projects / mesa.git / blob
? search:


5100ce0039c6b89a046a81eb75285ec893673ef7
[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_logic.h"
36 #include "gallivm/lp_bld_const.h"
37 #include "gallivm/lp_bld_swizzle.h"
38 #include "gallivm/lp_bld_struct.h"
39 #include "gallivm/lp_bld_type.h"
40 #include "gallivm/lp_bld_flow.h"
41 #include "gallivm/lp_bld_debug.h"
42 #include "gallivm/lp_bld_tgsi.h"
43 #include "gallivm/lp_bld_printf.h"
44 #include "gallivm/lp_bld_intr.h"
45 #include "gallivm/lp_bld_init.h"
46 #include "gallivm/lp_bld_type.h"
47 #include "gallivm/lp_bld_pack.h"
48 #include "gallivm/lp_bld_format.h"
49
50 #include "tgsi/tgsi_exec.h"
51 #include "tgsi/tgsi_dump.h"
52
53 #include "util/u_math.h"
54 #include "util/u_pointer.h"
55 #include "util/u_string.h"
56 #include "util/u_simple_list.h"
57
58
59 #define DEBUG_STORE 0
60
61
62 static void
63 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
64 boolean elts);
65
66
67 struct draw_gs_llvm_iface {
68 struct lp_build_tgsi_gs_iface base;
69
70 struct draw_gs_llvm_variant *variant;
71 LLVMValueRef input;
72 };
73
74 static INLINE const struct draw_gs_llvm_iface *
75 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
76 {
77 return (const struct draw_gs_llvm_iface *)iface;
78 }
79
80 /**
81 * Create LLVM type for struct draw_jit_texture
82 */
83 static LLVMTypeRef
84 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
85 {
86 LLVMTargetDataRef target = gallivm->target;
87 LLVMTypeRef texture_type;
88 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
89 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
90
91 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
92 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
93 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
94 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
95 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
96 elem_types[DRAW_JIT_TEXTURE_BASE] =
97 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
98 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
99 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
100 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
101 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
102
103 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
104 Elements(elem_types), 0);
105
106 #if HAVE_LLVM < 0x0300
107 LLVMAddTypeName(gallivm->module, struct_name, texture_type);
108
109 /* Make sure the target's struct layout cache doesn't return
110 * stale/invalid data.
111 */
112 LLVMInvalidateStructLayout(gallivm->target, texture_type);
113 #endif
114
115 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
116 target, texture_type,
117 DRAW_JIT_TEXTURE_WIDTH);
118 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
119 target, texture_type,
120 DRAW_JIT_TEXTURE_HEIGHT);
121 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
122 target, texture_type,
123 DRAW_JIT_TEXTURE_DEPTH);
124 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
125 target, texture_type,
126 DRAW_JIT_TEXTURE_FIRST_LEVEL);
127 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
128 target, texture_type,
129 DRAW_JIT_TEXTURE_LAST_LEVEL);
130 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
131 target, texture_type,
132 DRAW_JIT_TEXTURE_BASE);
133 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
134 target, texture_type,
135 DRAW_JIT_TEXTURE_ROW_STRIDE);
136 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
137 target, texture_type,
138 DRAW_JIT_TEXTURE_IMG_STRIDE);
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
140 target, texture_type,
141 DRAW_JIT_TEXTURE_MIP_OFFSETS);
142
143 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
144
145 return texture_type;
146 }
147
148
149 /**
150 * Create LLVM type for struct draw_jit_sampler
151 */
152 static LLVMTypeRef
153 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
154 {
155 LLVMTargetDataRef target = gallivm->target;
156 LLVMTypeRef sampler_type;
157 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
158
159 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
160 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
161 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
162 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
163 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
164
165 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
166 Elements(elem_types), 0);
167
168 #if HAVE_LLVM < 0x0300
169 LLVMAddTypeName(gallivm->module, struct_name, sampler_type);
170
171 /* Make sure the target's struct layout cache doesn't return
172 * stale/invalid data.
173 */
174 LLVMInvalidateStructLayout(gallivm->target, sampler_type);
175 #endif
176
177 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
178 target, sampler_type,
179 DRAW_JIT_SAMPLER_MIN_LOD);
180 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
181 target, sampler_type,
182 DRAW_JIT_SAMPLER_MAX_LOD);
183 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
184 target, sampler_type,
185 DRAW_JIT_SAMPLER_LOD_BIAS);
186 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
187 target, sampler_type,
188 DRAW_JIT_SAMPLER_BORDER_COLOR);
189
190 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
191
192 return sampler_type;
193 }
194
195
196 /**
197 * Create LLVM type for struct draw_jit_context
198 */
199 static LLVMTypeRef
200 create_jit_context_type(struct gallivm_state *gallivm,
201 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
202 const char *struct_name)
203 {
204 LLVMTargetDataRef target = gallivm->target;
205 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
206 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
207 LLVMTypeRef context_type;
208
209 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
210 LP_MAX_TGSI_CONST_BUFFERS);
211 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
212 DRAW_TOTAL_CLIP_PLANES), 0);
213 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
214 elem_types[3] = LLVMArrayType(texture_type,
215 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
216 elem_types[4] = LLVMArrayType(sampler_type,
217 PIPE_MAX_SAMPLERS); /* samplers */
218 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
219 Elements(elem_types), 0);
220 #if HAVE_LLVM < 0x0300
221 LLVMAddTypeName(gallivm->module, struct_name, context_type);
222
223 LLVMInvalidateStructLayout(gallivm->target, context_type);
224 #endif
225
226 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
227 target, context_type, DRAW_JIT_CTX_CONSTANTS);
228 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
229 target, context_type, DRAW_JIT_CTX_PLANES);
230 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
231 target, context_type, DRAW_JIT_CTX_VIEWPORT);
232 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
233 target, context_type,
234 DRAW_JIT_CTX_TEXTURES);
235 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
236 target, context_type,
237 DRAW_JIT_CTX_SAMPLERS);
238 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
239 target, context_type);
240
241 return context_type;
242 }
243
244
245 /**
246 * Create LLVM type for struct draw_gs_jit_context
247 */
248 static LLVMTypeRef
249 create_gs_jit_context_type(struct gallivm_state *gallivm,
250 unsigned vector_length,
251 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
252 const char *struct_name)
253 {
254 LLVMTargetDataRef target = gallivm->target;
255 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
256 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
257 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
258 LLVMTypeRef context_type;
259
260 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
261 LP_MAX_TGSI_CONST_BUFFERS);
262 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
263 DRAW_TOTAL_CLIP_PLANES), 0);
264 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
265
266 elem_types[3] = LLVMArrayType(texture_type,
267 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
268 elem_types[4] = LLVMArrayType(sampler_type,
269 PIPE_MAX_SAMPLERS); /* samplers */
270
271 elem_types[5] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
272 elem_types[6] = LLVMPointerType(LLVMVectorType(int_type,
273 vector_length), 0);
274 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
275 vector_length), 0);
276
277 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
