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 */
211 m0_vec = step0_ptr[i] > c0_vec
212 m1_vec = step1_ptr[i] > c1_vec
213 m2_vec = step2_ptr[i] > c2_vec
214 mask = m0_vec & m1_vec & m2_vec
216 struct lp_build_flow_context
*flow
;
217 struct lp_build_if_state ifctx
;
218 struct lp_type i32_type
;
219 LLVMTypeRef i32vec4_type
, mask_type
;
221 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
223 LLVMValueRef int_min_vec
;
224 LLVMValueRef not_draw_all
;
225 LLVMValueRef in_out_mask
;
229 /* int32 vector type */
230 memset(&i32_type
, 0, sizeof i32_type
);
231 i32_type
.floating
= FALSE
; /* values are integers */
232 i32_type
.sign
= TRUE
; /* values are signed */
233 i32_type
.norm
= FALSE
; /* values are not normalized */
234 i32_type
.width
= 32; /* 32-bit int values */
235 i32_type
.length
= 4; /* 4 elements per vector */
237 i32vec4_type
= lp_build_int32_vec4_type();
239 mask_type
= LLVMIntType(32 * 4);
241 /* int_min_vec = {INT_MIN, INT_MIN, INT_MIN, INT_MIN} */
242 int_min_vec
= lp_build_int_const_scalar(i32_type
, INT_MIN
);
245 /* c0_vec = {c0, c0, c0, c0}
246 * Note that we emit this code four times but LLVM optimizes away
247 * three instances of it.
249 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
250 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
251 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
252 lp_build_name(c0_vec
, "edgeconst0vec");
253 lp_build_name(c1_vec
, "edgeconst1vec");
254 lp_build_name(c2_vec
, "edgeconst2vec");
257 * Use a conditional here to do detailed pixel in/out testing.
258 * We only have to do this if c0 != {INT_MIN, INT_MIN, INT_MIN, INT_MIN}
260 flow
= lp_build_flow_create(builder
);
261 lp_build_flow_scope_begin(flow
);
263 #define OPTIMIZE_IN_OUT_TEST 0
264 #if OPTIMIZE_IN_OUT_TEST
265 in_out_mask
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_EQUAL
, c0_vec
, int_min_vec
);
266 lp_build_name(in_out_mask
, "inoutmaskvec");
268 not_draw_all
= LLVMBuildICmp(builder
,
270 LLVMBuildBitCast(builder
, in_out_mask
, mask_type
, ""),
271 LLVMConstNull(mask_type
),
274 lp_build_flow_scope_declare(flow
, &in_out_mask
);
276 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
279 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
280 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
281 LLVMValueRef index
, m
;
283 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
284 step0_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step0_ptr
, &index
, 1, ""), "");
285 step1_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step1_ptr
, &index
, 1, ""), "");
286 step2_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step2_ptr
, &index
, 1, ""), "");
288 lp_build_name(step0_vec
, "step0vec");
289 lp_build_name(step1_vec
, "step1vec");
290 lp_build_name(step2_vec
, "step2vec");
292 m0_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step0_vec
, c0_vec
);
293 m1_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step1_vec
, c1_vec
);
294 m2_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step2_vec
, c2_vec
);
296 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
297 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
298 lp_build_name(in_out_mask
, "inoutmaskvec");
300 /* This is the initial alive/dead pixel mask. Additional bits will get cleared
301 * when the Z test fails, etc.
304 #if OPTIMIZE_IN_OUT_TEST
305 lp_build_endif(&ifctx
);
308 lp_build_flow_scope_end(flow
);
309 lp_build_flow_destroy(flow
);
316 * Generate the fragment shader, depth/stencil test, and alpha tests.
