1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
38 * - depth/stencil test (stencil TBI)
41 * This file has only the glue to assembly the fragment pipeline. The actual
42 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
43 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
44 * muster the LLVM JIT execution engine to create a function that follows an
45 * established binary interface and that can be called from C directly.
47 * A big source of complexity here is that we often want to run different
48 * stages with different precisions and data types and precisions. For example,
49 * the fragment shader needs typically to be done in floats, but the
50 * depth/stencil test and blending is better done in the type that most closely
51 * matches the depth/stencil and color buffer respectively.
53 * Since the width of a SIMD vector register stays the same regardless of the
54 * element type, different types imply different number of elements, so we must
55 * code generate more instances of the stages with larger types to be able to
56 * feed/consume the stages with smaller types.
58 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_memory.h"
63 #include "util/u_format.h"
64 #include "util/u_debug_dump.h"
65 #include "pipe/internal/p_winsys_screen.h"
66 #include "pipe/p_shader_tokens.h"
67 #include "draw/draw_context.h"
68 #include "tgsi/tgsi_dump.h"
69 #include "tgsi/tgsi_scan.h"
70 #include "tgsi/tgsi_parse.h"
71 #include "lp_bld_type.h"
72 #include "lp_bld_conv.h"
73 #include "lp_bld_logic.h"
74 #include "lp_bld_depth.h"
75 #include "lp_bld_tgsi.h"
76 #include "lp_bld_alpha.h"
77 #include "lp_bld_blend.h"
78 #include "lp_bld_swizzle.h"
79 #include "lp_bld_flow.h"
80 #include "lp_bld_debug.h"
81 #include "lp_screen.h"
82 #include "lp_context.h"
87 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
88 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
92 * Generate the position vectors.
94 * TODO: This should be called only once per fragment pipeline, for the first
95 * quad, and the neighboring quad positions obtained by additions.
97 * Parameter x, y are the integer values with the quad upper left coordinates.
100 generate_pos(LLVMBuilderRef builder
,
104 LLVMValueRef dadx_ptr
,
105 LLVMValueRef dady_ptr
,
108 LLVMTypeRef int_elem_type
= LLVMInt32Type();
109 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
110 LLVMTypeRef elem_type
= LLVMFloatType();
111 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
112 LLVMValueRef x_offsets
[QUAD_SIZE
];
113 LLVMValueRef y_offsets
[QUAD_SIZE
];
118 * Derive from the quad's upper left scalar coordinates the coordinates for
119 * all other quad pixels
122 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
123 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
125 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
126 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
127 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
130 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
131 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
133 x
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
134 y
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
140 * Calculate z and w from the interpolation factors.
143 for(chan
= 2; chan
< NUM_CHANNELS
; ++chan
) {
144 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
145 LLVMValueRef a0
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, a0_ptr
, &index
, 1, ""), "");
146 LLVMValueRef dadx
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dadx_ptr
, &index
, 1, ""), "");
147 LLVMValueRef dady
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dady_ptr
, &index
, 1, ""), "");
149 a0
= lp_build_broadcast(builder
, vec_type
, a0
);
150 dadx
= lp_build_broadcast(builder
, vec_type
, dadx
);
151 dady
= lp_build_broadcast(builder
, vec_type
, dady
);
153 res
= LLVMBuildAdd(builder
, res
, LLVMBuildMul(builder
, dadx
, x
, ""), "");
154 res
= LLVMBuildAdd(builder
, res
, LLVMBuildMul(builder
, dady
, y
, ""), "");
158 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
159 lp_build_name(pos
[chan
], "pos.%c", "xyzw"[chan
]);
164 * Generate the depth test.
167 generate_depth(struct llvmpipe_context
*lp
,
168 LLVMBuilderRef builder
,
169 const struct pipe_depth_state
*state
,
170 union lp_type src_type
,
171 struct lp_build_mask_context
*mask
,
173 LLVMValueRef dst_ptr
)
175 const struct util_format_description
*format_desc
;
176 union lp_type dst_type
;
178 if(!lp
->framebuffer
.zsbuf
)
181 format_desc
= util_format_description(lp
->framebuffer
.zsbuf
->format
);
184 /* Pick the depth type. */
185 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
187 /* FIXME: Cope with a depth test type with a different bit width. */
188 assert(dst_type
.width
== src_type
.width
);
189 assert(dst_type
.length
== src_type
.length
);
192 src
= lp_build_clamped_float_to_unsigned_norm(builder
,
197 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
200 lp_build_depth_test(builder
,
211 * Generate the fragment shader, depth/stencil test, and alpha tests.
