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"
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
,
151 LLVMValueRef counter
)
153 const struct util_format_description
*format_desc
;
154 struct lp_type dst_type
;
156 if (!key
->depth
.enabled
&& !key
->stencil
[0].enabled
&& !key
->stencil
[1].enabled
)
159 format_desc
= util_format_description(key
->zsbuf_format
);
163 * Depths are expected to be between 0 and 1, even if they are stored in
164 * floats. Setting these bits here will ensure that the lp_build_conv() call
165 * below won't try to unnecessarily clamp the incoming values.
167 if(src_type
.floating
) {
168 src_type
.sign
= FALSE
;
169 src_type
.norm
= TRUE
;
172 assert(!src_type
.sign
);
173 assert(src_type
.norm
);
176 /* Pick the depth type. */
177 dst_type
= lp_depth_type(format_desc
, src_type
.width
*src_type
.length
);
179 /* FIXME: Cope with a depth test type with a different bit width. */
180 assert(dst_type
.width
== src_type
.width
);
181 assert(dst_type
.length
== src_type
.length
);
183 /* Convert fragment Z from float to integer */
184 lp_build_conv(builder
, src_type
, dst_type
, &src
, 1, &src
, 1);
186 dst_ptr
= LLVMBuildBitCast(builder
,
188 LLVMPointerType(lp_build_vec_type(dst_type
), 0), "");
189 lp_build_depth_stencil_test(builder
,
204 * Generate the code to do inside/outside triangle testing for the
205 * four pixels in a 2x2 quad. This will set the four elements of the
206 * quad mask vector to 0 or ~0.
207 * \param i which quad of the quad group to test, in [0,3]
210 generate_tri_edge_mask(LLVMBuilderRef builder
,
212 LLVMValueRef
*mask
, /* ivec4, out */
213 LLVMValueRef c0
, /* int32 */
214 LLVMValueRef c1
, /* int32 */
215 LLVMValueRef c2
, /* int32 */
216 LLVMValueRef step0_ptr
, /* ivec4 */
217 LLVMValueRef step1_ptr
, /* ivec4 */
218 LLVMValueRef step2_ptr
) /* ivec4 */
220 #define OPTIMIZE_IN_OUT_TEST 0
221 #if OPTIMIZE_IN_OUT_TEST
222 struct lp_build_if_state ifctx
;
223 LLVMValueRef not_draw_all
;
225 struct lp_build_flow_context
*flow
;
226 struct lp_type i32_type
;
227 LLVMTypeRef i32vec4_type
;
228 LLVMValueRef c0_vec
, c1_vec
, c2_vec
;
229 LLVMValueRef in_out_mask
;
233 /* int32 vector type */
234 memset(&i32_type
, 0, sizeof i32_type
);
235 i32_type
.floating
= FALSE
; /* values are integers */
236 i32_type
.sign
= TRUE
; /* values are signed */
237 i32_type
.norm
= FALSE
; /* values are not normalized */
238 i32_type
.width
= 32; /* 32-bit int values */
239 i32_type
.length
= 4; /* 4 elements per vector */
241 i32vec4_type
= lp_build_int32_vec4_type();
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
,
405 LLVMValueRef counter
)
407 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
408 LLVMTypeRef vec_type
;
409 LLVMValueRef consts_ptr
;
410 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][NUM_CHANNELS
];
411 LLVMValueRef z
= interp
->pos
[2];
412 LLVMValueRef stencil_refs
[2];
413 struct lp_build_flow_context
*flow
;
414 struct lp_build_mask_context mask
;
415 boolean early_depth_stencil_test
;
422 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(builder
, context_ptr
);
423 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(builder
, context_ptr
);
425 vec_type
= lp_build_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
, facing
, counter
);
473 lp_build_tgsi_soa(builder
, tokens
, type
, &mask
,
474 consts_ptr
, interp
->pos
, interp
->inputs
,
475 outputs
, sampler
, &shader
->info
);
477 /* loop over fragment shader outputs/results */
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.
