1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
34 * - triangle edge in/out testing
40 * - depth/stencil test (stencil TBI)
43 * This file has only the glue to 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_depth.h"
81 #include "gallivm/lp_bld_interp.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_alpha.h"
84 #include "gallivm/lp_bld_blend.h"
85 #include "gallivm/lp_bld_swizzle.h"
86 #include "gallivm/lp_bld_flow.h"
87 #include "gallivm/lp_bld_debug.h"
88 #include "lp_buffer.h"
89 #include "lp_context.h"
92 #include "lp_screen.h"
95 #include "lp_tex_sample.h"
98 #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
,
150 LLVMValueRef dst_ptr
)
152 const struct util_format_description
*format_desc
;
153 struct lp_type dst_type
;
155 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
158 format_desc
= util_format_description(key
->zsbuf_format
);
162 * Depths are expected to be between 0 and 1, even if they are stored in
163 * floats. Setting these bits here will ensure that the lp_build_conv() call
164 * below won't try to unnecessarily clamp the incoming values.
166 if(src_type
.floating
) {
167 src_type
.sign
= FALSE
;
168 src_type
.norm
= TRUE
;
171 assert(!src_type
.sign
);
172 assert(src_type
.norm
);
175 /* Pick the depth type. */
176 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
178 /* FIXME: Cope with a depth test type with a different bit width. */
179 assert(dst_type
.width
== src_type
.width
);
180 assert(dst_type
.length
== src_type
.length
);
182 /* Convert fragment Z from float to integer */
183 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
185 dst_ptr
= LLVMBuildBitCast(builder
,
187 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
188 lp_build_depth_stencil_test(builder
,
201 * Generate the code to do inside/outside triangle testing for the
202 * four pixels in a 2x2 quad. This will set the four elements of the
203 * quad mask vector to 0 or ~0.
204 * \param i which quad of the quad group to test, in [0,3]
207 generate_tri_edge_mask(LLVMBuilderRef builder
,
209 LLVMValueRef
*mask
, /* ivec4, out */
210 LLVMValueRef c0
, /* int32 */
211 LLVMValueRef c1
, /* int32 */
212 LLVMValueRef c2
, /* int32 */
213 LLVMValueRef step0_ptr
, /* ivec4 */
214 LLVMValueRef step1_ptr
, /* ivec4 */
215 LLVMValueRef step2_ptr
) /* ivec4 */
217 #define OPTIMIZE_IN_OUT_TEST 0
218 #if OPTIMIZE_IN_OUT_TEST
219 struct lp_build_if_state ifctx
;
220 LLVMValueRef not_draw_all
;
222 struct lp_build_flow_context
*flow
;
223 struct lp_type i32_type
;
224 LLVMTypeRef i32vec4_type
, mask_type
;
225 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
226 LLVMValueRef in_out_mask
;
230 /* int32 vector type */
231 memset(&i32_type
, 0, sizeof i32_type
);
232 i32_type
.floating
= FALSE
; /* values are integers */
233 i32_type
.sign
= TRUE
; /* values are signed */
234 i32_type
.norm
= FALSE
; /* values are not normalized */
235 i32_type
.width
= 32; /* 32-bit int values */
236 i32_type
.length
= 4; /* 4 elements per vector */
238 i32vec4_type
= lp_build_int32_vec4_type();
240 mask_type
= LLVMIntType(32 * 4);
243 * Use a conditional here to do detailed pixel in/out testing.
244 * We only have to do this if c0 != INT_MIN.
246 flow
= lp_build_flow_create(builder
);
247 lp_build_flow_scope_begin(flow
);
250 #if OPTIMIZE_IN_OUT_TEST
251 /* not_draw_all = (c0 != INT_MIN) */
252 not_draw_all
= LLVMBuildICmp(builder
,
255 LLVMConstInt(LLVMInt32Type(), INT_MIN
, 0),
258 in_out_mask
= lp_build_const_int_vec(i32_type
, ~0);
261 lp_build_flow_scope_declare(flow
, &in_out_mask
);
263 /* if (not_draw_all) {... */
264 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
267 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
268 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
269 LLVMValueRef index
, m
;
271 /* c0_vec = {c0, c0, c0, c0}
272 * Note that we emit this code four times but LLVM optimizes away
273 * three instances of it.