278 Elements(elem_types), 0);
279 #if HAVE_LLVM < 0x0300
280 LLVMAddTypeName(gallivm->module, struct_name, context_type);
281
282 LLVMInvalidateStructLayout(gallivm->target, context_type);
283 #endif
284
285 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
286 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
287 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
288 target, context_type, DRAW_GS_JIT_CTX_PLANES);
289 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
290 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
291 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
292 target, context_type,
293 DRAW_GS_JIT_CTX_TEXTURES);
294 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
295 target, context_type,
296 DRAW_GS_JIT_CTX_SAMPLERS);
297 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
298 target, context_type,
299 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
300 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
301 target, context_type,
302 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
303 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
304 target, context_type,
305 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
306 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
307 target, context_type);
308
309 return context_type;
310 }
311
312
313 static LLVMTypeRef
314 create_gs_jit_input_type(struct gallivm_state *gallivm)
315 {
316 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
317 LLVMTypeRef input_array;
318
319 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
320 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
321 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
322 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
323
324 return input_array;
325 }
326
327 /**
328 * Create LLVM type for struct pipe_vertex_buffer
329 */
330 static LLVMTypeRef
331 create_jit_vertex_buffer_type(struct gallivm_state *gallivm, const char *struct_name)
332 {
333 LLVMTargetDataRef target = gallivm->target;
334 LLVMTypeRef elem_types[4];
335 LLVMTypeRef vb_type;
336
337 elem_types[0] =
338 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
339 elem_types[2] =
340 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); /* vs_constants */
341
342 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
343 Elements(elem_types), 0);
344 #if HAVE_LLVM < 0x0300
345 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
346
347 LLVMInvalidateStructLayout(gallivm->target, vb_type);
348 #endif
349
350 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
351 target, vb_type, 0);
352 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
353 target, vb_type, 1);
354
355 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
356
357 return vb_type;
358 }
359
360
361 /**
362 * Create LLVM type for struct vertex_header;
363 */
364 static LLVMTypeRef
365 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
366 {
367 LLVMTargetDataRef target = gallivm->target;
368 LLVMTypeRef elem_types[4];
369 LLVMTypeRef vertex_header;
370 char struct_name[24];
371
372 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
373
374 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
375 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
376 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
377 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
378
379 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
380 Elements(elem_types), 0);
381 #if HAVE_LLVM < 0x0300
382 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
383
384 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
385 #endif
386
387 /* these are bit-fields and we can't take address of them
388 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
389 target, vertex_header,
390 DRAW_JIT_VERTEX_CLIPMASK);
391 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
392 target, vertex_header,
393 DRAW_JIT_VERTEX_EDGEFLAG);
394 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
395 target, vertex_header,
396 DRAW_JIT_VERTEX_PAD);
397 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
398 target, vertex_header,
399 DRAW_JIT_VERTEX_VERTEX_ID);
400 */
401 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
402 target, vertex_header,
403 DRAW_JIT_VERTEX_CLIP);
404 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
405 target, vertex_header,
406 DRAW_JIT_VERTEX_PRE_CLIP_POS);
407 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
408 target, vertex_header,
409 DRAW_JIT_VERTEX_DATA);
410
411 assert(LLVMABISizeOfType(target, vertex_header) ==
412 offsetof(struct vertex_header, data[data_elems]));
413
414 return vertex_header;
415 }
416
417
418 /**
419 * Create LLVM types for various structures.
420 */
421 static void
422 create_jit_types(struct draw_llvm_variant *variant)
423 {
424 struct gallivm_state *gallivm = variant->gallivm;
425 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type, vb_type;
426
427 texture_type = create_jit_texture_type(gallivm, "texture");
428 sampler_type = create_jit_sampler_type(gallivm, "sampler");
429
430 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
431 "draw_jit_context");
432 variant->context_ptr_type = LLVMPointerType(context_type, 0);
433
434 buffer_type = LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
435 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
436
437 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
438 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
439 }
440
441
442 static LLVMTypeRef
443 get_context_ptr_type(struct draw_llvm_variant *variant)
444 {
445 if (!variant->context_ptr_type)
446 create_jit_types(variant);
447 return variant->context_ptr_type;
448 }
449
450
451 static LLVMTypeRef
452 get_buffer_ptr_type(struct draw_llvm_variant *variant)
453 {
454 if (!variant->buffer_ptr_type)
455 create_jit_types(variant);
456 return variant->buffer_ptr_type;
457 }
458
459
460 static LLVMTypeRef
461 get_vb_ptr_type(struct draw_llvm_variant *variant)
462 {
463 if (!variant->vb_ptr_type)
464 create_jit_types(variant);
465 return variant->vb_ptr_type;
466 }
467
468 static LLVMTypeRef
469 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
470 {
471 if (!variant->vertex_header_ptr_type)
472 create_jit_types(variant);
473 return variant->vertex_header_ptr_type;
474 }
475
476
477 /**
478 * Create per-context LLVM info.
479 */
480 struct draw_llvm *
481 draw_llvm_create(struct draw_context *draw)
482 {
483 struct draw_llvm *llvm;
484
485 llvm = CALLOC_STRUCT( draw_llvm );
486 if (!llvm)
487 return NULL;
488
489 lp_build_init();
490
491 llvm->draw = draw;
492
493 llvm->nr_variants = 0;
494 make_empty_list(&llvm->vs_variants_list);
495
496 llvm->nr_gs_variants = 0;
497 make_empty_list(&llvm->gs_variants_list);
498
499 return llvm;
500 }
501
502
503 /**
504 * Free per-context LLVM info.
505 */
506 void
507 draw_llvm_destroy(struct draw_llvm *llvm)
508 {
509 /* XXX free other draw_llvm data? */
510 FREE(llvm);
511 }
512
513
514 /**
515 * Create LLVM-generated code for a vertex shader.
516 */
517 struct draw_llvm_variant *
518 draw_llvm_create_variant(struct draw_llvm *llvm,
519 unsigned num_inputs,
520 const struct draw_llvm_variant_key *key)
521 {
522 struct draw_llvm_variant *variant;
523 struct llvm_vertex_shader *shader =
524 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
525 LLVMTypeRef vertex_header;
526
527 variant = MALLOC(sizeof *variant +
528 shader->variant_key_size -
529 sizeof variant->key);
530 if (variant == NULL)
531 return NULL;
532
533 variant->llvm = llvm;
534
535 variant->gallivm = gallivm_create();
536
537 create_jit_types(variant);
538
539 memcpy(&variant->key, key, shader->variant_key_size);
540
541 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
542
543 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
544
545 draw_llvm_generate(llvm, variant, FALSE); /* linear */
546 draw_llvm_generate(llvm, variant, TRUE); /* elts */
547
548 gallivm_compile_module(variant->gallivm);
549
550 variant->jit_func = (draw_jit_vert_func)
551 gallivm_jit_function(variant->gallivm, variant->function);
552
553 variant->jit_func_elts = (draw_jit_vert_func_elts)
554 gallivm_jit_function(variant->gallivm, variant->function_elts);
555
556 variant->shader = shader;
557 variant->list_item_global.base = variant;
558 variant->list_item_local.base = variant;
559 /*variant->no = */shader->variants_created++;
