projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw/translate: fix instancing
[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_logic.h"
37 #include "gallivm/lp_bld_const.h"
38 #include "gallivm/lp_bld_swizzle.h"
39 #include "gallivm/lp_bld_struct.h"
40 #include "gallivm/lp_bld_type.h"
41 #include "gallivm/lp_bld_flow.h"
42 #include "gallivm/lp_bld_debug.h"
43 #include "gallivm/lp_bld_tgsi.h"
44 #include "gallivm/lp_bld_printf.h"
45 #include "gallivm/lp_bld_intr.h"
46 #include "gallivm/lp_bld_init.h"
47 #include "gallivm/lp_bld_type.h"
48 #include "gallivm/lp_bld_pack.h"
49 #include "gallivm/lp_bld_format.h"
50
51 #include "tgsi/tgsi_exec.h"
52 #include "tgsi/tgsi_dump.h"
53
54 #include "util/u_math.h"
55 #include "util/u_pointer.h"
56 #include "util/u_string.h"
57 #include "util/u_simple_list.h"
58
59
60 #define DEBUG_STORE 0
61
62
63 static void
64 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
65 boolean elts);
66
67
68 struct draw_gs_llvm_iface {
69 struct lp_build_tgsi_gs_iface base;
70
71 struct draw_gs_llvm_variant *variant;
72 LLVMValueRef input;
73 };
74
75 static INLINE const struct draw_gs_llvm_iface *
76 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
77 {
78 return (const struct draw_gs_llvm_iface *)iface;
79 }
80
81 /**
82 * Create LLVM type for draw_vertex_buffer.
83 */
84 static LLVMTypeRef
85 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
86 const char *struct_name)
87 {
88 LLVMTargetDataRef target = gallivm->target;
89 LLVMTypeRef dvbuffer_type;
90 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
91 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
92
93 elem_types[DRAW_JIT_DVBUFFER_MAP] =
94 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
95 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
96
97 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
98 Elements(elem_types), 0);
99
100 #if HAVE_LLVM < 0x0300
101 LLVMAddTypeName(gallivm->module, struct_name, dvbuffer_type);
102
103 /* Make sure the target's struct layout cache doesn't return
104 * stale/invalid data.
105 */
106 LLVMInvalidateStructLayout(gallivm->target, dvbuffer_type);
107 #endif
108
109 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
110 target, dvbuffer_type,
111 DRAW_JIT_DVBUFFER_MAP);
112 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
113 target, dvbuffer_type,
114 DRAW_JIT_DVBUFFER_SIZE);
115
116 return dvbuffer_type;
117 }
118
119 /**
120 * Create LLVM type for struct draw_jit_texture
121 */
122 static LLVMTypeRef
123 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
124 {
125 LLVMTargetDataRef target = gallivm->target;
126 LLVMTypeRef texture_type;
127 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
128 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
129
130 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
131 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
132 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
133 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
134 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
135 elem_types[DRAW_JIT_TEXTURE_BASE] =
136 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
137 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
138 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
139 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
140 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
141
142 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
143 Elements(elem_types), 0);
144
145 #if HAVE_LLVM < 0x0300
146 LLVMAddTypeName(gallivm->module, struct_name, texture_type);
147
148 /* Make sure the target's struct layout cache doesn't return
149 * stale/invalid data.
150 */
151 LLVMInvalidateStructLayout(gallivm->target, texture_type);
152 #endif
153
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_WIDTH);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_HEIGHT);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_DEPTH);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
164 target, texture_type,
165 DRAW_JIT_TEXTURE_FIRST_LEVEL);
166 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
167 target, texture_type,
168 DRAW_JIT_TEXTURE_LAST_LEVEL);
169 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
170 target, texture_type,
171 DRAW_JIT_TEXTURE_BASE);
172 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
173 target, texture_type,
174 DRAW_JIT_TEXTURE_ROW_STRIDE);
175 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
176 target, texture_type,
177 DRAW_JIT_TEXTURE_IMG_STRIDE);
178 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
179 target, texture_type,
180 DRAW_JIT_TEXTURE_MIP_OFFSETS);
181
182 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
183
184 return texture_type;
185 }
186
187
188 /**
189 * Create LLVM type for struct draw_jit_sampler
190 */
191 static LLVMTypeRef
192 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
193 {
194 LLVMTargetDataRef target = gallivm->target;
195 LLVMTypeRef sampler_type;
196 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
197
198 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
199 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
200 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
201 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
202 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
203
204 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
205 Elements(elem_types), 0);
206
207 #if HAVE_LLVM < 0x0300
208 LLVMAddTypeName(gallivm->module, struct_name, sampler_type);
209
210 /* Make sure the target's struct layout cache doesn't return
211 * stale/invalid data.
212 */
213 LLVMInvalidateStructLayout(gallivm->target, sampler_type);
214 #endif
215
216 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
217 target, sampler_type,
218 DRAW_JIT_SAMPLER_MIN_LOD);
219 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
220 target, sampler_type,
221 DRAW_JIT_SAMPLER_MAX_LOD);
222 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
223 target, sampler_type,
224 DRAW_JIT_SAMPLER_LOD_BIAS);
225 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
226 target, sampler_type,
227 DRAW_JIT_SAMPLER_BORDER_COLOR);
228
229 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
230
231 return sampler_type;
232 }
233
234
235 /**
236 * Create LLVM type for struct draw_jit_context
237 */
238 static LLVMTypeRef
239 create_jit_context_type(struct gallivm_state *gallivm,
240 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
241 const char *struct_name)
242 {
243 LLVMTargetDataRef target = gallivm->target;
244 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
245 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
246 LLVMTypeRef context_type;
247
248 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
249 LP_MAX_TGSI_CONST_BUFFERS);
250 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
251 DRAW_TOTAL_CLIP_PLANES), 0);
252 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
253 elem_types[3] = LLVMArrayType(texture_type,
254 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
255 elem_types[4] = LLVMArrayType(sampler_type,
256 PIPE_MAX_SAMPLERS); /* samplers */
257 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
258 Elements(elem_types), 0);
259 #if HAVE_LLVM < 0x0300
260 LLVMAddTypeName(gallivm->module, struct_name, context_type);
261
262 LLVMInvalidateStructLayout(gallivm->target, context_type);
263 #endif
264
265 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
266 target, context_type, DRAW_JIT_CTX_CONSTANTS);
267 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
268 target, context_type, DRAW_JIT_CTX_PLANES);
269 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
270 target, context_type, DRAW_JIT_CTX_VIEWPORT);
271 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
272 target, context_type,
273 DRAW_JIT_CTX_TEXTURES);
274 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
275 target, context_type,
276 DRAW_JIT_CTX_SAMPLERS);
277 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
278 target, context_type);
279
280 return context_type;
281 }
282
283
284 /**
285 * Create LLVM type for struct draw_gs_jit_context
286 */
287 static LLVMTypeRef
288 create_gs_jit_context_type(struct gallivm_state *gallivm,
289 unsigned vector_length,
290 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
291 const char *struct_name)
292 {
293 LLVMTargetDataRef target = gallivm->target;
294 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
295 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
296 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
297 LLVMTypeRef context_type;
298
299 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
300 LP_MAX_TGSI_CONST_BUFFERS);
301 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
302 DRAW_TOTAL_CLIP_PLANES), 0);
303 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
304
305 elem_types[3] = LLVMArrayType(texture_type,
306 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
307 elem_types[4] = LLVMArrayType(sampler_type,
308 PIPE_MAX_SAMPLERS); /* samplers */
309
310 elem_types[5] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
311 elem_types[6] = LLVMPointerType(LLVMVectorType(int_type,
312 vector_length), 0);
313 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
314 vector_length), 0);
315
316 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
317 Elements(elem_types), 0);
