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_coro.h"
41 #include "gallivm/lp_bld_swizzle.h"
42 #include "gallivm/lp_bld_struct.h"
43 #include "gallivm/lp_bld_type.h"
44 #include "gallivm/lp_bld_flow.h"
45 #include "gallivm/lp_bld_debug.h"
46 #include "gallivm/lp_bld_tgsi.h"
47 #include "gallivm/lp_bld_nir.h"
48 #include "gallivm/lp_bld_printf.h"
49 #include "gallivm/lp_bld_intr.h"
50 #include "gallivm/lp_bld_init.h"
51 #include "gallivm/lp_bld_type.h"
52 #include "gallivm/lp_bld_pack.h"
53 #include "gallivm/lp_bld_format.h"
54 #include "gallivm/lp_bld_misc.h"
55 #include "tgsi/tgsi_exec.h"
56 #include "tgsi/tgsi_dump.h"
58 #include "util/u_math.h"
59 #include "util/u_pointer.h"
60 #include "util/u_string.h"
61 #include "util/simple_list.h"
62 #include "nir_serialize.h"
63 #include "util/mesa-sha1.h"
68 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*var
);
71 struct draw_gs_llvm_iface
{
72 struct lp_build_gs_iface base
;
74 struct draw_gs_llvm_variant
*variant
;
78 static inline const struct draw_gs_llvm_iface
*
79 draw_gs_llvm_iface(const struct lp_build_gs_iface
*iface
)
81 return (const struct draw_gs_llvm_iface
*)iface
;
84 struct draw_tcs_llvm_iface
{
85 struct lp_build_tcs_iface base
;
87 struct draw_tcs_llvm_variant
*variant
;
92 static inline const struct draw_tcs_llvm_iface
*
93 draw_tcs_llvm_iface(const struct lp_build_tcs_iface
*iface
)
95 return (const struct draw_tcs_llvm_iface
*)iface
;
98 struct draw_tes_llvm_iface
{
99 struct lp_build_tes_iface base
;
101 struct draw_tes_llvm_variant
*variant
;
105 static inline const struct draw_tes_llvm_iface
*
106 draw_tes_llvm_iface(const struct lp_build_tes_iface
*iface
)
108 return (const struct draw_tes_llvm_iface
*)iface
;
112 * Create LLVM type for draw_vertex_buffer.
115 create_jit_dvbuffer_type(struct gallivm_state
*gallivm
,
116 const char *struct_name
)
118 LLVMTargetDataRef target
= gallivm
->target
;
119 LLVMTypeRef dvbuffer_type
;
120 LLVMTypeRef elem_types
[DRAW_JIT_DVBUFFER_NUM_FIELDS
];
121 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
123 elem_types
[DRAW_JIT_DVBUFFER_MAP
] =
124 LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
125 elem_types
[DRAW_JIT_DVBUFFER_SIZE
] = int32_type
;
127 dvbuffer_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
128 ARRAY_SIZE(elem_types
), 0);
130 (void) target
; /* silence unused var warning for non-debug build */
131 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, map
,
132 target
, dvbuffer_type
,
133 DRAW_JIT_DVBUFFER_MAP
);
134 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, size
,
135 target
, dvbuffer_type
,
136 DRAW_JIT_DVBUFFER_SIZE
);
138 return dvbuffer_type
;
142 * Create LLVM type for struct draw_jit_texture
145 create_jit_texture_type(struct gallivm_state
*gallivm
, const char *struct_name
)
147 LLVMTargetDataRef target
= gallivm
->target
;
148 LLVMTypeRef texture_type
;
149 LLVMTypeRef elem_types
[DRAW_JIT_TEXTURE_NUM_FIELDS
];
150 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
152 elem_types
[DRAW_JIT_TEXTURE_WIDTH
] =
153 elem_types
[DRAW_JIT_TEXTURE_HEIGHT
] =
154 elem_types
[DRAW_JIT_TEXTURE_DEPTH
] =
155 elem_types
[DRAW_JIT_TEXTURE_NUM_SAMPLES
] =
156 elem_types
[DRAW_JIT_TEXTURE_SAMPLE_STRIDE
] =
157 elem_types
[DRAW_JIT_TEXTURE_FIRST_LEVEL
] =
158 elem_types
[DRAW_JIT_TEXTURE_LAST_LEVEL
] = int32_type
;
159 elem_types
[DRAW_JIT_TEXTURE_BASE
] =
160 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
161 elem_types
[DRAW_JIT_TEXTURE_ROW_STRIDE
] =
162 elem_types
[DRAW_JIT_TEXTURE_IMG_STRIDE
] =
163 elem_types
[DRAW_JIT_TEXTURE_MIP_OFFSETS
] =
164 LLVMArrayType(int32_type
, PIPE_MAX_TEXTURE_LEVELS
);
166 texture_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
167 ARRAY_SIZE(elem_types
), 0);
169 (void) target
; /* silence unused var warning for non-debug build */
170 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, width
,
171 target
, texture_type
,
172 DRAW_JIT_TEXTURE_WIDTH
);
173 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, height
,
174 target
, texture_type
,
175 DRAW_JIT_TEXTURE_HEIGHT
);
176 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, depth
,
177 target
, texture_type
,
178 DRAW_JIT_TEXTURE_DEPTH
);
179 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, base
,
180 target
, texture_type
,
181 DRAW_JIT_TEXTURE_BASE
);
182 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, row_stride
,
183 target
, texture_type
,
184 DRAW_JIT_TEXTURE_ROW_STRIDE
);
185 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, img_stride
,
186 target
, texture_type
,
187 DRAW_JIT_TEXTURE_IMG_STRIDE
);
188 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, first_level
,
189 target
, texture_type
,
190 DRAW_JIT_TEXTURE_FIRST_LEVEL
);
191 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, last_level
,
192 target
, texture_type
,
193 DRAW_JIT_TEXTURE_LAST_LEVEL
);
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, mip_offsets
,
195 target
, texture_type
,
196 DRAW_JIT_TEXTURE_MIP_OFFSETS
);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, num_samples
,
198 target
, texture_type
,
199 DRAW_JIT_TEXTURE_NUM_SAMPLES
);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, sample_stride
,
201 target
, texture_type
,
202 DRAW_JIT_TEXTURE_SAMPLE_STRIDE
);
204 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture
, target
, texture_type
);
211 * Create LLVM type for struct draw_jit_sampler
214 create_jit_sampler_type(struct gallivm_state
*gallivm
, const char *struct_name
)
216 LLVMTargetDataRef target
= gallivm
->target
;
217 LLVMTypeRef sampler_type
;
218 LLVMTypeRef elem_types
[DRAW_JIT_SAMPLER_NUM_FIELDS
];
220 elem_types
[DRAW_JIT_SAMPLER_MIN_LOD
] =
221 elem_types
[DRAW_JIT_SAMPLER_MAX_LOD
] =
222 elem_types
[DRAW_JIT_SAMPLER_LOD_BIAS
] = LLVMFloatTypeInContext(gallivm
->context
);
223 elem_types
[DRAW_JIT_SAMPLER_BORDER_COLOR
] =
224 LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
226 sampler_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
227 ARRAY_SIZE(elem_types
), 0);
229 (void) target
; /* silence unused var warning for non-debug build */
230 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, min_lod
,
231 target
, sampler_type
,
232 DRAW_JIT_SAMPLER_MIN_LOD
);
233 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, max_lod
,
234 target
, sampler_type
,
235 DRAW_JIT_SAMPLER_MAX_LOD
);
236 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, lod_bias
,
237 target
, sampler_type
,
238 DRAW_JIT_SAMPLER_LOD_BIAS
);
239 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, border_color
,
240 target
, sampler_type
,
241 DRAW_JIT_SAMPLER_BORDER_COLOR
);
243 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler
, target
, sampler_type
);
249 * Create LLVM type for struct draw_jit_texture
252 create_jit_image_type(struct gallivm_state
*gallivm
, const char *struct_name
)
254 LLVMTargetDataRef target
= gallivm
->target
;
255 LLVMTypeRef image_type
;
256 LLVMTypeRef elem_types
[DRAW_JIT_IMAGE_NUM_FIELDS
];
257 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
259 elem_types
[DRAW_JIT_IMAGE_WIDTH
] =
260 elem_types
[DRAW_JIT_IMAGE_HEIGHT
] =
261 elem_types
[DRAW_JIT_IMAGE_DEPTH
] =
262 elem_types
[DRAW_JIT_IMAGE_ROW_STRIDE
] =
263 elem_types
[DRAW_JIT_IMAGE_IMG_STRIDE
] =
264 elem_types
[DRAW_JIT_IMAGE_NUM_SAMPLES
] =
265 elem_types
[DRAW_JIT_IMAGE_SAMPLE_STRIDE
] = int32_type
;
266 elem_types
[DRAW_JIT_IMAGE_BASE
] =
267 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
269 image_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
270 ARRAY_SIZE(elem_types
), 0);
272 (void) target
; /* silence unused var warning for non-debug build */
273 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, width
,
275 DRAW_JIT_IMAGE_WIDTH
);
276 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, height
,
278 DRAW_JIT_IMAGE_HEIGHT
);
279 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, depth
,
281 DRAW_JIT_IMAGE_DEPTH
);
282 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, base
,
284 DRAW_JIT_IMAGE_BASE
);
285 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, row_stride
,
287 DRAW_JIT_IMAGE_ROW_STRIDE
);
288 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, img_stride
,
290 DRAW_JIT_IMAGE_IMG_STRIDE
);
291 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, num_samples
,
293 DRAW_JIT_IMAGE_NUM_SAMPLES
);
294 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, sample_stride
,
296 DRAW_JIT_IMAGE_SAMPLE_STRIDE
);
298 LP_CHECK_STRUCT_SIZE(struct draw_jit_image
, target
, image_type
);
304 * Create LLVM type for struct draw_jit_context
307 create_jit_context_type(struct gallivm_state
*gallivm
,
308 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
309 LLVMTypeRef image_type
,
310 const char *struct_name
)
312 LLVMTargetDataRef target
= gallivm
->target
;
313 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
314 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
315 LLVMTypeRef elem_types
[DRAW_JIT_CTX_NUM_FIELDS
];
316 LLVMTypeRef context_type
;
318 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* vs_constants */
319 LP_MAX_TGSI_CONST_BUFFERS
);
320 elem_types
[1] = LLVMArrayType(int_type
, /* num_vs_constants */
321 LP_MAX_TGSI_CONST_BUFFERS
);
322 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
323 DRAW_TOTAL_CLIP_PLANES
), 0);
324 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
325 elem_types
[4] = LLVMArrayType(texture_type
,
326 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
327 elem_types
[5] = LLVMArrayType(sampler_type
,
328 PIPE_MAX_SAMPLERS
); /* samplers */
329 elem_types
[6] = LLVMArrayType(image_type
,
330 PIPE_MAX_SHADER_IMAGES
); /* images */
331 elem_types
[7] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* vs_ssbo */
332 LP_MAX_TGSI_SHADER_BUFFERS
);
333 elem_types
[8] = LLVMArrayType(int_type
, /* num_vs_ssbos */
334 LP_MAX_TGSI_SHADER_BUFFERS
);
335 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
336 ARRAY_SIZE(elem_types
), 0);
338 (void) target
; /* silence unused var warning for non-debug build */
339 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_constants
,
340 target
, context_type
, DRAW_JIT_CTX_CONSTANTS
);
341 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, num_vs_constants
,
342 target
, context_type
, DRAW_JIT_CTX_NUM_CONSTANTS
);
343 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, planes
,
344 target
, context_type
, DRAW_JIT_CTX_PLANES
);
345 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, viewports
,
346 target
, context_type
, DRAW_JIT_CTX_VIEWPORT
);
347 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, textures
,
348 target
, context_type
,
349 DRAW_JIT_CTX_TEXTURES
);
350 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, samplers
,
351 target
, context_type
,
352 DRAW_JIT_CTX_SAMPLERS
);
353 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, images
,
354 target
, context_type
, DRAW_JIT_CTX_IMAGES
);
355 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_ssbos
,
356 target
, context_type
, DRAW_JIT_CTX_SSBOS
);
357 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, num_vs_ssbos
,
358 target
, context_type
, DRAW_JIT_CTX_NUM_SSBOS
);
359 LP_CHECK_STRUCT_SIZE(struct draw_jit_context
,
360 target
, context_type
);
367 * Create LLVM type for struct draw_gs_jit_context
370 create_gs_jit_context_type(struct gallivm_state
*gallivm
,
371 unsigned vector_length
,
372 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
373 LLVMTypeRef image_type
,
374 const char *struct_name
)
376 LLVMTargetDataRef target
= gallivm
->target
;
377 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
378 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
379 LLVMTypeRef elem_types
[DRAW_GS_JIT_CTX_NUM_FIELDS
];
380 LLVMTypeRef context_type
;
382 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* constants */
383 LP_MAX_TGSI_CONST_BUFFERS
);
384 elem_types
[1] = LLVMArrayType(int_type
, /* num_constants */
385 LP_MAX_TGSI_CONST_BUFFERS
);
386 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
387 DRAW_TOTAL_CLIP_PLANES
), 0);
388 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
390 elem_types
[4] = LLVMArrayType(texture_type
,
391 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
392 elem_types
[5] = LLVMArrayType(sampler_type
,
393 PIPE_MAX_SAMPLERS
); /* samplers */
394 elem_types
[6] = LLVMArrayType(image_type
,
395 PIPE_MAX_SHADER_IMAGES
); /* images */
396 elem_types
[7] = LLVMPointerType(LLVMPointerType(int_type
, 0), 0);
397 elem_types
[8] = LLVMPointerType(LLVMVectorType(int_type
,
399 elem_types
[9] = LLVMPointerType(LLVMVectorType(int_type
,
402 elem_types
[10] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* ssbos */
403 LP_MAX_TGSI_SHADER_BUFFERS
);
404 elem_types
[11] = LLVMArrayType(int_type
, /* num_ssbos */
405 LP_MAX_TGSI_SHADER_BUFFERS
);
407 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
408 ARRAY_SIZE(elem_types
), 0);
410 (void) target
; /* silence unused var warning for non-debug build */
411 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, constants
,
412 target
, context_type
, DRAW_GS_JIT_CTX_CONSTANTS
);
413 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, num_constants
,
414 target
, context_type
, DRAW_GS_JIT_CTX_NUM_CONSTANTS
);
415 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, planes
,
416 target
, context_type
, DRAW_GS_JIT_CTX_PLANES
);
417 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, viewports
,
418 target
, context_type
, DRAW_GS_JIT_CTX_VIEWPORT
);
419 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, textures
,
420 target
, context_type
,
421 DRAW_GS_JIT_CTX_TEXTURES
);
422 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, samplers
,
423 target
, context_type
,
424 DRAW_GS_JIT_CTX_SAMPLERS
);
425 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, prim_lengths
,
426 target
, context_type
,
427 DRAW_GS_JIT_CTX_PRIM_LENGTHS
);
428 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_vertices
,
429 target
, context_type
,
430 DRAW_GS_JIT_CTX_EMITTED_VERTICES
);
431 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_prims
,
432 target
, context_type
,
433 DRAW_GS_JIT_CTX_EMITTED_PRIMS
);
434 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, ssbos
,
435 target
, context_type
, DRAW_GS_JIT_CTX_SSBOS
);
436 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, num_ssbos
,
437 target
, context_type
, DRAW_GS_JIT_CTX_NUM_SSBOS
);
438 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, images
,
439 target
, context_type
, DRAW_GS_JIT_CTX_IMAGES
);
440 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context
,
441 target
, context_type
);
448 create_gs_jit_input_type(struct gallivm_state
*gallivm
)
450 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
451 LLVMTypeRef input_array
;
453 input_array
= LLVMVectorType(float_type
, TGSI_NUM_CHANNELS
); /* num primitives */
454 input_array
= LLVMArrayType(input_array
, TGSI_NUM_CHANNELS
); /* num channels */
455 input_array
= LLVMArrayType(input_array
, PIPE_MAX_SHADER_INPUTS
); /* num attrs per vertex */
456 input_array
= LLVMPointerType(input_array
, 0); /* num vertices per prim */
462 * Create LLVM type for struct pipe_vertex_buffer
465 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
,
466 const char *struct_name
)
468 LLVMTargetDataRef target
= gallivm
->target
;
469 LLVMTypeRef elem_types
[4];
472 elem_types
[0] = LLVMInt16TypeInContext(gallivm
->context
);
473 elem_types
[1] = LLVMInt8TypeInContext(gallivm
->context
);
474 elem_types
[2] = LLVMInt32TypeInContext(gallivm
->context
);
475 elem_types
[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
477 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
478 ARRAY_SIZE(elem_types
), 0);
480 (void) target
; /* silence unused var warning for non-debug build */
481 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
483 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, is_user_buffer
,
485 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
487 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer
.resource
,
490 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
497 * Create LLVM type for struct vertex_header;
500 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
502 LLVMTargetDataRef target
= gallivm
->target
;
503 LLVMTypeRef elem_types
[3];
504 LLVMTypeRef vertex_header
;
505 char struct_name
[24];
507 snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
509 elem_types
[DRAW_JIT_VERTEX_VERTEX_ID
] = LLVMIntTypeInContext(gallivm
->context
, 32);
510 elem_types
[DRAW_JIT_VERTEX_CLIP_POS
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
511 elem_types
[DRAW_JIT_VERTEX_DATA
] = LLVMArrayType(elem_types
[1], data_elems
);
513 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
514 ARRAY_SIZE(elem_types
), 0);
516 /* these are bit-fields and we can't take address of them
517 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
518 target, vertex_header,
519 DRAW_JIT_VERTEX_CLIPMASK);
520 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
521 target, vertex_header,
522 DRAW_JIT_VERTEX_EDGEFLAG);
523 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
524 target, vertex_header,
525 DRAW_JIT_VERTEX_PAD);
526 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
527 target, vertex_header,
528 DRAW_JIT_VERTEX_VERTEX_ID);
530 (void) target
; /* silence unused var warning for non-debug build */
