projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm/swr: add stream_id to geom epilogue emit
[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_nir.h"
47 #include "gallivm/lp_bld_printf.h"
48 #include "gallivm/lp_bld_intr.h"
49 #include "gallivm/lp_bld_init.h"
50 #include "gallivm/lp_bld_type.h"
51 #include "gallivm/lp_bld_pack.h"
52 #include "gallivm/lp_bld_format.h"
53
54 #include "tgsi/tgsi_exec.h"
55 #include "tgsi/tgsi_dump.h"
56
57 #include "util/u_math.h"
58 #include "util/u_pointer.h"
59 #include "util/u_string.h"
60 #include "util/simple_list.h"
61
62
63 #define DEBUG_STORE 0
64
65
66 static void
67 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
68
69
70 struct draw_gs_llvm_iface {
71 struct lp_build_gs_iface base;
72
73 struct draw_gs_llvm_variant *variant;
74 LLVMValueRef input;
75 };
76
77 static inline const struct draw_gs_llvm_iface *
78 draw_gs_llvm_iface(const struct lp_build_gs_iface *iface)
79 {
80 return (const struct draw_gs_llvm_iface *)iface;
81 }
82
83 /**
84 * Create LLVM type for draw_vertex_buffer.
85 */
86 static LLVMTypeRef
87 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
88 const char *struct_name)
89 {
90 LLVMTargetDataRef target = gallivm->target;
91 LLVMTypeRef dvbuffer_type;
92 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
93 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
94
95 elem_types[DRAW_JIT_DVBUFFER_MAP] =
96 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
97 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
98
99 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
100 ARRAY_SIZE(elem_types), 0);
101
102 (void) target; /* silence unused var warning for non-debug build */
103 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
104 target, dvbuffer_type,
105 DRAW_JIT_DVBUFFER_MAP);
106 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
107 target, dvbuffer_type,
108 DRAW_JIT_DVBUFFER_SIZE);
109
110 return dvbuffer_type;
111 }
112
113 /**
114 * Create LLVM type for struct draw_jit_texture
115 */
116 static LLVMTypeRef
117 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
118 {
119 LLVMTargetDataRef target = gallivm->target;
120 LLVMTypeRef texture_type;
121 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
122 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
123
124 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
125 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
126 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
127 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
128 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
129 elem_types[DRAW_JIT_TEXTURE_BASE] =
130 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
131 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
132 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
133 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
134 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
135
136 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
137 ARRAY_SIZE(elem_types), 0);
138
139 (void) target; /* silence unused var warning for non-debug build */
140 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
141 target, texture_type,
142 DRAW_JIT_TEXTURE_WIDTH);
143 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
144 target, texture_type,
145 DRAW_JIT_TEXTURE_HEIGHT);
146 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
147 target, texture_type,
148 DRAW_JIT_TEXTURE_DEPTH);
149 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
150 target, texture_type,
151 DRAW_JIT_TEXTURE_BASE);
152 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
153 target, texture_type,
154 DRAW_JIT_TEXTURE_ROW_STRIDE);
155 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
156 target, texture_type,
157 DRAW_JIT_TEXTURE_IMG_STRIDE);
158 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
159 target, texture_type,
160 DRAW_JIT_TEXTURE_FIRST_LEVEL);
161 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
162 target, texture_type,
163 DRAW_JIT_TEXTURE_LAST_LEVEL);
164 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
165 target, texture_type,
166 DRAW_JIT_TEXTURE_MIP_OFFSETS);
167
168 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
169
170 return texture_type;
171 }
172
173
174 /**
175 * Create LLVM type for struct draw_jit_sampler
176 */
177 static LLVMTypeRef
178 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
179 {
180 LLVMTargetDataRef target = gallivm->target;
181 LLVMTypeRef sampler_type;
182 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
183
184 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
185 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
186 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
187 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
188 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
189
190 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
191 ARRAY_SIZE(elem_types), 0);
192
193 (void) target; /* silence unused var warning for non-debug build */
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
195 target, sampler_type,
196 DRAW_JIT_SAMPLER_MIN_LOD);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
198 target, sampler_type,
199 DRAW_JIT_SAMPLER_MAX_LOD);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
201 target, sampler_type,
202 DRAW_JIT_SAMPLER_LOD_BIAS);
203 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
204 target, sampler_type,
205 DRAW_JIT_SAMPLER_BORDER_COLOR);
206
207 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
208
209 return sampler_type;
210 }
211
212 /**
213 * Create LLVM type for struct draw_jit_texture
214 */
215 static LLVMTypeRef
216 create_jit_image_type(struct gallivm_state *gallivm, const char *struct_name)
217 {
218 LLVMTargetDataRef target = gallivm->target;
219 LLVMTypeRef image_type;
220 LLVMTypeRef elem_types[DRAW_JIT_IMAGE_NUM_FIELDS];
221 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
222
223 elem_types[DRAW_JIT_IMAGE_WIDTH] =
224 elem_types[DRAW_JIT_IMAGE_HEIGHT] =
225 elem_types[DRAW_JIT_IMAGE_DEPTH] =
226 elem_types[DRAW_JIT_IMAGE_ROW_STRIDE] =
227 elem_types[DRAW_JIT_IMAGE_IMG_STRIDE] = int32_type;
228 elem_types[DRAW_JIT_IMAGE_BASE] =
229 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
230
231 image_type = LLVMStructTypeInContext(gallivm->context, elem_types,
232 ARRAY_SIZE(elem_types), 0);
233
234 (void) target; /* silence unused var warning for non-debug build */
235 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, width,
236 target, image_type,
237 DRAW_JIT_IMAGE_WIDTH);
238 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, height,
239 target, image_type,
240 DRAW_JIT_IMAGE_HEIGHT);
241 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, depth,
242 target, image_type,
243 DRAW_JIT_IMAGE_DEPTH);
244 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, base,
245 target, image_type,
246 DRAW_JIT_IMAGE_BASE);
247 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, row_stride,
248 target, image_type,
249 DRAW_JIT_IMAGE_ROW_STRIDE);
250 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, img_stride,
251 target, image_type,
252 DRAW_JIT_IMAGE_IMG_STRIDE);
253
254 LP_CHECK_STRUCT_SIZE(struct draw_jit_image, target, image_type);
255
256 return image_type;
257 }
258
259 /**
260 * Create LLVM type for struct draw_jit_context
261 */
262 static LLVMTypeRef
263 create_jit_context_type(struct gallivm_state *gallivm,
264 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
265 LLVMTypeRef image_type,
266 const char *struct_name)
267 {
268 LLVMTargetDataRef target = gallivm->target;
269 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
270 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
271 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
272 LLVMTypeRef context_type;
273
274 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
275 LP_MAX_TGSI_CONST_BUFFERS);
276 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
277 LP_MAX_TGSI_CONST_BUFFERS);
278 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
279 DRAW_TOTAL_CLIP_PLANES), 0);
280 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
281 elem_types[4] = LLVMArrayType(texture_type,
282 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
283 elem_types[5] = LLVMArrayType(sampler_type,
284 PIPE_MAX_SAMPLERS); /* samplers */
285 elem_types[6] = LLVMArrayType(image_type,
286 PIPE_MAX_SHADER_IMAGES); /* images */
287 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
288 LP_MAX_TGSI_SHADER_BUFFERS);
289 elem_types[8] = LLVMArrayType(int_type, /* num_vs_ssbos */
290 LP_MAX_TGSI_SHADER_BUFFERS);
291 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
292 ARRAY_SIZE(elem_types), 0);
293
294 (void) target; /* silence unused var warning for non-debug build */
295 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
296 target, context_type, DRAW_JIT_CTX_CONSTANTS);
297 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
298 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
299 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
300 target, context_type, DRAW_JIT_CTX_PLANES);
301 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
302 target, context_type, DRAW_JIT_CTX_VIEWPORT);
303 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
304 target, context_type,
305 DRAW_JIT_CTX_TEXTURES);
306 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
307 target, context_type,
308 DRAW_JIT_CTX_SAMPLERS);
309 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, images,
310 target, context_type, DRAW_JIT_CTX_IMAGES);
311 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
312 target, context_type, DRAW_JIT_CTX_SSBOS);
313 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
314 target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
315 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
316 target, context_type);
317
318 return context_type;
319 }
320
321
322 /**
323 * Create LLVM type for struct draw_gs_jit_context
324 */
325 static LLVMTypeRef
326 create_gs_jit_context_type(struct gallivm_state *gallivm,
327 unsigned vector_length,
328 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
329 LLVMTypeRef image_type,
330 const char *struct_name)
331 {
332 LLVMTargetDataRef target = gallivm->target;
333 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
334 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
335 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
336 LLVMTypeRef context_type;
337
