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
[6];
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] = elem_types
[0]; /* gs_constants */
198 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
199 DRAW_TOTAL_CLIP_PLANES
), 0);
200 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewport */
201 elem_types
[4] = LLVMArrayType(texture_type
,
202 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
203 elem_types
[5] = LLVMArrayType(sampler_type
,
204 PIPE_MAX_SAMPLERS
); /* samplers */
205 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
206 Elements(elem_types
), 0);
207 #if HAVE_LLVM < 0x0300
208 LLVMAddTypeName(gallivm
->module
, struct_name
, context_type
);
210 LLVMInvalidateStructLayout(gallivm
->target
, context_type
);
213 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_constants
,
214 target
, context_type
, 0);
215 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, gs_constants
,
216 target
, context_type
, 1);
217 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, planes
,
218 target
, context_type
, 2);
219 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, viewport
,
220 target
, context_type
, 3);
221 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, textures
,
222 target
, context_type
,
223 DRAW_JIT_CTX_TEXTURES
);
224 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, samplers
,
225 target
, context_type
,
226 DRAW_JIT_CTX_SAMPLERS
);
227 LP_CHECK_STRUCT_SIZE(struct draw_jit_context
,
228 target
, context_type
);
235 * Create LLVM type for struct pipe_vertex_buffer
238 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
, const char *struct_name
)
240 LLVMTargetDataRef target
= gallivm
->target
;
241 LLVMTypeRef elem_types
[4];
245 elem_types
[1] = LLVMInt32TypeInContext(gallivm
->context
);
247 elem_types
[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0); /* vs_constants */
249 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
250 Elements(elem_types
), 0);
251 #if HAVE_LLVM < 0x0300
252 LLVMAddTypeName(gallivm
->module
, struct_name
, vb_type
);
254 LLVMInvalidateStructLayout(gallivm
->target
, vb_type
);
257 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
259 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
262 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
269 * Create LLVM type for struct vertex_header;
272 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
274 LLVMTargetDataRef target
= gallivm
->target
;
275 LLVMTypeRef elem_types
[4];
276 LLVMTypeRef vertex_header
;
277 char struct_name
[24];
279 util_snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
281 elem_types
[DRAW_JIT_VERTEX_VERTEX_ID
] = LLVMIntTypeInContext(gallivm
->context
, 32);
282 elem_types
[DRAW_JIT_VERTEX_CLIP
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
283 elem_types
[DRAW_JIT_VERTEX_PRE_CLIP_POS
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
284 elem_types
[DRAW_JIT_VERTEX_DATA
] = LLVMArrayType(elem_types
[1], data_elems
);
286 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
287 Elements(elem_types
), 0);
288 #if HAVE_LLVM < 0x0300
289 LLVMAddTypeName(gallivm
->module
, struct_name
, vertex_header
);
291 LLVMInvalidateStructLayout(gallivm
->target
, vertex_header
);
294 /* these are bit-fields and we can't take address of them
295 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
296 target, vertex_header,
297 DRAW_JIT_VERTEX_CLIPMASK);
298 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
299 target, vertex_header,
300 DRAW_JIT_VERTEX_EDGEFLAG);
301 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
302 target, vertex_header,
303 DRAW_JIT_VERTEX_PAD);
304 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
305 target, vertex_header,
306 DRAW_JIT_VERTEX_VERTEX_ID);
308 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, clip
,
309 target
, vertex_header
,
310 DRAW_JIT_VERTEX_CLIP
);
311 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, pre_clip_pos
,
312 target
, vertex_header
,
313 DRAW_JIT_VERTEX_PRE_CLIP_POS
);
314 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, data
,
315 target
, vertex_header
,
316 DRAW_JIT_VERTEX_DATA
);
318 assert(LLVMABISizeOfType(target
, vertex_header
) ==
319 offsetof(struct vertex_header
, data
[data_elems
]));
321 return vertex_header
;
326 * Create LLVM types for various structures.
329 create_jit_types(struct draw_llvm_variant
*variant
)
331 struct gallivm_state
*gallivm
= variant
->gallivm
;
332 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
, vb_type
;
334 texture_type
= create_jit_texture_type(gallivm
, "texture");
335 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
337 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
339 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
341 buffer_type
= LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
342 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
344 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
345 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
350 get_context_ptr_type(struct draw_llvm_variant
*variant
)
352 if (!variant
->context_ptr_type
)
353 create_jit_types(variant
);
354 return variant
->context_ptr_type
;
359 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
361 if (!variant
->buffer_ptr_type
)
362 create_jit_types(variant
);
363 return variant
->buffer_ptr_type
;
368 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
370 if (!variant
->vb_ptr_type
)
371 create_jit_types(variant
);
372 return variant
->vb_ptr_type
;
376 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
378 if (!variant
->vertex_header_ptr_type
)
379 create_jit_types(variant
);
380 return variant
->vertex_header_ptr_type
;
385 * Create per-context LLVM info.
