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"
105 /** Fragment shader number (for debugging) */
106 static unsigned fs_no
= 0;
110 * Expand the relevant bits of mask_input to a n*4-dword mask for the
111 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
112 * quad mask vector to 0 or ~0.
113 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
114 * quad arguments with fs length 8.
116 * \param first_quad which quad(s) of the quad group to test, in [0,3]
117 * \param mask_input bitwise mask for the whole 4x4 stamp
120 generate_quad_mask(struct gallivm_state
*gallivm
,
121 struct lp_type fs_type
,
123 LLVMValueRef mask_input
) /* int32 */
125 LLVMBuilderRef builder
= gallivm
->builder
;
126 struct lp_type mask_type
;
127 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
128 LLVMValueRef bits
[16];
133 * XXX: We'll need a different path for 16 x u8
135 assert(fs_type
.width
== 32);
136 assert(fs_type
.length
<= Elements(bits
));
137 mask_type
= lp_int_type(fs_type
);
140 * mask_input >>= (quad * 4)
142 switch (first_quad
) {
147 assert(fs_type
.length
== 4);
154 assert(fs_type
.length
== 4);
162 mask_input
= LLVMBuildLShr(builder
,
164 LLVMConstInt(i32t
, shift
, 0),
168 * mask = { mask_input & (1 << i), for i in [0,3] }
170 mask
= lp_build_broadcast(gallivm
,
171 lp_build_vec_type(gallivm
, mask_type
),
174 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
175 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
176 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
177 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
178 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
179 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
181 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
184 * mask = mask != 0 ? ~0 : 0
186 mask
= lp_build_compare(gallivm
,
187 mask_type
, PIPE_FUNC_NOTEQUAL
,
189 lp_build_const_int_vec(gallivm
, mask_type
, 0));
195 #define EARLY_DEPTH_TEST 0x1
196 #define LATE_DEPTH_TEST 0x2
197 #define EARLY_DEPTH_WRITE 0x4
198 #define LATE_DEPTH_WRITE 0x8
201 find_output_by_semantic( const struct tgsi_shader_info
*info
,
207 for (i
= 0; i
< info
->num_outputs
; i
++)
208 if (info
->output_semantic_name
[i
] == semantic
&&
209 info
->output_semantic_index
[i
] == index
)
217 * Generate the fragment shader, depth/stencil test, and alpha tests.
220 generate_fs_loop(struct gallivm_state
*gallivm
,
221 struct lp_fragment_shader
*shader
,
222 const struct lp_fragment_shader_variant_key
*key
,
223 LLVMBuilderRef builder
,
225 LLVMValueRef context_ptr
,
226 LLVMValueRef num_loop
,
227 struct lp_build_interp_soa_context
*interp
,
228 struct lp_build_sampler_soa
*sampler
,
229 LLVMValueRef mask_store
,
230 LLVMValueRef (*out_color
)[4],
231 LLVMValueRef depth_ptr
,
234 LLVMValueRef thread_data_ptr
)
236 const struct util_format_description
*zs_format_desc
= NULL
;
237 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
238 LLVMTypeRef vec_type
;
239 LLVMValueRef mask_ptr
, mask_val
;
240 LLVMValueRef consts_ptr
;
242 LLVMValueRef zs_value
= NULL
;
243 LLVMValueRef stencil_refs
[2];
244 LLVMValueRef depth_ptr_i
;
245 LLVMValueRef depth_offset
;
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
;
249 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
250 shader
->info
.base
.num_inputs
< 3 &&
251 shader
->info
.base
.num_instructions
< 8);
252 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
253 util_blend_state_is_dual(&key
->blend
, 0);
259 struct lp_bld_tgsi_system_values system_values
;
261 memset(&system_values
, 0, sizeof(system_values
));
263 if (key
->depth
.enabled
||
264 key
->stencil
[0].enabled
||
265 key
->stencil
[1].enabled
) {
267 zs_format_desc
= util_format_description(key
->zsbuf_format
);
268 assert(zs_format_desc
);
270 if (!shader
->info
.base
.writes_z
) {
271 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
)
272 /* With alpha test and kill, can do the depth test early
273 * and hopefully eliminate some quads. But need to do a
274 * special deferred depth write once the final mask value
277 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
279 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
282 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
285 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
286 !(key
->stencil
[0].enabled
&& key
->stencil
[0].writemask
))
287 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
294 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
295 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
297 vec_type
= lp_build_vec_type(gallivm
, type
);
299 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
301 lp_build_for_loop_begin(&loop_state
, gallivm
,
302 lp_build_const_int32(gallivm
, 0),
305 lp_build_const_int32(gallivm
, 1));
307 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
308 &loop_state
.counter
, 1, "mask_ptr");
309 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
311 depth_offset
= LLVMBuildMul(builder
, loop_state
.counter
,
312 lp_build_const_int32(gallivm
, depth_bits
* type
.length
),
315 depth_ptr_i
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset
, 1, "");
317 memset(outputs
, 0, sizeof outputs
);
319 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
320 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
321 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
322 lp_build_vec_type(gallivm
,
327 if (dual_source_blend
) {
328 assert(key
->nr_cbufs
<= 1);
329 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
330 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
331 lp_build_vec_type(gallivm
,
338 /* 'mask' will control execution based on quad's pixel alive/killed state */
339 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
341 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
342 lp_build_mask_check(&mask
);
344 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
347 if (depth_mode
& EARLY_DEPTH_TEST
) {
348 lp_build_depth_stencil_test(gallivm
,
360 if (depth_mode
& EARLY_DEPTH_WRITE
) {
361 lp_build_depth_write(gallivm
, type
, zs_format_desc
, depth_ptr_i
, zs_value
);
365 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
367 /* Build the actual shader */
368 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
369 consts_ptr
, &system_values
,
371 outputs
, sampler
, &shader
->info
.base
, NULL
);
374 if (key
->alpha
.enabled
) {
375 int color0
= find_output_by_semantic(&shader
->info
.base
,
379 if (color0
!= -1 && outputs
[color0
][3]) {
380 const struct util_format_description
*cbuf_format_desc
;
381 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
382 LLVMValueRef alpha_ref_value
;
384 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
385 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
387 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
389 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
390 &mask
, alpha
, alpha_ref_value
,
391 (depth_mode
& LATE_DEPTH_TEST
) != 0);
396 if (depth_mode
& LATE_DEPTH_TEST
) {
397 int pos0
= find_output_by_semantic(&shader
->info
.base
,
398 TGSI_SEMANTIC_POSITION
,
401 if (pos0
!= -1 && outputs
[pos0
][2]) {
402 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
405 lp_build_depth_stencil_test(gallivm
,
417 if (depth_mode
& LATE_DEPTH_WRITE
) {
418 lp_build_depth_write(gallivm
, type
, zs_format_desc
, depth_ptr_i
, zs_value
);
421 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
