projects / mesa.git / blob
? search:


1ee87697b4af6ebebaf0510b21ca4149efd0e4e2
[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 *draw_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 struct lp_build_sampler_soa *sampler = 0;
316
317 memset(&vs_type, 0, sizeof vs_type);
318 vs_type.floating = TRUE; /* floating point values */
319 vs_type.sign = TRUE; /* values are signed */
320 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
321 vs_type.width = 32; /* 32-bit float */
322 vs_type.length = 4; /* 4 elements per vector */
323 #if 0
324 num_vs = 4; /* number of vertices per block */
325 #endif
326
327 if (gallivm_debug & GALLIVM_DEBUG_IR) {
328 tgsi_dump(tokens, 0);
329 }
330
331 if (llvm->draw->num_sampler_views &&
332 llvm->draw->num_samplers)
333 sampler = draw_sampler;
334
335 lp_build_tgsi_soa(builder,
336 tokens,
337 vs_type,
338 NULL /*struct lp_build_mask_context *mask*/,
339 consts_ptr,
340 NULL /*pos*/,
341 inputs,
342 outputs,
343 sampler,
344 &llvm->draw->vs.vertex_shader->info);
345 }
346
347 #if DEBUG_STORE
348 static void print_vectorf(LLVMBuilderRef builder,
349 LLVMValueRef vec)
350 {
351 LLVMValueRef val[4];
352 val[0] = LLVMBuildExtractElement(builder, vec,
353 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
354 val[1] = LLVMBuildExtractElement(builder, vec,
355 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
356 val[2] = LLVMBuildExtractElement(builder, vec,
357 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
358 val[3] = LLVMBuildExtractElement(builder, vec,
359 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
360 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
361 val[0], val[1], val[2], val[3]);
362 }
363 #endif
364
365 static void
366 generate_fetch(LLVMBuilderRef builder,
367 LLVMValueRef vbuffers_ptr,
368 LLVMValueRef *res,
369 struct pipe_vertex_element *velem,
370 LLVMValueRef vbuf,
371 LLVMValueRef index,
372 LLVMValueRef instance_id)
373 {
374 LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
375 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
376 &indices, 1, "");
377 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
378 LLVMValueRef vb_max_index = draw_jit_vbuffer_max_index(builder, vbuf);
379 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
380 LLVMValueRef cond;
381 LLVMValueRef stride;
382
383 if (velem->instance_divisor) {
384 /* array index = instance_id / instance_divisor */
385 index = LLVMBuildUDiv(builder, instance_id,
386 LLVMConstInt(LLVMInt32Type(), velem->instance_divisor, 0),
387 "instance_divisor");
388 }
389
390 /* limit index to min(inex, vb_max_index) */
391 cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
392 index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
393
394 stride = LLVMBuildMul(builder, vb_stride, index, "");
395
396 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
397
398 stride = LLVMBuildAdd(builder, stride,
399 vb_buffer_offset,
400 "");
401 stride = LLVMBuildAdd(builder, stride,
402 LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
403 "");
404
405 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
406 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
407
408 *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
409 }
410
411 static LLVMValueRef
412 aos_to_soa(LLVMBuilderRef builder,
413 LLVMValueRef val0,
414 LLVMValueRef val1,
415 LLVMValueRef val2,
416 LLVMValueRef val3,
417 LLVMValueRef channel)
418 {
419 LLVMValueRef ex, res;
420
421 ex = LLVMBuildExtractElement(builder, val0,
422 channel, "");
423 res = LLVMBuildInsertElement(builder,
424 LLVMConstNull(LLVMTypeOf(val0)),
425 ex,
426 LLVMConstInt(LLVMInt32Type(), 0, 0),
427 "");
428
429 ex = LLVMBuildExtractElement(builder, val1,
430 channel, "");
431 res = LLVMBuildInsertElement(builder,
432 res, ex,
433 LLVMConstInt(LLVMInt32Type(), 1, 0),
434 "");
435
436 ex = LLVMBuildExtractElement(builder, val2,
437 channel, "");
438 res = LLVMBuildInsertElement(builder,
439 res, ex,
440 LLVMConstInt(LLVMInt32Type(), 2, 0),
441 "");
442
443 ex = LLVMBuildExtractElement(builder, val3,
444 channel, "");
445 res = LLVMBuildInsertElement(builder,
446 res, ex,