534 * \param rt the render target index (to index blend, colormask state)
535 * \param type the pixel color type
536 * \param context_ptr pointer to the runtime JIT context
537 * \param mask execution mask (active fragment/pixel mask)
538 * \param src colors from the fragment shader
539 * \param dst_ptr the destination color buffer pointer
542 generate_blend(const struct pipe_blend_state
*blend
,
544 LLVMBuilderRef builder
,
546 LLVMValueRef context_ptr
,
549 LLVMValueRef dst_ptr
)
551 struct lp_build_context bld
;
552 struct lp_build_flow_context
*flow
;
553 struct lp_build_mask_context mask_ctx
;
554 LLVMTypeRef vec_type
;
555 LLVMValueRef const_ptr
;
561 lp_build_context_init(&bld
, builder
, type
);
563 flow
= lp_build_flow_create(builder
);
565 /* we'll use this mask context to skip blending if all pixels are dead */
566 lp_build_mask_begin(&mask_ctx
, flow
, type
, mask
);
568 vec_type
= lp_build_vec_type(type
);
570 const_ptr
= lp_jit_context_blend_color(builder
, context_ptr
);
571 const_ptr
= LLVMBuildBitCast(builder
, const_ptr
,
572 LLVMPointerType(vec_type
, 0), "");
574 /* load constant blend color and colors from the dest color buffer */
575 for(chan
= 0; chan
< 4; ++chan
) {
576 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
577 con
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, const_ptr
, &index
, 1, ""), "");
579 dst
[chan
] = LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""), "");
581 lp_build_name(con
[chan
], "con.%c", "rgba"[chan
]);
582 lp_build_name(dst
[chan
], "dst.%c", "rgba"[chan
]);
586 lp_build_blend_soa(builder
, blend
, type
, rt
, src
, dst
, con
, res
);
588 /* store results to color buffer */
589 for(chan
= 0; chan
< 4; ++chan
) {
590 if(blend
->rt
[rt
].colormask
& (1 << chan
)) {
591 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), chan
, 0);
592 lp_build_name(res
[chan
], "res.%c", "rgba"[chan
]);
593 res
[chan
] = lp_build_select(&bld
, mask
, res
[chan
], dst
[chan
]);
594 LLVMBuildStore(builder
, res
[chan
], LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, ""));
598 lp_build_mask_end(&mask_ctx
);
599 lp_build_flow_destroy(flow
);
604 * Generate the runtime callable function for the whole fragment pipeline.
605 * Note that the function which we generate operates on a block of 16
606 * pixels at at time. The block contains 2x2 quads. Each quad contains
610 generate_fragment(struct llvmpipe_context
*lp
,
611 struct lp_fragment_shader
*shader
,
612 struct lp_fragment_shader_variant
*variant
,
613 unsigned do_tri_test
)
615 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
616 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
617 struct lp_type fs_type
;
618 struct lp_type blend_type
;
619 LLVMTypeRef fs_elem_type
;
620 LLVMTypeRef fs_int_vec_type
;
621 LLVMTypeRef blend_vec_type
;
622 LLVMTypeRef arg_types
[16];
623 LLVMTypeRef func_type
;
624 LLVMTypeRef int32_vec4_type
= lp_build_int32_vec4_type();
625 LLVMValueRef context_ptr
;
629 LLVMValueRef dadx_ptr
;
630 LLVMValueRef dady_ptr
;
631 LLVMValueRef color_ptr_ptr
;
632 LLVMValueRef depth_ptr
;
633 LLVMValueRef c0
, c1
, c2
, step0_ptr
, step1_ptr
, step2_ptr
, counter
= NULL
;
634 LLVMBasicBlockRef block
;
635 LLVMBuilderRef builder
;
638 struct lp_build_sampler_soa
*sampler
;
639 struct lp_build_interp_soa_context interp
;
640 LLVMValueRef fs_mask
[LP_MAX_VECTOR_LENGTH
];
641 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
][LP_MAX_VECTOR_LENGTH
];
642 LLVMValueRef blend_mask
;
643 LLVMValueRef function
;
651 /* TODO: actually pick these based on the fs and color buffer
652 * characteristics. */
654 memset(&fs_type
, 0, sizeof fs_type
);
655 fs_type
.floating
= TRUE
; /* floating point values */
656 fs_type
.sign
= TRUE
; /* values are signed */
657 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
658 fs_type
.width
= 32; /* 32-bit float */
659 fs_type
.length
= 4; /* 4 elements per vector */
660 num_fs
= 4; /* number of quads per block */
662 memset(&blend_type
, 0, sizeof blend_type
);
663 blend_type
.floating
= FALSE
; /* values are integers */
664 blend_type
.sign
= FALSE
; /* values are unsigned */
665 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
666 blend_type
.width
= 8; /* 8-bit ubyte values */
667 blend_type
.length
