projects / mesa.git / blob
? search:


59d5a440c03f001aa1ffb7c6873e5cc5c4f5165a
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
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 TUNGSTEN GRAPHICS 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 * - triangle edge in/out testing
35 * - scissor test
36 * - stipple (TBI)
37 * - early depth test
38 * - fragment shader
39 * - alpha test
40 * - depth/stencil test
41 * - blending
42 *
43 * This file has only the glue to assemble the fragment pipeline. The actual
44 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
45 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
46 * muster the LLVM JIT execution engine to create a function that follows an
47 * established binary interface and that can be called from C directly.
48 *
49 * A big source of complexity here is that we often want to run different
50 * stages with different precisions and data types and precisions. For example,
51 * the fragment shader needs typically to be done in floats, but the
52 * depth/stencil test and blending is better done in the type that most closely
53 * matches the depth/stencil and color buffer respectively.
54 *
55 * Since the width of a SIMD vector register stays the same regardless of the
56 * element type, different types imply different number of elements, so we must
57 * code generate more instances of the stages with larger types to be able to
58 * feed/consume the stages with smaller types.
59 *
60 * @author Jose Fonseca <jfonseca@vmware.com>
61 */
62
63 #include <limits.h>
64 #include "pipe/p_defines.h"
65 #include "util/u_inlines.h"
66 #include "util/u_memory.h"
67 #include "util/u_format.h"
68 #include "util/u_dump.h"
69 #include "os/os_time.h"
70 #include "pipe/p_shader_tokens.h"
71 #include "draw/draw_context.h"
72 #include "tgsi/tgsi_dump.h"
73 #include "tgsi/tgsi_scan.h"
74 #include "tgsi/tgsi_parse.h"
75 #include "gallivm/lp_bld_type.h"
76 #include "gallivm/lp_bld_const.h"
77 #include "gallivm/lp_bld_conv.h"
78 #include "gallivm/lp_bld_intr.h"
79 #include "gallivm/lp_bld_logic.h"
80 #include "gallivm/lp_bld_tgsi.h"
81 #include "gallivm/lp_bld_swizzle.h"
82 #include "gallivm/lp_bld_flow.h"
83 #include "gallivm/lp_bld_debug.h"
84
85 #include "lp_bld_alpha.h"
86 #include "lp_bld_blend.h"
87 #include "lp_bld_depth.h"
88 #include "lp_bld_interp.h"
89 #include "lp_context.h"
90 #include "lp_debug.h"
91 #include "lp_perf.h"
92 #include "lp_screen.h"
93 #include "lp_setup.h"
94 #include "lp_state.h"
95 #include "lp_tex_sample.h"
96
97
98 #include <llvm-c/Analysis.h>
99
100
101 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
102 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
103
104
105 /*
106 * Derive from the quad's upper left scalar coordinates the coordinates for
107 * all other quad pixels
108 */
109 static void
110 generate_pos0(LLVMBuilderRef builder,
111 LLVMValueRef x,
112 LLVMValueRef y,
113 LLVMValueRef *x0,
114 LLVMValueRef *y0)
115 {
116 LLVMTypeRef int_elem_type = LLVMInt32Type();
117 LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
118 LLVMTypeRef elem_type = LLVMFloatType();
119 LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
120 LLVMValueRef x_offsets[QUAD_SIZE];
121 LLVMValueRef y_offsets[QUAD_SIZE];
122 unsigned i;
123
124 x = lp_build_broadcast(builder, int_vec_type, x);
125 y = lp_build_broadcast(builder, int_vec_type, y);
126
127 for(i = 0; i < QUAD_SIZE; ++i) {
128 x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
129 y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
130 }
131
132 x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
133 y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
134
135 *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
136 *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
137 }
138
139
140 /**
141 * Generate the depth /stencil test code.
142 */
143 static void
144 generate_depth_stencil(LLVMBuilderRef builder,
145 const struct lp_fragment_shader_variant_key *key,
146 struct lp_type src_type,
147 struct lp_build_mask_context *mask,
148 LLVMValueRef stencil_refs[2],
149 LLVMValueRef src,
150 LLVMValueRef dst_ptr,
151 LLVMValueRef facing)
152 {
153 const struct util_format_description *format_desc;
154 struct lp_type dst_type;
155
156 if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
157 return;
158
159 format_desc = util_format_description(key->zsbuf_format);
160 assert(format_desc);
161
162 /*
163 * Depths are expected to be between 0 and 1, even if they are stored in
164 * floats. Setting these bits here will ensure that the lp_build_conv() call
165 * below won't try to unnecessarily clamp the incoming values.
