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:
37 * - depth/stencil test
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
57 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/u_simple_list.h"
69 #include "os/os_time.h"
70 #include "pipe/p_shader_tokens.h"
71 #include "draw/draw_context.h"
72 #include "tgsi/tgsi_dump.h"
73 #include "tgsi/tgsi_scan.h"
74 #include "tgsi/tgsi_parse.h"
75 #include "gallivm/lp_bld_type.h"
76 #include "gallivm/lp_bld_const.h"
77 #include "gallivm/lp_bld_conv.h"
78 #include "gallivm/lp_bld_init.h"
79 #include "gallivm/lp_bld_intr.h"
80 #include "gallivm/lp_bld_logic.h"
81 #include "gallivm/lp_bld_tgsi.h"
82 #include "gallivm/lp_bld_swizzle.h"
83 #include "gallivm/lp_bld_flow.h"
84 #include "gallivm/lp_bld_debug.h"
86 #include "lp_bld_alpha.h"
87 #include "lp_bld_blend.h"
88 #include "lp_bld_depth.h"
89 #include "lp_bld_interp.h"
90 #include "lp_context.h"
95 #include "lp_tex_sample.h"
97 #include "lp_state_fs.h"
100 #include <llvm-c/Analysis.h>
101 #include <llvm-c/BitWriter.h>
104 /** Fragment shader number (for debugging) */
105 static unsigned fs_no
= 0;
109 * Expand the relevent bits of mask_input to a 4-dword mask for the
110 * four pixels in a 2x2 quad. This will set the four elements of the
111 * quad mask vector to 0 or ~0.
113 * \param quad which quad of the quad group to test, in [0,3]
114 * \param mask_input bitwise mask for the whole 4x4 stamp
117 generate_quad_mask(struct gallivm_state
*gallivm
,
118 struct lp_type fs_type
,
120 LLVMValueRef mask_input
) /* int32 */
122 LLVMBuilderRef builder
= gallivm
->builder
;
123 struct lp_type mask_type
;
124 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
125 LLVMValueRef bits
[4];
130 * XXX: We'll need a different path for 16 x u8
132 assert(fs_type
.width
== 32);
133 assert(fs_type
.length
== 4);
134 mask_type
= lp_int_type(fs_type
);
137 * mask_input >>= (quad * 4)
157 mask_input
= LLVMBuildLShr(builder
,
159 LLVMConstInt(i32t
, shift
, 0),
163 * mask = { mask_input & (1 << i), for i in [0,3] }
165 mask
= lp_build_broadcast(gallivm
,
166 lp_build_vec_type(gallivm
, mask_type
),
169 bits
[0] = LLVMConstInt(i32t
, 1 << 0, 0);
170 bits
[1] = LLVMConstInt(i32t
, 1 << 1, 0);
171 bits
[2] = LLVMConstInt(i32t
, 1 << 4, 0);
172 bits
[3] = LLVMConstInt(i32t
, 1 << 5, 0);
174 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, 4), "");
177 * mask = mask != 0 ? ~0 : 0
179 mask
= lp_build_compare(gallivm
,
180 mask_type
, PIPE_FUNC_NOTEQUAL
,
182 lp_build_const_int_vec(gallivm
, mask_type
, 0));
188 #define EARLY_DEPTH_TEST 0x1
189 #define LATE_DEPTH_TEST 0x2
190 #define EARLY_DEPTH_WRITE 0x4
191 #define LATE_DEPTH_WRITE 0x8
194 find_output_by_semantic( const struct tgsi_shader_info
*info
,
200 for (i
= 0; i
< info
->num_outputs
; i
++)
201 if (info
->output_semantic_name
[i
] == semantic
&&
202 info
->output_semantic_index
[i
] == index
)
210 * Generate the fragment shader, depth/stencil test, and alpha tests.
