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;
108 * Generate the depth /stencil test code.
111 generate_depth_stencil(LLVMBuilderRef builder
,
112 const struct lp_fragment_shader_variant_key
*key
,
113 struct lp_type src_type
,
114 struct lp_build_mask_context
*mask
,
115 LLVMValueRef stencil_refs
[2],
117 LLVMValueRef dst_ptr
,
119 LLVMValueRef counter
,
122 const struct util_format_description
*format_desc
;
124 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
127 format_desc
= util_format_description(key
->zsbuf_format
);
130 lp_build_depth_stencil_test(builder
,
146 * Expand the relevent bits of mask_input to a 4-dword mask for the
147 * four pixels in a 2x2 quad. This will set the four elements of the
148 * quad mask vector to 0 or ~0.
150 * \param quad which quad of the quad group to test, in [0,3]
151 * \param mask_input bitwise mask for the whole 4x4 stamp
154 generate_quad_mask(LLVMBuilderRef builder
,
155 struct lp_type fs_type
,
157 LLVMValueRef mask_input
) /* int32 */
159 struct lp_type mask_type
;
160 LLVMTypeRef i32t
= LLVMInt32Type();
161 LLVMValueRef bits
[4];
166 * XXX: We'll need a different path for 16 x u8
168 assert(fs_type
.width
== 32);
169 assert(fs_type
.length
== 4);
170 mask_type
= lp_int_type(fs_type
);
173 * mask_input >>= (quad * 4)
194 mask_input
= LLVMBuildLShr(builder
,
196 LLVMConstInt(i32t
, shift
, 0),
200 * mask = { mask_input & (1 << i), for i in [0,3] }
203 mask
= lp_build_broadcast(builder
, lp_build_vec_type(mask_type
), mask_input
);
205 bits
[0] = LLVMConstInt(i32t
, 1 << 0, 0);
206 bits
[1] = LLVMConstInt(i32t
, 1 << 1, 0);
207 bits
[2] = LLVMConstInt(i32t
, 1 << 4, 0);
208 bits
[3] = LLVMConstInt(i32t
, 1 << 5, 0);
210 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, 4), "");
213 * mask = mask != 0 ? ~0 : 0
216 mask
= lp_build_compare(builder
,
217 mask_type
, PIPE_FUNC_NOTEQUAL
,
219 lp_build_const_int_vec(mask_type
, 0));
227 * Generate the fragment shader, depth/stencil test, and alpha tests.
228 * \param i which quad in the tile, in range [0,3]
229 * \param partial_mask if 1, do mask_input testing
232 generate_fs(struct llvmpipe_context
*lp
,
233 struct lp_fragment_shader
*shader
,
234 const struct lp_fragment_shader_variant_key
*key
,
235 LLVMBuilderRef builder
,
237 LLVMValueRef context_ptr
,
239 struct lp_build_interp_soa_context
*interp
,
240 struct lp_build_sampler_soa
*sampler
,
242 LLVMValueRef (*color
)[4],
243 LLVMValueRef depth_ptr
,
245 unsigned partial_mask
,
246 LLVMValueRef mask_input
,
247 LLVMValueRef counter
)
249 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
250 LLVMTypeRef vec_type
;
251 LLVMValueRef consts_ptr
;
252 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
254 LLVMValueRef stencil_refs
[2];
255 struct lp_build_flow_context
*flow
;
256 struct lp_build_mask_context mask
;
257 boolean early_depth_stencil_test
;
258 boolean simple_shader
= (shader
->info
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
259 shader
->info
.num_inputs
< 3 &&
260 shader
->info
.num_instructions
< 8);
267 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
268 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
270 vec_type
= lp_build_vec_type(type
);
272 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
274 flow
= lp_build_flow_create(builder
);
276 memset(outputs
, 0, sizeof outputs
);
278 lp_build_flow_scope_begin(flow
);
280 /* Declare the color and z variables */
281 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
282 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
283 color
[cbuf
][chan
] = lp_build_alloca(builder
, vec_type
, "color");
287 /* do triangle edge testing */
289 *pmask
= generate_quad_mask(builder
, type
,
293 *pmask
= lp_build_const_int_vec(type
, ~0);
297 early_depth_stencil_test
=
298 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
299 !key
->alpha
.enabled
&&
300 !shader
->info
.uses_kill
&&
301 !shader
->info
.writes_z