338 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
339 LP_MAX_TGSI_CONST_BUFFERS);
340 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
341 LP_MAX_TGSI_CONST_BUFFERS);
342 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
343 DRAW_TOTAL_CLIP_PLANES), 0);
344 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
345
346 elem_types[4] = LLVMArrayType(texture_type,
347 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
348 elem_types[5] = LLVMArrayType(sampler_type,
349 PIPE_MAX_SAMPLERS); /* samplers */
350 elem_types[6] = LLVMArrayType(image_type,
351 PIPE_MAX_SHADER_IMAGES); /* images */
352 elem_types[7] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
353 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
354 vector_length), 0);
355 elem_types[9] = LLVMPointerType(LLVMVectorType(int_type,
356 vector_length), 0);
357
358 elem_types[10] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
359 LP_MAX_TGSI_SHADER_BUFFERS);
360 elem_types[11] = LLVMArrayType(int_type, /* num_ssbos */
361 LP_MAX_TGSI_SHADER_BUFFERS);
362
363 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
364 ARRAY_SIZE(elem_types), 0);
365
366 (void) target; /* silence unused var warning for non-debug build */
367 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
368 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
369 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
370 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
371 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
372 target, context_type, DRAW_GS_JIT_CTX_PLANES);
373 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
374 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
375 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
376 target, context_type,
377 DRAW_GS_JIT_CTX_TEXTURES);
378 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
379 target, context_type,
380 DRAW_GS_JIT_CTX_SAMPLERS);
381 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
382 target, context_type,
383 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
384 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
385 target, context_type,
386 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
387 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
388 target, context_type,
389 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
390 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
391 target, context_type, DRAW_GS_JIT_CTX_SSBOS);
392 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
393 target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
394 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, images,
395 target, context_type, DRAW_GS_JIT_CTX_IMAGES);
396 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
397 target, context_type);
398
399 return context_type;
400 }
401
402
403 static LLVMTypeRef
404 create_gs_jit_input_type(struct gallivm_state *gallivm)
405 {
406 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
407 LLVMTypeRef input_array;
408
409 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
410 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
411 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
412 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
413
414 return input_array;
415 }
416
417 /**
418 * Create LLVM type for struct pipe_vertex_buffer
419 */
420 static LLVMTypeRef
421 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
422 const char *struct_name)
423 {
424 LLVMTargetDataRef target = gallivm->target;
425 LLVMTypeRef elem_types[4];
426 LLVMTypeRef vb_type;
427
428 elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
429 elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
430 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
431 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
432
433 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
434 ARRAY_SIZE(elem_types), 0);
435
436 (void) target; /* silence unused var warning for non-debug build */
437 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
438 target, vb_type, 0);
439 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
440 target, vb_type, 1);
441 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
442 target, vb_type, 2);
443 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
444 target, vb_type, 3);
445
446 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
447
448 return vb_type;
449 }
450
451
452 /**
453 * Create LLVM type for struct vertex_header;
454 */
455 static LLVMTypeRef
456 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
457 {
458 LLVMTargetDataRef target = gallivm->target;
459 LLVMTypeRef elem_types[3];
460 LLVMTypeRef vertex_header;
461 char struct_name[24];
462
463 snprintf(struct_name, 23, "vertex_header%d", data_elems);
464
465 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
466 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
467 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
468
469 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
470 ARRAY_SIZE(elem_types), 0);
471
472 /* these are bit-fields and we can't take address of them
473 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
474 target, vertex_header,
475 DRAW_JIT_VERTEX_CLIPMASK);
476 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
477 target, vertex_header,
478 DRAW_JIT_VERTEX_EDGEFLAG);
479 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
480 target, vertex_header,
481 DRAW_JIT_VERTEX_PAD);
482 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
483 target, vertex_header,
484 DRAW_JIT_VERTEX_VERTEX_ID);
485 */
486 (void) target; /* silence unused var warning for non-debug build */
487 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
488 target, vertex_header,
489 DRAW_JIT_VERTEX_CLIP_POS);
490 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
491 target, vertex_header,
492 DRAW_JIT_VERTEX_DATA);
493
494 assert(LLVMABISizeOfType(target, vertex_header) ==
495 offsetof(struct vertex_header, data[data_elems]));
496
497 return vertex_header;
498 }
499
500
501 /**
502 * Create LLVM types for various structures.
503 */
504 static void
505 create_jit_types(struct draw_llvm_variant *variant)
506 {
507 struct gallivm_state *gallivm = variant->gallivm;
508 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
509 vb_type, image_type;
510
511 texture_type = create_jit_texture_type(gallivm, "texture");
512 sampler_type = create_jit_sampler_type(gallivm, "sampler");
513 image_type = create_jit_image_type(gallivm, "image");
514
515 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
516 image_type,
517 "draw_jit_context");
518 variant->context_ptr_type = LLVMPointerType(context_type, 0);
519
520 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
521 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
522
523 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
524 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
525 }
526
527
528 static LLVMTypeRef
529 get_context_ptr_type(struct draw_llvm_variant *variant)
530 {
531 if (!variant->context_ptr_type)
532 create_jit_types(variant);
533 return variant->context_ptr_type;
534 }
535
536
537 static LLVMTypeRef
538 get_buffer_ptr_type(struct draw_llvm_variant *variant)
539 {
540 if (!variant->buffer_ptr_type)
541 create_jit_types(variant);
542 return variant->buffer_ptr_type;
543 }
544
545
546 static LLVMTypeRef
547 get_vb_ptr_type(struct draw_llvm_variant *variant)
548 {
549 if (!variant->vb_ptr_type)
550 create_jit_types(variant);
551 return variant->vb_ptr_type;
552 }
553
554 static LLVMTypeRef
555 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
556 {
557 if (!variant->vertex_header_ptr_type)
558 create_jit_types(variant);
559 return variant->vertex_header_ptr_type;
560 }
561
562
563 /**
564 * Create per-context LLVM info.
565 */
566 struct draw_llvm *
567 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
568 {
569 struct draw_llvm *llvm;
570
571 if (!lp_build_init())
572 return NULL;
573
574 llvm = CALLOC_STRUCT( draw_llvm );
575 if (!llvm)
576 return NULL;
577
578 llvm->draw = draw;
579
580 llvm->context = context;
581 if (!llvm->context) {
582 llvm->context = LLVMContextCreate();
583 llvm->context_owned = true;
584 }
585 if (!llvm->context)
586 goto fail;
587
588 llvm->nr_variants = 0;
589 make_empty_list(&llvm->vs_variants_list);
590
591 llvm->nr_gs_variants = 0;
592 make_empty_list(&llvm->gs_variants_list);
593
594 return llvm;
595
596 fail:
597 draw_llvm_destroy(llvm);
598 return NULL;
599 }
600
601
602 /**
603 * Free per-context LLVM info.
604 */
605 void
606 draw_llvm_destroy(struct draw_llvm *llvm)
607 {
608 if (llvm->context_owned)
609 LLVMContextDispose(llvm->context);
610 llvm->context = NULL;
611
612 /* XXX free other draw_llvm data? */
613 FREE(llvm);
614 }
615
616
617 /**
618 * Create LLVM-generated code for a vertex shader.
619 */
620 struct draw_llvm_variant *
621 draw_llvm_create_variant(struct draw_llvm *llvm,
622 unsigned num_inputs,
623 const struct draw_llvm_variant_key *key)
624 {
625 struct draw_llvm_variant *variant;
626 struct llvm_vertex_shader *shader =
627 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
628 LLVMTypeRef vertex_header;
629 char module_name[64];
630
631 variant = MALLOC(sizeof *variant +
632 shader->variant_key_size -
633 sizeof variant->key);
634 if (!variant)
635 return NULL;
636
637 variant->llvm = llvm;
638 variant->shader = shader;
639
640 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
641 variant->shader->variants_cached);
642
643 variant->gallivm = gallivm_create(module_name, llvm->context);
644
645 create_jit_types(variant);
646
647 memcpy(&variant->key, key, shader->variant_key_size);
648
649 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
650 if (llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_TGSI)
651 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
652 else
653 nir_print_shader(llvm->draw->vs.vertex_shader->state.ir.nir, stderr);
654 draw_llvm_dump_variant_key(&variant->key);
655 }
656
657 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
658
659 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
660
661 draw_llvm_generate(llvm, variant);
662
663 gallivm_compile_module(variant->gallivm);
664
665 variant->jit_func = (draw_jit_vert_func)
666 gallivm_jit_function(variant->gallivm, variant->function);
667
668 gallivm_free_ir(variant->gallivm);
669
670 variant->list_item_global.base = variant;
671 variant->list_item_local.base = variant;
672 /*variant->no = */shader->variants_created++;
