projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
tgsi/ureg: make the dst register match the src indirection
[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 inputs,
597 outputs,
598 sampler,
599 &llvm->draw->vs.vertex_shader->info,
600 NULL);
601
602 {
603 LLVMValueRef out;
604 unsigned chan, attrib;
605 struct lp_build_context bld;
606 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
607 lp_build_context_init(&bld, variant->gallivm, vs_type);
608
609 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
610 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
611 if (outputs[attrib][chan]) {
612 switch (info->output_semantic_name[attrib]) {
613 case TGSI_SEMANTIC_COLOR:
614 case TGSI_SEMANTIC_BCOLOR:
615 if (clamp_vertex_color) {
616 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
617 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
618 LLVMBuildStore(builder, out, outputs[attrib][chan]);
619 }
620 break;
621 case TGSI_SEMANTIC_FOG:
622 if (chan == 1 || chan == 2)
623 LLVMBuildStore(builder, bld.zero, outputs[attrib][chan]);
624 else if (chan == 3)
625 LLVMBuildStore(builder, bld.one, outputs[attrib][chan]);
626 break;
627 }
628 }
629 }
630 }
631 }
632 }
633
634
635 static void
636 generate_fetch(struct gallivm_state *gallivm,
637 LLVMValueRef vbuffers_ptr,
638 LLVMValueRef *res,
639 struct pipe_vertex_element *velem,
640 LLVMValueRef vbuf,
641 LLVMValueRef index,
642 LLVMValueRef instance_id)
643 {
644 const struct util_format_description *format_desc = util_format_description(velem->src_format);
645 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
646 LLVMBuilderRef builder = gallivm->builder;
647 LLVMValueRef indices =
648 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
649 velem->vertex_buffer_index, 0);
650 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
651 &indices, 1, "");
652 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
653 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
654 LLVMValueRef stride;
655
656 if (velem->instance_divisor) {
657 /* array index = instance_id / instance_divisor */
658 index = LLVMBuildUDiv(builder, instance_id,
659 lp_build_const_int32(gallivm, velem->instance_divisor),
660 "instance_divisor");
661 }
662
663 stride = LLVMBuildMul(builder, vb_stride, index, "");
664
665 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
666
667 stride = LLVMBuildAdd(builder, stride,
668 vb_buffer_offset,
669 "");
670 stride = LLVMBuildAdd(builder, stride,
671 lp_build_const_int32(gallivm, velem->src_offset),
672 "");
673
674 /* lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);*/
675 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
676
677 *res = lp_build_fetch_rgba_aos(gallivm,
678 format_desc,
679 lp_float32_vec4_type(),
680 vbuffer_ptr,
681 zero, zero, zero);
682 }
683
684 static void
685 convert_to_soa(struct gallivm_state *gallivm,
686 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
687 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
688 unsigned num_attribs, const struct lp_type soa_type)
689 {
690 unsigned i, j, k;
691 struct lp_type aos_channel_type = soa_type;
692
693 debug_assert(TGSI_NUM_CHANNELS == 4);
694 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
695
696 aos_channel_type.length >>= 1;
697
698 for (i = 0; i < num_attribs; ++i) {
699 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
700 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
701
702 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
703 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
704
705 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
706
707 for (k = 0; k < pixels_per_channel; ++k) {
708 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
709 }
710
711 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
712 }
713
714 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
715 }
716 }
717
718
719 static void
720 store_aos(struct gallivm_state *gallivm,
721 LLVMValueRef io_ptr,
722 LLVMValueRef index,
723 LLVMValueRef value)
724 {
725 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
726 LLVMBuilderRef builder = gallivm->builder;
727 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
728 LLVMValueRef indices[3];
729
730 indices[0] = lp_build_const_int32(gallivm, 0);
731 indices[1] = index;
732 indices[2] = lp_build_const_int32(gallivm, 0);
733
734 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
735 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
736
737 #if DEBUG_STORE
738 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
739 #endif
740
741 /* Unaligned store due to the vertex header */
742 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
743 }
744
745 /**
746 * Adjust the mask to architecture endianess. The mask will the store in struct:
747 *
748 * struct vertex_header {
749 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
750 * unsigned edgeflag:1;
751 * unsigned have_clipdist:1;
752 * unsigned vertex_id:16;
753 * [...]
