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"
93 #include "lp_screen.h"
96 #include "lp_tex_sample.h"
98 #include "lp_state_fs.h"
101 #include <llvm-c/Analysis.h>
102 #include <llvm-c/BitWriter.h>
105 static unsigned fs_no
= 0;
110 * Expand the relevent bits of mask_input to a 4-dword mask for the
111 * four pixels in a 2x2 quad. This will set the four elements of the
112 * quad mask vector to 0 or ~0.
114 * \param quad which quad of the quad group to test, in [0,3]
115 * \param mask_input bitwise mask for the whole 4x4 stamp
118 generate_quad_mask(LLVMBuilderRef builder
,
119 struct lp_type fs_type
,
121 LLVMValueRef mask_input
) /* int32 */
123 struct lp_type mask_type
;
124 LLVMTypeRef i32t
= LLVMInt32Type();
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)
158 mask_input
= LLVMBuildLShr(builder
,
160 LLVMConstInt(i32t
, shift
, 0),
164 * mask = { mask_input & (1 << i), for i in [0,3] }
167 mask
= lp_build_broadcast(builder
, lp_build_vec_type(mask_type
), mask_input
);
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
180 mask
= lp_build_compare(builder
,
181 mask_type
, PIPE_FUNC_NOTEQUAL
,
183 lp_build_const_int_vec(mask_type
, 0));
189 #define EARLY_DEPTH_TEST 0x1
190 #define LATE_DEPTH_TEST 0x2
191 #define EARLY_DEPTH_WRITE 0x4
192 #define LATE_DEPTH_WRITE 0x8
195 find_output_by_semantic( const struct tgsi_shader_info
*info
,
201 for (i
= 0; i
< info
->num_outputs
; i
++)
202 if (info
->output_semantic_name
[i
] == semantic
&&
203 info
->output_semantic_index
[i
] == index
)
211 * Generate the fragment shader, depth/stencil test, and alpha tests.
212 * \param i which quad in the tile, in range [0,3]
213 * \param partial_mask if 1, do mask_input testing
216 generate_fs(struct llvmpipe_context
*lp
,
217 struct lp_fragment_shader
*shader
,
218 const struct lp_fragment_shader_variant_key
*key
,
219 LLVMBuilderRef builder
,
221 LLVMValueRef context_ptr
,
223 struct lp_build_interp_soa_context
*interp
,
224 struct lp_build_sampler_soa
*sampler
,
226 LLVMValueRef (*color
)[4],
227 LLVMValueRef depth_ptr
,
229 unsigned partial_mask
,
230 LLVMValueRef mask_input
,
231 LLVMValueRef counter
)
233 const struct util_format_description
*zs_format_desc
= NULL
;
234 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
235 LLVMTypeRef vec_type
;
236 LLVMValueRef consts_ptr
;
237 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
239 LLVMValueRef zs_value
= NULL
;
240 LLVMValueRef stencil_refs
[2];
241 struct lp_build_mask_context mask
;
242 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
243 shader
->info
.base
.num_inputs
< 3 &&
244 shader
->info
.base
.num_instructions
< 8);
250 if (key
->depth
.enabled
||
251 key
->stencil
[0].enabled
||
252 key
->stencil
[1].enabled
) {
254 zs_format_desc
= util_format_description(key
->zsbuf_format
);
255 assert(zs_format_desc
);
257 if (!shader
->info
.base
.writes_z
) {
258 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
259 /* With alpha test and kill, can do the depth test early
260 * and hopefully eliminate some quads. But need to do a
261 * special deferred depth write once the final mask value
264 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
266 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
269 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
272 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
273 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
274 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
282 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
283 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
285 vec_type
= lp_build_vec_type(type
);
287 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
289 memset(outputs
, 0, sizeof outputs
);
291 /* Declare the color and z variables */
292 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
293 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
294 color
[cbuf
][chan
] = lp_build_alloca(builder
, vec_type
, "color");
298 /* do triangle edge testing */
300 *pmask
= generate_quad_mask(builder
, type
,
304 *pmask
= lp_build_const_int_vec(type
, ~0);
307 /* 'mask' will control execution based on quad's pixel alive/killed state */
308 lp_build_mask_begin(&mask
, builder
, type
, *pmask
);
310 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
311 lp_build_mask_check(&mask
);
313 lp_build_interp_soa_update_pos(interp
