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_context.h"
91 #include "lp_screen.h"
94 #include "lp_tex_sample.h"
97 #include <llvm-c/Analysis.h>
100 static const unsigned char quad_offset_x
[4] = {0, 1, 0, 1};
101 static const unsigned char quad_offset_y
[4] = {0, 0, 1, 1};
105 * Derive from the quad's upper left scalar coordinates the coordinates for
106 * all other quad pixels
109 generate_pos0(LLVMBuilderRef builder
,
115 LLVMTypeRef int_elem_type
= LLVMInt32Type();
116 LLVMTypeRef int_vec_type
= LLVMVectorType(int_elem_type
, QUAD_SIZE
);
117 LLVMTypeRef elem_type
= LLVMFloatType();
118 LLVMTypeRef vec_type
= LLVMVectorType(elem_type
, QUAD_SIZE
);
119 LLVMValueRef x_offsets
[QUAD_SIZE
];
120 LLVMValueRef y_offsets
[QUAD_SIZE
];
123 x
= lp_build_broadcast(builder
, int_vec_type
, x
);
124 y
= lp_build_broadcast(builder
, int_vec_type
, y
);
126 for(i
= 0; i
< QUAD_SIZE
; ++i
) {
127 x_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_x
[i
], 0);
128 y_offsets
[i
] = LLVMConstInt(int_elem_type
, quad_offset_y
[i
], 0);
131 x
= LLVMBuildAdd(builder
, x
, LLVMConstVector(x_offsets
, QUAD_SIZE
), "");
132 y
= LLVMBuildAdd(builder
, y
, LLVMConstVector(y_offsets
, QUAD_SIZE
), "");
134 *x0
= LLVMBuildSIToFP(builder
, x
, vec_type
, "");
135 *y0
= LLVMBuildSIToFP(builder
, y
, vec_type
, "");
140 * Generate the depth /stencil test code.
143 generate_depth_stencil(LLVMBuilderRef builder
,
144 const struct lp_fragment_shader_variant_key
*key
,
145 struct lp_type src_type
,
146 struct lp_build_mask_context
*mask
,
147 LLVMValueRef stencil_refs
[2],
149 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
,
202 * Generate the code to do inside/outside triangle testing for the
203 * four pixels in a 2x2 quad. This will set the four elements of the
204 * quad mask vector to 0 or ~0.
205 * \param i which quad of the quad group to test, in [0,3]
208 generate_tri_edge_mask(LLVMBuilderRef builder
,
210 LLVMValueRef
*mask
, /* ivec4, out */
211 LLVMValueRef c0
, /* int32 */
212 LLVMValueRef c1
, /* int32 */
213 LLVMValueRef c2
, /* int32 */
214 LLVMValueRef step0_ptr
, /* ivec4 */
215 LLVMValueRef step1_ptr
, /* ivec4 */
216 LLVMValueRef step2_ptr
) /* ivec4 */
218 #define OPTIMIZE_IN_OUT_TEST 0
219 #if OPTIMIZE_IN_OUT_TEST
220 struct lp_build_if_state ifctx
;
221 LLVMValueRef not_draw_all
;
223 struct lp_build_flow_context
*flow
;
224 struct lp_type i32_type
;
225 LLVMTypeRef i32vec4_type
, mask_type
;
226 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
227 LLVMValueRef in_out_mask
;
231 /* int32 vector type */
232 memset(&i32_type
, 0, sizeof i32_type
);
233 i32_type
.floating
= FALSE
; /* values are integers */
234 i32_type
.sign
= TRUE
; /* values are signed */
235 i32_type
.norm
= FALSE
; /* values are not normalized */
236 i32_type
.width
= 32; /* 32-bit int values */
237 i32_type
.length
= 4; /* 4 elements per vector */
239 i32vec4_type
= lp_build_int32_vec4_type();
241 mask_type
= LLVMIntType(32 * 4);
244 * Use a conditional here to do detailed pixel in/out testing.
245 * We only have to do this if c0 != INT_MIN.
