projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: rework lp_build_tgsi_soa to take a struct
[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 struct lp_build_tgsi_params params = {};
641
642 params.type = vs_type;
643 params.mask = bld_mask;
644 params.consts_ptr = consts_ptr;
645 params.const_sizes_ptr = num_consts_ptr;
646 params.system_values = system_values;
647 params.inputs = inputs;
648 params.context_ptr = context_ptr;
649 params.sampler = draw_sampler;
650 params.info = &llvm->draw->vs.vertex_shader->info;
651 params.ssbo_ptr = ssbos_ptr;
652 params.ssbo_sizes_ptr = num_ssbos_ptr;
653
654 lp_build_tgsi_soa(variant->gallivm,
655 tokens,
656 &params,
657 outputs);
658
659 {
660 LLVMValueRef out;
661 unsigned chan, attrib;
662 struct lp_build_context bld;
663 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
664 lp_build_context_init(&bld, variant->gallivm, vs_type);
665
666 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
667 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
668 if (outputs[attrib][chan]) {
669 switch (info->output_semantic_name[attrib]) {
670 case TGSI_SEMANTIC_COLOR:
671 case TGSI_SEMANTIC_BCOLOR:
672 if (clamp_vertex_color) {
673 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
674 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
675 LLVMBuildStore(builder, out, outputs[attrib][chan]);
676 }
677 break;
678 }
679 }
680 }
681 }
682 }
683 }
684
685
686 static void
687 fetch_instanced(struct gallivm_state *gallivm,
688 const struct util_format_description *format_desc,
689 struct lp_type vs_type,
690 LLVMValueRef vb_stride,
691 LLVMValueRef map_ptr,
692 LLVMValueRef buffer_size_adj,
693 LLVMValueRef *inputs,
694 LLVMValueRef index)
695 {
696 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
697 LLVMTypeRef aosf_t, aosi_t;
698 LLVMValueRef zero = LLVMConstNull(i32_t);
699 LLVMBuilderRef builder = gallivm->builder;
700 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
701 unsigned i;
702
703 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
704 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
705
706 /* This mul can overflow. Wraparound is ok. */
707 stride = LLVMBuildMul(builder, vb_stride, index, "");
708
709 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
710 stride, buffer_size_adj,
711 "buffer_overflowed");
712
713 if (0) {
714 lp_build_print_value(gallivm, " instance index = ", index);
715 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
716 }
717
718 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
719 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
720 stride = LLVMBuildAnd(builder, stride, index_valid, "");
721
722 aos = lp_build_fetch_rgba_aos(gallivm,
723 format_desc,
724 lp_float32_vec4_type(),
725 FALSE,
726 map_ptr,
727 stride, zero, zero,
728 NULL);
729
730 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
731 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
732 aos = LLVMBuildAnd(builder, aos, index_valid, "");
733 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
734
735 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
736 LLVMValueRef index = lp_build_const_int32(gallivm, i);
737 inputs[i] = lp_build_extract_broadcast(gallivm,
738 lp_float32_vec4_type(),
739 vs_type, aos, index);
740 }
741 }
742
743
744 static void
745 fetch_vector(struct gallivm_state *gallivm,
746 const struct util_format_description *format_desc,
747 struct lp_type vs_type,
748 LLVMValueRef vb_stride,
749 LLVMValueRef map_ptr,
750 LLVMValueRef buffer_size_adj,
751 LLVMValueRef *inputs,
752 LLVMValueRef indices)
753 {
754 LLVMBuilderRef builder = gallivm->builder;
755 struct lp_build_context blduivec;
756 struct lp_type fetch_type = vs_type;
757 LLVMValueRef offset, valid_mask;
758 unsigned i;
759
760 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
761
762 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
763 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
764
765 /* This mul can overflow. Wraparound is ok. */
766 offset = lp_build_mul(&blduivec, vb_stride, indices);
767
768 valid_mask = lp_build_compare(gallivm, blduivec.type,
769 PIPE_FUNC_LESS, offset, buffer_size_adj);
770
771 /* not valid elements use offset 0 */
772 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
773
774 if (0) {
775 lp_build_print_value(gallivm, " indices = ", indices);
776 lp_build_print_value(gallivm, " offsets = ", offset);
777 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
778 }
779
780 /*
781 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
782 * This should always produce better code.
783 */
784
785 /* The type handling is annoying here... */
786 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
787 format_desc->channel[0].pure_integer) {
788 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
789 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
790 }
791 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
792 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
793 }
794 }
795
796 lp_build_fetch_rgba_soa(gallivm, format_desc,
797 fetch_type, FALSE, map_ptr, offset,
798 blduivec.zero, blduivec.zero,
799 NULL, inputs);
800
801 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
802 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
803 lp_build_vec_type(gallivm, vs_type), "");
804 }
805
806 /* out-of-bound fetches return all zeros */
807 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
808 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
809 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
810 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
811 lp_build_vec_type(gallivm, vs_type), "");
812 }
813 }
814
815
816 static void
817 store_aos(struct gallivm_state *gallivm,
818 LLVMValueRef io_ptr,
819 LLVMValueRef index,
820 LLVMValueRef value)
821 {
822 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
823 LLVMBuilderRef builder = gallivm->builder;
824 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
825 LLVMValueRef indices[3];
826
827 indices[0] = lp_build_const_int32(gallivm, 0);
828 indices[1] = index;
829 indices[2] = lp_build_const_int32(gallivm, 0);
830
831 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
832 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
833
834 #if DEBUG_STORE
835 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
836 #endif
837
838 /* Unaligned store due to the vertex header */
839 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
840 }
841
842 /**
843 * Adjust the mask to architecture endianess. The mask will the store in struct:
844 *
845 * struct vertex_header {
846 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
847 * unsigned edgeflag:1;
848 * unsigned pad:1;
849 * unsigned vertex_id:16;
850 * [...]
