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
][TGSI_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
< TGSI_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
, NULL
, /* sys values array */
338 interp
->pos
, interp
->inputs
,
339 outputs
, sampler
, &shader
->info
.base
);
342 if (key
->alpha
.enabled
) {
343 int color0
= find_output_by_semantic(&shader
->info
.base
,
347 if (color0
!= -1 && outputs
[color0
][3]) {
348 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
349 LLVMValueRef alpha_ref_value
;
351 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
352 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
354 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
,
355 &mask
, alpha
, alpha_ref_value
,
356 (depth_mode
& LATE_DEPTH_TEST
) != 0);
361 if (depth_mode
& LATE_DEPTH_TEST
) {
362 int pos0
= find_output_by_semantic(&shader
->info
.base
,
363 TGSI_SEMANTIC_POSITION
,
366 if (pos0
!= -1 && outputs
[pos0
][2]) {
367 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
370 lp_build_depth_stencil_test(gallivm
,
382 if (depth_mode
& LATE_DEPTH_WRITE
) {
383 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
386 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
387 (depth_mode
& LATE_DEPTH_WRITE
))
389 /* Need to apply a reduced mask to the depth write. Reload the
390 * depth value, update from zs_value with the new mask value and
393 lp_build_deferred_depth_write(gallivm
,
403 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
405 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
&&
406 shader
->info
.base
.output_semantic_index
[attrib
] < key
->nr_cbufs
)
408 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
409 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
410 if(outputs
[attrib
][chan
]) {
411 /* XXX: just initialize outputs to point at colors[] and
414 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
415 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
416 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
423 lp_build_occlusion_count(gallivm
, type
,
424 lp_build_mask_value(&mask
), counter
);
426 *pmask
= lp_build_mask_end(&mask
);
431 * Generate color blending and color output.
432 * \param rt the render target index (to index blend, colormask state)
433 * \param type the pixel color type
434 * \param context_ptr pointer to the runtime JIT context
435 * \param mask execution mask (active fragment/pixel mask)
436 * \param src colors from the fragment shader
437 * \param dst_ptr the destination color buffer pointer
440 generate_blend(struct gallivm_state
*gallivm
,
441 const struct pipe_blend_state
*blend
,
443 LLVMBuilderRef builder
,
445 LLVMValueRef context_ptr
,
448 LLVMValueRef dst_ptr
,
451 struct lp_build_context bld
;
452 struct lp_build_mask_context mask_ctx
;
453 LLVMTypeRef vec_type
;
454 LLVMValueRef const_ptr
;
460 lp_build_context_init(&bld
, gallivm
, type
);
462 lp_build_mask_begin(&mask_ctx
, gallivm
, type
, mask
);
464 lp_build_mask_check(&mask_ctx
);
466 vec_type
= lp_build_vec_type(gallivm
, type
);
468 const_ptr
= lp_jit_context_blend_color(gallivm
, context_ptr
);
469 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
470 LLVMPointerType(vec_type
, 0), "");
472 /* load constant blend color and colors from the dest color buffer */
473 for(chan
= 0; chan
< 4; ++chan
) {
474 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
475 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
477 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
479 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
480 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
484 lp_build_blend_soa(gallivm
, blend
, type
, rt
, src
, dst
, con
, res
);
486 /* store results to color buffer */
487 for(chan
= 0; chan
< 4; ++chan
) {
488 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
489 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
490 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
491 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
492 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
496 lp_build_mask_end(&mask_ctx
);
501 * Generate the runtime callable function for the whole fragment pipeline.
