projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: add jit image type for vs/gs images.
[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, base,
149 target, texture_type,
150 DRAW_JIT_TEXTURE_BASE);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
152 target, texture_type,
153 DRAW_JIT_TEXTURE_ROW_STRIDE);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_IMG_STRIDE);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_FIRST_LEVEL);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_LAST_LEVEL);
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 * Create LLVM type for struct draw_jit_texture
213 */
214 static LLVMTypeRef
215 create_jit_image_type(struct gallivm_state *gallivm, const char *struct_name)
216 {
217 LLVMTargetDataRef target = gallivm->target;
218 LLVMTypeRef image_type;
219 LLVMTypeRef elem_types[DRAW_JIT_IMAGE_NUM_FIELDS];
220 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
221
222 elem_types[DRAW_JIT_IMAGE_WIDTH] =
223 elem_types[DRAW_JIT_IMAGE_HEIGHT] =
224 elem_types[DRAW_JIT_IMAGE_DEPTH] =
225 elem_types[DRAW_JIT_IMAGE_ROW_STRIDE] =
226 elem_types[DRAW_JIT_IMAGE_IMG_STRIDE] = int32_type;
227 elem_types[DRAW_JIT_IMAGE_BASE] =
228 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
229
230 image_type = LLVMStructTypeInContext(gallivm->context, elem_types,
231 ARRAY_SIZE(elem_types), 0);
232
233 (void) target; /* silence unused var warning for non-debug build */
234 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, width,
235 target, image_type,
236 DRAW_JIT_IMAGE_WIDTH);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, height,
238 target, image_type,
239 DRAW_JIT_IMAGE_HEIGHT);
240 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, depth,
241 target, image_type,
242 DRAW_JIT_IMAGE_DEPTH);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, base,
244 target, image_type,
245 DRAW_JIT_IMAGE_BASE);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, row_stride,
247 target, image_type,
248 DRAW_JIT_IMAGE_ROW_STRIDE);
249 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, img_stride,
250 target, image_type,
251 DRAW_JIT_IMAGE_IMG_STRIDE);
252
253 LP_CHECK_STRUCT_SIZE(struct draw_jit_image, target, image_type);
254
255 return image_type;
256 }
257
258 /**
259 * Create LLVM type for struct draw_jit_context
260 */
261 static LLVMTypeRef
262 create_jit_context_type(struct gallivm_state *gallivm,
263 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
264 LLVMTypeRef image_type,
265 const char *struct_name)
266 {
267 LLVMTargetDataRef target = gallivm->target;
268 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
269 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
270 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
271 LLVMTypeRef context_type;
272
273 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
274 LP_MAX_TGSI_CONST_BUFFERS);
275 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
276 LP_MAX_TGSI_CONST_BUFFERS);
277 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
278 DRAW_TOTAL_CLIP_PLANES), 0);
279 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
280 elem_types[4] = LLVMArrayType(texture_type,
281 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
282 elem_types[5] = LLVMArrayType(sampler_type,
283 PIPE_MAX_SAMPLERS); /* samplers */
284 elem_types[6] = LLVMArrayType(image_type,
285 PIPE_MAX_SHADER_IMAGES); /* images */
286 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
287 LP_MAX_TGSI_SHADER_BUFFERS);
288 elem_types[8] = LLVMArrayType(int_type, /* num_vs_ssbos */
289 LP_MAX_TGSI_SHADER_BUFFERS);
290 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
291 ARRAY_SIZE(elem_types), 0);
292
293 (void) target; /* silence unused var warning for non-debug build */
294 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
295 target, context_type, DRAW_JIT_CTX_CONSTANTS);
296 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
297 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
298 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
299 target, context_type, DRAW_JIT_CTX_PLANES);
300 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
301 target, context_type, DRAW_JIT_CTX_VIEWPORT);
302 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
303 target, context_type,
304 DRAW_JIT_CTX_TEXTURES);
305 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
306 target, context_type,
307 DRAW_JIT_CTX_SAMPLERS);
308 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, images,
309 target, context_type, DRAW_JIT_CTX_IMAGES);
310 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
311 target, context_type, DRAW_JIT_CTX_SSBOS);
312 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
313 target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
314 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
315 target, context_type);
316
317 return context_type;
318 }
319
320
321 /**
322 * Create LLVM type for struct draw_gs_jit_context
323 */
324 static LLVMTypeRef
325 create_gs_jit_context_type(struct gallivm_state *gallivm,
326 unsigned vector_length,
327 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
328 LLVMTypeRef image_type,
329 const char *struct_name)
330 {
331 LLVMTargetDataRef target = gallivm->target;
332 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
333 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
334 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
335 LLVMTypeRef context_type;
336
337 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
338 LP_MAX_TGSI_CONST_BUFFERS);
339 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
340 LP_MAX_TGSI_CONST_BUFFERS);
341 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
342 DRAW_TOTAL_CLIP_PLANES), 0);
343 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
344
345 elem_types[4] = LLVMArrayType(texture_type,
346 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
347 elem_types[5] = LLVMArrayType(sampler_type,
348 PIPE_MAX_SAMPLERS); /* samplers */
349 elem_types[6] = LLVMArrayType(image_type,
350 PIPE_MAX_SHADER_IMAGES); /* images */
351 elem_types[7] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
352 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
353 vector_length), 0);
354 elem_types[9] = LLVMPointerType(LLVMVectorType(int_type,
355 vector_length), 0);
356
357 elem_types[10] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
358 LP_MAX_TGSI_SHADER_BUFFERS);
359 elem_types[11] = LLVMArrayType(int_type, /* num_ssbos */
360 LP_MAX_TGSI_SHADER_BUFFERS);
361
362 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
363 ARRAY_SIZE(elem_types), 0);
364
365 (void) target; /* silence unused var warning for non-debug build */
366 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
367 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
368 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
369 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
370 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
371 target, context_type, DRAW_GS_JIT_CTX_PLANES);
372 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
373 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
374 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
375 target, context_type,
376 DRAW_GS_JIT_CTX_TEXTURES);
377 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
378 target, context_type,
379 DRAW_GS_JIT_CTX_SAMPLERS);
380 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
381 target, context_type,
382 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
383 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
384 target, context_type,
385 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
386 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
387 target, context_type,
388 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
389 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
390 target, context_type, DRAW_GS_JIT_CTX_SSBOS);
391 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
392 target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
393 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, images,
394 target, context_type, DRAW_GS_JIT_CTX_IMAGES);
395 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
396 target, context_type);
397
398 return context_type;
399 }
400
401
402 static LLVMTypeRef
403 create_gs_jit_input_type(struct gallivm_state *gallivm)
404 {
405 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
406 LLVMTypeRef input_array;
407
408 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
409 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
410 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
411 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
412
413 return input_array;
414 }
415
416 /**
417 * Create LLVM type for struct pipe_vertex_buffer
418 */
419 static LLVMTypeRef
420 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
421 const char *struct_name)
422 {
423 LLVMTargetDataRef target = gallivm->target;
424 LLVMTypeRef elem_types[4];
425 LLVMTypeRef vb_type;
426
427 elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
428 elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
429 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
430 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
431
432 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
433 ARRAY_SIZE(elem_types), 0);
434
435 (void) target; /* silence unused var warning for non-debug build */
436 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
437 target, vb_type, 0);
438 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
439 target, vb_type, 1);
440 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
441 target, vb_type, 2);
442 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
443 target, vb_type, 3);
444
445 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
446
447 return vb_type;
448 }
449
450
451 /**
452 * Create LLVM type for struct vertex_header;
453 */
454 static LLVMTypeRef
455 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
456 {
457 LLVMTargetDataRef target = gallivm->target;
458 LLVMTypeRef elem_types[3];
459 LLVMTypeRef vertex_header;
460 char struct_name[24];
461
462 snprintf(struct_name, 23, "vertex_header%d", data_elems);
463
464 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
465 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
466 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
467
468 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
469 ARRAY_SIZE(elem_types), 0);
470
471 /* these are bit-fields and we can't take address of them
472 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
473 target, vertex_header,
474 DRAW_JIT_VERTEX_CLIPMASK);
475 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
476 target, vertex_header,
477 DRAW_JIT_VERTEX_EDGEFLAG);
478 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
479 target, vertex_header,
480 DRAW_JIT_VERTEX_PAD);
481 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
482 target, vertex_header,
483 DRAW_JIT_VERTEX_VERTEX_ID);
484 */
485 (void) target; /* silence unused var warning for non-debug build */
486 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
487 target, vertex_header,
488 DRAW_JIT_VERTEX_CLIP_POS);
489 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
490 target, vertex_header,
491 DRAW_JIT_VERTEX_DATA);
492
493 assert(LLVMABISizeOfType(target, vertex_header) ==
494 offsetof(struct vertex_header, data[data_elems]));
495
496 return vertex_header;
497 }
498
499
500 /**
501 * Create LLVM types for various structures.