673 variant->list_item_global.base = variant;
674
675 return variant;
676 }
677
678
679 static void
680 generate_vs(struct draw_llvm_variant *variant,
681 LLVMBuilderRef builder,
682 struct lp_type vs_type,
683 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
684 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
685 const struct lp_bld_tgsi_system_values *system_values,
686 LLVMValueRef context_ptr,
687 const struct lp_build_sampler_soa *draw_sampler,
688 const struct lp_build_image_soa *draw_image,
689 boolean clamp_vertex_color,
690 struct lp_build_mask_context *bld_mask)
691 {
692 struct draw_llvm *llvm = variant->llvm;
693 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
694 LLVMValueRef consts_ptr =
695 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
696 LLVMValueRef num_consts_ptr =
697 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
698 LLVMValueRef ssbos_ptr =
699 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
700 LLVMValueRef num_ssbos_ptr =
701 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
702
703 struct lp_build_tgsi_params params;
704 memset(&params, 0, sizeof(params));
705
706 params.type = vs_type;
707 params.mask = bld_mask;
708 params.consts_ptr = consts_ptr;
709 params.const_sizes_ptr = num_consts_ptr;
710 params.system_values = system_values;
711 params.inputs = inputs;
712 params.context_ptr = context_ptr;
713 params.sampler = draw_sampler;
714 params.info = &llvm->draw->vs.vertex_shader->info;
715 params.ssbo_ptr = ssbos_ptr;
716 params.ssbo_sizes_ptr = num_ssbos_ptr;
717 params.image = draw_image;
718
719 if (llvm->draw->vs.vertex_shader->state.ir.nir &&
720 llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_NIR)
721 lp_build_nir_soa(variant->gallivm,
722 llvm->draw->vs.vertex_shader->state.ir.nir,
723 &params,
724 outputs);
725 else
726 lp_build_tgsi_soa(variant->gallivm,
727 tokens,
728 &params,
729 outputs);
730
731 {
732 LLVMValueRef out;
733 unsigned chan, attrib;
734 struct lp_build_context bld;
735 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
736 lp_build_context_init(&bld, variant->gallivm, vs_type);
737
738 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
739 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
740 if (outputs[attrib][chan]) {
741 switch (info->output_semantic_name[attrib]) {
742 case TGSI_SEMANTIC_COLOR:
743 case TGSI_SEMANTIC_BCOLOR:
744 if (clamp_vertex_color) {
745 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
746 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
747 LLVMBuildStore(builder, out, outputs[attrib][chan]);
748 }
749 break;
750 }
751 }
752 }
753 }
754 }
755 }
756
757
758 static void
759 fetch_instanced(struct gallivm_state *gallivm,
760 const struct util_format_description *format_desc,
761 struct lp_type vs_type,
762 LLVMValueRef vb_stride,
763 LLVMValueRef map_ptr,
764 LLVMValueRef buffer_size_adj,
765 LLVMValueRef *inputs,
766 LLVMValueRef index)
767 {
768 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
769 LLVMTypeRef aosf_t, aosi_t;
770 LLVMValueRef zero = LLVMConstNull(i32_t);
771 LLVMBuilderRef builder = gallivm->builder;
772 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
773 unsigned i;
774
775 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
776 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
777
778 /* This mul can overflow. Wraparound is ok. */
779 stride = LLVMBuildMul(builder, vb_stride, index, "");
780
781 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
782 stride, buffer_size_adj,
783 "buffer_overflowed");
784
785 if (0) {
786 lp_build_print_value(gallivm, " instance index = ", index);
787 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
788 }
789
790 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
791 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
792 stride = LLVMBuildAnd(builder, stride, index_valid, "");
793
794 aos = lp_build_fetch_rgba_aos(gallivm,
795 format_desc,
796 lp_float32_vec4_type(),
797 FALSE,
798 map_ptr,
799 stride, zero, zero,
800 NULL);
801
802 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
803 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
804 aos = LLVMBuildAnd(builder, aos, index_valid, "");
805 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
806
807 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
808 LLVMValueRef index = lp_build_const_int32(gallivm, i);
809 inputs[i] = lp_build_extract_broadcast(gallivm,
810 lp_float32_vec4_type(),
811 vs_type, aos, index);
812 }
813 }
814
815
816 static void
817 fetch_vector(struct gallivm_state *gallivm,
818 const struct util_format_description *format_desc,
819 struct lp_type vs_type,
820 LLVMValueRef vb_stride,
821 LLVMValueRef map_ptr,
822 LLVMValueRef buffer_size_adj,
823 LLVMValueRef *inputs,
824 LLVMValueRef indices)
825 {
826 LLVMBuilderRef builder = gallivm->builder;
827 struct lp_build_context blduivec;
828 struct lp_type fetch_type = vs_type;
829 LLVMValueRef offset, valid_mask;
830 unsigned i;
831
832 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
833
834 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
835 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
836
837 /* This mul can overflow. Wraparound is ok. */
838 offset = lp_build_mul(&blduivec, vb_stride, indices);
839
840 valid_mask = lp_build_compare(gallivm, blduivec.type,
841 PIPE_FUNC_LESS, offset, buffer_size_adj);
842
843 /* not valid elements use offset 0 */
844 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
845
846 if (0) {
847 lp_build_print_value(gallivm, " indices = ", indices);
848 lp_build_print_value(gallivm, " offsets = ", offset);
849 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
850 }
851
852 /*
853 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
854 * This should always produce better code.
855 */
856
857 /* The type handling is annoying here... */
858 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
859 format_desc->channel[0].pure_integer) {
860 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
861 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
862 }
863 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
864 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
865 }
866 }
867
868 lp_build_fetch_rgba_soa(gallivm, format_desc,
869 fetch_type, FALSE, map_ptr, offset,
870 blduivec.zero, blduivec.zero,
871 NULL, inputs);
872
873 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
874 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
875 lp_build_vec_type(gallivm, vs_type), "");
876 }
877
878 /* out-of-bound fetches return all zeros */
879 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
880 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
881 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
882 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
883 lp_build_vec_type(gallivm, vs_type), "");
884 }
885 }
886
887
888 static void
889 store_aos(struct gallivm_state *gallivm,
890 LLVMValueRef io_ptr,
891 LLVMValueRef index,
892 LLVMValueRef value)
893 {
894 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
895 LLVMBuilderRef builder = gallivm->builder;
896 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
897 LLVMValueRef indices[3];
898
899 indices[0] = lp_build_const_int32(gallivm, 0);
900 indices[1] = index;
901 indices[2] = lp_build_const_int32(gallivm, 0);
902
903 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
904 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
905
906 #if DEBUG_STORE
907 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
908 #endif
909
910 /* Unaligned store due to the vertex header */
911 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
912 }
913
914 /**
915 * Adjust the mask to architecture endianess. The mask will the store in struct:
916 *
917 * struct vertex_header {
918 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
919 * unsigned edgeflag:1;
920 * unsigned pad:1;
921 * unsigned vertex_id:16;
922 * [...]
923 * }
924 *
925 * On little-endian machine nothing needs to done, however on bit-endian machine
926 * the mask's fields need to be adjusted with the algorithm:
927 *
928 * uint32_t reverse (uint32_t x)
929 * {
930 * return (x >> 16) | // vertex_id
931 * ((x & 0x3fff) << 18) | // clipmask
932 * ((x & 0x4000) << 3) | // pad
933 * ((x & 0x8000) << 1); // edgeflag
934 * }
935 */
936 static LLVMValueRef
937 adjust_mask(struct gallivm_state *gallivm,
938 LLVMValueRef mask)
939 {
940 #if UTIL_ARCH_BIG_ENDIAN
941 LLVMBuilderRef builder = gallivm->builder;
942 LLVMValueRef vertex_id;
943 LLVMValueRef clipmask;
944 LLVMValueRef pad;
945 LLVMValueRef edgeflag;
946
947 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
948 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
949 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
950 if (0) {
951 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
952 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
953 }
954 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
955 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
956
957 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
958 if (0) {
959 mask = LLVMBuildOr(builder, mask, pad, "");
960 }
961 mask = LLVMBuildOr(builder, mask, edgeflag, "");
962 #endif
963 return mask;
964 }
965
966 static void
967 store_aos_array(struct gallivm_state *gallivm,
968 struct lp_type soa_type,
969 LLVMValueRef io_ptr,
970 LLVMValueRef *indices,
971 LLVMValueRef* aos,
972 int attrib,
973 int num_outputs,
974 LLVMValueRef clipmask,
975 boolean need_edgeflag)
976 {
977 LLVMBuilderRef builder = gallivm->builder;
978 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
979 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
980 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
981 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
982 int vector_length = soa_type.length;
983 int i;
984
985 debug_assert(TGSI_NUM_CHANNELS == 4);
986
987 for (i = 0; i < vector_length; i++) {
988 linear_inds[i] = lp_build_const_int32(gallivm, i);
989 if (indices) {
990 inds[i] = indices[i];
991 } else {
992 inds[i] = linear_inds[i];
993 }
994 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
995 }
996
997 if (attrib == 0) {
998 /* store vertex header for each of the n vertices */
999 LLVMValueRef val, cliptmp;
1000 int vertex_id_pad_edgeflag;
1001
1002 /* If this assertion fails, it means we need to update the bit twidding
1003 * code here. See struct vertex_header in draw_private.h.