422 (depth_mode
& LATE_DEPTH_WRITE
))
424 /* Need to apply a reduced mask to the depth write. Reload the
425 * depth value, update from zs_value with the new mask value and
428 lp_build_deferred_depth_write(gallivm
,
438 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
440 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
441 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
442 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
444 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
445 if(outputs
[attrib
][chan
]) {
446 /* XXX: just initialize outputs to point at colors[] and
449 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
450 LLVMValueRef color_ptr
;
451 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
452 &loop_state
.counter
, 1, "");
453 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
454 LLVMBuildStore(builder
, out
, color_ptr
);
460 if (key
->occlusion_count
) {
461 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
462 lp_build_name(counter
, "counter");
463 lp_build_occlusion_count(gallivm
, type
,
464 lp_build_mask_value(&mask
), counter
);
467 mask_val
= lp_build_mask_end(&mask
);
468 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
469 lp_build_for_loop_end(&loop_state
);
474 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
476 * Fragment Shader outputs pixels in small 2x2 blocks
477 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
479 * However in memory pixels are stored in rows
480 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
482 * @param type fragment shader type (4x or 8x float)
483 * @param num_fs number of fs_src
484 * @param dst_channels number of output channels
485 * @param fs_src output from fragment shader
486 * @param dst pointer to store result
487 * @param pad_inline is channel padding inline or at end of row
488 * @return the number of dsts
491 generate_fs_twiddle(struct gallivm_state
*gallivm
,
494 unsigned dst_channels
,
495 LLVMValueRef fs_src
[][4],
499 LLVMValueRef src
[16];
505 unsigned pixels
= num_fs
== 4 ? 1 : 2;
506 unsigned reorder_group
;
507 unsigned src_channels
;
511 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
512 src_count
= num_fs
* src_channels
;
514 assert(pixels
== 2 || num_fs
== 4);
515 assert(num_fs
* src_channels
<= Elements(src
));
518 * Transpose from SoA -> AoS
520 for (i
= 0; i
< num_fs
; ++i
) {
521 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
525 * Pick transformation options
532 if (dst_channels
== 1) {
538 } else if (dst_channels
== 2) {
542 } else if (dst_channels
> 2) {
549 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
555 * Split the src in half
558 for (i
= num_fs
; i
> 0; --i
) {
559 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
560 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
568 * Ensure pixels are in memory order
571 /* Twiddle pixels by reordering the array, e.g.:
573 * src_count = 8 -> 0 2 1 3 4 6 5 7
574 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
576 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
578 for (i
= 0; i
< src_count
; ++i
) {
579 unsigned group
= i
/ reorder_group
;
580 unsigned block
= (group
/ 4) * 4 * reorder_group
;
581 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
584 } else if (twiddle
) {
585 /* Twiddle pixels across elements of array */
586 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
589 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
593 * Moves any padding between pixels to the end
594 * e.g. RGBXRGBX -> RGBRGBXX
597 unsigned char swizzles
[16];
598 unsigned elems
= pixels
* dst_channels
;
600 for (i
= 0; i
< type
.length
; ++i
) {
602 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
604 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
607 for (i
= 0; i
< src_count
; ++i
) {
608 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
617 * Load an unswizzled block of pixels from memory
620 load_unswizzled_block(struct gallivm_state
*gallivm
,
621 LLVMValueRef base_ptr
,
623 unsigned block_width
,
624 unsigned block_height
,
626 struct lp_type dst_type
,
628 unsigned dst_alignment
)
630 LLVMBuilderRef builder
= gallivm
->builder
;
631 unsigned row_size
= dst_count
/ block_height
;
634 /* Ensure block exactly fits into dst */
635 assert((block_width
* block_height
) % dst_count
== 0);
637 for (i
= 0; i
< dst_count
; ++i
) {
638 unsigned x
= i
% row_size
;
639 unsigned y
= i
/ row_size
;
641 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
642 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
645 LLVMValueRef dst_ptr
;
647 gep
[0] = lp_build_const_int32(gallivm
, 0);
648 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
650 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
651 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
653 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
655 lp_set_load_alignment(dst
[i
], dst_alignment
);
661 * Store an unswizzled block of pixels to memory
664 store_unswizzled_block(struct gallivm_state
*gallivm
,
665 LLVMValueRef base_ptr
,
667 unsigned block_width
,
668 unsigned block_height
,
670 struct lp_type src_type
,
672 unsigned src_alignment
)
674 LLVMBuilderRef builder
= gallivm
->builder
;
675 unsigned row_size
= src_count
/ block_height
;
678 /* Ensure src exactly fits into block */
679 assert((block_width
* block_height
) % src_count
== 0);
681 for (i
= 0; i
< src_count
; ++i
) {
682 unsigned x
= i
% row_size
;
683 unsigned y
= i
/ row_size
;
685 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
686 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
689 LLVMValueRef src_ptr
;
691 gep
[0] = lp_build_const_int32(gallivm
, 0);
692 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
694 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
695 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
697 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
699 lp_set_store_alignment(src_ptr
, src_alignment
);
705 * Checks if a format description is an arithmetic format
707 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
709 static INLINE boolean
710 is_arithmetic_format(const struct util_format_description
*format_desc
)
712 boolean arith
= false;
715 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
716 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
717 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
725 * Retrieves the type representing the memory layout for a format
727 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
730 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
731 struct lp_type
* type
)
736 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
737 /* just make this a 32bit uint */
738 type
->floating
= false;
747 for (i
= 0; i
< 4; i
++)
748 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
752 memset(type
, 0, sizeof(struct lp_type
));
753 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
754 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
755 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
756 type
->norm
= format_desc
->channel
[chan
].normalized
;
758 if (is_arithmetic_format(format_desc
)) {
762 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
763 type
->width
+= format_desc
->channel
[i
].size
;
766 type
->width
= format_desc
->channel
[chan
].size
;
767 type
->length
= format_desc
->nr_channels
;
773 * Retrieves the type for a format which is usable in the blending code.
