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_const.h"
73 #include "lp_bld_conv.h"
74 #include "lp_bld_intr.h"
75 #include "lp_bld_logic.h"
76 #include "lp_bld_depth.h"
77 #include "lp_bld_interp.h"
78 #include "lp_bld_tgsi.h"
79 #include "lp_bld_alpha.h"
80 #include "lp_bld_blend.h"
81 #include "lp_bld_swizzle.h"
82 #include "lp_bld_flow.h"
83 #include "lp_bld_debug.h"
84 #include "lp_screen.h"
85 #include "lp_context.h"
88 #include "lp_tex_sample.h"
91 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
92 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
96 * Derive from the quad's upper left scalar coordinates the coordinates for
97 * all other quad pixels
100 generate_pos0(LLVMBuilderRef builder
,
106 LLVMTypeRef int_elem_type
= LLVMInt32Type();
107 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
108 LLVMTypeRef elem_type
= LLVMFloatType();
109 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
110 LLVMValueRef x_offsets
[QUAD_SIZE
];
111 LLVMValueRef y_offsets
[QUAD_SIZE
];
114 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
115 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
117 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
118 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
119 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
122 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
123 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
125 *x0
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
126 *y0
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
131 * Generate the depth test.
134 generate_depth(LLVMBuilderRef builder
,
135 const struct lp_fragment_shader_variant_key
*key
,
136 union lp_type src_type
,
137 struct lp_build_mask_context
*mask
,
139 LLVMValueRef dst_ptr
)
141 const struct util_format_description
*format_desc
;
142 union lp_type dst_type
;
144 if(!key
->depth
.enabled
)
147 format_desc
= util_format_description(key
->zsbuf_format
);
150 /* Pick the depth type. */
151 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
153 /* FIXME: Cope with a depth test type with a different bit width. */
154 assert(dst_type
.width
== src_type
.width
);
155 assert(dst_type
.length
== src_type
.length
);
158 src
= lp_build_clamped_float_to_unsigned_norm(builder
,
163 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
166 lp_build_depth_test(builder
,
177 * Generate the fragment shader, depth/stencil test, and alpha tests.
180 generate_fs(struct llvmpipe_context
*lp
,
181 struct lp_fragment_shader
*shader
,
182 const struct lp_fragment_shader_variant_key
*key
,
183 LLVMBuilderRef builder
,
185 LLVMValueRef context_ptr
,
187 const struct lp_build_interp_soa_context
*interp
,
188 struct lp_build_sampler_soa
*sampler
,
191 LLVMValueRef depth_ptr
)
193 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
194 LLVMTypeRef elem_type
;
195 LLVMTypeRef vec_type
;
196 LLVMTypeRef int_vec_type
;
197 LLVMValueRef consts_ptr
;
198 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
199 LLVMValueRef z
= interp
->pos
[2];
200 struct lp_build_flow_context
*flow
;
201 struct lp_build_mask_context mask
;
202 boolean early_depth_test
;
206 elem_type
= lp_build_elem_type(type
);
207 vec_type
= lp_build_vec_type(type
);
208 int_vec_type
= lp_build_int_vec_type(type
);
210 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
212 flow
= lp_build_flow_create(builder
);
214 memset(outputs
, 0, sizeof outputs
);
216 lp_build_flow_scope_begin(flow
);
218 /* Declare the color and z variables */
219 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
220 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
221 boolean declare
= FALSE
;
222 switch (shader
->info
.output_semantic_name
[attrib
]) {
223 case TGSI_SEMANTIC_COLOR
:
226 case TGSI_SEMANTIC_POSITION
:
232 outputs
[attrib
][chan
] = LLVMGetUndef(vec_type
);
233 lp_build_flow_scope_declare(flow
, &outputs
[attrib
][chan
]);
238 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
241 key
->depth
.enabled
&&
242 !key
->alpha
.enabled
&&
243 !shader
->info
.uses_kill
&&
244 !shader
->info
.writes_z
;
247 generate_depth(builder
, key
,
251 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
252 consts_ptr
, interp
->pos
, interp
->inputs
,
255 if(!early_depth_test
)
256 generate_depth(builder
, key
,
260 lp_build_mask_end(&mask
);
262 lp_build_flow_scope_end(flow
);
264 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
265 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
266 if(outputs
[attrib
][chan
]) {
267 lp_build_name(outputs
[attrib
][chan
], "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
269 switch (shader
->info
.output_semantic_name
[attrib
]) {
270 case TGSI_SEMANTIC_COLOR
:
272 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
274 lp_build_name(outputs
[attrib
][chan
], "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
277 /* XXX: should the alpha reference value be passed separately? */
278 if(cbuf
== 0 && chan
== 3) {
279 LLVMValueRef alpha
= outputs
[attrib
][chan
];
280 LLVMValueRef alpha_ref_value
;
281 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
282 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
283 lp_build_alpha_test(builder
, &key
->alpha
, type
,
284 &mask
, alpha
, alpha_ref_value
);
288 color
[chan
] = outputs
[attrib
][chan
];
293 case TGSI_SEMANTIC_POSITION
:
295 z
= outputs
[attrib
][chan
];
302 lp_build_flow_destroy(flow
);
310 * Generate color blending and color output.