388 draw_llvm_create(struct draw_context
*draw
)
390 struct draw_llvm
*llvm
;
392 llvm
= CALLOC_STRUCT( draw_llvm
);
400 llvm
->nr_variants
= 0;
401 make_empty_list(&llvm
->vs_variants_list
);
408 * Free per-context LLVM info.
411 draw_llvm_destroy(struct draw_llvm
*llvm
)
413 /* XXX free other draw_llvm data? */
419 * Create LLVM-generated code for a vertex shader.
421 struct draw_llvm_variant
*
422 draw_llvm_create_variant(struct draw_llvm
*llvm
,
424 const struct draw_llvm_variant_key
*key
)
426 struct draw_llvm_variant
*variant
;
427 struct llvm_vertex_shader
*shader
=
428 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
429 LLVMTypeRef vertex_header
;
431 variant
= MALLOC(sizeof *variant
+
432 shader
->variant_key_size
-
433 sizeof variant
->key
);
437 variant
->llvm
= llvm
;
439 variant
->gallivm
= gallivm_create();
441 create_jit_types(variant
);
443 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
445 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
447 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
449 draw_llvm_generate(llvm
, variant
, FALSE
); /* linear */
450 draw_llvm_generate(llvm
, variant
, TRUE
); /* elts */
452 gallivm_compile_module(variant
->gallivm
);
454 variant
->jit_func
= (draw_jit_vert_func
)
455 gallivm_jit_function(variant
->gallivm
, variant
->function
);
457 variant
->jit_func_elts
= (draw_jit_vert_func_elts
)
458 gallivm_jit_function(variant
->gallivm
, variant
->function_elts
);
460 variant
->shader
= shader
;
461 variant
->list_item_global
.base
= variant
;
462 variant
->list_item_local
.base
= variant
;
463 /*variant->no = */shader
->variants_created
++;
464 variant
->list_item_global
.base
= variant
;
471 generate_vs(struct draw_llvm_variant
*variant
,
472 LLVMBuilderRef builder
,
473 struct lp_type vs_type
,
474 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
475 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
476 const struct lp_bld_tgsi_system_values
*system_values
,
477 LLVMValueRef context_ptr
,
478 struct lp_build_sampler_soa
*draw_sampler
,
479 boolean clamp_vertex_color
)
481 struct draw_llvm
*llvm
= variant
->llvm
;
482 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
483 LLVMValueRef consts_ptr
= draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
484 struct lp_build_sampler_soa
*sampler
= 0;
486 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
487 tgsi_dump(tokens
, 0);
488 draw_llvm_dump_variant_key(&variant
->key
);
491 if (llvm
->draw
->num_sampler_views
&& llvm
->draw
->num_samplers
)
492 sampler
= draw_sampler
;
494 lp_build_tgsi_soa(variant
->gallivm
,
497 NULL
/*struct lp_build_mask_context *mask*/,
504 &llvm
->draw
->vs
.vertex_shader
->info
);
508 unsigned chan
, attrib
;
509 struct lp_build_context bld
;
510 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
511 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
513 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
514 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
515 if (outputs
[attrib
][chan
]) {
516 switch (info
->output_semantic_name
[attrib
]) {
517 case TGSI_SEMANTIC_COLOR
:
518 case TGSI_SEMANTIC_BCOLOR
:
519 if (clamp_vertex_color
) {
520 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
521 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
522 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
525 case TGSI_SEMANTIC_FOG
:
526 if (chan
== 1 || chan
== 2)
527 LLVMBuildStore(builder
, bld
.zero
, outputs
[attrib
][chan
]);
529 LLVMBuildStore(builder
, bld
.one
, outputs
[attrib
][chan
]);
540 generate_fetch(struct gallivm_state
*gallivm
,
541 LLVMValueRef vbuffers_ptr
,
543 struct pipe_vertex_element
*velem
,
546 LLVMValueRef instance_id
)
548 const struct util_format_description
*format_desc
= util_format_description(velem
->src_format
);
549 LLVMValueRef zero
= LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
));
550 LLVMBuilderRef builder
= gallivm
->builder
;
551 LLVMValueRef indices
=
552 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
),
553 velem
->vertex_buffer_index
, 0);
554 LLVMValueRef vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
,
556 LLVMValueRef vb_stride
= draw_jit_vbuffer_stride(gallivm
, vbuf
);
557 LLVMValueRef vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vbuf
);
560 if (velem
->instance_divisor
) {
561 /* array index = instance_id / instance_divisor */
562 index
= LLVMBuildUDiv(builder
, instance_id
,
563 lp_build_const_int32(gallivm
, velem
->instance_divisor
),
567 stride
= LLVMBuildMul(builder
