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 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
221 lp_llvm_viewport(LLVMValueRef context_ptr
,
222 struct gallivm_state
*gallivm
,
223 LLVMValueRef viewport_index
)
225 LLVMBuilderRef builder
= gallivm
->builder
;
228 struct lp_type viewport_type
=
229 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
231 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
232 ptr
= LLVMBuildPointerCast(builder
, ptr
,
233 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
235 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
242 * Generate the fragment shader, depth/stencil test, and alpha tests.
245 generate_fs_loop(struct gallivm_state
*gallivm
,
246 struct lp_fragment_shader
*shader
,
247 const struct lp_fragment_shader_variant_key
*key
,
248 LLVMBuilderRef builder
,
250 LLVMValueRef context_ptr
,
251 LLVMValueRef num_loop
,
252 struct lp_build_interp_soa_context
*interp
,
253 struct lp_build_sampler_soa
*sampler
,
254 LLVMValueRef mask_store
,
255 LLVMValueRef (*out_color
)[4],
256 LLVMValueRef depth_ptr
,
257 LLVMValueRef depth_stride
,
259 LLVMValueRef thread_data_ptr
)
261 const struct util_format_description
*zs_format_desc
= NULL
;
262 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
263 LLVMTypeRef vec_type
;
264 LLVMValueRef mask_ptr
, mask_val
;
265 LLVMValueRef consts_ptr
;
267 LLVMValueRef z_value
, s_value
;
268 LLVMValueRef z_fb
, s_fb
;
269 LLVMValueRef stencil_refs
[2];
270 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
271 struct lp_build_for_loop_state loop_state
;
272 struct lp_build_mask_context mask
;
274 * TODO: figure out if simple_shader optimization is really worthwile to
275 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
276 * code since tests tend to take another codepath than real shaders.
278 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
279 shader
->info
.base
.num_inputs
< 3 &&
280 shader
->info
.base
.num_instructions
< 8) && 0;
281 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
282 util_blend_state_is_dual(&key
->blend
, 0);
288 struct lp_bld_tgsi_system_values system_values
;
290 memset(&system_values
, 0, sizeof(system_values
));
292 if (key
->depth
.enabled
||
293 key
->stencil
[0].enabled
) {
295 zs_format_desc
= util_format_description(key
->zsbuf_format
);
296 assert(zs_format_desc
);
298 if (!shader
->info
.base
.writes_z
) {
299 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
) {
300 /* With alpha test and kill, can do the depth test early
301 * and hopefully eliminate some quads. But need to do a
302 * special deferred depth write once the final mask value
303 * is known. This only works though if there's either no
304 * stencil test or the stencil value isn't written.
306 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
307 (key
->stencil
[1].enabled
&&
308 key
->stencil
[1].writemask
)))
309 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
311 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
314 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
317 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
320 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
321 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
322 (key
->stencil
[1].enabled
&&
323 key
->stencil
[1].writemask
))))
324 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
331 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
332 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
334 vec_type
= lp_build_vec_type(gallivm
, type
);
336 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
338 lp_build_for_loop_begin(&loop_state
, gallivm
,
339 lp_build_const_int32(gallivm
, 0),
342 lp_build_const_int32(gallivm
, 1));
344 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
345 &loop_state
.counter
, 1, "mask_ptr");
346 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
348 memset(outputs
, 0, sizeof outputs
);
350 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
351 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
352 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
353 lp_build_vec_type(gallivm
,
358 if (dual_source_blend
) {
359 assert(key
->nr_cbufs
<= 1);
360 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
361 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
362 lp_build_vec_type(gallivm
,
369 /* 'mask' will control execution based on quad's pixel alive/killed state */
370 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
372 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
373 lp_build_mask_check(&mask
);
375 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
378 if (depth_mode
& EARLY_DEPTH_TEST
) {
379 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
380 zs_format_desc
, key
->resource_1d
,
381 depth_ptr
, depth_stride
,
382 &z_fb
, &s_fb
, loop_state
.counter
);
383 lp_build_depth_stencil_test(gallivm
,
395 if (depth_mode
& EARLY_DEPTH_WRITE
) {
396 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
397 zs_format_desc
, key
->resource_1d
,
398 NULL
, NULL
, NULL
, loop_state
.counter
,
399 depth_ptr
, depth_stride
,
403 * Note mask check if stencil is enabled must be after ds write not after
404 * stencil test otherwise new stencil values may not get written if all
405 * fragments got killed by depth/stencil test.
407 if (!simple_shader
&& key
->stencil
[0].enabled
)
408 lp_build_mask_check(&mask
);
411 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
413 /* Build the actual shader */
414 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
415 consts_ptr
, &system_values
,
417 outputs
, sampler
, &shader
->info
.base
, NULL
);
420 if (key
->alpha
.enabled
) {
421 int color0
= find_output_by_semantic(&shader
->info
.base
,
425 if (color0
!= -1 && outputs
[color0
][3]) {
426 const struct util_format_description
*cbuf_format_desc
;
427 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
428 LLVMValueRef alpha_ref_value
;
430 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
431 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
433 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
435 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
436 &mask
, alpha
, alpha_ref_value
,
437 (depth_mode
& LATE_DEPTH_TEST
) != 0);
442 if (depth_mode
& LATE_DEPTH_TEST
) {
443 int pos0
= find_output_by_semantic(&shader
->info
.base
,
444 TGSI_SEMANTIC_POSITION
,
447 if (pos0
!= -1 && outputs
[pos0
][2]) {
448 LLVMValueRef viewport
, min_depth
, max_depth
;
449 LLVMValueRef viewport_index
;
450 struct lp_build_context f32_bld
;
452 assert(type
.floating
);
453 lp_build_context_init(&f32_bld
, gallivm
, type
);
456 * Assumes clamping of the viewport index will occur in setup/gs. Value
457 * is passed through the rasterization stage via lp_rast_shader_inputs.