531 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, clip_pos
,
532 target
, vertex_header
,
533 DRAW_JIT_VERTEX_CLIP_POS
);
534 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, data
,
535 target
, vertex_header
,
536 DRAW_JIT_VERTEX_DATA
);
538 assert(LLVMABISizeOfType(target
, vertex_header
) ==
539 offsetof(struct vertex_header
, data
[data_elems
]));
541 return vertex_header
;
545 * Create LLVM type for struct draw_tcs_jit_context
548 create_tcs_jit_context_type(struct gallivm_state
*gallivm
,
549 unsigned vector_length
,
550 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
551 LLVMTypeRef image_type
,
552 const char *struct_name
)
554 LLVMTargetDataRef target
= gallivm
->target
;
555 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
556 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
557 LLVMTypeRef elem_types
[DRAW_TCS_JIT_CTX_NUM_FIELDS
];
558 LLVMTypeRef context_type
;
560 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* constants */
561 LP_MAX_TGSI_CONST_BUFFERS
);
562 elem_types
[1] = LLVMArrayType(int_type
, /* num_constants */
563 LP_MAX_TGSI_CONST_BUFFERS
);
564 elem_types
[2] = LLVMInt32TypeInContext(gallivm
->context
);
565 elem_types
[3] = LLVMInt32TypeInContext(gallivm
->context
);
567 elem_types
[4] = LLVMArrayType(texture_type
,
568 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
569 elem_types
[5] = LLVMArrayType(sampler_type
,
570 PIPE_MAX_SAMPLERS
); /* samplers */
571 elem_types
[6] = LLVMArrayType(image_type
,
572 PIPE_MAX_SHADER_IMAGES
); /* images */
574 elem_types
[7] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* ssbos */
575 LP_MAX_TGSI_SHADER_BUFFERS
);
576 elem_types
[8] = LLVMArrayType(int_type
, /* num_ssbos */
577 LP_MAX_TGSI_SHADER_BUFFERS
);
579 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
580 ARRAY_SIZE(elem_types
), 0);
582 (void) target
; /* silence unused var warning for non-debug build */
583 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, constants
,
584 target
, context_type
, DRAW_TCS_JIT_CTX_CONSTANTS
);
585 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, num_constants
,
586 target
, context_type
, DRAW_TCS_JIT_CTX_NUM_CONSTANTS
);
587 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, textures
,
588 target
, context_type
,
589 DRAW_TCS_JIT_CTX_TEXTURES
);
590 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, samplers
,
591 target
, context_type
,
592 DRAW_TCS_JIT_CTX_SAMPLERS
);
593 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, ssbos
,
594 target
, context_type
, DRAW_TCS_JIT_CTX_SSBOS
);
595 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, num_ssbos
,
596 target
, context_type
, DRAW_TCS_JIT_CTX_NUM_SSBOS
);
597 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context
, images
,
598 target
, context_type
, DRAW_TCS_JIT_CTX_IMAGES
);
599 LP_CHECK_STRUCT_SIZE(struct draw_tcs_jit_context
,
600 target
, context_type
);
606 create_tcs_jit_input_type(struct gallivm_state
*gallivm
)
608 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
609 LLVMTypeRef input_array
;
611 input_array
= LLVMArrayType(float_type
, TGSI_NUM_CHANNELS
); /* num channels */
612 input_array
= LLVMArrayType(input_array
, NUM_TCS_INPUTS
); /* num attrs per vertex */
613 input_array
= LLVMPointerType(input_array
, 0); /* num vertices per prim */
619 create_tcs_jit_output_type(struct gallivm_state
*gallivm
)
621 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
622 LLVMTypeRef output_array
;
624 output_array
= LLVMArrayType(float_type
, TGSI_NUM_CHANNELS
); /* num channels */
625 output_array
= LLVMArrayType(output_array
, PIPE_MAX_SHADER_INPUTS
); /* num attrs per vertex */
626 output_array
= LLVMPointerType(output_array
, 0); /* num vertices per prim */
632 create_tes_jit_input_type(struct gallivm_state
*gallivm
)
634 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
635 LLVMTypeRef input_array
;
637 input_array
= LLVMArrayType(float_type
, TGSI_NUM_CHANNELS
); /* num channels */
638 input_array
= LLVMArrayType(input_array
, PIPE_MAX_SHADER_INPUTS
); /* num attrs per vertex */
639 input_array
= LLVMPointerType(input_array
, 0); /* num vertices per prim */
645 * Create LLVM type for struct draw_tes_jit_context
648 create_tes_jit_context_type(struct gallivm_state
*gallivm
,
649 unsigned vector_length
,
650 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
651 LLVMTypeRef image_type
,
652 const char *struct_name
)
654 LLVMTargetDataRef target
= gallivm
->target
;
655 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
656 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
657 LLVMTypeRef elem_types
[DRAW_TCS_JIT_CTX_NUM_FIELDS
];
658 LLVMTypeRef context_type
;
660 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* constants */
661 LP_MAX_TGSI_CONST_BUFFERS
);
662 elem_types
[1] = LLVMArrayType(int_type
, /* num_constants */
663 LP_MAX_TGSI_CONST_BUFFERS
);
664 elem_types
[2] = LLVMInt32TypeInContext(gallivm
->context
);
665 elem_types
[3] = LLVMInt32TypeInContext(gallivm
->context
);
667 elem_types
[4] = LLVMArrayType(texture_type
,
668 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
669 elem_types
[5] = LLVMArrayType(sampler_type
,
670 PIPE_MAX_SAMPLERS
); /* samplers */
671 elem_types
[6] = LLVMArrayType(image_type
,
672 PIPE_MAX_SHADER_IMAGES
); /* images */
674 elem_types
[7] = LLVMArrayType(LLVMPointerType(int_type
, 0), /* ssbos */
675 LP_MAX_TGSI_SHADER_BUFFERS
);
676 elem_types
[8] = LLVMArrayType(int_type
, /* num_ssbos */
677 LP_MAX_TGSI_SHADER_BUFFERS
);
679 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
680 ARRAY_SIZE(elem_types
), 0);
682 (void) target
; /* silence unused var warning for non-debug build */
683 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, constants
,
684 target
, context_type
, DRAW_TCS_JIT_CTX_CONSTANTS
);
685 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, num_constants
,
686 target
, context_type
, DRAW_TCS_JIT_CTX_NUM_CONSTANTS
);
687 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, textures
,
688 target
, context_type
,
689 DRAW_TCS_JIT_CTX_TEXTURES
);
690 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, samplers
,
691 target
, context_type
,
692 DRAW_TCS_JIT_CTX_SAMPLERS
);
693 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, ssbos
,
694 target
, context_type
, DRAW_TCS_JIT_CTX_SSBOS
);
695 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, num_ssbos
,
696 target
, context_type
, DRAW_TCS_JIT_CTX_NUM_SSBOS
);
697 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context
, images
,
698 target
, context_type
, DRAW_TCS_JIT_CTX_IMAGES
);
699 LP_CHECK_STRUCT_SIZE(struct draw_tes_jit_context
,
700 target
, context_type
);
706 * Create LLVM types for various structures.
709 create_jit_types(struct draw_llvm_variant
*variant
)
711 struct gallivm_state
*gallivm
= variant
->gallivm
;
712 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
,
715 texture_type
= create_jit_texture_type(gallivm
, "texture");
716 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
717 image_type
= create_jit_image_type(gallivm
, "image");
719 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
722 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
724 buffer_type
= create_jit_dvbuffer_type(gallivm
, "draw_vertex_buffer");
725 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
727 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
728 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
733 get_context_ptr_type(struct draw_llvm_variant
*variant
)
735 if (!variant
->context_ptr_type
)
736 create_jit_types(variant
);
737 return variant
->context_ptr_type
;
742 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
744 if (!variant
->buffer_ptr_type
)
745 create_jit_types(variant
);
746 return variant
->buffer_ptr_type
;
751 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
753 if (!variant
->vb_ptr_type
)
754 create_jit_types(variant
);
755 return variant
->vb_ptr_type
;
759 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
761 if (!variant
->vertex_header_ptr_type
)
762 create_jit_types(variant
);
763 return variant
->vertex_header_ptr_type
;
768 * Create per-context LLVM info.
771 draw_llvm_create(struct draw_context
*draw
, LLVMContextRef context
)
773 struct draw_llvm
*llvm
;
775 if (!lp_build_init())
778 llvm
= CALLOC_STRUCT( draw_llvm
);
784 llvm
->context
= context
;
785 if (!llvm
->context
) {
786 llvm
->context
= LLVMContextCreate();
787 llvm
->context_owned
= true;
792 llvm
->nr_variants
= 0;
793 make_empty_list(&llvm
->vs_variants_list
);
795 llvm
->nr_gs_variants
= 0;
796 make_empty_list(&llvm
->gs_variants_list
);
798 llvm
->nr_tcs_variants
= 0;
799 make_empty_list(&llvm
->tcs_variants_list
);
801 llvm
->nr_tes_variants
= 0;
802 make_empty_list(&llvm
->tes_variants_list
);
807 draw_llvm_destroy(llvm
);
813 * Free per-context LLVM info.
816 draw_llvm_destroy(struct draw_llvm
*llvm
)
818 if (llvm
->context_owned
)
819 LLVMContextDispose(llvm
->context
);
820 llvm
->context
= NULL
;
822 /* XXX free other draw_llvm data? */
827 draw_get_ir_cache_key(struct nir_shader
*nir
,
828 const void *key
, size_t key_size
,
830 unsigned char ir_sha1_cache_key
[20])
832 struct blob blob
= { 0 };
837 nir_serialize(&blob
, nir
, true);
838 ir_binary
= blob
.data
;
841 struct mesa_sha1 ctx
;
842 _mesa_sha1_init(&ctx
);
843 _mesa_sha1_update(&ctx
, key
, key_size
);
844 _mesa_sha1_update(&ctx
, ir_binary
, ir_size
);
845 _mesa_sha1_update(&ctx
, &val_32bit
, 4);
846 _mesa_sha1_final(&ctx
, ir_sha1_cache_key
);
852 * Create LLVM-generated code for a vertex shader.
854 struct draw_llvm_variant
*
855 draw_llvm_create_variant(struct draw_llvm
*llvm
,
857 const struct draw_llvm_variant_key
*key
)
859 struct draw_llvm_variant
*variant
;
860 struct llvm_vertex_shader
*shader
=
861 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
862 LLVMTypeRef vertex_header
;
863 char module_name
[64];
864 unsigned char ir_sha1_cache_key
[20];
865 struct lp_cached_code cached
= { 0 };
866 bool needs_caching
= false;
867 variant
= MALLOC(sizeof *variant
+
868 shader
->variant_key_size
-
869 sizeof variant
->key
);
873 variant
->llvm
= llvm
;
874 variant
->shader
= shader
;
875 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
877 snprintf(module_name
, sizeof(module_name
), "draw_llvm_vs_variant%u",
878 variant
->shader
->variants_cached
);
880 if (shader
->base
.state
.ir
.nir
&& llvm
->draw
->disk_cache_cookie
) {
881 draw_get_ir_cache_key(shader
->base
.state
.ir
.nir
,
883 shader
->variant_key_size
,
887 llvm
->draw
->disk_cache_find_shader(llvm
->draw
->disk_cache_cookie
,
890 if (!cached
.data_size
)
891 needs_caching
= true;
893 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, &cached
);
895 create_jit_types(variant
);
897 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
898 if (llvm
->draw
->vs
.vertex_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
899 tgsi_dump(llvm
->draw
->vs
.vertex_shader
->state
.tokens
, 0);
901 nir_print_shader(llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
, stderr
);
902 draw_llvm_dump_variant_key(&variant
->key
);
905 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
907 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
909 draw_llvm_generate(llvm
, variant
);
911 gallivm_compile_module(variant
->gallivm
);
913 variant
->jit_func
= (draw_jit_vert_func
)
914 gallivm_jit_function(variant
->gallivm
, variant
->function
);
917 llvm
->draw
->disk_cache_insert_shader(llvm
->draw
->disk_cache_cookie
,
920 gallivm_free_ir(variant
->gallivm
);
922 variant
->list_item_global
.base
= variant
;
923 variant
->list_item_local
.base
= variant
;
924 /*variant->no = */shader
->variants_created
++;
925 variant
->list_item_global
.base
= variant
;
932 generate_vs(struct draw_llvm_variant
*variant
,
933 LLVMBuilderRef builder
,
934 struct lp_type vs_type
,
935 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
936 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
937 const struct lp_bld_tgsi_system_values
*system_values
,
938 LLVMValueRef context_ptr
,
939 const struct lp_build_sampler_soa
*draw_sampler
,
940 const struct lp_build_image_soa
*draw_image
,
941 boolean clamp_vertex_color
,
942 struct lp_build_mask_context
*bld_mask
)
944 struct draw_llvm
*llvm
= variant
->llvm
;
945 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
946 LLVMValueRef consts_ptr
=
947 draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
948 LLVMValueRef num_consts_ptr
=
949 draw_jit_context_num_vs_constants(variant
->gallivm
, context_ptr
);
950 LLVMValueRef ssbos_ptr
=
951 draw_jit_context_vs_ssbos(variant
->gallivm
, context_ptr
);
952 LLVMValueRef num_ssbos_ptr
=
953 draw_jit_context_num_vs_ssbos(variant
->gallivm
, context_ptr
);
955 struct lp_build_tgsi_params params
;
956 memset(¶ms
, 0, sizeof(params
));
958 params
.type
= vs_type
;
959 params
.mask
= bld_mask
;
960 params
.consts_ptr
= consts_ptr
;
961 params
.const_sizes_ptr
= num_consts_ptr
;
962 params
.system_values
= system_values
;
963 params
.inputs
= inputs
;
964 params
.context_ptr
= context_ptr
;
965 params
.sampler
= draw_sampler
;
966 params
.info
= &llvm
->draw
->vs
.vertex_shader
->info
;
967 params
.ssbo_ptr
= ssbos_ptr
;
968 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
969 params
.image
= draw_image
;
971 if (llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
&&
972 llvm
->draw
->vs
.vertex_shader
->state
.type
== PIPE_SHADER_IR_NIR
)
973 lp_build_nir_soa(variant
->gallivm
,
974 llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
,
978 lp_build_tgsi_soa(variant
->gallivm
,
985 unsigned chan
, attrib
;
986 struct lp_build_context bld
;
987 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
988 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
990 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
991 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
992 if (outputs
[attrib
][chan
]) {
993 switch (info
->output_semantic_name
[attrib
]) {
994 case TGSI_SEMANTIC_COLOR
:
995 case TGSI_SEMANTIC_BCOLOR
:
996 if (clamp_vertex_color
) {
997 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
998 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
999 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
1011 fetch_instanced(struct gallivm_state
*gallivm
,
1012 const struct util_format_description
*format_desc
,
1013 struct lp_type vs_type
,
1014 LLVMValueRef vb_stride
,
1015 LLVMValueRef map_ptr
,
1016 LLVMValueRef buffer_size_adj
,
1017 LLVMValueRef
*inputs
,
1020 LLVMTypeRef i32_t
= LLVMInt32TypeInContext(gallivm
->context
);
1021 LLVMTypeRef aosf_t
, aosi_t
;
1022 LLVMValueRef zero
= LLVMConstNull(i32_t
);
1023 LLVMBuilderRef builder
= gallivm
->builder
;
1024 LLVMValueRef stride
, buffer_overflowed
, aos
, index_valid
;
1027 aosf_t
= lp_build_vec_type(gallivm
, lp_float32_vec4_type());
1028 aosi_t
= lp_build_vec_type(gallivm
, lp_int32_vec4_type());
1030 /* This mul can overflow. Wraparound is ok. */
1031 stride
= LLVMBuildMul(builder
, vb_stride
, index
, "");
1033 buffer_overflowed
= LLVMBuildICmp(builder
, LLVMIntUGE
,
1034 stride
, buffer_size_adj
,
1035 "buffer_overflowed");
1038 lp_build_print_value(gallivm
, " instance index = ", index
);
1039 lp_build_print_value(gallivm
, " buffer overflowed = ", buffer_overflowed
);
1042 index_valid
= LLVMBuildNot(builder
, buffer_overflowed
, "");
1043 index_valid
= LLVMBuildSExt(builder
, index_valid
, i32_t
, "");
1044 stride
= LLVMBuildAnd(builder
, stride
, index_valid
, "");
1046 aos
= lp_build_fetch_rgba_aos(gallivm
,
1048 lp_float32_vec4_type(),
1054 index_valid
= lp_build_broadcast(gallivm
, aosi_t
, index_valid
);
1055 aos
= LLVMBuildBitCast(builder
, aos
, aosi_t
, "");
1056 aos
= LLVMBuildAnd(builder
, aos
, index_valid
, "");
1057 aos
= LLVMBuildBitCast(builder
, aos
, aosf_t
, "");
1059 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1060 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
1061 inputs
[i
] = lp_build_extract_broadcast(gallivm
,
1062 lp_float32_vec4_type(),
1063 vs_type
, aos
, index
);
1069 fetch_vector(struct gallivm_state
*gallivm
,
1070 const struct util_format_description
*format_desc
,
1071 struct lp_type vs_type
,
1072 LLVMValueRef vb_stride
,
1073 LLVMValueRef map_ptr
,
1074 LLVMValueRef buffer_size_adj
,
1075 LLVMValueRef
*inputs
,
1076 LLVMValueRef indices
)
1078 LLVMBuilderRef builder
= gallivm
->builder
;
1079 struct lp_build_context blduivec
;
1080 struct lp_type fetch_type
= vs_type
;
1081 LLVMValueRef offset
, valid_mask
;
1084 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
1086 vb_stride
= lp_build_broadcast_scalar(&blduivec
, vb_stride
);
1087 buffer_size_adj
= lp_build_broadcast_scalar(&blduivec
, buffer_size_adj
);
1089 /* This mul can overflow. Wraparound is ok. */
1090 offset
= lp_build_mul(&blduivec
, vb_stride
, indices
);
1092 valid_mask
= lp_build_compare(gallivm
, blduivec
.type
,
1093 PIPE_FUNC_LESS
, offset
, buffer_size_adj
);
1095 /* not valid elements use offset 0 */
1096 offset
= LLVMBuildAnd(builder
, offset
, valid_mask
, "");
1099 lp_build_print_value(gallivm
, " indices = ", indices
);
1100 lp_build_print_value(gallivm
, " offsets = ", offset
);
1101 lp_build_print_value(gallivm
, " valid_mask = ", valid_mask
);
1105 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
1106 * This should always produce better code.
