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:
38 * - depth/stencil test (stencil TBI)
41 * This file has only the glue to assembly the fragment pipeline. The actual
42 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
43 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
44 * muster the LLVM JIT execution engine to create a function that follows an
45 * established binary interface and that can be called from C directly.
47 * A big source of complexity here is that we often want to run different
48 * stages with different precisions and data types and precisions. For example,
49 * the fragment shader needs typically to be done in floats, but the
50 * depth/stencil test and blending is better done in the type that most closely
51 * matches the depth/stencil and color buffer respectively.
53 * Since the width of a SIMD vector register stays the same regardless of the
54 * element type, different types imply different number of elements, so we must
55 * code generate more instances of the stages with larger types to be able to
56 * feed/consume the stages with smaller types.
58 * @author Jose Fonseca <jfonseca@vmware.com>
62 #include "pipe/p_defines.h"
63 #include "util/u_memory.h"
64 #include "util/u_format.h"
65 #include "util/u_debug_dump.h"
66 #include "pipe/internal/p_winsys_screen.h"
67 #include "pipe/p_shader_tokens.h"
68 #include "draw/draw_context.h"
69 #include "tgsi/tgsi_dump.h"
70 #include "tgsi/tgsi_scan.h"
71 #include "tgsi/tgsi_parse.h"
72 #include "lp_bld_type.h"
73 #include "lp_bld_const.h"
74 #include "lp_bld_conv.h"
75 #include "lp_bld_intr.h"
76 #include "lp_bld_logic.h"
77 #include "lp_bld_depth.h"
78 #include "lp_bld_interp.h"
79 #include "lp_bld_tgsi.h"
80 #include "lp_bld_alpha.h"
81 #include "lp_bld_blend.h"
82 #include "lp_bld_swizzle.h"
83 #include "lp_bld_flow.h"
84 #include "lp_bld_debug.h"
85 #include "lp_screen.h"
86 #include "lp_context.h"
87 #include "lp_buffer.h"
90 #include "lp_tex_sample.h"
94 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
95 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
99 * Derive from the quad's upper left scalar coordinates the coordinates for
100 * all other quad pixels
103 generate_pos0(LLVMBuilderRef builder
,
109 LLVMTypeRef int_elem_type
= LLVMInt32Type();
110 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
111 LLVMTypeRef elem_type
= LLVMFloatType();
112 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
113 LLVMValueRef x_offsets
[QUAD_SIZE
];
114 LLVMValueRef y_offsets
[QUAD_SIZE
];
117 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
118 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
120 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
121 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
122 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
125 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
126 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
128 *x0
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
129 *y0
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
134 * Generate the depth test.
137 generate_depth(LLVMBuilderRef builder
,
138 const struct lp_fragment_shader_variant_key
*key
,
139 struct lp_type src_type
,
140 struct lp_build_mask_context
*mask
,
142 LLVMValueRef dst_ptr
)
144 const struct util_format_description
*format_desc
;
145 struct lp_type dst_type
;
147 if(!key
->depth
.enabled
)
150 format_desc
= util_format_description(key
->zsbuf_format
);
154 * Depths are expected to be between 0 and 1, even if they are stored in
155 * floats. Setting these bits here will ensure that the lp_build_conv() call
156 * below won't try to unnecessarily clamp the incoming values.
158 if(src_type
.floating
) {
159 src_type
.sign
= FALSE
;
160 src_type
.norm
= TRUE
;
163 assert(!src_type
.sign
);
164 assert(src_type
.norm
);
167 /* Pick the depth type. */
168 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
170 /* FIXME: Cope with a depth test type with a different bit width. */
171 assert(dst_type
.width
== src_type
.width
);
172 assert(dst_type
.length
== src_type
.length
);
174 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
176 dst_ptr
= LLVMBuildBitCast(builder
,
178 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
180 lp_build_depth_test(builder
,
191 * Generate the code to do inside/outside triangle testing for the
192 * four pixels in a 2x2 quad. This will set the four elements of the
193 * quad mask vector to 0 or ~0.