502 */
503 static void
504 create_jit_types(struct draw_llvm_variant *variant)
505 {
506 struct gallivm_state *gallivm = variant->gallivm;
507 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
508 vb_type, image_type;
509
510 texture_type = create_jit_texture_type(gallivm, "texture");
511 sampler_type = create_jit_sampler_type(gallivm, "sampler");
512 image_type = create_jit_image_type(gallivm, "image");
513
514 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
515 image_type,
516 "draw_jit_context");
517 variant->context_ptr_type = LLVMPointerType(context_type, 0);
518
519 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
520 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
521
522 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
523 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
524 }
525
526
527 static LLVMTypeRef
528 get_context_ptr_type(struct draw_llvm_variant *variant)
529 {
530 if (!variant->context_ptr_type)
531 create_jit_types(variant);
532 return variant->context_ptr_type;
533 }
534
535
536 static LLVMTypeRef
537 get_buffer_ptr_type(struct draw_llvm_variant *variant)
538 {
539 if (!variant->buffer_ptr_type)
540 create_jit_types(variant);
541 return variant->buffer_ptr_type;
542 }
543
544
545 static LLVMTypeRef
546 get_vb_ptr_type(struct draw_llvm_variant *variant)
547 {
548 if (!variant->vb_ptr_type)
549 create_jit_types(variant);
550 return variant->vb_ptr_type;
551 }
552
553 static LLVMTypeRef
554 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
555 {
556 if (!variant->vertex_header_ptr_type)
557 create_jit_types(variant);
558 return variant->vertex_header_ptr_type;
559 }
560
561
562 /**
563 * Create per-context LLVM info.
564 */
565 struct draw_llvm *
566 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
567 {
568 struct draw_llvm *llvm;
569
570 if (!lp_build_init())
571 return NULL;
572
573 llvm = CALLOC_STRUCT( draw_llvm );
574 if (!llvm)
575 return NULL;
576
577 llvm->draw = draw;
578
579 llvm->context = context;
580 if (!llvm->context) {
581 llvm->context = LLVMContextCreate();
582 llvm->context_owned = true;
583 }
584 if (!llvm->context)
585 goto fail;
586
587 llvm->nr_variants = 0;
588 make_empty_list(&llvm->vs_variants_list);
589
590 llvm->nr_gs_variants = 0;
591 make_empty_list(&llvm->gs_variants_list);
592
593 return llvm;
594
595 fail:
596 draw_llvm_destroy(llvm);
597 return NULL;
598 }
599
600
601 /**
602 * Free per-context LLVM info.
603 */
604 void
605 draw_llvm_destroy(struct draw_llvm *llvm)
606 {
607 if (llvm->context_owned)
608 LLVMContextDispose(llvm->context);
609 llvm->context = NULL;
610
611 /* XXX free other draw_llvm data? */
612 FREE(llvm);
613 }
614
615
616 /**
617 * Create LLVM-generated code for a vertex shader.
618 */
619 struct draw_llvm_variant *
620 draw_llvm_create_variant(struct draw_llvm *llvm,
621 unsigned num_inputs,
622 const struct draw_llvm_variant_key *key)
623 {
624 struct draw_llvm_variant *variant;
625 struct llvm_vertex_shader *shader =
626 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
627 LLVMTypeRef vertex_header;
628 char module_name[64];
629
630 variant = MALLOC(sizeof *variant +
631 shader->variant_key_size -
632 sizeof variant->key);
633 if (!variant)
634 return NULL;
635
636 variant->llvm = llvm;
637 variant->shader = shader;
638
639 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
640 variant->shader->variants_cached);
641
642 variant->gallivm = gallivm_create(module_name, llvm->context);
643
644 create_jit_types(variant);
645
646 memcpy(&variant->key, key, shader->variant_key_size);
647
648 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
649 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
650 draw_llvm_dump_variant_key(&variant->key);
651 }
652
653 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
654
655 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
656
657 draw_llvm_generate(llvm, variant);
658
659 gallivm_compile_module(variant->gallivm);
660
661 variant->jit_func = (draw_jit_vert_func)
662 gallivm_jit_function(variant->gallivm, variant->function);
663
664 gallivm_free_ir(variant->gallivm);
665
666 variant->list_item_global.base = variant;
667 variant->list_item_local.base = variant;
668 /*variant->no = */shader->variants_created++;
669 variant->list_item_global.base = variant;
670
671 return variant;
672 }
673
674
675 static void
676 generate_vs(struct draw_llvm_variant *variant,
677 LLVMBuilderRef builder,
678 struct lp_type vs_type,
679 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
680 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
681 const struct lp_bld_tgsi_system_values *system_values,
682 LLVMValueRef context_ptr,
683 const struct lp_build_sampler_soa *draw_sampler,
684 boolean clamp_vertex_color,
685 struct lp_build_mask_context *bld_mask)
686 {
687 struct draw_llvm *llvm = variant->llvm;
688 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
689 LLVMValueRef consts_ptr =
690 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
691 LLVMValueRef num_consts_ptr =
692 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
693 LLVMValueRef ssbos_ptr =
694 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
695 LLVMValueRef num_ssbos_ptr =
696 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
697
698 struct lp_build_tgsi_params params;
699 memset(&params, 0, sizeof(params));
700
701 params.type = vs_type;
702 params.mask = bld_mask;
703 params.consts_ptr = consts_ptr;
704 params.const_sizes_ptr = num_consts_ptr;
705 params.system_values = system_values;
706 params.inputs = inputs;
707 params.context_ptr = context_ptr;
708 params.sampler = draw_sampler;
709 params.info = &llvm->draw->vs.vertex_shader->info;
710 params.ssbo_ptr = ssbos_ptr;
711 params.ssbo_sizes_ptr = num_ssbos_ptr;
712
713 lp_build_tgsi_soa(variant->gallivm,
714 tokens,
715 &params,
716 outputs);
717
718 {
719 LLVMValueRef out;
720 unsigned chan, attrib;
721 struct lp_build_context bld;
722 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
723 lp_build_context_init(&bld, variant->gallivm, vs_type);
724
725 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
726 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
727 if (outputs[attrib][chan]) {
728 switch (info->output_semantic_name[attrib]) {
729 case TGSI_SEMANTIC_COLOR:
730 case TGSI_SEMANTIC_BCOLOR:
731 if (clamp_vertex_color) {
732 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
733 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
734 LLVMBuildStore(builder, out, outputs[attrib][chan]);
735 }
736 break;
737 }
738 }
739 }
740 }
741 }
742 }
743
744
745 static void
746 fetch_instanced(struct gallivm_state *gallivm,
747 const struct util_format_description *format_desc,
748 struct lp_type vs_type,
749 LLVMValueRef vb_stride,
750 LLVMValueRef map_ptr,
751 LLVMValueRef buffer_size_adj,
752 LLVMValueRef *inputs,
753 LLVMValueRef index)
754 {
755 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
756 LLVMTypeRef aosf_t, aosi_t;
757 LLVMValueRef zero = LLVMConstNull(i32_t);
758 LLVMBuilderRef builder = gallivm->builder;
759 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
760 unsigned i;
761
762 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
763 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
764
765 /* This mul can overflow. Wraparound is ok. */
766 stride = LLVMBuildMul(builder, vb_stride, index, "");
767
768 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
769 stride, buffer_size_adj,
770 "buffer_overflowed");
771
772 if (0) {
773 lp_build_print_value(gallivm, " instance index = ", index);
774 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
775 }
776
777 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
778 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
779 stride = LLVMBuildAnd(builder, stride, index_valid, "");
780
781 aos = lp_build_fetch_rgba_aos(gallivm,
782 format_desc,
783 lp_float32_vec4_type(),
784 FALSE,
785 map_ptr,
786 stride, zero, zero,
787 NULL);
788
789 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
790 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
791 aos = LLVMBuildAnd(builder, aos, index_valid, "");
792 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
793
794 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
795 LLVMValueRef index = lp_build_const_int32(gallivm, i);
796 inputs[i] = lp_build_extract_broadcast(gallivm,
797 lp_float32_vec4_type(),
798 vs_type, aos, index);
799 }
800 }
801
802
803 static void
804 fetch_vector(struct gallivm_state *gallivm,
805 const struct util_format_description *format_desc,
806 struct lp_type vs_type,
807 LLVMValueRef vb_stride,
808 LLVMValueRef map_ptr,
809 LLVMValueRef buffer_size_adj,
810 LLVMValueRef *inputs,
811 LLVMValueRef indices)
812 {
813 LLVMBuilderRef builder = gallivm->builder;
814 struct lp_build_context blduivec;
815 struct lp_type fetch_type = vs_type;
816 LLVMValueRef offset, valid_mask;
817 unsigned i;
818
819 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
820
821 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
822 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
823
824 /* This mul can overflow. Wraparound is ok. */
825 offset = lp_build_mul(&blduivec, vb_stride, indices);
826
827 valid_mask = lp_build_compare(gallivm, blduivec.type,
828 PIPE_FUNC_LESS, offset, buffer_size_adj);
829
830 /* not valid elements use offset 0 */
831 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
832
833 if (0) {
834 lp_build_print_value(gallivm, " indices = ", indices);
835 lp_build_print_value(gallivm, " offsets = ", offset);
836 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
837 }
838
839 /*
840 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
841 * This should always produce better code.