1109 /* The type handling is annoying here... */
1110 if (format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
&&
1111 format_desc
->channel
[0].pure_integer
) {
1112 if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_SIGNED
) {
1113 fetch_type
= lp_type_int_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
1115 else if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
1116 fetch_type
= lp_type_uint_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
1120 lp_build_fetch_rgba_soa(gallivm
, format_desc
,
1121 fetch_type
, FALSE
, map_ptr
, offset
,
1122 blduivec
.zero
, blduivec
.zero
,
1125 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1126 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
1127 lp_build_vec_type(gallivm
, vs_type
), "");
1130 /* out-of-bound fetches return all zeros */
1131 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1132 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
], blduivec
.vec_type
, "");
1133 inputs
[i
] = LLVMBuildAnd(builder
, inputs
[i
], valid_mask
, "");
1134 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
1135 lp_build_vec_type(gallivm
, vs_type
), "");
1141 store_aos(struct gallivm_state
*gallivm
,
1142 LLVMValueRef io_ptr
,
1146 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
1147 LLVMBuilderRef builder
= gallivm
->builder
;
1148 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
1149 LLVMValueRef indices
[3];
1151 indices
[0] = lp_build_const_int32(gallivm
, 0);
1153 indices
[2] = lp_build_const_int32(gallivm
, 0);
1155 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
1156 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
1159 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
1162 /* Unaligned store due to the vertex header */
1163 LLVMSetAlignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
1167 * Adjust the mask to architecture endianess. The mask will the store in struct:
1169 * struct vertex_header {
1170 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
1171 * unsigned edgeflag:1;
1173 * unsigned vertex_id:16;
1177 * On little-endian machine nothing needs to done, however on bit-endian machine
1178 * the mask's fields need to be adjusted with the algorithm:
1180 * uint32_t reverse (uint32_t x)
1182 * return (x >> 16) | // vertex_id
1183 * ((x & 0x3fff) << 18) | // clipmask
1184 * ((x & 0x4000) << 3) | // edgeflag
1185 * ((x & 0x8000) << 1); // pad
1189 adjust_mask(struct gallivm_state
*gallivm
,
1192 #if UTIL_ARCH_BIG_ENDIAN
1193 LLVMBuilderRef builder
= gallivm
->builder
;
1194 LLVMValueRef vertex_id
;
1195 LLVMValueRef clipmask
;
1197 LLVMValueRef edgeflag
;
1199 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
1200 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
1201 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
1203 pad
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
1204 pad
= LLVMBuildShl(builder
, pad
, lp_build_const_int32(gallivm
, 1), "");
1206 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
1207 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 3), "");
1209 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
1211 mask
= LLVMBuildOr(builder
, mask
, pad
, "");
1213 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
1219 store_aos_array(struct gallivm_state
*gallivm
,
1220 struct lp_type soa_type
,
1221 LLVMValueRef io_ptr
,
1222 LLVMValueRef
*indices
,
1226 LLVMValueRef clipmask
,
1227 boolean need_edgeflag
)
1229 LLVMBuilderRef builder
= gallivm
->builder
;
1230 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
1231 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1232 LLVMValueRef linear_inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1233 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1234 int vector_length
= soa_type
.length
;
1237 debug_assert(TGSI_NUM_CHANNELS
== 4);
1239 for (i
= 0; i
< vector_length
; i
++) {
1240 linear_inds
[i
] = lp_build_const_int32(gallivm
, i
);
1242 inds
[i
] = indices
[i
];
1244 inds
[i
] = linear_inds
[i
];
1246 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1250 /* store vertex header for each of the n vertices */
1251 LLVMValueRef val
, cliptmp
;
1252 int vertex_id_pad_edgeflag
;
1254 /* If this assertion fails, it means we need to update the bit twidding
1255 * code here. See struct vertex_header in draw_private.h.
1257 assert(DRAW_TOTAL_CLIP_PLANES
==14);
1258 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1259 if (!need_edgeflag
) {
1260 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
1263 vertex_id_pad_edgeflag
= (0xffff << 16);
1265 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
),
1266 vertex_id_pad_edgeflag
);
1267 /* OR with the clipmask */
1268 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
1269 for (i
= 0; i
< vector_length
; i
++) {
1270 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
1271 val
= LLVMBuildExtractElement(builder
, cliptmp
, linear_inds
[i
], "");
1272 val
= adjust_mask(gallivm
, val
);
1274 lp_build_printf(gallivm
, "io = %p, index %d, clipmask = %x\n",
1275 io_ptrs
[i
], inds
[i
], val
);
1277 LLVMBuildStore(builder
, val
, id_ptr
);
1281 /* store for each of the n vertices */
1282 for (i
= 0; i
< vector_length
; i
++) {
1283 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
1289 convert_to_aos(struct gallivm_state
*gallivm
,
1291 LLVMValueRef
*indices
,
1292 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1293 LLVMValueRef clipmask
,
1295 struct lp_type soa_type
,
1296 boolean need_edgeflag
)
1298 LLVMBuilderRef builder
= gallivm
->builder
;
1299 unsigned chan
, attrib
, i
;
1302 lp_build_printf(gallivm
, " # storing begin\n");
1304 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
1305 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
1306 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
1307 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1308 if (outputs
[attrib
][chan
]) {
1309 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
1310 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
1312 lp_build_printf(gallivm
, "output %d : %d ",
1313 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1315 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1317 lp_build_print_value(gallivm
, "val = ", out
);
1320 LLVMBuildBitCast(builder
, out
, lp_build_int_vec_type(gallivm
, soa_type
), "");
1322 lp_build_print_value(gallivm
, " ival = ", iv
);
1333 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
1334 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
1336 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
1338 for (i
= 0; i
< soa_type
.length
; ++i
) {
1339 aos
[i
] = lp_build_extract_range(gallivm
,
1340 soa
[i
% TGSI_NUM_CHANNELS
],
1341 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1346 store_aos_array(gallivm
,
1356 lp_build_printf(gallivm
, " # storing end\n");
1362 * Stores original vertex positions in clip coordinates
1365 store_clip(struct gallivm_state
*gallivm
,
1366 const struct lp_type vs_type
,
1367 LLVMValueRef io_ptr
,
1368 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1371 LLVMBuilderRef builder
= gallivm
->builder
;
1372 LLVMValueRef soa
[4];
1373 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
1374 LLVMValueRef indices
[2];
1375 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1376 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1377 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1378 LLVMTypeRef clip_ptr_type
=
1379 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
),
1384 indices
[1] = lp_build_const_int32(gallivm
, 0);
1386 for (i
= 0; i
< vs_type
.length
; i
++) {
1387 inds
[i
] = lp_build_const_int32(gallivm
, i
);
1388 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1391 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
1392 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
1393 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
1394 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
1396 for (i
= 0; i
< vs_type
.length
; i
++) {
1397 clip_ptrs
[i
] = draw_jit_header_clip_pos(gallivm
, io_ptrs
[i
]);
1400 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
1401 for (i
= 0; i
< vs_type
.length
; ++i
) {
1402 aos
[i
] = lp_build_extract_range(gallivm
,
1403 soa
[i
% TGSI_NUM_CHANNELS
],
1404 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1408 for (j
= 0; j
< vs_type
.length
; j
++) {
1409 LLVMValueRef clip_ptr
;
1411 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
1412 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
1414 /* Unaligned store */
1415 LLVMSetAlignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
1421 * Transforms the outputs for viewport mapping
1424 generate_viewport(struct draw_llvm_variant
*variant
,
1425 LLVMBuilderRef builder
,
1426 struct lp_type vs_type
,
1427 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1428 LLVMValueRef context_ptr
)
1431 struct gallivm_state
*gallivm
= variant
->gallivm
;
1432 struct lp_type f32_type
= vs_type
;
1433 const unsigned pos
= variant
->llvm
->draw
->vs
.position_output
;
1434 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1435 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn*/
1436 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
1437 LLVMValueRef vp_ptr
= draw_jit_context_viewports(gallivm
, context_ptr
);
1439 /* We treat pipe_viewport_state as a float array */
1440 const int scale_index_offset
= offsetof(struct pipe_viewport_state
, scale
) / sizeof(float);
1441 const int trans_index_offset
= offsetof(struct pipe_viewport_state
, translate
) / sizeof(float);
1443 /* for 1/w convention*/
1444 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
1445 LLVMBuildStore(builder
, out3
, outputs
[pos
][3]);
1447 /* Viewport Mapping */
1448 for (i
=0; i
<3; i
++) {
1449 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[pos
][i
], ""); /*x0 x1 .. xn*/
1452 LLVMValueRef scale_i
;
1453 LLVMValueRef trans_i
;
1456 index
= lp_build_const_int32(gallivm
, i
+ scale_index_offset
);
1457 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1459 index
= lp_build_const_int32(gallivm
, i
+ trans_index_offset
);
1460 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1462 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1463 LLVMBuildLoad(builder
, scale_i
, "scale"));
1464 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1465 LLVMBuildLoad(builder
, trans_i
, "trans"));
1468 out
= LLVMBuildFMul(builder
, out
, out3
, "");
1469 /* mult by scale, add translation */
1470 out
= lp_build_fmuladd(builder
, out
, scale
, trans
);
1472 /* store transformed outputs */
1473 LLVMBuildStore(builder
, out
, outputs
[pos
][i
]);
1480 * Returns clipmask as nxi32 bitmask for the n vertices
1483 generate_clipmask(struct draw_llvm
*llvm
,
1484 struct gallivm_state
*gallivm
,
1485 struct lp_type vs_type
,
1486 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1487 struct draw_llvm_variant_key
*key
,
1488 LLVMValueRef context_ptr
,
1489 boolean
*have_clipdist
)
1491 LLVMBuilderRef builder
= gallivm
->builder
;
1492 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
1493 LLVMValueRef test
, temp
;
1494 LLVMValueRef zero
, shift
;
1495 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
1496 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
1497 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
1498 struct lp_type f32_type
= vs_type
;
1499 struct lp_type i32_type
= lp_int_type(vs_type
);
1500 const unsigned pos
= llvm
->draw
->vs
.position_output
;
1501 const unsigned cv
= llvm
->draw
->vs
.clipvertex_output
;
1502 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
1503 boolean have_cd
= false;
1504 boolean clip_user
= key
->clip_user
;
1505 unsigned ucp_enable
= key
->ucp_enable
;
1508 cd
[0] = llvm
->draw
->vs
.ccdistance_output
[0];
1509 cd
[1] = llvm
->draw
->vs
.ccdistance_output
[1];
1511 if (cd
[0] != pos
|| cd
[1] != pos
)
1514 if (num_written_clipdistance
&& !clip_user
) {
1516 ucp_enable
= (1 << num_written_clipdistance
) - 1;
1519 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1520 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1521 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
1522 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
1525 * load clipvertex and position from correct locations.
1526 * if they are the same just load them once.
1528 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
1529 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
1530 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
1531 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
1533 if (clip_user
&& cv
!= pos
) {
1534 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
1535 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
1536 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
1537 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1546 * Be careful with the comparisons and NaNs (using llvm's unordered
1547 * comparisons here).
1549 /* Cliptest, for hardwired planes */
1551 * XXX should take guardband into account (currently not in key).
1552 * Otherwise might run the draw pipeline stages for nothing.
1556 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1558 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1562 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1563 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1564 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1565 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1566 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1569 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1570 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1571 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1572 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1575 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1576 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1577 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1578 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1579 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1583 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1584 if (key
->clip_halfz
) {
1586 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1587 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1588 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1592 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1593 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1594 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1595 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1598 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1599 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1600 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1601 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1605 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1606 LLVMValueRef indices
[3];
1607 LLVMValueRef is_nan_or_inf
;
1609 /* userclip planes */
1610 while (ucp_enable
) {
1611 unsigned plane_idx
= ffs(ucp_enable
)-1;
1612 ucp_enable
&= ~(1 << plane_idx
);
1615 if (have_cd
&& num_written_clipdistance
) {
1616 LLVMValueRef clipdist
;
1620 *have_clipdist
= TRUE
;
1622 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1624 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1626 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1627 is_nan_or_inf
= lp_build_is_inf_or_nan(gallivm
, vs_type
, clipdist
);
1628 test
= LLVMBuildOr(builder
, test
, is_nan_or_inf
, "");
1629 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1630 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1631 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1633 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1634 indices
[0] = lp_build_const_int32(gallivm
, 0);
1635 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1637 indices
[2] = lp_build_const_int32(gallivm
, 0);
1638 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1639 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1640 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1641 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1643 indices
[2] = lp_build_const_int32(gallivm
, 1);
1644 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1645 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1646 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1647 sum
= lp_build_fmuladd(builder
, planes
, cv_y
, sum
);
1649 indices
[2] = lp_build_const_int32(gallivm
, 2);
1650 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1651 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1652 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1653 sum
= lp_build_fmuladd(builder
, planes
, cv_z
, sum
);
1655 indices
[2] = lp_build_const_int32(gallivm
, 3);
1656 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1657 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1658 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1659 sum
= lp_build_fmuladd(builder
, planes
, cv_w
, sum
);
1661 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1662 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1663 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1664 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1668 if (key
->need_edgeflags
) {
1670 * This isn't really part of clipmask but stored the same in vertex
1671 * header later, so do it here.
1673 unsigned edge_attr
= llvm
->draw
->vs
.edgeflag_output
;
1674 LLVMValueRef one
= lp_build_const_vec(gallivm
, f32_type
, 1.0);
1675 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, outputs
[edge_attr
][0], "");
1676 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_EQUAL
, one
, edgeflag
);
1677 temp
= lp_build_const_int_vec(gallivm
, i32_type
,
1678 1LL << DRAW_TOTAL_CLIP_PLANES
);
1679 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1680 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1687 * Returns boolean if any clipping has occurred
1688 * Used zero/one i8 value to represent boolean
1691 clipmask_booli8(struct gallivm_state
*gallivm
,
1692 const struct lp_type vs_type
,
1693 LLVMValueRef clipmask_bool_ptr
,
1694 boolean edgeflag_in_clipmask
)
1696 LLVMBuilderRef builder
= gallivm
->builder
;
1697 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1698 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1700 struct lp_build_context bldivec
;
1702 lp_build_context_init(&bldivec
, gallivm
, lp_int_type(vs_type
));
1705 * We need to invert the edgeflag bit from the clipmask here
1706 * (because the result is really if we want to run the pipeline or not
1707 * and we (may) need it if edgeflag was 0).
1709 if (edgeflag_in_clipmask
) {
1710 LLVMValueRef edge
= lp_build_const_int_vec(gallivm
, bldivec
.type
,
1711 1LL << DRAW_TOTAL_CLIP_PLANES
);
1712 clipmask_bool
= LLVMBuildXor(builder
, clipmask_bool
, edge
, "");
1716 * XXX: probably should mask off bits from the mask which come from
1717 * vertices which were beyond the count (i.e. indices_valid for
1718 * linear fetches, for elts ones we don't have the correct mask
1719 * right now). Otherwise might run the pipeline for nothing,
1720 * though everything should still work.