211 * \param i which quad in the tile, in range [0,3]
212 * \param partial_mask if 1, do mask_input testing
215 generate_fs(struct gallivm_state
*gallivm
,
216 struct lp_fragment_shader
*shader
,
217 const struct lp_fragment_shader_variant_key
*key
,
218 LLVMBuilderRef builder
,
220 LLVMValueRef context_ptr
,
222 struct lp_build_interp_soa_context
*interp
,
223 struct lp_build_sampler_soa
*sampler
,
225 LLVMValueRef (*color
)[4],
226 LLVMValueRef depth_ptr
,
228 unsigned partial_mask
,
229 LLVMValueRef mask_input
,
230 LLVMValueRef counter
)
232 const struct util_format_description
*zs_format_desc
= NULL
;
233 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
234 LLVMTypeRef vec_type
;
235 LLVMValueRef consts_ptr
;
236 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
238 LLVMValueRef zs_value
= NULL
;
239 LLVMValueRef stencil_refs
[2];
240 struct lp_build_mask_context mask
;
241 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
242 shader
->info
.base
.num_inputs
< 3 &&
243 shader
->info
.base
.num_instructions
< 8);
249 if (key
->depth
.enabled
||
250 key
->stencil
[0].enabled
||
251 key
->stencil
[1].enabled
) {
253 zs_format_desc
= util_format_description(key
->zsbuf_format
);
254 assert(zs_format_desc
);
256 if (!shader
->info
.base
.writes_z
) {
257 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
258 /* With alpha test and kill, can do the depth test early
259 * and hopefully eliminate some quads. But need to do a
260 * special deferred depth write once the final mask value
263 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
265 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
268 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
271 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
272 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
273 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
281 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
282 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
284 vec_type
= lp_build_vec_type(gallivm
, type
);
286 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
288 memset(outputs
, 0, sizeof outputs
);
290 /* Declare the color and z variables */
291 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
292 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
293 color
[cbuf
][chan
] = lp_build_alloca(gallivm
, vec_type
, "color");
297 /* do triangle edge testing */
299 *pmask
= generate_quad_mask(gallivm
, type
,
303 *pmask
= lp_build_const_int_vec(gallivm
, type
, ~0);
306 /* 'mask' will control execution based on quad's pixel alive/killed state */
307 lp_build_mask_begin(&mask
, gallivm
, type
, *pmask
);
309 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
310 lp_build_mask_check(&mask
);
312 lp_build_interp_soa_update_pos(interp
, gallivm
, i
);
315 if (depth_mode
& EARLY_DEPTH_TEST
) {
316 lp_build_depth_stencil_test(gallivm
,
328 if (depth_mode
& EARLY_DEPTH_WRITE
) {
329 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
333 lp_build_interp_soa_update_inputs(interp
, gallivm
, i
);
335 /* Build the actual shader */
336 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
337 consts_ptr
, interp
->pos
, interp
->inputs
,
338 outputs
, sampler
, &shader
->info
.base
);
341 if (key
->alpha
.enabled
) {
342 int color0
= find_output_by_semantic(&shader
->info
.base
,
346 if (color0
!= -1 && outputs
[color0
][3]) {
347 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
348 LLVMValueRef alpha_ref_value
;
350 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
351 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
353 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
,
354 &mask
, alpha
, alpha_ref_value
,
355 (depth_mode
& LATE_DEPTH_TEST
) != 0);
360 if (depth_mode
& LATE_DEPTH_TEST
) {
361 int pos0
= find_output_by_semantic(&shader
->info
.base
,
362 TGSI_SEMANTIC_POSITION
,
365 if (pos0
!= -1 && outputs
[pos0
][2]) {
366 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
369 lp_build_depth_stencil_test(gallivm
,
381 if (depth_mode
& LATE_DEPTH_WRITE
) {
382 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
385 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
386 (depth_mode
& LATE_DEPTH_WRITE
))
388 /* Need to apply a reduced mask to the depth write. Reload the
389 * depth value, update from zs_value with the new mask value and
392 lp_build_deferred_depth_write(gallivm
,
402 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
404 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
&&
405 shader
->info
.base
.output_semantic_index
[attrib
] < key
->nr_cbufs
)
407 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
408 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
409 if(outputs
[attrib
][chan
]) {
410 /* XXX: just initialize outputs to point at colors[] and
413 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
414 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
415 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
422 lp_build_occlusion_count(gallivm
, type
,
423 lp_build_mask_value(&mask
), counter
);
425 *pmask
= lp_build_mask_end(&mask
);
430 * Generate color blending and color output.