1004 */
1005 assert(DRAW_TOTAL_CLIP_PLANES==14);
1006 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1007 if (!need_edgeflag) {
1008 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
1009 }
1010 else {
1011 vertex_id_pad_edgeflag = (0xffff << 16);
1012 }
1013 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
1014 vertex_id_pad_edgeflag);
1015 /* OR with the clipmask */
1016 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
1017 for (i = 0; i < vector_length; i++) {
1018 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
1019 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
1020 val = adjust_mask(gallivm, val);
1021 #if DEBUG_STORE
1022 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
1023 io_ptrs[i], inds[i], val);
1024 #endif
1025 LLVMBuildStore(builder, val, id_ptr);
1026 }
1027 }
1028
1029 /* store for each of the n vertices */
1030 for (i = 0; i < vector_length; i++) {
1031 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
1032 }
1033 }
1034
1035
1036 static void
1037 convert_to_aos(struct gallivm_state *gallivm,
1038 LLVMValueRef io,
1039 LLVMValueRef *indices,
1040 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1041 LLVMValueRef clipmask,
1042 int num_outputs,
1043 struct lp_type soa_type,
1044 boolean need_edgeflag)
1045 {
1046 LLVMBuilderRef builder = gallivm->builder;
1047 unsigned chan, attrib, i;
1048
1049 #if DEBUG_STORE
1050 lp_build_printf(gallivm, " # storing begin\n");
1051 #endif
1052 for (attrib = 0; attrib < num_outputs; ++attrib) {
1053 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1054 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1055 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1056 if (outputs[attrib][chan]) {
1057 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1058 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1059 #if DEBUG_STORE
1060 lp_build_printf(gallivm, "output %d : %d ",
1061 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1062 attrib, 0),
1063 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1064 chan, 0));
1065 lp_build_print_value(gallivm, "val = ", out);
1066 {
1067 LLVMValueRef iv =
1068 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1069
1070 lp_build_print_value(gallivm, " ival = ", iv);
1071 }
1072 #endif
1073 soa[chan] = out;
1074 }
1075 else {
1076 soa[chan] = 0;
1077 }
1078 }
1079
1080
1081 if (soa_type.length == TGSI_NUM_CHANNELS) {
1082 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1083 } else {
1084 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1085
1086 for (i = 0; i < soa_type.length; ++i) {
1087 aos[i] = lp_build_extract_range(gallivm,
1088 soa[i % TGSI_NUM_CHANNELS],
1089 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1090 TGSI_NUM_CHANNELS);
1091 }
1092 }
1093
1094 store_aos_array(gallivm,
1095 soa_type,
1096 io, indices,
1097 aos,
1098 attrib,
1099 num_outputs,
1100 clipmask,
1101 need_edgeflag);
1102 }
1103 #if DEBUG_STORE
1104 lp_build_printf(gallivm, " # storing end\n");
1105 #endif
1106 }
1107
1108
1109 /**
1110 * Stores original vertex positions in clip coordinates
1111 */
1112 static void
1113 store_clip(struct gallivm_state *gallivm,
1114 const struct lp_type vs_type,
1115 LLVMValueRef io_ptr,
1116 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1117 int idx)
1118 {
1119 LLVMBuilderRef builder = gallivm->builder;
1120 LLVMValueRef soa[4];
1121 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1122 LLVMValueRef indices[2];
1123 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1124 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1125 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1126 LLVMTypeRef clip_ptr_type =
1127 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1128 4), 0);
1129 int i, j;
1130
1131 indices[0] =
1132 indices[1] = lp_build_const_int32(gallivm, 0);
1133
1134 for (i = 0; i < vs_type.length; i++) {
1135 inds[i] = lp_build_const_int32(gallivm, i);
1136 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1137 }
1138
1139 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1140 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1141 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1142 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1143
1144 for (i = 0; i < vs_type.length; i++) {
1145 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1146 }
1147
1148 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1149 for (i = 0; i < vs_type.length; ++i) {
1150 aos[i] = lp_build_extract_range(gallivm,
1151 soa[i % TGSI_NUM_CHANNELS],
1152 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1153 TGSI_NUM_CHANNELS);
1154 }
1155
1156 for (j = 0; j < vs_type.length; j++) {
1157 LLVMValueRef clip_ptr;
1158
1159 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1160 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1161
1162 /* Unaligned store */
1163 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1164 }
1165 }
1166
1167
1168 /**
1169 * Transforms the outputs for viewport mapping
1170 */
1171 static void
1172 generate_viewport(struct draw_llvm_variant *variant,
1173 LLVMBuilderRef builder,
1174 struct lp_type vs_type,
1175 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1176 LLVMValueRef context_ptr)
1177 {
1178 int i;
1179 struct gallivm_state *gallivm = variant->gallivm;
1180 struct lp_type f32_type = vs_type;
1181 const unsigned pos = variant->llvm->draw->vs.position_output;
1182 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1183 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1184 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1185 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1186
1187 /* We treat pipe_viewport_state as a float array */
1188 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1189 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1190
1191 /* for 1/w convention*/
1192 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1193 LLVMBuildStore(builder, out3, outputs[pos][3]);
1194
1195 /* Viewport Mapping */
1196 for (i=0; i<3; i++) {
1197 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1198 LLVMValueRef scale;
1199 LLVMValueRef trans;
1200 LLVMValueRef scale_i;
1201 LLVMValueRef trans_i;
1202 LLVMValueRef index;
1203
1204 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1205 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1206
1207 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1208 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1209
1210 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1211 LLVMBuildLoad(builder, scale_i, "scale"));
1212 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1213 LLVMBuildLoad(builder, trans_i, "trans"));
1214
1215 /* divide by w */
1216 out = LLVMBuildFMul(builder, out, out3, "");
1217 /* mult by scale, add translation */
1218 out = lp_build_fmuladd(builder, out, scale, trans);
1219
1220 /* store transformed outputs */
1221 LLVMBuildStore(builder, out, outputs[pos][i]);
1222 }
1223
1224 }
1225
1226
1227 /**
1228 * Returns clipmask as nxi32 bitmask for the n vertices
1229 */
1230 static LLVMValueRef
1231 generate_clipmask(struct draw_llvm *llvm,
1232 struct gallivm_state *gallivm,
1233 struct lp_type vs_type,
1234 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1235 struct draw_llvm_variant_key *key,
1236 LLVMValueRef context_ptr,
1237 boolean *have_clipdist)
1238 {
1239 LLVMBuilderRef builder = gallivm->builder;
1240 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1241 LLVMValueRef test, temp;
1242 LLVMValueRef zero, shift;
1243 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1244 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1245 LLVMValueRef plane1, planes, plane_ptr, sum;
1246 struct lp_type f32_type = vs_type;
1247 struct lp_type i32_type = lp_int_type(vs_type);
1248 const unsigned pos = llvm->draw->vs.position_output;
1249 const unsigned cv = llvm->draw->vs.clipvertex_output;
1250 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1251 boolean have_cd = false;
1252 boolean clip_user = key->clip_user;
1253 unsigned ucp_enable = key->ucp_enable;
1254 unsigned cd[2];
1255
1256 cd[0] = llvm->draw->vs.ccdistance_output[0];
1257 cd[1] = llvm->draw->vs.ccdistance_output[1];
1258
1259 if (cd[0] != pos || cd[1] != pos)
1260 have_cd = true;
1261
1262 if (num_written_clipdistance && !clip_user) {
1263 clip_user = true;
1264 ucp_enable = (1 << num_written_clipdistance) - 1;
1265 }
1266
1267 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1268 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1269 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1270 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1271
1272 /*
1273 * load clipvertex and position from correct locations.
1274 * if they are the same just load them once.
1275 */
1276 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1277 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1278 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1279 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1280
1281 if (clip_user && cv != pos) {
1282 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1283 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1284 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1285 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1286 } else {
1287 cv_x = pos_x;
1288 cv_y = pos_y;
1289 cv_z = pos_z;
1290 cv_w = pos_w;
1291 }
1292
1293 /*
1294 * Be careful with the comparisons and NaNs (using llvm's unordered
1295 * comparisons here).
1296 */
1297 /* Cliptest, for hardwired planes */
1298 /*
1299 * XXX should take guardband into account (currently not in key).
1300 * Otherwise might run the draw pipeline stages for nothing.
1301 */
1302 if (key->clip_xy) {
1303 /* plane 1 */
1304 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1305 temp = shift;
1306 test = LLVMBuildAnd(builder, test, temp, "");
1307 mask = test;
1308
1309 /* plane 2 */
1310 test = LLVMBuildFAdd(builder, pos_x, 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 /* plane 3 */
1317 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1318 temp = LLVMBuildShl(builder, temp, shift, "");
1319 test = LLVMBuildAnd(builder, test, temp, "");
1320 mask = LLVMBuildOr(builder, mask, test, "");
1321
1322 /* plane 4 */
1323 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1324 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1325 temp = LLVMBuildShl(builder, temp, shift, "");
1326 test = LLVMBuildAnd(builder, test, temp, "");
1327 mask = LLVMBuildOr(builder, mask, test, "");
1328 }
1329
1330 if (key->clip_z) {
1331 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1332 if (key->clip_halfz) {
1333 /* plane 5 */
1334 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1335 test = LLVMBuildAnd(builder, test, temp, "");
1336 mask = LLVMBuildOr(builder, mask, test, "");
1337 }
1338 else {
1339 /* plane 5 */
1340 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1341 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1342 test = LLVMBuildAnd(builder, test, temp, "");
1343 mask = LLVMBuildOr(builder, mask, test, "");
1344 }
1345 /* plane 6 */
1346 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1347 temp = LLVMBuildShl(builder, temp, shift, "");
1348 test = LLVMBuildAnd(builder, test, temp, "");
1349 mask = LLVMBuildOr(builder, mask, test, "");
1350 }
1351
1352 if (clip_user) {
1353 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1354 LLVMValueRef indices[3];
1355 LLVMValueRef is_nan_or_inf;
1356
1357 /* userclip planes */
1358 while (ucp_enable) {
1359 unsigned plane_idx = ffs(ucp_enable)-1;
1360 ucp_enable &= ~(1 << plane_idx);
1361 plane_idx += 6;
1362
1363 if (have_cd && num_written_clipdistance) {
1364 LLVMValueRef clipdist;
1365 int i;
1366 i = plane_idx - 6;
1367
1368 *have_clipdist = TRUE;
1369 if (i < 4) {
1370 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1371 } else {
1372 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1373 }
1374 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1375 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1376 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1377 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1378 test = LLVMBuildAnd(builder, test, temp, "");
1379 mask = LLVMBuildOr(builder, mask, test, "");
1380 } else {
1381 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1382 indices[0] = lp_build_const_int32(gallivm, 0);
1383 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1384
1385 indices[2] = lp_build_const_int32(gallivm, 0);
1386 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1387 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1388 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1389 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1390
1391 indices[2] = lp_build_const_int32(gallivm, 1);
1392 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1393 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1394 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1395 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1396
1397 indices[2] = lp_build_const_int32(gallivm, 2);
1398 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1399 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1400 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1401 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1402
1403 indices[2] = lp_build_const_int32(gallivm, 3);
1404 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1405 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1406 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1407 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1408
1409 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1410 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1411 test = LLVMBuildAnd(builder, test, temp, "");
1412 mask = LLVMBuildOr(builder, mask, test, "");
1413 }
1414 }
1415 }
1416 if (key->need_edgeflags) {
1417 /*
1418 * This isn't really part of clipmask but stored the same in vertex
1419 * header later, so do it here.
1420 */
1421 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1422 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1423 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1424 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1425 temp = lp_build_const_int_vec(gallivm, i32_type,
1426 1LL << DRAW_TOTAL_CLIP_PLANES);
1427 test = LLVMBuildAnd(builder, test, temp, "");
1428 mask = LLVMBuildOr(builder, mask, test, "");
1429 }
1430 return mask;
1431 }
1432
1433
1434 /**
1435 * Returns boolean if any clipping has occurred
1436 * Used zero/one i8 value to represent boolean
1437 */
1438 static LLVMValueRef
1439 clipmask_booli8(struct gallivm_state *gallivm,
1440 const struct lp_type vs_type,
1441 LLVMValueRef clipmask_bool_ptr,
1442 boolean edgeflag_in_clipmask)
1443 {
1444 LLVMBuilderRef builder = gallivm->builder;
1445 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1446 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1447 LLVMValueRef ret;
1448 struct lp_build_context bldivec;
1449
1450 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1451
1452 /*
1453 * We need to invert the edgeflag bit from the clipmask here
1454 * (because the result is really if we want to run the pipeline or not
1455 * and we (may) need it if edgeflag was 0).
