projects / mesa.git / blob
? search:


42653d36ec4aa5d43a28853215e6810841f926e3
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_struct.h"
35 #include "gallivm/lp_bld_type.h"
36 #include "gallivm/lp_bld_flow.h"
37 #include "gallivm/lp_bld_debug.h"
38 #include "gallivm/lp_bld_tgsi.h"
39 #include "gallivm/lp_bld_printf.h"
40
41 #include "tgsi/tgsi_exec.h"
42 #include "tgsi/tgsi_dump.h"
43
44 #include "util/u_cpu_detect.h"
45 #include "util/u_pointer.h"
46 #include "util/u_string.h"
47
48 #include <llvm-c/Transforms/Scalar.h>
49
50 #define DEBUG_STORE 0
51
52 /* generates the draw jit function */
53 static void
54 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
55 static void
56 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *var);
57
58 static void
59 init_globals(struct draw_llvm *llvm)
60 {
61 LLVMTypeRef texture_type;
62
63 /* struct draw_jit_texture */
64 {
65 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
66
67 elem_types[DRAW_JIT_TEXTURE_WIDTH] = LLVMInt32Type();
68 elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
69 elem_types[DRAW_JIT_TEXTURE_DEPTH] = LLVMInt32Type();
70 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = LLVMInt32Type();
71 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
72 LLVMArrayType(LLVMInt32Type(), DRAW_MAX_TEXTURE_LEVELS);
73 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
74 LLVMArrayType(LLVMInt32Type(), DRAW_MAX_TEXTURE_LEVELS);
75 elem_types[DRAW_JIT_TEXTURE_DATA] =
76 LLVMArrayType(LLVMPointerType(LLVMInt8Type(), 0),
77 DRAW_MAX_TEXTURE_LEVELS);
78
79 texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
80
81 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
82 llvm->target, texture_type,
83 DRAW_JIT_TEXTURE_WIDTH);
84 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
85 llvm->target, texture_type,
86 DRAW_JIT_TEXTURE_HEIGHT);
87 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
88 llvm->target, texture_type,
89 DRAW_JIT_TEXTURE_DEPTH);
90 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
91 llvm->target, texture_type,
92 DRAW_JIT_TEXTURE_LAST_LEVEL);
93 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
94 llvm->target, texture_type,
95 DRAW_JIT_TEXTURE_ROW_STRIDE);
96 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
97 llvm->target, texture_type,
98 DRAW_JIT_TEXTURE_IMG_STRIDE);
99 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
100 llvm->target, texture_type,
101 DRAW_JIT_TEXTURE_DATA);
102 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
103 llvm->target, texture_type);
104
105 LLVMAddTypeName(llvm->module, "texture", texture_type);
106 }
107
108
109 /* struct draw_jit_context */
110 {
111 LLVMTypeRef elem_types[3];
112 LLVMTypeRef context_type;
113
114 elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
115 elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
116 elem_types[2] = LLVMArrayType(texture_type,
117 PIPE_MAX_VERTEX_SAMPLERS); /* textures */
118
119 context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
120
121 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
122 llvm->target, context_type, 0);
123 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
124 llvm->target, context_type, 1);
125 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
126 llvm->target, context_type,
127 DRAW_JIT_CTX_TEXTURES);
128 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
129 llvm->target, context_type);
130
131 LLVMAddTypeName(llvm->module, "draw_jit_context", context_type);
132
133 llvm->context_ptr_type = LLVMPointerType(context_type, 0);
134 }
135 {
136 LLVMTypeRef buffer_ptr = LLVMPointerType(LLVMIntType(8), 0);
137 llvm->buffer_ptr_type = LLVMPointerType(buffer_ptr, 0);
138 }
139 /* struct pipe_vertex_buffer */