, i
);
316 if (depth_mode
& EARLY_DEPTH_TEST
) {
317 lp_build_depth_stencil_test(builder
,
329 if (depth_mode
& EARLY_DEPTH_WRITE
) {
330 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
334 lp_build_interp_soa_update_inputs(interp
, i
);
336 /* Build the actual shader */
337 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
338 consts_ptr
, interp
->pos
, interp
->inputs
,
339 outputs
, sampler
, &shader
->info
.base
);
343 if (key
->alpha
.enabled
) {
344 int color0
= find_output_by_semantic(&shader
->info
.base
,
349 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
350 LLVMValueRef alpha_ref_value
;
352 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
353 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
355 lp_build_alpha_test(builder
, key
->alpha
.func
, type
,
356 &mask
, alpha
, alpha_ref_value
,
357 (depth_mode
& LATE_DEPTH_TEST
) != 0);
362 if (depth_mode
& LATE_DEPTH_TEST
) {
363 int pos0
= find_output_by_semantic(&shader
->info
.base
,
364 TGSI_SEMANTIC_POSITION
,
368 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "z");
369 lp_build_name(z
, "output%u.%u.%c", i
, pos0
, "xyzw"[chan
]);
372 lp_build_depth_stencil_test(builder
,
384 if (depth_mode
& LATE_DEPTH_WRITE
) {
385 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
388 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
389 (depth_mode
& LATE_DEPTH_WRITE
))
391 /* Need to apply a reduced mask to the depth write. Reload the
392 * depth value, update from zs_value with the new mask value and
395 lp_build_deferred_depth_write(builder
,
405 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
407 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
)
409 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
410 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
411 if(outputs
[attrib
][chan
]) {
412 /* XXX: just initialize outputs to point at colors[] and
415 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
416 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
417 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
424 lp_build_occlusion_count(builder
, type
,
425 lp_build_mask_value(&mask
), counter
);
427 *pmask
= lp_build_mask_end(&mask
);
432 * Generate color blending and color output.
433 * \param rt the render target index (to index blend, colormask state)
434 * \param type the pixel color type
435 * \param context_ptr pointer to the runtime JIT context
436 * \param mask execution mask (active fragment/pixel mask)
437 * \param src colors from the fragment shader
438 * \param dst_ptr the destination color buffer pointer
441 generate_blend(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
, builder
, type
);
462 lp_build_mask_begin(&mask_ctx
, builder
, type
, mask
);
464 lp_build_mask_check(&mask_ctx
);
466 vec_type
= lp_build_vec_type(type
);
468 const_ptr
= lp_jit_context_blend_color(builder
, 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
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
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(builder
, 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
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
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 llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
513 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
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 LLVMValueRef context_ptr
;
526 LLVMValueRef dadx_ptr
;
527 LLVMValueRef dady_ptr
;
528 LLVMValueRef color_ptr_ptr
;
529 LLVMValueRef depth_ptr
;
530 LLVMValueRef mask_input
;
531 LLVMValueRef counter
= NULL
;
532 LLVMBasicBlockRef block
;
533 LLVMBuilderRef builder
;
534 struct lp_build_sampler_soa
*sampler
;
535 struct lp_build_interp_soa_context interp
;
536 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
537 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
538 LLVMValueRef blend_mask
;
539 LLVMValueRef function
;
541 const struct util_format_description
*zs_format_desc
;
548 /* TODO: actually pick these based on the fs and color buffer
549 * characteristics. */
551 memset(&fs_type
, 0, sizeof fs_type
);
552 fs_type
.floating
= TRUE
; /* floating point values */
553 fs_type
.sign
= TRUE
; /* values are signed */
554 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
555 fs_type
.width
= 32; /* 32-bit float */
556 fs_type
.length
= 4; /* 4 elements per vector */
