projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: (trivial) dump tgsi for geometry shaders with GALLIVM_DEBUG_TGSI
[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 / divisor)
711 */
712 LLVMValueRef current_instance;
713 index = lp_build_const_int32(gallivm, draw->start_instance);
714 current_instance = LLVMBuildUDiv(builder, instance_id,
715 lp_build_const_int32(gallivm, velem->instance_divisor),
716 "instance_divisor");
717 index = lp_build_uadd_overflow(gallivm, index, current_instance, &ofbit);
718 }
719
720 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
721 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
722 stride = lp_build_uadd_overflow(
723 gallivm, stride,
724 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
725 needed_buffer_size = lp_build_uadd_overflow(
726 gallivm, stride,
727 lp_build_const_int32(gallivm,
728 util_format_get_blocksize(velem->src_format)),
729 &ofbit);
730
731 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
732 needed_buffer_size, buffer_size,
733 "buffer_overflowed");
734 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
735 #if 0
736 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
737 index, vb_stride);
738 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
739 vb_buffer_offset,
740 lp_build_const_int32(gallivm, velem->src_offset));
741 lp_build_print_value(gallivm, " blocksize = ",
742 lp_build_const_int32(
743 gallivm,
744 util_format_get_blocksize(velem->src_format)));
745 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
746 lp_build_printf(gallivm, " stride = %u\n", stride);
747 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
748 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
749 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
750 #endif
751
752 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
753 {
754 LLVMValueRef val =
755 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
756 LLVMBuildStore(builder, val, temp_ptr);
757 }
758 lp_build_else(&if_ctx);
759 {
760 LLVMValueRef val;
761 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
762
763 val = lp_build_fetch_rgba_aos(gallivm,
764 format_desc,
765 lp_float32_vec4_type(),
766 map_ptr,
767 zero, zero, zero);
768 LLVMBuildStore(builder, val, temp_ptr);
769 }
770 lp_build_endif(&if_ctx);
771
772 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
773 }
774
775 static void
776 convert_to_soa(struct gallivm_state *gallivm,
777 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
778 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
779 unsigned num_attribs, const struct lp_type soa_type)
780 {
781 unsigned i, j, k;
782 struct lp_type aos_channel_type = soa_type;
783
784 debug_assert(TGSI_NUM_CHANNELS == 4);
785 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
786
787 aos_channel_type.length >>= 1;
788
789 for (i = 0; i < num_attribs; ++i) {
790 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
791 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
792
793 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
794 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
795
796 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
797
798 for (k = 0; k < pixels_per_channel; ++k) {
799 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
800 }
801
802 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
803 }
804
805 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
806 }
807 }
808
809
810 static void
811 store_aos(struct gallivm_state *gallivm,
812 LLVMValueRef io_ptr,
813 LLVMValueRef index,
814 LLVMValueRef value)
815 {
816 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
817 LLVMBuilderRef builder = gallivm->builder;
818 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
819 LLVMValueRef indices[3];
820
821 indices[0] = lp_build_const_int32(gallivm, 0);
822 indices[1] = index;
823 indices[2] = lp_build_const_int32(gallivm, 0);
824
825 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
826 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
827
828 #if DEBUG_STORE
829 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
830 #endif
831
832 /* Unaligned store due to the vertex header */
833 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
834 }
835
836 /**
837 * Adjust the mask to architecture endianess. The mask will the store in struct:
838 *
839 * struct vertex_header {
840 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
841 * unsigned edgeflag:1;
842 * unsigned have_clipdist:1;
843 * unsigned vertex_id:16;
844 * [...]
845 * }
846 *
847 * On little-endian machine nothing needs to done, however on bit-endian machine
848 * the mask's fields need to be adjusted with the algorithm:
849 *
850 * uint32_t reverse (uint32_t x)
851 * {
852 * return (x >> 16) | // vertex_id
853 * ((x & 0x3fff) << 18) | // clipmask
854 * ((x & 0x4000) << 3) | // have_clipdist
855 * ((x & 0x8000) << 1); // edgeflag
856 * }
857 */
858 static LLVMValueRef
859 adjust_mask(struct gallivm_state *gallivm,
860 LLVMValueRef mask)
861 {
862 #ifdef PIPE_ARCH_BIG_ENDIAN
863 LLVMBuilderRef builder = gallivm->builder;
864 LLVMValueRef vertex_id;
865 LLVMValueRef clipmask;
866 LLVMValueRef have_clipdist;
867 LLVMValueRef edgeflag;
868
869 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
870 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
871 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
872 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
873 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
874 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
875 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
876
877 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
878 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
879 mask = LLVMBuildOr(builder, mask, edgeflag, "");
880 #endif
881 return mask;
882 }
883
884 static void
885 store_aos_array(struct gallivm_state *gallivm,
886 struct lp_type soa_type,
887 LLVMValueRef io_ptr,
888 LLVMValueRef *indices,
889 LLVMValueRef* aos,
890 int attrib,
891 int num_outputs,
892 LLVMValueRef clipmask,
893 boolean have_clipdist)
894 {
895 LLVMBuilderRef builder = gallivm->builder;
896 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
897 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
898 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
899 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
900 int vector_length = soa_type.length;
901 int i;
902
903 debug_assert(TGSI_NUM_CHANNELS == 4);
904
905 for (i = 0; i < vector_length; i++) {
906 linear_inds[i] = lp_build_const_int32(gallivm, i);
907 if (indices) {
908 inds[i] = indices[i];
909 } else {
910 inds[i] = linear_inds[i];
911 }
912 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
913 }
914
915 if (attrib == 0) {
916 /* store vertex header for each of the n vertices */
917 LLVMValueRef val, cliptmp;
918 int vertex_id_pad_edgeflag;
919
920 /* If this assertion fails, it means we need to update the bit twidding
921 * code here. See struct vertex_header in draw_private.h.