194 * \param i which quad of the quad group to test, in [0,3]
197 generate_tri_edge_mask(LLVMBuilderRef builder
,
199 LLVMValueRef
*mask
, /* ivec4, out */
200 LLVMValueRef c0
, /* int32 */
201 LLVMValueRef c1
, /* int32 */
202 LLVMValueRef c2
, /* int32 */
203 LLVMValueRef step0_ptr
, /* ivec4 */
204 LLVMValueRef step1_ptr
, /* ivec4 */
205 LLVMValueRef step2_ptr
) /* ivec4 */
207 #define OPTIMIZE_IN_OUT_TEST 0
208 #if OPTIMIZE_IN_OUT_TEST
209 struct lp_build_if_state ifctx
;
210 LLVMValueRef not_draw_all
;
212 struct lp_build_flow_context
*flow
;
213 struct lp_type i32_type
;
214 LLVMTypeRef i32vec4_type
, mask_type
;
215 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
216 LLVMValueRef in_out_mask
;
220 /* int32 vector type */
221 memset(&i32_type
, 0, sizeof i32_type
);
222 i32_type
.floating
= FALSE
; /* values are integers */
223 i32_type
.sign
= TRUE
; /* values are signed */
224 i32_type
.norm
= FALSE
; /* values are not normalized */
225 i32_type
.width
= 32; /* 32-bit int values */
226 i32_type
.length
= 4; /* 4 elements per vector */
228 i32vec4_type
= lp_build_int32_vec4_type();
230 mask_type
= LLVMIntType(32 * 4);
233 * Use a conditional here to do detailed pixel in/out testing.
234 * We only have to do this if c0 != INT_MIN.
236 flow
= lp_build_flow_create(builder
);
237 lp_build_flow_scope_begin(flow
);
240 #if OPTIMIZE_IN_OUT_TEST
241 /* not_draw_all = (c0 != INT_MIN) */
242 not_draw_all
= LLVMBuildICmp(builder
,
245 LLVMConstInt(LLVMInt32Type(), INT_MIN
, 0),
248 in_out_mask
= lp_build_int_const_scalar(i32_type
, ~0);
251 lp_build_flow_scope_declare(flow
, &in_out_mask
);
253 /* if (not_draw_all) {... */
254 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
257 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
258 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
259 LLVMValueRef index
, m
;
261 /* c0_vec = {c0, c0, c0, c0}
262 * Note that we emit this code four times but LLVM optimizes away
263 * three instances of it.
265 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
266 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
267 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
268 lp_build_name(c0_vec
, "edgeconst0vec");
269 lp_build_name(c1_vec
, "edgeconst1vec");
270 lp_build_name(c2_vec
, "edgeconst2vec");
272 /* load step0vec, step1, step2 vec from memory */
273 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
274 step0_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step0_ptr
, &index
, 1, ""), "");
275 step1_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step1_ptr
, &index
, 1, ""), "");
276 step2_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step2_ptr
, &index
, 1, ""), "");
277 lp_build_name(step0_vec
, "step0vec");
278 lp_build_name(step1_vec
, "step1vec");
279 lp_build_name(step2_vec
, "step2vec");
281 /* m0_vec = step0_ptr[i] > c0_vec */
282 m0_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step0_vec
, c0_vec
);
283 m1_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step1_vec
, c1_vec
);
284 m2_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step2_vec
, c2_vec
);
286 /* in_out_mask = m0_vec & m1_vec & m2_vec */
287 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
288 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
289 lp_build_name(in_out_mask
, "inoutmaskvec");
291 #if OPTIMIZE_IN_OUT_TEST
292 lp_build_endif(&ifctx
);
296 lp_build_flow_scope_end(flow
);
297 lp_build_flow_destroy(flow
);
299 /* This is the initial alive/dead pixel mask for a quad of four pixels.
300 * It's an int[4] vector with each word set to 0 or ~0.
301 * Words will get cleared when pixels faile the Z test, etc.
