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"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_logic.h"
35 #include "gallivm/lp_bld_const.h"
36 #include "gallivm/lp_bld_swizzle.h"
37 #include "gallivm/lp_bld_struct.h"
38 #include "gallivm/lp_bld_type.h"
39 #include "gallivm/lp_bld_flow.h"
40 #include "gallivm/lp_bld_debug.h"
41 #include "gallivm/lp_bld_tgsi.h"
42 #include "gallivm/lp_bld_printf.h"
43 #include "gallivm/lp_bld_intr.h"
44 #include "gallivm/lp_bld_init.h"
45 #include "gallivm/lp_bld_type.h"
46 #include "gallivm/lp_bld_pack.h"
47 #include "gallivm/lp_bld_format.h"
49 #include "tgsi/tgsi_exec.h"
50 #include "tgsi/tgsi_dump.h"
52 #include "util/u_math.h"
53 #include "util/u_pointer.h"
54 #include "util/u_string.h"
55 #include "util/u_simple_list.h"
62 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*var
,
67 * Create LLVM type for struct draw_jit_texture
70 create_jit_texture_type(struct gallivm_state
*gallivm
, const char *struct_name
)
72 LLVMTargetDataRef target
= gallivm
->target
;
73 LLVMTypeRef texture_type
;
74 LLVMTypeRef elem_types
[DRAW_JIT_TEXTURE_NUM_FIELDS
];
75 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
77 elem_types
[DRAW_JIT_TEXTURE_WIDTH
] =
78 elem_types
[DRAW_JIT_TEXTURE_HEIGHT
] =
79 elem_types
[DRAW_JIT_TEXTURE_DEPTH
] =
80 elem_types
[DRAW_JIT_TEXTURE_FIRST_LEVEL
] =
81 elem_types
[DRAW_JIT_TEXTURE_LAST_LEVEL
] = int32_type
;
82 elem_types
[DRAW_JIT_TEXTURE_BASE
] =
83 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
84 elem_types
[DRAW_JIT_TEXTURE_ROW_STRIDE
] =
85 elem_types
[DRAW_JIT_TEXTURE_IMG_STRIDE
] =
86 elem_types
[DRAW_JIT_TEXTURE_MIP_OFFSETS
] =
87 LLVMArrayType(int32_type
, PIPE_MAX_TEXTURE_LEVELS
);
89 texture_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
90 Elements(elem_types
), 0);
92 #if HAVE_LLVM < 0x0300
93 LLVMAddTypeName(gallivm
->module
, struct_name
, texture_type
);
95 /* Make sure the target's struct layout cache doesn't return
98 LLVMInvalidateStructLayout(gallivm
->target
, texture_type
);
101 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, width
,
102 target
, texture_type
,
103 DRAW_JIT_TEXTURE_WIDTH
);
104 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, height
,
105 target
, texture_type
,
106 DRAW_JIT_TEXTURE_HEIGHT
);
107 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, depth
,
108 target
, texture_type
,
109 DRAW_JIT_TEXTURE_DEPTH
);
110 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, first_level
,
111 target
, texture_type
,
112 DRAW_JIT_TEXTURE_FIRST_LEVEL
);
113 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, last_level
,
114 target
, texture_type
,
115 DRAW_JIT_TEXTURE_LAST_LEVEL
);
116 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, base
,
117 target
, texture_type
,
118 DRAW_JIT_TEXTURE_BASE
);
119 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, row_stride
,
120 target
, texture_type
,
121 DRAW_JIT_TEXTURE_ROW_STRIDE
);
122 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, img_stride
,
123 target
, texture_type
,
124 DRAW_JIT_TEXTURE_IMG_STRIDE
);
125 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, mip_offsets
,
126 target
, texture_type
,
127 DRAW_JIT_TEXTURE_MIP_OFFSETS
);
129 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture
, target
, texture_type
);
136 * Create LLVM type for struct draw_jit_sampler
139 create_jit_sampler_type(struct gallivm_state
*gallivm
, const char *struct_name
)
141 LLVMTargetDataRef target
= gallivm
->target
;
142 LLVMTypeRef sampler_type
;
143 LLVMTypeRef elem_types
[DRAW_JIT_SAMPLER_NUM_FIELDS
];
145 elem_types
[DRAW_JIT_SAMPLER_MIN_LOD
] =
146 elem_types
[DRAW_JIT_SAMPLER_MAX_LOD
] =
147 elem_types
[DRAW_JIT_SAMPLER_LOD_BIAS
] = LLVMFloatTypeInContext(gallivm
->context
);
148 elem_types
[DRAW_JIT_SAMPLER_BORDER_COLOR
] =
149 LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
151 sampler_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
152 Elements(elem_types
), 0);
154 #if HAVE_LLVM < 0x0300
155 LLVMAddTypeName(gallivm
->module
, struct_name
, sampler_type
);
157 /* Make sure the target's struct layout cache doesn't return
158 * stale/invalid data.