318 #if HAVE_LLVM < 0x0300
319 LLVMAddTypeName(gallivm->module, struct_name, context_type);
320
321 LLVMInvalidateStructLayout(gallivm->target, context_type);
322 #endif
323
324 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
325 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
326 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
327 target, context_type, DRAW_GS_JIT_CTX_PLANES);
328 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
329 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
330 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
331 target, context_type,
332 DRAW_GS_JIT_CTX_TEXTURES);
333 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
334 target, context_type,
335 DRAW_GS_JIT_CTX_SAMPLERS);
336 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
337 target, context_type,
338 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
339 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
340 target, context_type,
341 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
342 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
343 target, context_type,
344 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
345 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
346 target, context_type);
347
348 return context_type;
349 }
350
351
352 static LLVMTypeRef
353 create_gs_jit_input_type(struct gallivm_state *gallivm)
354 {
355 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
356 LLVMTypeRef input_array;
357
358 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
359 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
360 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
361 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
362
363 return input_array;
364 }
365
366 /**
367 * Create LLVM type for struct pipe_vertex_buffer
368 */
369 static LLVMTypeRef
370 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
371 const char *struct_name)
372 {
373 LLVMTargetDataRef target = gallivm->target;
374 LLVMTypeRef elem_types[4];
375 LLVMTypeRef vb_type;
376
377 elem_types[0] =
378 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
379 elem_types[2] =
380 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
381
382 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
383 Elements(elem_types), 0);
384 #if HAVE_LLVM < 0x0300
385 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
386
387 LLVMInvalidateStructLayout(gallivm->target, vb_type);
388 #endif
389
390 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
391 target, vb_type, 0);
392 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
393 target, vb_type, 1);
394
395 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
396
397 return vb_type;
398 }
399
400
401 /**
402 * Create LLVM type for struct vertex_header;
403 */
404 static LLVMTypeRef
405 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
406 {
407 LLVMTargetDataRef target = gallivm->target;
408 LLVMTypeRef elem_types[4];
409 LLVMTypeRef vertex_header;
410 char struct_name[24];
411
412 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
413
414 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
415 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
416 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
417 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
418
419 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
420 Elements(elem_types), 0);
421 #if HAVE_LLVM < 0x0300
422 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
423
424 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
425 #endif
426
427 /* these are bit-fields and we can't take address of them
428 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
429 target, vertex_header,
430 DRAW_JIT_VERTEX_CLIPMASK);
431 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
432 target, vertex_header,
433 DRAW_JIT_VERTEX_EDGEFLAG);
434 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
435 target, vertex_header,
436 DRAW_JIT_VERTEX_PAD);
437 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
438 target, vertex_header,
439 DRAW_JIT_VERTEX_VERTEX_ID);
440 */
441 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
442 target, vertex_header,
443 DRAW_JIT_VERTEX_CLIP);
444 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
445 target, vertex_header,
446 DRAW_JIT_VERTEX_PRE_CLIP_POS);
447 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
448 target, vertex_header,
449 DRAW_JIT_VERTEX_DATA);
450
451 assert(LLVMABISizeOfType(target, vertex_header) ==
452 offsetof(struct vertex_header, data[data_elems]));
453
454 return vertex_header;
455 }
456
457
458 /**
459 * Create LLVM types for various structures.
460 */
461 static void
462 create_jit_types(struct draw_llvm_variant *variant)
463 {
464 struct gallivm_state *gallivm = variant->gallivm;
465 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
466 vb_type;
467
468 texture_type = create_jit_texture_type(gallivm, "texture");
469 sampler_type = create_jit_sampler_type(gallivm, "sampler");
470
471 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
472 "draw_jit_context");
473 variant->context_ptr_type = LLVMPointerType(context_type, 0);
474
475 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
476 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
477
478 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
479 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
480 }
481
482
483 static LLVMTypeRef
484 get_context_ptr_type(struct draw_llvm_variant *variant)
485 {
486 if (!variant->context_ptr_type)
487 create_jit_types(variant);
488 return variant->context_ptr_type;
489 }
490
491
492 static LLVMTypeRef
493 get_buffer_ptr_type(struct draw_llvm_variant *variant)
494 {
495 if (!variant->buffer_ptr_type)
496 create_jit_types(variant);
497 return variant->buffer_ptr_type;
498 }
499
500
501 static LLVMTypeRef
502 get_vb_ptr_type(struct draw_llvm_variant *variant)
503 {
504 if (!variant->vb_ptr_type)
505 create_jit_types(variant);
506 return variant->vb_ptr_type;
507 }
508
509 static LLVMTypeRef
510 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
511 {
512 if (!variant->vertex_header_ptr_type)
513 create_jit_types(variant);
514 return variant->vertex_header_ptr_type;
515 }
516
517
518 /**
519 * Create per-context LLVM info.
520 */
521 struct draw_llvm *
522 draw_llvm_create(struct draw_context *draw)
523 {
524 struct draw_llvm *llvm;
525
526 llvm = CALLOC_STRUCT( draw_llvm );
527 if (!llvm)
528 return NULL;
529
530 lp_build_init();
531
532 llvm->draw = draw;
533
534 llvm->nr_variants = 0;
535 make_empty_list(&llvm->vs_variants_list);
536
537 llvm->nr_gs_variants = 0;
538 make_empty_list(&llvm->gs_variants_list);
539
540 return llvm;
541 }
542
543
544 /**
545 * Free per-context LLVM info.
546 */
547 void
548 draw_llvm_destroy(struct draw_llvm *llvm)
549 {
550 /* XXX free other draw_llvm data? */
551 FREE(llvm);
552 }
553
554
555 /**
556 * Create LLVM-generated code for a vertex shader.
557 */
558 struct draw_llvm_variant *
559 draw_llvm_create_variant(struct draw_llvm *llvm,
560 unsigned num_inputs,
561 const struct draw_llvm_variant_key *key)
562 {
563 struct draw_llvm_variant *variant;
564 struct llvm_vertex_shader *shader =
565 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
566 LLVMTypeRef vertex_header;
567
568 variant = MALLOC(sizeof *variant +
569 shader->variant_key_size -
570 sizeof variant->key);
571 if (variant == NULL)
572 return NULL;
573
574 variant->llvm = llvm;
575
576 variant->gallivm = gallivm_create();
577
578 create_jit_types(variant);
579
580 memcpy(&variant->key, key, shader->variant_key_size);
581
582 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
583
584 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
585
586 draw_llvm_generate(llvm, variant, FALSE); /* linear */
587 draw_llvm_generate(llvm, variant, TRUE); /* elts */
588
589 gallivm_compile_module(variant->gallivm);
590
591 variant->jit_func = (draw_jit_vert_func)
592 gallivm_jit_function(variant->gallivm, variant->function);
593
594 variant->jit_func_elts = (draw_jit_vert_func_elts)
595 gallivm_jit_function(variant->gallivm, variant->function_elts);
596
597 variant->shader = shader;
598 variant->list_item_global.base = variant;
599 variant->list_item_local.base = variant;
600 /*variant->no = */shader->variants_created++;
601 variant->list_item_global.base = variant;
602
603 return variant;
604 }
605
606
607 static void
608 generate_vs(struct draw_llvm_variant *variant,
609 LLVMBuilderRef builder,
610 struct lp_type vs_type,
611 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
612 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
613 const struct lp_bld_tgsi_system_values *system_values,
614 LLVMValueRef context_ptr,
615 struct lp_build_sampler_soa *draw_sampler,
616 boolean clamp_vertex_color)
617 {
618 struct draw_llvm *llvm = variant->llvm;
619 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
620 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(variant->gallivm, context_ptr);
621 struct lp_build_sampler_soa *sampler = 0;
622
623 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
624 tgsi_dump(tokens, 0);
625 draw_llvm_dump_variant_key(&variant->key);
626 }
627
628 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
629 sampler = draw_sampler;
630
631 lp_build_tgsi_soa(variant->gallivm,
632 tokens,
633 vs_type,
634 NULL /*struct lp_build_mask_context *mask*/,
635 consts_ptr,
636 system_values,
