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
43 * This file has only the glue to assemble 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_dump.h"
69 #include "os/os_time.h"
70 #include "pipe/p_shader_tokens.h"
71 #include "draw/draw_context.h"
72 #include "tgsi/tgsi_dump.h"
73 #include "tgsi/tgsi_scan.h"
74 #include "tgsi/tgsi_parse.h"
75 #include "gallivm/lp_bld_type.h"
76 #include "gallivm/lp_bld_const.h"
77 #include "gallivm/lp_bld_conv.h"
78 #include "gallivm/lp_bld_intr.h"
79 #include "gallivm/lp_bld_logic.h"
80 #include "gallivm/lp_bld_tgsi.h"
81 #include "gallivm/lp_bld_swizzle.h"
82 #include "gallivm/lp_bld_flow.h"
83 #include "gallivm/lp_bld_debug.h"
85 #include "lp_bld_alpha.h"
86 #include "lp_bld_blend.h"
87 #include "lp_bld_depth.h"
88 #include "lp_bld_interp.h"
89 #include "lp_context.h"
92 #include "lp_screen.h"
95 #include "lp_tex_sample.h"
98 #include <llvm-c/Analysis.h>
101 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
102 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
106 * Derive from the quad's upper left scalar coordinates the coordinates for
107 * all other quad pixels
110 generate_pos0(LLVMBuilderRef builder
,
116 LLVMTypeRef int_elem_type
= LLVMInt32Type();
117 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
118 LLVMTypeRef elem_type
= LLVMFloatType();
119 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
120 LLVMValueRef x_offsets
[QUAD_SIZE
];
121 LLVMValueRef y_offsets
[QUAD_SIZE
];
124 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
125 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
127 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
128 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
129 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
132 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
133 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
135 *x0
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
136 *y0
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
141 * Generate the depth /stencil test code.
144 generate_depth_stencil(LLVMBuilderRef builder
,
145 const struct lp_fragment_shader_variant_key
*key
,
146 struct lp_type src_type
,
147 struct lp_build_mask_context
*mask
,
148 LLVMValueRef stencil_refs
[2],
150 LLVMValueRef dst_ptr
,
152 LLVMValueRef counter
)
154 const struct util_format_description
*format_desc
;
155 struct lp_type dst_type
;
157 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
160 format_desc
= util_format_description(key
->zsbuf_format
);
164 * Depths are expected to be between 0 and 1, even if they are stored in
165 * floats. Setting these bits here will ensure that the lp_build_conv() call
166 * below won't try to unnecessarily clamp the incoming values.
168 if(src_type
.floating
) {
169 src_type
.sign
= FALSE
;
170 src_type
.norm
= TRUE
;
173 assert(!src_type
.sign
);
174 assert(src_type
.norm
);
177 /* Pick the depth type. */
178 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
180 /* FIXME: Cope with a depth test type with a different bit width. */
181 assert(dst_type
.width
== src_type
.width
);
182 assert(dst_type
.length
== src_type
.length
);
184 /* Convert fragment Z from float to integer */
185 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
187 dst_ptr
= LLVMBuildBitCast(builder
,
189 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
190 lp_build_depth_stencil_test(builder
,
205 * Generate the code to do inside/outside triangle testing for the
206 * four pixels in a 2x2 quad. This will set the four elements of the
207 * quad mask vector to 0 or ~0.
208 * \param i which quad of the quad group to test, in [0,3]
211 generate_tri_edge_mask(LLVMBuilderRef builder
,
213 LLVMValueRef
*mask
, /* ivec4, out */
214 LLVMValueRef c0
, /* int32 */
215 LLVMValueRef c1
, /* int32 */
216 LLVMValueRef c2
, /* int32 */
217 LLVMValueRef step0_ptr
, /* ivec4 */
218 LLVMValueRef step1_ptr
, /* ivec4 */
219 LLVMValueRef step2_ptr
) /* ivec4 */
221 #define OPTIMIZE_IN_OUT_TEST 0
222 #if OPTIMIZE_IN_OUT_TEST
223 struct lp_build_if_state ifctx
;
224 LLVMValueRef not_draw_all
;
226 struct lp_build_flow_context
*flow
;
227 struct lp_type i32_type
;
228 LLVMTypeRef i32vec4_type
;
229 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
230 LLVMValueRef in_out_mask
;
234 /* int32 vector type */
235 memset(&i32_type
, 0, sizeof i32_type
);
236 i32_type
.floating
= FALSE
; /* values are integers */
237 i32_type
.sign
= TRUE
; /* values are signed */
238 i32_type
.norm
= FALSE
; /* values are not normalized */
239 i32_type
.width
= 32; /* 32-bit int values */
240 i32_type
.length
= 4; /* 4 elements per vector */
242 i32vec4_type
= lp_build_int32_vec4_type();
245 * Use a conditional here to do detailed pixel in/out testing.