;
303 /* 'mask' will control execution based on quad's pixel alive/killed state */
304 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
306 if (!early_depth_stencil_test
&& !simple_shader
)
307 lp_build_mask_check(&mask
);
309 lp_build_interp_soa_update_pos(interp
, i
);
312 if (early_depth_stencil_test
)
313 generate_depth_stencil(builder
, key
,
320 lp_build_interp_soa_update_inputs(interp
, i
);
322 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
323 consts_ptr
, interp
->pos
, interp
->inputs
,
324 outputs
, sampler
, &shader
->info
);
326 /* loop over fragment shader outputs/results */
327 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
328 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
329 if(outputs
[attrib
][chan
]) {
330 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
331 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
333 switch (shader
->info
.output_semantic_name
[attrib
]) {
334 case TGSI_SEMANTIC_COLOR
:
336 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
338 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
341 /* XXX: should only test the final assignment to alpha */
342 if (cbuf
== 0 && chan
== 3 && key
->alpha
.enabled
) {
343 LLVMValueRef alpha
= out
;
344 LLVMValueRef alpha_ref_value
;
345 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
346 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
347 lp_build_alpha_test(builder
, key
->alpha
.func
, type
,
348 &mask
, alpha
, alpha_ref_value
, FALSE
);
351 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
355 case TGSI_SEMANTIC_POSITION
:
364 if (!early_depth_stencil_test
)
365 generate_depth_stencil(builder
, key
,
367 stencil_refs
, z
, depth_ptr
,
368 facing
, counter
, FALSE
);
370 lp_build_mask_end(&mask
);
372 lp_build_flow_scope_end(flow
);
374 lp_build_flow_destroy(flow
);
382 * Generate color blending and color output.
383 * \param rt the render target index (to index blend, colormask state)
384 * \param type the pixel color type
385 * \param context_ptr pointer to the runtime JIT context
386 * \param mask execution mask (active fragment/pixel mask)
387 * \param src colors from the fragment shader
388 * \param dst_ptr the destination color buffer pointer
391 generate_blend(const struct pipe_blend_state
*blend
,
393 LLVMBuilderRef builder
,
395 LLVMValueRef context_ptr
,
398 LLVMValueRef dst_ptr
,
401 struct lp_build_context bld
;
402 struct lp_build_flow_context
*flow
;
403 struct lp_build_mask_context mask_ctx
;
404 LLVMTypeRef vec_type
;
405 LLVMValueRef const_ptr
;
411 lp_build_context_init(&bld
, builder
, type
);
413 flow
= lp_build_flow_create(builder
);
414 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
416 lp_build_mask_check(&mask_ctx
);
418 vec_type
= lp_build_vec_type(type
);
420 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
421 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
422 LLVMPointerType(vec_type
, 0), "");
424 /* load constant blend color and colors from the dest color buffer */
425 for(chan
= 0; chan
< 4; ++chan
) {
426 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
427 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
429 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
431 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
432 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
436 lp_build_blend_soa(builder
, blend
, type
, rt
, src
, dst
, con
, res
);
438 /* store results to color buffer */
439 for(chan
= 0; chan
< 4; ++chan
) {
440 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
441 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
442 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
443 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
444 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
448 lp_build_mask_end(&mask_ctx
);
449 lp_build_flow_destroy(flow
);
454 * Generate the runtime callable function for the whole fragment pipeline.