557 num_fs
= 4; /* number of quads per block */
559 memset(&blend_type
, 0, sizeof blend_type
);
560 blend_type
.floating
= FALSE
; /* values are integers */
561 blend_type
.sign
= FALSE
; /* values are unsigned */
562 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
563 blend_type
.width
= 8; /* 8-bit ubyte values */
564 blend_type
.length
= 16; /* 16 elements per vector */
567 * Generate the function prototype. Any change here must be reflected in
568 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
571 fs_elem_type
= lp_build_elem_type(fs_type
);
572 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
574 blend_vec_type
= lp_build_vec_type(blend_type
);
576 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
577 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
579 arg_types
[0] = screen
->context_ptr_type
; /* context */
580 arg_types
[1] = LLVMInt32Type(); /* x */
581 arg_types
[2] = LLVMInt32Type(); /* y */
582 arg_types
[3] = LLVMFloatType(); /* facing */
583 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
584 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
585 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
586 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
587 arg_types
[8] = LLVMPointerType(LLVMInt8Type(), 0); /* depth */
588 arg_types
[9] = LLVMInt32Type(); /* mask_input */
589 arg_types
[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
591 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
593 function
= LLVMAddFunction(screen
->module
, func_name
, func_type
);
594 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
596 variant
->function
[partial_mask
] = function
;
599 /* XXX: need to propagate noalias down into color param now we are
600 * passing a pointer-to-pointer?
602 for(i
= 0; i
< Elements(arg_types
); ++i
)
603 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
604 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
606 context_ptr
= LLVMGetParam(function
, 0);
607 x
= LLVMGetParam(function
, 1);
608 y
= LLVMGetParam(function
, 2);
609 facing
= LLVMGetParam(function
, 3);
610 a0_ptr
= LLVMGetParam(function
, 4);
611 dadx_ptr
= LLVMGetParam(function
, 5);
612 dady_ptr
= LLVMGetParam(function
, 6);
613 color_ptr_ptr
= LLVMGetParam(function
, 7);
614 depth_ptr
= LLVMGetParam(function
, 8);
615 mask_input
= LLVMGetParam(function
, 9);
617 lp_build_name(context_ptr
, "context");
618 lp_build_name(x
, "x");
619 lp_build_name(y
, "y");
620 lp_build_name(a0_ptr
, "a0");
621 lp_build_name(dadx_ptr
, "dadx");
622 lp_build_name(dady_ptr
, "dady");
623 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
624 lp_build_name(depth_ptr
, "depth");
625 lp_build_name(mask_input
, "mask_input");
627 if (key
->occlusion_count
) {
628 counter
= LLVMGetParam(function
, 10);
629 lp_build_name(counter
, "counter");
636 block
= LLVMAppendBasicBlock(function
, "entry");
637 builder
= LLVMCreateBuilder();
638 LLVMPositionBuilderAtEnd(builder
, block
);
641 * The shader input interpolation info is not explicitely baked in the
642 * shader key, but everything it derives from (TGSI, and flatshade) is
643 * already included in the shader key.
645 lp_build_interp_soa_init(&interp
,
649 a0_ptr
, dadx_ptr
, dady_ptr
,
652 /* code generated texture sampling */
653 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
655 /* loop over quads in the block */
656 zs_format_desc
= util_format_description(key
->zsbuf_format
);
658 for(i
= 0; i
< num_fs
; ++i
) {
659 LLVMValueRef depth_offset
= LLVMConstInt(LLVMInt32Type(),
660 i
*fs_type
.length
*zs_format_desc
->block
.bits
/8,
662 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
663 LLVMValueRef depth_ptr_i
;
665 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
667 generate_fs(lp
, shader
, key
,
674 &fs_mask
[i
], /* output */
682 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
683 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
684 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
687 sampler
->destroy(sampler
);
689 /* Loop over color outputs / color buffers to do blending.
691 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
692 LLVMValueRef color_ptr
;
693 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
694 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
698 * Convert the fs's output color and mask to fit to the blending type.
