1 /**************************************************************************
3 * Copyright 2010 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
28 #include "draw_llvm.h"
30 #include "draw_context.h"
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_bitarit.h"
37 #include "gallivm/lp_bld_gather.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_const.h"
40 #include "gallivm/lp_bld_swizzle.h"
41 #include "gallivm/lp_bld_struct.h"
42 #include "gallivm/lp_bld_type.h"
43 #include "gallivm/lp_bld_flow.h"
44 #include "gallivm/lp_bld_debug.h"
45 #include "gallivm/lp_bld_tgsi.h"
46 #include "gallivm/lp_bld_printf.h"
47 #include "gallivm/lp_bld_intr.h"
48 #include "gallivm/lp_bld_init.h"
49 #include "gallivm/lp_bld_type.h"
50 #include "gallivm/lp_bld_pack.h"
51 #include "gallivm/lp_bld_format.h"
53 #include "tgsi/tgsi_exec.h"
54 #include "tgsi/tgsi_dump.h"
56 #include "util/u_math.h"
57 #include "util/u_pointer.h"
58 #include "util/u_string.h"
59 #include "util/simple_list.h"
66 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*var
);
69 struct draw_gs_llvm_iface
{
70 struct lp_build_gs_iface base
;
72 struct draw_gs_llvm_variant
*variant
;
76 static inline const struct draw_gs_llvm_iface
*
77 draw_gs_llvm_iface(const struct lp_build_gs_iface
*iface
)
79 return (const struct draw_gs_llvm_iface
*)iface
;
83 * Create LLVM type for draw_vertex_buffer.
86 create_jit_dvbuffer_type(struct gallivm_state
*gallivm
,
87 const char *struct_name
)
89 LLVMTargetDataRef target
= gallivm
->target
;
90 LLVMTypeRef dvbuffer_type
;
91 LLVMTypeRef elem_types
[DRAW_JIT_DVBUFFER_NUM_FIELDS
];
92 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
94 elem_types
[DRAW_JIT_DVBUFFER_MAP
] =
95 LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
96 elem_types
[DRAW_JIT_DVBUFFER_SIZE
] = int32_type
;
98 dvbuffer_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
99 ARRAY_SIZE(elem_types
), 0);
101 (void) target
; /* silence unused var warning for non-debug build */
102 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, map
,
103 target
, dvbuffer_type
,
104 DRAW_JIT_DVBUFFER_MAP
);
105 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, size
,
106 target
, dvbuffer_type
,
107 DRAW_JIT_DVBUFFER_SIZE
);
109 return dvbuffer_type
;
113 * Create LLVM type for struct draw_jit_texture
116 create_jit_texture_type(struct gallivm_state
*gallivm
, const char *struct_name
)
118 LLVMTargetDataRef target
= gallivm
->target
;
119 LLVMTypeRef texture_type
;
120 LLVMTypeRef elem_types
[DRAW_JIT_TEXTURE_NUM_FIELDS
];
121 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
123 elem_types
[DRAW_JIT_TEXTURE_WIDTH
] =
124 elem_types
[DRAW_JIT_TEXTURE_HEIGHT
] =
125 elem_types
[DRAW_JIT_TEXTURE_DEPTH
] =
126 elem_types
[DRAW_JIT_TEXTURE_FIRST_LEVEL
] =
127 elem_types
[DRAW_JIT_TEXTURE_LAST_LEVEL
] = int32_type
;
128 elem_types
[DRAW_JIT_TEXTURE_BASE
] =
129 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
130 elem_types
[DRAW_JIT_TEXTURE_ROW_STRIDE
] =
131 elem_types
[DRAW_JIT_TEXTURE_IMG_STRIDE
] =
132 elem_types
[DRAW_JIT_TEXTURE_MIP_OFFSETS
] =
133 LLVMArrayType(int32_type
, PIPE_MAX_TEXTURE_LEVELS
);
135 texture_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
136 ARRAY_SIZE(elem_types
), 0);
138 (void) target
; /* silence unused var warning for non-debug build */
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, width
,
140 target
, texture_type
,
141 DRAW_JIT_TEXTURE_WIDTH
);
142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, height
,
143 target
, texture_type
,
144 DRAW_JIT_TEXTURE_HEIGHT
);
145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, depth
,
146 target
, texture_type
,
147 DRAW_JIT_TEXTURE_DEPTH
);
148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, base
,
149 target
, texture_type
,
150 DRAW_JIT_TEXTURE_BASE
);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, row_stride
,
152 target
, texture_type
,
153 DRAW_JIT_TEXTURE_ROW_STRIDE
);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, img_stride
,
155 target
, texture_type
,
156 DRAW_JIT_TEXTURE_IMG_STRIDE
);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, first_level
,
158 target
, texture_type
,
159 DRAW_JIT_TEXTURE_FIRST_LEVEL
);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, last_level
,
161 target
, texture_type
,
162 DRAW_JIT_TEXTURE_LAST_LEVEL
);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, mip_offsets
,
164 target
, texture_type
,
165 DRAW_JIT_TEXTURE_MIP_OFFSETS
);
167 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture
, target
, texture_type
);
174 * Create LLVM type for struct draw_jit_sampler
177 create_jit_sampler_type(struct gallivm_state
*gallivm
, const char *struct_name
)
179 LLVMTargetDataRef target
= gallivm
->target
;
180 LLVMTypeRef sampler_type
;
181 LLVMTypeRef elem_types
[DRAW_JIT_SAMPLER_NUM_FIELDS
];
183 elem_types
[DRAW_JIT_SAMPLER_MIN_LOD
] =
184 elem_types
[DRAW_JIT_SAMPLER_MAX_LOD
] =
185 elem_types
[DRAW_JIT_SAMPLER_LOD_BIAS
] = LLVMFloatTypeInContext(gallivm
->context
);
186 elem_types
[DRAW_JIT_SAMPLER_BORDER_COLOR
] =
187 LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
189 sampler_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
190 ARRAY_SIZE(elem_types
), 0);
192 (void) target
; /* silence unused var warning for non-debug build */
193 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, min_lod
,
194 target
, sampler_type
,
195 DRAW_JIT_SAMPLER_MIN_LOD
);
196 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, max_lod
,
197 target
, sampler_type
,
198 DRAW_JIT_SAMPLER_MAX_LOD
);
199 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, lod_bias
,
200 target
, sampler_type
,
201 DRAW_JIT_SAMPLER_LOD_BIAS
);
202 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, border_color
,
203 target
, sampler_type
,
204 DRAW_JIT_SAMPLER_BORDER_COLOR
);
206 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler
, target
, sampler_type
);
212 * Create LLVM type for struct draw_jit_texture
215 create_jit_image_type(struct gallivm_state
*gallivm
, const char *struct_name
)
217 LLVMTargetDataRef target
= gallivm
->target
;
218 LLVMTypeRef image_type
;
219 LLVMTypeRef elem_types
[DRAW_JIT_IMAGE_NUM_FIELDS
];
220 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
222 elem_types
[DRAW_JIT_IMAGE_WIDTH
] =
223 elem_types
[DRAW_JIT_IMAGE_HEIGHT
] =
224 elem_types
[DRAW_JIT_IMAGE_DEPTH
] =
225 elem_types
[DRAW_JIT_IMAGE_ROW_STRIDE
] =
226 elem_types
[DRAW_JIT_IMAGE_IMG_STRIDE
] = int32_type
;
227 elem_types
[DRAW_JIT_IMAGE_BASE
] =
228 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
230 image_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
231 ARRAY_SIZE(elem_types
), 0);
233 (void) target
; /* silence unused var warning for non-debug build */
234 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, width
,
236 DRAW_JIT_IMAGE_WIDTH
);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, height
,
239 DRAW_JIT_IMAGE_HEIGHT
);
240 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, depth
,
242 DRAW_JIT_IMAGE_DEPTH
);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, base
,
245 DRAW_JIT_IMAGE_BASE
);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, row_stride
,
248 DRAW_JIT_IMAGE_ROW_STRIDE
);
249 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, img_stride
,
251 DRAW_JIT_IMAGE_IMG_STRIDE
);
253 LP_CHECK_STRUCT_SIZE(struct draw_jit_image
, target
, image_type
);
259 * Create LLVM type for struct draw_jit_context
262 create_jit_context_type(struct gallivm_state
*gallivm
,
263 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
264 LLVMTypeRef image_type
,
265 const char *struct_name
)
267 LLVMTargetDataRef target
= gallivm
->target
;
268 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
269 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
270 LLVMTypeRef elem_types
[DRAW_JIT_CTX_NUM_FIELDS
];
271 LLVMTypeRef context_type
;
273 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* vs_constants */
274 LP_MAX_TGSI_CONST_BUFFERS
);
275 elem_types
[1] = LLVMArrayType(int_type
, /* num_vs_constants */
276 LP_MAX_TGSI_CONST_BUFFERS
);
277 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
278 DRAW_TOTAL_CLIP_PLANES
), 0);
279 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
280 elem_types
[4] = LLVMArrayType(texture_type
,
281 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
282 elem_types
[5] = LLVMArrayType(sampler_type
,
283 PIPE_MAX_SAMPLERS
); /* samplers */
284 elem_types
[6] = LLVMArrayType(image_type
,
285 PIPE_MAX_SHADER_IMAGES
); /* images */
286 elem_types
[7] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* vs_ssbo */
287 LP_MAX_TGSI_SHADER_BUFFERS
);
288 elem_types
[8] = LLVMArrayType(int_type
, /* num_vs_ssbos */
289 LP_MAX_TGSI_SHADER_BUFFERS
);
290 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
291 ARRAY_SIZE(elem_types
), 0);
293 (void) target
; /* silence unused var warning for non-debug build */
294 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_constants
,
295 target
, context_type
, DRAW_JIT_CTX_CONSTANTS
);
296 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, num_vs_constants
,
297 target
, context_type
, DRAW_JIT_CTX_NUM_CONSTANTS
);
298 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, planes
,
299 target
, context_type
, DRAW_JIT_CTX_PLANES
);
300 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, viewports
,
301 target
, context_type
, DRAW_JIT_CTX_VIEWPORT
);
302 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, textures
