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 "util/u_dual_blend.h"
70 #include "os/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_swizzle.h"
84 #include "gallivm/lp_bld_flow.h"
85 #include "gallivm/lp_bld_debug.h"
86 #include "gallivm/lp_bld_arit.h"
87 #include "gallivm/lp_bld_pack.h"
88 #include "gallivm/lp_bld_format.h"
89 #include "gallivm/lp_bld_quad.h"
91 #include "lp_bld_alpha.h"
92 #include "lp_bld_blend.h"
93 #include "lp_bld_depth.h"
94 #include "lp_bld_interp.h"
95 #include "lp_context.h"
100 #include "lp_tex_sample.h"
101 #include "lp_flush.h"
102 #include "lp_state_fs.h"
106 /** Fragment shader number (for debugging) */
107 static unsigned fs_no
= 0;
111 * Expand the relevant bits of mask_input to a n*4-dword mask for the
112 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
113 * quad mask vector to 0 or ~0.
114 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
115 * quad arguments with fs length 8.
117 * \param first_quad which quad(s) of the quad group to test, in [0,3]
118 * \param mask_input bitwise mask for the whole 4x4 stamp
121 generate_quad_mask(struct gallivm_state
*gallivm
,
122 struct lp_type fs_type
,
124 LLVMValueRef mask_input
) /* int32 */
126 LLVMBuilderRef builder
= gallivm
->builder
;
127 struct lp_type mask_type
;
128 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
129 LLVMValueRef bits
[16];
134 * XXX: We'll need a different path for 16 x u8
136 assert(fs_type
.width
== 32);
137 assert(fs_type
.length
<= Elements(bits
));
138 mask_type
= lp_int_type(fs_type
);
141 * mask_input >>= (quad * 4)
143 switch (first_quad
) {
148 assert(fs_type
.length
== 4);
155 assert(fs_type
.length
== 4);
163 mask_input
= LLVMBuildLShr(builder
,
165 LLVMConstInt(i32t
, shift
, 0),
169 * mask = { mask_input & (1 << i), for i in [0,3] }
171 mask
= lp_build_broadcast(gallivm
,
172 lp_build_vec_type(gallivm
, mask_type
),
175 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
176 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
177 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
178 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
179 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
180 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
182 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
185 * mask = mask != 0 ? ~0 : 0
187 mask
= lp_build_compare(gallivm
,
188 mask_type
, PIPE_FUNC_NOTEQUAL
,
190 lp_build_const_int_vec(gallivm
, mask_type
, 0));
196 #define EARLY_DEPTH_TEST 0x1
197 #define LATE_DEPTH_TEST 0x2
198 #define EARLY_DEPTH_WRITE 0x4
199 #define LATE_DEPTH_WRITE 0x8
202 find_output_by_semantic( const struct tgsi_shader_info
*info
,
208 for (i
= 0; i
< info
->num_outputs
; i
++)
209 if (info
->output_semantic_name
[i
] == semantic
&&
210 info
->output_semantic_index
[i
] == index
)
218 * Generate the fragment shader, depth/stencil test, and alpha tests.
221 generate_fs_loop(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
,
227 LLVMValueRef num_loop
,
228 struct lp_build_interp_soa_context
*interp
,
229 struct lp_build_sampler_soa
*sampler
,
230 LLVMValueRef mask_store
,
231 LLVMValueRef (*out_color
)[4],
232 LLVMValueRef depth_ptr
,
233 LLVMValueRef depth_stride
,
235 LLVMValueRef thread_data_ptr
)
237 const struct util_format_description
*zs_format_desc
= NULL
;
238 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
239 LLVMTypeRef vec_type
;
240 LLVMValueRef mask_ptr
, mask_val
;
241 LLVMValueRef consts_ptr
;
243 LLVMValueRef z_value
, s_value
;
244 LLVMValueRef z_fb
, s_fb
;
245 LLVMValueRef stencil_refs
[2];
246 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
247 struct lp_build_for_loop_state loop_state
;
248 struct lp_build_mask_context mask
;
250 * TODO: figure out if simple_shader optimization is really worthwile to
251 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
252 * code since tests tend to take another codepath than real shaders.
254 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
255 shader
->info
.base
.num_inputs
< 3 &&
256 shader
->info
.base
.num_instructions
< 8) && 0;
257 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
258 util_blend_state_is_dual(&key
->blend
, 0);
264 struct lp_bld_tgsi_system_values system_values
;
266 memset(&system_values
, 0, sizeof(system_values
));
268 if (key
->depth
.enabled
||
269 key
->stencil
[0].enabled
) {
271 zs_format_desc
= util_format_description(key
->zsbuf_format
);
272 assert(zs_format_desc
);
274 if (!shader
->info
.base
.writes_z
) {
275 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
) {
276 /* With alpha test and kill, can do the depth test early
277 * and hopefully eliminate some quads. But need to do a
278 * special deferred depth write once the final mask value
279 * is known. This only works though if there's either no
280 * stencil test or the stencil value isn't written.
282 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
283 (key
->stencil
[1].enabled
&&
284 key
->stencil
[1].writemask
)))
285 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
287 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
290 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
293 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
296 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
297 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
298 (key
->stencil
[1].enabled
&&
299 key
->stencil
[1].writemask
))))
300 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
307 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
308 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
310 vec_type
= lp_build_vec_type(gallivm
, type
);
312 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
314 lp_build_for_loop_begin(&loop_state
, gallivm
,
315 lp_build_const_int32(gallivm
, 0),
318 lp_build_const_int32(gallivm
, 1));
320 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
321 &loop_state
.counter
, 1, "mask_ptr");
322 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
324 memset(outputs
, 0, sizeof outputs
);
326 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
327 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
328 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
329 lp_build_vec_type(gallivm
,
334 if (dual_source_blend
) {
335 assert(key
->nr_cbufs
<= 1);
336 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
337 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
338 lp_build_vec_type(gallivm
,
345 /* 'mask' will control execution based on quad's pixel alive/killed state */
346 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
348 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
349 lp_build_mask_check(&mask
);
351 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
354 if (depth_mode
& EARLY_DEPTH_TEST
) {
355 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
356 zs_format_desc
, key
->resource_1d
,
357 depth_ptr
, depth_stride
,
358 &z_fb
, &s_fb
, loop_state
.counter
);
359 lp_build_depth_stencil_test(gallivm
,
371 if (depth_mode
& EARLY_DEPTH_WRITE
) {
372 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
373 zs_format_desc
, key
->resource_1d
,
374 NULL
, NULL
, NULL
, loop_state
.counter
,
375 depth_ptr
, depth_stride
,
379 * Note mask check if stencil is enabled must be after ds write not after
380 * stencil test otherwise new stencil values may not get written if all
381 * fragments got killed by depth/stencil test.