637 inputs,
638 outputs,
639 sampler,
640 &llvm->draw->vs.vertex_shader->info,
641 NULL);
642
643 {
644 LLVMValueRef out;
645 unsigned chan, attrib;
646 struct lp_build_context bld;
647 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
648 lp_build_context_init(&bld, variant->gallivm, vs_type);
649
650 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
651 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
652 if (outputs[attrib][chan]) {
653 switch (info->output_semantic_name[attrib]) {
654 case TGSI_SEMANTIC_COLOR:
655 case TGSI_SEMANTIC_BCOLOR:
656 if (clamp_vertex_color) {
657 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
658 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
659 LLVMBuildStore(builder, out, outputs[attrib][chan]);
660 }
661 break;
662 case TGSI_SEMANTIC_FOG:
663 if (chan == 1 || chan == 2)
664 LLVMBuildStore(builder, bld.zero, outputs[attrib][chan]);
665 else if (chan == 3)
666 LLVMBuildStore(builder, bld.one, outputs[attrib][chan]);
667 break;
668 }
669 }
670 }
671 }
672 }
673 }
674
675 static void
676 generate_fetch(struct gallivm_state *gallivm,
677 struct draw_context *draw,
678 LLVMValueRef vbuffers_ptr,
679 LLVMValueRef *res,
680 struct pipe_vertex_element *velem,
681 LLVMValueRef vbuf,
682 LLVMValueRef index,
683 LLVMValueRef instance_id)
684 {
685 const struct util_format_description *format_desc =
686 util_format_description(velem->src_format);
687 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
688 LLVMBuilderRef builder = gallivm->builder;
689 LLVMValueRef indices =
690 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
691 velem->vertex_buffer_index, 0);
692 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
693 &indices, 1, "");
694 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
695 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
696 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
697 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
698 LLVMValueRef stride;
699 LLVMValueRef buffer_overflowed;
700 LLVMValueRef needed_buffer_size;
701 LLVMValueRef temp_ptr =
702 lp_build_alloca(gallivm,
703 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
704 LLVMValueRef ofbit = NULL;
705 struct lp_build_if_state if_ctx;
706
707 if (velem->instance_divisor) {
708 /* Index is equal to the start instance plus the number of current
709 * instance divided by the divisor. In this case we compute it as:
710 * index = start_instance + ((instance_id - start_instance) / divisor)
711 */
712 LLVMValueRef current_instance;
713 index = lp_build_const_int32(gallivm, draw->start_instance);
714 current_instance = LLVMBuildSub(builder, instance_id, index, "");
715 current_instance = LLVMBuildUDiv(builder, current_instance,
716 lp_build_const_int32(gallivm, velem->instance_divisor),
717 "instance_divisor");
718 index = LLVMBuildAdd(builder, index, current_instance, "instance");
719 }
720
721 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
722 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
723 stride = lp_build_uadd_overflow(
724 gallivm, stride,
725 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
726 needed_buffer_size = lp_build_uadd_overflow(
727 gallivm, stride,
728 lp_build_const_int32(gallivm,
729 util_format_get_blocksize(velem->src_format)),
730 &ofbit);
731
732 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
733 needed_buffer_size, buffer_size,
734 "buffer_overflowed");
735 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
736 #if 0
737 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
738 index, vb_stride);
739 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
740 vb_buffer_offset,
741 lp_build_const_int32(gallivm, velem->src_offset));
742 lp_build_print_value(gallivm, " blocksize = ",
743 lp_build_const_int32(
744 gallivm,
745 util_format_get_blocksize(velem->src_format)));
746 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
747 lp_build_printf(gallivm, " stride = %u\n", stride);
748 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
749 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
750 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
751 #endif
752
753 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
754 {
755 LLVMValueRef val =
756 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
757 LLVMBuildStore(builder, val, temp_ptr);
758 }
759 lp_build_else(&if_ctx);
760 {
761 LLVMValueRef val;
762 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
763
764 val = lp_build_fetch_rgba_aos(gallivm,
765 format_desc,
766 lp_float32_vec4_type(),
767 map_ptr,
768 zero, zero, zero);
769 LLVMBuildStore(builder, val, temp_ptr);
770 }
771 lp_build_endif(&if_ctx);
772
773 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
774 }
775
776 static void
777 convert_to_soa(struct gallivm_state *gallivm,
778 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
779 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
780 unsigned num_attribs, const struct lp_type soa_type)
781 {
782 unsigned i, j, k;
783 struct lp_type aos_channel_type = soa_type;
784
785 debug_assert(TGSI_NUM_CHANNELS == 4);
786 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
787
788 aos_channel_type.length >>= 1;
789
790 for (i = 0; i < num_attribs; ++i) {
791 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
792 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
793
794 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
795 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
796
797 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
798
799 for (k = 0; k < pixels_per_channel; ++k) {
800 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
801 }
802
803 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
804 }
805
806 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
807 }
808 }
809
810
811 static void
812 store_aos(struct gallivm_state *gallivm,
813 LLVMValueRef io_ptr,
814 LLVMValueRef index,
815 LLVMValueRef value)
816 {
817 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
818 LLVMBuilderRef builder = gallivm->builder;
819 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
820 LLVMValueRef indices[3];
821
822 indices[0] = lp_build_const_int32(gallivm, 0);
823 indices[1] = index;
824 indices[2] = lp_build_const_int32(gallivm, 0);
825
826 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
827 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
828
829 #if DEBUG_STORE
830 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
831 #endif
832
833 /* Unaligned store due to the vertex header */
834 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
835 }
836
837 /**
838 * Adjust the mask to architecture endianess. The mask will the store in struct:
839 *
840 * struct vertex_header {
841 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
842 * unsigned edgeflag:1;
843 * unsigned have_clipdist:1;
844 * unsigned vertex_id:16;
845 * [...]
846 * }
847 *
848 * On little-endian machine nothing needs to done, however on bit-endian machine
849 * the mask's fields need to be adjusted with the algorithm:
850 *
851 * uint32_t reverse (uint32_t x)
852 * {
853 * return (x >> 16) | // vertex_id
854 * ((x & 0x3fff) << 18) | // clipmask
855 * ((x & 0x4000) << 3) | // have_clipdist
856 * ((x & 0x8000) << 1); // edgeflag
857 * }
858 */
859 static LLVMValueRef
860 adjust_mask(struct gallivm_state *gallivm,
861 LLVMValueRef mask)
862 {
863 #ifdef PIPE_ARCH_BIG_ENDIAN
864 LLVMBuilderRef builder = gallivm->builder;
865 LLVMValueRef vertex_id;
866 LLVMValueRef clipmask;
867 LLVMValueRef have_clipdist;
868 LLVMValueRef edgeflag;
869
870 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
871 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
872 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
873 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
874 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
875 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
876 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
877
878 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
879 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
880 mask = LLVMBuildOr(builder, mask, edgeflag, "");
881 #endif
882 return mask;
883 }
884
885 static void
886 store_aos_array(struct gallivm_state *gallivm,
887 struct lp_type soa_type,
888 LLVMValueRef io_ptr,
889 LLVMValueRef *indices,
890 LLVMValueRef* aos,
891 int attrib,
892 int num_outputs,
893 LLVMValueRef clipmask,
894 boolean have_clipdist)
895 {
896 LLVMBuilderRef builder = gallivm->builder;
897 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
898 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
899 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
900 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
901 int vector_length = soa_type.length;
902 int i;
903
904 debug_assert(TGSI_NUM_CHANNELS == 4);
905
906 for (i = 0; i < vector_length; i++) {
907 linear_inds[i] = lp_build_const_int32(gallivm, i);
908 if (indices) {
909 inds[i] = indices[i];
910 } else {
911 inds[i] = linear_inds[i];
912 }
913 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
914 }
915
916 if (attrib == 0) {
917 /* store vertex header for each of the n vertices */
918 LLVMValueRef val, cliptmp;
919 int vertex_id_pad_edgeflag;
920
921 /* If this assertion fails, it means we need to update the bit twidding
922 * code here. See struct vertex_header in draw_private.h.