,
303 target
, context_type
,
304 DRAW_JIT_CTX_TEXTURES
);
305 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, samplers
,
306 target
, context_type
,
307 DRAW_JIT_CTX_SAMPLERS
);
308 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, images
,
309 target
, context_type
, DRAW_JIT_CTX_IMAGES
);
310 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_ssbos
,
311 target
, context_type
, DRAW_JIT_CTX_SSBOS
);
312 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, num_vs_ssbos
,
313 target
, context_type
, DRAW_JIT_CTX_NUM_SSBOS
);
314 LP_CHECK_STRUCT_SIZE(struct draw_jit_context
,
315 target
, context_type
);
322 * Create LLVM type for struct draw_gs_jit_context
325 create_gs_jit_context_type(struct gallivm_state
*gallivm
,
326 unsigned vector_length
,
327 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
328 LLVMTypeRef image_type
,
329 const char *struct_name
)
331 LLVMTargetDataRef target
= gallivm
->target
;
332 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
333 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
334 LLVMTypeRef elem_types
[DRAW_GS_JIT_CTX_NUM_FIELDS
];
335 LLVMTypeRef context_type
;
337 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* constants */
338 LP_MAX_TGSI_CONST_BUFFERS
);
339 elem_types
[1] = LLVMArrayType(int_type
, /* num_constants */
340 LP_MAX_TGSI_CONST_BUFFERS
);
341 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
342 DRAW_TOTAL_CLIP_PLANES
), 0);
343 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
345 elem_types
[4] = LLVMArrayType(texture_type
,
346 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
347 elem_types
[5] = LLVMArrayType(sampler_type
,
348 PIPE_MAX_SAMPLERS
); /* samplers */
349 elem_types
[6] = LLVMArrayType(image_type
,
350 PIPE_MAX_SHADER_IMAGES
); /* images */
351 elem_types
[7] = LLVMPointerType(LLVMPointerType(int_type
, 0), 0);
352 elem_types
[8] = LLVMPointerType(LLVMVectorType(int_type
,
354 elem_types
[9] = LLVMPointerType(LLVMVectorType(int_type
,
357 elem_types
[10] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* ssbos */
358 LP_MAX_TGSI_SHADER_BUFFERS
);
359 elem_types
[11] = LLVMArrayType(int_type
, /* num_ssbos */
360 LP_MAX_TGSI_SHADER_BUFFERS
);
362 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
363 ARRAY_SIZE(elem_types
), 0);
365 (void) target
; /* silence unused var warning for non-debug build */
366 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, constants
,
367 target
, context_type
, DRAW_GS_JIT_CTX_CONSTANTS
);
368 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, num_constants
,
369 target
, context_type
, DRAW_GS_JIT_CTX_NUM_CONSTANTS
);
370 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, planes
,
371 target
, context_type
, DRAW_GS_JIT_CTX_PLANES
);
372 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, viewports
,
373 target
, context_type
, DRAW_GS_JIT_CTX_VIEWPORT
);
374 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, textures
,
375 target
, context_type
,
376 DRAW_GS_JIT_CTX_TEXTURES
);
377 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, samplers
,
378 target
, context_type
,
379 DRAW_GS_JIT_CTX_SAMPLERS
);
380 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, prim_lengths
,
381 target
, context_type
,
382 DRAW_GS_JIT_CTX_PRIM_LENGTHS
);
383 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_vertices
,
384 target
, context_type
,
385 DRAW_GS_JIT_CTX_EMITTED_VERTICES
);
386 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_prims
,
387 target
, context_type
,
388 DRAW_GS_JIT_CTX_EMITTED_PRIMS
);
389 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, ssbos
,
390 target
, context_type
, DRAW_GS_JIT_CTX_SSBOS
);
391 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, num_ssbos
,
392 target
, context_type
, DRAW_GS_JIT_CTX_NUM_SSBOS
);
393 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, images
,
394 target
, context_type
, DRAW_GS_JIT_CTX_IMAGES
);
395 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context
,
396 target
, context_type
);
403 create_gs_jit_input_type(struct gallivm_state
*gallivm
)
405 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
406 LLVMTypeRef input_array
;
408 input_array
= LLVMVectorType(float_type
, TGSI_NUM_CHANNELS
); /* num primitives */
409 input_array
= LLVMArrayType(input_array
, TGSI_NUM_CHANNELS
); /* num channels */
410 input_array
= LLVMArrayType(input_array
, PIPE_MAX_SHADER_INPUTS
); /* num attrs per vertex */
411 input_array
= LLVMPointerType(input_array
, 0); /* num vertices per prim */
417 * Create LLVM type for struct pipe_vertex_buffer
420 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
,
421 const char *struct_name
)
423 LLVMTargetDataRef target
= gallivm
->target
;
424 LLVMTypeRef elem_types
[4];
427 elem_types
[0] = LLVMInt16TypeInContext(gallivm
->context
);
428 elem_types
[1] = LLVMInt8TypeInContext(gallivm
->context
);
429 elem_types
[2] = LLVMInt32TypeInContext(gallivm
->context
);
430 elem_types
[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
432 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
433 ARRAY_SIZE(elem_types
), 0);
435 (void) target
; /* silence unused var warning for non-debug build */
436 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
438 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, is_user_buffer
,
440 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
442 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer
.resource
,
445 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
452 * Create LLVM type for struct vertex_header;
455 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
457 LLVMTargetDataRef target
= gallivm
->target
;
458 LLVMTypeRef elem_types
[3];
459 LLVMTypeRef vertex_header
;
460 char struct_name
[24];
462 snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
464 elem_types
[DRAW_JIT_VERTEX_VERTEX_ID
] = LLVMIntTypeInContext(gallivm
->context
, 32);
465 elem_types
[DRAW_JIT_VERTEX_CLIP_POS
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
466 elem_types
[DRAW_JIT_VERTEX_DATA
] = LLVMArrayType(elem_types
[1], data_elems
);
468 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
469 ARRAY_SIZE(elem_types
), 0);
471 /* these are bit-fields and we can't take address of them
472 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
473 target, vertex_header,
474 DRAW_JIT_VERTEX_CLIPMASK);
475 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
476 target, vertex_header,
477 DRAW_JIT_VERTEX_EDGEFLAG);
478 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
479 target, vertex_header,
480 DRAW_JIT_VERTEX_PAD);
481 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
482 target, vertex_header,
483 DRAW_JIT_VERTEX_VERTEX_ID);
485 (void) target
; /* silence unused var warning for non-debug build */
486 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, clip_pos
,
487 target
, vertex_header
,
488 DRAW_JIT_VERTEX_CLIP_POS
);
489 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, data
,
490 target
, vertex_header
,
491 DRAW_JIT_VERTEX_DATA
);
493 assert(LLVMABISizeOfType(target
, vertex_header
) ==
494 offsetof(struct vertex_header
, data
[data_elems
]));
496 return vertex_header
;
501 * Create LLVM types for various structures.
504 create_jit_types(struct draw_llvm_variant
*variant
)
506 struct gallivm_state
*gallivm
= variant
->gallivm
;
507 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
,
510 texture_type
= create_jit_texture_type(gallivm
, "texture");
511 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
512 image_type
= create_jit_image_type(gallivm
, "image");
514 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
517 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
519 buffer_type
= create_jit_dvbuffer_type(gallivm
, "draw_vertex_buffer");
520 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
522 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
523 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
528 get_context_ptr_type(struct draw_llvm_variant
*variant
)
530 if (!variant
->context_ptr_type
)
531 create_jit_types(variant
);
532 return variant
->context_ptr_type
;
537 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
539 if (!variant
->buffer_ptr_type
)
540 create_jit_types(variant
);
541 return variant
->buffer_ptr_type
;
546 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
548 if (!variant
->vb_ptr_type
)
549 create_jit_types(variant
);
550 return variant
->vb_ptr_type
;
554 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
556 if (!variant
->vertex_header_ptr_type
)
557 create_jit_types(variant
);
558 return variant
->vertex_header_ptr_type
;
563 * Create per-context LLVM info.
566 draw_llvm_create(struct draw_context
*draw
, LLVMContextRef context
)
568 struct draw_llvm
*llvm
;
570 if (!lp_build_init())
573 llvm
= CALLOC_STRUCT( draw_llvm
);
579 llvm
->context
= context
;
580 if (!llvm
->context
) {
581 llvm
->context
= LLVMContextCreate();
582 llvm
->context_owned
= true;
587 llvm
->nr_variants
= 0;
588 make_empty_list(&llvm
->vs_variants_list
);
590 llvm
->nr_gs_variants
= 0;
591 make_empty_list(&llvm
->gs_variants_list
);
596 draw_llvm_destroy(llvm
);
602 * Free per-context LLVM info.
605 draw_llvm_destroy(struct draw_llvm
*llvm
)
607 if (llvm
->context_owned
)
608 LLVMContextDispose(llvm
->context
);
609 llvm
->context
= NULL
;
611 /* XXX free other draw_llvm data? */
617 * Create LLVM-generated code for a vertex shader.
