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
)
121 const struct util_format_description
*format_desc
;
123 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
126 format_desc
= util_format_description(key
->zsbuf_format
);
129 lp_build_depth_stencil_test(builder
,
144 * Expand the relevent bits of mask_input to a 4-dword mask for the
145 * four pixels in a 2x2 quad. This will set the four elements of the
146 * quad mask vector to 0 or ~0.
148 * \param quad which quad of the quad group to test, in [0,3]
149 * \param mask_input bitwise mask for the whole 4x4 stamp
152 generate_quad_mask(LLVMBuilderRef builder
,
153 struct lp_type fs_type
,
155 LLVMValueRef mask_input
) /* int32 */
157 struct lp_type mask_type
;
158 LLVMTypeRef i32t
= LLVMInt32Type();
159 LLVMValueRef bits
[4];
164 * XXX: We'll need a different path for 16 x u8
166 assert(fs_type
.width
== 32);
167 assert(fs_type
.length
== 4);
168 mask_type
= lp_int_type(fs_type
);
171 * mask_input >>= (quad * 4)
192 mask_input
= LLVMBuildLShr(builder
,
194 LLVMConstInt(i32t
, shift
, 0),
198 * mask = { mask_input & (1 << i), for i in [0,3] }
201 mask
= lp_build_broadcast(builder
, lp_build_vec_type(mask_type
), mask_input
);
203 bits
[0] = LLVMConstInt(i32t
, 1 << 0, 0);
204 bits
[1] = LLVMConstInt(i32t
, 1 << 1, 0);
205 bits
[2] = LLVMConstInt(i32t
, 1 << 4, 0);
206 bits
[3] = LLVMConstInt(i32t
, 1 << 5, 0);
208 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, 4), "");
211 * mask = mask != 0 ? ~0 : 0
214 mask
= lp_build_compare(builder
,
215 mask_type
, PIPE_FUNC_NOTEQUAL
,
217 lp_build_const_int_vec(mask_type
, 0));
225 * Generate the fragment shader, depth/stencil test, and alpha tests.
226 * \param i which quad in the tile, in range [0,3]
227 * \param partial_mask if 1, do mask_input testing
230 generate_fs(struct llvmpipe_context
*lp
,
231 struct lp_fragment_shader
*shader
,
232 const struct lp_fragment_shader_variant_key
*key
,
233 LLVMBuilderRef builder
,
235 LLVMValueRef context_ptr
,
237 struct lp_build_interp_soa_context
*interp
,
238 struct lp_build_sampler_soa
*sampler
,
240 LLVMValueRef (*color
)[4],
241 LLVMValueRef depth_ptr
,
243 unsigned partial_mask
,
244 LLVMValueRef mask_input
,
245 LLVMValueRef counter
)
247 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
248 LLVMTypeRef vec_type
;
249 LLVMValueRef consts_ptr
;
250 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
252 LLVMValueRef stencil_refs
[2];
253 struct lp_build_flow_context
*flow
;
254 struct lp_build_mask_context mask
;
255 boolean early_depth_stencil_test
;
262 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
263 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
265 vec_type
= lp_build_vec_type(type
);
267 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
269 flow
= lp_build_flow_create(builder
);
271 memset(outputs
, 0, sizeof outputs
);
273 lp_build_flow_scope_begin(flow
);
275 /* Declare the color and z variables */
276 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
277 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
278 color
[cbuf
][chan
] = lp_build_alloca(builder
, vec_type
, "color");
282 /* do triangle edge testing */
284 *pmask
= generate_quad_mask(builder
, type
,
288 *pmask
= lp_build_const_int_vec(type
, ~0);
291 /* 'mask' will control execution based on quad's pixel alive/killed state */
292 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
294 lp_build_interp_soa_update_pos(interp
, i
);
296 /* Try to avoid the 1/w for quads where mask is zero. TODO: avoid
297 * this for depth-fail quads also.
