projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: use SoA fetch, not AoS one
[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 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_bitarit.h"
37 #include "gallivm/lp_bld_gather.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_const.h"
40 #include "gallivm/lp_bld_swizzle.h"
41 #include "gallivm/lp_bld_struct.h"
42 #include "gallivm/lp_bld_type.h"
43 #include "gallivm/lp_bld_flow.h"
44 #include "gallivm/lp_bld_debug.h"
45 #include "gallivm/lp_bld_tgsi.h"
46 #include "gallivm/lp_bld_printf.h"
47 #include "gallivm/lp_bld_intr.h"
48 #include "gallivm/lp_bld_init.h"
49 #include "gallivm/lp_bld_type.h"
50 #include "gallivm/lp_bld_pack.h"
51 #include "gallivm/lp_bld_format.h"
52
53 #include "tgsi/tgsi_exec.h"
54 #include "tgsi/tgsi_dump.h"
55
56 #include "util/u_math.h"
57 #include "util/u_pointer.h"
58 #include "util/u_string.h"
59 #include "util/simple_list.h"
60
61
62 #define DEBUG_STORE 0
63
64
65 static void
66 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
67
68
69 struct draw_gs_llvm_iface {
70 struct lp_build_tgsi_gs_iface base;
71
72 struct draw_gs_llvm_variant *variant;
73 LLVMValueRef input;
74 };
75
76 static inline const struct draw_gs_llvm_iface *
77 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
78 {
79 return (const struct draw_gs_llvm_iface *)iface;
80 }
81
82 /**
83 * Create LLVM type for draw_vertex_buffer.
84 */
85 static LLVMTypeRef
86 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
87 const char *struct_name)
88 {
89 LLVMTargetDataRef target = gallivm->target;
90 LLVMTypeRef dvbuffer_type;
91 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
92 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
93
94 elem_types[DRAW_JIT_DVBUFFER_MAP] =
95 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
96 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
97
98 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
99 ARRAY_SIZE(elem_types), 0);
100
101 (void) target; /* silence unused var warning for non-debug build */
102 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
103 target, dvbuffer_type,
104 DRAW_JIT_DVBUFFER_MAP);
105 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
106 target, dvbuffer_type,
107 DRAW_JIT_DVBUFFER_SIZE);
108
109 return dvbuffer_type;
110 }
111
112 /**
113 * Create LLVM type for struct draw_jit_texture
114 */
115 static LLVMTypeRef
116 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
117 {
118 LLVMTargetDataRef target = gallivm->target;
119 LLVMTypeRef texture_type;
120 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
121 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
122
123 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
124 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
125 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
126 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
127 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
128 elem_types[DRAW_JIT_TEXTURE_BASE] =
129 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
130 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
131 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
132 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
133 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
134
135 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
136 ARRAY_SIZE(elem_types), 0);
137
138 (void) target; /* silence unused var warning for non-debug build */
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
140 target, texture_type,
141 DRAW_JIT_TEXTURE_WIDTH);
142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
143 target, texture_type,
144 DRAW_JIT_TEXTURE_HEIGHT);
145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
146 target, texture_type,
147 DRAW_JIT_TEXTURE_DEPTH);
148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
149 target, texture_type,
150 DRAW_JIT_TEXTURE_FIRST_LEVEL);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
152 target, texture_type,
153 DRAW_JIT_TEXTURE_LAST_LEVEL);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_BASE);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_ROW_STRIDE);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_IMG_STRIDE);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
164 target, texture_type,
165 DRAW_JIT_TEXTURE_MIP_OFFSETS);
166
167 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
168
169 return texture_type;
170 }
171
172
173 /**
174 * Create LLVM type for struct draw_jit_sampler
175 */
176 static LLVMTypeRef
177 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
178 {
179 LLVMTargetDataRef target = gallivm->target;
180 LLVMTypeRef sampler_type;
181 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
182
183 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
184 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
185 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
186 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
187 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
188
189 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
190 ARRAY_SIZE(elem_types), 0);
191
192 (void) target; /* silence unused var warning for non-debug build */
193 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
194 target, sampler_type,
195 DRAW_JIT_SAMPLER_MIN_LOD);
196 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
197 target, sampler_type,
198 DRAW_JIT_SAMPLER_MAX_LOD);
199 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
200 target, sampler_type,
201 DRAW_JIT_SAMPLER_LOD_BIAS);
202 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
203 target, sampler_type,
204 DRAW_JIT_SAMPLER_BORDER_COLOR);
205
206 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
207
208 return sampler_type;
209 }
210
211
212 /**
213 * Create LLVM type for struct draw_jit_context
214 */
215 static LLVMTypeRef
216 create_jit_context_type(struct gallivm_state *gallivm,
217 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
218 const char *struct_name)
219 {
220 LLVMTargetDataRef target = gallivm->target;
221 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
222 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
223 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
224 LLVMTypeRef context_type;
225
226 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
227 LP_MAX_TGSI_CONST_BUFFERS);
228 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
229 LP_MAX_TGSI_CONST_BUFFERS);
230 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
231 DRAW_TOTAL_CLIP_PLANES), 0);
232 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
233 elem_types[4] = LLVMArrayType(texture_type,
234 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
235 elem_types[5] = LLVMArrayType(sampler_type,
236 PIPE_MAX_SAMPLERS); /* samplers */
237 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
238 ARRAY_SIZE(elem_types), 0);
239
240 (void) target; /* silence unused var warning for non-debug build */
241 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
242 target, context_type, DRAW_JIT_CTX_CONSTANTS);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
244 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
245 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
246 target, context_type, DRAW_JIT_CTX_PLANES);
247 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
248 target, context_type, DRAW_JIT_CTX_VIEWPORT);
249 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
250 target, context_type,
251 DRAW_JIT_CTX_TEXTURES);
252 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
253 target, context_type,
254 DRAW_JIT_CTX_SAMPLERS);
255 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
256 target, context_type);
257
258 return context_type;
259 }
260
261
262 /**
263 * Create LLVM type for struct draw_gs_jit_context
264 */
265 static LLVMTypeRef
266 create_gs_jit_context_type(struct gallivm_state *gallivm,
267 unsigned vector_length,
268 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
269 const char *struct_name)
270 {
271 LLVMTargetDataRef target = gallivm->target;
272 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
273 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
274 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
275 LLVMTypeRef context_type;
276
277 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
278 LP_MAX_TGSI_CONST_BUFFERS);
279 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
280 LP_MAX_TGSI_CONST_BUFFERS);
281 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
282 DRAW_TOTAL_CLIP_PLANES), 0);
283 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
284
285 elem_types[4] = LLVMArrayType(texture_type,
286 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
287 elem_types[5] = LLVMArrayType(sampler_type,
288 PIPE_MAX_SAMPLERS); /* samplers */
289
290 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
291 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
292 vector_length), 0);
293 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
294 vector_length), 0);
295
296 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
297 ARRAY_SIZE(elem_types), 0);
298
299 (void) target; /* silence unused var warning for non-debug build */
300 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
301 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
302 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
303 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
304 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
305 target, context_type, DRAW_GS_JIT_CTX_PLANES);
306 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
307 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
308 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
309 target, context_type,
310 DRAW_GS_JIT_CTX_TEXTURES);
311 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
312 target, context_type,
313 DRAW_GS_JIT_CTX_SAMPLERS);
314 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
315 target, context_type,
316 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
317 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
318 target, context_type,
319 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
320 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
321 target, context_type,
322 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
323 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
324 target, context_type);
325
326 return context_type;
327 }
328
329
330 static LLVMTypeRef
331 create_gs_jit_input_type(struct gallivm_state *gallivm)
332 {
333 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
334 LLVMTypeRef input_array;
335
336 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
337 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
338 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
339 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
340
341 return input_array;
342 }
343
344 /**
345 * Create LLVM type for struct pipe_vertex_buffer
346 */
347 static LLVMTypeRef
348 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
349 const char *struct_name)
350 {
351 LLVMTargetDataRef target = gallivm->target;
352 LLVMTypeRef elem_types[4];
353 LLVMTypeRef vb_type;
354
355 elem_types[0] =
356 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
357 elem_types[2] =
358 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
359
360 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
361 ARRAY_SIZE(elem_types), 0);
362
363 (void) target; /* silence unused var warning for non-debug build */
364 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
365 target, vb_type, 0);
366 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
367 target, vb_type, 1);
368
369 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
370
371 return vb_type;
372 }
373
374
375 /**
376 * Create LLVM type for struct vertex_header;
377 */
378 static LLVMTypeRef
379 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
380 {
381 LLVMTargetDataRef target = gallivm->target;
382 LLVMTypeRef elem_types[3];
383 LLVMTypeRef vertex_header;
384 char struct_name[24];
385
386 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
387
388 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
389 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
390 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
391
392 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
393 ARRAY_SIZE(elem_types), 0);
394
395 /* these are bit-fields and we can't take address of them
396 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
397 target, vertex_header,
398 DRAW_JIT_VERTEX_CLIPMASK);
399 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
400 target, vertex_header,
401 DRAW_JIT_VERTEX_EDGEFLAG);
402 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
403 target, vertex_header,
404 DRAW_JIT_VERTEX_PAD);
405 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
406 target, vertex_header,
407 DRAW_JIT_VERTEX_VERTEX_ID);
408 */
409 (void) target; /* silence unused var warning for non-debug build */
410 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
411 target, vertex_header,
412 DRAW_JIT_VERTEX_CLIP_POS);
413 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
414 target, vertex_header,
415 DRAW_JIT_VERTEX_DATA);
416
417 assert(LLVMABISizeOfType(target, vertex_header) ==
418 offsetof(struct vertex_header, data[data_elems]));
419
420 return vertex_header;
421 }
422
423
424 /**
425 * Create LLVM types for various structures.