502 * Note that the function which we generate operates on a block of 16
503 * pixels at at time. The block contains 2x2 quads. Each quad contains
507 generate_fragment(struct llvmpipe_context
*lp
,
508 struct lp_fragment_shader
*shader
,
509 struct lp_fragment_shader_variant
*variant
,
510 unsigned partial_mask
)
512 struct gallivm_state
*gallivm
= lp
->gallivm
;
513 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
514 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
516 struct lp_type fs_type
;
517 struct lp_type blend_type
;
518 LLVMTypeRef fs_elem_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
][TGSI_NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
540 LLVMValueRef blend_mask
;
541 LLVMValueRef function
;
543 const struct util_format_description
*zs_format_desc
;
548 boolean cbuf0_write_all
;
550 /* Adjust color input interpolation according to flatshade state:
552 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
553 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
554 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
556 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
558 inputs
[i
].interp
= LP_INTERP_LINEAR
;
562 /* check if writes to cbuf[0] are to be copied to all cbufs */
563 cbuf0_write_all
= FALSE
;
564 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
565 if (shader
->info
.base
.properties
[i
].name
==
566 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
567 cbuf0_write_all
= TRUE
;
572 /* TODO: actually pick these based on the fs and color buffer
573 * characteristics. */
575 memset(&fs_type
, 0, sizeof fs_type
);
576 fs_type
.floating
= TRUE
; /* floating point values */
577 fs_type
.sign
= TRUE
; /* values are signed */
578 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
579 fs_type
.width
= 32; /* 32-bit float */
580 fs_type
.length
= 4; /* 4 elements per vector */
581 num_fs
= 4; /* number of quads per block */
583 memset(&blend_type
, 0, sizeof blend_type
);
584 blend_type
.floating
= FALSE
; /* values are integers */
585 blend_type
.sign
= FALSE
; /* values are unsigned */
586 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
587 blend_type
.width
= 8; /* 8-bit ubyte values */
588 blend_type
.length
= 16; /* 16 elements per vector */
591 * Generate the function prototype. Any change here must be reflected in
592 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
595 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
597 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
599 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
600 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
602 arg_types
[0] = lp_jit_get_context_type(lp
); /* context */
603 arg_types
[1] = int32_type
; /* x */
604 arg_types
[2] = int32_type
; /* y */
605 arg_types
[3] = int32_type
; /* facing */
606 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
607 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
608 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
609 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
610 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
611 arg_types
[9] = int32_type
; /* mask_input */
612 arg_types
[10] = LLVMPointerType(int32_type
, 0); /* counter */
614 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
615 arg_types
, Elements(arg_types
), 0);
617 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
618 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
620 variant
->function
[partial_mask
] = function
;
622 /* XXX: need to propagate noalias down into color param now we are
623 * passing a pointer-to-pointer?
625 for(i
= 0; i
< Elements(arg_types
); ++i
)
626 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
627 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
629 context_ptr
= LLVMGetParam(function
, 0);
630 x
= LLVMGetParam(function
, 1);
631 y
= LLVMGetParam(function
, 2);
632 facing
= LLVMGetParam(function
, 3);
633 a0_ptr
= LLVMGetParam(function
, 4);
634 dadx_ptr
= LLVMGetParam(function
, 5);
635 dady_ptr
= LLVMGetParam(function
, 6);
636 color_ptr_ptr
= LLVMGetParam(function
, 7);
637 depth_ptr
= LLVMGetParam(function
, 8);
638 mask_input
= LLVMGetParam(function
, 9);
640 lp_build_name(context_ptr
, "context");
641 lp_build_name(x
, "x");
642 lp_build_name(y
, "y");
643 lp_build_name(a0_ptr
, "a0");
644 lp_build_name(dadx_ptr
, "dadx");
645 lp_build_name(dady_ptr
, "dady");
646 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
647 lp_build_name(depth_ptr
, "depth");
648 lp_build_name(mask_input
, "mask_input");
650 if (key
->occlusion_count
) {
651 counter
= LLVMGetParam(function
, 10);
652 lp_build_name(counter
, "counter");
659 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
660 builder
= gallivm
->builder
;
662 LLVMPositionBuilderAtEnd(builder
, block
);
665 * The shader input interpolation info is not explicitely baked in the
666 * shader key, but everything it derives from (TGSI, and flatshade) is
667 * already included in the shader key.