140 {
141 LLVMTypeRef elem_types[4];
142 LLVMTypeRef vb_type;
143
144 elem_types[0] = LLVMInt32Type();
145 elem_types[1] = LLVMInt32Type();
146 elem_types[2] = LLVMInt32Type();
147 elem_types[3] = LLVMPointerType(LLVMOpaqueType(), 0); /* vs_constants */
148
149 vb_type = LLVMStructType(elem_types, Elements(elem_types), 0);
150
151 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
152 llvm->target, vb_type, 0);
153 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
154 llvm->target, vb_type, 2);
155 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer,
156 llvm->target, vb_type);
157
158 LLVMAddTypeName(llvm->module, "pipe_vertex_buffer", vb_type);
159
160 llvm->vb_ptr_type = LLVMPointerType(vb_type, 0);
161 }
162 }
163
164 static LLVMTypeRef
165 create_vertex_header(struct draw_llvm *llvm, int data_elems)
166 {
167 /* struct vertex_header */
168 LLVMTypeRef elem_types[3];
169 LLVMTypeRef vertex_header;
170 char struct_name[24];
171
172 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
173
174 elem_types[0] = LLVMIntType(32);
175 elem_types[1] = LLVMArrayType(LLVMFloatType(), 4);
176 elem_types[2] = LLVMArrayType(elem_types[1], data_elems);
177
178 vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
179
180 /* these are bit-fields and we can't take address of them
181 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
182 llvm->target, vertex_header,
183 DRAW_JIT_VERTEX_CLIPMASK);
184 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
185 llvm->target, vertex_header,
186 DRAW_JIT_VERTEX_EDGEFLAG);
187 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
188 llvm->target, vertex_header,
189 DRAW_JIT_VERTEX_PAD);
190 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
191 llvm->target, vertex_header,
192 DRAW_JIT_VERTEX_VERTEX_ID);
193 */
194 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
195 llvm->target, vertex_header,
196 DRAW_JIT_VERTEX_CLIP);
197 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
198 llvm->target, vertex_header,
199 DRAW_JIT_VERTEX_DATA);
200
201 LLVMAddTypeName(llvm->module, struct_name, vertex_header);
202
203 return LLVMPointerType(vertex_header, 0);
204 }
205
206 struct draw_llvm *
207 draw_llvm_create(struct draw_context *draw)
208 {
209 struct draw_llvm *llvm;
210
211 #ifdef PIPE_ARCH_X86
212 util_cpu_detect();
213 /* require SSE2 due to LLVM PR6960. */
214 if (!util_cpu_caps.has_sse2)
215 return NULL;
216 #endif
217
218 llvm = CALLOC_STRUCT( draw_llvm );
219 if (!llvm)
220 return NULL;
221
222 llvm->draw = draw;
223 llvm->engine = draw->engine;
224
225 debug_assert(llvm->engine);
226
227 llvm->module = LLVMModuleCreateWithName("draw_llvm");
228 llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
229
230 LLVMAddModuleProvider(llvm->engine, llvm->provider);
231
232 llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
233
234 llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
235 LLVMAddTargetData(llvm->target, llvm->pass);
236
237 if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
238 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
239 * but there are more on SVN. */
240 /* TODO: Add more passes */
241 LLVMAddCFGSimplificationPass(llvm->pass);
242 LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
243 LLVMAddConstantPropagationPass(llvm->pass);
244 if(util_cpu_caps.has_sse4_1) {
245 /* FIXME: There is a bug in this pass, whereby the combination of fptosi
246 * and sitofp (necessary for trunc/floor/ceil/round implementation)
247 * somehow becomes invalid code.
248 */
249 LLVMAddInstructionCombiningPass(llvm->pass);
250 }
251 LLVMAddGVNPass(llvm->pass);
252 } else {
253 /* We need at least this pass to prevent the backends to fail in
254 * unexpected ways.