754 * }
755 *
756 * On little-endian machine nothing needs to done, however on bit-endian machine
757 * the mask's fields need to be adjusted with the algorithm:
758 *
759 * uint32_t reverse (uint32_t x)
760 * {
761 * return (x >> 16) | // vertex_id
762 * ((x & 0x3fff) << 18) | // clipmask
763 * ((x & 0x4000) << 3) | // have_clipdist
764 * ((x & 0x8000) << 1); // edgeflag
765 * }
766 */
767 static LLVMValueRef
768 adjust_mask(struct gallivm_state *gallivm,
769 LLVMValueRef mask)
770 {
771 #ifdef PIPE_ARCH_BIG_ENDIAN
772 LLVMBuilderRef builder = gallivm->builder;
773 LLVMValueRef vertex_id;
774 LLVMValueRef clipmask;
775 LLVMValueRef have_clipdist;
776 LLVMValueRef edgeflag;
777
778 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
779 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
780 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
781 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
782 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
783 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
784 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
785
786 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
787 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
788 mask = LLVMBuildOr(builder, mask, edgeflag, "");
789 #endif
790 return mask;
791 }
792
793 static void
794 store_aos_array(struct gallivm_state *gallivm,
795 struct lp_type soa_type,
796 LLVMValueRef io_ptr,
797 LLVMValueRef *indices,
798 LLVMValueRef* aos,
799 int attrib,
800 int num_outputs,
801 LLVMValueRef clipmask,
802 boolean have_clipdist)
803 {
804 LLVMBuilderRef builder = gallivm->builder;
805 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
806 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
807 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
808 int vector_length = soa_type.length;
809 int i;
810
811 debug_assert(TGSI_NUM_CHANNELS == 4);
812
813 for (i = 0; i < vector_length; i++) {
814 if (indices) {
815 inds[i] = indices[i];
816 } else {
817 inds[i] = lp_build_const_int32(gallivm, i);
818 }
819 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
820 }
821
822 if (attrib == 0) {
823 /* store vertex header for each of the n vertices */
824 LLVMValueRef val, cliptmp;
825 int vertex_id_pad_edgeflag;
826
827 /* If this assertion fails, it means we need to update the bit twidding
828 * code here. See struct vertex_header in draw_private.h.
829 */
830 assert(DRAW_TOTAL_CLIP_PLANES==14);
831 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
832 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
833 if (have_clipdist)
834 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
835 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
836 /* OR with the clipmask */
837 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
838 for (i = 0; i < vector_length; i++) {
839 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
840 val = LLVMBuildExtractElement(builder, cliptmp, inds[i], "");
841 val = adjust_mask(gallivm, val);
842 LLVMBuildStore(builder, val, id_ptr);
843 #if DEBUG_STORE
844 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
845 io_ptrs[i], inds[i], val);
846 #endif
847 }
848 }
849
850 /* store for each of the n vertices */
851 for (i = 0; i < vector_length; i++) {
852 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
853 }
854 }
855
856
857 static void
858 convert_to_aos(struct gallivm_state *gallivm,
859 LLVMValueRef io,
860 LLVMValueRef *indices,
861 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
862 LLVMValueRef clipmask,
863 int num_outputs,
864 struct lp_type soa_type,
865 boolean have_clipdist)
866 {
867 LLVMBuilderRef builder = gallivm->builder;
868 unsigned chan, attrib, i;
869
870 #if DEBUG_STORE
871 lp_build_printf(gallivm, " # storing begin\n");
872 #endif
873 for (attrib = 0; attrib < num_outputs; ++attrib) {
874 LLVMValueRef soa[TGSI_NUM_CHANNELS];
875 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
876 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
877 if (outputs[attrib][chan]) {
878 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
879 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
880 #if DEBUG_STORE
881 lp_build_printf(gallivm, "output %d : %d ",
882 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
883 attrib, 0),
884 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
885 chan, 0));
886 lp_build_print_value(gallivm, "val = ", out);
887 #endif
888 soa[chan] = out;
889 }
890 else {
891 soa[chan] = 0;
892 }
893 }
894
895
896 if (soa_type.length == TGSI_NUM_CHANNELS) {
897 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
898 } else {
899 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
900
901 for (i = 0; i < soa_type.length; ++i) {
902 aos[i] = lp_build_extract_range(gallivm,
903 soa[i % TGSI_NUM_CHANNELS],
904 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
905 TGSI_NUM_CHANNELS);
906 }
907 }
908
909 store_aos_array(gallivm,
910 soa_type,
911 io, indices,
912 aos,
913 attrib,
914 num_outputs,
915 clipmask, have_clipdist);
916 }
917 #if DEBUG_STORE
918 lp_build_printf(gallivm, " # storing end\n");
919 #endif
920 }
921
922
923 /**
924 * Stores original vertex positions in clip coordinates
925 */
926 static void
927 store_clip(struct gallivm_state *gallivm,
928 const struct lp_type vs_type,
929 LLVMValueRef io_ptr,
930 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
931 boolean pre_clip_pos, int idx)
932 {
933 LLVMBuilderRef builder = gallivm->builder;
934 LLVMValueRef soa[4];
935 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
936 LLVMValueRef indices[2];