669 lp_build_interp_soa_init(&interp
,
671 shader
->info
.base
.num_inputs
,
674 a0_ptr
, dadx_ptr
, dady_ptr
,
677 /* code generated texture sampling */
678 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
680 /* loop over quads in the block */
681 zs_format_desc
= util_format_description(key
->zsbuf_format
);
683 for(i
= 0; i
< num_fs
; ++i
) {
684 LLVMValueRef depth_offset
= LLVMConstInt(int32_type
,
685 i
*fs_type
.length
*zs_format_desc
->block
.bits
/8,
687 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
688 LLVMValueRef depth_ptr_i
;
690 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
700 &fs_mask
[i
], /* output */
708 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
709 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
710 fs_out_color
[cbuf
][chan
][i
] =
711 out_color
[cbuf
* !cbuf0_write_all
][chan
];
714 sampler
->destroy(sampler
);
716 /* Loop over color outputs / color buffers to do blending.
718 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
719 LLVMValueRef color_ptr
;
720 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
721 LLVMValueRef blend_in_color
[TGSI_NUM_CHANNELS
];
725 * Convert the fs's output color and mask to fit to the blending type.
727 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
728 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
730 for (i
= 0; i
< num_fs
; i
++) {
732 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
735 lp_build_conv(gallivm
, fs_type
, blend_type
,
738 &blend_in_color
[chan
], 1);
740 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
743 if (partial_mask
|| !variant
->opaque
) {
744 lp_build_conv_mask(lp
->gallivm
, fs_type
, blend_type
,
748 blend_mask
= lp_build_const_int_vec(lp
->gallivm
, blend_type
, ~0);
751 color_ptr
= LLVMBuildLoad(builder
,
752 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
754 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
756 /* which blend/colormask state to use */
757 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
763 /* Could the 4x4 have been killed?
765 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
766 !key
->alpha
.enabled
&&
767 !shader
->info
.base
.uses_kill
);
769 generate_blend(lp
->gallivm
,
782 LLVMBuildRetVoid(builder
);
784 /* Verify the LLVM IR. If invalid, dump and abort */
786 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
788 lp_debug_dump_value(function
);
793 /* Apply optimizations to LLVM IR */
794 LLVMRunFunctionPassManager(gallivm
->passmgr
, function
);
796 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
797 /* Print the LLVM IR to stderr */
798 lp_debug_dump_value(function
);
802 /* Dump byte code to a file */
804 LLVMWriteBitcodeToFile(gallivm
->module
, "llvmpipe.bc");
807 variant
->nr_instrs
+= lp_build_count_instructions(function
);
809 * Translate the LLVM IR into machine code.
812 void *f
= LLVMGetPointerToGlobal(gallivm
->engine
, function
);
814 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
816 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
819 lp_func_delete_body(function
);
825 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
829 debug_printf("fs variant %p:\n", (void *) key
);
831 if (key
->flatshade
) {
832 debug_printf("flatshade = 1\n");
834 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
835 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
837 if (key
->depth
.enabled
) {
838 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
839 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
840 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
843 for (i
= 0; i
< 2; ++i
) {
844 if (key
->stencil
[i
].enabled
) {
845 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
846 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
847 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
848 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
849 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
850 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
854 if (key
->alpha
.enabled
) {
855 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
858 if (key
->occlusion_count
) {
859 debug_printf("occlusion_count = 1\n");
862 if (key
->blend
.logicop_enable
) {
863 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
865 else if (key
->blend
.rt
[0].blend_enable
) {
866 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
867 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
868 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
869 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
870 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
871 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
873 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
874 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
875 debug_printf("sampler[%u] = \n", i
);
876 debug_printf(" .format = %s\n",
877 util_format_name(key
->sampler
[i
].format
));
878 debug_printf(" .target = %s\n",
879 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
880 debug_printf(" .pot = %u %u %u\n",
881 key
->sampler
[i
].pot_width
,
882 key
->sampler
[i
].pot_height
,
883 key
->sampler
[i
].pot_depth
);
884 debug_printf(" .wrap = %s %s %s\n",
885 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
886 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
887 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
888 debug_printf(" .min_img_filter = %s\n",
889 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
890 debug_printf(" .min_mip_filter = %s\n",
891 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
892 debug_printf(" .mag_img_filter = %s\n",
893 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
894 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
895 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
896 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
897 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
898 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
899 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
900 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
906 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
908 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
909 variant
->shader
->no
, variant
->no
);
910 tgsi_dump(variant
->shader
->base
.tokens
, 0);
911 dump_fs_variant_key(&variant
->key
);
912 debug_printf("variant->opaque = %u\n", variant
->opaque
);
918 * Generate a new fragment shader variant from the shader code and
919 * other state indicated by the key.