166 */
167 if(src_type.floating) {
168 src_type.sign = FALSE;
169 src_type.norm = TRUE;
170 }
171 else {
172 assert(!src_type.sign);
173 assert(src_type.norm);
174 }
175
176 /* Pick the depth type. */
177 dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
178
179 /* FIXME: Cope with a depth test type with a different bit width. */
180 assert(dst_type.width == src_type.width);
181 assert(dst_type.length == src_type.length);
182
183 /* Convert fragment Z from float to integer */
184 lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
185
186 dst_ptr = LLVMBuildBitCast(builder,
187 dst_ptr,
188 LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
189 lp_build_depth_stencil_test(builder,
190 &key->depth,
191 key->stencil,
192 dst_type,
193 format_desc,
194 mask,
195 stencil_refs,
196 src,
197 dst_ptr,
198 facing);
199 }
200
201
202 /**
203 * Generate the code to do inside/outside triangle testing for the
204 * four pixels in a 2x2 quad. This will set the four elements of the
205 * quad mask vector to 0 or ~0.
206 * \param i which quad of the quad group to test, in [0,3]
207 */
208 static void
209 generate_tri_edge_mask(LLVMBuilderRef builder,
210 unsigned i,
211 LLVMValueRef *mask, /* ivec4, out */
212 LLVMValueRef c0, /* int32 */
213 LLVMValueRef c1, /* int32 */
214 LLVMValueRef c2, /* int32 */
215 LLVMValueRef step0_ptr, /* ivec4 */
216 LLVMValueRef step1_ptr, /* ivec4 */
217 LLVMValueRef step2_ptr) /* ivec4 */
218 {
219 #define OPTIMIZE_IN_OUT_TEST 0
220 #if OPTIMIZE_IN_OUT_TEST
221 struct lp_build_if_state ifctx;
222 LLVMValueRef not_draw_all;
223 #endif
224 struct lp_build_flow_context *flow;
225 struct lp_type i32_type;
226 LLVMTypeRef i32vec4_type;
227 LLVMValueRef c0_vec, c1_vec, c2_vec;
228 LLVMValueRef in_out_mask;
229
230 assert(i < 4);
231
232 /* int32 vector type */
233 memset(&i32_type, 0, sizeof i32_type);
234 i32_type.floating = FALSE; /* values are integers */
235 i32_type.sign = TRUE; /* values are signed */
236 i32_type.norm = FALSE; /* values are not normalized */
237 i32_type.width = 32; /* 32-bit int values */
238 i32_type.length = 4; /* 4 elements per vector */
239
240 i32vec4_type = lp_build_int32_vec4_type();
241
242 /*
243 * Use a conditional here to do detailed pixel in/out testing.
244 * We only have to do this if c0 != INT_MIN.
245 */
246 flow = lp_build_flow_create(builder);
247 lp_build_flow_scope_begin(flow);
248
249 {
250 #if OPTIMIZE_IN_OUT_TEST
251 /* not_draw_all = (c0 != INT_MIN) */
252 not_draw_all = LLVMBuildICmp(builder,
253 LLVMIntNE,
254 c0,
255 LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
256 "");
257
258 in_out_mask = lp_build_const_int_vec(i32_type, ~0);
259
260
261 lp_build_flow_scope_declare(flow, &in_out_mask);
262
263 /* if (not_draw_all) {... */
264 lp_build_if(&ifctx, flow, builder, not_draw_all);
265 #endif
266 {
267 LLVMValueRef step0_vec, step1_vec, step2_vec;
268 LLVMValueRef m0_vec, m1_vec, m2_vec;
269 LLVMValueRef index, m;
270
271 /* c0_vec = {c0, c0, c0, c0}
272 * Note that we emit this code four times but LLVM optimizes away
273 * three instances of it.
274 */
275 c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
276 c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
277 c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
278 lp_build_name(c0_vec, "edgeconst0vec");
279 lp_build_name(c1_vec, "edgeconst1vec");
280 lp_build_name(c2_vec, "edgeconst2vec");
281
282 /* load step0vec, step1, step2 vec from memory */
283 index = LLVMConstInt(LLVMInt32Type(), i, 0);
284 step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
285 step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
286 step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
287 lp_build_name(step0_vec, "step0vec");
288 lp_build_name(step1_vec, "step1vec");
289 lp_build_name(step2_vec, "step2vec");
290
291 /* m0_vec = step0_ptr[i] > c0_vec */
292 m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
293 m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
294 m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
295
296 /* in_out_mask = m0_vec & m1_vec & m2_vec */
297 m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
298 in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
299 lp_build_name(in_out_mask, "inoutmaskvec");
300 }
301 #if OPTIMIZE_IN_OUT_TEST
302 lp_build_endif(&ifctx);
303 #endif
304
305 }
306 lp_build_flow_scope_end(flow);
307 lp_build_flow_destroy(flow);
308
309 /* This is the initial alive/dead pixel mask for a quad of four pixels.