308 generate_scissor_test(LLVMBuilderRef builder
,
309 LLVMValueRef context_ptr
,
310 const struct lp_build_interp_soa_context
*interp
,
313 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
314 LLVMValueRef xpos
= interp
->pos
[0], ypos
= interp
->pos
[1];
315 LLVMValueRef xmin
, ymin
, xmax
, ymax
;
316 LLVMValueRef m0
, m1
, m2
, m3
, m
;
318 /* xpos, ypos contain the window coords for the four pixels in the quad */
322 /* get the current scissor bounds, convert to vectors */
323 xmin
= lp_jit_context_scissor_xmin_value(builder
, context_ptr
);
324 xmin
= lp_build_broadcast(builder
, vec_type
, xmin
);
326 ymin
= lp_jit_context_scissor_ymin_value(builder
, context_ptr
);
327 ymin
= lp_build_broadcast(builder
, vec_type
, ymin
);
329 xmax
= lp_jit_context_scissor_xmax_value(builder
, context_ptr
);
330 xmax
= lp_build_broadcast(builder
, vec_type
, xmax
);
332 ymax
= lp_jit_context_scissor_ymax_value(builder
, context_ptr
);
333 ymax
= lp_build_broadcast(builder
, vec_type
, ymax
);
335 /* compare the fragment's position coordinates against the scissor bounds */
336 m0
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, xpos
, xmin
);
337 m1
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, ypos
, ymin
);
338 m2
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, xpos
, xmax
);
339 m3
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, ypos
, ymax
);
341 /* AND all the masks together */
342 m
= LLVMBuildAnd(builder
, m0
, m1
, "");
343 m
= LLVMBuildAnd(builder
, m
, m2
, "");
344 m
= LLVMBuildAnd(builder
, m
, m3
, "");
346 lp_build_name(m
, "scissormask");
353 build_int32_vec_const(int value
)
355 struct lp_type i32_type
;
357 memset(&i32_type
, 0, sizeof i32_type
);
358 i32_type
.floating
= FALSE
; /* values are integers */
359 i32_type
.sign
= TRUE
; /* values are signed */
360 i32_type
.norm
= FALSE
; /* values are not normalized */
361 i32_type
.width
= 32; /* 32-bit int values */
362 i32_type
.length
= 4; /* 4 elements per vector */
363 return lp_build_int_const_scalar(i32_type
, value
);
369 * Generate the fragment shader, depth/stencil test, and alpha tests.
370 * \param i which quad in the tile, in range [0,3]
371 * \param do_tri_test if 1, do triangle edge in/out testing
374 generate_fs(struct llvmpipe_context
*lp
,
375 struct lp_fragment_shader
*shader
,
376 const struct lp_fragment_shader_variant_key
*key
,
377 LLVMBuilderRef builder
,
379 LLVMValueRef context_ptr
,
381 const struct lp_build_interp_soa_context
*interp
,
382 struct lp_build_sampler_soa
*sampler
,
384 LLVMValueRef (*color
)[4],
385 LLVMValueRef depth_ptr
,
386 unsigned do_tri_test
,
390 LLVMValueRef step0_ptr
,
391 LLVMValueRef step1_ptr
,
392 LLVMValueRef step2_ptr
)
394 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
395 LLVMTypeRef elem_type
;
396 LLVMTypeRef vec_type
;
397 LLVMTypeRef int_vec_type
;
398 LLVMValueRef consts_ptr
;
399 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
400 LLVMValueRef z
= interp
->pos
[2];
401 struct lp_build_flow_context
*flow
;
402 struct lp_build_mask_context mask
;
403 boolean early_depth_test
;
410 elem_type
= lp_build_elem_type(type
);
411 vec_type
= lp_build_vec_type(type
);
412 int_vec_type
= lp_build_int_vec_type(type
);
414 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
416 flow
= lp_build_flow_create(builder
);
418 memset(outputs
, 0, sizeof outputs
);
420 lp_build_flow_scope_begin(flow
);
422 /* Declare the color and z variables */
423 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
424 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
425 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
426 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
429 lp_build_flow_scope_declare(flow
, &z
);
431 /* do triangle edge testing */
433 generate_tri_edge_mask(builder
, i
, pmask
,
434 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
437 *pmask
= build_int32_vec_const(~0);
440 /* 'mask' will control execution based on quad's pixel alive/killed state */
441 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
445 generate_scissor_test(builder
, context_ptr
, interp
, type
);
446 lp_build_mask_update(&mask
, smask
);
450 key
->depth
.enabled
&&
451 !key
->alpha
.enabled
&&
452 !shader
->info
.uses_kill
&&
453 !shader
->info
.writes_z
;
456 generate_depth(builder
, key
,
460 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
461 consts_ptr
, interp
->pos
, interp
->inputs
,
464 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
465 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
466 if(outputs
[attrib
][chan
]) {
467 lp_build_name(outputs
[attrib
][chan
], "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
469 switch (shader
->info
.output_semantic_name
[attrib
]) {
470 case TGSI_SEMANTIC_COLOR
:
472 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
474 lp_build_name(outputs
[attrib
][chan
], "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
477 /* XXX: should the alpha reference value be passed separately? */
478 /* XXX: should only test the final assignment to alpha */
479 if(cbuf
== 0 && chan
== 3) {
480 LLVMValueRef alpha
= outputs
[attrib
][chan
];
481 LLVMValueRef alpha_ref_value
;
482 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
483 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
484 lp_build_alpha_test(builder
, &key
->alpha
, type
,
485 &mask
, alpha
, alpha_ref_value
);
488 color
[cbuf
][chan
] = outputs
[attrib
][chan
];
492 case TGSI_SEMANTIC_POSITION
:
494 z
= outputs
[attrib
][chan
];
501 if(!early_depth_test
)
502 generate_depth(builder
, key
,
506 lp_build_mask_end(&mask
);
508 lp_build_flow_scope_end(flow
);
510 lp_build_flow_destroy(flow
);
518 * Generate color blending and color output.