560 variant->list_item_global.base = variant;
561
562 return variant;
563 }
564
565
566 static void
567 generate_vs(struct draw_llvm_variant *variant,
568 LLVMBuilderRef builder,
569 struct lp_type vs_type,
570 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
571 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
572 const struct lp_bld_tgsi_system_values *system_values,
573 LLVMValueRef context_ptr,
574 struct lp_build_sampler_soa *draw_sampler,
575 boolean clamp_vertex_color)
576 {
577 struct draw_llvm *llvm = variant->llvm;
578 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
579 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(variant->gallivm, context_ptr);
580 struct lp_build_sampler_soa *sampler = 0;
581
582 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
583 tgsi_dump(tokens, 0);
584 draw_llvm_dump_variant_key(&variant->key);
585 }
586
587 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
588 sampler = draw_sampler;
589
590 lp_build_tgsi_soa(variant->gallivm,
591 tokens,
592 vs_type,
593 NULL /*struct lp_build_mask_context *mask*/,
594 consts_ptr,
595 system_values,
596 NULL /*pos*/,
597 inputs,
598 outputs,
599 sampler,
600 &llvm->draw->vs.vertex_shader->info,
601 NULL);
602
603 {
604 LLVMValueRef out;
605 unsigned chan, attrib;
606 struct lp_build_context bld;
607 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
608 lp_build_context_init(&bld, variant->gallivm, vs_type);
609
610 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
611 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
612 if (outputs[attrib][chan]) {
613 switch (info->output_semantic_name[attrib]) {
614 case TGSI_SEMANTIC_COLOR:
615 case TGSI_SEMANTIC_BCOLOR:
616 if (clamp_vertex_color) {
617 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
618 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
619 LLVMBuildStore(builder, out, outputs[attrib][chan]);
620 }
621 break;
622 case TGSI_SEMANTIC_FOG:
623 if (chan == 1 || chan == 2)
624 LLVMBuildStore(builder, bld.zero, outputs[attrib][chan]);
625 else if (chan == 3)
626 LLVMBuildStore(builder, bld.one, outputs[attrib][chan]);
627 break;
628 }
629 }
630 }
631 }
632 }
633 }
634
635
636 static void
637 generate_fetch(struct gallivm_state *gallivm,
638 LLVMValueRef vbuffers_ptr,
639 LLVMValueRef *res,
640 struct pipe_vertex_element *velem,
641 LLVMValueRef vbuf,
642 LLVMValueRef index,
643 LLVMValueRef instance_id)
644 {
645 const struct util_format_description *format_desc = util_format_description(velem->src_format);
646 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
647 LLVMBuilderRef builder = gallivm->builder;
648 LLVMValueRef indices =
649 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
650 velem->vertex_buffer_index, 0);
651 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
652 &indices, 1, "");
653 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
654 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
655 LLVMValueRef stride;
656
657 if (velem->instance_divisor) {
658 /* array index = instance_id / instance_divisor */
659 index = LLVMBuildUDiv(builder, instance_id,
660 lp_build_const_int32(gallivm, velem->instance_divisor),
661 "instance_divisor");
662 }
663
664 stride = LLVMBuildMul(builder, vb_stride, index, "");
665
666 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
667
668 stride = LLVMBuildAdd(builder, stride,
669 vb_buffer_offset,
670 "");
671 stride = LLVMBuildAdd(builder, stride,
672 lp_build_const_int32(gallivm, velem->src_offset),
673 "");
674
675 /* lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);*/
676 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
677
678 *res = lp_build_fetch_rgba_aos(gallivm,
679 format_desc,
680 lp_float32_vec4_type(),
681 vbuffer_ptr,
682 zero, zero, zero);
683 }
684
685 static void
686 convert_to_soa(struct gallivm_state *gallivm,
687 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
688 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
689 unsigned num_attribs, const struct lp_type soa_type)
690 {
691 unsigned i, j, k;
692 struct lp_type aos_channel_type = soa_type;
693
694 debug_assert(TGSI_NUM_CHANNELS == 4);
695 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
696
697 aos_channel_type.length >>= 1;
698
699 for (i = 0; i < num_attribs; ++i) {
700 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
701 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
702
703 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
704 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
705
706 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
707
708 for (k = 0; k < pixels_per_channel; ++k) {
709 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
710 }
711
712 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
713 }
714
715 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
716 }
717 }
718
719
720 static void
721 store_aos(struct gallivm_state *gallivm,
722 LLVMValueRef io_ptr,
723 LLVMValueRef index,
724 LLVMValueRef value)
725 {
726 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
727 LLVMBuilderRef builder = gallivm->builder;
728 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
729 LLVMValueRef indices[3];
730
731 indices[0] = lp_build_const_int32(gallivm, 0);
732 indices[1] = index;
733 indices[2] = lp_build_const_int32(gallivm, 0);
734
735 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
736 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
737
738 #if DEBUG_STORE
739 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
740 #endif
741
742 /* Unaligned store due to the vertex header */
743 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
744 }
745
746 /**
747 * Adjust the mask to architecture endianess. The mask will the store in struct:
748 *
749 * struct vertex_header {
750 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
751 * unsigned edgeflag:1;
752 * unsigned have_clipdist:1;
753 * unsigned vertex_id:16;
754 * [...]
755 * }
756 *
757 * On little-endian machine nothing needs to done, however on bit-endian machine
758 * the mask's fields need to be adjusted with the algorithm:
759 *
760 * uint32_t reverse (uint32_t x)
761 * {
762 * return (x >> 16) | // vertex_id
763 * ((x & 0x3fff) << 18) | // clipmask
764 * ((x & 0x4000) << 3) | // have_clipdist
765 * ((x & 0x8000) << 1); // edgeflag
766 * }
767 */
768 static LLVMValueRef
769 adjust_mask(struct gallivm_state *gallivm,
770 LLVMValueRef mask)
771 {
772 #ifdef PIPE_ARCH_BIG_ENDIAN
773 LLVMBuilderRef builder = gallivm->builder;
774 LLVMValueRef vertex_id;
775 LLVMValueRef clipmask;
776 LLVMValueRef have_clipdist;
777 LLVMValueRef edgeflag;
778
779 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
780 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
781 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
782 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
783 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
784 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
785 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
786
787 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
788 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
789 mask = LLVMBuildOr(builder, mask, edgeflag, "");
790 #endif
791 return mask;
792 }
793
794 static void
795 store_aos_array(struct gallivm_state *gallivm,
796 struct lp_type soa_type,
797 LLVMValueRef io_ptr,
798 LLVMValueRef *indices,
799 LLVMValueRef* aos,
800 int attrib,
801 int num_outputs,
802 LLVMValueRef clipmask,
803 boolean have_clipdist)
804 {
805 LLVMBuilderRef builder = gallivm->builder;
806 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
807 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
808 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
809 int vector_length = soa_type.length;
810 int i;
811
812 debug_assert(TGSI_NUM_CHANNELS == 4);
813
814 for (i = 0; i < vector_length; i++) {
815 if (indices) {
816 inds[i] = indices[i];
817 } else {
818 inds[i] = lp_build_const_int32(gallivm, i);
819 }
820 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
821 }
822
823 if (attrib == 0) {
824 /* store vertex header for each of the n vertices */
825 LLVMValueRef val, cliptmp;
826 int vertex_id_pad_edgeflag;
827
828 /* If this assertion fails, it means we need to update the bit twidding
829 * code here. See struct vertex_header in draw_private.h.
