projects / mesa.git / blob
? search:


da19af7bda94c07b0dfe0dc22883969d07afeb20
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_bitarit.h"
37 #include "gallivm/lp_bld_gather.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_const.h"
40 #include "gallivm/lp_bld_swizzle.h"
41 #include "gallivm/lp_bld_struct.h"
42 #include "gallivm/lp_bld_type.h"
43 #include "gallivm/lp_bld_flow.h"
44 #include "gallivm/lp_bld_debug.h"
45 #include "gallivm/lp_bld_tgsi.h"
46 #include "gallivm/lp_bld_printf.h"
47 #include "gallivm/lp_bld_intr.h"
48 #include "gallivm/lp_bld_init.h"
49 #include "gallivm/lp_bld_type.h"
50 #include "gallivm/lp_bld_pack.h"
51 #include "gallivm/lp_bld_format.h"
52
53 #include "tgsi/tgsi_exec.h"
54 #include "tgsi/tgsi_dump.h"
55
56 #include "util/u_math.h"
57 #include "util/u_pointer.h"
58 #include "util/u_string.h"
59 #include "util/simple_list.h"
60
61
62 #define DEBUG_STORE 0
63
64
65 static void
66 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
67
68
69 struct draw_gs_llvm_iface {
70 struct lp_build_tgsi_gs_iface base;
71
72 struct draw_gs_llvm_variant *variant;
73 LLVMValueRef input;
74 };
75
76 static inline const struct draw_gs_llvm_iface *
77 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
78 {
79 return (const struct draw_gs_llvm_iface *)iface;
80 }
81
82 /**
83 * Create LLVM type for draw_vertex_buffer.
84 */
85 static LLVMTypeRef
86 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
87 const char *struct_name)
88 {
89 LLVMTargetDataRef target = gallivm->target;
90 LLVMTypeRef dvbuffer_type;
91 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
92 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
93
94 elem_types[DRAW_JIT_DVBUFFER_MAP] =
95 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
96 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
97
98 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
99 ARRAY_SIZE(elem_types), 0);
100
101 (void) target; /* silence unused var warning for non-debug build */
102 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
103 target, dvbuffer_type,
104 DRAW_JIT_DVBUFFER_MAP);
105 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
106 target, dvbuffer_type,
107 DRAW_JIT_DVBUFFER_SIZE);
108
109 return dvbuffer_type;
110 }
111
112 /**
113 * Create LLVM type for struct draw_jit_texture
114 */
115 static LLVMTypeRef
116 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
117 {
118 LLVMTargetDataRef target = gallivm->target;
119 LLVMTypeRef texture_type;
120 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
121 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
122
123 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
124 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
125 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
126 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
127 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
128 elem_types[DRAW_JIT_TEXTURE_BASE] =
129 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
130 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
131 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
132 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
133 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
134
135 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
136 ARRAY_SIZE(elem_types), 0);
137
138 (void) target; /* silence unused var warning for non-debug build */
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
140 target, texture_type,
141 DRAW_JIT_TEXTURE_WIDTH);
142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
143 target, texture_type,
144 DRAW_JIT_TEXTURE_HEIGHT);
145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
146 target, texture_type,
147 DRAW_JIT_TEXTURE_DEPTH);
148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
149 target, texture_type,
150 DRAW_JIT_TEXTURE_FIRST_LEVEL);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
152 target, texture_type,
153 DRAW_JIT_TEXTURE_LAST_LEVEL);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_BASE);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_ROW_STRIDE);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_IMG_STRIDE);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
164 target, texture_type,
165 DRAW_JIT_TEXTURE_MIP_OFFSETS);
166
167 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
168
169 return texture_type;
170 }
171
172
173 /**
174 * Create LLVM type for struct draw_jit_sampler
175 */
176 static LLVMTypeRef
177 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
178 {
179 LLVMTargetDataRef target = gallivm->target;
180 LLVMTypeRef sampler_type;
181 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
182
183 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
184 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
185 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
186 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
187 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
188
189 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
190 ARRAY_SIZE(elem_types), 0);
191
192 (void) target; /* silence unused var warning for non-debug build */
193 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
194 target, sampler_type,
195 DRAW_JIT_SAMPLER_MIN_LOD);
196 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
197 target, sampler_type,
198 DRAW_JIT_SAMPLER_MAX_LOD);
199 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
200 target, sampler_type,
201 DRAW_JIT_SAMPLER_LOD_BIAS);
202 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
203 target, sampler_type,
204 DRAW_JIT_SAMPLER_BORDER_COLOR);
205
206 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
207
208 return sampler_type;
209 }
210
211
212 /**
213 * Create LLVM type for struct draw_jit_context
214 */
215 static LLVMTypeRef
216 create_jit_context_type(struct gallivm_state *gallivm,
217 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
218 const char *struct_name)
219 {
220 LLVMTargetDataRef target = gallivm->target;
221 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
222 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
223 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
224 LLVMTypeRef context_type;
225
226 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
227 LP_MAX_TGSI_CONST_BUFFERS);
228 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
229 LP_MAX_TGSI_CONST_BUFFERS);
230 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
231 DRAW_TOTAL_CLIP_PLANES), 0);
232 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
233 elem_types[4] = LLVMArrayType(texture_type,
234 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
235 elem_types[5] = LLVMArrayType(sampler_type,
236 PIPE_MAX_SAMPLERS); /* samplers */
237 elem_types[6] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
238 LP_MAX_TGSI_SHADER_BUFFERS);
239 elem_types[7] = LLVMArrayType(int_type, /* num_vs_ssbos */
240 LP_MAX_TGSI_SHADER_BUFFERS);
241 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
242 ARRAY_SIZE(elem_types), 0);
243
244 (void) target; /* silence unused var warning for non-debug build */
245 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
246 target, context_type, DRAW_JIT_CTX_CONSTANTS);
247 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
248 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
249 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
250 target, context_type, DRAW_JIT_CTX_PLANES);
251 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
252 target, context_type, DRAW_JIT_CTX_VIEWPORT);
253 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
254 target, context_type,
255 DRAW_JIT_CTX_TEXTURES);
256 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
257 target, context_type,
258 DRAW_JIT_CTX_SAMPLERS);
259 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
260 target, context_type, DRAW_JIT_CTX_SSBOS);
261 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
262 target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
263 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
264 target, context_type);
265
266 return context_type;
267 }
268
269
270 /**
271 * Create LLVM type for struct draw_gs_jit_context
272 */
273 static LLVMTypeRef
274 create_gs_jit_context_type(struct gallivm_state *gallivm,
275 unsigned vector_length,
276 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
277 const char *struct_name)
278 {
279 LLVMTargetDataRef target = gallivm->target;
280 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
281 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
282 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
283 LLVMTypeRef context_type;
284
285 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
286 LP_MAX_TGSI_CONST_BUFFERS);
287 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
288 LP_MAX_TGSI_CONST_BUFFERS);
289 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
290 DRAW_TOTAL_CLIP_PLANES), 0);
291 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
292
293 elem_types[4] = LLVMArrayType(texture_type,
294 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
295 elem_types[5] = LLVMArrayType(sampler_type,
296 PIPE_MAX_SAMPLERS); /* samplers */
297
298 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
299 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
300 vector_length), 0);
301 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
302 vector_length), 0);
303
304 elem_types[9] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
305 LP_MAX_TGSI_SHADER_BUFFERS);
306 elem_types[10] = LLVMArrayType(int_type, /* num_ssbos */
307 LP_MAX_TGSI_SHADER_BUFFERS);
308 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
309 ARRAY_SIZE(elem_types), 0);
310
311 (void) target; /* silence unused var warning for non-debug build */
312 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
313 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
314 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
315 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
316 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
317 target, context_type, DRAW_GS_JIT_CTX_PLANES);
318 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
319 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
320 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
321 target, context_type,
322 DRAW_GS_JIT_CTX_TEXTURES);
323 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
324 target, context_type,
325 DRAW_GS_JIT_CTX_SAMPLERS);
326 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
327 target, context_type,
328 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
329 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
330 target, context_type,
331 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
332 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
333 target, context_type,
334 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
335 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
336 target, context_type, DRAW_GS_JIT_CTX_SSBOS);
337 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
338 target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
339 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
340 target, context_type);
341
342 return context_type;
343 }
344
345
346 static LLVMTypeRef
347 create_gs_jit_input_type(struct gallivm_state *gallivm)
348 {
349 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
350 LLVMTypeRef input_array;
351
352 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
353 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
354 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
355 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
356
357 return input_array;
358 }
359
360 /**
361 * Create LLVM type for struct pipe_vertex_buffer
362 */
363 static LLVMTypeRef
364 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
365 const char *struct_name)
366 {
367 LLVMTargetDataRef target = gallivm->target;
368 LLVMTypeRef elem_types[4];
369 LLVMTypeRef vb_type;
370
371 elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
372 elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
373 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
374 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
375
376 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
377 ARRAY_SIZE(elem_types), 0);
378
379 (void) target; /* silence unused var warning for non-debug build */
380 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
381 target, vb_type, 0);
382 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
383 target, vb_type, 1);
384 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
385 target, vb_type, 2);
386 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
387 target, vb_type, 3);
388
389 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
390
391 return vb_type;
392 }
393
394
395 /**
396 * Create LLVM type for struct vertex_header;
397 */
398 static LLVMTypeRef
399 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
400 {
401 LLVMTargetDataRef target = gallivm->target;
402 LLVMTypeRef elem_types[3];
403 LLVMTypeRef vertex_header;
404 char struct_name[24];
405
406 snprintf(struct_name, 23, "vertex_header%d", data_elems);
407
408 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
409 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
410 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
411
412 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
413 ARRAY_SIZE(elem_types), 0);
414
415 /* these are bit-fields and we can't take address of them
416 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
417 target, vertex_header,
418 DRAW_JIT_VERTEX_CLIPMASK);
419 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
420 target, vertex_header,
421 DRAW_JIT_VERTEX_EDGEFLAG);
422 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
423 target, vertex_header,
424 DRAW_JIT_VERTEX_PAD);
425 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
426 target, vertex_header,
427 DRAW_JIT_VERTEX_VERTEX_ID);
428 */
429 (void) target; /* silence unused var warning for non-debug build */
430 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
431 target, vertex_header,
432 DRAW_JIT_VERTEX_CLIP_POS);
433 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
434 target, vertex_header,
435 DRAW_JIT_VERTEX_DATA);
436
437 assert(LLVMABISizeOfType(target, vertex_header) ==
438 offsetof(struct vertex_header, data[data_elems]));
439
440 return vertex_header;
441 }
442
443
444 /**
445 * Create LLVM types for various structures.