, vb_stride
, index
, "");
569 vbuffer_ptr
= LLVMBuildLoad(builder
, vbuffer_ptr
, "vbuffer");
571 stride
= LLVMBuildAdd(builder
, stride
,
574 stride
= LLVMBuildAdd(builder
, stride
,
575 lp_build_const_int32(gallivm
, velem
->src_offset
),
578 /* lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);*/
579 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffer_ptr
, &stride
, 1, "");
581 *res
= lp_build_fetch_rgba_aos(gallivm
,
583 lp_float32_vec4_type(),
589 convert_to_soa(struct gallivm_state
*gallivm
,
590 LLVMValueRef (*src_aos
)[LP_MAX_VECTOR_WIDTH
/ 32],
591 LLVMValueRef (*dst_soa
)[TGSI_NUM_CHANNELS
],
592 unsigned num_attribs
, const struct lp_type soa_type
)
595 struct lp_type aos_channel_type
= soa_type
;
597 debug_assert(TGSI_NUM_CHANNELS
== 4);
598 debug_assert((soa_type
.length
% TGSI_NUM_CHANNELS
) == 0);
600 aos_channel_type
.length
>>= 1;
602 for (i
= 0; i
< num_attribs
; ++i
) {
603 LLVMValueRef aos_channels
[TGSI_NUM_CHANNELS
];
604 unsigned pixels_per_channel
= soa_type
.length
/ TGSI_NUM_CHANNELS
;
606 for (j
= 0; j
< TGSI_NUM_CHANNELS
; ++j
) {
607 LLVMValueRef channel
[LP_MAX_VECTOR_LENGTH
] = { 0 };
609 assert(pixels_per_channel
<= LP_MAX_VECTOR_LENGTH
);
611 for (k
= 0; k
< pixels_per_channel
; ++k
) {
612 channel
[k
] = src_aos
[i
][j
+ TGSI_NUM_CHANNELS
* k
];
615 aos_channels
[j
] = lp_build_concat(gallivm
, channel
, aos_channel_type
, pixels_per_channel
);
618 lp_build_transpose_aos(gallivm
, soa_type
, aos_channels
, dst_soa
[i
]);
624 store_aos(struct gallivm_state
*gallivm
,
629 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
630 LLVMBuilderRef builder
= gallivm
->builder
;
631 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
632 LLVMValueRef indices
[3];
634 indices
[0] = lp_build_const_int32(gallivm
, 0);
636 indices
[2] = lp_build_const_int32(gallivm
, 0);
639 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
642 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
643 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
645 /* Unaligned store due to the vertex header */
646 lp_set_store_alignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
650 * Adjust the mask to architecture endianess. The mask will the store in struct:
652 * struct vertex_header {
653 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
654 * unsigned edgeflag:1;
655 * unsigned have_clipdist:1;
656 * unsigned vertex_id:16;
660 * On little-endian machine nothing needs to done, however on bit-endian machine
661 * the mask's fields need to be adjusted with the algorithm:
663 * uint32_t reverse (uint32_t x)
665 * return (x >> 16) | // vertex_id
666 * ((x & 0x3fff) << 18) | // clipmask
667 * ((x & 0x4000) << 3) | // have_clipdist
668 * ((x & 0x8000) << 1); // edgeflag
672 adjust_mask(struct gallivm_state
*gallivm
,
675 #ifdef PIPE_ARCH_BIG_ENDIAN
676 LLVMBuilderRef builder
= gallivm
->builder
;
677 LLVMValueRef vertex_id
;
678 LLVMValueRef clipmask
;
679 LLVMValueRef have_clipdist
;
680 LLVMValueRef edgeflag
;
682 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
683 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
684 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
685 have_clipdist
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
686 have_clipdist
= LLVMBuildShl(builder
, have_clipdist
, lp_build_const_int32(gallivm
, 3), "");
687 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
688 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 1), "");
690 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
691 mask
= LLVMBuildOr(builder
, mask
, have_clipdist
, "");
692 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
698 store_aos_array(struct gallivm_state
*gallivm
,
699 struct lp_type soa_type
,
704 LLVMValueRef clipmask
,
705 boolean have_clipdist
)
707 LLVMBuilderRef builder
= gallivm
->builder
;
708 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
709 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
710 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
711 int vector_length
= soa_type
.length
;
714 debug_assert(TGSI_NUM_CHANNELS
== 4);
716 for (i
= 0; i
< vector_length
; i
++) {
717 inds
[i
] = lp_build_const_int32(gallivm
, i
);
718 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
722 /* store vertex header for each of the n vertices */
723 LLVMValueRef val
, cliptmp
;
724 int vertex_id_pad_edgeflag
;
726 /* If this assertion fails, it means we need to update the bit twidding
727 * code here. See struct vertex_header in draw_private.h.