383 if (!simple_shader
&& key
->stencil
[0].enabled
)
384 lp_build_mask_check(&mask
);
387 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
389 /* Build the actual shader */
390 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
391 consts_ptr
, &system_values
,
393 outputs
, sampler
, &shader
->info
.base
, NULL
);
396 if (key
->alpha
.enabled
) {
397 int color0
= find_output_by_semantic(&shader
->info
.base
,
401 if (color0
!= -1 && outputs
[color0
][3]) {
402 const struct util_format_description
*cbuf_format_desc
;
403 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
404 LLVMValueRef alpha_ref_value
;
406 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
407 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
409 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
411 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
412 &mask
, alpha
, alpha_ref_value
,
413 (depth_mode
& LATE_DEPTH_TEST
) != 0);
418 if (depth_mode
& LATE_DEPTH_TEST
) {
419 int pos0
= find_output_by_semantic(&shader
->info
.base
,
420 TGSI_SEMANTIC_POSITION
,
423 if (pos0
!= -1 && outputs
[pos0
][2]) {
424 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
427 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
428 zs_format_desc
, key
->resource_1d
,
429 depth_ptr
, depth_stride
,
430 &z_fb
, &s_fb
, loop_state
.counter
);
432 lp_build_depth_stencil_test(gallivm
,
444 if (depth_mode
& LATE_DEPTH_WRITE
) {
445 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
446 zs_format_desc
, key
->resource_1d
,
447 NULL
, NULL
, NULL
, loop_state
.counter
,
448 depth_ptr
, depth_stride
,
452 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
453 (depth_mode
& LATE_DEPTH_WRITE
))
455 /* Need to apply a reduced mask to the depth write. Reload the
456 * depth value, update from zs_value with the new mask value and
459 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
460 zs_format_desc
, key
->resource_1d
,
461 &mask
, z_fb
, s_fb
, loop_state
.counter
,
462 depth_ptr
, depth_stride
,
468 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
470 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
471 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
472 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
474 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
475 if(outputs
[attrib
][chan
]) {
476 /* XXX: just initialize outputs to point at colors[] and
479 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
480 LLVMValueRef color_ptr
;
481 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
482 &loop_state
.counter
, 1, "");
483 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
484 LLVMBuildStore(builder
, out
, color_ptr
);
490 if (key
->occlusion_count
) {
491 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
492 lp_build_name(counter
, "counter");
493 lp_build_occlusion_count(gallivm
, type
,
494 lp_build_mask_value(&mask
), counter
);
497 mask_val
= lp_build_mask_end(&mask
);
498 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
499 lp_build_for_loop_end(&loop_state
);
504 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
506 * Fragment Shader outputs pixels in small 2x2 blocks
507 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
509 * However in memory pixels are stored in rows
510 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
512 * @param type fragment shader type (4x or 8x float)
513 * @param num_fs number of fs_src
514 * @param is_1d whether we're outputting to a 1d resource
515 * @param dst_channels number of output channels
516 * @param fs_src output from fragment shader
517 * @param dst pointer to store result
518 * @param pad_inline is channel padding inline or at end of row
519 * @return the number of dsts
522 generate_fs_twiddle(struct gallivm_state
*gallivm
,
525 unsigned dst_channels
,
526 LLVMValueRef fs_src
[][4],
530 LLVMValueRef src
[16];
536 unsigned pixels
= type
.length
/ 4;
537 unsigned reorder_group
;
538 unsigned src_channels
;
542 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
543 src_count
= num_fs
* src_channels
;
545 assert(pixels
== 2 || pixels
== 1);
546 assert(num_fs
* src_channels
<= Elements(src
));
549 * Transpose from SoA -> AoS
551 for (i
= 0; i
< num_fs
; ++i
) {
552 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
556 * Pick transformation options
563 if (dst_channels
== 1) {
569 } else if (dst_channels
== 2) {
573 } else if (dst_channels
> 2) {
580 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
586 * Split the src in half
589 for (i
= num_fs
; i
> 0; --i
) {
590 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
591 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
599 * Ensure pixels are in memory order
602 /* Twiddle pixels by reordering the array, e.g.:
604 * src_count = 8 -> 0 2 1 3 4 6 5 7
605 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
607 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
609 for (i
= 0; i
< src_count
; ++i
) {
610 unsigned group
= i
/ reorder_group
;
611 unsigned block
= (group
/ 4) * 4 * reorder_group
;
612 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
615 } else if (twiddle
) {
616 /* Twiddle pixels across elements of array */
617 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
620 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
624 * Moves any padding between pixels to the end
625 * e.g. RGBXRGBX -> RGBRGBXX
628 unsigned char swizzles
[16];
629 unsigned elems
= pixels
* dst_channels
;
631 for (i
= 0; i
< type
.length
; ++i
) {
633 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
635 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
638 for (i
= 0; i
< src_count
; ++i
) {
639 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
648 * Load an unswizzled block of pixels from memory
651 load_unswizzled_block(struct gallivm_state
*gallivm
,
652 LLVMValueRef base_ptr
,
654 unsigned block_width
,
655 unsigned block_height
,
657 struct lp_type dst_type
,
659 unsigned dst_alignment
)
661 LLVMBuilderRef builder
= gallivm
->builder
;
662 unsigned row_size
= dst_count
/ block_height
;
665 /* Ensure block exactly fits into dst */
666 assert((block_width
* block_height
) % dst_count
== 0);
668 for (i
= 0; i
< dst_count
; ++i
) {
669 unsigned x
= i
% row_size
;
670 unsigned y
= i
/ row_size
;
672 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
673 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
676 LLVMValueRef dst_ptr
;
678 gep
[0] = lp_build_const_int32(gallivm
, 0);
679 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
681 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
682 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
684 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
686 lp_set_load_alignment(dst
[i
], dst_alignment
);
692 * Store an unswizzled block of pixels to memory
695 store_unswizzled_block(struct gallivm_state
*gallivm
,
696 LLVMValueRef base_ptr
,
698 unsigned block_width
,
699 unsigned block_height
,
701 struct lp_type src_type
,
703 unsigned src_alignment
)
705 LLVMBuilderRef builder
= gallivm
->builder
;
706 unsigned row_size
= src_count
/ block_height
;
709 /* Ensure src exactly fits into block */
710 assert((block_width
* block_height
) % src_count
== 0);
712 for (i
= 0; i
< src_count
; ++i
) {
713 unsigned x
= i
% row_size
;
714 unsigned y
= i
/ row_size
;
716 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
717 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
720 LLVMValueRef src_ptr
;
722 gep
[0] = lp_build_const_int32(gallivm
, 0);
723 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
725 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
726 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
728 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
730 lp_set_store_alignment(src_ptr
, src_alignment
);
736 * Checks if a format description is an arithmetic format
738 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
740 static INLINE boolean
741 is_arithmetic_format(const struct util_format_description
*format_desc
)
743 boolean arith
= false;
746 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
747 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
748 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
756 * Checks if this format requires special handling due to required expansion
757 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
760 static INLINE boolean
761 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
763 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
764 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
772 * Retrieves the type representing the memory layout for a format
774 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
777 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
778 struct lp_type
* type
)
783 if (format_expands_to_float_soa(format_desc
)) {
784 /* just make this a 32bit uint */
785 type
->floating
= false;
794 for (i
= 0; i
< 4; i
++)
795 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
799 memset(type
, 0, sizeof(struct lp_type
));
800 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
801 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
802 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
803 type
->norm
= format_desc
->channel
[chan
].normalized
;
805 if (is_arithmetic_format(format_desc
)) {
809 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
810 type
->width
+= format_desc
->channel
[i
].size
;
813 type
->width
= format_desc
->channel
[chan
].size
;
814 type
->length
= format_desc
->nr_channels
;
820 * Retrieves the type for a format which is usable in the blending code.
