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 /** Fragment shader number (for debugging) */
101 static unsigned fs_no
= 0;
105 * Expand the relevant bits of mask_input to a n*4-dword mask for the
106 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
107 * quad mask vector to 0 or ~0.
108 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
109 * quad arguments with fs length 8.
111 * \param first_quad which quad(s) of the quad group to test, in [0,3]
112 * \param mask_input bitwise mask for the whole 4x4 stamp
115 generate_quad_mask(struct gallivm_state
*gallivm
,
116 struct lp_type fs_type
,
118 LLVMValueRef mask_input
) /* int32 */
120 LLVMBuilderRef builder
= gallivm
->builder
;
121 struct lp_type mask_type
;
122 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
123 LLVMValueRef bits
[16];
128 * XXX: We'll need a different path for 16 x u8
130 assert(fs_type
.width
== 32);
131 assert(fs_type
.length
<= Elements(bits
));
132 mask_type
= lp_int_type(fs_type
);
135 * mask_input >>= (quad * 4)
137 switch (first_quad
) {
142 assert(fs_type
.length
== 4);
149 assert(fs_type
.length
== 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 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
170 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
171 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
172 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
173 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
174 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
176 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
179 * mask = mask != 0 ? ~0 : 0
181 mask
= lp_build_compare(gallivm
,
182 mask_type
, PIPE_FUNC_NOTEQUAL
,
184 lp_build_const_int_vec(gallivm
, mask_type
, 0));
190 #define EARLY_DEPTH_TEST 0x1
191 #define LATE_DEPTH_TEST 0x2
192 #define EARLY_DEPTH_WRITE 0x4
193 #define LATE_DEPTH_WRITE 0x8
196 find_output_by_semantic( const struct tgsi_shader_info
*info
,
202 for (i
= 0; i
< info
->num_outputs
; i
++)
203 if (info
->output_semantic_name
[i
] == semantic
&&
204 info
->output_semantic_index
[i
] == index
)
212 * Generate the fragment shader, depth/stencil test, and alpha tests.
213 * \param i which quad in the tile, in range [0,3]
214 * \param partial_mask if 1, do mask_input testing
217 generate_fs(struct gallivm_state
*gallivm
,
218 struct lp_fragment_shader
*shader
,
219 const struct lp_fragment_shader_variant_key
*key
,
220 LLVMBuilderRef builder
,
222 LLVMValueRef context_ptr
,
224 struct lp_build_interp_soa_context
*interp
,
225 struct lp_build_sampler_soa
*sampler
,
227 LLVMValueRef (*color
)[4],
228 LLVMValueRef depth_ptr
,
230 unsigned partial_mask
,
231 LLVMValueRef mask_input
,
232 LLVMValueRef counter
)
234 const struct util_format_description
*zs_format_desc
= NULL
;
235 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
236 LLVMTypeRef vec_type
;
237 LLVMValueRef consts_ptr
;
238 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
240 LLVMValueRef zs_value
= NULL
;
241 LLVMValueRef stencil_refs
[2];
242 struct lp_build_mask_context mask
;
243 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
244 shader
->info
.base
.num_inputs
< 3 &&
245 shader
->info
.base
.num_instructions
< 8);
250 struct lp_bld_tgsi_system_values system_values
;
252 memset(&system_values
, 0, sizeof(system_values
));
254 if (key
->depth
.enabled
||
255 key
->stencil
[0].enabled
||
256 key
->stencil
[1].enabled
) {
258 zs_format_desc
= util_format_description(key
->zsbuf_format
);
259 assert(zs_format_desc
);
261 if (!shader
->info
.base
.writes_z
) {
262 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
263 /* With alpha test and kill, can do the depth test early
264 * and hopefully eliminate some quads. But need to do a
265 * special deferred depth write once the final mask value
268 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
270 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
273 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
276 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
277 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
278 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
286 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
287 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
289 vec_type
= lp_build_vec_type(gallivm
, type
);
291 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
293 memset(outputs
, 0, sizeof outputs
);
295 /* Declare the color and z variables */
296 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
297 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
298 color
[cbuf
][chan
] = lp_build_alloca(gallivm
, vec_type
, "color");
302 /* do triangle edge testing */
304 *pmask
= generate_quad_mask(gallivm
, type
,
305 i
*type
.length
/4, mask_input
);
308 *pmask
= lp_build_const_int_vec(gallivm