521 generate_blend(const struct pipe_blend_state
*blend
,
522 LLVMBuilderRef builder
,
524 LLVMValueRef context_ptr
,
527 LLVMValueRef dst_ptr
)
529 struct lp_build_context bld
;
530 struct lp_build_flow_context
*flow
;
531 struct lp_build_mask_context mask_ctx
;
532 LLVMTypeRef vec_type
;
533 LLVMTypeRef int_vec_type
;
534 LLVMValueRef const_ptr
;
540 lp_build_context_init(&bld
, builder
, type
);
542 flow
= lp_build_flow_create(builder
);
544 /* we'll use this mask context to skip blending if all pixels are dead */
545 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
547 vec_type
= lp_build_vec_type(type
);
548 int_vec_type
= lp_build_int_vec_type(type
);
550 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
551 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
552 LLVMPointerType(vec_type
, 0), "");
554 for(chan
= 0; chan
< 4; ++chan
) {
555 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
556 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
558 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
560 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
561 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
564 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
566 for(chan
= 0; chan
< 4; ++chan
) {
567 if(blend
->colormask
& (1 << chan
)) {
568 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
569 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
570 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
571 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
575 lp_build_mask_end(&mask_ctx
);
576 lp_build_flow_destroy(flow
);
581 * Generate the runtime callable function for the whole fragment pipeline.
582 * Note that the function which we generate operates on a block of 16
583 * pixels at at time. The block contains 2x2 quads. Each quad contains
587 generate_fragment(struct llvmpipe_context
*lp
,
588 struct lp_fragment_shader
*shader
,
589 struct lp_fragment_shader_variant
*variant
,
590 unsigned do_tri_test
)
592 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
593 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
594 struct lp_type fs_type
;
595 struct lp_type blend_type
;
596 LLVMTypeRef fs_elem_type
;
597 LLVMTypeRef fs_vec_type
;
598 LLVMTypeRef fs_int_vec_type
;
599 LLVMTypeRef blend_vec_type
;
600 LLVMTypeRef blend_int_vec_type
;
601 LLVMTypeRef arg_types
[14];
602 LLVMTypeRef func_type
;
603 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
604 LLVMValueRef context_ptr
;
608 LLVMValueRef dadx_ptr
;
609 LLVMValueRef dady_ptr
;
610 LLVMValueRef color_ptr_ptr
;
611 LLVMValueRef depth_ptr
;
612 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
;
613 LLVMBasicBlockRef block
;
614 LLVMBuilderRef builder
;
617 struct lp_build_sampler_soa
*sampler
;
618 struct lp_build_interp_soa_context interp
;
619 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
620 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
621 LLVMValueRef blend_mask
;
622 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
623 LLVMValueRef function
;
630 /* TODO: actually pick these based on the fs and color buffer
631 * characteristics. */
633 memset(&fs_type
, 0, sizeof fs_type
);
634 fs_type
.floating
= TRUE
; /* floating point values */
635 fs_type
.sign
= TRUE
; /* values are signed */
636 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
637 fs_type
.width
= 32; /* 32-bit float */
638 fs_type
.length
= 4; /* 4 elements per vector */
639 num_fs
= 4; /* number of quads per block */
641 memset(&blend_type
, 0, sizeof blend_type
);
642 blend_type
.floating
= FALSE
; /* values are integers */
643 blend_type
.sign
= FALSE
; /* values are unsigned */
644 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
645 blend_type
.width
= 8; /* 8-bit ubyte values */
646 blend_type
.length
= 16; /* 16 elements per vector */
649 * Generate the function prototype. Any change here must be reflected in
650 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
653 fs_elem_type
= lp_build_elem_type(fs_type
);
654 fs_vec_type
= lp_build_vec_type(fs_type
);
655 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
657 blend_vec_type
= lp_build_vec_type(blend_type
);
658 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
660 arg_types
[0] = screen
->context_ptr_type
; /* context */
661 arg_types
[1] = LLVMInt32Type(); /* x */
662 arg_types
[2] = LLVMInt32Type(); /* y */
663 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
664 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
665 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
666 arg_types
[6] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
667 arg_types
[7] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
668 arg_types
[8] = LLVMInt32Type(); /* c0 */
669 arg_types
[9] = LLVMInt32Type(); /* c1 */
670 arg_types
[10] = LLVMInt32Type(); /* c2 */
671 /* Note: the step arrays are built as int32[16] but we interpret
672 * them here as int32_vec4[4].