446 */
447 static void
448 create_jit_types(struct draw_llvm_variant *variant)
449 {
450 struct gallivm_state *gallivm = variant->gallivm;
451 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
452 vb_type;
453
454 texture_type = create_jit_texture_type(gallivm, "texture");
455 sampler_type = create_jit_sampler_type(gallivm, "sampler");
456
457 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
458 "draw_jit_context");
459 variant->context_ptr_type = LLVMPointerType(context_type, 0);
460
461 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
462 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
463
464 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
465 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
466 }
467
468
469 static LLVMTypeRef
470 get_context_ptr_type(struct draw_llvm_variant *variant)
471 {
472 if (!variant->context_ptr_type)
473 create_jit_types(variant);
474 return variant->context_ptr_type;
475 }
476
477
478 static LLVMTypeRef
479 get_buffer_ptr_type(struct draw_llvm_variant *variant)
480 {
481 if (!variant->buffer_ptr_type)
482 create_jit_types(variant);
483 return variant->buffer_ptr_type;
484 }
485
486
487 static LLVMTypeRef
488 get_vb_ptr_type(struct draw_llvm_variant *variant)
489 {
490 if (!variant->vb_ptr_type)
491 create_jit_types(variant);
492 return variant->vb_ptr_type;
493 }
494
495 static LLVMTypeRef
496 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
497 {
498 if (!variant->vertex_header_ptr_type)
499 create_jit_types(variant);
500 return variant->vertex_header_ptr_type;
501 }
502
503
504 /**
505 * Create per-context LLVM info.
506 */
507 struct draw_llvm *
508 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
509 {
510 struct draw_llvm *llvm;
511
512 if (!lp_build_init())
513 return NULL;
514
515 llvm = CALLOC_STRUCT( draw_llvm );
516 if (!llvm)
517 return NULL;
518
519 llvm->draw = draw;
520
521 llvm->context = context;
522 if (!llvm->context) {
523 llvm->context = LLVMContextCreate();
524 llvm->context_owned = true;
525 }
526 if (!llvm->context)
527 goto fail;
528
529 llvm->nr_variants = 0;
530 make_empty_list(&llvm->vs_variants_list);
531
532 llvm->nr_gs_variants = 0;
533 make_empty_list(&llvm->gs_variants_list);
534
535 return llvm;
536
537 fail:
538 draw_llvm_destroy(llvm);
539 return NULL;
540 }
541
542
543 /**
544 * Free per-context LLVM info.
545 */
546 void
547 draw_llvm_destroy(struct draw_llvm *llvm)
548 {
549 if (llvm->context_owned)
550 LLVMContextDispose(llvm->context);
551 llvm->context = NULL;
552
553 /* XXX free other draw_llvm data? */
554 FREE(llvm);
555 }
556
557
558 /**
559 * Create LLVM-generated code for a vertex shader.
560 */
561 struct draw_llvm_variant *
562 draw_llvm_create_variant(struct draw_llvm *llvm,
563 unsigned num_inputs,
564 const struct draw_llvm_variant_key *key)
565 {
566 struct draw_llvm_variant *variant;
567 struct llvm_vertex_shader *shader =
568 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
569 LLVMTypeRef vertex_header;
570 char module_name[64];
571
572 variant = MALLOC(sizeof *variant +
573 shader->variant_key_size -
574 sizeof variant->key);
575 if (!variant)
576 return NULL;
577
578 variant->llvm = llvm;
579 variant->shader = shader;
580
581 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
582 variant->shader->variants_cached);
583
584 variant->gallivm = gallivm_create(module_name, llvm->context);
585
586 create_jit_types(variant);
587
588 memcpy(&variant->key, key, shader->variant_key_size);
589
590 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
591 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
592 draw_llvm_dump_variant_key(&variant->key);
593 }
594
595 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
596
597 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
598
599 draw_llvm_generate(llvm, variant);
600
601 gallivm_compile_module(variant->gallivm);
602
603 variant->jit_func = (draw_jit_vert_func)
604 gallivm_jit_function(variant->gallivm, variant->function);
605
606 gallivm_free_ir(variant->gallivm);
607
608 variant->list_item_global.base = variant;
609 variant->list_item_local.base = variant;
610 /*variant->no = */shader->variants_created++;
611 variant->list_item_global.base = variant;
612
613 return variant;
614 }
615
616
617 static void
618 generate_vs(struct draw_llvm_variant *variant,
619 LLVMBuilderRef builder,
620 struct lp_type vs_type,
621 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
622 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
623 const struct lp_bld_tgsi_system_values *system_values,
624 LLVMValueRef context_ptr,
625 const struct lp_build_sampler_soa *draw_sampler,
626 boolean clamp_vertex_color,
627 struct lp_build_mask_context *bld_mask)
628 {
629 struct draw_llvm *llvm = variant->llvm;
630 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
631 LLVMValueRef consts_ptr =
632 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
633 LLVMValueRef num_consts_ptr =
634 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
635 LLVMValueRef ssbos_ptr =
636 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
637 LLVMValueRef num_ssbos_ptr =
638 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
639
640 lp_build_tgsi_soa(variant->gallivm,
641 tokens,
642 vs_type,
643 bld_mask, /*struct lp_build_mask_context *mask*/
644 consts_ptr,
645 num_consts_ptr,
646 system_values,
647 inputs,
648 outputs,
649 context_ptr,
650 NULL,
651 draw_sampler,
652 &llvm->draw->vs.vertex_shader->info,
653 NULL, ssbos_ptr, num_ssbos_ptr);
654
655 {
656 LLVMValueRef out;
657 unsigned chan, attrib;
658 struct lp_build_context bld;
659 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
660 lp_build_context_init(&bld, variant->gallivm, vs_type);
661
662 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
663 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
664 if (outputs[attrib][chan]) {
665 switch (info->output_semantic_name[attrib]) {
666 case TGSI_SEMANTIC_COLOR:
667 case TGSI_SEMANTIC_BCOLOR:
668 if (clamp_vertex_color) {
669 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
670 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
671 LLVMBuildStore(builder, out, outputs[attrib][chan]);
672 }
673 break;
674 }
675 }
676 }
677 }
678 }
679 }
680
681
682 static void
683 fetch_instanced(struct gallivm_state *gallivm,
684 const struct util_format_description *format_desc,
685 struct lp_type vs_type,
686 LLVMValueRef vb_stride,
687 LLVMValueRef map_ptr,
688 LLVMValueRef buffer_size_adj,
689 LLVMValueRef *inputs,
690 LLVMValueRef index)
691 {
692 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
693 LLVMTypeRef aosf_t, aosi_t;
694 LLVMValueRef zero = LLVMConstNull(i32_t);
695 LLVMBuilderRef builder = gallivm->builder;
696 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
697 unsigned i;
698
699 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
700 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
701
702 /* This mul can overflow. Wraparound is ok. */
703 stride = LLVMBuildMul(builder, vb_stride, index, "");
704
705 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
706 stride, buffer_size_adj,
707 "buffer_overflowed");
708
709 if (0) {
710 lp_build_print_value(gallivm, " instance index = ", index);
711 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
712 }
713
714 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
715 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
716 stride = LLVMBuildAnd(builder, stride, index_valid, "");
717
718 aos = lp_build_fetch_rgba_aos(gallivm,
719 format_desc,
720 lp_float32_vec4_type(),
721 FALSE,
722 map_ptr,
723 stride, zero, zero,
724 NULL);
725
726 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
727 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
728 aos = LLVMBuildAnd(builder, aos, index_valid, "");
729 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
730
731 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
732 LLVMValueRef index = lp_build_const_int32(gallivm, i);
733 inputs[i] = lp_build_extract_broadcast(gallivm,
734 lp_float32_vec4_type(),
735 vs_type, aos, index);
736 }
737 }
738
739
740 static void
741 fetch_vector(struct gallivm_state *gallivm,
742 const struct util_format_description *format_desc,
743 struct lp_type vs_type,
744 LLVMValueRef vb_stride,
745 LLVMValueRef map_ptr,
746 LLVMValueRef buffer_size_adj,
747 LLVMValueRef *inputs,
748 LLVMValueRef indices)
749 {
750 LLVMBuilderRef builder = gallivm->builder;
751 struct lp_build_context blduivec;
752 struct lp_type fetch_type = vs_type;
753 LLVMValueRef offset, valid_mask;
754 unsigned i;
755
756 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
757
758 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
759 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
760
761 /* This mul can overflow. Wraparound is ok. */
762 offset = lp_build_mul(&blduivec, vb_stride, indices);
763
764 valid_mask = lp_build_compare(gallivm, blduivec.type,
765 PIPE_FUNC_LESS, offset, buffer_size_adj);
766
767 /* not valid elements use offset 0 */
768 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
769
770 if (0) {
771 lp_build_print_value(gallivm, " indices = ", indices);
772 lp_build_print_value(gallivm, " offsets = ", offset);
773 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
774 }
775
776 /*
777 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
778 * This should always produce better code.
