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>
104 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(LLVMBuilderRef builder
,
118 struct lp_type fs_type
,
120 LLVMValueRef mask_input
) /* int32 */
122 struct lp_type mask_type
;
123 LLVMTypeRef i32t
= LLVMInt32Type();
124 LLVMValueRef bits
[4];
129 * XXX: We'll need a different path for 16 x u8
131 assert(fs_type
.width
== 32);
132 assert(fs_type
.length
== 4);
133 mask_type
= lp_int_type(fs_type
);
136 * 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] }
166 mask
= lp_build_broadcast(builder
, lp_build_vec_type(mask_type
), mask_input
);
168 bits
[0] = LLVMConstInt(i32t
, 1 << 0, 0);
169 bits
[1] = LLVMConstInt(i32t
, 1 << 1, 0);
170 bits
[2] = LLVMConstInt(i32t
, 1 << 4, 0);
171 bits
[3] = LLVMConstInt(i32t
, 1 << 5, 0);
173 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, 4), "");
176 * mask = mask != 0 ? ~0 : 0
179 mask
= lp_build_compare(builder
,
180 mask_type
, PIPE_FUNC_NOTEQUAL
,
182 lp_build_const_int_vec(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 llvmpipe_context
*lp
,
216 struct lp_fragment_shader
*shader
,
217 const struct lp_fragment_shader_variant_key
*key
,
218 LLVMBuilderRef builder
,
220 LLVMValueRef context_ptr
,
222 struct lp_build_interp_soa_context
*interp
,
223 struct lp_build_sampler_soa
*sampler
,
225 LLVMValueRef (*color
)[4],
226 LLVMValueRef depth_ptr
,
228 unsigned partial_mask
,
229 LLVMValueRef mask_input
,
230 LLVMValueRef counter
)
232 const struct util_format_description
*zs_format_desc
= NULL
;
233 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
234 LLVMTypeRef vec_type
;
235 LLVMValueRef consts_ptr
;
236 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
238 LLVMValueRef zs_value
= NULL
;
239 LLVMValueRef stencil_refs
[2];
240 struct lp_build_flow_context
*flow
;
241 struct lp_build_mask_context mask
;
242 boolean simple_shader
= (shader
->info
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
243 shader
->info
.num_inputs
< 3 &&
244 shader
->info
.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
.writes_z
) {
258 if (key
->alpha
.enabled
|| shader
->info
.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 flow
= lp_build_flow_create(builder
);
291 memset(outputs
, 0, sizeof outputs
);
293 /* Declare the color and z variables */
294 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
295 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
296 color
[cbuf
][chan
] = lp_build_alloca(builder
, vec_type
, "color");
300 /* do triangle edge testing */
302 *pmask
= generate_quad_mask(builder
, type
,
306 *pmask
= lp_build_const_int_vec(type
, ~0);
309 /* 'mask' will control execution based on quad's pixel alive/killed state */
310 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
312 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
313 lp_build_mask_check(&mask
);
315 lp_build_interp_soa_update_pos(interp
, i
);
318 if (depth_mode
& EARLY_DEPTH_TEST
) {
319 lp_build_depth_stencil_test(builder
,
331 if (depth_mode
& EARLY_DEPTH_WRITE
)
332 LLVMBuildStore(builder
, zs_value
, depth_ptr
);
335 lp_build_interp_soa_update_inputs(interp
, i
);
337 /* Build the actual shader */
338 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
339 consts_ptr
, interp
->pos
, interp
->inputs
,
340 outputs
, sampler
, &shader
->info
);
344 if (key
->alpha
.enabled
) {
345 int color0
= find_output_by_semantic(&shader
->info
,
350 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
351 LLVMValueRef alpha_ref_value
;
353 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
354 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
356 lp_build_alpha_test(builder
, key
->alpha
.func
, type
,
357 &mask
, alpha
, alpha_ref_value
,
358 (depth_mode
& LATE_DEPTH_TEST
) != 0);
363 if (depth_mode
& LATE_DEPTH_TEST
) {
364 int pos0
= find_output_by_semantic(&shader
->info
,
365 TGSI_SEMANTIC_POSITION
,
369 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "z");
370 lp_build_name(z
, "output%u.%u.%c", i
, pos0
, "xyzw"[chan
]);
373 lp_build_depth_stencil_test(builder
,
385 if (depth_mode
& LATE_DEPTH_WRITE
)
386 LLVMBuildStore(builder
, zs_value
, depth_ptr
);
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
.num_outputs
; ++attrib
)
407 if (shader
->info
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
)
409 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
410 for(chan
= 0; chan
< NUM_CHANNELS
; ++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
]);
423 lp_build_occlusion_count(builder
, type
,
424 lp_build_mask_value(&mask
), counter
);
426 *pmask
= lp_build_mask_end(&mask
);
428 lp_build_flow_destroy(flow
);
433 * Generate color blending and color output.