214 generate_fs(struct llvmpipe_context
*lp
,
215 struct lp_fragment_shader
*shader
,
216 const struct lp_fragment_shader_variant_key
*key
,
217 LLVMBuilderRef builder
,
219 LLVMValueRef context_ptr
,
224 LLVMValueRef dadx_ptr
,
225 LLVMValueRef dady_ptr
,
228 LLVMValueRef depth_ptr
)
230 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
231 LLVMTypeRef elem_type
;
232 LLVMTypeRef vec_type
;
233 LLVMTypeRef int_vec_type
;
234 LLVMValueRef consts_ptr
;
235 LLVMValueRef samplers_ptr
;
236 LLVMValueRef pos
[NUM_CHANNELS
];
237 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
238 struct lp_build_mask_context mask
;
239 boolean early_depth_test
;
243 elem_type
= lp_build_elem_type(type
);
244 vec_type
= lp_build_vec_type(type
);
245 int_vec_type
= lp_build_int_vec_type(type
);
247 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
248 samplers_ptr
= lp_jit_context_samplers(builder
, context_ptr
);
250 generate_pos(builder
, x
, y
, a0_ptr
, dadx_ptr
, dady_ptr
, pos
);
252 lp_build_mask_begin(&mask
, builder
, type
, *pmask
);
255 lp
->depth_stencil
->depth
.enabled
&&
256 lp
->framebuffer
.zsbuf
&&
257 !lp
->depth_stencil
->alpha
.enabled
&&
258 !lp
->fs
->info
.uses_kill
&&
259 !lp
->fs
->info
.writes_z
;
262 generate_depth(lp
, builder
, &key
->depth
,
266 memset(outputs
, 0, sizeof outputs
);
268 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
269 pos
, a0_ptr
, dadx_ptr
, dady_ptr
,
270 consts_ptr
, outputs
, samplers_ptr
);
272 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
273 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
274 if(outputs
[attrib
][chan
]) {
275 lp_build_name(outputs
[attrib
][chan
], "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
277 switch (shader
->info
.output_semantic_name
[attrib
]) {
278 case TGSI_SEMANTIC_COLOR
:
280 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
282 lp_build_name(outputs
[attrib
][chan
], "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
285 /* XXX: should the alpha reference value be passed separately? */
286 if(cbuf
== 0 && chan
== 3) {
287 LLVMValueRef alpha
= outputs
[attrib
][chan
];
288 LLVMValueRef alpha_ref_value
;
289 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
290 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
291 lp_build_alpha_test(builder
, &key
->alpha
, type
,
292 &mask
, alpha
, alpha_ref_value
);
296 color
[chan
] = outputs
[attrib
][chan
];
301 case TGSI_SEMANTIC_POSITION
:
303 pos
[2] = outputs
[attrib
][chan
];
310 if(!early_depth_test
)
311 generate_depth(lp
, builder
, &key
->depth
,
315 lp_build_mask_end(&mask
);
323 * Generate color blending and color output.
326 generate_blend(const struct pipe_blend_state
*blend
,
327 LLVMBuilderRef builder
,
329 LLVMValueRef context_ptr
,
332 LLVMValueRef dst_ptr
)
334 struct lp_build_context bld
;
335 LLVMTypeRef vec_type
;
336 LLVMTypeRef int_vec_type
;
337 LLVMValueRef const_ptr
;
343 vec_type
= lp_build_vec_type(type
);
344 int_vec_type
= lp_build_int_vec_type(type
);
346 lp_build_context_init(&bld
, builder
, type
);
348 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
349 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
350 LLVMPointerType(vec_type
, 0), "");
352 for(chan
= 0; chan
< 4; ++chan
) {
353 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
354 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
356 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
358 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
359 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
362 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
364 for(chan
= 0; chan
< 4; ++chan
) {
365 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
366 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
367 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
368 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
374 * Generate the runtime callable function for the whole fragment pipeline.