674 arg_types
[11] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
675 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
676 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
678 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
680 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
681 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
683 variant
->function
[do_tri_test
] = function
;
686 /* XXX: need to propagate noalias down into color param now we are
687 * passing a pointer-to-pointer?
689 for(i
= 0; i
< Elements(arg_types
); ++i
)
690 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
691 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
693 context_ptr
= LLVMGetParam(function
, 0);
694 x
= LLVMGetParam(function
, 1);
695 y
= LLVMGetParam(function
, 2);
696 a0_ptr
= LLVMGetParam(function
, 3);
697 dadx_ptr
= LLVMGetParam(function
, 4);
698 dady_ptr
= LLVMGetParam(function
, 5);
699 color_ptr_ptr
= LLVMGetParam(function
, 6);
700 depth_ptr
= LLVMGetParam(function
, 7);
701 c0
= LLVMGetParam(function
, 8);
702 c1
= LLVMGetParam(function
, 9);
703 c2
= LLVMGetParam(function
, 10);
704 step0_ptr
= LLVMGetParam(function
, 11);
705 step1_ptr
= LLVMGetParam(function
, 12);
706 step2_ptr
= LLVMGetParam(function
, 13);
708 lp_build_name(context_ptr
, "context");
709 lp_build_name(x
, "x");
710 lp_build_name(y
, "y");
711 lp_build_name(a0_ptr
, "a0");
712 lp_build_name(dadx_ptr
, "dadx");
713 lp_build_name(dady_ptr
, "dady");
714 lp_build_name(color_ptr_ptr
, "color_ptr");
715 lp_build_name(depth_ptr
, "depth");
716 lp_build_name(c0
, "c0");
717 lp_build_name(c1
, "c1");
718 lp_build_name(c2
, "c2");
719 lp_build_name(step0_ptr
, "step0");
720 lp_build_name(step1_ptr
, "step1");
721 lp_build_name(step2_ptr
, "step2");
727 block
= LLVMAppendBasicBlock(function
, "entry");
728 builder
= LLVMCreateBuilder();
729 LLVMPositionBuilderAtEnd(builder
, block
);
731 generate_pos0(builder
, x
, y
, &x0
, &y0
);
733 lp_build_interp_soa_init(&interp
,
737 a0_ptr
, dadx_ptr
, dady_ptr
,
740 /* code generated texture sampling */
741 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
743 /* loop over quads in the block */
744 for(i
= 0; i
< num_fs
; ++i
) {
745 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
746 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
747 LLVMValueRef depth_ptr_i
;
751 lp_build_interp_soa_update(&interp
, i
);
753 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
755 generate_fs(lp
, shader
, key
,
762 &fs_mask
[i
], /* output */
767 step0_ptr
, step1_ptr
, step2_ptr
);
769 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
770 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
771 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
774 sampler
->destroy(sampler
);
776 /* Loop over color outputs / color buffers to do blending.
778 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
779 LLVMValueRef color_ptr
;
780 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
783 * Convert the fs's output color and mask to fit to the blending type.