310 * It's an int[4] vector with each word set to 0 or ~0.
311 * Words will get cleared when pixels faile the Z test, etc.
312 */
313 *mask = in_out_mask;
314 }
315
316
317 static LLVMValueRef
318 generate_scissor_test(LLVMBuilderRef builder,
319 LLVMValueRef context_ptr,
320 const struct lp_build_interp_soa_context *interp,
321 struct lp_type type)
322 {
323 LLVMTypeRef vec_type = lp_build_vec_type(type);
324 LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
325 LLVMValueRef xmin, ymin, xmax, ymax;
326 LLVMValueRef m0, m1, m2, m3, m;
327
328 /* xpos, ypos contain the window coords for the four pixels in the quad */
329 assert(xpos);
330 assert(ypos);
331
332 /* get the current scissor bounds, convert to vectors */
333 xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
334 xmin = lp_build_broadcast(builder, vec_type, xmin);
335
336 ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
337 ymin = lp_build_broadcast(builder, vec_type, ymin);
338
339 xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
340 xmax = lp_build_broadcast(builder, vec_type, xmax);
341
342 ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
343 ymax = lp_build_broadcast(builder, vec_type, ymax);
344
345 /* compare the fragment's position coordinates against the scissor bounds */
346 m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
347 m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
348 m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
349 m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
350
351 /* AND all the masks together */
352 m = LLVMBuildAnd(builder, m0, m1, "");
353 m = LLVMBuildAnd(builder, m, m2, "");
354 m = LLVMBuildAnd(builder, m, m3, "");
355
356 lp_build_name(m, "scissormask");
357
358 return m;
359 }
360
361
362 static LLVMValueRef
363 build_int32_vec_const(int value)
364 {
365 struct lp_type i32_type;
366
367 memset(&i32_type, 0, sizeof i32_type);
368 i32_type.floating = FALSE; /* values are integers */
369 i32_type.sign = TRUE; /* values are signed */
370 i32_type.norm = FALSE; /* values are not normalized */
371 i32_type.width = 32; /* 32-bit int values */
372 i32_type.length = 4; /* 4 elements per vector */
373 return lp_build_const_int_vec(i32_type, value);
374 }
375
376
377
378 /**
379 * Generate the fragment shader, depth/stencil test, and alpha tests.
380 * \param i which quad in the tile, in range [0,3]
381 * \param do_tri_test if 1, do triangle edge in/out testing
382 */
383 static void
384 generate_fs(struct llvmpipe_context *lp,
385 struct lp_fragment_shader *shader,
386 const struct lp_fragment_shader_variant_key *key,
387 LLVMBuilderRef builder,
388 struct lp_type type,
389 LLVMValueRef context_ptr,
390 unsigned i,
391 const struct lp_build_interp_soa_context *interp,
392 struct lp_build_sampler_soa *sampler,
393 LLVMValueRef *pmask,
394 LLVMValueRef (*color)[4],
395 LLVMValueRef depth_ptr,
396 LLVMValueRef facing,
397 unsigned do_tri_test,
398 LLVMValueRef c0,
399 LLVMValueRef c1,
400 LLVMValueRef c2,
401 LLVMValueRef step0_ptr,
402 LLVMValueRef step1_ptr,
403 LLVMValueRef step2_ptr)
404 {
405 const struct tgsi_token *tokens = shader->base.tokens;
406 LLVMTypeRef elem_type;
407 LLVMTypeRef vec_type;
408 LLVMTypeRef int_vec_type;
409 LLVMValueRef consts_ptr;
410 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
411 LLVMValueRef z = interp->pos[2];
412 LLVMValueRef stencil_refs[2];
413 struct lp_build_flow_context *flow;
414 struct lp_build_mask_context mask;
415 boolean early_depth_stencil_test;
416 unsigned attrib;
417 unsigned chan;
418 unsigned cbuf;
419
420 assert(i < 4);
421
422 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
423 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
424
425 elem_type = lp_build_elem_type(type);
426 vec_type = lp_build_vec_type(type);
427 int_vec_type = lp_build_int_vec_type(type);
428
429 consts_ptr = lp_jit_context_constants(builder, context_ptr);
430
431 flow = lp_build_flow_create(builder);
432
433 memset(outputs, 0, sizeof outputs);
434
435 lp_build_flow_scope_begin(flow);
436
437 /* Declare the color and z variables */
438 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
439 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
440 color[cbuf][chan] = LLVMGetUndef(vec_type);
441 lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
442 }
443 }
444 lp_build_flow_scope_declare(flow, &z);
445
446 /* do triangle edge testing */
447 if (do_tri_test) {
448 generate_tri_edge_mask(builder, i, pmask,
449 c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