455 * Note that the function which we generate operates on a block of 16
456 * pixels at at time. The block contains 2x2 quads. Each quad contains
460 generate_fragment(struct llvmpipe_context
*lp
,
461 struct lp_fragment_shader
*shader
,
462 struct lp_fragment_shader_variant
*variant
,
463 unsigned partial_mask
)
465 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
466 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
468 struct lp_type fs_type
;
469 struct lp_type blend_type
;
470 LLVMTypeRef fs_elem_type
;
471 LLVMTypeRef fs_int_vec_type
;
472 LLVMTypeRef blend_vec_type
;
473 LLVMTypeRef arg_types
[11];
474 LLVMTypeRef func_type
;
475 LLVMValueRef context_ptr
;
479 LLVMValueRef dadx_ptr
;
480 LLVMValueRef dady_ptr
;
481 LLVMValueRef color_ptr_ptr
;
482 LLVMValueRef depth_ptr
;
483 LLVMValueRef mask_input
;
484 LLVMValueRef counter
= NULL
;
485 LLVMBasicBlockRef block
;
486 LLVMBuilderRef builder
;
487 struct lp_build_sampler_soa
*sampler
;
488 struct lp_build_interp_soa_context interp
;
489 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
490 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
491 LLVMValueRef blend_mask
;
492 LLVMValueRef function
;
500 /* TODO: actually pick these based on the fs and color buffer
501 * characteristics. */
503 memset(&fs_type
, 0, sizeof fs_type
);
504 fs_type
.floating
= TRUE
; /* floating point values */
505 fs_type
.sign
= TRUE
; /* values are signed */
506 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
507 fs_type
.width
= 32; /* 32-bit float */
508 fs_type
.length
= 4; /* 4 elements per vector */
509 num_fs
= 4; /* number of quads per block */
511 memset(&blend_type
, 0, sizeof blend_type
);
512 blend_type
.floating
= FALSE
; /* values are integers */
513 blend_type
.sign
= FALSE
; /* values are unsigned */
514 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
515 blend_type
.width
= 8; /* 8-bit ubyte values */
516 blend_type
.length
= 16; /* 16 elements per vector */
519 * Generate the function prototype. Any change here must be reflected in
520 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
523 fs_elem_type
= lp_build_elem_type(fs_type
);
524 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
526 blend_vec_type
= lp_build_vec_type(blend_type
);
528 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
529 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
531 arg_types
[0] = screen
->context_ptr_type
; /* context */
532 arg_types
[1] = LLVMInt32Type(); /* x */
533 arg_types
[2] = LLVMInt32Type(); /* y */
534 arg_types
[3] = LLVMFloatType(); /* facing */
535 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
536 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
537 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
538 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
539 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
540 arg_types
[9] = LLVMInt32Type(); /* mask_input */
541 arg_types
[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
543 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
545 function
= LLVMAddFunction(screen
->module
, func_name
, func_type
);
546 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
548 variant
->function
[partial_mask
] = function
;
551 /* XXX: need to propagate noalias down into color param now we are
552 * passing a pointer-to-pointer?
554 for(i
= 0; i
< Elements(arg_types
); ++i
)
555 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
556 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
558 context_ptr
= LLVMGetParam(function
, 0);
559 x
= LLVMGetParam(function
, 1);
560 y
= LLVMGetParam(function
, 2);
561 facing
= LLVMGetParam(function
, 3);
562 a0_ptr
= LLVMGetParam(function
, 4);
563 dadx_ptr
= LLVMGetParam(function
, 5);
564 dady_ptr
= LLVMGetParam(function
, 6);
565 color_ptr_ptr
= LLVMGetParam(function
, 7);
566 depth_ptr
= LLVMGetParam(function
, 8);
567 mask_input
= LLVMGetParam(function
, 9);
569 lp_build_name(context_ptr
, "context");
570 lp_build_name(x
, "x");
571 lp_build_name(y
, "y");
572 lp_build_name(a0_ptr
, "a0");
573 lp_build_name(dadx_ptr
, "dadx");
574 lp_build_name(dady_ptr
, "dady");
575 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
576 lp_build_name(depth_ptr
, "depth");
577 lp_build_name(mask_input
, "mask_input");
579 if (key
->occlusion_count
) {
580 counter
= LLVMGetParam(function
, 10);
581 lp_build_name(counter
, "counter");
588 block
= LLVMAppendBasicBlock(function
, "entry");
589 builder
= LLVMCreateBuilder();
590 LLVMPositionBuilderAtEnd(builder
, block
);
593 * The shader input interpolation info is not explicitely baked in the
594 * shader key, but everything it derives from (TGSI, and flatshade) is
595 * already included in the shader key.
597 lp_build_interp_soa_init(&interp
,
601 a0_ptr
, dadx_ptr
, dady_ptr
,
604 /* code generated texture sampling */
605 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
607 /* loop over quads in the block */
608 for(i
= 0; i
< num_fs
; ++i
) {
609 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
610 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
611 LLVMValueRef depth_ptr_i
;
613 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
615 generate_fs(lp
, shader
, key
,
622 &fs_mask
[i
], /* output */
630 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
631 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
632 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
635 sampler
->destroy(sampler
);
637 /* Loop over color outputs / color buffers to do blending.