937 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
938 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
939 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
940 int i, j;
941
942 indices[0] =
943 indices[1] = lp_build_const_int32(gallivm, 0);
944
945 for (i = 0; i < vs_type.length; i++) {
946 inds[i] = lp_build_const_int32(gallivm, i);
947 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
948 }
949
950 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
951 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
952 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
953 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
954
955 if (!pre_clip_pos) {
956 for (i = 0; i < vs_type.length; i++) {
957 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
958 }
959 } else {
960 for (i = 0; i < vs_type.length; i++) {
961 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
962 }
963 }
964
965 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
966 for (i = 0; i < vs_type.length; ++i) {
967 aos[i] = lp_build_extract_range(gallivm,
968 soa[i % TGSI_NUM_CHANNELS],
969 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
970 TGSI_NUM_CHANNELS);
971 }
972
973 for (j = 0; j < vs_type.length; j++) {
974 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
975 LLVMValueRef clip_ptr;
976
977 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
978 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
979
980 /* Unaligned store */
981 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
982 }
983 }
984
985
986 /**
987 * Transforms the outputs for viewport mapping
988 */
989 static void
990 generate_viewport(struct draw_llvm_variant *variant,
991 LLVMBuilderRef builder,
992 struct lp_type vs_type,
993 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
994 LLVMValueRef context_ptr)
995 {
996 int i;
997 struct gallivm_state *gallivm = variant->gallivm;
998 struct lp_type f32_type = vs_type;
999 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1000 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1001 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][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[pos][3]);
1008
1009 /* Viewport Mapping */
1010 for (i=0; i<3; i++) {
1011 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][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[pos][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 boolean is_indirect,
1257 LLVMValueRef vertex_index,
1258 LLVMValueRef attrib_index,
1259 LLVMValueRef swizzle_index)
1260 {
1261 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1262 struct gallivm_state *gallivm = bld_base->base.gallivm;
1263 LLVMBuilderRef builder = gallivm->builder;
1264 LLVMValueRef indices[3];
1265 LLVMValueRef res;
1266 struct lp_type type = bld_base->base.type;
1267
1268 if (is_indirect) {
1269 int i;
1270 res = bld_base->base.zero;
1271 for (i = 0; i < type.length; ++i) {
1272 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1273 LLVMValueRef vert_chan_index = LLVMBuildExtractElement(builder,
1274 vertex_index, idx, "");
1275 LLVMValueRef channel_vec, value;
1276 indices[0] = vert_chan_index;
1277 indices[1] = attrib_index;
1278 indices[2] = swizzle_index;
1279
1280 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1281 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1282 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1283
1284 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1285 }
1286 } else {
1287 indices[0] = vertex_index;
1288 indices[1] = attrib_index;
1289 indices[2] = swizzle_index;
1290
1291 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1292 res = LLVMBuildLoad(builder, res, "");
1293 }
1294
1295 return res;
1296 }
1297
1298 static void
1299 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1300 struct lp_build_tgsi_context * bld_base,
1301 LLVMValueRef (*outputs)[4],
1302 LLVMValueRef emitted_vertices_vec)
1303 {
1304 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1305 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1306 struct gallivm_state *gallivm = variant->gallivm;
1307 LLVMBuilderRef builder = gallivm->builder;
1308 struct lp_type gs_type = bld_base->base.type;
1309 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1310 lp_int_type(gs_type), 0);
1311 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1312 LLVMValueRef next_prim_offset =
1313 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1314 LLVMValueRef io = variant->io_ptr;
1315 unsigned i;
1316 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1317
1318 for (i = 0; i < gs_type.length; ++i) {
1319 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1320 LLVMValueRef currently_emitted =
1321 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1322 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1323 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1324 }
1325
1326 convert_to_aos(gallivm, io, indices,
1327 outputs, clipmask,
1328 gs_info->num_outputs, gs_type,
1329 FALSE);
1330 }
1331
1332 static void
1333 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1334 struct lp_build_tgsi_context * bld_base,
1335 LLVMValueRef verts_per_prim_vec,
1336 LLVMValueRef emitted_prims_vec)
1337 {
1338 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1339 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1340 struct gallivm_state *gallivm = variant->gallivm;
1341 LLVMBuilderRef builder = gallivm->builder;
1342 LLVMValueRef prim_lengts_ptr =
1343 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1344 unsigned i;
1345
1346 for (i = 0; i < bld_base->base.type.length; ++i) {