, type
, ~0);
311 /* 'mask' will control execution based on quad's pixel alive/killed state */
312 lp_build_mask_begin(&mask
, gallivm
, type
, *pmask
);
314 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
315 lp_build_mask_check(&mask
);
317 lp_build_interp_soa_update_pos(interp
, gallivm
, i
*type
.length
/4);
320 if (depth_mode
& EARLY_DEPTH_TEST
) {
321 lp_build_depth_stencil_test(gallivm
,
333 if (depth_mode
& EARLY_DEPTH_WRITE
) {
334 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
338 lp_build_interp_soa_update_inputs(interp
, gallivm
, i
*type
.length
/4);
340 /* Build the actual shader */
341 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
342 consts_ptr
, &system_values
,
343 interp
->pos
, interp
->inputs
,
344 outputs
, sampler
, &shader
->info
.base
);
347 if (key
->alpha
.enabled
) {
348 int color0
= find_output_by_semantic(&shader
->info
.base
,
352 if (color0
!= -1 && outputs
[color0
][3]) {
353 const struct util_format_description
*cbuf_format_desc
;
354 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
355 LLVMValueRef alpha_ref_value
;
357 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
358 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
360 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
362 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
363 &mask
, alpha
, alpha_ref_value
,
364 (depth_mode
& LATE_DEPTH_TEST
) != 0);
369 if (depth_mode
& LATE_DEPTH_TEST
) {
370 int pos0
= find_output_by_semantic(&shader
->info
.base
,
371 TGSI_SEMANTIC_POSITION
,
374 if (pos0
!= -1 && outputs
[pos0
][2]) {
375 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
378 lp_build_depth_stencil_test(gallivm
,
390 if (depth_mode
& LATE_DEPTH_WRITE
) {
391 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
394 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
395 (depth_mode
& LATE_DEPTH_WRITE
))
397 /* Need to apply a reduced mask to the depth write. Reload the
398 * depth value, update from zs_value with the new mask value and
401 lp_build_deferred_depth_write(gallivm
,
411 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
413 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
&&
414 shader
->info
.base
.output_semantic_index
[attrib
] < key
->nr_cbufs
)
416 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
417 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
418 if(outputs
[attrib
][chan
]) {
419 /* XXX: just initialize outputs to point at colors[] and
422 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
423 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
424 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
431 lp_build_occlusion_count(gallivm
, type
,
432 lp_build_mask_value(&mask
), counter
);
434 *pmask
= lp_build_mask_end(&mask
);
439 * Generate color blending and color output.
440 * \param rt the render target index (to index blend, colormask state)
441 * \param type the pixel color type
442 * \param context_ptr pointer to the runtime JIT context
443 * \param mask execution mask (active fragment/pixel mask)
444 * \param src colors from the fragment shader
445 * \param dst_ptr the destination color buffer pointer
448 generate_blend(struct gallivm_state
*gallivm
,
449 const struct pipe_blend_state
*blend
,
451 LLVMBuilderRef builder
,
453 LLVMValueRef context_ptr
,
456 LLVMValueRef dst_ptr
,
459 struct lp_build_context bld
;
460 struct lp_build_mask_context mask_ctx
;
461 LLVMTypeRef vec_type
;
462 LLVMValueRef const_ptr
;
468 lp_build_context_init(&bld
, gallivm
, type
);
470 lp_build_mask_begin(&mask_ctx
, gallivm
, type
, mask
);
472 lp_build_mask_check(&mask_ctx
);
474 vec_type
= lp_build_vec_type(gallivm
, type
);
476 const_ptr
= lp_jit_context_blend_color(gallivm
, context_ptr
);
477 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
478 LLVMPointerType(vec_type
, 0), "");
480 /* load constant blend color and colors from the dest color buffer */
481 for(chan
= 0; chan
< 4; ++chan
) {
482 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
483 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
485 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
487 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
488 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
492 lp_build_blend_soa(gallivm
, blend
, type
, rt
, src
, dst
, con
, res
);
494 /* store results to color buffer */
495 for(chan
= 0; chan
< 4; ++chan
) {
496 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
497 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
498 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
499 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
500 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
504 lp_build_mask_end(&mask_ctx
);
509 * Generate the runtime callable function for the whole fragment pipeline.