851 * }
852 *
853 * On little-endian machine nothing needs to done, however on bit-endian machine
854 * the mask's fields need to be adjusted with the algorithm:
855 *
856 * uint32_t reverse (uint32_t x)
857 * {
858 * return (x >> 16) | // vertex_id
859 * ((x & 0x3fff) << 18) | // clipmask
860 * ((x & 0x4000) << 3) | // pad
861 * ((x & 0x8000) << 1); // edgeflag
862 * }
863 */
864 static LLVMValueRef
865 adjust_mask(struct gallivm_state *gallivm,
866 LLVMValueRef mask)
867 {
868 #ifdef PIPE_ARCH_BIG_ENDIAN
869 LLVMBuilderRef builder = gallivm->builder;
870 LLVMValueRef vertex_id;
871 LLVMValueRef clipmask;
872 LLVMValueRef pad;
873 LLVMValueRef edgeflag;
874
875 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
876 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
877 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
878 if (0) {
879 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
880 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
881 }
882 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
883 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
884
885 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
886 if (0) {
887 mask = LLVMBuildOr(builder, mask, pad, "");
888 }
889 mask = LLVMBuildOr(builder, mask, edgeflag, "");
890 #endif
891 return mask;
892 }
893
894 static void
895 store_aos_array(struct gallivm_state *gallivm,
896 struct lp_type soa_type,
897 LLVMValueRef io_ptr,
898 LLVMValueRef *indices,
899 LLVMValueRef* aos,
900 int attrib,
901 int num_outputs,
902 LLVMValueRef clipmask,
903 boolean need_edgeflag)
904 {
905 LLVMBuilderRef builder = gallivm->builder;
906 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
907 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
908 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
909 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
910 int vector_length = soa_type.length;
911 int i;
912
913 debug_assert(TGSI_NUM_CHANNELS == 4);
914
915 for (i = 0; i < vector_length; i++) {
916 linear_inds[i] = lp_build_const_int32(gallivm, i);
917 if (indices) {
918 inds[i] = indices[i];
919 } else {
920 inds[i] = linear_inds[i];
921 }
922 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
923 }
924
925 if (attrib == 0) {
926 /* store vertex header for each of the n vertices */
927 LLVMValueRef val, cliptmp;
928 int vertex_id_pad_edgeflag;
929
930 /* If this assertion fails, it means we need to update the bit twidding
931 * code here. See struct vertex_header in draw_private.h.
932 */
933 assert(DRAW_TOTAL_CLIP_PLANES==14);
934 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
935 if (!need_edgeflag) {
936 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
937 }
938 else {
939 vertex_id_pad_edgeflag = (0xffff << 16);
940 }
941 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
942 vertex_id_pad_edgeflag);
943 /* OR with the clipmask */
944 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
945 for (i = 0; i < vector_length; i++) {
946 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
947 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
948 val = adjust_mask(gallivm, val);
949 #if DEBUG_STORE
950 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
951 io_ptrs[i], inds[i], val);
952 #endif
953 LLVMBuildStore(builder, val, id_ptr);
954 }
955 }
956
957 /* store for each of the n vertices */
958 for (i = 0; i < vector_length; i++) {
959 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
960 }
961 }
962
963
964 static void
965 convert_to_aos(struct gallivm_state *gallivm,
966 LLVMValueRef io,
967 LLVMValueRef *indices,
968 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
969 LLVMValueRef clipmask,
970 int num_outputs,
971 struct lp_type soa_type,
972 boolean need_edgeflag)
973 {
974 LLVMBuilderRef builder = gallivm->builder;
975 unsigned chan, attrib, i;
976
977 #if DEBUG_STORE
978 lp_build_printf(gallivm, " # storing begin\n");
979 #endif
980 for (attrib = 0; attrib < num_outputs; ++attrib) {
981 LLVMValueRef soa[TGSI_NUM_CHANNELS];
982 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
983 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
984 if (outputs[attrib][chan]) {
985 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
986 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
987 #if DEBUG_STORE
988 lp_build_printf(gallivm, "output %d : %d ",
989 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
990 attrib, 0),
991 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
992 chan, 0));
993 lp_build_print_value(gallivm, "val = ", out);
994 {
995 LLVMValueRef iv =
996 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
997
998 lp_build_print_value(gallivm, " ival = ", iv);
999 }
1000 #endif
1001 soa[chan] = out;
1002 }
1003 else {
1004 soa[chan] = 0;
1005 }
1006 }
1007
1008
1009 if (soa_type.length == TGSI_NUM_CHANNELS) {
1010 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1011 } else {
1012 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1013
1014 for (i = 0; i < soa_type.length; ++i) {
1015 aos[i] = lp_build_extract_range(gallivm,
1016 soa[i % TGSI_NUM_CHANNELS],
1017 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1018 TGSI_NUM_CHANNELS);
1019 }
1020 }
1021
1022 store_aos_array(gallivm,
1023 soa_type,
1024 io, indices,
1025 aos,
1026 attrib,
1027 num_outputs,
1028 clipmask,
1029 need_edgeflag);
1030 }
1031 #if DEBUG_STORE
1032 lp_build_printf(gallivm, " # storing end\n");
1033 #endif
1034 }
1035
1036
1037 /**
1038 * Stores original vertex positions in clip coordinates
1039 */
1040 static void
1041 store_clip(struct gallivm_state *gallivm,
1042 const struct lp_type vs_type,
1043 LLVMValueRef io_ptr,
1044 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1045 int idx)
1046 {
1047 LLVMBuilderRef builder = gallivm->builder;
1048 LLVMValueRef soa[4];
1049 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1050 LLVMValueRef indices[2];
1051 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1052 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1053 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1054 LLVMTypeRef clip_ptr_type =
1055 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1056 4), 0);
1057 int i, j;
1058
1059 indices[0] =
1060 indices[1] = lp_build_const_int32(gallivm, 0);
1061
1062 for (i = 0; i < vs_type.length; i++) {
1063 inds[i] = lp_build_const_int32(gallivm, i);
1064 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1065 }
1066
1067 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1068 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1069 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1070 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1071
1072 for (i = 0; i < vs_type.length; i++) {
1073 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1074 }
1075
1076 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1077 for (i = 0; i < vs_type.length; ++i) {
1078 aos[i] = lp_build_extract_range(gallivm,
1079 soa[i % TGSI_NUM_CHANNELS],
1080 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1081 TGSI_NUM_CHANNELS);
1082 }
1083
1084 for (j = 0; j < vs_type.length; j++) {
1085 LLVMValueRef clip_ptr;
1086
1087 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1088 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1089
1090 /* Unaligned store */
1091 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1092 }
1093 }
1094
1095
1096 /**
1097 * Transforms the outputs for viewport mapping
1098 */
1099 static void
1100 generate_viewport(struct draw_llvm_variant *variant,
1101 LLVMBuilderRef builder,
1102 struct lp_type vs_type,
1103 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1104 LLVMValueRef context_ptr)
1105 {
1106 int i;
1107 struct gallivm_state *gallivm = variant->gallivm;
1108 struct lp_type f32_type = vs_type;
1109 const unsigned pos = variant->llvm->draw->vs.position_output;
1110 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1111 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1112 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1113 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1114
1115 /* We treat pipe_viewport_state as a float array */
1116 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1117 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1118
1119 /* for 1/w convention*/
1120 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1121 LLVMBuildStore(builder, out3, outputs[pos][3]);
1122
1123 /* Viewport Mapping */
1124 for (i=0; i<3; i++) {
1125 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1126 LLVMValueRef scale;
1127 LLVMValueRef trans;
1128 LLVMValueRef scale_i;
1129 LLVMValueRef trans_i;
1130 LLVMValueRef index;
1131
1132 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1133 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1134
1135 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1136 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1137
1138 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1139 LLVMBuildLoad(builder, scale_i, "scale"));
1140 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1141 LLVMBuildLoad(builder, trans_i, "trans"));
1142
1143 /* divide by w */
1144 out = LLVMBuildFMul(builder, out, out3, "");
1145 /* mult by scale, add translation */
1146 out = lp_build_fmuladd(builder, out, scale, trans);
1147
1148 /* store transformed outputs */
1149 LLVMBuildStore(builder, out, outputs[pos][i]);
1150 }
1151
1152 }
1153
1154
1155 /**
1156 * Returns clipmask as nxi32 bitmask for the n vertices
1157 */
1158 static LLVMValueRef
1159 generate_clipmask(struct draw_llvm *llvm,
1160 struct gallivm_state *gallivm,
1161 struct lp_type vs_type,
1162 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1163 struct draw_llvm_variant_key *key,
1164 LLVMValueRef context_ptr,
1165 boolean *have_clipdist)
1166 {
1167 LLVMBuilderRef builder = gallivm->builder;
1168 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1169 LLVMValueRef test, temp;
1170 LLVMValueRef zero, shift;
1171 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1172 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1173 LLVMValueRef plane1, planes, plane_ptr, sum;
1174 struct lp_type f32_type = vs_type;
1175 struct lp_type i32_type = lp_int_type(vs_type);
1176 const unsigned pos = llvm->draw->vs.position_output;
1177 const unsigned cv = llvm->draw->vs.clipvertex_output;
1178 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1179 boolean have_cd = false;
1180 boolean clip_user = key->clip_user;
1181 unsigned ucp_enable = key->ucp_enable;
1182 unsigned cd[2];
1183
1184 cd[0] = llvm->draw->vs.ccdistance_output[0];
1185 cd[1] = llvm->draw->vs.ccdistance_output[1];
1186
1187 if (cd[0] != pos || cd[1] != pos)
1188 have_cd = true;
1189
1190 if (num_written_clipdistance && !clip_user) {
1191 clip_user = true;
1192 ucp_enable = (1 << num_written_clipdistance) - 1;
1193 }
1194
1195 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1196 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1197 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1198 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1199
1200 /*
1201 * load clipvertex and position from correct locations.