1347 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1348 LLVMValueRef prims_emitted =
1349 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1350 LLVMValueRef store_ptr;
1351 LLVMValueRef num_vertices =
1352 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1353
1354 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1355 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1356 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1357 LLVMBuildStore(builder, num_vertices, store_ptr);
1358 }
1359 }
1360
1361 static void
1362 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1363 struct lp_build_tgsi_context * bld_base,
1364 LLVMValueRef total_emitted_vertices_vec,
1365 LLVMValueRef emitted_prims_vec)
1366 {
1367 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1368 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1369 struct gallivm_state *gallivm = variant->gallivm;
1370 LLVMBuilderRef builder = gallivm->builder;
1371 LLVMValueRef emitted_verts_ptr =
1372 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1373 LLVMValueRef emitted_prims_ptr =
1374 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1375 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1376
1377 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1378 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1379
1380 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1381 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1382 }
1383
1384 static void
1385 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1386 boolean elts)
1387 {
1388 struct gallivm_state *gallivm = variant->gallivm;
1389 LLVMContextRef context = gallivm->context;
1390 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1391 LLVMTypeRef arg_types[8];
1392 LLVMTypeRef func_type;
1393 LLVMValueRef context_ptr;
1394 LLVMBasicBlockRef block;
1395 LLVMBuilderRef builder;
1396 struct lp_type vs_type;
1397 LLVMValueRef end, start;
1398 LLVMValueRef count, fetch_elts, fetch_count;
1399 LLVMValueRef stride, step, io_itr;
1400 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1401 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1402 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1403 struct draw_context *draw = llvm->draw;
1404 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1405 unsigned i, j;
1406 struct lp_build_context bld;
1407 struct lp_build_loop_state lp_loop;
1408 const int vector_length = lp_native_vector_width / 32;
1409 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1410 LLVMValueRef fetch_max;
1411 struct lp_build_sampler_soa *sampler = 0;
1412 LLVMValueRef ret, clipmask_bool_ptr;
1413 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1414 struct draw_llvm_variant_key *key = &variant->key;
1415 /* If geometry shader is present we need to skip both the viewport
1416 * transformation and clipping otherwise the inputs to the geometry
1417 * shader will be incorrect.
1418 */
1419 const boolean bypass_viewport = gs || key->bypass_viewport;
1420 const boolean enable_cliptest = !gs && (key->clip_xy ||
1421 key->clip_z ||
1422 key->clip_user);
1423 LLVMValueRef variant_func;
1424 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1425 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1426 boolean have_clipdist = FALSE;
1427 struct lp_bld_tgsi_system_values system_values;
1428
1429 memset(&system_values, 0, sizeof(system_values));
1430
1431 arg_types[0] = get_context_ptr_type(variant); /* context */
1432 arg_types[1] = get_vertex_header_ptr_type(variant); /* vertex_header */
1433 arg_types[2] = get_buffer_ptr_type(variant); /* vbuffers */
1434 if (elts)
1435 arg_types[3] = LLVMPointerType(int32_type, 0);/* fetch_elts * */
1436 else
1437 arg_types[3] = int32_type; /* start */
1438 arg_types[4] = int32_type; /* fetch_count / count */
1439 arg_types[5] = int32_type; /* stride */
1440 arg_types[6] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1441 arg_types[7] = int32_type; /* instance_id */
1442
1443 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1444
1445 variant_func = LLVMAddFunction(gallivm->module,
1446 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1447 func_type);
1448
1449 if (elts)
1450 variant->function_elts = variant_func;
1451 else
1452 variant->function = variant_func;
1453
1454 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1455 for (i = 0; i < Elements(arg_types); ++i)
1456 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1457 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1458 LLVMNoAliasAttribute);
1459
1460 context_ptr = LLVMGetParam(variant_func, 0);
1461 io_ptr = LLVMGetParam(variant_func, 1);
1462 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1463 stride = LLVMGetParam(variant_func, 5);
1464 vb_ptr = LLVMGetParam(variant_func, 6);
1465 system_values.instance_id = LLVMGetParam(variant_func, 7);
1466
1467 lp_build_name(context_ptr, "context");
1468 lp_build_name(io_ptr, "io");
1469 lp_build_name(vbuffers_ptr, "vbuffers");
1470 lp_build_name(stride, "stride");
1471 lp_build_name(vb_ptr, "vb");
1472 lp_build_name(system_values.instance_id, "instance_id");
1473
1474 if (elts) {
1475 fetch_elts = LLVMGetParam(variant_func, 3);
1476 fetch_count = LLVMGetParam(variant_func, 4);
1477 lp_build_name(fetch_elts, "fetch_elts");
1478 lp_build_name(fetch_count, "fetch_count");
1479 start = count = NULL;
1480 }
1481 else {
1482 start = LLVMGetParam(variant_func, 3);
1483 count = LLVMGetParam(variant_func, 4);
1484 lp_build_name(start, "start");
1485 lp_build_name(count, "count");
1486 fetch_elts = fetch_count = NULL;
1487 }
1488
1489 /*
1490 * Function body