275 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
276 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
277 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
278 lp_build_name(c0_vec
, "edgeconst0vec");
279 lp_build_name(c1_vec
, "edgeconst1vec");
280 lp_build_name(c2_vec
, "edgeconst2vec");
282 /* load step0vec, step1, step2 vec from memory */
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, ""), "");
287 lp_build_name(step0_vec
, "step0vec");
288 lp_build_name(step1_vec
, "step1vec");
289 lp_build_name(step2_vec
, "step2vec");
291 /* m0_vec = step0_ptr[i] > c0_vec */
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 /* in_out_mask = m0_vec & m1_vec & m2_vec */
297 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
298 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
299 lp_build_name(in_out_mask
, "inoutmaskvec");
301 #if OPTIMIZE_IN_OUT_TEST
302 lp_build_endif(&ifctx
);
306 lp_build_flow_scope_end(flow
);
307 lp_build_flow_destroy(flow
);
309 /* This is the initial alive/dead pixel mask for a quad of four pixels.
310 * It's an int[4] vector with each word set to 0 or ~0.
311 * Words will get cleared when pixels faile the Z test, etc.
318 generate_scissor_test(LLVMBuilderRef builder
,
319 LLVMValueRef context_ptr
,
320 const struct lp_build_interp_soa_context
*interp
,
323 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
324 LLVMValueRef xpos
= interp
->pos
[0], ypos
= interp
->pos
[1];
325 LLVMValueRef xmin
, ymin
, xmax
, ymax
;
326 LLVMValueRef m0
, m1
, m2
, m3
, m
;
328 /* xpos, ypos contain the window coords for the four pixels in the quad */
332 /* get the current scissor bounds, convert to vectors */
333 xmin
= lp_jit_context_scissor_xmin_value(builder
, context_ptr
);
334 xmin
= lp_build_broadcast(builder
, vec_type
, xmin
);
336 ymin
= lp_jit_context_scissor_ymin_value(builder
, context_ptr
);
337 ymin
= lp_build_broadcast(builder
, vec_type
, ymin
);
339 xmax
= lp_jit_context_scissor_xmax_value(builder
, context_ptr
);
340 xmax
= lp_build_broadcast(builder
, vec_type
, xmax
);
342 ymax
= lp_jit_context_scissor_ymax_value(builder
, context_ptr
);
343 ymax
= lp_build_broadcast(builder
, vec_type
, ymax
);
345 /* compare the fragment's position coordinates against the scissor bounds */
346 m0
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, xpos
, xmin
);
347 m1
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, ypos
, ymin
);
348 m2
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, xpos
, xmax
);
349 m3
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, ypos
, ymax
);
351 /* AND all the masks together */
352 m
= LLVMBuildAnd(builder
, m0
, m1
, "");
353 m
= LLVMBuildAnd(builder
, m
, m2
, "");
354 m
= LLVMBuildAnd(builder
, m
, m3
, "");
356 lp_build_name(m
, "scissormask");
363 build_int32_vec_const(int value
)
365 struct lp_type i32_type
;
367 memset(&i32_type
, 0, sizeof i32_type
);
368 i32_type
.floating
= FALSE
; /* values are integers */
369 i32_type
.sign
= TRUE
; /* values are signed */
370 i32_type
.norm
= FALSE
; /* values are not normalized */
371 i32_type
.width
= 32; /* 32-bit int values */
372 i32_type
.length
= 4; /* 4 elements per vector */
373 return lp_build_const_int_vec(i32_type
, value
);
379 * Generate the fragment shader, depth/stencil test, and alpha tests.