639 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
640 LLVMValueRef color_ptr
;
641 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
642 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
646 * Convert the fs's output color and mask to fit to the blending type.
648 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
649 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
651 for (i
= 0; i
< num_fs
; i
++) {
653 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
656 lp_build_conv(builder
, fs_type
, blend_type
,
659 &blend_in_color
[chan
], 1);
661 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
664 if (partial_mask
|| !variant
->opaque
) {
665 lp_build_conv_mask(builder
, fs_type
, blend_type
,
669 blend_mask
= lp_build_const_int_vec(blend_type
, ~0);
672 color_ptr
= LLVMBuildLoad(builder
,
673 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
675 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
677 /* which blend/colormask state to use */
678 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
684 /* Could the 4x4 have been killed?
686 boolean do_branch
= ((key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
687 !key
->alpha
.enabled
&&
688 !shader
->info
.uses_kill
);
690 generate_blend(&key
->blend
,
703 /* Avoid corrupting the FPU stack on 32bit OSes. */
704 lp_build_intrinsic(builder
, "llvm.x86.mmx.emms", LLVMVoidType(), NULL
, 0);
707 LLVMBuildRetVoid(builder
);
709 LLVMDisposeBuilder(builder
);
712 /* Verify the LLVM IR. If invalid, dump and abort */
714 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
716 lp_debug_dump_value(function
);
721 /* Apply optimizations to LLVM IR */
722 LLVMRunFunctionPassManager(screen
->pass
, function
);
724 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
725 /* Print the LLVM IR to stderr */
726 lp_debug_dump_value(function
);
731 * Translate the LLVM IR into machine code.
734 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
736 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
738 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
741 lp_func_delete_body(function
);
747 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
751 debug_printf("fs variant %p:\n", (void *) key
);
753 if (key
->flatshade
) {
754 debug_printf("flatshade = 1\n");
756 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
757 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
759 if (key
->depth
.enabled
) {
760 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
761 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
762 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
765 for (i
= 0; i
< 2; ++i
) {
766 if (key
->stencil
[i
].enabled
) {
767 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
768 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
769 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
770 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
771 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
772 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
776 if (key
->alpha
.enabled
) {
777 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
780 if (key
->occlusion_count
) {
781 debug_printf("occlusion_count = 1\n");
784 if (key
->blend
.logicop_enable
) {
785 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
787 else if (key
->blend
.rt
[0].blend_enable
) {
788 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
789 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
790 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
791 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
792 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
793 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
795 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
796 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
797 debug_printf("sampler[%u] = \n", i
);
798 debug_printf(" .format = %s\n",
799 util_format_name(key
->sampler
[i
].format
));
800 debug_printf(" .target = %s\n",
801 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
802 debug_printf(" .pot = %u %u %u\n",
803 key
->sampler
[i
].pot_width
,
804 key
->sampler
[i
].pot_height
,
805 key
->sampler
[i
].pot_depth
);
806 debug_printf(" .wrap = %s %s %s\n",
807 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
808 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
809 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
810 debug_printf(" .min_img_filter = %s\n",
811 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
812 debug_printf(" .min_mip_filter = %s\n",
813 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
814 debug_printf(" .mag_img_filter = %s\n",
815 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
816 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
817 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
818 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
819 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
820 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
821 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
822 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
828 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
830 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
831 variant
->shader
->no
, variant
->no
);
832 tgsi_dump(variant
->shader
->base
.tokens
, 0);
833 dump_fs_variant_key(&variant
->key
);
834 debug_printf("variant->opaque = %u\n", variant
->opaque
);
838 static struct lp_fragment_shader_variant
*
839 generate_variant(struct llvmpipe_context
*lp
,
840 struct lp_fragment_shader
*shader
,
841 const struct lp_fragment_shader_variant_key
*key
)
843 struct lp_fragment_shader_variant
*variant
;
844 boolean fullcolormask
;
846 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
850 variant
->shader
= shader
;
851 variant
->list_item_global
.base
= variant
;
852 variant
->list_item_local
.base
= variant
;
853 variant
->no
= shader
->variants_created
++;
855 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
858 * Determine whether we are touching all channels in the color buffer.