1456 */
1457 if (edgeflag_in_clipmask) {
1458 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1459 1LL << DRAW_TOTAL_CLIP_PLANES);
1460 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1461 }
1462
1463 /*
1464 * XXX: probably should mask off bits from the mask which come from
1465 * vertices which were beyond the count (i.e. indices_valid for
1466 * linear fetches, for elts ones we don't have the correct mask
1467 * right now). Otherwise might run the pipeline for nothing,
1468 * though everything should still work.
1469 */
1470 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1471 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1472 return ret;
1473 }
1474
1475 static LLVMValueRef
1476 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface *gs_iface,
1477 struct lp_build_context * bld,
1478 boolean is_vindex_indirect,
1479 LLVMValueRef vertex_index,
1480 boolean is_aindex_indirect,
1481 LLVMValueRef attrib_index,
1482 LLVMValueRef swizzle_index)
1483 {
1484 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1485 struct gallivm_state *gallivm = bld->gallivm;
1486 LLVMBuilderRef builder = gallivm->builder;
1487 LLVMValueRef indices[3];
1488 LLVMValueRef res;
1489 struct lp_type type = bld->type;
1490
1491 if (is_vindex_indirect || is_aindex_indirect) {
1492 int i;
1493 res = bld->zero;
1494 for (i = 0; i < type.length; ++i) {
1495 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1496 LLVMValueRef vert_chan_index = vertex_index;
1497 LLVMValueRef attr_chan_index = attrib_index;
1498 LLVMValueRef channel_vec, value;
1499
1500 if (is_vindex_indirect) {
1501 vert_chan_index = LLVMBuildExtractElement(builder,
1502 vertex_index, idx, "");
1503 }
1504 if (is_aindex_indirect) {
1505 attr_chan_index = LLVMBuildExtractElement(builder,
1506 attrib_index, idx, "");
1507 }
1508
1509 indices[0] = vert_chan_index;
1510 indices[1] = attr_chan_index;
1511 indices[2] = swizzle_index;
1512
1513 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1514 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1515 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1516
1517 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1518 }
1519 } else {
1520 indices[0] = vertex_index;
1521 indices[1] = attrib_index;
1522 indices[2] = swizzle_index;
1523
1524 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1525 res = LLVMBuildLoad(builder, res, "");
1526 }
1527
1528 return res;
1529 }
1530
1531 static void
1532 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface *gs_base,
1533 struct lp_build_context * bld,
1534 LLVMValueRef (*outputs)[4],
1535 LLVMValueRef emitted_vertices_vec,
1536 LLVMValueRef stream_id)
1537 {
1538 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1539 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1540 struct gallivm_state *gallivm = variant->gallivm;
1541 LLVMBuilderRef builder = gallivm->builder;
1542 struct lp_type gs_type = bld->type;
1543 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1544 lp_int_type(gs_type), 0);
1545 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1546 LLVMValueRef next_prim_offset =
1547 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1548 LLVMValueRef io = variant->io_ptr;
1549 unsigned i;
1550 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1551
1552 for (i = 0; i < gs_type.length; ++i) {
1553 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1554 LLVMValueRef currently_emitted =
1555 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1556 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1557 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1558 }
1559
1560 convert_to_aos(gallivm, io, indices,
1561 outputs, clipmask,
1562 gs_info->num_outputs, gs_type,
1563 FALSE);
1564 }
1565
1566 static void
1567 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface *gs_base,
1568 struct lp_build_context * bld,
1569 LLVMValueRef total_emitted_vertices_vec_ptr,
1570 LLVMValueRef verts_per_prim_vec,
1571 LLVMValueRef emitted_prims_vec,
1572 LLVMValueRef mask_vec)
1573 {
1574 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1575 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1576 struct gallivm_state *gallivm = variant->gallivm;
1577 LLVMBuilderRef builder = gallivm->builder;
1578 LLVMValueRef prim_lengts_ptr =
1579 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1580 unsigned i;
1581
1582 for (i = 0; i < bld->type.length; ++i) {
1583 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1584 LLVMValueRef prims_emitted =
1585 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1586 LLVMValueRef store_ptr;
1587 LLVMValueRef num_vertices =
1588 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1589
1590 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1591 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1592 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1593 LLVMBuildStore(builder, num_vertices, store_ptr);
1594 }
1595 }
1596
1597 static void
1598 draw_gs_llvm_epilogue(const struct lp_build_gs_iface *gs_base,
1599 LLVMValueRef total_emitted_vertices_vec,
1600 LLVMValueRef emitted_prims_vec, unsigned stream)
1601 {
1602 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1603 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1604 struct gallivm_state *gallivm = variant->gallivm;
1605 LLVMBuilderRef builder = gallivm->builder;
1606 LLVMValueRef emitted_verts_ptr =
1607 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1608 LLVMValueRef emitted_prims_ptr =
1609 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1610 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1611
1612 if (stream > 0)
1613 return;
1614
1615 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1616 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1617
1618 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1619 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1620 }
1621
1622 static void
1623 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1624 {
1625 struct gallivm_state *gallivm = variant->gallivm;
1626 LLVMContextRef context = gallivm->context;
1627 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1628 LLVMTypeRef arg_types[12];
1629 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1630 LLVMTypeRef func_type;
1631 LLVMValueRef context_ptr;
1632 LLVMBasicBlockRef block;
1633 LLVMBuilderRef builder;
1634 char func_name[64];
1635 struct lp_type vs_type;
1636 LLVMValueRef count, fetch_elts, start_or_maxelt;
1637 LLVMValueRef vertex_id_offset;
1638 LLVMValueRef stride, step, io_itr;
1639 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1640 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1641 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1642 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1643 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1644 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1645 LLVMValueRef fake_buf_ptr, fake_buf;
1646
1647 struct draw_context *draw = llvm->draw;
1648 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1649 unsigned i, j;
1650 struct lp_build_context bld, blduivec;
1651 struct lp_build_loop_state lp_loop;
1652 struct lp_build_if_state if_ctx;
1653 const int vector_length = lp_native_vector_width / 32;
1654 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1655 struct lp_build_sampler_soa *sampler = 0;
1656 struct lp_build_image_soa *image = NULL;
1657 LLVMValueRef ret, clipmask_bool_ptr;
1658 struct draw_llvm_variant_key *key = &variant->key;
1659 /* If geometry shader is present we need to skip both the viewport
1660 * transformation and clipping otherwise the inputs to the geometry
1661 * shader will be incorrect.
1662 * The code can't handle vp transform when vs writes vp index neither
1663 * (though this would be fixable here, but couldn't just broadcast
1664 * the values).
1665 */
1666 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1667 vs_info->writes_viewport_index;
1668 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1669 key->clip_z ||
1670 key->clip_user ||
1671 key->need_edgeflags);
1672 LLVMValueRef variant_func;
1673 const unsigned pos = draw->vs.position_output;
1674 const unsigned cv = draw->vs.clipvertex_output;
1675 boolean have_clipdist = FALSE;
1676 struct lp_bld_tgsi_system_values system_values;
1677
1678 memset(&system_values, 0, sizeof(system_values));
1679 memset(&outputs, 0, sizeof(outputs));
1680 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1681 variant->shader->variants_cached);
1682
1683 i = 0;
1684 arg_types[i++] = get_context_ptr_type(variant); /* context */
1685 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1686 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1687 arg_types[i++] = int32_type; /* count */
1688 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1689 arg_types[i++] = int32_type; /* stride */
1690 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1691 arg_types[i++] = int32_type; /* instance_id */
1692 arg_types[i++] = int32_type; /* vertex_id_offset */
1693 arg_types[i++] = int32_type; /* start_instance */
1694 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1695 arg_types[i++] = int32_type; /* draw_id */
1696
1697 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1698 arg_types, num_arg_types, 0);
1699
1700 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1701 variant->function = variant_func;
1702
1703 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1704 for (i = 0; i < num_arg_types; ++i)
1705 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1706 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1707
1708 context_ptr = LLVMGetParam(variant_func, 0);
1709 io_ptr = LLVMGetParam(variant_func, 1);
1710 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1711 count = LLVMGetParam(variant_func, 3);
1712 /*
1713 * XXX: the maxelt part is unused. Not really useful, since we cannot
1714 * get index buffer overflows due to vsplit (which provides its own
1715 * elts buffer, with a different size than what's passed in here).
1716 */
1717 start_or_maxelt = LLVMGetParam(variant_func, 4);
1718 /*
1719 * XXX: stride is actually unused. The stride we use is strictly calculated
1720 * from the number of outputs (including the draw_extra outputs).
1721 * Should probably fix some day (we need a new vs just because of extra
1722 * outputs which the generated vs won't touch).