426 */
427 static void
428 create_jit_types(struct draw_llvm_variant *variant)
429 {
430 struct gallivm_state *gallivm = variant->gallivm;
431 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
432 vb_type;
433
434 texture_type = create_jit_texture_type(gallivm, "texture");
435 sampler_type = create_jit_sampler_type(gallivm, "sampler");
436
437 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
438 "draw_jit_context");
439 variant->context_ptr_type = LLVMPointerType(context_type, 0);
440
441 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
442 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
443
444 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
445 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
446 }
447
448
449 static LLVMTypeRef
450 get_context_ptr_type(struct draw_llvm_variant *variant)
451 {
452 if (!variant->context_ptr_type)
453 create_jit_types(variant);
454 return variant->context_ptr_type;
455 }
456
457
458 static LLVMTypeRef
459 get_buffer_ptr_type(struct draw_llvm_variant *variant)
460 {
461 if (!variant->buffer_ptr_type)
462 create_jit_types(variant);
463 return variant->buffer_ptr_type;
464 }
465
466
467 static LLVMTypeRef
468 get_vb_ptr_type(struct draw_llvm_variant *variant)
469 {
470 if (!variant->vb_ptr_type)
471 create_jit_types(variant);
472 return variant->vb_ptr_type;
473 }
474
475 static LLVMTypeRef
476 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
477 {
478 if (!variant->vertex_header_ptr_type)
479 create_jit_types(variant);
480 return variant->vertex_header_ptr_type;
481 }
482
483
484 /**
485 * Create per-context LLVM info.
486 */
487 struct draw_llvm *
488 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
489 {
490 struct draw_llvm *llvm;
491
492 if (!lp_build_init())
493 return NULL;
494
495 llvm = CALLOC_STRUCT( draw_llvm );
496 if (!llvm)
497 return NULL;
498
499 llvm->draw = draw;
500
501 llvm->context = context;
502 if (!llvm->context) {
503 llvm->context = LLVMContextCreate();
504 llvm->context_owned = true;
505 }
506 if (!llvm->context)
507 goto fail;
508
509 llvm->nr_variants = 0;
510 make_empty_list(&llvm->vs_variants_list);
511
512 llvm->nr_gs_variants = 0;
513 make_empty_list(&llvm->gs_variants_list);
514
515 return llvm;
516
517 fail:
518 draw_llvm_destroy(llvm);
519 return NULL;
520 }
521
522
523 /**
524 * Free per-context LLVM info.
525 */
526 void
527 draw_llvm_destroy(struct draw_llvm *llvm)
528 {
529 if (llvm->context_owned)
530 LLVMContextDispose(llvm->context);
531 llvm->context = NULL;
532
533 /* XXX free other draw_llvm data? */
534 FREE(llvm);
535 }
536
537
538 /**
539 * Create LLVM-generated code for a vertex shader.
540 */
541 struct draw_llvm_variant *
542 draw_llvm_create_variant(struct draw_llvm *llvm,
543 unsigned num_inputs,
544 const struct draw_llvm_variant_key *key)
545 {
546 struct draw_llvm_variant *variant;
547 struct llvm_vertex_shader *shader =
548 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
549 LLVMTypeRef vertex_header;
550 char module_name[64];
551
552 variant = MALLOC(sizeof *variant +
553 shader->variant_key_size -
554 sizeof variant->key);
555 if (!variant)
556 return NULL;
557
558 variant->llvm = llvm;
559 variant->shader = shader;
560
561 util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
562 variant->shader->variants_cached);
563
564 variant->gallivm = gallivm_create(module_name, llvm->context);
565
566 create_jit_types(variant);
567
568 memcpy(&variant->key, key, shader->variant_key_size);
569
570 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
571 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
572 draw_llvm_dump_variant_key(&variant->key);
573 }
574
575 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
576
577 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
578
579 draw_llvm_generate(llvm, variant);
580
581 gallivm_compile_module(variant->gallivm);
582
583 variant->jit_func = (draw_jit_vert_func)
584 gallivm_jit_function(variant->gallivm, variant->function);
585
586 gallivm_free_ir(variant->gallivm);
587
588 variant->list_item_global.base = variant;
589 variant->list_item_local.base = variant;
590 /*variant->no = */shader->variants_created++;
591 variant->list_item_global.base = variant;
592
593 return variant;
594 }
595
596
597 static void
598 generate_vs(struct draw_llvm_variant *variant,
599 LLVMBuilderRef builder,
600 struct lp_type vs_type,
601 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
602 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
603 const struct lp_bld_tgsi_system_values *system_values,
604 LLVMValueRef context_ptr,
605 struct lp_build_sampler_soa *draw_sampler,
606 boolean clamp_vertex_color)
607 {
608 struct draw_llvm *llvm = variant->llvm;
609 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
610 LLVMValueRef consts_ptr =
611 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
612 LLVMValueRef num_consts_ptr =
613 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
614
615 lp_build_tgsi_soa(variant->gallivm,
616 tokens,
617 vs_type,
618 NULL /*struct lp_build_mask_context *mask*/,
619 consts_ptr,
620 num_consts_ptr,
621 system_values,
622 inputs,
623 outputs,
624 context_ptr,
625 NULL,
626 draw_sampler,
627 &llvm->draw->vs.vertex_shader->info,
628 NULL);
629
630 {
631 LLVMValueRef out;
632 unsigned chan, attrib;
633 struct lp_build_context bld;
634 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
635 lp_build_context_init(&bld, variant->gallivm, vs_type);
636
637 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
638 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
639 if (outputs[attrib][chan]) {
640 switch (info->output_semantic_name[attrib]) {
641 case TGSI_SEMANTIC_COLOR:
642 case TGSI_SEMANTIC_BCOLOR:
643 if (clamp_vertex_color) {
644 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
645 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
646 LLVMBuildStore(builder, out, outputs[attrib][chan]);
647 }
648 break;
649 }
650 }
651 }
652 }
653 }
654 }
655
656
657 static void
658 fetch_instanced(struct gallivm_state *gallivm,
659 const struct util_format_description *format_desc,
660 struct lp_type vs_type,
661 LLVMValueRef vb_stride,
662 LLVMValueRef map_ptr,
663 LLVMValueRef buffer_size_adj,
664 LLVMValueRef *inputs,
665 LLVMValueRef index)
666 {
667 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
668 LLVMTypeRef aosf_t, aosi_t;
669 LLVMValueRef zero = LLVMConstNull(i32_t);
670 LLVMBuilderRef builder = gallivm->builder;
671 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
672 unsigned i;
673
674 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
675 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
676
677 /* This mul can overflow. Wraparound is ok. */
678 stride = LLVMBuildMul(builder, vb_stride, index, "");
679
680 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
681 stride, buffer_size_adj,
682 "buffer_overflowed");
683
684 if (0) {
685 lp_build_print_value(gallivm, " instance index = ", index);
686 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
687 }
688
689 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
690 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
691 stride = LLVMBuildAnd(builder, stride, index_valid, "");
692
693 aos = lp_build_fetch_rgba_aos(gallivm,
694 format_desc,
695 lp_float32_vec4_type(),
696 FALSE,
697 map_ptr,
698 stride, zero, zero,
699 NULL);
700
701 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
702 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
703 aos = LLVMBuildAnd(builder, aos, index_valid, "");
704 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
705
706 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
707 LLVMValueRef index = lp_build_const_int32(gallivm, i);
708 inputs[i] = lp_build_extract_broadcast(gallivm,
709 lp_float32_vec4_type(),
710 vs_type, aos, index);
711 }
712 }
713
714
715 static void
716 fetch_vector(struct gallivm_state *gallivm,
717 const struct util_format_description *format_desc,
718 struct lp_type vs_type,
719 LLVMValueRef vb_stride,
720 LLVMValueRef map_ptr,
721 LLVMValueRef buffer_size_adj,
722 LLVMValueRef *inputs,
723 LLVMValueRef indices)
724 {
725 LLVMBuilderRef builder = gallivm->builder;
726 struct lp_build_context blduivec;
727 struct lp_type fetch_type = vs_type;
728 LLVMValueRef offset, valid_mask;
729 unsigned i;
730
731 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
732
733 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
734 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
735
736 /* This mul can overflow. Wraparound is ok. */
737 offset = lp_build_mul(&blduivec, vb_stride, indices);
738
739 valid_mask = lp_build_compare(gallivm, blduivec.type,
740 PIPE_FUNC_LESS, offset, buffer_size_adj);
741
742 /* not valid elements use offset 0 */
743 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
744
745 if (0) {
746 lp_build_print_value(gallivm, " indices = ", indices);
747 lp_build_print_value(gallivm, " offsets = ", offset);
748 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
749 }
750
751 /*
752 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
753 * This should always produce better code.