923 */
924 assert(DRAW_TOTAL_CLIP_PLANES==14);
925 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
926 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
927 if (have_clipdist)
928 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
929 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
930 /* OR with the clipmask */
931 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
932 for (i = 0; i < vector_length; i++) {
933 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
934 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
935 val = adjust_mask(gallivm, val);
936 LLVMBuildStore(builder, val, id_ptr);
937 #if DEBUG_STORE
938 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
939 io_ptrs[i], inds[i], val);
940 #endif
941 }
942 }
943
944 /* store for each of the n vertices */
945 for (i = 0; i < vector_length; i++) {
946 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
947 }
948 }
949
950
951 static void
952 convert_to_aos(struct gallivm_state *gallivm,
953 LLVMValueRef io,
954 LLVMValueRef *indices,
955 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
956 LLVMValueRef clipmask,
957 int num_outputs,
958 struct lp_type soa_type,
959 boolean have_clipdist)
960 {
961 LLVMBuilderRef builder = gallivm->builder;
962 unsigned chan, attrib, i;
963
964 #if DEBUG_STORE
965 lp_build_printf(gallivm, " # storing begin\n");
966 #endif
967 for (attrib = 0; attrib < num_outputs; ++attrib) {
968 LLVMValueRef soa[TGSI_NUM_CHANNELS];
969 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
970 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
971 if (outputs[attrib][chan]) {
972 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
973 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
974 #if DEBUG_STORE
975 lp_build_printf(gallivm, "output %d : %d ",
976 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
977 attrib, 0),
978 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
979 chan, 0));
980 lp_build_print_value(gallivm, "val = ", out);
981 #endif
982 soa[chan] = out;
983 }
984 else {
985 soa[chan] = 0;
986 }
987 }
988
989
990 if (soa_type.length == TGSI_NUM_CHANNELS) {
991 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
992 } else {
993 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
994
995 for (i = 0; i < soa_type.length; ++i) {
996 aos[i] = lp_build_extract_range(gallivm,
997 soa[i % TGSI_NUM_CHANNELS],
998 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
999 TGSI_NUM_CHANNELS);
1000 }
1001 }
1002
1003 store_aos_array(gallivm,
1004 soa_type,
1005 io, indices,
1006 aos,
1007 attrib,
1008 num_outputs,
1009 clipmask, have_clipdist);
1010 }
1011 #if DEBUG_STORE
1012 lp_build_printf(gallivm, " # storing end\n");
1013 #endif
1014 }
1015
1016
1017 /**
1018 * Stores original vertex positions in clip coordinates
1019 */
1020 static void
1021 store_clip(struct gallivm_state *gallivm,
1022 const struct lp_type vs_type,
1023 LLVMValueRef io_ptr,
1024 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1025 boolean pre_clip_pos, int idx)
1026 {
1027 LLVMBuilderRef builder = gallivm->builder;
1028 LLVMValueRef soa[4];
1029 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1030 LLVMValueRef indices[2];
1031 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1032 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1033 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1034 int i, j;
1035
1036 indices[0] =
1037 indices[1] = lp_build_const_int32(gallivm, 0);
1038
1039 for (i = 0; i < vs_type.length; i++) {
1040 inds[i] = lp_build_const_int32(gallivm, i);
1041 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1042 }
1043
1044 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1045 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1046 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1047 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1048
1049 if (!pre_clip_pos) {
1050 for (i = 0; i < vs_type.length; i++) {
1051 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1052 }
1053 } else {
1054 for (i = 0; i < vs_type.length; i++) {
1055 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1056 }
1057 }
1058
1059 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1060 for (i = 0; i < vs_type.length; ++i) {
1061 aos[i] = lp_build_extract_range(gallivm,
1062 soa[i % TGSI_NUM_CHANNELS],
1063 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1064 TGSI_NUM_CHANNELS);
1065 }
1066
1067 for (j = 0; j < vs_type.length; j++) {
1068 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
1069 LLVMValueRef clip_ptr;
1070
1071 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1072 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1073
1074 /* Unaligned store */
1075 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1076 }
1077 }
1078
1079
1080 /**
1081 * Transforms the outputs for viewport mapping
1082 */
1083 static void
1084 generate_viewport(struct draw_llvm_variant *variant,
1085 LLVMBuilderRef builder,
1086 struct lp_type vs_type,
1087 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1088 LLVMValueRef context_ptr)
1089 {
1090 int i;
1091 struct gallivm_state *gallivm = variant->gallivm;
1092 struct lp_type f32_type = vs_type;
1093 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1094 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1095 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1096 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1097 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1098
1099 /* for 1/w convention*/
1100 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1101 LLVMBuildStore(builder, out3, outputs[pos][3]);
1102
1103 /* Viewport Mapping */
1104 for (i=0; i<3; i++) {
1105 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1106 LLVMValueRef scale;
1107 LLVMValueRef trans;
1108 LLVMValueRef scale_i;
1109 LLVMValueRef trans_i;
1110 LLVMValueRef index;
1111
1112 index = lp_build_const_int32(gallivm, i);
1113 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1114
1115 index = lp_build_const_int32(gallivm, i+4);
1116 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1117
1118 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1119 LLVMBuildLoad(builder, scale_i, "scale"));
1120 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1121 LLVMBuildLoad(builder, trans_i, "trans"));
1122
1123 /* divide by w */
1124 out = LLVMBuildFMul(builder, out, out3, "");
1125 /* mult by scale */
1126 out = LLVMBuildFMul(builder, out, scale, "");
1127 /* add translation */
1128 out = LLVMBuildFAdd(builder, out, trans, "");
1129
1130 /* store transformed outputs */
1131 LLVMBuildStore(builder, out, outputs[pos][i]);
1132 }
1133
1134 }
1135
1136
1137 /**
1138 * Returns clipmask as nxi32 bitmask for the n vertices
1139 */
1140 static LLVMValueRef
1141 generate_clipmask(struct draw_llvm *llvm,
1142 struct gallivm_state *gallivm,
1143 struct lp_type vs_type,
1144 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1145 boolean clip_xy,
1146 boolean clip_z,
1147 boolean clip_user,
1148 boolean clip_halfz,
1149 unsigned ucp_enable,
1150 LLVMValueRef context_ptr,
1151 boolean *have_clipdist)
1152 {
1153 LLVMBuilderRef builder = gallivm->builder;
1154 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1155 LLVMValueRef test, temp;
1156 LLVMValueRef zero, shift;
1157 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1158 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1159 LLVMValueRef plane1, planes, plane_ptr, sum;
1160 struct lp_type f32_type = vs_type;
1161 struct lp_type i32_type = lp_int_type(vs_type);
1162 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1163 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1164 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1165 bool have_cd = false;
1166 unsigned cd[2];
1167
1168 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1169 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1170
1171 if (cd[0] != pos || cd[1] != pos)
1172 have_cd = true;
1173
1174 if (num_written_clipdistance && !clip_user) {
1175 clip_user = true;
1176 ucp_enable = (1 << num_written_clipdistance) - 1;
1177 }
1178
1179 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1180 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1181 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1182 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1183
1184 /*
1185 * load clipvertex and position from correct locations.
1186 * if they are the same just load them once.
