projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: add JIT context/functions for tess stages.
[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 * Create LLVM type for struct draw_tcs_jit_context
502 */
503 static LLVMTypeRef
504 create_tcs_jit_context_type(struct gallivm_state *gallivm,
505 unsigned vector_length,
506 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
507 LLVMTypeRef image_type,
508 const char *struct_name)
509 {
510 LLVMTargetDataRef target = gallivm->target;
511 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
512 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
513 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
514 LLVMTypeRef context_type;
515
516 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
517 LP_MAX_TGSI_CONST_BUFFERS);
518 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
519 LP_MAX_TGSI_CONST_BUFFERS);
520 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
521 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
522
523 elem_types[4] = LLVMArrayType(texture_type,
524 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
525 elem_types[5] = LLVMArrayType(sampler_type,
526 PIPE_MAX_SAMPLERS); /* samplers */
527 elem_types[6] = LLVMArrayType(image_type,
528 PIPE_MAX_SHADER_IMAGES); /* images */
529
530 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
531 LP_MAX_TGSI_SHADER_BUFFERS);
532 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
533 LP_MAX_TGSI_SHADER_BUFFERS);
534
535 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
536 ARRAY_SIZE(elem_types), 0);
537
538 (void) target; /* silence unused var warning for non-debug build */
539 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, constants,
540 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
541 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_constants,
542 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
543 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, textures,
544 target, context_type,
545 DRAW_TCS_JIT_CTX_TEXTURES);
546 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, samplers,
547 target, context_type,
548 DRAW_TCS_JIT_CTX_SAMPLERS);
549 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, ssbos,
550 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
551 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_ssbos,
552 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
553 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, images,
554 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
555 LP_CHECK_STRUCT_SIZE(struct draw_tcs_jit_context,
556 target, context_type);
557
558 return context_type;
559 }
560
561 static LLVMTypeRef
562 create_tcs_jit_input_type(struct gallivm_state *gallivm)
563 {
564 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
565 LLVMTypeRef input_array;
566
567 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
568 input_array = LLVMArrayType(input_array, NUM_TCS_INPUTS); /* num attrs per vertex */
569 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
570
571 return input_array;
572 }
573
574 static LLVMTypeRef
575 create_tcs_jit_output_type(struct gallivm_state *gallivm)
576 {
577 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
578 LLVMTypeRef output_array;
579
580 output_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
581 output_array = LLVMArrayType(output_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
582 output_array = LLVMPointerType(output_array, 0); /* num vertices per prim */
583
584 return output_array;
585 }
586
587 static LLVMTypeRef
588 create_tes_jit_input_type(struct gallivm_state *gallivm)
589 {
590 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
591 LLVMTypeRef input_array;
592
593 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
594 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
595 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
596
597 return input_array;
598 }
599
600 /**
601 * Create LLVM type for struct draw_tes_jit_context
602 */
603 static LLVMTypeRef
604 create_tes_jit_context_type(struct gallivm_state *gallivm,
605 unsigned vector_length,
606 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
607 LLVMTypeRef image_type,
608 const char *struct_name)
609 {
610 LLVMTargetDataRef target = gallivm->target;
611 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
612 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
613 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
614 LLVMTypeRef context_type;
615
616 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
617 LP_MAX_TGSI_CONST_BUFFERS);
618 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
619 LP_MAX_TGSI_CONST_BUFFERS);
620 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
621 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
622
623 elem_types[4] = LLVMArrayType(texture_type,
624 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
625 elem_types[5] = LLVMArrayType(sampler_type,
626 PIPE_MAX_SAMPLERS); /* samplers */
627 elem_types[6] = LLVMArrayType(image_type,
628 PIPE_MAX_SHADER_IMAGES); /* images */
629
630 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
631 LP_MAX_TGSI_SHADER_BUFFERS);
632 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
633 LP_MAX_TGSI_SHADER_BUFFERS);
634
635 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
636 ARRAY_SIZE(elem_types), 0);
637
638 (void) target; /* silence unused var warning for non-debug build */
639 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, constants,
640 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
641 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_constants,
642 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
643 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, textures,
644 target, context_type,
645 DRAW_TCS_JIT_CTX_TEXTURES);
646 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, samplers,
647 target, context_type,
648 DRAW_TCS_JIT_CTX_SAMPLERS);
649 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, ssbos,
650 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
651 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_ssbos,
652 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
653 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, images,
654 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
655 LP_CHECK_STRUCT_SIZE(struct draw_tes_jit_context,
656 target, context_type);
657
658 return context_type;
659 }
660
661 /**
662 * Create LLVM types for various structures.
663 */
664 static void
665 create_jit_types(struct draw_llvm_variant *variant)
666 {
667 struct gallivm_state *gallivm = variant->gallivm;
668 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
669 vb_type, image_type;
670
671 texture_type = create_jit_texture_type(gallivm, "texture");
672 sampler_type = create_jit_sampler_type(gallivm, "sampler");
673 image_type = create_jit_image_type(gallivm, "image");
674
675 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
676 image_type,
677 "draw_jit_context");
678 variant->context_ptr_type = LLVMPointerType(context_type, 0);
679
680 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
681 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
682
683 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
684 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
685 }
686
687
688 static LLVMTypeRef
689 get_context_ptr_type(struct draw_llvm_variant *variant)
690 {
691 if (!variant->context_ptr_type)
692 create_jit_types(variant);
693 return variant->context_ptr_type;
694 }
695
696
697 static LLVMTypeRef
698 get_buffer_ptr_type(struct draw_llvm_variant *variant)
699 {
700 if (!variant->buffer_ptr_type)
701 create_jit_types(variant);
702 return variant->buffer_ptr_type;
703 }
704
705
706 static LLVMTypeRef
707 get_vb_ptr_type(struct draw_llvm_variant *variant)
708 {
709 if (!variant->vb_ptr_type)
710 create_jit_types(variant);
711 return variant->vb_ptr_type;
712 }
713
714 static LLVMTypeRef
715 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
716 {
717 if (!variant->vertex_header_ptr_type)
718 create_jit_types(variant);
719 return variant->vertex_header_ptr_type;
720 }
721
722
723 /**
724 * Create per-context LLVM info.
725 */
726 struct draw_llvm *
727 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
728 {
729 struct draw_llvm *llvm;
730
731 if (!lp_build_init())
732 return NULL;
733
734 llvm = CALLOC_STRUCT( draw_llvm );
735 if (!llvm)
736 return NULL;
737
738 llvm->draw = draw;
739
740 llvm->context = context;
741 if (!llvm->context) {
742 llvm->context = LLVMContextCreate();
743 llvm->context_owned = true;
744 }
745 if (!llvm->context)
746 goto fail;
747
748 llvm->nr_variants = 0;
749 make_empty_list(&llvm->vs_variants_list);
750
751 llvm->nr_gs_variants = 0;
752 make_empty_list(&llvm->gs_variants_list);
753
754 return llvm;
755
756 fail:
757 draw_llvm_destroy(llvm);
758 return NULL;
759 }
760
761
762 /**
763 * Free per-context LLVM info.
764 */
765 void
766 draw_llvm_destroy(struct draw_llvm *llvm)
767 {
768 if (llvm->context_owned)
769 LLVMContextDispose(llvm->context);
770 llvm->context = NULL;
771
772 /* XXX free other draw_llvm data? */
773 FREE(llvm);
774 }
775
776
777 /**
778 * Create LLVM-generated code for a vertex shader.