842 */
843
844 /* The type handling is annoying here... */
845 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
846 format_desc->channel[0].pure_integer) {
847 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
848 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
849 }
850 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
851 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
852 }
853 }
854
855 lp_build_fetch_rgba_soa(gallivm, format_desc,
856 fetch_type, FALSE, map_ptr, offset,
857 blduivec.zero, blduivec.zero,
858 NULL, inputs);
859
860 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
861 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
862 lp_build_vec_type(gallivm, vs_type), "");
863 }
864
865 /* out-of-bound fetches return all zeros */
866 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
867 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
868 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
869 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
870 lp_build_vec_type(gallivm, vs_type), "");
871 }
872 }
873
874
875 static void
876 store_aos(struct gallivm_state *gallivm,
877 LLVMValueRef io_ptr,
878 LLVMValueRef index,
879 LLVMValueRef value)
880 {
881 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
882 LLVMBuilderRef builder = gallivm->builder;
883 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
884 LLVMValueRef indices[3];
885
886 indices[0] = lp_build_const_int32(gallivm, 0);
887 indices[1] = index;
888 indices[2] = lp_build_const_int32(gallivm, 0);
889
890 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
891 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
892
893 #if DEBUG_STORE
894 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
895 #endif
896
897 /* Unaligned store due to the vertex header */
898 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
899 }
900
901 /**
902 * Adjust the mask to architecture endianess. The mask will the store in struct:
903 *
904 * struct vertex_header {
905 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
906 * unsigned edgeflag:1;
907 * unsigned pad:1;
908 * unsigned vertex_id:16;
909 * [...]
910 * }
911 *
912 * On little-endian machine nothing needs to done, however on bit-endian machine
913 * the mask's fields need to be adjusted with the algorithm:
914 *
915 * uint32_t reverse (uint32_t x)
916 * {
917 * return (x >> 16) | // vertex_id
918 * ((x & 0x3fff) << 18) | // clipmask
919 * ((x & 0x4000) << 3) | // pad
920 * ((x & 0x8000) << 1); // edgeflag
921 * }
922 */
923 static LLVMValueRef
924 adjust_mask(struct gallivm_state *gallivm,
925 LLVMValueRef mask)
926 {
927 #ifdef PIPE_ARCH_BIG_ENDIAN
928 LLVMBuilderRef builder = gallivm->builder;
929 LLVMValueRef vertex_id;
930 LLVMValueRef clipmask;
931 LLVMValueRef pad;
932 LLVMValueRef edgeflag;
933
934 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
935 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
936 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
937 if (0) {
938 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
939 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
940 }
941 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
942 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
943
944 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
945 if (0) {
946 mask = LLVMBuildOr(builder, mask, pad, "");
947 }
948 mask = LLVMBuildOr(builder, mask, edgeflag, "");
949 #endif
950 return mask;
951 }
952
953 static void
954 store_aos_array(struct gallivm_state *gallivm,
955 struct lp_type soa_type,
956 LLVMValueRef io_ptr,
957 LLVMValueRef *indices,
958 LLVMValueRef* aos,
959 int attrib,
960 int num_outputs,
961 LLVMValueRef clipmask,
962 boolean need_edgeflag)
963 {
964 LLVMBuilderRef builder = gallivm->builder;
965 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
966 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
967 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
968 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
969 int vector_length = soa_type.length;
970 int i;
971
972 debug_assert(TGSI_NUM_CHANNELS == 4);
973
974 for (i = 0; i < vector_length; i++) {
975 linear_inds[i] = lp_build_const_int32(gallivm, i);
976 if (indices) {
977 inds[i] = indices[i];
978 } else {
979 inds[i] = linear_inds[i];
980 }
981 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
982 }
983
984 if (attrib == 0) {
985 /* store vertex header for each of the n vertices */
986 LLVMValueRef val, cliptmp;
987 int vertex_id_pad_edgeflag;
988
989 /* If this assertion fails, it means we need to update the bit twidding
990 * code here. See struct vertex_header in draw_private.h.
991 */
992 assert(DRAW_TOTAL_CLIP_PLANES==14);
993 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
994 if (!need_edgeflag) {
995 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
996 }
997 else {
998 vertex_id_pad_edgeflag = (0xffff << 16);
999 }
1000 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
1001 vertex_id_pad_edgeflag);
1002 /* OR with the clipmask */
1003 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
1004 for (i = 0; i < vector_length; i++) {
1005 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
1006 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
1007 val = adjust_mask(gallivm, val);
1008 #if DEBUG_STORE
1009 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
1010 io_ptrs[i], inds[i], val);
1011 #endif
1012 LLVMBuildStore(builder, val, id_ptr);
1013 }
1014 }
1015
1016 /* store for each of the n vertices */
1017 for (i = 0; i < vector_length; i++) {
1018 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
1019 }
1020 }
1021
1022
1023 static void
1024 convert_to_aos(struct gallivm_state *gallivm,
1025 LLVMValueRef io,
1026 LLVMValueRef *indices,
1027 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1028 LLVMValueRef clipmask,
1029 int num_outputs,
1030 struct lp_type soa_type,
1031 boolean need_edgeflag)
1032 {
1033 LLVMBuilderRef builder = gallivm->builder;
1034 unsigned chan, attrib, i;
1035
1036 #if DEBUG_STORE
1037 lp_build_printf(gallivm, " # storing begin\n");
1038 #endif
1039 for (attrib = 0; attrib < num_outputs; ++attrib) {
1040 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1041 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1042 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1043 if (outputs[attrib][chan]) {
1044 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1045 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1046 #if DEBUG_STORE
1047 lp_build_printf(gallivm, "output %d : %d ",
1048 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1049 attrib, 0),
1050 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1051 chan, 0));
1052 lp_build_print_value(gallivm, "val = ", out);
1053 {
1054 LLVMValueRef iv =
1055 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1056
1057 lp_build_print_value(gallivm, " ival = ", iv);
1058 }
1059 #endif
1060 soa[chan] = out;
1061 }
1062 else {
1063 soa[chan] = 0;
1064 }
1065 }
1066
1067
1068 if (soa_type.length == TGSI_NUM_CHANNELS) {
1069 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1070 } else {
1071 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1072
1073 for (i = 0; i < soa_type.length; ++i) {
1074 aos[i] = lp_build_extract_range(gallivm,
1075 soa[i % TGSI_NUM_CHANNELS],
1076 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1077 TGSI_NUM_CHANNELS);
1078 }
1079 }
1080
1081 store_aos_array(gallivm,
1082 soa_type,
1083 io, indices,
1084 aos,
1085 attrib,
1086 num_outputs,
1087 clipmask,
1088 need_edgeflag);
1089 }
1090 #if DEBUG_STORE
1091 lp_build_printf(gallivm, " # storing end\n");
1092 #endif
1093 }
1094
1095
1096 /**
1097 * Stores original vertex positions in clip coordinates
1098 */
1099 static void
1100 store_clip(struct gallivm_state *gallivm,
1101 const struct lp_type vs_type,
1102 LLVMValueRef io_ptr,
1103 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1104 int idx)
1105 {
1106 LLVMBuilderRef builder = gallivm->builder;
1107 LLVMValueRef soa[4];
1108 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1109 LLVMValueRef indices[2];
1110 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1111 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1112 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1113 LLVMTypeRef clip_ptr_type =
1114 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1115 4), 0);
1116 int i, j;
1117
1118 indices[0] =
1119 indices[1] = lp_build_const_int32(gallivm, 0);
1120
1121 for (i = 0; i < vs_type.length; i++) {
1122 inds[i] = lp_build_const_int32(gallivm, i);
1123 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1124 }
1125
1126 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1127 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1128 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1129 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1130
1131 for (i = 0; i < vs_type.length; i++) {
1132 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1133 }
1134
1135 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1136 for (i = 0; i < vs_type.length; ++i) {
1137 aos[i] = lp_build_extract_range(gallivm,
1138 soa[i % TGSI_NUM_CHANNELS],
1139 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1140 TGSI_NUM_CHANNELS);
1141 }
1142
1143 for (j = 0; j < vs_type.length; j++) {
1144 LLVMValueRef clip_ptr;
1145
1146 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1147 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1148
1149 /* Unaligned store */
1150 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1151 }
1152 }
1153
1154
1155 /**
1156 * Transforms the outputs for viewport mapping
1157 */
1158 static void
1159 generate_viewport(struct draw_llvm_variant *variant,
1160 LLVMBuilderRef builder,
1161 struct lp_type vs_type,
1162 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1163 LLVMValueRef context_ptr)
1164 {
1165 int i;
1166 struct gallivm_state *gallivm = variant->gallivm;
1167 struct lp_type f32_type = vs_type;
1168 const unsigned pos = variant->llvm->draw->vs.position_output;