619 struct draw_llvm_variant
*
620 draw_llvm_create_variant(struct draw_llvm
*llvm
,
622 const struct draw_llvm_variant_key
*key
)
624 struct draw_llvm_variant
*variant
;
625 struct llvm_vertex_shader
*shader
=
626 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
627 LLVMTypeRef vertex_header
;
628 char module_name
[64];
630 variant
= MALLOC(sizeof *variant
+
631 shader
->variant_key_size
-
632 sizeof variant
->key
);
636 variant
->llvm
= llvm
;
637 variant
->shader
= shader
;
639 snprintf(module_name
, sizeof(module_name
), "draw_llvm_vs_variant%u",
640 variant
->shader
->variants_cached
);
642 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
);
644 create_jit_types(variant
);
646 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
648 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
649 tgsi_dump(llvm
->draw
->vs
.vertex_shader
->state
.tokens
, 0);
650 draw_llvm_dump_variant_key(&variant
->key
);
653 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
655 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
657 draw_llvm_generate(llvm
, variant
);
659 gallivm_compile_module(variant
->gallivm
);
661 variant
->jit_func
= (draw_jit_vert_func
)
662 gallivm_jit_function(variant
->gallivm
, variant
->function
);
664 gallivm_free_ir(variant
->gallivm
);
666 variant
->list_item_global
.base
= variant
;
667 variant
->list_item_local
.base
= variant
;
668 /*variant->no = */shader
->variants_created
++;
669 variant
->list_item_global
.base
= variant
;
676 generate_vs(struct draw_llvm_variant
*variant
,
677 LLVMBuilderRef builder
,
678 struct lp_type vs_type
,
679 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
680 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
681 const struct lp_bld_tgsi_system_values
*system_values
,
682 LLVMValueRef context_ptr
,
683 const struct lp_build_sampler_soa
*draw_sampler
,
684 const struct lp_build_image_soa
*draw_image
,
685 boolean clamp_vertex_color
,
686 struct lp_build_mask_context
*bld_mask
)
688 struct draw_llvm
*llvm
= variant
->llvm
;
689 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
690 LLVMValueRef consts_ptr
=
691 draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
692 LLVMValueRef num_consts_ptr
=
693 draw_jit_context_num_vs_constants(variant
->gallivm
, context_ptr
);
694 LLVMValueRef ssbos_ptr
=
695 draw_jit_context_vs_ssbos(variant
->gallivm
, context_ptr
);
696 LLVMValueRef num_ssbos_ptr
=
697 draw_jit_context_num_vs_ssbos(variant
->gallivm
, context_ptr
);
699 struct lp_build_tgsi_params params
;
700 memset(¶ms
, 0, sizeof(params
));
702 params
.type
= vs_type
;
703 params
.mask
= bld_mask
;
704 params
.consts_ptr
= consts_ptr
;
705 params
.const_sizes_ptr
= num_consts_ptr
;
706 params
.system_values
= system_values
;
707 params
.inputs
= inputs
;
708 params
.context_ptr
= context_ptr
;
709 params
.sampler
= draw_sampler
;
710 params
.info
= &llvm
->draw
->vs
.vertex_shader
->info
;
711 params
.ssbo_ptr
= ssbos_ptr
;
712 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
713 params
.image
= draw_image
;
715 lp_build_tgsi_soa(variant
->gallivm
,
722 unsigned chan
, attrib
;
723 struct lp_build_context bld
;
724 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
725 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
727 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
728 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
729 if (outputs
[attrib
][chan
]) {
730 switch (info
->output_semantic_name
[attrib
]) {
731 case TGSI_SEMANTIC_COLOR
:
732 case TGSI_SEMANTIC_BCOLOR
:
733 if (clamp_vertex_color
) {
734 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
735 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
736 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
748 fetch_instanced(struct gallivm_state
*gallivm
,
749 const struct util_format_description
*format_desc
,
750 struct lp_type vs_type
,
751 LLVMValueRef vb_stride
,
752 LLVMValueRef map_ptr
,
753 LLVMValueRef buffer_size_adj
,
754 LLVMValueRef
*inputs
,
757 LLVMTypeRef i32_t
= LLVMInt32TypeInContext(gallivm
->context
);
758 LLVMTypeRef aosf_t
, aosi_t
;
759 LLVMValueRef zero
= LLVMConstNull(i32_t
);
760 LLVMBuilderRef builder
= gallivm
->builder
;
761 LLVMValueRef stride
, buffer_overflowed
, aos
, index_valid
;
764 aosf_t
= lp_build_vec_type(gallivm
, lp_float32_vec4_type());
765 aosi_t
= lp_build_vec_type(gallivm
, lp_int32_vec4_type());
767 /* This mul can overflow. Wraparound is ok. */
768 stride
= LLVMBuildMul(builder
, vb_stride
, index
, "");
770 buffer_overflowed
= LLVMBuildICmp(builder
, LLVMIntUGE
,
771 stride
, buffer_size_adj
,
772 "buffer_overflowed");
775 lp_build_print_value(gallivm
, " instance index = ", index
);
776 lp_build_print_value(gallivm
, " buffer overflowed = ", buffer_overflowed
);
779 index_valid
= LLVMBuildNot(builder
, buffer_overflowed
, "");
780 index_valid
= LLVMBuildSExt(builder
, index_valid
, i32_t
, "");
781 stride
= LLVMBuildAnd(builder
, stride
, index_valid
, "");
783 aos
= lp_build_fetch_rgba_aos(gallivm
,
785 lp_float32_vec4_type(),
791 index_valid
= lp_build_broadcast(gallivm
, aosi_t
, index_valid
);
792 aos
= LLVMBuildBitCast(builder
, aos
, aosi_t
, "");
793 aos
= LLVMBuildAnd(builder
, aos
, index_valid
, "");
794 aos
= LLVMBuildBitCast(builder
, aos
, aosf_t
, "");
796 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
797 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
798 inputs
[i
] = lp_build_extract_broadcast(gallivm
,
799 lp_float32_vec4_type(),
800 vs_type
, aos
, index
);
806 fetch_vector(struct gallivm_state
*gallivm
,
807 const struct util_format_description
*format_desc
,
808 struct lp_type vs_type
,
809 LLVMValueRef vb_stride
,
810 LLVMValueRef map_ptr
,
811 LLVMValueRef buffer_size_adj
,
812 LLVMValueRef
*inputs
,
813 LLVMValueRef indices
)
815 LLVMBuilderRef builder
= gallivm
->builder
;
816 struct lp_build_context blduivec
;
817 struct lp_type fetch_type
= vs_type
;
818 LLVMValueRef offset
, valid_mask
;
821 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
823 vb_stride
= lp_build_broadcast_scalar(&blduivec
, vb_stride
);
824 buffer_size_adj
= lp_build_broadcast_scalar(&blduivec
, buffer_size_adj
);
826 /* This mul can overflow. Wraparound is ok. */
827 offset
= lp_build_mul(&blduivec
, vb_stride
, indices
);
829 valid_mask
= lp_build_compare(gallivm
, blduivec
.type
,
830 PIPE_FUNC_LESS
, offset
, buffer_size_adj
);
832 /* not valid elements use offset 0 */
833 offset
= LLVMBuildAnd(builder
, offset
, valid_mask
, "");
836 lp_build_print_value(gallivm
, " indices = ", indices
);
837 lp_build_print_value(gallivm
, " offsets = ", offset
);
838 lp_build_print_value(gallivm
, " valid_mask = ", valid_mask
);
842 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
843 * This should always produce better code.
846 /* The type handling is annoying here... */
847 if (format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
&&
848 format_desc
->channel
[0].pure_integer
) {
849 if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_SIGNED
) {
850 fetch_type
= lp_type_int_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
852 else if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
853 fetch_type
= lp_type_uint_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
857 lp_build_fetch_rgba_soa(gallivm
, format_desc
,
858 fetch_type
, FALSE
, map_ptr
, offset
,
859 blduivec
.zero
, blduivec
.zero
,
862 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
863 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
864 lp_build_vec_type(gallivm
, vs_type
), "");
867 /* out-of-bound fetches return all zeros */
868 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
869 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
], blduivec
.vec_type
, "");
870 inputs
[i
] = LLVMBuildAnd(builder
, inputs
[i
], valid_mask
, "");
871 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
872 lp_build_vec_type(gallivm
, vs_type
), "");
878 store_aos(struct gallivm_state
*gallivm
,
883 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
884 LLVMBuilderRef builder
= gallivm
->builder
;
885 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
886 LLVMValueRef indices
[3];
888 indices
[0] = lp_build_const_int32(gallivm
, 0);
890 indices
[2] = lp_build_const_int32(gallivm
, 0);
892 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
893 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
896 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
899 /* Unaligned store due to the vertex header */
900 LLVMSetAlignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
904 * Adjust the mask to architecture endianess. The mask will the store in struct:
906 * struct vertex_header {
907 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
908 * unsigned edgeflag:1;
910 * unsigned vertex_id:16;
914 * On little-endian machine nothing needs to done, however on bit-endian machine
915 * the mask's fields need to be adjusted with the algorithm:
917 * uint32_t reverse (uint32_t x)
919 * return (x >> 16) | // vertex_id
920 * ((x & 0x3fff) << 18) | // clipmask
921 * ((x & 0x4000) << 3) | // pad
922 * ((x & 0x8000) << 1); // edgeflag
926 adjust_mask(struct gallivm_state
*gallivm
,
929 #ifdef PIPE_ARCH_BIG_ENDIAN
930 LLVMBuilderRef builder
= gallivm
->builder
;
931 LLVMValueRef vertex_id
;
932 LLVMValueRef clipmask
;
934 LLVMValueRef edgeflag
;
936 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
937 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
938 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
940 pad
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
941 pad
= LLVMBuildShl(builder
, pad
, lp_build_const_int32(gallivm
, 3), "");
943 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
944 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 1), "");
946 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
948 mask
= LLVMBuildOr(builder
, mask
, pad
, "");
950 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
956 store_aos_array(struct gallivm_state
*gallivm
,
957 struct lp_type soa_type
,
959 LLVMValueRef
*indices
,
963 LLVMValueRef clipmask
,
964 boolean need_edgeflag
)
966 LLVMBuilderRef builder
= gallivm
->builder
;
967 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
968 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
969 LLVMValueRef linear_inds
[LP_MAX_VECTOR_WIDTH
/ 32];
970 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
971 int vector_length
= soa_type
.length
;
974 debug_assert(TGSI_NUM_CHANNELS
== 4);
976 for (i
= 0; i
< vector_length
; i
++) {
977 linear_inds
[i
] = lp_build_const_int32(gallivm
, i
);
979 inds
[i
] = indices
[i
];
981 inds
[i
] = linear_inds
[i
];
983 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
987 /* store vertex header for each of the n vertices */
988 LLVMValueRef val
, cliptmp
;
989 int vertex_id_pad_edgeflag
;
991 /* If this assertion fails, it means we need to update the bit twidding
992 * code here. See struct vertex_header in draw_private.h.