775 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
778 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
779 struct lp_type
* type
)
784 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
785 /* always use ordinary floats for blending */
786 type
->floating
= true;
795 for (i
= 0; i
< 4; i
++)
796 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
800 memset(type
, 0, sizeof(struct lp_type
));
801 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
802 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
803 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
804 type
->norm
= format_desc
->channel
[chan
].normalized
;
805 type
->width
= format_desc
->channel
[chan
].size
;
806 type
->length
= format_desc
->nr_channels
;
808 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
809 if (format_desc
->channel
[i
].size
> type
->width
)
810 type
->width
= format_desc
->channel
[i
].size
;
813 if (type
->floating
) {
816 if (type
->width
<= 8) {
818 } else if (type
->width
<= 16) {
825 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
832 * Scale a normalized value from src_bits to dst_bits
834 static INLINE LLVMValueRef
835 scale_bits(struct gallivm_state
*gallivm
,
839 struct lp_type src_type
)
841 LLVMBuilderRef builder
= gallivm
->builder
;
842 LLVMValueRef result
= src
;
844 if (dst_bits
< src_bits
) {
845 /* Scale down by LShr */
846 result
= LLVMBuildLShr(builder
,
848 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- dst_bits
),
850 } else if (dst_bits
> src_bits
) {
852 int db
= dst_bits
- src_bits
;
854 /* Shift left by difference in bits */
855 result
= LLVMBuildShl(builder
,
857 lp_build_const_int_vec(gallivm
, src_type
, db
),
861 /* Enough bits in src to fill the remainder */
862 LLVMValueRef lower
= LLVMBuildLShr(builder
,
864 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
867 result
= LLVMBuildOr(builder
, result
, lower
, "");
868 } else if (db
> src_bits
) {
869 /* Need to repeatedly copy src bits to fill remainder in dst */
872 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
873 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
875 result
= LLVMBuildOr(builder
,
877 LLVMBuildLShr(builder
, result
, shuv
, ""),
888 * Convert from memory format to blending format
890 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
893 convert_to_blend_type(struct gallivm_state
*gallivm
,
894 const struct util_format_description
*src_fmt
,
895 struct lp_type src_type
,
896 struct lp_type dst_type
,
897 LLVMValueRef
* src
, // and dst
900 LLVMValueRef
*dst
= src
;
901 LLVMBuilderRef builder
= gallivm
->builder
;
902 struct lp_type blend_type
;
903 struct lp_type mem_type
;
905 unsigned pixels
= 16 / num_srcs
;
909 * full custom path for packed floats - none of the later functions would do
910 * anything useful, and given the lp_type representation they can't be fixed.
912 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
913 LLVMValueRef tmpsrc
[4];
915 * This is pretty suboptimal for this case blending in SoA would be much
916 * better, since conversion gets us SoA values so need to convert back.
918 assert(src_type
.width
== 32);
919 assert(dst_type
.floating
);
920 assert(dst_type
.width
== 32);
921 assert(dst_type
.length
% 4 == 0);
922 for (i
= 0; i
< 4; i
++) {
925 for (i
= 0; i
< num_srcs
/ 4; i
++) {
926 LLVMValueRef tmpsoa
[4];
927 LLVMValueRef tmps
= tmpsrc
[i
];
929 LLVMValueRef shuffles
[8];
931 /* fetch was 4 values but need 8-wide output values */
932 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
934 * for 8-wide aos transpose would give us wrong order not matching
935 * incoming converted fs values and mask. ARGH.
937 for (j
= 0; j
< 4; j
++) {
938 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
939 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
941 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
942 LLVMConstVector(shuffles
, 8), "");
944 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
945 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
950 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
951 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
953 /* Is the format arithmetic */
954 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
955 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
957 /* Pad if necessary */
958 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
959 for (i
= 0; i
< num_srcs
; ++i
) {
960 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
963 src_type
.length
= dst_type
.length
;
966 /* Special case for half-floats */
967 if (mem_type
.width
== 16 && mem_type
.floating
) {
968 assert(blend_type
.width
== 32 && blend_type
.floating
);
969 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
977 src_type
.width
= blend_type
.width
* blend_type
.length
;
978 blend_type
.length
*= pixels
;
979 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
981 for (i
= 0; i
< num_srcs
; ++i
) {
982 LLVMValueRef chans
[4];
983 LLVMValueRef res
= NULL
;
986 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
988 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
991 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
995 /* Extract bits from source */
996 chans
[j
] = LLVMBuildLShr(builder
,
998 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1001 chans
[j
] = LLVMBuildAnd(builder
,
1003 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1007 if (src_type
.norm
) {
1008 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1009 blend_type
.width
, chans
[j
], src_type
);
1012 /* Insert bits into correct position */
1013 chans
[j
] = LLVMBuildShl(builder
,
1015 lp_build_const_int_vec(gallivm
, src_type
, j
* blend_type
.width
),
1018 sa
+= src_fmt
->channel
[j
].size
;
1023 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1027 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1033 * Convert from blending format to memory format
1035 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1038 convert_from_blend_type(struct gallivm_state
*gallivm
,
1039 const struct util_format_description
*src_fmt
,
1040 struct lp_type src_type
,
1041 struct lp_type dst_type
,
1042 LLVMValueRef
* src
, // and dst
1045 LLVMValueRef
* dst
= src
;
1047 struct lp_type mem_type
;
1048 struct lp_type blend_type
;
1049 LLVMBuilderRef builder
= gallivm
->builder
;
1050 unsigned pixels
= 16 / num_srcs
;
1054 * full custom path for packed floats - none of the later functions would do
1055 * anything useful, and given the lp_type representation they can't be fixed.
1057 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1059 * This is pretty suboptimal for this case blending in SoA would be much
1060 * better - we need to transpose the AoS values back to SoA values for
1061 * conversion/packing.
1063 assert(src_type
.floating
);
1064 assert(src_type
.width
== 32);
1065 assert(src_type
.length
% 4 == 0);
1066 assert(dst_type
.width
== 32);
1067 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1068 LLVMValueRef tmpsoa
[4], tmpdst
;
1069 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1070 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1071 if (num_srcs
== 8) {
1072 LLVMValueRef tmpaos
, shuffles
[8];
1075 * for 8-wide aos transpose has given us wrong order not matching
1076 * output order. HMPF. Also need to split the output values manually.