1169 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1170 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1171 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1172 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1173
1174 /* We treat pipe_viewport_state as a float array */
1175 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1176 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1177
1178 /* for 1/w convention*/
1179 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1180 LLVMBuildStore(builder, out3, outputs[pos][3]);
1181
1182 /* Viewport Mapping */
1183 for (i=0; i<3; i++) {
1184 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1185 LLVMValueRef scale;
1186 LLVMValueRef trans;
1187 LLVMValueRef scale_i;
1188 LLVMValueRef trans_i;
1189 LLVMValueRef index;
1190
1191 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1192 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1193
1194 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1195 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1196
1197 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1198 LLVMBuildLoad(builder, scale_i, "scale"));
1199 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1200 LLVMBuildLoad(builder, trans_i, "trans"));
1201
1202 /* divide by w */
1203 out = LLVMBuildFMul(builder, out, out3, "");
1204 /* mult by scale, add translation */
1205 out = lp_build_fmuladd(builder, out, scale, trans);
1206
1207 /* store transformed outputs */
1208 LLVMBuildStore(builder, out, outputs[pos][i]);
1209 }
1210
1211 }
1212
1213
1214 /**
1215 * Returns clipmask as nxi32 bitmask for the n vertices
1216 */
1217 static LLVMValueRef
1218 generate_clipmask(struct draw_llvm *llvm,
1219 struct gallivm_state *gallivm,
1220 struct lp_type vs_type,
1221 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1222 struct draw_llvm_variant_key *key,
1223 LLVMValueRef context_ptr,
1224 boolean *have_clipdist)
1225 {
1226 LLVMBuilderRef builder = gallivm->builder;
1227 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1228 LLVMValueRef test, temp;
1229 LLVMValueRef zero, shift;
1230 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1231 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1232 LLVMValueRef plane1, planes, plane_ptr, sum;
1233 struct lp_type f32_type = vs_type;
1234 struct lp_type i32_type = lp_int_type(vs_type);
1235 const unsigned pos = llvm->draw->vs.position_output;
1236 const unsigned cv = llvm->draw->vs.clipvertex_output;
1237 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1238 boolean have_cd = false;
1239 boolean clip_user = key->clip_user;
1240 unsigned ucp_enable = key->ucp_enable;
1241 unsigned cd[2];
1242
1243 cd[0] = llvm->draw->vs.ccdistance_output[0];
1244 cd[1] = llvm->draw->vs.ccdistance_output[1];
1245
1246 if (cd[0] != pos || cd[1] != pos)
1247 have_cd = true;
1248
1249 if (num_written_clipdistance && !clip_user) {
1250 clip_user = true;
1251 ucp_enable = (1 << num_written_clipdistance) - 1;
1252 }
1253
1254 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1255 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1256 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1257 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1258
1259 /*
1260 * load clipvertex and position from correct locations.
1261 * if they are the same just load them once.
1262 */
1263 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1264 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1265 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1266 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1267
1268 if (clip_user && cv != pos) {
1269 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1270 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1271 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1272 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1273 } else {
1274 cv_x = pos_x;
1275 cv_y = pos_y;
1276 cv_z = pos_z;
1277 cv_w = pos_w;
1278 }
1279
1280 /*
1281 * Be careful with the comparisons and NaNs (using llvm's unordered
1282 * comparisons here).
1283 */
1284 /* Cliptest, for hardwired planes */
1285 /*
1286 * XXX should take guardband into account (currently not in key).
1287 * Otherwise might run the draw pipeline stages for nothing.
1288 */
1289 if (key->clip_xy) {
1290 /* plane 1 */
1291 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1292 temp = shift;
1293 test = LLVMBuildAnd(builder, test, temp, "");
1294 mask = test;
1295
1296 /* plane 2 */
1297 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1298 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1299 temp = LLVMBuildShl(builder, temp, shift, "");
1300 test = LLVMBuildAnd(builder, test, temp, "");
1301 mask = LLVMBuildOr(builder, mask, test, "");
1302
1303 /* plane 3 */
1304 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1305 temp = LLVMBuildShl(builder, temp, shift, "");
1306 test = LLVMBuildAnd(builder, test, temp, "");
1307 mask = LLVMBuildOr(builder, mask, test, "");
1308
1309 /* plane 4 */
1310 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1311 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1312 temp = LLVMBuildShl(builder, temp, shift, "");
1313 test = LLVMBuildAnd(builder, test, temp, "");
1314 mask = LLVMBuildOr(builder, mask, test, "");
1315 }
1316
1317 if (key->clip_z) {
1318 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1319 if (key->clip_halfz) {
1320 /* plane 5 */
1321 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1322 test = LLVMBuildAnd(builder, test, temp, "");
1323 mask = LLVMBuildOr(builder, mask, test, "");
1324 }
1325 else {
1326 /* plane 5 */
1327 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1328 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1329 test = LLVMBuildAnd(builder, test, temp, "");
1330 mask = LLVMBuildOr(builder, mask, test, "");
1331 }
1332 /* plane 6 */
1333 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1334 temp = LLVMBuildShl(builder, temp, shift, "");
1335 test = LLVMBuildAnd(builder, test, temp, "");
1336 mask = LLVMBuildOr(builder, mask, test, "");
1337 }
1338
1339 if (clip_user) {
1340 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1341 LLVMValueRef indices[3];
1342 LLVMValueRef is_nan_or_inf;
1343
1344 /* userclip planes */
1345 while (ucp_enable) {
1346 unsigned plane_idx = ffs(ucp_enable)-1;
1347 ucp_enable &= ~(1 << plane_idx);
1348 plane_idx += 6;
1349
1350 if (have_cd && num_written_clipdistance) {
1351 LLVMValueRef clipdist;
1352 int i;
1353 i = plane_idx - 6;
1354
1355 *have_clipdist = TRUE;
1356 if (i < 4) {
1357 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1358 } else {
1359 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1360 }
1361 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1362 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1363 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1364 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1365 test = LLVMBuildAnd(builder, test, temp, "");
1366 mask = LLVMBuildOr(builder, mask, test, "");
1367 } else {
1368 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1369 indices[0] = lp_build_const_int32(gallivm, 0);
1370 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1371
1372 indices[2] = lp_build_const_int32(gallivm, 0);
1373 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1374 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1375 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1376 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1377
1378 indices[2] = lp_build_const_int32(gallivm, 1);
1379 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1380 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1381 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1382 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1383
1384 indices[2] = lp_build_const_int32(gallivm, 2);
1385 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1386 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1387 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1388 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1389
1390 indices[2] = lp_build_const_int32(gallivm, 3);
1391 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1392 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1393 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1394 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1395
1396 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1397 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1398 test = LLVMBuildAnd(builder, test, temp, "");
1399 mask = LLVMBuildOr(builder, mask, test, "");
1400 }
1401 }
1402 }
1403 if (key->need_edgeflags) {
1404 /*
1405 * This isn't really part of clipmask but stored the same in vertex
1406 * header later, so do it here.
1407 */
1408 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1409 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1410 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1411 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1412 temp = lp_build_const_int_vec(gallivm, i32_type,
1413 1LL << DRAW_TOTAL_CLIP_PLANES);
1414 test = LLVMBuildAnd(builder, test, temp, "");
1415 mask = LLVMBuildOr(builder, mask, test, "");
1416 }
1417 return mask;
1418 }
1419
1420
1421 /**
1422 * Returns boolean if any clipping has occurred
1423 * Used zero/one i8 value to represent boolean
1424 */
1425 static LLVMValueRef
1426 clipmask_booli8(struct gallivm_state *gallivm,
1427 const struct lp_type vs_type,
1428 LLVMValueRef clipmask_bool_ptr,
1429 boolean edgeflag_in_clipmask)
1430 {
1431 LLVMBuilderRef builder = gallivm->builder;
1432 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1433 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1434 LLVMValueRef ret;
1435 struct lp_build_context bldivec;
1436
1437 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1438
1439 /*
1440 * We need to invert the edgeflag bit from the clipmask here
1441 * (because the result is really if we want to run the pipeline or not
1442 * and we (may) need it if edgeflag was 0).