822 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
825 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
826 struct lp_type
* type
)
831 if (format_expands_to_float_soa(format_desc
)) {
832 /* always use ordinary floats for blending */
833 type
->floating
= true;
842 for (i
= 0; i
< 4; i
++)
843 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
847 memset(type
, 0, sizeof(struct lp_type
));
848 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
849 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
850 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
851 type
->norm
= format_desc
->channel
[chan
].normalized
;
852 type
->width
= format_desc
->channel
[chan
].size
;
853 type
->length
= format_desc
->nr_channels
;
855 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
856 if (format_desc
->channel
[i
].size
> type
->width
)
857 type
->width
= format_desc
->channel
[i
].size
;
860 if (type
->floating
) {
863 if (type
->width
<= 8) {
865 } else if (type
->width
<= 16) {
872 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
879 * Scale a normalized value from src_bits to dst_bits
881 static INLINE LLVMValueRef
882 scale_bits(struct gallivm_state
*gallivm
,
886 struct lp_type src_type
)
888 LLVMBuilderRef builder
= gallivm
->builder
;
889 LLVMValueRef result
= src
;
891 if (dst_bits
< src_bits
) {
892 /* Scale down by LShr */
893 result
= LLVMBuildLShr(builder
,
895 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- dst_bits
),
897 } else if (dst_bits
> src_bits
) {
899 int db
= dst_bits
- src_bits
;
901 /* Shift left by difference in bits */
902 result
= LLVMBuildShl(builder
,
904 lp_build_const_int_vec(gallivm
, src_type
, db
),
908 /* Enough bits in src to fill the remainder */
909 LLVMValueRef lower
= LLVMBuildLShr(builder
,
911 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
914 result
= LLVMBuildOr(builder
, result
, lower
, "");
915 } else if (db
> src_bits
) {
916 /* Need to repeatedly copy src bits to fill remainder in dst */
919 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
920 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
922 result
= LLVMBuildOr(builder
,
924 LLVMBuildLShr(builder
, result
, shuv
, ""),
935 * Convert from memory format to blending format
937 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
940 convert_to_blend_type(struct gallivm_state
*gallivm
,
942 const struct util_format_description
*src_fmt
,
943 struct lp_type src_type
,
944 struct lp_type dst_type
,
945 LLVMValueRef
* src
, // and dst
948 LLVMValueRef
*dst
= src
;
949 LLVMBuilderRef builder
= gallivm
->builder
;
950 struct lp_type blend_type
;
951 struct lp_type mem_type
;
953 unsigned pixels
= block_size
/ num_srcs
;
957 * full custom path for packed floats and srgb formats - none of the later
958 * functions would do anything useful, and given the lp_type representation they
959 * can't be fixed. Should really have some SoA blend path for these kind of
960 * formats rather than hacking them in here.
962 if (format_expands_to_float_soa(src_fmt
)) {
963 LLVMValueRef tmpsrc
[4];
965 * This is pretty suboptimal for this case blending in SoA would be much
966 * better, since conversion gets us SoA values so need to convert back.
968 assert(src_type
.width
== 32);
969 assert(dst_type
.floating
);
970 assert(dst_type
.width
== 32);
971 assert(dst_type
.length
% 4 == 0);
972 assert(num_srcs
% 4 == 0);
974 for (i
= 0; i
< 4; i
++) {
977 for (i
= 0; i
< num_srcs
/ 4; i
++) {
978 LLVMValueRef tmpsoa
[4];
979 LLVMValueRef tmps
= tmpsrc
[i
];
980 if (dst_type
.length
== 8) {
981 LLVMValueRef shuffles
[8];
983 /* fetch was 4 values but need 8-wide output values */
984 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
986 * for 8-wide aos transpose would give us wrong order not matching
987 * incoming converted fs values and mask. ARGH.
989 for (j
= 0; j
< 4; j
++) {
990 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
991 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
993 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
994 LLVMConstVector(shuffles
, 8), "");
996 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
997 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1000 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1002 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1007 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1008 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1010 /* Is the format arithmetic */
1011 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1012 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1014 /* Pad if necessary */
1015 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1016 for (i
= 0; i
< num_srcs
; ++i
) {
1017 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1020 src_type
.length
= dst_type
.length
;
1023 /* Special case for half-floats */
1024 if (mem_type
.width
== 16 && mem_type
.floating
) {
1025 assert(blend_type
.width
== 32 && blend_type
.floating
);
1026 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1034 src_type
.width
= blend_type
.width
* blend_type
.length
;
1035 blend_type
.length
*= pixels
;
1036 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1038 for (i
= 0; i
< num_srcs
; ++i
) {
1039 LLVMValueRef chans
[4];
1040 LLVMValueRef res
= NULL
;
1042 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1044 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1046 unsigned sa
= src_fmt
->channel
[j
].shift
;
1047 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1048 unsigned from_lsb
= j
;
1050 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1053 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1057 /* Extract bits from source */
1058 chans
[j
] = LLVMBuildLShr(builder
,
1060 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1063 chans
[j
] = LLVMBuildAnd(builder
,
1065 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1069 if (src_type
.norm
) {
1070 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1071 blend_type
.width
, chans
[j
], src_type
);
1074 /* Insert bits into correct position */
1075 chans
[j
] = LLVMBuildShl(builder
,
1077 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1083 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1087 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1093 * Convert from blending format to memory format
1095 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1098 convert_from_blend_type(struct gallivm_state
*gallivm
,
1099 unsigned block_size
,
1100 const struct util_format_description
*src_fmt
,
1101 struct lp_type src_type
,
1102 struct lp_type dst_type
,
1103 LLVMValueRef
* src
, // and dst
1106 LLVMValueRef
* dst
= src
;
1108 struct lp_type mem_type
;
1109 struct lp_type blend_type
;
1110 LLVMBuilderRef builder
= gallivm
->builder
;
1111 unsigned pixels
= block_size
/ num_srcs
;
1115 * full custom path for packed floats and srgb formats - none of the later
1116 * functions would do anything useful, and given the lp_type representation they
1117 * can't be fixed. Should really have some SoA blend path for these kind of
1118 * formats rather than hacking them in here.
1120 if (format_expands_to_float_soa(src_fmt
)) {
1122 * This is pretty suboptimal for this case blending in SoA would be much
1123 * better - we need to transpose the AoS values back to SoA values for
1124 * conversion/packing.
1126 assert(src_type
.floating
);
1127 assert(src_type
.width
== 32);
1128 assert(src_type
.length
% 4 == 0);
1129 assert(dst_type
.width
== 32);
1131 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1132 LLVMValueRef tmpsoa
[4], tmpdst
;
1133 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1134 /* really really need SoA here */
1136 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1137 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1140 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1144 if (src_type
.length
== 8) {
1145 LLVMValueRef tmpaos
, shuffles
[8];
1148 * for 8-wide aos transpose has given us wrong order not matching
1149 * output order. HMPF. Also need to split the output values manually.