1078 for (j
= 0; j
< 4; j
++) {
1079 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1080 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1082 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1083 LLVMConstVector(shuffles
, 8), "");
1084 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1085 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1094 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1095 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1097 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1099 /* Special case for half-floats */
1100 if (mem_type
.width
== 16 && mem_type
.floating
) {
1101 int length
= dst_type
.length
;
1102 assert(blend_type
.width
== 32 && blend_type
.floating
);
1104 dst_type
.length
= src_type
.length
;
1106 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1108 dst_type
.length
= length
;
1112 /* Remove any padding */
1113 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1114 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1116 for (i
= 0; i
< num_srcs
; ++i
) {
1117 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1121 /* No bit arithmetic to do */
1126 src_type
.length
= pixels
;
1127 src_type
.width
= blend_type
.length
* blend_type
.width
;
1128 dst_type
.length
= pixels
;
1130 for (i
= 0; i
< num_srcs
; ++i
) {
1131 LLVMValueRef chans
[4];
1132 LLVMValueRef res
= NULL
;
1135 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1137 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1140 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1142 for (k
= 0; k
< blend_type
.width
; ++k
) {
1147 chans
[j
] = LLVMBuildLShr(builder
,
1149 lp_build_const_int_vec(gallivm
, src_type
, j
* blend_type
.width
),
1152 chans
[j
] = LLVMBuildAnd(builder
,
1154 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1157 /* Scale down bits */
1158 if (src_type
.norm
) {
1159 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1160 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1164 chans
[j
] = LLVMBuildShl(builder
,
1166 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1169 sa
+= src_fmt
->channel
[j
].size
;
1174 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1178 assert (dst_type
.width
!= 24);
1180 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1186 * Convert alpha to same blend type as src
1189 convert_alpha(struct gallivm_state
*gallivm
,
1190 struct lp_type row_type
,
1191 struct lp_type alpha_type
,
1192 const unsigned block_size
,
1193 const unsigned block_height
,
1194 const unsigned src_count
,
1195 const unsigned dst_channels
,
1196 const bool pad_inline
,
1197 LLVMValueRef
* src_alpha
)
1199 LLVMBuilderRef builder
= gallivm
->builder
;
1201 unsigned length
= row_type
.length
;
1202 row_type
.length
= alpha_type
.length
;
1204 /* Twiddle the alpha to match pixels */
1205 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, 4, src_alpha
);
1207 for (i
= 0; i
< 4; ++i
) {
1208 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1211 alpha_type
= row_type
;
1212 row_type
.length
= length
;
1214 /* If only one channel we can only need the single alpha value per pixel */
1215 if (src_count
== 1) {
1216 assert(dst_channels
== 1);
1218 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, 4, src_alpha
, src_count
);
1220 /* If there are more srcs than rows then we need to split alpha up */
1221 if (src_count
> block_height
) {
1222 for (i
= src_count
; i
> 0; --i
) {
1223 unsigned pixels
= block_size
/ src_count
;
1224 unsigned idx
= i
- 1;
1226 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4], (idx
* pixels
) % 4, pixels
);
1230 /* If there is a src for each pixel broadcast the alpha across whole row */
1231 if (src_count
== block_size
) {
1232 for (i
= 0; i
< src_count
; ++i
) {
1233 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1236 unsigned pixels
= block_size
/ src_count
;
1237 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1238 unsigned alpha_span
= 1;
1239 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1241 /* Check if we need 2 src_alphas for our shuffles */
1242 if (pixels
> alpha_type
.length
) {
1246 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1247 for (j
= 0; j
< row_type
.length
; ++j
) {
1248 if (j
< pixels
* channels
) {
1249 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1251 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1255 for (i
= 0; i
< src_count
; ++i
) {
1256 unsigned idx1
= i
, idx2
= i
;
1258 if (alpha_span
> 1){
1263 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1266 LLVMConstVector(shuffles
, row_type
.length
),
1275 * Generates the blend function for unswizzled colour buffers
1276 * Also generates the read & write from colour buffer
1279 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1281 struct lp_fragment_shader_variant
*variant
,
1282 enum pipe_format out_format
,
1283 unsigned int num_fs
,
1284 struct lp_type fs_type
,
1285 LLVMValueRef
* fs_mask
,
1286 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1287 LLVMValueRef context_ptr
,
1288 LLVMValueRef color_ptr
,
1289 LLVMValueRef stride
,
1290 unsigned partial_mask
,
1293 const unsigned alpha_channel
= 3;
1294 const unsigned block_width
= 4;
1295 const unsigned block_height
= 4;
1296 const unsigned block_size
= block_width
* block_height
;
1297 const unsigned lp_integer_vector_width
= 128;
1299 LLVMBuilderRef builder
= gallivm
->builder
;
1300 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1301 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1302 LLVMValueRef src_alpha
[4 * 4];
1303 LLVMValueRef src1_alpha
[4 * 4];
1304 LLVMValueRef src_mask
[4 * 4];
1305 LLVMValueRef src
[4 * 4];
1306 LLVMValueRef src1
[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 row_type
;
1317 struct lp_type dst_type
;
1319 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1320 unsigned vector_width
;
1321 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1322 unsigned dst_channels
;
1327 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1329 unsigned dst_alignment
;
1331 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1332 bool has_alpha
= false;
1333 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1334 util_blend_state_is_dual(&variant
->key
.blend
, 0);
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
, LP_BLD_SWIZZLE_DONTCARE
, 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 (out_format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1402 /* the code above can't work for layout_other */
1403 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1409 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1412 /* If 3 channels then pad to include alpha for 4 element transpose */
1413 if (dst_channels
== 3 && !has_alpha
) {
1414 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1415 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1418 if (out_format_desc
->nr_channels
== 4) {
1424 * Load shader output
1426 for (i
= 0; i
< num_fs
; ++i
) {
1427 /* Always load alpha for use in blending */
1428 LLVMValueRef alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1430 /* Load each channel */
1431 for (j
= 0; j
< dst_channels
; ++j
) {
1432 assert(swizzle
[j
] < 4);
1433 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1436 /* If 3 channels then pad to include alpha for 4 element transpose */
1438 * XXX If we include that here maybe could actually use it instead of
1439 * separate alpha for blending?