255 */
256 LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
257 }
258
259 init_globals(llvm);
260
261 if (gallivm_debug & GALLIVM_DEBUG_IR) {
262 LLVMDumpModule(llvm->module);
263 }
264
265 llvm->nr_variants = 0;
266 make_empty_list(&llvm->vs_variants_list);
267
268 return llvm;
269 }
270
271 void
272 draw_llvm_destroy(struct draw_llvm *llvm)
273 {
274 LLVMDisposePassManager(llvm->pass);
275
276 FREE(llvm);
277 }
278
279 struct draw_llvm_variant *
280 draw_llvm_create_variant(struct draw_llvm *llvm, int num_inputs)
281 {
282 struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
283 struct llvm_vertex_shader *shader =
284 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
285
286 variant->llvm = llvm;
287
288 draw_llvm_make_variant_key(llvm, &variant->key);
289
290 llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
291
292 draw_llvm_generate(llvm, variant);
293 draw_llvm_generate_elts(llvm, variant);
294
295 variant->shader = shader;
296 variant->list_item_global.base = variant;
297 variant->list_item_local.base = variant;
298 /*variant->no = */shader->variants_created++;
299 variant->list_item_global.base = variant;
300
301 return variant;
302 }
303
304 static void
305 generate_vs(struct draw_llvm *llvm,
306 LLVMBuilderRef builder,
307 LLVMValueRef (*outputs)[NUM_CHANNELS],
308 const LLVMValueRef (*inputs)[NUM_CHANNELS],
309 LLVMValueRef context_ptr,
310 struct lp_build_sampler_soa *sampler)
311 {
312 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
313 struct lp_type vs_type;
314 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
315
316 memset(&vs_type, 0, sizeof vs_type);
317 vs_type.floating = TRUE; /* floating point values */
318 vs_type.sign = TRUE; /* values are signed */
319 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
320 vs_type.width = 32; /* 32-bit float */
321 vs_type.length = 4; /* 4 elements per vector */
322 #if 0
323 num_vs = 4; /* number of vertices per block */
324 #endif
325
326 if (gallivm_debug & GALLIVM_DEBUG_IR) {
327 tgsi_dump(tokens, 0);
328 }
329
330 lp_build_tgsi_soa(builder,
331 tokens,
332 vs_type,
333 NULL /*struct lp_build_mask_context *mask*/,
334 consts_ptr,
335 NULL /*pos*/,
336 inputs,
337 outputs,
338 sampler,
339 &llvm->draw->vs.vertex_shader->info);
340 }
341
342 #if DEBUG_STORE
343 static void print_vectorf(LLVMBuilderRef builder,
344 LLVMValueRef vec)
345 {
346 LLVMValueRef val[4];
347 val[0] = LLVMBuildExtractElement(builder, vec,
348 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
349 val[1] = LLVMBuildExtractElement(builder, vec,
350 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
351 val[2] = LLVMBuildExtractElement(builder, vec,
352 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
353 val[3] = LLVMBuildExtractElement(builder, vec,
354 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
355 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
356 val[0], val[1], val[2], val[3]);
357 }
358 #endif
359
360 static void
361 generate_fetch(LLVMBuilderRef builder,
362 LLVMValueRef vbuffers_ptr,
363 LLVMValueRef *res,
364 struct pipe_vertex_element *velem,
365 LLVMValueRef vbuf,
366 LLVMValueRef index)
367 {
368 LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
369 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
370 &indices, 1, "");
371 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
372 LLVMValueRef vb_max_index = draw_jit_vbuffer_max_index(builder, vbuf);
373 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
374 LLVMValueRef cond;
375 LLVMValueRef stride;
376
377 cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
378
379 index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
380
381 stride = LLVMBuildMul(builder, vb_stride, index, "");
382
383 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
384
385 stride = LLVMBuildAdd(builder, stride,
386 vb_buffer_offset,
387 "");
388 stride = LLVMBuildAdd(builder, stride,