1151 for (j
= 0; j
< 4; j
++) {
1152 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1153 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1155 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1156 LLVMConstVector(shuffles
, 8), "");
1157 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1158 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1167 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1168 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1170 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1172 /* Special case for half-floats */
1173 if (mem_type
.width
== 16 && mem_type
.floating
) {
1174 int length
= dst_type
.length
;
1175 assert(blend_type
.width
== 32 && blend_type
.floating
);
1177 dst_type
.length
= src_type
.length
;
1179 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1181 dst_type
.length
= length
;
1185 /* Remove any padding */
1186 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1187 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1189 for (i
= 0; i
< num_srcs
; ++i
) {
1190 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1194 /* No bit arithmetic to do */
1199 src_type
.length
= pixels
;
1200 src_type
.width
= blend_type
.length
* blend_type
.width
;
1201 dst_type
.length
= pixels
;
1203 for (i
= 0; i
< num_srcs
; ++i
) {
1204 LLVMValueRef chans
[4];
1205 LLVMValueRef res
= NULL
;
1207 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1209 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1211 unsigned sa
= src_fmt
->channel
[j
].shift
;
1212 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1213 unsigned from_lsb
= j
;
1215 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1218 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1220 for (k
= 0; k
< blend_type
.width
; ++k
) {
1225 chans
[j
] = LLVMBuildLShr(builder
,
1227 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1230 chans
[j
] = LLVMBuildAnd(builder
,
1232 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1235 /* Scale down bits */
1236 if (src_type
.norm
) {
1237 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1238 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1242 chans
[j
] = LLVMBuildShl(builder
,
1244 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1247 sa
+= src_fmt
->channel
[j
].size
;
1252 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1256 assert (dst_type
.width
!= 24);
1258 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1264 * Convert alpha to same blend type as src
1267 convert_alpha(struct gallivm_state
*gallivm
,
1268 struct lp_type row_type
,
1269 struct lp_type alpha_type
,
1270 const unsigned block_size
,
1271 const unsigned block_height
,
1272 const unsigned src_count
,
1273 const unsigned dst_channels
,
1274 const bool pad_inline
,
1275 LLVMValueRef
* src_alpha
)
1277 LLVMBuilderRef builder
= gallivm
->builder
;
1279 unsigned length
= row_type
.length
;
1280 row_type
.length
= alpha_type
.length
;
1282 /* Twiddle the alpha to match pixels */
1283 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1286 * TODO this should use single lp_build_conv call for
1287 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1289 for (i
= 0; i
< block_height
; ++i
) {
1290 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1293 alpha_type
= row_type
;
1294 row_type
.length
= length
;
1296 /* If only one channel we can only need the single alpha value per pixel */
1297 if (src_count
== 1 && dst_channels
== 1) {
1299 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1301 /* If there are more srcs than rows then we need to split alpha up */
1302 if (src_count
> block_height
) {
1303 for (i
= src_count
; i
> 0; --i
) {
1304 unsigned pixels
= block_size
/ src_count
;
1305 unsigned idx
= i
- 1;
1307 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1308 (idx
* pixels
) % 4, pixels
);
1312 /* If there is a src for each pixel broadcast the alpha across whole row */
1313 if (src_count
== block_size
) {
1314 for (i
= 0; i
< src_count
; ++i
) {
1315 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1318 unsigned pixels
= block_size
/ src_count
;
1319 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1320 unsigned alpha_span
= 1;
1321 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1323 /* Check if we need 2 src_alphas for our shuffles */
1324 if (pixels
> alpha_type
.length
) {
1328 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1329 for (j
= 0; j
< row_type
.length
; ++j
) {
1330 if (j
< pixels
* channels
) {
1331 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1333 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1337 for (i
= 0; i
< src_count
; ++i
) {
1338 unsigned idx1
= i
, idx2
= i
;
1340 if (alpha_span
> 1){
1345 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1348 LLVMConstVector(shuffles
, row_type
.length
),
1357 * Generates the blend function for unswizzled colour buffers
1358 * Also generates the read & write from colour buffer
1361 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1363 struct lp_fragment_shader_variant
*variant
,
1364 enum pipe_format out_format
,
1365 unsigned int num_fs
,
1366 struct lp_type fs_type
,
1367 LLVMValueRef
* fs_mask
,
1368 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1369 LLVMValueRef context_ptr
,
1370 LLVMValueRef color_ptr
,
1371 LLVMValueRef stride
,
1372 unsigned partial_mask
,
1375 const unsigned alpha_channel
= 3;
1376 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1377 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1378 const unsigned block_size
= block_width
* block_height
;
1379 const unsigned lp_integer_vector_width
= 128;
1381 LLVMBuilderRef builder
= gallivm
->builder
;
1382 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1383 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1384 LLVMValueRef src_alpha
[4 * 4];
1385 LLVMValueRef src1_alpha
[4 * 4];
1386 LLVMValueRef src_mask
[4 * 4];
1387 LLVMValueRef src
[4 * 4];
1388 LLVMValueRef src1
[4 * 4];
1389 LLVMValueRef dst
[4 * 4];
1390 LLVMValueRef blend_color
;
1391 LLVMValueRef blend_alpha
;
1392 LLVMValueRef i32_zero
;
1393 LLVMValueRef check_mask
;
1394 LLVMValueRef undef_src_val
;
1396 struct lp_build_mask_context mask_ctx
;
1397 struct lp_type mask_type
;
1398 struct lp_type blend_type
;
1399 struct lp_type row_type
;
1400 struct lp_type dst_type
;
1402 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1403 unsigned vector_width
;
1404 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1405 unsigned dst_channels
;
1410 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1412 unsigned dst_alignment
;
1414 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1415 bool has_alpha
= false;
1416 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1417 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1419 const boolean is_1d
= variant
->key
.resource_1d
;
1420 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1422 mask_type
= lp_int32_vec4_type();
1423 mask_type
.length
= fs_type
.length
;
1425 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1426 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1429 /* Do not bother executing code when mask is empty.. */
1431 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1433 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1434 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1437 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1438 lp_build_mask_check(&mask_ctx
);
1441 partial_mask
|= !variant
->opaque
;
1442 i32_zero
= lp_build_const_int32(gallivm
, 0);
1444 #if HAVE_LLVM < 0x0302
1446 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1447 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1449 undef_src_val
= lp_build_zero(gallivm
, fs_type
);
1451 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1455 /* Get type from output format */
1456 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1457 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1459 row_type
.length
= fs_type
.length
;
1460 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1462 /* Compute correct swizzle and count channels */
1463 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1466 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1467 /* Ensure channel is used */
1468 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1472 /* Ensure not already written to (happens in case with GL_ALPHA) */
1473 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1477 /* Ensure we havn't already found all channels */
1478 if (dst_channels
>= out_format_desc
->nr_channels
) {
1482 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1485 if (i
== alpha_channel
) {
1490 if (format_expands_to_float_soa(out_format_desc
)) {
1492 * the code above can't work for layout_other
1493 * for srgb it would sort of work but we short-circuit swizzles, etc.
1494 * as that is done as part of unpack / pack.
1496 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1502 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1505 /* If 3 channels then pad to include alpha for 4 element transpose */
1506 if (dst_channels
== 3 && !has_alpha
) {
1507 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1508 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1511 if (out_format_desc
->nr_channels
== 4) {
1517 * Load shader output
1519 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1520 /* Always load alpha for use in blending */
1523 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1526 alpha
= undef_src_val
;
1529 /* Load each channel */
1530 for (j
= 0; j
< dst_channels
; ++j
) {
1531 assert(swizzle
[j
] < 4);
1533 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1536 fs_src
[i
][j
] = undef_src_val
;
1540 /* If 3 channels then pad to include alpha for 4 element transpose */
1542 * XXX If we include that here maybe could actually use it instead of
1543 * separate alpha for blending?