1441 if (dst_channels
== 3 && !has_alpha
) {
1442 fs_src
[i
][3] = alpha
;
1445 /* We split the row_mask and row_alpha as we want 128bit interleave */
1446 if (fs_type
.length
== 8) {
1447 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1448 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1450 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1451 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1453 src_mask
[i
] = fs_mask
[i
];
1454 src_alpha
[i
] = alpha
;
1457 if (dual_source_blend
) {
1458 /* same as above except different src/dst, skip masks and comments... */
1459 for (i
= 0; i
< num_fs
; ++i
) {
1460 LLVMValueRef alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1462 for (j
= 0; j
< dst_channels
; ++j
) {
1463 assert(swizzle
[j
] < 4);
1464 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1466 if (dst_channels
== 3 && !has_alpha
) {
1467 fs_src1
[i
][3] = alpha
;
1469 if (fs_type
.length
== 8) {
1470 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1471 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1473 src1_alpha
[i
] = alpha
;
1478 if (util_format_is_pure_integer(out_format
)) {
1480 * In this case fs_type was really ints or uints disguised as floats,
1483 fs_type
.floating
= 0;
1484 fs_type
.sign
= dst_type
.sign
;
1485 for (i
= 0; i
< num_fs
; ++i
) {
1486 for (j
= 0; j
< dst_channels
; ++j
) {
1487 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1488 lp_build_vec_type(gallivm
, fs_type
), "");
1490 if (dst_channels
== 3 && !has_alpha
) {
1491 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1492 lp_build_vec_type(gallivm
, fs_type
), "");
1498 * Pixel twiddle from fragment shader order to memory order
1500 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fs
, dst_channels
, fs_src
, src
, pad_inline
);
1501 if (dual_source_blend
) {
1502 generate_fs_twiddle(gallivm
, fs_type
, num_fs
, dst_channels
, fs_src1
, src1
, pad_inline
);
1505 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1506 if (src_count
!= num_fs
* src_channels
) {
1507 unsigned ds
= src_count
/ (num_fs
* src_channels
);
1508 row_type
.length
/= ds
;
1509 fs_type
.length
= row_type
.length
;
1512 blend_type
= row_type
;
1513 mask_type
.length
= 4;
1515 /* Convert src to row_type */
1516 if (dual_source_blend
) {
1517 struct lp_type old_row_type
= row_type
;
1518 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1519 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1522 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1525 /* If the rows are not an SSE vector, combine them to become SSE size! */
1526 if ((row_type
.width
* row_type
.length
) % 128) {
1527 unsigned bits
= row_type
.width
* row_type
.length
;
1530 dst_count
= src_count
/ (vector_width
/ bits
);
1531 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1532 if (dual_source_blend
) {
1533 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1536 row_type
.length
*= combined
;
1537 src_count
/= combined
;
1539 bits
= row_type
.width
* row_type
.length
;
1540 assert(bits
== 128 || bits
== 256);
1545 * Blend Colour conversion
1547 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1548 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1549 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1552 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1555 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1557 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1558 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1560 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1561 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1563 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1564 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1570 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], 4, &src_mask
[0]);
1572 if (src_count
< block_height
) {
1573 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1574 } else if (src_count
> block_height
) {
1575 for (i
= src_count
; i
> 0; --i
) {
1576 unsigned pixels
= block_size
/ src_count
;
1577 unsigned idx
= i
- 1;
1579 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4], (idx
* pixels
) % 4, pixels
);
1583 assert(mask_type
.width
== 32);
1585 for (i
= 0; i
< src_count
; ++i
) {
1586 unsigned pixels
= block_size
/ src_count
;
1587 unsigned pixel_width
= row_type
.width
* dst_channels
;
1589 if (pixel_width
== 24) {
1590 mask_type
.width
= 8;
1591 mask_type
.length
= vector_width
/ mask_type
.width
;
1593 mask_type
.length
= pixels
;
1594 mask_type
.width
= row_type
.width
* dst_channels
;
1596 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1598 mask_type
.length
*= dst_channels
;
1599 mask_type
.width
/= dst_channels
;
1602 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1603 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1610 struct lp_type alpha_type
= fs_type
;
1611 alpha_type
.length
= 4;
1612 convert_alpha(gallivm
, row_type
, alpha_type
,
1613 block_size
, block_height
,
1614 src_count
, dst_channels
,
1615 pad_inline
, src_alpha
);
1616 if (dual_source_blend
) {
1617 convert_alpha(gallivm
, row_type
, alpha_type
,
1618 block_size
, block_height
,
1619 src_count
, dst_channels
,
1620 pad_inline
, src1_alpha
);
1626 * Load dst from memory
1628 if (src_count
< block_height
) {
1629 dst_count
= block_height
;
1631 dst_count
= src_count
;
1634 dst_type
.length
*= 16 / dst_count
;
1636 if (out_format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1638 * we need multiple values at once for the conversion, so can as well
1639 * load them vectorized here too instead of concatenating later.
1640 * (Still need concatenation later for 8-wide vectors).
1642 dst_count
= block_height
;
1643 dst_type
.length
= block_width
;
1646 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1647 dst
, dst_type
, dst_count
, dst_alignment
);
1651 * Convert from dst/output format to src/blending format.
1653 * This is necessary as we can only read 1 row from memory at a time,
1654 * so the minimum dst_count will ever be at this point is 4.
1656 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1657 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1658 * on all 16 pixels in that single vector at once.
1660 if (dst_count
> src_count
) {
1661 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
1667 /* XXX this is broken for RGB8 formats -
1668 * they get expanded from 12 to 16 elements (to include alpha)
1669 * by convert_to_blend_type then reduced to 15 instead of 12
1670 * by convert_from_blend_type (a simple fix though breaks A8...).
1671 * R16G16B16 also crashes differently however something going wrong
1672 * inside llvm handling npot vector sizes seemingly.
1673 * It seems some cleanup could be done here (like skipping conversion/blend
1676 convert_to_blend_type(gallivm
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
1678 for (i
= 0; i
< src_count
; ++i
) {
1679 dst
[i
] = lp_build_blend_aos(gallivm
,
1680 &variant
->key
.blend
,
1685 has_alpha
? NULL
: src_alpha
[i
],
1687 has_alpha
? NULL
: src1_alpha
[i
],
1689 partial_mask
? src_mask
[i
] : NULL
,
1691 has_alpha
? NULL
: blend_alpha
,
1693 pad_inline
? 4 : dst_channels
);
1696 convert_from_blend_type(gallivm
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
1698 /* Split the blend rows back to memory rows */
1699 if (dst_count
> src_count
) {
1700 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
1702 if (src_count
== 1) {
1703 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1704 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1706 row_type
.length
/= 2;
1710 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
1711 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
1712 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
1713 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
1715 row_type
.length
/= 2;
1721 * Store blend result to memory
1723 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1724 dst
, dst_type
, dst_count
, dst_alignment
);
1727 lp_build_mask_end(&mask_ctx
);
1733 * Generate the runtime callable function for the whole fragment pipeline.