434 * \param rt the render target index (to index blend, colormask state)
435 * \param type the pixel color type
436 * \param context_ptr pointer to the runtime JIT context
437 * \param mask execution mask (active fragment/pixel mask)
438 * \param src colors from the fragment shader
439 * \param dst_ptr the destination color buffer pointer
442 generate_blend(const struct pipe_blend_state
*blend
,
444 LLVMBuilderRef builder
,
446 LLVMValueRef context_ptr
,
449 LLVMValueRef dst_ptr
,
452 struct lp_build_context bld
;
453 struct lp_build_flow_context
*flow
;
454 struct lp_build_mask_context mask_ctx
;
455 LLVMTypeRef vec_type
;
456 LLVMValueRef const_ptr
;
462 lp_build_context_init(&bld
, builder
, type
);
464 flow
= lp_build_flow_create(builder
);
465 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
467 lp_build_mask_check(&mask_ctx
);
469 vec_type
= lp_build_vec_type(type
);
471 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
472 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
473 LLVMPointerType(vec_type
, 0), "");
475 /* load constant blend color and colors from the dest color buffer */
476 for(chan
= 0; chan
< 4; ++chan
) {
477 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
478 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
480 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
482 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
483 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
487 lp_build_blend_soa(builder
, blend
, type
, rt
, src
, dst
, con
, res
);
489 /* store results to color buffer */
490 for(chan
= 0; chan
< 4; ++chan
) {
491 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
492 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
493 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
494 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
495 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
499 lp_build_mask_end(&mask_ctx
);
500 lp_build_flow_destroy(flow
);
505 * Generate the runtime callable function for the whole fragment pipeline.
506 * Note that the function which we generate operates on a block of 16
507 * pixels at at time. The block contains 2x2 quads. Each quad contains
511 generate_fragment(struct llvmpipe_context
*lp
,
512 struct lp_fragment_shader
*shader
,
513 struct lp_fragment_shader_variant
*variant
,
514 unsigned partial_mask
)
516 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
517 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
519 struct lp_type fs_type
;
520 struct lp_type blend_type
;
521 LLVMTypeRef fs_elem_type
;
522 LLVMTypeRef fs_int_vec_type
;
523 LLVMTypeRef blend_vec_type
;
524 LLVMTypeRef arg_types
[11];
525 LLVMTypeRef func_type
;
526 LLVMValueRef context_ptr
;
530 LLVMValueRef dadx_ptr
;
531 LLVMValueRef dady_ptr
;
532 LLVMValueRef color_ptr_ptr
;
533 LLVMValueRef depth_ptr
;
534 LLVMValueRef mask_input
;
535 LLVMValueRef counter
= NULL
;
536 LLVMBasicBlockRef block
;
537 LLVMBuilderRef builder
;
538 struct lp_build_sampler_soa
*sampler
;
539 struct lp_build_interp_soa_context interp
;
540 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
541 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
542 LLVMValueRef blend_mask
;
543 LLVMValueRef function
;
551 /* TODO: actually pick these based on the fs and color buffer
552 * characteristics. */
554 memset(&fs_type
, 0, sizeof fs_type
);
555 fs_type
.floating
= TRUE
; /* floating point values */
556 fs_type
.sign
= TRUE
; /* values are signed */
557 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
558 fs_type
.width
= 32; /* 32-bit float */
559 fs_type
.length
= 4; /* 4 elements per vector */
560 num_fs
= 4; /* number of quads per block */
562 memset(&blend_type
, 0, sizeof blend_type
);
563 blend_type
.floating
= FALSE
; /* values are integers */
564 blend_type
.sign
= FALSE
; /* values are unsigned */
565 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
566 blend_type