860 fullcolormask
= FALSE
;
861 if (key
->nr_cbufs
== 1) {
862 const struct util_format_description
*format_desc
;
863 format_desc
= util_format_description(key
->cbuf_format
[0]);
864 if ((~key
->blend
.rt
[0].colormask
&
865 util_format_colormask(format_desc
)) == 0) {
866 fullcolormask
= TRUE
;
871 !key
->blend
.logicop_enable
&&
872 !key
->blend
.rt
[0].blend_enable
&&
874 !key
->stencil
[0].enabled
&&
875 !key
->alpha
.enabled
&&
876 !key
->depth
.enabled
&&
877 !shader
->info
.uses_kill
881 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
882 lp_debug_fs_variant(variant
);
885 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
887 if (variant
->opaque
) {
888 /* Specialized shader, which doesn't need to read the color buffer. */
889 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
891 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
899 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
900 const struct pipe_shader_state
*templ
)
902 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
903 struct lp_fragment_shader
*shader
;
906 shader
= CALLOC_STRUCT(lp_fragment_shader
);
910 shader
->no
= fs_no
++;
911 make_empty_list(&shader
->variants
);
913 /* get/save the summary info for this shader */
914 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
916 /* we need to keep a local copy of the tokens */
917 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
919 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
920 if (shader
->draw_data
== NULL
) {
921 FREE((void *) shader
->base
.tokens
);
926 nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
928 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
929 sampler
[nr_samplers
]);
931 if (LP_DEBUG
& DEBUG_TGSI
) {
933 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader
->no
, (void *) shader
);
934 tgsi_dump(templ
->tokens
, 0);
935 debug_printf("usage masks:\n");
936 for (attrib
= 0; attrib
< shader
->info
.num_inputs
; ++attrib
) {
937 unsigned usage_mask
= shader
->info
.input_usage_mask
[attrib
];
938 debug_printf(" IN[%u].%s%s%s%s\n",
940 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
941 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
942 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
943 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
953 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
955 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
957 if (llvmpipe
->fs
== fs
)
960 draw_flush(llvmpipe
->draw
);
962 draw_bind_fragment_shader(llvmpipe
->draw
,
963 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
967 llvmpipe
->dirty
|= LP_NEW_FS
;
971 remove_shader_variant(struct llvmpipe_context
*lp
,
972 struct lp_fragment_shader_variant
*variant
)
974 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
977 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
978 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
979 variant
->shader
->no
, variant
->no
, variant
->shader
->variants_created
,
980 variant
->shader
->variants_cached
, lp
->nr_fs_variants
);
982 for (i
= 0; i
< Elements(variant
->function
); i
++) {
983 if (variant
->function
[i
]) {
984 if (variant
->jit_function
[i
])
985 LLVMFreeMachineCodeForFunction(screen
->engine
,
986 variant
->function
[i
]);
987 LLVMDeleteFunction(variant
->function
[i
]);
990 remove_from_list(&variant
->list_item_local
);
991 variant
->shader
->variants_cached
--;
992 remove_from_list(&variant
->list_item_global
);
993 lp
->nr_fs_variants
--;
998 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1000 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1001 struct lp_fragment_shader
*shader
= fs
;
1002 struct lp_fs_variant_list_item
*li
;
1004 assert(fs
!= llvmpipe
->fs
);
1008 * XXX: we need to flush the context until we have some sort of reference
1009 * counting in fragment shaders as they may still be binned
1010 * Flushing alone might not sufficient we need to wait on it too.
1013 llvmpipe_finish(pipe
, __FUNCTION__
);
1015 li
= first_elem(&shader
->variants
);
1016 while(!at_end(&shader
->variants
, li
)) {
1017 struct lp_fs_variant_list_item
*next
= next_elem(li
);
1018 remove_shader_variant(llvmpipe
, li
->base
);
1022 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1024 assert(shader
->variants_cached
== 0);
1025 FREE((void *) shader
->base
.tokens
);
1032 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1033 uint shader
, uint index
,
1034 struct pipe_resource
*constants
)
1036 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1037 unsigned size
= constants
? constants
->width0
: 0;
1038 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1040 assert(shader
< PIPE_SHADER_TYPES
);
1041 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1043 if(llvmpipe
->constants
[shader
][index
] == constants
)
1046 draw_flush(llvmpipe
->draw
);
1048 /* note: reference counting */
1049 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1051 if(shader
== PIPE_SHADER_VERTEX
||
1052 shader
== PIPE_SHADER_GEOMETRY
) {
1053 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1057 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1062 * Return the blend factor equivalent to a destination alpha of one.