431 * \param rt the render target index (to index blend, colormask state)
432 * \param type the pixel color type
433 * \param context_ptr pointer to the runtime JIT context
434 * \param mask execution mask (active fragment/pixel mask)
435 * \param src colors from the fragment shader
436 * \param dst_ptr the destination color buffer pointer
439 generate_blend(struct gallivm_state
*gallivm
,
440 const struct pipe_blend_state
*blend
,
442 LLVMBuilderRef builder
,
444 LLVMValueRef context_ptr
,
447 LLVMValueRef dst_ptr
,
450 struct lp_build_context bld
;
451 struct lp_build_mask_context mask_ctx
;
452 LLVMTypeRef vec_type
;
453 LLVMValueRef const_ptr
;
459 lp_build_context_init(&bld
, gallivm
, type
);
461 lp_build_mask_begin(&mask_ctx
, gallivm
, type
, mask
);
463 lp_build_mask_check(&mask_ctx
);
465 vec_type
= lp_build_vec_type(gallivm
, type
);
467 const_ptr
= lp_jit_context_blend_color(gallivm
, context_ptr
);
468 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
469 LLVMPointerType(vec_type
, 0), "");
471 /* load constant blend color and colors from the dest color buffer */
472 for(chan
= 0; chan
< 4; ++chan
) {
473 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
474 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
476 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
478 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
479 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
483 lp_build_blend_soa(gallivm
, blend
, type
, rt
, src
, dst
, con
, res
);
485 /* store results to color buffer */
486 for(chan
= 0; chan
< 4; ++chan
) {
487 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
488 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
489 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
490 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
491 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
495 lp_build_mask_end(&mask_ctx
);
500 * Generate the runtime callable function for the whole fragment pipeline.
501 * Note that the function which we generate operates on a block of 16
502 * pixels at at time. The block contains 2x2 quads. Each quad contains
506 generate_fragment(struct llvmpipe_context
*lp
,
507 struct lp_fragment_shader
*shader
,
508 struct lp_fragment_shader_variant
*variant
,
509 unsigned partial_mask
)
511 struct gallivm_state
*gallivm
= lp
->gallivm
;
512 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
513 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
515 struct lp_type fs_type
;
516 struct lp_type blend_type
;
517 LLVMTypeRef fs_elem_type
;
518 LLVMTypeRef fs_int_vec_type
;
519 LLVMTypeRef blend_vec_type
;
520 LLVMTypeRef arg_types
[11];
521 LLVMTypeRef func_type
;
522 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
523 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
524 LLVMValueRef context_ptr
;
528 LLVMValueRef dadx_ptr
;
529 LLVMValueRef dady_ptr
;
530 LLVMValueRef color_ptr_ptr
;
531 LLVMValueRef depth_ptr
;
532 LLVMValueRef mask_input
;
533 LLVMValueRef counter
= NULL
;
534 LLVMBasicBlockRef block
;
535 LLVMBuilderRef builder
;
536 struct lp_build_sampler_soa
*sampler
;
537 struct lp_build_interp_soa_context interp
;
538 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
539 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
540 LLVMValueRef blend_mask
;
541 LLVMValueRef function
;
543 const struct util_format_description
*zs_format_desc
;
549 /* Adjust color input interpolation according to flatshade state:
551 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
552 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
553 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
555 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
557 inputs
[i
].interp
= LP_INTERP_LINEAR
;
562 /* TODO: actually pick these based on the fs and color buffer
563 * characteristics. */
565 memset(&fs_type
, 0, sizeof fs_type
);
566 fs_type
.floating
= TRUE
; /* floating point values */
567 fs_type
.sign
= TRUE
; /* values are signed */
568 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
569 fs_type
.width
= 32; /* 32-bit float */
570 fs_type
.length
= 4; /* 4 elements per vector */
571 num_fs
= 4; /* number of quads per block */
573 memset(&blend_type
, 0, sizeof blend_type
);
574 blend_type
.floating
= FALSE
; /* values are integers */
575 blend_type
.sign
= FALSE
; /* values are unsigned */
576 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
577 blend_type
.width
= 8; /* 8-bit ubyte values */
578 blend_type
.length
= 16; /* 16 elements per vector */
581 * Generate the function prototype. Any change here must be reflected in
582 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
585 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
586 fs_int_vec_type
= lp_build_int_vec_type(gallivm
, fs_type
);
588 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
590 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
591 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
593 arg_types
[0] = lp_jit_get_context_type(lp
); /* context */
594 arg_types
[1] = int32_type
; /* x */
595 arg_types
[2] = int32_type
; /* y */
596 arg_types
[3] = int32_type
; /* facing */
597 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
598 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
599 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
600 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
601 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
602 arg_types
[9] = int32_type
; /* mask_input */
603 arg_types
[10] = LLVMPointerType(int32_type
, 0); /* counter */
605 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
606 arg_types
, Elements(arg_types
), 0);
608 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
609 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
611 variant
->function
[partial_mask
] = function
;
613 /* XXX: need to propagate noalias down into color param now we are
614 * passing a pointer-to-pointer?