1723 */
1724 stride = LLVMGetParam(variant_func, 5);
1725 vb_ptr = LLVMGetParam(variant_func, 6);
1726 system_values.instance_id = LLVMGetParam(variant_func, 7);
1727 vertex_id_offset = LLVMGetParam(variant_func, 8);
1728 system_values.base_instance = LLVMGetParam(variant_func, 9);
1729 fetch_elts = LLVMGetParam(variant_func, 10);
1730 system_values.draw_id = LLVMGetParam(variant_func, 11);
1731
1732 lp_build_name(context_ptr, "context");
1733 lp_build_name(io_ptr, "io");
1734 lp_build_name(vbuffers_ptr, "vbuffers");
1735 lp_build_name(count, "count");
1736 lp_build_name(start_or_maxelt, "start_or_maxelt");
1737 lp_build_name(stride, "stride");
1738 lp_build_name(vb_ptr, "vb");
1739 lp_build_name(system_values.instance_id, "instance_id");
1740 lp_build_name(vertex_id_offset, "vertex_id_offset");
1741 lp_build_name(system_values.base_instance, "start_instance");
1742 lp_build_name(fetch_elts, "fetch_elts");
1743 lp_build_name(system_values.draw_id, "draw_id");
1744
1745 /*
1746 * Function body
1747 */
1748
1749 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1750 builder = gallivm->builder;
1751 LLVMPositionBuilderAtEnd(builder, block);
1752
1753 memset(&vs_type, 0, sizeof vs_type);
1754 vs_type.floating = TRUE; /* floating point values */
1755 vs_type.sign = TRUE; /* values are signed */
1756 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1757 vs_type.width = 32; /* 32-bit float */
1758 vs_type.length = vector_length;
1759
1760 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1761 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1762
1763 /* hold temporary "bool" clipmask */
1764 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1765
1766 fake_buf = lp_build_alloca_undef(gallivm,
1767 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1768 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1769 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1770 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1771
1772 /* code generated texture sampling */
1773 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1774
1775 image = draw_llvm_image_soa_create(draw_llvm_variant_key_images(key));
1776
1777 step = lp_build_const_int32(gallivm, vector_length);
1778
1779 ind_vec = blduivec.undef;
1780 for (i = 0; i < vs_type.length; i++) {
1781 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1782 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1783 }
1784
1785 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1786 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1787
1788 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1789 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1790 /*
1791 * Only needed for non-indexed path.
1792 */
1793 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1794
1795 /*
1796 * Pre-calculate everything which is constant per shader invocation.
1797 */
1798 for (j = 0; j < key->nr_vertex_elements; ++j) {
1799 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1800 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1801 struct pipe_vertex_element *velem = &key->vertex_element[j];
1802 LLVMValueRef vb_index =
1803 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1804 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1805 util_format_get_blocksize(velem->src_format));
1806 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1807 velem->src_offset);
1808 struct lp_build_if_state if_ctx;
1809
1810 if (velem->src_format != PIPE_FORMAT_NONE) {
1811 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1812 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1813 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1814 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1815 LLVMInt32TypeInContext(context), "");
1816 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1817 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1818 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1819
1820 ofbit = NULL;
1821 /*
1822 * We'll set buffer_size_adj to zero if we have of, so it will
1823 * always overflow later automatically without having to keep ofbit.
1824 * Overflows (with normal wraparound) doing the actual offset
1825 * calculation should be ok, just not for the buffer size calc.
1826 * It would also be possible to detect such overflows and return
1827 * zeros if that happens, but this would be more complex.
1828 */
1829 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1830 tmp = lp_build_sub(&bld, bsize, bld.one);
1831 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1832 &ofbit);
1833 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1834 buf_offset, &ofbit);
1835
1836 /*
1837 * We can't easily set fake vertex buffers outside the generated code.
1838 * Hence, set fake vertex buffers here instead basically, so fetch
1839 * code can always fetch using offset 0, eliminating all control flow
1840 * inside the main loop.
1841 * (Alternatively, could have control flow per vector skipping fetch
1842 * if ofbit is true.)
1843 */
1844 if (velem->instance_divisor) {
1845 /*
1846 * Index is equal to the start instance plus the number of current
1847 * instance divided by the divisor. In this case we compute it as:
1848 * index = start_instance + (instance_id / divisor).
1849 * Note we could actually do the fetch here, outside the loop -
1850 * it's all constant, hopefully llvm recognizes this.
1851 */
1852 LLVMValueRef current_instance;
1853 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1854 lp_build_const_int32(gallivm,
1855 velem->instance_divisor),
1856 "instance_divisor");
1857 instance_index[j] = lp_build_uadd_overflow(gallivm, system_values.base_instance,
1858 current_instance, &ofbit);
1859 }
1860
1861 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1862 buffer_size_adj[j], "");
1863
1864 temp_ptr = lp_build_alloca_undef(gallivm,
1865 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1866
1867 lp_build_if(&if_ctx, gallivm, ofbit);
1868 {
1869 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1870 }
1871 lp_build_else(&if_ctx);
1872 {
1873 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1874 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1875 }
1876 lp_build_endif(&if_ctx);
1877 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1878
1879 if (0) {
1880 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1881 lp_build_const_int32(gallivm, j),
1882 vb_index, vb_stride[j]);
1883 lp_build_printf(gallivm,
1884 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1885 vb_buffer_offset, src_offset, buf_offset);
1886 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1887 buffer_size, bsize);
1888 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1889 }
1890 }
1891 }
1892
1893 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1894 {
1895 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1896 LLVMValueRef io;
1897 LLVMValueRef clipmask; /* holds the clipmask value */
1898 LLVMValueRef true_index_array, index_store;
1899 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1900
1901 io_itr = lp_loop.counter;
1902
1903 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1904 #if DEBUG_STORE
1905 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1906 io_itr, io, lp_loop.counter);
1907 #endif
1908
1909 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1910 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1911
1912 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
1913 /*
1914 * Limit indices to fetch_max, otherwise might try to access indices
1915 * beyond index buffer (or rather vsplit elt buffer) size.
1916 * Could probably safely (?) skip this for non-indexed draws and
1917 * simplify things minimally (by removing it could combine the ind_vec
1918 * and start_vec adds). I think the only effect for non-indexed draws will
1919 * be that for the invalid elements they will be all fetched from the
1920 * same location as the last valid one, but noone should really care.
1921 */
1922 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1923
1924 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1925
1926 lp_build_if(&if_ctx, gallivm, have_elts);
1927 {
1928 /*
1929 * Note: you'd expect some comparison/clamp against fetch_elt_max
1930 * here.
1931 * There used to be one here but it was incorrect: overflow was
1932 * detected if index > fetch_elt_max - but the correct condition
1933 * would be index >= fetch_elt_max (since this is just size of elts
1934 * buffer / element size).
1935 * Using the correct condition however will cause failures - due to
1936 * vsplit/vcache code which rebases indices. So, as an example, if
1937 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1938 * replace all invalid indices with 0 - which in case of elt_bias
1939 * not being zero will get a different fetch index than the valid
1940 * index 0. So, just rely on vsplit code preventing out-of-bounds
1941 * fetches. This is also why it's safe to do elts fetch even if there
1942 * was no index buffer bound - the real buffer is never seen here, at
1943 * least not if there are index buffer overflows...
1944 */
1945
1946 /*
1947 * XXX should not have to do this, as scale can be handled
1948 * natively by loads (hits asserts though).
1949 */
1950 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1951 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1952 LLVMPointerType(LLVMInt8TypeInContext(context),
1953 0), "");
1954 tmp = lp_build_gather(gallivm, vs_type.length,
1955 32, bld.type, TRUE,
1956 fetch_elts, tmp, FALSE);
1957 LLVMBuildStore(builder, tmp, index_store);
1958 }
1959 lp_build_else(&if_ctx);
1960 {
1961 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1962 LLVMBuildStore(builder, tmp, index_store);
1963 }
1964 lp_build_endif(&if_ctx);
1965
1966 true_index_array = LLVMBuildLoad(builder, index_store, "");
1967
1968 for (j = 0; j < key->nr_vertex_elements; ++j) {
1969 struct pipe_vertex_element *velem = &key->vertex_element[j];
1970 const struct util_format_description *format_desc =
1971 util_format_description(velem->src_format);
1972
1973 if (format_desc->format == PIPE_FORMAT_NONE) {
1974 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1975 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1976 }
1977 }
1978 else if (velem->instance_divisor) {
1979 fetch_instanced(gallivm, format_desc, vs_type,
1980 vb_stride[j], map_ptr[j],
1981 buffer_size_adj[j],
1982 inputs[j], instance_index[j]);
1983 }
1984 else {
1985 fetch_vector(gallivm, format_desc, vs_type,
1986 vb_stride[j], map_ptr[j],
1987 buffer_size_adj[j],
1988 inputs[j], true_index_array);
1989 }
1990 }
1991
1992 struct lp_build_mask_context mask;
1993
1994 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
1995 /* In the paths with elts vertex id has to be unaffected by the
1996 * index bias and because indices inside our elements array have
1997 * already had index bias applied we need to subtract it here to
1998 * get back to the original index.
1999 * in the linear paths vertex id has to be unaffected by the
2000 * original start index and because we abuse the 'start' variable
2001 * to either represent the actual start index or the index at which
2002 * the primitive was split (we split rendering into chunks of at
2003 * most 4095-vertices) we need to back out the original start
2004 * index out of our vertex id here.