1202 * if they are the same just load them once.
1203 */
1204 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1205 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1206 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1207 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1208
1209 if (clip_user && cv != pos) {
1210 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1211 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1212 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1213 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1214 } else {
1215 cv_x = pos_x;
1216 cv_y = pos_y;
1217 cv_z = pos_z;
1218 cv_w = pos_w;
1219 }
1220
1221 /*
1222 * Be careful with the comparisons and NaNs (using llvm's unordered
1223 * comparisons here).
1224 */
1225 /* Cliptest, for hardwired planes */
1226 /*
1227 * XXX should take guardband into account (currently not in key).
1228 * Otherwise might run the draw pipeline stages for nothing.
1229 */
1230 if (key->clip_xy) {
1231 /* plane 1 */
1232 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1233 temp = shift;
1234 test = LLVMBuildAnd(builder, test, temp, "");
1235 mask = test;
1236
1237 /* plane 2 */
1238 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1239 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1240 temp = LLVMBuildShl(builder, temp, shift, "");
1241 test = LLVMBuildAnd(builder, test, temp, "");
1242 mask = LLVMBuildOr(builder, mask, test, "");
1243
1244 /* plane 3 */
1245 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1246 temp = LLVMBuildShl(builder, temp, shift, "");
1247 test = LLVMBuildAnd(builder, test, temp, "");
1248 mask = LLVMBuildOr(builder, mask, test, "");
1249
1250 /* plane 4 */
1251 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1252 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1253 temp = LLVMBuildShl(builder, temp, shift, "");
1254 test = LLVMBuildAnd(builder, test, temp, "");
1255 mask = LLVMBuildOr(builder, mask, test, "");
1256 }
1257
1258 if (key->clip_z) {
1259 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1260 if (key->clip_halfz) {
1261 /* plane 5 */
1262 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1263 test = LLVMBuildAnd(builder, test, temp, "");
1264 mask = LLVMBuildOr(builder, mask, test, "");
1265 }
1266 else {
1267 /* plane 5 */
1268 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1269 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1270 test = LLVMBuildAnd(builder, test, temp, "");
1271 mask = LLVMBuildOr(builder, mask, test, "");
1272 }
1273 /* plane 6 */
1274 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1275 temp = LLVMBuildShl(builder, temp, shift, "");
1276 test = LLVMBuildAnd(builder, test, temp, "");
1277 mask = LLVMBuildOr(builder, mask, test, "");
1278 }
1279
1280 if (clip_user) {
1281 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1282 LLVMValueRef indices[3];
1283 LLVMValueRef is_nan_or_inf;
1284
1285 /* userclip planes */
1286 while (ucp_enable) {
1287 unsigned plane_idx = ffs(ucp_enable)-1;
1288 ucp_enable &= ~(1 << plane_idx);
1289 plane_idx += 6;
1290
1291 if (have_cd && num_written_clipdistance) {
1292 LLVMValueRef clipdist;
1293 int i;
1294 i = plane_idx - 6;
1295
1296 *have_clipdist = TRUE;
1297 if (i < 4) {
1298 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1299 } else {
1300 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1301 }
1302 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1303 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1304 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1305 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1306 test = LLVMBuildAnd(builder, test, temp, "");
1307 mask = LLVMBuildOr(builder, mask, test, "");
1308 } else {
1309 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1310 indices[0] = lp_build_const_int32(gallivm, 0);
1311 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1312
1313 indices[2] = lp_build_const_int32(gallivm, 0);
1314 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1315 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1316 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1317 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1318
1319 indices[2] = lp_build_const_int32(gallivm, 1);
1320 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1321 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1322 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1323 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1324
1325 indices[2] = lp_build_const_int32(gallivm, 2);
1326 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1327 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1328 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1329 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1330
1331 indices[2] = lp_build_const_int32(gallivm, 3);
1332 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1333 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1334 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1335 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1336
1337 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1338 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1339 test = LLVMBuildAnd(builder, test, temp, "");
1340 mask = LLVMBuildOr(builder, mask, test, "");
1341 }
1342 }
1343 }
1344 if (key->need_edgeflags) {
1345 /*
1346 * This isn't really part of clipmask but stored the same in vertex
1347 * header later, so do it here.
1348 */
1349 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1350 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1351 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1352 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1353 temp = lp_build_const_int_vec(gallivm, i32_type,
1354 1LL << DRAW_TOTAL_CLIP_PLANES);
1355 test = LLVMBuildAnd(builder, test, temp, "");
1356 mask = LLVMBuildOr(builder, mask, test, "");
1357 }
1358 return mask;
1359 }
1360
1361
1362 /**
1363 * Returns boolean if any clipping has occurred
1364 * Used zero/one i8 value to represent boolean
1365 */
1366 static LLVMValueRef
1367 clipmask_booli8(struct gallivm_state *gallivm,
1368 const struct lp_type vs_type,
1369 LLVMValueRef clipmask_bool_ptr,
1370 boolean edgeflag_in_clipmask)
1371 {
1372 LLVMBuilderRef builder = gallivm->builder;
1373 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1374 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1375 LLVMValueRef ret;
1376 struct lp_build_context bldivec;
1377
1378 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1379
1380 /*
1381 * We need to invert the edgeflag bit from the clipmask here
1382 * (because the result is really if we want to run the pipeline or not
1383 * and we (may) need it if edgeflag was 0).