510 * Note that the function which we generate operates on a block of 16
511 * pixels at at time. The block contains 2x2 quads. Each quad contains
515 generate_fragment(struct llvmpipe_context
*lp
,
516 struct lp_fragment_shader
*shader
,
517 struct lp_fragment_shader_variant
*variant
,
518 unsigned partial_mask
)
520 struct gallivm_state
*gallivm
= variant
->gallivm
;
521 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
522 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
524 struct lp_type fs_type
;
525 struct lp_type blend_type
;
526 LLVMTypeRef fs_elem_type
;
527 LLVMTypeRef blend_vec_type
;
528 LLVMTypeRef arg_types
[11];
529 LLVMTypeRef func_type
;
530 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
531 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
532 LLVMValueRef context_ptr
;
536 LLVMValueRef dadx_ptr
;
537 LLVMValueRef dady_ptr
;
538 LLVMValueRef color_ptr_ptr
;
539 LLVMValueRef depth_ptr
;
540 LLVMValueRef mask_input
;
541 LLVMValueRef counter
= NULL
;
542 LLVMBasicBlockRef block
;
543 LLVMBuilderRef builder
;
544 struct lp_build_sampler_soa
*sampler
;
545 struct lp_build_interp_soa_context interp
;
546 LLVMValueRef fs_mask
[16 / 4];
547 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
548 LLVMValueRef blend_mask
;
549 LLVMValueRef function
;
551 const struct util_format_description
*zs_format_desc
;
556 boolean cbuf0_write_all
;
558 assert(lp_native_vector_width
/ 32 >= 4);
560 /* Adjust color input interpolation according to flatshade state:
562 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
563 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
564 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
566 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
568 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
572 /* check if writes to cbuf[0] are to be copied to all cbufs */
573 cbuf0_write_all
= FALSE
;
574 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
575 if (shader
->info
.base
.properties
[i
].name
==
576 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
577 cbuf0_write_all
= TRUE
;
582 /* TODO: actually pick these based on the fs and color buffer
583 * characteristics. */
585 memset(&fs_type
, 0, sizeof fs_type
);
586 fs_type
.floating
= TRUE
; /* floating point values */
587 fs_type
.sign
= TRUE
; /* values are signed */
588 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
589 fs_type
.width
= 32; /* 32-bit float */
590 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
591 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
593 memset(&blend_type
, 0, sizeof blend_type
);
594 blend_type
.floating
= FALSE
; /* values are integers */
595 blend_type
.sign
= FALSE
; /* values are unsigned */
596 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
597 blend_type
.width
= 8; /* 8-bit ubyte values */
598 blend_type
.length
= 16; /* 16 elements per vector */
601 * Generate the function prototype. Any change here must be reflected in
602 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
605 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
607 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
609 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
610 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
612 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
613 arg_types
[1] = int32_type
; /* x */
614 arg_types
[2] = int32_type
; /* y */
615 arg_types
[3] = int32_type
; /* facing */
616 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
617 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
618 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
619 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
620 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
621 arg_types
[9] = int32_type
; /* mask_input */
622 arg_types
[10] = LLVMPointerType(int32_type
, 0); /* counter */
624 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
625 arg_types
, Elements(arg_types
), 0);
627 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
628 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
630 variant
->function
[partial_mask
] = function
;
632 /* XXX: need to propagate noalias down into color param now we are
633 * passing a pointer-to-pointer?