1545 if (dst_channels
== 3 && !has_alpha
) {
1546 fs_src
[i
][3] = alpha
;
1549 /* We split the row_mask and row_alpha as we want 128bit interleave */
1550 if (fs_type
.length
== 8) {
1551 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1552 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1554 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1555 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1557 src_mask
[i
] = fs_mask
[i
];
1558 src_alpha
[i
] = alpha
;
1561 if (dual_source_blend
) {
1562 /* same as above except different src/dst, skip masks and comments... */
1563 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1566 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1569 alpha
= undef_src_val
;
1572 for (j
= 0; j
< dst_channels
; ++j
) {
1573 assert(swizzle
[j
] < 4);
1575 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1578 fs_src1
[i
][j
] = undef_src_val
;
1581 if (dst_channels
== 3 && !has_alpha
) {
1582 fs_src1
[i
][3] = alpha
;
1584 if (fs_type
.length
== 8) {
1585 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1586 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1588 src1_alpha
[i
] = alpha
;
1593 if (util_format_is_pure_integer(out_format
)) {
1595 * In this case fs_type was really ints or uints disguised as floats,
1598 fs_type
.floating
= 0;
1599 fs_type
.sign
= dst_type
.sign
;
1600 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1601 for (j
= 0; j
< dst_channels
; ++j
) {
1602 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1603 lp_build_vec_type(gallivm
, fs_type
), "");
1605 if (dst_channels
== 3 && !has_alpha
) {
1606 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1607 lp_build_vec_type(gallivm
, fs_type
), "");
1613 * Pixel twiddle from fragment shader order to memory order
1615 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1616 dst_channels
, fs_src
, src
, pad_inline
);
1617 if (dual_source_blend
) {
1618 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1619 fs_src1
, src1
, pad_inline
);
1622 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1623 if (src_count
!= num_fullblock_fs
* src_channels
) {
1624 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1625 row_type
.length
/= ds
;
1626 fs_type
.length
= row_type
.length
;
1629 blend_type
= row_type
;
1630 mask_type
.length
= 4;
1632 /* Convert src to row_type */
1633 if (dual_source_blend
) {
1634 struct lp_type old_row_type
= row_type
;
1635 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1636 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1639 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1642 /* If the rows are not an SSE vector, combine them to become SSE size! */
1643 if ((row_type
.width
* row_type
.length
) % 128) {
1644 unsigned bits
= row_type
.width
* row_type
.length
;
1647 assert(src_count
>= (vector_width
/ bits
));
1649 dst_count
= src_count
/ (vector_width
/ bits
);
1651 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1652 if (dual_source_blend
) {
1653 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1656 row_type
.length
*= combined
;
1657 src_count
/= combined
;
1659 bits
= row_type
.width
* row_type
.length
;
1660 assert(bits
== 128 || bits
== 256);
1665 * Blend Colour conversion
1667 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1668 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1669 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1672 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1674 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1676 * since blending is done with floats, there was no conversion.
1677 * However, the rules according to fixed point renderbuffers still
1678 * apply, that is we must clamp inputs to 0.0/1.0.
1679 * (This would apply to separate alpha conversion too but we currently
1680 * force has_alpha to be true.)
1681 * TODO: should skip this with "fake" blend, since post-blend conversion
1682 * will clamp anyway.
1683 * TODO: could also skip this if fragment color clamping is enabled. We
1684 * don't support it natively so it gets baked into the shader however, so
1685 * can't really tell here.
1687 struct lp_build_context f32_bld
;
1688 assert(row_type
.floating
);
1689 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1690 for (i
= 0; i
< src_count
; i
++) {
1691 src
[i
] = lp_build_clamp(&f32_bld
, src
[i
], f32_bld
.zero
, f32_bld
.one
);
1693 if (dual_source_blend
) {
1694 for (i
= 0; i
< src_count
; i
++) {
1695 src1
[i
] = lp_build_clamp(&f32_bld
, src1
[i
], f32_bld
.zero
, f32_bld
.one
);
1698 /* probably can't be different than row_type but better safe than sorry... */
1699 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1700 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1704 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1706 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1707 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1709 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1710 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1712 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1713 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1719 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1721 if (src_count
< block_height
) {
1722 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1723 } else if (src_count
> block_height
) {
1724 for (i
= src_count
; i
> 0; --i
) {
1725 unsigned pixels
= block_size
/ src_count
;
1726 unsigned idx
= i
- 1;
1728 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1729 (idx
* pixels
) % 4, pixels
);
1733 assert(mask_type
.width
== 32);
1735 for (i
= 0; i
< src_count
; ++i
) {
1736 unsigned pixels
= block_size
/ src_count
;
1737 unsigned pixel_width
= row_type
.width
* dst_channels
;
1739 if (pixel_width
== 24) {
1740 mask_type
.width
= 8;
1741 mask_type
.length
= vector_width
/ mask_type
.width
;
1743 mask_type
.length
= pixels
;
1744 mask_type
.width
= row_type
.width
* dst_channels
;
1746 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1748 mask_type
.length
*= dst_channels
;
1749 mask_type
.width
/= dst_channels
;
1752 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1753 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1760 struct lp_type alpha_type
= fs_type
;
1761 alpha_type
.length
= 4;
1762 convert_alpha(gallivm
, row_type
, alpha_type
,
1763 block_size
, block_height
,
1764 src_count
, dst_channels
,
1765 pad_inline
, src_alpha
);
1766 if (dual_source_blend
) {
1767 convert_alpha(gallivm
, row_type
, alpha_type
,
1768 block_size
, block_height
,
1769 src_count
, dst_channels
,
1770 pad_inline
, src1_alpha
);
1776 * Load dst from memory
1778 if (src_count
< block_height
) {
1779 dst_count
= block_height
;
1781 dst_count
= src_count
;
1784 dst_type
.length
*= block_size
/ dst_count
;
1786 if (format_expands_to_float_soa(out_format_desc
)) {
1788 * we need multiple values at once for the conversion, so can as well
1789 * load them vectorized here too instead of concatenating later.
1790 * (Still need concatenation later for 8-wide vectors).
1792 dst_count
= block_height
;
1793 dst_type
.length
= block_width
;
1797 * Compute the alignment of the destination pointer in bytes
1798 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1799 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1800 * 1d tex but can't distinguish here) so need to stick with per-pixel
1801 * alignment in this case.
1804 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1807 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
1809 /* Force power-of-two alignment by extracting only the least-significant-bit */
1810 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
1812 * Resource base and stride pointers are aligned to 16 bytes, so that's
1813 * the maximum alignment we can guarantee
1815 dst_alignment
= MIN2(16, dst_alignment
);
1818 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
1819 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
1820 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
1821 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
1826 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1827 dst
, dst_type
, dst_count
, dst_alignment
);
1832 * Convert from dst/output format to src/blending format.
1834 * This is necessary as we can only read 1 row from memory at a time,
1835 * so the minimum dst_count will ever be at this point is 4.
1837 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1838 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1839 * on all 16 pixels in that single vector at once.
1841 if (dst_count
> src_count
) {
1842 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
1848 /* XXX this is broken for RGB8 formats -
1849 * they get expanded from 12 to 16 elements (to include alpha)
1850 * by convert_to_blend_type then reduced to 15 instead of 12
1851 * by convert_from_blend_type (a simple fix though breaks A8...).
1852 * R16G16B16 also crashes differently however something going wrong
1853 * inside llvm handling npot vector sizes seemingly.
1854 * It seems some cleanup could be done here (like skipping conversion/blend
1857 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
1860 * FIXME: Really should get logic ops / masks out of generic blend / row
1861 * format. Logic ops will definitely not work on the blend float format
1862 * used for SRGB here and I think OpenGL expects this to work as expected
1863 * (that is incoming values converted to srgb then logic op applied).
1865 for (i
= 0; i
< src_count
; ++i
) {
1866 dst
[i
] = lp_build_blend_aos(gallivm
,
1867 &variant
->key
.blend
,
1872 has_alpha
? NULL
: src_alpha
[i
],
1874 has_alpha
? NULL
: src1_alpha
[i
],
1876 partial_mask
? src_mask
[i
] : NULL
,
1878 has_alpha
? NULL
: blend_alpha
,
1880 pad_inline
? 4 : dst_channels
);
1883 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
1885 /* Split the blend rows back to memory rows */
1886 if (dst_count
> src_count
) {
1887 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
1889 if (src_count
== 1) {
1890 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1891 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1893 row_type
.length
/= 2;
1897 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
1898 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
1899 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1900 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1902 row_type
.length
/= 2;
1907 * Store blend result to memory
1910 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
1911 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
1914 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1915 dst
, dst_type
, dst_count
, dst_alignment
);
1919 lp_build_mask_end(&mask_ctx
);
1925 * Generate the runtime callable function for the whole fragment pipeline.