1187 */
1188 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1189 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1190 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1191 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1192
1193 if (clip_user && cv != pos) {
1194 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1195 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1196 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1197 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1198 } else {
1199 cv_x = pos_x;
1200 cv_y = pos_y;
1201 cv_z = pos_z;
1202 cv_w = pos_w;
1203 }
1204
1205 /* Cliptest, for hardwired planes */
1206 if (clip_xy) {
1207 /* plane 1 */
1208 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1209 temp = shift;
1210 test = LLVMBuildAnd(builder, test, temp, "");
1211 mask = test;
1212
1213 /* plane 2 */
1214 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1215 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1216 temp = LLVMBuildShl(builder, temp, shift, "");
1217 test = LLVMBuildAnd(builder, test, temp, "");
1218 mask = LLVMBuildOr(builder, mask, test, "");
1219
1220 /* plane 3 */
1221 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1222 temp = LLVMBuildShl(builder, temp, shift, "");
1223 test = LLVMBuildAnd(builder, test, temp, "");
1224 mask = LLVMBuildOr(builder, mask, test, "");
1225
1226 /* plane 4 */
1227 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1228 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1229 temp = LLVMBuildShl(builder, temp, shift, "");
1230 test = LLVMBuildAnd(builder, test, temp, "");
1231 mask = LLVMBuildOr(builder, mask, test, "");
1232 }
1233
1234 if (clip_z) {
1235 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1236 if (clip_halfz) {
1237 /* plane 5 */
1238 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1239 test = LLVMBuildAnd(builder, test, temp, "");
1240 mask = LLVMBuildOr(builder, mask, test, "");
1241 }
1242 else {
1243 /* plane 5 */
1244 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1245 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1246 test = LLVMBuildAnd(builder, test, temp, "");
1247 mask = LLVMBuildOr(builder, mask, test, "");
1248 }
1249 /* plane 6 */
1250 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1251 temp = LLVMBuildShl(builder, temp, shift, "");
1252 test = LLVMBuildAnd(builder, test, temp, "");
1253 mask = LLVMBuildOr(builder, mask, test, "");
1254 }
1255
1256 if (clip_user) {
1257 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1258 LLVMValueRef indices[3];
1259
1260 /* userclip planes */
1261 while (ucp_enable) {
1262 unsigned plane_idx = ffs(ucp_enable)-1;
1263 ucp_enable &= ~(1 << plane_idx);
1264 plane_idx += 6;
1265
1266 if (have_cd && num_written_clipdistance) {
1267 LLVMValueRef clipdist;
1268 int i;
1269 i = plane_idx - 6;
1270
1271 *have_clipdist = TRUE;
1272 if (i < 4) {
1273 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1274 } else {
1275 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1276 }
1277 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1278 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1279 test = LLVMBuildAnd(builder, test, temp, "");
1280 mask = LLVMBuildOr(builder, mask, test, "");
1281 } else {
1282 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1283 indices[0] = lp_build_const_int32(gallivm, 0);
1284 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1285
1286 indices[2] = lp_build_const_int32(gallivm, 0);
1287 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1288 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1289 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1290 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1291
1292 indices[2] = lp_build_const_int32(gallivm, 1);
1293 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1294 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1295 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1296 test = LLVMBuildFMul(builder, planes, cv_y, "");
1297 sum = LLVMBuildFAdd(builder, sum, test, "");
1298
1299 indices[2] = lp_build_const_int32(gallivm, 2);
1300 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1301 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1302 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1303 test = LLVMBuildFMul(builder, planes, cv_z, "");
1304 sum = LLVMBuildFAdd(builder, sum, test, "");
1305
1306 indices[2] = lp_build_const_int32(gallivm, 3);
1307 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1308 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1309 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1310 test = LLVMBuildFMul(builder, planes, cv_w, "");
1311 sum = LLVMBuildFAdd(builder, sum, test, "");
1312
1313 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1314 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1315 test = LLVMBuildAnd(builder, test, temp, "");
1316 mask = LLVMBuildOr(builder, mask, test, "");
1317 }
1318 }
1319 }
1320 return mask;
1321 }
1322
1323
1324 /**
1325 * Returns boolean if any clipping has occurred
1326 * Used zero/non-zero i32 value to represent boolean
1327 */
1328 static LLVMValueRef
1329 clipmask_booli32(struct gallivm_state *gallivm,
1330 const struct lp_type vs_type,
1331 LLVMValueRef clipmask_bool_ptr)
1332 {
1333 LLVMBuilderRef builder = gallivm->builder;
1334 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1335 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1336 LLVMValueRef ret = LLVMConstNull(int32_type);
1337 LLVMValueRef temp;
1338 int i;
1339
1340 /*
1341 * Can do this with log2(vector length) pack instructions and one extract
1342 * (as we don't actually need a or) with sse2 which would be way better.
1343 */
1344 for (i=0; i < vs_type.length; i++) {
1345 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1346 lp_build_const_int32(gallivm, i) , "");
1347 ret = LLVMBuildOr(builder, ret, temp, "");
1348 }
1349 return ret;
1350 }
1351
1352 static LLVMValueRef
1353 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1354 struct lp_build_tgsi_context * bld_base,
1355 boolean is_indirect,
1356 LLVMValueRef vertex_index,
1357 LLVMValueRef attrib_index,
1358 LLVMValueRef swizzle_index)
1359 {
1360 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1361 struct gallivm_state *gallivm = bld_base->base.gallivm;
1362 LLVMBuilderRef builder = gallivm->builder;
1363 LLVMValueRef indices[3];
1364 LLVMValueRef res;
1365 struct lp_type type = bld_base->base.type;
1366
1367 if (is_indirect) {
1368 int i;
1369 res = bld_base->base.zero;
1370 for (i = 0; i < type.length; ++i) {
1371 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1372 LLVMValueRef vert_chan_index = LLVMBuildExtractElement(builder,
1373 vertex_index, idx, "");
1374 LLVMValueRef channel_vec, value;
1375 indices[0] = vert_chan_index;
1376 indices[1] = attrib_index;
1377 indices[2] = swizzle_index;
1378
1379 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1380 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1381 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1382
1383 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1384 }
1385 } else {
1386 indices[0] = vertex_index;
1387 indices[1] = attrib_index;
1388 indices[2] = swizzle_index;
1389
1390 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1391 res = LLVMBuildLoad(builder, res, "");
1392 }
1393
1394 return res;
1395 }
1396
1397 static void
1398 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1399 struct lp_build_tgsi_context * bld_base,
1400 LLVMValueRef (*outputs)[4],
1401 LLVMValueRef emitted_vertices_vec)
1402 {
1403 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1404 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1405 struct gallivm_state *gallivm = variant->gallivm;
1406 LLVMBuilderRef builder = gallivm->builder;
1407 struct lp_type gs_type = bld_base->base.type;
1408 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1409 lp_int_type(gs_type), 0);
1410 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1411 LLVMValueRef next_prim_offset =
1412 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1413 LLVMValueRef io = variant->io_ptr;
1414 unsigned i;
1415 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1416
1417 for (i = 0; i < gs_type.length; ++i) {
1418 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1419 LLVMValueRef currently_emitted =
1420 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1421 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1422 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1423 }
1424
1425 convert_to_aos(gallivm, io, indices,
1426 outputs, clipmask,
1427 gs_info->num_outputs, gs_type,
1428 FALSE);
1429 }
1430
1431 static void
1432 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1433 struct lp_build_tgsi_context * bld_base,
1434 LLVMValueRef verts_per_prim_vec,
1435 LLVMValueRef emitted_prims_vec)
1436 {
1437 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1438 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1439 struct gallivm_state *gallivm = variant->gallivm;
1440 LLVMBuilderRef builder = gallivm->builder;
1441 LLVMValueRef prim_lengts_ptr =
1442 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1443 unsigned i;
1444
1445 for (i = 0; i < bld_base->base.type.length; ++i) {
1446 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1447 LLVMValueRef prims_emitted =
1448 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1449 LLVMValueRef store_ptr;
1450 LLVMValueRef num_vertices =
1451 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1452
1453 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1454 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1455 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1456 LLVMBuildStore(builder, num_vertices, store_ptr);
1457 }
1458 }
1459
1460 static void
1461 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1462 struct lp_build_tgsi_context * bld_base,
1463 LLVMValueRef total_emitted_vertices_vec,
1464 LLVMValueRef emitted_prims_vec)
1465 {
1466 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1467 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1468 struct gallivm_state *gallivm = variant->gallivm;
1469 LLVMBuilderRef builder = gallivm->builder;
1470 LLVMValueRef emitted_verts_ptr =
1471 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1472 LLVMValueRef emitted_prims_ptr =
1473 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1474 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1475
1476 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1477 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1478
1479 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1480 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1481 }
1482
1483 static void
1484 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1485 boolean elts)
1486 {
1487 struct gallivm_state *gallivm = variant->gallivm;
1488 LLVMContextRef context = gallivm->context;
1489 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1490 LLVMTypeRef arg_types[9];
1491 unsigned num_arg_types =
1492 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1493 LLVMTypeRef func_type;
1494 LLVMValueRef context_ptr;
1495 LLVMBasicBlockRef block;
1496 LLVMBuilderRef builder;
1497 struct lp_type vs_type;
1498 LLVMValueRef end, start;
1499 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1500 LLVMValueRef stride, step, io_itr;
1501 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1502 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1503 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1504 struct draw_context *draw = llvm->draw;
1505 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1506 unsigned i, j;
1507 struct lp_build_context bld;
1508 struct lp_build_loop_state lp_loop;
1509 const int vector_length = lp_native_vector_width / 32;
1510 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1511 LLVMValueRef fetch_max;
1512 struct lp_build_sampler_soa *sampler = 0;
1513 LLVMValueRef ret, clipmask_bool_ptr;
1514 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1515 struct draw_llvm_variant_key *key = &variant->key;
1516 /* If geometry shader is present we need to skip both the viewport
1517 * transformation and clipping otherwise the inputs to the geometry
1518 * shader will be incorrect.