301 early_depth_stencil_test
=
302 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
303 !key
->alpha
.enabled
&&
304 !shader
->info
.uses_kill
&&
305 !shader
->info
.writes_z
;
307 if (early_depth_stencil_test
)
308 generate_depth_stencil(builder
, key
,
310 stencil_refs
, z
, depth_ptr
, facing
, counter
);
312 lp_build_interp_soa_update_inputs(interp
, i
);
314 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
315 consts_ptr
, interp
->pos
, interp
->inputs
,
316 outputs
, sampler
, &shader
->info
);
318 /* loop over fragment shader outputs/results */
319 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
320 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
321 if(outputs
[attrib
][chan
]) {
322 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
323 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
325 switch (shader
->info
.output_semantic_name
[attrib
]) {
326 case TGSI_SEMANTIC_COLOR
:
328 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
330 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
333 /* XXX: should only test the final assignment to alpha */
334 if (cbuf
== 0 && chan
== 3 && key
->alpha
.enabled
) {
335 LLVMValueRef alpha
= out
;
336 LLVMValueRef alpha_ref_value
;
337 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
338 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
339 lp_build_alpha_test(builder
, key
->alpha
.func
, type
,
340 &mask
, alpha
, alpha_ref_value
);
343 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
347 case TGSI_SEMANTIC_POSITION
:
356 if (!early_depth_stencil_test
)
357 generate_depth_stencil(builder
, key
,
359 stencil_refs
, z
, depth_ptr
, facing
, counter
);
361 lp_build_mask_end(&mask
);
363 lp_build_flow_scope_end(flow
);
365 lp_build_flow_destroy(flow
);
373 * Generate color blending and color output.
374 * \param rt the render target index (to index blend, colormask state)
375 * \param type the pixel color type
376 * \param context_ptr pointer to the runtime JIT context
377 * \param mask execution mask (active fragment/pixel mask)
378 * \param src colors from the fragment shader
379 * \param dst_ptr the destination color buffer pointer
382 generate_blend(const struct pipe_blend_state
*blend
,
384 LLVMBuilderRef builder
,
386 LLVMValueRef context_ptr
,
389 LLVMValueRef dst_ptr
)
391 struct lp_build_context bld
;
392 struct lp_build_flow_context
*flow
;
393 struct lp_build_mask_context mask_ctx
;
394 LLVMTypeRef vec_type
;
395 LLVMValueRef const_ptr
;
401 lp_build_context_init(&bld
, builder
, type
);
403 flow
= lp_build_flow_create(builder
);
405 /* we'll use this mask context to skip blending if all pixels are dead */
406 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
408 vec_type
= lp_build_vec_type(type
);
410 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
411 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
412 LLVMPointerType(vec_type
, 0), "");
414 /* load constant blend color and colors from the dest color buffer */
415 for(chan
= 0; chan
< 4; ++chan
) {
416 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
417 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
419 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
421 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
422 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
426 lp_build_blend_soa(builder
, blend
, type
, rt
, src
, dst
, con
, res
);
428 /* store results to color buffer */
429 for(chan
= 0; chan
< 4; ++chan
) {
430 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
431 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
432 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
433 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
434 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
438 lp_build_mask_end(&mask_ctx
);
439 lp_build_flow_destroy(flow
);
444 * Generate the runtime callable function for the whole fragment pipeline.
445 * Note that the function which we generate operates on a block of 16
446 * pixels at at time. The block contains 2x2 quads. Each quad contains
450 generate_fragment(struct llvmpipe_context
*lp
,
451 struct lp_fragment_shader
*shader
,
452 struct lp_fragment_shader_variant
*variant
,
453 unsigned partial_mask
)
455 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
456 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
458 struct lp_type fs_type
;
459 struct lp_type blend_type
;
460 LLVMTypeRef fs_elem_type
;
461 LLVMTypeRef fs_int_vec_type
;
462 LLVMTypeRef blend_vec_type
;
463 LLVMTypeRef arg_types
[11];
464 LLVMTypeRef func_type
;
465 LLVMValueRef context_ptr
;
469 LLVMValueRef dadx_ptr
;
470 LLVMValueRef dady_ptr
;
471 LLVMValueRef color_ptr_ptr
;
472 LLVMValueRef depth_ptr
;
473 LLVMValueRef mask_input
;
474 LLVMValueRef counter
= NULL
;
475 LLVMBasicBlockRef block
;
476 LLVMBuilderRef builder
;
477 struct lp_build_sampler_soa