616 for(i
= 0; i
< Elements(arg_types
); ++i
)
617 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
618 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
620 context_ptr
= LLVMGetParam(function
, 0);
621 x
= LLVMGetParam(function
, 1);
622 y
= LLVMGetParam(function
, 2);
623 facing
= LLVMGetParam(function
, 3);
624 a0_ptr
= LLVMGetParam(function
, 4);
625 dadx_ptr
= LLVMGetParam(function
, 5);
626 dady_ptr
= LLVMGetParam(function
, 6);
627 color_ptr_ptr
= LLVMGetParam(function
, 7);
628 depth_ptr
= LLVMGetParam(function
, 8);
629 mask_input
= LLVMGetParam(function
, 9);
631 lp_build_name(context_ptr
, "context");
632 lp_build_name(x
, "x");
633 lp_build_name(y
, "y");
634 lp_build_name(a0_ptr
, "a0");
635 lp_build_name(dadx_ptr
, "dadx");
636 lp_build_name(dady_ptr
, "dady");
637 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
638 lp_build_name(depth_ptr
, "depth");
639 lp_build_name(mask_input
, "mask_input");
641 if (key
->occlusion_count
) {
642 counter
= LLVMGetParam(function
, 10);
643 lp_build_name(counter
, "counter");
650 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
651 builder
= gallivm
->builder
;
653 LLVMPositionBuilderAtEnd(builder
, block
);
656 * The shader input interpolation info is not explicitely baked in the
657 * shader key, but everything it derives from (TGSI, and flatshade) is
658 * already included in the shader key.
660 lp_build_interp_soa_init(&interp
,
662 shader
->info
.base
.num_inputs
,
665 a0_ptr
, dadx_ptr
, dady_ptr
,
668 /* code generated texture sampling */
669 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
671 /* loop over quads in the block */
672 zs_format_desc
= util_format_description(key
->zsbuf_format
);
674 for(i
= 0; i
< num_fs
; ++i
) {
675 LLVMValueRef depth_offset
= LLVMConstInt(int32_type
,
676 i
*fs_type
.length
*zs_format_desc
->block
.bits
/8,
678 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
679 LLVMValueRef depth_ptr_i
;
681 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
691 &fs_mask
[i
], /* output */
699 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
700 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
701 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
704 sampler
->destroy(sampler
);
706 /* Loop over color outputs / color buffers to do blending.
708 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
709 LLVMValueRef color_ptr
;
710 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
711 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
715 * Convert the fs's output color and mask to fit to the blending type.
717 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
718 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
720 for (i
= 0; i
< num_fs
; i
++) {
722 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
725 lp_build_conv(gallivm
, fs_type
, blend_type
,
728 &blend_in_color
[chan
], 1);
730 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
733 if (partial_mask
|| !variant
->opaque
) {
734 lp_build_conv_mask(lp
->gallivm
, fs_type
, blend_type
,
738 blend_mask
= lp_build_const_int_vec(lp
->gallivm
, blend_type
, ~0);
741 color_ptr
= LLVMBuildLoad(builder
,
742 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
744 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
746 /* which blend/colormask state to use */
747 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
753 /* Could the 4x4 have been killed?
755 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
756 !key
->alpha
.enabled
&&
757 !shader
->info
.base
.uses_kill
);
759 generate_blend(lp
->gallivm
,
772 LLVMBuildRetVoid(builder
);
774 /* Verify the LLVM IR. If invalid, dump and abort */
776 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
778 lp_debug_dump_value(function
);
783 /* Apply optimizations to LLVM IR */
784 LLVMRunFunctionPassManager(gallivm
->passmgr
, function
);
786 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
787 /* Print the LLVM IR to stderr */
788 lp_debug_dump_value(function
);
792 /* Dump byte code to a file */
794 LLVMWriteBitcodeToFile(gallivm
->module
, "llvmpipe.bc");
798 * Translate the LLVM IR into machine code.