922 */
923 assert(DRAW_TOTAL_CLIP_PLANES==14);
924 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
925 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
926 if (have_clipdist)
927 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
928 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
929 /* OR with the clipmask */
930 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
931 for (i = 0; i < vector_length; i++) {
932 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
933 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
934 val = adjust_mask(gallivm, val);
935 LLVMBuildStore(builder, val, id_ptr);
936 #if DEBUG_STORE
937 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
938 io_ptrs[i], inds[i], val);
939 #endif
940 }
941 }
942
943 /* store for each of the n vertices */
944 for (i = 0; i < vector_length; i++) {
945 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
946 }
947 }
948
949
950 static void
951 convert_to_aos(struct gallivm_state *gallivm,
952 LLVMValueRef io,
953 LLVMValueRef *indices,
954 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
955 LLVMValueRef clipmask,
956 int num_outputs,
957 struct lp_type soa_type,
958 boolean have_clipdist)
959 {
960 LLVMBuilderRef builder = gallivm->builder;
961 unsigned chan, attrib, i;
962
963 #if DEBUG_STORE
964 lp_build_printf(gallivm, " # storing begin\n");
965 #endif
966 for (attrib = 0; attrib < num_outputs; ++attrib) {
967 LLVMValueRef soa[TGSI_NUM_CHANNELS];
968 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
969 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
970 if (outputs[attrib][chan]) {
971 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
972 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
973 #if DEBUG_STORE
974 lp_build_printf(gallivm, "output %d : %d ",
975 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
976 attrib, 0),
977 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
978 chan, 0));
979 lp_build_print_value(gallivm, "val = ", out);
980 {
981 LLVMValueRef iv =
982 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
983
984 lp_build_print_value(gallivm, " ival = ", iv);
985 }
986 #endif
987 soa[chan] = out;
988 }
989 else {
990 soa[chan] = 0;
991 }
992 }
993
994
995 if (soa_type.length == TGSI_NUM_CHANNELS) {
996 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
997 } else {
998 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
999
1000 for (i = 0; i < soa_type.length; ++i) {
1001 aos[i] = lp_build_extract_range(gallivm,
1002 soa[i % TGSI_NUM_CHANNELS],
1003 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1004 TGSI_NUM_CHANNELS);
1005 }
1006 }
1007
1008 store_aos_array(gallivm,
1009 soa_type,
1010 io, indices,
1011 aos,
1012 attrib,
1013 num_outputs,
1014 clipmask, have_clipdist);
1015 }
1016 #if DEBUG_STORE
1017 lp_build_printf(gallivm, " # storing end\n");
1018 #endif
1019 }
1020
1021
1022 /**
1023 * Stores original vertex positions in clip coordinates
1024 */
1025 static void
1026 store_clip(struct gallivm_state *gallivm,
1027 const struct lp_type vs_type,
1028 LLVMValueRef io_ptr,
1029 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1030 boolean pre_clip_pos, int idx)
1031 {
1032 LLVMBuilderRef builder = gallivm->builder;
1033 LLVMValueRef soa[4];
1034 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1035 LLVMValueRef indices[2];
1036 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1037 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1038 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1039 int i, j;
1040
1041 indices[0] =
1042 indices[1] = lp_build_const_int32(gallivm, 0);
1043
1044 for (i = 0; i < vs_type.length; i++) {
1045 inds[i] = lp_build_const_int32(gallivm, i);
1046 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1047 }
1048
1049 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1050 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1051 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1052 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1053
1054 if (!pre_clip_pos) {
1055 for (i = 0; i < vs_type.length; i++) {
1056 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1057 }
1058 } else {
1059 for (i = 0; i < vs_type.length; i++) {
1060 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1061 }
1062 }
1063
1064 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1065 for (i = 0; i < vs_type.length; ++i) {
1066 aos[i] = lp_build_extract_range(gallivm,
1067 soa[i % TGSI_NUM_CHANNELS],
1068 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1069 TGSI_NUM_CHANNELS);
1070 }
1071
1072 for (j = 0; j < vs_type.length; j++) {
1073 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
1074 LLVMValueRef clip_ptr;
1075
1076 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1077 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1078
1079 /* Unaligned store */
1080 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1081 }
1082 }
1083
1084
1085 /**
1086 * Transforms the outputs for viewport mapping
1087 */
1088 static void
1089 generate_viewport(struct draw_llvm_variant *variant,
1090 LLVMBuilderRef builder,
1091 struct lp_type vs_type,
1092 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1093 LLVMValueRef context_ptr)
1094 {
1095 int i;
1096 struct gallivm_state *gallivm = variant->gallivm;
1097 struct lp_type f32_type = vs_type;
1098 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1099 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1100 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1101 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1102 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1103
1104 /* for 1/w convention*/
1105 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1106 LLVMBuildStore(builder, out3, outputs[pos][3]);
1107
1108 /* Viewport Mapping */
1109 for (i=0; i<3; i++) {
1110 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1111 LLVMValueRef scale;
1112 LLVMValueRef trans;
1113 LLVMValueRef scale_i;
1114 LLVMValueRef trans_i;
1115 LLVMValueRef index;
1116
1117 index = lp_build_const_int32(gallivm, i);
1118 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1119
1120 index = lp_build_const_int32(gallivm, i+4);
1121 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1122
1123 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1124 LLVMBuildLoad(builder, scale_i, "scale"));
1125 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1126 LLVMBuildLoad(builder, trans_i, "trans"));
1127
1128 /* divide by w */
1129 out = LLVMBuildFMul(builder, out, out3, "");
1130 /* mult by scale */
1131 out = LLVMBuildFMul(builder, out, scale, "");
1132 /* add translation */
1133 out = LLVMBuildFAdd(builder, out, trans, "");
1134
1135 /* store transformed outputs */
1136 LLVMBuildStore(builder, out, outputs[pos][i]);
1137 }
1138
1139 }
1140
1141
1142 /**
1143 * Returns clipmask as nxi32 bitmask for the n vertices
1144 */
1145 static LLVMValueRef
1146 generate_clipmask(struct draw_llvm *llvm,
1147 struct gallivm_state *gallivm,
1148 struct lp_type vs_type,
1149 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1150 boolean clip_xy,
1151 boolean clip_z,
1152 boolean clip_user,
1153 boolean clip_halfz,
1154 unsigned ucp_enable,
1155 LLVMValueRef context_ptr,
1156 boolean *have_clipdist)
1157 {
1158 LLVMBuilderRef builder = gallivm->builder;
1159 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1160 LLVMValueRef test, temp;
1161 LLVMValueRef zero, shift;
1162 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1163 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1164 LLVMValueRef plane1, planes, plane_ptr, sum;
1165 struct lp_type f32_type = vs_type;
1166 struct lp_type i32_type = lp_int_type(vs_type);
1167 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1168 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1169 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1170 bool have_cd = false;
1171 unsigned cd[2];
1172
1173 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1174 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1175
1176 if (cd[0] != pos || cd[1] != pos)
1177 have_cd = true;
1178
1179 if (num_written_clipdistance && !clip_user) {
1180 clip_user = true;
1181 ucp_enable = (1 << num_written_clipdistance) - 1;
1182 }
1183
1184 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1185 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1186 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1187 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1188
1189 /*
1190 * load clipvertex and position from correct locations.