729 assert(DRAW_TOTAL_CLIP_PLANES
==14);
730 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
731 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
733 vertex_id_pad_edgeflag
|= 1 << (DRAW_TOTAL_CLIP_PLANES
+1);
734 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
), vertex_id_pad_edgeflag
);
735 /* OR with the clipmask */
736 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
737 for (i
= 0; i
< vector_length
; i
++) {
738 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
739 val
= LLVMBuildExtractElement(builder
, cliptmp
, inds
[i
], "");
740 val
= adjust_mask(gallivm
, val
);
741 LLVMBuildStore(builder
, val
, id_ptr
);
743 lp_build_printf(gallivm
, "io = %p, index %d\n, clipmask = %x\n",
744 io_ptrs
[i
], inds
[i
], val
);
749 /* store for each of the n vertices */
750 for (i
= 0; i
< vector_length
; i
++) {
751 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
757 convert_to_aos(struct gallivm_state
*gallivm
,
759 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
760 LLVMValueRef clipmask
,
762 struct lp_type soa_type
,
763 boolean have_clipdist
)
765 LLVMBuilderRef builder
= gallivm
->builder
;
766 unsigned chan
, attrib
, i
;
769 lp_build_printf(gallivm
, " # storing begin\n");
771 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
772 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
773 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
774 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
775 if (outputs
[attrib
][chan
]) {
776 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
777 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
779 lp_build_printf(gallivm
, "output %d : %d ",
780 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
782 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
784 lp_build_print_value(gallivm
, "val = ", out
);
794 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
795 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
797 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
799 for (i
= 0; i
< soa_type
.length
; ++i
) {
800 aos
[i
] = lp_build_extract_range(gallivm
,
801 soa
[i
% TGSI_NUM_CHANNELS
],
802 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
807 store_aos_array(gallivm
,
813 clipmask
, have_clipdist
);
816 lp_build_printf(gallivm
, " # storing end\n");
822 * Stores original vertex positions in clip coordinates
825 store_clip(struct gallivm_state
*gallivm
,
826 const struct lp_type vs_type
,
828 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
829 boolean pre_clip_pos
, int idx
)
831 LLVMBuilderRef builder
= gallivm
->builder
;
833 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
834 LLVMValueRef indices
[2];
835 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
836 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
837 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
841 indices
[1] = lp_build_const_int32(gallivm
, 0);
843 for (i
= 0; i
< vs_type
.length
; i
++) {
844 inds
[i
] = lp_build_const_int32(gallivm
, i
);
845 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
848 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
849 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
850 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
851 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
854 for (i
= 0; i
< vs_type
.length
; i
++) {
855 clip_ptrs
[i
] = draw_jit_header_clip(gallivm
, io_ptrs
[i
]);
858 for (i
= 0; i
< vs_type
.length
; i
++) {
859 clip_ptrs
[i
] = draw_jit_header_pre_clip_pos(gallivm
, io_ptrs
[i
]);
863 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
864 for (i
= 0; i
< vs_type
.length
; ++i
) {
865 aos
[i
] = lp_build_extract_range(gallivm
,
866 soa
[i
% TGSI_NUM_CHANNELS
],
867 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
871 for (j
= 0; j
< vs_type
.length
; j
++) {
872 LLVMTypeRef clip_ptr_type
= LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4), 0);
873 LLVMValueRef clip_ptr
;
875 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
876 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
878 /* Unaligned store */
879 lp_set_store_alignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
885 * Transforms the outputs for viewport mapping
888 generate_viewport(struct draw_llvm_variant
*variant
,
889 LLVMBuilderRef builder
,
890 struct lp_type vs_type
,
891 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
892 LLVMValueRef context_ptr
)
895 struct gallivm_state
*gallivm
= variant
->gallivm
;
896 struct lp_type f32_type
= vs_type
;
897 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
898 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[0][3], ""); /*w0 w1 .. wn*/
899 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
900 LLVMValueRef vp_ptr
= draw_jit_context_viewport(gallivm
, context_ptr
);
902 /* for 1/w convention*/
903 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
904 LLVMBuildStore(builder
, out3
, outputs
[0][3]);
906 /* Viewport Mapping */
907 for (i
=0; i
<3; i
++) {
908 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[0][i
], ""); /*x0 x1 .. xn*/
911 LLVMValueRef scale_i
;
912 LLVMValueRef trans_i
;
915 index
= lp_build_const_int32(gallivm
, i
);
916 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
918 index
= lp_build_const_int32(gallivm
, i
+4);
919 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
921 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
922 LLVMBuildLoad(builder
, scale_i
, "scale"));
923 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
924 LLVMBuildLoad(builder
, trans_i
, "trans"));
927 out
= LLVMBuildFMul(builder
, out
, out3
, "");
929 out
= LLVMBuildFMul(builder
, out
, scale
, "");
930 /* add translation */
931 out
= LLVMBuildFAdd(builder
, out
, trans
, "");
933 /* store transformed outputs */
934 LLVMBuildStore(builder
, out
, outputs
[0][i
]);
941 * Returns clipmask as nxi32 bitmask for the n vertices
944 generate_clipmask(struct draw_llvm
*llvm
,
945 struct gallivm_state
*gallivm
,
946 struct lp_type vs_type
,
947 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
953 LLVMValueRef context_ptr
,
954 boolean
*have_clipdist
)
956 LLVMBuilderRef builder
= gallivm
->builder
;
957 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
958 LLVMValueRef test
, temp
;
