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:
37 * - depth/stencil test
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
57 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/u_simple_list.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_init.h"
79 #include "gallivm/lp_bld_intr.h"
80 #include "gallivm/lp_bld_logic.h"
81 #include "gallivm/lp_bld_tgsi.h"
82 #include "gallivm/lp_bld_swizzle.h"
83 #include "gallivm/lp_bld_flow.h"
84 #include "gallivm/lp_bld_debug.h"
85 #include "gallivm/lp_bld_arit.h"
86 #include "gallivm/lp_bld_pack.h"
87 #include "gallivm/lp_bld_format.h"
88 #include "gallivm/lp_bld_quad.h"
90 #include "lp_bld_alpha.h"
91 #include "lp_bld_blend.h"
92 #include "lp_bld_depth.h"
93 #include "lp_bld_interp.h"
94 #include "lp_context.h"
99 #include "lp_tex_sample.h"
100 #include "lp_flush.h"
101 #include "lp_state_fs.h"
104 /** Fragment shader number (for debugging) */
105 static unsigned fs_no
= 0;
109 * Expand the relevant bits of mask_input to a n*4-dword mask for the
110 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
111 * quad mask vector to 0 or ~0.
112 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
113 * quad arguments with fs length 8.
115 * \param first_quad which quad(s) of the quad group to test, in [0,3]
116 * \param mask_input bitwise mask for the whole 4x4 stamp
119 generate_quad_mask(struct gallivm_state
*gallivm
,
120 struct lp_type fs_type
,
122 LLVMValueRef mask_input
) /* int32 */
124 LLVMBuilderRef builder
= gallivm
->builder
;
125 struct lp_type mask_type
;
126 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
127 LLVMValueRef bits
[16];
132 * XXX: We'll need a different path for 16 x u8
134 assert(fs_type
.width
== 32);
135 assert(fs_type
.length
<= Elements(bits
));
136 mask_type
= lp_int_type(fs_type
);
139 * mask_input >>= (quad * 4)
141 switch (first_quad
) {
146 assert(fs_type
.length
== 4);
153 assert(fs_type
.length
== 4);
161 mask_input
= LLVMBuildLShr(builder
,
163 LLVMConstInt(i32t
, shift
, 0),
167 * mask = { mask_input & (1 << i), for i in [0,3] }
169 mask
= lp_build_broadcast(gallivm
,
170 lp_build_vec_type(gallivm
, mask_type
),
173 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
174 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
175 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
176 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
177 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
178 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
180 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
183 * mask = mask != 0 ? ~0 : 0
185 mask
= lp_build_compare(gallivm
,
186 mask_type
, PIPE_FUNC_NOTEQUAL
,
188 lp_build_const_int_vec(gallivm
, mask_type
, 0));
194 #define EARLY_DEPTH_TEST 0x1
195 #define LATE_DEPTH_TEST 0x2
196 #define EARLY_DEPTH_WRITE 0x4
197 #define LATE_DEPTH_WRITE 0x8
200 find_output_by_semantic( const struct tgsi_shader_info
*info
,
206 for (i
= 0; i
< info
->num_outputs
; i
++)
207 if (info
->output_semantic_name
[i
] == semantic
&&
208 info
->output_semantic_index
[i
] == index
)
216 * Generate the fragment shader, depth/stencil test, and alpha tests.
217 * \param i which quad in the tile, in range [0,3]
218 * \param partial_mask if 1, do mask_input testing
221 generate_fs(struct gallivm_state
*gallivm
,
222 struct lp_fragment_shader
*shader
,
223 const struct lp_fragment_shader_variant_key
*key
,
224 LLVMBuilderRef builder
,
226 LLVMValueRef context_ptr
,
228 struct lp_build_interp_soa_context
*interp
,
229 struct lp_build_sampler_soa
*sampler
,
231 LLVMValueRef (*color
)[4],
232 LLVMValueRef depth_ptr
,
234 unsigned partial_mask
,
235 LLVMValueRef mask_input
,
236 LLVMValueRef counter
)
238 const struct util_format_description
*zs_format_desc
= NULL
;
239 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
240 LLVMTypeRef vec_type
;
241 LLVMValueRef consts_ptr
;
242 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
244 LLVMValueRef zs_value
= NULL
;
245 LLVMValueRef stencil_refs
[2];
246 struct lp_build_mask_context mask
;
247 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
248 shader
->info
.base
.num_inputs
< 3 &&
249 shader
->info
.base
.num_instructions
< 8);
254 struct lp_bld_tgsi_system_values system_values
;
256 memset(&system_values
, 0, sizeof(system_values
));
258 if (key
->depth
.enabled
||
259 key
->stencil
[0].enabled
||
260 key
->stencil
[1].enabled
) {
262 zs_format_desc
= util_format_description(key
->zsbuf_format
);
263 assert(zs_format_desc
);
265 if (!shader
->info
.base
.writes_z
) {
266 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
267 /* With alpha test and kill, can do the depth test early
268 * and hopefully eliminate some quads. But need to do a
269 * special deferred depth write once the final mask value
272 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
274 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
277 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
280 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
281 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
282 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
290 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
291 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
293 vec_type
= lp_build_vec_type(gallivm
, type
);
295 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
297 memset(outputs
, 0, sizeof outputs
);
299 /* Declare the color and z variables */
300 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
301 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
302 color
[cbuf
][chan
] = lp_build_alloca(gallivm
, vec_type
, "color");
306 /* do triangle edge testing */
308 *pmask
= generate_quad_mask(gallivm
, type
,
309 i
*type
.length
/4, mask_input
);
312 *pmask
= lp_build_const_int_vec(gallivm
, type
, ~0);
315 /* 'mask' will control execution based on quad's pixel alive/killed state */
316 lp_build_mask_begin(&mask
, gallivm
, type
, *pmask
);
318 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
319 lp_build_mask_check(&mask
);
321 lp_build_interp_soa_update_pos(interp
, gallivm
, i
*type
.length
/4);
324 if (depth_mode
& EARLY_DEPTH_TEST
) {
325 lp_build_depth_stencil_test(gallivm
,
337 if (depth_mode
& EARLY_DEPTH_WRITE
) {
338 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
342 lp_build_interp_soa_update_inputs(interp
, gallivm
, i
*type
.length
/4);
344 /* Build the actual shader */
345 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
346 consts_ptr
, &system_values
,
347 interp
->pos
, interp
->inputs
,
348 outputs
, sampler
, &shader
->info
.base
);
351 if (key
->alpha
.enabled
) {
352 int color0
= find_output_by_semantic(&shader
->info
.base
,
356 if (color0
!= -1 && outputs
[color0
][3]) {
357 const struct util_format_description
*cbuf_format_desc
;
358 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
359 LLVMValueRef alpha_ref_value
;
361 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
362 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
364 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
366 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
367 &mask
, alpha
, alpha_ref_value
,
368 (depth_mode
& LATE_DEPTH_TEST
) != 0);
373 if (depth_mode
& LATE_DEPTH_TEST
) {
374 int pos0
= find_output_by_semantic(&shader
->info
.base
,
375 TGSI_SEMANTIC_POSITION
,
378 if (pos0
!= -1 && outputs
[pos0
][2]) {
379 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
382 lp_build_depth_stencil_test(gallivm
,
394 if (depth_mode
& LATE_DEPTH_WRITE
) {
395 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr
, zs_value
);
398 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
399 (depth_mode
& LATE_DEPTH_WRITE
))
401 /* Need to apply a reduced mask to the depth write. Reload the
402 * depth value, update from zs_value with the new mask value and
405 lp_build_deferred_depth_write(gallivm
,
415 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
417 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
&&
418 shader
->info
.base
.output_semantic_index
[attrib
] < key
->nr_cbufs
)
420 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
421 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
422 if(outputs
[attrib
][chan
]) {
423 /* XXX: just initialize outputs to point at colors[] and
426 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
427 lp_build_name(out
, "color%u.%u.%c", i
, attrib
, "rgba"[chan
]);
428 LLVMBuildStore(builder
, out
, color
[cbuf
][chan
]);
435 lp_build_occlusion_count(gallivm
, type
,
436 lp_build_mask_value(&mask
), counter
);
438 *pmask
= lp_build_mask_end(&mask
);
443 * Generate the fragment shader, depth/stencil test, and alpha tests.