779 */
780 struct draw_llvm_variant *
781 draw_llvm_create_variant(struct draw_llvm *llvm,
782 unsigned num_inputs,
783 const struct draw_llvm_variant_key *key)
784 {
785 struct draw_llvm_variant *variant;
786 struct llvm_vertex_shader *shader =
787 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
788 LLVMTypeRef vertex_header;
789 char module_name[64];
790
791 variant = MALLOC(sizeof *variant +
792 shader->variant_key_size -
793 sizeof variant->key);
794 if (!variant)
795 return NULL;
796
797 variant->llvm = llvm;
798 variant->shader = shader;
799
800 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
801 variant->shader->variants_cached);
802
803 variant->gallivm = gallivm_create(module_name, llvm->context);
804
805 create_jit_types(variant);
806
807 memcpy(&variant->key, key, shader->variant_key_size);
808
809 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
810 if (llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_TGSI)
811 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
812 else
813 nir_print_shader(llvm->draw->vs.vertex_shader->state.ir.nir, stderr);
814 draw_llvm_dump_variant_key(&variant->key);
815 }
816
817 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
818
819 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
820
821 draw_llvm_generate(llvm, variant);
822
823 gallivm_compile_module(variant->gallivm);
824
825 variant->jit_func = (draw_jit_vert_func)
826 gallivm_jit_function(variant->gallivm, variant->function);
827
828 gallivm_free_ir(variant->gallivm);
829
830 variant->list_item_global.base = variant;
831 variant->list_item_local.base = variant;
832 /*variant->no = */shader->variants_created++;
833 variant->list_item_global.base = variant;
834
835 return variant;
836 }
837
838
839 static void
840 generate_vs(struct draw_llvm_variant *variant,
841 LLVMBuilderRef builder,
842 struct lp_type vs_type,
843 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
844 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
845 const struct lp_bld_tgsi_system_values *system_values,
846 LLVMValueRef context_ptr,
847 const struct lp_build_sampler_soa *draw_sampler,
848 const struct lp_build_image_soa *draw_image,
849 boolean clamp_vertex_color,
850 struct lp_build_mask_context *bld_mask)
851 {
852 struct draw_llvm *llvm = variant->llvm;
853 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
854 LLVMValueRef consts_ptr =
855 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
856 LLVMValueRef num_consts_ptr =
857 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
858 LLVMValueRef ssbos_ptr =
859 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
860 LLVMValueRef num_ssbos_ptr =
861 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
862
863 struct lp_build_tgsi_params params;
864 memset(&params, 0, sizeof(params));
865
866 params.type = vs_type;
867 params.mask = bld_mask;
868 params.consts_ptr = consts_ptr;
869 params.const_sizes_ptr = num_consts_ptr;
870 params.system_values = system_values;
871 params.inputs = inputs;
872 params.context_ptr = context_ptr;
873 params.sampler = draw_sampler;
874 params.info = &llvm->draw->vs.vertex_shader->info;
875 params.ssbo_ptr = ssbos_ptr;
876 params.ssbo_sizes_ptr = num_ssbos_ptr;
877 params.image = draw_image;
878
879 if (llvm->draw->vs.vertex_shader->state.ir.nir &&
880 llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_NIR)
881 lp_build_nir_soa(variant->gallivm,
882 llvm->draw->vs.vertex_shader->state.ir.nir,
883 &params,
884 outputs);
885 else
886 lp_build_tgsi_soa(variant->gallivm,
887 tokens,
888 &params,
889 outputs);
890
891 {
892 LLVMValueRef out;
893 unsigned chan, attrib;
894 struct lp_build_context bld;
895 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
896 lp_build_context_init(&bld, variant->gallivm, vs_type);
897
898 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
899 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
900 if (outputs[attrib][chan]) {
901 switch (info->output_semantic_name[attrib]) {
902 case TGSI_SEMANTIC_COLOR:
903 case TGSI_SEMANTIC_BCOLOR:
904 if (clamp_vertex_color) {
905 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
906 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
907 LLVMBuildStore(builder, out, outputs[attrib][chan]);
908 }
909 break;
910 }
911 }
912 }
913 }
914 }
915 }
916
917
918 static void
919 fetch_instanced(struct gallivm_state *gallivm,
920 const struct util_format_description *format_desc,
921 struct lp_type vs_type,
922 LLVMValueRef vb_stride,
923 LLVMValueRef map_ptr,
924 LLVMValueRef buffer_size_adj,
925 LLVMValueRef *inputs,
926 LLVMValueRef index)
927 {
928 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
929 LLVMTypeRef aosf_t, aosi_t;
930 LLVMValueRef zero = LLVMConstNull(i32_t);
931 LLVMBuilderRef builder = gallivm->builder;
932 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
933 unsigned i;
934
935 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
936 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
937
938 /* This mul can overflow. Wraparound is ok. */
939 stride = LLVMBuildMul(builder, vb_stride, index, "");
940
941 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
942 stride, buffer_size_adj,
943 "buffer_overflowed");
944
945 if (0) {
946 lp_build_print_value(gallivm, " instance index = ", index);
947 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
948 }
949
950 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
951 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
952 stride = LLVMBuildAnd(builder, stride, index_valid, "");
953
954 aos = lp_build_fetch_rgba_aos(gallivm,
955 format_desc,
956 lp_float32_vec4_type(),
957 FALSE,
958 map_ptr,
959 stride, zero, zero,
960 NULL);
961
962 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
963 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
964 aos = LLVMBuildAnd(builder, aos, index_valid, "");
965 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
966
967 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
968 LLVMValueRef index = lp_build_const_int32(gallivm, i);
969 inputs[i] = lp_build_extract_broadcast(gallivm,
970 lp_float32_vec4_type(),
971 vs_type, aos, index);
972 }
973 }
974
975
976 static void
977 fetch_vector(struct gallivm_state *gallivm,
978 const struct util_format_description *format_desc,
979 struct lp_type vs_type,
980 LLVMValueRef vb_stride,
981 LLVMValueRef map_ptr,
982 LLVMValueRef buffer_size_adj,
983 LLVMValueRef *inputs,
984 LLVMValueRef indices)
985 {
986 LLVMBuilderRef builder = gallivm->builder;
987 struct lp_build_context blduivec;
988 struct lp_type fetch_type = vs_type;
989 LLVMValueRef offset, valid_mask;
990 unsigned i;
991
992 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
993
994 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
995 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
996
997 /* This mul can overflow. Wraparound is ok. */
998 offset = lp_build_mul(&blduivec, vb_stride, indices);
999
1000 valid_mask = lp_build_compare(gallivm, blduivec.type,
1001 PIPE_FUNC_LESS, offset, buffer_size_adj);
1002
1003 /* not valid elements use offset 0 */
1004 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
1005
1006 if (0) {
1007 lp_build_print_value(gallivm, " indices = ", indices);
1008 lp_build_print_value(gallivm, " offsets = ", offset);
1009 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
1010 }
1011
1012 /*
1013 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
1014 * This should always produce better code.
1015 */
1016
1017 /* The type handling is annoying here... */
1018 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1019 format_desc->channel[0].pure_integer) {
1020 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1021 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
1022 }
1023 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1024 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
1025 }
1026 }
1027
1028 lp_build_fetch_rgba_soa(gallivm, format_desc,
1029 fetch_type, FALSE, map_ptr, offset,
1030 blduivec.zero, blduivec.zero,
1031 NULL, inputs);
1032
1033 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1034 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1035 lp_build_vec_type(gallivm, vs_type), "");
1036 }
1037
1038 /* out-of-bound fetches return all zeros */
1039 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1040 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
1041 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
1042 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1043 lp_build_vec_type(gallivm, vs_type), "");
1044 }
1045 }
1046
1047
1048 static void
1049 store_aos(struct gallivm_state *gallivm,
1050 LLVMValueRef io_ptr,
1051 LLVMValueRef index,
1052 LLVMValueRef value)
1053 {
1054 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
1055 LLVMBuilderRef builder = gallivm->builder;
1056 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
1057 LLVMValueRef indices[3];
1058
1059 indices[0] = lp_build_const_int32(gallivm, 0);
1060 indices[1] = index;
1061 indices[2] = lp_build_const_int32(gallivm, 0);
1062
1063 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
1064 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
1065
1066 #if DEBUG_STORE
1067 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
1068 #endif
1069
1070 /* Unaligned store due to the vertex header */
1071 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
1072 }
1073
1074 /**
1075 * Adjust the mask to architecture endianess. The mask will the store in struct:
1076 *
1077 * struct vertex_header {
1078 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
1079 * unsigned edgeflag:1;
1080 * unsigned pad:1;
1081 * unsigned vertex_id:16;
1082 * [...]
1083 * }
1084 *
1085 * On little-endian machine nothing needs to done, however on bit-endian machine
1086 * the mask's fields need to be adjusted with the algorithm:
1087 *
1088 * uint32_t reverse (uint32_t x)
1089 * {
1090 * return (x >> 16) | // vertex_id
1091 * ((x & 0x3fff) << 18) | // clipmask
1092 * ((x & 0x4000) << 3) | // pad
1093 * ((x & 0x8000) << 1); // edgeflag
1094 * }
1095 */
1096 static LLVMValueRef
1097 adjust_mask(struct gallivm_state *gallivm,
1098 LLVMValueRef mask)
1099 {
1100 #if UTIL_ARCH_BIG_ENDIAN
1101 LLVMBuilderRef builder = gallivm->builder;
1102 LLVMValueRef vertex_id;
1103 LLVMValueRef clipmask;
1104 LLVMValueRef pad;
1105 LLVMValueRef edgeflag;
1106
1107 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
1108 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
1109 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
1110 if (0) {
1111 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
1112 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
1113 }
1114 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
1115 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
1116
1117 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
1118 if (0) {
1119 mask = LLVMBuildOr(builder, mask, pad, "");
1120 }
1121 mask = LLVMBuildOr(builder, mask, edgeflag, "");
1122 #endif
1123 return mask;
1124 }
1125
1126 static void
1127 store_aos_array(struct gallivm_state *gallivm,
1128 struct lp_type soa_type,
1129 LLVMValueRef io_ptr,
1130 LLVMValueRef *indices,
1131 LLVMValueRef* aos,
1132 int attrib,
1133 int num_outputs,
1134 LLVMValueRef clipmask,
1135 boolean need_edgeflag)
1136 {
1137 LLVMBuilderRef builder = gallivm->builder;
1138 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
1139 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1140 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
1141 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1142 int vector_length = soa_type.length;
1143 int i;
1144
1145 debug_assert(TGSI_NUM_CHANNELS == 4);
1146
1147 for (i = 0; i < vector_length; i++) {
1148 linear_inds[i] = lp_build_const_int32(gallivm, i);
1149 if (indices) {
1150 inds[i] = indices[i];
1151 } else {
1152 inds[i] = linear_inds[i];
1153 }
1154 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1155 }
1156
1157 if (attrib == 0) {
1158 /* store vertex header for each of the n vertices */
1159 LLVMValueRef val, cliptmp;
1160 int vertex_id_pad_edgeflag;
1161
1162 /* If this assertion fails, it means we need to update the bit twidding
1163 * code here. See struct vertex_header in draw_private.h.