1734 * Note that the function which we generate operates on a block of 16
1735 * pixels at at time. The block contains 2x2 quads. Each quad contains
1739 generate_fragment(struct llvmpipe_context
*lp
,
1740 struct lp_fragment_shader
*shader
,
1741 struct lp_fragment_shader_variant
*variant
,
1742 unsigned partial_mask
)
1744 struct gallivm_state
*gallivm
= variant
->gallivm
;
1745 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
1746 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
1747 char func_name
[256];
1748 struct lp_type fs_type
;
1749 struct lp_type blend_type
;
1750 LLVMTypeRef fs_elem_type
;
1751 LLVMTypeRef blend_vec_type
;
1752 LLVMTypeRef arg_types
[12];
1753 LLVMTypeRef func_type
;
1754 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1755 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1756 LLVMValueRef context_ptr
;
1759 LLVMValueRef a0_ptr
;
1760 LLVMValueRef dadx_ptr
;
1761 LLVMValueRef dady_ptr
;
1762 LLVMValueRef color_ptr_ptr
;
1763 LLVMValueRef stride_ptr
;
1764 LLVMValueRef depth_ptr
;
1765 LLVMValueRef mask_input
;
1766 LLVMValueRef thread_data_ptr
;
1767 LLVMBasicBlockRef block
;
1768 LLVMBuilderRef builder
;
1769 struct lp_build_sampler_soa
*sampler
;
1770 struct lp_build_interp_soa_context interp
;
1771 LLVMValueRef fs_mask
[16 / 4];
1772 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
1773 LLVMValueRef function
;
1774 LLVMValueRef facing
;
1775 const struct util_format_description
*zs_format_desc
;
1780 boolean cbuf0_write_all
;
1781 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
1782 util_blend_state_is_dual(&key
->blend
, 0);
1784 assert(lp_native_vector_width
/ 32 >= 4);
1786 /* Adjust color input interpolation according to flatshade state:
1788 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
1789 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
1790 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
1792 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
1794 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
1798 /* check if writes to cbuf[0] are to be copied to all cbufs */
1799 cbuf0_write_all
= FALSE
;
1800 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
1801 if (shader
->info
.base
.properties
[i
].name
==
1802 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
1803 cbuf0_write_all
= TRUE
;
1808 /* TODO: actually pick these based on the fs and color buffer
1809 * characteristics. */
1811 memset(&fs_type
, 0, sizeof fs_type
);
1812 fs_type
.floating
= TRUE
; /* floating point values */
1813 fs_type
.sign
= TRUE
; /* values are signed */
1814 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1815 fs_type
.width
= 32; /* 32-bit float */
1816 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
1817 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
1819 memset(&blend_type
, 0, sizeof blend_type
);
1820 blend_type
.floating
= FALSE
; /* values are integers */
1821 blend_type
.sign
= FALSE
; /* values are unsigned */
1822 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
1823 blend_type
.width
= 8; /* 8-bit ubyte values */
1824 blend_type
.length
= 16; /* 16 elements per vector */
1827 * Generate the function prototype. Any change here must be reflected in
1828 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1831 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
1833 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
1835 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
1836 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
1838 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
1839 arg_types
[1] = int32_type
; /* x */
1840 arg_types
[2] = int32_type
; /* y */
1841 arg_types
[3] = int32_type
; /* facing */
1842 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
1843 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
1844 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
1845 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
1846 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
1847 arg_types
[9] = int32_type
; /* mask_input */
1848 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
1849 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
1851 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
1852 arg_types
, Elements(arg_types
), 0);
1854 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1855 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
1857 variant
->function
[partial_mask
] = function
;
1859 /* XXX: need to propagate noalias down into color param now we are
1860 * passing a pointer-to-pointer?
1862 for(i
= 0; i
< Elements(arg_types
); ++i
)
1863 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1864 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
1866 context_ptr
= LLVMGetParam(function
, 0);
1867 x
= LLVMGetParam(function
, 1);
1868 y
= LLVMGetParam(function
, 2);
1869 facing
= LLVMGetParam(function
, 3);
1870 a0_ptr
= LLVMGetParam(function
, 4);
1871 dadx_ptr
= LLVMGetParam(function
, 5);
1872 dady_ptr
= LLVMGetParam(function
, 6);
1873 color_ptr_ptr
= LLVMGetParam(function
, 7);
1874 depth_ptr
= LLVMGetParam(function
, 8);
1875 mask_input
= LLVMGetParam(function
, 9);
1876 thread_data_ptr
= LLVMGetParam(function
, 10);
1877 stride_ptr
= LLVMGetParam(function
, 11);
1879 lp_build_name(context_ptr
, "context");
1880 lp_build_name(x
, "x");
1881 lp_build_name(y
, "y");
1882 lp_build_name(a0_ptr
, "a0");
1883 lp_build_name(dadx_ptr
, "dadx");
1884 lp_build_name(dady_ptr
, "dady");
1885 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
1886 lp_build_name(depth_ptr
, "depth");
1887 lp_build_name(thread_data_ptr
, "thread_data");
1888 lp_build_name(mask_input
, "mask_input");
1889 lp_build_name(stride_ptr
, "stride_ptr");
1895 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
1896 builder
= gallivm
->builder
;
1898 LLVMPositionBuilderAtEnd(builder
, block
);
1900 /* code generated texture sampling */
1901 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
1903 zs_format_desc
= util_format_description(key
->zsbuf_format
);
1906 unsigned depth_bits
= zs_format_desc
->block
.bits
/8;
1907 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
1908 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
1909 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
1910 num_loop
, "mask_store");
1911 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
1914 * The shader input interpolation info is not explicitely baked in the
1915 * shader key, but everything it derives from (TGSI, and flatshade) is
1916 * already included in the shader key.
1918 lp_build_interp_soa_init(&interp
,
1920 shader
->info
.base
.num_inputs
,
1922 shader
->info
.base
.pixel_center_integer
,
1924 a0_ptr
, dadx_ptr
, dady_ptr
,
1927 for (i
= 0; i
< num_fs
; i
++) {
1929 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
1930 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
1931 &indexi
, 1, "mask_ptr");
1934 mask
= generate_quad_mask(gallivm
, fs_type
,
1935 i
*fs_type
.length
/4, mask_input
);
1938 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
1940 LLVMBuildStore(builder
, mask
, mask_ptr
);
1943 generate_fs_loop(gallivm
,
1951 mask_store
, /* output */
1958 for (i
= 0; i
< num_fs
; i
++) {
1959 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
1960 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
1962 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
1963 /* This is fucked up need to reorganize things */
1964 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
1965 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1966 ptr
= LLVMBuildGEP(builder
,
1967 color_store
[cbuf
* !cbuf0_write_all
][chan
],
1969 fs_out_color
[cbuf
][chan
][i
] = ptr
;
1972 if (dual_source_blend
) {
1973 /* only support one dual source blend target hence always use output 1 */
1974 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1975 ptr
= LLVMBuildGEP(builder
,
1976 color_store
[1][chan
],
1978 fs_out_color
[1][chan
][i
] = ptr
;
1984 sampler
->destroy(sampler
);
1986 /* Loop over color outputs / color buffers to do blending.