380 * \param i which quad in the tile, in range [0,3]
381 * \param do_tri_test if 1, do triangle edge in/out testing
384 generate_fs(struct llvmpipe_context
*lp
,
385 struct lp_fragment_shader
*shader
,
386 const struct lp_fragment_shader_variant_key
*key
,
387 LLVMBuilderRef builder
,
389 LLVMValueRef context_ptr
,
391 const struct lp_build_interp_soa_context
*interp
,
392 struct lp_build_sampler_soa
*sampler
,
394 LLVMValueRef (*color
)[4],
395 LLVMValueRef depth_ptr
,
396 unsigned do_tri_test
,
400 LLVMValueRef step0_ptr
,
401 LLVMValueRef step1_ptr
,
402 LLVMValueRef step2_ptr
)
404 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
405 LLVMTypeRef elem_type
;
406 LLVMTypeRef vec_type
;
407 LLVMTypeRef int_vec_type
;
408 LLVMValueRef consts_ptr
;
409 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
410 LLVMValueRef z
= interp
->pos
[2];
411 LLVMValueRef stencil_refs
;
412 struct lp_build_flow_context
*flow
;
413 struct lp_build_mask_context mask
;
414 boolean early_depth_stencil_test
;
421 stencil_refs
= lp_jit_context_stencil_ref_values(builder
, context_ptr
);
423 elem_type
= lp_build_elem_type(type
);
424 vec_type
= lp_build_vec_type(type
);
425 int_vec_type
= lp_build_int_vec_type(type
);
427 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
429 flow
= lp_build_flow_create(builder
);
431 memset(outputs
, 0, sizeof outputs
);
433 lp_build_flow_scope_begin(flow
);
435 /* Declare the color and z variables */
436 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
437 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
438 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
439 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
442 lp_build_flow_scope_declare(flow
, &z
);
444 /* do triangle edge testing */
446 generate_tri_edge_mask(builder
, i
, pmask
,
447 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
450 *pmask
= build_int32_vec_const(~0);
453 /* 'mask' will control execution based on quad's pixel alive/killed state */
454 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
458 generate_scissor_test(builder
, context_ptr
, interp
, type
);
459 lp_build_mask_update(&mask
, smask
);
462 early_depth_stencil_test
=
463 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
464 !key
->alpha
.enabled
&&
465 !shader
->info
.uses_kill
&&
466 !shader
->info
.writes_z
;
468 if (early_depth_stencil_test
)
469 generate_depth_stencil(builder
, key
,
471 stencil_refs
, z
, depth_ptr
);
473 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
474 consts_ptr
, interp
->pos
, interp
->inputs
,
477 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
478 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
479 if(outputs
[attrib
][chan
]) {
480 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
481 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
483 switch (shader
->info
.output_semantic_name
[attrib
]) {
484 case TGSI_SEMANTIC_COLOR
:
486 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
488 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
491 /* XXX: should the alpha reference value be passed separately? */
492 /* XXX: should only test the final assignment to alpha */
493 if(cbuf
== 0 && chan
== 3) {
494 LLVMValueRef alpha
= out
;
495 LLVMValueRef alpha_ref_value
;
496 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
497 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
498 lp_build_alpha_test(builder
, &key
->alpha
, type
,
499 &mask
, alpha
, alpha_ref_value
);
502 color
[cbuf
][chan
] = out
;
506 case TGSI_SEMANTIC_POSITION
:
515 if (!early_depth_stencil_test
)
516 generate_depth_stencil(builder
, key
,
518 stencil_refs
, z
, depth_ptr
);
520 lp_build_mask_end(&mask
);
522 lp_build_flow_scope_end(flow
);
524 lp_build_flow_destroy(flow
);
532 * Generate color blending and color output.