389 LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
390 "");
391
392 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
393 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
394
395 *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
396 }
397
398 static LLVMValueRef
399 aos_to_soa(LLVMBuilderRef builder,
400 LLVMValueRef val0,
401 LLVMValueRef val1,
402 LLVMValueRef val2,
403 LLVMValueRef val3,
404 LLVMValueRef channel)
405 {
406 LLVMValueRef ex, res;
407
408 ex = LLVMBuildExtractElement(builder, val0,
409 channel, "");
410 res = LLVMBuildInsertElement(builder,
411 LLVMConstNull(LLVMTypeOf(val0)),
412 ex,
413 LLVMConstInt(LLVMInt32Type(), 0, 0),
414 "");
415
416 ex = LLVMBuildExtractElement(builder, val1,
417 channel, "");
418 res = LLVMBuildInsertElement(builder,
419 res, ex,
420 LLVMConstInt(LLVMInt32Type(), 1, 0),
421 "");
422
423 ex = LLVMBuildExtractElement(builder, val2,
424 channel, "");
425 res = LLVMBuildInsertElement(builder,
426 res, ex,
427 LLVMConstInt(LLVMInt32Type(), 2, 0),
428 "");
429
430 ex = LLVMBuildExtractElement(builder, val3,
431 channel, "");
432 res = LLVMBuildInsertElement(builder,
433 res, ex,
434 LLVMConstInt(LLVMInt32Type(), 3, 0),
435 "");
436
437 return res;
438 }
439
440 static void
441 soa_to_aos(LLVMBuilderRef builder,
442 LLVMValueRef soa[NUM_CHANNELS],
443 LLVMValueRef aos[NUM_CHANNELS])
444 {
445 LLVMValueRef comp;
446 int i = 0;
447
448 debug_assert(NUM_CHANNELS == 4);
449
450 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
451 aos[1] = aos[2] = aos[3] = aos[0];
452
453 for (i = 0; i < NUM_CHANNELS; ++i) {
454 LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
455
456 comp = LLVMBuildExtractElement(builder, soa[i],
457 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
458 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
459
460 comp = LLVMBuildExtractElement(builder, soa[i],
461 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
462 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
463
464 comp = LLVMBuildExtractElement(builder, soa[i],
465 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
466 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
467
468 comp = LLVMBuildExtractElement(builder, soa[i],
469 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
470 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
471
472 }
473 }
474
475 static void
476 convert_to_soa(LLVMBuilderRef builder,
477 LLVMValueRef (*aos)[NUM_CHANNELS],
478 LLVMValueRef (*soa)[NUM_CHANNELS],
479 int num_attribs)
480 {
481 int i;
482
483 debug_assert(NUM_CHANNELS == 4);
484
485 for (i = 0; i < num_attribs; ++i) {
486 LLVMValueRef val0 = aos[i][0];
487 LLVMValueRef val1 = aos[i][1];
488 LLVMValueRef val2 = aos[i][2];
489 LLVMValueRef val3 = aos[i][3];
490
491 soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
492 LLVMConstInt(LLVMInt32Type(), 0, 0));
493 soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
494 LLVMConstInt(LLVMInt32Type(), 1, 0));
495 soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
496 LLVMConstInt(LLVMInt32Type(), 2, 0));
497 soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
498 LLVMConstInt(LLVMInt32Type(), 3, 0));
499 }
500 }
501
502 static void
503 store_aos(LLVMBuilderRef builder,
504 LLVMValueRef io_ptr,
505 LLVMValueRef index,
506 LLVMValueRef value)
507 {
508 LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
509 LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
510 LLVMValueRef indices[3];
511
512 indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
513 indices[1] = index;
514 indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
515
516 /* undefined vertex */
517 LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
518 0xffff, 0), id_ptr);
519
520 #if DEBUG_STORE
521 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
522 #endif
523 #if 0
524 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
525 print_vectorf(builder, value);*/
526 data_ptr = LLVMBuildBitCast(builder, data_ptr,