246 * We only have to do this if c0 != INT_MIN.
248 flow
= lp_build_flow_create(builder
);
249 lp_build_flow_scope_begin(flow
);
252 #if OPTIMIZE_IN_OUT_TEST
253 /* not_draw_all = (c0 != INT_MIN) */
254 not_draw_all
= LLVMBuildICmp(builder
,
257 LLVMConstInt(LLVMInt32Type(), INT_MIN
, 0),
260 in_out_mask
= lp_build_const_int_vec(i32_type
, ~0);
263 lp_build_flow_scope_declare(flow
, &in_out_mask
);
265 /* if (not_draw_all) {... */
266 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
269 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
270 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
271 LLVMValueRef index
, m
;
273 /* c0_vec = {c0, c0, c0, c0}
274 * Note that we emit this code four times but LLVM optimizes away
275 * three instances of it.
277 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
278 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
279 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
280 lp_build_name(c0_vec
, "edgeconst0vec");
281 lp_build_name(c1_vec
, "edgeconst1vec");
282 lp_build_name(c2_vec
, "edgeconst2vec");
284 /* load step0vec, step1, step2 vec from memory */
285 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
286 step0_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step0_ptr
, &index
, 1, ""), "");
287 step1_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step1_ptr
, &index
, 1, ""), "");
288 step2_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step2_ptr
, &index
, 1, ""), "");
289 lp_build_name(step0_vec
, "step0vec");
290 lp_build_name(step1_vec
, "step1vec");
291 lp_build_name(step2_vec
, "step2vec");
293 /* m0_vec = step0_ptr[i] > c0_vec */
294 m0_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step0_vec
, c0_vec
);
295 m1_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step1_vec
, c1_vec
);
296 m2_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step2_vec
, c2_vec
);
298 /* in_out_mask = m0_vec & m1_vec & m2_vec */
299 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
300 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
301 lp_build_name(in_out_mask
, "inoutmaskvec");
303 #if OPTIMIZE_IN_OUT_TEST
304 lp_build_endif(&ifctx
);
308 lp_build_flow_scope_end(flow
);
309 lp_build_flow_destroy(flow
);
311 /* This is the initial alive/dead pixel mask for a quad of four pixels.
312 * It's an int[4] vector with each word set to 0 or ~0.
313 * Words will get cleared when pixels faile the Z test, etc.
320 generate_scissor_test(LLVMBuilderRef builder
,
321 LLVMValueRef context_ptr
,
322 const struct lp_build_interp_soa_context
*interp
,
325 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
326 LLVMValueRef xpos
= interp
->pos
[0], ypos
= interp
->pos
[1];
327 LLVMValueRef xmin
, ymin
, xmax
, ymax
;
328 LLVMValueRef m0
, m1
, m2
, m3
, m
;
330 /* xpos, ypos contain the window coords for the four pixels in the quad */
334 /* get the current scissor bounds, convert to vectors */
335 xmin
= lp_jit_context_scissor_xmin_value(builder
, context_ptr
);
336 xmin
= lp_build_broadcast(builder
, vec_type
, xmin
);
338 ymin
= lp_jit_context_scissor_ymin_value(builder
, context_ptr
);
339 ymin
= lp_build_broadcast(builder
, vec_type
, ymin
);
341 xmax
= lp_jit_context_scissor_xmax_value(builder
, context_ptr
);
342 xmax
= lp_build_broadcast(builder
, vec_type
, xmax
);
344 ymax
= lp_jit_context_scissor_ymax_value(builder
, context_ptr
);
345 ymax
= lp_build_broadcast(builder
, vec_type
, ymax
);
347 /* compare the fragment's position coordinates against the scissor bounds */
348 m0
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, xpos
, xmin
);
349 m1
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, ypos
, ymin
);
350 m2
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, xpos
, xmax
);
351 m3
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, ypos
, ymax
);
353 /* AND all the masks together */
354 m
= LLVMBuildAnd(builder
, m0
, m1
, "");
355 m
= LLVMBuildAnd(builder
, m
, m2
, "");
356 m
= LLVMBuildAnd(builder
, m
, m3
, "");
358 lp_build_name(m
, "scissormask");
365 build_int32_vec_const(int value
)
367 struct lp_type i32_type
;
369 memset(&i32_type
, 0, sizeof i32_type
);
370 i32_type
.floating
= FALSE
; /* values are integers */
371 i32_type
.sign
= TRUE
; /* values are signed */
372 i32_type
.norm
= FALSE
; /* values are not normalized */
373 i32_type
.width
= 32; /* 32-bit int values */
374 i32_type
.length
= 4; /* 4 elements per vector */
375 return lp_build_const_int_vec(i32_type
, value
);
381 * Generate the fragment shader, depth/stencil test, and alpha tests.