2005 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2006 */
2007 LLVMValueRef base_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, lp_build_const_int32(gallivm, 0));;
2008 system_values.basevertex = lp_build_broadcast_scalar(&blduivec, base_vertex);
2009 system_values.vertex_id = true_index_array;
2010 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
2011 lp_build_broadcast_scalar(&blduivec, vertex_id_offset), "");
2012
2013 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
2014 generate_vs(variant,
2015 builder,
2016 vs_type,
2017 outputs,
2018 ptr_aos,
2019 &system_values,
2020 context_ptr,
2021 sampler,
2022 image,
2023 key->clamp_vertex_color,
2024 &mask);
2025
2026 lp_build_mask_end(&mask);
2027 if (pos != -1 && cv != -1) {
2028 /* store original positions in clip before further manipulation */
2029 store_clip(gallivm, vs_type, io, outputs, pos);
2030
2031 /* do cliptest */
2032 if (enable_cliptest) {
2033 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
2034 /* allocate clipmask, assign it integer type */
2035 clipmask = generate_clipmask(llvm,
2036 gallivm,
2037 vs_type,
2038 outputs,
2039 key,
2040 context_ptr, &have_clipdist);
2041 temp = LLVMBuildOr(builder, clipmask, temp, "");
2042 /* store temporary clipping boolean value */
2043 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
2044 }
2045 else {
2046 clipmask = blduivec.zero;
2047 }
2048
2049 /* do viewport mapping */
2050 if (!bypass_viewport) {
2051 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
2052 }
2053 }
2054 else {
2055 clipmask = blduivec.zero;
2056 }
2057
2058 /* store clipmask in vertex header,
2059 * original positions in clip
2060 * and transformed positions in data
2061 */
2062 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
2063 vs_info->num_outputs, vs_type,
2064 enable_cliptest && key->need_edgeflags);
2065 }
2066 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2067
2068 sampler->destroy(sampler);
2069 image->destroy(image);
2070
2071 /* return clipping boolean value for function */
2072 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
2073 enable_cliptest && key->need_edgeflags);
2074
2075 LLVMBuildRet(builder, ret);
2076
2077 gallivm_verify_function(gallivm, variant_func);
2078 }
2079
2080
2081 struct draw_llvm_variant_key *
2082 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2083 {
2084 unsigned i;
2085 struct draw_llvm_variant_key *key;
2086 struct draw_sampler_static_state *draw_sampler;
2087 struct draw_image_static_state *draw_image;
2088
2089 key = (struct draw_llvm_variant_key *)store;
2090
2091 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2092
2093 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2094
2095 /* will have to rig this up properly later */
2096 key->clip_xy = llvm->draw->clip_xy;
2097 key->clip_z = llvm->draw->clip_z;
2098 key->clip_user = llvm->draw->clip_user;
2099 key->bypass_viewport = llvm->draw->bypass_viewport;
2100 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2101 /* XXX assumes edgeflag output not at 0 */
2102 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2103 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2104 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2105 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2106
2107 /* All variants of this shader will have the same value for
2108 * nr_samplers. Not yet trying to compact away holes in the
2109 * sampler array.
2110 */
2111 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2112 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2113 key->nr_sampler_views =
2114 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2115 }
2116 else {
2117 key->nr_sampler_views = key->nr_samplers;
2118 }
2119
2120 key->nr_images = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2121
2122 /* Presumably all variants of the shader should have the same
2123 * number of vertex elements - ie the number of shader inputs.
2124 * NOTE: we NEED to store the needed number of needed inputs
2125 * here, not the number of provided elements to match keysize
2126 * (and the offset of sampler state in the key).
2127 * If we have excess number of vertex elements, this is valid,
2128 * but the excess ones don't matter.
2129 * If we don't have enough vertex elements (which looks not really
2130 * valid but we'll handle it gracefully) fill out missing ones with
2131 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2132 */
2133 key->nr_vertex_elements =
2134 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2135
2136 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2137 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2138 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2139 memset(key->vertex_element, 0,
2140 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2141 }
2142 memcpy(key->vertex_element,
2143 llvm->draw->pt.vertex_element,
2144 sizeof(struct pipe_vertex_element) *
2145 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2146
2147 draw_sampler = draw_llvm_variant_key_samplers(key);
2148 memset(draw_sampler, 0,
2149 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2150
2151 for (i = 0 ; i < key->nr_samplers; i++) {
2152 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2153 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2154 }
2155 for (i = 0 ; i < key->nr_sampler_views; i++) {
2156 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2157 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2158 }
2159
2160 draw_image = draw_llvm_variant_key_images(key);
2161 memset(draw_image, 0,
2162 key->nr_images * sizeof *draw_image);
2163 for (i = 0; i < key->nr_images; i++) {
2164 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2165 llvm->draw->images[PIPE_SHADER_VERTEX][i]);
2166 }
2167 return key;
2168 }
2169
2170
2171 void
2172 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2173 {
2174 unsigned i;
2175 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2176 struct draw_image_static_state *image = draw_llvm_variant_key_images(key);
2177 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2178 debug_printf("clip_xy = %u\n", key->clip_xy);
2179 debug_printf("clip_z = %u\n", key->clip_z);
2180 debug_printf("clip_user = %u\n", key->clip_user);
2181 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2182 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2183 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2184 debug_printf("has_gs = %u\n", key->has_gs);
2185 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2186
2187 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2188 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2189 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2190 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2191 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2192 }
2193
2194 for (i = 0 ; i < key->nr_sampler_views; i++) {
2195 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2196 }
2197
2198 for (i = 0 ; i < key->nr_images; i++)
2199 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2200 }
2201
2202
2203 void
2204 draw_llvm_set_mapped_texture(struct draw_context *draw,
2205 enum pipe_shader_type shader_stage,
2206 unsigned sview_idx,
2207 uint32_t width, uint32_t height, uint32_t depth,
2208 uint32_t first_level, uint32_t last_level,
2209 const void *base_ptr,
2210 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2211 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2212 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2213 {
2214 unsigned j;
2215 struct draw_jit_texture *jit_tex;
2216
2217 assert(shader_stage == PIPE_SHADER_VERTEX ||
2218 shader_stage == PIPE_SHADER_GEOMETRY);
2219
2220 if (shader_stage == PIPE_SHADER_VERTEX) {
2221 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2222
2223 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2224 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2225 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2226
2227 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2228 } else {
2229 assert(0);
2230 return;
2231 }
2232
2233 jit_tex->width = width;
2234 jit_tex->height = height;
2235 jit_tex->depth = depth;
2236 jit_tex->first_level = first_level;
2237 jit_tex->last_level = last_level;
2238 jit_tex->base = base_ptr;
2239
2240 for (j = first_level; j <= last_level; j++) {
2241 jit_tex->mip_offsets[j] = mip_offsets[j];
2242 jit_tex->row_stride[j] = row_stride[j];
2243 jit_tex->img_stride[j] = img_stride[j];
2244 }
2245 }
2246
2247 void
2248 draw_llvm_set_mapped_image(struct draw_context *draw,
2249 enum pipe_shader_type shader_stage,
2250 unsigned idx,
2251 uint32_t width, uint32_t height, uint32_t depth,
2252 const void *base_ptr,
2253 uint32_t row_stride,
2254 uint32_t img_stride)
2255 {
2256 struct draw_jit_image *jit_image;
2257
2258 assert(shader_stage == PIPE_SHADER_VERTEX ||
2259 shader_stage == PIPE_SHADER_GEOMETRY);
2260
2261 if (shader_stage == PIPE_SHADER_VERTEX) {
2262 assert(idx < ARRAY_SIZE(draw->llvm->jit_context.images));
2263
2264 jit_image = &draw->llvm->jit_context.images[idx];
2265 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2266 assert(idx < ARRAY_SIZE(draw->llvm->gs_jit_context.images));
2267
2268 jit_image = &draw->llvm->gs_jit_context.images[idx];
2269 } else {
2270 assert(0);
2271 return;
2272 }
2273
2274 jit_image->width = width;
2275 jit_image->height = height;
2276 jit_image->depth = depth;
2277 jit_image->base = base_ptr;
2278
2279 jit_image->row_stride = row_stride;
2280 jit_image->img_stride = img_stride;
2281 }
2282
2283
2284 void
2285 draw_llvm_set_sampler_state(struct draw_context *draw,
2286 enum pipe_shader_type shader_type)
2287 {
2288 unsigned i;
2289
2290 if (shader_type == PIPE_SHADER_VERTEX) {
2291 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2292 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2293
2294 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2295 const struct pipe_sampler_state *s
2296 = draw->samplers[PIPE_SHADER_VERTEX][i];
2297 jit_sam->min_lod = s->min_lod;
2298 jit_sam->max_lod = s->max_lod;
2299 jit_sam->lod_bias = s->lod_bias;
2300 COPY_4V(jit_sam->border_color, s->border_color.f);
2301 }
2302 }
2303 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2304 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2305 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2306
2307 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2308 const struct pipe_sampler_state *s
2309 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2310 jit_sam->min_lod = s->min_lod;
2311 jit_sam->max_lod = s->max_lod;
2312 jit_sam->lod_bias = s->lod_bias;
2313 COPY_4V(jit_sam->border_color, s->border_color.f);
2314 }
2315 }
2316 }
2317 }
2318
2319
2320 void
2321 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2322 {
2323 struct draw_llvm *llvm = variant->llvm;
2324
2325 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2326 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2327 variant->shader->variants_cached, llvm->nr_variants);
2328 }
2329
2330 gallivm_destroy(variant->gallivm);
2331
2332 remove_from_list(&variant->list_item_local);
2333 variant->shader->variants_cached--;
2334 remove_from_list(&variant->list_item_global);
2335 llvm->nr_variants--;
2336 FREE(variant);
2337 }
2338
2339
2340 /**
2341 * Create LLVM types for various structures.
