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_intr.h"
79 #include "gallivm/lp_bld_logic.h"
80 #include "gallivm/lp_bld_tgsi.h"
81 #include "gallivm/lp_bld_swizzle.h"
82 #include "gallivm/lp_bld_flow.h"
83 #include "gallivm/lp_bld_debug.h"
85 #include "lp_bld_alpha.h"
86 #include "lp_bld_blend.h"
87 #include "lp_bld_depth.h"
88 #include "lp_bld_interp.h"
89 #include "lp_context.h"
92 #include "lp_screen.h"
95 #include "lp_tex_sample.h"
97 #include "lp_state_fs.h"
100 #include <llvm-c/Analysis.h>
103 static unsigned fs_no
= 0;
107 * Generate the depth /stencil test code.
110 generate_depth_stencil(LLVMBuilderRef builder
,
111 const struct lp_fragment_shader_variant_key
*key
,
112 struct lp_type src_type
,
113 struct lp_build_mask_context
*mask
,
114 LLVMValueRef stencil_refs
[2],
116 LLVMValueRef dst_ptr
,
118 LLVMValueRef counter
)
120 const struct util_format_description
*format_desc
;
121 struct lp_type dst_type
;
123 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
126 format_desc
= util_format_description(key
->zsbuf_format
);
130 * Depths are expected to be between 0 and 1, even if they are stored in
131 * floats. Setting these bits here will ensure that the lp_build_conv() call
132 * below won't try to unnecessarily clamp the incoming values.
134 if(src_type
.floating
) {
135 src_type
.sign
= FALSE
;
136 src_type
.norm
= TRUE
;
139 assert(!src_type
.sign
);
140 assert(src_type
.norm
);
143 /* Pick the depth type. */
144 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
146 /* FIXME: Cope with a depth test type with a different bit width. */
147 assert(dst_type
.width
== src_type
.width
);
148 assert(dst_type
.length
== src_type
.length
);
150 /* Convert fragment Z from float to integer */
151 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
153 dst_ptr
= LLVMBuildBitCast(builder
,
155 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
156 lp_build_depth_stencil_test(builder
,
171 * Expand the relevent bits of mask_input to a 4-dword mask for the
172 * four pixels in a 2x2 quad. This will set the four elements of the
173 * quad mask vector to 0 or ~0.
175 * \param quad which quad of the quad group to test, in [0,3]
176 * \param mask_input bitwise mask for the whole 4x4 stamp
179 generate_quad_mask(LLVMBuilderRef builder
,
180 struct lp_type fs_type
,
182 LLVMValueRef mask_input
) /* int32 */
184 struct lp_type mask_type
;
185 LLVMTypeRef i32t
= LLVMInt32Type();
186 LLVMValueRef bits
[4];
190 * XXX: We'll need a different path for 16 x u8
192 assert(fs_type
.width
== 32);
193 assert(fs_type
.length
== 4);
194 mask_type
= lp_int_type(fs_type
);
197 * mask_input >>= (quad * 4)
200 mask_input
= LLVMBuildLShr(builder
,
202 LLVMConstInt(i32t
, quad
* 4, 0),
206 * mask = { mask_input & (1 << i), for i in [0,3] }
209 mask
= lp_build_broadcast(builder
, lp_build_vec_type(mask_type
), mask_input
);
211 bits
[0] = LLVMConstInt(i32t
, 1 << 0, 0);
212 bits
[1] = LLVMConstInt(i32t
, 1 << 1, 0);
213 bits
[2] = LLVMConstInt(i32t
, 1 << 2, 0);
214 bits
[3] = LLVMConstInt(i32t
, 1 << 3, 0);
216 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, 4), "");
219 * mask = mask != 0 ? ~0 : 0
222 mask
= lp_build_compare(builder
,
223 mask_type
, PIPE_FUNC_NOTEQUAL
,
225 lp_build_const_int_vec(mask_type
, 0));
233 * Generate the fragment shader, depth/stencil test, and alpha tests.