446 generate_fs_loop(struct gallivm_state
*gallivm
,
447 struct lp_fragment_shader
*shader
,
448 const struct lp_fragment_shader_variant_key
*key
,
449 LLVMBuilderRef builder
,
451 LLVMValueRef context_ptr
,
452 LLVMValueRef num_loop
,
453 struct lp_build_interp_soa_context
*interp
,
454 struct lp_build_sampler_soa
*sampler
,
455 LLVMValueRef mask_store
,
456 LLVMValueRef (*out_color
)[4],
457 LLVMValueRef depth_ptr
,
460 LLVMValueRef counter
)
462 const struct util_format_description
*zs_format_desc
= NULL
;
463 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
464 LLVMTypeRef vec_type
;
465 LLVMValueRef mask_ptr
, mask_val
;
466 LLVMValueRef consts_ptr
;
468 LLVMValueRef zs_value
= NULL
;
469 LLVMValueRef stencil_refs
[2];
470 LLVMValueRef depth_ptr_i
;
471 LLVMValueRef depth_offset
;
472 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
473 struct lp_build_for_loop_state loop_state
;
474 struct lp_build_mask_context mask
;
475 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
476 shader
->info
.base
.num_inputs
< 3 &&
477 shader
->info
.base
.num_instructions
< 8);
483 struct lp_bld_tgsi_system_values system_values
;
485 memset(&system_values
, 0, sizeof(system_values
));
487 if (key
->depth
.enabled
||
488 key
->stencil
[0].enabled
||
489 key
->stencil
[1].enabled
) {
491 zs_format_desc
= util_format_description(key
->zsbuf_format
);
492 assert(zs_format_desc
);
494 if (!shader
->info
.base
.writes_z
) {
495 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
496 /* With alpha test and kill, can do the depth test early
497 * and hopefully eliminate some quads. But need to do a
498 * special deferred depth write once the final mask value
501 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
503 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
506 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
509 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
510 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
511 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
518 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
519 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
521 vec_type
= lp_build_vec_type(gallivm
, type
);
523 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
525 lp_build_for_loop_begin(&loop_state
, gallivm
,
526 lp_build_const_int32(gallivm
, 0),
529 lp_build_const_int32(gallivm
, 1));
531 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
532 &loop_state
.counter
, 1, "mask_ptr");
533 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
535 depth_offset
= LLVMBuildMul(builder
, loop_state
.counter
,
536 lp_build_const_int32(gallivm
, depth_bits
* type
.length
),
539 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
541 memset(outputs
, 0, sizeof outputs
);
543 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
544 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
545 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
546 lp_build_vec_type(gallivm
,
554 /* 'mask' will control execution based on quad's pixel alive/killed state */
555 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
557 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
558 lp_build_mask_check(&mask
);
560 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
563 if (depth_mode
& EARLY_DEPTH_TEST
) {
564 lp_build_depth_stencil_test(gallivm
,
576 if (depth_mode
& EARLY_DEPTH_WRITE
) {
577 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr_i
, zs_value
);
581 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
583 /* Build the actual shader */
584 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
585 consts_ptr
, &system_values
,
586 interp
->pos
, interp
->inputs
,
587 outputs
, sampler
, &shader
->info
.base
);
590 if (key
->alpha
.enabled
) {
591 int color0
= find_output_by_semantic(&shader
->info
.base
,
595 if (color0
!= -1 && outputs
[color0
][3]) {
596 const struct util_format_description
*cbuf_format_desc
;
597 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
598 LLVMValueRef alpha_ref_value
;
600 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
601 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
603 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
605 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
606 &mask
, alpha
, alpha_ref_value
,
607 (depth_mode
& LATE_DEPTH_TEST
) != 0);
612 if (depth_mode
& LATE_DEPTH_TEST
) {
613 int pos0
= find_output_by_semantic(&shader
->info
.base
,
614 TGSI_SEMANTIC_POSITION
,
617 if (pos0
!= -1 && outputs
[pos0
][2]) {
618 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
621 lp_build_depth_stencil_test(gallivm
,
633 if (depth_mode
& LATE_DEPTH_WRITE
) {
634 lp_build_depth_write(builder
, zs_format_desc
, depth_ptr_i
, zs_value
);
637 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
638 (depth_mode
& LATE_DEPTH_WRITE
))
640 /* Need to apply a reduced mask to the depth write. Reload the
641 * depth value, update from zs_value with the new mask value and
644 lp_build_deferred_depth_write(gallivm
,
654 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
656 if (shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
&&
657 shader
->info
.base
.output_semantic_index
[attrib
] < key
->nr_cbufs
)
659 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
660 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
661 if(outputs
[attrib
][chan
]) {
662 /* XXX: just initialize outputs to point at colors[] and
665 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
666 LLVMValueRef color_ptr
;
667 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
668 &loop_state
.counter
, 1, "");
669 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
670 LLVMBuildStore(builder
, out
, color_ptr
);
676 if (key
->occlusion_count
) {
677 lp_build_name(counter
, "counter");
678 lp_build_occlusion_count(gallivm
, type
,
679 lp_build_mask_value(&mask
), counter
);
682 mask_val
= lp_build_mask_end(&mask
);
683 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
684 lp_build_for_loop_end(&loop_state
);
689 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
691 * Fragment Shader outputs pixels in small 2x2 blocks
692 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
694 * However in memory pixels are stored in rows
695 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
697 * @param type fragment shader type (4x or 8x float)
698 * @param num_fs number of fs_src
699 * @param dst_channels number of output channels
700 * @param fs_src output from fragment shader
701 * @param dst pointer to store result
702 * @param pad_inline is channel padding inline or at end of row
703 * @return the number of dsts
706 generate_fs_twiddle(struct gallivm_state
*gallivm
,
709 unsigned dst_channels
,
710 LLVMValueRef fs_src
[][4],
714 LLVMValueRef src
[16];
720 unsigned pixels
= num_fs
== 4 ? 1 : 2;
721 unsigned reorder_group
;
722 unsigned src_channels
;
726 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
727 src_count
= num_fs
* src_channels
;
729 assert(pixels
== 2 || num_fs
== 4);
730 assert(num_fs
* src_channels
<= Elements(src
));
733 * Transpose from SoA -> AoS
735 for (i
= 0; i
< num_fs
; ++i
) {
736 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
740 * Pick transformation options
747 if (dst_channels
== 1) {
753 } else if (dst_channels
== 2) {
757 } else if (dst_channels
> 2) {
764 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
770 * Split the src in half
773 for (i
= num_fs
; i
> 0; --i
) {
774 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
775 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
783 * Ensure pixels are in memory order
786 /* Twiddle pixels by reordering the array, e.g.:
788 * src_count = 8 -> 0 2 1 3 4 6 5 7
789 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
791 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
793 for (i
= 0; i
< src_count
; ++i
) {
794 unsigned group
= i
/ reorder_group
;
795 unsigned block
= (group
/ 4) * 4 * reorder_group
;
796 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
799 } else if (twiddle
) {
800 /* Twiddle pixels across elements of array */
801 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
804 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
808 * Moves any padding between pixels to the end
809 * e.g. RGBXRGBX -> RGBRGBXX
812 unsigned char swizzles
[16];
813 unsigned elems
= pixels
* dst_channels
;
815 for (i
= 0; i
< type
.length
; ++i
) {
817 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
819 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
822 for (i
= 0; i
< src_count
; ++i
) {
823 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
832 * Load an unswizzled block of pixels from memory
