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:
34 * - triangle edge in/out testing
40 * - depth/stencil test (stencil TBI)
43 * This file has only the glue to assembly the fragment pipeline. The actual
44 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
45 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
46 * muster the LLVM JIT execution engine to create a function that follows an
47 * established binary interface and that can be called from C directly.
49 * A big source of complexity here is that we often want to run different
50 * stages with different precisions and data types and precisions. For example,
51 * the fragment shader needs typically to be done in floats, but the
52 * depth/stencil test and blending is better done in the type that most closely
53 * matches the depth/stencil and color buffer respectively.
55 * Since the width of a SIMD vector register stays the same regardless of the
56 * element type, different types imply different number of elements, so we must
57 * code generate more instances of the stages with larger types to be able to
58 * feed/consume the stages with smaller types.
60 * @author Jose Fonseca <jfonseca@vmware.com>
64 #include "pipe/p_defines.h"
65 #include "util/u_inlines.h"
66 #include "util/u_memory.h"
67 #include "util/u_format.h"
68 #include "util/u_debug_dump.h"
69 #include "util/u_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_depth.h"
81 #include "gallivm/lp_bld_interp.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_alpha.h"
84 #include "gallivm/lp_bld_blend.h"
85 #include "gallivm/lp_bld_swizzle.h"
86 #include "gallivm/lp_bld_flow.h"
87 #include "gallivm/lp_bld_debug.h"
88 #include "lp_buffer.h"
89 #include "lp_context.h"
92 #include "lp_screen.h"
95 #include "lp_tex_sample.h"
98 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
99 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
103 * Derive from the quad's upper left scalar coordinates the coordinates for
104 * all other quad pixels
107 generate_pos0(LLVMBuilderRef builder
,
113 LLVMTypeRef int_elem_type
= LLVMInt32Type();
114 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
115 LLVMTypeRef elem_type
= LLVMFloatType();
116 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
117 LLVMValueRef x_offsets
[QUAD_SIZE
];
118 LLVMValueRef y_offsets
[QUAD_SIZE
];
121 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
122 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
124 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
125 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
126 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
129 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
130 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
132 *x0
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
133 *y0
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
138 * Generate the depth test.
141 generate_depth(LLVMBuilderRef builder
,
142 const struct lp_fragment_shader_variant_key
*key
,
143 struct lp_type src_type
,
144 struct lp_build_mask_context
*mask
,
146 LLVMValueRef dst_ptr
)
148 const struct util_format_description
*format_desc
;
149 struct lp_type dst_type
;
151 if(!key
->depth
.enabled
)
154 format_desc
= util_format_description(key
->zsbuf_format
);
158 * Depths are expected to be between 0 and 1, even if they are stored in
159 * floats. Setting these bits here will ensure that the lp_build_conv() call
160 * below won't try to unnecessarily clamp the incoming values.
162 if(src_type
.floating
) {
163 src_type
.sign
= FALSE
;
164 src_type
.norm
= TRUE
;
167 assert(!src_type
.sign
);
168 assert(src_type
.norm
);
171 /* Pick the depth type. */
172 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
174 /* FIXME: Cope with a depth test type with a different bit width. */
175 assert(dst_type
.width
== src_type
.width
);
176 assert(dst_type
.length
== src_type
.length
);
178 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
180 dst_ptr
= LLVMBuildBitCast(builder
,
182 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
184 lp_build_depth_test(builder
,
195 * Generate the code to do inside/outside triangle testing for the
196 * four pixels in a 2x2 quad. This will set the four elements of the
197 * quad mask vector to 0 or ~0.