234 * \param i which quad in the tile, in range [0,3]
235 * \param partial_mask if 1, do mask_input testing
238 generate_fs(struct llvmpipe_context
*lp
,
239 struct lp_fragment_shader
*shader
,
240 const struct lp_fragment_shader_variant_key
*key
,
241 LLVMBuilderRef builder
,
243 LLVMValueRef context_ptr
,
245 const struct lp_build_interp_soa_context
*interp
,
246 struct lp_build_sampler_soa
*sampler
,
248 LLVMValueRef (*color
)[4],
249 LLVMValueRef depth_ptr
,
251 unsigned partial_mask
,
252 LLVMValueRef mask_input
,
253 LLVMValueRef counter
)
255 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
256 LLVMTypeRef vec_type
;
257 LLVMValueRef consts_ptr
;
258 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
259 LLVMValueRef z
= interp
->pos
[2];
260 LLVMValueRef stencil_refs
[2];
261 struct lp_build_flow_context
*flow
;
262 struct lp_build_mask_context mask
;
263 boolean early_depth_stencil_test
;
270 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
271 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
273 vec_type
= lp_build_vec_type(type
);
275 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
277 flow
= lp_build_flow_create(builder
);
279 memset(outputs
, 0, sizeof outputs
);
281 lp_build_flow_scope_begin(flow
);
283 /* Declare the color and z variables */
284 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
285 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
286 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
287 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
290 lp_build_flow_scope_declare(flow
, &z
);
292 /* do triangle edge testing */
294 *pmask
= generate_quad_mask(builder
, type
,
298 *pmask
= lp_build_const_int_vec(type
, ~0);
301 /* 'mask' will control execution based on quad's pixel alive/killed state */
302 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
304 early_depth_stencil_test
=
305 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
306 !key
->alpha
.enabled
&&
307 !shader
->info
.uses_kill
&&
308 !shader
->info
.writes_z
;
310 if (early_depth_stencil_test
)
311 generate_depth_stencil(builder
, key
,
313 stencil_refs
, z
, depth_ptr
, facing
, counter
);
315 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
316 consts_ptr
, interp
->pos
, interp
->inputs
,
317 outputs
, sampler
, &shader
->info
);
319 /* loop over fragment shader outputs/results */
320 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
321 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
322 if(outputs
[attrib
][chan
]) {
323 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
324 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
326 switch (shader
->info
.output_semantic_name
[attrib
]) {
327 case TGSI_SEMANTIC_COLOR
:
329 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
331 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
334 /* XXX: should the alpha reference value be passed separately? */
335 /* XXX: should only test the final assignment to alpha */
336 if(cbuf
== 0 && chan
== 3) {
337 LLVMValueRef alpha
= out
;
338 LLVMValueRef alpha_ref_value
;
339 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
340 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
341 lp_build_alpha_test(builder
, &key
->alpha
, type
,
342 &mask
, alpha
, alpha_ref_value
);
345 color
[cbuf
][chan
] = out
;
349 case TGSI_SEMANTIC_POSITION
:
358 if (!early_depth_stencil_test
)
359 generate_depth_stencil(builder
, key
,
361 stencil_refs
, z
, depth_ptr
, facing
, counter
);
363 lp_build_mask_end(&mask
);
365 lp_build_flow_scope_end(flow
);
367 lp_build_flow_destroy(flow
);
375 * Generate color blending and color output.