382 * \param i which quad in the tile, in range [0,3]
383 * \param do_tri_test if 1, do triangle edge in/out testing
386 generate_fs(struct llvmpipe_context
*lp
,
387 struct lp_fragment_shader
*shader
,
388 const struct lp_fragment_shader_variant_key
*key
,
389 LLVMBuilderRef builder
,
391 LLVMValueRef context_ptr
,
393 const struct lp_build_interp_soa_context
*interp
,
394 struct lp_build_sampler_soa
*sampler
,
396 LLVMValueRef (*color
)[4],
397 LLVMValueRef depth_ptr
,
399 unsigned do_tri_test
,
403 LLVMValueRef step0_ptr
,
404 LLVMValueRef step1_ptr
,
405 LLVMValueRef step2_ptr
,
406 LLVMValueRef counter
)
408 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
409 LLVMTypeRef vec_type
;
410 LLVMValueRef consts_ptr
;
411 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
412 LLVMValueRef z
= interp
->pos
[2];
413 LLVMValueRef stencil_refs
[2];
414 struct lp_build_flow_context
*flow
;
415 struct lp_build_mask_context mask
;
416 boolean early_depth_stencil_test
;
423 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
424 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
426 vec_type
= lp_build_vec_type(type
);
428 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
430 flow
= lp_build_flow_create(builder
);
432 memset(outputs
, 0, sizeof outputs
);
434 lp_build_flow_scope_begin(flow
);
436 /* Declare the color and z variables */
437 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
438 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
439 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
440 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
443 lp_build_flow_scope_declare(flow
, &z
);
445 /* do triangle edge testing */
447 generate_tri_edge_mask(builder
, i
, pmask
,
448 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
451 *pmask
= build_int32_vec_const(~0);
454 /* 'mask' will control execution based on quad's pixel alive/killed state */
455 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
459 generate_scissor_test(builder
, context_ptr
, interp
, type
);
460 lp_build_mask_update(&mask
, smask
);
463 early_depth_stencil_test
=
464 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
465 !key
->alpha
.enabled
&&
466 !shader
->info
.uses_kill
&&
467 !shader
->info
.writes_z
;
469 if (early_depth_stencil_test
)
470 generate_depth_stencil(builder
, key
,
472 stencil_refs
, z
, depth_ptr
, facing
, counter
);
474 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
475 consts_ptr
, interp
->pos
, interp
->inputs
,
476 outputs
, sampler
, &shader
->info
);
478 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
479 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
480 if(outputs
[attrib
][chan
]) {
481 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
482 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
484 switch (shader
->info
.output_semantic_name
[attrib
]) {
485 case TGSI_SEMANTIC_COLOR
:
487 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
489 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
492 /* XXX: should the alpha reference value be passed separately? */
493 /* XXX: should only test the final assignment to alpha */
494 if(cbuf
== 0 && chan
== 3) {
495 LLVMValueRef alpha
= out
;
496 LLVMValueRef alpha_ref_value
;
497 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
498 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
499 lp_build_alpha_test(builder
, &key
->alpha
, type
,
500 &mask
, alpha
, alpha_ref_value
);
503 color
[cbuf
][chan
] = out
;
507 case TGSI_SEMANTIC_POSITION
:
516 if (!early_depth_stencil_test
)
517 generate_depth_stencil(builder
, key
,
519 stencil_refs
, z
, depth_ptr
, facing
, counter
);
521 lp_build_mask_end(&mask
);
523 lp_build_flow_scope_end(flow
);
525 lp_build_flow_destroy(flow
);
533 * Generate color blending and color output.