785 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
786 lp_build_conv(builder
, fs_type
, blend_type
,
787 fs_out_color
[cbuf
][chan
], num_fs
,
788 &blend_in_color
[chan
], 1);
789 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
792 lp_build_conv_mask(builder
, fs_type
, blend_type
,
796 color_ptr
= LLVMBuildLoad(builder
,
797 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
799 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
804 generate_blend(&key
->blend
,
813 LLVMBuildRetVoid(builder
);
815 LLVMDisposeBuilder(builder
);
818 /* Verify the LLVM IR. If invalid, dump and abort */
820 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
822 LLVMDumpValue(function
);
827 /* Apply optimizations to LLVM IR */
829 LLVMRunFunctionPassManager(screen
->pass
, function
);
831 if (LP_DEBUG
& DEBUG_JIT
) {
832 /* Print the LLVM IR to stderr */
833 LLVMDumpValue(function
);
838 * Translate the LLVM IR into machine code.
840 variant
->jit_function
[do_tri_test
] = (lp_jit_frag_func
)LLVMGetPointerToGlobal(screen
->engine
, function
);
842 if (LP_DEBUG
& DEBUG_ASM
)
843 lp_disassemble(variant
->jit_function
[do_tri_test
]);
847 static struct lp_fragment_shader_variant
*
848 generate_variant(struct llvmpipe_context
*lp
,
849 struct lp_fragment_shader
*shader
,
850 const struct lp_fragment_shader_variant_key
*key
)
852 struct lp_fragment_shader_variant
*variant
;
854 if (LP_DEBUG
& DEBUG_JIT
) {
857 tgsi_dump(shader
->base
.tokens
, 0);
858 if(key
->depth
.enabled
) {
859 debug_printf("depth.format = %s\n", pf_name(key
->zsbuf_format
));
860 debug_printf("depth.func = %s\n", debug_dump_func(key
->depth
.func
, TRUE
));
861 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
863 if(key
->alpha
.enabled
) {
864 debug_printf("alpha.func = %s\n", debug_dump_func(key
->alpha
.func
, TRUE
));
865 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
867 if(key
->blend
.logicop_enable
) {
868 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
870 else if(key
->blend
.blend_enable
) {
871 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key
->blend
.rgb_func
, TRUE
));
872 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_src_factor
, TRUE
));
873 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_dst_factor
, TRUE
));
874 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key
->blend
.alpha_func
, TRUE
));
875 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_src_factor
, TRUE
));
876 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_dst_factor
, TRUE
));
878 debug_printf("blend.colormask = 0x%x\n", key
->blend
.colormask
);
879 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
880 if(key
->sampler
[i
].format
) {
881 debug_printf("sampler[%u] = \n", i
);
882 debug_printf(" .format = %s\n",
883 pf_name(key
->sampler
[i
].format
));
884 debug_printf(" .target = %s\n",
885 debug_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
886 debug_printf(" .pot = %u %u %u\n",
887 key
->sampler
[i
].pot_width
,
888 key
->sampler
[i
].pot_height
,
889 key
->sampler
[i
].pot_depth
);
890 debug_printf(" .wrap = %s %s %s\n",
891 debug_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
892 debug_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
893 debug_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
894 debug_printf(" .min_img_filter = %s\n",
895 debug_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
896 debug_printf(" .min_mip_filter = %s\n",
897 debug_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
898 debug_printf(" .mag_img_filter = %s\n",
899 debug_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
900 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
901 debug_printf(" .compare_func = %s\n", debug_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
902 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
903 debug_printf(" .prefilter = %u\n", key
->sampler
[i
].prefilter
);
908 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
912 variant
->shader
= shader
;
913 memcpy(&variant
->key
, key
, sizeof *key
);
915 generate_fragment(lp
, shader
, variant
, 0);
916 generate_fragment(lp
, shader
, variant
, 1);
918 /* insert new variant into linked list */
919 variant
->next
= shader
->variants
;
920 shader
->variants
= variant
;
927 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
928 const struct pipe_shader_state
*templ
)
930 struct lp_fragment_shader
*shader
;
932 shader
= CALLOC_STRUCT(lp_fragment_shader
);
936 /* get/save the summary info for this shader */
937 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
939 /* we need to keep a local copy of the tokens */
940 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