535 generate_blend(const struct pipe_blend_state
*blend
,
536 LLVMBuilderRef builder
,
538 LLVMValueRef context_ptr
,
541 LLVMValueRef dst_ptr
)
543 struct lp_build_context bld
;
544 struct lp_build_flow_context
*flow
;
545 struct lp_build_mask_context mask_ctx
;
546 LLVMTypeRef vec_type
;
547 LLVMTypeRef int_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
);
562 int_vec_type
= lp_build_int_vec_type(type
);
564 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
565 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
566 LLVMPointerType(vec_type
, 0), "");
568 for(chan
= 0; chan
< 4; ++chan
) {
569 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
570 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
572 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
574 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
575 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
578 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
580 for(chan
= 0; chan
< 4; ++chan
) {
581 if(blend
->rt
[0].colormask
& (1 << chan
)) {
582 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
583 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
584 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
585 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
589 lp_build_mask_end(&mask_ctx
);
590 lp_build_flow_destroy(flow
);
594 /** casting function to avoid compiler warnings */
595 static lp_jit_frag_func
596 cast_voidptr_to_lp_jit_frag_func(void *p
)
602 assert(sizeof(tmp
.v
) == sizeof(tmp
.f
));
609 * Generate the runtime callable function for the whole fragment pipeline.
610 * Note that the function which we generate operates on a block of 16
611 * pixels at at time. The block contains 2x2 quads. Each quad contains
615 generate_fragment(struct llvmpipe_context
*lp
,
616 struct lp_fragment_shader
*shader
,
617 struct lp_fragment_shader_variant
*variant
,
618 unsigned do_tri_test
)
620 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
621 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
622 struct lp_type fs_type
;
623 struct lp_type blend_type
;
624 LLVMTypeRef fs_elem_type
;
625 LLVMTypeRef fs_vec_type
;
626 LLVMTypeRef fs_int_vec_type
;
627 LLVMTypeRef blend_vec_type
;
628 LLVMTypeRef blend_int_vec_type
;
629 LLVMTypeRef arg_types
[14];
630 LLVMTypeRef func_type
;
631 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
632 LLVMValueRef context_ptr
;
636 LLVMValueRef dadx_ptr
;
637 LLVMValueRef dady_ptr
;
638 LLVMValueRef color_ptr_ptr
;
639 LLVMValueRef depth_ptr
;
640 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
;
641 LLVMBasicBlockRef block
;
642 LLVMBuilderRef builder
;
645 struct lp_build_sampler_soa
*sampler
;
646 struct lp_build_interp_soa_context interp
;
647 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
648 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
649 LLVMValueRef blend_mask
;
650 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
651 LLVMValueRef function
;
658 /* TODO: actually pick these based on the fs and color buffer
659 * characteristics. */
661 memset(&fs_type
, 0, sizeof fs_type
);
662 fs_type
.floating
= TRUE
; /* floating point values */
663 fs_type
.sign
= TRUE
; /* values are signed */
664 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
665 fs_type
.width
= 32; /* 32-bit float */
666 fs_type
.length
= 4; /* 4 elements per vector */
667 num_fs
= 4; /* number of quads per block */
669 memset(&blend_type
, 0, sizeof blend_type
);
670 blend_type
.floating
= FALSE
; /* values are integers */
671 blend_type
.sign
= FALSE
; /* values are unsigned */
672 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
673 blend_type
.width
= 8; /* 8-bit ubyte values */
674 blend_type
.length
= 16; /* 16 elements per vector */
677 * Generate the function prototype. Any change here must be reflected in
678 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
681 fs_elem_type
= lp_build_elem_type(fs_type
);
682 fs_vec_type
= lp_build_vec_type(fs_type
);
683 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
685 blend_vec_type
= lp_build_vec_type(blend_type
);
686 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
688 arg_types
[0] = screen
->context_ptr_type
; /* context */
689 arg_types
[1] = LLVMInt32Type(); /* x */
690 arg_types
[2] = LLVMInt32Type(); /* y */
691 arg_types
[3] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
692 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
693 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dady */
694 arg_types
[6] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
695 arg_types
[7] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
696 arg_types
[8] = LLVMInt32Type(); /* c0 */
697 arg_types
[9] = LLVMInt32Type(); /* c1 */
698 arg_types
[10] = LLVMInt32Type(); /* c2 */
699 /* Note: the step arrays are built as int32[16] but we interpret
700 * them here as int32_vec4[4].