376 * \param rt the render target index (to index blend, colormask state)
377 * \param type the pixel color type
378 * \param context_ptr pointer to the runtime JIT context
379 * \param mask execution mask (active fragment/pixel mask)
380 * \param src colors from the fragment shader
381 * \param dst_ptr the destination color buffer pointer
384 generate_blend(const struct pipe_blend_state
*blend
,
386 LLVMBuilderRef builder
,
388 LLVMValueRef context_ptr
,
391 LLVMValueRef dst_ptr
)
393 struct lp_build_context bld
;
394 struct lp_build_flow_context
*flow
;
395 struct lp_build_mask_context mask_ctx
;
396 LLVMTypeRef vec_type
;
397 LLVMValueRef const_ptr
;
403 lp_build_context_init(&bld
, builder
, type
);
405 flow
= lp_build_flow_create(builder
);
407 /* we'll use this mask context to skip blending if all pixels are dead */
408 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
410 vec_type
= lp_build_vec_type(type
);
412 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
413 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
414 LLVMPointerType(vec_type
, 0), "");
416 /* load constant blend color and colors from the dest color buffer */
417 for(chan
= 0; chan
< 4; ++chan
) {
418 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
419 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
421 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
423 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
424 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
428 lp_build_blend_soa(builder
, blend
, type
, rt
, src
, dst
, con
, res
);
430 /* store results to color buffer */
431 for(chan
= 0; chan
< 4; ++chan
) {
432 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
433 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
434 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
435 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
436 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
440 lp_build_mask_end(&mask_ctx
);
441 lp_build_flow_destroy(flow
);
446 * Generate the runtime callable function for the whole fragment pipeline.
447 * Note that the function which we generate operates on a block of 16
448 * pixels at at time. The block contains 2x2 quads. Each quad contains
452 generate_fragment(struct llvmpipe_context
*lp
,
453 struct lp_fragment_shader
*shader
,
454 struct lp_fragment_shader_variant
*variant
,
455 unsigned partial_mask
)
457 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
458 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
460 struct lp_type fs_type
;
461 struct lp_type blend_type
;
462 LLVMTypeRef fs_elem_type
;
463 LLVMTypeRef fs_int_vec_type
;
464 LLVMTypeRef blend_vec_type
;
465 LLVMTypeRef arg_types
[11];
466 LLVMTypeRef func_type
;
467 LLVMValueRef context_ptr
;
471 LLVMValueRef dadx_ptr
;
472 LLVMValueRef dady_ptr
;
473 LLVMValueRef color_ptr_ptr
;
474 LLVMValueRef depth_ptr
;
475 LLVMValueRef mask_input
;
476 LLVMValueRef counter
= NULL
;
477 LLVMBasicBlockRef block
;
478 LLVMBuilderRef builder
;
479 struct lp_build_sampler_soa
*sampler
;
480 struct lp_build_interp_soa_context interp
;
481 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
482 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
483 LLVMValueRef blend_mask
;
484 LLVMValueRef function
;
492 /* TODO: actually pick these based on the fs and color buffer
493 * characteristics. */
495 memset(&fs_type
, 0, sizeof fs_type
);
496 fs_type
.floating
= TRUE
; /* floating point values */
497 fs_type
.sign
= TRUE
; /* values are signed */
498 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
499 fs_type
.width
= 32; /* 32-bit float */
500 fs_type
.length
= 4; /* 4 elements per vector */
501 num_fs
= 4; /* number of quads per block */
503 memset(&blend_type
, 0, sizeof blend_type
);
504 blend_type
.floating
= FALSE
; /* values are integers */
505 blend_type
.sign
= FALSE
; /* values are unsigned */
506 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
507 blend_type
.width
= 8; /* 8-bit ubyte values */
508 blend_type
.length
= 16; /* 16 elements per vector */
511 * Generate the function prototype. Any change here must be reflected in
512 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
515 fs_elem_type
= lp_build_elem_type(fs_type
);
516 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
518 blend_vec_type
= lp_build_vec_type(blend_type
);
520 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
521 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
523 arg_types
[0] = screen
->context_ptr_type
; /* context */
524 arg_types
[1] = LLVMInt32Type(); /* x */
525 arg_types
[2] = LLVMInt32Type(); /* y */
526 arg_types
[3] = LLVMFloatType(); /* facing */
527 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
528 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
529 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
530 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
531 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
532 arg_types
[9] = LLVMInt32Type(); /* mask_input */
533 arg_types
[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
535 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
537 function
= LLVMAddFunction(screen
->module
, func_name
, func_type
);
538 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
540 variant
->function
[partial_mask
] = function
;
543 /* XXX: need to propagate noalias down into color param now we are
544 * passing a pointer-to-pointer?