779 */
780
781 /* The type handling is annoying here... */
782 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
783 format_desc->channel[0].pure_integer) {
784 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
785 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
786 }
787 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
788 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
789 }
790 }
791
792 lp_build_fetch_rgba_soa(gallivm, format_desc,
793 fetch_type, FALSE, map_ptr, offset,
794 blduivec.zero, blduivec.zero,
795 NULL, inputs);
796
797 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
798 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
799 lp_build_vec_type(gallivm, vs_type), "");
800 }
801
802 /* out-of-bound fetches return all zeros */
803 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
804 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
805 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
806 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
807 lp_build_vec_type(gallivm, vs_type), "");
808 }
809 }
810
811
812 static void
813 store_aos(struct gallivm_state *gallivm,
814 LLVMValueRef io_ptr,
815 LLVMValueRef index,
816 LLVMValueRef value)
817 {
818 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
819 LLVMBuilderRef builder = gallivm->builder;
820 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
821 LLVMValueRef indices[3];
822
823 indices[0] = lp_build_const_int32(gallivm, 0);
824 indices[1] = index;
825 indices[2] = lp_build_const_int32(gallivm, 0);
826
827 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
828 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
829
830 #if DEBUG_STORE
831 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
832 #endif
833
834 /* Unaligned store due to the vertex header */
835 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
836 }
837
838 /**
839 * Adjust the mask to architecture endianess. The mask will the store in struct:
840 *
841 * struct vertex_header {
842 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
843 * unsigned edgeflag:1;
844 * unsigned pad:1;
845 * unsigned vertex_id:16;
846 * [...]
847 * }
848 *
849 * On little-endian machine nothing needs to done, however on bit-endian machine
850 * the mask's fields need to be adjusted with the algorithm:
851 *
852 * uint32_t reverse (uint32_t x)
853 * {
854 * return (x >> 16) | // vertex_id
855 * ((x & 0x3fff) << 18) | // clipmask
856 * ((x & 0x4000) << 3) | // pad
857 * ((x & 0x8000) << 1); // edgeflag
858 * }
859 */
860 static LLVMValueRef
861 adjust_mask(struct gallivm_state *gallivm,
862 LLVMValueRef mask)
863 {
864 #ifdef PIPE_ARCH_BIG_ENDIAN
865 LLVMBuilderRef builder = gallivm->builder;
866 LLVMValueRef vertex_id;
867 LLVMValueRef clipmask;
868 LLVMValueRef pad;
869 LLVMValueRef edgeflag;
870
871 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
872 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
873 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
874 if (0) {
875 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
876 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
877 }
878 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
879 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
880
881 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
882 if (0) {
883 mask = LLVMBuildOr(builder, mask, pad, "");
884 }
885 mask = LLVMBuildOr(builder, mask, edgeflag, "");
886 #endif
887 return mask;
888 }
889
890 static void
891 store_aos_array(struct gallivm_state *gallivm,
892 struct lp_type soa_type,
893 LLVMValueRef io_ptr,
894 LLVMValueRef *indices,
895 LLVMValueRef* aos,
896 int attrib,
897 int num_outputs,
898 LLVMValueRef clipmask,
899 boolean need_edgeflag)
900 {
901 LLVMBuilderRef builder = gallivm->builder;
902 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
903 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
904 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
905 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
906 int vector_length = soa_type.length;
907 int i;
908
909 debug_assert(TGSI_NUM_CHANNELS == 4);
910
911 for (i = 0; i < vector_length; i++) {
912 linear_inds[i] = lp_build_const_int32(gallivm, i);
913 if (indices) {
914 inds[i] = indices[i];
915 } else {
916 inds[i] = linear_inds[i];
917 }
918 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
919 }
920
921 if (attrib == 0) {
922 /* store vertex header for each of the n vertices */
923 LLVMValueRef val, cliptmp;
924 int vertex_id_pad_edgeflag;
925
926 /* If this assertion fails, it means we need to update the bit twidding
927 * code here. See struct vertex_header in draw_private.h.
928 */
929 assert(DRAW_TOTAL_CLIP_PLANES==14);
930 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
931 if (!need_edgeflag) {
932 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
933 }
934 else {
935 vertex_id_pad_edgeflag = (0xffff << 16);
936 }
937 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
938 vertex_id_pad_edgeflag);
939 /* OR with the clipmask */
940 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
941 for (i = 0; i < vector_length; i++) {
942 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
943 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
944 val = adjust_mask(gallivm, val);
945 #if DEBUG_STORE
946 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
947 io_ptrs[i], inds[i], val);
948 #endif
949 LLVMBuildStore(builder, val, id_ptr);
950 }
951 }
952
953 /* store for each of the n vertices */
954 for (i = 0; i < vector_length; i++) {
955 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
956 }
957 }
958
959
960 static void
961 convert_to_aos(struct gallivm_state *gallivm,
962 LLVMValueRef io,
963 LLVMValueRef *indices,
964 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
965 LLVMValueRef clipmask,
966 int num_outputs,
967 struct lp_type soa_type,
968 boolean need_edgeflag)
969 {
970 LLVMBuilderRef builder = gallivm->builder;
971 unsigned chan, attrib, i;
972
973 #if DEBUG_STORE
974 lp_build_printf(gallivm, " # storing begin\n");
975 #endif
976 for (attrib = 0; attrib < num_outputs; ++attrib) {
977 LLVMValueRef soa[TGSI_NUM_CHANNELS];
978 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
979 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
980 if (outputs[attrib][chan]) {
981 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
982 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
983 #if DEBUG_STORE
984 lp_build_printf(gallivm, "output %d : %d ",
985 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
986 attrib, 0),
987 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
988 chan, 0));
989 lp_build_print_value(gallivm, "val = ", out);
990 {
991 LLVMValueRef iv =
992 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
993
994 lp_build_print_value(gallivm, " ival = ", iv);
995 }
996 #endif
997 soa[chan] = out;
998 }
999 else {
1000 soa[chan] = 0;
1001 }
1002 }
1003
1004
1005 if (soa_type.length == TGSI_NUM_CHANNELS) {
1006 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1007 } else {
1008 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1009
1010 for (i = 0; i < soa_type.length; ++i) {
1011 aos[i] = lp_build_extract_range(gallivm,
1012 soa[i % TGSI_NUM_CHANNELS],
1013 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1014 TGSI_NUM_CHANNELS);
1015 }
1016 }
1017
1018 store_aos_array(gallivm,
1019 soa_type,
1020 io, indices,
1021 aos,
1022 attrib,
1023 num_outputs,
1024 clipmask,
1025 need_edgeflag);
1026 }
1027 #if DEBUG_STORE
1028 lp_build_printf(gallivm, " # storing end\n");
1029 #endif
1030 }
1031
1032
1033 /**
1034 * Stores original vertex positions in clip coordinates
1035 */
1036 static void
1037 store_clip(struct gallivm_state *gallivm,
1038 const struct lp_type vs_type,
1039 LLVMValueRef io_ptr,
1040 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1041 int idx)
1042 {
1043 LLVMBuilderRef builder = gallivm->builder;
1044 LLVMValueRef soa[4];
1045 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1046 LLVMValueRef indices[2];
1047 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1048 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1049 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1050 LLVMTypeRef clip_ptr_type =
1051 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1052 4), 0);
1053 int i, j;
1054
1055 indices[0] =
1056 indices[1] = lp_build_const_int32(gallivm, 0);
1057
1058 for (i = 0; i < vs_type.length; i++) {
1059 inds[i] = lp_build_const_int32(gallivm, i);
1060 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1061 }
1062
1063 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1064 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1065 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1066 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1067
1068 for (i = 0; i < vs_type.length; i++) {
1069 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1070 }
1071
1072 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1073 for (i = 0; i < vs_type.length; ++i) {
1074 aos[i] = lp_build_extract_range(gallivm,
1075 soa[i % TGSI_NUM_CHANNELS],
1076 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1077 TGSI_NUM_CHANNELS);
1078 }
1079
1080 for (j = 0; j < vs_type.length; j++) {
1081 LLVMValueRef clip_ptr;
1082
1083 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1084 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1085