1722 ret
= lp_build_any_true_range(&bldivec
, vs_type
.length
, clipmask_bool
);
1723 ret
= LLVMBuildZExt(builder
, ret
, int8_type
, "");
1728 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface
*gs_iface
,
1729 struct lp_build_context
* bld
,
1730 boolean is_vindex_indirect
,
1731 LLVMValueRef vertex_index
,
1732 boolean is_aindex_indirect
,
1733 LLVMValueRef attrib_index
,
1734 LLVMValueRef swizzle_index
)
1736 const struct draw_gs_llvm_iface
*gs
= draw_gs_llvm_iface(gs_iface
);
1737 struct gallivm_state
*gallivm
= bld
->gallivm
;
1738 LLVMBuilderRef builder
= gallivm
->builder
;
1739 LLVMValueRef indices
[3];
1741 struct lp_type type
= bld
->type
;
1743 if (is_vindex_indirect
|| is_aindex_indirect
) {
1746 for (i
= 0; i
< type
.length
; ++i
) {
1747 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
1748 LLVMValueRef vert_chan_index
= vertex_index
;
1749 LLVMValueRef attr_chan_index
= attrib_index
;
1750 LLVMValueRef channel_vec
, value
;
1752 if (is_vindex_indirect
) {
1753 vert_chan_index
= LLVMBuildExtractElement(builder
,
1754 vertex_index
, idx
, "");
1756 if (is_aindex_indirect
) {
1757 attr_chan_index
= LLVMBuildExtractElement(builder
,
1758 attrib_index
, idx
, "");
1761 indices
[0] = vert_chan_index
;
1762 indices
[1] = attr_chan_index
;
1763 indices
[2] = swizzle_index
;
1765 channel_vec
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1766 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
1767 value
= LLVMBuildExtractElement(builder
, channel_vec
, idx
, "");
1769 res
= LLVMBuildInsertElement(builder
, res
, value
, idx
, "");
1772 indices
[0] = vertex_index
;
1773 indices
[1] = attrib_index
;
1774 indices
[2] = swizzle_index
;
1776 res
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1777 res
= LLVMBuildLoad(builder
, res
, "");
1784 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface
*gs_base
,
1785 struct lp_build_context
* bld
,
1786 LLVMValueRef (*outputs
)[4],
1787 LLVMValueRef emitted_vertices_vec
,
1788 LLVMValueRef stream_id
)
1790 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1791 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1792 struct gallivm_state
*gallivm
= variant
->gallivm
;
1793 LLVMBuilderRef builder
= gallivm
->builder
;
1794 struct lp_type gs_type
= bld
->type
;
1795 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
1796 lp_int_type(gs_type
), 0);
1797 LLVMValueRef indices
[LP_MAX_VECTOR_LENGTH
];
1798 LLVMValueRef next_prim_offset
=
1799 lp_build_const_int32(gallivm
, variant
->shader
->base
.primitive_boundary
);
1800 LLVMValueRef io
= variant
->io_ptr
;
1802 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
1804 for (i
= 0; i
< gs_type
.length
; ++i
) {
1805 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1806 LLVMValueRef currently_emitted
=
1807 LLVMBuildExtractElement(builder
, emitted_vertices_vec
, ind
, "");
1808 indices
[i
] = LLVMBuildMul(builder
, ind
, next_prim_offset
, "");
1809 indices
[i
] = LLVMBuildAdd(builder
, indices
[i
], currently_emitted
, "");
1812 LLVMValueRef stream_idx
= LLVMBuildExtractElement(builder
, stream_id
, lp_build_const_int32(gallivm
, 0), "");
1813 LLVMValueRef cnd
= LLVMBuildICmp(builder
, LLVMIntULT
, stream_idx
, lp_build_const_int32(gallivm
, variant
->shader
->base
.num_vertex_streams
), "");
1814 struct lp_build_if_state if_ctx
;
1815 lp_build_if(&if_ctx
, gallivm
, cnd
);
1816 io
= lp_build_pointer_get(builder
, io
, LLVMBuildExtractElement(builder
, stream_id
, lp_build_const_int32(gallivm
, 0), ""));
1818 convert_to_aos(gallivm
, io
, indices
,
1820 gs_info
->num_outputs
, gs_type
,
1822 lp_build_endif(&if_ctx
);
1826 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface
*gs_base
,
1827 struct lp_build_context
* bld
,
1828 LLVMValueRef total_emitted_vertices_vec_ptr
,
1829 LLVMValueRef verts_per_prim_vec
,
1830 LLVMValueRef emitted_prims_vec
,
1831 LLVMValueRef mask_vec
)
1833 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1834 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1835 struct gallivm_state
*gallivm
= variant
->gallivm
;
1836 LLVMBuilderRef builder
= gallivm
->builder
;
1837 LLVMValueRef prim_lengts_ptr
=
1838 draw_gs_jit_prim_lengths(variant
->gallivm
, variant
->context_ptr
);
1841 LLVMValueRef cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, mask_vec
, lp_build_const_int_vec(gallivm
, bld
->type
, 0), "");
1842 for (i
= 0; i
< bld
->type
.length
; ++i
) {
1843 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1844 LLVMValueRef prims_emitted
=
1845 LLVMBuildExtractElement(builder
, emitted_prims_vec
, ind
, "");
1846 LLVMValueRef store_ptr
;
1847 LLVMValueRef num_vertices
=
1848 LLVMBuildExtractElement(builder
, verts_per_prim_vec
, ind
, "");
1850 LLVMValueRef this_cond
= LLVMBuildExtractElement(gallivm
->builder
, cond
, ind
, "");
1851 struct lp_build_if_state ifthen
;
1852 lp_build_if(&ifthen
, gallivm
, this_cond
);
1853 store_ptr
= LLVMBuildGEP(builder
, prim_lengts_ptr
, &prims_emitted
, 1, "");
1854 store_ptr
= LLVMBuildLoad(builder
, store_ptr
, "");
1855 store_ptr
= LLVMBuildGEP(builder
, store_ptr
, &ind
, 1, "");
1856 LLVMBuildStore(builder
, num_vertices
, store_ptr
);
1857 lp_build_endif(&ifthen
);
1862 draw_gs_llvm_epilogue(const struct lp_build_gs_iface
*gs_base
,
1863 LLVMValueRef total_emitted_vertices_vec
,
1864 LLVMValueRef emitted_prims_vec
, unsigned stream
)
1866 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1867 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1868 struct gallivm_state
*gallivm
= variant
->gallivm
;
1869 LLVMBuilderRef builder
= gallivm
->builder
;
1870 LLVMValueRef emitted_verts_ptr
=
1871 draw_gs_jit_emitted_vertices(gallivm
, variant
->context_ptr
);
1872 LLVMValueRef emitted_prims_ptr
=
1873 draw_gs_jit_emitted_prims(gallivm
, variant
->context_ptr
);
1874 LLVMValueRef stream_val
= lp_build_const_int32(gallivm
, stream
);
1876 emitted_verts_ptr
= LLVMBuildGEP(builder
, emitted_verts_ptr
, &stream_val
, 1, "");
1877 emitted_prims_ptr
= LLVMBuildGEP(builder
, emitted_prims_ptr
, &stream_val
, 1, "");
1879 LLVMBuildStore(builder
, total_emitted_vertices_vec
, emitted_verts_ptr
);
1880 LLVMBuildStore(builder
, emitted_prims_vec
, emitted_prims_ptr
);
1884 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
)
1886 struct gallivm_state
*gallivm
= variant
->gallivm
;
1887 LLVMContextRef context
= gallivm
->context
;
1888 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1889 LLVMTypeRef arg_types
[12];
1890 unsigned num_arg_types
= ARRAY_SIZE(arg_types
);
1891 LLVMTypeRef func_type
;
1892 LLVMValueRef context_ptr
;
1893 LLVMBasicBlockRef block
;
1894 LLVMBuilderRef builder
;
1896 struct lp_type vs_type
;
1897 LLVMValueRef count
, fetch_elts
, start_or_maxelt
;
1898 LLVMValueRef vertex_id_offset
;
1899 LLVMValueRef stride
, step
, io_itr
;
1900 LLVMValueRef ind_vec
, start_vec
, have_elts
, fetch_max
, tmp
;
1901 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1902 LLVMValueRef vb_stride
[PIPE_MAX_ATTRIBS
];
1903 LLVMValueRef map_ptr
[PIPE_MAX_ATTRIBS
];
1904 LLVMValueRef buffer_size_adj
[PIPE_MAX_ATTRIBS
];
1905 LLVMValueRef instance_index
[PIPE_MAX_ATTRIBS
];
1906 LLVMValueRef fake_buf_ptr
, fake_buf
;
1908 struct draw_context
*draw
= llvm
->draw
;
1909 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1911 struct lp_build_context bld
, blduivec
;
1912 struct lp_build_loop_state lp_loop
;
1913 struct lp_build_if_state if_ctx
;
1914 const int vector_length
= lp_native_vector_width
/ 32;
1915 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1916 struct lp_build_sampler_soa
*sampler
= 0;
1917 struct lp_build_image_soa
*image
= NULL
;
1918 LLVMValueRef ret
, clipmask_bool_ptr
;
1919 struct draw_llvm_variant_key
*key
= &variant
->key
;
1920 /* If geometry shader is present we need to skip both the viewport
1921 * transformation and clipping otherwise the inputs to the geometry
1922 * shader will be incorrect.
1923 * The code can't handle vp transform when vs writes vp index neither
1924 * (though this would be fixable here, but couldn't just broadcast
1927 const boolean bypass_viewport
= key
->has_gs_or_tes
|| key
->bypass_viewport
||
1928 vs_info
->writes_viewport_index
;
1929 const boolean enable_cliptest
= !key
->has_gs_or_tes
&& (key
->clip_xy
||
1932 key
->need_edgeflags
);
1933 LLVMValueRef variant_func
;
1934 const unsigned pos
= draw
->vs
.position_output
;
1935 const unsigned cv
= draw
->vs
.clipvertex_output
;
1936 boolean have_clipdist
= FALSE
;
1937 struct lp_bld_tgsi_system_values system_values
;
1939 memset(&system_values
, 0, sizeof(system_values
));
1940 memset(&outputs
, 0, sizeof(outputs
));
1941 snprintf(func_name
, sizeof(func_name
), "draw_llvm_vs_variant");
1944 arg_types
[i
++] = get_context_ptr_type(variant
); /* context */
1945 arg_types
[i
++] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1946 arg_types
[i
++] = get_buffer_ptr_type(variant
); /* vbuffers */
1947 arg_types
[i
++] = int32_type
; /* count */
1948 arg_types
[i
++] = int32_type
; /* start/fetch_elt_max */
1949 arg_types
[i
++] = int32_type
; /* stride */
1950 arg_types
[i
++] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1951 arg_types
[i
++] = int32_type
; /* instance_id */
1952 arg_types
[i
++] = int32_type
; /* vertex_id_offset */
1953 arg_types
[i
++] = int32_type
; /* start_instance */
1954 arg_types
[i
++] = LLVMPointerType(int32_type
, 0); /* fetch_elts */
1955 arg_types
[i
++] = int32_type
; /* draw_id */
1957 func_type
= LLVMFunctionType(LLVMInt8TypeInContext(context
),
1958 arg_types
, num_arg_types
, 0);
1960 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1961 variant
->function
= variant_func
;
1963 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1964 for (i
= 0; i
< num_arg_types
; ++i
)
1965 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1966 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
1968 if (gallivm
->cache
&& gallivm
->cache
->data_size
)
1970 context_ptr
= LLVMGetParam(variant_func
, 0);
1971 io_ptr
= LLVMGetParam(variant_func
, 1);
1972 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1973 count
= LLVMGetParam(variant_func
, 3);
1975 * XXX: the maxelt part is unused. Not really useful, since we cannot
1976 * get index buffer overflows due to vsplit (which provides its own
1977 * elts buffer, with a different size than what's passed in here).
1979 start_or_maxelt
= LLVMGetParam(variant_func
, 4);
1981 * XXX: stride is actually unused. The stride we use is strictly calculated
1982 * from the number of outputs (including the draw_extra outputs).
1983 * Should probably fix some day (we need a new vs just because of extra
1984 * outputs which the generated vs won't touch).
1986 stride
= LLVMGetParam(variant_func
, 5);
1987 vb_ptr
= LLVMGetParam(variant_func
, 6);
1988 system_values
.instance_id
= LLVMGetParam(variant_func
, 7);
1989 vertex_id_offset
= LLVMGetParam(variant_func
, 8);
1990 system_values
.base_instance
= LLVMGetParam(variant_func
, 9);
1991 fetch_elts
= LLVMGetParam(variant_func
, 10);
1992 system_values
.draw_id
= LLVMGetParam(variant_func
, 11);
1994 lp_build_name(context_ptr
, "context");
1995 lp_build_name(io_ptr
, "io");
1996 lp_build_name(vbuffers_ptr
, "vbuffers");
1997 lp_build_name(count
, "count");
1998 lp_build_name(start_or_maxelt
, "start_or_maxelt");
1999 lp_build_name(stride
, "stride");
2000 lp_build_name(vb_ptr
, "vb");
2001 lp_build_name(system_values
.instance_id
, "instance_id");
2002 lp_build_name(vertex_id_offset
, "vertex_id_offset");
2003 lp_build_name(system_values
.base_instance
, "start_instance");
2004 lp_build_name(fetch_elts
, "fetch_elts");
2005 lp_build_name(system_values
.draw_id
, "draw_id");
2011 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
2012 builder
= gallivm
->builder
;
2013 LLVMPositionBuilderAtEnd(builder
, block
);
2015 memset(&vs_type
, 0, sizeof vs_type
);
2016 vs_type
.floating
= TRUE
; /* floating point values */
2017 vs_type
.sign
= TRUE
; /* values are signed */
2018 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2019 vs_type
.width
= 32; /* 32-bit float */
2020 vs_type
.length
= vector_length
;
2022 lp_build_context_init(&bld
, gallivm
, lp_type_uint(32));
2023 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
2025 /* hold temporary "bool" clipmask */
2026 clipmask_bool_ptr
= lp_build_alloca(gallivm
, blduivec
.vec_type
, "");
2028 fake_buf
= lp_build_alloca_undef(gallivm
,
2029 LLVMVectorType(LLVMInt64TypeInContext(context
), 4), "");
2030 fake_buf
= LLVMBuildBitCast(builder
, fake_buf
,
2031 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
2032 fake_buf_ptr
= LLVMBuildGEP(builder
, fake_buf
, &bld
.zero
, 1, "");
2034 /* code generated texture sampling */
2035 sampler
= draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key
), key
->nr_samplers
);
2037 image
= draw_llvm_image_soa_create(draw_llvm_variant_key_images(key
),
2040 step
= lp_build_const_int32(gallivm
, vector_length
);
2042 ind_vec
= blduivec
.undef
;
2043 for (i
= 0; i
< vs_type
.length
; i
++) {
2044 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
2045 ind_vec
= LLVMBuildInsertElement(builder
, ind_vec
, index
, index
, "");
2048 have_elts
= LLVMBuildICmp(builder
, LLVMIntNE
,
2049 LLVMConstPointerNull(arg_types
[10]), fetch_elts
, "");
2051 fetch_max
= LLVMBuildSub(builder
, count
, bld
.one
, "fetch_max");
2052 fetch_max
= lp_build_broadcast_scalar(&blduivec
, fetch_max
);
2054 * Only needed for non-indexed path.
2056 start_vec
= lp_build_broadcast_scalar(&blduivec
, start_or_maxelt
);
2059 * Pre-calculate everything which is constant per shader invocation.
2061 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
2062 LLVMValueRef vb_buffer_offset
, buffer_size
, temp_ptr
;
2063 LLVMValueRef vb_info
, vbuffer_ptr
, buf_offset
, ofbit
;
2064 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
2065 LLVMValueRef vb_index
=
2066 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
2067 LLVMValueRef bsize
= lp_build_const_int32(gallivm
,
2068 util_format_get_blocksize(velem
->src_format
));
2069 LLVMValueRef src_offset
= lp_build_const_int32(gallivm
,
2071 struct lp_build_if_state if_ctx
;
2073 if (velem
->src_format
!= PIPE_FORMAT_NONE
) {
2074 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
, &vb_index
, 1, "");
2075 vb_info
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
2076 vb_stride
[j
] = draw_jit_vbuffer_stride(gallivm
, vb_info
);
2077 vb_stride
[j
] = LLVMBuildZExt(gallivm
->builder
, vb_stride
[j
],
2078 LLVMInt32TypeInContext(context
), "");
2079 vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vb_info
);
2080 map_ptr
[j
] = draw_jit_dvbuffer_map(gallivm
, vbuffer_ptr
);
2081 buffer_size
= draw_jit_dvbuffer_size(gallivm
, vbuffer_ptr
);
2085 * We'll set buffer_size_adj to zero if we have of, so it will
2086 * always overflow later automatically without having to keep ofbit.
2087 * Overflows (with normal wraparound) doing the actual offset
2088 * calculation should be ok, just not for the buffer size calc.
2089 * It would also be possible to detect such overflows and return
2090 * zeros if that happens, but this would be more complex.
2092 buf_offset
= lp_build_add(&bld
, vb_buffer_offset
, src_offset
);
2093 tmp
= lp_build_sub(&bld
, bsize
, bld
.one
);
2094 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size
, tmp
,
2096 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size_adj
[j
],
2097 buf_offset
, &ofbit
);
2100 * We can't easily set fake vertex buffers outside the generated code.
2101 * Hence, set fake vertex buffers here instead basically, so fetch
2102 * code can always fetch using offset 0, eliminating all control flow
2103 * inside the main loop.
2104 * (Alternatively, could have control flow per vector skipping fetch
2105 * if ofbit is true.)
2107 if (velem
->instance_divisor
) {
2109 * Index is equal to the start instance plus the number of current
2110 * instance divided by the divisor. In this case we compute it as:
2111 * index = start_instance + (instance_id / divisor).
2112 * Note we could actually do the fetch here, outside the loop -
2113 * it's all constant, hopefully llvm recognizes this.
2115 LLVMValueRef current_instance
;
2116 current_instance
= LLVMBuildUDiv(builder
, system_values
.instance_id
,
2117 lp_build_const_int32(gallivm
,
2118 velem
->instance_divisor
),
2119 "instance_divisor");
2120 instance_index
[j
] = lp_build_uadd_overflow(gallivm
, system_values
.base_instance
,
2121 current_instance
, &ofbit
);
2124 buffer_size_adj
[j
] = LLVMBuildSelect(builder
, ofbit
, bld
.zero
,
2125 buffer_size_adj
[j
], "");
2127 temp_ptr
= lp_build_alloca_undef(gallivm
,
2128 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
2130 lp_build_if(&if_ctx
, gallivm
, ofbit
);
2132 LLVMBuildStore(builder
, fake_buf_ptr
, temp_ptr
);
2134 lp_build_else(&if_ctx
);
2136 map_ptr
[j
] = LLVMBuildGEP(builder
, map_ptr
[j
], &buf_offset
, 1, "");
2137 LLVMBuildStore(builder
, map_ptr
[j
], temp_ptr
);
2139 lp_build_endif(&if_ctx
);
2140 map_ptr
[j
] = LLVMBuildLoad(builder
, temp_ptr
, "map_ptr");
2143 lp_build_printf(gallivm
, "velem %d, vbuf index = %u, vb_stride = %u\n",
2144 lp_build_const_int32(gallivm
, j
),
2145 vb_index
, vb_stride
[j
]);
2146 lp_build_printf(gallivm
,
2147 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
2148 vb_buffer_offset
, src_offset
, buf_offset
);
2149 lp_build_printf(gallivm
, " buffer size = %u, blocksize = %u\n",
2150 buffer_size
, bsize
);
2151 lp_build_printf(gallivm
, " instance_id = %u\n", system_values
.instance_id
);
2156 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
2158 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
2160 LLVMValueRef clipmask
; /* holds the clipmask value */
2161 LLVMValueRef true_index_array
, index_store
;
2162 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
2164 io_itr
= lp_loop
.counter
;
2166 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
2168 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
2169 io_itr
, io
, lp_loop
.counter
);
2172 true_index_array
= lp_build_broadcast_scalar(&blduivec
, lp_loop
.counter
);
2173 true_index_array
= LLVMBuildAdd(builder
, true_index_array
, ind_vec
, "");
2175 LLVMValueRef exec_mask
= lp_build_cmp(&blduivec
, PIPE_FUNC_LEQUAL
, true_index_array
, fetch_max
);
2177 * Limit indices to fetch_max, otherwise might try to access indices
2178 * beyond index buffer (or rather vsplit elt buffer) size.