536 generate_blend(const struct pipe_blend_state
*blend
,
537 LLVMBuilderRef builder
,
539 LLVMValueRef context_ptr
,
542 LLVMValueRef dst_ptr
)
544 struct lp_build_context bld
;
545 struct lp_build_flow_context
*flow
;
546 struct lp_build_mask_context mask_ctx
;
547 LLVMTypeRef vec_type
;
548 LLVMValueRef const_ptr
;
554 lp_build_context_init(&bld
, builder
, type
);
556 flow
= lp_build_flow_create(builder
);
558 /* we'll use this mask context to skip blending if all pixels are dead */
559 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
561 vec_type
= lp_build_vec_type(type
);
563 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
564 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
565 LLVMPointerType(vec_type
, 0), "");
567 for(chan
= 0; chan
< 4; ++chan
) {
568 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
569 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
571 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
573 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
574 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
577 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
579 for(chan
= 0; chan
< 4; ++chan
) {
580 if(blend
->rt
[0].colormask
& (1 << chan
)) {
581 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
582 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
583 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
584 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
588 lp_build_mask_end(&mask_ctx
);
589 lp_build_flow_destroy(flow
);
593 /** casting function to avoid compiler warnings */
594 static lp_jit_frag_func
595 cast_voidptr_to_lp_jit_frag_func(void *p
)
601 assert(sizeof(tmp
.v
) == sizeof(tmp
.f
));
608 * Generate the runtime callable function for the whole fragment pipeline.
609 * Note that the function which we generate operates on a block of 16
610 * pixels at at time. The block contains 2x2 quads. Each quad contains
614 generate_fragment(struct llvmpipe_context
*lp
,
615 struct lp_fragment_shader
*shader
,
616 struct lp_fragment_shader_variant
*variant
,
617 unsigned do_tri_test
)
619 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
620 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
621 struct lp_type fs_type
;
622 struct lp_type blend_type
;
623 LLVMTypeRef fs_elem_type
;
624 LLVMTypeRef fs_int_vec_type
;
625 LLVMTypeRef blend_vec_type
;
626 LLVMTypeRef arg_types
[16];
627 LLVMTypeRef func_type
;
628 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
629 LLVMValueRef context_ptr
;
633 LLVMValueRef dadx_ptr
;
634 LLVMValueRef dady_ptr
;
635 LLVMValueRef color_ptr_ptr
;
636 LLVMValueRef depth_ptr
;
637 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
, counter
= NULL
;
638 LLVMBasicBlockRef block
;
639 LLVMBuilderRef builder
;
642 struct lp_build_sampler_soa
*sampler
;
643 struct lp_build_interp_soa_context interp
;
644 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
645 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
646 LLVMValueRef blend_mask
;
647 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
648 LLVMValueRef function
;
656 /* TODO: actually pick these based on the fs and color buffer
657 * characteristics. */
659 memset(&fs_type
, 0, sizeof fs_type
);
660 fs_type
.floating
= TRUE
; /* floating point values */
661 fs_type
.sign
= TRUE
; /* values are signed */
662 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
663 fs_type
.width
= 32; /* 32-bit float */
664 fs_type
.length
= 4; /* 4 elements per vector */
665 num_fs
= 4; /* number of quads per block */
667 memset(&blend_type
, 0, sizeof blend_type
);
668 blend_type
.floating
= FALSE
; /* values are integers */
669 blend_type
.sign
= FALSE
; /* values are unsigned */
670 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
671 blend_type
.width
= 8; /* 8-bit ubyte values */
672 blend_type
.length
= 16; /* 16 elements per vector */
675 * Generate the function prototype. Any change here must be reflected in
676 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
679 fs_elem_type
= lp_build_elem_type(fs_type
);
680 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
682 blend_vec_type
= lp_build_vec_type(blend_type
);
684 arg_types
[0] = screen
->context_ptr_type
; /* context */
685 arg_types
[1] = LLVMInt32Type(); /* x */
686 arg_types
[2] = LLVMInt32Type(); /* y */
687 arg_types
[3] = LLVMFloatType(); /* facing */
688 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
689 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
690 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
691 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
692 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
693 arg_types
[9] = LLVMInt32Type(); /* c0 */
694 arg_types
[10] = LLVMInt32Type(); /* c1 */
695 arg_types
[11] = LLVMInt32Type(); /* c2 */
696 /* Note: the step arrays are built as int32[16] but we interpret
697 * them here as int32_vec4[4].