527 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
528 "datavec");
529 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
530
531 LLVMBuildStore(builder, value, data_ptr);
532 #else
533 {
534 LLVMValueRef x, y, z, w;
535 LLVMValueRef idx0, idx1, idx2, idx3;
536 LLVMValueRef gep0, gep1, gep2, gep3;
537 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
538
539 idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
540 idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
541 idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
542 idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
543
544 x = LLVMBuildExtractElement(builder, value,
545 idx0, "");
546 y = LLVMBuildExtractElement(builder, value,
547 idx1, "");
548 z = LLVMBuildExtractElement(builder, value,
549 idx2, "");
550 w = LLVMBuildExtractElement(builder, value,
551 idx3, "");
552
553 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
554 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
555 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
556 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
557
558 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
559 x, gep0, y, gep1, z, gep2, w, gep3);*/
560 LLVMBuildStore(builder, x, gep0);
561 LLVMBuildStore(builder, y, gep1);
562 LLVMBuildStore(builder, z, gep2);
563 LLVMBuildStore(builder, w, gep3);
564 }
565 #endif
566 }
567
568 static void
569 store_aos_array(LLVMBuilderRef builder,
570 LLVMValueRef io_ptr,
571 LLVMValueRef aos[NUM_CHANNELS],
572 int attrib,
573 int num_outputs)
574 {
575 LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
576 LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
577 LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
578 LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
579 LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
580 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
581
582 debug_assert(NUM_CHANNELS == 4);
583
584 io0_ptr = LLVMBuildGEP(builder, io_ptr,
585 &ind0, 1, "");
586 io1_ptr = LLVMBuildGEP(builder, io_ptr,
587 &ind1, 1, "");
588 io2_ptr = LLVMBuildGEP(builder, io_ptr,
589 &ind2, 1, "");
590 io3_ptr = LLVMBuildGEP(builder, io_ptr,
591 &ind3, 1, "");
592
593 #if DEBUG_STORE
594 lp_build_printf(builder, " io = %p, indexes[%d, %d, %d, %d]\n",
595 io_ptr, ind0, ind1, ind2, ind3);
596 #endif
597
598 store_aos(builder, io0_ptr, attr_index, aos[0]);
599 store_aos(builder, io1_ptr, attr_index, aos[1]);
600 store_aos(builder, io2_ptr, attr_index, aos[2]);
601 store_aos(builder, io3_ptr, attr_index, aos[3]);
602 }
603
604 static void
605 convert_to_aos(LLVMBuilderRef builder,
606 LLVMValueRef io,
607 LLVMValueRef (*outputs)[NUM_CHANNELS],
608 int num_outputs,
609 int max_vertices)
610 {
611 unsigned chan, attrib;
612
613 #if DEBUG_STORE
614 lp_build_printf(builder, " # storing begin\n");
615 #endif
616 for (attrib = 0; attrib < num_outputs; ++attrib) {
617 LLVMValueRef soa[4];
618 LLVMValueRef aos[4];
619 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
620 if(outputs[attrib][chan]) {
621 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
622 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
623 /*lp_build_printf(builder, "output %d : %d ",
624 LLVMConstInt(LLVMInt32Type(), attrib, 0),
625 LLVMConstInt(LLVMInt32Type(), chan, 0));
626 print_vectorf(builder, out);*/
627 soa[chan] = out;
628 } else
629 soa[chan] = 0;
630 }
631 soa_to_aos(builder, soa, aos);
632 store_aos_array(builder,
633 io,
634 aos,
635 attrib,
636 num_outputs);
637 }
638 #if DEBUG_STORE
639 lp_build_printf(builder, " # storing end\n");
640 #endif
641 }
642
643 static void
644 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
645 {
646 LLVMTypeRef arg_types[7];
647 LLVMTypeRef func_type;
648 LLVMValueRef context_ptr;
649 LLVMBasicBlockRef block;
650 LLVMBuilderRef builder;
651 LLVMValueRef start, end, count, stride, step, io_itr;
652 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