= 16; /* 16 elements per vector */
670 * Generate the function prototype. Any change here must be reflected in
671 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
674 fs_elem_type
= lp_build_elem_type(fs_type
);
675 fs_int_vec_type
= lp_build_int_vec_type(fs_type
);
677 blend_vec_type
= lp_build_vec_type(blend_type
);
679 arg_types
[0] = screen
->context_ptr_type
; /* context */
680 arg_types
[1] = LLVMInt32Type(); /* x */
681 arg_types
[2] = LLVMInt32Type(); /* y */
682 arg_types
[3] = LLVMFloatType(); /* facing */
683 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
684 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
685 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
686 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
687 arg_types
[8] = LLVMPointerType(fs_int_vec_type
, 0); /* depth */
688 arg_types
[9] = LLVMInt32Type(); /* c0 */
689 arg_types
[10] = LLVMInt32Type(); /* c1 */
690 arg_types
[11] = LLVMInt32Type(); /* c2 */
691 /* Note: the step arrays are built as int32[16] but we interpret
692 * them here as int32_vec4[4].
694 arg_types
[12] = LLVMPointerType(int32_vec4_type
, 0);/* step0 */
695 arg_types
[13] = LLVMPointerType(int32_vec4_type
, 0);/* step1 */
696 arg_types
[14] = LLVMPointerType(int32_vec4_type
, 0);/* step2 */
697 arg_types
[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
699 func_type
= LLVMFunctionType(LLVMVoidType(), arg_types
, Elements(arg_types
), 0);
701 function
= LLVMAddFunction(screen
->module
, "shader", func_type
);
702 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
704 variant
->function
[do_tri_test
] = function
;
707 /* XXX: need to propagate noalias down into color param now we are
708 * passing a pointer-to-pointer?
710 for(i
= 0; i
< Elements(arg_types
); ++i
)
711 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
712 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
714 context_ptr
= LLVMGetParam(function
, 0);
715 x
= LLVMGetParam(function
, 1);
716 y
= LLVMGetParam(function
, 2);
717 facing
= LLVMGetParam(function
, 3);
718 a0_ptr
= LLVMGetParam(function
, 4);
719 dadx_ptr
= LLVMGetParam(function
, 5);
720 dady_ptr
= LLVMGetParam(function
, 6);
721 color_ptr_ptr
= LLVMGetParam(function
, 7);
722 depth_ptr
= LLVMGetParam(function
, 8);
723 c0
= LLVMGetParam(function
, 9);
724 c1
= LLVMGetParam(function
, 10);
725 c2
= LLVMGetParam(function
, 11);
726 step0_ptr
= LLVMGetParam(function
, 12);
727 step1_ptr
= LLVMGetParam(function
, 13);
728 step2_ptr
= LLVMGetParam(function
, 14);
730 lp_build_name(context_ptr
, "context");
731 lp_build_name(x
, "x");
732 lp_build_name(y
, "y");
733 lp_build_name(a0_ptr
, "a0");
734 lp_build_name(dadx_ptr
, "dadx");
735 lp_build_name(dady_ptr
, "dady");
736 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
737 lp_build_name(depth_ptr
, "depth");
738 lp_build_name(c0
, "c0");
739 lp_build_name(c1
, "c1");
740 lp_build_name(c2
, "c2");
741 lp_build_name(step0_ptr
, "step0");
742 lp_build_name(step1_ptr
, "step1");
743 lp_build_name(step2_ptr
, "step2");
745 if (key
->occlusion_count
) {
746 counter
= LLVMGetParam(function
, 15);
747 lp_build_name(counter
, "counter");
754 block
= LLVMAppendBasicBlock(function
, "entry");
755 builder
= LLVMCreateBuilder();
756 LLVMPositionBuilderAtEnd(builder
, block
);
758 generate_pos0(builder
, x
, y
, &x0
, &y0
);
760 lp_build_interp_soa_init(&interp
,
764 a0_ptr
, dadx_ptr
, dady_ptr
,
767 /* code generated texture sampling */
768 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
770 /* loop over quads in the block */
771 for(i
= 0; i
< num_fs
; ++i
) {
772 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
773 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][NUM_CHANNELS
];
774 LLVMValueRef depth_ptr_i
;
777 lp_build_interp_soa_update(&interp
, i
);
779 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &index
, 1, "");
781 generate_fs(lp
, shader
, key
,
788 &fs_mask
[i
], /* output */
794 step0_ptr
, step1_ptr
, step2_ptr
, counter
);
796 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
797 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
)
798 fs_out_color
[cbuf
][chan
][i
] = out_color
[cbuf
][chan
];
801 sampler
->destroy(sampler
);
803 /* Loop over color outputs / color buffers to do blending.