994 assert(DRAW_TOTAL_CLIP_PLANES
==14);
995 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
996 if (!need_edgeflag
) {
997 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
1000 vertex_id_pad_edgeflag
= (0xffff << 16);
1002 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
),
1003 vertex_id_pad_edgeflag
);
1004 /* OR with the clipmask */
1005 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
1006 for (i
= 0; i
< vector_length
; i
++) {
1007 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
1008 val
= LLVMBuildExtractElement(builder
, cliptmp
, linear_inds
[i
], "");
1009 val
= adjust_mask(gallivm
, val
);
1011 lp_build_printf(gallivm
, "io = %p, index %d, clipmask = %x\n",
1012 io_ptrs
[i
], inds
[i
], val
);
1014 LLVMBuildStore(builder
, val
, id_ptr
);
1018 /* store for each of the n vertices */
1019 for (i
= 0; i
< vector_length
; i
++) {
1020 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
1026 convert_to_aos(struct gallivm_state
*gallivm
,
1028 LLVMValueRef
*indices
,
1029 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1030 LLVMValueRef clipmask
,
1032 struct lp_type soa_type
,
1033 boolean need_edgeflag
)
1035 LLVMBuilderRef builder
= gallivm
->builder
;
1036 unsigned chan
, attrib
, i
;
1039 lp_build_printf(gallivm
, " # storing begin\n");
1041 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
1042 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
1043 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
1044 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1045 if (outputs
[attrib
][chan
]) {
1046 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
1047 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
1049 lp_build_printf(gallivm
, "output %d : %d ",
1050 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1052 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1054 lp_build_print_value(gallivm
, "val = ", out
);
1057 LLVMBuildBitCast(builder
, out
, lp_build_int_vec_type(gallivm
, soa_type
), "");
1059 lp_build_print_value(gallivm
, " ival = ", iv
);
1070 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
1071 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
1073 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
1075 for (i
= 0; i
< soa_type
.length
; ++i
) {
1076 aos
[i
] = lp_build_extract_range(gallivm
,
1077 soa
[i
% TGSI_NUM_CHANNELS
],
1078 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1083 store_aos_array(gallivm
,
1093 lp_build_printf(gallivm
, " # storing end\n");
1099 * Stores original vertex positions in clip coordinates
1102 store_clip(struct gallivm_state
*gallivm
,
1103 const struct lp_type vs_type
,
1104 LLVMValueRef io_ptr
,
1105 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1108 LLVMBuilderRef builder
= gallivm
->builder
;
1109 LLVMValueRef soa
[4];
1110 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
1111 LLVMValueRef indices
[2];
1112 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1113 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1114 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1115 LLVMTypeRef clip_ptr_type
=
1116 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
),
1121 indices
[1] = lp_build_const_int32(gallivm
, 0);
1123 for (i
= 0; i
< vs_type
.length
; i
++) {
1124 inds
[i
] = lp_build_const_int32(gallivm
, i
);
1125 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1128 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
1129 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
1130 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
1131 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
1133 for (i
= 0; i
< vs_type
.length
; i
++) {
1134 clip_ptrs
[i
] = draw_jit_header_clip_pos(gallivm
, io_ptrs
[i
]);
1137 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
1138 for (i
= 0; i
< vs_type
.length
; ++i
) {
1139 aos
[i
] = lp_build_extract_range(gallivm
,
1140 soa
[i
% TGSI_NUM_CHANNELS
],
1141 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1145 for (j
= 0; j
< vs_type
.length
; j
++) {
1146 LLVMValueRef clip_ptr
;
1148 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
1149 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
1151 /* Unaligned store */
1152 LLVMSetAlignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
1158 * Transforms the outputs for viewport mapping
1161 generate_viewport(struct draw_llvm_variant
*variant
,
1162 LLVMBuilderRef builder
,
1163 struct lp_type vs_type
,
1164 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1165 LLVMValueRef context_ptr
)
1168 struct gallivm_state
*gallivm
= variant
->gallivm
;
1169 struct lp_type f32_type
= vs_type
;
1170 const unsigned pos
= variant
->llvm
->draw
->vs
.position_output
;
1171 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1172 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn*/
1173 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
1174 LLVMValueRef vp_ptr
= draw_jit_context_viewports(gallivm
, context_ptr
);
1176 /* We treat pipe_viewport_state as a float array */
1177 const int scale_index_offset
= offsetof(struct pipe_viewport_state
, scale
) / sizeof(float);
1178 const int trans_index_offset
= offsetof(struct pipe_viewport_state
, translate
) / sizeof(float);
1180 /* for 1/w convention*/
1181 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
1182 LLVMBuildStore(builder
, out3
, outputs
[pos
][3]);
1184 /* Viewport Mapping */
1185 for (i
=0; i
<3; i
++) {
1186 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[pos
][i
], ""); /*x0 x1 .. xn*/
1189 LLVMValueRef scale_i
;
1190 LLVMValueRef trans_i
;
1193 index
= lp_build_const_int32(gallivm
, i
+ scale_index_offset
);
1194 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1196 index
= lp_build_const_int32(gallivm
, i
+ trans_index_offset
);
1197 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1199 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1200 LLVMBuildLoad(builder
, scale_i
, "scale"));
1201 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1202 LLVMBuildLoad(builder
, trans_i
, "trans"));
1205 out
= LLVMBuildFMul(builder
, out
, out3
, "");
1206 /* mult by scale, add translation */
1207 out
= lp_build_fmuladd(builder
, out
, scale
, trans
);
1209 /* store transformed outputs */
1210 LLVMBuildStore(builder
, out
, outputs
[pos
][i
]);
1217 * Returns clipmask as nxi32 bitmask for the n vertices
1220 generate_clipmask(struct draw_llvm
*llvm
,
1221 struct gallivm_state
*gallivm
,
1222 struct lp_type vs_type
,
1223 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1224 struct draw_llvm_variant_key
*key
,
1225 LLVMValueRef context_ptr
,
1226 boolean
*have_clipdist
)
1228 LLVMBuilderRef builder
= gallivm
->builder
;
1229 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
1230 LLVMValueRef test
, temp
;
1231 LLVMValueRef zero
, shift
;
1232 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
1233 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
1234 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
1235 struct lp_type f32_type
= vs_type
;
1236 struct lp_type i32_type
= lp_int_type(vs_type
);
1237 const unsigned pos
= llvm
->draw
->vs
.position_output
;
1238 const unsigned cv
= llvm
->draw
->vs
.clipvertex_output
;
1239 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
1240 boolean have_cd
= false;
1241 boolean clip_user
= key
->clip_user
;
1242 unsigned ucp_enable
= key
->ucp_enable
;
1245 cd
[0] = llvm
->draw
->vs
.ccdistance_output
[0];
1246 cd
[1] = llvm
->draw
->vs
.ccdistance_output
[1];
1248 if (cd
[0] != pos
|| cd
[1] != pos
)
1251 if (num_written_clipdistance
&& !clip_user
) {
1253 ucp_enable
= (1 << num_written_clipdistance
) - 1;
1256 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1257 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1258 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
1259 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
1262 * load clipvertex and position from correct locations.
1263 * if they are the same just load them once.
1265 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
1266 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
1267 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
1268 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
1270 if (clip_user
&& cv
!= pos
) {
1271 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
1272 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
1273 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
1274 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1283 * Be careful with the comparisons and NaNs (using llvm's unordered
1284 * comparisons here).
1286 /* Cliptest, for hardwired planes */
1288 * XXX should take guardband into account (currently not in key).
1289 * Otherwise might run the draw pipeline stages for nothing.
1293 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1295 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1299 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1300 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1301 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1302 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1303 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1306 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1307 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1308 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1309 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1312 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1313 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1314 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1315 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1316 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1320 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1321 if (key
->clip_halfz
) {
1323 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1324 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1325 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1329 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1330 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1331 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1332 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1335 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1336 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1337 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1338 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1342 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1343 LLVMValueRef indices
[3];
1344 LLVMValueRef is_nan_or_inf
;
1346 /* userclip planes */
1347 while (ucp_enable
) {
1348 unsigned plane_idx
= ffs(ucp_enable
)-1;
1349 ucp_enable
&= ~(1 << plane_idx
);
1352 if (have_cd
&& num_written_clipdistance
) {
1353 LLVMValueRef clipdist
;
1357 *have_clipdist
= TRUE
;
1359 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1361 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1363 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1364 is_nan_or_inf
= lp_build_is_inf_or_nan(gallivm
, vs_type
, clipdist
);
1365 test
= LLVMBuildOr(builder
, test
, is_nan_or_inf
, "");
1366 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1367 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1368 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1370 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1371 indices
[0] = lp_build_const_int32(gallivm
, 0);
1372 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1374 indices
[2] = lp_build_const_int32(gallivm
, 0);
1375 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1376 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1377 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1378 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1380 indices
[2] = lp_build_const_int32(gallivm
, 1);
1381 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1382 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1383 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1384 sum
= lp_build_fmuladd(builder
, planes
, cv_y
, sum
);
1386 indices
[2] = lp_build_const_int32(gallivm
, 2);
1387 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1388 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1389 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1390 sum
= lp_build_fmuladd(builder
, planes
, cv_z
, sum
);
1392 indices
[2] = lp_build_const_int32(gallivm
, 3);
1393 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1394 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1395 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1396 sum
= lp_build_fmuladd(builder
, planes
, cv_w
, sum
);
1398 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1399 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1400 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1401 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1405 if (key
->need_edgeflags
) {
1407 * This isn't really part of clipmask but stored the same in vertex
1408 * header later, so do it here.
1410 unsigned edge_attr
= llvm
->draw
->vs
.edgeflag_output
;
1411 LLVMValueRef one
= lp_build_const_vec(gallivm
, f32_type
, 1.0);
1412 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, outputs
[edge_attr
][0], "");
1413 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_EQUAL
, one
, edgeflag
);
1414 temp
= lp_build_const_int_vec(gallivm
, i32_type
,
1415 1LL << DRAW_TOTAL_CLIP_PLANES
);
1416 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1417 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1424 * Returns boolean if any clipping has occurred
1425 * Used zero/one i8 value to represent boolean
1428 clipmask_booli8(struct gallivm_state
*gallivm
,
1429 const struct lp_type vs_type
,
1430 LLVMValueRef clipmask_bool_ptr
,
1431 boolean edgeflag_in_clipmask
)
1433 LLVMBuilderRef builder
= gallivm
->builder
;
1434 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1435 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1437 struct lp_build_context bldivec
;
1439 lp_build_context_init(&bldivec
, gallivm
, lp_int_type(vs_type
));
1442 * We need to invert the edgeflag bit from the clipmask here
1443 * (because the result is really if we want to run the pipeline or not
1444 * and we (may) need it if edgeflag was 0).
1446 if (edgeflag_in_clipmask
) {
1447 LLVMValueRef edge
= lp_build_const_int_vec(gallivm
, bldivec
.type
,
1448 1LL << DRAW_TOTAL_CLIP_PLANES
);
1449 clipmask_bool
= LLVMBuildXor(builder
, clipmask_bool
, edge
, "");
1453 * XXX: probably should mask off bits from the mask which come from
1454 * vertices which were beyond the count (i.e. indices_valid for
1455 * linear fetches, for elts ones we don't have the correct mask
1456 * right now). Otherwise might run the pipeline for nothing,
1457 * though everything should still work.