*sampler
;
478 struct lp_build_interp_soa_context interp
;
479 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
480 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
481 LLVMValueRef blend_mask
;
482 LLVMValueRef function
;
490 /* TODO: actually pick these based on the fs and color buffer
491 * characteristics. */
493 memset(&fs_type
, 0, sizeof fs_type
);
494 fs_type
.floating
= TRUE
; /* floating point values */
495 fs_type
.sign
= TRUE
; /* values are signed */
496 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
497 fs_type
.width
= 32; /* 32-bit float */
498 fs_type
.length
= 4; /* 4 elements per vector */
499 num_fs
= 4; /* number of quads per block */
501 memset(&blend_type
, 0, sizeof blend_type
);
502 blend_type
.floating
= FALSE
; /* values are integers */
503 blend_type
.sign
= FALSE
; /* values are unsigned */
504 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
505 blend_type
.width
= 8; /* 8-bit ubyte values */
506 blend_type
.length
= 16; /* 16 elements per vector */
509 * Generate the function prototype. Any change here must be reflected in
510 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
513 fs_elem_type
= lp_build_elem_type(fs_type
);
514 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
516 blend_vec_type
= lp_build_vec_type(blend_type
);
518 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
519 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
521 arg_types
[0] = screen
->context_ptr_type
; /* context */
522 arg_types
[1] = LLVMInt32Type(); /* x */
523 arg_types
[2] = LLVMInt32Type(); /* y */
524 arg_types
[3] = LLVMFloatType(); /* facing */
525 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
526 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
527 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
528 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
529 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
530 arg_types
[9] = LLVMInt32Type(); /* mask_input */
531 arg_types
[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
533 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
535 function
= LLVMAddFunction(screen
->module
, func_name
, func_type
);
536 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
538 variant
->function
[partial_mask
] = function
;
541 /* XXX: need to propagate noalias down into color param now we are
542 * passing a pointer-to-pointer?
544 for(i
= 0; i
< Elements(arg_types
); ++i
)
545 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
546 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
548 context_ptr
= LLVMGetParam(function
, 0);
549 x
= LLVMGetParam(function
, 1);
550 y
= LLVMGetParam(function
, 2);
551 facing
= LLVMGetParam(function
, 3);
552 a0_ptr
= LLVMGetParam(function
, 4);
553 dadx_ptr
= LLVMGetParam(function
, 5);
554 dady_ptr
= LLVMGetParam(function
, 6);
555 color_ptr_ptr
= LLVMGetParam(function
, 7);
556 depth_ptr
= LLVMGetParam(function
, 8);
557 mask_input
= LLVMGetParam(function
, 9);
559 lp_build_name(context_ptr
, "context");
560 lp_build_name(x
, "x");
561 lp_build_name(y
, "y");
562 lp_build_name(a0_ptr
, "a0");
563 lp_build_name(dadx_ptr
, "dadx");
564 lp_build_name(dady_ptr
, "dady");
565 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
566 lp_build_name(depth_ptr
, "depth");
567 lp_build_name(mask_input
, "mask_input");
569 if (key
->occlusion_count
) {
570 counter
= LLVMGetParam(function
, 10);
571 lp_build_name(counter
, "counter");
578 block
= LLVMAppendBasicBlock(function
, "entry");
579 builder
= LLVMCreateBuilder();
580 LLVMPositionBuilderAtEnd(builder
, block
);
583 * The shader input interpolation info is not explicitely baked in the
584 * shader key, but everything it derives from (TGSI, and flatshade) is
585 * already included in the shader key.
587 lp_build_interp_soa_init(&interp
,
591 a0_ptr
, dadx_ptr
, dady_ptr
,
594 /* code generated texture sampling */
595 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
597 /* loop over quads in the block */
598 for(i
= 0; i
< num_fs
; ++i
) {
599 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
600 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
601 LLVMValueRef depth_ptr_i
;
603 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
605 generate_fs(lp
, shader
, key
,
612 &fs_mask
[i
], /* output */
620 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
621 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
622 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
625 sampler
->destroy(sampler
);
627 /* Loop over color outputs / color buffers to do blending.
629 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
630 LLVMValueRef color_ptr
;
631 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
632 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
636 * Convert the fs's output color and mask to fit to the blending type.