450 }
451 else {
452 *pmask = build_int32_vec_const(~0);
453 }
454
455 /* 'mask' will control execution based on quad's pixel alive/killed state */
456 lp_build_mask_begin(&mask, flow, type, *pmask);
457
458 if (key->scissor) {
459 LLVMValueRef smask =
460 generate_scissor_test(builder, context_ptr, interp, type);
461 lp_build_mask_update(&mask, smask);
462 }
463
464 early_depth_stencil_test =
465 (key->depth.enabled || key->stencil[0].enabled) &&
466 !key->alpha.enabled &&
467 !shader->info.uses_kill &&
468 !shader->info.writes_z;
469
470 if (early_depth_stencil_test)
471 generate_depth_stencil(builder, key,
472 type, &mask,
473 stencil_refs, z, depth_ptr, facing);
474
475 lp_build_tgsi_soa(builder, tokens, type, &mask,
476 consts_ptr, interp->pos, interp->inputs,
477 outputs, sampler);
478
479 for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
480 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
481 if(outputs[attrib][chan]) {
482 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
483 lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
484
485 switch (shader->info.output_semantic_name[attrib]) {
486 case TGSI_SEMANTIC_COLOR:
487 {
488 unsigned cbuf = shader->info.output_semantic_index[attrib];
489
490 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
491
492 /* Alpha test */
493 /* XXX: should the alpha reference value be passed separately? */
494 /* XXX: should only test the final assignment to alpha */
495 if(cbuf == 0 && chan == 3) {
496 LLVMValueRef alpha = out;
497 LLVMValueRef alpha_ref_value;
498 alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
499 alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
500 lp_build_alpha_test(builder, &key->alpha, type,
501 &mask, alpha, alpha_ref_value);
502 }
503
504 color[cbuf][chan] = out;
505 break;
506 }
507
508 case TGSI_SEMANTIC_POSITION:
509 if(chan == 2)
510 z = out;
511 break;
512 }
513 }
514 }
515 }
516
517 if (!early_depth_stencil_test)
518 generate_depth_stencil(builder, key,
519 type, &mask,
520 stencil_refs, z, depth_ptr, facing);
521
522 lp_build_mask_end(&mask);
523
524 lp_build_flow_scope_end(flow);
525
526 lp_build_flow_destroy(flow);
527
528 *pmask = mask.value;
529
530 }
531
532
533 /**
534 * Generate color blending and color output.
535 */
536 static void
537 generate_blend(const struct pipe_blend_state *blend,
538 LLVMBuilderRef builder,
539 struct lp_type type,
540 LLVMValueRef context_ptr,
541 LLVMValueRef mask,
542 LLVMValueRef *src,
543 LLVMValueRef dst_ptr)
544 {
545 struct lp_build_context bld;
546 struct lp_build_flow_context *flow;
547 struct lp_build_mask_context mask_ctx;
548 LLVMTypeRef vec_type;
549 LLVMTypeRef int_vec_type;
550 LLVMValueRef const_ptr;
551 LLVMValueRef con[4];
552 LLVMValueRef dst[4];
553 LLVMValueRef res[4];
554 unsigned chan;
555
556 lp_build_context_init(&bld, builder, type);
557
558 flow = lp_build_flow_create(builder);
559
560 /* we'll use this mask context to skip blending if all pixels are dead */
561 lp_build_mask_begin(&mask_ctx, flow, type, mask);
562
563 vec_type = lp_build_vec_type(type);
564 int_vec_type = lp_build_int_vec_type(type);
565
566 const_ptr = lp_jit_context_blend_color(builder, context_ptr);
567 const_ptr = LLVMBuildBitCast(builder, const_ptr,
568 LLVMPointerType(vec_type, 0), "");
569
570 for(chan = 0; chan < 4; ++chan) {
571 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
572 con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
573
574 dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
575
576 lp_build_name(con[chan], "con.%c", "rgba"[chan]);
577 lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
578 }
579
580 lp_build_blend_soa(builder, blend, type, src, dst, con, res);
581
582 for(chan = 0; chan < 4; ++chan) {
583 if(blend->rt[0].colormask & (1 << chan)) {
584 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
585 lp_build_name(res[chan], "res.%c", "rgba"[chan]);
586 res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
587 LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
588 }
589 }
590
591 lp_build_mask_end(&mask_ctx);
592 lp_build_flow_destroy(flow);
593 }
594
595
596 /** casting function to avoid compiler warnings */
597 static lp_jit_frag_func
598 cast_voidptr_to_lp_jit_frag_func(void *p)
599 {
600 union {
601 void *v;
602 lp_jit_frag_func f;
603 } tmp;
604 assert(sizeof(tmp.v) == sizeof(tmp.f));
605 tmp.v = p;
606 return tmp.f;
607 }
608
609
610 /**
611 * Generate the runtime callable function for the whole fragment pipeline.