959 LLVMValueRef zero
, shift
;
960 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
961 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
962 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
963 struct lp_type f32_type
= vs_type
;
964 struct lp_type i32_type
= lp_int_type(vs_type
);
965 const unsigned pos
= draw_current_shader_position_output(llvm
->draw
);
966 const unsigned cv
= draw_current_shader_clipvertex_output(llvm
->draw
);
967 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
968 bool have_cd
= false;
971 cd
[0] = draw_current_shader_clipdistance_output(llvm
->draw
, 0);
972 cd
[1] = draw_current_shader_clipdistance_output(llvm
->draw
, 1);
974 if (cd
[0] != pos
|| cd
[1] != pos
)
977 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
978 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
979 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
980 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
983 * load clipvertex and position from correct locations.
984 * if they are the same just load them once.
986 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
987 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
988 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
989 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
991 if (clip_user
&& cv
!= pos
) {
992 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
993 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
994 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
995 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1003 /* Cliptest, for hardwired planes */
1006 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1008 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1012 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1013 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1014 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1015 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1016 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1019 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1020 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1021 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1022 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1025 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1026 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1027 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1028 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1029 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1033 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1036 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1037 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1038 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1042 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1043 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1044 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1045 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1048 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1049 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1050 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1051 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1055 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1056 LLVMValueRef indices
[3];
1058 /* userclip planes */
1059 while (ucp_enable
) {
1060 unsigned plane_idx
= ffs(ucp_enable
)-1;
1061 ucp_enable
&= ~(1 << plane_idx
);
1064 if (have_cd
&& num_written_clipdistance
) {
1065 LLVMValueRef clipdist
;
1069 *have_clipdist
= TRUE
;
1071 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1073 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1075 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1076 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1 << plane_idx
);
1077 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1078 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1080 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1081 indices
[0] = lp_build_const_int32(gallivm
, 0);
1082 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1084 indices
[2] = lp_build_const_int32(gallivm
, 0);
1085 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1086 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1087 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1088 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1090 indices
[2] = lp_build_const_int32(gallivm
, 1);
1091 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1092 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1093 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1094 test
= LLVMBuildFMul(builder
, planes
, cv_y
, "");
1095 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1097 indices
[2] = lp_build_const_int32(gallivm
, 2);
1098 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1099 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1100 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1101 test
= LLVMBuildFMul(builder
, planes
, cv_z
, "");
1102 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1104 indices
[2] = lp_build_const_int32(gallivm
, 3);
1105 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1106 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1107 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1108 test
= LLVMBuildFMul(builder
, planes
, cv_w
, "");
1109 sum
= LLVMBuildFAdd(builder
, sum
, test
, "");
1111 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1112 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1 << plane_idx
);
1113 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1114 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1123 * Returns boolean if any clipping has occurred
1124 * Used zero/non-zero i32 value to represent boolean
1127 clipmask_booli32(struct gallivm_state
*gallivm
,
1128 const struct lp_type vs_type
,
1129 LLVMValueRef clipmask_bool_ptr
)
1131 LLVMBuilderRef builder
= gallivm
->builder
;
1132 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1133 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1134 LLVMValueRef ret
= LLVMConstNull(int32_type
);
1139 * Can do this with log2(vector length) pack instructions and one extract
1140 * (as we don't actually need a or) with sse2 which would be way better.