247 flow
= lp_build_flow_create(builder
);
248 lp_build_flow_scope_begin(flow
);
251 #if OPTIMIZE_IN_OUT_TEST
252 /* not_draw_all = (c0 != INT_MIN) */
253 not_draw_all
= LLVMBuildICmp(builder
,
256 LLVMConstInt(LLVMInt32Type(), INT_MIN
, 0),
259 in_out_mask
= lp_build_const_int_vec(i32_type
, ~0);
262 lp_build_flow_scope_declare(flow
, &in_out_mask
);
264 /* if (not_draw_all) {... */
265 lp_build_if(&ifctx
, flow
, builder
, not_draw_all
);
268 LLVMValueRef step0_vec
, step1_vec
, step2_vec
;
269 LLVMValueRef m0_vec
, m1_vec
, m2_vec
;
270 LLVMValueRef index
, m
;
272 /* c0_vec = {c0, c0, c0, c0}
273 * Note that we emit this code four times but LLVM optimizes away
274 * three instances of it.
276 c0_vec
= lp_build_broadcast(builder
, i32vec4_type
, c0
);
277 c1_vec
= lp_build_broadcast(builder
, i32vec4_type
, c1
);
278 c2_vec
= lp_build_broadcast(builder
, i32vec4_type
, c2
);
279 lp_build_name(c0_vec
, "edgeconst0vec");
280 lp_build_name(c1_vec
, "edgeconst1vec");
281 lp_build_name(c2_vec
, "edgeconst2vec");
283 /* load step0vec, step1, step2 vec from memory */
284 index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
285 step0_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step0_ptr
, &index
, 1, ""), "");
286 step1_vec
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, step1_ptr
, &index
, 1, ""), "");
287 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 = step0_ptr[i] > c0_vec */
293 m0_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step0_vec
, c0_vec
);
294 m1_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step1_vec
, c1_vec
);
295 m2_vec
= lp_build_compare(builder
, i32_type
, PIPE_FUNC_GREATER
, step2_vec
, c2_vec
);
297 /* in_out_mask = m0_vec & m1_vec & m2_vec */
298 m
= LLVMBuildAnd(builder
, m0_vec
, m1_vec
, "");
299 in_out_mask
= LLVMBuildAnd(builder
, m
, m2_vec
, "");
300 lp_build_name(in_out_mask
, "inoutmaskvec");
302 #if OPTIMIZE_IN_OUT_TEST
303 lp_build_endif(&ifctx
);
307 lp_build_flow_scope_end(flow
);
308 lp_build_flow_destroy(flow
);
310 /* This is the initial alive/dead pixel mask for a quad of four pixels.
311 * It's an int[4] vector with each word set to 0 or ~0.
312 * Words will get cleared when pixels faile the Z test, etc.
319 generate_scissor_test(LLVMBuilderRef builder
,
320 LLVMValueRef context_ptr
,
321 const struct lp_build_interp_soa_context
*interp
,
324 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
325 LLVMValueRef xpos
= interp
->pos
[0], ypos
= interp
->pos
[1];
326 LLVMValueRef xmin
, ymin
, xmax
, ymax
;
327 LLVMValueRef m0
, m1
, m2
, m3
, m
;
329 /* xpos, ypos contain the window coords for the four pixels in the quad */
333 /* get the current scissor bounds, convert to vectors */
334 xmin
= lp_jit_context_scissor_xmin_value(builder
, context_ptr
);
335 xmin
= lp_build_broadcast(builder
, vec_type
, xmin
);
337 ymin
= lp_jit_context_scissor_ymin_value(builder
, context_ptr
);
338 ymin
= lp_build_broadcast(builder
, vec_type
, ymin
);
340 xmax
= lp_jit_context_scissor_xmax_value(builder
, context_ptr
);
341 xmax
= lp_build_broadcast(builder
, vec_type
, xmax
);
343 ymax
= lp_jit_context_scissor_ymax_value(builder
, context_ptr
);
344 ymax
= lp_build_broadcast(builder
, vec_type
, ymax
);
346 /* compare the fragment's position coordinates against the scissor bounds */
347 m0
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, xpos
, xmin
);
348 m1
= lp_build_compare(builder
, type
, PIPE_FUNC_GEQUAL
, ypos
, ymin
);
349 m2
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, xpos
, xmax
);
350 m3
= lp_build_compare(builder
, type
, PIPE_FUNC_LESS
, ypos
, ymax
);
352 /* AND all the masks together */
353 m
= LLVMBuildAnd(builder
, m0
, m1
, "");
354 m
= LLVMBuildAnd(builder
, m
, m2
, "");
355 m
= LLVMBuildAnd(builder
, m
, m3
, "");
357 lp_build_name(m
, "scissormask");
364 build_int32_vec_const(int value
)
366 struct lp_type i32_type
;
368 memset(&i32_type
, 0, sizeof i32_type
);
369 i32_type
.floating
= FALSE
; /* values are integers */
370 i32_type
.sign
= TRUE
; /* values are signed */
371 i32_type
.norm
= FALSE
; /* values are not normalized */
372 i32_type
.width
= 32; /* 32-bit int values */
373 i32_type
.length
= 4; /* 4 elements per vector */
374 return lp_build_const_int_vec(i32_type
, value
);
380 * Generate the fragment shader, depth/stencil test, and alpha tests.