546 for(i
= 0; i
< Elements(arg_types
); ++i
)
547 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
548 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
550 context_ptr
= LLVMGetParam(function
, 0);
551 x
= LLVMGetParam(function
, 1);
552 y
= LLVMGetParam(function
, 2);
553 facing
= LLVMGetParam(function
, 3);
554 a0_ptr
= LLVMGetParam(function
, 4);
555 dadx_ptr
= LLVMGetParam(function
, 5);
556 dady_ptr
= LLVMGetParam(function
, 6);
557 color_ptr_ptr
= LLVMGetParam(function
, 7);
558 depth_ptr
= LLVMGetParam(function
, 8);
559 mask_input
= LLVMGetParam(function
, 9);
561 lp_build_name(context_ptr
, "context");
562 lp_build_name(x
, "x");
563 lp_build_name(y
, "y");
564 lp_build_name(a0_ptr
, "a0");
565 lp_build_name(dadx_ptr
, "dadx");
566 lp_build_name(dady_ptr
, "dady");
567 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
568 lp_build_name(depth_ptr
, "depth");
569 lp_build_name(mask_input
, "mask_input");
571 if (key
->occlusion_count
) {
572 counter
= LLVMGetParam(function
, 10);
573 lp_build_name(counter
, "counter");
580 block
= LLVMAppendBasicBlock(function
, "entry");
581 builder
= LLVMCreateBuilder();
582 LLVMPositionBuilderAtEnd(builder
, block
);
585 * The shader input interpolation info is not explicitely baked in the
586 * shader key, but everything it derives from (TGSI, and flatshade) is
587 * already included in the shader key.
589 lp_build_interp_soa_init(&interp
,
593 a0_ptr
, dadx_ptr
, dady_ptr
,
596 /* code generated texture sampling */
597 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
599 /* loop over quads in the block */
600 for(i
= 0; i
< num_fs
; ++i
) {
601 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
602 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
603 LLVMValueRef depth_ptr_i
;
606 lp_build_interp_soa_update(&interp
, i
);
608 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
610 generate_fs(lp
, shader
, key
,
617 &fs_mask
[i
], /* output */
625 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
626 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
627 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
630 sampler
->destroy(sampler
);
632 /* Loop over color outputs / color buffers to do blending.
634 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
635 LLVMValueRef color_ptr
;
636 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
637 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
641 * Convert the fs's output color and mask to fit to the blending type.
643 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
644 lp_build_conv(builder
, fs_type
, blend_type
,
645 fs_out_color
[cbuf
][chan
], num_fs
,
646 &blend_in_color
[chan
], 1);
647 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
650 if (partial_mask
|| !variant
->opaque
) {
651 lp_build_conv_mask(builder
, fs_type
, blend_type
,
655 blend_mask
= lp_build_const_int_vec(blend_type
, ~0);
658 color_ptr
= LLVMBuildLoad(builder
,
659 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
661 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
663 /* which blend/colormask state to use */
664 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
669 generate_blend(&key
->blend
,
680 /* Avoid corrupting the FPU stack on 32bit OSes. */
681 lp_build_intrinsic(builder
, "llvm.x86.mmx.emms", LLVMVoidType(), NULL
, 0);
684 LLVMBuildRetVoid(builder
);
686 LLVMDisposeBuilder(builder
);
689 /* Verify the LLVM IR. If invalid, dump and abort */
691 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
693 lp_debug_dump_value(function
);
698 /* Apply optimizations to LLVM IR */
699 LLVMRunFunctionPassManager(screen
->pass
, function
);
701 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
702 /* Print the LLVM IR to stderr */
703 lp_debug_dump_value(function
);
708 * Translate the LLVM IR into machine code.