198 * \param i which quad of the quad group to test, in [0,3]
201 generate_tri_edge_mask(LLVMBuilderRef builder
,
203 LLVMValueRef
*mask
, /* ivec4, out */
204 LLVMValueRef c0
, /* int32 */
205 LLVMValueRef c1
, /* int32 */
206 LLVMValueRef c2
, /* int32 */
207 LLVMValueRef step0_ptr
, /* ivec4 */
208 LLVMValueRef step1_ptr
, /* ivec4 */
209 LLVMValueRef step2_ptr
) /* ivec4 */
211 #define OPTIMIZE_IN_OUT_TEST 0
212 #if OPTIMIZE_IN_OUT_TEST
213 struct lp_build_if_state ifctx
;
214 LLVMValueRef not_draw_all
;
216 struct lp_build_flow_context
*flow
;
217 struct lp_type i32_type
;
218 LLVMTypeRef i32vec4_type
, mask_type
;
219 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
220 LLVMValueRef in_out_mask
;
224 /* int32 vector type */
225 memset(&i32_type
, 0, sizeof i32_type
);
226 i32_type
.floating
= FALSE
; /* values are integers */
227 i32_type
.sign
= TRUE
; /* values are signed */
228 i32_type
.norm
= FALSE
; /* values are not normalized */
229 i32_type
.width
= 32; /* 32-bit int values */
230 i32_type
.length
= 4; /* 4 elements per vector */
232 i32vec4_type
= lp_build_int32_vec4_type();
234 mask_type
= LLVMIntType(32 * 4);
237 * Use a conditional here to do detailed pixel in/out testing.
238 * We only have to do this if c0 != INT_MIN.
240 flow
= lp_build_flow_create(builder
);
241 lp_build_flow_scope_begin(flow
);
244 #if OPTIMIZE_IN_OUT_TEST
245 /* not_draw_all = (c0 != INT_MIN) */
246 not_draw_all
= LLVMBuildICmp(builder
,
249 LLVMConstInt(LLVMInt32Type(), INT_MIN
, 0),
252 in_out_mask
= lp_build_int_const_scalar(i32_type
, ~0);
255 lp_build_flow_scope_declare(flow
, &in_out_mask
);
257 /* if (not_draw_all) {... */
258 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
261 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
262 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
263 LLVMValueRef index
, m
;
265 /* c0_vec = {c0, c0, c0, c0}
266 * Note that we emit this code four times but LLVM optimizes away
267 * three instances of it.
269 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
270 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
271 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
272 lp_build_name(c0_vec
, "edgeconst0vec");
273 lp_build_name(c1_vec
, "edgeconst1vec");
274 lp_build_name(c2_vec
, "edgeconst2vec");
276 /* load step0vec, step1, step2 vec from memory */
277 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
278 step0_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step0_ptr
, &index
, 1, ""), "");
279 step1_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step1_ptr
, &index
, 1, ""), "");
280 step2_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step2_ptr
, &index
, 1, ""), "");
281 lp_build_name(step0_vec
, "step0vec");
282 lp_build_name(step1_vec
, "step1vec");
283 lp_build_name(step2_vec
, "step2vec");
285 /* m0_vec = step0_ptr[i] > c0_vec */
286 m0_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step0_vec
, c0_vec
);
287 m1_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step1_vec
, c1_vec
);
288 m2_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step2_vec
, c2_vec
);
290 /* in_out_mask = m0_vec & m1_vec & m2_vec */
291 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
292 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
293 lp_build_name(in_out_mask
, "inoutmaskvec");
295 #if OPTIMIZE_IN_OUT_TEST
296 lp_build_endif(&ifctx
);
300 lp_build_flow_scope_end(flow
);
301 lp_build_flow_destroy(flow
);
303 /* This is the initial alive/dead pixel mask for a quad of four pixels.
304 * It's an int[4] vector with each word set to 0 or ~0.
305 * Words will get cleared when pixels faile the Z test, etc.