702 arg_types
[11] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
703 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
704 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
706 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
708 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
709 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
711 variant
->function
[do_tri_test
] = function
;
714 /* XXX: need to propagate noalias down into color param now we are
715 * passing a pointer-to-pointer?
717 for(i
= 0; i
< Elements(arg_types
); ++i
)
718 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
719 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
721 context_ptr
= LLVMGetParam(function
, 0);
722 x
= LLVMGetParam(function
, 1);
723 y
= LLVMGetParam(function
, 2);
724 a0_ptr
= LLVMGetParam(function
, 3);
725 dadx_ptr
= LLVMGetParam(function
, 4);
726 dady_ptr
= LLVMGetParam(function
, 5);
727 color_ptr_ptr
= LLVMGetParam(function
, 6);
728 depth_ptr
= LLVMGetParam(function
, 7);
729 c0
= LLVMGetParam(function
, 8);
730 c1
= LLVMGetParam(function
, 9);
731 c2
= LLVMGetParam(function
, 10);
732 step0_ptr
= LLVMGetParam(function
, 11);
733 step1_ptr
= LLVMGetParam(function
, 12);
734 step2_ptr
= LLVMGetParam(function
, 13);
736 lp_build_name(context_ptr
, "context");
737 lp_build_name(x
, "x");
738 lp_build_name(y
, "y");
739 lp_build_name(a0_ptr
, "a0");
740 lp_build_name(dadx_ptr
, "dadx");
741 lp_build_name(dady_ptr
, "dady");
742 lp_build_name(color_ptr_ptr
, "color_ptr");
743 lp_build_name(depth_ptr
, "depth");
744 lp_build_name(c0
, "c0");
745 lp_build_name(c1
, "c1");
746 lp_build_name(c2
, "c2");
747 lp_build_name(step0_ptr
, "step0");
748 lp_build_name(step1_ptr
, "step1");
749 lp_build_name(step2_ptr
, "step2");
755 block
= LLVMAppendBasicBlock(function
, "entry");
756 builder
= LLVMCreateBuilder();
757 LLVMPositionBuilderAtEnd(builder
, block
);
759 generate_pos0(builder
, x
, y
, &x0
, &y0
);
761 lp_build_interp_soa_init(&interp
,
765 a0_ptr
, dadx_ptr
, dady_ptr
,
768 /* code generated texture sampling */
769 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
771 /* loop over quads in the block */
772 for(i
= 0; i
< num_fs
; ++i
) {
773 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
774 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
775 LLVMValueRef depth_ptr_i
;
779 lp_build_interp_soa_update(&interp
, i
);
781 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
783 generate_fs(lp
, shader
, key
,
790 &fs_mask
[i
], /* output */
795 step0_ptr
, step1_ptr
, step2_ptr
);
797 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
798 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
799 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
802 sampler
->destroy(sampler
);
804 /* Loop over color outputs / color buffers to do blending.
806 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
807 LLVMValueRef color_ptr
;
808 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
811 * Convert the fs's output color and mask to fit to the blending type.
813 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
814 lp_build_conv(builder
, fs_type
, blend_type
,
815 fs_out_color
[cbuf
][chan
], num_fs
,
816 &blend_in_color
[chan
], 1);
817 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
820 lp_build_conv_mask(builder
, fs_type
, blend_type
,
824 color_ptr
= LLVMBuildLoad(builder
,
825 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
827 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
832 generate_blend(&key
->blend
,
841 LLVMBuildRetVoid(builder
);
843 LLVMDisposeBuilder(builder
);
846 /* Verify the LLVM IR. If invalid, dump and abort */
848 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
850 LLVMDumpValue(function
);
855 /* Apply optimizations to LLVM IR */
857 LLVMRunFunctionPassManager(screen
->pass
, function
);
859 if (LP_DEBUG
& DEBUG_JIT
) {
860 /* Print the LLVM IR to stderr */
861 LLVMDumpValue(function
);
866 * Translate the LLVM IR into machine code.