160 LLVMInvalidateStructLayout(gallivm
->target
, sampler_type
);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, min_lod
,
164 target
, sampler_type
,
165 DRAW_JIT_SAMPLER_MIN_LOD
);
166 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, max_lod
,
167 target
, sampler_type
,
168 DRAW_JIT_SAMPLER_MAX_LOD
);
169 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, lod_bias
,
170 target
, sampler_type
,
171 DRAW_JIT_SAMPLER_LOD_BIAS
);
172 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, border_color
,
173 target
, sampler_type
,
174 DRAW_JIT_SAMPLER_BORDER_COLOR
);
176 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler
, target
, sampler_type
);
183 * Create LLVM type for struct draw_jit_context
186 create_jit_context_type(struct gallivm_state
*gallivm
,
187 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
188 const char *struct_name
)
190 LLVMTargetDataRef target
= gallivm
->target
;
191 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
192 LLVMTypeRef elem_types
[5];
193 LLVMTypeRef context_type
;
195 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* vs_constants */
196 LP_MAX_TGSI_CONST_BUFFERS
);
197 elem_types
[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
198 DRAW_TOTAL_CLIP_PLANES
), 0);
199 elem_types
[2] = LLVMPointerType(float_type
, 0); /* viewport */
200 elem_types
[3] = LLVMArrayType(texture_type
,
201 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
202 elem_types
[4] = LLVMArrayType(sampler_type
,
203 PIPE_MAX_SAMPLERS
); /* samplers */
204 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
205 Elements(elem_types
), 0);
206 #if HAVE_LLVM < 0x0300
207 LLVMAddTypeName(gallivm
->module
, struct_name
, context_type
);
209 LLVMInvalidateStructLayout(gallivm
->target
, context_type
);
212 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_constants
,
213 target
, context_type
, 0);
214 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, planes
,
215 target
, context_type
, 1);
216 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, viewport
,
217 target
, context_type
, 2);
218 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, textures
,
219 target
, context_type
,
220 DRAW_JIT_CTX_TEXTURES
);
221 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, samplers
,
222 target
, context_type
,
223 DRAW_JIT_CTX_SAMPLERS
);
224 LP_CHECK_STRUCT_SIZE(struct draw_jit_context
,
225 target
, context_type
);
232 * Create LLVM type for struct pipe_vertex_buffer
235 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
, const char *struct_name
)
237 LLVMTargetDataRef target
= gallivm
->target
;
238 LLVMTypeRef elem_types
[4];
242 elem_types
[1] = LLVMInt32TypeInContext(gallivm
->context
);
244 elem_types
[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0); /* vs_constants */
246 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
247 Elements(elem_types
), 0);
248 #if HAVE_LLVM < 0x0300
249 LLVMAddTypeName(gallivm
->module
, struct_name
, vb_type
);
251 LLVMInvalidateStructLayout(gallivm
->target
, vb_type
);
254 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
256 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
259 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
266 * Create LLVM type for struct vertex_header;
269 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
271 LLVMTargetDataRef target
= gallivm
->target
;
272 LLVMTypeRef elem_types
[4];
273 LLVMTypeRef vertex_header
;
274 char struct_name
[24];
276 util_snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
278 elem_types
[DRAW_JIT_VERTEX_VERTEX_ID
] = LLVMIntTypeInContext(gallivm
->context
, 32);
279 elem_types
[DRAW_JIT_VERTEX_CLIP
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
280 elem_types
[DRAW_JIT_VERTEX_PRE_CLIP_POS
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
281 elem_types
[DRAW_JIT_VERTEX_DATA
] = LLVMArrayType(elem_types
[1], data_elems
);
283 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
284 Elements(elem_types
), 0);
285 #if HAVE_LLVM < 0x0300
286 LLVMAddTypeName(gallivm
->module
, struct_name
, vertex_header
);
288 LLVMInvalidateStructLayout(gallivm
->target
, vertex_header
);
291 /* these are bit-fields and we can't take address of them
292 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
293 target, vertex_header,
294 DRAW_JIT_VERTEX_CLIPMASK);
295 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
296 target, vertex_header,
297 DRAW_JIT_VERTEX_EDGEFLAG);
298 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
299 target, vertex_header,
300 DRAW_JIT_VERTEX_PAD);
301 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
302 target, vertex_header,
303 DRAW_JIT_VERTEX_VERTEX_ID);
305 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, clip
,
306 target
, vertex_header
,
307 DRAW_JIT_VERTEX_CLIP
);
308 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, pre_clip_pos
,
309 target
, vertex_header
,
310 DRAW_JIT_VERTEX_PRE_CLIP_POS
);
311 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, data
,
312 target
, vertex_header
,
313 DRAW_JIT_VERTEX_DATA
);
315 assert(LLVMABISizeOfType(target
, vertex_header
) ==
316 offsetof(struct vertex_header
, data
[data_elems
]));
318 return vertex_header
;
323 * Create LLVM types for various structures.
326 create_jit_types(struct draw_llvm_variant
*variant
)
328 struct gallivm_state
*gallivm
= variant
->gallivm
;
329 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
, vb_type
;
331 texture_type
= create_jit_texture_type(gallivm
, "texture");
332 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
334 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
336 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
338 buffer_type
= LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
339 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
341 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
342 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
347 get_context_ptr_type(struct draw_llvm_variant
*variant
)
349 if (!variant
->context_ptr_type
)
350 create_jit_types(variant
);
351 return variant
->context_ptr_type
;
356 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
358 if (!variant
->buffer_ptr_type
)
359 create_jit_types(variant
);
360 return variant
->buffer_ptr_type
;
365 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
367 if (!variant
->vb_ptr_type
)
368 create_jit_types(variant
);
369 return variant
->vb_ptr_type
;
373 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
375 if (!variant
->vertex_header_ptr_type
)
376 create_jit_types(variant
);
377 return variant
->vertex_header_ptr_type
;
382 * Create per-context LLVM info.