754 */
755
756 /* The type handling is annoying here... */
757 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
758 format_desc->channel[0].pure_integer) {
759 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
760 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
761 }
762 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
763 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
764 }
765 }
766
767 lp_build_fetch_rgba_soa(gallivm, format_desc,
768 fetch_type, FALSE, map_ptr, offset,
769 blduivec.zero, blduivec.zero,
770 NULL, inputs);
771
772 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
773 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
774 lp_build_vec_type(gallivm, vs_type), "");
775 }
776
777 /* out-of-bound fetches return all zeros */
778 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
779 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
780 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
781 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
782 lp_build_vec_type(gallivm, vs_type), "");
783 }
784 }
785
786
787 static void
788 store_aos(struct gallivm_state *gallivm,
789 LLVMValueRef io_ptr,
790 LLVMValueRef index,
791 LLVMValueRef value)
792 {
793 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
794 LLVMBuilderRef builder = gallivm->builder;
795 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
796 LLVMValueRef indices[3];
797
798 indices[0] = lp_build_const_int32(gallivm, 0);
799 indices[1] = index;
800 indices[2] = lp_build_const_int32(gallivm, 0);
801
802 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
803 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
804
805 #if DEBUG_STORE
806 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
807 #endif
808
809 /* Unaligned store due to the vertex header */
810 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
811 }
812
813 /**
814 * Adjust the mask to architecture endianess. The mask will the store in struct:
815 *
816 * struct vertex_header {
817 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
818 * unsigned edgeflag:1;
819 * unsigned pad:1;
820 * unsigned vertex_id:16;
821 * [...]
822 * }
823 *
824 * On little-endian machine nothing needs to done, however on bit-endian machine
825 * the mask's fields need to be adjusted with the algorithm:
826 *
827 * uint32_t reverse (uint32_t x)
828 * {
829 * return (x >> 16) | // vertex_id
830 * ((x & 0x3fff) << 18) | // clipmask
831 * ((x & 0x4000) << 3) | // pad
832 * ((x & 0x8000) << 1); // edgeflag
833 * }
834 */
835 static LLVMValueRef
836 adjust_mask(struct gallivm_state *gallivm,
837 LLVMValueRef mask)
838 {
839 #ifdef PIPE_ARCH_BIG_ENDIAN
840 LLVMBuilderRef builder = gallivm->builder;
841 LLVMValueRef vertex_id;
842 LLVMValueRef clipmask;
843 LLVMValueRef pad;
844 LLVMValueRef edgeflag;
845
846 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
847 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
848 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
849 if (0) {
850 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
851 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
852 }
853 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
854 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
855
856 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
857 if (0) {
858 mask = LLVMBuildOr(builder, mask, pad, "");
859 }
860 mask = LLVMBuildOr(builder, mask, edgeflag, "");
861 #endif
862 return mask;
863 }
864
865 static void
866 store_aos_array(struct gallivm_state *gallivm,
867 struct lp_type soa_type,
868 LLVMValueRef io_ptr,
869 LLVMValueRef *indices,
870 LLVMValueRef* aos,
871 int attrib,
872 int num_outputs,
873 LLVMValueRef clipmask,
874 boolean need_edgeflag)
875 {
876 LLVMBuilderRef builder = gallivm->builder;
877 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
878 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
879 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
880 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
881 int vector_length = soa_type.length;
882 int i;
883
884 debug_assert(TGSI_NUM_CHANNELS == 4);
885
886 for (i = 0; i < vector_length; i++) {
887 linear_inds[i] = lp_build_const_int32(gallivm, i);
888 if (indices) {
889 inds[i] = indices[i];
890 } else {
891 inds[i] = linear_inds[i];
892 }
893 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
894 }
895
896 if (attrib == 0) {
897 /* store vertex header for each of the n vertices */
898 LLVMValueRef val, cliptmp;
899 int vertex_id_pad_edgeflag;
900
901 /* If this assertion fails, it means we need to update the bit twidding
902 * code here. See struct vertex_header in draw_private.h.
903 */
904 assert(DRAW_TOTAL_CLIP_PLANES==14);
905 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
906 if (!need_edgeflag) {
907 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
908 }
909 else {
910 vertex_id_pad_edgeflag = (0xffff << 16);
911 }
912 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
913 vertex_id_pad_edgeflag);
914 /* OR with the clipmask */
915 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
916 for (i = 0; i < vector_length; i++) {
917 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
918 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
919 val = adjust_mask(gallivm, val);
920 #if DEBUG_STORE
921 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
922 io_ptrs[i], inds[i], val);
923 #endif
924 LLVMBuildStore(builder, val, id_ptr);
925 }
926 }
927
928 /* store for each of the n vertices */
929 for (i = 0; i < vector_length; i++) {
930 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
931 }
932 }
933
934
935 static void
936 convert_to_aos(struct gallivm_state *gallivm,
937 LLVMValueRef io,
938 LLVMValueRef *indices,
939 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
940 LLVMValueRef clipmask,
941 int num_outputs,
942 struct lp_type soa_type,
943 boolean need_edgeflag)
944 {
945 LLVMBuilderRef builder = gallivm->builder;
946 unsigned chan, attrib, i;
947
948 #if DEBUG_STORE
949 lp_build_printf(gallivm, " # storing begin\n");
950 #endif
951 for (attrib = 0; attrib < num_outputs; ++attrib) {
952 LLVMValueRef soa[TGSI_NUM_CHANNELS];
953 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
954 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
955 if (outputs[attrib][chan]) {
956 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
957 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
958 #if DEBUG_STORE
959 lp_build_printf(gallivm, "output %d : %d ",
960 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
961 attrib, 0),
962 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
963 chan, 0));
964 lp_build_print_value(gallivm, "val = ", out);
965 {
966 LLVMValueRef iv =
967 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
968
969 lp_build_print_value(gallivm, " ival = ", iv);
970 }
971 #endif
972 soa[chan] = out;
973 }
974 else {
975 soa[chan] = 0;
976 }
977 }
978
979
980 if (soa_type.length == TGSI_NUM_CHANNELS) {
981 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
982 } else {
983 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
984
985 for (i = 0; i < soa_type.length; ++i) {
986 aos[i] = lp_build_extract_range(gallivm,
987 soa[i % TGSI_NUM_CHANNELS],
988 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
989 TGSI_NUM_CHANNELS);
990 }
991 }
992
993 store_aos_array(gallivm,
994 soa_type,
995 io, indices,
996 aos,
997 attrib,
998 num_outputs,
999 clipmask,
1000 need_edgeflag);
1001 }
1002 #if DEBUG_STORE
1003 lp_build_printf(gallivm, " # storing end\n");
1004 #endif
1005 }
1006
1007
1008 /**
1009 * Stores original vertex positions in clip coordinates
1010 */
1011 static void
1012 store_clip(struct gallivm_state *gallivm,
1013 const struct lp_type vs_type,
1014 LLVMValueRef io_ptr,
1015 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1016 int idx)
1017 {
1018 LLVMBuilderRef builder = gallivm->builder;
1019 LLVMValueRef soa[4];
1020 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1021 LLVMValueRef indices[2];
1022 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1023 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1024 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1025 LLVMTypeRef clip_ptr_type =
1026 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1027 4), 0);
1028 int i, j;
1029
1030 indices[0] =
1031 indices[1] = lp_build_const_int32(gallivm, 0);
1032
1033 for (i = 0; i < vs_type.length; i++) {
1034 inds[i] = lp_build_const_int32(gallivm, i);
1035 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1036 }
1037
1038 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1039 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1040 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1041 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1042
1043 for (i = 0; i < vs_type.length; i++) {
1044 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1045 }
1046
1047 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1048 for (i = 0; i < vs_type.length; ++i) {
1049 aos[i] = lp_build_extract_range(gallivm,
1050 soa[i % TGSI_NUM_CHANNELS],
1051 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1052 TGSI_NUM_CHANNELS);
1053 }
1054
1055 for (j = 0; j < vs_type.length; j++) {
1056 LLVMValueRef clip_ptr;
1057
1058 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1059 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1060