376 static struct lp_fragment_shader_variant
*
377 generate_fragment(struct llvmpipe_context
*lp
,
378 struct lp_fragment_shader
*shader
,
379 const struct lp_fragment_shader_variant_key
*key
)
381 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
382 struct lp_fragment_shader_variant
*variant
;
383 union lp_type fs_type
;
384 union lp_type blend_type
;
385 LLVMTypeRef fs_elem_type
;
386 LLVMTypeRef fs_vec_type
;
387 LLVMTypeRef fs_int_vec_type
;
388 LLVMTypeRef blend_vec_type
;
389 LLVMTypeRef blend_int_vec_type
;
390 LLVMTypeRef arg_types
[9];
391 LLVMTypeRef func_type
;
392 LLVMValueRef context_ptr
;
396 LLVMValueRef dadx_ptr
;
397 LLVMValueRef dady_ptr
;
398 LLVMValueRef mask_ptr
;
399 LLVMValueRef color_ptr
;
400 LLVMValueRef depth_ptr
;
401 LLVMBasicBlockRef block
;
402 LLVMBuilderRef builder
;
403 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
404 LLVMValueRef fs_out_color
[NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
405 LLVMValueRef blend_mask
;
406 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
412 tgsi_dump(shader
->base
.tokens
, 0);
413 if(key
->depth
.enabled
) {
414 debug_printf("depth.func = %s\n", debug_dump_func(key
->depth
.func
, TRUE
));
415 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
416 debug_printf("depth.occlusion_count = %u\n", key
->depth
.occlusion_count
);
418 if(key
->alpha
.enabled
) {
419 debug_printf("alpha.func = %s\n", debug_dump_func(key
->alpha
.func
, TRUE
));
420 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
422 if(key
->blend
.logicop_enable
) {
423 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
425 else if(key
->blend
.blend_enable
) {
426 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key
->blend
.rgb_func
, TRUE
));
427 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_src_factor
, TRUE
));
428 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_dst_factor
, TRUE
));
429 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key
->blend
.alpha_func
, TRUE
));
430 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_src_factor
, TRUE
));
431 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_dst_factor
, TRUE
));
433 debug_printf("blend.colormask = 0x%x\n", key
->blend
.colormask
);
436 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
440 variant
->shader
= shader
;
441 memcpy(&variant
->key
, key
, sizeof *key
);
443 /* TODO: actually pick these based on the fs and color buffer
444 * characteristics. */
447 fs_type
.floating
= TRUE
; /* floating point values */
448 fs_type
.sign
= TRUE
; /* values are signed */
449 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
450 fs_type
.width
= 32; /* 32-bit float */
451 fs_type
.length
= 4; /* 4 element per vector */
454 blend_type
.value
= 0;
455 blend_type
.floating
= FALSE
; /* values are integers */
456 blend_type
.sign
= FALSE
; /* values are unsigned */
457 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
458 blend_type
.width
= 8; /* 8-bit ubyte values */
459 blend_type
.length
= 16; /* 16 elements per vector */
462 * Generate the function prototype. Any change here must be reflected in
463 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
466 fs_elem_type
= lp_build_elem_type(fs_type
);
467 fs_vec_type
= lp_build_vec_type(fs_type
);
468 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
470 blend_vec_type
= lp_build_vec_type(blend_type
);
471 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
473 arg_types
[0] = screen
->context_ptr_type
; /* context */
474 arg_types
[1] = LLVMInt32Type(); /* x */
475 arg_types
[2] = LLVMInt32Type(); /* y */
476 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
477 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
478 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
479 arg_types
[6] = LLVMPointerType(fs_int_vec_type
, 0); /* mask */
480 arg_types
[7] = LLVMPointerType(blend_vec_type
, 0); /* color */
481 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
483 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
485 variant
->function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
486 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
487 for(i
= 0; i
< Elements(arg_types
); ++i
)
488 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
489 LLVMAddAttribute(LLVMGetParam(variant
->function
, i
), LLVMNoAliasAttribute
);
491 context_ptr
= LLVMGetParam(variant
->function
, 0);
492 x
= LLVMGetParam(variant
->function
, 1);
493 y
= LLVMGetParam(variant
->function
, 2);
494 a0_ptr
= LLVMGetParam(variant
->function
, 3);
495 dadx_ptr
= LLVMGetParam(variant
->function
, 4);
496 dady_ptr
= LLVMGetParam(variant
->function
, 5);
497 mask_ptr
= LLVMGetParam(variant
->function
, 6);
498 color_ptr
= LLVMGetParam(variant
->function
, 7);
499 depth_ptr
= LLVMGetParam(variant
->function
, 8);
501 lp_build_name(context_ptr
, "context");
502 lp_build_name(x
, "x");
503 lp_build_name(y
, "y");
504 lp_build_name(a0_ptr
, "a0");
505 lp_build_name(dadx_ptr
, "dadx");
506 lp_build_name(dady_ptr
, "dady");
507 lp_build_name(mask_ptr
, "mask");
508 lp_build_name(color_ptr
, "color");
509 lp_build_name(depth_ptr
, "depth");
515 block
= LLVMAppendBasicBlock(variant
->function
, "entry");
516 builder
= LLVMCreateBuilder();
517 LLVMPositionBuilderAtEnd(builder
, block
);
519 for(i
= 0; i
< num_fs
; ++i
) {
520 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
521 LLVMValueRef out_color
[NUM_CHANNELS
];
523 LLVMValueRef depth_ptr_i
;
525 /* TODO: Reuse position interpolation */
526 x_i
= LLVMBuildAdd(builder
, x
, LLVMConstInt(LLVMInt32Type(), 2*i
, 0), "");
528 fs_mask
[i
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, mask_ptr
, &index
, 1, ""), "");
529 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
531 generate_fs(lp
, shader
, key
,
537 a0_ptr
, dadx_ptr
, dady_ptr
,
542 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
543 fs_out_color
[chan
][i
] = out_color
[chan
];
547 * Convert the fs's output color and mask to fit to the blending type.