1086 /* Unaligned store */
1087 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1088 }
1089 }
1090
1091
1092 /**
1093 * Transforms the outputs for viewport mapping
1094 */
1095 static void
1096 generate_viewport(struct draw_llvm_variant *variant,
1097 LLVMBuilderRef builder,
1098 struct lp_type vs_type,
1099 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1100 LLVMValueRef context_ptr)
1101 {
1102 int i;
1103 struct gallivm_state *gallivm = variant->gallivm;
1104 struct lp_type f32_type = vs_type;
1105 const unsigned pos = variant->llvm->draw->vs.position_output;
1106 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1107 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1108 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1109 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1110
1111 /* We treat pipe_viewport_state as a float array */
1112 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1113 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1114
1115 /* for 1/w convention*/
1116 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1117 LLVMBuildStore(builder, out3, outputs[pos][3]);
1118
1119 /* Viewport Mapping */
1120 for (i=0; i<3; i++) {
1121 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1122 LLVMValueRef scale;
1123 LLVMValueRef trans;
1124 LLVMValueRef scale_i;
1125 LLVMValueRef trans_i;
1126 LLVMValueRef index;
1127
1128 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1129 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1130
1131 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1132 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1133
1134 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1135 LLVMBuildLoad(builder, scale_i, "scale"));
1136 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1137 LLVMBuildLoad(builder, trans_i, "trans"));
1138
1139 /* divide by w */
1140 out = LLVMBuildFMul(builder, out, out3, "");
1141 /* mult by scale, add translation */
1142 out = lp_build_fmuladd(builder, out, scale, trans);
1143
1144 /* store transformed outputs */
1145 LLVMBuildStore(builder, out, outputs[pos][i]);
1146 }
1147
1148 }
1149
1150
1151 /**
1152 * Returns clipmask as nxi32 bitmask for the n vertices
1153 */
1154 static LLVMValueRef
1155 generate_clipmask(struct draw_llvm *llvm,
1156 struct gallivm_state *gallivm,
1157 struct lp_type vs_type,
1158 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1159 struct draw_llvm_variant_key *key,
1160 LLVMValueRef context_ptr,
1161 boolean *have_clipdist)
1162 {
1163 LLVMBuilderRef builder = gallivm->builder;
1164 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1165 LLVMValueRef test, temp;
1166 LLVMValueRef zero, shift;
1167 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1168 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1169 LLVMValueRef plane1, planes, plane_ptr, sum;
1170 struct lp_type f32_type = vs_type;
1171 struct lp_type i32_type = lp_int_type(vs_type);
1172 const unsigned pos = llvm->draw->vs.position_output;
1173 const unsigned cv = llvm->draw->vs.clipvertex_output;
1174 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1175 boolean have_cd = false;
1176 boolean clip_user = key->clip_user;
1177 unsigned ucp_enable = key->ucp_enable;
1178 unsigned cd[2];
1179
1180 cd[0] = llvm->draw->vs.ccdistance_output[0];
1181 cd[1] = llvm->draw->vs.ccdistance_output[1];
1182
1183 if (cd[0] != pos || cd[1] != pos)
1184 have_cd = true;
1185
1186 if (num_written_clipdistance && !clip_user) {
1187 clip_user = true;
1188 ucp_enable = (1 << num_written_clipdistance) - 1;
1189 }
1190
1191 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1192 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1193 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1194 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1195
1196 /*
1197 * load clipvertex and position from correct locations.
1198 * if they are the same just load them once.
1199 */
1200 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1201 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1202 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1203 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1204
1205 if (clip_user && cv != pos) {
1206 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1207 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1208 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1209 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1210 } else {
1211 cv_x = pos_x;
1212 cv_y = pos_y;
1213 cv_z = pos_z;
1214 cv_w = pos_w;
1215 }
1216
1217 /*
1218 * Be careful with the comparisons and NaNs (using llvm's unordered
1219 * comparisons here).
1220 */
1221 /* Cliptest, for hardwired planes */
1222 /*
1223 * XXX should take guardband into account (currently not in key).
1224 * Otherwise might run the draw pipeline stages for nothing.
1225 */
1226 if (key->clip_xy) {
1227 /* plane 1 */
1228 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1229 temp = shift;
1230 test = LLVMBuildAnd(builder, test, temp, "");
1231 mask = test;
1232
1233 /* plane 2 */
1234 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1235 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1236 temp = LLVMBuildShl(builder, temp, shift, "");
1237 test = LLVMBuildAnd(builder, test, temp, "");
1238 mask = LLVMBuildOr(builder, mask, test, "");
1239
1240 /* plane 3 */
1241 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1242 temp = LLVMBuildShl(builder, temp, shift, "");
1243 test = LLVMBuildAnd(builder, test, temp, "");
1244 mask = LLVMBuildOr(builder, mask, test, "");
1245
1246 /* plane 4 */
1247 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1248 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1249 temp = LLVMBuildShl(builder, temp, shift, "");
1250 test = LLVMBuildAnd(builder, test, temp, "");
1251 mask = LLVMBuildOr(builder, mask, test, "");
1252 }
1253
1254 if (key->clip_z) {
1255 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1256 if (key->clip_halfz) {
1257 /* plane 5 */
1258 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1259 test = LLVMBuildAnd(builder, test, temp, "");
1260 mask = LLVMBuildOr(builder, mask, test, "");
1261 }
1262 else {
1263 /* plane 5 */
1264 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1265 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1266 test = LLVMBuildAnd(builder, test, temp, "");
1267 mask = LLVMBuildOr(builder, mask, test, "");
1268 }
1269 /* plane 6 */
1270 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1271 temp = LLVMBuildShl(builder, temp, shift, "");
1272 test = LLVMBuildAnd(builder, test, temp, "");
1273 mask = LLVMBuildOr(builder, mask, test, "");
1274 }
1275
1276 if (clip_user) {
1277 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1278 LLVMValueRef indices[3];
1279 LLVMValueRef is_nan_or_inf;
1280
1281 /* userclip planes */
1282 while (ucp_enable) {
1283 unsigned plane_idx = ffs(ucp_enable)-1;
1284 ucp_enable &= ~(1 << plane_idx);
1285 plane_idx += 6;
1286
1287 if (have_cd && num_written_clipdistance) {
1288 LLVMValueRef clipdist;
1289 int i;
1290 i = plane_idx - 6;
1291
1292 *have_clipdist = TRUE;
1293 if (i < 4) {
1294 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1295 } else {
1296 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1297 }
1298 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1299 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1300 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1301 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1302 test = LLVMBuildAnd(builder, test, temp, "");
1303 mask = LLVMBuildOr(builder, mask, test, "");
1304 } else {
1305 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1306 indices[0] = lp_build_const_int32(gallivm, 0);
1307 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1308
1309 indices[2] = lp_build_const_int32(gallivm, 0);
1310 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1311 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1312 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1313 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1314
1315 indices[2] = lp_build_const_int32(gallivm, 1);
1316 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1317 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1318 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1319 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1320
1321 indices[2] = lp_build_const_int32(gallivm, 2);
1322 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1323 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1324 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1325 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1326
1327 indices[2] = lp_build_const_int32(gallivm, 3);
1328 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1329 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1330 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1331 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1332
1333 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1334 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1335 test = LLVMBuildAnd(builder, test, temp, "");
1336 mask = LLVMBuildOr(builder, mask, test, "");
1337 }
1338 }
1339 }
1340 if (key->need_edgeflags) {
1341 /*
1342 * This isn't really part of clipmask but stored the same in vertex
1343 * header later, so do it here.