312 generate_scissor_test(LLVMBuilderRef builder
,
313 LLVMValueRef context_ptr
,
314 const struct lp_build_interp_soa_context
*interp
,
317 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
318 LLVMValueRef xpos
= interp
->pos
[0], ypos
= interp
->pos
[1];
319 LLVMValueRef xmin
, ymin
, xmax
, ymax
;
320 LLVMValueRef m0
, m1
, m2
, m3
, m
;
322 /* xpos, ypos contain the window coords for the four pixels in the quad */
326 /* get the current scissor bounds, convert to vectors */
327 xmin
= lp_jit_context_scissor_xmin_value(builder
, context_ptr
);
328 xmin
= lp_build_broadcast(builder
, vec_type
, xmin
);
330 ymin
= lp_jit_context_scissor_ymin_value(builder
, context_ptr
);
331 ymin
= lp_build_broadcast(builder
, vec_type
, ymin
);
333 xmax
= lp_jit_context_scissor_xmax_value(builder
, context_ptr
);
334 xmax
= lp_build_broadcast(builder
, vec_type
, xmax
);
336 ymax
= lp_jit_context_scissor_ymax_value(builder
, context_ptr
);
337 ymax
= lp_build_broadcast(builder
, vec_type
, ymax
);
339 /* compare the fragment's position coordinates against the scissor bounds */
340 m0
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, xpos
, xmin
);
341 m1
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, ypos
, ymin
);
342 m2
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, xpos
, xmax
);
343 m3
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, ypos
, ymax
);
345 /* AND all the masks together */
346 m
= LLVMBuildAnd(builder
, m0
, m1
, "");
347 m
= LLVMBuildAnd(builder
, m
, m2
, "");
348 m
= LLVMBuildAnd(builder
, m
, m3
, "");
350 lp_build_name(m
, "scissormask");
357 build_int32_vec_const(int value
)
359 struct lp_type i32_type
;
361 memset(&i32_type
, 0, sizeof i32_type
);
362 i32_type
.floating
= FALSE
; /* values are integers */
363 i32_type
.sign
= TRUE
; /* values are signed */
364 i32_type
.norm
= FALSE
; /* values are not normalized */
365 i32_type
.width
= 32; /* 32-bit int values */
366 i32_type
.length
= 4; /* 4 elements per vector */
367 return lp_build_int_const_scalar(i32_type
, value
);
373 * Generate the fragment shader, depth/stencil test, and alpha tests.
374 * \param i which quad in the tile, in range [0,3]
375 * \param do_tri_test if 1, do triangle edge in/out testing
378 generate_fs(struct llvmpipe_context
*lp
,
379 struct lp_fragment_shader
*shader
,
380 const struct lp_fragment_shader_variant_key
*key
,
381 LLVMBuilderRef builder
,
383 LLVMValueRef context_ptr
,
385 const struct lp_build_interp_soa_context
*interp
,
386 struct lp_build_sampler_soa
*sampler
,
388 LLVMValueRef (*color
)[4],
389 LLVMValueRef depth_ptr
,
390 unsigned do_tri_test
,
394 LLVMValueRef step0_ptr
,
395 LLVMValueRef step1_ptr
,
396 LLVMValueRef step2_ptr
)
398 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
399 LLVMTypeRef elem_type
;
400 LLVMTypeRef vec_type
;
401 LLVMTypeRef int_vec_type
;
402 LLVMValueRef consts_ptr
;
403 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
404 LLVMValueRef z
= interp
->pos
[2];
405 struct lp_build_flow_context
*flow
;
406 struct lp_build_mask_context mask
;
407 boolean early_depth_test
;
414 elem_type
= lp_build_elem_type(type
);
415 vec_type
= lp_build_vec_type(type
);
416 int_vec_type
= lp_build_int_vec_type(type
);
418 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
420 flow
= lp_build_flow_create(builder
);
422 memset(outputs
, 0, sizeof outputs
);
424 lp_build_flow_scope_begin(flow
);
426 /* Declare the color and z variables */
427 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
428 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
429 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
430 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
433 lp_build_flow_scope_declare(flow
, &z
);
435 /* do triangle edge testing */
437 generate_tri_edge_mask(builder
, i
, pmask
,
438 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
441 *pmask
= build_int32_vec_const(~0);
444 /* 'mask' will control execution based on quad's pixel alive/killed state */
445 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
449 generate_scissor_test(builder
, context_ptr
, interp
, type
);
450 lp_build_mask_update(&mask
, smask
);
454 key
->depth
.enabled
&&
455 !key
->alpha
.enabled
&&
456 !shader
->info
.uses_kill
&&
457 !shader
->info
.writes_z
;
460 generate_depth(builder
, key
,
464 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
465 consts_ptr
, interp
->pos
, interp
->inputs
,
468 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
469 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
470 if(outputs
[attrib
][chan
]) {
471 lp_build_name(outputs
[attrib
][chan
], "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
473 switch (shader
->info
.output_semantic_name
[attrib
]) {
474 case TGSI_SEMANTIC_COLOR
:
476 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
478 lp_build_name(outputs
[attrib
][chan
], "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
481 /* XXX: should the alpha reference value be passed separately? */
482 /* XXX: should only test the final assignment to alpha */
483 if(cbuf
== 0 && chan
== 3) {
484 LLVMValueRef alpha
= outputs
[attrib
][chan
];
485 LLVMValueRef alpha_ref_value
;
486 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
487 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
488 lp_build_alpha_test(builder
, &key
->alpha
, type
,
489 &mask
, alpha
, alpha_ref_value
);
492 color
[cbuf
][chan
] = outputs
[attrib
][chan
];
496 case TGSI_SEMANTIC_POSITION
:
498 z
= outputs
[attrib
][chan
];
505 if(!early_depth_test
)
506 generate_depth(builder
, key
,
510 lp_build_mask_end(&mask
);
512 lp_build_flow_scope_end(flow
);
514 lp_build_flow_destroy(flow
);
522 * Generate color blending and color output.