700 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
701 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
703 for (i
= 0; i
< num_fs
; i
++) {
705 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
708 lp_build_conv(builder
, fs_type
, blend_type
,
711 &blend_in_color
[chan
], 1);
713 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
716 if (partial_mask
|| !variant
->opaque
) {
717 lp_build_conv_mask(builder
, fs_type
, blend_type
,
721 blend_mask
= lp_build_const_int_vec(blend_type
, ~0);
724 color_ptr
= LLVMBuildLoad(builder
,
725 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
727 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
729 /* which blend/colormask state to use */
730 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
736 /* Could the 4x4 have been killed?
738 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
739 !key
->alpha
.enabled
&&
740 !shader
->info
.base
.uses_kill
);
742 generate_blend(&key
->blend
,
755 /* Avoid corrupting the FPU stack on 32bit OSes. */
756 lp_build_intrinsic(builder
, "llvm.x86.mmx.emms", LLVMVoidType(), NULL
, 0);
759 LLVMBuildRetVoid(builder
);
761 LLVMDisposeBuilder(builder
);
764 /* Verify the LLVM IR. If invalid, dump and abort */
766 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
768 lp_debug_dump_value(function
);
773 /* Apply optimizations to LLVM IR */
774 LLVMRunFunctionPassManager(screen
->pass
, function
);
776 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
777 /* Print the LLVM IR to stderr */
778 lp_debug_dump_value(function
);
782 /* Dump byte code to a file */
784 LLVMWriteBitcodeToFile(lp_build_module
, "llvmpipe.bc");
788 * Translate the LLVM IR into machine code.
791 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
793 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
795 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
798 lp_func_delete_body(function
);
804 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
808 debug_printf("fs variant %p:\n", (void *) key
);
810 if (key
->flatshade
) {
811 debug_printf("flatshade = 1\n");
813 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
814 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
816 if (key
->depth
.enabled
) {
817 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
818 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
819 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
822 for (i
= 0; i
< 2; ++i
) {
823 if (key
->stencil
[i
].enabled
) {
824 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
825 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
826 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
827 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
828 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
829 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
833 if (key
->alpha
.enabled
) {
834 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
837 if (key
->occlusion_count
) {
838 debug_printf("occlusion_count = 1\n");
841 if (key
->blend
.logicop_enable
) {
842 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
844 else if (key
->blend
.rt
[0].blend_enable
) {
845 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
846 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
847 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
848 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
849 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
850 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
852 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
853 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
854 debug_printf("sampler[%u] = \n", i
);
855 debug_printf(" .format = %s\n",
856 util_format_name(key
->sampler
[i
].format
));
857 debug_printf(" .target = %s\n",
858 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
859 debug_printf(" .pot = %u %u %u\n",
860 key
->sampler
[i
].pot_width
,
861 key
->sampler
[i
].pot_height
,
862 key
->sampler
[i
].pot_depth
);
863 debug_printf(" .wrap = %s %s %s\n",
864 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
865 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
866 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
867 debug_printf(" .min_img_filter = %s\n",
868 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
869 debug_printf(" .min_mip_filter = %s\n",
870 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
871 debug_printf(" .mag_img_filter = %s\n",
872 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
873 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
874 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
875 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
876 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
877 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
878 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
879 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
885 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
887 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
888 variant
->shader
->no
, variant
->no
);
889 tgsi_dump(variant
->shader
->base
.tokens
, 0);
890 dump_fs_variant_key(&variant
->key
);
891 debug_printf("variant->opaque = %u\n", variant
->opaque
);
895 static struct lp_fragment_shader_variant
*
896 generate_variant(struct llvmpipe_context
*lp
,
897 struct lp_fragment_shader
*shader
,
898 const struct lp_fragment_shader_variant_key
*key
)
900 struct lp_fragment_shader_variant
*variant
;
901 boolean fullcolormask
;
903 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
907 variant
->shader
= shader
;
908 variant
->list_item_global
.base
= variant
;
909 variant
->list_item_local
.base
= variant
;
910 variant
->no
= shader
->variants_created
++;
912 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
915 * Determine whether we are touching all channels in the color buffer.