638 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
639 LLVMValueRef fs_color_vals
[LP_MAX_VECTOR_LENGTH
];
641 for (i
= 0; i
< num_fs
; i
++) {
643 LLVMBuildLoad(builder
, fs_out_color
[cbuf
][chan
][i
], "fs_color_vals");
646 lp_build_conv(builder
, fs_type
, blend_type
,
649 &blend_in_color
[chan
], 1);
651 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
654 if (partial_mask
|| !variant
->opaque
) {
655 lp_build_conv_mask(builder
, fs_type
, blend_type
,
659 blend_mask
= lp_build_const_int_vec(blend_type
, ~0);
662 color_ptr
= LLVMBuildLoad(builder
,
663 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
665 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
667 /* which blend/colormask state to use */
668 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
673 generate_blend(&key
->blend
,
684 /* Avoid corrupting the FPU stack on 32bit OSes. */
685 lp_build_intrinsic(builder
, "llvm.x86.mmx.emms", LLVMVoidType(), NULL
, 0);
688 LLVMBuildRetVoid(builder
);
690 LLVMDisposeBuilder(builder
);
693 /* Verify the LLVM IR. If invalid, dump and abort */
695 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
697 lp_debug_dump_value(function
);
702 /* Apply optimizations to LLVM IR */
703 LLVMRunFunctionPassManager(screen
->pass
, function
);
705 if ((gallivm_debug
& GALLIVM_DEBUG_IR
) || (LP_DEBUG
& DEBUG_FS
)) {
706 /* Print the LLVM IR to stderr */
707 lp_debug_dump_value(function
);
712 * Translate the LLVM IR into machine code.
715 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
717 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
719 if ((gallivm_debug
& GALLIVM_DEBUG_ASM
) || (LP_DEBUG
& DEBUG_FS
)) {
722 lp_func_delete_body(function
);
728 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
732 debug_printf("fs variant %p:\n", (void *) key
);
734 if (key
->flatshade
) {
735 debug_printf("flatshade = 1\n");
737 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
738 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
740 if (key
->depth
.enabled
) {
741 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
742 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
743 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
746 for (i
= 0; i
< 2; ++i
) {
747 if (key
->stencil
[i
].enabled
) {
748 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
749 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
750 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
751 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
752 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
753 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
757 if (key
->alpha
.enabled
) {
758 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
761 if (key
->occlusion_count
) {
762 debug_printf("occlusion_count = 1\n");
765 if (key
->blend
.logicop_enable
) {
766 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
768 else if (key
->blend
.rt
[0].blend_enable
) {
769 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
770 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
771 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
772 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
773 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
774 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
776 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
777 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
778 debug_printf("sampler[%u] = \n", i
);
779 debug_printf(" .format = %s\n",
780 util_format_name(key
->sampler
[i
].format
));
781 debug_printf(" .target = %s\n",
782 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
783 debug_printf(" .pot = %u %u %u\n",
784 key
->sampler
[i
].pot_width
,
785 key
->sampler
[i
].pot_height
,
786 key
->sampler
[i
].pot_depth
);
787 debug_printf(" .wrap = %s %s %s\n",
788 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
789 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
790 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
791 debug_printf(" .min_img_filter = %s\n",
792 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
793 debug_printf(" .min_mip_filter = %s\n",
794 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
795 debug_printf(" .mag_img_filter = %s\n",
796 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
797 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
798 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
799 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
800 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
801 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
802 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
803 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
809 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
811 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
812 variant
->shader
->no
, variant
->no
);
813 tgsi_dump(variant
->shader
->base
.tokens
, 0);
814 dump_fs_variant_key(&variant
->key
);
815 debug_printf("variant->opaque = %u\n", variant
->opaque
);
819 static struct lp_fragment_shader_variant
*
820 generate_variant(struct llvmpipe_context
*lp
,
821 struct lp_fragment_shader
*shader
,
822 const struct lp_fragment_shader_variant_key
*key
)
824 struct lp_fragment_shader_variant
*variant
;
825 boolean fullcolormask
;
827 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
831 variant
->shader
= shader
;
832 variant
->list_item_global
.base
= variant
;
833 variant
->list_item_local
.base
= variant
;
834 variant
->no
= shader
->variants_created
++;
836 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
839 * Determine whether we are touching all channels in the color buffer.