1164 */
1165 assert(DRAW_TOTAL_CLIP_PLANES==14);
1166 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1167 if (!need_edgeflag) {
1168 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
1169 }
1170 else {
1171 vertex_id_pad_edgeflag = (0xffff << 16);
1172 }
1173 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
1174 vertex_id_pad_edgeflag);
1175 /* OR with the clipmask */
1176 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
1177 for (i = 0; i < vector_length; i++) {
1178 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
1179 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
1180 val = adjust_mask(gallivm, val);
1181 #if DEBUG_STORE
1182 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
1183 io_ptrs[i], inds[i], val);
1184 #endif
1185 LLVMBuildStore(builder, val, id_ptr);
1186 }
1187 }
1188
1189 /* store for each of the n vertices */
1190 for (i = 0; i < vector_length; i++) {
1191 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
1192 }
1193 }
1194
1195
1196 static void
1197 convert_to_aos(struct gallivm_state *gallivm,
1198 LLVMValueRef io,
1199 LLVMValueRef *indices,
1200 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1201 LLVMValueRef clipmask,
1202 int num_outputs,
1203 struct lp_type soa_type,
1204 boolean need_edgeflag)
1205 {
1206 LLVMBuilderRef builder = gallivm->builder;
1207 unsigned chan, attrib, i;
1208
1209 #if DEBUG_STORE
1210 lp_build_printf(gallivm, " # storing begin\n");
1211 #endif
1212 for (attrib = 0; attrib < num_outputs; ++attrib) {
1213 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1214 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1215 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1216 if (outputs[attrib][chan]) {
1217 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1218 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1219 #if DEBUG_STORE
1220 lp_build_printf(gallivm, "output %d : %d ",
1221 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1222 attrib, 0),
1223 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1224 chan, 0));
1225 lp_build_print_value(gallivm, "val = ", out);
1226 {
1227 LLVMValueRef iv =
1228 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1229
1230 lp_build_print_value(gallivm, " ival = ", iv);
1231 }
1232 #endif
1233 soa[chan] = out;
1234 }
1235 else {
1236 soa[chan] = 0;
1237 }
1238 }
1239
1240
1241 if (soa_type.length == TGSI_NUM_CHANNELS) {
1242 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1243 } else {
1244 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1245
1246 for (i = 0; i < soa_type.length; ++i) {
1247 aos[i] = lp_build_extract_range(gallivm,
1248 soa[i % TGSI_NUM_CHANNELS],
1249 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1250 TGSI_NUM_CHANNELS);
1251 }
1252 }
1253
1254 store_aos_array(gallivm,
1255 soa_type,
1256 io, indices,
1257 aos,
1258 attrib,
1259 num_outputs,
1260 clipmask,
1261 need_edgeflag);
1262 }
1263 #if DEBUG_STORE
1264 lp_build_printf(gallivm, " # storing end\n");
1265 #endif
1266 }
1267
1268
1269 /**
1270 * Stores original vertex positions in clip coordinates
1271 */
1272 static void
1273 store_clip(struct gallivm_state *gallivm,
1274 const struct lp_type vs_type,
1275 LLVMValueRef io_ptr,
1276 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1277 int idx)
1278 {
1279 LLVMBuilderRef builder = gallivm->builder;
1280 LLVMValueRef soa[4];
1281 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1282 LLVMValueRef indices[2];
1283 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1284 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1285 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1286 LLVMTypeRef clip_ptr_type =
1287 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1288 4), 0);
1289 int i, j;
1290
1291 indices[0] =
1292 indices[1] = lp_build_const_int32(gallivm, 0);
1293
1294 for (i = 0; i < vs_type.length; i++) {
1295 inds[i] = lp_build_const_int32(gallivm, i);
1296 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1297 }
1298
1299 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1300 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1301 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1302 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1303
1304 for (i = 0; i < vs_type.length; i++) {
1305 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1306 }
1307
1308 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1309 for (i = 0; i < vs_type.length; ++i) {
1310 aos[i] = lp_build_extract_range(gallivm,
1311 soa[i % TGSI_NUM_CHANNELS],
1312 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1313 TGSI_NUM_CHANNELS);
1314 }
1315
1316 for (j = 0; j < vs_type.length; j++) {
1317 LLVMValueRef clip_ptr;
1318
1319 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1320 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1321
1322 /* Unaligned store */
1323 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1324 }
1325 }
1326
1327
1328 /**
1329 * Transforms the outputs for viewport mapping
1330 */
1331 static void
1332 generate_viewport(struct draw_llvm_variant *variant,
1333 LLVMBuilderRef builder,
1334 struct lp_type vs_type,
1335 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1336 LLVMValueRef context_ptr)
1337 {
1338 int i;
1339 struct gallivm_state *gallivm = variant->gallivm;
1340 struct lp_type f32_type = vs_type;
1341 const unsigned pos = variant->llvm->draw->vs.position_output;
1342 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1343 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1344 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1345 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1346
1347 /* We treat pipe_viewport_state as a float array */
1348 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1349 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1350
1351 /* for 1/w convention*/
1352 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1353 LLVMBuildStore(builder, out3, outputs[pos][3]);
1354
1355 /* Viewport Mapping */
1356 for (i=0; i<3; i++) {
1357 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1358 LLVMValueRef scale;
1359 LLVMValueRef trans;
1360 LLVMValueRef scale_i;
1361 LLVMValueRef trans_i;
1362 LLVMValueRef index;
1363
1364 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1365 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1366
1367 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1368 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1369
1370 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1371 LLVMBuildLoad(builder, scale_i, "scale"));
1372 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1373 LLVMBuildLoad(builder, trans_i, "trans"));
1374
1375 /* divide by w */
1376 out = LLVMBuildFMul(builder, out, out3, "");
1377 /* mult by scale, add translation */
1378 out = lp_build_fmuladd(builder, out, scale, trans);
1379
1380 /* store transformed outputs */
1381 LLVMBuildStore(builder, out, outputs[pos][i]);
1382 }
1383
1384 }
1385
1386
1387 /**
1388 * Returns clipmask as nxi32 bitmask for the n vertices
1389 */
1390 static LLVMValueRef
1391 generate_clipmask(struct draw_llvm *llvm,
1392 struct gallivm_state *gallivm,
1393 struct lp_type vs_type,
1394 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1395 struct draw_llvm_variant_key *key,
1396 LLVMValueRef context_ptr,
1397 boolean *have_clipdist)
1398 {
1399 LLVMBuilderRef builder = gallivm->builder;
1400 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1401 LLVMValueRef test, temp;
1402 LLVMValueRef zero, shift;
1403 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1404 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1405 LLVMValueRef plane1, planes, plane_ptr, sum;
1406 struct lp_type f32_type = vs_type;
1407 struct lp_type i32_type = lp_int_type(vs_type);
1408 const unsigned pos = llvm->draw->vs.position_output;
1409 const unsigned cv = llvm->draw->vs.clipvertex_output;
1410 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1411 boolean have_cd = false;
1412 boolean clip_user = key->clip_user;
1413 unsigned ucp_enable = key->ucp_enable;
1414 unsigned cd[2];
1415
1416 cd[0] = llvm->draw->vs.ccdistance_output[0];
1417 cd[1] = llvm->draw->vs.ccdistance_output[1];
1418
1419 if (cd[0] != pos || cd[1] != pos)
1420 have_cd = true;
1421
1422 if (num_written_clipdistance && !clip_user) {
1423 clip_user = true;
1424 ucp_enable = (1 << num_written_clipdistance) - 1;
1425 }
1426
1427 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1428 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1429 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1430 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1431
1432 /*
1433 * load clipvertex and position from correct locations.
1434 * if they are the same just load them once.
1435 */
1436 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1437 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1438 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1439 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1440
1441 if (clip_user && cv != pos) {
1442 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1443 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1444 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1445 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1446 } else {
1447 cv_x = pos_x;
1448 cv_y = pos_y;
1449 cv_z = pos_z;
1450 cv_w = pos_w;
1451 }
1452
1453 /*
1454 * Be careful with the comparisons and NaNs (using llvm's unordered
1455 * comparisons here).