525 generate_blend(const struct pipe_blend_state
*blend
,
526 LLVMBuilderRef builder
,
528 LLVMValueRef context_ptr
,
531 LLVMValueRef dst_ptr
)
533 struct lp_build_context bld
;
534 struct lp_build_flow_context
*flow
;
535 struct lp_build_mask_context mask_ctx
;
536 LLVMTypeRef vec_type
;
537 LLVMTypeRef int_vec_type
;
538 LLVMValueRef const_ptr
;
544 lp_build_context_init(&bld
, builder
, type
);
546 flow
= lp_build_flow_create(builder
);
548 /* we'll use this mask context to skip blending if all pixels are dead */
549 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
551 vec_type
= lp_build_vec_type(type
);
552 int_vec_type
= lp_build_int_vec_type(type
);
554 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
555 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
556 LLVMPointerType(vec_type
, 0), "");
558 for(chan
= 0; chan
< 4; ++chan
) {
559 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
560 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
562 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
564 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
565 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
568 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
570 for(chan
= 0; chan
< 4; ++chan
) {
571 if(blend
->rt
[0].colormask
& (1 << chan
)) {
572 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
573 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
574 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
575 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
579 lp_build_mask_end(&mask_ctx
);
580 lp_build_flow_destroy(flow
);
585 * Generate the runtime callable function for the whole fragment pipeline.
586 * Note that the function which we generate operates on a block of 16
587 * pixels at at time. The block contains 2x2 quads. Each quad contains
591 generate_fragment(struct llvmpipe_context
*lp
,
592 struct lp_fragment_shader
*shader
,
593 struct lp_fragment_shader_variant
*variant
,
594 unsigned do_tri_test
)
596 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
597 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
598 struct lp_type fs_type
;
599 struct lp_type blend_type
;
600 LLVMTypeRef fs_elem_type
;
601 LLVMTypeRef fs_vec_type
;
602 LLVMTypeRef fs_int_vec_type
;
603 LLVMTypeRef blend_vec_type
;
604 LLVMTypeRef blend_int_vec_type
;
605 LLVMTypeRef arg_types
[14];
606 LLVMTypeRef func_type
;
607 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
608 LLVMValueRef context_ptr
;
612 LLVMValueRef dadx_ptr
;
613 LLVMValueRef dady_ptr
;
614 LLVMValueRef color_ptr_ptr
;
615 LLVMValueRef depth_ptr
;
616 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
;
617 LLVMBasicBlockRef block
;
618 LLVMBuilderRef builder
;
621 struct lp_build_sampler_soa
*sampler
;
622 struct lp_build_interp_soa_context interp
;
623 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
624 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
625 LLVMValueRef blend_mask
;
626 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
627 LLVMValueRef function
;
634 /* TODO: actually pick these based on the fs and color buffer
635 * characteristics. */
637 memset(&fs_type
, 0, sizeof fs_type
);
638 fs_type
.floating
= TRUE
; /* floating point values */
639 fs_type
.sign
= TRUE
; /* values are signed */
640 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
641 fs_type
.width
= 32; /* 32-bit float */
642 fs_type
.length
= 4; /* 4 elements per vector */
643 num_fs
= 4; /* number of quads per block */
645 memset(&blend_type
, 0, sizeof blend_type
);
646 blend_type
.floating
= FALSE
; /* values are integers */
647 blend_type
.sign
= FALSE
; /* values are unsigned */
648 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
649 blend_type
.width
= 8; /* 8-bit ubyte values */
650 blend_type
.length
= 16; /* 16 elements per vector */
653 * Generate the function prototype. Any change here must be reflected in
654 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
657 fs_elem_type
= lp_build_elem_type(fs_type
);
658 fs_vec_type
= lp_build_vec_type(fs_type
);
659 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
661 blend_vec_type
= lp_build_vec_type(blend_type
);
662 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
664 arg_types
[0] = screen
->context_ptr_type
; /* context */
665 arg_types
[1] = LLVMInt32Type(); /* x */
666 arg_types
[2] = LLVMInt32Type(); /* y */
667 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
668 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
669 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
670 arg_types
[6] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
671 arg_types
[7] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
672 arg_types
[8] = LLVMInt32Type(); /* c0 */
673 arg_types
[9] = LLVMInt32Type(); /* c1 */
674 arg_types
[10] = LLVMInt32Type(); /* c2 */
675 /* Note: the step arrays are built as int32[16] but we interpret
676 * them here as int32_vec4[4].