.width
= 8; /* 8-bit ubyte values */
567 blend_type
.length
= 16; /* 16 elements per vector */
570 * Generate the function prototype. Any change here must be reflected in
571 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
574 fs_elem_type
= lp_build_elem_type(fs_type
);
575 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
577 blend_vec_type
= lp_build_vec_type(blend_type
);
579 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
580 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
582 arg_types
[0] = screen
->context_ptr_type
; /* context */
583 arg_types
[1] = LLVMInt32Type(); /* x */
584 arg_types
[2] = LLVMInt32Type(); /* y */
585 arg_types
[3] = LLVMFloatType(); /* facing */
586 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
587 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
588 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
589 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
590 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
591 arg_types
[9] = LLVMInt32Type(); /* mask_input */
592 arg_types
[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
594 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
596 function
= LLVMAddFunction(screen
->module
, func_name
, func_type
);
597 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
599 variant
->function
[partial_mask
] = function
;
602 /* XXX: need to propagate noalias down into color param now we are
603 * passing a pointer-to-pointer?
605 for(i
= 0; i
< Elements(arg_types
); ++i
)
606 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
607 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
609 context_ptr
= LLVMGetParam(function
, 0);
610 x
= LLVMGetParam(function
, 1);
611 y
= LLVMGetParam(function
, 2);
612 facing
= LLVMGetParam(function
, 3);
613 a0_ptr
= LLVMGetParam(function
, 4);
614 dadx_ptr
= LLVMGetParam(function
, 5);
615 dady_ptr
= LLVMGetParam(function
, 6);
616 color_ptr_ptr
= LLVMGetParam(function
, 7);
617 depth_ptr
= LLVMGetParam(function
, 8);
618 mask_input
= LLVMGetParam(function
, 9);
620 lp_build_name(context_ptr
, "context");
621 lp_build_name(x
, "x");
622 lp_build_name(y
, "y");
623 lp_build_name(a0_ptr
, "a0");
624 lp_build_name(dadx_ptr
, "dadx");
625 lp_build_name(dady_ptr
, "dady");
626 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
627 lp_build_name(depth_ptr
, "depth");
628 lp_build_name(mask_input
, "mask_input");
630 if (key
->occlusion_count
) {
631 counter
= LLVMGetParam(function
, 10);
632 lp_build_name(counter
, "counter");
639 block
= LLVMAppendBasicBlock(function
, "entry");
640 builder
= LLVMCreateBuilder();
641 LLVMPositionBuilderAtEnd(builder
, block
);
644 * The shader input interpolation info is not explicitely baked in the
645 * shader key, but everything it derives from (TGSI, and flatshade) is
646 * already included in the shader key.
648 lp_build_interp_soa_init(&interp
,
652 a0_ptr
, dadx_ptr
, dady_ptr
,
655 /* code generated texture sampling */
656 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
658 /* loop over quads in the block */
659 for(i
= 0; i
< num_fs
; ++i
) {
660 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
661 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
662 LLVMValueRef depth_ptr_i
;
664 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
666 generate_fs(lp
, shader
, key
,
673 &fs_mask
[i
], /* output */
681 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
682 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
683 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
686 sampler
->destroy(sampler
);
688 /* Loop over color outputs / color buffers to do blending.
690 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
691 LLVMValueRef color_ptr
;
692 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
693 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
697 * Convert the fs's output color and mask to fit to the blending type.