1456 */
1457 /* Cliptest, for hardwired planes */
1458 /*
1459 * XXX should take guardband into account (currently not in key).
1460 * Otherwise might run the draw pipeline stages for nothing.
1461 */
1462 if (key->clip_xy) {
1463 /* plane 1 */
1464 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1465 temp = shift;
1466 test = LLVMBuildAnd(builder, test, temp, "");
1467 mask = test;
1468
1469 /* plane 2 */
1470 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1471 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1472 temp = LLVMBuildShl(builder, temp, shift, "");
1473 test = LLVMBuildAnd(builder, test, temp, "");
1474 mask = LLVMBuildOr(builder, mask, test, "");
1475
1476 /* plane 3 */
1477 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1478 temp = LLVMBuildShl(builder, temp, shift, "");
1479 test = LLVMBuildAnd(builder, test, temp, "");
1480 mask = LLVMBuildOr(builder, mask, test, "");
1481
1482 /* plane 4 */
1483 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1484 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1485 temp = LLVMBuildShl(builder, temp, shift, "");
1486 test = LLVMBuildAnd(builder, test, temp, "");
1487 mask = LLVMBuildOr(builder, mask, test, "");
1488 }
1489
1490 if (key->clip_z) {
1491 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1492 if (key->clip_halfz) {
1493 /* plane 5 */
1494 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1495 test = LLVMBuildAnd(builder, test, temp, "");
1496 mask = LLVMBuildOr(builder, mask, test, "");
1497 }
1498 else {
1499 /* plane 5 */
1500 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1501 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1502 test = LLVMBuildAnd(builder, test, temp, "");
1503 mask = LLVMBuildOr(builder, mask, test, "");
1504 }
1505 /* plane 6 */
1506 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1507 temp = LLVMBuildShl(builder, temp, shift, "");
1508 test = LLVMBuildAnd(builder, test, temp, "");
1509 mask = LLVMBuildOr(builder, mask, test, "");
1510 }
1511
1512 if (clip_user) {
1513 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1514 LLVMValueRef indices[3];
1515 LLVMValueRef is_nan_or_inf;
1516
1517 /* userclip planes */
1518 while (ucp_enable) {
1519 unsigned plane_idx = ffs(ucp_enable)-1;
1520 ucp_enable &= ~(1 << plane_idx);
1521 plane_idx += 6;
1522
1523 if (have_cd && num_written_clipdistance) {
1524 LLVMValueRef clipdist;
1525 int i;
1526 i = plane_idx - 6;
1527
1528 *have_clipdist = TRUE;
1529 if (i < 4) {
1530 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1531 } else {
1532 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1533 }
1534 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1535 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1536 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1537 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1538 test = LLVMBuildAnd(builder, test, temp, "");
1539 mask = LLVMBuildOr(builder, mask, test, "");
1540 } else {
1541 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1542 indices[0] = lp_build_const_int32(gallivm, 0);
1543 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1544
1545 indices[2] = lp_build_const_int32(gallivm, 0);
1546 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1547 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1548 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1549 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1550
1551 indices[2] = lp_build_const_int32(gallivm, 1);
1552 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1553 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1554 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1555 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1556
1557 indices[2] = lp_build_const_int32(gallivm, 2);
1558 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1559 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1560 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1561 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1562
1563 indices[2] = lp_build_const_int32(gallivm, 3);
1564 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1565 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1566 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1567 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1568
1569 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1570 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1571 test = LLVMBuildAnd(builder, test, temp, "");
1572 mask = LLVMBuildOr(builder, mask, test, "");
1573 }
1574 }
1575 }
1576 if (key->need_edgeflags) {
1577 /*
1578 * This isn't really part of clipmask but stored the same in vertex
1579 * header later, so do it here.
1580 */
1581 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1582 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1583 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1584 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1585 temp = lp_build_const_int_vec(gallivm, i32_type,
1586 1LL << DRAW_TOTAL_CLIP_PLANES);
1587 test = LLVMBuildAnd(builder, test, temp, "");
1588 mask = LLVMBuildOr(builder, mask, test, "");
1589 }
1590 return mask;
1591 }
1592
1593
1594 /**
1595 * Returns boolean if any clipping has occurred
1596 * Used zero/one i8 value to represent boolean
1597 */
1598 static LLVMValueRef
1599 clipmask_booli8(struct gallivm_state *gallivm,
1600 const struct lp_type vs_type,
1601 LLVMValueRef clipmask_bool_ptr,
1602 boolean edgeflag_in_clipmask)
1603 {
1604 LLVMBuilderRef builder = gallivm->builder;
1605 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1606 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1607 LLVMValueRef ret;
1608 struct lp_build_context bldivec;
1609
1610 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1611
1612 /*
1613 * We need to invert the edgeflag bit from the clipmask here
1614 * (because the result is really if we want to run the pipeline or not
1615 * and we (may) need it if edgeflag was 0).
1616 */
1617 if (edgeflag_in_clipmask) {
1618 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1619 1LL << DRAW_TOTAL_CLIP_PLANES);
1620 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1621 }
1622
1623 /*
1624 * XXX: probably should mask off bits from the mask which come from
1625 * vertices which were beyond the count (i.e. indices_valid for
1626 * linear fetches, for elts ones we don't have the correct mask
1627 * right now). Otherwise might run the pipeline for nothing,
1628 * though everything should still work.
1629 */
1630 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1631 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1632 return ret;
1633 }
1634
1635 static LLVMValueRef
1636 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface *gs_iface,
1637 struct lp_build_context * bld,
1638 boolean is_vindex_indirect,
1639 LLVMValueRef vertex_index,
1640 boolean is_aindex_indirect,
1641 LLVMValueRef attrib_index,
1642 LLVMValueRef swizzle_index)
1643 {
1644 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1645 struct gallivm_state *gallivm = bld->gallivm;
1646 LLVMBuilderRef builder = gallivm->builder;
1647 LLVMValueRef indices[3];
1648 LLVMValueRef res;
1649 struct lp_type type = bld->type;
1650
1651 if (is_vindex_indirect || is_aindex_indirect) {
1652 int i;
1653 res = bld->zero;
1654 for (i = 0; i < type.length; ++i) {
1655 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1656 LLVMValueRef vert_chan_index = vertex_index;
1657 LLVMValueRef attr_chan_index = attrib_index;
1658 LLVMValueRef channel_vec, value;
1659
1660 if (is_vindex_indirect) {
1661 vert_chan_index = LLVMBuildExtractElement(builder,
1662 vertex_index, idx, "");
1663 }
1664 if (is_aindex_indirect) {
1665 attr_chan_index = LLVMBuildExtractElement(builder,
1666 attrib_index, idx, "");
1667 }
1668
1669 indices[0] = vert_chan_index;
1670 indices[1] = attr_chan_index;
1671 indices[2] = swizzle_index;
1672
1673 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1674 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1675 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1676
1677 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1678 }
1679 } else {
1680 indices[0] = vertex_index;
1681 indices[1] = attrib_index;
1682 indices[2] = swizzle_index;
1683
1684 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1685 res = LLVMBuildLoad(builder, res, "");
1686 }
1687
1688 return res;
1689 }
1690
1691 static void
1692 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface *gs_base,
1693 struct lp_build_context * bld,
1694 LLVMValueRef (*outputs)[4],
1695 LLVMValueRef emitted_vertices_vec,
1696 LLVMValueRef stream_id)
1697 {
1698 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1699 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1700 struct gallivm_state *gallivm = variant->gallivm;
1701 LLVMBuilderRef builder = gallivm->builder;
1702 struct lp_type gs_type = bld->type;
1703 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1704 lp_int_type(gs_type), 0);
1705 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1706 LLVMValueRef next_prim_offset =
1707 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1708 LLVMValueRef io = variant->io_ptr;
1709 unsigned i;
1710 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1711
1712 for (i = 0; i < gs_type.length; ++i) {
1713 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1714 LLVMValueRef currently_emitted =
1715 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1716 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1717 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1718 }
1719
1720 LLVMValueRef stream_idx = LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), "");
1721 LLVMValueRef cnd = LLVMBuildICmp(builder, LLVMIntULT, stream_idx, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1722 struct lp_build_if_state if_ctx;
1723 lp_build_if(&if_ctx, gallivm, cnd);
1724 io = lp_build_pointer_get(builder, io, LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), ""));
1725
1726 convert_to_aos(gallivm, io, indices,
1727 outputs, clipmask,
1728 gs_info->num_outputs, gs_type,
1729 FALSE);
1730 lp_build_endif(&if_ctx);
1731 }
1732
1733 static void
1734 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface *gs_base,
1735 struct lp_build_context * bld,
1736 LLVMValueRef total_emitted_vertices_vec_ptr,
1737 LLVMValueRef verts_per_prim_vec,
1738 LLVMValueRef emitted_prims_vec,
1739 LLVMValueRef mask_vec)
1740 {
1741 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1742 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1743 struct gallivm_state *gallivm = variant->gallivm;
1744 LLVMBuilderRef builder = gallivm->builder;
1745 LLVMValueRef prim_lengts_ptr =
1746 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1747 unsigned i;