1344 */
1345 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1346 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1347 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1348 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1349 temp = lp_build_const_int_vec(gallivm, i32_type,
1350 1LL << DRAW_TOTAL_CLIP_PLANES);
1351 test = LLVMBuildAnd(builder, test, temp, "");
1352 mask = LLVMBuildOr(builder, mask, test, "");
1353 }
1354 return mask;
1355 }
1356
1357
1358 /**
1359 * Returns boolean if any clipping has occurred
1360 * Used zero/one i8 value to represent boolean
1361 */
1362 static LLVMValueRef
1363 clipmask_booli8(struct gallivm_state *gallivm,
1364 const struct lp_type vs_type,
1365 LLVMValueRef clipmask_bool_ptr,
1366 boolean edgeflag_in_clipmask)
1367 {
1368 LLVMBuilderRef builder = gallivm->builder;
1369 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1370 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1371 LLVMValueRef ret;
1372 struct lp_build_context bldivec;
1373
1374 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1375
1376 /*
1377 * We need to invert the edgeflag bit from the clipmask here
1378 * (because the result is really if we want to run the pipeline or not
1379 * and we (may) need it if edgeflag was 0).
1380 */
1381 if (edgeflag_in_clipmask) {
1382 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1383 1LL << DRAW_TOTAL_CLIP_PLANES);
1384 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1385 }
1386
1387 /*
1388 * XXX: probably should mask off bits from the mask which come from
1389 * vertices which were beyond the count (i.e. indices_valid for
1390 * linear fetches, for elts ones we don't have the correct mask
1391 * right now). Otherwise might run the pipeline for nothing,
1392 * though everything should still work.
1393 */
1394 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1395 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1396 return ret;
1397 }
1398
1399 static LLVMValueRef
1400 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1401 struct lp_build_tgsi_context * bld_base,
1402 boolean is_vindex_indirect,
1403 LLVMValueRef vertex_index,
1404 boolean is_aindex_indirect,
1405 LLVMValueRef attrib_index,
1406 LLVMValueRef swizzle_index)
1407 {
1408 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1409 struct gallivm_state *gallivm = bld_base->base.gallivm;
1410 LLVMBuilderRef builder = gallivm->builder;
1411 LLVMValueRef indices[3];
1412 LLVMValueRef res;
1413 struct lp_type type = bld_base->base.type;
1414
1415 if (is_vindex_indirect || is_aindex_indirect) {
1416 int i;
1417 res = bld_base->base.zero;
1418 for (i = 0; i < type.length; ++i) {
1419 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1420 LLVMValueRef vert_chan_index = vertex_index;
1421 LLVMValueRef attr_chan_index = attrib_index;
1422 LLVMValueRef channel_vec, value;
1423
1424 if (is_vindex_indirect) {
1425 vert_chan_index = LLVMBuildExtractElement(builder,
1426 vertex_index, idx, "");
1427 }
1428 if (is_aindex_indirect) {
1429 attr_chan_index = LLVMBuildExtractElement(builder,
1430 attrib_index, idx, "");
1431 }
1432
1433 indices[0] = vert_chan_index;
1434 indices[1] = attr_chan_index;
1435 indices[2] = swizzle_index;
1436
1437 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1438 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1439 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1440
1441 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1442 }
1443 } else {
1444 indices[0] = vertex_index;
1445 indices[1] = attrib_index;
1446 indices[2] = swizzle_index;
1447
1448 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1449 res = LLVMBuildLoad(builder, res, "");
1450 }
1451
1452 return res;
1453 }
1454
1455 static void
1456 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1457 struct lp_build_tgsi_context * bld_base,
1458 LLVMValueRef (*outputs)[4],
1459 LLVMValueRef emitted_vertices_vec)
1460 {
1461 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1462 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1463 struct gallivm_state *gallivm = variant->gallivm;
1464 LLVMBuilderRef builder = gallivm->builder;
1465 struct lp_type gs_type = bld_base->base.type;
1466 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1467 lp_int_type(gs_type), 0);
1468 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1469 LLVMValueRef next_prim_offset =
1470 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1471 LLVMValueRef io = variant->io_ptr;
1472 unsigned i;
1473 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1474
1475 for (i = 0; i < gs_type.length; ++i) {
1476 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1477 LLVMValueRef currently_emitted =
1478 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1479 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1480 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1481 }
1482
1483 convert_to_aos(gallivm, io, indices,
1484 outputs, clipmask,
1485 gs_info->num_outputs, gs_type,
1486 FALSE);
1487 }
1488
1489 static void
1490 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1491 struct lp_build_tgsi_context * bld_base,
1492 LLVMValueRef verts_per_prim_vec,
1493 LLVMValueRef emitted_prims_vec)
1494 {
1495 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1496 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1497 struct gallivm_state *gallivm = variant->gallivm;
1498 LLVMBuilderRef builder = gallivm->builder;
1499 LLVMValueRef prim_lengts_ptr =
1500 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1501 unsigned i;
1502
1503 for (i = 0; i < bld_base->base.type.length; ++i) {
1504 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1505 LLVMValueRef prims_emitted =
1506 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1507 LLVMValueRef store_ptr;
1508 LLVMValueRef num_vertices =
1509 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1510
1511 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1512 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1513 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1514 LLVMBuildStore(builder, num_vertices, store_ptr);
1515 }
1516 }
1517
1518 static void
1519 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1520 struct lp_build_tgsi_context * bld_base,
1521 LLVMValueRef total_emitted_vertices_vec,
1522 LLVMValueRef emitted_prims_vec)
1523 {
1524 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1525 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1526 struct gallivm_state *gallivm = variant->gallivm;
1527 LLVMBuilderRef builder = gallivm->builder;
1528 LLVMValueRef emitted_verts_ptr =
1529 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1530 LLVMValueRef emitted_prims_ptr =
1531 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1532 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1533
1534 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1535 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1536
1537 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1538 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1539 }
1540
1541 static void
1542 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1543 {
1544 struct gallivm_state *gallivm = variant->gallivm;
1545 LLVMContextRef context = gallivm->context;
1546 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1547 LLVMTypeRef arg_types[11];
1548 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1549 LLVMTypeRef func_type;
1550 LLVMValueRef context_ptr;
1551 LLVMBasicBlockRef block;
1552 LLVMBuilderRef builder;
1553 char func_name[64];
1554 struct lp_type vs_type;
1555 LLVMValueRef count, fetch_elts, start_or_maxelt;
1556 LLVMValueRef vertex_id_offset, start_instance;
1557 LLVMValueRef stride, step, io_itr;
1558 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1559 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1560 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1561 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1562 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1563 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1564 LLVMValueRef fake_buf_ptr, fake_buf;
1565
1566 struct draw_context *draw = llvm->draw;
1567 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1568 unsigned i, j;
1569 struct lp_build_context bld, blduivec;
1570 struct lp_build_loop_state lp_loop;
1571 struct lp_build_if_state if_ctx;
1572 const int vector_length = lp_native_vector_width / 32;
1573 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1574 struct lp_build_sampler_soa *sampler = 0;
1575 LLVMValueRef ret, clipmask_bool_ptr;
1576 struct draw_llvm_variant_key *key = &variant->key;
1577 /* If geometry shader is present we need to skip both the viewport
1578 * transformation and clipping otherwise the inputs to the geometry
1579 * shader will be incorrect.
1580 * The code can't handle vp transform when vs writes vp index neither
1581 * (though this would be fixable here, but couldn't just broadcast
1582 * the values).
1583 */
1584 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1585 vs_info->writes_viewport_index;
1586 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1587 key->clip_z ||
1588 key->clip_user ||
1589 key->need_edgeflags);
1590 LLVMValueRef variant_func;
1591 const unsigned pos = draw->vs.position_output;
1592 const unsigned cv = draw->vs.clipvertex_output;
1593 boolean have_clipdist = FALSE;
1594 struct lp_bld_tgsi_system_values system_values;
1595
1596 memset(&system_values, 0, sizeof(system_values));
1597
1598 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1599 variant->shader->variants_cached);
1600
1601 i = 0;
1602 arg_types[i++] = get_context_ptr_type(variant); /* context */
1603 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1604 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1605 arg_types[i++] = int32_type; /* count */
1606 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1607 arg_types[i++] = int32_type; /* stride */
1608 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1609 arg_types[i++] = int32_type; /* instance_id */
1610 arg_types[i++] = int32_type; /* vertex_id_offset */
1611 arg_types[i++] = int32_type; /* start_instance */
1612 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1613
1614 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1615 arg_types, num_arg_types, 0);
1616
1617 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1618 variant->function = variant_func;
1619
1620 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1621 for (i = 0; i < num_arg_types; ++i)
1622 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1623 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1624
1625 context_ptr = LLVMGetParam(variant_func, 0);
1626 io_ptr = LLVMGetParam(variant_func, 1);
1627 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1628 count = LLVMGetParam(variant_func, 3);
1629 /*
1630 * XXX: the maxelt part is unused. Not really useful, since we cannot
1631 * get index buffer overflows due to vsplit (which provides its own
1632 * elts buffer, with a different size than what's passed in here).