317 * \param i which quad in the tile, in range [0,3]
320 generate_fs(struct llvmpipe_context
*lp
,
321 struct lp_fragment_shader
*shader
,
322 const struct lp_fragment_shader_variant_key
*key
,
323 LLVMBuilderRef builder
,
325 LLVMValueRef context_ptr
,
327 const struct lp_build_interp_soa_context
*interp
,
328 struct lp_build_sampler_soa
*sampler
,
330 LLVMValueRef (*color
)[4],
331 LLVMValueRef depth_ptr
,
335 LLVMValueRef step0_ptr
,
336 LLVMValueRef step1_ptr
,
337 LLVMValueRef step2_ptr
)
339 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
340 LLVMTypeRef elem_type
;
341 LLVMTypeRef vec_type
;
342 LLVMTypeRef int_vec_type
;
343 LLVMValueRef consts_ptr
;
344 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
345 LLVMValueRef z
= interp
->pos
[2];
346 struct lp_build_flow_context
*flow
;
347 struct lp_build_mask_context mask
;
348 boolean early_depth_test
;
355 elem_type
= lp_build_elem_type(type
);
356 vec_type
= lp_build_vec_type(type
);
357 int_vec_type
= lp_build_int_vec_type(type
);
359 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
361 flow
= lp_build_flow_create(builder
);
363 memset(outputs
, 0, sizeof outputs
);
365 lp_build_flow_scope_begin(flow
);
367 /* Declare the color and z variables */
368 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
369 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
370 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
371 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
374 lp_build_flow_scope_declare(flow
, &z
);
376 /* do triangle edge testing */
377 generate_tri_edge_mask(builder
, i
, pmask
,
378 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
380 /* 'mask' will control execution based on quad's pixel alive/killed state */
381 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
385 key
->depth
.enabled
&&
386 !key
->alpha
.enabled
&&
387 !shader
->info
.uses_kill
&&
388 !shader
->info
.writes_z
;
391 generate_depth(builder
, key
,
395 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
396 consts_ptr
, interp
->pos
, interp
->inputs
,
399 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
400 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
401 if(outputs
[attrib
][chan
]) {
402 lp_build_name(outputs
[attrib
][chan
], "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
404 switch (shader
->info
.output_semantic_name
[attrib
]) {
405 case TGSI_SEMANTIC_COLOR
:
407 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
409 lp_build_name(outputs
[attrib
][chan
], "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
412 /* XXX: should the alpha reference value be passed separately? */
413 /* XXX: should only test the final assignment to alpha */
414 if(cbuf
== 0 && chan
== 3) {
415 LLVMValueRef alpha
= outputs
[attrib
][chan
];
416 LLVMValueRef alpha_ref_value
;
417 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
418 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
419 lp_build_alpha_test(builder
, &key
->alpha
, type
,
420 &mask
, alpha
, alpha_ref_value
);
423 color
[cbuf
][chan
] = outputs
[attrib
][chan
];
427 case TGSI_SEMANTIC_POSITION
:
429 z
= outputs
[attrib
][chan
];
436 if(!early_depth_test
)
437 generate_depth(builder
, key
,
441 lp_build_mask_end(&mask
);
443 lp_build_flow_scope_end(flow
);
445 lp_build_flow_destroy(flow
);
453 * Generate color blending and color output.
456 generate_blend(const struct pipe_blend_state
*blend
,
457 LLVMBuilderRef builder
,
459 LLVMValueRef context_ptr
,
462 LLVMValueRef dst_ptr
)
464 struct lp_build_context bld
;
465 struct lp_build_flow_context
*flow
;
466 struct lp_build_mask_context mask_ctx
;
467 LLVMTypeRef vec_type
;
468 LLVMTypeRef int_vec_type
;
469 LLVMValueRef const_ptr
;
475 lp_build_context_init(&bld
, builder
, type
);
477 flow
= lp_build_flow_create(builder
);
479 /* we'll use this mask context to skip blending if all pixels are dead */
480 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
482 vec_type
= lp_build_vec_type(type
);
483 int_vec_type
= lp_build_int_vec_type(type
);
485 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
486 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
487 LLVMPointerType(vec_type
, 0), "");
489 for(chan
= 0; chan
< 4; ++chan
) {
490 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
491 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
493 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
495 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
496 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
499 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
501 for(chan
= 0; chan
< 4; ++chan
) {
502 if(blend
->colormask
& (1 << chan
)) {
503 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
504 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
505 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
506 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
510 lp_build_mask_end(&mask_ctx
);
511 lp_build_flow_destroy(flow
);
516 * Generate the runtime callable function for the whole fragment pipeline.