1748
1749 for (i = 0; i < bld->type.length; ++i) {
1750 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1751 LLVMValueRef prims_emitted =
1752 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1753 LLVMValueRef store_ptr;
1754 LLVMValueRef num_vertices =
1755 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1756
1757 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1758 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1759 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1760 LLVMBuildStore(builder, num_vertices, store_ptr);
1761 }
1762 }
1763
1764 static void
1765 draw_gs_llvm_epilogue(const struct lp_build_gs_iface *gs_base,
1766 LLVMValueRef total_emitted_vertices_vec,
1767 LLVMValueRef emitted_prims_vec, unsigned stream)
1768 {
1769 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1770 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1771 struct gallivm_state *gallivm = variant->gallivm;
1772 LLVMBuilderRef builder = gallivm->builder;
1773 LLVMValueRef emitted_verts_ptr =
1774 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1775 LLVMValueRef emitted_prims_ptr =
1776 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1777 LLVMValueRef stream_val = lp_build_const_int32(gallivm, stream);
1778
1779 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &stream_val, 1, "");
1780 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &stream_val, 1, "");
1781
1782 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1783 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1784 }
1785
1786 static void
1787 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1788 {
1789 struct gallivm_state *gallivm = variant->gallivm;
1790 LLVMContextRef context = gallivm->context;
1791 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1792 LLVMTypeRef arg_types[12];
1793 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1794 LLVMTypeRef func_type;
1795 LLVMValueRef context_ptr;
1796 LLVMBasicBlockRef block;
1797 LLVMBuilderRef builder;
1798 char func_name[64];
1799 struct lp_type vs_type;
1800 LLVMValueRef count, fetch_elts, start_or_maxelt;
1801 LLVMValueRef vertex_id_offset;
1802 LLVMValueRef stride, step, io_itr;
1803 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1804 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1805 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1806 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1807 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1808 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1809 LLVMValueRef fake_buf_ptr, fake_buf;
1810
1811 struct draw_context *draw = llvm->draw;
1812 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1813 unsigned i, j;
1814 struct lp_build_context bld, blduivec;
1815 struct lp_build_loop_state lp_loop;
1816 struct lp_build_if_state if_ctx;
1817 const int vector_length = lp_native_vector_width / 32;
1818 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1819 struct lp_build_sampler_soa *sampler = 0;
1820 struct lp_build_image_soa *image = NULL;
1821 LLVMValueRef ret, clipmask_bool_ptr;
1822 struct draw_llvm_variant_key *key = &variant->key;
1823 /* If geometry shader is present we need to skip both the viewport
1824 * transformation and clipping otherwise the inputs to the geometry
1825 * shader will be incorrect.
1826 * The code can't handle vp transform when vs writes vp index neither
1827 * (though this would be fixable here, but couldn't just broadcast
1828 * the values).
1829 */
1830 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1831 vs_info->writes_viewport_index;
1832 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1833 key->clip_z ||
1834 key->clip_user ||
1835 key->need_edgeflags);
1836 LLVMValueRef variant_func;
1837 const unsigned pos = draw->vs.position_output;
1838 const unsigned cv = draw->vs.clipvertex_output;
1839 boolean have_clipdist = FALSE;
1840 struct lp_bld_tgsi_system_values system_values;
1841
1842 memset(&system_values, 0, sizeof(system_values));
1843 memset(&outputs, 0, sizeof(outputs));
1844 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1845 variant->shader->variants_cached);
1846
1847 i = 0;
1848 arg_types[i++] = get_context_ptr_type(variant); /* context */
1849 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1850 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1851 arg_types[i++] = int32_type; /* count */
1852 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1853 arg_types[i++] = int32_type; /* stride */
1854 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1855 arg_types[i++] = int32_type; /* instance_id */
1856 arg_types[i++] = int32_type; /* vertex_id_offset */
1857 arg_types[i++] = int32_type; /* start_instance */
1858 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1859 arg_types[i++] = int32_type; /* draw_id */
1860
1861 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1862 arg_types, num_arg_types, 0);
1863
1864 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1865 variant->function = variant_func;
1866
1867 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1868 for (i = 0; i < num_arg_types; ++i)
1869 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1870 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1871
1872 context_ptr = LLVMGetParam(variant_func, 0);
1873 io_ptr = LLVMGetParam(variant_func, 1);
1874 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1875 count = LLVMGetParam(variant_func, 3);
1876 /*
1877 * XXX: the maxelt part is unused. Not really useful, since we cannot
1878 * get index buffer overflows due to vsplit (which provides its own
1879 * elts buffer, with a different size than what's passed in here).
1880 */
1881 start_or_maxelt = LLVMGetParam(variant_func, 4);
1882 /*
1883 * XXX: stride is actually unused. The stride we use is strictly calculated
1884 * from the number of outputs (including the draw_extra outputs).
1885 * Should probably fix some day (we need a new vs just because of extra
1886 * outputs which the generated vs won't touch).
1887 */
1888 stride = LLVMGetParam(variant_func, 5);
1889 vb_ptr = LLVMGetParam(variant_func, 6);
1890 system_values.instance_id = LLVMGetParam(variant_func, 7);
1891 vertex_id_offset = LLVMGetParam(variant_func, 8);
1892 system_values.base_instance = LLVMGetParam(variant_func, 9);
1893 fetch_elts = LLVMGetParam(variant_func, 10);
1894 system_values.draw_id = LLVMGetParam(variant_func, 11);
1895
1896 lp_build_name(context_ptr, "context");
1897 lp_build_name(io_ptr, "io");
1898 lp_build_name(vbuffers_ptr, "vbuffers");
1899 lp_build_name(count, "count");
1900 lp_build_name(start_or_maxelt, "start_or_maxelt");
1901 lp_build_name(stride, "stride");
1902 lp_build_name(vb_ptr, "vb");
1903 lp_build_name(system_values.instance_id, "instance_id");
1904 lp_build_name(vertex_id_offset, "vertex_id_offset");
1905 lp_build_name(system_values.base_instance, "start_instance");
1906 lp_build_name(fetch_elts, "fetch_elts");
1907 lp_build_name(system_values.draw_id, "draw_id");
1908
1909 /*
1910 * Function body
1911 */
1912
1913 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1914 builder = gallivm->builder;
1915 LLVMPositionBuilderAtEnd(builder, block);
1916
1917 memset(&vs_type, 0, sizeof vs_type);
1918 vs_type.floating = TRUE; /* floating point values */
1919 vs_type.sign = TRUE; /* values are signed */
1920 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1921 vs_type.width = 32; /* 32-bit float */
1922 vs_type.length = vector_length;
1923
1924 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1925 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1926
1927 /* hold temporary "bool" clipmask */
1928 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1929
1930 fake_buf = lp_build_alloca_undef(gallivm,
1931 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1932 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1933 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1934 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1935
1936 /* code generated texture sampling */
1937 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1938
1939 image = draw_llvm_image_soa_create(draw_llvm_variant_key_images(key));
1940
1941 step = lp_build_const_int32(gallivm, vector_length);
1942
1943 ind_vec = blduivec.undef;
1944 for (i = 0; i < vs_type.length; i++) {
1945 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1946 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1947 }
1948
1949 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1950 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1951
1952 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1953 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1954 /*
1955 * Only needed for non-indexed path.
1956 */
1957 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1958
1959 /*
1960 * Pre-calculate everything which is constant per shader invocation.
1961 */
1962 for (j = 0; j < key->nr_vertex_elements; ++j) {
1963 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1964 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1965 struct pipe_vertex_element *velem = &key->vertex_element[j];
1966 LLVMValueRef vb_index =
1967 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1968 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1969 util_format_get_blocksize(velem->src_format));
1970 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1971 velem->src_offset);
1972 struct lp_build_if_state if_ctx;
1973
1974 if (velem->src_format != PIPE_FORMAT_NONE) {
1975 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1976 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1977 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1978 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1979 LLVMInt32TypeInContext(context), "");
1980 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1981 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1982 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1983
1984 ofbit = NULL;
1985 /*
1986 * We'll set buffer_size_adj to zero if we have of, so it will
1987 * always overflow later automatically without having to keep ofbit.
1988 * Overflows (with normal wraparound) doing the actual offset
1989 * calculation should be ok, just not for the buffer size calc.
1990 * It would also be possible to detect such overflows and return
1991 * zeros if that happens, but this would be more complex.
1992 */
1993 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1994 tmp = lp_build_sub(&bld, bsize, bld.one);
1995 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1996 &ofbit);
1997 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1998 buf_offset, &ofbit);
1999
2000 /*
2001 * We can't easily set fake vertex buffers outside the generated code.