313 generate_blend(const struct pipe_blend_state
*blend
,
314 LLVMBuilderRef builder
,
316 LLVMValueRef context_ptr
,
319 LLVMValueRef dst_ptr
)
321 struct lp_build_context bld
;
322 LLVMTypeRef vec_type
;
323 LLVMTypeRef int_vec_type
;
324 LLVMValueRef const_ptr
;
330 vec_type
= lp_build_vec_type(type
);
331 int_vec_type
= lp_build_int_vec_type(type
);
333 lp_build_context_init(&bld
, builder
, type
);
335 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
336 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
337 LLVMPointerType(vec_type
, 0), "");
339 for(chan
= 0; chan
< 4; ++chan
) {
340 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
341 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
343 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
345 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
346 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
349 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
351 for(chan
= 0; chan
< 4; ++chan
) {
352 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
353 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
354 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
355 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
361 * Generate the runtime callable function for the whole fragment pipeline.
363 static struct lp_fragment_shader_variant
*
364 generate_fragment(struct llvmpipe_context
*lp
,
365 struct lp_fragment_shader
*shader
,
366 const struct lp_fragment_shader_variant_key
*key
)
368 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
369 struct lp_fragment_shader_variant
*variant
;
370 union lp_type fs_type
;
371 union lp_type blend_type
;
372 LLVMTypeRef fs_elem_type
;
373 LLVMTypeRef fs_vec_type
;
374 LLVMTypeRef fs_int_vec_type
;
375 LLVMTypeRef blend_vec_type
;
376 LLVMTypeRef blend_int_vec_type
;
377 LLVMTypeRef arg_types
[9];
378 LLVMTypeRef func_type
;
379 LLVMValueRef context_ptr
;
383 LLVMValueRef dadx_ptr
;
384 LLVMValueRef dady_ptr
;
385 LLVMValueRef mask_ptr
;
386 LLVMValueRef color_ptr
;
387 LLVMValueRef depth_ptr
;
388 LLVMBasicBlockRef block
;
389 LLVMBuilderRef builder
;
392 struct lp_build_sampler_soa
*sampler
;
393 struct lp_build_interp_soa_context interp
;
394 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
395 LLVMValueRef fs_out_color
[NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
396 LLVMValueRef blend_mask
;
397 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
403 tgsi_dump(shader
->base
.tokens
, 0);
404 if(key
->depth
.enabled
) {
405 debug_printf("depth.func = %s\n", debug_dump_func(key
->depth
.func
, TRUE
));
406 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
407 debug_printf("depth.occlusion_count = %u\n", key
->depth
.occlusion_count
);
409 if(key
->alpha
.enabled
) {
410 debug_printf("alpha.func = %s\n", debug_dump_func(key
->alpha
.func
, TRUE
));
411 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
413 if(key
->blend
.logicop_enable
) {
414 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
416 else if(key
->blend
.blend_enable
) {
417 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key
->blend
.rgb_func
, TRUE
));
418 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_src_factor
, TRUE
));
419 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_dst_factor
, TRUE
));
420 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key
->blend
.alpha_func
, TRUE
));
421 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_src_factor
, TRUE
));
422 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_dst_factor
, TRUE
));
424 debug_printf("blend.colormask = 0x%x\n", key
->blend
.colormask
);
427 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
431 variant
->shader
= shader
;
432 memcpy(&variant
->key
, key
, sizeof *key
);
434 /* TODO: actually pick these based on the fs and color buffer