699 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
700 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
701 arg_types
[14] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
702 arg_types
[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
704 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
706 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
707 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
709 variant
->function
[do_tri_test
] = function
;
712 /* XXX: need to propagate noalias down into color param now we are
713 * passing a pointer-to-pointer?
715 for(i
= 0; i
< Elements(arg_types
); ++i
)
716 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
717 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
719 context_ptr
= LLVMGetParam(function
, 0);
720 x
= LLVMGetParam(function
, 1);
721 y
= LLVMGetParam(function
, 2);
722 facing
= LLVMGetParam(function
, 3);
723 a0_ptr
= LLVMGetParam(function
, 4);
724 dadx_ptr
= LLVMGetParam(function
, 5);
725 dady_ptr
= LLVMGetParam(function
, 6);
726 color_ptr_ptr
= LLVMGetParam(function
, 7);
727 depth_ptr
= LLVMGetParam(function
, 8);
728 c0
= LLVMGetParam(function
, 9);
729 c1
= LLVMGetParam(function
, 10);
730 c2
= LLVMGetParam(function
, 11);
731 step0_ptr
= LLVMGetParam(function
, 12);
732 step1_ptr
= LLVMGetParam(function
, 13);
733 step2_ptr
= LLVMGetParam(function
, 14);
735 lp_build_name(context_ptr
, "context");
736 lp_build_name(x
, "x");
737 lp_build_name(y
, "y");
738 lp_build_name(a0_ptr
, "a0");
739 lp_build_name(dadx_ptr
, "dadx");
740 lp_build_name(dady_ptr
, "dady");
741 lp_build_name(color_ptr_ptr
, "color_ptr");
742 lp_build_name(depth_ptr
, "depth");
743 lp_build_name(c0
, "c0");
744 lp_build_name(c1
, "c1");
745 lp_build_name(c2
, "c2");
746 lp_build_name(step0_ptr
, "step0");
747 lp_build_name(step1_ptr
, "step1");
748 lp_build_name(step2_ptr
, "step2");
750 if (key
->occlusion_count
) {
751 counter
= LLVMGetParam(function
, 15);
752 lp_build_name(counter
, "counter");
759 block
= LLVMAppendBasicBlock(function
, "entry");
760 builder
= LLVMCreateBuilder();
761 LLVMPositionBuilderAtEnd(builder
, block
);
763 generate_pos0(builder
, x
, y
, &x0
, &y0
);
765 lp_build_interp_soa_init(&interp
,
769 a0_ptr
, dadx_ptr
, dady_ptr
,
772 /* code generated texture sampling */
773 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
775 /* loop over quads in the block */
776 for(i
= 0; i
< num_fs
; ++i
) {
777 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
778 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
779 LLVMValueRef depth_ptr_i
;
782 lp_build_interp_soa_update(&interp
, i
);
784 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
786 generate_fs(lp
, shader
, key
,
793 &fs_mask
[i
], /* output */
799 step0_ptr
, step1_ptr
, step2_ptr
, counter
);
801 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
802 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
803 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
806 sampler
->destroy(sampler
);
808 /* Loop over color outputs / color buffers to do blending.
810 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
811 LLVMValueRef color_ptr
;
812 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
815 * Convert the fs's output color and mask to fit to the blending type.
817 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
818 lp_build_conv(builder
, fs_type
, blend_type
,
819 fs_out_color
[cbuf
][chan
], num_fs
,
820 &blend_in_color
[chan
], 1);
821 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
824 lp_build_conv_mask(builder
, fs_type
, blend_type
,
828 color_ptr
= LLVMBuildLoad(builder
,
829 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
831 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
836 generate_blend(&key
->blend
,
845 LLVMBuildRetVoid(builder
);
847 LLVMDisposeBuilder(builder
);
850 /* Verify the LLVM IR. If invalid, dump and abort */
852 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
854 lp_debug_dump_value(function
);
859 /* Apply optimizations to LLVM IR */
861 LLVMRunFunctionPassManager(screen
->pass
, function
);
863 if (LP_DEBUG
& DEBUG_JIT
) {
864 /* Print the LLVM IR to stderr */
865 lp_debug_dump_value(function
);
870 * Translate the LLVM IR into machine code.