1384 */
1385 if (edgeflag_in_clipmask) {
1386 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1387 1LL << DRAW_TOTAL_CLIP_PLANES);
1388 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1389 }
1390
1391 /*
1392 * XXX: probably should mask off bits from the mask which come from
1393 * vertices which were beyond the count (i.e. indices_valid for
1394 * linear fetches, for elts ones we don't have the correct mask
1395 * right now). Otherwise might run the pipeline for nothing,
1396 * though everything should still work.
1397 */
1398 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1399 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1400 return ret;
1401 }
1402
1403 static LLVMValueRef
1404 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1405 struct lp_build_tgsi_context * bld_base,
1406 boolean is_vindex_indirect,
1407 LLVMValueRef vertex_index,
1408 boolean is_aindex_indirect,
1409 LLVMValueRef attrib_index,
1410 LLVMValueRef swizzle_index)
1411 {
1412 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1413 struct gallivm_state *gallivm = bld_base->base.gallivm;
1414 LLVMBuilderRef builder = gallivm->builder;
1415 LLVMValueRef indices[3];
1416 LLVMValueRef res;
1417 struct lp_type type = bld_base->base.type;
1418
1419 if (is_vindex_indirect || is_aindex_indirect) {
1420 int i;
1421 res = bld_base->base.zero;
1422 for (i = 0; i < type.length; ++i) {
1423 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1424 LLVMValueRef vert_chan_index = vertex_index;
1425 LLVMValueRef attr_chan_index = attrib_index;
1426 LLVMValueRef channel_vec, value;
1427
1428 if (is_vindex_indirect) {
1429 vert_chan_index = LLVMBuildExtractElement(builder,
1430 vertex_index, idx, "");
1431 }
1432 if (is_aindex_indirect) {
1433 attr_chan_index = LLVMBuildExtractElement(builder,
1434 attrib_index, idx, "");
1435 }
1436
1437 indices[0] = vert_chan_index;
1438 indices[1] = attr_chan_index;
1439 indices[2] = swizzle_index;
1440
1441 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1442 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1443 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1444
1445 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1446 }
1447 } else {
1448 indices[0] = vertex_index;
1449 indices[1] = attrib_index;
1450 indices[2] = swizzle_index;
1451
1452 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1453 res = LLVMBuildLoad(builder, res, "");
1454 }
1455
1456 return res;
1457 }
1458
1459 static void
1460 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1461 struct lp_build_tgsi_context * bld_base,
1462 LLVMValueRef (*outputs)[4],
1463 LLVMValueRef emitted_vertices_vec)
1464 {
1465 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1466 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1467 struct gallivm_state *gallivm = variant->gallivm;
1468 LLVMBuilderRef builder = gallivm->builder;
1469 struct lp_type gs_type = bld_base->base.type;
1470 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1471 lp_int_type(gs_type), 0);
1472 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1473 LLVMValueRef next_prim_offset =
1474 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1475 LLVMValueRef io = variant->io_ptr;
1476 unsigned i;
1477 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1478
1479 for (i = 0; i < gs_type.length; ++i) {
1480 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1481 LLVMValueRef currently_emitted =
1482 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1483 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1484 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1485 }
1486
1487 convert_to_aos(gallivm, io, indices,
1488 outputs, clipmask,
1489 gs_info->num_outputs, gs_type,
1490 FALSE);
1491 }
1492
1493 static void
1494 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1495 struct lp_build_tgsi_context * bld_base,
1496 LLVMValueRef verts_per_prim_vec,
1497 LLVMValueRef emitted_prims_vec)
1498 {
1499 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1500 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1501 struct gallivm_state *gallivm = variant->gallivm;
1502 LLVMBuilderRef builder = gallivm->builder;
1503 LLVMValueRef prim_lengts_ptr =
1504 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1505 unsigned i;
1506
1507 for (i = 0; i < bld_base->base.type.length; ++i) {
1508 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1509 LLVMValueRef prims_emitted =
1510 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1511 LLVMValueRef store_ptr;
1512 LLVMValueRef num_vertices =
1513 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1514
1515 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1516 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1517 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1518 LLVMBuildStore(builder, num_vertices, store_ptr);
1519 }
1520 }
1521
1522 static void
1523 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1524 struct lp_build_tgsi_context * bld_base,
1525 LLVMValueRef total_emitted_vertices_vec,
1526 LLVMValueRef emitted_prims_vec)
1527 {
1528 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1529 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1530 struct gallivm_state *gallivm = variant->gallivm;
1531 LLVMBuilderRef builder = gallivm->builder;
1532 LLVMValueRef emitted_verts_ptr =
1533 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1534 LLVMValueRef emitted_prims_ptr =
1535 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1536 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1537
1538 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1539 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1540
1541 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1542 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1543 }
1544
1545 static void
1546 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1547 {
1548 struct gallivm_state *gallivm = variant->gallivm;
1549 LLVMContextRef context = gallivm->context;
1550 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1551 LLVMTypeRef arg_types[11];
1552 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1553 LLVMTypeRef func_type;
1554 LLVMValueRef context_ptr;
1555 LLVMBasicBlockRef block;
1556 LLVMBuilderRef builder;
1557 char func_name[64];
1558 struct lp_type vs_type;
1559 LLVMValueRef count, fetch_elts, start_or_maxelt;
1560 LLVMValueRef vertex_id_offset, start_instance;
1561 LLVMValueRef stride, step, io_itr;
1562 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1563 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1564 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1565 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1566 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1567 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1568 LLVMValueRef fake_buf_ptr, fake_buf;
1569
1570 struct draw_context *draw = llvm->draw;
1571 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1572 unsigned i, j;
1573 struct lp_build_context bld, blduivec;
1574 struct lp_build_loop_state lp_loop;
1575 struct lp_build_if_state if_ctx;
1576 const int vector_length = lp_native_vector_width / 32;
1577 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1578 struct lp_build_sampler_soa *sampler = 0;
1579 LLVMValueRef ret, clipmask_bool_ptr;
1580 struct draw_llvm_variant_key *key = &variant->key;
1581 /* If geometry shader is present we need to skip both the viewport
1582 * transformation and clipping otherwise the inputs to the geometry
1583 * shader will be incorrect.
1584 * The code can't handle vp transform when vs writes vp index neither
1585 * (though this would be fixable here, but couldn't just broadcast
1586 * the values).