1061 /* Unaligned store */
1062 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1063 }
1064 }
1065
1066
1067 /**
1068 * Transforms the outputs for viewport mapping
1069 */
1070 static void
1071 generate_viewport(struct draw_llvm_variant *variant,
1072 LLVMBuilderRef builder,
1073 struct lp_type vs_type,
1074 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1075 LLVMValueRef context_ptr)
1076 {
1077 int i;
1078 struct gallivm_state *gallivm = variant->gallivm;
1079 struct lp_type f32_type = vs_type;
1080 const unsigned pos = variant->llvm->draw->vs.position_output;
1081 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1082 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1083 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1084 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1085
1086 /* We treat pipe_viewport_state as a float array */
1087 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1088 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1089
1090 /* for 1/w convention*/
1091 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1092 LLVMBuildStore(builder, out3, outputs[pos][3]);
1093
1094 /* Viewport Mapping */
1095 for (i=0; i<3; i++) {
1096 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1097 LLVMValueRef scale;
1098 LLVMValueRef trans;
1099 LLVMValueRef scale_i;
1100 LLVMValueRef trans_i;
1101 LLVMValueRef index;
1102
1103 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1104 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1105
1106 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1107 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1108
1109 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1110 LLVMBuildLoad(builder, scale_i, "scale"));
1111 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1112 LLVMBuildLoad(builder, trans_i, "trans"));
1113
1114 /* divide by w */
1115 out = LLVMBuildFMul(builder, out, out3, "");
1116 /* mult by scale, add translation */
1117 out = lp_build_fmuladd(builder, out, scale, trans);
1118
1119 /* store transformed outputs */
1120 LLVMBuildStore(builder, out, outputs[pos][i]);
1121 }
1122
1123 }
1124
1125
1126 /**
1127 * Returns clipmask as nxi32 bitmask for the n vertices
1128 */
1129 static LLVMValueRef
1130 generate_clipmask(struct draw_llvm *llvm,
1131 struct gallivm_state *gallivm,
1132 struct lp_type vs_type,
1133 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1134 struct draw_llvm_variant_key *key,
1135 LLVMValueRef context_ptr,
1136 boolean *have_clipdist)
1137 {
1138 LLVMBuilderRef builder = gallivm->builder;
1139 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1140 LLVMValueRef test, temp;
1141 LLVMValueRef zero, shift;
1142 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1143 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1144 LLVMValueRef plane1, planes, plane_ptr, sum;
1145 struct lp_type f32_type = vs_type;
1146 struct lp_type i32_type = lp_int_type(vs_type);
1147 const unsigned pos = llvm->draw->vs.position_output;
1148 const unsigned cv = llvm->draw->vs.clipvertex_output;
1149 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1150 boolean have_cd = false;
1151 boolean clip_user = key->clip_user;
1152 unsigned ucp_enable = key->ucp_enable;
1153 unsigned cd[2];
1154
1155 cd[0] = llvm->draw->vs.ccdistance_output[0];
1156 cd[1] = llvm->draw->vs.ccdistance_output[1];
1157
1158 if (cd[0] != pos || cd[1] != pos)
1159 have_cd = true;
1160
1161 if (num_written_clipdistance && !clip_user) {
1162 clip_user = true;
1163 ucp_enable = (1 << num_written_clipdistance) - 1;
1164 }
1165
1166 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1167 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1168 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1169 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1170
1171 /*
1172 * load clipvertex and position from correct locations.
1173 * if they are the same just load them once.
1174 */
1175 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1176 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1177 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1178 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1179
1180 if (clip_user && cv != pos) {
1181 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1182 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1183 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1184 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1185 } else {
1186 cv_x = pos_x;
1187 cv_y = pos_y;
1188 cv_z = pos_z;
1189 cv_w = pos_w;
1190 }
1191
1192 /*
1193 * Be careful with the comparisons and NaNs (using llvm's unordered
1194 * comparisons here).
1195 */
1196 /* Cliptest, for hardwired planes */
1197 /*
1198 * XXX should take guardband into account (currently not in key).
1199 * Otherwise might run the draw pipeline stages for nothing.
1200 */
1201 if (key->clip_xy) {
1202 /* plane 1 */
1203 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1204 temp = shift;
1205 test = LLVMBuildAnd(builder, test, temp, "");
1206 mask = test;
1207
1208 /* plane 2 */
1209 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1210 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1211 temp = LLVMBuildShl(builder, temp, shift, "");
1212 test = LLVMBuildAnd(builder, test, temp, "");
1213 mask = LLVMBuildOr(builder, mask, test, "");
1214
1215 /* plane 3 */
1216 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1217 temp = LLVMBuildShl(builder, temp, shift, "");
1218 test = LLVMBuildAnd(builder, test, temp, "");
1219 mask = LLVMBuildOr(builder, mask, test, "");
1220
1221 /* plane 4 */
1222 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1223 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1224 temp = LLVMBuildShl(builder, temp, shift, "");
1225 test = LLVMBuildAnd(builder, test, temp, "");
1226 mask = LLVMBuildOr(builder, mask, test, "");
1227 }
1228
1229 if (key->clip_z) {
1230 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1231 if (key->clip_halfz) {
1232 /* plane 5 */
1233 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1234 test = LLVMBuildAnd(builder, test, temp, "");
1235 mask = LLVMBuildOr(builder, mask, test, "");
1236 }
1237 else {
1238 /* plane 5 */
1239 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1240 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1241 test = LLVMBuildAnd(builder, test, temp, "");
1242 mask = LLVMBuildOr(builder, mask, test, "");
1243 }
1244 /* plane 6 */
1245 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1246 temp = LLVMBuildShl(builder, temp, shift, "");
1247 test = LLVMBuildAnd(builder, test, temp, "");
1248 mask = LLVMBuildOr(builder, mask, test, "");
1249 }
1250
1251 if (clip_user) {
1252 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1253 LLVMValueRef indices[3];
1254 LLVMValueRef is_nan_or_inf;
1255
1256 /* userclip planes */
1257 while (ucp_enable) {
1258 unsigned plane_idx = ffs(ucp_enable)-1;
1259 ucp_enable &= ~(1 << plane_idx);
1260 plane_idx += 6;
1261
1262 if (have_cd && num_written_clipdistance) {
1263 LLVMValueRef clipdist;
1264 int i;
1265 i = plane_idx - 6;
1266
1267 *have_clipdist = TRUE;
1268 if (i < 4) {
1269 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1270 } else {
1271 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1272 }
1273 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1274 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1275 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1276 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1277 test = LLVMBuildAnd(builder, test, temp, "");
1278 mask = LLVMBuildOr(builder, mask, test, "");
1279 } else {
1280 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1281 indices[0] = lp_build_const_int32(gallivm, 0);
1282 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1283
1284 indices[2] = lp_build_const_int32(gallivm, 0);
1285 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1286 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1287 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1288 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1289
1290 indices[2] = lp_build_const_int32(gallivm, 1);
1291 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1292 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1293 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1294 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1295
1296 indices[2] = lp_build_const_int32(gallivm, 2);
1297 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1298 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1299 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1300 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1301
1302 indices[2] = lp_build_const_int32(gallivm, 3);
1303 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1304 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1305 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1306 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1307
1308 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1309 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1310 test = LLVMBuildAnd(builder, test, temp, "");
1311 mask = LLVMBuildOr(builder, mask, test, "");
1312 }
1313 }
1314 }
1315 if (key->need_edgeflags) {
1316 /*
1317 * This isn't really part of clipmask but stored the same in vertex
1318 * header later, so do it here.