1491 */
1492
1493 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1494 builder = gallivm->builder;
1495 LLVMPositionBuilderAtEnd(builder, block);
1496
1497 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1498
1499 memset(&vs_type, 0, sizeof vs_type);
1500 vs_type.floating = TRUE; /* floating point values */
1501 vs_type.sign = TRUE; /* values are signed */
1502 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1503 vs_type.width = 32; /* 32-bit float */
1504 vs_type.length = vector_length;
1505
1506 /* hold temporary "bool" clipmask */
1507 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1508 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1509
1510 /* code generated texture sampling */
1511 sampler = draw_llvm_sampler_soa_create(
1512 draw_llvm_variant_key_samplers(key),
1513 context_ptr);
1514
1515 if (elts) {
1516 start = zero;
1517 end = fetch_count;
1518 }
1519 else {
1520 end = lp_build_add(&bld, start, count);
1521 }
1522
1523 step = lp_build_const_int32(gallivm, vector_length);
1524
1525 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1526
1527 lp_build_loop_begin(&lp_loop, gallivm, start);
1528 {
1529 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1530 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1531 LLVMValueRef io;
1532 LLVMValueRef clipmask; /* holds the clipmask value */
1533 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1534
1535 if (elts)
1536 io_itr = lp_loop.counter;
1537 else
1538 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
1539
1540 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1541 #if DEBUG_STORE
1542 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1543 io_itr, io, lp_loop.counter);
1544 #endif
1545 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1546 for (i = 0; i < vector_length; ++i) {
1547 LLVMValueRef true_index =
1548 LLVMBuildAdd(builder,
1549 lp_loop.counter,
1550 lp_build_const_int32(gallivm, i), "");
1551
1552 /* make sure we're not out of bounds which can happen
1553 * if fetch_count % 4 != 0, because on the last iteration
1554 * a few of the 4 vertex fetches will be out of bounds */
1555 true_index = lp_build_min(&bld, true_index, fetch_max);
1556
1557 if (elts) {
1558 LLVMValueRef fetch_ptr;
1559 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1560 &true_index, 1, "");
1561 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
1562 }
1563
1564 system_values.vertex_id = LLVMBuildInsertElement(gallivm->builder,
1565 system_values.vertex_id, true_index,
1566 lp_build_const_int32(gallivm, i), "");
1567 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1568 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1569 LLVMValueRef vb_index =
1570 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1571 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1572 generate_fetch(gallivm, vbuffers_ptr,
1573 &aos_attribs[j][i], velem, vb, true_index,
1574 system_values.instance_id);
1575 }
1576 }
1577 convert_to_soa(gallivm, aos_attribs, inputs,
1578 draw->pt.nr_vertex_elements, vs_type);
1579
1580 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1581 generate_vs(variant,
1582 builder,
1583 vs_type,
1584 outputs,
1585 ptr_aos,
1586 &system_values,
1587 context_ptr,
1588 sampler,
1589 key->clamp_vertex_color);
1590
1591 if (pos != -1 && cv != -1) {
1592 /* store original positions in clip before further manipulation */
1593 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1594 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1595
1596 /* do cliptest */
1597 if (enable_cliptest) {
1598 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1599 /* allocate clipmask, assign it integer type */
1600 clipmask = generate_clipmask(llvm,
1601 gallivm,
1602 vs_type,
1603 outputs,
1604 key->clip_xy,
1605 key->clip_z,
1606 key->clip_user,
1607 key->clip_halfz,
1608 key->ucp_enable,
1609 context_ptr, &have_clipdist);
1610 temp = LLVMBuildOr(builder, clipmask, temp, "");
1611 /* store temporary clipping boolean value */
1612 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1613 }
1614 else {
1615 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1616 }
1617
1618 /* do viewport mapping */
1619 if (!bypass_viewport) {
1620 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1621 }
1622 }
1623 else {
1624 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1625 }
1626
1627 /* store clipmask in vertex header,
1628 * original positions in clip
1629 * and transformed positions in data
1630 */
1631 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1632 vs_info->num_outputs, vs_type,
1633 have_clipdist);
1634 }
1635
1636 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE);
1637
1638 sampler->destroy(sampler);
1639
1640 /* return clipping boolean value for function */
1641 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1642
1643 LLVMBuildRet(builder, ret);
1644
1645 gallivm_verify_function(gallivm, variant_func);
1646 }
1647
1648
1649 struct draw_llvm_variant_key *
1650 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1651 {
1652 unsigned i;
1653 struct draw_llvm_variant_key *key;
1654 struct draw_sampler_static_state *draw_sampler;
1655
1656 key = (struct draw_llvm_variant_key *)store;
1657
1658 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1659
1660 /* Presumably all variants of the shader should have the same
1661 * number of vertex elements - ie the number of shader inputs.
1662 * NOTE: we NEED to store the needed number of needed inputs
1663 * here, not the number of provided elements to match keysize
1664 * (and the offset of sampler state in the key).