1988 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
1989 LLVMValueRef color_ptr
;
1990 LLVMValueRef stride
;
1991 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
1993 boolean do_branch
= ((key
->depth
.enabled
1994 || key
->stencil
[0].enabled
1995 || key
->alpha
.enabled
)
1996 && !shader
->info
.base
.uses_kill
);
1998 color_ptr
= LLVMBuildLoad(builder
,
1999 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2002 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2004 stride
= LLVMBuildLoad(builder
,
2005 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2008 generate_unswizzled_blend(gallivm
, cbuf
, variant
, key
->cbuf_format
[cbuf
],
2009 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2010 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2013 LLVMBuildRetVoid(builder
);
2015 gallivm_verify_function(gallivm
, function
);
2017 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2022 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2026 debug_printf("fs variant %p:\n", (void *) key
);
2028 if (key
->flatshade
) {
2029 debug_printf("flatshade = 1\n");
2031 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2032 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2034 if (key
->depth
.enabled
) {
2035 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2036 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2037 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2040 for (i
= 0; i
< 2; ++i
) {
2041 if (key
->stencil
[i
].enabled
) {
2042 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2043 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2044 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2045 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2046 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2047 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2051 if (key
->alpha
.enabled
) {
2052 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2055 if (key
->occlusion_count
) {
2056 debug_printf("occlusion_count = 1\n");
2059 if (key
->blend
.logicop_enable
) {
2060 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2062 else if (key
->blend
.rt
[0].blend_enable
) {
2063 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2064 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2065 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2066 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2067 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2068 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2070 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2071 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2072 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2073 debug_printf("sampler[%u] = \n", i
);
2074 debug_printf(" .wrap = %s %s %s\n",
2075 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2076 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2077 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2078 debug_printf(" .min_img_filter = %s\n",
2079 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2080 debug_printf(" .min_mip_filter = %s\n",
2081 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2082 debug_printf(" .mag_img_filter = %s\n",
2083 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2084 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2085 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2086 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2087 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2088 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2089 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2090 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2092 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2093 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2094 debug_printf("texture[%u] = \n", i
);
2095 debug_printf(" .format = %s\n",
2096 util_format_name(texture
->format
));
2097 debug_printf(" .target = %s\n",
2098 util_dump_tex_target(texture
->target
, TRUE
));
2099 debug_printf(" .level_zero_only = %u\n",
2100 texture
->level_zero_only
);
2101 debug_printf(" .pot = %u %u %u\n",
2103 texture
->pot_height
,
2104 texture
->pot_depth
);
2110 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2112 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2113 variant
->shader
->no
, variant
->no
);
2114 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2115 dump_fs_variant_key(&variant
->key
);
2116 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2122 * Generate a new fragment shader variant from the shader code and
2123 * other state indicated by the key.
2125 static struct lp_fragment_shader_variant
*
2126 generate_variant(struct llvmpipe_context
*lp
,
2127 struct lp_fragment_shader
*shader
,
2128 const struct lp_fragment_shader_variant_key
*key
)
2130 struct lp_fragment_shader_variant
*variant
;
2131 const struct util_format_description
*cbuf0_format_desc
;
2132 boolean fullcolormask
;
2134 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2138 variant
->gallivm
= gallivm_create();
2139 if (!variant
->gallivm
) {
2144 variant
->shader
= shader
;
2145 variant
->list_item_global
.base
= variant
;
2146 variant
->list_item_local
.base
= variant
;
2147 variant
->no
= shader
->variants_created
++;
2149 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2152 * Determine whether we are touching all channels in the color buffer.
2154 fullcolormask
= FALSE
;
2155 if (key
->nr_cbufs
== 1) {
2156 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2157 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2161 !key
->blend
.logicop_enable
&&
2162 !key
->blend
.rt
[0].blend_enable
&&
2164 !key
->stencil
[0].enabled
&&
2165 !key
->alpha
.enabled
&&
2166 !key
->depth
.enabled
&&
2167 !shader
->info
.base
.uses_kill
2170 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2171 lp_debug_fs_variant(variant
);
2174 lp_jit_init_types(variant
);
2176 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2177 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2179 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2180 if (variant
->opaque
) {
2181 /* Specialized shader, which doesn't need to read the color buffer. */
2182 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2187 * Compile everything
2190 gallivm_compile_module(variant
->gallivm
);
2192 if (variant
->function
[RAST_EDGE_TEST
]) {
2193 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2194 gallivm_jit_function(variant
->gallivm
,
2195 variant
->function
[RAST_EDGE_TEST
]);
2198 if (variant
->function
[RAST_WHOLE
]) {
2199 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2200 gallivm_jit_function(variant
->gallivm
,
2201 variant
->function
[RAST_WHOLE
]);
2202 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2203 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2211 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2212 const struct pipe_shader_state
*templ
)
2214 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2215 struct lp_fragment_shader
*shader
;
2217 int nr_sampler_views
;
2220 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2224 shader
->no
= fs_no
++;
2225 make_empty_list(&shader
->variants
);
2227 /* get/save the summary info for this shader */
2228 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2230 /* we need to keep a local copy of the tokens */
2231 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2233 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2234 if (shader
->draw_data
== NULL
) {
2235 FREE((void *) shader
->base
.tokens
);
2240 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2241 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2243 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2244 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2246 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2247 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2248 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2250 switch (shader
->info
.base
.input_interpolate
[i
]) {
2251 case TGSI_INTERPOLATE_CONSTANT
:
2252 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2254 case TGSI_INTERPOLATE_LINEAR
:
2255 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2257 case TGSI_INTERPOLATE_PERSPECTIVE
:
2258 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2260 case TGSI_INTERPOLATE_COLOR
:
2261 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2268 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2269 case TGSI_SEMANTIC_FACE
:
2270 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2272 case TGSI_SEMANTIC_POSITION
:
2273 /* Position was already emitted above
2275 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2276 shader
->inputs
[i
].src_index
= 0;
2280 shader
->inputs
[i
].src_index
= i
+1;
2283 if (LP_DEBUG
& DEBUG_TGSI
) {
2285 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2286 shader
->no
, (void *) shader
);
2287 tgsi_dump(templ
->tokens
, 0);
2288 debug_printf("usage masks:\n");
2289 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2290 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2291 debug_printf(" IN[%u].%s%s%s%s\n",
2293 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2294 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2295 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2296 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2306 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2308 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2310 if (llvmpipe
->fs
== fs
)
2313 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2315 draw_bind_fragment_shader(llvmpipe
->draw
,
2316 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2318 llvmpipe
->dirty
|= LP_NEW_FS
;
2323 * Remove shader variant from two lists: the shader's variant list
2324 * and the context's variant list.
2327 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2328 struct lp_fragment_shader_variant
*variant
)
2332 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2333 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2334 " #%u v total cached #%u\n",
2335 variant
->shader
->no
,
2337 variant
->shader
->variants_created
,
2338 variant
->shader
->variants_cached
,
2339 lp
->nr_fs_variants
);
2342 /* free all the variant's JIT'd functions */
2343 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2344 if (variant
->function
[i
]) {
2345 gallivm_free_function(variant
->gallivm
,
2346 variant
->function
[i
],
2347 variant
->jit_function
[i
]);
2351 gallivm_destroy(variant
->gallivm
);
2353 /* remove from shader's list */
2354 remove_from_list(&variant
->list_item_local
);
2355 variant
->shader
->variants_cached
--;
2357 /* remove from context's list */
2358 remove_from_list(&variant
->list_item_global
);
2359 lp
->nr_fs_variants
--;
2360 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2367 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2369 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2370 struct lp_fragment_shader
*shader
= fs
;
2371 struct lp_fs_variant_list_item
*li
;
2373 assert(fs
!= llvmpipe
->fs
);
2376 * XXX: we need to flush the context until we have some sort of reference
2377 * counting in fragment shaders as they may still be binned
2378 * Flushing alone might not sufficient we need to wait on it too.