805 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
806 LLVMValueRef color_ptr
;
807 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), cbuf
, 0);
808 LLVMValueRef blend_in_color
[NUM_CHANNELS
];
812 * Convert the fs's output color and mask to fit to the blending type.
814 for(chan
= 0; chan
< NUM_CHANNELS
; ++chan
) {
815 lp_build_conv(builder
, fs_type
, blend_type
,
816 fs_out_color
[cbuf
][chan
], num_fs
,
817 &blend_in_color
[chan
], 1);
818 lp_build_name(blend_in_color
[chan
], "color%d.%c", cbuf
, "rgba"[chan
]);
821 lp_build_conv_mask(builder
, fs_type
, blend_type
,
825 color_ptr
= LLVMBuildLoad(builder
,
826 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
828 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
830 /* which blend/colormask state to use */
831 rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
836 generate_blend(&key
->blend
,
846 LLVMBuildRetVoid(builder
);
848 LLVMDisposeBuilder(builder
);
851 /* Verify the LLVM IR. If invalid, dump and abort */
853 if(LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
855 lp_debug_dump_value(function
);
860 /* Apply optimizations to LLVM IR */
862 LLVMRunFunctionPassManager(screen
->pass
, function
);
864 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
865 /* Print the LLVM IR to stderr */
866 lp_debug_dump_value(function
);
871 * Translate the LLVM IR into machine code.
874 void *f
= LLVMGetPointerToGlobal(screen
->engine
, function
);
876 variant
->jit_function
[do_tri_test
] = cast_voidptr_to_lp_jit_frag_func(f
);
878 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
885 static struct lp_fragment_shader_variant
*
886 generate_variant(struct llvmpipe_context
*lp
,
887 struct lp_fragment_shader
*shader
,
888 const struct lp_fragment_shader_variant_key
*key
)
890 struct lp_fragment_shader_variant
*variant
;
892 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
895 tgsi_dump(shader
->base
.tokens
, 0);
896 if(key
->depth
.enabled
) {
897 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
898 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
899 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
901 for (i
= 0; i
< 2; ++i
) {
902 if(key
->stencil
[i
].enabled
) {
903 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
904 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
905 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
906 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
907 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
908 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
911 if(key
->alpha
.enabled
) {
912 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
913 debug_printf("alpha.ref_value = %f\n", key
->alpha
.ref_value
);
915 if(key
->blend
.logicop_enable
) {
916 debug_printf("blend.logicop_func = %u\n", key
->blend
.logicop_func
);
918 else if(key
->blend
.rt
[0].blend_enable
) {
919 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
920 debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
921 debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
922 debug_printf("alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
923 debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
924 debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
926 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
927 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
) {
928 if(key
->sampler
[i
].format
) {
929 debug_printf("sampler[%u] = \n", i
);
930 debug_printf(" .format = %s\n",
931 util_format_name(key
->sampler
[i
].format
));
932 debug_printf(" .target = %s\n",
933 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
934 debug_printf(" .pot = %u %u %u\n",
935 key
->sampler
[i
].pot_width
,
936 key
->sampler
[i
].pot_height
,
937 key
->sampler
[i
].pot_depth
);
938 debug_printf(" .wrap = %s %s %s\n",
939 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
940 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
941 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
942 debug_printf(" .min_img_filter = %s\n",
943 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
944 debug_printf(" .min_mip_filter = %s\n",
945 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