447 LLVMConstInt(LLVMInt32Type(), 3, 0),
448 "");
449
450 return res;
451 }
452
453 static void
454 soa_to_aos(LLVMBuilderRef builder,
455 LLVMValueRef soa[NUM_CHANNELS],
456 LLVMValueRef aos[NUM_CHANNELS])
457 {
458 LLVMValueRef comp;
459 int i = 0;
460
461 debug_assert(NUM_CHANNELS == 4);
462
463 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
464 aos[1] = aos[2] = aos[3] = aos[0];
465
466 for (i = 0; i < NUM_CHANNELS; ++i) {
467 LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
468
469 comp = LLVMBuildExtractElement(builder, soa[i],
470 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
471 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
472
473 comp = LLVMBuildExtractElement(builder, soa[i],
474 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
475 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
476
477 comp = LLVMBuildExtractElement(builder, soa[i],
478 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
479 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
480
481 comp = LLVMBuildExtractElement(builder, soa[i],
482 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
483 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
484
485 }
486 }
487
488 static void
489 convert_to_soa(LLVMBuilderRef builder,
490 LLVMValueRef (*aos)[NUM_CHANNELS],
491 LLVMValueRef (*soa)[NUM_CHANNELS],
492 int num_attribs)
493 {
494 int i;
495
496 debug_assert(NUM_CHANNELS == 4);
497
498 for (i = 0; i < num_attribs; ++i) {
499 LLVMValueRef val0 = aos[i][0];
500 LLVMValueRef val1 = aos[i][1];
501 LLVMValueRef val2 = aos[i][2];
502 LLVMValueRef val3 = aos[i][3];
503
504 soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
505 LLVMConstInt(LLVMInt32Type(), 0, 0));
506 soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
507 LLVMConstInt(LLVMInt32Type(), 1, 0));
508 soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
509 LLVMConstInt(LLVMInt32Type(), 2, 0));
510 soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
511 LLVMConstInt(LLVMInt32Type(), 3, 0));
512 }
513 }
514
515 static void
516 store_aos(LLVMBuilderRef builder,
517 LLVMValueRef io_ptr,
518 LLVMValueRef index,
519 LLVMValueRef value)
520 {
521 LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
522 LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
523 LLVMValueRef indices[3];
524
525 indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
526 indices[1] = index;
527 indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
528
529 /* undefined vertex */
530 LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
531 0xffff, 0), id_ptr);
532
533 #if DEBUG_STORE
534 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
535 #endif
536 #if 0
537 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
538 print_vectorf(builder, value);*/
539 data_ptr = LLVMBuildBitCast(builder, data_ptr,
540 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
541 "datavec");
542 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
543
544 LLVMBuildStore(builder, value, data_ptr);
545 #else
546 {
547 LLVMValueRef x, y, z, w;
548 LLVMValueRef idx0, idx1, idx2, idx3;
549 LLVMValueRef gep0, gep1, gep2, gep3;
550 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
551
552 idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
553 idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
554 idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
555 idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
556
557 x = LLVMBuildExtractElement(builder, value,
558 idx0, "");
559 y = LLVMBuildExtractElement(builder, value,
560 idx1, "");
561 z = LLVMBuildExtractElement(builder, value,
562 idx2, "");
563 w = LLVMBuildExtractElement(builder, value,
564 idx3, "");
565
566 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
567 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
568 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
569 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
570
571 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
572 x, gep0, y, gep1, z, gep2, w, gep3);*/
573 LLVMBuildStore(builder, x, gep0);
574 LLVMBuildStore(builder, y, gep1);
575 LLVMBuildStore(builder, z, gep2);