2179 * Could probably safely (?) skip this for non-indexed draws and
2180 * simplify things minimally (by removing it could combine the ind_vec
2181 * and start_vec adds). I think the only effect for non-indexed draws will
2182 * be that for the invalid elements they will be all fetched from the
2183 * same location as the last valid one, but noone should really care.
2185 true_index_array
= lp_build_min(&blduivec
, true_index_array
, fetch_max
);
2187 index_store
= lp_build_alloca_undef(gallivm
, blduivec
.vec_type
, "index_store");
2189 lp_build_if(&if_ctx
, gallivm
, have_elts
);
2192 * Note: you'd expect some comparison/clamp against fetch_elt_max
2194 * There used to be one here but it was incorrect: overflow was
2195 * detected if index > fetch_elt_max - but the correct condition
2196 * would be index >= fetch_elt_max (since this is just size of elts
2197 * buffer / element size).
2198 * Using the correct condition however will cause failures - due to
2199 * vsplit/vcache code which rebases indices. So, as an example, if
2200 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
2201 * replace all invalid indices with 0 - which in case of elt_bias
2202 * not being zero will get a different fetch index than the valid
2203 * index 0. So, just rely on vsplit code preventing out-of-bounds
2204 * fetches. This is also why it's safe to do elts fetch even if there
2205 * was no index buffer bound - the real buffer is never seen here, at
2206 * least not if there are index buffer overflows...
2210 * XXX should not have to do this, as scale can be handled
2211 * natively by loads (hits asserts though).
2213 tmp
= lp_build_shl_imm(&blduivec
, true_index_array
, 2);
2214 fetch_elts
= LLVMBuildBitCast(builder
, fetch_elts
,
2215 LLVMPointerType(LLVMInt8TypeInContext(context
),
2217 tmp
= lp_build_gather(gallivm
, vs_type
.length
,
2219 fetch_elts
, tmp
, FALSE
);
2220 LLVMBuildStore(builder
, tmp
, index_store
);
2222 lp_build_else(&if_ctx
);
2224 tmp
= LLVMBuildAdd(builder
, true_index_array
, start_vec
, "");
2225 LLVMBuildStore(builder
, tmp
, index_store
);
2227 lp_build_endif(&if_ctx
);
2229 true_index_array
= LLVMBuildLoad(builder
, index_store
, "");
2231 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
2232 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
2233 const struct util_format_description
*format_desc
=
2234 util_format_description(velem
->src_format
);
2236 if (format_desc
->format
== PIPE_FORMAT_NONE
) {
2237 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2238 inputs
[j
][i
] = lp_build_zero(gallivm
, vs_type
);
2241 else if (velem
->instance_divisor
) {
2242 fetch_instanced(gallivm
, format_desc
, vs_type
,
2243 vb_stride
[j
], map_ptr
[j
],
2245 inputs
[j
], instance_index
[j
]);
2248 fetch_vector(gallivm
, format_desc
, vs_type
,
2249 vb_stride
[j
], map_ptr
[j
],
2251 inputs
[j
], true_index_array
);
2255 struct lp_build_mask_context mask
;
2257 lp_build_mask_begin(&mask
, gallivm
, vs_type
, exec_mask
);
2258 /* In the paths with elts vertex id has to be unaffected by the
2259 * index bias and because indices inside our elements array have
2260 * already had index bias applied we need to subtract it here to
2261 * get back to the original index.
2262 * in the linear paths vertex id has to be unaffected by the
2263 * original start index and because we abuse the 'start' variable
2264 * to either represent the actual start index or the index at which
2265 * the primitive was split (we split rendering into chunks of at
2266 * most 4095-vertices) we need to back out the original start
2267 * index out of our vertex id here.
2268 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2270 LLVMValueRef base_vertex
= lp_build_select(&bld
, have_elts
, vertex_id_offset
, lp_build_const_int32(gallivm
, 0));;
2271 system_values
.basevertex
= lp_build_broadcast_scalar(&blduivec
, base_vertex
);
2272 system_values
.vertex_id
= true_index_array
;
2273 system_values
.vertex_id_nobase
= LLVMBuildSub(builder
, true_index_array
,
2274 lp_build_broadcast_scalar(&blduivec
, vertex_id_offset
), "");
2276 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
2277 generate_vs(variant
,
2286 key
->clamp_vertex_color
,
2289 lp_build_mask_end(&mask
);
2290 if (pos
!= -1 && cv
!= -1) {
2291 /* store original positions in clip before further manipulation */
2292 store_clip(gallivm
, vs_type
, io
, outputs
, pos
);
2295 if (enable_cliptest
) {
2296 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
2297 /* allocate clipmask, assign it integer type */
2298 clipmask
= generate_clipmask(llvm
,
2303 context_ptr
, &have_clipdist
);
2304 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
2305 /* store temporary clipping boolean value */
2306 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
2309 clipmask
= blduivec
.zero
;
2312 /* do viewport mapping */
2313 if (!bypass_viewport
) {
2314 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
2318 clipmask
= blduivec
.zero
;
2321 /* store clipmask in vertex header,
2322 * original positions in clip
2323 * and transformed positions in data
2325 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
2326 vs_info
->num_outputs
, vs_type
,
2327 enable_cliptest
&& key
->need_edgeflags
);
2329 lp_build_loop_end_cond(&lp_loop
, count
, step
, LLVMIntUGE
);
2331 sampler
->destroy(sampler
);
2332 image
->destroy(image
);
2334 /* return clipping boolean value for function */
2335 ret
= clipmask_booli8(gallivm
, vs_type
, clipmask_bool_ptr
,
2336 enable_cliptest
&& key
->need_edgeflags
);
2338 LLVMBuildRet(builder
, ret
);
2340 gallivm_verify_function(gallivm
, variant_func
);
2344 struct draw_llvm_variant_key
*
2345 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2348 struct draw_llvm_variant_key
*key
;
2349 struct draw_sampler_static_state
*draw_sampler
;
2350 struct draw_image_static_state
*draw_image
;
2352 key
= (struct draw_llvm_variant_key
*)store
;
2354 memset(key
, 0, offsetof(struct draw_llvm_variant_key
, vertex_element
[0]));
2356 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
2358 /* will have to rig this up properly later */
2359 key
->clip_xy
= llvm
->draw
->clip_xy
;
2360 key
->clip_z
= llvm
->draw
->clip_z
;
2361 key
->clip_user
= llvm
->draw
->clip_user
;
2362 key
->bypass_viewport
= llvm
->draw
->bypass_viewport
;
2363 key
->clip_halfz
= llvm
->draw
->rasterizer
->clip_halfz
;
2364 /* XXX assumes edgeflag output not at 0 */
2365 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
2366 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
2367 key
->has_gs_or_tes
= llvm
->draw
->gs
.geometry_shader
!= NULL
|| llvm
->draw
->tes
.tess_eval_shader
!= NULL
;
2368 key
->num_outputs
= draw_total_vs_outputs(llvm
->draw
);
2370 /* All variants of this shader will have the same value for
2371 * nr_samplers. Not yet trying to compact away holes in the
2374 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2375 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2376 key
->nr_sampler_views
=
2377 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2380 key
->nr_sampler_views
= key
->nr_samplers
;
2383 key
->nr_images
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2385 /* Presumably all variants of the shader should have the same
2386 * number of vertex elements - ie the number of shader inputs.
2387 * NOTE: we NEED to store the needed number of needed inputs
2388 * here, not the number of provided elements to match keysize
2389 * (and the offset of sampler state in the key).
2390 * If we have excess number of vertex elements, this is valid,
2391 * but the excess ones don't matter.
2392 * If we don't have enough vertex elements (which looks not really
2393 * valid but we'll handle it gracefully) fill out missing ones with
2394 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2396 key
->nr_vertex_elements
=
2397 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
2399 if (llvm
->draw
->pt
.nr_vertex_elements
< key
->nr_vertex_elements
) {
2400 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2401 key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
);
2402 memset(key
->vertex_element
, 0,
2403 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
2405 memcpy(key
->vertex_element
,
2406 llvm
->draw
->pt
.vertex_element
,
2407 sizeof(struct pipe_vertex_element
) *
2408 MIN2(key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
));
2410 draw_sampler
= draw_llvm_variant_key_samplers(key
);
2411 memset(draw_sampler
, 0,
2412 MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2414 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2415 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2416 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
2418 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2419 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2420 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
2423 draw_image
= draw_llvm_variant_key_images(key
);
2424 memset(draw_image
, 0,
2425 key
->nr_images
* sizeof *draw_image
);
2426 for (i
= 0; i
< key
->nr_images
; i
++) {
2427 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
2428 llvm
->draw
->images
[PIPE_SHADER_VERTEX
][i
]);
2435 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
2438 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
2439 struct draw_image_static_state
*image
= draw_llvm_variant_key_images(key
);
2440 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
2441 debug_printf("clip_xy = %u\n", key
->clip_xy
);
2442 debug_printf("clip_z = %u\n", key
->clip_z
);
2443 debug_printf("clip_user = %u\n", key
->clip_user
);
2444 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
2445 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
2446 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
2447 debug_printf("has_gs_or_tes = %u\n", key
->has_gs_or_tes
);
2448 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
2450 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
2451 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
2452 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
2453 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
2454 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
2457 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2458 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
2461 for (i
= 0 ; i
< key
->nr_images
; i
++)
2462 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
2467 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
2468 enum pipe_shader_type shader_stage
,
2470 uint32_t width
, uint32_t height
, uint32_t depth
,
2471 uint32_t first_level
, uint32_t last_level
,
2472 uint32_t num_samples
,
2473 uint32_t sample_stride
,
2474 const void *base_ptr
,
2475 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2476 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2477 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
2480 struct draw_jit_texture
*jit_tex
;
2482 switch (shader_stage
) {
2483 case PIPE_SHADER_VERTEX
:
2484 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.textures
));
2485 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
2487 case PIPE_SHADER_GEOMETRY
:
2488 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.textures
));
2489 jit_tex
= &draw
->llvm
->gs_jit_context
.textures
[sview_idx
];
2491 case PIPE_SHADER_TESS_CTRL
:
2492 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->tcs_jit_context
.textures
));
2493 jit_tex
= &draw
->llvm
->tcs_jit_context
.textures
[sview_idx
];
2495 case PIPE_SHADER_TESS_EVAL
:
2496 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->tes_jit_context
.textures
));
2497 jit_tex
= &draw
->llvm
->tes_jit_context
.textures
[sview_idx
];
2504 jit_tex
->width
= width
;
2505 jit_tex
->height
= height
;
2506 jit_tex
->depth
= depth
;
2507 jit_tex
->first_level
= first_level
;
2508 jit_tex
->last_level
= last_level
;
2509 jit_tex
->base
= base_ptr
;
2510 jit_tex
->num_samples
= num_samples
;
2511 jit_tex
->sample_stride
= sample_stride
;
2513 for (j
= first_level
; j
<= last_level
; j
++) {
2514 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
2515 jit_tex
->row_stride
[j
] = row_stride
[j
];
2516 jit_tex
->img_stride
[j
] = img_stride
[j
];
2521 draw_llvm_set_mapped_image(struct draw_context
*draw
,
2522 enum pipe_shader_type shader_stage
,
2524 uint32_t width
, uint32_t height
, uint32_t depth
,
2525 const void *base_ptr
,
2526 uint32_t row_stride
,
2527 uint32_t img_stride
,
2528 uint32_t num_samples
,
2529 uint32_t sample_stride
)
2531 struct draw_jit_image
*jit_image
;
2533 switch (shader_stage
) {
2534 case PIPE_SHADER_VERTEX
:
2535 assert(idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.images
));
2536 jit_image
= &draw
->llvm
->jit_context
.images
[idx
];
2538 case PIPE_SHADER_GEOMETRY
:
2539 assert(idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.images
));
2540 jit_image
= &draw
->llvm
->gs_jit_context
.images
[idx
];
2542 case PIPE_SHADER_TESS_CTRL
:
2543 assert(idx
< ARRAY_SIZE(draw
->llvm
->tcs_jit_context
.images
));
2544 jit_image
= &draw
->llvm
->tcs_jit_context
.images
[idx
];
2546 case PIPE_SHADER_TESS_EVAL
:
2547 assert(idx
< ARRAY_SIZE(draw
->llvm
->tes_jit_context
.images
));
2548 jit_image
= &draw
->llvm
->tes_jit_context
.images
[idx
];
2555 jit_image
->width
= width
;
2556 jit_image
->height
= height
;
2557 jit_image
->depth
= depth
;
2558 jit_image
->base
= base_ptr
;
2560 jit_image
->row_stride
= row_stride
;
2561 jit_image
->img_stride
= img_stride
;
2562 jit_image
->num_samples
= num_samples
;
2563 jit_image
->sample_stride
= sample_stride
;
2568 draw_llvm_set_sampler_state(struct draw_context
*draw
,
2569 enum pipe_shader_type shader_type
)
2573 switch (shader_type
) {
2574 case PIPE_SHADER_VERTEX
:
2575 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
2576 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
2578 if (draw
->samplers
[PIPE_SHADER_VERTEX
][i
]) {
2579 const struct pipe_sampler_state
*s
2580 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
2581 jit_sam
->min_lod
= s
->min_lod
;
2582 jit_sam
->max_lod
= s
->max_lod
;
2583 jit_sam
->lod_bias
= s
->lod_bias
;
2584 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2588 case PIPE_SHADER_GEOMETRY
:
2589 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_GEOMETRY
]; i
++) {
2590 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->gs_jit_context
.samplers
[i
];
2592 if (draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]) {
2593 const struct pipe_sampler_state
*s
2594 = draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
];
2595 jit_sam
->min_lod
= s
->min_lod
;
2596 jit_sam
->max_lod
= s
->max_lod
;
2597 jit_sam
->lod_bias
= s
->lod_bias
;
2598 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2602 case PIPE_SHADER_TESS_CTRL
:
2603 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_TESS_CTRL
]; i
++) {
2604 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->tcs_jit_context
.samplers
[i
];
2606 if (draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
]) {
2607 const struct pipe_sampler_state
*s
2608 = draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
];
2609 jit_sam
->min_lod
= s
->min_lod
;
2610 jit_sam
->max_lod
= s
->max_lod
;
2611 jit_sam
->lod_bias
= s
->lod_bias
;
2612 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2616 case PIPE_SHADER_TESS_EVAL
:
2617 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_TESS_EVAL
]; i
++) {
2618 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->tes_jit_context
.samplers
[i
];
2620 if (draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
]) {
2621 const struct pipe_sampler_state
*s
2622 = draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
];
2623 jit_sam
->min_lod
= s
->min_lod
;
2624 jit_sam
->max_lod
= s
->max_lod
;
2625 jit_sam
->lod_bias
= s
->lod_bias
;
2626 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2638 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
2640 struct draw_llvm
*llvm
= variant
->llvm
;
2642 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2643 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2644 variant
->shader
->variants_cached
, llvm
->nr_variants
);
2647 gallivm_destroy(variant
->gallivm
);
2649 remove_from_list(&variant
->list_item_local
);
2650 variant
->shader
->variants_cached
--;
2651 remove_from_list(&variant
->list_item_global
);
2652 llvm
->nr_variants
--;
2658 * Create LLVM types for various structures.