835 load_unswizzled_block(struct gallivm_state
*gallivm
,
836 LLVMValueRef base_ptr
,
838 unsigned block_width
,
839 unsigned block_height
,
841 struct lp_type dst_type
,
843 unsigned dst_alignment
)
845 LLVMBuilderRef builder
= gallivm
->builder
;
846 unsigned row_size
= dst_count
/ block_height
;
849 /* Ensure block exactly fits into dst */
850 assert((block_width
* block_height
) % dst_count
== 0);
852 for (i
= 0; i
< dst_count
; ++i
) {
853 unsigned x
= i
% row_size
;
854 unsigned y
= i
/ row_size
;
856 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
857 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
860 LLVMValueRef dst_ptr
;
862 gep
[0] = lp_build_const_int32(gallivm
, 0);
863 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
865 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
866 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
868 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
870 lp_set_load_alignment(dst
[i
], dst_alignment
);
876 * Store an unswizzled block of pixels to memory
879 store_unswizzled_block(struct gallivm_state
*gallivm
,
880 LLVMValueRef base_ptr
,
882 unsigned block_width
,
883 unsigned block_height
,
885 struct lp_type src_type
,
887 unsigned src_alignment
)
889 LLVMBuilderRef builder
= gallivm
->builder
;
890 unsigned row_size
= src_count
/ block_height
;
893 /* Ensure src exactly fits into block */
894 assert((block_width
* block_height
) % src_count
== 0);
896 for (i
= 0; i
< src_count
; ++i
) {
897 unsigned x
= i
% row_size
;
898 unsigned y
= i
/ row_size
;
900 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
901 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
904 LLVMValueRef src_ptr
;
906 gep
[0] = lp_build_const_int32(gallivm
, 0);
907 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
909 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
910 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
912 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
914 lp_set_store_alignment(src_ptr
, src_alignment
);
920 * Checks if a format description is an arithmetic format
922 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
924 static INLINE boolean
925 is_arithmetic_format(const struct util_format_description
*format_desc
)
927 boolean arith
= false;
930 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
931 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
932 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
940 * Retrieves the type representing the memory layout for a format
942 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
945 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
946 struct lp_type
* type
)
950 memset(type
, 0, sizeof(struct lp_type
));
951 type
->floating
= format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
;
952 type
->fixed
= format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_FIXED
;
953 type
->sign
= format_desc
->channel
[0].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
954 type
->norm
= format_desc
->channel
[0].normalized
;
956 if (is_arithmetic_format(format_desc
)) {
960 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
961 type
->width
+= format_desc
->channel
[i
].size
;
964 type
->width
= format_desc
->channel
[0].size
;
965 type
->length
= format_desc
->nr_channels
;
971 * Retrieves the type for a format which is usable in the blending code.
973 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
976 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
977 struct lp_type
* type
)
981 memset(type
, 0, sizeof(struct lp_type
));
982 type
->floating
= format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
;
983 type
->fixed
= format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_FIXED
;
984 type
->sign
= format_desc
->channel
[0].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
985 type
->norm
= format_desc
->channel
[0].normalized
;
986 type
->width
= format_desc
->channel
[0].size
;
987 type
->length
= format_desc
->nr_channels
;
989 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
990 if (format_desc
->channel
[i
].size
> type
->width
)
991 type
->width
= format_desc
->channel
[i
].size
;
994 if (type
->floating
) {
997 if (type
->width
<= 8) {
999 } else if (type
->width
<= 16) {
1006 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1013 * Scale a normalised value from src_bits to dst_bits
1015 static INLINE LLVMValueRef
1016 scale_bits(struct gallivm_state
*gallivm
,
1020 struct lp_type src_type
)
1022 LLVMBuilderRef builder
= gallivm
->builder
;
1023 LLVMValueRef result
= src
;
1025 if (dst_bits
< src_bits
) {
1026 /* Scale down by LShr */
1027 result
= LLVMBuildLShr(builder
,
1029 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- dst_bits
),
1031 } else if (dst_bits
> src_bits
) {
1033 int db
= dst_bits
- src_bits
;
1035 /* Shift left by difference in bits */
1036 result
= LLVMBuildShl(builder
,
1038 lp_build_const_int_vec(gallivm
, src_type
, db
),
1041 if (db
< src_bits
) {
1042 /* Enough bits in src to fill the remainder */
1043 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1045 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1048 result
= LLVMBuildOr(builder
, result
, lower
, "");
1049 } else if (db
> src_bits
) {
1050 /* Need to repeatedely copy src bits to fill remainder in dst */
1053 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1054 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1056 result
= LLVMBuildOr(builder
,
1058 LLVMBuildLShr(builder
, result
, shuv
, ""),
1069 * Convert from memory format to blending format
1071 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1074 convert_to_blend_type(struct gallivm_state
*gallivm
,
1075 const struct util_format_description
*src_fmt
,
1076 struct lp_type src_type
,
1077 struct lp_type dst_type
,
1078 LLVMValueRef
* src
, // and dst
1081 LLVMValueRef
*dst
= src
;
1082 LLVMBuilderRef builder
= gallivm
->builder
;
1083 struct lp_type blend_type
;
1084 struct lp_type mem_type
;
1086 unsigned pixels
= 16 / num_srcs
;
1089 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1090 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1092 /* Is the format arithmetic */
1093 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1094 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1096 /* Pad if necessary */
1097 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1098 for (i
= 0; i
< num_srcs
; ++i
) {
1099 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1102 src_type
.length
= dst_type
.length
;
1105 /* Special case for half-floats */
1106 if (mem_type
.width
== 16 && mem_type
.floating
) {
1107 assert(blend_type
.width
== 32 && blend_type
.floating
);
1108 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1116 src_type
.width
= blend_type
.width
* blend_type
.length
;
1117 blend_type
.length
*= pixels
;
1118 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1120 for (i
= 0; i
< num_srcs
; ++i
) {
1121 LLVMValueRef chans
[4];
1125 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1127 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1130 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1134 /* Extract bits from source */
1135 chans
[j
] = LLVMBuildLShr(builder
,
1137 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1140 chans
[j
] = LLVMBuildAnd(builder
,
1142 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1146 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
, blend_type
.width
, chans
[j
], src_type
);
1148 /* Insert bits into correct position */
1149 chans
[j
] = LLVMBuildShl(builder
,
1151 lp_build_const_int_vec(gallivm
, src_type
, j
* blend_type
.width
),
1154 sa
+= src_fmt
->channel
[j
].size
;
1159 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1163 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1169 * Convert from blending format to memory format
1171 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1174 convert_from_blend_type(struct gallivm_state
*gallivm
,
1175 const struct util_format_description
*src_fmt
,
1176 struct lp_type src_type
,
1177 struct lp_type dst_type
,
1178 LLVMValueRef
* src
, // and dst
1181 LLVMValueRef
* dst
= src
;
1183 struct lp_type mem_type
;
1184 struct lp_type blend_type
;
1185 LLVMBuilderRef builder
= gallivm
->builder
;
1186 unsigned pixels
= 16 / num_srcs
;
1189 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1190 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1192 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1194 /* Special case for half-floats */
1195 if (mem_type
.width
== 16 && mem_type
.floating
) {
1196 int length
= dst_type
.length
;
1197 assert(blend_type
.width
== 32 && blend_type
.floating
);
1199 dst_type
.length