381 * \param i which quad in the tile, in range [0,3]
382 * \param do_tri_test if 1, do triangle edge in/out testing
385 generate_fs(struct llvmpipe_context
*lp
,
386 struct lp_fragment_shader
*shader
,
387 const struct lp_fragment_shader_variant_key
*key
,
388 LLVMBuilderRef builder
,
390 LLVMValueRef context_ptr
,
392 const struct lp_build_interp_soa_context
*interp
,
393 struct lp_build_sampler_soa
*sampler
,
395 LLVMValueRef (*color
)[4],
396 LLVMValueRef depth_ptr
,
398 unsigned do_tri_test
,
402 LLVMValueRef step0_ptr
,
403 LLVMValueRef step1_ptr
,
404 LLVMValueRef step2_ptr
)
406 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
407 LLVMTypeRef elem_type
;
408 LLVMTypeRef vec_type
;
409 LLVMTypeRef int_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 elem_type
= lp_build_elem_type(type
);
427 vec_type
= lp_build_vec_type(type
);
428 int_vec_type
= lp_build_int_vec_type(type
);
430 consts_ptr
= lp_jit_context_constants(builder
, context_ptr
);
432 flow
= lp_build_flow_create(builder
);
434 memset(outputs
, 0, sizeof outputs
);
436 lp_build_flow_scope_begin(flow
);
438 /* Declare the color and z variables */
439 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
440 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
441 color
[cbuf
][chan
] = LLVMGetUndef(vec_type
);
442 lp_build_flow_scope_declare(flow
, &color
[cbuf
][chan
]);
445 lp_build_flow_scope_declare(flow
, &z
);
447 /* do triangle edge testing */
449 generate_tri_edge_mask(builder
, i
, pmask
,
450 c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
);
453 *pmask
= build_int32_vec_const(~0);
456 /* 'mask' will control execution based on quad's pixel alive/killed state */
457 lp_build_mask_begin(&mask
, flow
, type
, *pmask
);
461 generate_scissor_test(builder
, context_ptr
, interp
, type
);
462 lp_build_mask_update(&mask
, smask
);
465 early_depth_stencil_test
=
466 (key
->depth
.enabled
|| key
->stencil
[0].enabled
) &&
467 !key
->alpha
.enabled
&&
468 !shader
->info
.uses_kill
&&
469 !shader
->info
.writes_z
;
471 if (early_depth_stencil_test
)
472 generate_depth_stencil(builder
, key
,
474 stencil_refs
, z
, depth_ptr
, facing
);
476 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
477 consts_ptr
, interp
->pos
, interp
->inputs
,
480 for (attrib
= 0; attrib
< shader
->info
.num_outputs
; ++attrib
) {
481 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
482 if(outputs
[attrib
][chan
]) {
483 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
484 lp_build_name(out
, "output%u.%u.%c", i
, attrib
, "xyzw"[chan
]);
486 switch (shader
->info
.output_semantic_name
[attrib
]) {
487 case TGSI_SEMANTIC_COLOR
:
489 unsigned cbuf
= shader
->info
.output_semantic_index
[attrib
];
491 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
494 /* XXX: should the alpha reference value be passed separately? */
495 /* XXX: should only test the final assignment to alpha */
496 if(cbuf
== 0 && chan
== 3) {
497 LLVMValueRef alpha
= out
;
498 LLVMValueRef alpha_ref_value
;
499 alpha_ref_value
= lp_jit_context_alpha_ref_value(builder
, context_ptr
);
500 alpha_ref_value
= lp_build_broadcast(builder
, vec_type
, alpha_ref_value
);
501 lp_build_alpha_test(builder
, &key
->alpha
, type
,
502 &mask
, alpha
, alpha_ref_value
);
505 color
[cbuf
][chan
] = out
;
509 case TGSI_SEMANTIC_POSITION
:
518 if (!early_depth_stencil_test
)
519 generate_depth_stencil(builder
, key
,
521 stencil_refs
, z
, depth_ptr
, facing
);
523 lp_build_mask_end(&mask
);
525 lp_build_flow_scope_end(flow
);
527 lp_build_flow_destroy(flow
);
535 * Generate color blending and color output.