612 * Note that the function which we generate operates on a block of 16
613 * pixels at at time. The block contains 2x2 quads. Each quad contains
614 * 2x2 pixels.
615 */
616 static void
617 generate_fragment(struct llvmpipe_context *lp,
618 struct lp_fragment_shader *shader,
619 struct lp_fragment_shader_variant *variant,
620 unsigned do_tri_test)
621 {
622 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
623 const struct lp_fragment_shader_variant_key *key = &variant->key;
624 struct lp_type fs_type;
625 struct lp_type blend_type;
626 LLVMTypeRef fs_elem_type;
627 LLVMTypeRef fs_vec_type;
628 LLVMTypeRef fs_int_vec_type;
629 LLVMTypeRef blend_vec_type;
630 LLVMTypeRef blend_int_vec_type;
631 LLVMTypeRef arg_types[15];
632 LLVMTypeRef func_type;
633 LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
634 LLVMValueRef context_ptr;
635 LLVMValueRef x;
636 LLVMValueRef y;
637 LLVMValueRef a0_ptr;
638 LLVMValueRef dadx_ptr;
639 LLVMValueRef dady_ptr;
640 LLVMValueRef color_ptr_ptr;
641 LLVMValueRef depth_ptr;
642 LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
643 LLVMBasicBlockRef block;
644 LLVMBuilderRef builder;
645 LLVMValueRef x0;
646 LLVMValueRef y0;
647 struct lp_build_sampler_soa *sampler;
648 struct lp_build_interp_soa_context interp;
649 LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
650 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
651 LLVMValueRef blend_mask;
652 LLVMValueRef blend_in_color[NUM_CHANNELS];
653 LLVMValueRef function;
654 LLVMValueRef facing;
655 unsigned num_fs;
656 unsigned i;
657 unsigned chan;
658 unsigned cbuf;
659
660
661 /* TODO: actually pick these based on the fs and color buffer
662 * characteristics. */
663
664 memset(&fs_type, 0, sizeof fs_type);
665 fs_type.floating = TRUE; /* floating point values */
666 fs_type.sign = TRUE; /* values are signed */
667 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
668 fs_type.width = 32; /* 32-bit float */
669 fs_type.length = 4; /* 4 elements per vector */
670 num_fs = 4; /* number of quads per block */
671
672 memset(&blend_type, 0, sizeof blend_type);
673 blend_type.floating = FALSE; /* values are integers */
674 blend_type.sign = FALSE; /* values are unsigned */
675 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
676 blend_type.width = 8; /* 8-bit ubyte values */
677 blend_type.length = 16; /* 16 elements per vector */
678
679 /*
680 * Generate the function prototype. Any change here must be reflected in
681 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
682 */
683
684 fs_elem_type = lp_build_elem_type(fs_type);
685 fs_vec_type = lp_build_vec_type(fs_type);
686 fs_int_vec_type = lp_build_int_vec_type(fs_type);
687
688 blend_vec_type = lp_build_vec_type(blend_type);
689 blend_int_vec_type = lp_build_int_vec_type(blend_type);
690
691 arg_types[0] = screen->context_ptr_type; /* context */
692 arg_types[1] = LLVMInt32Type(); /* x */
693 arg_types[2] = LLVMInt32Type(); /* y */
694 arg_types[3] = LLVMFloatType(); /* facing */
695 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
696 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
697 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
698 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
699 arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
700 arg_types[9] = LLVMInt32Type(); /* c0 */
701 arg_types[10] = LLVMInt32Type(); /* c1 */
702 arg_types[11] = LLVMInt32Type(); /* c2 */
703 /* Note: the step arrays are built as int32[16] but we interpret
704 * them here as int32_vec4[4].
705 */
706 arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
707 arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
708 arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
709
710 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
711
712 function = LLVMAddFunction(screen->module, "shader", func_type);
713 LLVMSetFunctionCallConv(function, LLVMCCallConv);
714
715 variant->function[do_tri_test] = function;
716
717
718 /* XXX: need to propagate noalias down into color param now we are
719 * passing a pointer-to-pointer?