635 for(i
= 0; i
< Elements(arg_types
); ++i
)
636 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
637 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
639 context_ptr
= LLVMGetParam(function
, 0);
640 x
= LLVMGetParam(function
, 1);
641 y
= LLVMGetParam(function
, 2);
642 facing
= LLVMGetParam(function
, 3);
643 a0_ptr
= LLVMGetParam(function
, 4);
644 dadx_ptr
= LLVMGetParam(function
, 5);
645 dady_ptr
= LLVMGetParam(function
, 6);
646 color_ptr_ptr
= LLVMGetParam(function
, 7);
647 depth_ptr
= LLVMGetParam(function
, 8);
648 mask_input
= LLVMGetParam(function
, 9);
650 lp_build_name(context_ptr
, "context");
651 lp_build_name(x
, "x");
652 lp_build_name(y
, "y");
653 lp_build_name(a0_ptr
, "a0");
654 lp_build_name(dadx_ptr
, "dadx");
655 lp_build_name(dady_ptr
, "dady");
656 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
657 lp_build_name(depth_ptr
, "depth");
658 lp_build_name(mask_input
, "mask_input");
660 if (key
->occlusion_count
) {
661 counter
= LLVMGetParam(function
, 10);
662 lp_build_name(counter
, "counter");
669 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
670 builder
= gallivm
->builder
;
672 LLVMPositionBuilderAtEnd(builder
, block
);
675 * The shader input interpolation info is not explicitely baked in the
676 * shader key, but everything it derives from (TGSI, and flatshade) is
677 * already included in the shader key.
679 lp_build_interp_soa_init(&interp
,
681 shader
->info
.base
.num_inputs
,
684 a0_ptr
, dadx_ptr
, dady_ptr
,
687 /* code generated texture sampling */
688 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
690 /* loop over quads in the block */
691 zs_format_desc
= util_format_description(key
->zsbuf_format
);
693 for(i
= 0; i
< num_fs
; ++i
) {
694 LLVMValueRef depth_offset
= LLVMConstInt(int32_type
,
695 i
*fs_type
.length
*zs_format_desc
->block
.bits
/8,
697 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
698 LLVMValueRef depth_ptr_i
;
700 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
710 &fs_mask
[i
], /* output */
718 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
719 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
720 fs_out_color
[cbuf
][chan
][i
] =
721 out_color
[cbuf
* !cbuf0_write_all
][chan
];
724 sampler
->destroy(sampler
);
726 /* Loop over color outputs / color buffers to do blending.
728 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
729 LLVMValueRef color_ptr
;
730 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
731 LLVMValueRef blend_in_color
[TGSI_NUM_CHANNELS
];
735 * Convert the fs's output color and mask to fit to the blending type.
737 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
738 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
740 for (i
= 0; i
< num_fs
; i
++) {
742 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
745 lp_build_conv(gallivm
, fs_type
, blend_type
,
748 &blend_in_color
[chan
], 1);
750 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
753 if (partial_mask
|| !variant
->opaque
) {
754 lp_build_conv_mask(variant
->gallivm
, fs_type
, blend_type
,
758 blend_mask
= lp_build_const_int_vec(variant
->gallivm
, blend_type
, ~0);
761 color_ptr
= LLVMBuildLoad(builder
,
762 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
764 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
766 /* which blend/colormask state to use */
767 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
773 /* Could the 4x4 have been killed?
775 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
776 !key
->alpha
.enabled
&&
777 !shader
->info
.base
.uses_kill
);
779 generate_blend(variant
->gallivm
,
792 LLVMBuildRetVoid(builder
);
794 gallivm_verify_function(gallivm
, function
);
796 variant
->nr_instrs
+= lp_build_count_instructions(function
);
801 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
805 debug_printf("fs variant %p:\n", (void *) key
);
807 if (key
->flatshade
) {
808 debug_printf("flatshade = 1\n");
810 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
811 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
813 if (key
->depth
.enabled
) {
814 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
815 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
816 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
819 for (i
= 0; i
< 2; ++i
) {
820 if (key
->stencil
[i
].enabled
) {
821 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
822 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
823 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
824 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
825 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
826 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
830 if (key
->alpha
.enabled
) {
831 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
834 if (key
->occlusion_count
) {
835 debug_printf("occlusion_count = 1\n");
838 if (key
->blend
.logicop_enable
) {
839 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
841 else if (key
->blend
.rt
[0].blend_enable
) {
842 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
843 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
844 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
845 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
846 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
847 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
849 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
850 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
851 debug_printf("sampler[%u] = \n", i
);
852 debug_printf(" .format = %s\n",
853 util_format_name(key
->sampler
[i
].format
));
854 debug_printf(" .target = %s\n",
855 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
856 debug_printf(" .pot = %u %u %u\n",
857 key
->sampler
[i
].pot_width
,
858 key
->sampler
[i
].pot_height
,
859 key
->sampler
[i
].pot_depth
);
860 debug_printf(" .wrap = %s %s %s\n",
861 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
862 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
863 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
864 debug_printf(" .min_img_filter = %s\n",
865 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
866 debug_printf(" .min_mip_filter = %s\n",
867 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
868 debug_printf(" .mag_img_filter = %s\n",
869 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
870 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
871 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
872 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
873 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
874 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
875 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
876 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
882 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
884 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
885 variant
->shader
->no
, variant
->no
);
886 tgsi_dump(variant
->shader
->base
.tokens
, 0);
887 dump_fs_variant_key(&variant
->key
);
888 debug_printf("variant->opaque = %u\n", variant
->opaque
);
894 * Generate a new fragment shader variant from the shader code and
895 * other state indicated by the key.