1665 */
1666 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1667 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1668
1669 /* will have to rig this up properly later */
1670 key->clip_xy = llvm->draw->clip_xy;
1671 key->clip_z = llvm->draw->clip_z;
1672 key->clip_user = llvm->draw->clip_user;
1673 key->bypass_viewport = llvm->draw->identity_viewport;
1674 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1675 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1676 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1677 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1678 key->pad1 = 0;
1679
1680 /* All variants of this shader will have the same value for
1681 * nr_samplers. Not yet trying to compact away holes in the
1682 * sampler array.
1683 */
1684 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1685 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1686 key->nr_sampler_views =
1687 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1688 }
1689 else {
1690 key->nr_sampler_views = key->nr_samplers;
1691 }
1692
1693 draw_sampler = draw_llvm_variant_key_samplers(key);
1694
1695 memcpy(key->vertex_element,
1696 llvm->draw->pt.vertex_element,
1697 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1698
1699 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1700
1701 for (i = 0 ; i < key->nr_samplers; i++) {
1702 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1703 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1704 }
1705 for (i = 0 ; i < key->nr_sampler_views; i++) {
1706 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1707 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1708 }
1709
1710 return key;
1711 }
1712
1713
1714 void
1715 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1716 {
1717 unsigned i;
1718 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1719
1720 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1721 debug_printf("clip_xy = %u\n", key->clip_xy);
1722 debug_printf("clip_z = %u\n", key->clip_z);
1723 debug_printf("clip_user = %u\n", key->clip_user);
1724 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1725 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1726 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1727 debug_printf("has_gs = %u\n", key->has_gs);
1728 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1729
1730 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1731 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1732 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1733 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1734 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1735 }
1736
1737 for (i = 0 ; i < key->nr_sampler_views; i++) {
1738 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1739 }
1740 }
1741
1742
1743 void
1744 draw_llvm_set_mapped_texture(struct draw_context *draw,
1745 unsigned shader_stage,
1746 unsigned sview_idx,
1747 uint32_t width, uint32_t height, uint32_t depth,
1748 uint32_t first_level, uint32_t last_level,
1749 const void *base_ptr,
1750 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1751 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1752 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1753 {
1754 unsigned j;
1755 struct draw_jit_texture *jit_tex;
1756
1757 assert(shader_stage == PIPE_SHADER_VERTEX ||
1758 shader_stage == PIPE_SHADER_GEOMETRY);
1759
1760 if (shader_stage == PIPE_SHADER_VERTEX) {
1761 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1762
1763 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1764 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1765 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1766
1767 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1768 } else {
1769 assert(0);
1770 return;
1771 }
1772
1773 jit_tex->width = width;
1774 jit_tex->height = height;
1775 jit_tex->depth = depth;
1776 jit_tex->first_level = first_level;
1777 jit_tex->last_level = last_level;
1778 jit_tex->base = base_ptr;
1779
1780 for (j = first_level; j <= last_level; j++) {
1781 jit_tex->mip_offsets[j] = mip_offsets[j];
1782 jit_tex->row_stride[j] = row_stride[j];
1783 jit_tex->img_stride[j] = img_stride[j];
1784 }
1785 }
1786
1787
1788 void
1789 draw_llvm_set_sampler_state(struct draw_context *draw,
1790 unsigned shader_type)
1791 {
1792 unsigned i;
1793
1794 if (shader_type == PIPE_SHADER_VERTEX) {
1795 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1796 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1797
1798 if (draw->samplers[i]) {
1799 const struct pipe_sampler_state *s
1800 = draw->samplers[PIPE_SHADER_VERTEX][i];
1801 jit_sam->min_lod = s->min_lod;
1802 jit_sam->max_lod = s->max_lod;
1803 jit_sam->lod_bias = s->lod_bias;
1804 COPY_4V(jit_sam->border_color, s->border_color.f);
1805 }
1806 }
1807 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1808 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1809 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1810
1811 if (draw->samplers[i]) {
1812 const struct pipe_sampler_state *s
1813 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1814 jit_sam->min_lod = s->min_lod;
1815 jit_sam->max_lod = s->max_lod;
1816 jit_sam->lod_bias = s->lod_bias;
1817 COPY_4V(jit_sam->border_color, s->border_color.f);
1818 }
1819 }
1820 }
1821 }
1822
1823
1824 void
1825 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1826 {
1827 struct draw_llvm *llvm = variant->llvm;
1828
1829 if (variant->function_elts) {
1830 gallivm_free_function(variant->gallivm,
1831 variant->function_elts, variant->jit_func_elts);
1832 }
1833
1834 if (variant->function) {
1835 gallivm_free_function(variant->gallivm,
1836 variant->function, variant->jit_func);
1837 }
1838
1839 gallivm_destroy(variant->gallivm);
1840
1841 remove_from_list(&variant->list_item_local);
1842 variant->shader->variants_cached--;
1843 remove_from_list(&variant->list_item_global);
1844 llvm->nr_variants--;
1845 FREE(variant);
1846 }
1847
1848
1849 /**
1850 * Create LLVM types for various structures.