1633 */
1634 start_or_maxelt = LLVMGetParam(variant_func, 4);
1635 /*
1636 * XXX: stride is actually unused. The stride we use is strictly calculated
1637 * from the number of outputs (including the draw_extra outputs).
1638 * Should probably fix some day (we need a new vs just because of extra
1639 * outputs which the generated vs won't touch).
1640 */
1641 stride = LLVMGetParam(variant_func, 5);
1642 vb_ptr = LLVMGetParam(variant_func, 6);
1643 system_values.instance_id = LLVMGetParam(variant_func, 7);
1644 vertex_id_offset = LLVMGetParam(variant_func, 8);
1645 start_instance = LLVMGetParam(variant_func, 9);
1646 fetch_elts = LLVMGetParam(variant_func, 10);
1647
1648 lp_build_name(context_ptr, "context");
1649 lp_build_name(io_ptr, "io");
1650 lp_build_name(vbuffers_ptr, "vbuffers");
1651 lp_build_name(count, "count");
1652 lp_build_name(start_or_maxelt, "start_or_maxelt");
1653 lp_build_name(stride, "stride");
1654 lp_build_name(vb_ptr, "vb");
1655 lp_build_name(system_values.instance_id, "instance_id");
1656 lp_build_name(vertex_id_offset, "vertex_id_offset");
1657 lp_build_name(start_instance, "start_instance");
1658 lp_build_name(fetch_elts, "fetch_elts");
1659
1660 /*
1661 * Function body
1662 */
1663
1664 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1665 builder = gallivm->builder;
1666 LLVMPositionBuilderAtEnd(builder, block);
1667
1668 memset(&vs_type, 0, sizeof vs_type);
1669 vs_type.floating = TRUE; /* floating point values */
1670 vs_type.sign = TRUE; /* values are signed */
1671 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1672 vs_type.width = 32; /* 32-bit float */
1673 vs_type.length = vector_length;
1674
1675 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1676 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1677
1678 /* hold temporary "bool" clipmask */
1679 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1680
1681 fake_buf = lp_build_alloca_undef(gallivm,
1682 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1683 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1684 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1685 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1686
1687 /* code generated texture sampling */
1688 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1689
1690 step = lp_build_const_int32(gallivm, vector_length);
1691
1692 ind_vec = blduivec.undef;
1693 for (i = 0; i < vs_type.length; i++) {
1694 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1695 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1696 }
1697
1698 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1699 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1700
1701 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1702 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1703 /*
1704 * Only needed for non-indexed path.
1705 */
1706 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1707
1708 /*
1709 * Pre-calculate everything which is constant per shader invocation.
1710 */
1711 for (j = 0; j < key->nr_vertex_elements; ++j) {
1712 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1713 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1714 struct pipe_vertex_element *velem = &key->vertex_element[j];
1715 LLVMValueRef vb_index =
1716 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1717 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1718 util_format_get_blocksize(velem->src_format));
1719 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1720 velem->src_offset);
1721 struct lp_build_if_state if_ctx;
1722
1723 if (velem->src_format != PIPE_FORMAT_NONE) {
1724 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1725 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1726 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1727 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1728 LLVMInt32TypeInContext(context), "");
1729 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1730 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1731 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1732
1733 ofbit = NULL;
1734 /*
1735 * We'll set buffer_size_adj to zero if we have of, so it will
1736 * always overflow later automatically without having to keep ofbit.
1737 * Overflows (with normal wraparound) doing the actual offset
1738 * calculation should be ok, just not for the buffer size calc.
1739 * It would also be possible to detect such overflows and return
1740 * zeros if that happens, but this would be more complex.
1741 */
1742 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1743 tmp = lp_build_sub(&bld, bsize, bld.one);
1744 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1745 &ofbit);
1746 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1747 buf_offset, &ofbit);
1748
1749 /*
1750 * We can't easily set fake vertex buffers outside the generated code.
1751 * Hence, set fake vertex buffers here instead basically, so fetch
1752 * code can always fetch using offset 0, eliminating all control flow
1753 * inside the main loop.
1754 * (Alternatively, could have control flow per vector skipping fetch
1755 * if ofbit is true.)
1756 */
1757 if (velem->instance_divisor) {
1758 /*
1759 * Index is equal to the start instance plus the number of current
1760 * instance divided by the divisor. In this case we compute it as:
1761 * index = start_instance + (instance_id / divisor).
1762 * Note we could actually do the fetch here, outside the loop -
1763 * it's all constant, hopefully llvm recognizes this.
1764 */
1765 LLVMValueRef current_instance;
1766 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1767 lp_build_const_int32(gallivm,
1768 velem->instance_divisor),
1769 "instance_divisor");
1770 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1771 current_instance, &ofbit);
1772 }
1773
1774 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1775 buffer_size_adj[j], "");
1776
1777 temp_ptr = lp_build_alloca_undef(gallivm,
1778 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1779
1780 lp_build_if(&if_ctx, gallivm, ofbit);
1781 {
1782 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1783 }
1784 lp_build_else(&if_ctx);
1785 {
1786 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1787 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1788 }
1789 lp_build_endif(&if_ctx);
1790 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1791
1792 if (0) {
1793 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1794 lp_build_const_int32(gallivm, j),
1795 vb_index, vb_stride[j]);
1796 lp_build_printf(gallivm,
1797 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1798 vb_buffer_offset, src_offset, buf_offset);
1799 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1800 buffer_size, bsize);
1801 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1802 }
1803 }
1804 }
1805
1806 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1807 {
1808 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1809 LLVMValueRef io;
1810 LLVMValueRef clipmask; /* holds the clipmask value */
1811 LLVMValueRef true_index_array, index_store;
1812 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1813
1814 io_itr = lp_loop.counter;
1815
1816 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1817 #if DEBUG_STORE
1818 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1819 io_itr, io, lp_loop.counter);
1820 #endif
1821
1822 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1823 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1824
1825 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
1826 /*
1827 * Limit indices to fetch_max, otherwise might try to access indices
1828 * beyond index buffer (or rather vsplit elt buffer) size.
1829 * Could probably safely (?) skip this for non-indexed draws and
1830 * simplify things minimally (by removing it could combine the ind_vec
1831 * and start_vec adds). I think the only effect for non-indexed draws will
1832 * be that for the invalid elements they will be all fetched from the
1833 * same location as the last valid one, but noone should really care.
1834 */
1835 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1836
1837 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1838
1839 lp_build_if(&if_ctx, gallivm, have_elts);
1840 {
1841 /*
1842 * Note: you'd expect some comparison/clamp against fetch_elt_max
1843 * here.
1844 * There used to be one here but it was incorrect: overflow was
1845 * detected if index > fetch_elt_max - but the correct condition
1846 * would be index >= fetch_elt_max (since this is just size of elts
1847 * buffer / element size).
1848 * Using the correct condition however will cause failures - due to
1849 * vsplit/vcache code which rebases indices. So, as an example, if
1850 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1851 * replace all invalid indices with 0 - which in case of elt_bias
1852 * not being zero will get a different fetch index than the valid
1853 * index 0. So, just rely on vsplit code preventing out-of-bounds
1854 * fetches. This is also why it's safe to do elts fetch even if there
1855 * was no index buffer bound - the real buffer is never seen here, at
1856 * least not if there are index buffer overflows...
1857 */
1858
1859 /*
1860 * XXX should not have to do this, as scale can be handled
1861 * natively by loads (hits asserts though).
1862 */
1863 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1864 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1865 LLVMPointerType(LLVMInt8TypeInContext(context),
1866 0), "");
1867 tmp = lp_build_gather(gallivm, vs_type.length,
1868 32, bld.type, TRUE,
1869 fetch_elts, tmp, FALSE);
1870 LLVMBuildStore(builder, tmp, index_store);
1871 }
1872 lp_build_else(&if_ctx);
1873 {
1874 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1875 LLVMBuildStore(builder, tmp, index_store);
1876 }
1877 lp_build_endif(&if_ctx);
1878
1879 true_index_array = LLVMBuildLoad(builder, index_store, "");
1880
1881 for (j = 0; j < key->nr_vertex_elements; ++j) {
1882 struct pipe_vertex_element *velem = &key->vertex_element[j];
1883 const struct util_format_description *format_desc =
1884 util_format_description(velem->src_format);
1885
1886 if (format_desc->format == PIPE_FORMAT_NONE) {
1887 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1888 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1889 }
1890 }
1891 else if (velem->instance_divisor) {
1892 fetch_instanced(gallivm, format_desc, vs_type,
1893 vb_stride[j], map_ptr[j],
1894 buffer_size_adj[j],
1895 inputs[j], instance_index[j]);
1896 }
1897 else {
1898 fetch_vector(gallivm, format_desc, vs_type,
1899 vb_stride[j], map_ptr[j],
1900 buffer_size_adj[j],
1901 inputs[j], true_index_array);
1902 }
1903 }
1904
1905 struct lp_build_mask_context mask;
1906
1907 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
1908 /* In the paths with elts vertex id has to be unaffected by the
1909 * index bias and because indices inside our elements array have
1910 * already had index bias applied we need to subtract it here to
1911 * get back to the original index.