= src_type
.length
;
1201 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1203 dst_type
.length
= length
;
1207 /* Remove any padding */
1208 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1209 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1211 for (i
= 0; i
< num_srcs
; ++i
) {
1212 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1216 /* No bit arithmitic to do */
1221 src_type
.length
= pixels
;
1222 src_type
.width
= blend_type
.length
* blend_type
.width
;
1223 dst_type
.length
= pixels
;
1225 for (i
= 0; i
< num_srcs
; ++i
) {
1226 LLVMValueRef chans
[4];
1230 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1232 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1235 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1237 for (k
= 0; k
< blend_type
.width
; ++k
) {
1242 chans
[j
] = LLVMBuildLShr(builder
,
1244 lp_build_const_int_vec(gallivm
, src_type
, j
* blend_type
.width
),
1247 chans
[j
] = LLVMBuildAnd(builder
,
1249 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1252 /* Scale down bits */
1253 chans
[j
] = scale_bits(gallivm
, blend_type
.width
, src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1256 chans
[j
] = LLVMBuildShl(builder
,
1258 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1261 sa
+= src_fmt
->channel
[j
].size
;
1266 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1270 assert (dst_type
.width
!= 24);
1272 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1278 * Generates the blend function for unswizzled colour buffers
1279 * Also generates the read & write from colour buffer
1282 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1284 struct lp_fragment_shader_variant
*variant
,
1285 enum pipe_format out_format
,
1286 unsigned int num_fs
,
1287 struct lp_type fs_type
,
1288 LLVMValueRef
* fs_mask
,
1289 LLVMValueRef fs_out_color
[TGSI_NUM_CHANNELS
][4],
1290 LLVMValueRef context_ptr
,
1291 LLVMValueRef color_ptr
,
1292 LLVMValueRef stride
,
1293 unsigned partial_mask
,
1296 const unsigned alpha_channel
= 3;
1297 const unsigned block_width
= 4;
1298 const unsigned block_height
= 4;
1299 const unsigned block_size
= block_width
* block_height
;
1300 const unsigned lp_integer_vector_width
= 128;
1302 LLVMBuilderRef builder
= gallivm
->builder
;
1303 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1304 LLVMValueRef src_alpha
[4 * 4];
1305 LLVMValueRef src_mask
[4 * 4];
1306 LLVMValueRef src
[4 * 4];
1307 LLVMValueRef dst
[4 * 4];
1308 LLVMValueRef blend_color
;
1309 LLVMValueRef blend_alpha
;
1310 LLVMValueRef i32_zero
;
1311 LLVMValueRef check_mask
;
1313 struct lp_build_mask_context mask_ctx
;
1314 struct lp_type mask_type
;
1315 struct lp_type blend_type
;
1316 struct lp_type alpha_type
;
1317 struct lp_type row_type
;
1318 struct lp_type dst_type
;
1320 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1321 unsigned vector_width
;
1322 unsigned dst_channels
;
1323 unsigned src_channels
;
1328 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1330 unsigned dst_alignment
;
1332 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1333 bool has_alpha
= false;
1335 src_channels
= TGSI_NUM_CHANNELS
;
1336 mask_type
= lp_int32_vec4_type();
1337 mask_type
.length
= fs_type
.length
;
1339 /* Compute the alignment of the destination pointer in bytes */
1341 dst_alignment
= (block_width
* out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1343 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1344 dst_alignment
= (1 * out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1346 /* Force power-of-two alignment by extracting only the least-significant-bit */
1347 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
1348 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1349 dst_alignment
= MIN2(dst_alignment
, 16);
1351 /* Do not bother executing code when mask is empty.. */
1353 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1355 for (i
= 0; i
< num_fs
; ++i
) {
1356 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1359 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1360 lp_build_mask_check(&mask_ctx
);
1363 partial_mask
|= !variant
->opaque
;
1364 i32_zero
= lp_build_const_int32(gallivm
, 0);
1366 /* Get type from output format */
1367 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1368 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1370 row_type
.length
= fs_type
.length
;
1371 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1373 /* Compute correct swizzle and count channels */
1374 memset(swizzle
, 0xFF, TGSI_NUM_CHANNELS
);
1377 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1378 /* Ensure channel is used */
1379 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1383 /* Ensure not already written to (happens in case with GL_ALPHA) */
1384 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1388 /* Ensure we havn't already found all channels */
1389 if (dst_channels
>= out_format_desc
->nr_channels
) {
1393 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1396 if (i
== alpha_channel
) {
1401 /* If 3 channels then pad to include alpha for 4 element transpose */
1402 if (dst_channels
== 3 && !has_alpha
) {
1405 if (out_format_desc
->nr_channels
== 4) {
1411 * Load shader output
1413 for (i
= 0; i
< num_fs
; ++i
) {
1414 /* Always load alpha for use in blending */
1415 LLVMValueRef alpha
= LLVMBuildLoad(builder
, fs_out_color
[alpha_channel
][i
], "");
1417 /* Load each channel */
1418 for (j
= 0; j
< dst_channels
; ++j
) {
1419 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[swizzle
[j
]][i
], "");
1422 /* If 3 channels then pad to include alpha for 4 element transpose */
1423 if (dst_channels
== 3 && !has_alpha
) {
1424 fs_src
[i
][3] = alpha
;
1428 /* We split the row_mask and row_alpha as we want 128bit interleave */
1429 if (fs_type
.length
== 8) {
1430 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1431 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1433 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1434 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1436 src_mask
[i
] = fs_mask
[i
];
1437 src_alpha
[i
] = alpha
;
1443 * Pixel twiddle from fragment shader order to memory order
1445 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fs
, dst_channels
, fs_src
, src
, pad_inline
);
1446 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1447 if (src_count
!= num_fs
* src_channels
) {
1448 unsigned ds
= src_count
/ (num_fs
* src_channels
);
1449 row_type
.length
/= ds
;
1450 fs_type
.length
= row_type
.length
;
1453 blend_type
= row_type
;
1454 alpha_type
= fs_type
;
1455 alpha_type
.length
= 4;
1456 mask_type
.length
= 4;
1458 /* Convert src to row_type */
1459 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1461 /* If the rows are not an SSE vector, combine them to become SSE size! */
1462 if ((row_type
.width
* row_type
.length
) % 128) {
1463 unsigned bits
= row_type
.width
* row_type
.length
;
1466 dst_count
= src_count
/ (vector_width
/ bits
);
1467 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1469 row_type
.length
*= combined
;
1470 src_count
/= combined
;
1472 bits
= row_type
.width
* row_type
.length
;
1473 assert(bits
== 128 || bits
== 256);
1478 * Blend Colour conversion
1480 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1481 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1482 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1485 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1488 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1490 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1491 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1493 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1494 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1496 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1497 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1503 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], 4, &src_mask
[0]);
1505 if (src_count
< block_height
) {
1506 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1507 } else if (src_count
> block_height
) {
1508 for (i
= src_count
; i
> 0; --i
) {
1509 unsigned pixels
= block_size
/ src_count
;
1510 unsigned idx
= i
- 1;
1512 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4], (idx
* pixels
) % 4, pixels
);
1516 assert(mask_type
.width
== 32);
1518 for (i
= 0; i
< src_count
; ++i
) {
1519 unsigned pixels
= block_size
/ src_count
;
1520 unsigned pixel_width
= row_type
.width
* dst_channels
;
1522 if (pixel_width
== 24) {
1523 mask_type
.width
= 8;
1524 mask_type
.length
= vector_width
/ mask_type
.width
;
1526 mask_type
.length
= pixels
;
1527 mask_type
.width
= row_type
.width
* dst_channels
;