921 static struct lp_fragment_shader_variant
*
922 generate_variant(struct llvmpipe_context
*lp
,
923 struct lp_fragment_shader
*shader
,
924 const struct lp_fragment_shader_variant_key
*key
)
926 struct lp_fragment_shader_variant
*variant
;
927 boolean fullcolormask
;
929 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
933 variant
->shader
= shader
;
934 variant
->list_item_global
.base
= variant
;
935 variant
->list_item_local
.base
= variant
;
936 variant
->no
= shader
->variants_created
++;
938 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
941 * Determine whether we are touching all channels in the color buffer.
943 fullcolormask
= FALSE
;
944 if (key
->nr_cbufs
== 1) {
945 const struct util_format_description
*format_desc
;
946 format_desc
= util_format_description(key
->cbuf_format
[0]);
947 if ((~key
->blend
.rt
[0].colormask
&
948 util_format_colormask(format_desc
)) == 0) {
949 fullcolormask
= TRUE
;
954 !key
->blend
.logicop_enable
&&
955 !key
->blend
.rt
[0].blend_enable
&&
957 !key
->stencil
[0].enabled
&&
958 !key
->alpha
.enabled
&&
959 !key
->depth
.enabled
&&
960 !shader
->info
.base
.uses_kill
964 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
965 lp_debug_fs_variant(variant
);
968 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
970 if (variant
->opaque
) {
971 /* Specialized shader, which doesn't need to read the color buffer. */
972 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
974 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
982 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
983 const struct pipe_shader_state
*templ
)
985 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
986 struct lp_fragment_shader
*shader
;
990 shader
= CALLOC_STRUCT(lp_fragment_shader
);
994 shader
->no
= fs_no
++;
995 make_empty_list(&shader
->variants
);
997 /* get/save the summary info for this shader */
998 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
1000 /* we need to keep a local copy of the tokens */
1001 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
1003 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
1004 if (shader
->draw_data
== NULL
) {
1005 FREE((void *) shader
->base
.tokens
);
1010 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1012 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
1013 sampler
[nr_samplers
]);
1015 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1016 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
1018 switch (shader
->info
.base
.input_interpolate
[i
]) {
1019 case TGSI_INTERPOLATE_CONSTANT
:
1020 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1022 case TGSI_INTERPOLATE_LINEAR
:
1023 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
1025 case TGSI_INTERPOLATE_PERSPECTIVE
:
1026 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1028 case TGSI_INTERPOLATE_COLOR
:
1029 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
1036 switch (shader
->info
.base
.input_semantic_name
[i
]) {
1037 case TGSI_SEMANTIC_FACE
:
1038 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
1040 case TGSI_SEMANTIC_POSITION
:
1041 /* Position was already emitted above
1043 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
1044 shader
->inputs
[i
].src_index
= 0;
1048 shader
->inputs
[i
].src_index
= i
+1;
1051 if (LP_DEBUG
& DEBUG_TGSI
) {
1053 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
1054 shader
->no
, (void *) shader
);
1055 tgsi_dump(templ
->tokens
, 0);
1056 debug_printf("usage masks:\n");
1057 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
1058 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
1059 debug_printf(" IN[%u].%s%s%s%s\n",
1061 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
1062 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
1063 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
1064 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
1074 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1076 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1078 if (llvmpipe
->fs
== fs
)
1081 draw_flush(llvmpipe
->draw
);
1083 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
1085 draw_bind_fragment_shader(llvmpipe
->draw
,
1086 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
1088 llvmpipe
->dirty
|= LP_NEW_FS
;
1093 * Remove shader variant from two lists: the shader's variant list
1094 * and the context's variant list.