830 */
831 assert(DRAW_TOTAL_CLIP_PLANES==14);
832 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
833 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
834 if (have_clipdist)
835 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
836 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
837 /* OR with the clipmask */
838 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
839 for (i = 0; i < vector_length; i++) {
840 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
841 val = LLVMBuildExtractElement(builder, cliptmp, inds[i], "");
842 val = adjust_mask(gallivm, val);
843 LLVMBuildStore(builder, val, id_ptr);
844 #if DEBUG_STORE
845 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
846 io_ptrs[i], inds[i], val);
847 #endif
848 }
849 }
850
851 /* store for each of the n vertices */
852 for (i = 0; i < vector_length; i++) {
853 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
854 }
855 }
856
857
858 static void
859 convert_to_aos(struct gallivm_state *gallivm,
860 LLVMValueRef io,
861 LLVMValueRef *indices,
862 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
863 LLVMValueRef clipmask,
864 int num_outputs,
865 struct lp_type soa_type,
866 boolean have_clipdist)
867 {
868 LLVMBuilderRef builder = gallivm->builder;
869 unsigned chan, attrib, i;
870
871 #if DEBUG_STORE
872 lp_build_printf(gallivm, " # storing begin\n");
873 #endif
874 for (attrib = 0; attrib < num_outputs; ++attrib) {
875 LLVMValueRef soa[TGSI_NUM_CHANNELS];
876 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
877 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
878 if (outputs[attrib][chan]) {
879 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
880 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
881 #if DEBUG_STORE
882 lp_build_printf(gallivm, "output %d : %d ",
883 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
884 attrib, 0),
885 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
886 chan, 0));
887 lp_build_print_value(gallivm, "val = ", out);
888 #endif
889 soa[chan] = out;
890 }
891 else {
892 soa[chan] = 0;
893 }
894 }
895
896
897 if (soa_type.length == TGSI_NUM_CHANNELS) {
898 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
899 } else {
900 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
901
902 for (i = 0; i < soa_type.length; ++i) {
903 aos[i] = lp_build_extract_range(gallivm,
904 soa[i % TGSI_NUM_CHANNELS],
905 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
906 TGSI_NUM_CHANNELS);
907 }
908 }
909
910 store_aos_array(gallivm,
911 soa_type,
912 io, indices,
913 aos,
914 attrib,
915 num_outputs,
916 clipmask, have_clipdist);
917 }
918 #if DEBUG_STORE
919 lp_build_printf(gallivm, " # storing end\n");
920 #endif
921 }
922
923
924 /**
925 * Stores original vertex positions in clip coordinates
926 */
927 static void
928 store_clip(struct gallivm_state *gallivm,
929 const struct lp_type vs_type,
930 LLVMValueRef io_ptr,
931 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
932 boolean pre_clip_pos, int idx)
933 {
934 LLVMBuilderRef builder = gallivm->builder;
935 LLVMValueRef soa[4];
936 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
937 LLVMValueRef indices[2];
938 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
939 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
940 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
941 int i, j;
942
943 indices[0] =
944 indices[1] = lp_build_const_int32(gallivm, 0);
945
946 for (i = 0; i < vs_type.length; i++) {
947 inds[i] = lp_build_const_int32(gallivm, i);
948 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
949 }
950
951 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
952 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
953 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
954 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
955
956 if (!pre_clip_pos) {
957 for (i = 0; i < vs_type.length; i++) {
958 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
959 }
960 } else {
961 for (i = 0; i < vs_type.length; i++) {
962 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
963 }
964 }
965
966 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
967 for (i = 0; i < vs_type.length; ++i) {
968 aos[i] = lp_build_extract_range(gallivm,
969 soa[i % TGSI_NUM_CHANNELS],
970 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
971 TGSI_NUM_CHANNELS);
972 }
973
974 for (j = 0; j < vs_type.length; j++) {
975 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
976 LLVMValueRef clip_ptr;
977
978 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
979 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
980
981 /* Unaligned store */
982 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
983 }
984 }
985
986
987 /**
988 * Transforms the outputs for viewport mapping
989 */
990 static void
991 generate_viewport(struct draw_llvm_variant *variant,
992 LLVMBuilderRef builder,
993 struct lp_type vs_type,
994 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
995 LLVMValueRef context_ptr)
996 {
997 int i;
998 struct gallivm_state *gallivm = variant->gallivm;
999 struct lp_type f32_type = vs_type;
1000 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1001 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 .. wn*/
1002 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1003 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1004
1005 /* for 1/w convention*/
1006 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1007 LLVMBuildStore(builder, out3, outputs[0][3]);
1008
1009 /* Viewport Mapping */
1010 for (i=0; i<3; i++) {
1011 LLVMValueRef out = LLVMBuildLoad(builder, outputs[0][i], ""); /*x0 x1 .. xn*/
1012 LLVMValueRef scale;
1013 LLVMValueRef trans;
1014 LLVMValueRef scale_i;
1015 LLVMValueRef trans_i;
1016 LLVMValueRef index;
1017
1018 index = lp_build_const_int32(gallivm, i);
1019 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1020
1021 index = lp_build_const_int32(gallivm, i+4);
1022 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1023
1024 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1025 LLVMBuildLoad(builder, scale_i, "scale"));
1026 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1027 LLVMBuildLoad(builder, trans_i, "trans"));
1028
1029 /* divide by w */
1030 out = LLVMBuildFMul(builder, out, out3, "");
1031 /* mult by scale */
1032 out = LLVMBuildFMul(builder, out, scale, "");
1033 /* add translation */
1034 out = LLVMBuildFAdd(builder, out, trans, "");
1035
1036 /* store transformed outputs */
1037 LLVMBuildStore(builder, out, outputs[0][i]);
1038 }
1039
1040 }
1041
1042
1043 /**
1044 * Returns clipmask as nxi32 bitmask for the n vertices
1045 */
1046 static LLVMValueRef
1047 generate_clipmask(struct draw_llvm *llvm,
1048 struct gallivm_state *gallivm,
1049 struct lp_type vs_type,
1050 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1051 boolean clip_xy,
1052 boolean clip_z,
1053 boolean clip_user,
1054 boolean clip_halfz,
1055 unsigned ucp_enable,
1056 LLVMValueRef context_ptr,
1057 boolean *have_clipdist)
1058 {
1059 LLVMBuilderRef builder = gallivm->builder;
1060 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1061 LLVMValueRef test, temp;
1062 LLVMValueRef zero, shift;
1063 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1064 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1065 LLVMValueRef plane1, planes, plane_ptr, sum;
1066 struct lp_type f32_type = vs_type;
1067 struct lp_type i32_type = lp_int_type(vs_type);
1068 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1069 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1070 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1071 bool have_cd = false;
1072 unsigned cd[2];
1073
1074 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1075 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1076
1077 if (cd[0] != pos || cd[1] != pos)
1078 have_cd = true;
1079
1080 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1081 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1082 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1083 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1084
1085 /*
1086 * load clipvertex and position from correct locations.
1087 * if they are the same just load them once.