385 draw_llvm_create(struct draw_context
*draw
)
387 struct draw_llvm
*llvm
;
389 llvm
= CALLOC_STRUCT( draw_llvm
);
397 llvm
->nr_variants
= 0;
398 make_empty_list(&llvm
->vs_variants_list
);
405 * Free per-context LLVM info.
408 draw_llvm_destroy(struct draw_llvm
*llvm
)
410 /* XXX free other draw_llvm data? */
416 * Create LLVM-generated code for a vertex shader.
418 struct draw_llvm_variant
*
419 draw_llvm_create_variant(struct draw_llvm
*llvm
,
421 const struct draw_llvm_variant_key
*key
)
423 struct draw_llvm_variant
*variant
;
424 struct llvm_vertex_shader
*shader
=
425 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
426 LLVMTypeRef vertex_header
;
428 variant
= MALLOC(sizeof *variant
+
429 shader
->variant_key_size
-
430 sizeof variant
->key
);
434 variant
->llvm
= llvm
;
436 variant
->gallivm
= gallivm_create();
438 create_jit_types(variant
);
440 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
442 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
444 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
446 draw_llvm_generate(llvm
, variant
, FALSE
); /* linear */
447 draw_llvm_generate(llvm
, variant
, TRUE
); /* elts */
449 gallivm_compile_module(variant
->gallivm
);
451 variant
->jit_func
= (draw_jit_vert_func
)
452 gallivm_jit_function(variant
->gallivm
, variant
->function
);
454 variant
->jit_func_elts
= (draw_jit_vert_func_elts
)
455 gallivm_jit_function(variant
->gallivm
, variant
->function_elts
);
457 variant
->shader
= shader
;
458 variant
->list_item_global
.base
= variant
;
459 variant
->list_item_local
.base
= variant
;
460 /*variant->no = */shader
->variants_created
++;
461 variant
->list_item_global
.base
= variant
;
468 generate_vs(struct draw_llvm_variant
*variant
,
469 LLVMBuilderRef builder
,
470 struct lp_type vs_type
,
471 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
472 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
473 const struct lp_bld_tgsi_system_values
*system_values
,
474 LLVMValueRef context_ptr
,
475 struct lp_build_sampler_soa
*draw_sampler
,
476 boolean clamp_vertex_color
)
478 struct draw_llvm
*llvm
= variant
->llvm
;
479 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
480 LLVMValueRef consts_ptr
= draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
481 struct lp_build_sampler_soa
*sampler
= 0;
483 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
484 tgsi_dump(tokens
, 0);
485 draw_llvm_dump_variant_key(&variant
->key
);
488 if (llvm
->draw
->num_sampler_views
&& llvm
->draw
->num_samplers
)
489 sampler
= draw_sampler
;
491 lp_build_tgsi_soa(variant
->gallivm
,
494 NULL
/*struct lp_build_mask_context *mask*/,
501 &llvm
->draw
->vs
.vertex_shader
->info
);
505 unsigned chan
, attrib
;
506 struct lp_build_context bld
;
507 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
508 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
510 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
511 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
512 if (outputs
[attrib
][chan
]) {
513 switch (info
->output_semantic_name
[attrib
]) {
514 case TGSI_SEMANTIC_COLOR
:
515 case TGSI_SEMANTIC_BCOLOR
:
516 if (clamp_vertex_color
) {
517 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
518 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
519 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
522 case TGSI_SEMANTIC_FOG
:
523 if (chan
== 1 || chan
== 2)
524 LLVMBuildStore(builder
, bld
.zero
, outputs
[attrib
][chan
]);
526 LLVMBuildStore(builder
, bld
.one
, outputs
[attrib
][chan
]);
537 generate_fetch(struct gallivm_state
*gallivm
,
538 LLVMValueRef vbuffers_ptr
,
540 struct pipe_vertex_element
*velem
,
543 LLVMValueRef instance_id
)
545 const struct util_format_description
*format_desc
= util_format_description(velem
->src_format
);
546 LLVMValueRef zero
= LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
));
547 LLVMBuilderRef builder
= gallivm
->builder
;
548 LLVMValueRef indices
=
549 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
),
550 velem
->vertex_buffer_index
, 0);
551 LLVMValueRef vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
,
553 LLVMValueRef vb_stride
= draw_jit_vbuffer_stride(gallivm
, vbuf
);
554 LLVMValueRef vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vbuf
);
557 if (velem
->instance_divisor
) {
558 /* array index = instance_id / instance_divisor */
559 index
= LLVMBuildUDiv(builder
, instance_id
,
560 lp_build_const_int32(gallivm
, velem
->instance_divisor
),
564 stride
= LLVMBuildMul(builder