1142 for (i
=0; i
< vs_type
.length
; i
++) {
1143 temp
= LLVMBuildExtractElement(builder
, clipmask_bool
,
1144 lp_build_const_int32(gallivm
, i
) , "");
1145 ret
= LLVMBuildOr(builder
, ret
, temp
, "");
1152 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
,
1155 struct gallivm_state
*gallivm
= variant
->gallivm
;
1156 LLVMContextRef context
= gallivm
->context
;
1157 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1158 LLVMTypeRef arg_types
[8];
1159 LLVMTypeRef func_type
;
1160 LLVMValueRef context_ptr
;
1161 LLVMBasicBlockRef block
;
1162 LLVMBuilderRef builder
;
1163 struct lp_type vs_type
;
1164 LLVMValueRef end
, start
;
1165 LLVMValueRef count
, fetch_elts
, fetch_count
;
1166 LLVMValueRef stride
, step
, io_itr
;
1167 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1168 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
1169 LLVMValueRef one
= lp_build_const_int32(gallivm
, 1);
1170 struct draw_context
*draw
= llvm
->draw
;
1171 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1173 struct lp_build_context bld
;
1174 struct lp_build_loop_state lp_loop
;
1175 const int vector_length
= lp_native_vector_width
/ 32;
1176 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1177 LLVMValueRef fetch_max
;
1178 struct lp_build_sampler_soa
*sampler
= 0;
1179 LLVMValueRef ret
, clipmask_bool_ptr
;
1180 const boolean bypass_viewport
= variant
->key
.bypass_viewport
;
1181 const boolean enable_cliptest
= variant
->key
.clip_xy
||
1182 variant
->key
.clip_z
||
1183 variant
->key
.clip_user
;
1184 LLVMValueRef variant_func
;
1185 const unsigned pos
= draw_current_shader_position_output(llvm
->draw
);
1186 const unsigned cv
= draw_current_shader_clipvertex_output(llvm
->draw
);
1187 boolean have_clipdist
= FALSE
;
1188 struct lp_bld_tgsi_system_values system_values
;
1190 memset(&system_values
, 0, sizeof(system_values
));
1192 arg_types
[0] = get_context_ptr_type(variant
); /* context */
1193 arg_types
[1] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1194 arg_types
[2] = get_buffer_ptr_type(variant
); /* vbuffers */
1196 arg_types
[3] = LLVMPointerType(int32_type
, 0);/* fetch_elts * */
1198 arg_types
[3] = int32_type
; /* start */
1199 arg_types
[4] = int32_type
; /* fetch_count / count */
1200 arg_types
[5] = int32_type
; /* stride */
1201 arg_types
[6] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1202 arg_types
[7] = int32_type
; /* instance_id */
1204 func_type
= LLVMFunctionType(int32_type
, arg_types
, Elements(arg_types
), 0);
1206 variant_func
= LLVMAddFunction(gallivm
->module
,
1207 elts
? "draw_llvm_shader_elts" : "draw_llvm_shader",
1211 variant
->function_elts
= variant_func
;
1213 variant
->function
= variant_func
;
1215 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1216 for (i
= 0; i
< Elements(arg_types
); ++i
)
1217 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1218 LLVMAddAttribute(LLVMGetParam(variant_func
, i
),
1219 LLVMNoAliasAttribute
);
1221 context_ptr
= LLVMGetParam(variant_func
, 0);
1222 io_ptr
= LLVMGetParam(variant_func
, 1);
1223 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1224 stride
= LLVMGetParam(variant_func
, 5);
1225 vb_ptr
= LLVMGetParam(variant_func
, 6);
1226 system_values
.instance_id
= LLVMGetParam(variant_func
, 7);
1228 lp_build_name(context_ptr
, "context");
1229 lp_build_name(io_ptr
, "io");
1230 lp_build_name(vbuffers_ptr
, "vbuffers");
1231 lp_build_name(stride
, "stride");
1232 lp_build_name(vb_ptr
, "vb");
1233 lp_build_name(system_values
.instance_id
, "instance_id");
1236 fetch_elts
= LLVMGetParam(variant_func
, 3);
1237 fetch_count
= LLVMGetParam(variant_func
, 4);
1238 lp_build_name(fetch_elts
, "fetch_elts");
1239 lp_build_name(fetch_count
, "fetch_count");
1240 start
= count
= NULL
;
1243 start
= LLVMGetParam(variant_func
, 3);
1244 count
= LLVMGetParam(variant_func
, 4);
1245 lp_build_name(start
, "start");
1246 lp_build_name(count
, "count");
1247 fetch_elts
= fetch_count
= NULL
;
1254 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
1255 builder
= gallivm
->builder
;
1256 LLVMPositionBuilderAtEnd(builder
, block
);
1258 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
1260 memset(&vs_type
, 0, sizeof vs_type
);
1261 vs_type
.floating
= TRUE
; /* floating point values */
1262 vs_type
.sign
= TRUE
; /* values are signed */
1263 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1264 vs_type
.width
= 32; /* 32-bit float */
1265 vs_type
.length
= vector_length
;
1267 /* hold temporary "bool" clipmask */