538 generate_blend(const struct pipe_blend_state
*blend
,
539 LLVMBuilderRef builder
,
541 LLVMValueRef context_ptr
,
544 LLVMValueRef dst_ptr
)
546 struct lp_build_context bld
;
547 struct lp_build_flow_context
*flow
;
548 struct lp_build_mask_context mask_ctx
;
549 LLVMTypeRef vec_type
;
550 LLVMTypeRef int_vec_type
;
551 LLVMValueRef const_ptr
;
557 lp_build_context_init(&bld
, builder
, type
);
559 flow
= lp_build_flow_create(builder
);
561 /* we'll use this mask context to skip blending if all pixels are dead */
562 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
564 vec_type
= lp_build_vec_type(type
);
565 int_vec_type
= lp_build_int_vec_type(type
);
567 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
568 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
569 LLVMPointerType(vec_type
, 0), "");
571 for(chan
= 0; chan
< 4; ++chan
) {
572 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
573 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
575 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
577 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
578 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
581 lp_build_blend_soa(builder
, blend
, type
, src
, dst
, con
, res
);
583 for(chan
= 0; chan
< 4; ++chan
) {
584 if(blend
->rt
[0].colormask
& (1 << chan
)) {
585 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
586 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
587 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
588 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
592 lp_build_mask_end(&mask_ctx
);
593 lp_build_flow_destroy(flow
);
597 /** casting function to avoid compiler warnings */
598 static lp_jit_frag_func
599 cast_voidptr_to_lp_jit_frag_func(void *p
)
605 assert(sizeof(tmp
.v
) == sizeof(tmp
.f
));
612 * Generate the runtime callable function for the whole fragment pipeline.
613 * Note that the function which we generate operates on a block of 16
614 * pixels at at time. The block contains 2x2 quads. Each quad contains
618 generate_fragment(struct llvmpipe_context
*lp
,
619 struct lp_fragment_shader
*shader
,
620 struct lp_fragment_shader_variant
*variant
,
621 unsigned do_tri_test
)
623 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
624 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
625 struct lp_type fs_type
;
626 struct lp_type blend_type
;
627 LLVMTypeRef fs_elem_type
;
628 LLVMTypeRef fs_vec_type
;
629 LLVMTypeRef fs_int_vec_type
;
630 LLVMTypeRef blend_vec_type
;
631 LLVMTypeRef blend_int_vec_type
;
632 LLVMTypeRef arg_types
[15];
633 LLVMTypeRef func_type
;
634 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
635 LLVMValueRef context_ptr
;
639 LLVMValueRef dadx_ptr
;
640 LLVMValueRef dady_ptr
;
641 LLVMValueRef color_ptr_ptr
;
642 LLVMValueRef depth_ptr
;
643 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
;
644 LLVMBasicBlockRef block
;
645 LLVMBuilderRef builder
;
648 struct lp_build_sampler_soa
*sampler
;
649 struct lp_build_interp_soa_context interp
;
650 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
651 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
652 LLVMValueRef blend_mask
;
653 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
654 LLVMValueRef function
;
662 /* TODO: actually pick these based on the fs and color buffer
663 * characteristics. */
665 memset(&fs_type
, 0, sizeof fs_type
);
666 fs_type
.floating
= TRUE
; /* floating point values */
667 fs_type
.sign
= TRUE
; /* values are signed */
668 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
669 fs_type
.width
= 32; /* 32-bit float */
670 fs_type
.length
= 4; /* 4 elements per vector */
671 num_fs
= 4; /* number of quads per block */
673 memset(&blend_type
, 0, sizeof blend_type
);
674 blend_type
.floating
= FALSE
; /* values are integers */
675 blend_type
.sign
= FALSE
; /* values are unsigned */
676 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
677 blend_type
.width
= 8; /* 8-bit ubyte values */
678 blend_type
.length
= 16; /* 16 elements per vector */