1912 * in the linear paths vertex id has to be unaffected by the
1913 * original start index and because we abuse the 'start' variable
1914 * to either represent the actual start index or the index at which
1915 * the primitive was split (we split rendering into chunks of at
1916 * most 4095-vertices) we need to back out the original start
1917 * index out of our vertex id here.
1918 */
1919 system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
1920 vertex_id_offset);
1921 system_values.vertex_id = true_index_array;
1922 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1923 system_values.basevertex, "");
1924
1925 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1926 generate_vs(variant,
1927 builder,
1928 vs_type,
1929 outputs,
1930 ptr_aos,
1931 &system_values,
1932 context_ptr,
1933 sampler,
1934 key->clamp_vertex_color,
1935 &mask);
1936
1937 lp_build_mask_end(&mask);
1938 if (pos != -1 && cv != -1) {
1939 /* store original positions in clip before further manipulation */
1940 store_clip(gallivm, vs_type, io, outputs, pos);
1941
1942 /* do cliptest */
1943 if (enable_cliptest) {
1944 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1945 /* allocate clipmask, assign it integer type */
1946 clipmask = generate_clipmask(llvm,
1947 gallivm,
1948 vs_type,
1949 outputs,
1950 key,
1951 context_ptr, &have_clipdist);
1952 temp = LLVMBuildOr(builder, clipmask, temp, "");
1953 /* store temporary clipping boolean value */
1954 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1955 }
1956 else {
1957 clipmask = blduivec.zero;
1958 }
1959
1960 /* do viewport mapping */
1961 if (!bypass_viewport) {
1962 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1963 }
1964 }
1965 else {
1966 clipmask = blduivec.zero;
1967 }
1968
1969 /* store clipmask in vertex header,
1970 * original positions in clip
1971 * and transformed positions in data
1972 */
1973 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1974 vs_info->num_outputs, vs_type,
1975 enable_cliptest && key->need_edgeflags);
1976 }
1977 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1978
1979 sampler->destroy(sampler);
1980
1981 /* return clipping boolean value for function */
1982 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
1983 enable_cliptest && key->need_edgeflags);
1984
1985 LLVMBuildRet(builder, ret);
1986
1987 gallivm_verify_function(gallivm, variant_func);
1988 }
1989
1990
1991 struct draw_llvm_variant_key *
1992 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1993 {
1994 unsigned i;
1995 struct draw_llvm_variant_key *key;
1996 struct draw_sampler_static_state *draw_sampler;
1997
1998 key = (struct draw_llvm_variant_key *)store;
1999
2000 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2001
2002 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2003
2004 /* will have to rig this up properly later */
2005 key->clip_xy = llvm->draw->clip_xy;
2006 key->clip_z = llvm->draw->clip_z;
2007 key->clip_user = llvm->draw->clip_user;
2008 key->bypass_viewport = llvm->draw->bypass_viewport;
2009 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2010 /* XXX assumes edgeflag output not at 0 */
2011 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2012 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2013 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2014 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2015
2016 /* All variants of this shader will have the same value for
2017 * nr_samplers. Not yet trying to compact away holes in the
2018 * sampler array.
2019 */
2020 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2021 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2022 key->nr_sampler_views =
2023 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2024 }
2025 else {
2026 key->nr_sampler_views = key->nr_samplers;
2027 }
2028
2029 /* Presumably all variants of the shader should have the same
2030 * number of vertex elements - ie the number of shader inputs.
2031 * NOTE: we NEED to store the needed number of needed inputs
2032 * here, not the number of provided elements to match keysize
2033 * (and the offset of sampler state in the key).
2034 * If we have excess number of vertex elements, this is valid,
2035 * but the excess ones don't matter.
2036 * If we don't have enough vertex elements (which looks not really
2037 * valid but we'll handle it gracefully) fill out missing ones with
2038 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2039 */
2040 key->nr_vertex_elements =
2041 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2042
2043 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2044 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2045 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2046 memset(key->vertex_element, 0,
2047 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2048 }
2049 memcpy(key->vertex_element,
2050 llvm->draw->pt.vertex_element,
2051 sizeof(struct pipe_vertex_element) *
2052 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2053
2054 draw_sampler = draw_llvm_variant_key_samplers(key);
2055 memset(draw_sampler, 0,
2056 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2057
2058 for (i = 0 ; i < key->nr_samplers; i++) {
2059 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2060 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2061 }
2062 for (i = 0 ; i < key->nr_sampler_views; i++) {
2063 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2064 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2065 }
2066
2067 return key;
2068 }
2069
2070
2071 void
2072 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2073 {
2074 unsigned i;
2075 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2076
2077 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2078 debug_printf("clip_xy = %u\n", key->clip_xy);
2079 debug_printf("clip_z = %u\n", key->clip_z);
2080 debug_printf("clip_user = %u\n", key->clip_user);
2081 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2082 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2083 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2084 debug_printf("has_gs = %u\n", key->has_gs);
2085 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2086
2087 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2088 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2089 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2090 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2091 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2092 }
2093
2094 for (i = 0 ; i < key->nr_sampler_views; i++) {
2095 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2096 }
2097 }
2098
2099
2100 void
2101 draw_llvm_set_mapped_texture(struct draw_context *draw,
2102 enum pipe_shader_type shader_stage,
2103 unsigned sview_idx,
2104 uint32_t width, uint32_t height, uint32_t depth,
2105 uint32_t first_level, uint32_t last_level,
2106 const void *base_ptr,
2107 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2108 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2109 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2110 {
2111 unsigned j;
2112 struct draw_jit_texture *jit_tex;
2113
2114 assert(shader_stage == PIPE_SHADER_VERTEX ||
2115 shader_stage == PIPE_SHADER_GEOMETRY);
2116
2117 if (shader_stage == PIPE_SHADER_VERTEX) {
2118 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2119
2120 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2121 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2122 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2123
2124 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2125 } else {
2126 assert(0);
2127 return;
2128 }
2129
2130 jit_tex->width = width;
2131 jit_tex->height = height;
2132 jit_tex->depth = depth;
2133 jit_tex->first_level = first_level;
2134 jit_tex->last_level = last_level;
2135 jit_tex->base = base_ptr;
2136
2137 for (j = first_level; j <= last_level; j++) {
2138 jit_tex->mip_offsets[j] = mip_offsets[j];
2139 jit_tex->row_stride[j] = row_stride[j];
2140 jit_tex->img_stride[j] = img_stride[j];
2141 }
2142 }
2143
2144
2145 void
2146 draw_llvm_set_sampler_state(struct draw_context *draw,
2147 enum pipe_shader_type shader_type)
2148 {
2149 unsigned i;
2150
2151 if (shader_type == PIPE_SHADER_VERTEX) {
2152 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2153 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2154
2155 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2156 const struct pipe_sampler_state *s
2157 = draw->samplers[PIPE_SHADER_VERTEX][i];
2158 jit_sam->min_lod = s->min_lod;
2159 jit_sam->max_lod = s->max_lod;
2160 jit_sam->lod_bias = s->lod_bias;
2161 COPY_4V(jit_sam->border_color, s->border_color.f);
2162 }
2163 }
2164 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2165 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2166 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2167
2168 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2169 const struct pipe_sampler_state *s
2170 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2171 jit_sam->min_lod = s->min_lod;
2172 jit_sam->max_lod = s->max_lod;
2173 jit_sam->lod_bias = s->lod_bias;
2174 COPY_4V(jit_sam->border_color, s->border_color.f);
2175 }
2176 }
2177 }
2178 }
2179
2180
2181 void
2182 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2183 {
2184 struct draw_llvm *llvm = variant->llvm;
2185
2186 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2187 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2188 variant->shader->variants_cached, llvm->nr_variants);
2189 }
2190
2191 gallivm_destroy(variant->gallivm);
2192
2193 remove_from_list(&variant->list_item_local);
2194 variant->shader->variants_cached--;
2195 remove_from_list(&variant->list_item_global);
2196 llvm->nr_variants--;
2197 FREE(variant);
2198 }
2199
2200
2201 /**
2202 * Create LLVM types for various structures.