2002 * Hence, set fake vertex buffers here instead basically, so fetch
2003 * code can always fetch using offset 0, eliminating all control flow
2004 * inside the main loop.
2005 * (Alternatively, could have control flow per vector skipping fetch
2006 * if ofbit is true.)
2007 */
2008 if (velem->instance_divisor) {
2009 /*
2010 * Index is equal to the start instance plus the number of current
2011 * instance divided by the divisor. In this case we compute it as:
2012 * index = start_instance + (instance_id / divisor).
2013 * Note we could actually do the fetch here, outside the loop -
2014 * it's all constant, hopefully llvm recognizes this.
2015 */
2016 LLVMValueRef current_instance;
2017 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
2018 lp_build_const_int32(gallivm,
2019 velem->instance_divisor),
2020 "instance_divisor");
2021 instance_index[j] = lp_build_uadd_overflow(gallivm, system_values.base_instance,
2022 current_instance, &ofbit);
2023 }
2024
2025 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
2026 buffer_size_adj[j], "");
2027
2028 temp_ptr = lp_build_alloca_undef(gallivm,
2029 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
2030
2031 lp_build_if(&if_ctx, gallivm, ofbit);
2032 {
2033 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
2034 }
2035 lp_build_else(&if_ctx);
2036 {
2037 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
2038 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
2039 }
2040 lp_build_endif(&if_ctx);
2041 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
2042
2043 if (0) {
2044 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
2045 lp_build_const_int32(gallivm, j),
2046 vb_index, vb_stride[j]);
2047 lp_build_printf(gallivm,
2048 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
2049 vb_buffer_offset, src_offset, buf_offset);
2050 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
2051 buffer_size, bsize);
2052 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
2053 }
2054 }
2055 }
2056
2057 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
2058 {
2059 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
2060 LLVMValueRef io;
2061 LLVMValueRef clipmask; /* holds the clipmask value */
2062 LLVMValueRef true_index_array, index_store;
2063 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
2064
2065 io_itr = lp_loop.counter;
2066
2067 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
2068 #if DEBUG_STORE
2069 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
2070 io_itr, io, lp_loop.counter);
2071 #endif
2072
2073 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
2074 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
2075
2076 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
2077 /*
2078 * Limit indices to fetch_max, otherwise might try to access indices
2079 * beyond index buffer (or rather vsplit elt buffer) size.
2080 * Could probably safely (?) skip this for non-indexed draws and
2081 * simplify things minimally (by removing it could combine the ind_vec
2082 * and start_vec adds). I think the only effect for non-indexed draws will
2083 * be that for the invalid elements they will be all fetched from the
2084 * same location as the last valid one, but noone should really care.
2085 */
2086 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
2087
2088 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
2089
2090 lp_build_if(&if_ctx, gallivm, have_elts);
2091 {
2092 /*
2093 * Note: you'd expect some comparison/clamp against fetch_elt_max
2094 * here.
2095 * There used to be one here but it was incorrect: overflow was
2096 * detected if index > fetch_elt_max - but the correct condition
2097 * would be index >= fetch_elt_max (since this is just size of elts
2098 * buffer / element size).
2099 * Using the correct condition however will cause failures - due to
2100 * vsplit/vcache code which rebases indices. So, as an example, if
2101 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
2102 * replace all invalid indices with 0 - which in case of elt_bias
2103 * not being zero will get a different fetch index than the valid
2104 * index 0. So, just rely on vsplit code preventing out-of-bounds
2105 * fetches. This is also why it's safe to do elts fetch even if there
2106 * was no index buffer bound - the real buffer is never seen here, at
2107 * least not if there are index buffer overflows...
2108 */
2109
2110 /*
2111 * XXX should not have to do this, as scale can be handled
2112 * natively by loads (hits asserts though).
2113 */
2114 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
2115 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
2116 LLVMPointerType(LLVMInt8TypeInContext(context),
2117 0), "");
2118 tmp = lp_build_gather(gallivm, vs_type.length,
2119 32, bld.type, TRUE,
2120 fetch_elts, tmp, FALSE);
2121 LLVMBuildStore(builder, tmp, index_store);
2122 }
2123 lp_build_else(&if_ctx);
2124 {
2125 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
2126 LLVMBuildStore(builder, tmp, index_store);
2127 }
2128 lp_build_endif(&if_ctx);
2129
2130 true_index_array = LLVMBuildLoad(builder, index_store, "");
2131
2132 for (j = 0; j < key->nr_vertex_elements; ++j) {
2133 struct pipe_vertex_element *velem = &key->vertex_element[j];
2134 const struct util_format_description *format_desc =
2135 util_format_description(velem->src_format);
2136
2137 if (format_desc->format == PIPE_FORMAT_NONE) {
2138 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2139 inputs[j][i] = lp_build_zero(gallivm, vs_type);
2140 }
2141 }
2142 else if (velem->instance_divisor) {
2143 fetch_instanced(gallivm, format_desc, vs_type,
2144 vb_stride[j], map_ptr[j],
2145 buffer_size_adj[j],
2146 inputs[j], instance_index[j]);
2147 }
2148 else {
2149 fetch_vector(gallivm, format_desc, vs_type,
2150 vb_stride[j], map_ptr[j],
2151 buffer_size_adj[j],
2152 inputs[j], true_index_array);
2153 }
2154 }
2155
2156 struct lp_build_mask_context mask;
2157
2158 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
2159 /* In the paths with elts vertex id has to be unaffected by the
2160 * index bias and because indices inside our elements array have
2161 * already had index bias applied we need to subtract it here to
2162 * get back to the original index.
2163 * in the linear paths vertex id has to be unaffected by the
2164 * original start index and because we abuse the 'start' variable
2165 * to either represent the actual start index or the index at which
2166 * the primitive was split (we split rendering into chunks of at
2167 * most 4095-vertices) we need to back out the original start
2168 * index out of our vertex id here.
2169 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2170 */
2171 LLVMValueRef base_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, lp_build_const_int32(gallivm, 0));;
2172 system_values.basevertex = lp_build_broadcast_scalar(&blduivec, base_vertex);
2173 system_values.vertex_id = true_index_array;
2174 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
2175 lp_build_broadcast_scalar(&blduivec, vertex_id_offset), "");
2176
2177 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
2178 generate_vs(variant,
2179 builder,
2180 vs_type,
2181 outputs,
2182 ptr_aos,
2183 &system_values,
2184 context_ptr,
2185 sampler,
2186 image,
2187 key->clamp_vertex_color,
2188 &mask);
2189
2190 lp_build_mask_end(&mask);
2191 if (pos != -1 && cv != -1) {
2192 /* store original positions in clip before further manipulation */
2193 store_clip(gallivm, vs_type, io, outputs, pos);
2194
2195 /* do cliptest */
2196 if (enable_cliptest) {
2197 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
2198 /* allocate clipmask, assign it integer type */
2199 clipmask = generate_clipmask(llvm,
2200 gallivm,
2201 vs_type,
2202 outputs,
2203 key,
2204 context_ptr, &have_clipdist);
2205 temp = LLVMBuildOr(builder, clipmask, temp, "");
2206 /* store temporary clipping boolean value */
2207 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
2208 }
2209 else {
2210 clipmask = blduivec.zero;
2211 }
2212
2213 /* do viewport mapping */
2214 if (!bypass_viewport) {
2215 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
2216 }
2217 }
2218 else {
2219 clipmask = blduivec.zero;
2220 }
2221
2222 /* store clipmask in vertex header,
2223 * original positions in clip
2224 * and transformed positions in data
2225 */
2226 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
2227 vs_info->num_outputs, vs_type,
2228 enable_cliptest && key->need_edgeflags);
2229 }
2230 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2231
2232 sampler->destroy(sampler);
2233 image->destroy(image);
2234
2235 /* return clipping boolean value for function */
2236 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
2237 enable_cliptest && key->need_edgeflags);
2238
2239 LLVMBuildRet(builder, ret);
2240
2241 gallivm_verify_function(gallivm, variant_func);
2242 }
2243
2244
2245 struct draw_llvm_variant_key *
2246 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2247 {
2248 unsigned i;
2249 struct draw_llvm_variant_key *key;
2250 struct draw_sampler_static_state *draw_sampler;
2251 struct draw_image_static_state *draw_image;
2252
2253 key = (struct draw_llvm_variant_key *)store;
2254
2255 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2256
2257 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2258
2259 /* will have to rig this up properly later */
2260 key->clip_xy = llvm->draw->clip_xy;
2261 key->clip_z = llvm->draw->clip_z;
2262 key->clip_user = llvm->draw->clip_user;
2263 key->bypass_viewport = llvm->draw->bypass_viewport;
2264 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2265 /* XXX assumes edgeflag output not at 0 */
2266 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2267 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2268 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2269 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2270
2271 /* All variants of this shader will have the same value for
2272 * nr_samplers. Not yet trying to compact away holes in the
2273 * sampler array.