1587 */
1588 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1589 vs_info->writes_viewport_index;
1590 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1591 key->clip_z ||
1592 key->clip_user ||
1593 key->need_edgeflags);
1594 LLVMValueRef variant_func;
1595 const unsigned pos = draw->vs.position_output;
1596 const unsigned cv = draw->vs.clipvertex_output;
1597 boolean have_clipdist = FALSE;
1598 struct lp_bld_tgsi_system_values system_values;
1599
1600 memset(&system_values, 0, sizeof(system_values));
1601
1602 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1603 variant->shader->variants_cached);
1604
1605 i = 0;
1606 arg_types[i++] = get_context_ptr_type(variant); /* context */
1607 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1608 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1609 arg_types[i++] = int32_type; /* count */
1610 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1611 arg_types[i++] = int32_type; /* stride */
1612 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1613 arg_types[i++] = int32_type; /* instance_id */
1614 arg_types[i++] = int32_type; /* vertex_id_offset */
1615 arg_types[i++] = int32_type; /* start_instance */
1616 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1617
1618 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1619 arg_types, num_arg_types, 0);
1620
1621 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1622 variant->function = variant_func;
1623
1624 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1625 for (i = 0; i < num_arg_types; ++i)
1626 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1627 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1628
1629 context_ptr = LLVMGetParam(variant_func, 0);
1630 io_ptr = LLVMGetParam(variant_func, 1);
1631 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1632 count = LLVMGetParam(variant_func, 3);
1633 /*
1634 * XXX: the maxelt part is unused. Not really useful, since we cannot
1635 * get index buffer overflows due to vsplit (which provides its own
1636 * elts buffer, with a different size than what's passed in here).
1637 */
1638 start_or_maxelt = LLVMGetParam(variant_func, 4);
1639 /*
1640 * XXX: stride is actually unused. The stride we use is strictly calculated
1641 * from the number of outputs (including the draw_extra outputs).
1642 * Should probably fix some day (we need a new vs just because of extra
1643 * outputs which the generated vs won't touch).
1644 */
1645 stride = LLVMGetParam(variant_func, 5);
1646 vb_ptr = LLVMGetParam(variant_func, 6);
1647 system_values.instance_id = LLVMGetParam(variant_func, 7);
1648 vertex_id_offset = LLVMGetParam(variant_func, 8);
1649 start_instance = LLVMGetParam(variant_func, 9);
1650 fetch_elts = LLVMGetParam(variant_func, 10);
1651
1652 lp_build_name(context_ptr, "context");
1653 lp_build_name(io_ptr, "io");
1654 lp_build_name(vbuffers_ptr, "vbuffers");
1655 lp_build_name(count, "count");
1656 lp_build_name(start_or_maxelt, "start_or_maxelt");
1657 lp_build_name(stride, "stride");
1658 lp_build_name(vb_ptr, "vb");
1659 lp_build_name(system_values.instance_id, "instance_id");
1660 lp_build_name(vertex_id_offset, "vertex_id_offset");
1661 lp_build_name(start_instance, "start_instance");
1662 lp_build_name(fetch_elts, "fetch_elts");
1663
1664 /*
1665 * Function body
1666 */
1667
1668 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1669 builder = gallivm->builder;
1670 LLVMPositionBuilderAtEnd(builder, block);
1671
1672 memset(&vs_type, 0, sizeof vs_type);
1673 vs_type.floating = TRUE; /* floating point values */
1674 vs_type.sign = TRUE; /* values are signed */
1675 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1676 vs_type.width = 32; /* 32-bit float */
1677 vs_type.length = vector_length;
1678
1679 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1680 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1681
1682 /* hold temporary "bool" clipmask */
1683 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1684
1685 fake_buf = lp_build_alloca_undef(gallivm,
1686 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1687 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1688 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1689 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1690
1691 /* code generated texture sampling */
1692 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1693
1694 step = lp_build_const_int32(gallivm, vector_length);
1695
1696 ind_vec = blduivec.undef;
1697 for (i = 0; i < vs_type.length; i++) {
1698 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1699 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1700 }
1701
1702 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1703 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1704
1705 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1706 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1707 /*
1708 * Only needed for non-indexed path.
1709 */
1710 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1711
1712 /*
1713 * Pre-calculate everything which is constant per shader invocation.
1714 */
1715 for (j = 0; j < key->nr_vertex_elements; ++j) {
1716 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1717 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1718 struct pipe_vertex_element *velem = &key->vertex_element[j];
1719 LLVMValueRef vb_index =
1720 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1721 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1722 util_format_get_blocksize(velem->src_format));
1723 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1724 velem->src_offset);
1725 struct lp_build_if_state if_ctx;
1726
1727 if (velem->src_format != PIPE_FORMAT_NONE) {
1728 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1729 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1730 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1731 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1732 LLVMInt32TypeInContext(context), "");
1733 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1734 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1735 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1736
1737 ofbit = NULL;
1738 /*
1739 * We'll set buffer_size_adj to zero if we have of, so it will
1740 * always overflow later automatically without having to keep ofbit.
1741 * Overflows (with normal wraparound) doing the actual offset
1742 * calculation should be ok, just not for the buffer size calc.
1743 * It would also be possible to detect such overflows and return
1744 * zeros if that happens, but this would be more complex.
1745 */
1746 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1747 tmp = lp_build_sub(&bld, bsize, bld.one);
1748 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1749 &ofbit);
1750 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1751 buf_offset, &ofbit);
1752
1753 /*
1754 * We can't easily set fake vertex buffers outside the generated code.
1755 * Hence, set fake vertex buffers here instead basically, so fetch
1756 * code can always fetch using offset 0, eliminating all control flow
1757 * inside the main loop.
1758 * (Alternatively, could have control flow per vector skipping fetch
1759 * if ofbit is true.)
1760 */
1761 if (velem->instance_divisor) {
1762 /*
1763 * Index is equal to the start instance plus the number of current
1764 * instance divided by the divisor. In this case we compute it as:
1765 * index = start_instance + (instance_id / divisor).
1766 * Note we could actually do the fetch here, outside the loop -
1767 * it's all constant, hopefully llvm recognizes this.
1768 */
1769 LLVMValueRef current_instance;
1770 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1771 lp_build_const_int32(gallivm,
1772 velem->instance_divisor),
1773 "instance_divisor");
1774 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1775 current_instance, &ofbit);
1776 }
1777
1778 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1779 buffer_size_adj[j], "");
1780
1781 temp_ptr = lp_build_alloca_undef(gallivm,
1782 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1783
1784 lp_build_if(&if_ctx, gallivm, ofbit);
1785 {
1786 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1787 }
1788 lp_build_else(&if_ctx);
1789 {
1790 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1791 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1792 }
1793 lp_build_endif(&if_ctx);
1794 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1795
1796 if (0) {
1797 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1798 lp_build_const_int32(gallivm, j),
1799 vb_index, vb_stride[j]);
1800 lp_build_printf(gallivm,
1801 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1802 vb_buffer_offset, src_offset, buf_offset);
1803 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1804 buffer_size, bsize);
1805 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1806 }
1807 }
1808 }
1809
1810 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1811 {
1812 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1813 LLVMValueRef io;
1814 LLVMValueRef clipmask; /* holds the clipmask value */
1815 LLVMValueRef true_index_array, index_store;
1816 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1817
1818 io_itr = lp_loop.counter;
1819
1820 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1821 #if DEBUG_STORE
1822 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1823 io_itr, io, lp_loop.counter);
1824 #endif
1825
1826 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1827 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1828
1829 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
1830 /*
1831 * Limit indices to fetch_max, otherwise might try to access indices
1832 * beyond index buffer (or rather vsplit elt buffer) size.