681 * Generate the function prototype. Any change here must be reflected in
682 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
685 fs_elem_type
= lp_build_elem_type(fs_type
);
686 fs_vec_type
= lp_build_vec_type(fs_type
);
687 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
689 blend_vec_type
= lp_build_vec_type(blend_type
);
690 blend_int_vec_type
= lp_build_int_vec_type(blend_type
);
692 arg_types
[0] = screen
->context_ptr_type
; /* context */
693 arg_types
[1] = LLVMInt32Type(); /* x */
694 arg_types
[2] = LLVMInt32Type(); /* y */
695 arg_types
[3] = LLVMFloatType(); /* facing */
696 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
697 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
698 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
699 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
700 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
701 arg_types
[9] = LLVMInt32Type(); /* c0 */
702 arg_types
[10] = LLVMInt32Type(); /* c1 */
703 arg_types
[11] = LLVMInt32Type(); /* c2 */
704 /* Note: the step arrays are built as int32[16] but we interpret
705 * them here as int32_vec4[4].
707 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
708 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
709 arg_types
[14] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
711 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
713 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
714 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
716 variant
->function
[do_tri_test
] = function
;
719 /* XXX: need to propagate noalias down into color param now we are
720 * passing a pointer-to-pointer?
722 for(i
= 0; i
< Elements(arg_types
); ++i
)
723 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
724 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
726 context_ptr
= LLVMGetParam(function
, 0);
727 x
= LLVMGetParam(function
, 1);
728 y
= LLVMGetParam(function
, 2);
729 facing
= LLVMGetParam(function
, 3);
730 a0_ptr
= LLVMGetParam(function
, 4);
731 dadx_ptr
= LLVMGetParam(function
, 5);
732 dady_ptr
= LLVMGetParam(function
, 6);
733 color_ptr_ptr
= LLVMGetParam(function
, 7);
734 depth_ptr
= LLVMGetParam(function
, 8);
735 c0
= LLVMGetParam(function
, 9);
736 c1
= LLVMGetParam(function
, 10);
737 c2
= LLVMGetParam(function
, 11);
738 step0_ptr
= LLVMGetParam(function
, 12);
739 step1_ptr
= LLVMGetParam(function
, 13);
740 step2_ptr
= LLVMGetParam(function
, 14);
742 lp_build_name(context_ptr
, "context");
743 lp_build_name(x
, "x");
744 lp_build_name(y
, "y");
745 lp_build_name(a0_ptr
, "a0");
746 lp_build_name(dadx_ptr
, "dadx");
747 lp_build_name(dady_ptr
, "dady");
748 lp_build_name(color_ptr_ptr
, "color_ptr");
749 lp_build_name(depth_ptr
, "depth");
750 lp_build_name(c0
, "c0");
751 lp_build_name(c1
, "c1");
752 lp_build_name(c2
, "c2");
753 lp_build_name(step0_ptr
, "step0");
754 lp_build_name(step1_ptr
, "step1");
755 lp_build_name(step2_ptr
, "step2");
761 block
= LLVMAppendBasicBlock(function
, "entry");
762 builder
= LLVMCreateBuilder();
763 LLVMPositionBuilderAtEnd(builder
, block
);
765 generate_pos0(builder
, x
, y
, &x0
, &y0
);
767 lp_build_interp_soa_init(&interp
,
771 a0_ptr
, dadx_ptr
, dady_ptr
,
774 /* code generated texture sampling */
775 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
777 /* loop over quads in the block */
778 for(i
= 0; i
< num_fs
; ++i
) {
779 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
780 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
781 LLVMValueRef depth_ptr_i
;
785 lp_build_interp_soa_update(&interp
, i
);
787 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
789 generate_fs(lp
, shader
, key
,
796 &fs_mask
[i
], /* output */
802 step0_ptr
, step1_ptr
, step2_ptr
);
804 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
805 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
806 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
809 sampler
->destroy(sampler
);
811 /* Loop over color outputs / color buffers to do blending.