576 LLVMBuildStore(builder, w, gep3);
577 }
578 #endif
579 }
580
581 static void
582 store_aos_array(LLVMBuilderRef builder,
583 LLVMValueRef io_ptr,
584 LLVMValueRef aos[NUM_CHANNELS],
585 int attrib,
586 int num_outputs)
587 {
588 LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
589 LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
590 LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
591 LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
592 LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
593 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
594
595 debug_assert(NUM_CHANNELS == 4);
596
597 io0_ptr = LLVMBuildGEP(builder, io_ptr,
598 &ind0, 1, "");
599 io1_ptr = LLVMBuildGEP(builder, io_ptr,
600 &ind1, 1, "");
601 io2_ptr = LLVMBuildGEP(builder, io_ptr,
602 &ind2, 1, "");
603 io3_ptr = LLVMBuildGEP(builder, io_ptr,
604 &ind3, 1, "");
605
606 #if DEBUG_STORE
607 lp_build_printf(builder, " io = %p, indexes[%d, %d, %d, %d]\n",
608 io_ptr, ind0, ind1, ind2, ind3);
609 #endif
610
611 store_aos(builder, io0_ptr, attr_index, aos[0]);
612 store_aos(builder, io1_ptr, attr_index, aos[1]);
613 store_aos(builder, io2_ptr, attr_index, aos[2]);
614 store_aos(builder, io3_ptr, attr_index, aos[3]);
615 }
616
617 static void
618 convert_to_aos(LLVMBuilderRef builder,
619 LLVMValueRef io,
620 LLVMValueRef (*outputs)[NUM_CHANNELS],
621 int num_outputs,
622 int max_vertices)
623 {
624 unsigned chan, attrib;
625
626 #if DEBUG_STORE
627 lp_build_printf(builder, " # storing begin\n");
628 #endif
629 for (attrib = 0; attrib < num_outputs; ++attrib) {
630 LLVMValueRef soa[4];
631 LLVMValueRef aos[4];
632 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
633 if(outputs[attrib][chan]) {
634 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
635 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
636 /*lp_build_printf(builder, "output %d : %d ",
637 LLVMConstInt(LLVMInt32Type(), attrib, 0),
638 LLVMConstInt(LLVMInt32Type(), chan, 0));
639 print_vectorf(builder, out);*/
640 soa[chan] = out;
641 } else
642 soa[chan] = 0;
643 }
644 soa_to_aos(builder, soa, aos);
645 store_aos_array(builder,
646 io,
647 aos,
648 attrib,
649 num_outputs);
650 }
651 #if DEBUG_STORE
652 lp_build_printf(builder, " # storing end\n");
653 #endif
654 }
655
656 static void
657 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
658 {
659 LLVMTypeRef arg_types[8];
660 LLVMTypeRef func_type;
661 LLVMValueRef context_ptr;
662 LLVMBasicBlockRef block;
663 LLVMBuilderRef builder;
664 LLVMValueRef start, end, count, stride, step, io_itr;
665 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
666 LLVMValueRef instance_id;
667 struct draw_context *draw = llvm->draw;
668 unsigned i, j;
669 struct lp_build_context bld;
670 struct lp_build_loop_state lp_loop;
671 struct lp_type vs_type = lp_type_float_vec(32);
672 const int max_vertices = 4;
673 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
674 void *code;
675 struct lp_build_sampler_soa *sampler = 0;
676
677 arg_types[0] = llvm->context_ptr_type; /* context */
678 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
679 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
680 arg_types[3] = LLVMInt32Type(); /* start */
681 arg_types[4] = LLVMInt32Type(); /* count */
682 arg_types[5] = LLVMInt32Type(); /* stride */
683 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
684 arg_types[7] = LLVMInt32Type(); /* instance_id */
685
686 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
687
688 variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
689 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
690 for(i = 0; i < Elements(arg_types); ++i)
691 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
692 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
693
694 context_ptr = LLVMGetParam(variant->function, 0);
695 io_ptr = LLVMGetParam(variant->function, 1);
696 vbuffers_ptr = LLVMGetParam(variant->function, 2);
697 start = LLVMGetParam(variant->function, 3);
698 count = LLVMGetParam(variant->function, 4);
699 stride = LLVMGetParam(variant->function, 5);
700 vb_ptr = LLVMGetParam(variant->function, 6);
701 instance_id = LLVMGetParam(variant->function, 7);