1833 * Could probably safely (?) skip this for non-indexed draws and
1834 * simplify things minimally (by removing it could combine the ind_vec
1835 * and start_vec adds). I think the only effect for non-indexed draws will
1836 * be that for the invalid elements they will be all fetched from the
1837 * same location as the last valid one, but noone should really care.
1838 */
1839 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1840
1841 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1842
1843 lp_build_if(&if_ctx, gallivm, have_elts);
1844 {
1845 /*
1846 * Note: you'd expect some comparison/clamp against fetch_elt_max
1847 * here.
1848 * There used to be one here but it was incorrect: overflow was
1849 * detected if index > fetch_elt_max - but the correct condition
1850 * would be index >= fetch_elt_max (since this is just size of elts
1851 * buffer / element size).
1852 * Using the correct condition however will cause failures - due to
1853 * vsplit/vcache code which rebases indices. So, as an example, if
1854 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1855 * replace all invalid indices with 0 - which in case of elt_bias
1856 * not being zero will get a different fetch index than the valid
1857 * index 0. So, just rely on vsplit code preventing out-of-bounds
1858 * fetches. This is also why it's safe to do elts fetch even if there
1859 * was no index buffer bound - the real buffer is never seen here, at
1860 * least not if there are index buffer overflows...
1861 */
1862
1863 /*
1864 * XXX should not have to do this, as scale can be handled
1865 * natively by loads (hits asserts though).
1866 */
1867 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1868 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1869 LLVMPointerType(LLVMInt8TypeInContext(context),
1870 0), "");
1871 tmp = lp_build_gather(gallivm, vs_type.length,
1872 32, bld.type, TRUE,
1873 fetch_elts, tmp, FALSE);
1874 LLVMBuildStore(builder, tmp, index_store);
1875 }
1876 lp_build_else(&if_ctx);
1877 {
1878 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1879 LLVMBuildStore(builder, tmp, index_store);
1880 }
1881 lp_build_endif(&if_ctx);
1882
1883 true_index_array = LLVMBuildLoad(builder, index_store, "");
1884
1885 for (j = 0; j < key->nr_vertex_elements; ++j) {
1886 struct pipe_vertex_element *velem = &key->vertex_element[j];
1887 const struct util_format_description *format_desc =
1888 util_format_description(velem->src_format);
1889
1890 if (format_desc->format == PIPE_FORMAT_NONE) {
1891 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1892 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1893 }
1894 }
1895 else if (velem->instance_divisor) {
1896 fetch_instanced(gallivm, format_desc, vs_type,
1897 vb_stride[j], map_ptr[j],
1898 buffer_size_adj[j],
1899 inputs[j], instance_index[j]);
1900 }
1901 else {
1902 fetch_vector(gallivm, format_desc, vs_type,
1903 vb_stride[j], map_ptr[j],
1904 buffer_size_adj[j],
1905 inputs[j], true_index_array);
1906 }
1907 }
1908
1909 struct lp_build_mask_context mask;
1910
1911 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
1912 /* In the paths with elts vertex id has to be unaffected by the
1913 * index bias and because indices inside our elements array have
1914 * already had index bias applied we need to subtract it here to
1915 * get back to the original index.
1916 * in the linear paths vertex id has to be unaffected by the
1917 * original start index and because we abuse the 'start' variable
1918 * to either represent the actual start index or the index at which
1919 * the primitive was split (we split rendering into chunks of at
1920 * most 4095-vertices) we need to back out the original start
1921 * index out of our vertex id here.
1922 */
1923 system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
1924 vertex_id_offset);
1925 system_values.vertex_id = true_index_array;
1926 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1927 system_values.basevertex, "");
1928
1929 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1930 generate_vs(variant,
1931 builder,
1932 vs_type,
1933 outputs,
1934 ptr_aos,
1935 &system_values,
1936 context_ptr,
1937 sampler,
1938 key->clamp_vertex_color,
1939 &mask);
1940
1941 lp_build_mask_end(&mask);
1942 if (pos != -1 && cv != -1) {
1943 /* store original positions in clip before further manipulation */
1944 store_clip(gallivm, vs_type, io, outputs, pos);
1945
1946 /* do cliptest */
1947 if (enable_cliptest) {
1948 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1949 /* allocate clipmask, assign it integer type */
1950 clipmask = generate_clipmask(llvm,
1951 gallivm,
1952 vs_type,
1953 outputs,
1954 key,
1955 context_ptr, &have_clipdist);
1956 temp = LLVMBuildOr(builder, clipmask, temp, "");
1957 /* store temporary clipping boolean value */
1958 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1959 }
1960 else {
1961 clipmask = blduivec.zero;
1962 }
1963
1964 /* do viewport mapping */
1965 if (!bypass_viewport) {
1966 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1967 }
1968 }
1969 else {
1970 clipmask = blduivec.zero;
1971 }
1972
1973 /* store clipmask in vertex header,
1974 * original positions in clip
1975 * and transformed positions in data
1976 */
1977 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1978 vs_info->num_outputs, vs_type,
1979 enable_cliptest && key->need_edgeflags);
1980 }
1981 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1982
1983 sampler->destroy(sampler);
1984
1985 /* return clipping boolean value for function */
1986 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
1987 enable_cliptest && key->need_edgeflags);
1988
1989 LLVMBuildRet(builder, ret);
1990
1991 gallivm_verify_function(gallivm, variant_func);
1992 }
1993
1994
1995 struct draw_llvm_variant_key *
1996 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1997 {
1998 unsigned i;
1999 struct draw_llvm_variant_key *key;
2000 struct draw_sampler_static_state *draw_sampler;
2001
2002 key = (struct draw_llvm_variant_key *)store;
2003
2004 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2005
2006 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2007
2008 /* will have to rig this up properly later */
2009 key->clip_xy = llvm->draw->clip_xy;
2010 key->clip_z = llvm->draw->clip_z;
2011 key->clip_user = llvm->draw->clip_user;
2012 key->bypass_viewport = llvm->draw->bypass_viewport;
2013 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2014 /* XXX assumes edgeflag output not at 0 */
2015 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2016 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2017 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2018 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2019
2020 /* All variants of this shader will have the same value for
2021 * nr_samplers. Not yet trying to compact away holes in the
2022 * sampler array.
2023 */
2024 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2025 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2026 key->nr_sampler_views =
2027 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2028 }
2029 else {
2030 key->nr_sampler_views = key->nr_samplers;
2031 }
2032
2033 /* Presumably all variants of the shader should have the same
2034 * number of vertex elements - ie the number of shader inputs.
2035 * NOTE: we NEED to store the needed number of needed inputs
2036 * here, not the number of provided elements to match keysize
2037 * (and the offset of sampler state in the key).
2038 * If we have excess number of vertex elements, this is valid,
2039 * but the excess ones don't matter.