, vb_stride
, index
, "");
566 vbuffer_ptr
= LLVMBuildLoad(builder
, vbuffer_ptr
, "vbuffer");
568 stride
= LLVMBuildAdd(builder
, stride
,
571 stride
= LLVMBuildAdd(builder
, stride
,
572 lp_build_const_int32(gallivm
, velem
->src_offset
),
575 /* lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);*/
576 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffer_ptr
, &stride
, 1, "");
578 *res
= lp_build_fetch_rgba_aos(gallivm
,
580 lp_float32_vec4_type(),
586 convert_to_soa(struct gallivm_state
*gallivm
,
587 LLVMValueRef (*src_aos
)[LP_MAX_VECTOR_WIDTH
/ 32],
588 LLVMValueRef (*dst_soa
)[TGSI_NUM_CHANNELS
],
589 unsigned num_attribs
, const struct lp_type soa_type
)
592 struct lp_type aos_channel_type
= soa_type
;
594 debug_assert(TGSI_NUM_CHANNELS
== 4);
595 debug_assert((soa_type
.length
% TGSI_NUM_CHANNELS
) == 0);
597 aos_channel_type
.length
>>= 1;
599 for (i
= 0; i
< num_attribs
; ++i
) {
600 LLVMValueRef aos_channels
[TGSI_NUM_CHANNELS
];
601 unsigned pixels_per_channel
= soa_type
.length
/ TGSI_NUM_CHANNELS
;
603 for (j
= 0; j
< TGSI_NUM_CHANNELS
; ++j
) {
604 LLVMValueRef channel
[LP_MAX_VECTOR_LENGTH
] = { 0 };
606 assert(pixels_per_channel
<= LP_MAX_VECTOR_LENGTH
);
608 for (k
= 0; k
< pixels_per_channel
; ++k
) {
609 channel
[k
] = src_aos
[i
][j
+ TGSI_NUM_CHANNELS
* k
];
612 aos_channels
[j
] = lp_build_concat(gallivm
, channel
, aos_channel_type
, pixels_per_channel
);
615 lp_build_transpose_aos(gallivm
, soa_type
, aos_channels
, dst_soa
[i
]);
621 store_aos(struct gallivm_state
*gallivm
,
626 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
627 LLVMBuilderRef builder
= gallivm
->builder
;
628 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
629 LLVMValueRef indices
[3];
631 indices
[0] = lp_build_const_int32(gallivm
, 0);
633 indices
[2] = lp_build_const_int32(gallivm
, 0);
636 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
639 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
640 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
642 /* Unaligned store due to the vertex header */
643 lp_set_store_alignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
647 * Adjust the mask to architecture endianess. The mask will the store in struct:
649 * struct vertex_header {
650 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
651 * unsigned edgeflag:1;
652 * unsigned have_clipdist:1;
653 * unsigned vertex_id:16;
657 * On little-endian machine nothing needs to done, however on bit-endian machine
658 * the mask's fields need to be adjusted with the algorithm:
660 * uint32_t reverse (uint32_t x)
662 * return (x >> 16) | // vertex_id
663 * ((x & 0x3fff) << 18) | // clipmask
664 * ((x & 0x4000) << 3) | // have_clipdist
665 * ((x & 0x8000) << 1); // edgeflag
669 adjust_mask(struct gallivm_state
*gallivm
,
672 #ifdef PIPE_ARCH_BIG_ENDIAN
673 LLVMBuilderRef builder
= gallivm
->builder
;
674 LLVMValueRef vertex_id
;
675 LLVMValueRef clipmask
;
676 LLVMValueRef have_clipdist
;
677 LLVMValueRef edgeflag
;
679 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
680 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
681 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
682 have_clipdist
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
683 have_clipdist
= LLVMBuildShl(builder
, have_clipdist
, lp_build_const_int32(gallivm
, 3), "");
684 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
685 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 1), "");
687 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
688 mask
= LLVMBuildOr(builder
, mask
, have_clipdist
, "");
689 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
695 store_aos_array(struct gallivm_state
*gallivm
,
696 struct lp_type soa_type
,
701 LLVMValueRef clipmask
,
702 boolean have_clipdist
)
704 LLVMBuilderRef builder
= gallivm
->builder
;
705 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
706 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
707 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
708 int vector_length
= soa_type
.length
;
711 debug_assert(TGSI_NUM_CHANNELS
== 4);
713 for (i
= 0; i
< vector_length
; i
++) {
714 inds
[i
] = lp_build_const_int32(gallivm
, i
);
715 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
719 /* store vertex header for each of the n vertices */
720 LLVMValueRef val
, cliptmp
;
721 int vertex_id_pad_edgeflag
;
723 /* If this assertion fails, it means we need to update the bit twidding
724 * code here. See struct vertex_header in draw_private.h.