1519 */
1520 const boolean bypass_viewport = gs || key->bypass_viewport;
1521 const boolean enable_cliptest = !gs && (key->clip_xy ||
1522 key->clip_z ||
1523 key->clip_user);
1524 LLVMValueRef variant_func;
1525 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1526 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1527 boolean have_clipdist = FALSE;
1528 struct lp_bld_tgsi_system_values system_values;
1529
1530 memset(&system_values, 0, sizeof(system_values));
1531
1532 i = 0;
1533 arg_types[i++] = get_context_ptr_type(variant); /* context */
1534 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1535 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1536 if (elts) {
1537 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1538 arg_types[i++] = int32_type; /* fetch_elt_max */
1539 } else
1540 arg_types[i++] = int32_type; /* start */
1541 arg_types[i++] = int32_type; /* fetch_count / count */
1542 arg_types[i++] = int32_type; /* stride */
1543 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1544 arg_types[i++] = int32_type; /* instance_id */
1545
1546 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1547
1548 variant_func = LLVMAddFunction(gallivm->module,
1549 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1550 func_type);
1551
1552 if (elts)
1553 variant->function_elts = variant_func;
1554 else
1555 variant->function = variant_func;
1556
1557 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1558 for (i = 0; i < num_arg_types; ++i)
1559 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1560 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1561 LLVMNoAliasAttribute);
1562
1563 context_ptr = LLVMGetParam(variant_func, 0);
1564 io_ptr = LLVMGetParam(variant_func, 1);
1565 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1566 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1567 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1568 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1569
1570 lp_build_name(context_ptr, "context");
1571 lp_build_name(io_ptr, "io");
1572 lp_build_name(vbuffers_ptr, "vbuffers");
1573 lp_build_name(stride, "stride");
1574 lp_build_name(vb_ptr, "vb");
1575 lp_build_name(system_values.instance_id, "instance_id");
1576
1577 if (elts) {
1578 fetch_elts = LLVMGetParam(variant_func, 3);
1579 fetch_elt_max = LLVMGetParam(variant_func, 4);
1580 fetch_count = LLVMGetParam(variant_func, 5);
1581 lp_build_name(fetch_elts, "fetch_elts");
1582 lp_build_name(fetch_elt_max, "fetch_elt_max");
1583 lp_build_name(fetch_count, "fetch_count");
1584 start = count = NULL;
1585 }
1586 else {
1587 start = LLVMGetParam(variant_func, 3);
1588 count = LLVMGetParam(variant_func, 4);
1589 lp_build_name(start, "start");
1590 lp_build_name(count, "count");
1591 fetch_elts = fetch_count = NULL;
1592 }
1593
1594 /*
1595 * Function body
1596 */
1597
1598 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1599 builder = gallivm->builder;
1600 LLVMPositionBuilderAtEnd(builder, block);
1601
1602 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1603
1604 memset(&vs_type, 0, sizeof vs_type);
1605 vs_type.floating = TRUE; /* floating point values */
1606 vs_type.sign = TRUE; /* values are signed */
1607 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1608 vs_type.width = 32; /* 32-bit float */
1609 vs_type.length = vector_length;
1610
1611 /* hold temporary "bool" clipmask */
1612 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1613 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1614
1615 /* code generated texture sampling */
1616 sampler = draw_llvm_sampler_soa_create(
1617 draw_llvm_variant_key_samplers(key),
1618 context_ptr);
1619
1620 if (elts) {
1621 start = zero;
1622 end = fetch_count;
1623 count = fetch_count;
1624 }
1625 else {
1626 end = lp_build_add(&bld, start, count);
1627 }
1628
1629 step = lp_build_const_int32(gallivm, vector_length);
1630
1631 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1632
1633 lp_build_loop_begin(&lp_loop, gallivm, zero);
1634 {
1635 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1636 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1637 LLVMValueRef io;
1638 LLVMValueRef clipmask; /* holds the clipmask value */
1639 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1640
1641 io_itr = lp_loop.counter;
1642
1643 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1644 #if DEBUG_STORE
1645 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1646 io_itr, io, lp_loop.counter);
1647 #endif
1648 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1649 for (i = 0; i < vector_length; ++i) {
1650 LLVMValueRef true_index =
1651 LLVMBuildAdd(builder,
1652 lp_loop.counter,
1653 lp_build_const_int32(gallivm, i), "");
1654 true_index = LLVMBuildAdd(builder, start, true_index, "");
1655
1656 /* make sure we're not out of bounds which can happen
1657 * if fetch_count % 4 != 0, because on the last iteration
1658 * a few of the 4 vertex fetches will be out of bounds */
1659 true_index = lp_build_min(&bld, true_index, fetch_max);
1660
1661 system_values.vertex_id = LLVMBuildInsertElement(
1662 gallivm->builder,
1663 system_values.vertex_id, true_index,
1664 lp_build_const_int32(gallivm, i), "");
1665
1666 if (elts) {
1667 LLVMValueRef fetch_ptr;
1668 LLVMValueRef index_overflowed;
1669 LLVMValueRef index_ptr =
1670 lp_build_alloca(
1671 gallivm,
1672 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1673 struct lp_build_if_state if_ctx;
1674 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
1675 true_index, fetch_elt_max,
1676 "index_overflowed");
1677
1678 lp_build_if(&if_ctx, gallivm, index_overflowed);
1679 {
1680 /* Generate maximum possible index so that
1681 * generate_fetch can treat it just like
1682 * any other overflow and return zeros.