1926 * Note that the function which we generate operates on a block of 16
1927 * pixels at at time. The block contains 2x2 quads. Each quad contains
1931 generate_fragment(struct llvmpipe_context
*lp
,
1932 struct lp_fragment_shader
*shader
,
1933 struct lp_fragment_shader_variant
*variant
,
1934 unsigned partial_mask
)
1936 struct gallivm_state
*gallivm
= variant
->gallivm
;
1937 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
1938 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
1939 char func_name
[256];
1940 struct lp_type fs_type
;
1941 struct lp_type blend_type
;
1942 LLVMTypeRef fs_elem_type
;
1943 LLVMTypeRef blend_vec_type
;
1944 LLVMTypeRef arg_types
[13];
1945 LLVMTypeRef func_type
;
1946 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1947 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1948 LLVMValueRef context_ptr
;
1951 LLVMValueRef a0_ptr
;
1952 LLVMValueRef dadx_ptr
;
1953 LLVMValueRef dady_ptr
;
1954 LLVMValueRef color_ptr_ptr
;
1955 LLVMValueRef stride_ptr
;
1956 LLVMValueRef depth_ptr
;
1957 LLVMValueRef depth_stride
;
1958 LLVMValueRef mask_input
;
1959 LLVMValueRef thread_data_ptr
;
1960 LLVMBasicBlockRef block
;
1961 LLVMBuilderRef builder
;
1962 struct lp_build_sampler_soa
*sampler
;
1963 struct lp_build_interp_soa_context interp
;
1964 LLVMValueRef fs_mask
[16 / 4];
1965 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
1966 LLVMValueRef function
;
1967 LLVMValueRef facing
;
1972 boolean cbuf0_write_all
;
1973 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
1974 util_blend_state_is_dual(&key
->blend
, 0);
1976 assert(lp_native_vector_width
/ 32 >= 4);
1978 /* Adjust color input interpolation according to flatshade state:
1980 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
1981 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1982 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
1984 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1986 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1990 /* check if writes to cbuf[0] are to be copied to all cbufs */
1991 cbuf0_write_all
= FALSE
;
1992 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
1993 if (shader
->info
.base
.properties
[i
].name
==
1994 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
1995 cbuf0_write_all
= TRUE
;
2000 /* TODO: actually pick these based on the fs and color buffer
2001 * characteristics. */
2003 memset(&fs_type
, 0, sizeof fs_type
);
2004 fs_type
.floating
= TRUE
; /* floating point values */
2005 fs_type
.sign
= TRUE
; /* values are signed */
2006 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2007 fs_type
.width
= 32; /* 32-bit float */
2008 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2010 memset(&blend_type
, 0, sizeof blend_type
);
2011 blend_type
.floating
= FALSE
; /* values are integers */
2012 blend_type
.sign
= FALSE
; /* values are unsigned */
2013 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2014 blend_type
.width
= 8; /* 8-bit ubyte values */
2015 blend_type
.length
= 16; /* 16 elements per vector */
2018 * Generate the function prototype. Any change here must be reflected in
2019 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2022 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2024 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2026 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2027 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2029 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2030 arg_types
[1] = int32_type
; /* x */
2031 arg_types
[2] = int32_type
; /* y */
2032 arg_types
[3] = int32_type
; /* facing */
2033 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2034 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2035 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2036 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2037 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2038 arg_types
[9] = int32_type
; /* mask_input */
2039 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2040 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2041 arg_types
[12] = int32_type
; /* depth_stride */
2043 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2044 arg_types
, Elements(arg_types
), 0);
2046 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2047 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2049 variant
->function
[partial_mask
] = function
;
2051 /* XXX: need to propagate noalias down into color param now we are
2052 * passing a pointer-to-pointer?
2054 for(i
= 0; i
< Elements(arg_types
); ++i
)
2055 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2056 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2058 context_ptr
= LLVMGetParam(function
, 0);
2059 x
= LLVMGetParam(function
, 1);
2060 y
= LLVMGetParam(function
, 2);
2061 facing
= LLVMGetParam(function
, 3);
2062 a0_ptr
= LLVMGetParam(function
, 4);
2063 dadx_ptr
= LLVMGetParam(function
, 5);
2064 dady_ptr
= LLVMGetParam(function
, 6);
2065 color_ptr_ptr
= LLVMGetParam(function
, 7);
2066 depth_ptr
= LLVMGetParam(function
, 8);
2067 mask_input
= LLVMGetParam(function
, 9);
2068 thread_data_ptr
= LLVMGetParam(function
, 10);
2069 stride_ptr
= LLVMGetParam(function
, 11);
2070 depth_stride
= LLVMGetParam(function
, 12);
2072 lp_build_name(context_ptr
, "context");
2073 lp_build_name(x
, "x");
2074 lp_build_name(y
, "y");
2075 lp_build_name(a0_ptr
, "a0");
2076 lp_build_name(dadx_ptr
, "dadx");
2077 lp_build_name(dady_ptr
, "dady");
2078 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2079 lp_build_name(depth_ptr
, "depth");
2080 lp_build_name(thread_data_ptr
, "thread_data");
2081 lp_build_name(mask_input
, "mask_input");
2082 lp_build_name(stride_ptr
, "stride_ptr");
2083 lp_build_name(depth_stride
, "depth_stride");
2089 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2090 builder
= gallivm
->builder
;
2092 LLVMPositionBuilderAtEnd(builder
, block
);
2094 /* code generated texture sampling */
2095 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
2097 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2098 /* for 1d resources only run "upper half" of stamp */
2099 if (key
->resource_1d
)
2103 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2104 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2105 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2106 num_loop
, "mask_store");
2107 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2110 * The shader input interpolation info is not explicitely baked in the
2111 * shader key, but everything it derives from (TGSI, and flatshade) is
2112 * already included in the shader key.
2114 lp_build_interp_soa_init(&interp
,
2116 shader
->info
.base
.num_inputs
,
2118 shader
->info
.base
.pixel_center_integer
,
2120 a0_ptr
, dadx_ptr
, dady_ptr
,
2123 for (i
= 0; i
< num_fs
; i
++) {
2125 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2126 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2127 &indexi
, 1, "mask_ptr");
2130 mask
= generate_quad_mask(gallivm
, fs_type
,
2131 i
*fs_type
.length
/4, mask_input
);
2134 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2136 LLVMBuildStore(builder
, mask
, mask_ptr
);
2139 generate_fs_loop(gallivm
,
2147 mask_store
, /* output */
2154 for (i
= 0; i
< num_fs
; i
++) {
2155 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2156 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2158 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2159 /* This is fucked up need to reorganize things */
2160 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2161 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2162 ptr
= LLVMBuildGEP(builder
,
2163 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2165 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2168 if (dual_source_blend
) {
2169 /* only support one dual source blend target hence always use output 1 */
2170 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2171 ptr
= LLVMBuildGEP(builder
,
2172 color_store
[1][chan
],
2174 fs_out_color
[1][chan
][i
] = ptr
;
2180 sampler
->destroy(sampler
);
2182 /* Loop over color outputs / color buffers to do blending.