2342 */
2343 static void
2344 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2345 {
2346 struct gallivm_state *gallivm = var->gallivm;
2347 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
2348
2349 texture_type = create_jit_texture_type(gallivm, "texture");
2350 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2351 image_type = create_jit_image_type(gallivm, "image");
2352
2353 context_type = create_gs_jit_context_type(gallivm,
2354 var->shader->base.vector_length,
2355 texture_type, sampler_type,
2356 image_type,
2357 "draw_gs_jit_context");
2358 var->context_ptr_type = LLVMPointerType(context_type, 0);
2359
2360 var->input_array_type = create_gs_jit_input_type(gallivm);
2361 }
2362
2363 static LLVMTypeRef
2364 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2365 {
2366 if (!variant->context_ptr_type)
2367 create_gs_jit_types(variant);
2368 return variant->context_ptr_type;
2369 }
2370
2371 static LLVMValueRef
2372 generate_mask_value(struct draw_gs_llvm_variant *variant,
2373 struct lp_type gs_type)
2374 {
2375 struct gallivm_state *gallivm = variant->gallivm;
2376 LLVMBuilderRef builder = gallivm->builder;
2377 struct lp_type mask_type = lp_int_type(gs_type);
2378 LLVMValueRef num_prims;
2379 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2380 unsigned i;
2381
2382 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2383 variant->num_prims);
2384 for (i = 0; i < gs_type.length; i++) {
2385 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2386 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2387 }
2388 mask_val = lp_build_compare(gallivm, mask_type,
2389 PIPE_FUNC_GREATER, num_prims, mask_val);
2390
2391 return mask_val;
2392 }
2393
2394 static void
2395 draw_gs_llvm_generate(struct draw_llvm *llvm,
2396 struct draw_gs_llvm_variant *variant)
2397 {
2398 struct gallivm_state *gallivm = variant->gallivm;
2399 LLVMContextRef context = gallivm->context;
2400 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2401 LLVMTypeRef arg_types[7];
2402 LLVMTypeRef func_type;
2403 LLVMValueRef variant_func;
2404 LLVMValueRef context_ptr;
2405 LLVMValueRef prim_id_ptr;
2406 LLVMBasicBlockRef block;
2407 LLVMBuilderRef builder;
2408 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2409 struct lp_build_sampler_soa *sampler = 0;
2410 struct lp_build_image_soa *image = NULL;
2411 struct lp_build_context bld;
2412 struct lp_bld_tgsi_system_values system_values;
2413 char func_name[64];
2414 struct lp_type gs_type;
2415 unsigned i;
2416 struct draw_gs_llvm_iface gs_iface;
2417 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2418 LLVMValueRef consts_ptr, num_consts_ptr;
2419 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2420 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2421 struct lp_build_mask_context mask;
2422 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2423 unsigned vector_length = variant->shader->base.vector_length;
2424
2425 memset(&system_values, 0, sizeof(system_values));
2426 memset(&outputs, 0, sizeof(outputs));
2427
2428 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2429 variant->shader->variants_cached);
2430
2431 assert(variant->vertex_header_ptr_type);
2432
2433 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2434 arg_types[1] = variant->input_array_type; /* input */
2435 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2436 arg_types[3] = int32_type; /* num_prims */
2437 arg_types[4] = int32_type; /* instance_id */
2438 arg_types[5] = LLVMPointerType(
2439 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2440 arg_types[6] = int32_type;
2441
2442 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2443
2444 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2445
2446 variant->function = variant_func;
2447
2448 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2449
2450 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2451 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2452 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2453
2454 context_ptr = LLVMGetParam(variant_func, 0);
2455 input_array = LLVMGetParam(variant_func, 1);
2456 io_ptr = LLVMGetParam(variant_func, 2);
2457 num_prims = LLVMGetParam(variant_func, 3);
2458 system_values.instance_id = LLVMGetParam(variant_func, 4);
2459 prim_id_ptr = LLVMGetParam(variant_func, 5);
2460 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2461
2462 lp_build_name(context_ptr, "context");
2463 lp_build_name(input_array, "input");
2464 lp_build_name(io_ptr, "io");
2465 lp_build_name(num_prims, "num_prims");
2466 lp_build_name(system_values.instance_id, "instance_id");
2467 lp_build_name(prim_id_ptr, "prim_id_ptr");
2468 lp_build_name(system_values.invocation_id, "invocation_id");
2469
2470 variant->context_ptr = context_ptr;
2471 variant->io_ptr = io_ptr;
2472 variant->num_prims = num_prims;
2473
2474 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2475 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2476 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2477 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2478 gs_iface.input = input_array;
2479 gs_iface.variant = variant;
2480
2481 /*
2482 * Function body
2483 */
2484
2485 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2486 builder = gallivm->builder;
2487 LLVMPositionBuilderAtEnd(builder, block);
2488
2489 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2490
2491 memset(&gs_type, 0, sizeof gs_type);
2492 gs_type.floating = TRUE; /* floating point values */
2493 gs_type.sign = TRUE; /* values are signed */
2494 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2495 gs_type.width = 32; /* 32-bit float */
2496 gs_type.length = vector_length;
2497
2498 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2499 num_consts_ptr =
2500 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2501
2502 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2503 num_ssbos_ptr =
2504 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2505
2506 /* code generated texture sampling */
2507 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2508 image = draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant->key));
2509 mask_val = generate_mask_value(variant, gs_type);
2510 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2511
2512 if (gs_info->uses_primid) {
2513 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2514 }
2515
2516 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2517 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2518 tgsi_dump(tokens, 0);
2519 else
2520 nir_print_shader(llvm->draw->gs.geometry_shader->state.ir.nir, stderr);
2521 draw_gs_llvm_dump_variant_key(&variant->key);
2522 }
2523
2524 struct lp_build_tgsi_params params;
2525 memset(&params, 0, sizeof(params));
2526
2527 params.type = gs_type;
2528 params.mask = &mask;
2529 params.consts_ptr = consts_ptr;
2530 params.const_sizes_ptr = num_consts_ptr;
2531 params.system_values = &system_values;
2532 params.context_ptr = context_ptr;
2533 params.sampler = sampler;
2534 params.info = &llvm->draw->gs.geometry_shader->info;
2535 params.gs_iface = (const struct lp_build_gs_iface *)&gs_iface;
2536 params.ssbo_ptr = ssbos_ptr;
2537 params.ssbo_sizes_ptr = num_ssbos_ptr;
2538 params.image = image;
2539
2540 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2541 lp_build_tgsi_soa(variant->gallivm,
2542 tokens,
2543 &params,
2544 outputs);
2545 else
2546 lp_build_nir_soa(variant->gallivm,
2547 llvm->draw->gs.geometry_shader->state.ir.nir,
2548 &params,
2549 outputs);
2550
2551 sampler->destroy(sampler);
2552 image->destroy(image);
2553
2554 lp_build_mask_end(&mask);
2555
2556 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2557
2558 gallivm_verify_function(gallivm, variant_func);
2559 }
2560
2561
2562 struct draw_gs_llvm_variant *
2563 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2564 unsigned num_outputs,
2565 const struct draw_gs_llvm_variant_key *key)
2566 {
2567 struct draw_gs_llvm_variant *variant;
2568 struct llvm_geometry_shader *shader =
2569 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2570 LLVMTypeRef vertex_header;
2571 char module_name[64];
2572
2573 variant = MALLOC(sizeof *variant +
2574 shader->variant_key_size -
2575 sizeof variant->key);
2576 if (!variant)
2577 return NULL;
2578
2579 variant->llvm = llvm;
2580 variant->shader = shader;
2581
2582 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2583 variant->shader->variants_cached);
2584
2585 variant->gallivm = gallivm_create(module_name, llvm->context);
2586
2587 create_gs_jit_types(variant);
2588
2589 memcpy(&variant->key, key, shader->variant_key_size);
2590
2591 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2592
2593 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2594
2595 draw_gs_llvm_generate(llvm, variant);
2596
2597 gallivm_compile_module(variant->gallivm);
2598
2599 variant->jit_func = (draw_gs_jit_func)
2600 gallivm_jit_function(variant->gallivm, variant->function);
2601
2602 gallivm_free_ir(variant->gallivm);
2603
2604 variant->list_item_global.base = variant;
2605 variant->list_item_local.base = variant;
2606 /*variant->no = */shader->variants_created++;
2607 variant->list_item_global.base = variant;
2608
2609 return variant;
2610 }
2611
2612 void
2613 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2614 {
2615 struct draw_llvm *llvm = variant->llvm;
2616
2617 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2618 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2619 variant->shader->variants_cached, llvm->nr_gs_variants);
2620 }
2621
2622 gallivm_destroy(variant->gallivm);
2623
2624 remove_from_list(&variant->list_item_local);
2625 variant->shader->variants_cached--;
2626 remove_from_list(&variant->list_item_global);
2627 llvm->nr_gs_variants--;
2628 FREE(variant);
2629 }
2630
2631 struct draw_gs_llvm_variant_key *
2632 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2633 {
2634 unsigned i;
2635 struct draw_gs_llvm_variant_key *key;
2636 struct draw_sampler_static_state *draw_sampler;
2637 struct draw_image_static_state *draw_image;
2638
2639 key = (struct draw_gs_llvm_variant_key *)store;
2640
2641 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2642
2643 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2644
2645 /* All variants of this shader will have the same value for
2646 * nr_samplers. Not yet trying to compact away holes in the
2647 * sampler array.
2648 */
2649 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2650 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2651 key->nr_sampler_views =
2652 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2653 }
2654 else {
2655 key->nr_sampler_views = key->nr_samplers;
2656 }
2657
2658 key->nr_images = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2659
2660 draw_sampler = key->samplers;
2661
2662 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2663
2664 for (i = 0 ; i < key->nr_samplers; i++) {
2665 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2666 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2667 }
2668 for (i = 0 ; i < key->nr_sampler_views; i++) {
2669 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2670 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2671 }
2672
2673 draw_image = draw_gs_llvm_variant_key_images(key);
2674 memset(draw_image, 0,
2675 key->nr_images * sizeof *draw_image);
2676 for (i = 0; i < key->nr_images; i++) {
2677 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2678 llvm->draw->images[PIPE_SHADER_GEOMETRY][i]);
2679 }
2680 return key;
2681 }
2682
2683 void
2684 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2685 {
2686 unsigned i;
2687 struct draw_sampler_static_state *sampler = key->samplers;
2688 struct draw_image_static_state *image = draw_gs_llvm_variant_key_images(key);
2689 for (i = 0 ; i < key->nr_sampler_views; i++) {
2690 debug_printf("sampler[%i].src_format = %s\n", i,
2691 util_format_name(sampler[i].texture_state.format));
2692 }
2693
2694 for (i = 0 ; i < key->nr_images; i++)
2695 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2696
2697 }