1191 * if they are the same just load them once.
1192 */
1193 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1194 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1195 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1196 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1197
1198 if (clip_user && cv != pos) {
1199 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1200 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1201 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1202 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1203 } else {
1204 cv_x = pos_x;
1205 cv_y = pos_y;
1206 cv_z = pos_z;
1207 cv_w = pos_w;
1208 }
1209
1210 /* Cliptest, for hardwired planes */
1211 if (clip_xy) {
1212 /* plane 1 */
1213 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1214 temp = shift;
1215 test = LLVMBuildAnd(builder, test, temp, "");
1216 mask = test;
1217
1218 /* plane 2 */
1219 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1220 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1221 temp = LLVMBuildShl(builder, temp, shift, "");
1222 test = LLVMBuildAnd(builder, test, temp, "");
1223 mask = LLVMBuildOr(builder, mask, test, "");
1224
1225 /* plane 3 */
1226 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1227 temp = LLVMBuildShl(builder, temp, shift, "");
1228 test = LLVMBuildAnd(builder, test, temp, "");
1229 mask = LLVMBuildOr(builder, mask, test, "");
1230
1231 /* plane 4 */
1232 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1233 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1234 temp = LLVMBuildShl(builder, temp, shift, "");
1235 test = LLVMBuildAnd(builder, test, temp, "");
1236 mask = LLVMBuildOr(builder, mask, test, "");
1237 }
1238
1239 if (clip_z) {
1240 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1241 if (clip_halfz) {
1242 /* plane 5 */
1243 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1244 test = LLVMBuildAnd(builder, test, temp, "");
1245 mask = LLVMBuildOr(builder, mask, test, "");
1246 }
1247 else {
1248 /* plane 5 */
1249 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1250 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1251 test = LLVMBuildAnd(builder, test, temp, "");
1252 mask = LLVMBuildOr(builder, mask, test, "");
1253 }
1254 /* plane 6 */
1255 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1256 temp = LLVMBuildShl(builder, temp, shift, "");
1257 test = LLVMBuildAnd(builder, test, temp, "");
1258 mask = LLVMBuildOr(builder, mask, test, "");
1259 }
1260
1261 if (clip_user) {
1262 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1263 LLVMValueRef indices[3];
1264
1265 /* userclip planes */
1266 while (ucp_enable) {
1267 unsigned plane_idx = ffs(ucp_enable)-1;
1268 ucp_enable &= ~(1 << plane_idx);
1269 plane_idx += 6;
1270
1271 if (have_cd && num_written_clipdistance) {
1272 LLVMValueRef clipdist;
1273 int i;
1274 i = plane_idx - 6;
1275
1276 *have_clipdist = TRUE;
1277 if (i < 4) {
1278 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1279 } else {
1280 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1281 }
1282 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1283 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1284 test = LLVMBuildAnd(builder, test, temp, "");
1285 mask = LLVMBuildOr(builder, mask, test, "");
1286 } else {
1287 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1288 indices[0] = lp_build_const_int32(gallivm, 0);
1289 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1290
1291 indices[2] = lp_build_const_int32(gallivm, 0);
1292 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1293 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1294 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1295 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1296
1297 indices[2] = lp_build_const_int32(gallivm, 1);
1298 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1299 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1300 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1301 test = LLVMBuildFMul(builder, planes, cv_y, "");
1302 sum = LLVMBuildFAdd(builder, sum, test, "");
1303
1304 indices[2] = lp_build_const_int32(gallivm, 2);
1305 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1306 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1307 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1308 test = LLVMBuildFMul(builder, planes, cv_z, "");
1309 sum = LLVMBuildFAdd(builder, sum, test, "");
1310
1311 indices[2] = lp_build_const_int32(gallivm, 3);
1312 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1313 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1314 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1315 test = LLVMBuildFMul(builder, planes, cv_w, "");
1316 sum = LLVMBuildFAdd(builder, sum, test, "");
1317
1318 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1319 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1320 test = LLVMBuildAnd(builder, test, temp, "");
1321 mask = LLVMBuildOr(builder, mask, test, "");
1322 }
1323 }
1324 }
1325 return mask;
1326 }
1327
1328
1329 /**
1330 * Returns boolean if any clipping has occurred
1331 * Used zero/non-zero i32 value to represent boolean
1332 */
1333 static LLVMValueRef
1334 clipmask_booli32(struct gallivm_state *gallivm,
1335 const struct lp_type vs_type,
1336 LLVMValueRef clipmask_bool_ptr)
1337 {
1338 LLVMBuilderRef builder = gallivm->builder;
1339 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1340 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1341 LLVMValueRef ret = LLVMConstNull(int32_type);
1342 LLVMValueRef temp;
1343 int i;
1344
1345 /*
1346 * Can do this with log2(vector length) pack instructions and one extract
1347 * (as we don't actually need a or) with sse2 which would be way better.