517 * Note that the function which we generate operates on a block of 16
518 * pixels at at time. The block contains 2x2 quads. Each quad contains
521 static struct lp_fragment_shader_variant
*
522 generate_fragment(struct llvmpipe_context
*lp
,
523 struct lp_fragment_shader
*shader
,
524 const struct lp_fragment_shader_variant_key
*key
)
526 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
527 struct lp_fragment_shader_variant
*variant
;
528 struct lp_type fs_type
;
529 struct lp_type blend_type
;
530 LLVMTypeRef fs_elem_type
;
531 LLVMTypeRef fs_vec_type
;
532 LLVMTypeRef fs_int_vec_type
;
533 LLVMTypeRef blend_vec_type
;
534 LLVMTypeRef blend_int_vec_type
;
535 LLVMTypeRef arg_types
[14];
536 LLVMTypeRef func_type
;
537 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
538 LLVMValueRef context_ptr
;
542 LLVMValueRef dadx_ptr
;
543 LLVMValueRef dady_ptr
;
544 LLVMValueRef color_ptr_ptr
;
545 LLVMValueRef depth_ptr
;
546 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
;
547 LLVMBasicBlockRef block
;
548 LLVMBuilderRef builder
;
551 struct lp_build_sampler_soa
*sampler
;
552 struct lp_build_interp_soa_context interp
;
553 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
554 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
555 LLVMValueRef blend_mask
;
556 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
562 if (LP_DEBUG
& DEBUG_JIT
) {
563 tgsi_dump(shader
->base
.tokens
, 0);
564 if(key
->depth
.enabled
) {
565 debug_printf("depth.format = %s\n", pf_name(key
->zsbuf_format
));
566 debug_printf("depth.func = %s\n", debug_dump_func(key
->depth
.func
, TRUE
));
567 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
569 if(key
->alpha
.enabled
) {
570 debug_printf("alpha.func = %s\n", debug_dump_func(key
->alpha
.func
, TRUE
));
571 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
573 if(key
->blend
.logicop_enable
) {
574 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
576 else if(key
->blend
.blend_enable
) {
577 debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key
->blend
.rgb_func
, TRUE
));
578 debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_src_factor
, TRUE
));
579 debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.rgb_dst_factor
, TRUE
));
580 debug_printf("alpha_func = %s\n", debug_dump_blend_func (key
->blend
.alpha_func
, TRUE
));
581 debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_src_factor
, TRUE
));
582 debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key
->blend
.alpha_dst_factor
, TRUE
));
584 debug_printf("blend.colormask = 0x%x\n", key
->blend
.colormask
);
585 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
586 if(key
->sampler
[i
].format
) {
587 debug_printf("sampler[%u] = \n", i
);
588 debug_printf(" .format = %s\n",
589 pf_name(key
->sampler
[i
].format
));
590 debug_printf(" .target = %s\n",
591 debug_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
592 debug_printf(" .pot = %u %u %u\n",
593 key
->sampler
[i
].pot_width
,
594 key
->sampler
[i
].pot_height
,
595 key
->sampler
[i
].pot_depth
);
596 debug_printf(" .wrap = %s %s %s\n",
597 debug_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
598 debug_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
599 debug_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
600 debug_printf(" .min_img_filter = %s\n",
601 debug_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
602 debug_printf(" .min_mip_filter = %s\n",
603 debug_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
604 debug_printf(" .mag_img_filter = %s\n",
605 debug_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
606 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
607 debug_printf(" .compare_func = %s\n", debug_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
608 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
609 debug_printf(" .prefilter = %u\n", key
->sampler
[i
].prefilter
);
614 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
618 variant
->shader
= shader
;
619 memcpy(&variant
->key
, key
, sizeof *key
);
621 /* TODO: actually pick these based on the fs and color buffer
622 * characteristics. */
624 memset(&fs_type
, 0, sizeof fs_type
);
625 fs_type
.floating
= TRUE
; /* floating point values */
626 fs_type
.sign
= TRUE
; /* values are signed */
627 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
628 fs_type
.width
= 32; /* 32-bit float */
629 fs_type
.length
= 4; /* 4 elements per vector */
630 num_fs
= 4; /* number of quads per block */
632 memset(&blend_type
, 0, sizeof blend_type
);
633 blend_type
.floating
= FALSE
; /* values are integers */
634 blend_type
.sign
= FALSE
; /* values are unsigned */
635 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
636 blend_type
.width
= 8; /* 8-bit ubyte values */
637 blend_type
.length
= 16; /* 16 elements per vector */
640 * Generate the function prototype. Any change here must be reflected in
641 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
644 fs_elem_type
= lp_build_elem_type(fs_type
);
645 fs_vec_type
= lp_build_vec_type(fs_type
);
646 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
648 blend_vec_type
= lp_build_vec_type(blend_type
);
649 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
651 arg_types
[0] = screen
->context_ptr_type
; /* context */
652 arg_types
[1] = LLVMInt32Type(); /* x */
653 arg_types
[2] = LLVMInt32Type(); /* y */
654 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
655 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
656 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
657 arg_types
[6] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
658 arg_types
[7] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
659 arg_types
[8] = LLVMInt32Type(); /* c0 */
660 arg_types
[9] = LLVMInt32Type(); /* c1 */
661 arg_types
[10] = LLVMInt32Type(); /* c2 */
662 /* Note: the step arrays are built as int32[16] but we interpret
663 * them here as int32_vec4[4].