678 arg_types
[11] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
679 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
680 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
682 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
684 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
685 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
687 variant
->function
[do_tri_test
] = function
;
690 /* XXX: need to propagate noalias down into color param now we are
691 * passing a pointer-to-pointer?
693 for(i
= 0; i
< Elements(arg_types
); ++i
)
694 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
695 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
697 context_ptr
= LLVMGetParam(function
, 0);
698 x
= LLVMGetParam(function
, 1);
699 y
= LLVMGetParam(function
, 2);
700 a0_ptr
= LLVMGetParam(function
, 3);
701 dadx_ptr
= LLVMGetParam(function
, 4);
702 dady_ptr
= LLVMGetParam(function
, 5);
703 color_ptr_ptr
= LLVMGetParam(function
, 6);
704 depth_ptr
= LLVMGetParam(function
, 7);
705 c0
= LLVMGetParam(function
, 8);
706 c1
= LLVMGetParam(function
, 9);
707 c2
= LLVMGetParam(function
, 10);
708 step0_ptr
= LLVMGetParam(function
, 11);
709 step1_ptr
= LLVMGetParam(function
, 12);
710 step2_ptr
= LLVMGetParam(function
, 13);
712 lp_build_name(context_ptr
, "context");
713 lp_build_name(x
, "x");
714 lp_build_name(y
, "y");
715 lp_build_name(a0_ptr
, "a0");
716 lp_build_name(dadx_ptr
, "dadx");
717 lp_build_name(dady_ptr
, "dady");
718 lp_build_name(color_ptr_ptr
, "color_ptr");
719 lp_build_name(depth_ptr
, "depth");
720 lp_build_name(c0
, "c0");
721 lp_build_name(c1
, "c1");
722 lp_build_name(c2
, "c2");
723 lp_build_name(step0_ptr
, "step0");
724 lp_build_name(step1_ptr
, "step1");
725 lp_build_name(step2_ptr
, "step2");
731 block
= LLVMAppendBasicBlock(function
, "entry");
732 builder
= LLVMCreateBuilder();
733 LLVMPositionBuilderAtEnd(builder
, block
);
735 generate_pos0(builder
, x
, y
, &x0
, &y0
);
737 lp_build_interp_soa_init(&interp
,
741 a0_ptr
, dadx_ptr
, dady_ptr
,
744 /* code generated texture sampling */
745 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
747 /* loop over quads in the block */
748 for(i
= 0; i
< num_fs
; ++i
) {
749 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
750 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
751 LLVMValueRef depth_ptr_i
;
755 lp_build_interp_soa_update(&interp
, i
);
757 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
759 generate_fs(lp
, shader
, key
,
766 &fs_mask
[i
], /* output */
771 step0_ptr
, step1_ptr
, step2_ptr
);
773 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
774 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
775 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
778 sampler
->destroy(sampler
);
780 /* Loop over color outputs / color buffers to do blending.
782 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
783 LLVMValueRef color_ptr
;
784 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
787 * Convert the fs's output color and mask to fit to the blending type.