947 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
949 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
951 if (llvmpipe
->fs
== fs
)
954 draw_flush(llvmpipe
->draw
);
958 llvmpipe
->dirty
|= LP_NEW_FS
;
963 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
965 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
966 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
967 struct lp_fragment_shader
*shader
= fs
;
968 struct lp_fragment_shader_variant
*variant
;
970 assert(fs
!= llvmpipe
->fs
);
973 variant
= shader
->variants
;
975 struct lp_fragment_shader_variant
*next
= variant
->next
;
978 for (i
= 0; i
< Elements(variant
->function
); i
++) {
979 if (variant
->function
[i
]) {
980 if (variant
->jit_function
[i
])
981 LLVMFreeMachineCodeForFunction(screen
->engine
,
982 variant
->function
[i
]);
983 LLVMDeleteFunction(variant
->function
[i
]);
992 FREE((void *) shader
->base
.tokens
);
999 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1000 uint shader
, uint index
,
1001 const struct pipe_constant_buffer
*constants
)
1003 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1004 struct pipe_buffer
*buffer
= constants
? constants
->buffer
: NULL
;
1005 unsigned size
= buffer
? buffer
->size
: 0;
1006 const void *data
= buffer
? llvmpipe_buffer(buffer
)->data
: NULL
;
1008 assert(shader
< PIPE_SHADER_TYPES
);
1011 if(llvmpipe
->constants
[shader
].buffer
== buffer
)
1014 draw_flush(llvmpipe
->draw
);
1016 /* note: reference counting */
1017 pipe_buffer_reference(&llvmpipe
->constants
[shader
].buffer
, buffer
);
1019 if(shader
== PIPE_SHADER_VERTEX
) {
1020 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
,
1024 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1029 * We need to generate several variants of the fragment pipeline to match
1030 * all the combinations of the contributing state atoms.
1032 * TODO: there is actually no reason to tie this to context state -- the
1033 * generated code could be cached globally in the screen.
1036 make_variant_key(struct llvmpipe_context
*lp
,
1037 struct lp_fragment_shader
*shader
,
1038 struct lp_fragment_shader_variant_key
*key
)
1042 memset(key
, 0, sizeof *key
);
1044 if(lp
->framebuffer
.zsbuf
&&
1045 lp
->depth_stencil
->depth
.enabled
) {
1046 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1047 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1050 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1051 if(key
->alpha
.enabled
)
1052 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1053 /* alpha.ref_value is passed in jit_context */
1055 key
->flatshade
= lp
->rasterizer
->flatshade
;
1056 key
->scissor
= lp
->rasterizer
->scissor
;
1058 if (lp
->framebuffer
.nr_cbufs
) {
1059 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1062 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1063 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1064 const struct util_format_description
*format_desc
;
1067 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1068 assert(format_desc
->layout
== UTIL_FORMAT_COLORSPACE_RGB
||
1069 format_desc
->layout
== UTIL_FORMAT_COLORSPACE_SRGB
);
1071 /* mask out color channels not present in the color buffer.
1072 * Should be simple to incorporate per-cbuf writemasks:
1074 for(chan
= 0; chan
< 4; ++chan
) {
1075 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1077 if(swizzle
<= UTIL_FORMAT_SWIZZLE_W
)
1078 key
->cbuf_blend
[i
].colormask
|= (1 << chan
);
1082 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1083 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1084 lp_sampler_static_state(&key
->sampler
[i
], lp
->texture
[i
], lp
->sampler
[i
]);
1089 * Update fragment state. This is called just prior to drawing
1090 * something when some fragment-related state has changed.
1093 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1095 struct lp_fragment_shader
*shader
= lp
->fs
;
1096 struct lp_fragment_shader_variant_key key
;
1097 struct lp_fragment_shader_variant
*variant
;
1100 make_variant_key(lp
, shader
, &key
);
1102 variant
= shader
->variants
;
1104 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1107 variant
= variant
->next
;
1111 variant
= generate_variant(lp
, shader
, &key
);
1113 shader
->current
= variant
;
1115 /* TODO: put this in the variant */
1116 /* TODO: most of these can be relaxed, in particular the colormask */
1117 opaque
= !key
.blend
.logicop_enable
&&
1118 !key
.blend
.blend_enable
&&
1119 key
.blend
.colormask
== 0xf &&
1120 !key
.alpha
.enabled
&&
1121 !key
.depth
.enabled
&&
1123 !shader
->info
.uses_kill
1126 lp_setup_set_fs_functions(lp
->setup
,
1127 shader
->current
->jit_function
[0],
1128 shader
->current
->jit_function
[1],