2661 create_gs_jit_types(struct draw_gs_llvm_variant
*var
)
2663 struct gallivm_state
*gallivm
= var
->gallivm
;
2664 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
2666 texture_type
= create_jit_texture_type(gallivm
, "texture");
2667 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
2668 image_type
= create_jit_image_type(gallivm
, "image");
2670 context_type
= create_gs_jit_context_type(gallivm
,
2671 var
->shader
->base
.vector_length
,
2672 texture_type
, sampler_type
,
2674 "draw_gs_jit_context");
2675 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
2677 var
->input_array_type
= create_gs_jit_input_type(gallivm
);
2681 get_gs_context_ptr_type(struct draw_gs_llvm_variant
*variant
)
2683 if (!variant
->context_ptr_type
)
2684 create_gs_jit_types(variant
);
2685 return variant
->context_ptr_type
;
2689 generate_mask_value(struct draw_gs_llvm_variant
*variant
,
2690 struct lp_type gs_type
)
2692 struct gallivm_state
*gallivm
= variant
->gallivm
;
2693 LLVMBuilderRef builder
= gallivm
->builder
;
2694 struct lp_type mask_type
= lp_int_type(gs_type
);
2695 LLVMValueRef num_prims
;
2696 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
2699 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
),
2700 variant
->num_prims
);
2701 for (i
= 0; i
< gs_type
.length
; i
++) {
2702 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
2703 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, idx
, idx
, "");
2705 mask_val
= lp_build_compare(gallivm
, mask_type
,
2706 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
2712 draw_gs_llvm_generate(struct draw_llvm
*llvm
,
2713 struct draw_gs_llvm_variant
*variant
)
2715 struct gallivm_state
*gallivm
= variant
->gallivm
;
2716 LLVMContextRef context
= gallivm
->context
;
2717 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
2718 LLVMTypeRef arg_types
[7];
2719 LLVMTypeRef func_type
;
2720 LLVMValueRef variant_func
;
2721 LLVMValueRef context_ptr
;
2722 LLVMValueRef prim_id_ptr
;
2723 LLVMBasicBlockRef block
;
2724 LLVMBuilderRef builder
;
2725 LLVMValueRef io_ptr
, input_array
, num_prims
, mask_val
;
2726 struct lp_build_sampler_soa
*sampler
= 0;
2727 struct lp_build_image_soa
*image
= NULL
;
2728 struct lp_build_context bld
;
2729 struct lp_bld_tgsi_system_values system_values
;
2731 struct lp_type gs_type
;
2733 struct draw_gs_llvm_iface gs_iface
;
2734 const struct tgsi_token
*tokens
= variant
->shader
->base
.state
.tokens
;
2735 LLVMValueRef consts_ptr
, num_consts_ptr
;
2736 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
2737 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
2738 struct lp_build_mask_context mask
;
2739 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
2740 unsigned vector_length
= variant
->shader
->base
.vector_length
;
2742 memset(&system_values
, 0, sizeof(system_values
));
2743 memset(&outputs
, 0, sizeof(outputs
));
2745 snprintf(func_name
, sizeof(func_name
), "draw_llvm_gs_variant");
2747 assert(variant
->vertex_header_ptr_type
);
2749 arg_types
[0] = get_gs_context_ptr_type(variant
); /* context */
2750 arg_types
[1] = variant
->input_array_type
; /* input */
2751 arg_types
[2] = LLVMPointerType(variant
->vertex_header_ptr_type
, 0); /* vertex_header */
2752 arg_types
[3] = int32_type
; /* num_prims */
2753 arg_types
[4] = int32_type
; /* instance_id */
2754 arg_types
[5] = LLVMPointerType(
2755 LLVMVectorType(int32_type
, vector_length
), 0); /* prim_id_ptr */
2756 arg_types
[6] = int32_type
;
2758 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
2760 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2762 variant
->function
= variant_func
;
2764 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
2766 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2767 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2768 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2770 if (gallivm
->cache
&& gallivm
->cache
->data_size
)
2772 context_ptr
= LLVMGetParam(variant_func
, 0);
2773 input_array
= LLVMGetParam(variant_func
, 1);
2774 io_ptr
= LLVMGetParam(variant_func
, 2);
2775 num_prims
= LLVMGetParam(variant_func
, 3);
2776 system_values
.instance_id
= LLVMGetParam(variant_func
, 4);
2777 prim_id_ptr
= LLVMGetParam(variant_func
, 5);
2778 system_values
.invocation_id
= LLVMGetParam(variant_func
, 6);
2780 lp_build_name(context_ptr
, "context");
2781 lp_build_name(input_array
, "input");
2782 lp_build_name(io_ptr
, "io");
2783 lp_build_name(num_prims
, "num_prims");
2784 lp_build_name(system_values
.instance_id
, "instance_id");
2785 lp_build_name(prim_id_ptr
, "prim_id_ptr");
2786 lp_build_name(system_values
.invocation_id
, "invocation_id");
2788 variant
->context_ptr
= context_ptr
;
2789 variant
->io_ptr
= io_ptr
;
2790 variant
->num_prims
= num_prims
;
2792 gs_iface
.base
.fetch_input
= draw_gs_llvm_fetch_input
;
2793 gs_iface
.base
.emit_vertex
= draw_gs_llvm_emit_vertex
;
2794 gs_iface
.base
.end_primitive
= draw_gs_llvm_end_primitive
;
2795 gs_iface
.base
.gs_epilogue
= draw_gs_llvm_epilogue
;
2796 gs_iface
.input
= input_array
;
2797 gs_iface
.variant
= variant
;
2803 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
2804 builder
= gallivm
->builder
;
2805 LLVMPositionBuilderAtEnd(builder
, block
);
2807 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
2809 memset(&gs_type
, 0, sizeof gs_type
);
2810 gs_type
.floating
= TRUE
; /* floating point values */
2811 gs_type
.sign
= TRUE
; /* values are signed */
2812 gs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2813 gs_type
.width
= 32; /* 32-bit float */
2814 gs_type
.length
= vector_length
;
2816 consts_ptr
= draw_gs_jit_context_constants(variant
->gallivm
, context_ptr
);
2818 draw_gs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
2820 ssbos_ptr
= draw_gs_jit_context_ssbos(variant
->gallivm
, context_ptr
);
2822 draw_gs_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
2824 /* code generated texture sampling */
2825 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
, variant
->key
.nr_samplers
);
2826 image
= draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant
->key
),
2827 variant
->key
.nr_images
);
2828 mask_val
= generate_mask_value(variant
, gs_type
);
2829 lp_build_mask_begin(&mask
, gallivm
, gs_type
, mask_val
);
2831 if (gs_info
->uses_primid
) {
2832 system_values
.prim_id
= LLVMBuildLoad(builder
, prim_id_ptr
, "prim_id");
2835 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2836 if (llvm
->draw
->gs
.geometry_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
2837 tgsi_dump(tokens
, 0);
2839 nir_print_shader(llvm
->draw
->gs
.geometry_shader
->state
.ir
.nir
, stderr
);
2840 draw_gs_llvm_dump_variant_key(&variant
->key
);
2843 struct lp_build_tgsi_params params
;
2844 memset(¶ms
, 0, sizeof(params
));
2846 params
.type
= gs_type
;
2847 params
.mask
= &mask
;
2848 params
.consts_ptr
= consts_ptr
;
2849 params
.const_sizes_ptr
= num_consts_ptr
;
2850 params
.system_values
= &system_values
;
2851 params
.context_ptr
= context_ptr
;
2852 params
.sampler
= sampler
;
2853 params
.info
= &llvm
->draw
->gs
.geometry_shader
->info
;
2854 params
.gs_iface
= (const struct lp_build_gs_iface
*)&gs_iface
;
2855 params
.ssbo_ptr
= ssbos_ptr
;
2856 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
2857 params
.image
= image
;
2859 if (llvm
->draw
->gs
.geometry_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
2860 lp_build_tgsi_soa(variant
->gallivm
,
2865 lp_build_nir_soa(variant
->gallivm
,
2866 llvm
->draw
->gs
.geometry_shader
->state
.ir
.nir
,
2870 sampler
->destroy(sampler
);
2871 image
->destroy(image
);
2873 lp_build_mask_end(&mask
);
2875 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
2877 gallivm_verify_function(gallivm
, variant_func
);
2880 struct draw_gs_llvm_variant
*
2881 draw_gs_llvm_create_variant(struct draw_llvm
*llvm
,
2882 unsigned num_outputs
,
2883 const struct draw_gs_llvm_variant_key
*key
)
2885 struct draw_gs_llvm_variant
*variant
;
2886 struct llvm_geometry_shader
*shader
=
2887 llvm_geometry_shader(llvm
->draw
->gs
.geometry_shader
);
2888 LLVMTypeRef vertex_header
;
2889 char module_name
[64];
2890 unsigned char ir_sha1_cache_key
[20];
2891 struct lp_cached_code cached
= { 0 };
2892 bool needs_caching
= false;
2894 variant
= MALLOC(sizeof *variant
+
2895 shader
->variant_key_size
-
2896 sizeof variant
->key
);
2900 variant
->llvm
= llvm
;
2901 variant
->shader
= shader
;
2903 snprintf(module_name
, sizeof(module_name
), "draw_llvm_gs_variant%u",
2904 variant
->shader
->variants_cached
);
2906 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2908 if (shader
->base
.state
.ir
.nir
&& llvm
->draw
->disk_cache_cookie
) {
2909 draw_get_ir_cache_key(shader
->base
.state
.ir
.nir
,
2911 shader
->variant_key_size
,
2915 llvm
->draw
->disk_cache_find_shader(llvm
->draw
->disk_cache_cookie
,
2918 if (!cached
.data_size
)
2919 needs_caching
= true;
2921 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, &cached
);
2923 create_gs_jit_types(variant
);
2925 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
2927 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
2929 draw_gs_llvm_generate(llvm
, variant
);
2931 gallivm_compile_module(variant
->gallivm
);
2933 variant
->jit_func
= (draw_gs_jit_func
)
2934 gallivm_jit_function(variant
->gallivm
, variant
->function
);
2937 llvm
->draw
->disk_cache_insert_shader(llvm
->draw
->disk_cache_cookie
,
2940 gallivm_free_ir(variant
->gallivm
);
2942 variant
->list_item_global
.base
= variant
;
2943 variant
->list_item_local
.base
= variant
;
2944 /*variant->no = */shader
->variants_created
++;
2945 variant
->list_item_global
.base
= variant
;
2951 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant
*variant
)
2953 struct draw_llvm
*llvm
= variant
->llvm
;
2955 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2956 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2957 variant
->shader
->variants_cached
, llvm
->nr_gs_variants
);
2960 gallivm_destroy(variant
->gallivm
);
2962 remove_from_list(&variant
->list_item_local
);
2963 variant
->shader
->variants_cached
--;
2964 remove_from_list(&variant
->list_item_global
);
2965 llvm
->nr_gs_variants
--;
2969 struct draw_gs_llvm_variant_key
*
2970 draw_gs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2973 struct draw_gs_llvm_variant_key
*key
;
2974 struct draw_sampler_static_state
*draw_sampler
;
2975 struct draw_image_static_state
*draw_image
;
2977 key
= (struct draw_gs_llvm_variant_key
*)store
;
2979 memset(key
, 0, offsetof(struct draw_gs_llvm_variant_key
, samplers
[0]));
2981 key
->num_outputs
= draw_total_gs_outputs(llvm
->draw
);
2983 /* All variants of this shader will have the same value for
2984 * nr_samplers. Not yet trying to compact away holes in the
2987 key
->nr_samplers
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2988 if (llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2989 key
->nr_sampler_views
=
2990 llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2993 key
->nr_sampler_views
= key
->nr_samplers
;
2996 key
->nr_images
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2998 draw_sampler
= key
->samplers
;
3000 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
3002 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
3003 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
3004 llvm
->draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]);
3006 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3007 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
3008 llvm
->draw
->sampler_views
[PIPE_SHADER_GEOMETRY
][i
]);
3011 draw_image
= draw_gs_llvm_variant_key_images(key
);
3012 memset(draw_image
, 0,
3013 key
->nr_images
* sizeof *draw_image
);
3014 for (i
= 0; i
< key
->nr_images
; i
++) {
3015 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
3016 llvm
->draw
->images
[PIPE_SHADER_GEOMETRY
][i
]);
3022 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key
*key
)
3025 struct draw_sampler_static_state
*sampler
= key
->samplers
;
3026 struct draw_image_static_state
*image
= draw_gs_llvm_variant_key_images(key
);
3027 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3028 debug_printf("sampler[%i].src_format = %s\n", i
,
3029 util_format_name(sampler
[i
].texture_state
.format
));
3032 for (i
= 0 ; i
< key
->nr_images
; i
++)
3033 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
3038 create_tcs_jit_types(struct draw_tcs_llvm_variant
*var
)
3040 struct gallivm_state
*gallivm
= var
->gallivm
;
3041 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
3043 texture_type
= create_jit_texture_type(gallivm
, "texture");
3044 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
3045 image_type
= create_jit_image_type(gallivm
, "image");
3047 context_type
= create_tcs_jit_context_type(gallivm
,
3049 texture_type
, sampler_type
,
3051 "draw_tcs_jit_context");
3052 var
->input_array_type
= create_tcs_jit_input_type(gallivm
);
3053 var
->output_array_type
= create_tcs_jit_output_type(gallivm
);
3054 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
3058 get_tcs_context_ptr_type(struct draw_tcs_llvm_variant
*variant
)
3060 if (!variant
->context_ptr_type
)
3061 create_tcs_jit_types(variant
);
3062 return variant
->context_ptr_type
;
3066 draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface
*tes_iface
,
3067 struct lp_build_context
*bld
,
3068 boolean is_vindex_indirect
,
3069 LLVMValueRef vertex_index
,
3070 boolean is_aindex_indirect
,
3071 LLVMValueRef attrib_index
,
3072 LLVMValueRef swizzle_index
)
3074 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3075 struct gallivm_state
*gallivm
= bld
->gallivm
;
3076 LLVMBuilderRef builder
= gallivm
->builder
;
3077 LLVMValueRef indices
[3];
3079 struct lp_type type
= bld
->type
;
3081 if (is_vindex_indirect
|| is_aindex_indirect
) {
3085 for (i
= 0; i
< type
.length
; ++i
) {
3086 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3087 LLVMValueRef vert_chan_index
= vertex_index
;
3088 LLVMValueRef attr_chan_index
= attrib_index
;
3089 LLVMValueRef channel_vec
;
3091 if (is_vindex_indirect
) {
3092 vert_chan_index
= LLVMBuildExtractElement(builder
,
3093 vertex_index
, idx
, "");
3095 if (is_aindex_indirect
) {
3096 attr_chan_index
= LLVMBuildExtractElement(builder
,
3097 attrib_index
, idx
, "");
3100 indices
[0] = vert_chan_index
;
3101 indices
[1] = attr_chan_index
;
3102 indices
[2] = swizzle_index
;
3104 channel_vec
= LLVMBuildGEP(builder
, tcs
->input
, indices
, 3, "");
3105 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3107 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3110 indices
[0] = vertex_index
;
3111 indices
[1] = attrib_index
;
3112 indices
[2] = swizzle_index
;
3114 res
= LLVMBuildGEP(builder
, tcs
->input
, indices
, 3, "");
3115 res
= LLVMBuildLoad(builder
, res
, "");
3116 res
= lp_build_broadcast_scalar(bld
, res
);
3122 draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface
*tes_iface
,
3123 struct lp_build_context
*bld
,
3124 boolean is_vindex_indirect
,
3125 LLVMValueRef vertex_index
,
3126 boolean is_aindex_indirect
,
3127 LLVMValueRef attrib_index
,
3128 LLVMValueRef swizzle_index
,
3131 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3132 struct gallivm_state
*gallivm
= bld
->gallivm
;
3133 LLVMBuilderRef builder
= gallivm
->builder
;
3134 LLVMValueRef indices
[3];
3136 struct lp_type type
= bld
->type
;
3138 if (is_vindex_indirect
|| is_aindex_indirect
) {
3142 for (i
= 0; i
< type
.length
; ++i
) {
3143 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3144 LLVMValueRef vert_chan_index
= vertex_index
;
3145 LLVMValueRef attr_chan_index
= attrib_index
;
3146 LLVMValueRef channel_vec
;
3148 if (is_vindex_indirect
) {
3149 vert_chan_index
= LLVMBuildExtractElement(builder
,
3150 vertex_index
, idx
, "");
3152 if (is_aindex_indirect
) {
3153 attr_chan_index
= LLVMBuildExtractElement(builder
,
3154 attrib_index
, idx
, "");
3157 indices
[0] = vert_chan_index
;
3158 indices
[1] = attr_chan_index
;
3159 indices
[2] = swizzle_index
;
3161 channel_vec
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3162 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3164 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3167 indices
[0] = vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3168 indices
[1] = attrib_index
;
3169 indices
[2] = swizzle_index
;
3171 res
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3172 res
= LLVMBuildLoad(builder
, res
, "");
3173 res
= lp_build_broadcast_scalar(bld
, res
);
3179 draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface
*tes_iface
,
3180 struct lp_build_context
*bld
,
3182 boolean is_vindex_indirect
,
3183 LLVMValueRef vertex_index
,
3184 boolean is_aindex_indirect
,
3185 LLVMValueRef attrib_index
,
3186 LLVMValueRef swizzle_index
,
3188 LLVMValueRef mask_vec
)
3190 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3191 struct gallivm_state
*gallivm
= bld
->gallivm
;
3192 LLVMBuilderRef builder
= gallivm
->builder
;
3193 LLVMValueRef indices
[3];
3195 struct lp_type type
= bld
->type
;
3197 if (is_vindex_indirect
|| is_aindex_indirect
) {
3200 for (i
= 0; i
< type
.length
; ++i
) {
3201 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3202 LLVMValueRef vert_chan_index
= vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3203 LLVMValueRef attr_chan_index
= attrib_index
;
3204 LLVMValueRef channel_vec
;
3206 if (is_vindex_indirect
) {
3207 vert_chan_index
= LLVMBuildExtractElement(builder
,
3208 vertex_index
, idx
, "");
3210 if (is_aindex_indirect
) {
3211 attr_chan_index
= LLVMBuildExtractElement(builder
,
3212 attrib_index
, idx
, "");
3215 indices
[0] = vert_chan_index
;
3216 indices
[1] = attr_chan_index
;
3217 indices
[2] = swizzle_index
;
3219 channel_vec
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3221 res
= LLVMBuildExtractElement(builder
, value
, idx
, "");
3223 struct lp_build_if_state ifthen
;
3224 LLVMValueRef cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, mask_vec
, lp_build_const_int_vec(gallivm
, bld
->type
, 0), "");
3225 cond
= LLVMBuildExtractElement(gallivm
->builder
, cond
, idx
, "");
3226 lp_build_if(&ifthen
, gallivm
, cond
);
3227 LLVMBuildStore(builder
, res
, channel_vec
);
3228 lp_build_endif(&ifthen
);
3231 indices
[0] = vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3232 indices
[1] = attrib_index
;
3233 indices
[2] = swizzle_index
;
3235 res
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3236 for (unsigned i
= 0; i
< type
.length
; ++i
) {
3237 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3238 LLVMValueRef val
= LLVMBuildExtractElement(builder
, value
, idx
, "");
3240 struct lp_build_if_state ifthen
;
3241 LLVMValueRef cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, mask_vec
, lp_build_const_int_vec(gallivm
, bld
->type
, 0), "");
3242 cond
= LLVMBuildExtractElement(gallivm
->builder
, cond
, idx
, "");
3243 lp_build_if(&ifthen
, gallivm
, cond
);
3244 LLVMBuildStore(builder
, val
, res
);
3245 lp_build_endif(&ifthen
);
3252 generate_tcs_mask_value(struct draw_tcs_llvm_variant
*variant
,
3253 struct lp_type tcs_type
, LLVMValueRef limit
, LLVMValueRef loop_counter
)
3255 struct gallivm_state
*gallivm
= variant
->gallivm
;
3256 LLVMBuilderRef builder
= gallivm
->builder
;
3257 struct lp_type mask_type
= lp_int_type(tcs_type
);
3258 LLVMValueRef num_vecs
;
3259 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
3262 num_vecs
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
), limit
);
3263 for (i
= 0; i
< tcs_type
.length
; i
++) {
3264 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3265 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, LLVMBuildAdd(builder
, loop_counter
, idx
, ""), idx
, "");
3267 mask_val
= lp_build_compare(gallivm
, mask_type
,
3268 PIPE_FUNC_GREATER
, num_vecs
, mask_val
);
3274 draw_tcs_llvm_generate(struct draw_llvm
*llvm
,
3275 struct draw_tcs_llvm_variant
*variant
)
3277 struct gallivm_state
*gallivm
= variant
->gallivm
;
3278 LLVMContextRef context
= gallivm
->context
;
3279 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
3280 LLVMTypeRef arg_types
[6];
3281 LLVMTypeRef func_type
, coro_func_type
;
3282 LLVMValueRef variant_func
, variant_coro
;
3283 LLVMValueRef context_ptr
;
3284 LLVMValueRef input_array
, output_array
, prim_id
, patch_vertices_in
;
3285 LLVMValueRef mask_val
;
3286 LLVMBasicBlockRef block
;
3287 LLVMBuilderRef builder
;
3288 struct lp_build_context bld
, bldvec
;
3289 struct lp_build_sampler_soa
*sampler
= 0;
3290 struct lp_build_image_soa
*image
= NULL
;
3291 struct lp_bld_tgsi_system_values system_values
;
3292 char func_name
[64], func_name_coro
[64];
3294 struct draw_tcs_llvm_iface tcs_iface
;
3295 struct lp_build_mask_context mask
;
3296 LLVMValueRef consts_ptr
, num_consts_ptr
;
3297 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
3298 struct lp_type tcs_type
;
3299 unsigned vector_length
= variant
->shader
->base
.vector_length
;
3301 memset(&system_values
, 0, sizeof(system_values
));
3303 snprintf(func_name
, sizeof(func_name
), "draw_llvm_tcs_variant");
3305 snprintf(func_name_coro
, sizeof(func_name_coro
), "draw_llvm_tcs_coro_variant");
3307 arg_types
[0] = get_tcs_context_ptr_type(variant
); /* context */
3308 arg_types
[1] = variant
->input_array_type
; /* input */
3309 arg_types
[2] = variant
->output_array_type
;
3310 arg_types
[3] = int32_type
;
3311 arg_types
[4] = int32_type
;
3312 arg_types
[5] = int32_type
; /* coroutine only */
3314 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
) - 1, 0);
3316 coro_func_type
= LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), arg_types
, ARRAY_SIZE(arg_types
), 0);
3318 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
3320 variant_coro
= LLVMAddFunction(gallivm
->module
, func_name_coro
, coro_func_type
);
3322 variant
->function
= variant_func
;
3323 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
3325 LLVMSetFunctionCallConv(variant_coro
, LLVMCCallConv
);
3327 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
) {
3328 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
) {
3329 lp_add_function_attr(variant_coro
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3330 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3334 if (gallivm
->cache
&& gallivm
->cache
->data_size
)
3336 context_ptr
= LLVMGetParam(variant_func
, 0);
3337 input_array
= LLVMGetParam(variant_func
, 1);
3338 output_array
= LLVMGetParam(variant_func
, 2);
3339 prim_id
= LLVMGetParam(variant_func
, 3);
3340 patch_vertices_in
= LLVMGetParam(variant_func
, 4);
3342 lp_build_name(context_ptr
, "context");
3343 lp_build_name(input_array
, "input");
3344 lp_build_name(output_array
, "output");
3345 lp_build_name(prim_id
, "prim_id");
3346 lp_build_name(patch_vertices_in
, "patch_vertices_in");
3348 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
3349 builder
= gallivm
->builder
;
3350 LLVMPositionBuilderAtEnd(builder
, block
);
3352 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
3354 memset(&tcs_type
, 0, sizeof tcs_type
);
3355 tcs_type
.floating
= TRUE
; /* floating point values */
3356 tcs_type
.sign
= TRUE
; /* values are signed */
3357 tcs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
3358 tcs_type
.width
= 32; /* 32-bit float */
3359 tcs_type
.length
= vector_length
;
3361 lp_build_context_init(&bldvec
, variant
->gallivm
, lp_int_type(tcs_type
));
3363 LLVMValueRef count
= lp_build_const_int32(gallivm
, variant
->shader
->base
.vertices_out
);
3364 LLVMValueRef step
= lp_build_const_int32(gallivm
, vector_length
);
3366 struct lp_build_loop_state loop_state
[2];
3367 LLVMValueRef num_inner_loop
;
3368 unsigned count_align
= util_align_npot(variant
->shader
->base
.vertices_out
, tcs_type
.length
);
3369 num_inner_loop
= lp_build_const_int32(gallivm
, count_align
/ tcs_type
.length
);
3370 LLVMTypeRef hdl_ptr_type
= LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
3371 LLVMValueRef coro_hdls
= LLVMBuildArrayAlloca(gallivm
->builder
, hdl_ptr_type
, num_inner_loop
, "coro_hdls");
3372 unsigned end_coroutine
= INT_MAX
;
3373 lp_build_loop_begin(&loop_state
[1], gallivm
,
3374 lp_build_const_int32(gallivm
, 0)); /* coroutine reentry loop */
3375 lp_build_loop_begin(&loop_state
[0], gallivm
,
3376 lp_build_const_int32(gallivm
, 0)); /* inner loop */
3378 LLVMValueRef args
[6];
3379 args
[0] = context_ptr
;
3380 args
[1] = input_array
;
3381 args
[2] = output_array
;
3383 args
[4] = patch_vertices_in
;
3384 args
[5] = loop_state
[0].counter
;
3385 LLVMValueRef coro_entry
= LLVMBuildGEP(builder
, coro_hdls
, &loop_state
[0].counter
, 1, "");
3386 LLVMValueRef coro_hdl
= LLVMBuildLoad(builder
, coro_entry
, "coro_hdl");
3388 struct lp_build_if_state ifstate
;
3389 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, loop_state
[1].counter
,
3390 lp_build_const_int32(gallivm
, 0), "");
3391 /* first time here - call the coroutine function entry point */
3392 lp_build_if(&ifstate
, gallivm
, cmp
);
3393 LLVMValueRef coro_ret
= LLVMBuildCall(builder
, variant_coro
, args
, 6, "");
3394 LLVMBuildStore(builder
, coro_ret
, coro_entry
);
3395 lp_build_else(&ifstate
);
3396 /* subsequent calls for this invocation - check if done. */
3397 LLVMValueRef coro_done
= lp_build_coro_done(gallivm
, coro_hdl
);
3398 struct lp_build_if_state ifstate2
;
3399 lp_build_if(&ifstate2
, gallivm
, coro_done
);
3400 /* if done destroy and force loop exit */
3401 lp_build_coro_destroy(gallivm
, coro_hdl
);
3402 lp_build_loop_force_set_counter(&loop_state
[1], lp_build_const_int32(gallivm
, end_coroutine
- 1));
3403 lp_build_else(&ifstate2
);
3404 /* otherwise resume the coroutine */
3405 lp_build_coro_resume(gallivm
, coro_hdl
);
3406 lp_build_endif(&ifstate2
);
3407 lp_build_endif(&ifstate
);
3408 lp_build_loop_force_reload_counter(&loop_state
[1]);
3410 lp_build_loop_end_cond(&loop_state
[0],
3413 lp_build_loop_end_cond(&loop_state
[1],
3414 lp_build_const_int32(gallivm
, end_coroutine
),
3416 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
3418 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "entry");
3419 LLVMPositionBuilderAtEnd(builder
, block
);
3421 context_ptr
= LLVMGetParam(variant_coro
, 0);
3422 input_array
= LLVMGetParam(variant_coro
, 1);
3423 output_array
= LLVMGetParam(variant_coro
, 2);
3424 prim_id
= LLVMGetParam(variant_coro
, 3);
3425 patch_vertices_in
= LLVMGetParam(variant_coro
, 4);
3427 consts_ptr
= draw_tcs_jit_context_constants(variant
->gallivm
, context_ptr
);
3429 draw_tcs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
3431 ssbos_ptr
= draw_tcs_jit_context_ssbos(variant
->gallivm
, context_ptr
);
3433 draw_tcs_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
3434 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
, variant
->key
.nr_samplers
);
3435 image
= draw_llvm_image_soa_create(draw_tcs_llvm_variant_key_images(&variant
->key
),
3436 variant
->key
.nr_images
);
3438 LLVMValueRef counter
= LLVMGetParam(variant_coro
, 5);
3439 LLVMValueRef invocvec
= LLVMGetUndef(LLVMVectorType(int32_type
, vector_length
));
3440 for (i
= 0; i
< vector_length
; i
++) {
3441 LLVMValueRef idx
= LLVMBuildAdd(builder
, LLVMBuildMul(builder
, counter
, step
, ""), lp_build_const_int32(gallivm
, i
), "");
3442 invocvec
= LLVMBuildInsertElement(builder
, invocvec
, idx
, idx
, "");
3445 system_values
.invocation_id
= invocvec
;
3446 system_values
.prim_id
= lp_build_broadcast_scalar(&bldvec
, prim_id
);
3447 system_values
.vertices_in
= lp_build_broadcast_scalar(&bldvec
, patch_vertices_in
);
3448 tcs_iface
.input
= input_array
;
3449 tcs_iface
.output
= output_array
;
3450 tcs_iface
.base
.emit_fetch_input
= draw_tcs_llvm_emit_fetch_input
;
3451 tcs_iface
.base
.emit_fetch_output
= draw_tcs_llvm_emit_fetch_output
;
3452 tcs_iface
.base
.emit_store_output
= draw_tcs_llvm_emit_store_output
;
3456 LLVMValueRef coro_id
= lp_build_coro_id(gallivm
);
3457 LLVMValueRef coro_hdl
= lp_build_coro_begin_alloc_mem(gallivm
, coro_id
);
3459 mask_val
= generate_tcs_mask_value(variant
, tcs_type
, count
, LLVMBuildMul(builder
, counter
, step
, ""));
3460 lp_build_mask_begin(&mask
, gallivm
, tcs_type
, mask_val
);
3462 struct lp_build_coro_suspend_info coro_info
;
3464 LLVMBasicBlockRef sus_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "suspend");
3465 LLVMBasicBlockRef clean_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "cleanup");
3467 coro_info
.suspend
= sus_block
;
3468 coro_info
.cleanup
= clean_block
;
3470 struct lp_build_tgsi_params params
;
3471 memset(¶ms
, 0, sizeof(params
));
3473 params
.type
= tcs_type
;
3474 params
.mask
= &mask
;
3475 params
.consts_ptr
= consts_ptr
;
3476 params
.const_sizes_ptr
= num_consts_ptr
;
3477 params
.system_values
= &system_values
;
3478 params
.context_ptr
= context_ptr
;
3479 params
.sampler
= sampler
;
3480 params
.info
= &llvm
->draw
->tcs
.tess_ctrl_shader
->info
;
3481 params
.ssbo_ptr
= ssbos_ptr
;
3482 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
3483 params
.image
= image
;
3484 params
.coro
= &coro_info
;
3485 params
.tcs_iface
= &tcs_iface
.base
;
3487 lp_build_nir_soa(variant
->gallivm
,
3488 llvm
->draw
->tcs
.tess_ctrl_shader
->state
.ir
.nir
,
3491 lp_build_mask_end(&mask
);
3493 lp_build_coro_suspend_switch(gallivm
, &coro_info
, NULL
, true);
3494 LLVMPositionBuilderAtEnd(builder
, clean_block
);
3496 lp_build_coro_free_mem(gallivm
, coro_id
, coro_hdl
);
3498 LLVMBuildBr(builder
, sus_block
);
3499 LLVMPositionBuilderAtEnd(builder
, sus_block
);
3501 lp_build_coro_end(gallivm
, coro_hdl
);
3502 LLVMBuildRet(builder
, coro_hdl
);
3505 sampler
->destroy(sampler
);
3506 image
->destroy(image
);
3507 gallivm_verify_function(gallivm
, variant_func
);
3508 gallivm_verify_function(gallivm
, variant_coro
);
3511 struct draw_tcs_llvm_variant
*
3512 draw_tcs_llvm_create_variant(struct draw_llvm
*llvm
,
3513 unsigned num_outputs
,
3514 const struct draw_tcs_llvm_variant_key
*key
)
3516 struct draw_tcs_llvm_variant
*variant
;
3517 struct llvm_tess_ctrl_shader
*shader
= llvm_tess_ctrl_shader(llvm
->draw
->tcs
.tess_ctrl_shader
);
3518 char module_name
[64];
3519 unsigned char ir_sha1_cache_key
[20];
3520 struct lp_cached_code cached
= { 0 };
3521 bool needs_caching
= false;
3523 variant
= MALLOC(sizeof *variant
+
3524 shader
->variant_key_size
- sizeof variant
->key
);
3528 variant
->llvm
= llvm
;
3529 variant
->shader
= shader
;
3531 snprintf(module_name
, sizeof(module_name
), "draw_llvm_tcs_variant%u",
3532 variant
->shader
->variants_cached
);
3534 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3536 if (shader
->base
.state
.ir
.nir
&& llvm
->draw
->disk_cache_cookie
) {
3537 draw_get_ir_cache_key(shader
->base
.state
.ir
.nir
,
3539 shader
->variant_key_size
,
3543 llvm
->draw
->disk_cache_find_shader(llvm
->draw
->disk_cache_cookie
,
3546 if (!cached
.data_size
)
3547 needs_caching
= true;
3550 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, &cached
);
3552 create_tcs_jit_types(variant
);
3554 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3555 nir_print_shader(llvm
->draw
->tcs
.tess_ctrl_shader
->state
.ir
.nir
, stderr
);
3556 draw_tcs_llvm_dump_variant_key(&variant
->key
);
3559 lp_build_coro_declare_malloc_hooks(variant
->gallivm
);
3560 draw_tcs_llvm_generate(llvm
, variant
);
3562 gallivm_compile_module(variant
->gallivm
);
3564 lp_build_coro_add_malloc_hooks(variant
->gallivm
);
3565 variant
->jit_func
= (draw_tcs_jit_func
)
3566 gallivm_jit_function(variant
->gallivm
, variant
->function
);
3569 llvm
->draw
->disk_cache_insert_shader(llvm
->draw
->disk_cache_cookie
,
3572 gallivm_free_ir(variant
->gallivm
);
3574 variant
->list_item_global
.base
= variant
;
3575 variant
->list_item_local
.base
= variant
;
3576 /*variant->no = */shader
->variants_created
++;
3577 variant
->list_item_global
.base
= variant
;
3583 draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant
*variant
)
3585 struct draw_llvm
*llvm
= variant
->llvm
;
3587 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3588 debug_printf("Deleting TCS variant: %u tcs variants,\t%u total variants\n",
3589 variant
->shader
->variants_cached
, llvm
->nr_tcs_variants
);
3592 gallivm_destroy(variant
->gallivm
);
3594 remove_from_list(&variant
->list_item_local
);
3595 variant
->shader
->variants_cached
--;
3596 remove_from_list(&variant
->list_item_global
);
3597 llvm
->nr_tcs_variants
--;
3601 struct draw_tcs_llvm_variant_key
*
3602 draw_tcs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
3605 struct draw_tcs_llvm_variant_key
*key
;
3606 struct draw_sampler_static_state
*draw_sampler
;
3607 struct draw_image_static_state
*draw_image
;
3609 key
= (struct draw_tcs_llvm_variant_key
*)store
;
3611 memset(key
, 0, offsetof(struct draw_tcs_llvm_variant_key
, samplers
[0]));
3613 /* All variants of this shader will have the same value for
3614 * nr_samplers. Not yet trying to compact away holes in the
3617 key
->nr_samplers
= llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3618 if (llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3619 key
->nr_sampler_views
=
3620 llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3623 key
->nr_sampler_views
= key
->nr_samplers
;
3626 key
->nr_images
= llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
3628 draw_sampler
= key
->samplers
;
3630 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
3632 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
3633 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
3634 llvm
->draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
]);
3636 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3637 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
3638 llvm
->draw
->sampler_views
[PIPE_SHADER_TESS_CTRL
][i
]);
3641 draw_image
= draw_tcs_llvm_variant_key_images(key
);
3642 memset(draw_image
, 0,
3643 key
->nr_images
* sizeof *draw_image
);
3644 for (i
= 0; i
< key
->nr_images
; i
++) {
3645 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
3646 llvm
->draw
->images
[PIPE_SHADER_TESS_CTRL
][i
]);
3652 draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key
*key
)
3655 struct draw_sampler_static_state
*sampler
= key
->samplers
;
3656 struct draw_image_static_state
*image
= draw_tcs_llvm_variant_key_images(key
);
3657 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3658 debug_printf("sampler[%i].src_format = %s\n", i
,
3659 util_format_name(sampler
[i
].texture_state
.format
));
3662 for (i
= 0 ; i
< key
->nr_images
; i
++)
3663 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