869 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
871 variant
->jit_function
[do_tri_test
] = cast_voidptr_to_lp_jit_frag_func(f
);
873 if (LP_DEBUG
& DEBUG_ASM
)
879 static struct lp_fragment_shader_variant
*
880 generate_variant(struct llvmpipe_context
*lp
,
881 struct lp_fragment_shader
*shader
,
882 const struct lp_fragment_shader_variant_key
*key
)
884 struct lp_fragment_shader_variant
*variant
;
886 if (LP_DEBUG
& DEBUG_JIT
) {
889 tgsi_dump(shader
->base
.tokens
, 0);
890 if(key
->depth
.enabled
) {
891 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
892 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
893 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
895 if(key
->alpha
.enabled
) {
896 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
897 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
899 if(key
->blend
.logicop_enable
) {
900 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
902 else if(key
->blend
.rt
[0].blend_enable
) {
903 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
904 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
905 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
906 debug_printf("alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
907 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
908 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
910 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
911 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
912 if(key
->sampler
[i
].format
) {
913 debug_printf("sampler[%u] = \n", i
);
914 debug_printf(" .format = %s\n",
915 util_format_name(key
->sampler
[i
].format
));
916 debug_printf(" .target = %s\n",
917 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
918 debug_printf(" .pot = %u %u %u\n",
919 key
->sampler
[i
].pot_width
,
920 key
->sampler
[i
].pot_height
,
921 key
->sampler
[i
].pot_depth
);
922 debug_printf(" .wrap = %s %s %s\n",
923 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
924 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
925 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
926 debug_printf(" .min_img_filter = %s\n",
927 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
928 debug_printf(" .min_mip_filter = %s\n",
929 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
930 debug_printf(" .mag_img_filter = %s\n",
931 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
932 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
933 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
934 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
939 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
943 variant
->shader
= shader
;
944 memcpy(&variant
->key
, key
, sizeof *key
);
946 generate_fragment(lp
, shader
, variant
, 0);
947 generate_fragment(lp
, shader
, variant
, 1);
949 /* insert new variant into linked list */
950 variant
->next
= shader
->variants
;
951 shader
->variants
= variant
;
958 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
959 const struct pipe_shader_state
*templ
)
961 struct lp_fragment_shader
*shader
;
963 shader
= CALLOC_STRUCT(lp_fragment_shader
);
967 /* get/save the summary info for this shader */
968 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
970 /* we need to keep a local copy of the tokens */
971 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
978 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
980 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
982 if (llvmpipe
->fs
== fs
)
985 draw_flush(llvmpipe
->draw
);
989 llvmpipe
->dirty
|= LP_NEW_FS
;
994 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
996 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
997 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
998 struct lp_fragment_shader
*shader
= fs
;
999 struct lp_fragment_shader_variant
*variant
;
1001 assert(fs
!= llvmpipe
->fs
);
1005 * XXX: we need to flush the context until we have some sort of reference
1006 * counting in fragment shaders as they may still be binned
1008 draw_flush(llvmpipe
->draw
);
1009 lp_setup_flush(llvmpipe
->setup
, 0);
1011 variant
= shader
->variants
;
1013 struct lp_fragment_shader_variant
*next
= variant
->next
;
1016 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1017 if (variant
->function
[i
]) {
1018 if (variant
->jit_function
[i
])
1019 LLVMFreeMachineCodeForFunction(screen
->engine
,
1020 variant
->function
[i
]);
1021 LLVMDeleteFunction(variant
->function
[i
]);
1030 FREE((void *) shader
->base
.tokens
);
1037 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1038 uint shader
, uint index
,
1039 struct pipe_buffer
*constants
)
1041 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1042 unsigned size
= constants
? constants
->size
: 0;
1043 const void *data
= constants
? llvmpipe_buffer(constants
)->data
: NULL
;
1045 assert(shader
< PIPE_SHADER_TYPES
);
1048 if(llvmpipe
->constants
[shader
] == constants
)
1051 draw_flush(llvmpipe
->draw
);
1053 /* note: reference counting */
1054 pipe_buffer_reference(&llvmpipe
->constants
[shader
], constants
);
1056 if(shader
== PIPE_SHADER_VERTEX
) {
1057 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
, 0,
1061 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1066 * We need to generate several variants of the fragment pipeline to match
1067 * all the combinations of the contributing state atoms.