2184 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2185 LLVMValueRef color_ptr
;
2186 LLVMValueRef stride
;
2187 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2189 boolean do_branch
= ((key
->depth
.enabled
2190 || key
->stencil
[0].enabled
2191 || key
->alpha
.enabled
)
2192 && !shader
->info
.base
.uses_kill
);
2194 color_ptr
= LLVMBuildLoad(builder
,
2195 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2198 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2200 stride
= LLVMBuildLoad(builder
,
2201 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2204 generate_unswizzled_blend(gallivm
, cbuf
, variant
, key
->cbuf_format
[cbuf
],
2205 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2206 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2209 LLVMBuildRetVoid(builder
);
2211 gallivm_verify_function(gallivm
, function
);
2213 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2218 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2222 debug_printf("fs variant %p:\n", (void *) key
);
2224 if (key
->flatshade
) {
2225 debug_printf("flatshade = 1\n");
2227 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2228 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2230 if (key
->depth
.enabled
) {
2231 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2232 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2233 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2236 for (i
= 0; i
< 2; ++i
) {
2237 if (key
->stencil
[i
].enabled
) {
2238 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2239 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2240 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2241 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2242 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2243 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2247 if (key
->alpha
.enabled
) {
2248 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2251 if (key
->occlusion_count
) {
2252 debug_printf("occlusion_count = 1\n");
2255 if (key
->blend
.logicop_enable
) {
2256 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2258 else if (key
->blend
.rt
[0].blend_enable
) {
2259 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2260 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2261 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2262 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2263 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2264 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2266 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2267 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2268 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2269 debug_printf("sampler[%u] = \n", i
);
2270 debug_printf(" .wrap = %s %s %s\n",
2271 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2272 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2273 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2274 debug_printf(" .min_img_filter = %s\n",
2275 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2276 debug_printf(" .min_mip_filter = %s\n",
2277 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2278 debug_printf(" .mag_img_filter = %s\n",
2279 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2280 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2281 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2282 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2283 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2284 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2285 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2286 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2288 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2289 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2290 debug_printf("texture[%u] = \n", i
);
2291 debug_printf(" .format = %s\n",
2292 util_format_name(texture
->format
));
2293 debug_printf(" .target = %s\n",
2294 util_dump_tex_target(texture
->target
, TRUE
));
2295 debug_printf(" .level_zero_only = %u\n",
2296 texture
->level_zero_only
);
2297 debug_printf(" .pot = %u %u %u\n",
2299 texture
->pot_height
,
2300 texture
->pot_depth
);
2306 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2308 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2309 variant
->shader
->no
, variant
->no
);
2310 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2311 dump_fs_variant_key(&variant
->key
);
2312 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2318 * Generate a new fragment shader variant from the shader code and
2319 * other state indicated by the key.
2321 static struct lp_fragment_shader_variant
*
2322 generate_variant(struct llvmpipe_context
*lp
,
2323 struct lp_fragment_shader
*shader
,
2324 const struct lp_fragment_shader_variant_key
*key
)
2326 struct lp_fragment_shader_variant
*variant
;
2327 const struct util_format_description
*cbuf0_format_desc
;
2328 boolean fullcolormask
;
2330 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2334 variant
->gallivm
= gallivm_create();
2335 if (!variant
->gallivm
) {
2340 variant
->shader
= shader
;
2341 variant
->list_item_global
.base
= variant
;
2342 variant
->list_item_local
.base
= variant
;
2343 variant
->no
= shader
->variants_created
++;
2345 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2348 * Determine whether we are touching all channels in the color buffer.
2350 fullcolormask
= FALSE
;
2351 if (key
->nr_cbufs
== 1) {
2352 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2353 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2357 !key
->blend
.logicop_enable
&&
2358 !key
->blend
.rt
[0].blend_enable
&&
2360 !key
->stencil
[0].enabled
&&
2361 !key
->alpha
.enabled
&&
2362 !key
->depth
.enabled
&&
2363 !shader
->info
.base
.uses_kill
2366 if ((!shader
|| shader
->info
.base
.num_tokens
<= 1) &&
2367 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2368 variant
->ps_inv_multiplier
= 0;
2370 variant
->ps_inv_multiplier
= 1;
2373 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2374 lp_debug_fs_variant(variant
);
2377 lp_jit_init_types(variant
);
2379 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2380 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2382 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2383 if (variant
->opaque
) {
2384 /* Specialized shader, which doesn't need to read the color buffer. */
2385 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2390 * Compile everything
2393 gallivm_compile_module(variant
->gallivm
);
2395 if (variant
->function
[RAST_EDGE_TEST
]) {
2396 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2397 gallivm_jit_function(variant
->gallivm
,
2398 variant
->function
[RAST_EDGE_TEST
]);
2401 if (variant
->function
[RAST_WHOLE
]) {
2402 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2403 gallivm_jit_function(variant
->gallivm
,
2404 variant
->function
[RAST_WHOLE
]);
2405 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2406 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2414 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2415 const struct pipe_shader_state
*templ
)
2417 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2418 struct lp_fragment_shader
*shader
;
2420 int nr_sampler_views
;
2423 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2427 shader
->no
= fs_no
++;
2428 make_empty_list(&shader
->variants
);
2430 /* get/save the summary info for this shader */
2431 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2433 /* we need to keep a local copy of the tokens */
2434 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2436 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2437 if (shader
->draw_data
== NULL
) {
2438 FREE((void *) shader
->base
.tokens
);
2443 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2444 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2446 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2447 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2449 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2450 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2451 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2453 switch (shader
->info
.base
.input_interpolate
[i
]) {
2454 case TGSI_INTERPOLATE_CONSTANT
:
2455 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2457 case TGSI_INTERPOLATE_LINEAR
:
2458 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2460 case TGSI_INTERPOLATE_PERSPECTIVE
:
2461 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2463 case TGSI_INTERPOLATE_COLOR
:
2464 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2471 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2472 case TGSI_SEMANTIC_FACE
:
2473 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2475 case TGSI_SEMANTIC_POSITION
:
2476 /* Position was already emitted above
2478 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2479 shader
->inputs
[i
].src_index
= 0;
2483 shader
->inputs
[i
].src_index
= i
+1;
2486 if (LP_DEBUG
& DEBUG_TGSI
) {
2488 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2489 shader
->no
, (void *) shader
);
2490 tgsi_dump(templ
->tokens
, 0);
2491 debug_printf("usage masks:\n");
2492 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2493 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2494 debug_printf(" IN[%u].%s%s%s%s\n",
2496 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2497 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2498 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2499 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2509 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2511 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2513 if (llvmpipe
->fs
== fs
)
2516 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2518 draw_bind_fragment_shader(llvmpipe
->draw
,
2519 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2521 llvmpipe
->dirty
|= LP_NEW_FS
;
2526 * Remove shader variant from two lists: the shader's variant list
2527 * and the context's variant list.
2530 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2531 struct lp_fragment_shader_variant
*variant
)
2535 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2536 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2537 " #%u v total cached #%u\n",
2538 variant
->shader
->no
,
2540 variant
->shader
->variants_created
,
2541 variant
->shader
->variants_cached
,
2542 lp
->nr_fs_variants
);
2545 /* free all the variant's JIT'd functions */
2546 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2547 if (variant
->function
[i
]) {
2548 gallivm_free_function(variant
->gallivm
,
2549 variant
->function
[i
],
2550 variant
->jit_function
[i
]);
2554 gallivm_destroy(variant
->gallivm
);
2556 /* remove from shader's list */
2557 remove_from_list(&variant
->list_item_local
);
2558 variant
->shader
->variants_cached
--;
2560 /* remove from context's list */
2561 remove_from_list(&variant
->list_item_global
);
2562 lp
->nr_fs_variants
--;
2563 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2570 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2572 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2573 struct lp_fragment_shader
*shader
= fs
;
2574 struct lp_fs_variant_list_item
*li
;
2576 assert(fs
!= llvmpipe
->fs
);
2579 * XXX: we need to flush the context until we have some sort of reference
2580 * counting in fragment shaders as they may still be binned
2581 * Flushing alone might not sufficient we need to wait on it too.