2274 */
2275 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2276 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2277 key->nr_sampler_views =
2278 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2279 }
2280 else {
2281 key->nr_sampler_views = key->nr_samplers;
2282 }
2283
2284 key->nr_images = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2285
2286 /* Presumably all variants of the shader should have the same
2287 * number of vertex elements - ie the number of shader inputs.
2288 * NOTE: we NEED to store the needed number of needed inputs
2289 * here, not the number of provided elements to match keysize
2290 * (and the offset of sampler state in the key).
2291 * If we have excess number of vertex elements, this is valid,
2292 * but the excess ones don't matter.
2293 * If we don't have enough vertex elements (which looks not really
2294 * valid but we'll handle it gracefully) fill out missing ones with
2295 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2296 */
2297 key->nr_vertex_elements =
2298 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2299
2300 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2301 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2302 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2303 memset(key->vertex_element, 0,
2304 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2305 }
2306 memcpy(key->vertex_element,
2307 llvm->draw->pt.vertex_element,
2308 sizeof(struct pipe_vertex_element) *
2309 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2310
2311 draw_sampler = draw_llvm_variant_key_samplers(key);
2312 memset(draw_sampler, 0,
2313 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2314
2315 for (i = 0 ; i < key->nr_samplers; i++) {
2316 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2317 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2318 }
2319 for (i = 0 ; i < key->nr_sampler_views; i++) {
2320 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2321 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2322 }
2323
2324 draw_image = draw_llvm_variant_key_images(key);
2325 memset(draw_image, 0,
2326 key->nr_images * sizeof *draw_image);
2327 for (i = 0; i < key->nr_images; i++) {
2328 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2329 llvm->draw->images[PIPE_SHADER_VERTEX][i]);
2330 }
2331 return key;
2332 }
2333
2334
2335 void
2336 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2337 {
2338 unsigned i;
2339 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2340 struct draw_image_static_state *image = draw_llvm_variant_key_images(key);
2341 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2342 debug_printf("clip_xy = %u\n", key->clip_xy);
2343 debug_printf("clip_z = %u\n", key->clip_z);
2344 debug_printf("clip_user = %u\n", key->clip_user);
2345 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2346 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2347 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2348 debug_printf("has_gs = %u\n", key->has_gs);
2349 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2350
2351 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2352 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2353 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2354 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2355 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2356 }
2357
2358 for (i = 0 ; i < key->nr_sampler_views; i++) {
2359 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2360 }
2361
2362 for (i = 0 ; i < key->nr_images; i++)
2363 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2364 }
2365
2366
2367 void
2368 draw_llvm_set_mapped_texture(struct draw_context *draw,
2369 enum pipe_shader_type shader_stage,
2370 unsigned sview_idx,
2371 uint32_t width, uint32_t height, uint32_t depth,
2372 uint32_t first_level, uint32_t last_level,
2373 const void *base_ptr,
2374 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2375 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2376 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2377 {
2378 unsigned j;
2379 struct draw_jit_texture *jit_tex;
2380
2381 assert(shader_stage == PIPE_SHADER_VERTEX ||
2382 shader_stage == PIPE_SHADER_GEOMETRY);
2383
2384 if (shader_stage == PIPE_SHADER_VERTEX) {
2385 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2386
2387 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2388 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2389 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2390
2391 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2392 } else {
2393 assert(0);
2394 return;
2395 }
2396
2397 jit_tex->width = width;
2398 jit_tex->height = height;
2399 jit_tex->depth = depth;
2400 jit_tex->first_level = first_level;
2401 jit_tex->last_level = last_level;
2402 jit_tex->base = base_ptr;
2403
2404 for (j = first_level; j <= last_level; j++) {
2405 jit_tex->mip_offsets[j] = mip_offsets[j];
2406 jit_tex->row_stride[j] = row_stride[j];
2407 jit_tex->img_stride[j] = img_stride[j];
2408 }
2409 }
2410
2411 void
2412 draw_llvm_set_mapped_image(struct draw_context *draw,
2413 enum pipe_shader_type shader_stage,
2414 unsigned idx,
2415 uint32_t width, uint32_t height, uint32_t depth,
2416 const void *base_ptr,
2417 uint32_t row_stride,
2418 uint32_t img_stride)
2419 {
2420 struct draw_jit_image *jit_image;
2421
2422 assert(shader_stage == PIPE_SHADER_VERTEX ||
2423 shader_stage == PIPE_SHADER_GEOMETRY);
2424
2425 if (shader_stage == PIPE_SHADER_VERTEX) {
2426 assert(idx < ARRAY_SIZE(draw->llvm->jit_context.images));
2427
2428 jit_image = &draw->llvm->jit_context.images[idx];
2429 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2430 assert(idx < ARRAY_SIZE(draw->llvm->gs_jit_context.images));
2431
2432 jit_image = &draw->llvm->gs_jit_context.images[idx];
2433 } else {
2434 assert(0);
2435 return;
2436 }
2437
2438 jit_image->width = width;
2439 jit_image->height = height;
2440 jit_image->depth = depth;
2441 jit_image->base = base_ptr;
2442
2443 jit_image->row_stride = row_stride;
2444 jit_image->img_stride = img_stride;
2445 }
2446
2447
2448 void
2449 draw_llvm_set_sampler_state(struct draw_context *draw,
2450 enum pipe_shader_type shader_type)
2451 {
2452 unsigned i;
2453
2454 if (shader_type == PIPE_SHADER_VERTEX) {
2455 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2456 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2457
2458 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2459 const struct pipe_sampler_state *s
2460 = draw->samplers[PIPE_SHADER_VERTEX][i];
2461 jit_sam->min_lod = s->min_lod;
2462 jit_sam->max_lod = s->max_lod;
2463 jit_sam->lod_bias = s->lod_bias;
2464 COPY_4V(jit_sam->border_color, s->border_color.f);
2465 }
2466 }
2467 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2468 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2469 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2470
2471 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2472 const struct pipe_sampler_state *s
2473 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2474 jit_sam->min_lod = s->min_lod;
2475 jit_sam->max_lod = s->max_lod;
2476 jit_sam->lod_bias = s->lod_bias;
2477 COPY_4V(jit_sam->border_color, s->border_color.f);
2478 }
2479 }
2480 }
2481 }
2482
2483
2484 void
2485 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2486 {
2487 struct draw_llvm *llvm = variant->llvm;
2488
2489 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2490 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2491 variant->shader->variants_cached, llvm->nr_variants);
2492 }
2493
2494 gallivm_destroy(variant->gallivm);
2495
2496 remove_from_list(&variant->list_item_local);
2497 variant->shader->variants_cached--;
2498 remove_from_list(&variant->list_item_global);
2499 llvm->nr_variants--;
2500 FREE(variant);
2501 }
2502
2503
2504 /**
2505 * Create LLVM types for various structures.