1683 * We don't have to worry about the restart
1684 * primitive index because it has already been
1685 * handled
1686 */
1687 LLVMValueRef val =
1688 lp_build_const_int32(gallivm, 0xffffffff);
1689 LLVMBuildStore(builder, val, index_ptr);
1690 }
1691 lp_build_else(&if_ctx);
1692 {
1693 LLVMValueRef val;
1694 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1695 &true_index, 1, "");
1696 val = LLVMBuildLoad(builder, fetch_ptr, "");
1697 LLVMBuildStore(builder, val, index_ptr);
1698 }
1699 lp_build_endif(&if_ctx);
1700 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1701 }
1702
1703 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1704 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1705 LLVMValueRef vb_index =
1706 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1707 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1708 generate_fetch(gallivm, draw, vbuffers_ptr,
1709 &aos_attribs[j][i], velem, vb, true_index,
1710 system_values.instance_id);
1711 }
1712 }
1713 convert_to_soa(gallivm, aos_attribs, inputs,
1714 draw->pt.nr_vertex_elements, vs_type);
1715
1716 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1717 generate_vs(variant,
1718 builder,
1719 vs_type,
1720 outputs,
1721 ptr_aos,
1722 &system_values,
1723 context_ptr,
1724 sampler,
1725 key->clamp_vertex_color);
1726
1727 if (pos != -1 && cv != -1) {
1728 /* store original positions in clip before further manipulation */
1729 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1730 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1731
1732 /* do cliptest */
1733 if (enable_cliptest) {
1734 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1735 /* allocate clipmask, assign it integer type */
1736 clipmask = generate_clipmask(llvm,
1737 gallivm,
1738 vs_type,
1739 outputs,
1740 key->clip_xy,
1741 key->clip_z,
1742 key->clip_user,
1743 key->clip_halfz,
1744 key->ucp_enable,
1745 context_ptr, &have_clipdist);
1746 temp = LLVMBuildOr(builder, clipmask, temp, "");
1747 /* store temporary clipping boolean value */
1748 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1749 }
1750 else {
1751 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1752 }
1753
1754 /* do viewport mapping */
1755 if (!bypass_viewport) {
1756 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1757 }
1758 }
1759 else {
1760 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1761 }
1762
1763 /* store clipmask in vertex header,
1764 * original positions in clip
1765 * and transformed positions in data
1766 */
1767 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1768 vs_info->num_outputs, vs_type,
1769 have_clipdist);
1770 }
1771 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1772
1773 sampler->destroy(sampler);
1774
1775 /* return clipping boolean value for function */
1776 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1777
1778 LLVMBuildRet(builder, ret);
1779
1780 gallivm_verify_function(gallivm, variant_func);
1781 }
1782
1783
1784 struct draw_llvm_variant_key *
1785 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1786 {
1787 unsigned i;
1788 struct draw_llvm_variant_key *key;
1789 struct draw_sampler_static_state *draw_sampler;
1790
1791 key = (struct draw_llvm_variant_key *)store;
1792
1793 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1794
1795 /* Presumably all variants of the shader should have the same
1796 * number of vertex elements - ie the number of shader inputs.
1797 * NOTE: we NEED to store the needed number of needed inputs
1798 * here, not the number of provided elements to match keysize
1799 * (and the offset of sampler state in the key).
1800 */
1801 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1802 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1803
1804 /* will have to rig this up properly later */
1805 key->clip_xy = llvm->draw->clip_xy;
1806 key->clip_z = llvm->draw->clip_z;
1807 key->clip_user = llvm->draw->clip_user;
1808 key->bypass_viewport = llvm->draw->identity_viewport;
1809 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1810 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1811 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1812 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1813 key->pad1 = 0;
1814
1815 /* All variants of this shader will have the same value for
1816 * nr_samplers. Not yet trying to compact away holes in the
1817 * sampler array.
1818 */
1819 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1820 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1821 key->nr_sampler_views =
1822 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1823 }
1824 else {
1825 key->nr_sampler_views = key->nr_samplers;
1826 }
1827
1828 draw_sampler = draw_llvm_variant_key_samplers(key);
1829
1830 memcpy(key->vertex_element,
1831 llvm->draw->pt.vertex_element,
1832 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1833
1834 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1835
1836 for (i = 0 ; i < key->nr_samplers; i++) {
1837 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1838 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1839 }
1840 for (i = 0 ; i < key->nr_sampler_views; i++) {
1841 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1842 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1843 }
1844
1845 return key;
1846 }
1847
1848
1849 void
1850 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1851 {
1852 unsigned i;
1853 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1854
1855 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1856 debug_printf("clip_xy = %u\n", key->clip_xy);
1857 debug_printf("clip_z = %u\n", key->clip_z);
1858 debug_printf("clip_user = %u\n", key->clip_user);
1859 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1860 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1861 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1862 debug_printf("has_gs = %u\n", key->has_gs);
1863 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1864
1865 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1866 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1867 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1868 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1869 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1870 }
1871
1872 for (i = 0 ; i < key->nr_sampler_views; i++) {
1873 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1874 }
1875 }
1876
1877
1878 void
1879 draw_llvm_set_mapped_texture(struct draw_context *draw,
1880 unsigned shader_stage,
1881 unsigned sview_idx,
1882 uint32_t width, uint32_t height, uint32_t depth,
1883 uint32_t first_level, uint32_t last_level,
1884 const void *base_ptr,
1885 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1886 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1887 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1888 {
1889 unsigned j;
1890 struct draw_jit_texture *jit_tex;
1891
1892 assert(shader_stage == PIPE_SHADER_VERTEX ||
1893 shader_stage == PIPE_SHADER_GEOMETRY);
1894
1895 if (shader_stage == PIPE_SHADER_VERTEX) {
1896 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1897
1898 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1899 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1900 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1901
1902 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1903 } else {
1904 assert(0);
1905 return;
1906 }
1907
1908 jit_tex->width = width;
1909 jit_tex->height = height;
1910 jit_tex->depth = depth;
1911 jit_tex->first_level = first_level;
1912 jit_tex->last_level = last_level;
1913 jit_tex->base = base_ptr;
1914
1915 for (j = first_level; j <= last_level; j++) {
1916 jit_tex->mip_offsets[j] = mip_offsets[j];
1917 jit_tex->row_stride[j] = row_stride[j];
1918 jit_tex->img_stride[j] = img_stride[j];
1919 }
1920 }
1921
1922
1923 void
1924 draw_llvm_set_sampler_state(struct draw_context *draw,
1925 unsigned shader_type)
1926 {
1927 unsigned i;
1928
1929 if (shader_type == PIPE_SHADER_VERTEX) {
1930 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1931 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1932
1933 if (draw->samplers[i]) {
1934 const struct pipe_sampler_state *s
1935 = draw->samplers[PIPE_SHADER_VERTEX][i];
1936 jit_sam->min_lod = s->min_lod;
1937 jit_sam->max_lod = s->max_lod;
1938 jit_sam->lod_bias = s->lod_bias;
1939 COPY_4V(jit_sam->border_color, s->border_color.f);
1940 }
1941 }
1942 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1943 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1944 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1945
1946 if (draw->samplers[i]) {
1947 const struct pipe_sampler_state *s
1948 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1949 jit_sam->min_lod = s->min_lod;
1950 jit_sam->max_lod = s->max_lod;
1951 jit_sam->lod_bias = s->lod_bias;
1952 COPY_4V(jit_sam->border_color, s->border_color.f);
1953 }
1954 }
1955 }
1956 }
1957
1958
1959 void
1960 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1961 {
1962 struct draw_llvm *llvm = variant->llvm;
1963
1964 if (variant->function_elts) {
1965 gallivm_free_function(variant->gallivm,
1966 variant->function_elts, variant->jit_func_elts);
1967 }
1968
1969 if (variant->function) {
1970 gallivm_free_function(variant->gallivm,
1971 variant->function, variant->jit_func);
1972 }
1973
1974 gallivm_destroy(variant->gallivm);
1975
1976 remove_from_list(&variant->list_item_local);
1977 variant->shader->variants_cached--;
1978 remove_from_list(&variant->list_item_global);
1979 llvm->nr_variants--;
1980 FREE(variant);
1981 }
1982
1983
1984 /**
1985 * Create LLVM types for various structures.
