035 LLVMValueRef ptr;
3036 for (unsigned s = 0; s < key->coverage_samples; s++) {
3037 int idx = (i + (s * num_fs));
3038 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
3039 ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
3040
3041 fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
3042 }
3043
3044 for (unsigned s = 0; s < key->min_samples; s++) {
3045 /* This is fucked up need to reorganize things */
3046 int idx = s * num_fs + i;
3047 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
3048 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
3049 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3050 ptr = LLVMBuildGEP(builder,
3051 color_store[cbuf * !cbuf0_write_all][chan],
3052 &sindexi, 1, "");
3053 fs_out_color[s][cbuf][chan][i] = ptr;
3054 }
3055 }
3056 if (dual_source_blend) {
3057 /* only support one dual source blend target hence always use output 1 */
3058 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3059 ptr = LLVMBuildGEP(builder,
3060 color_store[1][chan],
3061 &sindexi, 1, "");
3062 fs_out_color[s][1][chan][i] = ptr;
3063 }
3064 }
3065 }
3066 }
3067 }
3068
3069 sampler->destroy(sampler);
3070 image->destroy(image);
3071 /* Loop over color outputs / color buffers to do blending.
3072 */
3073 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
3074 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
3075 LLVMValueRef color_ptr;
3076 LLVMValueRef stride;
3077 LLVMValueRef sample_stride = NULL;
3078 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
3079
3080 boolean do_branch = ((key->depth.enabled
3081 || key->stencil[0].enabled
3082 || key->alpha.enabled)
3083 && !shader->info.base.uses_kill);
3084
3085 color_ptr = LLVMBuildLoad(builder,
3086 LLVMBuildGEP(builder, color_ptr_ptr,
3087 &index, 1, ""),
3088 "");
3089
3090 stride = LLVMBuildLoad(builder,
3091 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
3092 "");
3093
3094 if (key->multisample)
3095 sample_stride = LLVMBuildLoad(builder,
3096 LLVMBuildGEP(builder, color_sample_stride_ptr,
3097 &index, 1, ""), "");
3098
3099 for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3100 unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3101 unsigned out_idx = key->min_samples == 1 ? 0 : s;
3102 LLVMValueRef out_ptr = color_ptr;;
3103
3104 if (key->multisample) {
3105 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3106 out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3107 }
3108 out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3109
3110 lp_build_name(out_ptr, "color_ptr%d", cbuf);
3111
3112 generate_unswizzled_blend(gallivm, cbuf, variant,
3113 key->cbuf_format[cbuf],
3114 num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],
3115 context_ptr, out_ptr, stride,
3116 partial_mask, do_branch);
3117 }
3118 }
3119 }
3120
3121 LLVMBuildRetVoid(builder);
3122
3123 gallivm_verify_function(gallivm, function);
3124 }
3125
3126
3127 static void
3128 dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3129 {
3130 unsigned i;
3131
3132 debug_printf("fs variant %p:\n", (void *) key);
3133
3134 if (key->flatshade) {
3135 debug_printf("flatshade = 1\n");
3136 }
3137 if (key->multisample) {
3138 debug_printf("multisample = 1\n");
3139 debug_printf("coverage samples = %d\n", key->coverage_samples);
3140 debug_printf("min samples = %d\n", key->min_samples);
3141 }
3142 for (i = 0; i < key->nr_cbufs; ++i) {
3143 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3144 debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3145 }
3146 if (key->depth.enabled || key->stencil[0].enabled) {
3147 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3148 debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3149 }
3150 if (key->depth.enabled) {
3151 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3152 debug_printf("depth.writemask = %u\n", key->depth.writemask);
3153 }
3154
3155 for (i = 0; i < 2; ++i) {
3156 if (key->stencil[i].enabled) {
3157 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3158 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3159 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3160 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3161 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3162 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3163 }
3164 }
3165
3166 if (key->alpha.enabled) {
3167 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3168 }
3169
3170 if (key->occlusion_count) {
3171 debug_printf("occlusion_count = 1\n");
3172 }
3173
3174 if (key->blend.logicop_enable) {
3175 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3176 }
3177 else if (key->blend.rt[0].blend_enable) {
3178 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
3179 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3180 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3181 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
3182 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3183 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3184 }
3185 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3186 if (key->blend.alpha_to_coverage) {
3187 debug_printf("blend.alpha_to_coverage is enabled\n");
3188 }
3189 for (i = 0; i < key->nr_samplers; ++i) {