711 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
713 variant
->jit_function
[partial_mask
] = (lp_jit_frag_func
)pointer_to_func(f
);
715 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
718 lp_func_delete_body(function
);
724 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
728 debug_printf("fs variant %p:\n", (void *) key
);
730 if (key
->depth
.enabled
) {
731 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
732 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
733 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
736 for (i
= 0; i
< 2; ++i
) {
737 if (key
->stencil
[i
].enabled
) {
738 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
739 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
740 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
741 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
742 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
743 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
747 if (key
->alpha
.enabled
) {
748 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
749 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
752 if (key
->blend
.logicop_enable
) {
753 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
755 else if (key
->blend
.rt
[0].blend_enable
) {
756 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
757 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
758 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
759 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
760 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
761 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
763 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
764 for (i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
765 if (key
->sampler
[i
].format
) {
766 debug_printf("sampler[%u] = \n", i
);
767 debug_printf(" .format = %s\n",
768 util_format_name(key
->sampler
[i
].format
));
769 debug_printf(" .target = %s\n",
770 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
771 debug_printf(" .pot = %u %u %u\n",
772 key
->sampler
[i
].pot_width
,
773 key
->sampler
[i
].pot_height
,
774 key
->sampler
[i
].pot_depth
);
775 debug_printf(" .wrap = %s %s %s\n",
776 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
777 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
778 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
779 debug_printf(" .min_img_filter = %s\n",
780 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
781 debug_printf(" .min_mip_filter = %s\n",
782 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
783 debug_printf(" .mag_img_filter = %s\n",
784 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
785 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
786 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
787 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
794 static struct lp_fragment_shader_variant
*
795 generate_variant(struct llvmpipe_context
*lp
,
796 struct lp_fragment_shader
*shader
,
797 const struct lp_fragment_shader_variant_key
*key
)
799 struct lp_fragment_shader_variant
*variant
;
801 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
805 variant
->shader
= shader
;
806 variant
->list_item_global
.base
= variant
;
807 variant
->list_item_local
.base
= variant
;
808 variant
->no
= shader
->variants_created
++;
810 memcpy(&variant
->key
, key
, sizeof *key
);
812 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
813 debug_printf("llvmpipe: Creating fragment shader #%u variant #%u:\n",
814 shader
->no
, variant
->no
);
815 tgsi_dump(shader
->base
.tokens
, 0);
816 dump_fs_variant_key(key
);
819 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
820 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
822 /* TODO: most of these can be relaxed, in particular the colormask */
824 !key
->blend
.logicop_enable
&&
825 !key
->blend
.rt
[0].blend_enable
&&
826 key
->blend
.rt
[0].colormask
== 0xf &&
827 !key
->stencil
[0].enabled
&&
828 !key
->alpha
.enabled
&&
829 !key
->depth
.enabled
&&
830 !shader
->info
.uses_kill
838 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
839 const struct pipe_shader_state
*templ
)
841 struct lp_fragment_shader
*shader
;
843 shader
= CALLOC_STRUCT(lp_fragment_shader
);
847 shader
->no
= fs_no
++;
848 make_empty_list(&shader
->variants
);
850 /* get/save the summary info for this shader */
851 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