1986 */
1987 static void
1988 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1989 {
1990 struct gallivm_state *gallivm = var->gallivm;
1991 LLVMTypeRef texture_type, sampler_type, context_type;
1992
1993 texture_type = create_jit_texture_type(gallivm, "texture");
1994 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1995
1996 context_type = create_gs_jit_context_type(gallivm,
1997 var->shader->base.vector_length,
1998 texture_type, sampler_type,
1999 "draw_gs_jit_context");
2000 var->context_ptr_type = LLVMPointerType(context_type, 0);
2001
2002 var->input_array_type = create_gs_jit_input_type(gallivm);
2003 }
2004
2005 static LLVMTypeRef
2006 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2007 {
2008 if (!variant->context_ptr_type)
2009 create_gs_jit_types(variant);
2010 return variant->context_ptr_type;
2011 }
2012
2013 static LLVMValueRef
2014 generate_mask_value(struct draw_gs_llvm_variant *variant,
2015 struct lp_type gs_type)
2016 {
2017 struct gallivm_state *gallivm = variant->gallivm;
2018 LLVMBuilderRef builder = gallivm->builder;
2019 LLVMValueRef bits[16];
2020 struct lp_type mask_type = lp_int_type(gs_type);
2021 struct lp_type mask_elem_type = lp_elem_type(mask_type);
2022 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
2023 mask_type,
2024 0);
2025 unsigned i;
2026
2027 assert(gs_type.length <= Elements(bits));
2028
2029 for (i = gs_type.length; i >= 1; --i) {
2030 int idx = i - 1;
2031 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2032 bits[idx] = lp_build_compare(gallivm,
2033 mask_elem_type, PIPE_FUNC_GEQUAL,
2034 variant->num_prims, ind);
2035 }
2036 for (i = 0; i < gs_type.length; ++i) {
2037 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2038 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
2039 }
2040 mask_val = lp_build_compare(gallivm,
2041 mask_type, PIPE_FUNC_NOTEQUAL,
2042 mask_val,
2043 lp_build_const_int_vec(gallivm, mask_type, 0));
2044
2045 return mask_val;
2046 }
2047
2048 static void
2049 draw_gs_llvm_generate(struct draw_llvm *llvm,
2050 struct draw_gs_llvm_variant *variant)
2051 {
2052 struct gallivm_state *gallivm = variant->gallivm;
2053 LLVMContextRef context = gallivm->context;
2054 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2055 LLVMTypeRef arg_types[6];
2056 LLVMTypeRef func_type;
2057 LLVMValueRef variant_func;
2058 LLVMValueRef context_ptr;
2059 LLVMValueRef prim_id_ptr;
2060 LLVMBasicBlockRef block;
2061 LLVMBuilderRef builder;
2062 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2063 struct lp_build_sampler_soa *sampler = 0;
2064 struct lp_build_context bld;
2065 struct lp_bld_tgsi_system_values system_values;
2066 struct lp_type gs_type;
2067 unsigned i;
2068 struct draw_gs_llvm_iface gs_iface;
2069 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2070 LLVMValueRef consts_ptr;
2071 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2072 struct lp_build_mask_context mask;
2073 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2074 unsigned vector_length = variant->shader->base.vector_length;
2075
2076 memset(&system_values, 0, sizeof(system_values));
2077
2078 assert(variant->vertex_header_ptr_type);
2079
2080 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2081 arg_types[1] = variant->input_array_type; /* input */
2082 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2083 arg_types[3] = int32_type; /* num_prims */
2084 arg_types[4] = int32_type; /* instance_id */
2085 arg_types[5] = LLVMPointerType(
2086 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2087
2088 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2089
2090 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2091 func_type);
2092 variant->function = variant_func;
2093
2094 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2095
2096 for (i = 0; i < Elements(arg_types); ++i)
2097 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2098 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2099 LLVMNoAliasAttribute);
2100
2101 context_ptr = LLVMGetParam(variant_func, 0);
2102 input_array = LLVMGetParam(variant_func, 1);
2103 io_ptr = LLVMGetParam(variant_func, 2);
2104 num_prims = LLVMGetParam(variant_func, 3);
2105 system_values.instance_id = LLVMGetParam(variant_func, 4);
2106 prim_id_ptr = LLVMGetParam(variant_func, 5);
2107
2108 lp_build_name(context_ptr, "context");
2109 lp_build_name(input_array, "input");
2110 lp_build_name(io_ptr, "io");
2111 lp_build_name(num_prims, "num_prims");
2112 lp_build_name(system_values.instance_id, "instance_id");
2113 lp_build_name(prim_id_ptr, "prim_id_ptr");
2114
2115 variant->context_ptr = context_ptr;
2116 variant->io_ptr = io_ptr;
2117 variant->num_prims = num_prims;
2118
2119 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2120 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2121 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2122 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2123 gs_iface.input = input_array;
2124 gs_iface.variant = variant;
2125
2126 /*
2127 * Function body
2128 */
2129
2130 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2131 builder = gallivm->builder;
2132 LLVMPositionBuilderAtEnd(builder, block);
2133
2134 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2135
2136 memset(&gs_type, 0, sizeof gs_type);
2137 gs_type.floating = TRUE; /* floating point values */
2138 gs_type.sign = TRUE; /* values are signed */
2139 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2140 gs_type.width = 32; /* 32-bit float */
2141 gs_type.length = vector_length;
2142
2143 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2144
2145 /* code generated texture sampling */
2146 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2147 context_ptr);
2148
2149 mask_val = generate_mask_value(variant, gs_type);
2150 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2151
2152 if (gs_info->uses_primid) {
2153 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
2154 }
2155
2156 lp_build_tgsi_soa(variant->gallivm,
2157 tokens,
2158 gs_type,
2159 &mask,
2160 consts_ptr,
2161 &system_values,
2162 NULL,
2163 outputs,
2164 sampler,
2165 &llvm->draw->gs.geometry_shader->info,
2166 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2167
2168 sampler->destroy(sampler);
2169
2170 lp_build_mask_end(&mask);
2171
2172 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2173
2174 gallivm_verify_function(gallivm, variant_func);
2175 }
2176
2177
2178 struct draw_gs_llvm_variant *
2179 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2180 unsigned num_outputs,
2181 const struct draw_gs_llvm_variant_key *key)
2182 {
2183 struct draw_gs_llvm_variant *variant;
2184 struct llvm_geometry_shader *shader =
2185 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2186 LLVMTypeRef vertex_header;
2187
2188 variant = MALLOC(sizeof *variant +
2189 shader->variant_key_size -
2190 sizeof variant->key);
2191 if (variant == NULL)
2192 return NULL;
2193
2194 variant->llvm = llvm;
2195 variant->shader = shader;
2196
2197 variant->gallivm = gallivm_create();
2198
2199 create_gs_jit_types(variant);
2200
2201 memcpy(&variant->key, key, shader->variant_key_size);
2202
2203 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2204
2205 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2206
2207 draw_gs_llvm_generate(llvm, variant);
2208
2209 gallivm_compile_module(variant->gallivm);
2210
2211 variant->jit_func = (draw_gs_jit_func)
2212 gallivm_jit_function(variant->gallivm, variant->function);
2213
2214 variant->list_item_global.base = variant;
2215 variant->list_item_local.base = variant;
2216 /*variant->no = */shader->variants_created++;
2217 variant->list_item_global.base = variant;
2218
2219 return variant;
2220 }
2221
2222 void
2223 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2224 {
2225 struct draw_llvm *llvm = variant->llvm;
2226
2227 if (variant->function) {
2228 gallivm_free_function(variant->gallivm,
2229 variant->function, variant->jit_func);
2230 }
2231
2232 gallivm_destroy(variant->gallivm);
2233
2234 remove_from_list(&variant->list_item_local);
2235 variant->shader->variants_cached--;
2236 remove_from_list(&variant->list_item_global);
2237 llvm->nr_gs_variants--;
2238 FREE(variant);
2239 }
2240
2241 struct draw_gs_llvm_variant_key *
2242 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2243 {
2244 unsigned i;
2245 struct draw_gs_llvm_variant_key *key;
2246 struct draw_sampler_static_state *draw_sampler;
2247
2248 key = (struct draw_gs_llvm_variant_key *)store;
2249
2250 /* All variants of this shader will have the same value for
2251 * nr_samplers. Not yet trying to compact away holes in the
2252 * sampler array.
2253 */
2254 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2255 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2256 key->nr_sampler_views =
2257 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2258 }
2259 else {
2260 key->nr_sampler_views = key->nr_samplers;
2261 }
2262
2263 draw_sampler = key->samplers;
2264
2265 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2266
2267 for (i = 0 ; i < key->nr_samplers; i++) {
2268 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2269 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2270 }
2271 for (i = 0 ; i < key->nr_sampler_views; i++) {
2272 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2273 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2274 }
2275
2276 return key;
2277 }
2278
2279 void
2280 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2281 {
2282 unsigned i;
2283 struct draw_sampler_static_state *sampler = key->samplers;
2284
2285 for (i = 0 ; i < key->nr_sampler_views; i++) {
2286 debug_printf("sampler[%i].src_format = %s\n", i,
2287 util_format_name(sampler[i].texture_state.format));
2288 }
2289 }