699 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
700 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
702 for (i
= 0; i
< num_fs
; i
++) {
704 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
707 lp_build_conv(builder
, fs_type
, blend_type
,
710 &blend_in_color
[chan
], 1);
712 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
715 if (partial_mask
|| !variant
->opaque
) {
716 lp_build_conv_mask(builder
, fs_type
, blend_type
,
720 blend_mask
= lp_build_const_int_vec(blend_type
, ~0);
723 color_ptr
= LLVMBuildLoad(builder
,
724 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
726 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
728 /* which blend/colormask state to use */
729 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
735 /* Could the 4x4 have been killed?
737 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
738 !key
->alpha
.enabled
&&
739 !shader
->info
.uses_kill
);
741 generate_blend(&key
->blend
,
754 /* Avoid corrupting the FPU stack on 32bit OSes. */
755 lp_build_intrinsic(builder
, "llvm.x86.mmx.emms", LLVMVoidType(), NULL
, 0);
758 LLVMBuildRetVoid(builder
);
760 LLVMDisposeBuilder(builder
);
763 /* Verify the LLVM IR. If invalid, dump and abort */
765 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
767 lp_debug_dump_value(function
);
772 /* Apply optimizations to LLVM IR */
773 LLVMRunFunctionPassManager(screen
->pass
, function
);
775 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
776 /* Print the LLVM IR to stderr */
777 lp_debug_dump_value(function
);
782 * Translate the LLVM IR into machine code.
785 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
787 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
789 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
792 lp_func_delete_body(function
);
798 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
802 debug_printf("fs variant %p:\n", (void *) key
);
804 if (key
->flatshade
) {
805 debug_printf("flatshade = 1\n");
807 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
808 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
810 if (key
->depth
.enabled
) {
811 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
812 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
813 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
816 for (i
= 0; i
< 2; ++i
) {
817 if (key
->stencil
[i
].enabled
) {
818 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
819 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
820 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
821 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
822 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
823 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
827 if (key
->alpha
.enabled
) {
828 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
831 if (key
->occlusion_count
) {
832 debug_printf("occlusion_count = 1\n");
835 if (key
->blend
.logicop_enable
) {
836 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
838 else if (key
->blend
.rt
[0].blend_enable
) {
839 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
840 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
841 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
842 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
843 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
844 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
846 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
847 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
848 debug_printf("sampler[%u] = \n", i
);
849 debug_printf(" .format = %s\n",
850 util_format_name(key
->sampler
[i
].format
));
851 debug_printf(" .target = %s\n",
852 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
853 debug_printf(" .pot = %u %u %u\n",
854 key
->sampler
[i
].pot_width
,
855 key
->sampler
[i
].pot_height
,
856 key
->sampler
[i
].pot_depth
);
857 debug_printf(" .wrap = %s %s %s\n",
858 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
859 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
860 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
861 debug_printf(" .min_img_filter = %s\n",
862 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
863 debug_printf(" .min_mip_filter = %s\n",
864 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
865 debug_printf(" .mag_img_filter = %s\n",
866 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
867 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
868 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
869 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
870 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
871 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
872 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
873 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
879 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
881 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
882 variant
->shader
->no
, variant
->no
);
883 tgsi_dump(variant
->shader
->base
.tokens
, 0);
884 dump_fs_variant_key(&variant
->key
);
885 debug_printf("variant->opaque = %u\n", variant
->opaque
);
889 static struct lp_fragment_shader_variant
*
890 generate_variant(struct llvmpipe_context
*lp
,
891 struct lp_fragment_shader
*shader
,
892 const struct lp_fragment_shader_variant_key
*key
)
894 struct lp_fragment_shader_variant
*variant
;
895 boolean fullcolormask
;
897 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
901 variant
->shader
= shader
;
902 variant
->list_item_global
.base
= variant
;
903 variant
->list_item_local
.base
= variant
;
904 variant
->no
= shader
->variants_created
++;
906 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
909 * Determine whether we are touching all channels in the color buffer.