3668 create_tes_jit_types(struct draw_tes_llvm_variant
*var
)
3670 struct gallivm_state
*gallivm
= var
->gallivm
;
3671 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
3673 texture_type
= create_jit_texture_type(gallivm
, "texture");
3674 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
3675 image_type
= create_jit_image_type(gallivm
, "image");
3677 context_type
= create_tes_jit_context_type(gallivm
,
3679 texture_type
, sampler_type
,
3681 "draw_tes_jit_context");
3682 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
3684 var
->input_array_type
= create_tes_jit_input_type(gallivm
);
3688 get_tes_context_ptr_type(struct draw_tes_llvm_variant
*variant
)
3690 if (!variant
->context_ptr_type
)
3691 create_tes_jit_types(variant
);
3692 return variant
->context_ptr_type
;
3696 generate_tes_mask_value(struct draw_tes_llvm_variant
*variant
,
3697 struct lp_type tes_type
, LLVMValueRef limit
, LLVMValueRef loop_counter
)
3699 struct gallivm_state
*gallivm
= variant
->gallivm
;
3700 LLVMBuilderRef builder
= gallivm
->builder
;
3701 struct lp_type mask_type
= lp_int_type(tes_type
);
3702 LLVMValueRef num_prims
;
3703 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
3706 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
), limit
);
3707 for (i
= 0; i
< tes_type
.length
; i
++) {
3708 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3709 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, LLVMBuildAdd(builder
, loop_counter
, idx
, ""), idx
, "");
3711 mask_val
= lp_build_compare(gallivm
, mask_type
,
3712 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
3718 draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface
*tes_iface
,
3719 struct lp_build_context
*bld
,
3720 boolean is_vindex_indirect
,
3721 LLVMValueRef vertex_index
,
3722 boolean is_aindex_indirect
,
3723 LLVMValueRef attrib_index
,
3724 LLVMValueRef swizzle_index
)
3726 const struct draw_tes_llvm_iface
*tes
= draw_tes_llvm_iface(tes_iface
);
3727 struct gallivm_state
*gallivm
= bld
->gallivm
;
3728 LLVMBuilderRef builder
= gallivm
->builder
;
3729 LLVMValueRef indices
[3];
3731 struct lp_type type
= bld
->type
;
3733 if (is_vindex_indirect
|| is_aindex_indirect
) {
3738 for (i
= 0; i
< type
.length
; ++i
) {
3739 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3740 LLVMValueRef vert_chan_index
= vertex_index
;
3741 LLVMValueRef attr_chan_index
= attrib_index
;
3742 LLVMValueRef channel_vec
;
3744 if (is_vindex_indirect
) {
3745 vert_chan_index
= LLVMBuildExtractElement(builder
,
3746 vertex_index
, idx
, "");
3748 if (is_aindex_indirect
) {
3749 attr_chan_index
= LLVMBuildExtractElement(builder
,
3750 attrib_index
, idx
, "");
3753 indices
[0] = vert_chan_index
;
3754 indices
[1] = attr_chan_index
;
3755 indices
[2] = swizzle_index
;
3757 channel_vec
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3758 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3760 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3763 indices
[0] = vertex_index
;
3764 indices
[1] = attrib_index
;
3765 indices
[2] = swizzle_index
;
3767 res
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3768 res
= LLVMBuildLoad(builder
, res
, "");
3769 res
= lp_build_broadcast_scalar(bld
, res
);
3775 draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface
*tes_iface
,
3776 struct lp_build_context
*bld
,
3777 boolean is_aindex_indirect
,
3778 LLVMValueRef attrib_index
,
3779 LLVMValueRef swizzle_index
)
3781 const struct draw_tes_llvm_iface
*tes
= draw_tes_llvm_iface(tes_iface
);
3782 struct gallivm_state
*gallivm
= bld
->gallivm
;
3783 LLVMBuilderRef builder
= gallivm
->builder
;
3784 LLVMValueRef indices
[3];
3786 struct lp_type type
= bld
->type
;
3788 if (is_aindex_indirect
) {
3793 for (i
= 0; i
< type
.length
; ++i
) {
3794 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3795 LLVMValueRef attr_chan_index
= attrib_index
;
3796 LLVMValueRef channel_vec
;
3798 if (is_aindex_indirect
) {
3799 attr_chan_index
= LLVMBuildExtractElement(builder
,
3800 attrib_index
, idx
, "");
3803 indices
[0] = lp_build_const_int32(gallivm
, 0);
3804 indices
[1] = attr_chan_index
;
3805 indices
[2] = swizzle_index
;
3807 channel_vec
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3808 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3810 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3813 indices
[0] = lp_build_const_int32(gallivm
, 0);
3814 indices
[1] = attrib_index
;
3815 indices
[2] = swizzle_index
;
3817 res
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3818 res
= LLVMBuildLoad(builder
, res
, "");
3819 res
= lp_build_broadcast_scalar(bld
, res
);
3825 draw_tes_llvm_generate(struct draw_llvm
*llvm
,
3826 struct draw_tes_llvm_variant
*variant
)
3828 struct gallivm_state
*gallivm
= variant
->gallivm
;
3829 LLVMContextRef context
= gallivm
->context
;
3830 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
3831 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(context
);
3832 LLVMTypeRef arg_types
[10];
3833 LLVMTypeRef func_type
;
3834 LLVMValueRef variant_func
;
3835 LLVMValueRef context_ptr
;
3836 LLVMValueRef tess_coord
[2], io_ptr
, input_array
, num_tess_coord
;
3837 LLVMValueRef tess_inner
, tess_outer
, prim_id
, patch_vertices_in
;
3838 LLVMBasicBlockRef block
;
3839 LLVMBuilderRef builder
;
3840 LLVMValueRef mask_val
;
3841 struct lp_build_context bld
, bldvec
;
3842 struct lp_build_sampler_soa
*sampler
= 0;
3843 struct lp_build_image_soa
*image
= NULL
;
3844 struct lp_bld_tgsi_system_values system_values
;
3847 struct draw_tes_llvm_iface tes_iface
;
3848 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
3849 struct lp_build_mask_context mask
;
3850 LLVMValueRef consts_ptr
, num_consts_ptr
;
3851 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
3853 struct lp_type tes_type
;
3854 unsigned vector_length
= variant
->shader
->base
.vector_length
;
3856 memset(&system_values
, 0, sizeof(system_values
));
3857 memset(&outputs
, 0, sizeof(outputs
));
3859 snprintf(func_name
, sizeof(func_name
), "draw_llvm_tes_variant");
3861 arg_types
[0] = get_tes_context_ptr_type(variant
); /* context */
3862 arg_types
[1] = variant
->input_array_type
; /* input */
3863 arg_types
[2] = variant
->vertex_header_ptr_type
;
3864 arg_types
[3] = int32_type
;
3865 arg_types
[4] = int32_type
;
3866 arg_types
[5] = LLVMPointerType(flt_type
, 0);
3867 arg_types
[6] = LLVMPointerType(flt_type
, 0);
3868 arg_types
[7] = LLVMPointerType(LLVMArrayType(flt_type
, 4), 0);
3869 arg_types
[8] = LLVMPointerType(LLVMArrayType(flt_type
, 2), 0);
3870 arg_types
[9] = int32_type
;
3872 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
3873 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
3875 variant
->function
= variant_func
;
3876 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
3878 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
3879 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
3880 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3882 if (gallivm
->cache
&& gallivm
->cache
->data_size
)
3884 context_ptr
= LLVMGetParam(variant_func
, 0);
3885 input_array
= LLVMGetParam(variant_func
, 1);
3886 io_ptr
= LLVMGetParam(variant_func
, 2);
3887 prim_id
= LLVMGetParam(variant_func
, 3);
3888 num_tess_coord
= LLVMGetParam(variant_func
, 4);
3889 tess_coord
[0] = LLVMGetParam(variant_func
, 5);
3890 tess_coord
[1] = LLVMGetParam(variant_func
, 6);
3891 tess_outer
= LLVMGetParam(variant_func
, 7);
3892 tess_inner
= LLVMGetParam(variant_func
, 8);
3893 patch_vertices_in
= LLVMGetParam(variant_func
, 9);
3895 lp_build_name(context_ptr
, "context");
3896 lp_build_name(input_array
, "input");
3897 lp_build_name(io_ptr
, "io");
3898 lp_build_name(prim_id
, "prim_id");
3899 lp_build_name(num_tess_coord
, "num_tess_coord");
3900 lp_build_name(tess_coord
[0], "tess_coord[0]");
3901 lp_build_name(tess_coord
[1], "tess_coord[1]");
3902 lp_build_name(tess_outer
, "tess_outer");
3903 lp_build_name(tess_inner
, "tess_inner");
3904 lp_build_name(patch_vertices_in
, "patch_vertices_in");
3906 tes_iface
.base
.fetch_vertex_input
= draw_tes_llvm_fetch_vertex_input
;
3907 tes_iface
.base
.fetch_patch_input
= draw_tes_llvm_fetch_patch_input
;
3908 tes_iface
.input
= input_array
;
3909 tes_iface
.variant
= variant
;
3911 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
3912 builder
= gallivm
->builder
;
3913 LLVMPositionBuilderAtEnd(builder
, block
);
3915 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
3917 memset(&tes_type
, 0, sizeof tes_type
);
3918 tes_type
.floating
= TRUE
; /* floating point values */
3919 tes_type
.sign
= TRUE
; /* values are signed */
3920 tes_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
3921 tes_type
.width
= 32; /* 32-bit float */
3922 tes_type
.length
= vector_length
;
3924 lp_build_context_init(&bldvec
, variant
->gallivm
, lp_int_type(tes_type
));
3925 consts_ptr
= draw_tes_jit_context_constants(variant
->gallivm
, context_ptr
);
3927 draw_tes_jit_context_num_constants(variant
->gallivm
, context_ptr
);
3929 ssbos_ptr
= draw_tes_jit_context_ssbos(variant
->gallivm
, context_ptr
);
3931 draw_tes_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
3932 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
, variant
->key
.nr_samplers
);
3933 image
= draw_llvm_image_soa_create(draw_tes_llvm_variant_key_images(&variant
->key
),
3934 variant
->key
.nr_images
);
3935 step
= lp_build_const_int32(gallivm
, vector_length
);
3937 system_values
.tess_outer
= LLVMBuildLoad(builder
, tess_outer
, "");
3938 system_values
.tess_inner
= LLVMBuildLoad(builder
, tess_inner
, "");
3940 system_values
.prim_id
= lp_build_broadcast_scalar(&bldvec
, prim_id
);
3942 system_values
.vertices_in
= lp_build_broadcast_scalar(&bldvec
, patch_vertices_in
);
3943 struct lp_build_loop_state lp_loop
;
3944 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
3948 io
= LLVMBuildGEP(builder
, io_ptr
, &lp_loop
.counter
, 1, "");
3949 mask_val
= generate_tes_mask_value(variant
, tes_type
, num_tess_coord
, lp_loop
.counter
);
3950 lp_build_mask_begin(&mask
, gallivm
, tes_type
, mask_val
);
3952 system_values
.tess_coord
= LLVMGetUndef(LLVMArrayType(LLVMVectorType(flt_type
, vector_length
), 3));
3953 for (i
= 0; i
< 3; i
++) {
3954 LLVMValueRef tess_coord_chan
= LLVMGetUndef(LLVMVectorType(flt_type
, vector_length
));
3955 for (unsigned j
= 0; j
< vector_length
; j
++) {
3956 LLVMValueRef idx
= LLVMBuildAdd(builder
, lp_loop
.counter
, lp_build_const_int32(gallivm
, j
), "");
3957 LLVMValueRef tc_val
;
3959 if (variant
->shader
->base
.prim_mode
== PIPE_PRIM_TRIANGLES
) {
3960 tc_val
= lp_build_const_float(gallivm
, 1.0);
3961 tc_val
= LLVMBuildFSub(builder
, tc_val
, lp_build_pointer_get(builder
, tess_coord
[0], idx
), "");
3962 tc_val
= LLVMBuildFSub(builder
, tc_val
, lp_build_pointer_get(builder
, tess_coord
[1], idx
), "");
3964 tc_val
= lp_build_const_float(gallivm
, 0.0);
3966 tc_val
= lp_build_pointer_get(builder
, tess_coord
[i
], idx
);
3968 tess_coord_chan
= LLVMBuildInsertElement(builder
, tess_coord_chan
, tc_val
, lp_build_const_int32(gallivm
, j
), "");
3970 system_values
.tess_coord
= LLVMBuildInsertValue(builder
, system_values
.tess_coord
, tess_coord_chan
, i
, "");
3973 struct lp_build_tgsi_params params
;
3974 memset(¶ms
, 0, sizeof(params
));
3976 params
.type
= tes_type
;
3977 params
.mask
= &mask
;
3978 params
.consts_ptr
= consts_ptr
;
3979 params
.const_sizes_ptr
= num_consts_ptr
;
3980 params
.system_values
= &system_values
;
3981 params
.context_ptr
= context_ptr
;
3982 params
.sampler
= sampler
;
3983 params
.info
= &llvm
->draw
->tes
.tess_eval_shader
->info
;
3984 params
.ssbo_ptr
= ssbos_ptr
;
3985 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
3986 params
.image
= image
;
3987 params
.tes_iface
= &tes_iface
.base
;
3989 lp_build_nir_soa(variant
->gallivm
,
3990 llvm
->draw
->tes
.tess_eval_shader
->state
.ir
.nir
,
3994 lp_build_mask_end(&mask
);
3995 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
3996 lp_int_type(tes_type
), 0);
3998 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
3999 params
.info
->num_outputs
, tes_type
, FALSE
);
4001 lp_build_loop_end_cond(&lp_loop
, num_tess_coord
, step
, LLVMIntUGE
);
4002 sampler
->destroy(sampler
);
4003 image
->destroy(image
);
4005 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
4006 gallivm_verify_function(gallivm
, variant_func
);
4009 struct draw_tes_llvm_variant
*
4010 draw_tes_llvm_create_variant(struct draw_llvm
*llvm
,
4011 unsigned num_outputs
,
4012 const struct draw_tes_llvm_variant_key
*key
)
4014 struct draw_tes_llvm_variant
*variant
;
4015 struct llvm_tess_eval_shader
*shader
= llvm_tess_eval_shader(llvm
->draw
->tes
.tess_eval_shader
);
4016 LLVMTypeRef vertex_header
;
4017 char module_name
[64];
4018 unsigned char ir_sha1_cache_key
[20];
4019 struct lp_cached_code cached
= { 0 };
4020 bool needs_caching
= false;
4022 variant
= MALLOC(sizeof *variant
+
4023 shader
->variant_key_size
- sizeof variant
->key
);
4027 variant
->llvm
= llvm
;
4028 variant
->shader
= shader
;
4030 snprintf(module_name
, sizeof(module_name
), "draw_llvm_tes_variant%u",
4031 variant
->shader
->variants_cached
);
4033 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
4034 if (shader
->base
.state
.ir
.nir
&& llvm
->draw
->disk_cache_cookie
) {
4035 draw_get_ir_cache_key(shader
->base
.state
.ir
.nir
,
4037 shader
->variant_key_size
,
4041 llvm
->draw
->disk_cache_find_shader(llvm
->draw
->disk_cache_cookie
,
4044 if (!cached
.data_size
)
4045 needs_caching
= true;
4047 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, &cached
);
4049 create_tes_jit_types(variant
);
4051 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
4053 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
4055 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
4056 nir_print_shader(llvm
->draw
->tes
.tess_eval_shader
->state
.ir
.nir
, stderr
);
4057 draw_tes_llvm_dump_variant_key(&variant
->key
);
4060 draw_tes_llvm_generate(llvm
, variant
);
4062 gallivm_compile_module(variant
->gallivm
);
4064 variant
->jit_func
= (draw_tes_jit_func
)
4065 gallivm_jit_function(variant
->gallivm
, variant
->function
);
4068 llvm
->draw
->disk_cache_insert_shader(llvm
->draw
->disk_cache_cookie
,
4071 gallivm_free_ir(variant
->gallivm
);
4073 variant
->list_item_global
.base
= variant
;
4074 variant
->list_item_local
.base
= variant
;
4075 /*variant->no = */shader
->variants_created
++;
4076 variant
->list_item_global
.base
= variant
;
4082 draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant
*variant
)
4084 struct draw_llvm
*llvm
= variant
->llvm
;
4086 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
4087 debug_printf("Deleting TES variant: %u tes variants,\t%u total variants\n",
4088 variant
->shader
->variants_cached
, llvm
->nr_tes_variants
);
4091 gallivm_destroy(variant
->gallivm
);
4093 remove_from_list(&variant
->list_item_local
);
4094 variant
->shader
->variants_cached
--;
4095 remove_from_list(&variant
->list_item_global
);
4096 llvm
->nr_tes_variants
--;
4100 struct draw_tes_llvm_variant_key
*
4101 draw_tes_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
4104 struct draw_tes_llvm_variant_key
*key
;
4105 struct draw_sampler_static_state
*draw_sampler
;
4106 struct draw_image_static_state
*draw_image
;
4108 key
= (struct draw_tes_llvm_variant_key
*)store
;
4110 memset(key
, 0, offsetof(struct draw_tes_llvm_variant_key
, samplers
[0]));
4112 /* All variants of this shader will have the same value for
4113 * nr_samplers. Not yet trying to compact away holes in the
4116 key
->nr_samplers
= llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
4117 if (llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
4118 key
->nr_sampler_views
=
4119 llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
4122 key
->nr_sampler_views
= key
->nr_samplers
;
4125 key
->nr_images
= llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
4127 draw_sampler
= key
->samplers
;
4129 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
4131 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
4132 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
4133 llvm
->draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
]);
4135 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
4136 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
4137 llvm
->draw
->sampler_views
[PIPE_SHADER_TESS_EVAL
][i
]);
4140 draw_image
= draw_tes_llvm_variant_key_images(key
);
4141 memset(draw_image
, 0,
4142 key
->nr_images
* sizeof *draw_image
);
4143 for (i
= 0; i
< key
->nr_images
; i
++) {
4144 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
4145 llvm
->draw
->images
[PIPE_SHADER_TESS_EVAL
][i
]);
4151 draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key
*key
)
4154 struct draw_sampler_static_state
*sampler
= key
->samplers
;
4155 struct draw_image_static_state
*image
= draw_tes_llvm_variant_key_images(key
);
4156 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
4157 debug_printf("sampler[%i].src_format = %s\n", i
,
4158 util_format_name(sampler
[i
].texture_state
.format
));
4161 for (i
= 0 ; i
< key
->nr_images
; i
++)
4162 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));