3190 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
3191 debug_printf("sampler[%u] = \n", i);
3192 debug_printf(" .wrap = %s %s %s\n",
3193 util_str_tex_wrap(sampler->wrap_s, TRUE),
3194 util_str_tex_wrap(sampler->wrap_t, TRUE),
3195 util_str_tex_wrap(sampler->wrap_r, TRUE));
3196 debug_printf(" .min_img_filter = %s\n",
3197 util_str_tex_filter(sampler->min_img_filter, TRUE));
3198 debug_printf(" .min_mip_filter = %s\n",
3199 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3200 debug_printf(" .mag_img_filter = %s\n",
3201 util_str_tex_filter(sampler->mag_img_filter, TRUE));
3202 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3203 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3204 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
3205 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3206 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3207 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
3208 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
3209 }
3210 for (i = 0; i < key->nr_sampler_views; ++i) {
3211 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
3212 debug_printf("texture[%u] = \n", i);
3213 debug_printf(" .format = %s\n",
3214 util_format_name(texture->format));
3215 debug_printf(" .target = %s\n",
3216 util_str_tex_target(texture->target, TRUE));
3217 debug_printf(" .level_zero_only = %u\n",
3218 texture->level_zero_only);
3219 debug_printf(" .pot = %u %u %u\n",
3220 texture->pot_width,
3221 texture->pot_height,
3222 texture->pot_depth);
3223 }
3224 struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3225 for (i = 0; i < key->nr_images; ++i) {
3226 const struct lp_static_texture_state *image = &images[i].image_state;
3227 debug_printf("image[%u] = \n", i);
3228 debug_printf(" .format = %s\n",
3229 util_format_name(image->format));
3230 debug_printf(" .target = %s\n",
3231 util_str_tex_target(image->target, TRUE));
3232 debug_printf(" .level_zero_only = %u\n",
3233 image->level_zero_only);
3234 debug_printf(" .pot = %u %u %u\n",
3235 image->pot_width,
3236 image->pot_height,
3237 image->pot_depth);
3238 }
3239 }
3240
3241
3242 void
3243 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3244 {
3245 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3246 variant->shader->no, variant->no);
3247 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3248 tgsi_dump(variant->shader->base.tokens, 0);
3249 else
3250 nir_print_shader(variant->shader->base.ir.nir, stderr);
3251 dump_fs_variant_key(&variant->key);
3252 debug_printf("variant->opaque = %u\n", variant->opaque);
3253 debug_printf("\n");
3254 }
3255
3256 static void
3257 lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant *variant,
3258 unsigned char ir_sha1_cache_key[20])
3259 {
3260 struct blob blob = { 0 };
3261 unsigned ir_size;
3262 void *ir_binary;
3263
3264 blob_init(&blob);
3265 nir_serialize(&blob, variant->shader->base.ir.nir, true);
3266 ir_binary = blob.data;
3267 ir_size = blob.size;
3268
3269 struct mesa_sha1 ctx;
3270 _mesa_sha1_init(&ctx);
3271 _mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);
3272 _mesa_sha1_update(&ctx, ir_binary, ir_size);
3273 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
3274
3275 blob_finish(&blob);
3276 }
3277
3278 /**
3279 * Generate a new fragment shader variant from the shader code and
3280 * other state indicated by the key.
3281 */
3282 static struct lp_fragment_shader_variant *
3283 generate_variant(struct llvmpipe_context *lp,
3284 struct lp_fragment_shader *shader,
3285 const struct lp_fragment_shader_variant_key *key)
3286 {
3287 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
3288 struct lp_fragment_shader_variant *variant;
3289 const struct util_format_description *cbuf0_format_desc = NULL;
3290 boolean fullcolormask;
3291 char module_name[64];
3292 unsigned char ir_sha1_cache_key[20];
3293 struct lp_cached_code cached = { 0 };
3294 bool needs_caching = false;
3295 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3296 if (!variant)
3297 return NULL;
3298
3299 memset(variant, 0, sizeof(*variant));
3300 snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3301 shader->no, shader->variants_created);
3302
3303 variant->shader = shader;
3304 memcpy(&variant->key, key, shader->variant_key_size);
3305
3306 if (shader->base.ir.nir) {
3307 lp_fs_get_ir_cache_key(variant, ir_sha1_cache_key);
3308
3309 lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);
3310 if (!cached.data_size)
3311 needs_caching = true;
3312 }
3313 variant->gallivm = gallivm_create(module_name, lp->context, &cached);
3314 if (!variant->gallivm) {
3315 FREE(variant);
3316 return NULL;
3317 }
3318
3319 variant->list_item_global.base = variant;
3320 variant->list_item_local.base = variant;
3321 variant->no = shader->variants_created++;
3322
3323
3324
3325 /*
3326 * Determine whether we are touching all channels in the color buffer.
3327 */
3328 fullcolormask = FALSE;
3329 if (key->nr_cbufs == 1) {
3330 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3331 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3332 }
3333
3334 variant->opaque =
3335 !key->blend.logicop_enable &&
3336 !key->blend.rt[0].blend_enable &&
3337 fullcolormask &&
3338 !key->stencil[0].enabled &&
3339 !key->alpha.enabled &&