665 arg_types
[11] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
666 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
667 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
669 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
671 variant
->function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
672 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
674 /* XXX: need to propagate noalias down into color param now we are
675 * passing a pointer-to-pointer?
677 for(i
= 0; i
< Elements(arg_types
); ++i
)
678 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
679 LLVMAddAttribute(LLVMGetParam(variant
->function
, i
), LLVMNoAliasAttribute
);
681 context_ptr
= LLVMGetParam(variant
->function
, 0);
682 x
= LLVMGetParam(variant
->function
, 1);
683 y
= LLVMGetParam(variant
->function
, 2);
684 a0_ptr
= LLVMGetParam(variant
->function
, 3);
685 dadx_ptr
= LLVMGetParam(variant
->function
, 4);
686 dady_ptr
= LLVMGetParam(variant
->function
, 5);
687 color_ptr_ptr
= LLVMGetParam(variant
->function
, 6);
688 depth_ptr
= LLVMGetParam(variant
->function
, 7);
689 c0
= LLVMGetParam(variant
->function
, 8);
690 c1
= LLVMGetParam(variant
->function
, 9);
691 c2
= LLVMGetParam(variant
->function
, 10);
692 step0_ptr
= LLVMGetParam(variant
->function
, 11);
693 step1_ptr
= LLVMGetParam(variant
->function
, 12);
694 step2_ptr
= LLVMGetParam(variant
->function
, 13);
696 lp_build_name(context_ptr
, "context");
697 lp_build_name(x
, "x");
698 lp_build_name(y
, "y");
699 lp_build_name(a0_ptr
, "a0");
700 lp_build_name(dadx_ptr
, "dadx");
701 lp_build_name(dady_ptr
, "dady");
702 lp_build_name(color_ptr_ptr
, "color_ptr");
703 lp_build_name(depth_ptr
, "depth");
704 lp_build_name(c0
, "c0");
705 lp_build_name(c1
, "c1");
706 lp_build_name(c2
, "c2");
707 lp_build_name(step0_ptr
, "step0");
708 lp_build_name(step1_ptr
, "step1");
709 lp_build_name(step2_ptr
, "step2");
715 block
= LLVMAppendBasicBlock(variant
->function
, "entry");
716 builder
= LLVMCreateBuilder();
717 LLVMPositionBuilderAtEnd(builder
, block
);
719 generate_pos0(builder
, x
, y
, &x0
, &y0
);
721 lp_build_interp_soa_init(&interp
, shader
->base
.tokens
, builder
, fs_type
,
722 a0_ptr
, dadx_ptr
, dady_ptr
,
725 /* code generated texture sampling */
726 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
728 /* loop over quads in the block */
729 for(i
= 0; i
< num_fs
; ++i
) {
730 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
731 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
732 LLVMValueRef depth_ptr_i
;
736 lp_build_interp_soa_update(&interp
, i
);
738 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
740 generate_fs(lp
, shader
, key
,
747 &fs_mask
[i
], /* output */
751 step0_ptr
, step1_ptr
, step2_ptr
);
753 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
754 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
755 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
758 sampler
->destroy(sampler
);
760 /* Loop over color outputs / color buffers to do blending.
762 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
763 LLVMValueRef color_ptr
;
764 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
767 * Convert the fs's output color and mask to fit to the blending type.
769 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
770 lp_build_conv(builder
, fs_type
, blend_type
,
771 fs_out_color
[cbuf
][chan
], num_fs
,
772 &blend_in_color
[chan
], 1);
773 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
776 lp_build_conv_mask(builder
, fs_type
, blend_type
,
780 color_ptr
= LLVMBuildLoad(builder
,
781 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
783 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
788 generate_blend(&key
->blend
,
797 LLVMBuildRetVoid(builder
);
799 LLVMDisposeBuilder(builder
);
802 /* Verify the LLVM IR. If invalid, dump and abort */
804 if(LLVMVerifyFunction(variant
->function
, LLVMPrintMessageAction
)) {
806 LLVMDumpValue(variant
->function
);
811 /* Apply optimizations to LLVM IR */
813 LLVMRunFunctionPassManager(screen
->pass
, variant
->function
);
815 if (LP_DEBUG
& DEBUG_JIT
) {
816 /* Print the LLVM IR to stderr */
817 LLVMDumpValue(variant
->function
);
822 * Translate the LLVM IR into machine code.