726 assert(DRAW_TOTAL_CLIP_PLANES
==14);
727 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
728 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
730 vertex_id_pad_edgeflag
|= 1 << (DRAW_TOTAL_CLIP_PLANES
+1);
731 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
), vertex_id_pad_edgeflag
);
732 /* OR with the clipmask */
733 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
734 for (i
= 0; i
< vector_length
; i
++) {
735 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
736 val
= LLVMBuildExtractElement(builder
, cliptmp
, inds
[i
], "");
737 val
= adjust_mask(gallivm
, val
);
738 LLVMBuildStore(builder
, val
, id_ptr
);
740 lp_build_printf(gallivm
, "io = %p, index %d\n, clipmask = %x\n",
741 io_ptrs
[i
], inds
[i
], val
);
746 /* store for each of the n vertices */
747 for (i
= 0; i
< vector_length
; i
++) {
748 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
754 convert_to_aos(struct gallivm_state
*gallivm
,
756 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
757 LLVMValueRef clipmask
,
759 struct lp_type soa_type
,
760 boolean have_clipdist
)
762 LLVMBuilderRef builder
= gallivm
->builder
;
763 unsigned chan
, attrib
, i
;
766 lp_build_printf(gallivm
, " # storing begin\n");
768 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
769 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
770 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
771 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
772 if (outputs
[attrib
][chan
]) {
773 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
774 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
776 lp_build_printf(gallivm
, "output %d : %d ",
777 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
779 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
781 lp_build_print_value(gallivm
, "val = ", out
);
791 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
792 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
794 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
796 for (i
= 0; i
< soa_type
.length
; ++i
) {
797 aos
[i
] = lp_build_extract_range(gallivm
,
798 soa
[i
% TGSI_NUM_CHANNELS
],
799 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
804 store_aos_array(gallivm
,
810 clipmask
, have_clipdist
);
813 lp_build_printf(gallivm
, " # storing end\n");
819 * Stores original vertex positions in clip coordinates
822 store_clip(struct gallivm_state
*gallivm
,
823 const struct lp_type vs_type
,
825 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
826 boolean pre_clip_pos
, int idx
)
828 LLVMBuilderRef builder
= gallivm
->builder
;
830 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
831 LLVMValueRef indices
[2];
832 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
833 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
834 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
838 indices
[1] = lp_build_const_int32(gallivm
, 0);
840 for (i
= 0; i
< vs_type
.length
; i
++) {
841 inds
[i
] = lp_build_const_int32(gallivm
, i
);
842 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
845 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
846 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
847 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
848 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
851 for (i
= 0; i
< vs_type
.length
; i
++) {
852 clip_ptrs
[i
] = draw_jit_header_clip(gallivm
, io_ptrs
[i
]);
855 for (i
= 0; i
< vs_type
.length
; i
++) {
856 clip_ptrs
[i
] = draw_jit_header_pre_clip_pos(gallivm
, io_ptrs
[i
]);
860 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
861 for (i
= 0; i
< vs_type
.length
; ++i
) {
862 aos
[i
] = lp_build_extract_range(gallivm
,
863 soa
[i
% TGSI_NUM_CHANNELS
],
864 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
868 for (j
= 0; j
< vs_type
.length
; j
++) {
869 LLVMTypeRef clip_ptr_type
= LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4), 0);
870 LLVMValueRef clip_ptr
;
872 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
873 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
875 /* Unaligned store */
876 lp_set_store_alignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
882 * Transforms the outputs for viewport mapping
885 generate_viewport(struct draw_llvm_variant
*variant
,
886 LLVMBuilderRef builder
,
887 struct lp_type vs_type
,
888 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
889 LLVMValueRef context_ptr
)
892 struct gallivm_state
*gallivm
= variant
->gallivm
;
893 struct lp_type f32_type
= vs_type
;
894 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
895 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[0][3], ""); /*w0 w1 .. wn*/
896 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
897 LLVMValueRef vp_ptr
= draw_jit_context_viewport(gallivm
, context_ptr
);
899 /* for 1/w convention*/
900 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
901 LLVMBuildStore(builder
, out3
, outputs
[0][3]);
903 /* Viewport Mapping */
904 for (i
=0; i
<3; i
++) {
905 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[0][i
], ""); /*x0 x1 .. xn*/
908 LLVMValueRef scale_i
;
909 LLVMValueRef trans_i
;
912 index
= lp_build_const_int32(gallivm
, i
);
913 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
915 index
= lp_build_const_int32(gallivm
, i
+4);
916 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
918 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
919 LLVMBuildLoad(builder
, scale_i
, "scale"));
920 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
921 LLVMBuildLoad(builder
, trans_i
, "trans"));
924 out
= LLVMBuildFMul(builder
, out
, out3
, "");
926 out
= LLVMBuildFMul(builder
, out
, scale
, "");
927 /* add translation */
928 out
= LLVMBuildFAdd(builder
, out
, trans
, "");
930 /* store transformed outputs */
931 LLVMBuildStore(builder
, out
, outputs
[0][i
]);
938 * Returns clipmask as nxi32 bitmask for the n vertices
941 generate_clipmask(struct draw_llvm
*llvm
,
942 struct gallivm_state
*gallivm
,
943 struct lp_type vs_type
,
944 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
950 LLVMValueRef context_ptr
,
951 boolean
*have_clipdist
)
953 LLVMBuilderRef builder
= gallivm
->builder
;
954 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
955 LLVMValueRef test
, temp
;
956 LLVMValueRef zero
, shift
;
957 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
958 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
959 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
960 struct lp_type f32_type
= vs_type
;
961 struct lp_type i32_type
= lp_int_type(vs_type
);
962 const unsigned pos
= draw_current_shader_position_output(llvm
->draw
);
963 const unsigned cv
= draw_current_shader_clipvertex_output(llvm
->draw
);
964 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
965 bool have_cd
= false;
968 cd
[0] = draw_current_shader_clipdistance_output(llvm
->draw
, 0);
969 cd
[1] = draw_current_shader_clipdistance_output(llvm
->draw
, 1);
971 if (cd
[0] != pos
|| cd
[1] != pos
)
974 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
975 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
976 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
977 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
980 * load clipvertex and position from correct locations.
981 * if they are the same just load them once.