1319 */
1320 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1321 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1322 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1323 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1324 temp = lp_build_const_int_vec(gallivm, i32_type,
1325 1LL << DRAW_TOTAL_CLIP_PLANES);
1326 test = LLVMBuildAnd(builder, test, temp, "");
1327 mask = LLVMBuildOr(builder, mask, test, "");
1328 }
1329 return mask;
1330 }
1331
1332
1333 /**
1334 * Returns boolean if any clipping has occurred
1335 * Used zero/one i8 value to represent boolean
1336 */
1337 static LLVMValueRef
1338 clipmask_booli8(struct gallivm_state *gallivm,
1339 const struct lp_type vs_type,
1340 LLVMValueRef clipmask_bool_ptr,
1341 boolean edgeflag_in_clipmask)
1342 {
1343 LLVMBuilderRef builder = gallivm->builder;
1344 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1345 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1346 LLVMValueRef ret;
1347 struct lp_build_context bldivec;
1348
1349 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1350
1351 /*
1352 * We need to invert the edgeflag bit from the clipmask here
1353 * (because the result is really if we want to run the pipeline or not
1354 * and we (may) need it if edgeflag was 0).
1355 */
1356 if (edgeflag_in_clipmask) {
1357 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1358 1LL << DRAW_TOTAL_CLIP_PLANES);
1359 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1360 }
1361
1362 /*
1363 * XXX: probably should mask off bits from the mask which come from
1364 * vertices which were beyond the count (i.e. indices_valid for
1365 * linear fetches, for elts ones we don't have the correct mask
1366 * right now). Otherwise might run the pipeline for nothing,
1367 * though everything should still work.
1368 */
1369 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1370 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1371 return ret;
1372 }
1373
1374 static LLVMValueRef
1375 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1376 struct lp_build_tgsi_context * bld_base,
1377 boolean is_vindex_indirect,
1378 LLVMValueRef vertex_index,
1379 boolean is_aindex_indirect,
1380 LLVMValueRef attrib_index,
1381 LLVMValueRef swizzle_index)
1382 {
1383 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1384 struct gallivm_state *gallivm = bld_base->base.gallivm;
1385 LLVMBuilderRef builder = gallivm->builder;
1386 LLVMValueRef indices[3];
1387 LLVMValueRef res;
1388 struct lp_type type = bld_base->base.type;
1389
1390 if (is_vindex_indirect || is_aindex_indirect) {
1391 int i;
1392 res = bld_base->base.zero;
1393 for (i = 0; i < type.length; ++i) {
1394 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1395 LLVMValueRef vert_chan_index = vertex_index;
1396 LLVMValueRef attr_chan_index = attrib_index;
1397 LLVMValueRef channel_vec, value;
1398
1399 if (is_vindex_indirect) {
1400 vert_chan_index = LLVMBuildExtractElement(builder,
1401 vertex_index, idx, "");
1402 }
1403 if (is_aindex_indirect) {
1404 attr_chan_index = LLVMBuildExtractElement(builder,
1405 attrib_index, idx, "");
1406 }
1407
1408 indices[0] = vert_chan_index;
1409 indices[1] = attr_chan_index;
1410 indices[2] = swizzle_index;
1411
1412 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1413 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1414 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1415
1416 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1417 }
1418 } else {
1419 indices[0] = vertex_index;
1420 indices[1] = attrib_index;
1421 indices[2] = swizzle_index;
1422
1423 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1424 res = LLVMBuildLoad(builder, res, "");
1425 }
1426
1427 return res;
1428 }
1429
1430 static void
1431 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1432 struct lp_build_tgsi_context * bld_base,
1433 LLVMValueRef (*outputs)[4],
1434 LLVMValueRef emitted_vertices_vec)
1435 {
1436 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1437 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1438 struct gallivm_state *gallivm = variant->gallivm;
1439 LLVMBuilderRef builder = gallivm->builder;
1440 struct lp_type gs_type = bld_base->base.type;
1441 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1442 lp_int_type(gs_type), 0);
1443 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1444 LLVMValueRef next_prim_offset =
1445 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1446 LLVMValueRef io = variant->io_ptr;
1447 unsigned i;
1448 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1449
1450 for (i = 0; i < gs_type.length; ++i) {
1451 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1452 LLVMValueRef currently_emitted =
1453 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1454 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1455 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1456 }
1457
1458 convert_to_aos(gallivm, io, indices,
1459 outputs, clipmask,
1460 gs_info->num_outputs, gs_type,
1461 FALSE);
1462 }
1463
1464 static void
1465 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1466 struct lp_build_tgsi_context * bld_base,
1467 LLVMValueRef verts_per_prim_vec,
1468 LLVMValueRef emitted_prims_vec)
1469 {
1470 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1471 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1472 struct gallivm_state *gallivm = variant->gallivm;
1473 LLVMBuilderRef builder = gallivm->builder;
1474 LLVMValueRef prim_lengts_ptr =
1475 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1476 unsigned i;
1477
1478 for (i = 0; i < bld_base->base.type.length; ++i) {
1479 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1480 LLVMValueRef prims_emitted =
1481 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1482 LLVMValueRef store_ptr;
1483 LLVMValueRef num_vertices =
1484 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1485
1486 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1487 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1488 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1489 LLVMBuildStore(builder, num_vertices, store_ptr);
1490 }
1491 }
1492
1493 static void
1494 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1495 struct lp_build_tgsi_context * bld_base,
1496 LLVMValueRef total_emitted_vertices_vec,
1497 LLVMValueRef emitted_prims_vec)
1498 {
1499 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1500 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1501 struct gallivm_state *gallivm = variant->gallivm;
1502 LLVMBuilderRef builder = gallivm->builder;
1503 LLVMValueRef emitted_verts_ptr =
1504 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1505 LLVMValueRef emitted_prims_ptr =
1506 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1507 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1508
1509 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1510 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1511
1512 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1513 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1514 }
1515
1516 static void
1517 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1518 {
1519 struct gallivm_state *gallivm = variant->gallivm;
1520 LLVMContextRef context = gallivm->context;
1521 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1522 LLVMTypeRef arg_types[11];
1523 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1524 LLVMTypeRef func_type;
1525 LLVMValueRef context_ptr;
1526 LLVMBasicBlockRef block;
1527 LLVMBuilderRef builder;
1528 char func_name[64];
1529 struct lp_type vs_type;
1530 LLVMValueRef count, fetch_elts, start_or_maxelt;
1531 LLVMValueRef vertex_id_offset, start_instance;
1532 LLVMValueRef stride, step, io_itr;
1533 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1534 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1535 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1536 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1537 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1538 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1539 LLVMValueRef fake_buf_ptr, fake_buf;
1540
1541 struct draw_context *draw = llvm->draw;
1542 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1543 unsigned i, j;
1544 struct lp_build_context bld, blduivec;
1545 struct lp_build_loop_state lp_loop;
1546 struct lp_build_if_state if_ctx;
1547 const int vector_length = lp_native_vector_width / 32;
1548 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1549 struct lp_build_sampler_soa *sampler = 0;
1550 LLVMValueRef ret, clipmask_bool_ptr;
1551 struct draw_llvm_variant_key *key = &variant->key;
1552 /* If geometry shader is present we need to skip both the viewport
1553 * transformation and clipping otherwise the inputs to the geometry
1554 * shader will be incorrect.
1555 * The code can't handle vp transform when vs writes vp index neither
1556 * (though this would be fixable here, but couldn't just broadcast
1557 * the values).
1558 */
1559 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1560 vs_info->writes_viewport_index;
1561 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1562 key->clip_z ||
1563 key->clip_user ||
1564 key->need_edgeflags);
1565 LLVMValueRef variant_func;
1566 const unsigned pos = draw->vs.position_output;
1567 const unsigned cv = draw->vs.clipvertex_output;
1568 boolean have_clipdist = FALSE;
1569 struct lp_bld_tgsi_system_values system_values;
1570
1571 memset(&system_values, 0, sizeof(system_values));
1572
1573 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1574 variant->shader->variants_cached);
1575
1576 i = 0;
1577 arg_types[i++] = get_context_ptr_type(variant); /* context */
1578 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1579 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1580 arg_types[i++] = int32_type; /* count */
1581 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1582 arg_types[i++] = int32_type; /* stride */
1583 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1584 arg_types[i++] = int32_type; /* instance_id */
1585 arg_types[i++] = int32_type; /* vertex_id_offset */
1586 arg_types[i++] = int32_type; /* start_instance */
1587 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1588
1589 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1590 arg_types, num_arg_types, 0);
1591
1592 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1593 variant->function = variant_func;
1594
1595 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1596 for (i = 0; i < num_arg_types; ++i)
1597 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1598 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1599
1600 context_ptr = LLVMGetParam(variant_func, 0);
1601 io_ptr = LLVMGetParam(variant_func, 1);
1602 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1603 count = LLVMGetParam(variant_func, 3);
1604 /*
1605 * XXX: the maxelt part is unused. Not really useful, since we cannot
1606 * get index buffer overflows due to vsplit (which provides its own
1607 * elts buffer, with a different size than what's passed in here).