1348 */
1349 for (i=0; i < vs_type.length; i++) {
1350 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1351 lp_build_const_int32(gallivm, i) , "");
1352 ret = LLVMBuildOr(builder, ret, temp, "");
1353 }
1354 return ret;
1355 }
1356
1357 static LLVMValueRef
1358 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1359 struct lp_build_tgsi_context * bld_base,
1360 boolean is_indirect,
1361 LLVMValueRef vertex_index,
1362 LLVMValueRef attrib_index,
1363 LLVMValueRef swizzle_index)
1364 {
1365 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1366 struct gallivm_state *gallivm = bld_base->base.gallivm;
1367 LLVMBuilderRef builder = gallivm->builder;
1368 LLVMValueRef indices[3];
1369 LLVMValueRef res;
1370 struct lp_type type = bld_base->base.type;
1371
1372 if (is_indirect) {
1373 int i;
1374 res = bld_base->base.zero;
1375 for (i = 0; i < type.length; ++i) {
1376 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1377 LLVMValueRef vert_chan_index = LLVMBuildExtractElement(builder,
1378 vertex_index, idx, "");
1379 LLVMValueRef channel_vec, value;
1380 indices[0] = vert_chan_index;
1381 indices[1] = attrib_index;
1382 indices[2] = swizzle_index;
1383
1384 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1385 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1386 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1387
1388 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1389 }
1390 } else {
1391 indices[0] = vertex_index;
1392 indices[1] = attrib_index;
1393 indices[2] = swizzle_index;
1394
1395 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1396 res = LLVMBuildLoad(builder, res, "");
1397 }
1398
1399 return res;
1400 }
1401
1402 static void
1403 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1404 struct lp_build_tgsi_context * bld_base,
1405 LLVMValueRef (*outputs)[4],
1406 LLVMValueRef emitted_vertices_vec)
1407 {
1408 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1409 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1410 struct gallivm_state *gallivm = variant->gallivm;
1411 LLVMBuilderRef builder = gallivm->builder;
1412 struct lp_type gs_type = bld_base->base.type;
1413 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1414 lp_int_type(gs_type), 0);
1415 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1416 LLVMValueRef next_prim_offset =
1417 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1418 LLVMValueRef io = variant->io_ptr;
1419 unsigned i;
1420 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1421
1422 for (i = 0; i < gs_type.length; ++i) {
1423 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1424 LLVMValueRef currently_emitted =
1425 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1426 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1427 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1428 }
1429
1430 convert_to_aos(gallivm, io, indices,
1431 outputs, clipmask,
1432 gs_info->num_outputs, gs_type,
1433 FALSE);
1434 }
1435
1436 static void
1437 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1438 struct lp_build_tgsi_context * bld_base,
1439 LLVMValueRef verts_per_prim_vec,
1440 LLVMValueRef emitted_prims_vec)
1441 {
1442 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1443 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1444 struct gallivm_state *gallivm = variant->gallivm;
1445 LLVMBuilderRef builder = gallivm->builder;
1446 LLVMValueRef prim_lengts_ptr =
1447 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1448 unsigned i;
1449
1450 for (i = 0; i < bld_base->base.type.length; ++i) {
1451 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1452 LLVMValueRef prims_emitted =
1453 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1454 LLVMValueRef store_ptr;
1455 LLVMValueRef num_vertices =
1456 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1457
1458 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1459 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1460 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1461 LLVMBuildStore(builder, num_vertices, store_ptr);
1462 }
1463 }
1464
1465 static void
1466 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1467 struct lp_build_tgsi_context * bld_base,
1468 LLVMValueRef total_emitted_vertices_vec,
1469 LLVMValueRef emitted_prims_vec)
1470 {
1471 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1472 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1473 struct gallivm_state *gallivm = variant->gallivm;
1474 LLVMBuilderRef builder = gallivm->builder;
1475 LLVMValueRef emitted_verts_ptr =
1476 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1477 LLVMValueRef emitted_prims_ptr =
1478 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1479 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1480
1481 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1482 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1483
1484 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1485 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1486 }
1487
1488 static void
1489 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1490 boolean elts)
1491 {
1492 struct gallivm_state *gallivm = variant->gallivm;
1493 LLVMContextRef context = gallivm->context;
1494 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1495 LLVMTypeRef arg_types[10];
1496 unsigned num_arg_types =
1497 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1498 LLVMTypeRef func_type;
1499 LLVMValueRef context_ptr;
1500 LLVMBasicBlockRef block;
1501 LLVMBuilderRef builder;
1502 struct lp_type vs_type;
1503 LLVMValueRef end, start;
1504 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1505 LLVMValueRef vertex_id_offset;
1506 LLVMValueRef stride, step, io_itr;
1507 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1508 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1509 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1510 struct draw_context *draw = llvm->draw;
1511 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1512 unsigned i, j;
1513 struct lp_build_context bld;
1514 struct lp_build_loop_state lp_loop;
1515 const int vector_length = lp_native_vector_width / 32;
1516 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1517 LLVMValueRef fetch_max;
1518 struct lp_build_sampler_soa *sampler = 0;
1519 LLVMValueRef ret, clipmask_bool_ptr;
1520 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1521 struct draw_llvm_variant_key *key = &variant->key;
1522 /* If geometry shader is present we need to skip both the viewport
1523 * transformation and clipping otherwise the inputs to the geometry
1524 * shader will be incorrect.