983 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
984 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
985 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
986 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
988 if (clip_user
&& cv
!= pos
) {
989 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
990 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
991 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
992 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1000 /* Cliptest, for hardwired planes */
1003 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1005 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1009 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1010 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1011 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1012 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1013 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1016 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1017 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1018 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1019 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1022 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1023 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1024 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1025 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1026 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1030 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1033 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1034 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1035 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1039 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1040 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1041 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1042 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1045 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1046 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1047 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1048 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1052 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1053 LLVMValueRef indices
[3];
1055 /* userclip planes */
1056 while (ucp_enable
) {
1057 unsigned plane_idx
= ffs(ucp_enable
)-1;
1058 ucp_enable
&= ~(1 << plane_idx
);
1061 if (have_cd
&& num_written_clipdistance
) {
1062 LLVMValueRef clipdist
;
1066 *have_clipdist
= TRUE
;
1068 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1070 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1072 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1073 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1 << plane_idx
);
1074 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1075 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1077 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1078 indices
[0] = lp_build_const_int32(gallivm
, 0);
1079 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1081 indices
[2] = lp_build_const_int32(gallivm
, 0);
1082 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1083 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1084 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1085 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1087 indices
[2] = lp_build_const_int32(gallivm
, 1);
1088 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1089 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1090 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1091 test
= LLVMBuildFMul(builder
, planes
, cv_y
, "");
1092 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1094 indices
[2] = lp_build_const_int32(gallivm
, 2);
1095 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1096 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1097 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1098 test
= LLVMBuildFMul(builder
, planes
, cv_z
, "");
1099 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1101 indices
[2] = lp_build_const_int32(gallivm
, 3);
1102 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1103 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1104 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1105 test
= LLVMBuildFMul(builder
, planes
, cv_w
, "");
1106 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1108 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1109 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1 << plane_idx
);
1110 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1111 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1120 * Returns boolean if any clipping has occurred
1121 * Used zero/non-zero i32 value to represent boolean
1124 clipmask_booli32(struct gallivm_state
*gallivm
,
1125 const struct lp_type vs_type
,
1126 LLVMValueRef clipmask_bool_ptr
)
1128 LLVMBuilderRef builder
= gallivm
->builder
;
1129 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1130 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1131 LLVMValueRef ret
= LLVMConstNull(int32_type
);
1136 * Can do this with log2(vector length) pack instructions and one extract
1137 * (as we don't actually need a or) with sse2 which would be way better.
1139 for (i
=0; i
< vs_type
.length
; i
++) {
1140 temp
= LLVMBuildExtractElement(builder
, clipmask_bool
,
1141 lp_build_const_int32(gallivm
, i
) , "");
1142 ret
= LLVMBuildOr(builder
, ret
, temp
, "");
1149 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
,
1152 struct gallivm_state
*gallivm
= variant
->gallivm
;
1153 LLVMContextRef context
= gallivm
->context
;
1154 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1155 LLVMTypeRef arg_types
[8];
1156 LLVMTypeRef func_type
;
1157 LLVMValueRef context_ptr
;
1158 LLVMBasicBlockRef block
;
1159 LLVMBuilderRef builder
;
1160 struct lp_type vs_type
;
1161 LLVMValueRef end
, start
;
1162 LLVMValueRef count
, fetch_elts
, fetch_count
;
1163 LLVMValueRef stride
, step
, io_itr
;
1164 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1165 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
1166 LLVMValueRef one
= lp_build_const_int32(gallivm
, 1);
1167 struct draw_context
*draw
= llvm
->draw
;
1168 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1170 struct lp_build_context bld
;
1171 struct lp_build_loop_state lp_loop
;
1172 const int vector_length
= lp_native_vector_width
/ 32;
1173 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1174 LLVMValueRef fetch_max
;