1268 clipmask_bool_ptr
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, vs_type
), "");
1269 LLVMBuildStore(builder
, lp_build_zero(gallivm
, lp_int_type(vs_type
)), clipmask_bool_ptr
);
1271 /* code generated texture sampling */
1272 sampler
= draw_llvm_sampler_soa_create(
1273 draw_llvm_variant_key_samplers(&variant
->key
),
1281 end
= lp_build_add(&bld
, start
, count
);
1284 step
= lp_build_const_int32(gallivm
, vector_length
);
1286 fetch_max
= LLVMBuildSub(builder
, end
, one
, "fetch_max");
1288 lp_build_loop_begin(&lp_loop
, gallivm
, start
);
1290 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
1291 LLVMValueRef aos_attribs
[PIPE_MAX_SHADER_INPUTS
][LP_MAX_VECTOR_WIDTH
/ 32] = { { 0 } };
1293 LLVMValueRef clipmask
; /* holds the clipmask value */
1294 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
1297 io_itr
= lp_loop
.counter
;
1299 io_itr
= LLVMBuildSub(builder
, lp_loop
.counter
, start
, "");
1301 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
1303 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
1304 io_itr
, io
, lp_loop
.counter
);
1306 system_values
.vertex_id
= lp_build_zero(gallivm
, lp_type_uint_vec(32, 32*vector_length
));
1307 for (i
= 0; i
< vector_length
; ++i
) {
1308 LLVMValueRef true_index
=
1309 LLVMBuildAdd(builder
,
1311 lp_build_const_int32(gallivm
, i
), "");
1313 /* make sure we're not out of bounds which can happen
1314 * if fetch_count % 4 != 0, because on the last iteration
1315 * a few of the 4 vertex fetches will be out of bounds */
1316 true_index
= lp_build_min(&bld
, true_index
, fetch_max
);
1319 LLVMValueRef fetch_ptr
;
1320 fetch_ptr
= LLVMBuildGEP(builder
, fetch_elts
,
1321 &true_index
, 1, "");
1322 true_index
= LLVMBuildLoad(builder
, fetch_ptr
, "fetch_elt");
1325 system_values
.vertex_id
= LLVMBuildInsertElement(gallivm
->builder
,
1326 system_values
.vertex_id
, true_index
,
1327 lp_build_const_int32(gallivm
, i
), "");
1328 for (j
= 0; j
< draw
->pt
.nr_vertex_elements
; ++j
) {
1329 struct pipe_vertex_element
*velem
= &draw
->pt
.vertex_element
[j
];
1330 LLVMValueRef vb_index
=
1331 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
1332 LLVMValueRef vb
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
1333 generate_fetch(gallivm
, vbuffers_ptr
,
1334 &aos_attribs
[j
][i
], velem
, vb
, true_index
,
1335 system_values
.instance_id
);
1338 convert_to_soa(gallivm
, aos_attribs
, inputs
,
1339 draw
->pt
.nr_vertex_elements
, vs_type
);
1341 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
1342 generate_vs(variant
,
1350 variant
->key
.clamp_vertex_color
);
1352 /* store original positions in clip before further manipulation */
1353 store_clip(gallivm
, vs_type
, io
, outputs
, 0, cv
);
1354 store_clip(gallivm
, vs_type
, io
, outputs
, 1, pos
);
1357 if (enable_cliptest
) {
1358 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1359 /* allocate clipmask, assign it integer type */
1360 clipmask
= generate_clipmask(llvm
,
1364 variant
->key
.clip_xy
,
1365 variant
->key
.clip_z
,
1366 variant
->key
.clip_user
,
1367 variant
->key
.clip_halfz
,
1368 variant
->key
.ucp_enable
,
1369 context_ptr
, &have_clipdist
);
1370 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
1371 /* store temporary clipping boolean value */
1372 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
1375 clipmask
= lp_build_const_int_vec(gallivm
, lp_int_type(vs_type
), 0);
1378 /* do viewport mapping */
1379 if (!bypass_viewport
) {
1380 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
1383 /* store clipmask in vertex header,
1384 * original positions in clip
1385 * and transformed positions in data
1387 convert_to_aos(gallivm
, io
, outputs
, clipmask
,
1388 vs_info
->num_outputs
, vs_type
,
1392 lp_build_loop_end_cond(&lp_loop
, end
, step
, LLVMIntUGE
);
1394 sampler
->destroy(sampler
);
1396 /* return clipping boolean value for function */
1397 ret
= clipmask_booli32(gallivm
, vs_type
, clipmask_bool_ptr
);
1399 LLVMBuildRet(builder
, ret
);
1401 gallivm_verify_function(gallivm
, variant_func
);
1405 struct draw_llvm_variant_key
*
1406 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
1409 struct draw_llvm_variant_key
*key
;
1410 struct draw_sampler_static_state
*draw_sampler
;
1412 key
= (struct draw_llvm_variant_key
*)store
;
1414 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
1416 /* Presumably all variants of the shader should have the same
1417 * number of vertex elements - ie the number of shader inputs.