824 variant
->jit_function
= (lp_jit_frag_func
)LLVMGetPointerToGlobal(screen
->engine
, variant
->function
);
826 if (LP_DEBUG
& DEBUG_ASM
)
827 lp_disassemble(variant
->jit_function
);
829 variant
->next
= shader
->variants
;
830 shader
->variants
= variant
;
837 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
838 const struct pipe_shader_state
*templ
)
840 struct lp_fragment_shader
*shader
;
842 shader
= CALLOC_STRUCT(lp_fragment_shader
);
846 /* get/save the summary info for this shader */
847 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
849 /* we need to keep a local copy of the tokens */
850 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
857 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
859 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
861 if (llvmpipe
->fs
== fs
)
864 draw_flush(llvmpipe
->draw
);
868 llvmpipe
->dirty
|= LP_NEW_FS
;
873 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
875 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
876 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
877 struct lp_fragment_shader
*shader
= fs
;
878 struct lp_fragment_shader_variant
*variant
;
880 assert(fs
!= llvmpipe
->fs
);
883 variant
= shader
->variants
;
885 struct lp_fragment_shader_variant
*next
= variant
->next
;
887 if(variant
->function
) {
888 if(variant
->jit_function
)
889 LLVMFreeMachineCodeForFunction(screen
->engine
, variant
->function
);
890 LLVMDeleteFunction(variant
->function
);
898 FREE((void *) shader
->base
.tokens
);
905 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
906 uint shader
, uint index
,
907 const struct pipe_constant_buffer
*constants
)
909 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
910 struct pipe_buffer
*buffer
= constants
? constants
->buffer
: NULL
;
911 unsigned size
= buffer
? buffer
->size
: 0;
912 const void *data
= buffer
? llvmpipe_buffer(buffer
)->data
: NULL
;
914 assert(shader
< PIPE_SHADER_TYPES
);
917 if(llvmpipe
->constants
[shader
].buffer
== buffer
)
920 draw_flush(llvmpipe
->draw
);
922 /* note: reference counting */
923 pipe_buffer_reference(&llvmpipe
->constants
[shader
].buffer
, buffer
);
925 if(shader
== PIPE_SHADER_VERTEX
) {
926 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
,
930 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
935 * We need to generate several variants of the fragment pipeline to match
936 * all the combinations of the contributing state atoms.
938 * TODO: there is actually no reason to tie this to context state -- the
939 * generated code could be cached globally in the screen.
942 make_variant_key(struct llvmpipe_context
*lp
,
943 struct lp_fragment_shader
*shader
,
944 struct lp_fragment_shader_variant_key
*key
)
948 memset(key
, 0, sizeof *key
);
950 if(lp
->framebuffer
.zsbuf
&&
951 lp
->depth_stencil
->depth
.enabled
) {
952 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
953 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
956 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
957 if(key
->alpha
.enabled
)
958 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
959 /* alpha.ref_value is passed in jit_context */
961 if (lp
->framebuffer
.nr_cbufs
) {
962 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
965 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
966 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
967 const struct util_format_description
*format_desc
;
970 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
971 assert(format_desc
->layout
== UTIL_FORMAT_COLORSPACE_RGB
||
972 format_desc
->layout
== UTIL_FORMAT_COLORSPACE_SRGB
);
974 /* mask out color channels not present in the color buffer.
975 * Should be simple to incorporate per-cbuf writemasks:
977 for(chan
= 0; chan
< 4; ++chan
) {
978 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
980 if(swizzle
<= UTIL_FORMAT_SWIZZLE_W
)
981 key
->cbuf_blend
[i
].colormask
|= (1 << chan
);
985 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
986 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
987 lp_sampler_static_state(&key
->sampler
[i
], lp
->texture
[i
], lp
->sampler
[i
]);
992 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
994 struct lp_fragment_shader
*shader
= lp
->fs
;
995 struct lp_fragment_shader_variant_key key
;
996 struct lp_fragment_shader_variant
*variant
;
998 make_variant_key(lp
, shader
, &key
);
1000 variant
= shader
->variants
;
1002 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1005 variant
= variant
->next
;
1009 variant
= generate_fragment(lp
, shader
, &key
);
1011 shader
->current
= variant
;
1013 lp_setup_set_fs_function(lp
->setup
,
1014 shader
->current
->jit_function
);