841 fullcolormask
= FALSE
;
842 if (key
->nr_cbufs
== 1) {
843 const struct util_format_description
*format_desc
;
844 format_desc
= util_format_description(key
->cbuf_format
[0]);
845 if ((~key
->blend
.rt
[0].colormask
&
846 util_format_colormask(format_desc
)) == 0) {
847 fullcolormask
= TRUE
;
852 !key
->blend
.logicop_enable
&&
853 !key
->blend
.rt
[0].blend_enable
&&
855 !key
->stencil
[0].enabled
&&
856 !key
->alpha
.enabled
&&
857 !key
->depth
.enabled
&&
858 !shader
->info
.uses_kill
862 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
863 lp_debug_fs_variant(variant
);
866 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
868 if (variant
->opaque
) {
869 /* Specialized shader, which doesn't need to read the color buffer. */
870 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
872 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
880 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
881 const struct pipe_shader_state
*templ
)
883 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
884 struct lp_fragment_shader
*shader
;
887 shader
= CALLOC_STRUCT(lp_fragment_shader
);
891 shader
->no
= fs_no
++;
892 make_empty_list(&shader
->variants
);
894 /* get/save the summary info for this shader */
895 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
897 /* we need to keep a local copy of the tokens */
898 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
900 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
901 if (shader
->draw_data
== NULL
) {
902 FREE((void *) shader
->base
.tokens
);
907 nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
909 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
910 sampler
[nr_samplers
]);
912 if (LP_DEBUG
& DEBUG_TGSI
) {
914 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader
->no
, (void *) shader
);
915 tgsi_dump(templ
->tokens
, 0);
916 debug_printf("usage masks:\n");
917 for (attrib
= 0; attrib
< shader
->info
.num_inputs
; ++attrib
) {
918 unsigned usage_mask
= shader
->info
.input_usage_mask
[attrib
];
919 debug_printf(" IN[%u].%s%s%s%s\n",
921 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
922 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
923 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
924 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
934 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
936 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
938 if (llvmpipe
->fs
== fs
)
941 draw_flush(llvmpipe
->draw
);
943 draw_bind_fragment_shader(llvmpipe
->draw
,
944 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
948 llvmpipe
->dirty
|= LP_NEW_FS
;
952 remove_shader_variant(struct llvmpipe_context
*lp
,
953 struct lp_fragment_shader_variant
*variant
)
955 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
958 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
959 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
960 variant
->shader
->no
, variant
->no
, variant
->shader
->variants_created
,
961 variant
->shader
->variants_cached
, lp
->nr_fs_variants
);
963 for (i
= 0; i
< Elements(variant
->function
); i
++) {
964 if (variant
->function
[i
]) {
965 if (variant
->jit_function
[i
])
966 LLVMFreeMachineCodeForFunction(screen
->engine
,
967 variant
->function
[i
]);
968 LLVMDeleteFunction(variant
->function
[i
]);
971 remove_from_list(&variant
->list_item_local
);
972 variant
->shader
->variants_cached
--;
973 remove_from_list(&variant
->list_item_global
);
974 lp
->nr_fs_variants
--;
979 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
981 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
982 struct lp_fragment_shader
*shader
= fs
;
983 struct lp_fs_variant_list_item
*li
;
985 assert(fs
!= llvmpipe
->fs
);
989 * XXX: we need to flush the context until we have some sort of reference
990 * counting in fragment shaders as they may still be binned
991 * Flushing alone might not sufficient we need to wait on it too.
994 llvmpipe_finish(pipe
, __FUNCTION__
);
996 li
= first_elem(&shader
->variants
);
997 while(!at_end(&shader
->variants
, li
)) {
998 struct lp_fs_variant_list_item
*next
= next_elem(li
);
999 remove_shader_variant(llvmpipe
, li
->base
);
1003 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
1005 assert(shader
->variants_cached
== 0);
1006 FREE((void *) shader
->base
.tokens
);
1013 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1014 uint shader
, uint index
,
1015 struct pipe_resource
*constants
)
1017 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1018 unsigned size
= constants
? constants
->width0
: 0;
1019 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1021 assert(shader
< PIPE_SHADER_TYPES
);
1022 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
1024 if(llvmpipe
->constants
[shader
][index
] == constants
)
1027 draw_flush(llvmpipe
->draw
);
1029 /* note: reference counting */
1030 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
1032 if(shader
== PIPE_SHADER_VERTEX
||
1033 shader
== PIPE_SHADER_GEOMETRY
) {
1034 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
1038 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1043 * Return the blend factor equivalent to a destination alpha of one.