1088 */
1089 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1090 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1091 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1092 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1093
1094 if (clip_user && cv != pos) {
1095 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1096 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1097 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1098 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1099 } else {
1100 cv_x = pos_x;
1101 cv_y = pos_y;
1102 cv_z = pos_z;
1103 cv_w = pos_w;
1104 }
1105
1106 /* Cliptest, for hardwired planes */
1107 if (clip_xy) {
1108 /* plane 1 */
1109 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1110 temp = shift;
1111 test = LLVMBuildAnd(builder, test, temp, "");
1112 mask = test;
1113
1114 /* plane 2 */
1115 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1116 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1117 temp = LLVMBuildShl(builder, temp, shift, "");
1118 test = LLVMBuildAnd(builder, test, temp, "");
1119 mask = LLVMBuildOr(builder, mask, test, "");
1120
1121 /* plane 3 */
1122 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1123 temp = LLVMBuildShl(builder, temp, shift, "");
1124 test = LLVMBuildAnd(builder, test, temp, "");
1125 mask = LLVMBuildOr(builder, mask, test, "");
1126
1127 /* plane 4 */
1128 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1129 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1130 temp = LLVMBuildShl(builder, temp, shift, "");
1131 test = LLVMBuildAnd(builder, test, temp, "");
1132 mask = LLVMBuildOr(builder, mask, test, "");
1133 }
1134
1135 if (clip_z) {
1136 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1137 if (clip_halfz) {
1138 /* plane 5 */
1139 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1140 test = LLVMBuildAnd(builder, test, temp, "");
1141 mask = LLVMBuildOr(builder, mask, test, "");
1142 }
1143 else {
1144 /* plane 5 */
1145 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1146 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1147 test = LLVMBuildAnd(builder, test, temp, "");
1148 mask = LLVMBuildOr(builder, mask, test, "");
1149 }
1150 /* plane 6 */
1151 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1152 temp = LLVMBuildShl(builder, temp, shift, "");
1153 test = LLVMBuildAnd(builder, test, temp, "");
1154 mask = LLVMBuildOr(builder, mask, test, "");
1155 }
1156
1157 if (clip_user) {
1158 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1159 LLVMValueRef indices[3];
1160
1161 /* userclip planes */
1162 while (ucp_enable) {
1163 unsigned plane_idx = ffs(ucp_enable)-1;
1164 ucp_enable &= ~(1 << plane_idx);
1165 plane_idx += 6;
1166
1167 if (have_cd && num_written_clipdistance) {
1168 LLVMValueRef clipdist;
1169 int i;
1170 i = plane_idx - 6;
1171
1172 *have_clipdist = TRUE;
1173 if (i < 4) {
1174 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1175 } else {
1176 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1177 }
1178 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1179 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1180 test = LLVMBuildAnd(builder, test, temp, "");
1181 mask = LLVMBuildOr(builder, mask, test, "");
1182 } else {
1183 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1184 indices[0] = lp_build_const_int32(gallivm, 0);
1185 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1186
1187 indices[2] = lp_build_const_int32(gallivm, 0);
1188 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1189 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1190 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1191 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1192
1193 indices[2] = lp_build_const_int32(gallivm, 1);
1194 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1195 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1196 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1197 test = LLVMBuildFMul(builder, planes, cv_y, "");
1198 sum = LLVMBuildFAdd(builder, sum, test, "");
1199
1200 indices[2] = lp_build_const_int32(gallivm, 2);
1201 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1202 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1203 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1204 test = LLVMBuildFMul(builder, planes, cv_z, "");
1205 sum = LLVMBuildFAdd(builder, sum, test, "");
1206
1207 indices[2] = lp_build_const_int32(gallivm, 3);
1208 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1209 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1210 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1211 test = LLVMBuildFMul(builder, planes, cv_w, "");
1212 sum = LLVMBuildFAdd(builder, sum, test, "");
1213
1214 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1215 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1216 test = LLVMBuildAnd(builder, test, temp, "");
1217 mask = LLVMBuildOr(builder, mask, test, "");
1218 }
1219 }
1220 }
1221 return mask;
1222 }
1223
1224
1225 /**
1226 * Returns boolean if any clipping has occurred
1227 * Used zero/non-zero i32 value to represent boolean
1228 */
1229 static LLVMValueRef
1230 clipmask_booli32(struct gallivm_state *gallivm,
1231 const struct lp_type vs_type,
1232 LLVMValueRef clipmask_bool_ptr)
1233 {
1234 LLVMBuilderRef builder = gallivm->builder;
1235 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1236 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1237 LLVMValueRef ret = LLVMConstNull(int32_type);
1238 LLVMValueRef temp;
1239 int i;
1240
1241 /*
1242 * Can do this with log2(vector length) pack instructions and one extract
1243 * (as we don't actually need a or) with sse2 which would be way better.
1244 */
1245 for (i=0; i < vs_type.length; i++) {
1246 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1247 lp_build_const_int32(gallivm, i) , "");
1248 ret = LLVMBuildOr(builder, ret, temp, "");
1249 }
1250 return ret;
1251 }
1252
1253 static LLVMValueRef
1254 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1255 struct lp_build_tgsi_context * bld_base,
1256 LLVMValueRef vertex_index,
1257 LLVMValueRef attrib_index,
1258 LLVMValueRef swizzle_index)
1259 {
1260 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1261 struct gallivm_state *gallivm = bld_base->base.gallivm;
1262 LLVMBuilderRef builder = gallivm->builder;
1263 LLVMValueRef indices[3];
1264 LLVMValueRef res;
1265
1266 indices[0] = vertex_index;
1267 indices[1] = attrib_index;
1268 indices[2] = swizzle_index;
1269
1270 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1271 res = LLVMBuildLoad(builder, res, "");
1272
1273 return res;
1274 }
1275
1276 static void
1277 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1278 struct lp_build_tgsi_context * bld_base,
1279 LLVMValueRef (*outputs)[4],
1280 LLVMValueRef emitted_vertices_vec)
1281 {
1282 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1283 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1284 struct gallivm_state *gallivm = variant->gallivm;
1285 LLVMBuilderRef builder = gallivm->builder;
1286 struct lp_type gs_type = bld_base->base.type;
1287 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1288 lp_int_type(gs_type), 0);
1289 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1290 LLVMValueRef max_output_vertices =
1291 lp_build_const_int32(gallivm, variant->shader->base.max_output_vertices);
1292 LLVMValueRef io = variant->io_ptr;
1293 unsigned i;
1294 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1295
1296 for (i = 0; i < gs_type.length; ++i) {
1297 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1298 LLVMValueRef currently_emitted =
1299 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1300 indices[i] = LLVMBuildMul(builder, ind, max_output_vertices, "");
1301 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1302 }
1303
1304 convert_to_aos(gallivm, io, indices,
1305 outputs, clipmask,
1306 gs_info->num_outputs, gs_type,
1307 FALSE);
1308 }
1309
1310 static void
1311 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1312 struct lp_build_tgsi_context * bld_base,
1313 LLVMValueRef verts_per_prim_vec,
1314 LLVMValueRef emitted_prims_vec)
1315 {
1316 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1317 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1318 struct gallivm_state *gallivm = variant->gallivm;
1319 LLVMBuilderRef builder = gallivm->builder;
1320 LLVMValueRef prim_lengts_ptr =
1321 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1322 unsigned i;
1323
1324 for (i = 0; i < bld_base->base.type.length; ++i) {
1325 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1326 LLVMValueRef prims_emitted =
1327 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1328 LLVMValueRef store_ptr;
1329 LLVMValueRef num_vertices =
1330 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1331
1332 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1333 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1334 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1335 LLVMBuildStore(builder, num_vertices, store_ptr);
1336 }
1337 }
1338
1339 static void
1340 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1341 struct lp_build_tgsi_context * bld_base,
1342 LLVMValueRef total_emitted_vertices_vec,
1343 LLVMValueRef emitted_prims_vec)
1344 {
1345 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1346 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1347 struct gallivm_state *gallivm = variant->gallivm;
1348 LLVMBuilderRef builder = gallivm->builder;
1349 LLVMValueRef emitted_verts_ptr =
1350 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1351 LLVMValueRef emitted_prims_ptr =
1352 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1353 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1354
1355 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1356 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1357
1358 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1359 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1360 }
1361
1362 static void
1363 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1364 boolean elts)
1365 {
1366 struct gallivm_state *gallivm = variant->gallivm;
1367 LLVMContextRef context = gallivm->context;
1368 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1369 LLVMTypeRef arg_types[8];
1370 LLVMTypeRef func_type;
1371 LLVMValueRef context_ptr;
1372 LLVMBasicBlockRef block;
1373 LLVMBuilderRef builder;
1374 struct lp_type vs_type;
1375 LLVMValueRef end, start;
1376 LLVMValueRef count, fetch_elts, fetch_count;
1377 LLVMValueRef stride, step, io_itr;
1378 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1379 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1380 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1381 struct draw_context *draw = llvm->draw;
1382 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1383 unsigned i, j;
1384 struct lp_build_context bld;
1385 struct lp_build_loop_state lp_loop;
1386 const int vector_length = lp_native_vector_width / 32;
1387 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1388 LLVMValueRef fetch_max;
1389 struct lp_build_sampler_soa *sampler = 0;
1390 LLVMValueRef ret, clipmask_bool_ptr;
1391 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1392 /* If geometry shader is present we need to skip both the viewport
1393 * transformation and clipping otherwise the inputs to the geometry
1394 * shader will be incorrect.