720 */
721 for(i = 0; i < Elements(arg_types); ++i)
722 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
723 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
724
725 context_ptr = LLVMGetParam(function, 0);
726 x = LLVMGetParam(function, 1);
727 y = LLVMGetParam(function, 2);
728 facing = LLVMGetParam(function, 3);
729 a0_ptr = LLVMGetParam(function, 4);
730 dadx_ptr = LLVMGetParam(function, 5);
731 dady_ptr = LLVMGetParam(function, 6);
732 color_ptr_ptr = LLVMGetParam(function, 7);
733 depth_ptr = LLVMGetParam(function, 8);
734 c0 = LLVMGetParam(function, 9);
735 c1 = LLVMGetParam(function, 10);
736 c2 = LLVMGetParam(function, 11);
737 step0_ptr = LLVMGetParam(function, 12);
738 step1_ptr = LLVMGetParam(function, 13);
739 step2_ptr = LLVMGetParam(function, 14);
740
741 lp_build_name(context_ptr, "context");
742 lp_build_name(x, "x");
743 lp_build_name(y, "y");
744 lp_build_name(a0_ptr, "a0");
745 lp_build_name(dadx_ptr, "dadx");
746 lp_build_name(dady_ptr, "dady");
747 lp_build_name(color_ptr_ptr, "color_ptr");
748 lp_build_name(depth_ptr, "depth");
749 lp_build_name(c0, "c0");
750 lp_build_name(c1, "c1");
751 lp_build_name(c2, "c2");
752 lp_build_name(step0_ptr, "step0");
753 lp_build_name(step1_ptr, "step1");
754 lp_build_name(step2_ptr, "step2");
755
756 /*
757 * Function body
758 */
759
760 block = LLVMAppendBasicBlock(function, "entry");
761 builder = LLVMCreateBuilder();
762 LLVMPositionBuilderAtEnd(builder, block);
763
764 generate_pos0(builder, x, y, &x0, &y0);
765
766 lp_build_interp_soa_init(&interp,
767 shader->base.tokens,
768 key->flatshade,
769 builder, fs_type,
770 a0_ptr, dadx_ptr, dady_ptr,
771 x0, y0);
772
773 /* code generated texture sampling */
774 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
775
776 /* loop over quads in the block */
777 for(i = 0; i < num_fs; ++i) {
778 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
779 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
780 LLVMValueRef depth_ptr_i;
781
782 if(i != 0)
783 lp_build_interp_soa_update(&interp, i);
784
785 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
786
787 generate_fs(lp, shader, key,
788 builder,
789 fs_type,
790 context_ptr,
791 i,
792 &interp,
793 sampler,
794 &fs_mask[i], /* output */
795 out_color,
796 depth_ptr_i,
797 facing,
798 do_tri_test,
799 c0, c1, c2,
800 step0_ptr, step1_ptr, step2_ptr);
801
802 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
803 for(chan = 0; chan < NUM_CHANNELS; ++chan)
804 fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
805 }
806
807 sampler->destroy(sampler);
808
809 /* Loop over color outputs / color buffers to do blending.
810 */
811 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
812 LLVMValueRef color_ptr;
813 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
814
815 /*
816 * Convert the fs's output color and mask to fit to the blending type.
817 */
818 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
819 lp_build_conv(builder, fs_type, blend_type,
820 fs_out_color[cbuf][chan], num_fs,
821 &blend_in_color[chan], 1);
822 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
823 }
824
825 lp_build_conv_mask(builder, fs_type, blend_type,
826 fs_mask, num_fs,
827 &blend_mask, 1);
828
829 color_ptr = LLVMBuildLoad(builder,
830 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
831 "");
832 lp_build_name(color_ptr, "color_ptr%d", cbuf);
833
834 /*
835 * Blending.
836 */
837 generate_blend(&key->blend,
838 builder,
839 blend_type,
840 context_ptr,
841 blend_mask,
842 blend_in_color,
843 color_ptr);
844 }
845
846 LLVMBuildRetVoid(builder);
847
848 LLVMDisposeBuilder(builder);
849
850
851 /* Verify the LLVM IR. If invalid, dump and abort */
852 #ifdef DEBUG
853 if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
854 if (1)
855 LLVMDumpValue(function);
856 abort();
857 }
858 #endif
859
860 /* Apply optimizations to LLVM IR */
861 if (1)
862 LLVMRunFunctionPassManager(screen->pass, function);
863
864 if (LP_DEBUG & DEBUG_JIT) {
865 /* Print the LLVM IR to stderr */
866 LLVMDumpValue(function);
867 debug_printf("\n");
868 }
869
870 /*
871 * Translate the LLVM IR into machine code.