853 /* we need to keep a local copy of the tokens */
854 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
856 if (LP_DEBUG
& DEBUG_TGSI
) {
858 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader
->no
, (void *) shader
);
859 tgsi_dump(templ
->tokens
, 0);
860 debug_printf("usage masks:\n");
861 for (attrib
= 0; attrib
< shader
->info
.num_inputs
; ++attrib
) {
862 unsigned usage_mask
= shader
->info
.input_usage_mask
[attrib
];
863 debug_printf(" IN[%u].%s%s%s%s\n",
865 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
866 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
867 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
868 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
878 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
880 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
882 if (llvmpipe
->fs
== fs
)
885 draw_flush(llvmpipe
->draw
);
889 llvmpipe
->dirty
|= LP_NEW_FS
;
893 remove_shader_variant(struct llvmpipe_context
*lp
,
894 struct lp_fragment_shader_variant
*variant
)
896 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
899 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
900 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
901 variant
->shader
->no
, variant
->no
, variant
->shader
->variants_created
,
902 variant
->shader
->variants_cached
, lp
->nr_fs_variants
);
904 for (i
= 0; i
< Elements(variant
->function
); i
++) {
905 if (variant
->function
[i
]) {
906 if (variant
->jit_function
[i
])
907 LLVMFreeMachineCodeForFunction(screen
->engine
,
908 variant
->function
[i
]);
909 LLVMDeleteFunction(variant
->function
[i
]);
912 remove_from_list(&variant
->list_item_local
);
913 variant
->shader
->variants_cached
--;
914 remove_from_list(&variant
->list_item_global
);
915 lp
->nr_fs_variants
--;
920 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
922 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
923 struct pipe_fence_handle
*fence
= NULL
;
924 struct lp_fragment_shader
*shader
= fs
;
925 struct lp_fs_variant_list_item
*li
;
927 assert(fs
!= llvmpipe
->fs
);
931 * XXX: we need to flush the context until we have some sort of reference
932 * counting in fragment shaders as they may still be binned
933 * Flushing alone might not sufficient we need to wait on it too.
936 llvmpipe_flush(pipe
, 0, &fence
);
939 pipe
->screen
->fence_finish(pipe
->screen
, fence
, 0);
940 pipe
->screen
->fence_reference(pipe
->screen
, &fence
, NULL
);
943 li
= first_elem(&shader
->variants
);
944 while(!at_end(&shader
->variants
, li
)) {
945 struct lp_fs_variant_list_item
*next
= next_elem(li
);
946 remove_shader_variant(llvmpipe
, li
->base
);
950 assert(shader
->variants_cached
== 0);
951 FREE((void *) shader
->base
.tokens
);
958 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
959 uint shader
, uint index
,
960 struct pipe_resource
*constants
)
962 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
963 unsigned size
= constants
? constants
->width0
: 0;
964 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
966 assert(shader
< PIPE_SHADER_TYPES
);
967 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
969 if(llvmpipe
->constants
[shader
][index
] == constants
)
972 draw_flush(llvmpipe
->draw
);
974 /* note: reference counting */
975 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
977 if(shader
== PIPE_SHADER_VERTEX
||
978 shader
== PIPE_SHADER_GEOMETRY
) {
979 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
983 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
988 * Return the blend factor equivalent to a destination alpha of one.
990 static INLINE
unsigned
991 force_dst_alpha_one(unsigned factor
, boolean alpha
)
994 case PIPE_BLENDFACTOR_DST_ALPHA
:
995 return PIPE_BLENDFACTOR_ONE
;
996 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
997 return PIPE_BLENDFACTOR_ZERO
;
998 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
999 return PIPE_BLENDFACTOR_ZERO
;
1004 case PIPE_BLENDFACTOR_DST_COLOR
:
1005 return PIPE_BLENDFACTOR_ONE
;
1006 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
1007 return PIPE_BLENDFACTOR_ZERO
;
1016 * We need to generate several variants of the fragment pipeline to match
1017 * all the combinations of the contributing state atoms.
1019 * TODO: there is actually no reason to tie this to context state -- the
1020 * generated code could be cached globally in the screen.