2040 * If we don't have enough vertex elements (which looks not really
2041 * valid but we'll handle it gracefully) fill out missing ones with
2042 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2043 */
2044 key->nr_vertex_elements =
2045 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2046
2047 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2048 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2049 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2050 memset(key->vertex_element, 0,
2051 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2052 }
2053 memcpy(key->vertex_element,
2054 llvm->draw->pt.vertex_element,
2055 sizeof(struct pipe_vertex_element) *
2056 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2057
2058 draw_sampler = draw_llvm_variant_key_samplers(key);
2059 memset(draw_sampler, 0,
2060 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2061
2062 for (i = 0 ; i < key->nr_samplers; i++) {
2063 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2064 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2065 }
2066 for (i = 0 ; i < key->nr_sampler_views; i++) {
2067 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2068 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2069 }
2070
2071 return key;
2072 }
2073
2074
2075 void
2076 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2077 {
2078 unsigned i;
2079 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2080
2081 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2082 debug_printf("clip_xy = %u\n", key->clip_xy);
2083 debug_printf("clip_z = %u\n", key->clip_z);
2084 debug_printf("clip_user = %u\n", key->clip_user);
2085 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2086 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2087 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2088 debug_printf("has_gs = %u\n", key->has_gs);
2089 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2090
2091 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2092 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2093 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2094 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2095 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2096 }
2097
2098 for (i = 0 ; i < key->nr_sampler_views; i++) {
2099 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2100 }
2101 }
2102
2103
2104 void
2105 draw_llvm_set_mapped_texture(struct draw_context *draw,
2106 enum pipe_shader_type shader_stage,
2107 unsigned sview_idx,
2108 uint32_t width, uint32_t height, uint32_t depth,
2109 uint32_t first_level, uint32_t last_level,
2110 const void *base_ptr,
2111 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2112 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2113 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2114 {
2115 unsigned j;
2116 struct draw_jit_texture *jit_tex;
2117
2118 assert(shader_stage == PIPE_SHADER_VERTEX ||
2119 shader_stage == PIPE_SHADER_GEOMETRY);
2120
2121 if (shader_stage == PIPE_SHADER_VERTEX) {
2122 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2123
2124 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2125 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2126 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2127
2128 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2129 } else {
2130 assert(0);
2131 return;
2132 }
2133
2134 jit_tex->width = width;
2135 jit_tex->height = height;
2136 jit_tex->depth = depth;
2137 jit_tex->first_level = first_level;
2138 jit_tex->last_level = last_level;
2139 jit_tex->base = base_ptr;
2140
2141 for (j = first_level; j <= last_level; j++) {
2142 jit_tex->mip_offsets[j] = mip_offsets[j];
2143 jit_tex->row_stride[j] = row_stride[j];
2144 jit_tex->img_stride[j] = img_stride[j];
2145 }
2146 }
2147
2148
2149 void
2150 draw_llvm_set_sampler_state(struct draw_context *draw,
2151 enum pipe_shader_type shader_type)
2152 {
2153 unsigned i;
2154
2155 if (shader_type == PIPE_SHADER_VERTEX) {
2156 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2157 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2158
2159 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2160 const struct pipe_sampler_state *s
2161 = draw->samplers[PIPE_SHADER_VERTEX][i];
2162 jit_sam->min_lod = s->min_lod;
2163 jit_sam->max_lod = s->max_lod;
2164 jit_sam->lod_bias = s->lod_bias;
2165 COPY_4V(jit_sam->border_color, s->border_color.f);
2166 }
2167 }
2168 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2169 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2170 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2171
2172 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2173 const struct pipe_sampler_state *s
2174 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2175 jit_sam->min_lod = s->min_lod;
2176 jit_sam->max_lod = s->max_lod;
2177 jit_sam->lod_bias = s->lod_bias;
2178 COPY_4V(jit_sam->border_color, s->border_color.f);
2179 }
2180 }
2181 }
2182 }
2183
2184
2185 void
2186 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2187 {
2188 struct draw_llvm *llvm = variant->llvm;
2189
2190 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2191 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2192 variant->shader->variants_cached, llvm->nr_variants);
2193 }
2194
2195 gallivm_destroy(variant->gallivm);
2196
2197 remove_from_list(&variant->list_item_local);
2198 variant->shader->variants_cached--;
2199 remove_from_list(&variant->list_item_global);
2200 llvm->nr_variants--;
2201 FREE(variant);
2202 }
2203
2204
2205 /**
2206 * Create LLVM types for various structures.
2207 */
2208 static void
2209 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2210 {
2211 struct gallivm_state *gallivm = var->gallivm;
2212 LLVMTypeRef texture_type, sampler_type, context_type;
2213
2214 texture_type = create_jit_texture_type(gallivm, "texture");
2215 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2216
2217 context_type = create_gs_jit_context_type(gallivm,
2218 var->shader->base.vector_length,
2219 texture_type, sampler_type,
2220 "draw_gs_jit_context");
2221 var->context_ptr_type = LLVMPointerType(context_type, 0);
2222
2223 var->input_array_type = create_gs_jit_input_type(gallivm);
2224 }
2225
2226 static LLVMTypeRef
2227 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2228 {
2229 if (!variant->context_ptr_type)
2230 create_gs_jit_types(variant);
2231 return variant->context_ptr_type;
2232 }
2233
2234 static LLVMValueRef
2235 generate_mask_value(struct draw_gs_llvm_variant *variant,
2236 struct lp_type gs_type)
2237 {
2238 struct gallivm_state *gallivm = variant->gallivm;
2239 LLVMBuilderRef builder = gallivm->builder;
2240 struct lp_type mask_type = lp_int_type(gs_type);
2241 LLVMValueRef num_prims;
2242 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2243 unsigned i;
2244
2245 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2246 variant->num_prims);
2247 for (i = 0; i < gs_type.length; i++) {
2248 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2249 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2250 }
2251 mask_val = lp_build_compare(gallivm, mask_type,
2252 PIPE_FUNC_GREATER, num_prims, mask_val);
2253
2254 return mask_val;
2255 }
2256
2257 static void
2258 draw_gs_llvm_generate(struct draw_llvm *llvm,
2259 struct draw_gs_llvm_variant *variant)
2260 {
2261 struct gallivm_state *gallivm = variant->gallivm;
2262 LLVMContextRef context = gallivm->context;
2263 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2264 LLVMTypeRef arg_types[7];
2265 LLVMTypeRef func_type;
2266 LLVMValueRef variant_func;
2267 LLVMValueRef context_ptr;
2268 LLVMValueRef prim_id_ptr;
2269 LLVMBasicBlockRef block;
2270 LLVMBuilderRef builder;
2271 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2272 struct lp_build_sampler_soa *sampler = 0;
2273 struct lp_build_context bld;
2274 struct lp_bld_tgsi_system_values system_values;
2275 char func_name[64];
2276 struct lp_type gs_type;
2277 unsigned i;
2278 struct draw_gs_llvm_iface gs_iface;
2279 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2280 LLVMValueRef consts_ptr, num_consts_ptr;
2281 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2282 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2283 struct lp_build_mask_context mask;