801 void *f
= LLVMGetPointerToGlobal(gallivm
->engine
, function
);
803 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
805 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
808 lp_func_delete_body(function
);
814 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
818 debug_printf("fs variant %p:\n", (void *) key
);
820 if (key
->flatshade
) {
821 debug_printf("flatshade = 1\n");
823 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
824 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
826 if (key
->depth
.enabled
) {
827 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
828 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
829 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
832 for (i
= 0; i
< 2; ++i
) {
833 if (key
->stencil
[i
].enabled
) {
834 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
835 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
836 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
837 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
838 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
839 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
843 if (key
->alpha
.enabled
) {
844 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
847 if (key
->occlusion_count
) {
848 debug_printf("occlusion_count = 1\n");
851 if (key
->blend
.logicop_enable
) {
852 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
854 else if (key
->blend
.rt
[0].blend_enable
) {
855 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
856 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
857 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
858 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
859 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
860 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
862 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
863 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
864 debug_printf("sampler[%u] = \n", i
);
865 debug_printf(" .format = %s\n",
866 util_format_name(key
->sampler
[i
].format
));
867 debug_printf(" .target = %s\n",
868 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
869 debug_printf(" .pot = %u %u %u\n",
870 key
->sampler
[i
].pot_width
,
871 key
->sampler
[i
].pot_height
,
872 key
->sampler
[i
].pot_depth
);
873 debug_printf(" .wrap = %s %s %s\n",
874 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
875 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
876 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
877 debug_printf(" .min_img_filter = %s\n",
878 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
879 debug_printf(" .min_mip_filter = %s\n",
880 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
881 debug_printf(" .mag_img_filter = %s\n",
882 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
883 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
884 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
885 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
886 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
887 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
888 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
889 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
895 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
897 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
898 variant
->shader
->no
, variant
->no
);
899 tgsi_dump(variant
->shader
->base
.tokens
, 0);
900 dump_fs_variant_key(&variant
->key
);
901 debug_printf("variant->opaque = %u\n", variant
->opaque
);
907 * Generate a new fragment shader variant from the shader code and
908 * other state indicated by the key.
910 static struct lp_fragment_shader_variant
*
911 generate_variant(struct llvmpipe_context
*lp
,
912 struct lp_fragment_shader
*shader
,
913 const struct lp_fragment_shader_variant_key
*key
)
915 struct lp_fragment_shader_variant
*variant
;
916 boolean fullcolormask
;
918 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
922 variant
->shader
= shader
;
923 variant
->list_item_global
.base
= variant
;
924 variant
->list_item_local
.base
= variant
;
925 variant
->no
= shader
->variants_created
++;
927 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
930 * Determine whether we are touching all channels in the color buffer.
932 fullcolormask
= FALSE
;
933 if (key
->nr_cbufs
== 1) {
934 const struct util_format_description
*format_desc
;
935 format_desc
= util_format_description(key
->cbuf_format
[0]);
936 if ((~key
->blend
.rt
[0].colormask
&
937 util_format_colormask(format_desc
)) == 0) {
938 fullcolormask
= TRUE
;
943 !key
->blend
.logicop_enable
&&
944 !key
->blend
.rt
[0].blend_enable
&&
946 !key
->stencil
[0].enabled
&&
947 !key
->alpha
.enabled
&&
948 !key
->depth
.enabled
&&
949 !shader
->info
.base
.uses_kill
953 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
954 lp_debug_fs_variant(variant
);
957 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
959 if (variant
->opaque
) {
960 /* Specialized shader, which doesn't need to read the color buffer. */
961 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
963 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
971 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
972 const struct pipe_shader_state
*templ
)
974 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
975 struct lp_fragment_shader
*shader
;
979 shader
= CALLOC_STRUCT(lp_fragment_shader
);
983 shader
->no
= fs_no
++;
984 make_empty_list(&shader
->variants
);
986 /* get/save the summary info for this shader */
987 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
989 /* we need to keep a local copy of the tokens */
990 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
992 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
993 if (shader
->draw_data
== NULL
) {
994 FREE((void *) shader
->base
.tokens
);
999 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1001 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
1002 sampler
[nr_samplers
]);
1004 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1005 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
1007 switch (shader
->info
.base
.input_interpolate
[i
]) {
1008 case TGSI_INTERPOLATE_CONSTANT
:
1009 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1011 case TGSI_INTERPOLATE_LINEAR
:
1012 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
1014 case TGSI_INTERPOLATE_PERSPECTIVE
:
1015 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1022 switch (shader
->info
.base
.input_semantic_name
[i
]) {
1023 case TGSI_SEMANTIC_COLOR
:
1024 /* Colors may be either linearly or constant interpolated in
1025 * the fragment shader, but that information isn't available
1026 * here. Mark color inputs and fix them up later.