1851 */
1852 static void
1853 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1854 {
1855 struct gallivm_state *gallivm = var->gallivm;
1856 LLVMTypeRef texture_type, sampler_type, context_type;
1857
1858 texture_type = create_jit_texture_type(gallivm, "texture");
1859 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1860
1861 context_type = create_gs_jit_context_type(gallivm,
1862 var->shader->base.vector_length,
1863 texture_type, sampler_type,
1864 "draw_gs_jit_context");
1865 var->context_ptr_type = LLVMPointerType(context_type, 0);
1866
1867 var->input_array_type = create_gs_jit_input_type(gallivm);
1868 }
1869
1870 static LLVMTypeRef
1871 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
1872 {
1873 if (!variant->context_ptr_type)
1874 create_gs_jit_types(variant);
1875 return variant->context_ptr_type;
1876 }
1877
1878 static LLVMValueRef
1879 generate_mask_value(struct draw_gs_llvm_variant *variant,
1880 struct lp_type gs_type)
1881 {
1882 struct gallivm_state *gallivm = variant->gallivm;
1883 LLVMBuilderRef builder = gallivm->builder;
1884 LLVMValueRef bits[16];
1885 struct lp_type mask_type = lp_int_type(gs_type);
1886 struct lp_type mask_elem_type = lp_elem_type(mask_type);
1887 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
1888 mask_type,
1889 0);
1890 unsigned i;
1891
1892 assert(gs_type.length <= Elements(bits));
1893
1894 for (i = gs_type.length; i >= 1; --i) {
1895 int idx = i - 1;
1896 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1897 bits[idx] = lp_build_compare(gallivm,
1898 mask_elem_type, PIPE_FUNC_GEQUAL,
1899 variant->num_prims, ind);
1900 }
1901 for (i = 0; i < gs_type.length; ++i) {
1902 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1903 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
1904 }
1905 mask_val = lp_build_compare(gallivm,
1906 mask_type, PIPE_FUNC_NOTEQUAL,
1907 mask_val,
1908 lp_build_const_int_vec(gallivm, mask_type, 0));
1909
1910 return mask_val;
1911 }
1912
1913 static void
1914 draw_gs_llvm_generate(struct draw_llvm *llvm,
1915 struct draw_gs_llvm_variant *variant)
1916 {
1917 struct gallivm_state *gallivm = variant->gallivm;
1918 LLVMContextRef context = gallivm->context;
1919 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1920 LLVMTypeRef arg_types[6];
1921 LLVMTypeRef func_type;
1922 LLVMValueRef variant_func;
1923 LLVMValueRef context_ptr;
1924 LLVMValueRef prim_id_ptr;
1925 LLVMBasicBlockRef block;
1926 LLVMBuilderRef builder;
1927 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
1928 struct lp_build_sampler_soa *sampler = 0;
1929 struct lp_build_context bld;
1930 struct lp_bld_tgsi_system_values system_values;
1931 struct lp_type gs_type;
1932 unsigned i;
1933 struct draw_gs_llvm_iface gs_iface;
1934 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
1935 LLVMValueRef consts_ptr;
1936 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1937 struct lp_build_mask_context mask;
1938 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1939 unsigned vector_length = variant->shader->base.vector_length;
1940
1941 memset(&system_values, 0, sizeof(system_values));
1942
1943 assert(variant->vertex_header_ptr_type);
1944
1945 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
1946 arg_types[1] = variant->input_array_type; /* input */
1947 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
1948 arg_types[3] = int32_type; /* num_prims */
1949 arg_types[4] = int32_type; /* instance_id */
1950 arg_types[5] = LLVMPointerType(
1951 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
1952
1953 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1954
1955 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
1956 func_type);
1957 variant->function = variant_func;
1958
1959 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1960
1961 for (i = 0; i < Elements(arg_types); ++i)
1962 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1963 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1964 LLVMNoAliasAttribute);
1965
1966 context_ptr = LLVMGetParam(variant_func, 0);
1967 input_array = LLVMGetParam(variant_func, 1);
1968 io_ptr = LLVMGetParam(variant_func, 2);
1969 num_prims = LLVMGetParam(variant_func, 3);
1970 system_values.instance_id = LLVMGetParam(variant_func, 4);
1971 prim_id_ptr = LLVMGetParam(variant_func, 5);
1972
1973 lp_build_name(context_ptr, "context");
1974 lp_build_name(input_array, "input");
1975 lp_build_name(io_ptr, "io");
1976 lp_build_name(io_ptr, "num_prims");
1977 lp_build_name(system_values.instance_id, "instance_id");
1978 lp_build_name(prim_id_ptr, "prim_id_ptr");
1979
1980 variant->context_ptr = context_ptr;
1981 variant->io_ptr = io_ptr;
1982 variant->num_prims = num_prims;
1983
1984 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
1985 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
1986 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