1097 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
1098 struct lp_fragment_shader_variant
*variant
)
1102 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
1103 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
1104 " #%u v total cached #%u\n",
1105 variant
->shader
->no
,
1107 variant
->shader
->variants_created
,
1108 variant
->shader
->variants_cached
,
1109 lp
->nr_fs_variants
);
1112 /* free all the variant's JIT'd functions */
1113 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1114 if (variant
->function
[i
]) {
1115 if (variant
->jit_function
[i
])
1116 LLVMFreeMachineCodeForFunction(lp
->gallivm
->engine
,
1117 variant
->function
[i
]);
1118 LLVMDeleteFunction(variant
->function
[i
]);
1122 /* remove from shader's list */
1123 remove_from_list(&variant
->list_item_local
);
1124 variant
->shader
->variants_cached
--;
1126 /* remove from context's list */
1127 remove_from_list(&variant
->list_item_global
);
1128 lp
->nr_fs_variants
--;
1129 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
1136 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1138 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1139 struct lp_fragment_shader
*shader
= fs
;
1140 struct lp_fs_variant_list_item
*li
;
1142 assert(fs
!= llvmpipe
->fs
);
1145 * XXX: we need to flush the context until we have some sort of reference
1146 * counting in fragment shaders as they may still be binned
1147 * Flushing alone might not sufficient we need to wait on it too.
1149 llvmpipe_finish(pipe
, __FUNCTION__
);
1151 /* Delete all the variants */
1152 li
= first_elem(&shader
->variants
);
1153 while(!at_end(&shader
->variants
, li
)) {
1154 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1155 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
1159 /* Delete draw module's data */
1160 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1162 assert(shader
->variants_cached
== 0);
1163 FREE((void *) shader
->base
.tokens
);
1170 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1171 uint shader
, uint index
,
1172 struct pipe_resource
*constants
)
1174 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1175 unsigned size
= constants
? constants
->width0
: 0;
1176 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1178 assert(shader
< PIPE_SHADER_TYPES
);
1179 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1181 if(llvmpipe
->constants
[shader
][index
] == constants
)
1184 draw_flush(llvmpipe
->draw
);
1186 /* note: reference counting */
1187 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1189 if(shader
== PIPE_SHADER_VERTEX
||
1190 shader
== PIPE_SHADER_GEOMETRY
) {
1191 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1195 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1200 * Return the blend factor equivalent to a destination alpha of one.
1202 static INLINE
unsigned
1203 force_dst_alpha_one(unsigned factor
)
1206 case PIPE_BLENDFACTOR_DST_ALPHA
:
1207 return PIPE_BLENDFACTOR_ONE
;
1208 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1209 return PIPE_BLENDFACTOR_ZERO
;
1210 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1211 return PIPE_BLENDFACTOR_ZERO
;
1219 * We need to generate several variants of the fragment pipeline to match
1220 * all the combinations of the contributing state atoms.
1222 * TODO: there is actually no reason to tie this to context state -- the
1223 * generated code could be cached globally in the screen.