911 fullcolormask
= FALSE
;
912 if (key
->nr_cbufs
== 1) {
913 const struct util_format_description
*format_desc
;
914 format_desc
= util_format_description(key
->cbuf_format
[0]);
915 if ((~key
->blend
.rt
[0].colormask
&
916 util_format_colormask(format_desc
)) == 0) {
917 fullcolormask
= TRUE
;
922 !key
->blend
.logicop_enable
&&
923 !key
->blend
.rt
[0].blend_enable
&&
925 !key
->stencil
[0].enabled
&&
926 !key
->alpha
.enabled
&&
927 !key
->depth
.enabled
&&
928 !shader
->info
.uses_kill
932 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
933 lp_debug_fs_variant(variant
);
936 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
938 if (variant
->opaque
) {
939 /* Specialized shader, which doesn't need to read the color buffer. */
940 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
942 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
950 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
951 const struct pipe_shader_state
*templ
)
953 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
954 struct lp_fragment_shader
*shader
;
957 shader
= CALLOC_STRUCT(lp_fragment_shader
);
961 shader
->no
= fs_no
++;
962 make_empty_list(&shader
->variants
);
964 /* get/save the summary info for this shader */
965 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
967 /* we need to keep a local copy of the tokens */
968 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
970 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
971 if (shader
->draw_data
== NULL
) {
972 FREE((void *) shader
->base
.tokens
);
977 nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
979 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
980 sampler
[nr_samplers
]);
982 if (LP_DEBUG
& DEBUG_TGSI
) {
984 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader
->no
, (void *) shader
);
985 tgsi_dump(templ
->tokens
, 0);
986 debug_printf("usage masks:\n");
987 for (attrib
= 0; attrib
< shader
->info
.num_inputs
; ++attrib
) {
988 unsigned usage_mask
= shader
->info
.input_usage_mask
[attrib
];
989 debug_printf(" IN[%u].%s%s%s%s\n",
991 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
992 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
993 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
994 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
1004 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1006 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1008 if (llvmpipe
->fs
== fs
)
1011 draw_flush(llvmpipe
->draw
);
1013 draw_bind_fragment_shader(llvmpipe
->draw
,
1014 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
1018 llvmpipe
->dirty
|= LP_NEW_FS
;
1022 remove_shader_variant(struct llvmpipe_context
*lp
,
1023 struct lp_fragment_shader_variant
*variant
)
1025 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
1028 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
1029 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1030 variant
->shader
->no
, variant
->no
, variant
->shader
->variants_created
,
1031 variant
->shader
->variants_cached
, lp
->nr_fs_variants
);
1033 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1034 if (variant
->function
[i
]) {
1035 if (variant
->jit_function
[i
])
1036 LLVMFreeMachineCodeForFunction(screen
->engine
,
1037 variant
->function
[i
]);
1038 LLVMDeleteFunction(variant
->function
[i
]);
1041 remove_from_list(&variant
->list_item_local
);
1042 variant
->shader
->variants_cached
--;
1043 remove_from_list(&variant
->list_item_global
);
1044 lp
->nr_fs_variants
--;
1049 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1051 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1052 struct lp_fragment_shader
*shader
= fs
;
1053 struct lp_fs_variant_list_item
*li
;
1055 assert(fs
!= llvmpipe
->fs
);
1059 * XXX: we need to flush the context until we have some sort of reference
1060 * counting in fragment shaders as they may still be binned
1061 * Flushing alone might not sufficient we need to wait on it too.
1064 llvmpipe_finish(pipe
, __FUNCTION__
);
1066 li
= first_elem(&shader
->variants
);
1067 while(!at_end(&shader
->variants
, li
)) {
1068 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1069 remove_shader_variant(llvmpipe
, li
->base
);
1073 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1075 assert(shader
->variants_cached
== 0);
1076 FREE((void *) shader
->base
.tokens
);
1083 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1084 uint shader
, uint index
,
1085 struct pipe_resource
*constants
)
1087 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1088 unsigned size
= constants
? constants
->width0
: 0;
1089 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1091 assert(shader
< PIPE_SHADER_TYPES
);
1092 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1094 if(llvmpipe
->constants
[shader
][index
] == constants
)
1097 draw_flush(llvmpipe
->draw
);
1099 /* note: reference counting */
1100 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1102 if(shader
== PIPE_SHADER_VERTEX
||
1103 shader
== PIPE_SHADER_GEOMETRY
) {
1104 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1108 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1113 * Return the blend factor equivalent to a destination alpha of one.