2380 llvmpipe_finish(pipe
, __FUNCTION__
);
2382 /* Delete all the variants */
2383 li
= first_elem(&shader
->variants
);
2384 while(!at_end(&shader
->variants
, li
)) {
2385 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2386 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2390 /* Delete draw module's data */
2391 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2393 assert(shader
->variants_cached
== 0);
2394 FREE((void *) shader
->base
.tokens
);
2401 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2402 uint shader
, uint index
,
2403 struct pipe_constant_buffer
*cb
)
2405 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2406 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2408 assert(shader
< PIPE_SHADER_TYPES
);
2409 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2411 /* note: reference counting */
2412 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2414 if (shader
== PIPE_SHADER_VERTEX
||
2415 shader
== PIPE_SHADER_GEOMETRY
) {
2416 /* Pass the constants to the 'draw' module */
2417 const unsigned size
= cb
? cb
->buffer_size
: 0;
2421 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2423 else if (cb
&& cb
->user_buffer
) {
2424 data
= (ubyte
*) cb
->user_buffer
;
2431 data
+= cb
->buffer_offset
;
2433 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2437 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2439 if (cb
&& cb
->user_buffer
) {
2440 pipe_resource_reference(&constants
, NULL
);
2446 * Return the blend factor equivalent to a destination alpha of one.
2448 static INLINE
unsigned
2449 force_dst_alpha_one(unsigned factor
)
2452 case PIPE_BLENDFACTOR_DST_ALPHA
:
2453 return PIPE_BLENDFACTOR_ONE
;
2454 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2455 return PIPE_BLENDFACTOR_ZERO
;
2456 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2457 return PIPE_BLENDFACTOR_ZERO
;
2465 * We need to generate several variants of the fragment pipeline to match
2466 * all the combinations of the contributing state atoms.
2468 * TODO: there is actually no reason to tie this to context state -- the
2469 * generated code could be cached globally in the screen.
2472 make_variant_key(struct llvmpipe_context
*lp
,
2473 struct lp_fragment_shader
*shader
,
2474 struct lp_fragment_shader_variant_key
*key
)
2478 memset(key
, 0, shader
->variant_key_size
);
2480 if (lp
->framebuffer
.zsbuf
) {
2481 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2482 const struct util_format_description
*zsbuf_desc
=
2483 util_format_description(zsbuf_format
);
2485 if (lp
->depth_stencil
->depth
.enabled
&&
2486 util_format_has_depth(zsbuf_desc
)) {
2487 key
->zsbuf_format
= zsbuf_format
;
2488 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2490 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2491 util_format_has_stencil(zsbuf_desc
)) {
2492 key
->zsbuf_format
= zsbuf_format
;
2493 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2497 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2498 if (!lp
->framebuffer
.nr_cbufs
||
2499 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2500 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2502 if(key
->alpha
.enabled
)
2503 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2504 /* alpha.ref_value is passed in jit_context */
2506 key
->flatshade
= lp
->rasterizer
->flatshade
;
2507 if (lp
->active_occlusion_query
) {
2508 key
->occlusion_count
= TRUE
;
2511 if (lp
->framebuffer
.nr_cbufs
) {
2512 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2515 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2517 if (!key
->blend
.independent_blend_enable
) {
2518 /* we always need independent blend otherwise the fixups below won't work */
2519 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2520 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2522 key
->blend
.independent_blend_enable
= 1;
2525 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2526 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2527 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2528 const struct util_format_description
*format_desc
;
2530 key
->cbuf_format
[i
] = format
;
2532 format_desc
= util_format_description(format
);
2533 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2534 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2537 * Mask out color channels not present in the color buffer.
2539 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2542 * Disable blend for integer formats.
2544 if (util_format_is_pure_integer(format
)) {
2545 blend_rt
->blend_enable
= 0;
2549 * Our swizzled render tiles always have an alpha channel, but the linear
2550 * render target format often does not, so force here the dst alpha to be
2553 * This is not a mere optimization. Wrong results will be produced if the
2554 * dst alpha is used, the dst format does not have alpha, and the previous
2555 * rendering was not flushed from the swizzled to linear buffer. For
2556 * example, NonPowTwo DCT.
2558 * TODO: This should be generalized to all channels for better
2559 * performance, but only alpha causes correctness issues.
2561 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2563 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2564 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2565 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
);
2566 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
);
2567 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2568 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2569 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2573 /* This value will be the same for all the variants of a given shader:
2575 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2577 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2578 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2579 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
2580 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
2585 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2586 * are dx10-style? Can't really have mixed opcodes, at least not
2587 * if we want to skip the holes here (without rescanning tgsi).
2589 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2590 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2591 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2592 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
2593 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2594 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2599 key
->nr_sampler_views
= key
->nr_samplers
;
2600 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2601 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2602 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2603 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2612 * Update fragment shader state. This is called just prior to drawing
2613 * something when some fragment-related state has changed.
2616 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
2618 struct lp_fragment_shader
*shader
= lp
->fs
;
2619 struct lp_fragment_shader_variant_key key
;
2620 struct lp_fragment_shader_variant
*variant
= NULL
;
2621 struct lp_fs_variant_list_item
*li
;
2623 make_variant_key(lp
, shader
, &key
);
2625 /* Search the variants for one which matches the key */
2626 li
= first_elem(&shader
->variants
);
2627 while(!at_end(&shader
->variants
, li
)) {
2628 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
2636 /* Move this variant to the head of the list to implement LRU
2637 * deletion of shader's when we have too many.
2639 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2642 /* variant not found, create it now */
2645 unsigned variants_to_cull
;
2648 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2651 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
2654 /* First, check if we've exceeded the max number of shader variants.
2655 * If so, free 25% of them (the least recently used ones).
2657 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
2659 if (variants_to_cull
||
2660 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
2661 struct pipe_context
*pipe
= &lp
->pipe
;
2664 * XXX: we need to flush the context until we have some sort of
2665 * reference counting in fragment shaders as they may still be binned
2666 * Flushing alone might not be sufficient we need to wait on it too.
2668 llvmpipe_finish(pipe
, __FUNCTION__
);
2671 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2672 * number of shader variants (potentially all of them) could be
2673 * pending for destruction on flush.
2676 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
2677 struct lp_fs_variant_list_item
*item
;
2678 if (is_empty_list(&lp
->fs_variants_list
)) {
2681 item
= last_elem(&lp
->fs_variants_list
);
2684 llvmpipe_remove_shader_variant(lp
, item
->base
);
2689 * Generate the new variant.
2692 variant
= generate_variant(lp
, shader
, &key
);
2695 LP_COUNT_ADD(llvm_compile_time
, dt
);
2696 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
2698 llvmpipe_variant_count
++;
2700 /* Put the new variant into the list */
2702 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
2703 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
2704 lp
->nr_fs_variants
++;
2705 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
2706 shader
->variants_cached
++;
2710 /* Bind this variant */
2711 lp_setup_set_fs_variant(lp
->setup
, variant
);
2719 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
2721 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
2722 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
2723 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
2725 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;