813 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
814 LLVMValueRef color_ptr
;
815 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
818 * Convert the fs's output color and mask to fit to the blending type.
820 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
821 lp_build_conv(builder
, fs_type
, blend_type
,
822 fs_out_color
[cbuf
][chan
], num_fs
,
823 &blend_in_color
[chan
], 1);
824 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
827 lp_build_conv_mask(builder
, fs_type
, blend_type
,
831 color_ptr
= LLVMBuildLoad(builder
,
832 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
834 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
839 generate_blend(&key
->blend
,
848 LLVMBuildRetVoid(builder
);
850 LLVMDisposeBuilder(builder
);
853 /* Verify the LLVM IR. If invalid, dump and abort */
855 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
857 LLVMDumpValue(function
);
862 /* Apply optimizations to LLVM IR */
864 LLVMRunFunctionPassManager(screen
->pass
, function
);
866 if (LP_DEBUG
& DEBUG_JIT
) {
867 /* Print the LLVM IR to stderr */
868 LLVMDumpValue(function
);
873 * Translate the LLVM IR into machine code.
876 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
878 variant
->jit_function
[do_tri_test
] = cast_voidptr_to_lp_jit_frag_func(f
);
880 if (LP_DEBUG
& DEBUG_ASM
)
886 static struct lp_fragment_shader_variant
*
887 generate_variant(struct llvmpipe_context
*lp
,
888 struct lp_fragment_shader
*shader
,
889 const struct lp_fragment_shader_variant_key
*key
)
891 struct lp_fragment_shader_variant
*variant
;
893 if (LP_DEBUG
& DEBUG_JIT
) {
896 tgsi_dump(shader
->base
.tokens
, 0);
897 if(key
->depth
.enabled
) {
898 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
899 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
900 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
902 if(key
->alpha
.enabled
) {
903 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
904 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
906 if(key
->blend
.logicop_enable
) {
907 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
909 else if(key
->blend
.rt
[0].blend_enable
) {
910 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
911 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
912 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
913 debug_printf("alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
914 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
915 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
917 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
918 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
919 if(key
->sampler
[i
].format
) {
920 debug_printf("sampler[%u] = \n", i
);
921 debug_printf(" .format = %s\n",
922 util_format_name(key
->sampler
[i
].format
));
923 debug_printf(" .target = %s\n",
924 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
925 debug_printf(" .pot = %u %u %u\n",
926 key
->sampler
[i
].pot_width
,
927 key
->sampler
[i
].pot_height
,
928 key
->sampler
[i
].pot_depth
);
929 debug_printf(" .wrap = %s %s %s\n",
930 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
931 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
932 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
933 debug_printf(" .min_img_filter = %s\n",
934 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
935 debug_printf(" .min_mip_filter = %s\n",
936 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
937 debug_printf(" .mag_img_filter = %s\n",
938 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
939 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
940 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
941 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
946 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
950 variant
->shader
= shader
;
951 memcpy(&variant
->key
, key
, sizeof *key
);
953 generate_fragment(lp
, shader
, variant
, 0);
954 generate_fragment(lp
, shader
, variant
, 1);
956 /* insert new variant into linked list */
957 variant
->next
= shader
->variants
;
958 shader
->variants
= variant
;
965 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
966 const struct pipe_shader_state
*templ
)
968 struct lp_fragment_shader
*shader
;
970 shader
= CALLOC_STRUCT(lp_fragment_shader
);
974 /* get/save the summary info for this shader */
975 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
977 /* we need to keep a local copy of the tokens */
978 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
985 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
987 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
989 if (llvmpipe
->fs
== fs
)
992 draw_flush(llvmpipe
->draw
);
996 llvmpipe
->dirty
|= LP_NEW_FS
;
1001 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1003 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1004 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
1005 struct lp_fragment_shader
*shader
= fs
;
1006 struct lp_fragment_shader_variant
*variant
;
1008 assert(fs
!= llvmpipe
->fs
);
1012 * XXX: we need to flush the context until we have some sort of reference
1013 * counting in fragment shaders as they may still be binned
1015 draw_flush(llvmpipe
->draw
);
1016 lp_setup_flush(llvmpipe
->setup
, 0);
1018 variant
= shader
->variants
;
1020 struct lp_fragment_shader_variant
*next
= variant
->next
;
1023 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1024 if (variant
->function
[i
]) {
1025 if (variant
->jit_function
[i
])
1026 LLVMFreeMachineCodeForFunction(screen
->engine
,
1027 variant
->function
[i
]);
1028 LLVMDeleteFunction(variant
->function
[i
]);
1037 FREE((void *) shader
->base
.tokens
);
1044 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1045 uint shader
, uint index
,
1046 struct pipe_resource
*constants
)
1048 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1049 unsigned size
= constants
? constants
->width0
: 0;
1050 const void *data
= constants
? llvmpipe_resource(constants
)->data
: NULL
;
1052 assert(shader
< PIPE_SHADER_TYPES
);
1055 if(llvmpipe
->constants
[shader
] == constants
)
1058 draw_flush(llvmpipe
->draw
);
1060 /* note: reference counting */
1061 pipe_resource_reference(&llvmpipe
->constants
[shader
], constants
);
1063 if(shader
== PIPE_SHADER_VERTEX
) {
1064 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
, 0,
1068 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1073 * We need to generate several variants of the fragment pipeline to match
1074 * all the combinations of the contributing state atoms.