1115 static INLINE
unsigned
1116 force_dst_alpha_one(unsigned factor
)
1119 case PIPE_BLENDFACTOR_DST_ALPHA
:
1120 return PIPE_BLENDFACTOR_ONE
;
1121 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1122 return PIPE_BLENDFACTOR_ZERO
;
1123 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1124 return PIPE_BLENDFACTOR_ZERO
;
1132 * We need to generate several variants of the fragment pipeline to match
1133 * all the combinations of the contributing state atoms.
1135 * TODO: there is actually no reason to tie this to context state -- the
1136 * generated code could be cached globally in the screen.
1139 make_variant_key(struct llvmpipe_context
*lp
,
1140 struct lp_fragment_shader
*shader
,
1141 struct lp_fragment_shader_variant_key
*key
)
1145 memset(key
, 0, shader
->variant_key_size
);
1147 if (lp
->framebuffer
.zsbuf
) {
1148 if (lp
->depth_stencil
->depth
.enabled
) {
1149 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1150 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1152 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1153 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1154 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1158 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1159 if(key
->alpha
.enabled
)
1160 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1161 /* alpha.ref_value is passed in jit_context */
1163 key
->flatshade
= lp
->rasterizer
->flatshade
;
1164 if (lp
->active_query_count
) {
1165 key
->occlusion_count
= TRUE
;
1168 if (lp
->framebuffer
.nr_cbufs
) {
1169 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1172 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1173 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1174 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1175 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1176 const struct util_format_description
*format_desc
;
1178 key
->cbuf_format
[i
] = format
;
1180 format_desc
= util_format_description(format
);
1181 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1182 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1184 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1187 * Mask out color channels not present in the color buffer.
1189 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1192 * Our swizzled render tiles always have an alpha channel, but the linear
1193 * render target format often does not, so force here the dst alpha to be
1196 * This is not a mere optimization. Wrong results will be produced if the
1197 * dst alpha is used, the dst format does not have alpha, and the previous
1198 * rendering was not flushed from the swizzled to linear buffer. For
1199 * example, NonPowTwo DCT.
1201 * TODO: This should be generalized to all channels for better
1202 * performance, but only alpha causes correctness issues.
1204 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1206 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1207 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1208 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1209 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1210 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1211 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1215 /* This value will be the same for all the variants of a given shader:
1217 key
->nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1219 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1220 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1221 lp_sampler_static_state(&key
->sampler
[i
],
1222 lp
->fragment_sampler_views
[i
],
1229 * Update fragment state. This is called just prior to drawing
1230 * something when some fragment-related state has changed.
1233 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1235 struct lp_fragment_shader
*shader
= lp
->fs
;
1236 struct lp_fragment_shader_variant_key key
;
1237 struct lp_fragment_shader_variant
*variant
= NULL
;
1238 struct lp_fs_variant_list_item
*li
;
1240 make_variant_key(lp
, shader
, &key
);
1242 li
= first_elem(&shader
->variants
);
1243 while(!at_end(&shader
->variants
, li
)) {
1244 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1252 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1258 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1259 struct pipe_context
*pipe
= &lp
->pipe
;
1262 * XXX: we need to flush the context until we have some sort of reference
1263 * counting in fragment shaders as they may still be binned
1264 * Flushing alone might not be sufficient we need to wait on it too.
1266 llvmpipe_finish(pipe
, __FUNCTION__
);
1268 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1269 struct lp_fs_variant_list_item
*item
= last_elem(&lp
->fs_variants_list
);
1270 remove_shader_variant(lp
, item
->base
);
1275 variant
= generate_variant(lp
, shader
, &key
);
1279 LP_COUNT_ADD(llvm_compile_time
, dt
);
1280 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1283 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1284 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1285 lp
->nr_fs_variants
++;
1286 shader
->variants_cached
++;
1290 lp_setup_set_fs_variant(lp
->setup
, variant
);
1296 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1298 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1299 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1300 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1302 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;