1064 static INLINE
unsigned
1065 force_dst_alpha_one(unsigned factor
)
1068 case PIPE_BLENDFACTOR_DST_ALPHA
:
1069 return PIPE_BLENDFACTOR_ONE
;
1070 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1071 return PIPE_BLENDFACTOR_ZERO
;
1072 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1073 return PIPE_BLENDFACTOR_ZERO
;
1081 * We need to generate several variants of the fragment pipeline to match
1082 * all the combinations of the contributing state atoms.
1084 * TODO: there is actually no reason to tie this to context state -- the
1085 * generated code could be cached globally in the screen.
1088 make_variant_key(struct llvmpipe_context
*lp
,
1089 struct lp_fragment_shader
*shader
,
1090 struct lp_fragment_shader_variant_key
*key
)
1094 memset(key
, 0, shader
->variant_key_size
);
1096 if (lp
->framebuffer
.zsbuf
) {
1097 if (lp
->depth_stencil
->depth
.enabled
) {
1098 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1099 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1101 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1102 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1103 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1107 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1108 if(key
->alpha
.enabled
)
1109 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1110 /* alpha.ref_value is passed in jit_context */
1112 key
->flatshade
= lp
->rasterizer
->flatshade
;
1113 if (lp
->active_query_count
) {
1114 key
->occlusion_count
= TRUE
;
1117 if (lp
->framebuffer
.nr_cbufs
) {
1118 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1121 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1122 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1123 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1124 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1125 const struct util_format_description
*format_desc
;
1127 key
->cbuf_format
[i
] = format
;
1129 format_desc
= util_format_description(format
);
1130 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1131 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1133 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1136 * Mask out color channels not present in the color buffer.
1138 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1141 * Our swizzled render tiles always have an alpha channel, but the linear
1142 * render target format often does not, so force here the dst alpha to be
1145 * This is not a mere optimization. Wrong results will be produced if the
1146 * dst alpha is used, the dst format does not have alpha, and the previous
1147 * rendering was not flushed from the swizzled to linear buffer. For
1148 * example, NonPowTwo DCT.
1150 * TODO: This should be generalized to all channels for better
1151 * performance, but only alpha causes correctness issues.
1153 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1155 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1156 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1157 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1158 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1159 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1160 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1164 /* This value will be the same for all the variants of a given shader:
1166 key
->nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1168 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1169 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1170 lp_sampler_static_state(&key
->sampler
[i
],
1171 lp
->fragment_sampler_views
[i
],
1178 * Update fragment state. This is called just prior to drawing
1179 * something when some fragment-related state has changed.
1182 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1184 struct lp_fragment_shader
*shader
= lp
->fs
;
1185 struct lp_fragment_shader_variant_key key
;
1186 struct lp_fragment_shader_variant
*variant
= NULL
;
1187 struct lp_fs_variant_list_item
*li
;
1189 make_variant_key(lp
, shader
, &key
);
1191 li
= first_elem(&shader
->variants
);
1192 while(!at_end(&shader
->variants
, li
)) {
1193 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1201 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1207 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1208 struct pipe_context
*pipe
= &lp
->pipe
;
1211 * XXX: we need to flush the context until we have some sort of reference
1212 * counting in fragment shaders as they may still be binned
1213 * Flushing alone might not be sufficient we need to wait on it too.
1215 llvmpipe_finish(pipe
, __FUNCTION__
);
1217 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1218 struct lp_fs_variant_list_item
*item
= last_elem(&lp
->fs_variants_list
);
1219 remove_shader_variant(lp
, item
->base
);
1224 variant
= generate_variant(lp
, shader
, &key
);
1228 LP_COUNT_ADD(llvm_compile_time
, dt
);
1229 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1232 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1233 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1234 lp
->nr_fs_variants
++;
1235 shader
->variants_cached
++;
1239 lp_setup_set_fs_variant(lp
->setup
, variant
);
1245 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1247 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1248 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1249 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1251 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;