435 * characteristics. */
438 fs_type
.floating
= TRUE
; /* floating point values */
439 fs_type
.sign
= TRUE
; /* values are signed */
440 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
441 fs_type
.width
= 32; /* 32-bit float */
442 fs_type
.length
= 4; /* 4 element per vector */
445 blend_type
.value
= 0;
446 blend_type
.floating
= FALSE
; /* values are integers */
447 blend_type
.sign
= FALSE
; /* values are unsigned */
448 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
449 blend_type
.width
= 8; /* 8-bit ubyte values */
450 blend_type
.length
= 16; /* 16 elements per vector */
453 * Generate the function prototype. Any change here must be reflected in
454 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
457 fs_elem_type
= lp_build_elem_type(fs_type
);
458 fs_vec_type
= lp_build_vec_type(fs_type
);
459 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
461 blend_vec_type
= lp_build_vec_type(blend_type
);
462 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
464 arg_types
[0] = screen
->context_ptr_type
; /* context */
465 arg_types
[1] = LLVMInt32Type(); /* x */
466 arg_types
[2] = LLVMInt32Type(); /* y */
467 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
468 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
469 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
470 arg_types
[6] = LLVMPointerType(fs_int_vec_type
, 0); /* mask */
471 arg_types
[7] = LLVMPointerType(blend_vec_type
, 0); /* color */
472 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
474 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
476 variant
->function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
477 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
478 for(i
= 0; i
< Elements(arg_types
); ++i
)
479 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
480 LLVMAddAttribute(LLVMGetParam(variant
->function
, i
), LLVMNoAliasAttribute
);
482 context_ptr
= LLVMGetParam(variant
->function
, 0);
483 x
= LLVMGetParam(variant
->function
, 1);
484 y
= LLVMGetParam(variant
->function
, 2);
485 a0_ptr
= LLVMGetParam(variant
->function
, 3);
486 dadx_ptr
= LLVMGetParam(variant
->function
, 4);
487 dady_ptr
= LLVMGetParam(variant
->function
, 5);
488 mask_ptr
= LLVMGetParam(variant
->function
, 6);
489 color_ptr
= LLVMGetParam(variant
->function
, 7);
490 depth_ptr
= LLVMGetParam(variant
->function
, 8);
492 lp_build_name(context_ptr
, "context");
493 lp_build_name(x
, "x");
494 lp_build_name(y
, "y");
495 lp_build_name(a0_ptr
, "a0");
496 lp_build_name(dadx_ptr
, "dadx");
497 lp_build_name(dady_ptr
, "dady");
498 lp_build_name(mask_ptr
, "mask");
499 lp_build_name(color_ptr
, "color");
500 lp_build_name(depth_ptr
, "depth");
506 block
= LLVMAppendBasicBlock(variant
->function
, "entry");
507 builder
= LLVMCreateBuilder();
508 LLVMPositionBuilderAtEnd(builder
, block
);
510 generate_pos0(builder
, x
, y
, &x0
, &y0
);
512 lp_build_interp_soa_init(&interp
, shader
->base
.tokens
, builder
, fs_type
,
513 a0_ptr
, dadx_ptr
, dady_ptr
,
517 /* C texture sampling */
518 sampler
= lp_c_sampler_soa_create(context_ptr
);
520 /* code generated texture sampling */
521 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
524 for(i
= 0; i
< num_fs
; ++i
) {
525 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
526 LLVMValueRef out_color
[NUM_CHANNELS
];
527 LLVMValueRef depth_ptr_i
;
530 lp_build_interp_soa_update(&interp
);
532 fs_mask
[i
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, mask_ptr
, &index
, 1, ""), "");
533 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
535 generate_fs(lp
, shader
, key
,
546 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
547 fs_out_color
[chan
][i
] = out_color
[chan
];
550 sampler
->destroy(sampler
);
553 * Convert the fs's output color and mask to fit to the blending type.