1608 */
1609 start_or_maxelt = LLVMGetParam(variant_func, 4);
1610 /*
1611 * XXX: stride is actually unused. The stride we use is strictly calculated
1612 * from the number of outputs (including the draw_extra outputs).
1613 * Should probably fix some day (we need a new vs just because of extra
1614 * outputs which the generated vs won't touch).
1615 */
1616 stride = LLVMGetParam(variant_func, 5);
1617 vb_ptr = LLVMGetParam(variant_func, 6);
1618 system_values.instance_id = LLVMGetParam(variant_func, 7);
1619 vertex_id_offset = LLVMGetParam(variant_func, 8);
1620 start_instance = LLVMGetParam(variant_func, 9);
1621 fetch_elts = LLVMGetParam(variant_func, 10);
1622
1623 lp_build_name(context_ptr, "context");
1624 lp_build_name(io_ptr, "io");
1625 lp_build_name(vbuffers_ptr, "vbuffers");
1626 lp_build_name(count, "count");
1627 lp_build_name(start_or_maxelt, "start_or_maxelt");
1628 lp_build_name(stride, "stride");
1629 lp_build_name(vb_ptr, "vb");
1630 lp_build_name(system_values.instance_id, "instance_id");
1631 lp_build_name(vertex_id_offset, "vertex_id_offset");
1632 lp_build_name(start_instance, "start_instance");
1633 lp_build_name(fetch_elts, "fetch_elts");
1634
1635 /*
1636 * Function body
1637 */
1638
1639 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1640 builder = gallivm->builder;
1641 LLVMPositionBuilderAtEnd(builder, block);
1642
1643 memset(&vs_type, 0, sizeof vs_type);
1644 vs_type.floating = TRUE; /* floating point values */
1645 vs_type.sign = TRUE; /* values are signed */
1646 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1647 vs_type.width = 32; /* 32-bit float */
1648 vs_type.length = vector_length;
1649
1650 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1651 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1652
1653 /* hold temporary "bool" clipmask */
1654 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1655
1656 fake_buf = lp_build_alloca_undef(gallivm,
1657 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1658 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1659 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1660 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1661
1662 /* code generated texture sampling */
1663 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1664
1665 step = lp_build_const_int32(gallivm, vector_length);
1666
1667 ind_vec = blduivec.undef;
1668 for (i = 0; i < vs_type.length; i++) {
1669 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1670 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1671 }
1672
1673 fetch_max = lp_build_alloca(gallivm, int32_type, "fetch_max");
1674
1675 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1676 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1677
1678 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1679 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1680 /*
1681 * Only needed for non-indexed path.
1682 */
1683 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1684
1685 /*
1686 * Pre-calculate everything which is constant per shader invocation.
1687 */
1688 for (j = 0; j < key->nr_vertex_elements; ++j) {
1689 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1690 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1691 struct pipe_vertex_element *velem = &key->vertex_element[j];
1692 LLVMValueRef vb_index =
1693 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1694 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1695 util_format_get_blocksize(velem->src_format));
1696 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1697 velem->src_offset);
1698 struct lp_build_if_state if_ctx;
1699
1700 if (velem->src_format != PIPE_FORMAT_NONE) {
1701 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1702 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1703 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1704 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1705 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1706 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1707
1708 ofbit = NULL;
1709 /*
1710 * We'll set buffer_size_adj to zero if we have of, so it will
1711 * always overflow later automatically without having to keep ofbit.
1712 * Overflows (with normal wraparound) doing the actual offset
1713 * calculation should be ok, just not for the buffer size calc.
1714 * It would also be possible to detect such overflows and return
1715 * zeros if that happens, but this would be more complex.
1716 */
1717 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1718 tmp = lp_build_sub(&bld, bsize, bld.one);
1719 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1720 &ofbit);
1721 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1722 buf_offset, &ofbit);
1723
1724 /*
1725 * We can't easily set fake vertex buffers outside the generated code.
1726 * Hence, set fake vertex buffers here instead basically, so fetch
1727 * code can always fetch using offset 0, eliminating all control flow
1728 * inside the main loop.
1729 * (Alternatively, could have control flow per vector skipping fetch
1730 * if ofbit is true.)
1731 */
1732 if (velem->instance_divisor) {
1733 /*
1734 * Index is equal to the start instance plus the number of current
1735 * instance divided by the divisor. In this case we compute it as:
1736 * index = start_instance + (instance_id / divisor).
1737 * Note we could actually do the fetch here, outside the loop -
1738 * it's all constant, hopefully llvm recognizes this.
1739 */
1740 LLVMValueRef current_instance;
1741 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1742 lp_build_const_int32(gallivm,
1743 velem->instance_divisor),
1744 "instance_divisor");
1745 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1746 current_instance, &ofbit);
1747 }
1748
1749 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1750 buffer_size_adj[j], "");
1751
1752 temp_ptr = lp_build_alloca_undef(gallivm,
1753 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1754
1755 lp_build_if(&if_ctx, gallivm, ofbit);
1756 {
1757 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1758 }
1759 lp_build_else(&if_ctx);
1760 {
1761 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1762 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1763 }
1764 lp_build_endif(&if_ctx);
1765 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1766
1767 if (0) {
1768 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1769 lp_build_const_int32(gallivm, j),
1770 vb_index, vb_stride[j]);
1771 lp_build_printf(gallivm,
1772 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1773 vb_buffer_offset, src_offset, buf_offset);
1774 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1775 buffer_size, bsize);
1776 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1777 }
1778 }
1779 }
1780
1781 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1782 {
1783 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1784 LLVMValueRef io;
1785 LLVMValueRef clipmask; /* holds the clipmask value */
1786 LLVMValueRef true_index_array, index_store;
1787 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1788
1789 io_itr = lp_loop.counter;
1790
1791 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1792 #if DEBUG_STORE
1793 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1794 io_itr, io, lp_loop.counter);
1795 #endif
1796
1797 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1798 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1799
1800 /*
1801 * Limit indices to fetch_max, otherwise might try to access indices
1802 * beyond index buffer (or rather vsplit elt buffer) size.
1803 * Could probably safely (?) skip this for non-indexed draws and
1804 * simplify things minimally (by removing it could combine the ind_vec
1805 * and start_vec adds). I think the only effect for non-indexed draws will
1806 * be that for the invalid elements they will be all fetched from the
1807 * same location as the last valid one, but noone should really care.
1808 */
1809 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1810
1811 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1812
1813 lp_build_if(&if_ctx, gallivm, have_elts);
1814 {
1815 /*
1816 * Note: you'd expect some comparison/clamp against fetch_elt_max
1817 * here.
1818 * There used to be one here but it was incorrect: overflow was
1819 * detected if index > fetch_elt_max - but the correct condition
1820 * would be index >= fetch_elt_max (since this is just size of elts
1821 * buffer / element size).
1822 * Using the correct condition however will cause failures - due to
1823 * vsplit/vcache code which rebases indices. So, as an example, if
1824 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1825 * replace all invalid indices with 0 - which in case of elt_bias
1826 * not being zero will get a different fetch index than the valid
1827 * index 0. So, just rely on vsplit code preventing out-of-bounds
1828 * fetches. This is also why it's safe to do elts fetch even if there
1829 * was no index buffer bound - the real buffer is never seen here, at
1830 * least not if there are index buffer overflows...
1831 */
1832
1833 /*
1834 * XXX should not have to do this, as scale can be handled
1835 * natively by loads (hits asserts though).
1836 */
1837 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1838 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1839 LLVMPointerType(LLVMInt8TypeInContext(context),
1840 0), "");
1841 tmp = lp_build_gather(gallivm, vs_type.length,
1842 32, bld.type, TRUE,
1843 fetch_elts, tmp, FALSE);
1844 LLVMBuildStore(builder, tmp, index_store);
1845 }
1846 lp_build_else(&if_ctx);
1847 {
1848 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1849 LLVMBuildStore(builder, tmp, index_store);
1850 }
1851 lp_build_endif(&if_ctx);
1852
1853 true_index_array = LLVMBuildLoad(builder, index_store, "");
1854
1855 for (j = 0; j < key->nr_vertex_elements; ++j) {
1856 struct pipe_vertex_element *velem = &key->vertex_element[j];
1857 const struct util_format_description *format_desc =
1858 util_format_description(velem->src_format);
1859
1860 if (format_desc->format == PIPE_FORMAT_NONE) {
1861 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1862 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1863 }
1864 }
1865 else if (velem->instance_divisor) {
1866 fetch_instanced(gallivm, format_desc, vs_type,
1867 vb_stride[j], map_ptr[j],
1868 buffer_size_adj[j],
1869 inputs[j], instance_index[j]);
1870 }
1871 else {
1872 fetch_vector(gallivm, format_desc, vs_type,
1873 vb_stride[j], map_ptr[j],
1874 buffer_size_adj[j],
1875 inputs[j], true_index_array);
1876 }
1877 }
1878
1879 /* In the paths with elts vertex id has to be unaffected by the
1880 * index bias and because indices inside our elements array have
1881 * already had index bias applied we need to subtract it here to
1882 * get back to the original index.