702
703 lp_build_name(context_ptr, "context");
704 lp_build_name(io_ptr, "io");
705 lp_build_name(vbuffers_ptr, "vbuffers");
706 lp_build_name(start, "start");
707 lp_build_name(count, "count");
708 lp_build_name(stride, "stride");
709 lp_build_name(vb_ptr, "vb");
710 lp_build_name(instance_id, "instance_id");
711
712 /*
713 * Function body
714 */
715
716 block = LLVMAppendBasicBlock(variant->function, "entry");
717 builder = LLVMCreateBuilder();
718 LLVMPositionBuilderAtEnd(builder, block);
719
720 lp_build_context_init(&bld, builder, vs_type);
721
722 end = lp_build_add(&bld, start, count);
723
724 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
725
726 /* code generated texture sampling */
727 sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
728 context_ptr);
729
730 #if DEBUG_STORE
731 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
732 start, end, step);
733 #endif
734 lp_build_loop_begin(builder, start, &lp_loop);
735 {
736 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
737 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
738 LLVMValueRef io;
739 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
740
741 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
742 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
743 #if DEBUG_STORE
744 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
745 io_itr, io, lp_loop.counter);
746 #endif
747 for (i = 0; i < NUM_CHANNELS; ++i) {
748 LLVMValueRef true_index = LLVMBuildAdd(
749 builder,
750 lp_loop.counter,
751 LLVMConstInt(LLVMInt32Type(), i, 0), "");
752 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
753 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
754 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
755 velem->vertex_buffer_index,
756 0);
757 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
758 &vb_index, 1, "");
759 generate_fetch(builder, vbuffers_ptr,
760 &aos_attribs[j][i], velem, vb, true_index,
761 instance_id);
762 }
763 }
764 convert_to_soa(builder, aos_attribs, inputs,
765 draw->pt.nr_vertex_elements);
766
767 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
768 generate_vs(llvm,
769 builder,
770 outputs,
771 ptr_aos,
772 context_ptr,
773 sampler);
774
775 convert_to_aos(builder, io, outputs,
776 draw->vs.vertex_shader->info.num_outputs,
777 max_vertices);
778 }
779 lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
780
781 sampler->destroy(sampler);
782
783 LLVMBuildRetVoid(builder);
784
785 LLVMDisposeBuilder(builder);
786
787 /*
788 * Translate the LLVM IR into machine code.
789 */
790 #ifdef DEBUG
791 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
792 lp_debug_dump_value(variant->function);
793 assert(0);
794 }
795 #endif
796
797 LLVMRunFunctionPassManager(llvm->pass, variant->function);
798
799 if (gallivm_debug & GALLIVM_DEBUG_IR) {
800 lp_debug_dump_value(variant->function);
801 debug_printf("\n");
802 }
803
804 code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
805 variant->jit_func = (draw_jit_vert_func)pointer_to_func(code);
806
807 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
808 lp_disassemble(code);
809 }
810 }
811
812
813 static void
814 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
815 {
816 LLVMTypeRef arg_types[8];
817 LLVMTypeRef func_type;
818 LLVMValueRef context_ptr;
819 LLVMBasicBlockRef block;
820 LLVMBuilderRef builder;
821 LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
822 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
823 LLVMValueRef instance_id;
824 struct draw_context *draw = llvm->draw;
825 unsigned i, j;
826 struct lp_build_context bld;
827 struct lp_build_context bld_int;
828 struct lp_build_loop_state lp_loop;
829 struct lp_type vs_type = lp_type_float_vec(32);
830 const int max_vertices = 4;
831 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
832 LLVMValueRef fetch_max;
833 void *code;
834 struct lp_build_sampler_soa *sampler = 0;
835
836 arg_types[0] = llvm->context_ptr_type; /* context */
837 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
838 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