946 debug_printf(" .mag_img_filter = %s\n",
947 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
948 if(key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
949 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
950 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
955 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
959 memcpy(&variant
->key
, key
, sizeof *key
);
961 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
962 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
964 /* TODO: most of these can be relaxed, in particular the colormask */
966 !key
->blend
.logicop_enable
&&
967 !key
->blend
.rt
[0].blend_enable
&&
968 key
->blend
.rt
[0].colormask
== 0xf &&
969 !key
->stencil
[0].enabled
&&
970 !key
->alpha
.enabled
&&
971 !key
->depth
.enabled
&&
973 !shader
->info
.uses_kill
976 /* insert new variant into linked list */
977 variant
->next
= shader
->variants
;
978 shader
->variants
= variant
;
985 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
986 const struct pipe_shader_state
*templ
)
988 struct lp_fragment_shader
*shader
;
990 shader
= CALLOC_STRUCT(lp_fragment_shader
);
994 /* get/save the summary info for this shader */
995 tgsi_scan_shader(templ
->tokens
, &shader
->info
);
997 /* we need to keep a local copy of the tokens */
998 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
1000 if (gallivm_debug
& GALLIVM_DEBUG_TGSI
) {
1001 debug_printf("llvmpipe: Create fragment shader %p:\n", (void *) shader
);
1002 tgsi_dump(templ
->tokens
, 0);
1010 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
1012 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1014 if (llvmpipe
->fs
== fs
)
1017 draw_flush(llvmpipe
->draw
);
1021 llvmpipe
->dirty
|= LP_NEW_FS
;
1026 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
1028 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1029 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
1030 struct lp_fragment_shader
*shader
= fs
;
1031 struct lp_fragment_shader_variant
*variant
;
1033 assert(fs
!= llvmpipe
->fs
);
1037 * XXX: we need to flush the context until we have some sort of reference
1038 * counting in fragment shaders as they may still be binned
1040 draw_flush(llvmpipe
->draw
);
1041 lp_setup_flush(llvmpipe
->setup
, 0);
1043 variant
= shader
->variants
;
1045 struct lp_fragment_shader_variant
*next
= variant
->next
;
1048 for (i
= 0; i
< Elements(variant
->function
); i
++) {
1049 if (variant
->function
[i
]) {
1050 if (variant
->jit_function
[i
])
1051 LLVMFreeMachineCodeForFunction(screen
->engine
,
1052 variant
->function
[i
]);
1053 LLVMDeleteFunction(variant
->function
[i
]);
1062 FREE((void *) shader
->base
.tokens
);
1069 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
1070 uint shader
, uint index
,
1071 struct pipe_resource
*constants
)
1073 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1074 unsigned size
= constants
? constants
->width0
: 0;
1075 const void *data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
1077 assert(shader
< PIPE_SHADER_TYPES
);
1080 if(llvmpipe
->constants
[shader
] == constants
)
1083 draw_flush(llvmpipe
->draw
);
1085 /* note: reference counting */
1086 pipe_resource_reference(&llvmpipe
->constants
[shader
], constants
);
1088 if(shader
== PIPE_SHADER_VERTEX
) {
1089 draw_set_mapped_constant_buffer(llvmpipe
->draw
, PIPE_SHADER_VERTEX
, 0,
1093 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
1098 * Return the blend factor equivalent to a destination alpha of one.
1100 static INLINE
unsigned
1101 force_dst_alpha_one(unsigned factor
, boolean alpha
)
1104 case PIPE_BLENDFACTOR_DST_ALPHA
:
1105 return PIPE_BLENDFACTOR_ONE
;
1106 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
1107 return PIPE_BLENDFACTOR_ZERO
;
1108 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
1109 return PIPE_BLENDFACTOR_ZERO
;
1114 case PIPE_BLENDFACTOR_DST_COLOR
:
1115 return PIPE_BLENDFACTOR_ONE
;
1116 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
1117 return PIPE_BLENDFACTOR_ZERO
;
1126 * We need to generate several variants of the fragment pipeline to match
1127 * all the combinations of the contributing state atoms.