1459 ret
= lp_build_any_true_range(&bldivec
, vs_type
.length
, clipmask_bool
);
1460 ret
= LLVMBuildZExt(builder
, ret
, int8_type
, "");
1465 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface
*gs_iface
,
1466 struct lp_build_context
* bld
,
1467 boolean is_vindex_indirect
,
1468 LLVMValueRef vertex_index
,
1469 boolean is_aindex_indirect
,
1470 LLVMValueRef attrib_index
,
1471 LLVMValueRef swizzle_index
)
1473 const struct draw_gs_llvm_iface
*gs
= draw_gs_llvm_iface(gs_iface
);
1474 struct gallivm_state
*gallivm
= bld
->gallivm
;
1475 LLVMBuilderRef builder
= gallivm
->builder
;
1476 LLVMValueRef indices
[3];
1478 struct lp_type type
= bld
->type
;
1480 if (is_vindex_indirect
|| is_aindex_indirect
) {
1483 for (i
= 0; i
< type
.length
; ++i
) {
1484 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
1485 LLVMValueRef vert_chan_index
= vertex_index
;
1486 LLVMValueRef attr_chan_index
= attrib_index
;
1487 LLVMValueRef channel_vec
, value
;
1489 if (is_vindex_indirect
) {
1490 vert_chan_index
= LLVMBuildExtractElement(builder
,
1491 vertex_index
, idx
, "");
1493 if (is_aindex_indirect
) {
1494 attr_chan_index
= LLVMBuildExtractElement(builder
,
1495 attrib_index
, idx
, "");
1498 indices
[0] = vert_chan_index
;
1499 indices
[1] = attr_chan_index
;
1500 indices
[2] = swizzle_index
;
1502 channel_vec
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1503 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
1504 value
= LLVMBuildExtractElement(builder
, channel_vec
, idx
, "");
1506 res
= LLVMBuildInsertElement(builder
, res
, value
, idx
, "");
1509 indices
[0] = vertex_index
;
1510 indices
[1] = attrib_index
;
1511 indices
[2] = swizzle_index
;
1513 res
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1514 res
= LLVMBuildLoad(builder
, res
, "");
1521 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface
*gs_base
,
1522 struct lp_build_context
* bld
,
1523 LLVMValueRef (*outputs
)[4],
1524 LLVMValueRef emitted_vertices_vec
)
1526 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1527 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1528 struct gallivm_state
*gallivm
= variant
->gallivm
;
1529 LLVMBuilderRef builder
= gallivm
->builder
;
1530 struct lp_type gs_type
= bld
->type
;
1531 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
1532 lp_int_type(gs_type
), 0);
1533 LLVMValueRef indices
[LP_MAX_VECTOR_LENGTH
];
1534 LLVMValueRef next_prim_offset
=
1535 lp_build_const_int32(gallivm
, variant
->shader
->base
.primitive_boundary
);
1536 LLVMValueRef io
= variant
->io_ptr
;
1538 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
1540 for (i
= 0; i
< gs_type
.length
; ++i
) {
1541 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1542 LLVMValueRef currently_emitted
=
1543 LLVMBuildExtractElement(builder
, emitted_vertices_vec
, ind
, "");
1544 indices
[i
] = LLVMBuildMul(builder
, ind
, next_prim_offset
, "");
1545 indices
[i
] = LLVMBuildAdd(builder
, indices
[i
], currently_emitted
, "");
1548 convert_to_aos(gallivm
, io
, indices
,
1550 gs_info
->num_outputs
, gs_type
,
1555 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface
*gs_base
,
1556 struct lp_build_context
* bld
,
1557 LLVMValueRef total_emitted_vertices_vec_ptr
,
1558 LLVMValueRef verts_per_prim_vec
,
1559 LLVMValueRef emitted_prims_vec
,
1560 LLVMValueRef mask_vec
)
1562 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1563 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1564 struct gallivm_state
*gallivm
= variant
->gallivm
;
1565 LLVMBuilderRef builder
= gallivm
->builder
;
1566 LLVMValueRef prim_lengts_ptr
=
1567 draw_gs_jit_prim_lengths(variant
->gallivm
, variant
->context_ptr
);
1570 for (i
= 0; i
< bld
->type
.length
; ++i
) {
1571 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1572 LLVMValueRef prims_emitted
=
1573 LLVMBuildExtractElement(builder
, emitted_prims_vec
, ind
, "");
1574 LLVMValueRef store_ptr
;
1575 LLVMValueRef num_vertices
=
1576 LLVMBuildExtractElement(builder
, verts_per_prim_vec
, ind
, "");
1578 store_ptr
= LLVMBuildGEP(builder
, prim_lengts_ptr
, &prims_emitted
, 1, "");
1579 store_ptr
= LLVMBuildLoad(builder
, store_ptr
, "");
1580 store_ptr
= LLVMBuildGEP(builder
, store_ptr
, &ind
, 1, "");
1581 LLVMBuildStore(builder
, num_vertices
, store_ptr
);
1586 draw_gs_llvm_epilogue(const struct lp_build_gs_iface
*gs_base
,
1587 LLVMValueRef total_emitted_vertices_vec
,
1588 LLVMValueRef emitted_prims_vec
)
1590 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1591 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1592 struct gallivm_state
*gallivm
= variant
->gallivm
;
1593 LLVMBuilderRef builder
= gallivm
->builder
;
1594 LLVMValueRef emitted_verts_ptr
=
1595 draw_gs_jit_emitted_vertices(gallivm
, variant
->context_ptr
);
1596 LLVMValueRef emitted_prims_ptr
=
1597 draw_gs_jit_emitted_prims(gallivm
, variant
->context_ptr
);
1598 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
1600 emitted_verts_ptr
= LLVMBuildGEP(builder
, emitted_verts_ptr
, &zero
, 0, "");
1601 emitted_prims_ptr
= LLVMBuildGEP(builder
, emitted_prims_ptr
, &zero
, 0, "");
1603 LLVMBuildStore(builder
, total_emitted_vertices_vec
, emitted_verts_ptr
);
1604 LLVMBuildStore(builder
, emitted_prims_vec
, emitted_prims_ptr
);
1608 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
)
1610 struct gallivm_state
*gallivm
= variant
->gallivm
;
1611 LLVMContextRef context
= gallivm
->context
;
1612 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1613 LLVMTypeRef arg_types
[11];
1614 unsigned num_arg_types
= ARRAY_SIZE(arg_types
);
1615 LLVMTypeRef func_type
;
1616 LLVMValueRef context_ptr
;
1617 LLVMBasicBlockRef block
;
1618 LLVMBuilderRef builder
;
1620 struct lp_type vs_type
;
1621 LLVMValueRef count
, fetch_elts
, start_or_maxelt
;
1622 LLVMValueRef vertex_id_offset
, start_instance
;
1623 LLVMValueRef stride
, step
, io_itr
;
1624 LLVMValueRef ind_vec
, start_vec
, have_elts
, fetch_max
, tmp
;
1625 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1626 LLVMValueRef vb_stride
[PIPE_MAX_ATTRIBS
];
1627 LLVMValueRef map_ptr
[PIPE_MAX_ATTRIBS
];
1628 LLVMValueRef buffer_size_adj
[PIPE_MAX_ATTRIBS
];
1629 LLVMValueRef instance_index
[PIPE_MAX_ATTRIBS
];
1630 LLVMValueRef fake_buf_ptr
, fake_buf
;
1632 struct draw_context
*draw
= llvm
->draw
;
1633 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1635 struct lp_build_context bld
, blduivec
;
1636 struct lp_build_loop_state lp_loop
;
1637 struct lp_build_if_state if_ctx
;
1638 const int vector_length
= lp_native_vector_width
/ 32;
1639 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1640 struct lp_build_sampler_soa
*sampler
= 0;
1641 struct lp_build_image_soa
*image
= NULL
;
1642 LLVMValueRef ret
, clipmask_bool_ptr
;
1643 struct draw_llvm_variant_key
*key
= &variant
->key
;
1644 /* If geometry shader is present we need to skip both the viewport
1645 * transformation and clipping otherwise the inputs to the geometry
1646 * shader will be incorrect.
1647 * The code can't handle vp transform when vs writes vp index neither
1648 * (though this would be fixable here, but couldn't just broadcast
1651 const boolean bypass_viewport
= key
->has_gs
|| key
->bypass_viewport
||
1652 vs_info
->writes_viewport_index
;
1653 const boolean enable_cliptest
= !key
->has_gs
&& (key
->clip_xy
||
1656 key
->need_edgeflags
);
1657 LLVMValueRef variant_func
;
1658 const unsigned pos
= draw
->vs
.position_output
;
1659 const unsigned cv
= draw
->vs
.clipvertex_output
;
1660 boolean have_clipdist
= FALSE
;
1661 struct lp_bld_tgsi_system_values system_values
;
1663 memset(&system_values
, 0, sizeof(system_values
));
1665 snprintf(func_name
, sizeof(func_name
), "draw_llvm_vs_variant%u",
1666 variant
->shader
->variants_cached
);
1669 arg_types
[i
++] = get_context_ptr_type(variant
); /* context */
1670 arg_types
[i
++] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1671 arg_types
[i
++] = get_buffer_ptr_type(variant
); /* vbuffers */
1672 arg_types
[i
++] = int32_type
; /* count */
1673 arg_types
[i
++] = int32_type
; /* start/fetch_elt_max */
1674 arg_types
[i
++] = int32_type
; /* stride */
1675 arg_types
[i
++] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1676 arg_types
[i
++] = int32_type
; /* instance_id */
1677 arg_types
[i
++] = int32_type
; /* vertex_id_offset */
1678 arg_types
[i
++] = int32_type
; /* start_instance */
1679 arg_types
[i
++] = LLVMPointerType(int32_type
, 0); /* fetch_elts */
1681 func_type
= LLVMFunctionType(LLVMInt8TypeInContext(context
),
1682 arg_types
, num_arg_types
, 0);
1684 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1685 variant
->function
= variant_func
;
1687 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1688 for (i
= 0; i
< num_arg_types
; ++i
)
1689 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1690 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
1692 context_ptr
= LLVMGetParam(variant_func
, 0);
1693 io_ptr
= LLVMGetParam(variant_func
, 1);
1694 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1695 count
= LLVMGetParam(variant_func
, 3);
1697 * XXX: the maxelt part is unused. Not really useful, since we cannot
1698 * get index buffer overflows due to vsplit (which provides its own
1699 * elts buffer, with a different size than what's passed in here).
1701 start_or_maxelt
= LLVMGetParam(variant_func
, 4);
1703 * XXX: stride is actually unused. The stride we use is strictly calculated
1704 * from the number of outputs (including the draw_extra outputs).
1705 * Should probably fix some day (we need a new vs just because of extra
1706 * outputs which the generated vs won't touch).