550 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
551 lp_build_conv(builder
, fs_type
, blend_type
,
552 fs_out_color
[chan
], num_fs
,
553 &blend_in_color
[chan
], 1);
554 lp_build_name(blend_in_color
[chan
], "color.%c", "rgba"[chan
]);
558 lp_build_conv_mask(builder
, fs_type
, blend_type
,
566 generate_blend(&key
->blend
,
574 LLVMBuildRetVoid(builder
);
576 LLVMDisposeBuilder(builder
);
579 * Translate the LLVM IR into machine code.
582 LLVMRunFunctionPassManager(screen
->pass
, variant
->function
);
585 LLVMDumpValue(variant
->function
);
589 if(LLVMVerifyFunction(variant
->function
, LLVMPrintMessageAction
)) {
590 LLVMDumpValue(variant
->function
);
594 variant
->jit_function
= (lp_jit_frag_func
)LLVMGetPointerToGlobal(screen
->engine
, variant
->function
);
597 lp_disassemble(variant
->jit_function
);
600 variant
->next
= shader
->variants
;
601 shader
->variants
= variant
;
608 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
609 const struct pipe_shader_state
*templ
)
611 struct lp_fragment_shader
*shader
;
613 shader
= CALLOC_STRUCT(lp_fragment_shader
);
617 /* get/save the summary info for this shader */
618 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
620 /* we need to keep a local copy of the tokens */
621 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
628 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
630 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
632 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
634 llvmpipe
->dirty
|= LP_NEW_FS
;
639 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
641 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
642 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
643 struct lp_fragment_shader
*shader
= fs
;
644 struct lp_fragment_shader_variant
*variant
;
646 assert(fs
!= llvmpipe
->fs
);
648 variant
= shader
->variants
;
650 struct lp_fragment_shader_variant
*next
= variant
->next
;
652 if(variant
->function
) {
653 if(variant
->jit_function
)
654 LLVMFreeMachineCodeForFunction(screen
->engine
, variant
->function
);
655 LLVMDeleteFunction(variant
->function
);
663 FREE((void *) shader
->base
.tokens
);
670 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
671 uint shader
, uint index
,
672 const struct pipe_constant_buffer
*buf
)
674 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
676 assert(shader
< PIPE_SHADER_TYPES
);
679 /* note: reference counting */
680 pipe_buffer_reference(&llvmpipe
->constants
[shader
].buffer
,
681 buf
? buf
->buffer
: NULL
);
683 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
688 * We need to generate several variants of the fragment pipeline to match
689 * all the combinations of the contributing state atoms.
691 * TODO: there is actually no reason to tie this to context state -- the
692 * generated code could be cached globally in the screen.
695 make_variant_key(struct llvmpipe_context
*lp
,
696 struct lp_fragment_shader_variant_key
*key
)
698 memset(key
, 0, sizeof *key
);
700 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof &key
->depth
);
702 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
703 if(key
->alpha
.enabled
)
704 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
705 /* alpha.ref_value is passed in jit_context */
707 memcpy(&key
->blend
, lp
->blend
, sizeof &key
->blend
);
712 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
714 struct lp_fragment_shader
*shader
= lp
->fs
;
715 struct lp_fragment_shader_variant_key key
;
716 struct lp_fragment_shader_variant
*variant
;
718 make_variant_key(lp
, &key
);
720 variant
= shader
->variants
;
722 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
725 variant
= variant
->next
;
729 variant
= generate_fragment(lp
, shader
, &key
);
731 shader
->current
= variant
;