556 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
557 lp_build_conv(builder
, fs_type
, blend_type
,
558 fs_out_color
[chan
], num_fs
,
559 &blend_in_color
[chan
], 1);
560 lp_build_name(blend_in_color
[chan
], "color.%c", "rgba"[chan
]);
564 lp_build_conv_mask(builder
, fs_type
, blend_type
,
572 generate_blend(&key
->blend
,
580 LLVMBuildRetVoid(builder
);
582 LLVMDisposeBuilder(builder
);
585 * Translate the LLVM IR into machine code.
588 LLVMRunFunctionPassManager(screen
->pass
, variant
->function
);
591 LLVMDumpValue(variant
->function
);
595 if(LLVMVerifyFunction(variant
->function
, LLVMPrintMessageAction
)) {
596 LLVMDumpValue(variant
->function
);
600 variant
->jit_function
= (lp_jit_frag_func
)LLVMGetPointerToGlobal(screen
->engine
, variant
->function
);
603 lp_disassemble(variant
->jit_function
);
606 variant
->next
= shader
->variants
;
607 shader
->variants
= variant
;
614 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
615 const struct pipe_shader_state
*templ
)
617 struct lp_fragment_shader
*shader
;
619 shader
= CALLOC_STRUCT(lp_fragment_shader
);
623 /* get/save the summary info for this shader */
624 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
626 /* we need to keep a local copy of the tokens */
627 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
634 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
636 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
638 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
640 llvmpipe
->dirty
|= LP_NEW_FS
;
645 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
647 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
648 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
649 struct lp_fragment_shader
*shader
= fs
;
650 struct lp_fragment_shader_variant
*variant
;
652 assert(fs
!= llvmpipe
->fs
);
654 variant
= shader
->variants
;
656 struct lp_fragment_shader_variant
*next
= variant
->next
;
658 if(variant
->function
) {
659 if(variant
->jit_function
)
660 LLVMFreeMachineCodeForFunction(screen
->engine
, variant
->function
);
661 LLVMDeleteFunction(variant
->function
);
669 FREE((void *) shader
->base
.tokens
);
676 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
677 uint shader
, uint index
,
678 const struct pipe_constant_buffer
*buf
)
680 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
682 assert(shader
< PIPE_SHADER_TYPES
);
685 /* note: reference counting */
686 pipe_buffer_reference(&llvmpipe
->constants
[shader
].buffer
,
687 buf
? buf
->buffer
: NULL
);
689 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
694 * We need to generate several variants of the fragment pipeline to match
695 * all the combinations of the contributing state atoms.
697 * TODO: there is actually no reason to tie this to context state -- the
698 * generated code could be cached globally in the screen.
701 make_variant_key(struct llvmpipe_context
*lp
,
702 struct lp_fragment_shader
*shader
,
703 struct lp_fragment_shader_variant_key
*key
)
707 memset(key
, 0, sizeof *key
);
709 if(lp
->framebuffer
.zsbuf
&&
710 lp
->depth_stencil
->depth
.enabled
) {
711 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
712 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
715 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
716 if(key
->alpha
.enabled
)
717 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
718 /* alpha.ref_value is passed in jit_context */
720 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
722 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
723 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
724 lp_sampler_static_state(&key
->sampler
[i
], lp
->texture
[i
], lp
->sampler
[i
]);
729 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
731 struct lp_fragment_shader
*shader
= lp
->fs
;
732 struct lp_fragment_shader_variant_key key
;
733 struct lp_fragment_shader_variant
*variant
;
735 make_variant_key(lp
, shader
, &key
);
737 variant
= shader
->variants
;
739 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
742 variant
= variant
->next
;
746 variant
= generate_fragment(lp
, shader
, &key
);
748 shader
->current
= variant
;