1175 struct lp_build_sampler_soa
*sampler
= 0;
1176 LLVMValueRef ret
, clipmask_bool_ptr
;
1177 const struct draw_geometry_shader
*gs
= draw
->gs
.geometry_shader
;
1178 /* If geometry shader is present we need to skip both the viewport
1179 * transformation and clipping otherwise the inputs to the geometry
1180 * shader will be incorrect.
1182 const boolean bypass_viewport
= gs
|| variant
->key
.bypass_viewport
;
1183 const boolean enable_cliptest
= !gs
&& (variant
->key
.clip_xy
||
1184 variant
->key
.clip_z
||
1185 variant
->key
.clip_user
);
1186 LLVMValueRef variant_func
;
1187 const unsigned pos
= draw_current_shader_position_output(llvm
->draw
);
1188 const unsigned cv
= draw_current_shader_clipvertex_output(llvm
->draw
);
1189 boolean have_clipdist
= FALSE
;
1190 struct lp_bld_tgsi_system_values system_values
;
1192 memset(&system_values
, 0, sizeof(system_values
));
1194 arg_types
[0] = get_context_ptr_type(variant
); /* context */
1195 arg_types
[1] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1196 arg_types
[2] = get_buffer_ptr_type(variant
); /* vbuffers */
1198 arg_types
[3] = LLVMPointerType(int32_type
, 0);/* fetch_elts * */
1200 arg_types
[3] = int32_type
; /* start */
1201 arg_types
[4] = int32_type
; /* fetch_count / count */
1202 arg_types
[5] = int32_type
; /* stride */
1203 arg_types
[6] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1204 arg_types
[7] = int32_type
; /* instance_id */
1206 func_type
= LLVMFunctionType(int32_type
, arg_types
, Elements(arg_types
), 0);
1208 variant_func
= LLVMAddFunction(gallivm
->module
,
1209 elts
? "draw_llvm_shader_elts" : "draw_llvm_shader",
1213 variant
->function_elts
= variant_func
;
1215 variant
->function
= variant_func
;
1217 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1218 for (i
= 0; i
< Elements(arg_types
); ++i
)
1219 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1220 LLVMAddAttribute(LLVMGetParam(variant_func
, i
),
1221 LLVMNoAliasAttribute
);
1223 context_ptr
= LLVMGetParam(variant_func
, 0);
1224 io_ptr
= LLVMGetParam(variant_func
, 1);
1225 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1226 stride
= LLVMGetParam(variant_func
, 5);
1227 vb_ptr
= LLVMGetParam(variant_func
, 6);
1228 system_values
.instance_id
= LLVMGetParam(variant_func
, 7);
1230 lp_build_name(context_ptr
, "context");
1231 lp_build_name(io_ptr
, "io");
1232 lp_build_name(vbuffers_ptr
, "vbuffers");
1233 lp_build_name(stride
, "stride");
1234 lp_build_name(vb_ptr
, "vb");
1235 lp_build_name(system_values
.instance_id
, "instance_id");
1238 fetch_elts
= LLVMGetParam(variant_func
, 3);
1239 fetch_count
= LLVMGetParam(variant_func
, 4);
1240 lp_build_name(fetch_elts
, "fetch_elts");
1241 lp_build_name(fetch_count
, "fetch_count");
1242 start
= count
= NULL
;
1245 start
= LLVMGetParam(variant_func
, 3);
1246 count
= LLVMGetParam(variant_func
, 4);
1247 lp_build_name(start
, "start");
1248 lp_build_name(count
, "count");
1249 fetch_elts
= fetch_count
= NULL
;
1256 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
1257 builder
= gallivm
->builder
;
1258 LLVMPositionBuilderAtEnd(builder
, block
);
1260 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
1262 memset(&vs_type
, 0, sizeof vs_type
);
1263 vs_type
.floating
= TRUE
; /* floating point values */
1264 vs_type
.sign
= TRUE
; /* values are signed */
1265 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1266 vs_type
.width
= 32; /* 32-bit float */
1267 vs_type
.length
= vector_length
;
1269 /* hold temporary "bool" clipmask */
1270 clipmask_bool_ptr
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, vs_type
), "");
1271 LLVMBuildStore(builder
, lp_build_zero(gallivm
, lp_int_type(vs_type
)), clipmask_bool_ptr
);
1273 /* code generated texture sampling */
1274 sampler
= draw_llvm_sampler_soa_create(
1275 draw_llvm_variant_key_samplers(&variant
->key
),
1283 end
= lp_build_add(&bld
, start
, count
);
1286 step
= lp_build_const_int32(gallivm
, vector_length
);
1288 fetch_max
= LLVMBuildSub(builder
, end
, one
, "fetch_max");
1290 lp_build_loop_begin(&lp_loop
, gallivm
, start
);
1292 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
1293 LLVMValueRef aos_attribs
[PIPE_MAX_SHADER_INPUTS
][LP_MAX_VECTOR_WIDTH
/ 32] = { { 0 } };
1295 LLVMValueRef clipmask
; /* holds the clipmask value */
1296 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
1299 io_itr
= lp_loop
.counter
;
1301 io_itr
= LLVMBuildSub(builder
, lp_loop
.counter
, start
, "");
1303 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
1305 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
1306 io_itr
, io
, lp_loop
.counter
);
1308 system_values
.vertex_id
= lp_build_zero(gallivm
, lp_type_uint_vec(32, 32*vector_length
));
1309 for (i
= 0; i
< vector_length
; ++i
) {
1310 LLVMValueRef true_index
=
1311 LLVMBuildAdd(builder
,
1313 lp_build_const_int32(gallivm
, i
), "");
1315 /* make sure we're not out of bounds which can happen
1316 * if fetch_count % 4 != 0, because on the last iteration
1317 * a few of the 4 vertex fetches will be out of bounds */
1318 true_index
= lp_build_min(&bld
, true_index
, fetch_max
);
1321 LLVMValueRef fetch_ptr
;
1322 fetch_ptr
= LLVMBuildGEP(builder
, fetch_elts
,
1323 &true_index
, 1, "");
1324 true_index
= LLVMBuildLoad(builder
, fetch_ptr
, "fetch_elt");
1327 system_values
.vertex_id
= LLVMBuildInsertElement(gallivm
->builder
,
1328 system_values
.vertex_id
, true_index
,
1329 lp_build_const_int32(gallivm
, i
), "");
1330 for (j
= 0; j
< draw
->pt
.nr_vertex_elements
; ++j
) {
1331 struct pipe_vertex_element
*velem
= &draw
->pt
.vertex_element
[j
];
1332 LLVMValueRef vb_index
=
1333 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
1334 LLVMValueRef vb
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
1335 generate_fetch(gallivm
, vbuffers_ptr
,
1336 &aos_attribs
[j
][i
], velem
, vb
, true_index
,
1337 system_values
.instance_id
);
1340 convert_to_soa(gallivm
, aos_attribs
, inputs
,
1341 draw
->pt
.nr_vertex_elements
, vs_type
);
1343 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
1344 generate_vs(variant
,
1352 variant
->key
.clamp_vertex_color
);
1354 if (pos
!= -1 && cv
!= -1) {
1355 /* store original positions in clip before further manipulation */
1356 store_clip(gallivm
, vs_type
, io
, outputs
, 0, cv
);
1357 store_clip(gallivm
, vs_type
, io
, outputs
, 1, pos
);
1360 if (enable_cliptest
) {
1361 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1362 /* allocate clipmask, assign it integer type */