897 static struct lp_fragment_shader_variant
*
898 generate_variant(struct llvmpipe_context
*lp
,
899 struct lp_fragment_shader
*shader
,
900 const struct lp_fragment_shader_variant_key
*key
)
902 struct lp_fragment_shader_variant
*variant
;
903 const struct util_format_description
*cbuf0_format_desc
;
904 boolean fullcolormask
;
906 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
910 variant
->gallivm
= gallivm_create();
911 if (!variant
->gallivm
) {
916 variant
->shader
= shader
;
917 variant
->list_item_global
.base
= variant
;
918 variant
->list_item_local
.base
= variant
;
919 variant
->no
= shader
->variants_created
++;
921 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
924 * Determine whether we are touching all channels in the color buffer.
926 fullcolormask
= FALSE
;
927 if (key
->nr_cbufs
== 1) {
928 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
929 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
933 !key
->blend
.logicop_enable
&&
934 !key
->blend
.rt
[0].blend_enable
&&
936 !key
->stencil
[0].enabled
&&
937 !key
->alpha
.enabled
&&
938 !key
->depth
.enabled
&&
939 !shader
->info
.base
.uses_kill
943 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
944 lp_debug_fs_variant(variant
);
947 lp_jit_init_types(variant
);
949 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
950 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
952 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
953 if (variant
->opaque
) {
954 /* Specialized shader, which doesn't need to read the color buffer. */
955 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
963 gallivm_compile_module(variant
->gallivm
);
965 if (variant
->function
[RAST_EDGE_TEST
]) {
966 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
967 gallivm_jit_function(variant
->gallivm
,
968 variant
->function
[RAST_EDGE_TEST
]);
971 if (variant
->function
[RAST_WHOLE
]) {
972 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
973 gallivm_jit_function(variant
->gallivm
,
974 variant
->function
[RAST_WHOLE
]);
975 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
976 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
984 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
985 const struct pipe_shader_state
*templ
)
987 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
988 struct lp_fragment_shader
*shader
;
992 shader
= CALLOC_STRUCT(lp_fragment_shader
);
996 shader
->no
= fs_no
++;
997 make_empty_list(&shader
->variants
);
999 /* get/save the summary info for this shader */
1000 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
1002 /* we need to keep a local copy of the tokens */
1003 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
1005 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
1006 if (shader
->draw_data
== NULL
) {
1007 FREE((void *) shader
->base
.tokens
);
1012 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1014 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
1015 sampler
[nr_samplers
]);
1017 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1018 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
1019 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
1021 switch (shader
->info
.base
.input_interpolate
[i
]) {
1022 case TGSI_INTERPOLATE_CONSTANT
:
1023 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1025 case TGSI_INTERPOLATE_LINEAR
:
1026 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
1028 case TGSI_INTERPOLATE_PERSPECTIVE
:
1029 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1031 case TGSI_INTERPOLATE_COLOR
:
1032 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
1039 switch (shader
->info
.base
.input_semantic_name
[i
]) {
1040 case TGSI_SEMANTIC_FACE
:
1041 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
1043 case TGSI_SEMANTIC_POSITION
:
1044 /* Position was already emitted above
1046 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
1047 shader
->inputs
[i
].src_index
= 0;
1051 shader
->inputs
[i
].src_index
= i
+1;
1054 if (LP_DEBUG
& DEBUG_TGSI
) {
1056 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
1057 shader
->no
, (void *) shader
);
1058 tgsi_dump(templ
->tokens
, 0);
1059 debug_printf("usage masks:\n");
1060 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
1061 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
1062 debug_printf(" IN[%u].%s%s%s%s\n",
1064 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
1065 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
1066 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
1067 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
1077 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1079 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1081 if (llvmpipe
->fs
== fs
)
1084 draw_flush(llvmpipe
->draw
);
1086 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
1088 draw_bind_fragment_shader(llvmpipe
->draw
,
1089 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
1091 llvmpipe
->dirty
|= LP_NEW_FS
;
1096 * Remove shader variant from two lists: the shader's variant list
1097 * and the context's variant list.