1076 * TODO: there is actually no reason to tie this to context state -- the
1077 * generated code could be cached globally in the screen.
1080 make_variant_key(struct llvmpipe_context
*lp
,
1081 struct lp_fragment_shader
*shader
,
1082 struct lp_fragment_shader_variant_key
*key
)
1086 memset(key
, 0, sizeof *key
);
1088 if (lp
->framebuffer
.zsbuf
) {
1089 if (lp
->depth_stencil
->depth
.enabled
) {
1090 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1091 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1093 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1094 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1095 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1099 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1100 if(key
->alpha
.enabled
)
1101 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1102 /* alpha.ref_value is passed in jit_context */
1104 key
->flatshade
= lp
->rasterizer
->flatshade
;
1105 key
->scissor
= lp
->rasterizer
->scissor
;
1107 if (lp
->framebuffer
.nr_cbufs
) {
1108 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1111 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1112 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1113 const struct util_format_description
*format_desc
;
1116 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1117 assert(format_desc
->layout
== UTIL_FORMAT_COLORSPACE_RGB
||
1118 format_desc
->layout
== UTIL_FORMAT_COLORSPACE_SRGB
);
1120 /* mask out color channels not present in the color buffer.
1121 * Should be simple to incorporate per-cbuf writemasks:
1123 for(chan
= 0; chan
< 4; ++chan
) {
1124 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1126 if(swizzle
<= UTIL_FORMAT_SWIZZLE_W
)
1127 key
->blend
.rt
[0].colormask
|= (1 << chan
);
1131 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1132 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1133 lp_sampler_static_state(&key
->sampler
[i
], lp
->fragment_sampler_views
[i
]->texture
, lp
->sampler
[i
]);
1138 * Update fragment state. This is called just prior to drawing
1139 * something when some fragment-related state has changed.
1142 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1144 struct lp_fragment_shader
*shader
= lp
->fs
;
1145 struct lp_fragment_shader_variant_key key
;
1146 struct lp_fragment_shader_variant
*variant
;
1149 make_variant_key(lp
, shader
, &key
);
1151 variant
= shader
->variants
;
1153 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1156 variant
= variant
->next
;
1164 variant
= generate_variant(lp
, shader
, &key
);
1168 LP_COUNT_ADD(llvm_compile_time
, dt
);
1169 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1172 shader
->current
= variant
;
1174 /* TODO: put this in the variant */
1175 /* TODO: most of these can be relaxed, in particular the colormask */
1176 opaque
= !key
.blend
.logicop_enable
&&
1177 !key
.blend
.rt
[0].blend_enable
&&
1178 key
.blend
.rt
[0].colormask
== 0xf &&
1179 !key
.stencil
[0].enabled
&&
1180 !key
.alpha
.enabled
&&
1181 !key
.depth
.enabled
&&
1183 !shader
->info
.uses_kill
1186 lp_setup_set_fs_functions(lp
->setup
,
1187 shader
->current
->jit_function
[RAST_WHOLE
],
1188 shader
->current
->jit_function
[RAST_EDGE_TEST
],