839 arg_types[3] = LLVMPointerType(LLVMInt32Type(), 0); /* fetch_elts * */
840 arg_types[4] = LLVMInt32Type(); /* fetch_count */
841 arg_types[5] = LLVMInt32Type(); /* stride */
842 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
843 arg_types[7] = LLVMInt32Type(); /* instance_id */
844
845 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
846
847 variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts",
848 func_type);
849 LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
850 for(i = 0; i < Elements(arg_types); ++i)
851 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
852 LLVMAddAttribute(LLVMGetParam(variant->function_elts, i),
853 LLVMNoAliasAttribute);
854
855 context_ptr = LLVMGetParam(variant->function_elts, 0);
856 io_ptr = LLVMGetParam(variant->function_elts, 1);
857 vbuffers_ptr = LLVMGetParam(variant->function_elts, 2);
858 fetch_elts = LLVMGetParam(variant->function_elts, 3);
859 fetch_count = LLVMGetParam(variant->function_elts, 4);
860 stride = LLVMGetParam(variant->function_elts, 5);
861 vb_ptr = LLVMGetParam(variant->function_elts, 6);
862 instance_id = LLVMGetParam(variant->function_elts, 7);
863
864 lp_build_name(context_ptr, "context");
865 lp_build_name(io_ptr, "io");
866 lp_build_name(vbuffers_ptr, "vbuffers");
867 lp_build_name(fetch_elts, "fetch_elts");
868 lp_build_name(fetch_count, "fetch_count");
869 lp_build_name(stride, "stride");
870 lp_build_name(vb_ptr, "vb");
871 lp_build_name(instance_id, "instance_id");
872
873 /*
874 * Function body
875 */
876
877 block = LLVMAppendBasicBlock(variant->function_elts, "entry");
878 builder = LLVMCreateBuilder();
879 LLVMPositionBuilderAtEnd(builder, block);
880
881 lp_build_context_init(&bld, builder, vs_type);
882 lp_build_context_init(&bld_int, builder, lp_type_int(32));
883
884 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
885
886 /* code generated texture sampling */
887 sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
888 context_ptr);
889
890 fetch_max = LLVMBuildSub(builder, fetch_count,
891 LLVMConstInt(LLVMInt32Type(), 1, 0),
892 "fetch_max");
893
894 lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 0, 0), &lp_loop);
895 {
896 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
897 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
898 LLVMValueRef io;
899 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
900
901 io_itr = lp_loop.counter;
902 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
903 #if DEBUG_STORE
904 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
905 io_itr, io, lp_loop.counter);
906 #endif
907 for (i = 0; i < NUM_CHANNELS; ++i) {
908 LLVMValueRef true_index = LLVMBuildAdd(
909 builder,
910 lp_loop.counter,
911 LLVMConstInt(LLVMInt32Type(), i, 0), "");
912 LLVMValueRef fetch_ptr;
913
914 /* make sure we're not out of bounds which can happen
915 * if fetch_count % 4 != 0, because on the last iteration
916 * a few of the 4 vertex fetches will be out of bounds */
917 true_index = lp_build_min(&bld_int, true_index, fetch_max);
918
919 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
920 &true_index, 1, "");
921 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
922 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
923 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
924 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
925 velem->vertex_buffer_index,
926 0);
927 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
928 &vb_index, 1, "");
929 generate_fetch(builder, vbuffers_ptr,
930 &aos_attribs[j][i], velem, vb, true_index,
931 instance_id);
932 }
933 }
934 convert_to_soa(builder, aos_attribs, inputs,
935 draw->pt.nr_vertex_elements);
936
937 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
938 generate_vs(llvm,
939 builder,
940 outputs,
941 ptr_aos,
942 context_ptr,
943 sampler);
944
945 convert_to_aos(builder, io, outputs,
946 draw->vs.vertex_shader->info.num_outputs,
947 max_vertices);
948 }
949 lp_build_loop_end_cond(builder, fetch_count, step, LLVMIntUGE, &lp_loop);
950
951 sampler->destroy(sampler);
952
953 LLVMBuildRetVoid(builder);
954
955 LLVMDisposeBuilder(builder);
956
957 /*
958 * Translate the LLVM IR into machine code.