1525 */
1526 const boolean bypass_viewport = gs || key->bypass_viewport;
1527 const boolean enable_cliptest = !gs && (key->clip_xy ||
1528 key->clip_z ||
1529 key->clip_user);
1530 LLVMValueRef variant_func;
1531 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1532 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1533 boolean have_clipdist = FALSE;
1534 struct lp_bld_tgsi_system_values system_values;
1535
1536 memset(&system_values, 0, sizeof(system_values));
1537
1538 i = 0;
1539 arg_types[i++] = get_context_ptr_type(variant); /* context */
1540 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1541 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1542 if (elts) {
1543 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1544 arg_types[i++] = int32_type; /* fetch_elt_max */
1545 } else
1546 arg_types[i++] = int32_type; /* start */
1547 arg_types[i++] = int32_type; /* fetch_count / count */
1548 arg_types[i++] = int32_type; /* stride */
1549 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1550 arg_types[i++] = int32_type; /* instance_id */
1551 arg_types[i++] = int32_type; /* vertex_id_offset */
1552
1553 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1554
1555 variant_func = LLVMAddFunction(gallivm->module,
1556 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1557 func_type);
1558
1559 if (elts)
1560 variant->function_elts = variant_func;
1561 else
1562 variant->function = variant_func;
1563
1564 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1565 for (i = 0; i < num_arg_types; ++i)
1566 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1567 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1568 LLVMNoAliasAttribute);
1569
1570 context_ptr = LLVMGetParam(variant_func, 0);
1571 io_ptr = LLVMGetParam(variant_func, 1);
1572 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1573 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1574 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1575 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1576 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1577
1578 lp_build_name(context_ptr, "context");
1579 lp_build_name(io_ptr, "io");
1580 lp_build_name(vbuffers_ptr, "vbuffers");
1581 lp_build_name(stride, "stride");
1582 lp_build_name(vb_ptr, "vb");
1583 lp_build_name(system_values.instance_id, "instance_id");
1584 lp_build_name(vertex_id_offset, "vertex_id_offset");
1585
1586 if (elts) {
1587 fetch_elts = LLVMGetParam(variant_func, 3);
1588 fetch_elt_max = LLVMGetParam(variant_func, 4);
1589 fetch_count = LLVMGetParam(variant_func, 5);
1590 lp_build_name(fetch_elts, "fetch_elts");
1591 lp_build_name(fetch_elt_max, "fetch_elt_max");
1592 lp_build_name(fetch_count, "fetch_count");
1593 start = count = NULL;
1594 }
1595 else {
1596 start = LLVMGetParam(variant_func, 3);
1597 count = LLVMGetParam(variant_func, 4);
1598 lp_build_name(start, "start");
1599 lp_build_name(count, "count");
1600 fetch_elts = fetch_count = NULL;
1601 }
1602
1603 /*
1604 * Function body
1605 */
1606
1607 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1608 builder = gallivm->builder;
1609 LLVMPositionBuilderAtEnd(builder, block);
1610
1611 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1612
1613 memset(&vs_type, 0, sizeof vs_type);
1614 vs_type.floating = TRUE; /* floating point values */
1615 vs_type.sign = TRUE; /* values are signed */
1616 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1617 vs_type.width = 32; /* 32-bit float */
1618 vs_type.length = vector_length;
1619
1620 /* hold temporary "bool" clipmask */
1621 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1622 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1623
1624 /* code generated texture sampling */
1625 sampler = draw_llvm_sampler_soa_create(
1626 draw_llvm_variant_key_samplers(key),
1627 context_ptr);
1628
1629 if (elts) {
1630 start = zero;
1631 end = fetch_count;
1632 count = fetch_count;
1633 }
1634 else {
1635 end = lp_build_add(&bld, start, count);
1636 }
1637
1638 step = lp_build_const_int32(gallivm, vector_length);
1639
1640 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1641
1642 lp_build_loop_begin(&lp_loop, gallivm, zero);
1643 {
1644 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1645 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1646 LLVMValueRef io;
1647 LLVMValueRef clipmask; /* holds the clipmask value */
1648 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1649
1650 io_itr = lp_loop.counter;
1651
1652 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1653 #if DEBUG_STORE
1654 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1655 io_itr, io, lp_loop.counter);
1656 #endif
1657 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1658 for (i = 0; i < vector_length; ++i) {
1659 LLVMValueRef vert_index =
1660 LLVMBuildAdd(builder,
1661 lp_loop.counter,
1662 lp_build_const_int32(gallivm, i), "");
1663 LLVMValueRef true_index =
1664 LLVMBuildAdd(builder, start, vert_index, "");
1665 LLVMValueRef vertex_id;
1666
1667 /* make sure we're not out of bounds which can happen
1668 * if fetch_count % 4 != 0, because on the last iteration
1669 * a few of the 4 vertex fetches will be out of bounds */
1670 true_index = lp_build_min(&bld, true_index, fetch_max);
1671
1672 if (elts) {
1673 LLVMValueRef fetch_ptr;
1674 LLVMValueRef index_overflowed;
1675 LLVMValueRef index_ptr =
1676 lp_build_alloca(
1677 gallivm,
1678 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1679 struct lp_build_if_state if_ctx;
1680 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1681 true_index, fetch_elt_max,
1682 "index_overflowed");
1683
1684 lp_build_if(&if_ctx, gallivm, index_overflowed);
1685 {
1686 /* Generate maximum possible index so that
1687 * generate_fetch can treat it just like
1688 * any other overflow and return zeros.
1689 * We don't have to worry about the restart
1690 * primitive index because it has already been
1691 * handled
1692 */
1693 LLVMValueRef val =
1694 lp_build_const_int32(gallivm, 0xffffffff);
1695 LLVMBuildStore(builder, val, index_ptr);
1696 }
1697 lp_build_else(&if_ctx);
1698 {
1699 LLVMValueRef val;
1700 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1701 &true_index, 1, "");
1702 val = LLVMBuildLoad(builder, fetch_ptr, "");
1703 LLVMBuildStore(builder, val, index_ptr);
1704 }
1705 lp_build_endif(&if_ctx);
1706 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1707 }
1708 /* in the paths with elts vertex id has to be unaffected by the
1709 * index bias and because indices inside our elements array have
1710 * already had index bias applied we need to subtract it here to
1711 * get back to the original index.