1023 make_variant_key(struct llvmpipe_context
*lp
,
1024 struct lp_fragment_shader
*shader
,
1025 struct lp_fragment_shader_variant_key
*key
)
1029 memset(key
, 0, sizeof *key
);
1031 if (lp
->framebuffer
.zsbuf
) {
1032 if (lp
->depth_stencil
->depth
.enabled
) {
1033 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1034 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1036 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1037 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1038 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1042 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1043 if(key
->alpha
.enabled
)
1044 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1045 /* alpha.ref_value is passed in jit_context */
1047 key
->flatshade
= lp
->rasterizer
->flatshade
;
1048 if (lp
->active_query_count
) {
1049 key
->occlusion_count
= TRUE
;
1052 if (lp
->framebuffer
.nr_cbufs
) {
1053 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1056 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1057 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1058 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1059 const struct util_format_description
*format_desc
;
1062 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1063 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1064 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1066 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1068 /* mask out color channels not present in the color buffer.
1069 * Should be simple to incorporate per-cbuf writemasks:
1071 for(chan
= 0; chan
< 4; ++chan
) {
1072 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1074 if(swizzle
> UTIL_FORMAT_SWIZZLE_W
)
1075 blend_rt
->colormask
&= ~(1 << chan
);
1079 * Our swizzled render tiles always have an alpha channel, but the linear
1080 * render target format often does not, so force here the dst alpha to be
1083 * This is not a mere optimization. Wrong results will be produced if the
1084 * dst alpha is used, the dst format does not have alpha, and the previous
1085 * rendering was not flushed from the swizzled to linear buffer. For
1086 * example, NonPowTwo DCT.
1088 * TODO: This should be generalized to all channels for better
1089 * performance, but only alpha causes correctness issues.
1091 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1092 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
, FALSE
);
1093 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
, FALSE
);
1094 blend_rt
->alpha_src_factor
= force_dst_alpha_one(blend_rt
->alpha_src_factor
, TRUE
);
1095 blend_rt
->alpha_dst_factor
= force_dst_alpha_one(blend_rt
->alpha_dst_factor
, TRUE
);
1099 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1100 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1101 lp_sampler_static_state(&key
->sampler
[i
], lp
->fragment_sampler_views
[i
], lp
->sampler
[i
]);
1105 * Update fragment state. This is called just prior to drawing
1106 * something when some fragment-related state has changed.
1109 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1111 struct lp_fragment_shader
*shader
= lp
->fs
;
1112 struct lp_fragment_shader_variant_key key
;
1113 struct lp_fragment_shader_variant
*variant
= NULL
;
1114 struct lp_fs_variant_list_item
*li
;
1116 make_variant_key(lp
, shader
, &key
);
1118 li
= first_elem(&shader
->variants
);
1119 while(!at_end(&shader
->variants
, li
)) {
1120 if(memcmp(&li
->base
->key
, &key
, sizeof key
) == 0) {
1128 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1134 if (lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
) {
1135 struct pipe_context
*pipe
= &lp
->pipe
;
1136 struct pipe_fence_handle
*fence
= NULL
;
1139 * XXX: we need to flush the context until we have some sort of reference
1140 * counting in fragment shaders as they may still be binned
1141 * Flushing alone might not be sufficient we need to wait on it too.
1143 llvmpipe_flush(pipe
, 0, &fence
);
1146 pipe
->screen
->fence_finish(pipe
->screen
, fence
, 0);
1147 pipe
->screen
->fence_reference(pipe
->screen
, &fence
, NULL
);
1149 for (i
= 0; i
< LP_MAX_SHADER_VARIANTS
/ 4; i
++) {
1150 struct lp_fs_variant_list_item
*item
= last_elem(&lp
->fs_variants_list
);
1151 remove_shader_variant(lp
, item
->base
);
1156 variant
= generate_variant(lp
, shader
, &key
);
1160 LP_COUNT_ADD(llvm_compile_time
, dt
);
1161 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1164 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
1165 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
1166 lp
->nr_fs_variants
++;
1167 shader
->variants_cached
++;
1171 lp_setup_set_fs_variant(lp
->setup
, variant
);
1177 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1179 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1180 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1181 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1183 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;