2583 llvmpipe_finish(pipe
, __FUNCTION__
);
2585 /* Delete all the variants */
2586 li
= first_elem(&shader
->variants
);
2587 while(!at_end(&shader
->variants
, li
)) {
2588 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2589 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2593 /* Delete draw module's data */
2594 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2596 assert(shader
->variants_cached
== 0);
2597 FREE((void *) shader
->base
.tokens
);
2604 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2605 uint shader
, uint index
,
2606 struct pipe_constant_buffer
*cb
)
2608 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2609 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2611 assert(shader
< PIPE_SHADER_TYPES
);
2612 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2614 /* note: reference counting */
2615 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2617 if (shader
== PIPE_SHADER_VERTEX
||
2618 shader
== PIPE_SHADER_GEOMETRY
) {
2619 /* Pass the constants to the 'draw' module */
2620 const unsigned size
= cb
? cb
->buffer_size
: 0;
2624 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2626 else if (cb
&& cb
->user_buffer
) {
2627 data
= (ubyte
*) cb
->user_buffer
;
2634 data
+= cb
->buffer_offset
;
2636 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2640 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2642 if (cb
&& cb
->user_buffer
) {
2643 pipe_resource_reference(&constants
, NULL
);
2649 * Return the blend factor equivalent to a destination alpha of one.
2651 static INLINE
unsigned
2652 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2655 case PIPE_BLENDFACTOR_DST_ALPHA
:
2656 return PIPE_BLENDFACTOR_ONE
;
2657 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2658 return PIPE_BLENDFACTOR_ZERO
;
2659 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2661 return PIPE_BLENDFACTOR_ZERO
;
2663 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2671 * We need to generate several variants of the fragment pipeline to match
2672 * all the combinations of the contributing state atoms.
2674 * TODO: there is actually no reason to tie this to context state -- the
2675 * generated code could be cached globally in the screen.
2678 make_variant_key(struct llvmpipe_context
*lp
,
2679 struct lp_fragment_shader
*shader
,
2680 struct lp_fragment_shader_variant_key
*key
)
2684 memset(key
, 0, shader
->variant_key_size
);
2686 if (lp
->framebuffer
.zsbuf
) {
2687 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2688 const struct util_format_description
*zsbuf_desc
=
2689 util_format_description(zsbuf_format
);
2691 if (lp
->depth_stencil
->depth
.enabled
&&
2692 util_format_has_depth(zsbuf_desc
)) {
2693 key
->zsbuf_format
= zsbuf_format
;
2694 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2696 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2697 util_format_has_stencil(zsbuf_desc
)) {
2698 key
->zsbuf_format
= zsbuf_format
;
2699 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2701 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2702 key
->resource_1d
= TRUE
;
2706 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2707 if (!lp
->framebuffer
.nr_cbufs
||
2708 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2709 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2711 if(key
->alpha
.enabled
)
2712 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2713 /* alpha.ref_value is passed in jit_context */
2715 key
->flatshade
= lp
->rasterizer
->flatshade
;
2716 if (lp
->active_occlusion_queries
) {
2717 key
->occlusion_count
= TRUE
;
2720 if (lp
->framebuffer
.nr_cbufs
) {
2721 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2724 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2726 if (!key
->blend
.independent_blend_enable
) {
2727 /* we always need independent blend otherwise the fixups below won't work */
2728 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2729 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2731 key
->blend
.independent_blend_enable
= 1;
2734 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2735 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2736 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2737 const struct util_format_description
*format_desc
;
2739 key
->cbuf_format
[i
] = format
;
2742 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2743 * mixing of 2d textures with height 1 and 1d textures, so make sure
2744 * we pick 1d if any cbuf or zsbuf is 1d.
2746 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[0]->texture
)) {
2747 key
->resource_1d
= TRUE
;
2750 format_desc
= util_format_description(format
);
2751 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2752 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2755 * Mask out color channels not present in the color buffer.
2757 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2760 * Disable blend for integer formats.
2762 if (util_format_is_pure_integer(format
)) {
2763 blend_rt
->blend_enable
= 0;
2767 * Our swizzled render tiles always have an alpha channel, but the linear
2768 * render target format often does not, so force here the dst alpha to be
2771 * This is not a mere optimization. Wrong results will be produced if the
2772 * dst alpha is used, the dst format does not have alpha, and the previous
2773 * rendering was not flushed from the swizzled to linear buffer. For
2774 * example, NonPowTwo DCT.
2776 * TODO: This should be generalized to all channels for better
2777 * performance, but only alpha causes correctness issues.
2779 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2781 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2782 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2783 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
2784 boolean clamped_zero
= !util_format_is_float(format
) &&
2785 !util_format_is_snorm(format
);
2786 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
,
2788 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
,
2790 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2791 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2792 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2796 /* This value will be the same for all the variants of a given shader:
2798 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2800 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2801 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2802 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
2803 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
2808 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2809 * are dx10-style? Can't really have mixed opcodes, at least not
2810 * if we want to skip the holes here (without rescanning tgsi).
2812 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2813 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2814 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2815 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
2816 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2817 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2822 key
->nr_sampler_views
= key
->nr_samplers
;
2823 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2824 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2825 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2826 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2835 * Update fragment shader state. This is called just prior to drawing
2836 * something when some fragment-related state has changed.
2839 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
2841 struct lp_fragment_shader
*shader
= lp
->fs
;
2842 struct lp_fragment_shader_variant_key key
;
2843 struct lp_fragment_shader_variant
*variant
= NULL
;
2844 struct lp_fs_variant_list_item
*li
;
2846 make_variant_key(lp
, shader
, &key
);
2848 /* Search the variants for one which matches the key */
2849 li
= first_elem(&shader
->variants
);
2850 while(!at_end(&shader
->variants
, li
)) {
2851 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
2859 /* Move this variant to the head of the list to implement LRU
2860 * deletion of shader's when we have too many.
2862 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2865 /* variant not found, create it now */
2868 unsigned variants_to_cull
;
2871 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2874 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
2877 /* First, check if we've exceeded the max number of shader variants.
2878 * If so, free 25% of them (the least recently used ones).
2880 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
2882 if (variants_to_cull
||
2883 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
2884 struct pipe_context
*pipe
= &lp
->pipe
;
2887 * XXX: we need to flush the context until we have some sort of
2888 * reference counting in fragment shaders as they may still be binned
2889 * Flushing alone might not be sufficient we need to wait on it too.
2891 llvmpipe_finish(pipe
, __FUNCTION__
);
2894 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2895 * number of shader variants (potentially all of them) could be
2896 * pending for destruction on flush.
2899 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
2900 struct lp_fs_variant_list_item
*item
;
2901 if (is_empty_list(&lp
->fs_variants_list
)) {
2904 item
= last_elem(&lp
->fs_variants_list
);
2907 llvmpipe_remove_shader_variant(lp
, item
->base
);
2912 * Generate the new variant.
2915 variant
= generate_variant(lp
, shader
, &key
);
2918 LP_COUNT_ADD(llvm_compile_time
, dt
);
2919 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
2921 llvmpipe_variant_count
++;
2923 /* Put the new variant into the list */
2925 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
2926 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2927 lp
->nr_fs_variants
++;
2928 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
2929 shader
->variants_cached
++;
2933 /* Bind this variant */
2934 lp_setup_set_fs_variant(lp
->setup
, variant
);
2942 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
2944 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
2945 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
2946 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
2948 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
2952 * Rasterization is disabled if there is no pixel shader and
2953 * both depth and stencil testing are disabled:
2954 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
2957 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
2959 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
2962 !lp
->depth_stencil
->depth
.enabled
&&
2963 !lp
->depth_stencil
->stencil
[0].enabled
);