1712 * in the linear paths vertex id has to be unaffected by the
1713 * original start index and because we abuse the 'start' variable
1714 * to either represent the actual start index or the index at which
1715 * the primitive was split (we split rendering into chunks of at
1716 * most 4095-vertices) we need to back out the original start
1717 * index out of our vertex id here.
1718 */
1719 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1720
1721 system_values.vertex_id = LLVMBuildInsertElement(
1722 gallivm->builder,
1723 system_values.vertex_id, vertex_id,
1724 lp_build_const_int32(gallivm, i), "");
1725
1726 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1727 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1728 LLVMValueRef vb_index =
1729 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1730 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1731 generate_fetch(gallivm, draw, vbuffers_ptr,
1732 &aos_attribs[j][i], velem, vb, true_index,
1733 system_values.instance_id);
1734 }
1735 }
1736 convert_to_soa(gallivm, aos_attribs, inputs,
1737 draw->pt.nr_vertex_elements, vs_type);
1738
1739 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1740 generate_vs(variant,
1741 builder,
1742 vs_type,
1743 outputs,
1744 ptr_aos,
1745 &system_values,
1746 context_ptr,
1747 sampler,
1748 key->clamp_vertex_color);
1749
1750 if (pos != -1 && cv != -1) {
1751 /* store original positions in clip before further manipulation */
1752 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1753 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1754
1755 /* do cliptest */
1756 if (enable_cliptest) {
1757 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1758 /* allocate clipmask, assign it integer type */
1759 clipmask = generate_clipmask(llvm,
1760 gallivm,
1761 vs_type,
1762 outputs,
1763 key->clip_xy,
1764 key->clip_z,
1765 key->clip_user,
1766 key->clip_halfz,
1767 key->ucp_enable,
1768 context_ptr, &have_clipdist);
1769 temp = LLVMBuildOr(builder, clipmask, temp, "");
1770 /* store temporary clipping boolean value */
1771 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1772 }
1773 else {
1774 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1775 }
1776
1777 /* do viewport mapping */
1778 if (!bypass_viewport) {
1779 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1780 }
1781 }
1782 else {
1783 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1784 }
1785
1786 /* store clipmask in vertex header,
1787 * original positions in clip
1788 * and transformed positions in data
1789 */
1790 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1791 vs_info->num_outputs, vs_type,
1792 have_clipdist);
1793 }
1794 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1795
1796 sampler->destroy(sampler);
1797
1798 /* return clipping boolean value for function */
1799 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1800
1801 LLVMBuildRet(builder, ret);
1802
1803 gallivm_verify_function(gallivm, variant_func);
1804 }
1805
1806
1807 struct draw_llvm_variant_key *
1808 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1809 {
1810 unsigned i;
1811 struct draw_llvm_variant_key *key;
1812 struct draw_sampler_static_state *draw_sampler;
1813
1814 key = (struct draw_llvm_variant_key *)store;
1815
1816 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1817
1818 /* Presumably all variants of the shader should have the same
1819 * number of vertex elements - ie the number of shader inputs.
1820 * NOTE: we NEED to store the needed number of needed inputs
1821 * here, not the number of provided elements to match keysize
1822 * (and the offset of sampler state in the key).
1823 */
1824 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1825 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1826
1827 /* will have to rig this up properly later */
1828 key->clip_xy = llvm->draw->clip_xy;
1829 key->clip_z = llvm->draw->clip_z;
1830 key->clip_user = llvm->draw->clip_user;
1831 key->bypass_viewport = llvm->draw->identity_viewport;
1832 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1833 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1834 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1835 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1836 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1837 key->pad1 = 0;
1838
1839 /* All variants of this shader will have the same value for
1840 * nr_samplers. Not yet trying to compact away holes in the
1841 * sampler array.
1842 */
1843 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1844 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1845 key->nr_sampler_views =
1846 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1847 }
1848 else {
1849 key->nr_sampler_views = key->nr_samplers;
1850 }
1851
1852 draw_sampler = draw_llvm_variant_key_samplers(key);
1853
1854 memcpy(key->vertex_element,
1855 llvm->draw->pt.vertex_element,
1856 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1857
1858 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1859
1860 for (i = 0 ; i < key->nr_samplers; i++) {
1861 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1862 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1863 }
1864 for (i = 0 ; i < key->nr_sampler_views; i++) {
1865 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1866 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1867 }
1868
1869 return key;
1870 }
1871
1872
1873 void
1874 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1875 {
1876 unsigned i;
1877 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1878
1879 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1880 debug_printf("clip_xy = %u\n", key->clip_xy);
1881 debug_printf("clip_z = %u\n", key->clip_z);
1882 debug_printf("clip_user = %u\n", key->clip_user);
1883 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1884 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1885 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1886 debug_printf("has_gs = %u\n", key->has_gs);
1887 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1888
1889 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1890 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1891 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1892 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1893 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1894 }
1895
1896 for (i = 0 ; i < key->nr_sampler_views; i++) {
1897 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1898 }
1899 }
1900
1901
1902 void
1903 draw_llvm_set_mapped_texture(struct draw_context *draw,