1129 * TODO: there is actually no reason to tie this to context state -- the
1130 * generated code could be cached globally in the screen.
1133 make_variant_key(struct llvmpipe_context
*lp
,
1134 struct lp_fragment_shader
*shader
,
1135 struct lp_fragment_shader_variant_key
*key
)
1139 memset(key
, 0, sizeof *key
);
1141 if (lp
->framebuffer
.zsbuf
) {
1142 if (lp
->depth_stencil
->depth
.enabled
) {
1143 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1144 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
1146 if (lp
->depth_stencil
->stencil
[0].enabled
) {
1147 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
1148 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
1152 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
1153 if(key
->alpha
.enabled
)
1154 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
1155 /* alpha.ref_value is passed in jit_context */
1157 key
->flatshade
= lp
->rasterizer
->flatshade
;
1158 key
->scissor
= lp
->rasterizer
->scissor
;
1159 if (lp
->active_query_count
) {
1160 key
->occlusion_count
= TRUE
;
1163 if (lp
->framebuffer
.nr_cbufs
) {
1164 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
1167 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
1168 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
1169 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
1170 const struct util_format_description
*format_desc
;
1173 format_desc
= util_format_description(lp
->framebuffer
.cbufs
[i
]->format
);
1174 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
1175 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
1177 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
1179 /* mask out color channels not present in the color buffer.
1180 * Should be simple to incorporate per-cbuf writemasks:
1182 for(chan
= 0; chan
< 4; ++chan
) {
1183 enum util_format_swizzle swizzle
= format_desc
->swizzle
[chan
];
1185 if(swizzle
> UTIL_FORMAT_SWIZZLE_W
)
1186 blend_rt
->colormask
&= ~(1 << chan
);
1190 * Our swizzled render tiles always have an alpha channel, but the linear
1191 * render target format often does not, so force here the dst alpha to be
1194 * This is not a mere optimization. Wrong results will be produced if the
1195 * dst alpha is used, the dst format does not have alpha, and the previous
1196 * rendering was not flushed from the swizzled to linear buffer. For
1197 * example, NonPowTwo DCT.
1199 * TODO: This should be generalized to all channels for better
1200 * performance, but only alpha causes correctness issues.
1202 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
) {
1203 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
, FALSE
);
1204 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
, FALSE
);
1205 blend_rt
->alpha_src_factor
= force_dst_alpha_one(blend_rt
->alpha_src_factor
, TRUE
);
1206 blend_rt
->alpha_dst_factor
= force_dst_alpha_one(blend_rt
->alpha_dst_factor
, TRUE
);
1210 for(i
= 0; i
< PIPE_MAX_SAMPLERS
; ++i
)
1211 if(shader
->info
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
))
1212 lp_sampler_static_state(&key
->sampler
[i
], lp
->fragment_sampler_views
[i
], lp
->sampler
[i
]);
1217 * Update fragment state. This is called just prior to drawing
1218 * something when some fragment-related state has changed.
1221 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
1223 struct lp_fragment_shader
*shader
= lp
->fs
;
1224 struct lp_fragment_shader_variant_key key
;
1225 struct lp_fragment_shader_variant
*variant
;
1227 make_variant_key(lp
, shader
, &key
);
1229 variant
= shader
->variants
;
1231 if(memcmp(&variant
->key
, &key
, sizeof key
) == 0)
1234 variant
= variant
->next
;
1242 variant
= generate_variant(lp
, shader
, &key
);
1246 LP_COUNT_ADD(llvm_compile_time
, dt
);
1247 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
1250 lp_setup_set_fs_functions(lp
->setup
,
1251 variant
->jit_function
[RAST_WHOLE
],
1252 variant
->jit_function
[RAST_EDGE_TEST
],
1259 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
1261 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
1262 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
1263 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
1265 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;