1363 clipmask
= generate_clipmask(llvm
,
1367 variant
->key
.clip_xy
,
1368 variant
->key
.clip_z
,
1369 variant
->key
.clip_user
,
1370 variant
->key
.clip_halfz
,
1371 variant
->key
.ucp_enable
,
1372 context_ptr
, &have_clipdist
);
1373 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
1374 /* store temporary clipping boolean value */
1375 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
1378 clipmask
= lp_build_const_int_vec(gallivm
, lp_int_type(vs_type
), 0);
1381 /* do viewport mapping */
1382 if (!bypass_viewport
) {
1383 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
1387 clipmask
= lp_build_const_int_vec(gallivm
, lp_int_type(vs_type
), 0);
1390 /* store clipmask in vertex header,
1391 * original positions in clip
1392 * and transformed positions in data
1394 convert_to_aos(gallivm
, io
, outputs
, clipmask
,
1395 vs_info
->num_outputs
, vs_type
,
1399 lp_build_loop_end_cond(&lp_loop
, end
, step
, LLVMIntUGE
);
1401 sampler
->destroy(sampler
);
1403 /* return clipping boolean value for function */
1404 ret
= clipmask_booli32(gallivm
, vs_type
, clipmask_bool_ptr
);
1406 LLVMBuildRet(builder
, ret
);
1408 gallivm_verify_function(gallivm
, variant_func
);
1412 struct draw_llvm_variant_key
*
1413 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
1416 struct draw_llvm_variant_key
*key
;
1417 struct draw_sampler_static_state
*draw_sampler
;
1419 key
= (struct draw_llvm_variant_key
*)store
;
1421 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
1423 /* Presumably all variants of the shader should have the same
1424 * number of vertex elements - ie the number of shader inputs.
1425 * NOTE: we NEED to store the needed number of needed inputs
1426 * here, not the number of provided elements to match keysize
1427 * (and the offset of sampler state in the key).
1429 key
->nr_vertex_elements
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
1430 assert(key
->nr_vertex_elements
<= llvm
->draw
->pt
.nr_vertex_elements
);
1432 /* will have to rig this up properly later */
1433 key
->clip_xy
= llvm
->draw
->clip_xy
;
1434 key
->clip_z
= llvm
->draw
->clip_z
;
1435 key
->clip_user
= llvm
->draw
->clip_user
;
1436 key
->bypass_viewport
= llvm
->draw
->identity_viewport
;
1437 key
->clip_halfz
= !llvm
->draw
->rasterizer
->gl_rasterization_rules
;
1438 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
1439 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
1443 /* All variants of this shader will have the same value for
1444 * nr_samplers. Not yet trying to compact away holes in the
1447 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1448 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
1449 key
->nr_sampler_views
=
1450 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
1453 key
->nr_sampler_views
= key
->nr_samplers
;
1456 draw_sampler
= draw_llvm_variant_key_samplers(key
);
1458 memcpy(key
->vertex_element
,
1459 llvm
->draw
->pt
.vertex_element
,
1460 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
1462 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
1464 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
1465 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
1466 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
1468 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
1469 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
1470 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
1478 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
1481 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
1483 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
1484 debug_printf("clip_xy = %u\n", key
->clip_xy
);
1485 debug_printf("clip_z = %u\n", key
->clip_z
);
1486 debug_printf("clip_user = %u\n", key
->clip_user
);
1487 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
1488 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
1489 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
1490 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
1492 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
1493 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
1494 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
1495 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
1496 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
1499 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
1500 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
1506 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
1508 uint32_t width
, uint32_t height
, uint32_t depth
,
1509 uint32_t first_level
, uint32_t last_level
,
1510 const void *base_ptr
,
1511 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
1512 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
1513 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
1516 struct draw_jit_texture
*jit_tex
;
1518 assert(sview_idx
< Elements(draw
->llvm
->jit_context
.textures
));
1520 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
1522 jit_tex
->width
= width
;
1523 jit_tex
->height
= height
;
1524 jit_tex
->depth
= depth
;
1525 jit_tex
->first_level
= first_level
;
1526 jit_tex
->last_level
= last_level
;
1527 jit_tex
->base
= base_ptr
;
1529 for (j
= first_level
; j
<= last_level
; j
++) {
1530 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
1531 jit_tex
->row_stride
[j
] = row_stride
[j
];
1532 jit_tex
->img_stride
[j
] = img_stride
[j
];
1538 draw_llvm_set_sampler_state(struct draw_context
*draw
)
1542 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
1543 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
1545 if (draw
->samplers
[i
]) {
1546 const struct pipe_sampler_state
*s
1547 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
1548 jit_sam
->min_lod
= s
->min_lod
;
1549 jit_sam
->max_lod
= s
->max_lod
;
1550 jit_sam
->lod_bias
= s
->lod_bias
;
1551 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
1558 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
1560 struct draw_llvm
*llvm
= variant
->llvm
;
1562 if (variant
->function_elts
) {
1563 gallivm_free_function(variant
->gallivm
,
1564 variant
->function_elts
, variant
->jit_func_elts
);
1567 if (variant
->function
) {
1568 gallivm_free_function(variant
->gallivm
,
1569 variant
->function
, variant
->jit_func
);
1572 gallivm_destroy(variant
->gallivm
);
1574 remove_from_list(&variant
->list_item_local
);
1575 variant
->shader
->variants_cached
--;
1576 remove_from_list(&variant
->list_item_global
);
1577 llvm
->nr_variants
--;