1443 */
1444 if (edgeflag_in_clipmask) {
1445 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1446 1LL << DRAW_TOTAL_CLIP_PLANES);
1447 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1448 }
1449
1450 /*
1451 * XXX: probably should mask off bits from the mask which come from
1452 * vertices which were beyond the count (i.e. indices_valid for
1453 * linear fetches, for elts ones we don't have the correct mask
1454 * right now). Otherwise might run the pipeline for nothing,
1455 * though everything should still work.
1456 */
1457 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1458 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1459 return ret;
1460 }
1461
1462 static LLVMValueRef
1463 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1464 struct lp_build_tgsi_context * bld_base,
1465 boolean is_vindex_indirect,
1466 LLVMValueRef vertex_index,
1467 boolean is_aindex_indirect,
1468 LLVMValueRef attrib_index,
1469 LLVMValueRef swizzle_index)
1470 {
1471 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1472 struct gallivm_state *gallivm = bld_base->base.gallivm;
1473 LLVMBuilderRef builder = gallivm->builder;
1474 LLVMValueRef indices[3];
1475 LLVMValueRef res;
1476 struct lp_type type = bld_base->base.type;
1477
1478 if (is_vindex_indirect || is_aindex_indirect) {
1479 int i;
1480 res = bld_base->base.zero;
1481 for (i = 0; i < type.length; ++i) {
1482 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1483 LLVMValueRef vert_chan_index = vertex_index;
1484 LLVMValueRef attr_chan_index = attrib_index;
1485 LLVMValueRef channel_vec, value;
1486
1487 if (is_vindex_indirect) {
1488 vert_chan_index = LLVMBuildExtractElement(builder,
1489 vertex_index, idx, "");
1490 }
1491 if (is_aindex_indirect) {
1492 attr_chan_index = LLVMBuildExtractElement(builder,
1493 attrib_index, idx, "");
1494 }
1495
1496 indices[0] = vert_chan_index;
1497 indices[1] = attr_chan_index;
1498 indices[2] = swizzle_index;
1499
1500 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1501 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1502 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1503
1504 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1505 }
1506 } else {
1507 indices[0] = vertex_index;
1508 indices[1] = attrib_index;
1509 indices[2] = swizzle_index;
1510
1511 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1512 res = LLVMBuildLoad(builder, res, "");
1513 }
1514
1515 return res;
1516 }
1517
1518 static void
1519 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1520 struct lp_build_tgsi_context * bld_base,
1521 LLVMValueRef (*outputs)[4],
1522 LLVMValueRef emitted_vertices_vec)
1523 {
1524 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1525 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1526 struct gallivm_state *gallivm = variant->gallivm;
1527 LLVMBuilderRef builder = gallivm->builder;
1528 struct lp_type gs_type = bld_base->base.type;
1529 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1530 lp_int_type(gs_type), 0);
1531 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1532 LLVMValueRef next_prim_offset =
1533 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1534 LLVMValueRef io = variant->io_ptr;
1535 unsigned i;
1536 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1537
1538 for (i = 0; i < gs_type.length; ++i) {
1539 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1540 LLVMValueRef currently_emitted =
1541 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1542 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1543 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1544 }
1545
1546 convert_to_aos(gallivm, io, indices,
1547 outputs, clipmask,
1548 gs_info->num_outputs, gs_type,
1549 FALSE);
1550 }
1551
1552 static void
1553 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1554 struct lp_build_tgsi_context * bld_base,
1555 LLVMValueRef verts_per_prim_vec,
1556 LLVMValueRef emitted_prims_vec)
1557 {
1558 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1559 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1560 struct gallivm_state *gallivm = variant->gallivm;
1561 LLVMBuilderRef builder = gallivm->builder;
1562 LLVMValueRef prim_lengts_ptr =
1563 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1564 unsigned i;
1565
1566 for (i = 0; i < bld_base->base.type.length; ++i) {
1567 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1568 LLVMValueRef prims_emitted =
1569 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1570 LLVMValueRef store_ptr;
1571 LLVMValueRef num_vertices =
1572 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1573
1574 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1575 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1576 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1577 LLVMBuildStore(builder, num_vertices, store_ptr);
1578 }
1579 }
1580
1581 static void
1582 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1583 struct lp_build_tgsi_context * bld_base,
1584 LLVMValueRef total_emitted_vertices_vec,
1585 LLVMValueRef emitted_prims_vec)
1586 {
1587 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1588 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1589 struct gallivm_state *gallivm = variant->gallivm;
1590 LLVMBuilderRef builder = gallivm->builder;
1591 LLVMValueRef emitted_verts_ptr =
1592 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1593 LLVMValueRef emitted_prims_ptr =
1594 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1595 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1596
1597 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1598 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1599
1600 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1601 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1602 }
1603
1604 static void
1605 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1606 {
1607 struct gallivm_state *gallivm = variant->gallivm;
1608 LLVMContextRef context = gallivm->context;
1609 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1610 LLVMTypeRef arg_types[11];
1611 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1612 LLVMTypeRef func_type;
1613 LLVMValueRef context_ptr;
1614 LLVMBasicBlockRef block;
1615 LLVMBuilderRef builder;
1616 char func_name[64];
1617 struct lp_type vs_type;
1618 LLVMValueRef count, fetch_elts, start_or_maxelt;
1619 LLVMValueRef vertex_id_offset, start_instance;
1620 LLVMValueRef stride, step, io_itr;
1621 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1622 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1623 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1624 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1625 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1626 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1627 LLVMValueRef fake_buf_ptr, fake_buf;
1628
1629 struct draw_context *draw = llvm->draw;
1630 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1631 unsigned i, j;
1632 struct lp_build_context bld, blduivec;
1633 struct lp_build_loop_state lp_loop;
1634 struct lp_build_if_state if_ctx;
1635 const int vector_length = lp_native_vector_width / 32;
1636 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1637 struct lp_build_sampler_soa *sampler = 0;
1638 LLVMValueRef ret, clipmask_bool_ptr;
1639 struct draw_llvm_variant_key *key = &variant->key;
1640 /* If geometry shader is present we need to skip both the viewport
1641 * transformation and clipping otherwise the inputs to the geometry
1642 * shader will be incorrect.
1643 * The code can't handle vp transform when vs writes vp index neither
1644 * (though this would be fixable here, but couldn't just broadcast
1645 * the values).
1646 */
1647 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1648 vs_info->writes_viewport_index;
1649 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1650 key->clip_z ||
1651 key->clip_user ||
1652 key->need_edgeflags);
1653 LLVMValueRef variant_func;
1654 const unsigned pos = draw->vs.position_output;
1655 const unsigned cv = draw->vs.clipvertex_output;
1656 boolean have_clipdist = FALSE;
1657 struct lp_bld_tgsi_system_values system_values;
1658
1659 memset(&system_values, 0, sizeof(system_values));
1660
1661 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1662 variant->shader->variants_cached);
1663
1664 i = 0;
1665 arg_types[i++] = get_context_ptr_type(variant); /* context */
1666 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1667 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1668 arg_types[i++] = int32_type; /* count */
1669 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1670 arg_types[i++] = int32_type; /* stride */
1671 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1672 arg_types[i++] = int32_type; /* instance_id */
1673 arg_types[i++] = int32_type; /* vertex_id_offset */
1674 arg_types[i++] = int32_type; /* start_instance */
1675 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1676
1677 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1678 arg_types, num_arg_types, 0);
1679
1680 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1681 variant->function = variant_func;
1682
1683 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1684 for (i = 0; i < num_arg_types; ++i)
1685 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1686 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1687
1688 context_ptr = LLVMGetParam(variant_func, 0);
1689 io_ptr = LLVMGetParam(variant_func, 1);
1690 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1691 count = LLVMGetParam(variant_func, 3);
1692 /*
1693 * XXX: the maxelt part is unused. Not really useful, since we cannot
1694 * get index buffer overflows due to vsplit (which provides its own
1695 * elts buffer, with a different size than what's passed in here).
1696 */
1697 start_or_maxelt = LLVMGetParam(variant_func, 4);
1698 /*
1699 * XXX: stride is actually unused. The stride we use is strictly calculated
1700 * from the number of outputs (including the draw_extra outputs).
1701 * Should probably fix some day (we need a new vs just because of extra
1702 * outputs which the generated vs won't touch).