653 struct draw_context *draw = llvm->draw;
654 unsigned i, j;
655 struct lp_build_context bld;
656 struct lp_build_loop_state lp_loop;
657 struct lp_type vs_type = lp_type_float_vec(32);
658 const int max_vertices = 4;
659 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
660 void *code;
661 struct lp_build_sampler_soa *sampler = 0;
662
663 arg_types[0] = llvm->context_ptr_type; /* context */
664 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
665 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
666 arg_types[3] = LLVMInt32Type(); /* start */
667 arg_types[4] = LLVMInt32Type(); /* count */
668 arg_types[5] = LLVMInt32Type(); /* stride */
669 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
670
671 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
672
673 variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
674 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
675 for(i = 0; i < Elements(arg_types); ++i)
676 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
677 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
678
679 context_ptr = LLVMGetParam(variant->function, 0);
680 io_ptr = LLVMGetParam(variant->function, 1);
681 vbuffers_ptr = LLVMGetParam(variant->function, 2);
682 start = LLVMGetParam(variant->function, 3);
683 count = LLVMGetParam(variant->function, 4);
684 stride = LLVMGetParam(variant->function, 5);
685 vb_ptr = LLVMGetParam(variant->function, 6);
686
687 lp_build_name(context_ptr, "context");
688 lp_build_name(io_ptr, "io");
689 lp_build_name(vbuffers_ptr, "vbuffers");
690 lp_build_name(start, "start");
691 lp_build_name(count, "count");
692 lp_build_name(stride, "stride");
693 lp_build_name(vb_ptr, "vb");
694
695 /*
696 * Function body
697 */
698
699 block = LLVMAppendBasicBlock(variant->function, "entry");
700 builder = LLVMCreateBuilder();
701 LLVMPositionBuilderAtEnd(builder, block);
702
703 lp_build_context_init(&bld, builder, vs_type);
704
705 end = lp_build_add(&bld, start, count);
706
707 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
708
709 /* code generated texture sampling */
710 sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
711 context_ptr);
712
713 #if DEBUG_STORE
714 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
715 start, end, step);
716 #endif
717 lp_build_loop_begin(builder, start, &lp_loop);
718 {
719 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
720 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
721 LLVMValueRef io;
722 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
723
724 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
725 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
726 #if DEBUG_STORE
727 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
728 io_itr, io, lp_loop.counter);
729 #endif
730 for (i = 0; i < NUM_CHANNELS; ++i) {
731 LLVMValueRef true_index = LLVMBuildAdd(
732 builder,
733 lp_loop.counter,
734 LLVMConstInt(LLVMInt32Type(), i, 0), "");
735 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
736 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
737 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
738 velem->vertex_buffer_index,
739 0);
740 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
741 &vb_index, 1, "");
742 generate_fetch(builder, vbuffers_ptr,
743 &aos_attribs[j][i], velem, vb, true_index);
744 }
745 }
746 convert_to_soa(builder, aos_attribs, inputs,
747 draw->pt.nr_vertex_elements);
748
749 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
750 generate_vs(llvm,
751 builder,
752 outputs,
753 ptr_aos,
754 context_ptr,
755 sampler);
756
757 convert_to_aos(builder, io, outputs,
758 draw->vs.vertex_shader->info.num_outputs,
759 max_vertices);
760 }
761 lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
762
763 sampler->destroy(sampler);
764
765 LLVMBuildRetVoid(builder);
766
767 LLVMDisposeBuilder(builder);
768
769 /*
770 * Translate the LLVM IR into machine code.