1987 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
1988 gs_iface.input = input_array;
1989 gs_iface.variant = variant;
1990
1991 /*
1992 * Function body
1993 */
1994
1995 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1996 builder = gallivm->builder;
1997 LLVMPositionBuilderAtEnd(builder, block);
1998
1999 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2000
2001 memset(&gs_type, 0, sizeof gs_type);
2002 gs_type.floating = TRUE; /* floating point values */
2003 gs_type.sign = TRUE; /* values are signed */
2004 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2005 gs_type.width = 32; /* 32-bit float */
2006 gs_type.length = vector_length;
2007
2008 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2009
2010 /* code generated texture sampling */
2011 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2012 context_ptr);
2013
2014 mask_val = generate_mask_value(variant, gs_type);
2015 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2016
2017 if (gs_info->uses_primid) {
2018 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
2019 }
2020
2021 lp_build_tgsi_soa(variant->gallivm,
2022 tokens,
2023 gs_type,
2024 &mask,
2025 consts_ptr,
2026 &system_values,
2027 NULL,
2028 outputs,
2029 sampler,
2030 &llvm->draw->gs.geometry_shader->info,
2031 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2032
2033 sampler->destroy(sampler);
2034
2035 lp_build_mask_end(&mask);
2036
2037 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2038
2039 gallivm_verify_function(gallivm, variant_func);
2040 }
2041
2042
2043 struct draw_gs_llvm_variant *
2044 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2045 unsigned num_outputs,
2046 const struct draw_gs_llvm_variant_key *key)
2047 {
2048 struct draw_gs_llvm_variant *variant;
2049 struct llvm_geometry_shader *shader =
2050 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2051 LLVMTypeRef vertex_header;
2052
2053 variant = MALLOC(sizeof *variant +
2054 shader->variant_key_size -
2055 sizeof variant->key);
2056 if (variant == NULL)
2057 return NULL;
2058
2059 variant->llvm = llvm;
2060 variant->shader = shader;
2061
2062 variant->gallivm = gallivm_create();
2063
2064 create_gs_jit_types(variant);
2065
2066 memcpy(&variant->key, key, shader->variant_key_size);
2067
2068 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2069
2070 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2071
2072 draw_gs_llvm_generate(llvm, variant);
2073
2074 gallivm_compile_module(variant->gallivm);
2075
2076 variant->jit_func = (draw_gs_jit_func)
2077 gallivm_jit_function(variant->gallivm, variant->function);
2078
2079 variant->list_item_global.base = variant;
2080 variant->list_item_local.base = variant;
2081 /*variant->no = */shader->variants_created++;
2082 variant->list_item_global.base = variant;
2083
2084 return variant;
2085 }
2086
2087 void
2088 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2089 {
2090 struct draw_llvm *llvm = variant->llvm;
2091
2092 if (variant->function) {
2093 gallivm_free_function(variant->gallivm,
2094 variant->function, variant->jit_func);
2095 }
2096
2097 gallivm_destroy(variant->gallivm);
2098
2099 remove_from_list(&variant->list_item_local);
2100 variant->shader->variants_cached--;
2101 remove_from_list(&variant->list_item_global);
2102 llvm->nr_gs_variants--;
2103 FREE(variant);
2104 }
2105
2106 struct draw_gs_llvm_variant_key *
2107 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2108 {
2109 unsigned i;
2110 struct draw_gs_llvm_variant_key *key;
2111 struct draw_sampler_static_state *draw_sampler;
2112
2113 key = (struct draw_gs_llvm_variant_key *)store;
2114
2115 /* All variants of this shader will have the same value for
2116 * nr_samplers. Not yet trying to compact away holes in the
2117 * sampler array.
2118 */
2119 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2120 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2121 key->nr_sampler_views =
2122 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2123 }
2124 else {
2125 key->nr_sampler_views = key->nr_samplers;
2126 }
2127
2128 draw_sampler = key->samplers;
2129
2130 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2131
2132 for (i = 0 ; i < key->nr_samplers; i++) {
2133 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2134 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2135 }
2136 for (i = 0 ; i < key->nr_sampler_views; i++) {
2137 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2138 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2139 }
2140
2141 return key;
2142 }
2143
2144 void
2145 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2146 {
2147 unsigned i;
2148 struct draw_sampler_static_state *sampler = key->samplers;
2149
2150 for (i = 0 ; i < key->nr_sampler_views; i++) {
2151 debug_printf("sampler[%i].src_format = %s\n", i,
2152 util_format_name(sampler[i].texture_state.format));
2153 }
2154 }