1529 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1531 mask_type
.length
*= dst_channels
;
1532 mask_type
.width
/= dst_channels
;
1535 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1536 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1543 unsigned length
= row_type
.length
;
1544 row_type
.length
= alpha_type
.length
;
1546 /* Twiddle the alpha to match pixels */
1547 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, 4, src_alpha
);
1549 for (i
= 0; i
< 4; ++i
) {
1550 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1553 alpha_type
= row_type
;
1554 row_type
.length
= length
;
1556 /* If only one channel we can only need the single alpha value per pixel */
1557 if (src_count
== 1) {
1558 assert(dst_channels
== 1);
1560 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, 4, src_alpha
, src_count
);
1562 /* If there are more srcs than rows then we need to split alpha up */
1563 if (src_count
> block_height
) {
1564 for (i
= src_count
; i
> 0; --i
) {
1565 unsigned pixels
= block_size
/ src_count
;
1566 unsigned idx
= i
- 1;
1568 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4], (idx
* pixels
) % 4, pixels
);
1572 /* If there is a src for each pixel broadcast the alpha across whole row */
1573 if (src_count
== block_size
) {
1574 for (i
= 0; i
< src_count
; ++i
) {
1575 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1578 unsigned pixels
= block_size
/ src_count
;
1579 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1580 unsigned alpha_span
= 1;
1582 /* Check if we need 2 src_alphas for our shuffles */
1583 if (pixels
> alpha_type
.length
) {
1587 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1588 for (i
= 0; i
< src_count
; ++i
) {
1589 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1590 unsigned idx1
= i
, idx2
= i
;
1592 if (alpha_span
> 1){
1597 for (j
= 0; j
< row_type
.length
; ++j
) {
1598 if (j
< pixels
* channels
) {
1599 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1601 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1605 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1608 LLVMConstVector(shuffles
, row_type
.length
),
1617 * Load dst from memory
1619 if (src_count
< block_height
) {
1620 dst_count
= block_height
;
1622 dst_count
= src_count
;
1625 dst_type
.length
*= 16 / dst_count
;
1627 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1628 dst
, dst_type
, dst_count
, dst_alignment
);
1632 * Convert from dst/output format to src/blending format.
1634 * This is necessary as we can only read 1 row from memory at a time,
1635 * so the minimum dst_count will ever be at this point is 4.
1637 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1638 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1639 * on all 16 pixels in that single vector at once.
1641 if (dst_count
> src_count
) {
1642 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
1648 convert_to_blend_type(gallivm
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
1650 for (i
= 0; i
< src_count
; ++i
) {
1651 dst
[i
] = lp_build_blend_aos(gallivm
,
1652 &variant
->key
.blend
,
1653 variant
->key
.cbuf_format
,
1657 has_alpha
? NULL
: src_alpha
[i
],
1659 partial_mask
? src_mask
[i
] : NULL
,
1661 has_alpha
? NULL
: blend_alpha
,
1663 pad_inline
? 4 : dst_channels
);
1666 convert_from_blend_type(gallivm
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
1668 /* Split the blend rows back to memory rows */
1669 if (dst_count
> src_count
) {
1670 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
1672 if (src_count
== 1) {
1673 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1674 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1676 row_type
.length
/= 2;
1680 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
1681 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
1682 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1683 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1685 row_type
.length
/= 2;
1691 * Store blend result to memory
1693 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1694 dst
, dst_type
, dst_count
, dst_alignment
);
1697 lp_build_mask_end(&mask_ctx
);
1703 * Generate the runtime callable function for the whole fragment pipeline.
1704 * Note that the function which we generate operates on a block of 16
1705 * pixels at at time. The block contains 2x2 quads. Each quad contains
1709 generate_fragment(struct llvmpipe_context
*lp
,
1710 struct lp_fragment_shader
*shader
,
1711 struct lp_fragment_shader_variant
*variant
,
1712 unsigned partial_mask
)
1714 struct gallivm_state
*gallivm
= variant
->gallivm
;
1715 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
1716 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
1717 char func_name
[256];
1718 struct lp_type fs_type
;
1719 struct lp_type blend_type
;
1720 LLVMTypeRef fs_elem_type
;
1721 LLVMTypeRef blend_vec_type
;
1722 LLVMTypeRef arg_types
[12];
1723 LLVMTypeRef func_type
;
1724 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1725 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1726 LLVMValueRef context_ptr
;
1729 LLVMValueRef a0_ptr
;
1730 LLVMValueRef dadx_ptr
;
1731 LLVMValueRef dady_ptr
;
1732 LLVMValueRef color_ptr_ptr
;
1733 LLVMValueRef stride_ptr
;
1734 LLVMValueRef depth_ptr
;
1735 LLVMValueRef mask_input
;
1736 LLVMValueRef counter
= NULL
;
1737 LLVMBasicBlockRef block
;
1738 LLVMBuilderRef builder
;
1739 struct lp_build_sampler_soa
*sampler
;
1740 struct lp_build_interp_soa_context interp
;
1741 LLVMValueRef fs_mask
[16 / 4];
1742 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
1743 LLVMValueRef function
;
1744 LLVMValueRef facing
;
1745 const struct util_format_description
*zs_format_desc
;
1750 boolean cbuf0_write_all
;
1751 boolean try_loop
= TRUE
;
1753 assert(lp_native_vector_width
/ 32 >= 4);
1755 /* Adjust color input interpolation according to flatshade state:
1757 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
1758 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1759 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
1761 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1763 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1767 /* check if writes to cbuf[0] are to be copied to all cbufs */
1768 cbuf0_write_all
= FALSE
;
1769 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
1770 if (shader
->info
.base
.properties
[i
].name
==
1771 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
1772 cbuf0_write_all
= TRUE
;
1777 /* TODO: actually pick these based on the fs and color buffer
1778 * characteristics. */
1780 memset(&fs_type
, 0, sizeof fs_type
);
1781 fs_type
.floating
= TRUE
; /* floating point values */
1782 fs_type
.sign
= TRUE
; /* values are signed */
1783 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1784 fs_type
.width
= 32; /* 32-bit float */
1785 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
1786 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
1788 memset(&blend_type
, 0, sizeof blend_type
);
1789 blend_type
.floating
= FALSE
; /* values are integers */
1790 blend_type
.sign
= FALSE
; /* values are unsigned */
1791 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
1792 blend_type
.width
= 8; /* 8-bit ubyte values */
1793 blend_type
.length
= 16; /* 16 elements per vector */
1796 * Generate the function prototype. Any change here must be reflected in
1797 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1800 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
1802 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
1804 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
1805 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
1807 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
1808 arg_types
[1] = int32_type
; /* x */
1809 arg_types
[2] = int32_type
; /* y */
1810 arg_types
[3] = int32_type
; /* facing */
1811 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
1812 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
1813 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
1814 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
1815 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
1816 arg_types
[9] = int32_type
; /* mask_input */
1817 arg_types
[10] = LLVMPointerType(int32_type
, 0); /* counter */
1818 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
1820 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
1821 arg_types
, Elements(arg_types
), 0);
1823 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1824 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
1826 variant
->function
[partial_mask
] = function
;
1828 /* XXX: need to propagate noalias down into color param now we are
1829 * passing a pointer-to-pointer?