1904 unsigned shader_stage,
1905 unsigned sview_idx,
1906 uint32_t width, uint32_t height, uint32_t depth,
1907 uint32_t first_level, uint32_t last_level,
1908 const void *base_ptr,
1909 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1910 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1911 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1912 {
1913 unsigned j;
1914 struct draw_jit_texture *jit_tex;
1915
1916 assert(shader_stage == PIPE_SHADER_VERTEX ||
1917 shader_stage == PIPE_SHADER_GEOMETRY);
1918
1919 if (shader_stage == PIPE_SHADER_VERTEX) {
1920 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1921
1922 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1923 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1924 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1925
1926 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1927 } else {
1928 assert(0);
1929 return;
1930 }
1931
1932 jit_tex->width = width;
1933 jit_tex->height = height;
1934 jit_tex->depth = depth;
1935 jit_tex->first_level = first_level;
1936 jit_tex->last_level = last_level;
1937 jit_tex->base = base_ptr;
1938
1939 for (j = first_level; j <= last_level; j++) {
1940 jit_tex->mip_offsets[j] = mip_offsets[j];
1941 jit_tex->row_stride[j] = row_stride[j];
1942 jit_tex->img_stride[j] = img_stride[j];
1943 }
1944 }
1945
1946
1947 void
1948 draw_llvm_set_sampler_state(struct draw_context *draw,
1949 unsigned shader_type)
1950 {
1951 unsigned i;
1952
1953 if (shader_type == PIPE_SHADER_VERTEX) {
1954 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1955 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1956
1957 if (draw->samplers[i]) {
1958 const struct pipe_sampler_state *s
1959 = draw->samplers[PIPE_SHADER_VERTEX][i];
1960 jit_sam->min_lod = s->min_lod;
1961 jit_sam->max_lod = s->max_lod;
1962 jit_sam->lod_bias = s->lod_bias;
1963 COPY_4V(jit_sam->border_color, s->border_color.f);
1964 }
1965 }
1966 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1967 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1968 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1969
1970 if (draw->samplers[i]) {
1971 const struct pipe_sampler_state *s
1972 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1973 jit_sam->min_lod = s->min_lod;
1974 jit_sam->max_lod = s->max_lod;
1975 jit_sam->lod_bias = s->lod_bias;
1976 COPY_4V(jit_sam->border_color, s->border_color.f);
1977 }
1978 }
1979 }
1980 }
1981
1982
1983 void
1984 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1985 {
1986 struct draw_llvm *llvm = variant->llvm;
1987
1988 if (variant->function_elts) {
1989 gallivm_free_function(variant->gallivm,
1990 variant->function_elts, variant->jit_func_elts);
1991 }
1992
1993 if (variant->function) {
1994 gallivm_free_function(variant->gallivm,
1995 variant->function, variant->jit_func);
1996 }
1997
1998 gallivm_destroy(variant->gallivm);
1999
2000 remove_from_list(&variant->list_item_local);
2001 variant->shader->variants_cached--;
2002 remove_from_list(&variant->list_item_global);
2003 llvm->nr_variants--;
2004 FREE(variant);
2005 }
2006
2007
2008 /**
2009 * Create LLVM types for various structures.
2010 */
2011 static void
2012 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2013 {
2014 struct gallivm_state *gallivm = var->gallivm;
2015 LLVMTypeRef texture_type, sampler_type, context_type;
2016
2017 texture_type = create_jit_texture_type(gallivm, "texture");
2018 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2019
2020 context_type = create_gs_jit_context_type(gallivm,
2021 var->shader->base.vector_length,
2022 texture_type, sampler_type,
2023 "draw_gs_jit_context");
2024 var->context_ptr_type = LLVMPointerType(context_type, 0);
2025
2026 var->input_array_type = create_gs_jit_input_type(gallivm);
2027 }
2028
2029 static LLVMTypeRef
2030 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2031 {
2032 if (!variant->context_ptr_type)
2033 create_gs_jit_types(variant);
2034 return variant->context_ptr_type;
2035 }
2036
2037 static LLVMValueRef
2038 generate_mask_value(struct draw_gs_llvm_variant *variant,
2039 struct lp_type gs_type)
2040 {
2041 struct gallivm_state *gallivm = variant->gallivm;
2042 LLVMBuilderRef builder = gallivm->builder;
2043 LLVMValueRef bits[16];
2044 struct lp_type mask_type = lp_int_type(gs_type);
2045 struct lp_type mask_elem_type = lp_elem_type(mask_type);
2046 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
2047 mask_type,
2048 0);
2049 unsigned i;
2050
2051 assert(gs_type.length <= Elements(bits));
2052
2053 for (i = gs_type.length; i >= 1; --i) {
2054 int idx = i - 1;
2055 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2056 bits[idx] = lp_build_compare(gallivm,
2057 mask_elem_type, PIPE_FUNC_GEQUAL,
2058 variant->num_prims, ind);
2059 }
2060 for (i = 0; i < gs_type.length; ++i) {
2061 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2062 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
2063 }
2064 mask_val = lp_build_compare(gallivm,
2065 mask_type, PIPE_FUNC_NOTEQUAL,
2066 mask_val,
2067 lp_build_const_int_vec(gallivm, mask_type, 0));
2068
2069 return mask_val;
2070 }
2071
2072 static void
2073 draw_gs_llvm_generate(struct draw_llvm *llvm,
2074 struct draw_gs_llvm_variant *variant)
2075 {
2076 struct gallivm_state *gallivm = variant->gallivm;
2077 LLVMContextRef context = gallivm->context;
2078 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2079 LLVMTypeRef arg_types[6];
2080 LLVMTypeRef func_type;
2081 LLVMValueRef variant_func;
2082 LLVMValueRef context_ptr;
2083 LLVMValueRef prim_id_ptr;
2084 LLVMBasicBlockRef block;
2085 LLVMBuilderRef builder;
2086 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2087 struct lp_build_sampler_soa *sampler = 0;
2088 struct lp_build_context bld;
2089 struct lp_bld_tgsi_system_values system_values;
2090 struct lp_type gs_type;
2091 unsigned i;
2092 struct draw_gs_llvm_iface gs_iface;
2093 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2094 LLVMValueRef consts_ptr;
2095 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2096 struct lp_build_mask_context mask;
2097 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2098 unsigned vector_length = variant->shader->base.vector_length;
2099
2100 memset(&system_values, 0, sizeof(system_values));
2101
2102 assert(variant->vertex_header_ptr_type);
2103
2104 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2105 arg_types[1] = variant->input_array_type; /* input */
2106 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2107 arg_types[3] = int32_type; /* num_prims */