2506 */
2507 static void
2508 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2509 {
2510 struct gallivm_state *gallivm = var->gallivm;
2511 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
2512
2513 texture_type = create_jit_texture_type(gallivm, "texture");
2514 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2515 image_type = create_jit_image_type(gallivm, "image");
2516
2517 context_type = create_gs_jit_context_type(gallivm,
2518 var->shader->base.vector_length,
2519 texture_type, sampler_type,
2520 image_type,
2521 "draw_gs_jit_context");
2522 var->context_ptr_type = LLVMPointerType(context_type, 0);
2523
2524 var->input_array_type = create_gs_jit_input_type(gallivm);
2525 }
2526
2527 static LLVMTypeRef
2528 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2529 {
2530 if (!variant->context_ptr_type)
2531 create_gs_jit_types(variant);
2532 return variant->context_ptr_type;
2533 }
2534
2535 static LLVMValueRef
2536 generate_mask_value(struct draw_gs_llvm_variant *variant,
2537 struct lp_type gs_type)
2538 {
2539 struct gallivm_state *gallivm = variant->gallivm;
2540 LLVMBuilderRef builder = gallivm->builder;
2541 struct lp_type mask_type = lp_int_type(gs_type);
2542 LLVMValueRef num_prims;
2543 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2544 unsigned i;
2545
2546 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2547 variant->num_prims);
2548 for (i = 0; i < gs_type.length; i++) {
2549 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2550 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2551 }
2552 mask_val = lp_build_compare(gallivm, mask_type,
2553 PIPE_FUNC_GREATER, num_prims, mask_val);
2554
2555 return mask_val;
2556 }
2557
2558 static void
2559 draw_gs_llvm_generate(struct draw_llvm *llvm,
2560 struct draw_gs_llvm_variant *variant)
2561 {
2562 struct gallivm_state *gallivm = variant->gallivm;
2563 LLVMContextRef context = gallivm->context;
2564 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2565 LLVMTypeRef arg_types[7];
2566 LLVMTypeRef func_type;
2567 LLVMValueRef variant_func;
2568 LLVMValueRef context_ptr;
2569 LLVMValueRef prim_id_ptr;
2570 LLVMBasicBlockRef block;
2571 LLVMBuilderRef builder;
2572 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2573 struct lp_build_sampler_soa *sampler = 0;
2574 struct lp_build_image_soa *image = NULL;
2575 struct lp_build_context bld;
2576 struct lp_bld_tgsi_system_values system_values;
2577 char func_name[64];
2578 struct lp_type gs_type;
2579 unsigned i;
2580 struct draw_gs_llvm_iface gs_iface;
2581 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2582 LLVMValueRef consts_ptr, num_consts_ptr;
2583 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2584 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2585 struct lp_build_mask_context mask;
2586 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2587 unsigned vector_length = variant->shader->base.vector_length;
2588
2589 memset(&system_values, 0, sizeof(system_values));
2590 memset(&outputs, 0, sizeof(outputs));
2591
2592 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2593 variant->shader->variants_cached);
2594
2595 assert(variant->vertex_header_ptr_type);
2596
2597 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2598 arg_types[1] = variant->input_array_type; /* input */
2599 arg_types[2] = LLVMPointerType(variant->vertex_header_ptr_type, 0); /* vertex_header */
2600 arg_types[3] = int32_type; /* num_prims */
2601 arg_types[4] = int32_type; /* instance_id */
2602 arg_types[5] = LLVMPointerType(
2603 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2604 arg_types[6] = int32_type;
2605
2606 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2607
2608 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2609
2610 variant->function = variant_func;
2611
2612 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2613
2614 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2615 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2616 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2617
2618 context_ptr = LLVMGetParam(variant_func, 0);
2619 input_array = LLVMGetParam(variant_func, 1);
2620 io_ptr = LLVMGetParam(variant_func, 2);
2621 num_prims = LLVMGetParam(variant_func, 3);
2622 system_values.instance_id = LLVMGetParam(variant_func, 4);
2623 prim_id_ptr = LLVMGetParam(variant_func, 5);
2624 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2625
2626 lp_build_name(context_ptr, "context");
2627 lp_build_name(input_array, "input");
2628 lp_build_name(io_ptr, "io");
2629 lp_build_name(num_prims, "num_prims");
2630 lp_build_name(system_values.instance_id, "instance_id");
2631 lp_build_name(prim_id_ptr, "prim_id_ptr");
2632 lp_build_name(system_values.invocation_id, "invocation_id");
2633
2634 variant->context_ptr = context_ptr;
2635 variant->io_ptr = io_ptr;
2636 variant->num_prims = num_prims;
2637
2638 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2639 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2640 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2641 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2642 gs_iface.input = input_array;
2643 gs_iface.variant = variant;
2644
2645 /*
2646 * Function body
2647 */
2648
2649 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2650 builder = gallivm->builder;
2651 LLVMPositionBuilderAtEnd(builder, block);
2652
2653 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2654
2655 memset(&gs_type, 0, sizeof gs_type);
2656 gs_type.floating = TRUE; /* floating point values */
2657 gs_type.sign = TRUE; /* values are signed */
2658 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2659 gs_type.width = 32; /* 32-bit float */
2660 gs_type.length = vector_length;
2661
2662 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2663 num_consts_ptr =
2664 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2665
2666 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2667 num_ssbos_ptr =
2668 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2669
2670 /* code generated texture sampling */
2671 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2672 image = draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant->key));
2673 mask_val = generate_mask_value(variant, gs_type);
2674 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2675
2676 if (gs_info->uses_primid) {
2677 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2678 }
2679
2680 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2681 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2682 tgsi_dump(tokens, 0);
2683 else
2684 nir_print_shader(llvm->draw->gs.geometry_shader->state.ir.nir, stderr);
2685 draw_gs_llvm_dump_variant_key(&variant->key);
2686 }
2687
2688 struct lp_build_tgsi_params params;
2689 memset(&params, 0, sizeof(params));
2690
2691 params.type = gs_type;
2692 params.mask = &mask;
2693 params.consts_ptr = consts_ptr;
2694 params.const_sizes_ptr = num_consts_ptr;
2695 params.system_values = &system_values;
2696 params.context_ptr = context_ptr;
2697 params.sampler = sampler;
2698 params.info = &llvm->draw->gs.geometry_shader->info;
2699 params.gs_iface = (const struct lp_build_gs_iface *)&gs_iface;
2700 params.ssbo_ptr = ssbos_ptr;
2701 params.ssbo_sizes_ptr = num_ssbos_ptr;
2702 params.image = image;
2703
2704 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2705 lp_build_tgsi_soa(variant->gallivm,
2706 tokens,
2707 &params,
2708 outputs);
2709 else
2710 lp_build_nir_soa(variant->gallivm,
2711 llvm->draw->gs.geometry_shader->state.ir.nir,
2712 &params,
2713 outputs);
2714
2715 sampler->destroy(sampler);
2716 image->destroy(image);
2717
2718 lp_build_mask_end(&mask);
2719
2720 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2721
2722 gallivm_verify_function(gallivm, variant_func);
2723 }
2724
2725
2726 struct draw_gs_llvm_variant *
2727 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2728 unsigned num_outputs,
2729 const struct draw_gs_llvm_variant_key *key)
2730 {
2731 struct draw_gs_llvm_variant *variant;
2732 struct llvm_geometry_shader *shader =
2733 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2734 LLVMTypeRef vertex_header;
2735 char module_name[64];
2736
2737 variant = MALLOC(sizeof *variant +
2738 shader->variant_key_size -
2739 sizeof variant->key);
2740 if (!variant)
2741 return NULL;
2742
2743 variant->llvm = llvm;
2744 variant->shader = shader;
2745
2746 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2747 variant->shader->variants_cached);
2748
2749 variant->gallivm = gallivm_create(module_name, llvm->context);
2750
2751 create_gs_jit_types(variant);
2752
2753 memcpy(&variant->key, key, shader->variant_key_size);
2754
2755 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2756
2757 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2758
2759 draw_gs_llvm_generate(llvm, variant);
2760
2761 gallivm_compile_module(variant->gallivm);
2762
2763 variant->jit_func = (draw_gs_jit_func)
2764 gallivm_jit_function(variant->gallivm, variant->function);
2765
2766 gallivm_free_ir(variant->gallivm);
2767
2768 variant->list_item_global.base = variant;
2769 variant->list_item_local.base = variant;
2770 /*variant->no = */shader->variants_created++;
2771 variant->list_item_global.base = variant;
2772
2773 return variant;
2774 }
2775
2776 void
2777 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2778 {
2779 struct draw_llvm *llvm = variant->llvm;
2780
2781 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2782 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2783 variant->shader->variants_cached, llvm->nr_gs_variants);
2784 }
2785
2786 gallivm_destroy(variant->gallivm);
2787
2788 remove_from_list(&variant->list_item_local);
2789 variant->shader->variants_cached--;
2790 remove_from_list(&variant->list_item_global);
2791 llvm->nr_gs_variants--;
2792 FREE(variant);
2793 }
2794
2795 struct draw_gs_llvm_variant_key *
2796 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2797 {
2798 unsigned i;
2799 struct draw_gs_llvm_variant_key *key;
2800 struct draw_sampler_static_state *draw_sampler;
2801 struct draw_image_static_state *draw_image;
2802
2803 key = (struct draw_gs_llvm_variant_key *)store;
2804
2805 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2806
2807 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2808
2809 /* All variants of this shader will have the same value for
2810 * nr_samplers. Not yet trying to compact away holes in the
2811 * sampler array.
2812 */
2813 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2814 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2815 key->nr_sampler_views =
2816 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2817 }
2818 else {
2819 key->nr_sampler_views = key->nr_samplers;
2820 }
2821
2822 key->nr_images = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2823
2824 draw_sampler = key->samplers;
2825
2826 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2827
2828 for (i = 0 ; i < key->nr_samplers; i++) {
2829 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2830 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2831 }
2832 for (i = 0 ; i < key->nr_sampler_views; i++) {
2833 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2834 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2835 }
2836
2837 draw_image = draw_gs_llvm_variant_key_images(key);
2838 memset(draw_image, 0,
2839 key->nr_images * sizeof *draw_image);
2840 for (i = 0; i < key->nr_images; i++) {
2841 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2842 llvm->draw->images[PIPE_SHADER_GEOMETRY][i]);
2843 }
2844 return key;
2845 }
2846
2847 void
2848 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2849 {
2850 unsigned i;
2851 struct draw_sampler_static_state *sampler = key->samplers;
2852 struct draw_image_static_state *image = draw_gs_llvm_variant_key_images(key);
2853 for (i = 0 ; i < key->nr_sampler_views; i++) {
2854 debug_printf("sampler[%i].src_format = %s\n", i,
2855 util_format_name(sampler[i].texture_state.format));
2856 }
2857
2858 for (i = 0 ; i < key->nr_images; i++)
2859 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2860
2861 }