1708 stride
= LLVMGetParam(variant_func
, 5);
1709 vb_ptr
= LLVMGetParam(variant_func
, 6);
1710 system_values
.instance_id
= LLVMGetParam(variant_func
, 7);
1711 vertex_id_offset
= LLVMGetParam(variant_func
, 8);
1712 start_instance
= LLVMGetParam(variant_func
, 9);
1713 fetch_elts
= LLVMGetParam(variant_func
, 10);
1715 lp_build_name(context_ptr
, "context");
1716 lp_build_name(io_ptr
, "io");
1717 lp_build_name(vbuffers_ptr
, "vbuffers");
1718 lp_build_name(count
, "count");
1719 lp_build_name(start_or_maxelt
, "start_or_maxelt");
1720 lp_build_name(stride
, "stride");
1721 lp_build_name(vb_ptr
, "vb");
1722 lp_build_name(system_values
.instance_id
, "instance_id");
1723 lp_build_name(vertex_id_offset
, "vertex_id_offset");
1724 lp_build_name(start_instance
, "start_instance");
1725 lp_build_name(fetch_elts
, "fetch_elts");
1731 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
1732 builder
= gallivm
->builder
;
1733 LLVMPositionBuilderAtEnd(builder
, block
);
1735 memset(&vs_type
, 0, sizeof vs_type
);
1736 vs_type
.floating
= TRUE
; /* floating point values */
1737 vs_type
.sign
= TRUE
; /* values are signed */
1738 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1739 vs_type
.width
= 32; /* 32-bit float */
1740 vs_type
.length
= vector_length
;
1742 lp_build_context_init(&bld
, gallivm
, lp_type_uint(32));
1743 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
1745 /* hold temporary "bool" clipmask */
1746 clipmask_bool_ptr
= lp_build_alloca(gallivm
, blduivec
.vec_type
, "");
1748 fake_buf
= lp_build_alloca_undef(gallivm
,
1749 LLVMVectorType(LLVMInt64TypeInContext(context
), 4), "");
1750 fake_buf
= LLVMBuildBitCast(builder
, fake_buf
,
1751 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
1752 fake_buf_ptr
= LLVMBuildGEP(builder
, fake_buf
, &bld
.zero
, 1, "");
1754 /* code generated texture sampling */
1755 sampler
= draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key
));
1757 image
= draw_llvm_image_soa_create(draw_llvm_variant_key_images(key
));
1759 step
= lp_build_const_int32(gallivm
, vector_length
);
1761 ind_vec
= blduivec
.undef
;
1762 for (i
= 0; i
< vs_type
.length
; i
++) {
1763 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
1764 ind_vec
= LLVMBuildInsertElement(builder
, ind_vec
, index
, index
, "");
1767 have_elts
= LLVMBuildICmp(builder
, LLVMIntNE
,
1768 LLVMConstPointerNull(arg_types
[10]), fetch_elts
, "");
1770 fetch_max
= LLVMBuildSub(builder
, count
, bld
.one
, "fetch_max");
1771 fetch_max
= lp_build_broadcast_scalar(&blduivec
, fetch_max
);
1773 * Only needed for non-indexed path.
1775 start_vec
= lp_build_broadcast_scalar(&blduivec
, start_or_maxelt
);
1778 * Pre-calculate everything which is constant per shader invocation.
1780 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
1781 LLVMValueRef vb_buffer_offset
, buffer_size
, temp_ptr
;
1782 LLVMValueRef vb_info
, vbuffer_ptr
, buf_offset
, ofbit
;
1783 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
1784 LLVMValueRef vb_index
=
1785 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
1786 LLVMValueRef bsize
= lp_build_const_int32(gallivm
,
1787 util_format_get_blocksize(velem
->src_format
));
1788 LLVMValueRef src_offset
= lp_build_const_int32(gallivm
,
1790 struct lp_build_if_state if_ctx
;
1792 if (velem
->src_format
!= PIPE_FORMAT_NONE
) {
1793 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
, &vb_index
, 1, "");
1794 vb_info
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
1795 vb_stride
[j
] = draw_jit_vbuffer_stride(gallivm
, vb_info
);
1796 vb_stride
[j
] = LLVMBuildZExt(gallivm
->builder
, vb_stride
[j
],
1797 LLVMInt32TypeInContext(context
), "");
1798 vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vb_info
);
1799 map_ptr
[j
] = draw_jit_dvbuffer_map(gallivm
, vbuffer_ptr
);
1800 buffer_size
= draw_jit_dvbuffer_size(gallivm
, vbuffer_ptr
);
1804 * We'll set buffer_size_adj to zero if we have of, so it will
1805 * always overflow later automatically without having to keep ofbit.
1806 * Overflows (with normal wraparound) doing the actual offset
1807 * calculation should be ok, just not for the buffer size calc.
1808 * It would also be possible to detect such overflows and return
1809 * zeros if that happens, but this would be more complex.
1811 buf_offset
= lp_build_add(&bld
, vb_buffer_offset
, src_offset
);
1812 tmp
= lp_build_sub(&bld
, bsize
, bld
.one
);
1813 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size
, tmp
,
1815 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size_adj
[j
],
1816 buf_offset
, &ofbit
);
1819 * We can't easily set fake vertex buffers outside the generated code.
1820 * Hence, set fake vertex buffers here instead basically, so fetch
1821 * code can always fetch using offset 0, eliminating all control flow
1822 * inside the main loop.
1823 * (Alternatively, could have control flow per vector skipping fetch
1824 * if ofbit is true.)
1826 if (velem
->instance_divisor
) {
1828 * Index is equal to the start instance plus the number of current
1829 * instance divided by the divisor. In this case we compute it as:
1830 * index = start_instance + (instance_id / divisor).
1831 * Note we could actually do the fetch here, outside the loop -
1832 * it's all constant, hopefully llvm recognizes this.
1834 LLVMValueRef current_instance
;
1835 current_instance
= LLVMBuildUDiv(builder
, system_values
.instance_id
,
1836 lp_build_const_int32(gallivm
,
1837 velem
->instance_divisor
),
1838 "instance_divisor");
1839 instance_index
[j
] = lp_build_uadd_overflow(gallivm
, start_instance
,
1840 current_instance
, &ofbit
);
1843 buffer_size_adj
[j
] = LLVMBuildSelect(builder
, ofbit
, bld
.zero
,
1844 buffer_size_adj
[j
], "");
1846 temp_ptr
= lp_build_alloca_undef(gallivm
,
1847 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
1849 lp_build_if(&if_ctx
, gallivm
, ofbit
);
1851 LLVMBuildStore(builder
, fake_buf_ptr
, temp_ptr
);
1853 lp_build_else(&if_ctx
);
1855 map_ptr
[j
] = LLVMBuildGEP(builder
, map_ptr
[j
], &buf_offset
, 1, "");
1856 LLVMBuildStore(builder
, map_ptr
[j
], temp_ptr
);
1858 lp_build_endif(&if_ctx
);
1859 map_ptr
[j
] = LLVMBuildLoad(builder
, temp_ptr
, "map_ptr");
1862 lp_build_printf(gallivm
, "velem %d, vbuf index = %u, vb_stride = %u\n",
1863 lp_build_const_int32(gallivm
, j
),
1864 vb_index
, vb_stride
[j
]);
1865 lp_build_printf(gallivm
,
1866 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1867 vb_buffer_offset
, src_offset
, buf_offset
);
1868 lp_build_printf(gallivm
, " buffer size = %u, blocksize = %u\n",
1869 buffer_size
, bsize
);
1870 lp_build_printf(gallivm
, " instance_id = %u\n", system_values
.instance_id
);
1875 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
1877 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
1879 LLVMValueRef clipmask
; /* holds the clipmask value */
1880 LLVMValueRef true_index_array
, index_store
;
1881 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
1883 io_itr
= lp_loop
.counter
;
1885 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
1887 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
1888 io_itr
, io
, lp_loop
.counter
);
1891 true_index_array
= lp_build_broadcast_scalar(&blduivec
, lp_loop
.counter
);
1892 true_index_array
= LLVMBuildAdd(builder
, true_index_array
, ind_vec
, "");
1894 LLVMValueRef exec_mask
= lp_build_cmp(&blduivec
, PIPE_FUNC_LEQUAL
, true_index_array
, fetch_max
);
1896 * Limit indices to fetch_max, otherwise might try to access indices
1897 * beyond index buffer (or rather vsplit elt buffer) size.
1898 * Could probably safely (?) skip this for non-indexed draws and
1899 * simplify things minimally (by removing it could combine the ind_vec
1900 * and start_vec adds). I think the only effect for non-indexed draws will
1901 * be that for the invalid elements they will be all fetched from the
1902 * same location as the last valid one, but noone should really care.
1904 true_index_array
= lp_build_min(&blduivec
, true_index_array
, fetch_max
);
1906 index_store
= lp_build_alloca_undef(gallivm
, blduivec
.vec_type
, "index_store");
1908 lp_build_if(&if_ctx
, gallivm
, have_elts
);
1911 * Note: you'd expect some comparison/clamp against fetch_elt_max
1913 * There used to be one here but it was incorrect: overflow was
1914 * detected if index > fetch_elt_max - but the correct condition
1915 * would be index >= fetch_elt_max (since this is just size of elts
1916 * buffer / element size).
1917 * Using the correct condition however will cause failures - due to
1918 * vsplit/vcache code which rebases indices. So, as an example, if
1919 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1920 * replace all invalid indices with 0 - which in case of elt_bias
1921 * not being zero will get a different fetch index than the valid
1922 * index 0. So, just rely on vsplit code preventing out-of-bounds
1923 * fetches. This is also why it's safe to do elts fetch even if there
1924 * was no index buffer bound - the real buffer is never seen here, at
1925 * least not if there are index buffer overflows...
1929 * XXX should not have to do this, as scale can be handled
1930 * natively by loads (hits asserts though).
1932 tmp
= lp_build_shl_imm(&blduivec
, true_index_array
, 2);
1933 fetch_elts
= LLVMBuildBitCast(builder
, fetch_elts
,
1934 LLVMPointerType(LLVMInt8TypeInContext(context
),
1936 tmp
= lp_build_gather(gallivm
, vs_type
.length
,
1938 fetch_elts
, tmp
, FALSE
);
1939 LLVMBuildStore(builder
, tmp
, index_store
);
1941 lp_build_else(&if_ctx
);
1943 tmp
= LLVMBuildAdd(builder
, true_index_array
, start_vec
, "");
1944 LLVMBuildStore(builder
, tmp
, index_store
);
1946 lp_build_endif(&if_ctx
);
1948 true_index_array
= LLVMBuildLoad(builder
, index_store
, "");
1950 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
1951 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
1952 const struct util_format_description
*format_desc
=
1953 util_format_description(velem
->src_format
);
1955 if (format_desc
->format
== PIPE_FORMAT_NONE
) {
1956 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1957 inputs
[j
][i
] = lp_build_zero(gallivm
, vs_type
);
1960 else if (velem
->instance_divisor
) {
1961 fetch_instanced(gallivm
, format_desc
, vs_type
,
1962 vb_stride
[j
], map_ptr
[j
],
1964 inputs
[j
], instance_index
[j
]);
1967 fetch_vector(gallivm
, format_desc
, vs_type
,
1968 vb_stride
[j
], map_ptr
[j
],
1970 inputs
[j
], true_index_array
);
1974 struct lp_build_mask_context mask
;
1976 lp_build_mask_begin(&mask
, gallivm
, vs_type
, exec_mask
);
1977 /* In the paths with elts vertex id has to be unaffected by the
1978 * index bias and because indices inside our elements array have
1979 * already had index bias applied we need to subtract it here to
1980 * get back to the original index.