789 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
790 lp_build_conv(builder
, fs_type
, blend_type
,
791 fs_out_color
[cbuf
][chan
], num_fs
,
792 &blend_in_color
[chan
], 1);
793 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
796 lp_build_conv_mask(builder
, fs_type
, blend_type
,
800 color_ptr
= LLVMBuildLoad(builder
,
801 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
803 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
808 generate_blend(&key
->blend
,
817 LLVMBuildRetVoid(builder
);
819 LLVMDisposeBuilder(builder
);
822 /* Verify the LLVM IR. If invalid, dump and abort */
824 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
826 LLVMDumpValue(function
);
831 /* Apply optimizations to LLVM IR */
833 LLVMRunFunctionPassManager(screen
->pass
, function
);
835 if (LP_DEBUG
& DEBUG_JIT
) {
836 /* Print the LLVM IR to stderr */
837 LLVMDumpValue(function
);
842 * Translate the LLVM IR into machine code.
844 variant
->jit_function
[do_tri_test
] = (lp_jit_frag_func
)LLVMGetPointerToGlobal(screen
->engine
, function
);
846 if (LP_DEBUG
& DEBUG_ASM
)
847 lp_disassemble(variant
->jit_function
[do_tri_test
]);
851 static struct lp_fragment_shader_variant
*
852 generate_variant(struct llvmpipe_context
*lp
,
853 struct lp_fragment_shader
*shader
,
854 const struct lp_fragment_shader_variant_key
*key
)
856 struct lp_fragment_shader_variant
*variant
;
858 if (LP_DEBUG
& DEBUG_JIT
) {
861 tgsi_dump(shader
->base
.tokens
, 0);
862 if(key
->depth
.enabled
) {
863 debug_printf("depth.format = %s\n", pf_name(key
->zsbuf_format
));
864 debug_printf("depth.func = %s\n", debug_dump_func(key
->depth
.func
, TRUE
));
865 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
867 if(key
->alpha
.enabled
) {
868 debug_printf("alpha.func = %s\n", debug_dump_func(key
->alpha
.func
, TRUE
));
869 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
871 if(key
->blend
.logicop_enable
) {
872 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
874 else if(key
->blend
.rt
[0].blend_enable
) {
875 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
876 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
877 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
878 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
879 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
880 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
882 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
883 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
884 if(key
->sampler
[i
].format
) {
885 debug_printf("sampler[%u] = \n", i
);
886 debug_printf(" .format = %s\n",
887 pf_name(key
->sampler
[i
].format
));
888 debug_printf(" .target = %s\n",
889 debug_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
890 debug_printf(" .pot = %u %u %u\n",
891 key
->sampler
[i
].pot_width
,
892 key
->sampler
[i
].pot_height
,
893 key
->sampler
[i
].pot_depth
);
894 debug_printf(" .wrap = %s %s %s\n",
895 debug_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
896 debug_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
897 debug_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
898 debug_printf(" .min_img_filter = %s\n",
899 debug_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
900 debug_printf(" .min_mip_filter = %s\n",
901 debug_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
902 debug_printf(" .mag_img_filter = %s\n",
903 debug_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
904 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
905 debug_printf(" .compare_func = %s\n", debug_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
906 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
907 debug_printf(" .prefilter = %u\n", key
->sampler
[i
].prefilter
);
912 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
916 variant
->shader
= shader
;
917 memcpy(&variant
->key
, key
, sizeof *key
);
919 generate_fragment(lp
, shader
, variant
, 0);
920 generate_fragment(lp
, shader
, variant
, 1);
922 /* insert new variant into linked list */
923 variant
->next
= shader
->variants
;
924 shader
->variants
= variant
;
931 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
932 const struct pipe_shader_state
*templ
)
934 struct lp_fragment_shader
*shader
;
936 shader
= CALLOC_STRUCT(lp_fragment_shader
);
940 /* get/save the summary info for this shader */
941 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
943 /* we need to keep a local copy of the tokens */
944 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
951 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
953 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
955 if (llvmpipe
->fs
== fs
)
958 draw_flush(llvmpipe
->draw
);
962 llvmpipe
->dirty
|= LP_NEW_FS
;
967 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
969 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
970 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
971 struct lp_fragment_shader
*shader
= fs
;
972 struct lp_fragment_shader_variant
*variant
;
974 assert(fs
!= llvmpipe
->fs
);
977 variant
= shader
->variants
;
979 struct lp_fragment_shader_variant
*next
= variant
->next
;
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
]);
996 FREE((void *) shader
->base
.tokens
);
1003 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1004 uint shader
, uint index
,
1005 struct pipe_buffer
*constants
)
1007 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1008 unsigned size
= constants
? constants
->size
: 0;
1009 const void *data
= constants
? llvmpipe_buffer(constants
)->data
: NULL
;
1011 assert(shader
< PIPE_SHADER_TYPES
);
1014 if(llvmpipe
->constants
[shader
] == constants
)
1017 draw_flush(llvmpipe
->draw
);
1019 /* note: reference counting */
1020 pipe_buffer_reference(&llvmpipe
->constants
[shader
], constants
);
1022 if(shader
== PIPE_SHADER_VERTEX
) {
1023 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
, 0,
1027 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1032 * We need to generate several variants of the fragment pipeline to match
1033 * all the combinations of the contributing state atoms.