1395 */
1396 const boolean bypass_viewport = gs || variant->key.bypass_viewport;
1397 const boolean enable_cliptest = !gs && (variant->key.clip_xy ||
1398 variant->key.clip_z ||
1399 variant->key.clip_user);
1400 LLVMValueRef variant_func;
1401 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1402 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1403 boolean have_clipdist = FALSE;
1404 struct lp_bld_tgsi_system_values system_values;
1405
1406 memset(&system_values, 0, sizeof(system_values));
1407
1408 arg_types[0] = get_context_ptr_type(variant); /* context */
1409 arg_types[1] = get_vertex_header_ptr_type(variant); /* vertex_header */
1410 arg_types[2] = get_buffer_ptr_type(variant); /* vbuffers */
1411 if (elts)
1412 arg_types[3] = LLVMPointerType(int32_type, 0);/* fetch_elts * */
1413 else
1414 arg_types[3] = int32_type; /* start */
1415 arg_types[4] = int32_type; /* fetch_count / count */
1416 arg_types[5] = int32_type; /* stride */
1417 arg_types[6] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1418 arg_types[7] = int32_type; /* instance_id */
1419
1420 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1421
1422 variant_func = LLVMAddFunction(gallivm->module,
1423 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1424 func_type);
1425
1426 if (elts)
1427 variant->function_elts = variant_func;
1428 else
1429 variant->function = variant_func;
1430
1431 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1432 for (i = 0; i < Elements(arg_types); ++i)
1433 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1434 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1435 LLVMNoAliasAttribute);
1436
1437 context_ptr = LLVMGetParam(variant_func, 0);
1438 io_ptr = LLVMGetParam(variant_func, 1);
1439 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1440 stride = LLVMGetParam(variant_func, 5);
1441 vb_ptr = LLVMGetParam(variant_func, 6);
1442 system_values.instance_id = LLVMGetParam(variant_func, 7);
1443
1444 lp_build_name(context_ptr, "context");
1445 lp_build_name(io_ptr, "io");
1446 lp_build_name(vbuffers_ptr, "vbuffers");
1447 lp_build_name(stride, "stride");
1448 lp_build_name(vb_ptr, "vb");
1449 lp_build_name(system_values.instance_id, "instance_id");
1450
1451 if (elts) {
1452 fetch_elts = LLVMGetParam(variant_func, 3);
1453 fetch_count = LLVMGetParam(variant_func, 4);
1454 lp_build_name(fetch_elts, "fetch_elts");
1455 lp_build_name(fetch_count, "fetch_count");
1456 start = count = NULL;
1457 }
1458 else {
1459 start = LLVMGetParam(variant_func, 3);
1460 count = LLVMGetParam(variant_func, 4);
1461 lp_build_name(start, "start");
1462 lp_build_name(count, "count");
1463 fetch_elts = fetch_count = NULL;
1464 }
1465
1466 /*
1467 * Function body
1468 */
1469
1470 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1471 builder = gallivm->builder;
1472 LLVMPositionBuilderAtEnd(builder, block);
1473
1474 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1475
1476 memset(&vs_type, 0, sizeof vs_type);
1477 vs_type.floating = TRUE; /* floating point values */
1478 vs_type.sign = TRUE; /* values are signed */
1479 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1480 vs_type.width = 32; /* 32-bit float */
1481 vs_type.length = vector_length;
1482
1483 /* hold temporary "bool" clipmask */
1484 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1485 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1486
1487 /* code generated texture sampling */
1488 sampler = draw_llvm_sampler_soa_create(
1489 draw_llvm_variant_key_samplers(&variant->key),
1490 context_ptr);
1491
1492 if (elts) {
1493 start = zero;
1494 end = fetch_count;
1495 }
1496 else {
1497 end = lp_build_add(&bld, start, count);
1498 }
1499
1500 step = lp_build_const_int32(gallivm, vector_length);
1501
1502 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1503
1504 lp_build_loop_begin(&lp_loop, gallivm, start);
1505 {
1506 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1507 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1508 LLVMValueRef io;
1509 LLVMValueRef clipmask; /* holds the clipmask value */
1510 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1511
1512 if (elts)
1513 io_itr = lp_loop.counter;
1514 else
1515 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
1516
1517 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1518 #if DEBUG_STORE
1519 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1520 io_itr, io, lp_loop.counter);
1521 #endif
1522 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1523 for (i = 0; i < vector_length; ++i) {
1524 LLVMValueRef true_index =
1525 LLVMBuildAdd(builder,
1526 lp_loop.counter,
1527 lp_build_const_int32(gallivm, i), "");
1528
1529 /* make sure we're not out of bounds which can happen
1530 * if fetch_count % 4 != 0, because on the last iteration
1531 * a few of the 4 vertex fetches will be out of bounds */
1532 true_index = lp_build_min(&bld, true_index, fetch_max);
1533
1534 if (elts) {
1535 LLVMValueRef fetch_ptr;
1536 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1537 &true_index, 1, "");
1538 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
1539 }
1540
1541 system_values.vertex_id = LLVMBuildInsertElement(gallivm->builder,
1542 system_values.vertex_id, true_index,
1543 lp_build_const_int32(gallivm, i), "");
1544 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1545 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1546 LLVMValueRef vb_index =
1547 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1548 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1549 generate_fetch(gallivm, vbuffers_ptr,
1550 &aos_attribs[j][i], velem, vb, true_index,
1551 system_values.instance_id);
1552 }
1553 }
1554 convert_to_soa(gallivm, aos_attribs, inputs,
1555 draw->pt.nr_vertex_elements, vs_type);
1556
1557 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1558 generate_vs(variant,
1559 builder,
1560 vs_type,
1561 outputs,
1562 ptr_aos,
1563 &system_values,
1564 context_ptr,
1565 sampler,
1566 variant->key.clamp_vertex_color);
1567
1568 if (pos != -1 && cv != -1) {
1569 /* store original positions in clip before further manipulation */
1570 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1571 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1572
1573 /* do cliptest */
1574 if (enable_cliptest) {
1575 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1576 /* allocate clipmask, assign it integer type */
1577 clipmask = generate_clipmask(llvm,
1578 gallivm,
1579 vs_type,
1580 outputs,
1581 variant->key.clip_xy,
1582 variant->key.clip_z,
1583 variant->key.clip_user,
1584 variant->key.clip_halfz,
1585 variant->key.ucp_enable,
1586 context_ptr, &have_clipdist);
1587 temp = LLVMBuildOr(builder, clipmask, temp, "");
1588 /* store temporary clipping boolean value */
1589 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1590 }
1591 else {
1592 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1593 }
1594
1595 /* do viewport mapping */
1596 if (!bypass_viewport) {
1597 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1598 }
1599 }
1600 else {
1601 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1602 }
1603
1604 /* store clipmask in vertex header,
1605 * original positions in clip
1606 * and transformed positions in data
1607 */
1608 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1609 vs_info->num_outputs, vs_type,
1610 have_clipdist);
1611 }
1612
1613 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE);
1614
1615 sampler->destroy(sampler);
1616
1617 /* return clipping boolean value for function */
1618 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1619
1620 LLVMBuildRet(builder, ret);
1621
1622 gallivm_verify_function(gallivm, variant_func);
1623 }
1624
1625
1626 struct draw_llvm_variant_key *
1627 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1628 {
1629 unsigned i;
1630 struct draw_llvm_variant_key *key;
1631 struct draw_sampler_static_state *draw_sampler;
1632
1633 key = (struct draw_llvm_variant_key *)store;
1634
1635 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1636
1637 /* Presumably all variants of the shader should have the same
1638 * number of vertex elements - ie the number of shader inputs.