1703 */
1704 stride = LLVMGetParam(variant_func, 5);
1705 vb_ptr = LLVMGetParam(variant_func, 6);
1706 system_values.instance_id = LLVMGetParam(variant_func, 7);
1707 vertex_id_offset = LLVMGetParam(variant_func, 8);
1708 start_instance = LLVMGetParam(variant_func, 9);
1709 fetch_elts = LLVMGetParam(variant_func, 10);
1710
1711 lp_build_name(context_ptr, "context");
1712 lp_build_name(io_ptr, "io");
1713 lp_build_name(vbuffers_ptr, "vbuffers");
1714 lp_build_name(count, "count");
1715 lp_build_name(start_or_maxelt, "start_or_maxelt");
1716 lp_build_name(stride, "stride");
1717 lp_build_name(vb_ptr, "vb");
1718 lp_build_name(system_values.instance_id, "instance_id");
1719 lp_build_name(vertex_id_offset, "vertex_id_offset");
1720 lp_build_name(start_instance, "start_instance");
1721 lp_build_name(fetch_elts, "fetch_elts");
1722
1723 /*
1724 * Function body
1725 */
1726
1727 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1728 builder = gallivm->builder;
1729 LLVMPositionBuilderAtEnd(builder, block);
1730
1731 memset(&vs_type, 0, sizeof vs_type);
1732 vs_type.floating = TRUE; /* floating point values */
1733 vs_type.sign = TRUE; /* values are signed */
1734 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1735 vs_type.width = 32; /* 32-bit float */
1736 vs_type.length = vector_length;
1737
1738 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1739 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1740
1741 /* hold temporary "bool" clipmask */
1742 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1743
1744 fake_buf = lp_build_alloca_undef(gallivm,
1745 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1746 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1747 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1748 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1749
1750 /* code generated texture sampling */
1751 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1752
1753 step = lp_build_const_int32(gallivm, vector_length);
1754
1755 ind_vec = blduivec.undef;
1756 for (i = 0; i < vs_type.length; i++) {
1757 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1758 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1759 }
1760
1761 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1762 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1763
1764 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1765 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1766 /*
1767 * Only needed for non-indexed path.
1768 */
1769 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1770
1771 /*
1772 * Pre-calculate everything which is constant per shader invocation.
1773 */
1774 for (j = 0; j < key->nr_vertex_elements; ++j) {
1775 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1776 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1777 struct pipe_vertex_element *velem = &key->vertex_element[j];
1778 LLVMValueRef vb_index =
1779 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1780 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1781 util_format_get_blocksize(velem->src_format));
1782 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1783 velem->src_offset);
1784 struct lp_build_if_state if_ctx;
1785
1786 if (velem->src_format != PIPE_FORMAT_NONE) {
1787 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1788 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1789 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1790 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1791 LLVMInt32TypeInContext(context), "");
1792 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1793 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1794 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1795
1796 ofbit = NULL;
1797 /*
1798 * We'll set buffer_size_adj to zero if we have of, so it will
1799 * always overflow later automatically without having to keep ofbit.
1800 * Overflows (with normal wraparound) doing the actual offset
1801 * calculation should be ok, just not for the buffer size calc.
1802 * It would also be possible to detect such overflows and return
1803 * zeros if that happens, but this would be more complex.
1804 */
1805 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1806 tmp = lp_build_sub(&bld, bsize, bld.one);
1807 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1808 &ofbit);
1809 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1810 buf_offset, &ofbit);
1811
1812 /*
1813 * We can't easily set fake vertex buffers outside the generated code.
1814 * Hence, set fake vertex buffers here instead basically, so fetch
1815 * code can always fetch using offset 0, eliminating all control flow
1816 * inside the main loop.
1817 * (Alternatively, could have control flow per vector skipping fetch
1818 * if ofbit is true.)
1819 */
1820 if (velem->instance_divisor) {
1821 /*
1822 * Index is equal to the start instance plus the number of current
1823 * instance divided by the divisor. In this case we compute it as:
1824 * index = start_instance + (instance_id / divisor).
1825 * Note we could actually do the fetch here, outside the loop -
1826 * it's all constant, hopefully llvm recognizes this.
1827 */
1828 LLVMValueRef current_instance;
1829 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1830 lp_build_const_int32(gallivm,
1831 velem->instance_divisor),
1832 "instance_divisor");
1833 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1834 current_instance, &ofbit);
1835 }
1836
1837 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1838 buffer_size_adj[j], "");
1839
1840 temp_ptr = lp_build_alloca_undef(gallivm,
1841 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1842
1843 lp_build_if(&if_ctx, gallivm, ofbit);
1844 {
1845 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1846 }
1847 lp_build_else(&if_ctx);
1848 {
1849 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1850 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1851 }
1852 lp_build_endif(&if_ctx);
1853 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1854
1855 if (0) {
1856 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1857 lp_build_const_int32(gallivm, j),
1858 vb_index, vb_stride[j]);
1859 lp_build_printf(gallivm,
1860 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1861 vb_buffer_offset, src_offset, buf_offset);
1862 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1863 buffer_size, bsize);
1864 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1865 }
1866 }
1867 }
1868
1869 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1870 {
1871 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1872 LLVMValueRef io;
1873 LLVMValueRef clipmask; /* holds the clipmask value */
1874 LLVMValueRef true_index_array, index_store;
1875 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1876
1877 io_itr = lp_loop.counter;
1878
1879 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1880 #if DEBUG_STORE
1881 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1882 io_itr, io, lp_loop.counter);
1883 #endif
1884
1885 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1886 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1887
1888 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
1889 /*
1890 * Limit indices to fetch_max, otherwise might try to access indices
1891 * beyond index buffer (or rather vsplit elt buffer) size.
1892 * Could probably safely (?) skip this for non-indexed draws and
1893 * simplify things minimally (by removing it could combine the ind_vec
1894 * and start_vec adds). I think the only effect for non-indexed draws will
1895 * be that for the invalid elements they will be all fetched from the
1896 * same location as the last valid one, but noone should really care.
1897 */
1898 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1899
1900 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1901
1902 lp_build_if(&if_ctx, gallivm, have_elts);
1903 {
1904 /*
1905 * Note: you'd expect some comparison/clamp against fetch_elt_max
1906 * here.
1907 * There used to be one here but it was incorrect: overflow was
1908 * detected if index > fetch_elt_max - but the correct condition
1909 * would be index >= fetch_elt_max (since this is just size of elts
1910 * buffer / element size).
1911 * Using the correct condition however will cause failures - due to
1912 * vsplit/vcache code which rebases indices. So, as an example, if
1913 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1914 * replace all invalid indices with 0 - which in case of elt_bias
1915 * not being zero will get a different fetch index than the valid
1916 * index 0. So, just rely on vsplit code preventing out-of-bounds
1917 * fetches. This is also why it's safe to do elts fetch even if there
1918 * was no index buffer bound - the real buffer is never seen here, at
1919 * least not if there are index buffer overflows...
1920 */
1921
1922 /*
1923 * XXX should not have to do this, as scale can be handled
1924 * natively by loads (hits asserts though).
1925 */
1926 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1927 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1928 LLVMPointerType(LLVMInt8TypeInContext(context),
1929 0), "");
1930 tmp = lp_build_gather(gallivm, vs_type.length,
1931 32, bld.type, TRUE,
1932 fetch_elts, tmp, FALSE);
1933 LLVMBuildStore(builder, tmp, index_store);
1934 }
1935 lp_build_else(&if_ctx);
1936 {
1937 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1938 LLVMBuildStore(builder, tmp, index_store);
1939 }
1940 lp_build_endif(&if_ctx);
1941
1942 true_index_array = LLVMBuildLoad(builder, index_store, "");
1943
1944 for (j = 0; j < key->nr_vertex_elements; ++j) {
1945 struct pipe_vertex_element *velem = &key->vertex_element[j];
1946 const struct util_format_description *format_desc =
1947 util_format_description(velem->src_format);
1948
1949 if (format_desc->format == PIPE_FORMAT_NONE) {
1950 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1951 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1952 }
1953 }
1954 else if (velem->instance_divisor) {
1955 fetch_instanced(gallivm, format_desc, vs_type,
1956 vb_stride[j], map_ptr[j],
1957 buffer_size_adj[j],
1958 inputs[j], instance_index[j]);
1959 }
1960 else {
1961 fetch_vector(gallivm, format_desc, vs_type,
1962 vb_stride[j], map_ptr[j],
1963 buffer_size_adj[j],
1964 inputs[j], true_index_array);
1965 }
1966 }
1967
1968 struct lp_build_mask_context mask;
1969
1970 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
1971 /* In the paths with elts vertex id has to be unaffected by the
1972 * index bias and because indices inside our elements array have
1973 * already had index bias applied we need to subtract it here to
1974 * get back to the original index.
1975 * in the linear paths vertex id has to be unaffected by the
1976 * original start index and because we abuse the 'start' variable
1977 * to either represent the actual start index or the index at which
1978 * the primitive was split (we split rendering into chunks of at
1979 * most 4095-vertices) we need to back out the original start
1980 * index out of our vertex id here.