771 */
772 #ifdef DEBUG
773 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
774 lp_debug_dump_value(variant->function);
775 assert(0);
776 }
777 #endif
778
779 LLVMRunFunctionPassManager(llvm->pass, variant->function);
780
781 if (gallivm_debug & GALLIVM_DEBUG_IR) {
782 lp_debug_dump_value(variant->function);
783 debug_printf("\n");
784 }
785
786 code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
787 variant->jit_func = (draw_jit_vert_func)pointer_to_func(code);
788
789 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
790 lp_disassemble(code);
791 }
792 }
793
794
795 static void
796 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
797 {
798 LLVMTypeRef arg_types[7];
799 LLVMTypeRef func_type;
800 LLVMValueRef context_ptr;
801 LLVMBasicBlockRef block;
802 LLVMBuilderRef builder;
803 LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
804 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
805 struct draw_context *draw = llvm->draw;
806 unsigned i, j;
807 struct lp_build_context bld;
808 struct lp_build_context bld_int;
809 struct lp_build_loop_state lp_loop;
810 struct lp_type vs_type = lp_type_float_vec(32);
811 const int max_vertices = 4;
812 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
813 LLVMValueRef fetch_max;
814 void *code;
815 struct lp_build_sampler_soa *sampler = 0;
816
817 arg_types[0] = llvm->context_ptr_type; /* context */
818 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
819 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
820 arg_types[3] = LLVMPointerType(LLVMInt32Type(), 0); /* fetch_elts * */
821 arg_types[4] = LLVMInt32Type(); /* fetch_count */
822 arg_types[5] = LLVMInt32Type(); /* stride */
823 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
824
825 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
826
827 variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts", func_type);
828 LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
829 for(i = 0; i < Elements(arg_types); ++i)
830 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
831 LLVMAddAttribute(LLVMGetParam(variant->function_elts, i), LLVMNoAliasAttribute);
832
833 context_ptr = LLVMGetParam(variant->function_elts, 0);
834 io_ptr = LLVMGetParam(variant->function_elts, 1);
835 vbuffers_ptr = LLVMGetParam(variant->function_elts, 2);
836 fetch_elts = LLVMGetParam(variant->function_elts, 3);
837 fetch_count = LLVMGetParam(variant->function_elts, 4);
838 stride = LLVMGetParam(variant->function_elts, 5);
839 vb_ptr = LLVMGetParam(variant->function_elts, 6);
840
841 lp_build_name(context_ptr, "context");
842 lp_build_name(io_ptr, "io");
843 lp_build_name(vbuffers_ptr, "vbuffers");
844 lp_build_name(fetch_elts, "fetch_elts");
845 lp_build_name(fetch_count, "fetch_count");
846 lp_build_name(stride, "stride");
847 lp_build_name(vb_ptr, "vb");
848
849 /*
850 * Function body
851 */
852
853 block = LLVMAppendBasicBlock(variant->function_elts, "entry");
854 builder = LLVMCreateBuilder();
855 LLVMPositionBuilderAtEnd(builder, block);
856
857 lp_build_context_init(&bld, builder, vs_type);
858 lp_build_context_init(&bld_int, builder, lp_type_int(32));
859
860 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
861
862 /* code generated texture sampling */
863 sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
864 context_ptr);
865
866 fetch_max = LLVMBuildSub(builder, fetch_count,
867 LLVMConstInt(LLVMInt32Type(), 1, 0),
868 "fetch_max");
869
870 lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 0, 0), &lp_loop);
871 {
872 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
873 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
874 LLVMValueRef io;
875 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
876
877 io_itr = lp_loop.counter;
878 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
879 #if DEBUG_STORE
880 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
881 io_itr, io, lp_loop.counter);
882 #endif
883 for (i = 0; i < NUM_CHANNELS; ++i) {
884 LLVMValueRef true_index = LLVMBuildAdd(
885 builder,
886 lp_loop.counter,
887 LLVMConstInt(LLVMInt32Type(), i, 0), "");
888 LLVMValueRef fetch_ptr;
889
890 /* make sure we're not out of bounds which can happen
891 * if fetch_count % 4 != 0, because on the last iteration
892 * a few of the 4 vertex fetches will be out of bounds */
893 true_index = lp_build_min(&bld_int, true_index, fetch_max);
894
895 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
896 &true_index, 1, "");
897 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
898 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
899 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
900 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
901 velem->vertex_buffer_index,
902 0);
903 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
904 &vb_index, 1, "");
905 generate_fetch(builder, vbuffers_ptr,
906 &aos_attribs[j][i], velem, vb, true_index);
907 }
908 }
909 convert_to_soa(builder, aos_attribs, inputs,
910 draw->pt.nr_vertex_elements);
911
912 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
913 generate_vs(llvm,
914 builder,
915 outputs,
916 ptr_aos,
917 context_ptr,
918 sampler);
919
920 convert_to_aos(builder, io, outputs,
921 draw->vs.vertex_shader->info.num_outputs,
922 max_vertices);
923 }
924 lp_build_loop_end_cond(builder, fetch_count, step, LLVMIntUGE, &lp_loop);
925
926 sampler->destroy(sampler);
927
928 LLVMBuildRetVoid(builder);
929
930 LLVMDisposeBuilder(builder);
931
932 /*
933 * Translate the LLVM IR into machine code.