1028 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
1030 case TGSI_SEMANTIC_FACE
:
1031 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
1033 case TGSI_SEMANTIC_POSITION
:
1034 /* Position was already emitted above
1036 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
1037 shader
->inputs
[i
].src_index
= 0;
1041 shader
->inputs
[i
].src_index
= i
+1;
1044 if (LP_DEBUG
& DEBUG_TGSI
) {
1046 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
1047 shader
->no
, (void *) shader
);
1048 tgsi_dump(templ
->tokens
, 0);
1049 debug_printf("usage masks:\n");
1050 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
1051 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
1052 debug_printf(" IN[%u].%s%s%s%s\n",
1054 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
1055 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
1056 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
1057 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
1067 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1069 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1071 if (llvmpipe
->fs
== fs
)
1074 draw_flush(llvmpipe
->draw
);
1076 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
1078 draw_bind_fragment_shader(llvmpipe
->draw
,
1079 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
1081 llvmpipe
->dirty
|= LP_NEW_FS
;
1086 * Remove shader variant from two lists: the shader's variant list
1087 * and the context's variant list.
1090 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
1091 struct lp_fragment_shader_variant
*variant
)
1095 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
1096 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
1097 " #%u v total cached #%u\n",
1098 variant
->shader
->no
,
1100 variant
->shader
->variants_created
,
1101 variant
->shader
->variants_cached
,
1102 lp
->nr_fs_variants
);
1105 /* free all the variant's JIT'd functions */
1106 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1107 if (variant
->function
[i
]) {
1108 if (variant
->jit_function
[i
])
1109 LLVMFreeMachineCodeForFunction(lp
->gallivm
->engine
,
1110 variant
->function
[i
]);
1111 LLVMDeleteFunction(variant
->function
[i
]);
1115 /* remove from shader's list */
1116 remove_from_list(&variant
->list_item_local
);
1117 variant
->shader
->variants_cached
--;
1119 /* remove from context's list */
1120 remove_from_list(&variant
->list_item_global
);
1121 lp
->nr_fs_variants
--;
1128 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1130 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1131 struct lp_fragment_shader
*shader
= fs
;
1132 struct lp_fs_variant_list_item
*li
;
1134 assert(fs
!= llvmpipe
->fs
);
1137 * XXX: we need to flush the context until we have some sort of reference
1138 * counting in fragment shaders as they may still be binned
1139 * Flushing alone might not sufficient we need to wait on it too.
1141 llvmpipe_finish(pipe
, __FUNCTION__
);
1143 /* Delete all the variants */
1144 li
= first_elem(&shader
->variants
);
1145 while(!at_end(&shader
->variants
, li
)) {
1146 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1147 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
1151 /* Delete draw module's data */
1152 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1154 assert(shader
->variants_cached
== 0);
1155 FREE((void *) shader
->base
.tokens
);
1162 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1163 uint shader
, uint index
,
1164 struct pipe_resource
*constants
)
1166 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1167 unsigned size
= constants
? constants
->width0
: 0;
1168 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1170 assert(shader
< PIPE_SHADER_TYPES
);
1171 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1173 if(llvmpipe
->constants
[shader
][index
] == constants
)
1176 draw_flush(llvmpipe
->draw
);
1178 /* note: reference counting */
1179 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1181 if(shader
== PIPE_SHADER_VERTEX
||
1182 shader
== PIPE_SHADER_GEOMETRY
) {
1183 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1187 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1192 * Return the blend factor equivalent to a destination alpha of one.