1981 */
1982 system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
1983 vertex_id_offset);
1984 system_values.vertex_id = true_index_array;
1985 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1986 system_values.basevertex, "");
1987
1988 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1989 generate_vs(variant,
1990 builder,
1991 vs_type,
1992 outputs,
1993 ptr_aos,
1994 &system_values,
1995 context_ptr,
1996 sampler,
1997 key->clamp_vertex_color,
1998 &mask);
1999
2000 lp_build_mask_end(&mask);
2001 if (pos != -1 && cv != -1) {
2002 /* store original positions in clip before further manipulation */
2003 store_clip(gallivm, vs_type, io, outputs, pos);
2004
2005 /* do cliptest */
2006 if (enable_cliptest) {
2007 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
2008 /* allocate clipmask, assign it integer type */
2009 clipmask = generate_clipmask(llvm,
2010 gallivm,
2011 vs_type,
2012 outputs,
2013 key,
2014 context_ptr, &have_clipdist);
2015 temp = LLVMBuildOr(builder, clipmask, temp, "");
2016 /* store temporary clipping boolean value */
2017 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
2018 }
2019 else {
2020 clipmask = blduivec.zero;
2021 }
2022
2023 /* do viewport mapping */
2024 if (!bypass_viewport) {
2025 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
2026 }
2027 }
2028 else {
2029 clipmask = blduivec.zero;
2030 }
2031
2032 /* store clipmask in vertex header,
2033 * original positions in clip
2034 * and transformed positions in data
2035 */
2036 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
2037 vs_info->num_outputs, vs_type,
2038 enable_cliptest && key->need_edgeflags);
2039 }
2040 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2041
2042 sampler->destroy(sampler);
2043
2044 /* return clipping boolean value for function */
2045 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
2046 enable_cliptest && key->need_edgeflags);
2047
2048 LLVMBuildRet(builder, ret);
2049
2050 gallivm_verify_function(gallivm, variant_func);
2051 }
2052
2053
2054 struct draw_llvm_variant_key *
2055 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2056 {
2057 unsigned i;
2058 struct draw_llvm_variant_key *key;
2059 struct draw_sampler_static_state *draw_sampler;
2060
2061 key = (struct draw_llvm_variant_key *)store;
2062
2063 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2064
2065 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2066
2067 /* will have to rig this up properly later */
2068 key->clip_xy = llvm->draw->clip_xy;
2069 key->clip_z = llvm->draw->clip_z;
2070 key->clip_user = llvm->draw->clip_user;
2071 key->bypass_viewport = llvm->draw->bypass_viewport;
2072 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2073 /* XXX assumes edgeflag output not at 0 */
2074 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2075 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2076 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2077 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2078
2079 /* All variants of this shader will have the same value for
2080 * nr_samplers. Not yet trying to compact away holes in the
2081 * sampler array.
2082 */
2083 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2084 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2085 key->nr_sampler_views =
2086 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2087 }
2088 else {
2089 key->nr_sampler_views = key->nr_samplers;
2090 }
2091
2092 /* Presumably all variants of the shader should have the same
2093 * number of vertex elements - ie the number of shader inputs.
2094 * NOTE: we NEED to store the needed number of needed inputs
2095 * here, not the number of provided elements to match keysize
2096 * (and the offset of sampler state in the key).
2097 * If we have excess number of vertex elements, this is valid,
2098 * but the excess ones don't matter.
2099 * If we don't have enough vertex elements (which looks not really
2100 * valid but we'll handle it gracefully) fill out missing ones with
2101 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2102 */
2103 key->nr_vertex_elements =
2104 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2105
2106 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2107 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2108 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2109 memset(key->vertex_element, 0,
2110 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2111 }
2112 memcpy(key->vertex_element,
2113 llvm->draw->pt.vertex_element,
2114 sizeof(struct pipe_vertex_element) *
2115 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2116
2117 draw_sampler = draw_llvm_variant_key_samplers(key);
2118 memset(draw_sampler, 0,
2119 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2120
2121 for (i = 0 ; i < key->nr_samplers; i++) {
2122 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2123 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2124 }
2125 for (i = 0 ; i < key->nr_sampler_views; i++) {
2126 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2127 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2128 }
2129
2130 return key;
2131 }
2132
2133
2134 void
2135 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2136 {
2137 unsigned i;
2138 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2139
2140 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2141 debug_printf("clip_xy = %u\n", key->clip_xy);
2142 debug_printf("clip_z = %u\n", key->clip_z);
2143 debug_printf("clip_user = %u\n", key->clip_user);
2144 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2145 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2146 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2147 debug_printf("has_gs = %u\n", key->has_gs);
2148 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2149
2150 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2151 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2152 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2153 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2154 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2155 }
2156
2157 for (i = 0 ; i < key->nr_sampler_views; i++) {
2158 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2159 }
2160 }
2161
2162
2163 void
2164 draw_llvm_set_mapped_texture(struct draw_context *draw,
2165 enum pipe_shader_type shader_stage,
2166 unsigned sview_idx,
2167 uint32_t width, uint32_t height, uint32_t depth,
2168 uint32_t first_level, uint32_t last_level,
2169 const void *base_ptr,
2170 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2171 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2172 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2173 {
2174 unsigned j;
2175 struct draw_jit_texture *jit_tex;
2176
2177 assert(shader_stage == PIPE_SHADER_VERTEX ||
2178 shader_stage == PIPE_SHADER_GEOMETRY);
2179
2180 if (shader_stage == PIPE_SHADER_VERTEX) {
2181 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2182
2183 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2184 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2185 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2186
2187 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2188 } else {
2189 assert(0);
2190 return;
2191 }
2192
2193 jit_tex->width = width;
2194 jit_tex->height = height;
2195 jit_tex->depth = depth;
2196 jit_tex->first_level = first_level;
2197 jit_tex->last_level = last_level;
2198 jit_tex->base = base_ptr;
2199
2200 for (j = first_level; j <= last_level; j++) {
2201 jit_tex->mip_offsets[j] = mip_offsets[j];
2202 jit_tex->row_stride[j] = row_stride[j];
2203 jit_tex->img_stride[j] = img_stride[j];
2204 }
2205 }
2206
2207
2208 void
2209 draw_llvm_set_sampler_state(struct draw_context *draw,
2210 enum pipe_shader_type shader_type)
2211 {
2212 unsigned i;
2213
2214 if (shader_type == PIPE_SHADER_VERTEX) {
2215 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2216 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2217
2218 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2219 const struct pipe_sampler_state *s
2220 = draw->samplers[PIPE_SHADER_VERTEX][i];
2221 jit_sam->min_lod = s->min_lod;
2222 jit_sam->max_lod = s->max_lod;
2223 jit_sam->lod_bias = s->lod_bias;
2224 COPY_4V(jit_sam->border_color, s->border_color.f);
2225 }
2226 }
2227 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2228 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2229 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2230
2231 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2232 const struct pipe_sampler_state *s
2233 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2234 jit_sam->min_lod = s->min_lod;
2235 jit_sam->max_lod = s->max_lod;
2236 jit_sam->lod_bias = s->lod_bias;
2237 COPY_4V(jit_sam->border_color, s->border_color.f);
2238 }
2239 }
2240 }
2241 }
2242
2243
2244 void
2245 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2246 {
2247 struct draw_llvm *llvm = variant->llvm;
2248
2249 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2250 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2251 variant->shader->variants_cached, llvm->nr_variants);
2252 }
2253
2254 gallivm_destroy(variant->gallivm);
2255
2256 remove_from_list(&variant->list_item_local);
2257 variant->shader->variants_cached--;
2258 remove_from_list(&variant->list_item_global);
2259 llvm->nr_variants--;
2260 FREE(variant);
2261 }
2262
2263
2264 /**
2265 * Create LLVM types for various structures.