1981 * in the linear paths vertex id has to be unaffected by the
1982 * original start index and because we abuse the 'start' variable
1983 * to either represent the actual start index or the index at which
1984 * the primitive was split (we split rendering into chunks of at
1985 * most 4095-vertices) we need to back out the original start
1986 * index out of our vertex id here.
1988 system_values
.basevertex
= lp_build_broadcast_scalar(&blduivec
,
1990 system_values
.vertex_id
= true_index_array
;
1991 system_values
.vertex_id_nobase
= LLVMBuildSub(builder
, true_index_array
,
1992 system_values
.basevertex
, "");
1994 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
1995 generate_vs(variant
,
2004 key
->clamp_vertex_color
,
2007 lp_build_mask_end(&mask
);
2008 if (pos
!= -1 && cv
!= -1) {
2009 /* store original positions in clip before further manipulation */
2010 store_clip(gallivm
, vs_type
, io
, outputs
, pos
);
2013 if (enable_cliptest
) {
2014 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
2015 /* allocate clipmask, assign it integer type */
2016 clipmask
= generate_clipmask(llvm
,
2021 context_ptr
, &have_clipdist
);
2022 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
2023 /* store temporary clipping boolean value */
2024 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
2027 clipmask
= blduivec
.zero
;
2030 /* do viewport mapping */
2031 if (!bypass_viewport
) {
2032 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
2036 clipmask
= blduivec
.zero
;
2039 /* store clipmask in vertex header,
2040 * original positions in clip
2041 * and transformed positions in data
2043 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
2044 vs_info
->num_outputs
, vs_type
,
2045 enable_cliptest
&& key
->need_edgeflags
);
2047 lp_build_loop_end_cond(&lp_loop
, count
, step
, LLVMIntUGE
);
2049 sampler
->destroy(sampler
);
2050 image
->destroy(image
);
2052 /* return clipping boolean value for function */
2053 ret
= clipmask_booli8(gallivm
, vs_type
, clipmask_bool_ptr
,
2054 enable_cliptest
&& key
->need_edgeflags
);
2056 LLVMBuildRet(builder
, ret
);
2058 gallivm_verify_function(gallivm
, variant_func
);
2062 struct draw_llvm_variant_key
*
2063 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2066 struct draw_llvm_variant_key
*key
;
2067 struct draw_sampler_static_state
*draw_sampler
;
2068 struct draw_image_static_state
*draw_image
;
2070 key
= (struct draw_llvm_variant_key
*)store
;
2072 memset(key
, 0, offsetof(struct draw_llvm_variant_key
, vertex_element
[0]));
2074 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
2076 /* will have to rig this up properly later */
2077 key
->clip_xy
= llvm
->draw
->clip_xy
;
2078 key
->clip_z
= llvm
->draw
->clip_z
;
2079 key
->clip_user
= llvm
->draw
->clip_user
;
2080 key
->bypass_viewport
= llvm
->draw
->bypass_viewport
;
2081 key
->clip_halfz
= llvm
->draw
->rasterizer
->clip_halfz
;
2082 /* XXX assumes edgeflag output not at 0 */
2083 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
2084 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
2085 key
->has_gs
= llvm
->draw
->gs
.geometry_shader
!= NULL
;
2086 key
->num_outputs
= draw_total_vs_outputs(llvm
->draw
);
2088 /* All variants of this shader will have the same value for
2089 * nr_samplers. Not yet trying to compact away holes in the
2092 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2093 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2094 key
->nr_sampler_views
=
2095 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2098 key
->nr_sampler_views
= key
->nr_samplers
;
2101 key
->nr_images
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2103 /* Presumably all variants of the shader should have the same
2104 * number of vertex elements - ie the number of shader inputs.
2105 * NOTE: we NEED to store the needed number of needed inputs
2106 * here, not the number of provided elements to match keysize
2107 * (and the offset of sampler state in the key).
2108 * If we have excess number of vertex elements, this is valid,
2109 * but the excess ones don't matter.
2110 * If we don't have enough vertex elements (which looks not really
2111 * valid but we'll handle it gracefully) fill out missing ones with
2112 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2114 key
->nr_vertex_elements
=
2115 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
2117 if (llvm
->draw
->pt
.nr_vertex_elements
< key
->nr_vertex_elements
) {
2118 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2119 key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
);
2120 memset(key
->vertex_element
, 0,
2121 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
2123 memcpy(key
->vertex_element
,
2124 llvm
->draw
->pt
.vertex_element
,
2125 sizeof(struct pipe_vertex_element
) *
2126 MIN2(key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
));
2128 draw_sampler
= draw_llvm_variant_key_samplers(key
);
2129 memset(draw_sampler
, 0,
2130 MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2132 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2133 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2134 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
2136 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2137 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2138 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
2141 draw_image
= draw_llvm_variant_key_images(key
);
2142 memset(draw_image
, 0,
2143 key
->nr_images
* sizeof *draw_image
);
2144 for (i
= 0; i
< key
->nr_images
; i
++) {
2145 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
2146 llvm
->draw
->images
[PIPE_SHADER_VERTEX
][i
]);
2153 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
2156 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
2157 struct draw_image_static_state
*image
= draw_llvm_variant_key_images(key
);
2158 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
2159 debug_printf("clip_xy = %u\n", key
->clip_xy
);
2160 debug_printf("clip_z = %u\n", key
->clip_z
);
2161 debug_printf("clip_user = %u\n", key
->clip_user
);
2162 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
2163 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
2164 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
2165 debug_printf("has_gs = %u\n", key
->has_gs
);
2166 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
2168 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
2169 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
2170 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
2171 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
2172 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
2175 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2176 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
2179 for (i
= 0 ; i
< key
->nr_images
; i
++)
2180 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
2185 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
2186 enum pipe_shader_type shader_stage
,
2188 uint32_t width
, uint32_t height
, uint32_t depth
,
2189 uint32_t first_level
, uint32_t last_level
,
2190 const void *base_ptr
,
2191 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2192 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2193 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
2196 struct draw_jit_texture
*jit_tex
;
2198 assert(shader_stage
== PIPE_SHADER_VERTEX
||
2199 shader_stage
== PIPE_SHADER_GEOMETRY
);
2201 if (shader_stage
== PIPE_SHADER_VERTEX
) {
2202 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.textures
));
2204 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
2205 } else if (shader_stage
== PIPE_SHADER_GEOMETRY
) {
2206 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.textures
));
2208 jit_tex
= &draw
->llvm
->gs_jit_context
.textures
[sview_idx
];
2214 jit_tex
->width
= width
;
2215 jit_tex
->height
= height
;
2216 jit_tex
->depth
= depth
;
2217 jit_tex
->first_level
= first_level
;
2218 jit_tex
->last_level
= last_level
;
2219 jit_tex
->base
= base_ptr
;
2221 for (j
= first_level
; j
<= last_level
; j
++) {
2222 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
2223 jit_tex
->row_stride
[j
] = row_stride
[j
];
2224 jit_tex
->img_stride
[j
] = img_stride
[j
];
2229 draw_llvm_set_mapped_image(struct draw_context
*draw
,
2230 enum pipe_shader_type shader_stage
,
2232 uint32_t width
, uint32_t height
, uint32_t depth
,
2233 const void *base_ptr
,
2234 uint32_t row_stride
,
2235 uint32_t img_stride
)
2237 struct draw_jit_image
*jit_image
;
2239 assert(shader_stage
== PIPE_SHADER_VERTEX
||
2240 shader_stage
== PIPE_SHADER_GEOMETRY
);
2242 if (shader_stage
== PIPE_SHADER_VERTEX
) {
2243 assert(idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.images
));
2245 jit_image
= &draw
->llvm
->jit_context
.images
[idx
];
2246 } else if (shader_stage
== PIPE_SHADER_GEOMETRY
) {
2247 assert(idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.images
));
2249 jit_image
= &draw
->llvm
->gs_jit_context
.images
[idx
];
2255 jit_image
->width
= width
;
2256 jit_image
->height
= height
;
2257 jit_image
->depth
= depth
;
2258 jit_image
->base
= base_ptr
;
2260 jit_image
->row_stride
= row_stride
;
2261 jit_image
->img_stride
= img_stride
;
2266 draw_llvm_set_sampler_state(struct draw_context
*draw
,
2267 enum pipe_shader_type shader_type
)
2271 if (shader_type
== PIPE_SHADER_VERTEX
) {
2272 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
2273 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
2275 if (draw
->samplers
[PIPE_SHADER_VERTEX
][i
]) {
2276 const struct pipe_sampler_state
*s
2277 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
2278 jit_sam
->min_lod
= s
->min_lod
;
2279 jit_sam
->max_lod
= s
->max_lod
;
2280 jit_sam
->lod_bias
= s
->lod_bias
;
2281 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2284 } else if (shader_type
== PIPE_SHADER_GEOMETRY
) {
2285 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_GEOMETRY
]; i
++) {
2286 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->gs_jit_context
.samplers
[i
];
2288 if (draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]) {
2289 const struct pipe_sampler_state
*s
2290 = draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
];
2291 jit_sam
->min_lod
= s
->min_lod
;
2292 jit_sam
->max_lod
= s
->max_lod
;
2293 jit_sam
->lod_bias
= s
->lod_bias
;
2294 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2302 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
2304 struct draw_llvm
*llvm
= variant
->llvm
;
2306 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2307 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2308 variant
->shader
->variants_cached
, llvm
->nr_variants
);
2311 gallivm_destroy(variant
->gallivm
);
2313 remove_from_list(&variant
->list_item_local
);
2314 variant
->shader
->variants_cached
--;
2315 remove_from_list(&variant
->list_item_global
);
2316 llvm
->nr_variants
--;
2322 * Create LLVM types for various structures.