1194 static INLINE
unsigned
1195 force_dst_alpha_one(unsigned factor
)
1198 case PIPE_BLENDFACTOR_DST_ALPHA
:
1199 return PIPE_BLENDFACTOR_ONE
;
1200 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1201 return PIPE_BLENDFACTOR_ZERO
;
1202 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1203 return PIPE_BLENDFACTOR_ZERO
;
1211 * We need to generate several variants of the fragment pipeline to match
1212 * all the combinations of the contributing state atoms.
1214 * TODO: there is actually no reason to tie this to context state -- the
1215 * generated code could be cached globally in the screen.
1218 make_variant_key(struct llvmpipe_context
*lp
,
1219 struct lp_fragment_shader
*shader
,
1220 struct lp_fragment_shader_variant_key
*key
)
1224 memset(key
, 0, shader
->variant_key_size
);
1226 if (lp
->framebuffer
.zsbuf
) {
1227 if (lp
->depth_stencil
->depth
.enabled
) {
1228 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1229 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1231 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1232 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1233 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1237 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1238 if(key
->alpha
.enabled
)
1239 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1240 /* alpha.ref_value is passed in jit_context */
1242 key
->flatshade
= lp
->rasterizer
->flatshade
;
1243 if (lp
->active_query_count
) {
1244 key
->occlusion_count
= TRUE
;
1247 if (lp
->framebuffer
.nr_cbufs
) {
1248 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1251 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1252 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1253 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1254 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1255 const struct util_format_description
*format_desc
;
1257 key
->cbuf_format
[i
] = format
;
1259 format_desc
= util_format_description(format
);
1260 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1261 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1263 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1266 * Mask out color channels not present in the color buffer.
1268 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1271 * Our swizzled render tiles always have an alpha channel, but the linear
1272 * render target format often does not, so force here the dst alpha to be
1275 * This is not a mere optimization. Wrong results will be produced if the
1276 * dst alpha is used, the dst format does not have alpha, and the previous
1277 * rendering was not flushed from the swizzled to linear buffer. For
1278 * example, NonPowTwo DCT.
1280 * TODO: This should be generalized to all channels for better
1281 * performance, but only alpha causes correctness issues.
1283 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1285 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1286 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1287 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1288 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1289 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1290 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1294 /* This value will be the same for all the variants of a given shader:
1296 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1298 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1299 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1300 lp_sampler_static_state(&key
->sampler
[i
],
1301 lp
->fragment_sampler_views
[i
],
1310 * Update fragment shader state. This is called just prior to drawing
1311 * something when some fragment-related state has changed.
1314 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1316 struct lp_fragment_shader
*shader
= lp
->fs
;
1317 struct lp_fragment_shader_variant_key key
;
1318 struct lp_fragment_shader_variant
*variant
= NULL
;
1319 struct lp_fs_variant_list_item
*li
;
1321 make_variant_key(lp
, shader
, &key
);
1323 /* Search the variants for one which matches the key */
1324 li
= first_elem(&shader
->variants
);
1325 while(!at_end(&shader
->variants
, li
)) {
1326 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1334 /* Move this variant to the head of the list to implement LRU
1335 * deletion of shader's when we have too many.
1337 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1340 /* variant not found, create it now */
1344 /* First, check if we've exceeded the max number of shader variants.
1345 * If so, free 25% of them (the least recently used ones).
1347 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1348 struct pipe_context
*pipe
= &lp
->pipe
;
1351 * XXX: we need to flush the context until we have some sort of
1352 * reference counting in fragment shaders as they may still be binned
1353 * Flushing alone might not be sufficient we need to wait on it too.
1355 llvmpipe_finish(pipe
, __FUNCTION__
);
1357 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1358 struct lp_fs_variant_list_item
*item
;
1359 item
= last_elem(&lp
->fs_variants_list
);
1360 llvmpipe_remove_shader_variant(lp
, item
->base
);
1365 * Generate the new variant.
1368 variant
= generate_variant(lp
, shader
, &key
);
1371 LP_COUNT_ADD(llvm_compile_time
, dt
);
1372 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1374 llvmpipe_variant_count
++;
1376 /* Put the new variant into the list */
1378 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1379 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1380 lp
->nr_fs_variants
++;
1381 shader
->variants_cached
++;
1385 /* Bind this variant */
1386 lp_setup_set_fs_variant(lp
->setup
, variant
);
1396 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1398 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1399 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1400 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1402 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;