2266 */
2267 static void
2268 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2269 {
2270 struct gallivm_state *gallivm = var->gallivm;
2271 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
2272
2273 texture_type = create_jit_texture_type(gallivm, "texture");
2274 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2275 image_type = create_jit_image_type(gallivm, "image");
2276
2277 context_type = create_gs_jit_context_type(gallivm,
2278 var->shader->base.vector_length,
2279 texture_type, sampler_type,
2280 image_type,
2281 "draw_gs_jit_context");
2282 var->context_ptr_type = LLVMPointerType(context_type, 0);
2283
2284 var->input_array_type = create_gs_jit_input_type(gallivm);
2285 }
2286
2287 static LLVMTypeRef
2288 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2289 {
2290 if (!variant->context_ptr_type)
2291 create_gs_jit_types(variant);
2292 return variant->context_ptr_type;
2293 }
2294
2295 static LLVMValueRef
2296 generate_mask_value(struct draw_gs_llvm_variant *variant,
2297 struct lp_type gs_type)
2298 {
2299 struct gallivm_state *gallivm = variant->gallivm;
2300 LLVMBuilderRef builder = gallivm->builder;
2301 struct lp_type mask_type = lp_int_type(gs_type);
2302 LLVMValueRef num_prims;
2303 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2304 unsigned i;
2305
2306 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2307 variant->num_prims);
2308 for (i = 0; i < gs_type.length; i++) {
2309 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2310 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2311 }
2312 mask_val = lp_build_compare(gallivm, mask_type,
2313 PIPE_FUNC_GREATER, num_prims, mask_val);
2314
2315 return mask_val;
2316 }
2317
2318 static void
2319 draw_gs_llvm_generate(struct draw_llvm *llvm,
2320 struct draw_gs_llvm_variant *variant)
2321 {
2322 struct gallivm_state *gallivm = variant->gallivm;
2323 LLVMContextRef context = gallivm->context;
2324 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2325 LLVMTypeRef arg_types[7];
2326 LLVMTypeRef func_type;
2327 LLVMValueRef variant_func;
2328 LLVMValueRef context_ptr;
2329 LLVMValueRef prim_id_ptr;
2330 LLVMBasicBlockRef block;
2331 LLVMBuilderRef builder;
2332 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2333 struct lp_build_sampler_soa *sampler = 0;
2334 struct lp_build_context bld;
2335 struct lp_bld_tgsi_system_values system_values;
2336 char func_name[64];
2337 struct lp_type gs_type;
2338 unsigned i;
2339 struct draw_gs_llvm_iface gs_iface;
2340 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2341 LLVMValueRef consts_ptr, num_consts_ptr;
2342 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2343 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2344 struct lp_build_mask_context mask;
2345 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2346 unsigned vector_length = variant->shader->base.vector_length;
2347
2348 memset(&system_values, 0, sizeof(system_values));
2349
2350 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2351 variant->shader->variants_cached);
2352
2353 assert(variant->vertex_header_ptr_type);
2354
2355 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2356 arg_types[1] = variant->input_array_type; /* input */
2357 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2358 arg_types[3] = int32_type; /* num_prims */
2359 arg_types[4] = int32_type; /* instance_id */
2360 arg_types[5] = LLVMPointerType(
2361 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2362 arg_types[6] = int32_type;
2363
2364 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2365
2366 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2367
2368 variant->function = variant_func;
2369
2370 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2371
2372 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2373 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2374 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2375
2376 context_ptr = LLVMGetParam(variant_func, 0);
2377 input_array = LLVMGetParam(variant_func, 1);
2378 io_ptr = LLVMGetParam(variant_func, 2);
2379 num_prims = LLVMGetParam(variant_func, 3);
2380 system_values.instance_id = LLVMGetParam(variant_func, 4);
2381 prim_id_ptr = LLVMGetParam(variant_func, 5);
2382 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2383
2384 lp_build_name(context_ptr, "context");
2385 lp_build_name(input_array, "input");
2386 lp_build_name(io_ptr, "io");
2387 lp_build_name(num_prims, "num_prims");
2388 lp_build_name(system_values.instance_id, "instance_id");
2389 lp_build_name(prim_id_ptr, "prim_id_ptr");
2390 lp_build_name(system_values.invocation_id, "invocation_id");
2391
2392 variant->context_ptr = context_ptr;
2393 variant->io_ptr = io_ptr;
2394 variant->num_prims = num_prims;
2395
2396 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2397 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2398 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2399 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2400 gs_iface.input = input_array;
2401 gs_iface.variant = variant;
2402
2403 /*
2404 * Function body
2405 */
2406
2407 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2408 builder = gallivm->builder;
2409 LLVMPositionBuilderAtEnd(builder, block);
2410
2411 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2412
2413 memset(&gs_type, 0, sizeof gs_type);
2414 gs_type.floating = TRUE; /* floating point values */
2415 gs_type.sign = TRUE; /* values are signed */
2416 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2417 gs_type.width = 32; /* 32-bit float */
2418 gs_type.length = vector_length;
2419
2420 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2421 num_consts_ptr =
2422 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2423
2424 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2425 num_ssbos_ptr =
2426 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2427
2428 /* code generated texture sampling */
2429 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2430
2431 mask_val = generate_mask_value(variant, gs_type);
2432 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2433
2434 if (gs_info->uses_primid) {
2435 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2436 }
2437
2438 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2439 tgsi_dump(tokens, 0);
2440 draw_gs_llvm_dump_variant_key(&variant->key);
2441 }
2442
2443 struct lp_build_tgsi_params params;
2444 memset(&params, 0, sizeof(params));
2445
2446 params.type = gs_type;
2447 params.mask = &mask;
2448 params.consts_ptr = consts_ptr;
2449 params.const_sizes_ptr = num_consts_ptr;
2450 params.system_values = &system_values;
2451 params.context_ptr = context_ptr;
2452 params.sampler = sampler;
2453 params.info = &llvm->draw->gs.geometry_shader->info;
2454 params.gs_iface = (const struct lp_build_tgsi_gs_iface *)&gs_iface;
2455 params.ssbo_ptr = ssbos_ptr;
2456 params.ssbo_sizes_ptr = num_ssbos_ptr;
2457
2458 lp_build_tgsi_soa(variant->gallivm,
2459 tokens,
2460 &params,
2461 outputs);
2462
2463 sampler->destroy(sampler);
2464
2465 lp_build_mask_end(&mask);
2466
2467 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2468
2469 gallivm_verify_function(gallivm, variant_func);
2470 }
2471
2472
2473 struct draw_gs_llvm_variant *
2474 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2475 unsigned num_outputs,
2476 const struct draw_gs_llvm_variant_key *key)
2477 {
2478 struct draw_gs_llvm_variant *variant;
2479 struct llvm_geometry_shader *shader =
2480 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2481 LLVMTypeRef vertex_header;
2482 char module_name[64];
2483
2484 variant = MALLOC(sizeof *variant +
2485 shader->variant_key_size -
2486 sizeof variant->key);
2487 if (!variant)
2488 return NULL;
2489
2490 variant->llvm = llvm;
2491 variant->shader = shader;
2492
2493 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2494 variant->shader->variants_cached);
2495
2496 variant->gallivm = gallivm_create(module_name, llvm->context);
2497
2498 create_gs_jit_types(variant);
2499
2500 memcpy(&variant->key, key, shader->variant_key_size);
2501
2502 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2503
2504 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2505
2506 draw_gs_llvm_generate(llvm, variant);
2507
2508 gallivm_compile_module(variant->gallivm);
2509
2510 variant->jit_func = (draw_gs_jit_func)
2511 gallivm_jit_function(variant->gallivm, variant->function);
2512
2513 gallivm_free_ir(variant->gallivm);
2514
2515 variant->list_item_global.base = variant;
2516 variant->list_item_local.base = variant;
2517 /*variant->no = */shader->variants_created++;
2518 variant->list_item_global.base = variant;
2519
2520 return variant;
2521 }
2522
2523 void
2524 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2525 {
2526 struct draw_llvm *llvm = variant->llvm;
2527
2528 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2529 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2530 variant->shader->variants_cached, llvm->nr_gs_variants);
2531 }
2532
2533 gallivm_destroy(variant->gallivm);
2534
2535 remove_from_list(&variant->list_item_local);
2536 variant->shader->variants_cached--;
2537 remove_from_list(&variant->list_item_global);
2538 llvm->nr_gs_variants--;
2539 FREE(variant);
2540 }
2541
2542 struct draw_gs_llvm_variant_key *
2543 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2544 {
2545 unsigned i;
2546 struct draw_gs_llvm_variant_key *key;
2547 struct draw_sampler_static_state *draw_sampler;
2548
2549 key = (struct draw_gs_llvm_variant_key *)store;
2550
2551 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2552
2553 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2554
2555 /* All variants of this shader will have the same value for
2556 * nr_samplers. Not yet trying to compact away holes in the
2557 * sampler array.
2558 */
2559 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2560 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2561 key->nr_sampler_views =
2562 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2563 }
2564 else {
2565 key->nr_sampler_views = key->nr_samplers;
2566 }
2567
2568 draw_sampler = key->samplers;
2569
2570 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2571
2572 for (i = 0 ; i < key->nr_samplers; i++) {
2573 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2574 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2575 }
2576 for (i = 0 ; i < key->nr_sampler_views; i++) {
2577 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2578 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2579 }
2580
2581 return key;
2582 }
2583
2584 void
2585 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2586 {
2587 unsigned i;
2588 struct draw_sampler_static_state *sampler = key->samplers;
2589
2590 for (i = 0 ; i < key->nr_sampler_views; i++) {
2591 debug_printf("sampler[%i].src_format = %s\n", i,
2592 util_format_name(sampler[i].texture_state.format));
2593 }
2594 }