1831 for(i
= 0; i
< Elements(arg_types
); ++i
)
1832 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1833 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
1835 context_ptr
= LLVMGetParam(function
, 0);
1836 x
= LLVMGetParam(function
, 1);
1837 y
= LLVMGetParam(function
, 2);
1838 facing
= LLVMGetParam(function
, 3);
1839 a0_ptr
= LLVMGetParam(function
, 4);
1840 dadx_ptr
= LLVMGetParam(function
, 5);
1841 dady_ptr
= LLVMGetParam(function
, 6);
1842 color_ptr_ptr
= LLVMGetParam(function
, 7);
1843 depth_ptr
= LLVMGetParam(function
, 8);
1844 mask_input
= LLVMGetParam(function
, 9);
1845 stride_ptr
= LLVMGetParam(function
, 11);
1847 lp_build_name(context_ptr
, "context");
1848 lp_build_name(x
, "x");
1849 lp_build_name(y
, "y");
1850 lp_build_name(a0_ptr
, "a0");
1851 lp_build_name(dadx_ptr
, "dadx");
1852 lp_build_name(dady_ptr
, "dady");
1853 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
1854 lp_build_name(depth_ptr
, "depth");
1855 lp_build_name(mask_input
, "mask_input");
1856 lp_build_name(stride_ptr
, "stride_ptr");
1858 if (key
->occlusion_count
) {
1859 counter
= LLVMGetParam(function
, 10);
1860 lp_build_name(counter
, "counter");
1867 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
1868 builder
= gallivm
->builder
;
1870 LLVMPositionBuilderAtEnd(builder
, block
);
1872 /* code generated texture sampling */
1873 sampler
= lp_llvm_sampler_soa_create(key
->sampler
, context_ptr
);
1875 zs_format_desc
= util_format_description(key
->zsbuf_format
);
1879 * The shader input interpolation info is not explicitely baked in the
1880 * shader key, but everything it derives from (TGSI, and flatshade) is
1881 * already included in the shader key.
1883 lp_build_interp_soa_init(&interp
,
1885 shader
->info
.base
.num_inputs
,
1889 a0_ptr
, dadx_ptr
, dady_ptr
,
1892 /* loop over quads in the block */
1893 for(i
= 0; i
< num_fs
; ++i
) {
1894 LLVMValueRef depth_offset
= LLVMConstInt(int32_type
,
1895 i
*fs_type
.length
*zs_format_desc
->block
.bits
/8,
1897 LLVMValueRef out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
1898 LLVMValueRef depth_ptr_i
;
1900 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
1902 generate_fs(gallivm
,
1910 &fs_mask
[i
], /* output */
1918 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++)
1919 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1920 fs_out_color
[cbuf
][chan
][i
] =
1921 out_color
[cbuf
* !cbuf0_write_all
][chan
];
1925 unsigned depth_bits
= zs_format_desc
->block
.bits
/8;
1926 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
1927 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
1928 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
1929 num_loop
, "mask_store");
1930 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
1933 * The shader input interpolation info is not explicitely baked in the
1934 * shader key, but everything it derives from (TGSI, and flatshade) is
1935 * already included in the shader key.
1937 lp_build_interp_soa_init(&interp
,
1939 shader
->info
.base
.num_inputs
,
1943 a0_ptr
, dadx_ptr
, dady_ptr
,
1946 for (i
= 0; i
< num_fs
; i
++) {
1948 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
1949 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
1950 &indexi
, 1, "mask_ptr");
1953 mask
= generate_quad_mask(gallivm
, fs_type
,
1954 i
*fs_type
.length
/4, mask_input
);
1957 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
1959 LLVMBuildStore(builder
, mask
, mask_ptr
);
1962 generate_fs_loop(gallivm
,
1970 mask_store
, /* output */
1977 for (i
= 0; i
< num_fs
; i
++) {
1978 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
1979 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
1981 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
1982 /* This is fucked up need to reorganize things */
1983 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
1984 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1985 ptr
= LLVMBuildGEP(builder
,
1986 color_store
[cbuf
* !cbuf0_write_all
][chan
],
1988 fs_out_color
[cbuf
][chan
][i
] = ptr
;
1994 sampler
->destroy(sampler
);
1996 /* Loop over color outputs / color buffers to do blending.