1883 * in the linear paths vertex id has to be unaffected by the
1884 * original start index and because we abuse the 'start' variable
1885 * to either represent the actual start index or the index at which
1886 * the primitive was split (we split rendering into chunks of at
1887 * most 4095-vertices) we need to back out the original start
1888 * index out of our vertex id here.
1889 */
1890 system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
1891 vertex_id_offset);
1892 system_values.vertex_id = true_index_array;
1893 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1894 system_values.basevertex, "");
1895
1896 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1897 generate_vs(variant,
1898 builder,
1899 vs_type,
1900 outputs,
1901 ptr_aos,
1902 &system_values,
1903 context_ptr,
1904 sampler,
1905 key->clamp_vertex_color);
1906
1907 if (pos != -1 && cv != -1) {
1908 /* store original positions in clip before further manipulation */
1909 store_clip(gallivm, vs_type, io, outputs, pos);
1910
1911 /* do cliptest */
1912 if (enable_cliptest) {
1913 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1914 /* allocate clipmask, assign it integer type */
1915 clipmask = generate_clipmask(llvm,
1916 gallivm,
1917 vs_type,
1918 outputs,
1919 key,
1920 context_ptr, &have_clipdist);
1921 temp = LLVMBuildOr(builder, clipmask, temp, "");
1922 /* store temporary clipping boolean value */
1923 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1924 }
1925 else {
1926 clipmask = blduivec.zero;
1927 }
1928
1929 /* do viewport mapping */
1930 if (!bypass_viewport) {
1931 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1932 }
1933 }
1934 else {
1935 clipmask = blduivec.zero;
1936 }
1937
1938 /* store clipmask in vertex header,
1939 * original positions in clip
1940 * and transformed positions in data
1941 */
1942 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1943 vs_info->num_outputs, vs_type,
1944 enable_cliptest && key->need_edgeflags);
1945 }
1946 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1947
1948 sampler->destroy(sampler);
1949
1950 /* return clipping boolean value for function */
1951 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
1952 enable_cliptest && key->need_edgeflags);
1953
1954 LLVMBuildRet(builder, ret);
1955
1956 gallivm_verify_function(gallivm, variant_func);
1957 }
1958
1959
1960 struct draw_llvm_variant_key *
1961 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1962 {
1963 unsigned i;
1964 struct draw_llvm_variant_key *key;
1965 struct draw_sampler_static_state *draw_sampler;
1966
1967 key = (struct draw_llvm_variant_key *)store;
1968
1969 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
1970
1971 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1972
1973 /* will have to rig this up properly later */
1974 key->clip_xy = llvm->draw->clip_xy;
1975 key->clip_z = llvm->draw->clip_z;
1976 key->clip_user = llvm->draw->clip_user;
1977 key->bypass_viewport = llvm->draw->bypass_viewport;
1978 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1979 /* XXX assumes edgeflag output not at 0 */
1980 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1981 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1982 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1983 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1984
1985 /* All variants of this shader will have the same value for
1986 * nr_samplers. Not yet trying to compact away holes in the
1987 * sampler array.
1988 */
1989 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1990 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1991 key->nr_sampler_views =
1992 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1993 }
1994 else {
1995 key->nr_sampler_views = key->nr_samplers;
1996 }
1997
1998 /* Presumably all variants of the shader should have the same
1999 * number of vertex elements - ie the number of shader inputs.
2000 * NOTE: we NEED to store the needed number of needed inputs
2001 * here, not the number of provided elements to match keysize
2002 * (and the offset of sampler state in the key).
2003 * If we have excess number of vertex elements, this is valid,
2004 * but the excess ones don't matter.
2005 * If we don't have enough vertex elements (which looks not really
2006 * valid but we'll handle it gracefully) fill out missing ones with
2007 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2008 */
2009 key->nr_vertex_elements =
2010 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2011
2012 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2013 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2014 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2015 memset(key->vertex_element, 0,
2016 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2017 }
2018 memcpy(key->vertex_element,
2019 llvm->draw->pt.vertex_element,
2020 sizeof(struct pipe_vertex_element) *
2021 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2022
2023 draw_sampler = draw_llvm_variant_key_samplers(key);
2024 memset(draw_sampler, 0,
2025 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2026
2027 for (i = 0 ; i < key->nr_samplers; i++) {
2028 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2029 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2030 }
2031 for (i = 0 ; i < key->nr_sampler_views; i++) {
2032 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2033 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2034 }
2035
2036 return key;
2037 }
2038
2039
2040 void
2041 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2042 {
2043 unsigned i;
2044 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2045
2046 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2047 debug_printf("clip_xy = %u\n", key->clip_xy);
2048 debug_printf("clip_z = %u\n", key->clip_z);
2049 debug_printf("clip_user = %u\n", key->clip_user);
2050 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2051 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2052 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2053 debug_printf("has_gs = %u\n", key->has_gs);
2054 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2055
2056 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2057 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2058 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2059 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2060 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2061 }
2062
2063 for (i = 0 ; i < key->nr_sampler_views; i++) {
2064 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2065 }
2066 }
2067
2068
2069 void
2070 draw_llvm_set_mapped_texture(struct draw_context *draw,
2071 unsigned shader_stage,
2072 unsigned sview_idx,
2073 uint32_t width, uint32_t height, uint32_t depth,
2074 uint32_t first_level, uint32_t last_level,
2075 const void *base_ptr,
2076 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2077 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2078 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2079 {
2080 unsigned j;
2081 struct draw_jit_texture *jit_tex;
2082
2083 assert(shader_stage == PIPE_SHADER_VERTEX ||
2084 shader_stage == PIPE_SHADER_GEOMETRY);
2085
2086 if (shader_stage == PIPE_SHADER_VERTEX) {
2087 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2088
2089 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2090 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2091 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2092
2093 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2094 } else {
2095 assert(0);
2096 return;
2097 }
2098
2099 jit_tex->width = width;
2100 jit_tex->height = height;
2101 jit_tex->depth = depth;
2102 jit_tex->first_level = first_level;
2103 jit_tex->last_level = last_level;
2104 jit_tex->base = base_ptr;
2105
2106 for (j = first_level; j <= last_level; j++) {
2107 jit_tex->mip_offsets[j] = mip_offsets[j];
2108 jit_tex->row_stride[j] = row_stride[j];
2109 jit_tex->img_stride[j] = img_stride[j];
2110 }
2111 }
2112
2113
2114 void
2115 draw_llvm_set_sampler_state(struct draw_context *draw,
2116 unsigned shader_type)
2117 {
2118 unsigned i;
2119
2120 if (shader_type == PIPE_SHADER_VERTEX) {
2121 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2122 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2123
2124 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2125 const struct pipe_sampler_state *s
2126 = draw->samplers[PIPE_SHADER_VERTEX][i];
2127 jit_sam->min_lod = s->min_lod;
2128 jit_sam->max_lod = s->max_lod;
2129 jit_sam->lod_bias = s->lod_bias;
2130 COPY_4V(jit_sam->border_color, s->border_color.f);
2131 }
2132 }
2133 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2134 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2135 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2136
2137 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2138 const struct pipe_sampler_state *s
2139 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2140 jit_sam->min_lod = s->min_lod;
2141 jit_sam->max_lod = s->max_lod;
2142 jit_sam->lod_bias = s->lod_bias;
2143 COPY_4V(jit_sam->border_color, s->border_color.f);
2144 }
2145 }
2146 }
2147 }
2148
2149
2150 void
2151 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2152 {
2153 struct draw_llvm *llvm = variant->llvm;
2154
2155 gallivm_destroy(variant->gallivm);
2156
2157 remove_from_list(&variant->list_item_local);
2158 variant->shader->variants_cached--;
2159 remove_from_list(&variant->list_item_global);
2160 llvm->nr_variants--;
2161 FREE(variant);
2162 }
2163
2164
2165 /**
2166 * Create LLVM types for various structures.