2284 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2285 unsigned vector_length = variant->shader->base.vector_length;
2286
2287 memset(&system_values, 0, sizeof(system_values));
2288
2289 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2290 variant->shader->variants_cached);
2291
2292 assert(variant->vertex_header_ptr_type);
2293
2294 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2295 arg_types[1] = variant->input_array_type; /* input */
2296 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2297 arg_types[3] = int32_type; /* num_prims */
2298 arg_types[4] = int32_type; /* instance_id */
2299 arg_types[5] = LLVMPointerType(
2300 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2301 arg_types[6] = int32_type;
2302
2303 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2304
2305 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2306
2307 variant->function = variant_func;
2308
2309 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2310
2311 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2312 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2313 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2314
2315 context_ptr = LLVMGetParam(variant_func, 0);
2316 input_array = LLVMGetParam(variant_func, 1);
2317 io_ptr = LLVMGetParam(variant_func, 2);
2318 num_prims = LLVMGetParam(variant_func, 3);
2319 system_values.instance_id = LLVMGetParam(variant_func, 4);
2320 prim_id_ptr = LLVMGetParam(variant_func, 5);
2321 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2322
2323 lp_build_name(context_ptr, "context");
2324 lp_build_name(input_array, "input");
2325 lp_build_name(io_ptr, "io");
2326 lp_build_name(num_prims, "num_prims");
2327 lp_build_name(system_values.instance_id, "instance_id");
2328 lp_build_name(prim_id_ptr, "prim_id_ptr");
2329 lp_build_name(system_values.invocation_id, "invocation_id");
2330
2331 variant->context_ptr = context_ptr;
2332 variant->io_ptr = io_ptr;
2333 variant->num_prims = num_prims;
2334
2335 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2336 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2337 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2338 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2339 gs_iface.input = input_array;
2340 gs_iface.variant = variant;
2341
2342 /*
2343 * Function body
2344 */
2345
2346 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2347 builder = gallivm->builder;
2348 LLVMPositionBuilderAtEnd(builder, block);
2349
2350 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2351
2352 memset(&gs_type, 0, sizeof gs_type);
2353 gs_type.floating = TRUE; /* floating point values */
2354 gs_type.sign = TRUE; /* values are signed */
2355 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2356 gs_type.width = 32; /* 32-bit float */
2357 gs_type.length = vector_length;
2358
2359 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2360 num_consts_ptr =
2361 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2362
2363 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2364 num_ssbos_ptr =
2365 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2366
2367 /* code generated texture sampling */
2368 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2369
2370 mask_val = generate_mask_value(variant, gs_type);
2371 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2372
2373 if (gs_info->uses_primid) {
2374 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2375 }
2376
2377 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2378 tgsi_dump(tokens, 0);
2379 draw_gs_llvm_dump_variant_key(&variant->key);
2380 }
2381
2382 struct lp_build_tgsi_params params = {};
2383
2384 params.type = gs_type;
2385 params.mask = &mask;
2386 params.consts_ptr = consts_ptr;
2387 params.const_sizes_ptr = num_consts_ptr;
2388 params.system_values = &system_values;
2389 params.context_ptr = context_ptr;
2390 params.sampler = sampler;
2391 params.info = &llvm->draw->gs.geometry_shader->info;
2392 params.gs_iface = (const struct lp_build_tgsi_gs_iface *)&gs_iface;
2393 params.ssbo_ptr = ssbos_ptr;
2394 params.ssbo_sizes_ptr = num_ssbos_ptr;
2395
2396 lp_build_tgsi_soa(variant->gallivm,
2397 tokens,
2398 &params,
2399 outputs);
2400
2401 sampler->destroy(sampler);
2402
2403 lp_build_mask_end(&mask);
2404
2405 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2406
2407 gallivm_verify_function(gallivm, variant_func);
2408 }
2409
2410
2411 struct draw_gs_llvm_variant *
2412 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2413 unsigned num_outputs,
2414 const struct draw_gs_llvm_variant_key *key)
2415 {
2416 struct draw_gs_llvm_variant *variant;
2417 struct llvm_geometry_shader *shader =
2418 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2419 LLVMTypeRef vertex_header;
2420 char module_name[64];
2421
2422 variant = MALLOC(sizeof *variant +
2423 shader->variant_key_size -
2424 sizeof variant->key);
2425 if (!variant)
2426 return NULL;
2427
2428 variant->llvm = llvm;
2429 variant->shader = shader;
2430
2431 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2432 variant->shader->variants_cached);
2433
2434 variant->gallivm = gallivm_create(module_name, llvm->context);
2435
2436 create_gs_jit_types(variant);
2437
2438 memcpy(&variant->key, key, shader->variant_key_size);
2439
2440 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2441
2442 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2443
2444 draw_gs_llvm_generate(llvm, variant);
2445
2446 gallivm_compile_module(variant->gallivm);
2447
2448 variant->jit_func = (draw_gs_jit_func)
2449 gallivm_jit_function(variant->gallivm, variant->function);
2450
2451 gallivm_free_ir(variant->gallivm);
2452
2453 variant->list_item_global.base = variant;
2454 variant->list_item_local.base = variant;
2455 /*variant->no = */shader->variants_created++;
2456 variant->list_item_global.base = variant;
2457
2458 return variant;
2459 }
2460
2461 void
2462 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2463 {
2464 struct draw_llvm *llvm = variant->llvm;
2465
2466 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2467 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2468 variant->shader->variants_cached, llvm->nr_gs_variants);
2469 }
2470
2471 gallivm_destroy(variant->gallivm);
2472
2473 remove_from_list(&variant->list_item_local);
2474 variant->shader->variants_cached--;
2475 remove_from_list(&variant->list_item_global);
2476 llvm->nr_gs_variants--;
2477 FREE(variant);
2478 }
2479
2480 struct draw_gs_llvm_variant_key *
2481 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2482 {
2483 unsigned i;
2484 struct draw_gs_llvm_variant_key *key;
2485 struct draw_sampler_static_state *draw_sampler;
2486
2487 key = (struct draw_gs_llvm_variant_key *)store;
2488
2489 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2490
2491 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2492
2493 /* All variants of this shader will have the same value for
2494 * nr_samplers. Not yet trying to compact away holes in the
2495 * sampler array.
2496 */
2497 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2498 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2499 key->nr_sampler_views =
2500 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2501 }
2502 else {
2503 key->nr_sampler_views = key->nr_samplers;
2504 }
2505
2506 draw_sampler = key->samplers;
2507
2508 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2509
2510 for (i = 0 ; i < key->nr_samplers; i++) {
2511 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2512 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2513 }
2514 for (i = 0 ; i < key->nr_sampler_views; i++) {
2515 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2516 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2517 }
2518
2519 return key;
2520 }
2521
2522 void
2523 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2524 {
2525 unsigned i;
2526 struct draw_sampler_static_state *sampler = key->samplers;
2527
2528 for (i = 0 ; i < key->nr_sampler_views; i++) {
2529 debug_printf("sampler[%i].src_format = %s\n", i,
2530 util_format_name(sampler[i].texture_state.format));
2531 }
2532 }