1100 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
1101 struct lp_fragment_shader_variant
*variant
)
1105 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
1106 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
1107 " #%u v total cached #%u\n",
1108 variant
->shader
->no
,
1110 variant
->shader
->variants_created
,
1111 variant
->shader
->variants_cached
,
1112 lp
->nr_fs_variants
);
1115 /* free all the variant's JIT'd functions */
1116 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1117 if (variant
->function
[i
]) {
1118 gallivm_free_function(variant
->gallivm
,
1119 variant
->function
[i
],
1120 variant
->jit_function
[i
]);
1124 gallivm_destroy(variant
->gallivm
);
1126 /* remove from shader's list */
1127 remove_from_list(&variant
->list_item_local
);
1128 variant
->shader
->variants_cached
--;
1130 /* remove from context's list */
1131 remove_from_list(&variant
->list_item_global
);
1132 lp
->nr_fs_variants
--;
1133 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
1140 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1142 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1143 struct lp_fragment_shader
*shader
= fs
;
1144 struct lp_fs_variant_list_item
*li
;
1146 assert(fs
!= llvmpipe
->fs
);
1149 * XXX: we need to flush the context until we have some sort of reference
1150 * counting in fragment shaders as they may still be binned
1151 * Flushing alone might not sufficient we need to wait on it too.
1153 llvmpipe_finish(pipe
, __FUNCTION__
);
1155 /* Delete all the variants */
1156 li
= first_elem(&shader
->variants
);
1157 while(!at_end(&shader
->variants
, li
)) {
1158 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1159 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
1163 /* Delete draw module's data */
1164 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1166 assert(shader
->variants_cached
== 0);
1167 FREE((void *) shader
->base
.tokens
);
1174 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1175 uint shader
, uint index
,
1176 struct pipe_constant_buffer
*cb
)
1178 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1179 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
1183 if (cb
&& cb
->user_buffer
) {
1184 constants
= llvmpipe_user_buffer_create(pipe
->screen
,
1185 (void *) cb
->user_buffer
,
1187 PIPE_BIND_CONSTANT_BUFFER
);
1190 size
= constants
? constants
->width0
: 0;
1191 data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1193 assert(shader
< PIPE_SHADER_TYPES
);
1194 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1196 if(llvmpipe
->constants
[shader
][index
] == constants
)
1199 draw_flush(llvmpipe
->draw
);
1201 /* note: reference counting */
1202 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1204 if(shader
== PIPE_SHADER_VERTEX
||
1205 shader
== PIPE_SHADER_GEOMETRY
) {
1206 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1210 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1212 if (cb
&& cb
->user_buffer
) {
1213 pipe_resource_reference(&constants
, NULL
);
1219 * Return the blend factor equivalent to a destination alpha of one.
1221 static INLINE
unsigned
1222 force_dst_alpha_one(unsigned factor
)
1225 case PIPE_BLENDFACTOR_DST_ALPHA
:
1226 return PIPE_BLENDFACTOR_ONE
;
1227 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1228 return PIPE_BLENDFACTOR_ZERO
;
1229 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1230 return PIPE_BLENDFACTOR_ZERO
;
1238 * We need to generate several variants of the fragment pipeline to match
1239 * all the combinations of the contributing state atoms.
1241 * TODO: there is actually no reason to tie this to context state -- the
1242 * generated code could be cached globally in the screen.