1639 * NOTE: we NEED to store the needed number of needed inputs
1640 * here, not the number of provided elements to match keysize
1641 * (and the offset of sampler state in the key).
1642 */
1643 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1644 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1645
1646 /* will have to rig this up properly later */
1647 key->clip_xy = llvm->draw->clip_xy;
1648 key->clip_z = llvm->draw->clip_z;
1649 key->clip_user = llvm->draw->clip_user;
1650 key->bypass_viewport = llvm->draw->identity_viewport;
1651 key->clip_halfz = !llvm->draw->rasterizer->gl_rasterization_rules;
1652 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1653 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1654 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1655 key->pad1 = 0;
1656
1657 /* All variants of this shader will have the same value for
1658 * nr_samplers. Not yet trying to compact away holes in the
1659 * sampler array.
1660 */
1661 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1662 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1663 key->nr_sampler_views =
1664 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1665 }
1666 else {
1667 key->nr_sampler_views = key->nr_samplers;
1668 }
1669
1670 draw_sampler = draw_llvm_variant_key_samplers(key);
1671
1672 memcpy(key->vertex_element,
1673 llvm->draw->pt.vertex_element,
1674 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1675
1676 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1677
1678 for (i = 0 ; i < key->nr_samplers; i++) {
1679 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1680 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1681 }
1682 for (i = 0 ; i < key->nr_sampler_views; i++) {
1683 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1684 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1685 }
1686
1687 return key;
1688 }
1689
1690
1691 void
1692 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1693 {
1694 unsigned i;
1695 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1696
1697 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1698 debug_printf("clip_xy = %u\n", key->clip_xy);
1699 debug_printf("clip_z = %u\n", key->clip_z);
1700 debug_printf("clip_user = %u\n", key->clip_user);
1701 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1702 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1703 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1704 debug_printf("has_gs = %u\n", key->has_gs);
1705 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1706
1707 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1708 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1709 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1710 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1711 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1712 }
1713
1714 for (i = 0 ; i < key->nr_sampler_views; i++) {
1715 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1716 }
1717 }
1718
1719
1720 void
1721 draw_llvm_set_mapped_texture(struct draw_context *draw,
1722 unsigned shader_stage,
1723 unsigned sview_idx,
1724 uint32_t width, uint32_t height, uint32_t depth,
1725 uint32_t first_level, uint32_t last_level,
1726 const void *base_ptr,
1727 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1728 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1729 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1730 {
1731 unsigned j;
1732 struct draw_jit_texture *jit_tex;
1733
1734 assert(shader_stage == PIPE_SHADER_VERTEX ||
1735 shader_stage == PIPE_SHADER_GEOMETRY);
1736
1737 if (shader_stage == PIPE_SHADER_VERTEX) {
1738 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1739
1740 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1741 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1742 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1743
1744 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1745 }
1746
1747 jit_tex->width = width;
1748 jit_tex->height = height;
1749 jit_tex->depth = depth;
1750 jit_tex->first_level = first_level;
1751 jit_tex->last_level = last_level;
1752 jit_tex->base = base_ptr;
1753
1754 for (j = first_level; j <= last_level; j++) {
1755 jit_tex->mip_offsets[j] = mip_offsets[j];
1756 jit_tex->row_stride[j] = row_stride[j];
1757 jit_tex->img_stride[j] = img_stride[j];
1758 }
1759 }
1760
1761
1762 void
1763 draw_llvm_set_sampler_state(struct draw_context *draw,
1764 unsigned shader_type)
1765 {
1766 unsigned i;
1767
1768 if (shader_type == PIPE_SHADER_VERTEX) {
1769 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1770 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1771
1772 if (draw->samplers[i]) {
1773 const struct pipe_sampler_state *s
1774 = draw->samplers[PIPE_SHADER_VERTEX][i];
1775 jit_sam->min_lod = s->min_lod;
1776 jit_sam->max_lod = s->max_lod;
1777 jit_sam->lod_bias = s->lod_bias;
1778 COPY_4V(jit_sam->border_color, s->border_color.f);
1779 }
1780 }
1781 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1782 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1783 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1784
1785 if (draw->samplers[i]) {
1786 const struct pipe_sampler_state *s
1787 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1788 jit_sam->min_lod = s->min_lod;
1789 jit_sam->max_lod = s->max_lod;
1790 jit_sam->lod_bias = s->lod_bias;
1791 COPY_4V(jit_sam->border_color, s->border_color.f);
1792 }
1793 }
1794 }
1795 }
1796
1797
1798 void
1799 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1800 {
1801 struct draw_llvm *llvm = variant->llvm;
1802
1803 if (variant->function_elts) {
1804 gallivm_free_function(variant->gallivm,
1805 variant->function_elts, variant->jit_func_elts);
1806 }
1807
1808 if (variant->function) {
1809 gallivm_free_function(variant->gallivm,
1810 variant->function, variant->jit_func);
1811 }
1812
1813 gallivm_destroy(variant->gallivm);
1814
1815 remove_from_list(&variant->list_item_local);
1816 variant->shader->variants_cached--;
1817 remove_from_list(&variant->list_item_global);
1818 llvm->nr_variants--;
1819 FREE(variant);
1820 }
1821
1822
1823 /**
1824 * Create LLVM types for various structures.