959 */
960 #ifdef DEBUG
961 if(LLVMVerifyFunction(variant->function_elts, LLVMPrintMessageAction)) {
962 lp_debug_dump_value(variant->function_elts);
963 assert(0);
964 }
965 #endif
966
967 LLVMRunFunctionPassManager(llvm->pass, variant->function_elts);
968
969 if (gallivm_debug & GALLIVM_DEBUG_IR) {
970 lp_debug_dump_value(variant->function_elts);
971 debug_printf("\n");
972 }
973
974 code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function_elts);
975 variant->jit_func_elts = (draw_jit_vert_func_elts)pointer_to_func(code);
976
977 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
978 lp_disassemble(code);
979 }
980 }
981
982 void
983 draw_llvm_make_variant_key(struct draw_llvm *llvm,
984 struct draw_llvm_variant_key *key)
985 {
986 unsigned i;
987
988 memset(key, 0, sizeof(struct draw_llvm_variant_key));
989
990 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
991
992 memcpy(key->vertex_element,
993 llvm->draw->pt.vertex_element,
994 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
995
996 memcpy(&key->vs,
997 &llvm->draw->vs.vertex_shader->state,
998 sizeof(struct pipe_shader_state));
999
1000 /* if the driver implemented the sampling hooks then
1001 * setup our sampling state */
1002 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers) {
1003 for(i = 0; i < PIPE_MAX_VERTEX_SAMPLERS; ++i) {
1004 struct draw_vertex_shader *shader = llvm->draw->vs.vertex_shader;
1005 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1006 lp_sampler_static_state(&key->sampler[i],
1007 llvm->draw->sampler_views[i],
1008 llvm->draw->samplers[i]);
1009 }
1010 }
1011 }
1012
1013 void
1014 draw_llvm_set_mapped_texture(struct draw_context *draw,
1015 unsigned sampler_idx,
1016 uint32_t width, uint32_t height, uint32_t depth,
1017 uint32_t last_level,
1018 uint32_t row_stride[DRAW_MAX_TEXTURE_LEVELS],
1019 uint32_t img_stride[DRAW_MAX_TEXTURE_LEVELS],
1020 const void *data[DRAW_MAX_TEXTURE_LEVELS])
1021 {
1022 unsigned j;
1023 struct draw_jit_texture *jit_tex;
1024
1025 assert(sampler_idx < PIPE_MAX_VERTEX_SAMPLERS);
1026
1027
1028 jit_tex = &draw->llvm->jit_context.textures[sampler_idx];
1029
1030 jit_tex->width = width;
1031 jit_tex->height = height;
1032 jit_tex->depth = depth;
1033 jit_tex->last_level = last_level;
1034
1035 for (j = 0; j <= last_level; j++) {
1036 jit_tex->data[j] = data[j];
1037 jit_tex->row_stride[j] = row_stride[j];
1038 jit_tex->img_stride[j] = img_stride[j];
1039 }
1040 }
1041
1042 void
1043 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1044 {
1045 struct draw_llvm *llvm = variant->llvm;
1046 struct draw_context *draw = llvm->draw;
1047
1048 if (variant->function_elts) {
1049 if (variant->function_elts)
1050 LLVMFreeMachineCodeForFunction(draw->engine,
1051 variant->function_elts);
1052 LLVMDeleteFunction(variant->function_elts);
1053 }
1054
1055 if (variant->function) {
1056 if (variant->function)
1057 LLVMFreeMachineCodeForFunction(draw->engine,
1058 variant->function);
1059 LLVMDeleteFunction(variant->function);
1060 }
1061
1062 remove_from_list(&variant->list_item_local);
1063 variant->shader->variants_cached--;
1064 remove_from_list(&variant->list_item_global);
1065 llvm->nr_variants--;
1066 FREE(variant);
1067 }