1035 * TODO: there is actually no reason to tie this to context state -- the
1036 * generated code could be cached globally in the screen.
1039 make_variant_key(struct llvmpipe_context
*lp
,
1040 struct lp_fragment_shader
*shader
,
1041 struct lp_fragment_shader_variant_key
*key
)
1045 memset(key
, 0, sizeof *key
);
1047 if(lp
->framebuffer
.zsbuf
&&
1048 lp
->depth_stencil
->depth
.enabled
) {
1049 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1050 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1053 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1054 if(key
->alpha
.enabled
)
1055 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1056 /* alpha.ref_value is passed in jit_context */
1058 key
->flatshade
= lp
->rasterizer
->flatshade
;
1059 key
->scissor
= lp
->rasterizer
->scissor
;
1061 if (lp
->framebuffer
.nr_cbufs
) {
1062 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1065 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1066 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1067 const struct util_format_description
*format_desc
;
1070 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1071 assert(format_desc
->layout
== UTIL_FORMAT_COLORSPACE_RGB
||
1072 format_desc
->layout
== UTIL_FORMAT_COLORSPACE_SRGB
);
1074 /* mask out color channels not present in the color buffer.
1075 * Should be simple to incorporate per-cbuf writemasks:
1077 for(chan
= 0; chan
< 4; ++chan
) {
1078 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1080 if(swizzle
<= UTIL_FORMAT_SWIZZLE_W
)
1081 key
->blend
.rt
[0].colormask
|= (1 << chan
);
1085 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1086 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1087 lp_sampler_static_state(&key
->sampler
[i
], lp
->texture
[i
], lp
->sampler
[i
]);
1092 * Update fragment state. This is called just prior to drawing
1093 * something when some fragment-related state has changed.
1096 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1098 struct lp_fragment_shader
*shader
= lp
->fs
;
1099 struct lp_fragment_shader_variant_key key
;
1100 struct lp_fragment_shader_variant
*variant
;
1103 make_variant_key(lp
, shader
, &key
);
1105 variant
= shader
->variants
;
1107 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1110 variant
= variant
->next
;
1114 struct util_time t0
, t1
;
1118 variant
= generate_variant(lp
, shader
, &key
);
1121 dt
= util_time_diff(&t0
, &t1
);
1122 LP_COUNT_ADD(llvm_compile_time
, dt
);
1123 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1126 shader
->current
= variant
;
1128 /* TODO: put this in the variant */
1129 /* TODO: most of these can be relaxed, in particular the colormask */
1130 opaque
= !key
.blend
.logicop_enable
&&
1131 !key
.blend
.rt
[0].blend_enable
&&
1132 key
.blend
.rt
[0].colormask
== 0xf &&
1133 !key
.alpha
.enabled
&&
1134 !key
.depth
.enabled
&&
1136 !shader
->info
.uses_kill
1139 lp_setup_set_fs_functions(lp
->setup
,
1140 shader
->current
->jit_function
[0],
1141 shader
->current
->jit_function
[1],