1825 */
1826 static void
1827 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1828 {
1829 struct gallivm_state *gallivm = var->gallivm;
1830 LLVMTypeRef texture_type, sampler_type, context_type;
1831
1832 texture_type = create_jit_texture_type(gallivm, "texture");
1833 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1834
1835 context_type = create_gs_jit_context_type(gallivm,
1836 var->shader->base.vector_length,
1837 texture_type, sampler_type,
1838 "draw_gs_jit_context");
1839 var->context_ptr_type = LLVMPointerType(context_type, 0);
1840
1841 var->input_array_type = create_gs_jit_input_type(gallivm);
1842 }
1843
1844 static LLVMTypeRef
1845 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
1846 {
1847 if (!variant->context_ptr_type)
1848 create_gs_jit_types(variant);
1849 return variant->context_ptr_type;
1850 }
1851
1852 static LLVMValueRef
1853 generate_mask_value(struct draw_gs_llvm_variant *variant,
1854 struct lp_type gs_type)
1855 {
1856 struct gallivm_state *gallivm = variant->gallivm;
1857 LLVMBuilderRef builder = gallivm->builder;
1858 LLVMValueRef bits[16];
1859 struct lp_type mask_type = lp_int_type(gs_type);
1860 struct lp_type mask_elem_type = lp_elem_type(mask_type);
1861 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
1862 mask_type,
1863 0);
1864 unsigned i;
1865
1866 assert(gs_type.length <= Elements(bits));
1867
1868 for (i = gs_type.length; i >= 1; --i) {
1869 int idx = i - 1;
1870 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1871 bits[idx] = lp_build_compare(gallivm,
1872 mask_elem_type, PIPE_FUNC_GEQUAL,
1873 variant->num_prims, ind);
1874 }
1875 for (i = 0; i < gs_type.length; ++i) {
1876 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1877 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
1878 }
1879 mask_val = lp_build_compare(gallivm,
1880 mask_type, PIPE_FUNC_NOTEQUAL,
1881 mask_val,
1882 lp_build_const_int_vec(gallivm, mask_type, 0));
1883
1884 return mask_val;
1885 }
1886
1887 static void
1888 draw_gs_llvm_generate(struct draw_llvm *llvm,
1889 struct draw_gs_llvm_variant *variant)
1890 {
1891 struct gallivm_state *gallivm = variant->gallivm;
1892 LLVMContextRef context = gallivm->context;
1893 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1894 LLVMTypeRef arg_types[6];
1895 LLVMTypeRef func_type;
1896 LLVMValueRef variant_func;
1897 LLVMValueRef context_ptr;
1898 LLVMValueRef prim_id_ptr;
1899 LLVMBasicBlockRef block;
1900 LLVMBuilderRef builder;
1901 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
1902 struct lp_build_sampler_soa *sampler = 0;
1903 struct lp_build_context bld;
1904 struct lp_bld_tgsi_system_values system_values;
1905 struct lp_type gs_type;
1906 unsigned i;
1907 struct draw_gs_llvm_iface gs_iface;
1908 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
1909 LLVMValueRef consts_ptr;
1910 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1911 struct lp_build_mask_context mask;
1912 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1913 unsigned vector_length = variant->shader->base.vector_length;
1914
1915 memset(&system_values, 0, sizeof(system_values));
1916
1917 assert(variant->vertex_header_ptr_type);
1918
1919 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
1920 arg_types[1] = variant->input_array_type; /* input */
1921 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
1922 arg_types[3] = int32_type; /* num_prims */
1923 arg_types[4] = int32_type; /* instance_id */
1924 arg_types[5] = LLVMPointerType(
1925 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
1926
1927 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1928
1929 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
1930 func_type);
1931 variant->function = variant_func;
1932
1933 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1934
1935 for (i = 0; i < Elements(arg_types); ++i)
1936 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1937 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1938 LLVMNoAliasAttribute);
1939
1940 context_ptr = LLVMGetParam(variant_func, 0);
1941 input_array = LLVMGetParam(variant_func, 1);
1942 io_ptr = LLVMGetParam(variant_func, 2);
1943 num_prims = LLVMGetParam(variant_func, 3);
1944 system_values.instance_id = LLVMGetParam(variant_func, 4);
1945 prim_id_ptr = LLVMGetParam(variant_func, 5);
1946
1947 lp_build_name(context_ptr, "context");
1948 lp_build_name(input_array, "input");
1949 lp_build_name(io_ptr, "io");
1950 lp_build_name(io_ptr, "num_prims");
1951 lp_build_name(system_values.instance_id, "instance_id");
1952 lp_build_name(prim_id_ptr, "prim_id_ptr");
1953
1954 variant->context_ptr = context_ptr;
1955 variant->io_ptr = io_ptr;
1956 variant->num_prims = num_prims;
1957
1958 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
1959 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
1960 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
1961 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
1962 gs_iface.input = input_array;
1963 gs_iface.variant = variant;
1964
1965 /*
1966 * Function body
1967 */
1968
1969 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1970 builder = gallivm->builder;
1971 LLVMPositionBuilderAtEnd(builder, block);
1972
1973 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1974
1975 memset(&gs_type, 0, sizeof gs_type);
1976 gs_type.floating = TRUE; /* floating point values */
1977 gs_type.sign = TRUE; /* values are signed */
1978 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1979 gs_type.width = 32; /* 32-bit float */
1980 gs_type.length = vector_length;
1981
1982 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
1983
1984 /* code generated texture sampling */
1985 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
1986 context_ptr);
1987
1988 mask_val = generate_mask_value(variant, gs_type);
1989 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
1990
1991 if (gs_info->uses_primid) {
1992 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
1993 }
1994
1995 lp_build_tgsi_soa(variant->gallivm,
1996 tokens,
1997 gs_type,
1998 &mask,
1999 consts_ptr,
2000 &system_values,
2001 NULL /*pos*/,
2002 NULL,
2003 outputs,
2004 sampler,
2005 &llvm->draw->gs.geometry_shader->info,
2006 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2007
2008 sampler->destroy(sampler);
2009
2010 lp_build_mask_end(&mask);
2011
2012 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2013
2014 gallivm_verify_function(gallivm, variant_func);
2015 }
2016
2017
2018 struct draw_gs_llvm_variant *
2019 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2020 unsigned num_outputs,
2021 const struct draw_gs_llvm_variant_key *key)
2022 {
2023 struct draw_gs_llvm_variant *variant;
2024 struct llvm_geometry_shader *shader =
2025 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2026 LLVMTypeRef vertex_header;
2027
2028 variant = MALLOC(sizeof *variant +
2029 shader->variant_key_size -
2030 sizeof variant->key);
2031 if (variant == NULL)
2032 return NULL;
2033
2034 variant->llvm = llvm;
2035 variant->shader = shader;
2036
2037 variant->gallivm = gallivm_create();
2038
2039 create_gs_jit_types(variant);
2040
2041 memcpy(&variant->key, key, shader->variant_key_size);
2042
2043 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2044
2045 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2046
2047 draw_gs_llvm_generate(llvm, variant);
2048
2049 gallivm_compile_module(variant->gallivm);
2050
2051 variant->jit_func = (draw_gs_jit_func)
2052 gallivm_jit_function(variant->gallivm, variant->function);
2053
2054 variant->list_item_global.base = variant;
2055 variant->list_item_local.base = variant;
2056 /*variant->no = */shader->variants_created++;
2057 variant->list_item_global.base = variant;
2058
2059 return variant;
2060 }
2061
2062 void
2063 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2064 {
2065 struct draw_llvm *llvm = variant->llvm;
2066
2067 if (variant->function) {
2068 gallivm_free_function(variant->gallivm,
2069 variant->function, variant->jit_func);
2070 }
2071
2072 gallivm_destroy(variant->gallivm);
2073
2074 remove_from_list(&variant->list_item_local);
2075 variant->shader->variants_cached--;
2076 remove_from_list(&variant->list_item_global);
2077 llvm->nr_gs_variants--;
2078 FREE(variant);
2079 }
2080
2081 struct draw_gs_llvm_variant_key *
2082 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2083 {
2084 unsigned i;
2085 struct draw_gs_llvm_variant_key *key;
2086 struct draw_sampler_static_state *draw_sampler;
2087
2088 key = (struct draw_gs_llvm_variant_key *)store;
2089
2090 /* All variants of this shader will have the same value for
2091 * nr_samplers. Not yet trying to compact away holes in the
2092 * sampler array.
2093 */
2094 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2095 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2096 key->nr_sampler_views =
2097 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2098 }
2099 else {
2100 key->nr_sampler_views = key->nr_samplers;
2101 }
2102
2103 draw_sampler = key->samplers;
2104
2105 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2106
2107 for (i = 0 ; i < key->nr_samplers; i++) {
2108 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2109 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2110 }
2111 for (i = 0 ; i < key->nr_sampler_views; i++) {
2112 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2113 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2114 }
2115
2116 return key;
2117 }
2118
2119 void
2120 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2121 {
2122 unsigned i;
2123 struct draw_sampler_static_state *sampler = key->samplers;
2124
2125 for (i = 0 ; i < key->nr_sampler_views; i++) {
2126 debug_printf("sampler[%i].src_format = %s\n", i,
2127 util_format_name(sampler[i].texture_state.format));
2128 }
2129 }