1045 static INLINE
unsigned
1046 force_dst_alpha_one(unsigned factor
)
1049 case PIPE_BLENDFACTOR_DST_ALPHA
:
1050 return PIPE_BLENDFACTOR_ONE
;
1051 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1052 return PIPE_BLENDFACTOR_ZERO
;
1053 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1054 return PIPE_BLENDFACTOR_ZERO
;
1062 * We need to generate several variants of the fragment pipeline to match
1063 * all the combinations of the contributing state atoms.
1065 * TODO: there is actually no reason to tie this to context state -- the
1066 * generated code could be cached globally in the screen.
1069 make_variant_key(struct llvmpipe_context
*lp
,
1070 struct lp_fragment_shader
*shader
,
1071 struct lp_fragment_shader_variant_key
*key
)
1075 memset(key
, 0, shader
->variant_key_size
);
1077 if (lp
->framebuffer
.zsbuf
) {
1078 if (lp
->depth_stencil
->depth
.enabled
) {
1079 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1080 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1082 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1083 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1084 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1088 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1089 if(key
->alpha
.enabled
)
1090 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1091 /* alpha.ref_value is passed in jit_context */
1093 key
->flatshade
= lp
->rasterizer
->flatshade
;
1094 if (lp
->active_query_count
) {
1095 key
->occlusion_count
= TRUE
;
1098 if (lp
->framebuffer
.nr_cbufs
) {
1099 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1102 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1103 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1104 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
1105 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1106 const struct util_format_description
*format_desc
;
1108 key
->cbuf_format
[i
] = format
;
1110 format_desc
= util_format_description(format
);
1111 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1112 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1114 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1117 * Mask out color channels not present in the color buffer.
1119 blend_rt
->colormask
&= util_format_colormask(format_desc
);
1122 * Our swizzled render tiles always have an alpha channel, but the linear
1123 * render target format often does not, so force here the dst alpha to be
1126 * This is not a mere optimization. Wrong results will be produced if the
1127 * dst alpha is used, the dst format does not have alpha, and the previous
1128 * rendering was not flushed from the swizzled to linear buffer. For
1129 * example, NonPowTwo DCT.
1131 * TODO: This should be generalized to all channels for better
1132 * performance, but only alpha causes correctness issues.
1134 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1136 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1137 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
1138 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
1139 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
1140 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
1141 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
1145 /* This value will be the same for all the variants of a given shader:
1147 key
->nr_samplers
= shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
1149 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
1150 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
1151 lp_sampler_static_state(&key
->sampler
[i
],
1152 lp
->fragment_sampler_views
[i
],
1159 * Update fragment state. This is called just prior to drawing
1160 * something when some fragment-related state has changed.
1163 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1165 struct lp_fragment_shader
*shader
= lp
->fs
;
1166 struct lp_fragment_shader_variant_key key
;
1167 struct lp_fragment_shader_variant
*variant
= NULL
;
1168 struct lp_fs_variant_list_item
*li
;
1170 make_variant_key(lp
, shader
, &key
);
1172 li
= first_elem(&shader
->variants
);
1173 while(!at_end(&shader
->variants
, li
)) {
1174 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
1182 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1188 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1189 struct pipe_context
*pipe
= &lp
->pipe
;
1192 * XXX: we need to flush the context until we have some sort of reference
1193 * counting in fragment shaders as they may still be binned
1194 * Flushing alone might not be sufficient we need to wait on it too.
1196 llvmpipe_finish(pipe
, __FUNCTION__
);
1198 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1199 struct lp_fs_variant_list_item
*item
= last_elem(&lp
->fs_variants_list
);
1200 remove_shader_variant(lp
, item
->base
);
1205 variant
= generate_variant(lp
, shader
, &key
);
1209 LP_COUNT_ADD(llvm_compile_time
, dt
);
1210 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1213 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1214 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1215 lp
->nr_fs_variants
++;
1216 shader
->variants_cached
++;
1220 lp_setup_set_fs_variant(lp
->setup
, variant
);
1226 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1228 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1229 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1230 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1232 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;