917 fullcolormask
= FALSE
;
918 if (key
->nr_cbufs
== 1) {
919 const struct util_format_description
*format_desc
;
920 format_desc
= util_format_description(key
->cbuf_format
[0]);
921 if ((~key
->blend
.rt
[0].colormask
&
922 util_format_colormask(format_desc
)) == 0) {
923 fullcolormask
= TRUE
;
928 !key
->blend
.logicop_enable
&&
929 !key
->blend
.rt
[0].blend_enable
&&
931 !key
->stencil
[0].enabled
&&
932 !key
->alpha
.enabled
&&
933 !key
->depth
.enabled
&&
934 !shader
->info
.base
.uses_kill
938 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
939 lp_debug_fs_variant(variant
);
942 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
944 if (variant
->opaque
) {
945 /* Specialized shader, which doesn't need to read the color buffer. */
946 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
948 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
956 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
957 const struct pipe_shader_state
*templ
)
959 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
960 struct lp_fragment_shader
*shader
;
963 shader
= CALLOC_STRUCT(lp_fragment_shader
);
967 shader
->no
= fs_no
++;
968 make_empty_list(&shader
->variants
);
970 /* get/save the summary info for this shader */
971 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
973 /* we need to keep a local copy of the tokens */
974 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
976 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
977 if (shader
->draw_data
== NULL
) {
978 FREE((void *) shader
->base
.tokens
);
983 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
985 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
986 sampler
[nr_samplers
]);
988 if (LP_DEBUG
& DEBUG_TGSI
) {
990 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader
->no
, (void *) shader
);
991 tgsi_dump(templ
->tokens
, 0);
992 debug_printf("usage masks:\n");
993 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
994 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
995 debug_printf(" IN[%u].%s%s%s%s\n",
997 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
998 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
999 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
1000 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
1010 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1012 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1014 if (llvmpipe
->fs
== fs
)
1017 draw_flush(llvmpipe
->draw
);
1019 draw_bind_fragment_shader(llvmpipe
->draw
,
1020 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
1024 llvmpipe
->dirty
|= LP_NEW_FS
;
1028 remove_shader_variant(struct llvmpipe_context
*lp
,
1029 struct lp_fragment_shader_variant
*variant
)
1031 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
1034 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
1035 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1036 variant
->shader
->no
, variant
->no
, variant
->shader
->variants_created
,
1037 variant
->shader
->variants_cached
, lp
->nr_fs_variants
);
1039 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1040 if (variant
->function
[i
]) {
1041 if (variant
->jit_function
[i
])
1042 LLVMFreeMachineCodeForFunction(screen
->engine
,
1043 variant
->function
[i
]);
1044 LLVMDeleteFunction(variant
->function
[i
]);
1047 remove_from_list(&variant
->list_item_local
);
1048 variant
->shader
->variants_cached
--;
1049 remove_from_list(&variant
->list_item_global
);
1050 lp
->nr_fs_variants
--;
1055 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1057 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1058 struct lp_fragment_shader
*shader
= fs
;
1059 struct lp_fs_variant_list_item
*li
;
1061 assert(fs
!= llvmpipe
->fs
);
1065 * XXX: we need to flush the context until we have some sort of reference
1066 * counting in fragment shaders as they may still be binned
1067 * Flushing alone might not sufficient we need to wait on it too.
1070 llvmpipe_finish(pipe
, __FUNCTION__
);
1072 li
= first_elem(&shader
->variants
);
1073 while(!at_end(&shader
->variants
, li
)) {
1074 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1075 remove_shader_variant(llvmpipe
, li
->base
);
1079 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1081 assert(shader
->variants_cached
== 0);
1082 FREE((void *) shader
->base
.tokens
);
1089 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1090 uint shader
, uint index
,
1091 struct pipe_resource
*constants
)
1093 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1094 unsigned size
= constants
? constants
->width0
: 0;
1095 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1097 assert(shader
< PIPE_SHADER_TYPES
);
1098 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1100 if(llvmpipe
->constants
[shader
][index
] == constants
)
1103 draw_flush(llvmpipe
->draw
);
1105 /* note: reference counting */
1106 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1108 if(shader
== PIPE_SHADER_VERTEX
||
1109 shader
== PIPE_SHADER_GEOMETRY
) {
1110 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1114 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1119 * Return the blend factor equivalent to a destination alpha of one.