2203 */
2204 static void
2205 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2206 {
2207 struct gallivm_state *gallivm = var->gallivm;
2208 LLVMTypeRef texture_type, sampler_type, context_type;
2209
2210 texture_type = create_jit_texture_type(gallivm, "texture");
2211 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2212
2213 context_type = create_gs_jit_context_type(gallivm,
2214 var->shader->base.vector_length,
2215 texture_type, sampler_type,
2216 "draw_gs_jit_context");
2217 var->context_ptr_type = LLVMPointerType(context_type, 0);
2218
2219 var->input_array_type = create_gs_jit_input_type(gallivm);
2220 }
2221
2222 static LLVMTypeRef
2223 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2224 {
2225 if (!variant->context_ptr_type)
2226 create_gs_jit_types(variant);
2227 return variant->context_ptr_type;
2228 }
2229
2230 static LLVMValueRef
2231 generate_mask_value(struct draw_gs_llvm_variant *variant,
2232 struct lp_type gs_type)
2233 {
2234 struct gallivm_state *gallivm = variant->gallivm;
2235 LLVMBuilderRef builder = gallivm->builder;
2236 struct lp_type mask_type = lp_int_type(gs_type);
2237 LLVMValueRef num_prims;
2238 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2239 unsigned i;
2240
2241 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2242 variant->num_prims);
2243 for (i = 0; i < gs_type.length; i++) {
2244 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2245 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2246 }
2247 mask_val = lp_build_compare(gallivm, mask_type,
2248 PIPE_FUNC_GREATER, num_prims, mask_val);
2249
2250 return mask_val;
2251 }
2252
2253 static void
2254 draw_gs_llvm_generate(struct draw_llvm *llvm,
2255 struct draw_gs_llvm_variant *variant)
2256 {
2257 struct gallivm_state *gallivm = variant->gallivm;
2258 LLVMContextRef context = gallivm->context;
2259 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2260 LLVMTypeRef arg_types[7];
2261 LLVMTypeRef func_type;
2262 LLVMValueRef variant_func;
2263 LLVMValueRef context_ptr;
2264 LLVMValueRef prim_id_ptr;
2265 LLVMBasicBlockRef block;
2266 LLVMBuilderRef builder;
2267 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2268 struct lp_build_sampler_soa *sampler = 0;
2269 struct lp_build_context bld;
2270 struct lp_bld_tgsi_system_values system_values;
2271 char func_name[64];
2272 struct lp_type gs_type;
2273 unsigned i;
2274 struct draw_gs_llvm_iface gs_iface;
2275 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2276 LLVMValueRef consts_ptr, num_consts_ptr;
2277 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2278 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2279 struct lp_build_mask_context mask;
2280 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2281 unsigned vector_length = variant->shader->base.vector_length;
2282
2283 memset(&system_values, 0, sizeof(system_values));
2284
2285 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2286 variant->shader->variants_cached);
2287
2288 assert(variant->vertex_header_ptr_type);
2289
2290 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2291 arg_types[1] = variant->input_array_type; /* input */
2292 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2293 arg_types[3] = int32_type; /* num_prims */
2294 arg_types[4] = int32_type; /* instance_id */
2295 arg_types[5] = LLVMPointerType(
2296 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2297 arg_types[6] = int32_type;
2298
2299 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2300
2301 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2302
2303 variant->function = variant_func;
2304
2305 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2306
2307 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2308 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2309 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2310
2311 context_ptr = LLVMGetParam(variant_func, 0);
2312 input_array = LLVMGetParam(variant_func, 1);
2313 io_ptr = LLVMGetParam(variant_func, 2);
2314 num_prims = LLVMGetParam(variant_func, 3);
2315 system_values.instance_id = LLVMGetParam(variant_func, 4);
2316 prim_id_ptr = LLVMGetParam(variant_func, 5);
2317 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2318
2319 lp_build_name(context_ptr, "context");
2320 lp_build_name(input_array, "input");
2321 lp_build_name(io_ptr, "io");
2322 lp_build_name(num_prims, "num_prims");
2323 lp_build_name(system_values.instance_id, "instance_id");
2324 lp_build_name(prim_id_ptr, "prim_id_ptr");
2325 lp_build_name(system_values.invocation_id, "invocation_id");
2326
2327 variant->context_ptr = context_ptr;
2328 variant->io_ptr = io_ptr;
2329 variant->num_prims = num_prims;
2330
2331 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2332 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2333 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2334 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2335 gs_iface.input = input_array;
2336 gs_iface.variant = variant;
2337
2338 /*
2339 * Function body
2340 */
2341
2342 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2343 builder = gallivm->builder;
2344 LLVMPositionBuilderAtEnd(builder, block);
2345
2346 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2347
2348 memset(&gs_type, 0, sizeof gs_type);
2349 gs_type.floating = TRUE; /* floating point values */
2350 gs_type.sign = TRUE; /* values are signed */
2351 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2352 gs_type.width = 32; /* 32-bit float */
2353 gs_type.length = vector_length;
2354
2355 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2356 num_consts_ptr =
2357 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2358
2359 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2360 num_ssbos_ptr =
2361 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2362
2363 /* code generated texture sampling */
2364 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2365
2366 mask_val = generate_mask_value(variant, gs_type);
2367 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2368
2369 if (gs_info->uses_primid) {
2370 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2371 }
2372
2373 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2374 tgsi_dump(tokens, 0);
2375 draw_gs_llvm_dump_variant_key(&variant->key);
2376 }
2377
2378 lp_build_tgsi_soa(variant->gallivm,
2379 tokens,
2380 gs_type,
2381 &mask,
2382 consts_ptr,
2383 num_consts_ptr,
2384 &system_values,
2385 NULL,
2386 outputs,
2387 context_ptr,
2388 NULL,
2389 sampler,
2390 &llvm->draw->gs.geometry_shader->info,
2391 (const struct lp_build_tgsi_gs_iface *)&gs_iface, ssbos_ptr, num_ssbos_ptr);
2392
2393 sampler->destroy(sampler);
2394
2395 lp_build_mask_end(&mask);
2396
2397 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2398
2399 gallivm_verify_function(gallivm, variant_func);
2400 }
2401
2402
2403 struct draw_gs_llvm_variant *
2404 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2405 unsigned num_outputs,
2406 const struct draw_gs_llvm_variant_key *key)
2407 {
2408 struct draw_gs_llvm_variant *variant;
2409 struct llvm_geometry_shader *shader =
2410 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2411 LLVMTypeRef vertex_header;
2412 char module_name[64];
2413
2414 variant = MALLOC(sizeof *variant +
2415 shader->variant_key_size -
2416 sizeof variant->key);
2417 if (!variant)
2418 return NULL;
2419
2420 variant->llvm = llvm;
2421 variant->shader = shader;
2422
2423 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2424 variant->shader->variants_cached);
2425
2426 variant->gallivm = gallivm_create(module_name, llvm->context);
2427
2428 create_gs_jit_types(variant);
2429
2430 memcpy(&variant->key, key, shader->variant_key_size);
2431
2432 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2433
2434 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2435
2436 draw_gs_llvm_generate(llvm, variant);
2437
2438 gallivm_compile_module(variant->gallivm);
2439
2440 variant->jit_func = (draw_gs_jit_func)
2441 gallivm_jit_function(variant->gallivm, variant->function);
2442
2443 gallivm_free_ir(variant->gallivm);
2444
2445 variant->list_item_global.base = variant;
2446 variant->list_item_local.base = variant;
2447 /*variant->no = */shader->variants_created++;
2448 variant->list_item_global.base = variant;
2449
2450 return variant;
2451 }
2452
2453 void
2454 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2455 {
2456 struct draw_llvm *llvm = variant->llvm;
2457
2458 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2459 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2460 variant->shader->variants_cached, llvm->nr_gs_variants);
2461 }
2462
2463 gallivm_destroy(variant->gallivm);
2464
2465 remove_from_list(&variant->list_item_local);
2466 variant->shader->variants_cached--;
2467 remove_from_list(&variant->list_item_global);
2468 llvm->nr_gs_variants--;
2469 FREE(variant);
2470 }
2471
2472 struct draw_gs_llvm_variant_key *
2473 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2474 {
2475 unsigned i;
2476 struct draw_gs_llvm_variant_key *key;
2477 struct draw_sampler_static_state *draw_sampler;
2478
2479 key = (struct draw_gs_llvm_variant_key *)store;
2480
2481 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2482
2483 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2484
2485 /* All variants of this shader will have the same value for
2486 * nr_samplers. Not yet trying to compact away holes in the
2487 * sampler array.
2488 */
2489 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2490 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2491 key->nr_sampler_views =
2492 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2493 }
2494 else {
2495 key->nr_sampler_views = key->nr_samplers;
2496 }
2497
2498 draw_sampler = key->samplers;
2499
2500 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2501
2502 for (i = 0 ; i < key->nr_samplers; i++) {
2503 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2504 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2505 }
2506 for (i = 0 ; i < key->nr_sampler_views; i++) {
2507 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2508 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2509 }
2510
2511 return key;
2512 }
2513
2514 void
2515 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2516 {
2517 unsigned i;
2518 struct draw_sampler_static_state *sampler = key->samplers;
2519
2520 for (i = 0 ; i < key->nr_sampler_views; i++) {
2521 debug_printf("sampler[%i].src_format = %s\n", i,
2522 util_format_name(sampler[i].texture_state.format));
2523 }
2524 }