1226 make_variant_key(struct llvmpipe_context
*lp
,
1227 struct lp_fragment_shader
*shader
,
1228 struct lp_fragment_shader_variant_key
*key
)
1232 memset(key
, 0, shader
->variant_key_size
);
1234 if (lp
->framebuffer
.zsbuf
) {
1235 if (lp
->depth_stencil
->depth
.enabled
) {
1236 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1237 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1239 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1240 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1241 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1245 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1246 if(key
->alpha
.enabled
)
1247 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1248 /* alpha.ref_value is passed in jit_context */
1250 key
->flatshade
= lp
->rasterizer
->flatshade
;
1251 if (lp
->active_query_count
) {
1252 key
->occlusion_count
= TRUE
;
1255 if (lp
->framebuffer
.nr_cbufs
) {
1256 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1259 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1260 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1261 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1262 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1263 const struct util_format_description
*format_desc
;
1265 key
->cbuf_format
[i
] = format
;
1267 format_desc
= util_format_description(format
);
1268 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1269 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1271 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1274 * Mask out color channels not present in the color buffer.
1276 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1279 * Our swizzled render tiles always have an alpha channel, but the linear
1280 * render target format often does not, so force here the dst alpha to be
1283 * This is not a mere optimization. Wrong results will be produced if the
1284 * dst alpha is used, the dst format does not have alpha, and the previous
1285 * rendering was not flushed from the swizzled to linear buffer. For
1286 * example, NonPowTwo DCT.
1288 * TODO: This should be generalized to all channels for better
1289 * performance, but only alpha causes correctness issues.
1291 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1293 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
1294 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
1295 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1296 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1297 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1298 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1299 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1303 /* This value will be the same for all the variants of a given shader:
1305 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1307 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1308 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1309 lp_sampler_static_state(&key
->sampler
[i
],
1310 lp
->fragment_sampler_views
[i
],
1319 * Update fragment shader state. This is called just prior to drawing
1320 * something when some fragment-related state has changed.
1323 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1325 struct lp_fragment_shader
*shader
= lp
->fs
;
1326 struct lp_fragment_shader_variant_key key
;
1327 struct lp_fragment_shader_variant
*variant
= NULL
;
1328 struct lp_fs_variant_list_item
*li
;
1330 make_variant_key(lp
, shader
, &key
);
1332 /* Search the variants for one which matches the key */
1333 li
= first_elem(&shader
->variants
);
1334 while(!at_end(&shader
->variants
, li
)) {
1335 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1343 /* Move this variant to the head of the list to implement LRU
1344 * deletion of shader's when we have too many.
1346 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1349 /* variant not found, create it now */
1352 unsigned variants_to_cull
;
1355 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
1358 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
1361 /* First, check if we've exceeded the max number of shader variants.
1362 * If so, free 25% of them (the least recently used ones).
1364 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
1366 if (variants_to_cull
||
1367 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
1368 struct pipe_context
*pipe
= &lp
->pipe
;
1371 * XXX: we need to flush the context until we have some sort of
1372 * reference counting in fragment shaders as they may still be binned
1373 * Flushing alone might not be sufficient we need to wait on it too.
1375 llvmpipe_finish(pipe
, __FUNCTION__
);
1378 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
1379 * number of shader variants (potentially all of them) could be
1380 * pending for destruction on flush.
1383 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
1384 struct lp_fs_variant_list_item
*item
;
1385 if (is_empty_list(&lp
->fs_variants_list
)) {
1388 item
= last_elem(&lp
->fs_variants_list
);
1391 llvmpipe_remove_shader_variant(lp
, item
->base
);
1396 * Generate the new variant.
1399 variant
= generate_variant(lp
, shader
, &key
);
1402 LP_COUNT_ADD(llvm_compile_time
, dt
);
1403 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1405 llvmpipe_variant_count
++;
1407 /* Put the new variant into the list */
1409 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1410 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1411 lp
->nr_fs_variants
++;
1412 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
1413 shader
->variants_cached
++;
1417 /* Bind this variant */
1418 lp_setup_set_fs_variant(lp
->setup
, variant
);
1428 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1430 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1431 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1432 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1434 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;