1998 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
1999 LLVMValueRef color_ptr
;
2000 LLVMValueRef stride
;
2001 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2002 unsigned rt
= key
->blend
.independent_blend_enable
? cbuf
: 0;
2004 boolean do_branch
= ((key
->depth
.enabled
2005 || key
->stencil
[0].enabled
2006 || key
->alpha
.enabled
)
2007 && !shader
->info
.base
.uses_kill
);
2009 color_ptr
= LLVMBuildLoad(builder
,
2010 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2013 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2015 stride
= LLVMBuildLoad(builder
,
2016 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2019 generate_unswizzled_blend(gallivm
, rt
, variant
, key
->cbuf_format
[cbuf
],
2020 num_fs
, fs_type
, fs_mask
, fs_out_color
[cbuf
],
2021 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2024 LLVMBuildRetVoid(builder
);
2026 gallivm_verify_function(gallivm
, function
);
2028 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2033 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2037 debug_printf("fs variant %p:\n", (void *) key
);
2039 if (key
->flatshade
) {
2040 debug_printf("flatshade = 1\n");
2042 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2043 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2045 if (key
->depth
.enabled
) {
2046 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2047 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2048 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2051 for (i
= 0; i
< 2; ++i
) {
2052 if (key
->stencil
[i
].enabled
) {
2053 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2054 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2055 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2056 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2057 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2058 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2062 if (key
->alpha
.enabled
) {
2063 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2066 if (key
->occlusion_count
) {
2067 debug_printf("occlusion_count = 1\n");
2070 if (key
->blend
.logicop_enable
) {
2071 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2073 else if (key
->blend
.rt
[0].blend_enable
) {
2074 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2075 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2076 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2077 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2078 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2079 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2081 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2082 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2083 debug_printf("sampler[%u] = \n", i
);
2084 debug_printf(" .format = %s\n",
2085 util_format_name(key
->sampler
[i
].format
));
2086 debug_printf(" .target = %s\n",
2087 util_dump_tex_target(key
->sampler
[i
].target
, TRUE
));
2088 debug_printf(" .pot = %u %u %u\n",
2089 key
->sampler
[i
].pot_width
,
2090 key
->sampler
[i
].pot_height
,
2091 key
->sampler
[i
].pot_depth
);
2092 debug_printf(" .wrap = %s %s %s\n",
2093 util_dump_tex_wrap(key
->sampler
[i
].wrap_s
, TRUE
),
2094 util_dump_tex_wrap(key
->sampler
[i
].wrap_t
, TRUE
),
2095 util_dump_tex_wrap(key
->sampler
[i
].wrap_r
, TRUE
));
2096 debug_printf(" .min_img_filter = %s\n",
2097 util_dump_tex_filter(key
->sampler
[i
].min_img_filter
, TRUE
));
2098 debug_printf(" .min_mip_filter = %s\n",
2099 util_dump_tex_mipfilter(key
->sampler
[i
].min_mip_filter
, TRUE
));
2100 debug_printf(" .mag_img_filter = %s\n",
2101 util_dump_tex_filter(key
->sampler
[i
].mag_img_filter
, TRUE
));
2102 if (key
->sampler
[i
].compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2103 debug_printf(" .compare_func = %s\n", util_dump_func(key
->sampler
[i
].compare_func
, TRUE
));
2104 debug_printf(" .normalized_coords = %u\n", key
->sampler
[i
].normalized_coords
);
2105 debug_printf(" .min_max_lod_equal = %u\n", key
->sampler
[i
].min_max_lod_equal
);
2106 debug_printf(" .lod_bias_non_zero = %u\n", key
->sampler
[i
].lod_bias_non_zero
);
2107 debug_printf(" .apply_min_lod = %u\n", key
->sampler
[i
].apply_min_lod
);
2108 debug_printf(" .apply_max_lod = %u\n", key
->sampler
[i
].apply_max_lod
);
2114 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2116 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2117 variant
->shader
->no
, variant
->no
);
2118 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2119 dump_fs_variant_key(&variant
->key
);
2120 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2126 * Generate a new fragment shader variant from the shader code and
2127 * other state indicated by the key.
2129 static struct lp_fragment_shader_variant
*
2130 generate_variant(struct llvmpipe_context
*lp
,
2131 struct lp_fragment_shader
*shader
,
2132 const struct lp_fragment_shader_variant_key
*key
)
2134 struct lp_fragment_shader_variant
*variant
;
2135 const struct util_format_description
*cbuf0_format_desc
;
2136 boolean fullcolormask
;
2138 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2142 variant
->gallivm
= gallivm_create();
2143 if (!variant
->gallivm
) {
2148 variant
->shader
= shader
;
2149 variant
->list_item_global
.base
= variant
;
2150 variant
->list_item_local
.base
= variant
;
2151 variant
->no
= shader
->variants_created
++;
2153 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2156 * Determine whether we are touching all channels in the color buffer.
2158 fullcolormask
= FALSE
;
2159 if (key
->nr_cbufs
== 1) {
2160 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2161 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2165 !key
->blend
.logicop_enable
&&
2166 !key
->blend
.rt
[0].blend_enable
&&
2168 !key
->stencil
[0].enabled
&&
2169 !key
->alpha
.enabled
&&
2170 !key
->depth
.enabled
&&
2171 !shader
->info
.base
.uses_kill
2174 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2175 lp_debug_fs_variant(variant
);
2178 lp_jit_init_types(variant
);
2180 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2181 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2183 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2184 if (variant
->opaque
) {
2185 /* Specialized shader, which doesn't need to read the color buffer. */
2186 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2191 * Compile everything
2194 gallivm_compile_module(variant
->gallivm
);
2196 if (variant
->function
[RAST_EDGE_TEST
]) {
2197 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2198 gallivm_jit_function(variant
->gallivm
,
2199 variant
->function
[RAST_EDGE_TEST
]);
2202 if (variant
->function
[RAST_WHOLE
]) {
2203 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2204 gallivm_jit_function(variant
->gallivm
,
2205 variant
->function
[RAST_WHOLE
]);
2206 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2207 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2215 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2216 const struct pipe_shader_state
*templ
)
2218 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2219 struct lp_fragment_shader
*shader
;
2223 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2227 shader
->no
= fs_no
++;
2228 make_empty_list(&shader
->variants
);
2230 /* get/save the summary info for this shader */
2231 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2233 /* we need to keep a local copy of the tokens */
2234 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2236 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2237 if (shader
->draw_data
== NULL
) {
2238 FREE((void *) shader
->base
.tokens
);
2243 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2245 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2246 sampler
[nr_samplers
]);
2248 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2249 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2250 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2252 switch (shader
->info
.base
.input_interpolate
[i
]) {
2253 case TGSI_INTERPOLATE_CONSTANT
:
2254 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2256 case TGSI_INTERPOLATE_LINEAR
:
2257 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2259 case TGSI_INTERPOLATE_PERSPECTIVE
:
2260 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2262 case TGSI_INTERPOLATE_COLOR
:
2263 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2270 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2271 case TGSI_SEMANTIC_FACE
:
2272 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2274 case TGSI_SEMANTIC_POSITION
:
2275 /* Position was already emitted above
2277 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2278 shader
->inputs
[i
].src_index
= 0;
2282 shader
->inputs
[i
].src_index
= i
+1;
2285 if (LP_DEBUG
& DEBUG_TGSI
) {
2287 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2288 shader
->no
, (void *) shader
);
2289 tgsi_dump(templ
->tokens
, 0);
2290 debug_printf("usage masks:\n");
2291 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2292 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2293 debug_printf(" IN[%u].%s%s%s%s\n",
2295 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2296 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2297 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2298 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2308 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2310 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2312 if (llvmpipe
->fs
== fs
)
2315 draw_flush(llvmpipe
->draw
);
2317 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2319 draw_bind_fragment_shader(llvmpipe
->draw
,
2320 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2322 llvmpipe
->dirty
|= LP_NEW_FS
;
2327 * Remove shader variant from two lists: the shader's variant list
2328 * and the context's variant list.