872 */
873 {
874 void *f = LLVMGetPointerToGlobal(screen->engine, function);
875
876 variant->jit_function[do_tri_test] = cast_voidptr_to_lp_jit_frag_func(f);
877
878 if (LP_DEBUG & DEBUG_ASM)
879 lp_disassemble(f);
880 }
881 }
882
883
884 static struct lp_fragment_shader_variant *
885 generate_variant(struct llvmpipe_context *lp,
886 struct lp_fragment_shader *shader,
887 const struct lp_fragment_shader_variant_key *key)
888 {
889 struct lp_fragment_shader_variant *variant;
890
891 if (LP_DEBUG & DEBUG_JIT) {
892 unsigned i;
893
894 tgsi_dump(shader->base.tokens, 0);
895 if(key->depth.enabled) {
896 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
897 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
898 debug_printf("depth.writemask = %u\n", key->depth.writemask);
899 }
900 if(key->alpha.enabled) {
901 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
902 debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
903 }
904 if(key->blend.logicop_enable) {
905 debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
906 }
907 else if(key->blend.rt[0].blend_enable) {
908 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
909 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
910 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
911 debug_printf("alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
912 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
913 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
914 }
915 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
916 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
917 if(key->sampler[i].format) {
918 debug_printf("sampler[%u] = \n", i);
919 debug_printf(" .format = %s\n",
920 util_format_name(key->sampler[i].format));
921 debug_printf(" .target = %s\n",
922 util_dump_tex_target(key->sampler[i].target, TRUE));
923 debug_printf(" .pot = %u %u %u\n",
924 key->sampler[i].pot_width,
925 key->sampler[i].pot_height,
926 key->sampler[i].pot_depth);
927 debug_printf(" .wrap = %s %s %s\n",
928 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
929 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
930 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
931 debug_printf(" .min_img_filter = %s\n",
932 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
933 debug_printf(" .min_mip_filter = %s\n",
934 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
935 debug_printf(" .mag_img_filter = %s\n",
936 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
937 if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
938 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
939 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
940 }
941 }
942 }
943
944 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
945 if(!variant)
946 return NULL;
947
948 variant->shader = shader;
949 memcpy(&variant->key, key, sizeof *key);
950
951 generate_fragment(lp, shader, variant, 0);
952 generate_fragment(lp, shader, variant, 1);
953
954 /* insert new variant into linked list */
955 variant->next = shader->variants;
956 shader->variants = variant;
957
958 return variant;
959 }
960
961
962 void *
963 llvmpipe_create_fs_state(struct pipe_context *pipe,
964 const struct pipe_shader_state *templ)
965 {
966 struct lp_fragment_shader *shader;
967
968 shader = CALLOC_STRUCT(lp_fragment_shader);
969 if (!shader)
970 return NULL;
971
972 /* get/save the summary info for this shader */
973 tgsi_scan_shader(templ->tokens, &shader->info);
974
975 /* we need to keep a local copy of the tokens */
976 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
977
978 return shader;
979 }
980
981
982 void
983 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
984 {
985 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
986
987 if (llvmpipe->fs == fs)
988 return;
989
990 draw_flush(llvmpipe->draw);
991
992 llvmpipe->fs = fs;
993
994 llvmpipe->dirty |= LP_NEW_FS;
995 }
996
997
998 void
999 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1000 {
1001 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1002 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
1003 struct lp_fragment_shader *shader = fs;
1004 struct lp_fragment_shader_variant *variant;
1005
1006 assert(fs != llvmpipe->fs);
1007 (void) llvmpipe;
1008
1009 /*
1010 * XXX: we need to flush the context until we have some sort of reference
1011 * counting in fragment shaders as they may still be binned
1012 */
1013 draw_flush(llvmpipe->draw);
1014 lp_setup_flush(llvmpipe->setup, 0);
1015
1016 variant = shader->variants;
1017 while(variant) {
1018 struct lp_fragment_shader_variant *next = variant->next;
1019 unsigned i;
1020
1021 for (i = 0; i < Elements(variant->function); i++) {
1022 if (variant->function[i]) {
1023 if (variant->jit_function[i])
1024 LLVMFreeMachineCodeForFunction(screen->engine,
1025 variant->function[i]);
1026 LLVMDeleteFunction(variant->function[i]);
1027 }
1028 }
1029
1030 FREE(variant);
1031
1032 variant = next;
1033 }
1034
1035 FREE((void *) shader->base.tokens);
1036 FREE(shader);
1037 }
1038
1039
1040
1041 void
1042 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1043 uint shader, uint index,
1044 struct pipe_resource *constants)
1045 {
1046 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1047 unsigned size = constants ? constants->width0 : 0;
1048 const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1049
1050 assert(shader < PIPE_SHADER_TYPES);
1051 assert(index == 0);
1052
1053 if(llvmpipe->constants[shader] == constants)
1054 return;
1055
1056 draw_flush(llvmpipe->draw);
1057
1058 /* note: reference counting */
1059 pipe_resource_reference(&llvmpipe->constants[shader], constants);
1060
1061 if(shader == PIPE_SHADER_VERTEX) {
1062 draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, 0,
1063 data, size);
1064 }
1065
1066 llvmpipe->dirty |= LP_NEW_CONSTANTS;
1067 }
1068
1069
1070 /**
1071 * We need to generate several variants of the fragment pipeline to match
1072 * all the combinations of the contributing state atoms.