1121 static INLINE
unsigned
1122 force_dst_alpha_one(unsigned factor
)
1125 case PIPE_BLENDFACTOR_DST_ALPHA
:
1126 return PIPE_BLENDFACTOR_ONE
;
1127 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1128 return PIPE_BLENDFACTOR_ZERO
;
1129 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1130 return PIPE_BLENDFACTOR_ZERO
;
1138 * We need to generate several variants of the fragment pipeline to match
1139 * all the combinations of the contributing state atoms.
1141 * TODO: there is actually no reason to tie this to context state -- the
1142 * generated code could be cached globally in the screen.
1145 make_variant_key(struct llvmpipe_context
*lp
,
1146 struct lp_fragment_shader
*shader
,
1147 struct lp_fragment_shader_variant_key
*key
)
1151 memset(key
, 0, shader
->variant_key_size
);
1153 if (lp
->framebuffer
.zsbuf
) {
1154 if (lp
->depth_stencil
->depth
.enabled
) {
1155 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1156 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1158 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1159 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1160 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1164 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1165 if(key
->alpha
.enabled
)
1166 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1167 /* alpha.ref_value is passed in jit_context */
1169 key
->flatshade
= lp
->rasterizer
->flatshade
;
1170 if (lp
->active_query_count
) {
1171 key
->occlusion_count
= TRUE
;
1174 if (lp
->framebuffer
.nr_cbufs
) {
1175 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1178 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1179 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1180 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1181 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1182 const struct util_format_description
*format_desc
;
1184 key
->cbuf_format
[i
] = format
;
1186 format_desc
= util_format_description(format
);
1187 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1188 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1190 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1193 * Mask out color channels not present in the color buffer.
1195 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1198 * Our swizzled render tiles always have an alpha channel, but the linear
1199 * render target format often does not, so force here the dst alpha to be
1202 * This is not a mere optimization. Wrong results will be produced if the
1203 * dst alpha is used, the dst format does not have alpha, and the previous
1204 * rendering was not flushed from the swizzled to linear buffer. For
1205 * example, NonPowTwo DCT.
1207 * TODO: This should be generalized to all channels for better
1208 * performance, but only alpha causes correctness issues.
1210 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1212 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1213 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1214 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1215 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1216 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1217 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1221 /* This value will be the same for all the variants of a given shader:
1223 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1225 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1226 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1227 lp_sampler_static_state(&key
->sampler
[i
],
1228 lp
->fragment_sampler_views
[i
],
1235 * Update fragment state. This is called just prior to drawing
1236 * something when some fragment-related state has changed.
1239 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1241 struct lp_fragment_shader
*shader
= lp
->fs
;
1242 struct lp_fragment_shader_variant_key key
;
1243 struct lp_fragment_shader_variant
*variant
= NULL
;
1244 struct lp_fs_variant_list_item
*li
;
1246 make_variant_key(lp
, shader
, &key
);
1248 li
= first_elem(&shader
->variants
);
1249 while(!at_end(&shader
->variants
, li
)) {
1250 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1258 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1264 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1265 struct pipe_context
*pipe
= &lp
->pipe
;
1268 * XXX: we need to flush the context until we have some sort of reference
1269 * counting in fragment shaders as they may still be binned
1270 * Flushing alone might not be sufficient we need to wait on it too.
1272 llvmpipe_finish(pipe
, __FUNCTION__
);
1274 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1275 struct lp_fs_variant_list_item
*item
= last_elem(&lp
->fs_variants_list
);
1276 remove_shader_variant(lp
, item
->base
);
1281 variant
= generate_variant(lp
, shader
, &key
);
1285 LP_COUNT_ADD(llvm_compile_time
, dt
);
1286 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1289 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1290 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1291 lp
->nr_fs_variants
++;
1292 shader
->variants_cached
++;
1296 lp_setup_set_fs_variant(lp
->setup
, variant
);
1302 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1304 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1305 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1306 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1308 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;