2325 create_gs_jit_types(struct draw_gs_llvm_variant
*var
)
2327 struct gallivm_state
*gallivm
= var
->gallivm
;
2328 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
2330 texture_type
= create_jit_texture_type(gallivm
, "texture");
2331 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
2332 image_type
= create_jit_image_type(gallivm
, "image");
2334 context_type
= create_gs_jit_context_type(gallivm
,
2335 var
->shader
->base
.vector_length
,
2336 texture_type
, sampler_type
,
2338 "draw_gs_jit_context");
2339 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
2341 var
->input_array_type
= create_gs_jit_input_type(gallivm
);
2345 get_gs_context_ptr_type(struct draw_gs_llvm_variant
*variant
)
2347 if (!variant
->context_ptr_type
)
2348 create_gs_jit_types(variant
);
2349 return variant
->context_ptr_type
;
2353 generate_mask_value(struct draw_gs_llvm_variant
*variant
,
2354 struct lp_type gs_type
)
2356 struct gallivm_state
*gallivm
= variant
->gallivm
;
2357 LLVMBuilderRef builder
= gallivm
->builder
;
2358 struct lp_type mask_type
= lp_int_type(gs_type
);
2359 LLVMValueRef num_prims
;
2360 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
2363 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
),
2364 variant
->num_prims
);
2365 for (i
= 0; i
< gs_type
.length
; i
++) {
2366 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
2367 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, idx
, idx
, "");
2369 mask_val
= lp_build_compare(gallivm
, mask_type
,
2370 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
2376 draw_gs_llvm_generate(struct draw_llvm
*llvm
,
2377 struct draw_gs_llvm_variant
*variant
)
2379 struct gallivm_state
*gallivm
= variant
->gallivm
;
2380 LLVMContextRef context
= gallivm
->context
;
2381 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
2382 LLVMTypeRef arg_types
[7];
2383 LLVMTypeRef func_type
;
2384 LLVMValueRef variant_func
;
2385 LLVMValueRef context_ptr
;
2386 LLVMValueRef prim_id_ptr
;
2387 LLVMBasicBlockRef block
;
2388 LLVMBuilderRef builder
;
2389 LLVMValueRef io_ptr
, input_array
, num_prims
, mask_val
;
2390 struct lp_build_sampler_soa
*sampler
= 0;
2391 struct lp_build_image_soa
*image
= NULL
;
2392 struct lp_build_context bld
;
2393 struct lp_bld_tgsi_system_values system_values
;
2395 struct lp_type gs_type
;
2397 struct draw_gs_llvm_iface gs_iface
;
2398 const struct tgsi_token
*tokens
= variant
->shader
->base
.state
.tokens
;
2399 LLVMValueRef consts_ptr
, num_consts_ptr
;
2400 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
2401 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
2402 struct lp_build_mask_context mask
;
2403 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
2404 unsigned vector_length
= variant
->shader
->base
.vector_length
;
2406 memset(&system_values
, 0, sizeof(system_values
));
2408 snprintf(func_name
, sizeof(func_name
), "draw_llvm_gs_variant%u",
2409 variant
->shader
->variants_cached
);
2411 assert(variant
->vertex_header_ptr_type
);
2413 arg_types
[0] = get_gs_context_ptr_type(variant
); /* context */
2414 arg_types
[1] = variant
->input_array_type
; /* input */
2415 arg_types
[2] = variant
->vertex_header_ptr_type
; /* vertex_header */
2416 arg_types
[3] = int32_type
; /* num_prims */
2417 arg_types
[4] = int32_type
; /* instance_id */
2418 arg_types
[5] = LLVMPointerType(
2419 LLVMVectorType(int32_type
, vector_length
), 0); /* prim_id_ptr */
2420 arg_types
[6] = int32_type
;
2422 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
2424 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2426 variant
->function
= variant_func
;
2428 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
2430 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2431 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2432 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2434 context_ptr
= LLVMGetParam(variant_func
, 0);
2435 input_array
= LLVMGetParam(variant_func
, 1);
2436 io_ptr
= LLVMGetParam(variant_func
, 2);
2437 num_prims
= LLVMGetParam(variant_func
, 3);
2438 system_values
.instance_id
= LLVMGetParam(variant_func
, 4);
2439 prim_id_ptr
= LLVMGetParam(variant_func
, 5);
2440 system_values
.invocation_id
= LLVMGetParam(variant_func
, 6);
2442 lp_build_name(context_ptr
, "context");
2443 lp_build_name(input_array
, "input");
2444 lp_build_name(io_ptr
, "io");
2445 lp_build_name(num_prims
, "num_prims");
2446 lp_build_name(system_values
.instance_id
, "instance_id");
2447 lp_build_name(prim_id_ptr
, "prim_id_ptr");
2448 lp_build_name(system_values
.invocation_id
, "invocation_id");
2450 variant
->context_ptr
= context_ptr
;
2451 variant
->io_ptr
= io_ptr
;
2452 variant
->num_prims
= num_prims
;
2454 gs_iface
.base
.fetch_input
= draw_gs_llvm_fetch_input
;
2455 gs_iface
.base
.emit_vertex
= draw_gs_llvm_emit_vertex
;
2456 gs_iface
.base
.end_primitive
= draw_gs_llvm_end_primitive
;
2457 gs_iface
.base
.gs_epilogue
= draw_gs_llvm_epilogue
;
2458 gs_iface
.input
= input_array
;
2459 gs_iface
.variant
= variant
;
2465 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
2466 builder
= gallivm
->builder
;
2467 LLVMPositionBuilderAtEnd(builder
, block
);
2469 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
2471 memset(&gs_type
, 0, sizeof gs_type
);
2472 gs_type
.floating
= TRUE
; /* floating point values */
2473 gs_type
.sign
= TRUE
; /* values are signed */
2474 gs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2475 gs_type
.width
= 32; /* 32-bit float */
2476 gs_type
.length
= vector_length
;
2478 consts_ptr
= draw_gs_jit_context_constants(variant
->gallivm
, context_ptr
);
2480 draw_gs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
2482 ssbos_ptr
= draw_gs_jit_context_ssbos(variant
->gallivm
, context_ptr
);
2484 draw_gs_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
2486 /* code generated texture sampling */
2487 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
);
2488 image
= draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant
->key
));
2489 mask_val
= generate_mask_value(variant
, gs_type
);
2490 lp_build_mask_begin(&mask
, gallivm
, gs_type
, mask_val
);
2492 if (gs_info
->uses_primid
) {
2493 system_values
.prim_id
= LLVMBuildLoad(builder
, prim_id_ptr
, "prim_id");
2496 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2497 tgsi_dump(tokens
, 0);
2498 draw_gs_llvm_dump_variant_key(&variant
->key
);
2501 struct lp_build_tgsi_params params
;
2502 memset(¶ms
, 0, sizeof(params
));
2504 params
.type
= gs_type
;
2505 params
.mask
= &mask
;
2506 params
.consts_ptr
= consts_ptr
;
2507 params
.const_sizes_ptr
= num_consts_ptr
;
2508 params
.system_values
= &system_values
;
2509 params
.context_ptr
= context_ptr
;
2510 params
.sampler
= sampler
;
2511 params
.info
= &llvm
->draw
->gs
.geometry_shader
->info
;
2512 params
.gs_iface
= (const struct lp_build_gs_iface
*)&gs_iface
;
2513 params
.ssbo_ptr
= ssbos_ptr
;
2514 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
2515 params
.image
= image
;
2517 lp_build_tgsi_soa(variant
->gallivm
,
2522 sampler
->destroy(sampler
);
2523 image
->destroy(image
);
2525 lp_build_mask_end(&mask
);
2527 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
2529 gallivm_verify_function(gallivm
, variant_func
);
2533 struct draw_gs_llvm_variant
*
2534 draw_gs_llvm_create_variant(struct draw_llvm
*llvm
,
2535 unsigned num_outputs
,
2536 const struct draw_gs_llvm_variant_key
*key
)
2538 struct draw_gs_llvm_variant
*variant
;
2539 struct llvm_geometry_shader
*shader
=
2540 llvm_geometry_shader(llvm
->draw
->gs
.geometry_shader
);
2541 LLVMTypeRef vertex_header
;
2542 char module_name
[64];
2544 variant
= MALLOC(sizeof *variant
+
2545 shader
->variant_key_size
-
2546 sizeof variant
->key
);
2550 variant
->llvm
= llvm
;
2551 variant
->shader
= shader
;
2553 snprintf(module_name
, sizeof(module_name
), "draw_llvm_gs_variant%u",
2554 variant
->shader
->variants_cached
);
2556 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
);
2558 create_gs_jit_types(variant
);
2560 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2562 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
2564 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
2566 draw_gs_llvm_generate(llvm
, variant
);
2568 gallivm_compile_module(variant
->gallivm
);
2570 variant
->jit_func
= (draw_gs_jit_func
)
2571 gallivm_jit_function(variant
->gallivm
, variant
->function
);
2573 gallivm_free_ir(variant
->gallivm
);
2575 variant
->list_item_global
.base
= variant
;
2576 variant
->list_item_local
.base
= variant
;
2577 /*variant->no = */shader
->variants_created
++;
2578 variant
->list_item_global
.base
= variant
;
2584 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant
*variant
)
2586 struct draw_llvm
*llvm
= variant
->llvm
;
2588 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2589 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2590 variant
->shader
->variants_cached
, llvm
->nr_gs_variants
);
2593 gallivm_destroy(variant
->gallivm
);
2595 remove_from_list(&variant
->list_item_local
);
2596 variant
->shader
->variants_cached
--;
2597 remove_from_list(&variant
->list_item_global
);
2598 llvm
->nr_gs_variants
--;
2602 struct draw_gs_llvm_variant_key
*
2603 draw_gs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2606 struct draw_gs_llvm_variant_key
*key
;
2607 struct draw_sampler_static_state
*draw_sampler
;
2608 struct draw_image_static_state
*draw_image
;
2610 key
= (struct draw_gs_llvm_variant_key
*)store
;
2612 memset(key
, 0, offsetof(struct draw_gs_llvm_variant_key
, samplers
[0]));
2614 key
->num_outputs
= draw_total_gs_outputs(llvm
->draw
);
2616 /* All variants of this shader will have the same value for
2617 * nr_samplers. Not yet trying to compact away holes in the
2620 key
->nr_samplers
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2621 if (llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2622 key
->nr_sampler_views
=
2623 llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2626 key
->nr_sampler_views
= key
->nr_samplers
;
2629 key
->nr_images
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2631 draw_sampler
= key
->samplers
;
2633 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2635 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2636 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2637 llvm
->draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]);
2639 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2640 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2641 llvm
->draw
->sampler_views
[PIPE_SHADER_GEOMETRY
][i
]);
2644 draw_image
= draw_gs_llvm_variant_key_images(key
);
2645 memset(draw_image
, 0,
2646 key
->nr_images
* sizeof *draw_image
);
2647 for (i
= 0; i
< key
->nr_images
; i
++) {
2648 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
2649 llvm
->draw
->images
[PIPE_SHADER_GEOMETRY
][i
]);
2655 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key
*key
)
2658 struct draw_sampler_static_state
*sampler
= key
->samplers
;
2659 struct draw_image_static_state
*image
= draw_gs_llvm_variant_key_images(key
);
2660 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2661 debug_printf("sampler[%i].src_format = %s\n", i
,
2662 util_format_name(sampler
[i
].texture_state
.format
));
2665 for (i
= 0 ; i
< key
->nr_images
; i
++)
2666 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));