2167 */
2168 static void
2169 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2170 {
2171 struct gallivm_state *gallivm = var->gallivm;
2172 LLVMTypeRef texture_type, sampler_type, context_type;
2173
2174 texture_type = create_jit_texture_type(gallivm, "texture");
2175 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2176
2177 context_type = create_gs_jit_context_type(gallivm,
2178 var->shader->base.vector_length,
2179 texture_type, sampler_type,
2180 "draw_gs_jit_context");
2181 var->context_ptr_type = LLVMPointerType(context_type, 0);
2182
2183 var->input_array_type = create_gs_jit_input_type(gallivm);
2184 }
2185
2186 static LLVMTypeRef
2187 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2188 {
2189 if (!variant->context_ptr_type)
2190 create_gs_jit_types(variant);
2191 return variant->context_ptr_type;
2192 }
2193
2194 static LLVMValueRef
2195 generate_mask_value(struct draw_gs_llvm_variant *variant,
2196 struct lp_type gs_type)
2197 {
2198 struct gallivm_state *gallivm = variant->gallivm;
2199 LLVMBuilderRef builder = gallivm->builder;
2200 struct lp_type mask_type = lp_int_type(gs_type);
2201 LLVMValueRef num_prims;
2202 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2203 unsigned i;
2204
2205 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2206 variant->num_prims);
2207 for (i = 0; i < gs_type.length; i++) {
2208 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2209 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2210 }
2211 mask_val = lp_build_compare(gallivm, mask_type,
2212 PIPE_FUNC_GREATER, num_prims, mask_val);
2213
2214 return mask_val;
2215 }
2216
2217 static void
2218 draw_gs_llvm_generate(struct draw_llvm *llvm,
2219 struct draw_gs_llvm_variant *variant)
2220 {
2221 struct gallivm_state *gallivm = variant->gallivm;
2222 LLVMContextRef context = gallivm->context;
2223 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2224 LLVMTypeRef arg_types[7];
2225 LLVMTypeRef func_type;
2226 LLVMValueRef variant_func;
2227 LLVMValueRef context_ptr;
2228 LLVMValueRef prim_id_ptr;
2229 LLVMBasicBlockRef block;
2230 LLVMBuilderRef builder;
2231 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2232 struct lp_build_sampler_soa *sampler = 0;
2233 struct lp_build_context bld;
2234 struct lp_bld_tgsi_system_values system_values;
2235 char func_name[64];
2236 struct lp_type gs_type;
2237 unsigned i;
2238 struct draw_gs_llvm_iface gs_iface;
2239 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2240 LLVMValueRef consts_ptr, num_consts_ptr;
2241 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2242 struct lp_build_mask_context mask;
2243 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2244 unsigned vector_length = variant->shader->base.vector_length;
2245
2246 memset(&system_values, 0, sizeof(system_values));
2247
2248 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2249 variant->shader->variants_cached);
2250
2251 assert(variant->vertex_header_ptr_type);
2252
2253 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2254 arg_types[1] = variant->input_array_type; /* input */
2255 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2256 arg_types[3] = int32_type; /* num_prims */
2257 arg_types[4] = int32_type; /* instance_id */
2258 arg_types[5] = LLVMPointerType(
2259 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2260 arg_types[6] = int32_type;
2261
2262 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2263
2264 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2265
2266 variant->function = variant_func;
2267
2268 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2269
2270 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2271 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2272 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2273
2274 context_ptr = LLVMGetParam(variant_func, 0);
2275 input_array = LLVMGetParam(variant_func, 1);
2276 io_ptr = LLVMGetParam(variant_func, 2);
2277 num_prims = LLVMGetParam(variant_func, 3);
2278 system_values.instance_id = LLVMGetParam(variant_func, 4);
2279 prim_id_ptr = LLVMGetParam(variant_func, 5);
2280 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2281
2282 lp_build_name(context_ptr, "context");
2283 lp_build_name(input_array, "input");
2284 lp_build_name(io_ptr, "io");
2285 lp_build_name(num_prims, "num_prims");
2286 lp_build_name(system_values.instance_id, "instance_id");
2287 lp_build_name(prim_id_ptr, "prim_id_ptr");
2288 lp_build_name(system_values.invocation_id, "invocation_id");
2289
2290 variant->context_ptr = context_ptr;
2291 variant->io_ptr = io_ptr;
2292 variant->num_prims = num_prims;
2293
2294 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2295 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2296 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2297 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2298 gs_iface.input = input_array;
2299 gs_iface.variant = variant;
2300
2301 /*
2302 * Function body
2303 */
2304
2305 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2306 builder = gallivm->builder;
2307 LLVMPositionBuilderAtEnd(builder, block);
2308
2309 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2310
2311 memset(&gs_type, 0, sizeof gs_type);
2312 gs_type.floating = TRUE; /* floating point values */
2313 gs_type.sign = TRUE; /* values are signed */
2314 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2315 gs_type.width = 32; /* 32-bit float */
2316 gs_type.length = vector_length;
2317
2318 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2319 num_consts_ptr =
2320 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2321
2322 /* code generated texture sampling */
2323 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2324
2325 mask_val = generate_mask_value(variant, gs_type);
2326 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2327
2328 if (gs_info->uses_primid) {
2329 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2330 }
2331
2332 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2333 tgsi_dump(tokens, 0);
2334 draw_gs_llvm_dump_variant_key(&variant->key);
2335 }
2336
2337 lp_build_tgsi_soa(variant->gallivm,
2338 tokens,
2339 gs_type,
2340 &mask,
2341 consts_ptr,
2342 num_consts_ptr,
2343 &system_values,
2344 NULL,
2345 outputs,
2346 context_ptr,
2347 NULL,
2348 sampler,
2349 &llvm->draw->gs.geometry_shader->info,
2350 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2351
2352 sampler->destroy(sampler);
2353
2354 lp_build_mask_end(&mask);
2355
2356 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2357
2358 gallivm_verify_function(gallivm, variant_func);
2359 }
2360
2361
2362 struct draw_gs_llvm_variant *
2363 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2364 unsigned num_outputs,
2365 const struct draw_gs_llvm_variant_key *key)
2366 {
2367 struct draw_gs_llvm_variant *variant;
2368 struct llvm_geometry_shader *shader =
2369 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2370 LLVMTypeRef vertex_header;
2371 char module_name[64];
2372
2373 variant = MALLOC(sizeof *variant +
2374 shader->variant_key_size -
2375 sizeof variant->key);
2376 if (!variant)
2377 return NULL;
2378
2379 variant->llvm = llvm;
2380 variant->shader = shader;
2381
2382 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2383 variant->shader->variants_cached);
2384
2385 variant->gallivm = gallivm_create(module_name, llvm->context);
2386
2387 create_gs_jit_types(variant);
2388
2389 memcpy(&variant->key, key, shader->variant_key_size);
2390
2391 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2392
2393 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2394
2395 draw_gs_llvm_generate(llvm, variant);
2396
2397 gallivm_compile_module(variant->gallivm);
2398
2399 variant->jit_func = (draw_gs_jit_func)
2400 gallivm_jit_function(variant->gallivm, variant->function);
2401
2402 gallivm_free_ir(variant->gallivm);
2403
2404 variant->list_item_global.base = variant;
2405 variant->list_item_local.base = variant;
2406 /*variant->no = */shader->variants_created++;
2407 variant->list_item_global.base = variant;
2408
2409 return variant;
2410 }
2411
2412 void
2413 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2414 {
2415 struct draw_llvm *llvm = variant->llvm;
2416
2417 gallivm_destroy(variant->gallivm);
2418
2419 remove_from_list(&variant->list_item_local);
2420 variant->shader->variants_cached--;
2421 remove_from_list(&variant->list_item_global);
2422 llvm->nr_gs_variants--;
2423 FREE(variant);
2424 }
2425
2426 struct draw_gs_llvm_variant_key *
2427 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2428 {
2429 unsigned i;
2430 struct draw_gs_llvm_variant_key *key;
2431 struct draw_sampler_static_state *draw_sampler;
2432
2433 key = (struct draw_gs_llvm_variant_key *)store;
2434
2435 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2436
2437 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2438
2439 /* All variants of this shader will have the same value for
2440 * nr_samplers. Not yet trying to compact away holes in the
2441 * sampler array.
2442 */
2443 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2444 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2445 key->nr_sampler_views =
2446 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2447 }
2448 else {
2449 key->nr_sampler_views = key->nr_samplers;
2450 }
2451
2452 draw_sampler = key->samplers;
2453
2454 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2455
2456 for (i = 0 ; i < key->nr_samplers; i++) {
2457 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2458 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2459 }
2460 for (i = 0 ; i < key->nr_sampler_views; i++) {
2461 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2462 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2463 }
2464
2465 return key;
2466 }
2467
2468 void
2469 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2470 {
2471 unsigned i;
2472 struct draw_sampler_static_state *sampler = key->samplers;
2473
2474 for (i = 0 ; i < key->nr_sampler_views; i++) {
2475 debug_printf("sampler[%i].src_format = %s\n", i,
2476 util_format_name(sampler[i].texture_state.format));
2477 }
2478 }