2108 arg_types[4] = int32_type; /* instance_id */
2109 arg_types[5] = LLVMPointerType(
2110 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2111
2112 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2113
2114 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2115 func_type);
2116 variant->function = variant_func;
2117
2118 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2119
2120 for (i = 0; i < Elements(arg_types); ++i)
2121 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2122 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2123 LLVMNoAliasAttribute);
2124
2125 context_ptr = LLVMGetParam(variant_func, 0);
2126 input_array = LLVMGetParam(variant_func, 1);
2127 io_ptr = LLVMGetParam(variant_func, 2);
2128 num_prims = LLVMGetParam(variant_func, 3);
2129 system_values.instance_id = LLVMGetParam(variant_func, 4);
2130 prim_id_ptr = LLVMGetParam(variant_func, 5);
2131
2132 lp_build_name(context_ptr, "context");
2133 lp_build_name(input_array, "input");
2134 lp_build_name(io_ptr, "io");
2135 lp_build_name(num_prims, "num_prims");
2136 lp_build_name(system_values.instance_id, "instance_id");
2137 lp_build_name(prim_id_ptr, "prim_id_ptr");
2138
2139 variant->context_ptr = context_ptr;
2140 variant->io_ptr = io_ptr;
2141 variant->num_prims = num_prims;
2142
2143 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2144 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2145 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2146 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2147 gs_iface.input = input_array;
2148 gs_iface.variant = variant;
2149
2150 /*
2151 * Function body
2152 */
2153
2154 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2155 builder = gallivm->builder;
2156 LLVMPositionBuilderAtEnd(builder, block);
2157
2158 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2159
2160 memset(&gs_type, 0, sizeof gs_type);
2161 gs_type.floating = TRUE; /* floating point values */
2162 gs_type.sign = TRUE; /* values are signed */
2163 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2164 gs_type.width = 32; /* 32-bit float */
2165 gs_type.length = vector_length;
2166
2167 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2168
2169 /* code generated texture sampling */
2170 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2171 context_ptr);
2172
2173 mask_val = generate_mask_value(variant, gs_type);
2174 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2175
2176 if (gs_info->uses_primid) {
2177 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
2178 }
2179
2180 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2181 tgsi_dump(tokens, 0);
2182 draw_gs_llvm_dump_variant_key(&variant->key);
2183 }
2184
2185 lp_build_tgsi_soa(variant->gallivm,
2186 tokens,
2187 gs_type,
2188 &mask,
2189 consts_ptr,
2190 &system_values,
2191 NULL,
2192 outputs,
2193 sampler,
2194 &llvm->draw->gs.geometry_shader->info,
2195 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2196
2197 sampler->destroy(sampler);
2198
2199 lp_build_mask_end(&mask);
2200
2201 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2202
2203 gallivm_verify_function(gallivm, variant_func);
2204 }
2205
2206
2207 struct draw_gs_llvm_variant *
2208 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2209 unsigned num_outputs,
2210 const struct draw_gs_llvm_variant_key *key)
2211 {
2212 struct draw_gs_llvm_variant *variant;
2213 struct llvm_geometry_shader *shader =
2214 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2215 LLVMTypeRef vertex_header;
2216
2217 variant = MALLOC(sizeof *variant +
2218 shader->variant_key_size -
2219 sizeof variant->key);
2220 if (variant == NULL)
2221 return NULL;
2222
2223 variant->llvm = llvm;
2224 variant->shader = shader;
2225
2226 variant->gallivm = gallivm_create();
2227
2228 create_gs_jit_types(variant);
2229
2230 memcpy(&variant->key, key, shader->variant_key_size);
2231
2232 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2233
2234 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2235
2236 draw_gs_llvm_generate(llvm, variant);
2237
2238 gallivm_compile_module(variant->gallivm);
2239
2240 variant->jit_func = (draw_gs_jit_func)
2241 gallivm_jit_function(variant->gallivm, variant->function);
2242
2243 variant->list_item_global.base = variant;
2244 variant->list_item_local.base = variant;
2245 /*variant->no = */shader->variants_created++;
2246 variant->list_item_global.base = variant;
2247
2248 return variant;
2249 }
2250
2251 void
2252 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2253 {
2254 struct draw_llvm *llvm = variant->llvm;
2255
2256 if (variant->function) {
2257 gallivm_free_function(variant->gallivm,
2258 variant->function, variant->jit_func);
2259 }
2260
2261 gallivm_destroy(variant->gallivm);
2262
2263 remove_from_list(&variant->list_item_local);
2264 variant->shader->variants_cached--;
2265 remove_from_list(&variant->list_item_global);
2266 llvm->nr_gs_variants--;
2267 FREE(variant);
2268 }
2269
2270 struct draw_gs_llvm_variant_key *
2271 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2272 {
2273 unsigned i;
2274 struct draw_gs_llvm_variant_key *key;
2275 struct draw_sampler_static_state *draw_sampler;
2276
2277 key = (struct draw_gs_llvm_variant_key *)store;
2278
2279 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2280
2281 /* All variants of this shader will have the same value for
2282 * nr_samplers. Not yet trying to compact away holes in the
2283 * sampler array.
2284 */
2285 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2286 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2287 key->nr_sampler_views =
2288 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2289 }
2290 else {
2291 key->nr_sampler_views = key->nr_samplers;
2292 }
2293
2294 draw_sampler = key->samplers;
2295
2296 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2297
2298 for (i = 0 ; i < key->nr_samplers; i++) {
2299 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2300 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2301 }
2302 for (i = 0 ; i < key->nr_sampler_views; i++) {
2303 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2304 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2305 }
2306
2307 return key;
2308 }
2309
2310 void
2311 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2312 {
2313 unsigned i;
2314 struct draw_sampler_static_state *sampler = key->samplers;
2315
2316 for (i = 0 ; i < key->nr_sampler_views; i++) {
2317 debug_printf("sampler[%i].src_format = %s\n", i,
2318 util_format_name(sampler[i].texture_state.format));
2319 }
2320 }