1418 * NOTE: we NEED to store the needed number of needed inputs
1419 * here, not the number of provided elements to match keysize
1420 * (and the offset of sampler state in the key).
1422 key
->nr_vertex_elements
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
1423 assert(key
->nr_vertex_elements
<= llvm
->draw
->pt
.nr_vertex_elements
);
1425 /* will have to rig this up properly later */
1426 key
->clip_xy
= llvm
->draw
->clip_xy
;
1427 key
->clip_z
= llvm
->draw
->clip_z
;
1428 key
->clip_user
= llvm
->draw
->clip_user
;
1429 key
->bypass_viewport
= llvm
->draw
->identity_viewport
;
1430 key
->clip_halfz
= !llvm
->draw
->rasterizer
->gl_rasterization_rules
;
1431 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
1432 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
1436 /* All variants of this shader will have the same value for
1437 * nr_samplers. Not yet trying to compact away holes in the
1440 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1441 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
1442 key
->nr_sampler_views
=
1443 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
1446 key
->nr_sampler_views
= key
->nr_samplers
;
1449 draw_sampler
= draw_llvm_variant_key_samplers(key
);
1451 memcpy(key
->vertex_element
,
1452 llvm
->draw
->pt
.vertex_element
,
1453 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
1455 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
1457 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
1458 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
1459 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
1461 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
1462 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
1463 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
1471 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
1474 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
1476 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
1477 debug_printf("clip_xy = %u\n", key
->clip_xy
);
1478 debug_printf("clip_z = %u\n", key
->clip_z
);
1479 debug_printf("clip_user = %u\n", key
->clip_user
);
1480 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
1481 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
1482 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
1483 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
1485 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
1486 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
1487 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
1488 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
1489 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
1492 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
1493 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
1499 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
1501 uint32_t width
, uint32_t height
, uint32_t depth
,
1502 uint32_t first_level
, uint32_t last_level
,
1503 const void *base_ptr
,
1504 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
1505 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
1506 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
1509 struct draw_jit_texture
*jit_tex
;
1511 assert(sview_idx
< Elements(draw
->llvm
->jit_context
.textures
));
1513 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
1515 jit_tex
->width
= width
;
1516 jit_tex
->height
= height
;
1517 jit_tex
->depth
= depth
;
1518 jit_tex
->first_level
= first_level
;
1519 jit_tex
->last_level
= last_level
;
1520 jit_tex
->base
= base_ptr
;
1522 for (j
= first_level
; j
<= last_level
; j
++) {
1523 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
1524 jit_tex
->row_stride
[j
] = row_stride
[j
];
1525 jit_tex
->img_stride
[j
] = img_stride
[j
];
1531 draw_llvm_set_sampler_state(struct draw_context
*draw
)
1535 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
1536 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
1538 if (draw
->samplers
[i
]) {
1539 const struct pipe_sampler_state
*s
1540 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
1541 jit_sam
->min_lod
= s
->min_lod
;
1542 jit_sam
->max_lod
= s
->max_lod
;
1543 jit_sam
->lod_bias
= s
->lod_bias
;
1544 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
1551 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
1553 struct draw_llvm
*llvm
= variant
->llvm
;
1555 if (variant
->function_elts
) {
1556 gallivm_free_function(variant
->gallivm
,
1557 variant
->function_elts
, variant
->jit_func_elts
);
1560 if (variant
->function
) {
1561 gallivm_free_function(variant
->gallivm
,
1562 variant
->function
, variant
->jit_func
);
1565 gallivm_destroy(variant
->gallivm
);
1567 remove_from_list(&variant
->list_item_local
);
1568 variant
->shader
->variants_cached
--;
1569 remove_from_list(&variant
->list_item_global
);
1570 llvm
->nr_variants
--;