1069 * TODO: there is actually no reason to tie this to context state -- the
1070 * generated code could be cached globally in the screen.
1073 make_variant_key(struct llvmpipe_context
*lp
,
1074 struct lp_fragment_shader
*shader
,
1075 struct lp_fragment_shader_variant_key
*key
)
1079 memset(key
, 0, sizeof *key
);
1081 if (lp
->framebuffer
.zsbuf
) {
1082 if (lp
->depth_stencil
->depth
.enabled
) {
1083 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1084 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1086 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1087 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1088 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1092 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1093 if(key
->alpha
.enabled
)
1094 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1095 /* alpha.ref_value is passed in jit_context */
1097 key
->flatshade
= lp
->rasterizer
->flatshade
;
1098 key
->scissor
= lp
->rasterizer
->scissor
;
1100 if (lp
->framebuffer
.nr_cbufs
) {
1101 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1104 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1105 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1106 const struct util_format_description
*format_desc
;
1109 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1110 assert(format_desc
->layout
== UTIL_FORMAT_COLORSPACE_RGB
||
1111 format_desc
->layout
== UTIL_FORMAT_COLORSPACE_SRGB
);
1113 /* mask out color channels not present in the color buffer.
1114 * Should be simple to incorporate per-cbuf writemasks:
1116 for(chan
= 0; chan
< 4; ++chan
) {
1117 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1119 if(swizzle
<= UTIL_FORMAT_SWIZZLE_W
)
1120 key
->blend
.rt
[0].colormask
|= (1 << chan
);
1124 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1125 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1126 lp_sampler_static_state(&key
->sampler
[i
], lp
->fragment_sampler_views
[i
]->texture
, lp
->sampler
[i
]);
1131 * Update fragment state. This is called just prior to drawing
1132 * something when some fragment-related state has changed.
1135 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1137 struct lp_fragment_shader
*shader
= lp
->fs
;
1138 struct lp_fragment_shader_variant_key key
;
1139 struct lp_fragment_shader_variant
*variant
;
1142 make_variant_key(lp
, shader
, &key
);
1144 variant
= shader
->variants
;
1146 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1149 variant
= variant
->next
;
1157 variant
= generate_variant(lp
, shader
, &key
);
1161 LP_COUNT_ADD(llvm_compile_time
, dt
);
1162 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1165 shader
->current
= variant
;
1167 /* TODO: put this in the variant */
1168 /* TODO: most of these can be relaxed, in particular the colormask */
1169 opaque
= !key
.blend
.logicop_enable
&&
1170 !key
.blend
.rt
[0].blend_enable
&&
1171 key
.blend
.rt
[0].colormask
== 0xf &&
1172 !key
.stencil
[0].enabled
&&
1173 !key
.alpha
.enabled
&&
1174 !key
.depth
.enabled
&&
1176 !shader
->info
.uses_kill
1179 lp_setup_set_fs_functions(lp
->setup
,
1180 shader
->current
->jit_function
[0],
1181 shader
->current
->jit_function
[1],