1245 make_variant_key(struct llvmpipe_context
*lp
,
1246 struct lp_fragment_shader
*shader
,
1247 struct lp_fragment_shader_variant_key
*key
)
1251 memset(key
, 0, shader
->variant_key_size
);
1253 if (lp
->framebuffer
.zsbuf
) {
1254 if (lp
->depth_stencil
->depth
.enabled
) {
1255 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1256 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1258 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1259 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1260 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1264 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1265 if(key
->alpha
.enabled
)
1266 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1267 /* alpha.ref_value is passed in jit_context */
1269 key
->flatshade
= lp
->rasterizer
->flatshade
;
1270 if (lp
->active_query_count
) {
1271 key
->occlusion_count
= TRUE
;
1274 if (lp
->framebuffer
.nr_cbufs
) {
1275 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1278 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1279 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1280 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1281 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1282 const struct util_format_description
*format_desc
;
1284 key
->cbuf_format
[i
] = format
;
1286 format_desc
= util_format_description(format
);
1287 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1288 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1290 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1293 * Mask out color channels not present in the color buffer.
1295 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1298 * Our swizzled render tiles always have an alpha channel, but the linear
1299 * render target format often does not, so force here the dst alpha to be
1302 * This is not a mere optimization. Wrong results will be produced if the
1303 * dst alpha is used, the dst format does not have alpha, and the previous
1304 * rendering was not flushed from the swizzled to linear buffer. For
1305 * example, NonPowTwo DCT.
1307 * TODO: This should be generalized to all channels for better
1308 * performance, but only alpha causes correctness issues.
1310 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1312 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
1313 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
1314 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1315 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1316 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1317 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1318 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1322 /* This value will be the same for all the variants of a given shader:
1324 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1326 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1327 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1328 lp_sampler_static_state(&key
->sampler
[i
],
1329 lp
->fragment_sampler_views
[i
],
1338 * Update fragment shader state. This is called just prior to drawing
1339 * something when some fragment-related state has changed.
1342 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1344 struct lp_fragment_shader
*shader
= lp
->fs
;
1345 struct lp_fragment_shader_variant_key key
;
1346 struct lp_fragment_shader_variant
*variant
= NULL
;
1347 struct lp_fs_variant_list_item
*li
;
1349 make_variant_key(lp
, shader
, &key
);
1351 /* Search the variants for one which matches the key */
1352 li
= first_elem(&shader
->variants
);
1353 while(!at_end(&shader
->variants
, li
)) {
1354 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1362 /* Move this variant to the head of the list to implement LRU
1363 * deletion of shader's when we have too many.
1365 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1368 /* variant not found, create it now */
1371 unsigned variants_to_cull
;
1374 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
1377 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
1380 /* First, check if we've exceeded the max number of shader variants.
1381 * If so, free 25% of them (the least recently used ones).
1383 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
1385 if (variants_to_cull
||
1386 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
1387 struct pipe_context
*pipe
= &lp
->pipe
;
1390 * XXX: we need to flush the context until we have some sort of
1391 * reference counting in fragment shaders as they may still be binned
1392 * Flushing alone might not be sufficient we need to wait on it too.
1394 llvmpipe_finish(pipe
, __FUNCTION__
);
1397 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
1398 * number of shader variants (potentially all of them) could be
1399 * pending for destruction on flush.
1402 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
1403 struct lp_fs_variant_list_item
*item
;
1404 if (is_empty_list(&lp
->fs_variants_list
)) {
1407 item
= last_elem(&lp
->fs_variants_list
);
1410 llvmpipe_remove_shader_variant(lp
, item
->base
);
1415 * Generate the new variant.
1418 variant
= generate_variant(lp
, shader
, &key
);
1421 LP_COUNT_ADD(llvm_compile_time
, dt
);
1422 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1424 llvmpipe_variant_count
++;
1426 /* Put the new variant into the list */
1428 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1429 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1430 lp
->nr_fs_variants
++;
1431 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
1432 shader
->variants_cached
++;
1436 /* Bind this variant */
1437 lp_setup_set_fs_variant(lp
->setup
, variant
);
1447 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1449 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1450 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1451 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1453 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;