1073 *
1074 * TODO: there is actually no reason to tie this to context state -- the
1075 * generated code could be cached globally in the screen.
1076 */
1077 static void
1078 make_variant_key(struct llvmpipe_context *lp,
1079 struct lp_fragment_shader *shader,
1080 struct lp_fragment_shader_variant_key *key)
1081 {
1082 unsigned i;
1083
1084 memset(key, 0, sizeof *key);
1085
1086 if (lp->framebuffer.zsbuf) {
1087 if (lp->depth_stencil->depth.enabled) {
1088 key->zsbuf_format = lp->framebuffer.zsbuf->format;
1089 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1090 }
1091 if (lp->depth_stencil->stencil[0].enabled) {
1092 key->zsbuf_format = lp->framebuffer.zsbuf->format;
1093 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1094 }
1095 }
1096
1097 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1098 if(key->alpha.enabled)
1099 key->alpha.func = lp->depth_stencil->alpha.func;
1100 /* alpha.ref_value is passed in jit_context */
1101
1102 key->flatshade = lp->rasterizer->flatshade;
1103 key->scissor = lp->rasterizer->scissor;
1104
1105 if (lp->framebuffer.nr_cbufs) {
1106 memcpy(&key->blend, lp->blend, sizeof key->blend);
1107 }
1108
1109 key->nr_cbufs = lp->framebuffer.nr_cbufs;
1110 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1111 const struct util_format_description *format_desc;
1112 unsigned chan;
1113
1114 format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1115 assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
1116 format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
1117
1118 key->blend.rt[i].colormask = lp->blend->rt[i].colormask;
1119
1120 /* mask out color channels not present in the color buffer.
1121 * Should be simple to incorporate per-cbuf writemasks:
1122 */
1123 for(chan = 0; chan < 4; ++chan) {
1124 enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1125
1126 if(swizzle > UTIL_FORMAT_SWIZZLE_W)
1127 key->blend.rt[i].colormask &= ~(1 << chan);
1128 }
1129 }
1130
1131 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1132 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1133 lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i], lp->sampler[i]);
1134 }
1135
1136
1137 /**
1138 * Update fragment state. This is called just prior to drawing
1139 * something when some fragment-related state has changed.
1140 */
1141 void
1142 llvmpipe_update_fs(struct llvmpipe_context *lp)
1143 {
1144 struct lp_fragment_shader *shader = lp->fs;
1145 struct lp_fragment_shader_variant_key key;
1146 struct lp_fragment_shader_variant *variant;
1147 boolean opaque;
1148
1149 make_variant_key(lp, shader, &key);
1150
1151 variant = shader->variants;
1152 while(variant) {
1153 if(memcmp(&variant->key, &key, sizeof key) == 0)
1154 break;
1155
1156 variant = variant->next;
1157 }
1158
1159 if (!variant) {
1160 int64_t t0, t1;
1161 int64_t dt;
1162 t0 = os_time_get();
1163
1164 variant = generate_variant(lp, shader, &key);
1165
1166 t1 = os_time_get();
1167 dt = t1 - t0;
1168 LP_COUNT_ADD(llvm_compile_time, dt);
1169 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
1170 }
1171
1172 shader->current = variant;
1173
1174 /* TODO: put this in the variant */
1175 /* TODO: most of these can be relaxed, in particular the colormask */
1176 opaque = !key.blend.logicop_enable &&
1177 !key.blend.rt[0].blend_enable &&
1178 key.blend.rt[0].colormask == 0xf &&
1179 !key.stencil[0].enabled &&
1180 !key.alpha.enabled &&
1181 !key.depth.enabled &&
1182 !key.scissor &&
1183 !shader->info.uses_kill
1184 ? TRUE : FALSE;
1185
1186 lp_setup_set_fs_functions(lp->setup,
1187 shader->current->jit_function[RAST_WHOLE],
1188 shader->current->jit_function[RAST_EDGE_TEST],
1189 opaque);
1190 }