873 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
875 variant
->jit_function
[do_tri_test
] = cast_voidptr_to_lp_jit_frag_func(f
);
877 if (LP_DEBUG
& DEBUG_ASM
)
883 static struct lp_fragment_shader_variant
*
884 generate_variant(struct llvmpipe_context
*lp
,
885 struct lp_fragment_shader
*shader
,
886 const struct lp_fragment_shader_variant_key
*key
)
888 struct lp_fragment_shader_variant
*variant
;
890 if (LP_DEBUG
& DEBUG_JIT
) {
893 tgsi_dump(shader
->base
.tokens
, 0);
894 if(key
->depth
.enabled
) {
895 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
896 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
897 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
899 if(key
->alpha
.enabled
) {
900 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
901 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
903 if(key
->blend
.logicop_enable
) {
904 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
906 else if(key
->blend
.rt
[0].blend_enable
) {
907 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
908 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
909 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
910 debug_printf("alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
911 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
912 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
914 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
915 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
916 if(key
->sampler
[i
].format
) {
917 debug_printf("sampler[%u] = \n", i
);
918 debug_printf(" .format = %s\n",
919 util_format_name(key
->sampler
[i
].format
));
920 debug_printf(" .target = %s\n",
921 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
922 debug_printf(" .pot = %u %u %u\n",
923 key
->sampler
[i
].pot_width
,
924 key
->sampler
[i
].pot_height
,
925 key
->sampler
[i
].pot_depth
);
926 debug_printf(" .wrap = %s %s %s\n",
927 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
928 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
929 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
930 debug_printf(" .min_img_filter = %s\n",
931 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
932 debug_printf(" .min_mip_filter = %s\n",
933 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
934 debug_printf(" .mag_img_filter = %s\n",
935 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
936 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
937 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
938 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
943 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
947 memcpy(&variant
->key
, key
, sizeof *key
);
949 generate_fragment(lp
, shader
, variant
, 0);
950 generate_fragment(lp
, shader
, variant
, 1);
952 /* insert new variant into linked list */
953 variant
->next
= shader
->variants
;
954 shader
->variants
= variant
;
961 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
962 const struct pipe_shader_state
*templ
)
964 struct lp_fragment_shader
*shader
;
966 shader
= CALLOC_STRUCT(lp_fragment_shader
);
970 /* get/save the summary info for this shader */
971 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
973 /* we need to keep a local copy of the tokens */
974 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
976 if (LP_DEBUG
& DEBUG_TGSI
) {
977 debug_printf("llvmpipe: Create fragment shader %p:\n", (void *) shader
);
978 tgsi_dump(templ
->tokens
, 0);
986 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
988 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
990 if (llvmpipe
->fs
== fs
)
993 draw_flush(llvmpipe
->draw
);
997 llvmpipe
->dirty
|= LP_NEW_FS
;
1002 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1004 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1005 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
1006 struct lp_fragment_shader
*shader
= fs
;
1007 struct lp_fragment_shader_variant
*variant
;
1009 assert(fs
!= llvmpipe
->fs
);
1013 * XXX: we need to flush the context until we have some sort of reference
1014 * counting in fragment shaders as they may still be binned
1016 draw_flush(llvmpipe
->draw
);
1017 lp_setup_flush(llvmpipe
->setup
, 0);
1019 variant
= shader
->variants
;
1021 struct lp_fragment_shader_variant
*next
= variant
->next
;
1024 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1025 if (variant
->function
[i
]) {
1026 if (variant
->jit_function
[i
])
1027 LLVMFreeMachineCodeForFunction(screen
->engine
,
1028 variant
->function
[i
]);
1029 LLVMDeleteFunction(variant
->function
[i
]);
1038 FREE((void *) shader
->base
.tokens
);
1045 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1046 uint shader
, uint index
,
1047 struct pipe_resource
*constants
)
1049 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1050 unsigned size
= constants
? constants
->width0
: 0;
1051 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1053 assert(shader
< PIPE_SHADER_TYPES
);
1056 if(llvmpipe
->constants
[shader
] == constants
)
1059 draw_flush(llvmpipe
->draw
);
1061 /* note: reference counting */
1062 pipe_resource_reference(&llvmpipe
->constants
[shader
], constants
);
1064 if(shader
== PIPE_SHADER_VERTEX
) {
1065 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
, 0,
1069 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1074 * Return the blend factor equivalent to a destination alpha of one.