2331 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2332 struct lp_fragment_shader_variant
*variant
)
2336 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2337 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2338 " #%u v total cached #%u\n",
2339 variant
->shader
->no
,
2341 variant
->shader
->variants_created
,
2342 variant
->shader
->variants_cached
,
2343 lp
->nr_fs_variants
);
2346 /* free all the variant's JIT'd functions */
2347 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2348 if (variant
->function
[i
]) {
2349 gallivm_free_function(variant
->gallivm
,
2350 variant
->function
[i
],
2351 variant
->jit_function
[i
]);
2355 gallivm_destroy(variant
->gallivm
);
2357 /* remove from shader's list */
2358 remove_from_list(&variant
->list_item_local
);
2359 variant
->shader
->variants_cached
--;
2361 /* remove from context's list */
2362 remove_from_list(&variant
->list_item_global
);
2363 lp
->nr_fs_variants
--;
2364 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2371 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2373 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2374 struct lp_fragment_shader
*shader
= fs
;
2375 struct lp_fs_variant_list_item
*li
;
2377 assert(fs
!= llvmpipe
->fs
);
2380 * XXX: we need to flush the context until we have some sort of reference
2381 * counting in fragment shaders as they may still be binned
2382 * Flushing alone might not sufficient we need to wait on it too.
2384 llvmpipe_finish(pipe
, __FUNCTION__
);
2386 /* Delete all the variants */
2387 li
= first_elem(&shader
->variants
);
2388 while(!at_end(&shader
->variants
, li
)) {
2389 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2390 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2394 /* Delete draw module's data */
2395 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2397 assert(shader
->variants_cached
== 0);
2398 FREE((void *) shader
->base
.tokens
);
2405 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2406 uint shader
, uint index
,
2407 struct pipe_constant_buffer
*cb
)
2409 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2410 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2414 if (cb
&& cb
->user_buffer
) {
2415 constants
= llvmpipe_user_buffer_create(pipe
->screen
,
2416 (void *) cb
->user_buffer
,
2418 PIPE_BIND_CONSTANT_BUFFER
);
2421 size
= constants
? constants
->width0
: 0;
2422 data
= constants
? llvmpipe_resource_data(constants
) : NULL
;
2424 assert(shader
< PIPE_SHADER_TYPES
);
2425 assert(index
< PIPE_MAX_CONSTANT_BUFFERS
);
2427 if(llvmpipe
->constants
[shader
][index
] == constants
)
2430 draw_flush(llvmpipe
->draw
);
2432 /* note: reference counting */
2433 pipe_resource_reference(&llvmpipe
->constants
[shader
][index
], constants
);
2435 if(shader
== PIPE_SHADER_VERTEX
||
2436 shader
== PIPE_SHADER_GEOMETRY
) {
2437 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2441 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2443 if (cb
&& cb
->user_buffer
) {
2444 pipe_resource_reference(&constants
, NULL
);
2450 * Return the blend factor equivalent to a destination alpha of one.
2452 static INLINE
unsigned
2453 force_dst_alpha_one(unsigned factor
)
2456 case PIPE_BLENDFACTOR_DST_ALPHA
:
2457 return PIPE_BLENDFACTOR_ONE
;
2458 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2459 return PIPE_BLENDFACTOR_ZERO
;
2460 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2461 return PIPE_BLENDFACTOR_ZERO
;
2469 * We need to generate several variants of the fragment pipeline to match
2470 * all the combinations of the contributing state atoms.
2472 * TODO: there is actually no reason to tie this to context state -- the
2473 * generated code could be cached globally in the screen.
2476 make_variant_key(struct llvmpipe_context
*lp
,
2477 struct lp_fragment_shader
*shader
,
2478 struct lp_fragment_shader_variant_key
*key
)
2482 memset(key
, 0, shader
->variant_key_size
);
2484 if (lp
->framebuffer
.zsbuf
) {
2485 if (lp
->depth_stencil
->depth
.enabled
) {
2486 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2487 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2489 if (lp
->depth_stencil
->stencil
[0].enabled
) {
2490 key
->zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2491 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2495 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2496 if(key
->alpha
.enabled
)
2497 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2498 /* alpha.ref_value is passed in jit_context */
2500 key
->flatshade
= lp
->rasterizer
->flatshade
;
2501 if (lp
->active_occlusion_query
) {
2502 key
->occlusion_count
= TRUE
;
2505 if (lp
->framebuffer
.nr_cbufs
) {
2506 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2509 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2510 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2511 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2512 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2513 const struct util_format_description
*format_desc
;
2515 key
->cbuf_format
[i
] = format
;
2517 format_desc
= util_format_description(format
);
2518 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2519 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2521 blend_rt
->colormask
= lp
->blend
->rt
[i
].colormask
;
2524 * Mask out color channels not present in the color buffer.
2526 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2529 * Our swizzled render tiles always have an alpha channel, but the linear
2530 * render target format often does not, so force here the dst alpha to be
2533 * This is not a mere optimization. Wrong results will be produced if the
2534 * dst alpha is used, the dst format does not have alpha, and the previous
2535 * rendering was not flushed from the swizzled to linear buffer. For
2536 * example, NonPowTwo DCT.
2538 * TODO: This should be generalized to all channels for better
2539 * performance, but only alpha causes correctness issues.
2541 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2543 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2544 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2545 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
2546 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
2547 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2548 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2549 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2553 /* This value will be the same for all the variants of a given shader:
2555 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2557 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2558 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2559 lp_sampler_static_state(&key
->sampler
[i
],
2560 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
],
2561 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
2569 * Update fragment shader state. This is called just prior to drawing
2570 * something when some fragment-related state has changed.
2573 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
2575 struct lp_fragment_shader
*shader
= lp
->fs
;
2576 struct lp_fragment_shader_variant_key key
;
2577 struct lp_fragment_shader_variant
*variant
= NULL
;
2578 struct lp_fs_variant_list_item
*li
;
2580 make_variant_key(lp
, shader
, &key
);
2582 /* Search the variants for one which matches the key */
2583 li
= first_elem(&shader
->variants
);
2584 while(!at_end(&shader
->variants
, li
)) {
2585 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
2593 /* Move this variant to the head of the list to implement LRU
2594 * deletion of shader's when we have too many.
2596 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2599 /* variant not found, create it now */
2602 unsigned variants_to_cull
;
2605 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2608 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
2611 /* First, check if we've exceeded the max number of shader variants.
2612 * If so, free 25% of them (the least recently used ones).
2614 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
2616 if (variants_to_cull
||
2617 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
2618 struct pipe_context
*pipe
= &lp
->pipe
;
2621 * XXX: we need to flush the context until we have some sort of
2622 * reference counting in fragment shaders as they may still be binned
2623 * Flushing alone might not be sufficient we need to wait on it too.
2625 llvmpipe_finish(pipe
, __FUNCTION__
);
2628 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2629 * number of shader variants (potentially all of them) could be
2630 * pending for destruction on flush.
2633 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
2634 struct lp_fs_variant_list_item
*item
;
2635 if (is_empty_list(&lp
->fs_variants_list
)) {
2638 item
= last_elem(&lp
->fs_variants_list
);
2641 llvmpipe_remove_shader_variant(lp
, item
->base
);
2646 * Generate the new variant.
2649 variant
= generate_variant(lp
, shader
, &key
);
2652 LP_COUNT_ADD(llvm_compile_time
, dt
);
2653 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
2655 llvmpipe_variant_count
++;
2657 /* Put the new variant into the list */
2659 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
2660 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2661 lp
->nr_fs_variants
++;
2662 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
2663 shader
->variants_cached
++;
2667 /* Bind this variant */
2668 lp_setup_set_fs_variant(lp
->setup
, variant
);
2678 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
2680 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
2681 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
2682 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
2684 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;