934 */
935 #ifdef DEBUG
936 if(LLVMVerifyFunction(variant->function_elts, LLVMPrintMessageAction)) {
937 lp_debug_dump_value(variant->function_elts);
938 assert(0);
939 }
940 #endif
941
942 LLVMRunFunctionPassManager(llvm->pass, variant->function_elts);
943
944 if (gallivm_debug & GALLIVM_DEBUG_IR) {
945 lp_debug_dump_value(variant->function_elts);
946 debug_printf("\n");
947 }
948
949 code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function_elts);
950 variant->jit_func_elts = (draw_jit_vert_func_elts)pointer_to_func(code);
951
952 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
953 lp_disassemble(code);
954 }
955 }
956
957 void
958 draw_llvm_make_variant_key(struct draw_llvm *llvm,
959 struct draw_llvm_variant_key *key)
960 {
961 unsigned i;
962
963 memset(key, 0, sizeof(struct draw_llvm_variant_key));
964
965 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
966
967 memcpy(key->vertex_element,
968 llvm->draw->pt.vertex_element,
969 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
970
971 memcpy(&key->vs,
972 &llvm->draw->vs.vertex_shader->state,
973 sizeof(struct pipe_shader_state));
974
975 for(i = 0; i < PIPE_MAX_VERTEX_SAMPLERS; ++i) {
976 struct draw_vertex_shader *shader = llvm->draw->vs.vertex_shader;
977 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
978 lp_sampler_static_state(&key->sampler[i],
979 llvm->draw->sampler_views[i],
980 llvm->draw->samplers[i]);
981 }
982 }
983
984 void
985 draw_llvm_set_mapped_texture(struct draw_context *draw,
986 unsigned sampler_idx,
987 uint32_t width, uint32_t height, uint32_t depth,
988 uint32_t last_level,
989 uint32_t row_stride[DRAW_MAX_TEXTURE_LEVELS],
990 uint32_t img_stride[DRAW_MAX_TEXTURE_LEVELS],
991 const void *data[DRAW_MAX_TEXTURE_LEVELS])
992 {
993 unsigned j;
994 struct draw_jit_texture *jit_tex;
995
996 assert(sampler_idx <= PIPE_MAX_VERTEX_SAMPLERS);
997
998
999 jit_tex = &draw->llvm->jit_context.textures[sampler_idx];
1000
1001 jit_tex->width = width;
1002 jit_tex->height = height;
1003 jit_tex->depth = depth;
1004 jit_tex->last_level = last_level;
1005
1006 for (j = 0; j <= last_level; j++) {
1007 jit_tex->data[j] = data[j];
1008 jit_tex->row_stride[j] = row_stride[j];
1009 jit_tex->img_stride[j] = img_stride[j];
1010 }
1011 }
1012
1013 void
1014 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1015 {
1016 struct draw_llvm *llvm = variant->llvm;
1017 struct draw_context *draw = llvm->draw;
1018
1019 if (variant->function_elts) {
1020 if (variant->function_elts)
1021 LLVMFreeMachineCodeForFunction(draw->engine,
1022 variant->function_elts);
1023 LLVMDeleteFunction(variant->function_elts);
1024 }
1025
1026 if (variant->function) {
1027 if (variant->function)
1028 LLVMFreeMachineCodeForFunction(draw->engine,
1029 variant->function);
1030 LLVMDeleteFunction(variant->function);
1031 }
1032
1033 remove_from_list(&variant->list_item_local);
1034 variant->shader->variants_cached--;
1035 remove_from_list(&variant->list_item_global);
1036 llvm->nr_variants--;
1037 FREE(variant);
1038 }