1076 static INLINE
unsigned
1077 force_dst_alpha_one(unsigned factor
, boolean alpha
)
1080 case PIPE_BLENDFACTOR_DST_ALPHA
:
1081 return PIPE_BLENDFACTOR_ONE
;
1082 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1083 return PIPE_BLENDFACTOR_ZERO
;
1084 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1085 return PIPE_BLENDFACTOR_ZERO
;
1090 case PIPE_BLENDFACTOR_DST_COLOR
:
1091 return PIPE_BLENDFACTOR_ONE
;
1092 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
1093 return PIPE_BLENDFACTOR_ZERO
;
1102 * We need to generate several variants of the fragment pipeline to match
1103 * all the combinations of the contributing state atoms.
1105 * TODO: there is actually no reason to tie this to context state -- the
1106 * generated code could be cached globally in the screen.
1109 make_variant_key(struct llvmpipe_context
*lp
,
1110 struct lp_fragment_shader
*shader
,
1111 struct lp_fragment_shader_variant_key
*key
)
1115 memset(key
, 0, sizeof *key
);
1117 if (lp
->framebuffer
.zsbuf
) {
1118 if (lp
->depth_stencil
->depth
.enabled
) {
1119 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1120 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1122 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1123 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1124 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1128 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1129 if(key
->alpha
.enabled
)
1130 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1131 /* alpha.ref_value is passed in jit_context */
1133 key
->flatshade
= lp
->rasterizer
->flatshade
;
1134 key
->scissor
= lp
->rasterizer
->scissor
;
1135 if (lp
->active_query_count
) {
1136 key
->occlusion_count
= TRUE
;
1139 if (lp
->framebuffer
.nr_cbufs
) {
1140 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1143 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1144 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1145 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1146 const struct util_format_description
*format_desc
;
1149 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1150 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1151 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1153 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1155 /* mask out color channels not present in the color buffer.
1156 * Should be simple to incorporate per-cbuf writemasks:
1158 for(chan
= 0; chan
< 4; ++chan
) {
1159 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1161 if(swizzle
> UTIL_FORMAT_SWIZZLE_W
)
1162 blend_rt
->colormask
&= ~(1 << chan
);
1166 * Our swizzled render tiles always have an alpha channel, but the linear
1167 * render target format often does not, so force here the dst alpha to be
1170 * This is not a mere optimization. Wrong results will be produced if the
1171 * dst alpha is used, the dst format does not have alpha, and the previous
1172 * rendering was not flushed from the swizzled to linear buffer. For
1173 * example, NonPowTwo DCT.
1175 * TODO: This should be generalized to all channels for better
1176 * performance, but only alpha causes correctness issues.
1178 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1179 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
, FALSE
);
1180 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
, FALSE
);
1181 blend_rt
->alpha_src_factor
= force_dst_alpha_one(blend_rt
->alpha_src_factor
, TRUE
);
1182 blend_rt
->alpha_dst_factor
= force_dst_alpha_one(blend_rt
->alpha_dst_factor
, TRUE
);
1186 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1187 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1188 lp_sampler_static_state(&key
->sampler
[i
], lp
->fragment_sampler_views
[i
], lp
->sampler
[i
]);
1193 * Update fragment state. This is called just prior to drawing
1194 * something when some fragment-related state has changed.
1197 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1199 struct lp_fragment_shader
*shader
= lp
->fs
;
1200 struct lp_fragment_shader_variant_key key
;
1201 struct lp_fragment_shader_variant
*variant
;
1204 make_variant_key(lp
, shader
, &key
);
1206 variant
= shader
->variants
;
1208 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1211 variant
= variant
->next
;
1219 variant
= generate_variant(lp
, shader
, &key
);
1223 LP_COUNT_ADD(llvm_compile_time
, dt
);
1224 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1227 /* TODO: put this in the variant */
1228 /* TODO: most of these can be relaxed, in particular the colormask */
1229 opaque
= !key
.blend
.logicop_enable
&&
1230 !key
.blend
.rt
[0].blend_enable
&&
1231 key
.blend
.rt
[0].colormask
== 0xf &&
1232 !key
.stencil
[0].enabled
&&
1233 !key
.alpha
.enabled
&&
1234 !key
.depth
.enabled
&&
1236 !shader
->info
.uses_kill
1239 lp_setup_set_fs_functions(lp
->setup
,
1240 variant
->jit_function
[RAST_WHOLE
],
1241 variant
->jit_function
[RAST_EDGE_TEST
],
1248 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1250 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1251 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1252 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1254 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;