459 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
462 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
466 * Load the min and max depth from the lp_jit_context.viewports
467 * array of lp_jit_viewport structures.
469 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
471 /* viewports[viewport_index].min_depth */
472 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
473 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
),
475 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
477 /* viewports[viewport_index].max_depth */
478 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
479 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
),
481 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
483 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
486 * Clamp to the min and max depth values for the given viewport.
488 z
= lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
491 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
492 zs_format_desc
, key
->resource_1d
,
493 depth_ptr
, depth_stride
,
494 &z_fb
, &s_fb
, loop_state
.counter
);
496 lp_build_depth_stencil_test(gallivm
,
508 if (depth_mode
& LATE_DEPTH_WRITE
) {
509 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
510 zs_format_desc
, key
->resource_1d
,
511 NULL
, NULL
, NULL
, loop_state
.counter
,
512 depth_ptr
, depth_stride
,
516 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
517 (depth_mode
& LATE_DEPTH_WRITE
))
519 /* Need to apply a reduced mask to the depth write. Reload the
520 * depth value, update from zs_value with the new mask value and
523 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
524 zs_format_desc
, key
->resource_1d
,
525 &mask
, z_fb
, s_fb
, loop_state
.counter
,
526 depth_ptr
, depth_stride
,
532 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
534 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
535 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
536 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
538 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
539 if(outputs
[attrib
][chan
]) {
540 /* XXX: just initialize outputs to point at colors[] and
543 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
544 LLVMValueRef color_ptr
;
545 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
546 &loop_state
.counter
, 1, "");
547 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
548 LLVMBuildStore(builder
, out
, color_ptr
);
554 if (key
->occlusion_count
) {
555 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
556 lp_build_name(counter
, "counter");
557 lp_build_occlusion_count(gallivm
, type
,
558 lp_build_mask_value(&mask
), counter
);
561 mask_val
= lp_build_mask_end(&mask
);
562 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
563 lp_build_for_loop_end(&loop_state
);
568 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
570 * Fragment Shader outputs pixels in small 2x2 blocks
571 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
573 * However in memory pixels are stored in rows
574 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
576 * @param type fragment shader type (4x or 8x float)
577 * @param num_fs number of fs_src
578 * @param is_1d whether we're outputting to a 1d resource
579 * @param dst_channels number of output channels
580 * @param fs_src output from fragment shader
581 * @param dst pointer to store result
582 * @param pad_inline is channel padding inline or at end of row
583 * @return the number of dsts
586 generate_fs_twiddle(struct gallivm_state
*gallivm
,
589 unsigned dst_channels
,
590 LLVMValueRef fs_src
[][4],
594 LLVMValueRef src
[16];
600 unsigned pixels
= type
.length
/ 4;
601 unsigned reorder_group
;
602 unsigned src_channels
;
606 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
607 src_count
= num_fs
* src_channels
;
609 assert(pixels
== 2 || pixels
== 1);
610 assert(num_fs
* src_channels
<= Elements(src
));
613 * Transpose from SoA -> AoS
615 for (i
= 0; i
< num_fs
; ++i
) {
616 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
620 * Pick transformation options
627 if (dst_channels
== 1) {
633 } else if (dst_channels
== 2) {
637 } else if (dst_channels
> 2) {
644 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
650 * Split the src in half
653 for (i
= num_fs
; i
> 0; --i
) {
654 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
655 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
663 * Ensure pixels are in memory order
666 /* Twiddle pixels by reordering the array, e.g.:
668 * src_count = 8 -> 0 2 1 3 4 6 5 7
669 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
671 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
673 for (i
= 0; i
< src_count
; ++i
) {
674 unsigned group
= i
/ reorder_group
;
675 unsigned block
= (group
/ 4) * 4 * reorder_group
;
676 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
679 } else if (twiddle
) {
680 /* Twiddle pixels across elements of array */
681 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
684 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
688 * Moves any padding between pixels to the end
689 * e.g. RGBXRGBX -> RGBRGBXX
692 unsigned char swizzles
[16];
693 unsigned elems
= pixels
* dst_channels
;
695 for (i
= 0; i
< type
.length
; ++i
) {
697 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
699 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
702 for (i
= 0; i
< src_count
; ++i
) {
703 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
712 * Load an unswizzled block of pixels from memory
715 load_unswizzled_block(struct gallivm_state
*gallivm
,
716 LLVMValueRef base_ptr
,
718 unsigned block_width
,
719 unsigned block_height
,
721 struct lp_type dst_type
,
723 unsigned dst_alignment
)
725 LLVMBuilderRef builder
= gallivm
->builder
;
726 unsigned row_size
= dst_count
/ block_height
;
729 /* Ensure block exactly fits into dst */
730 assert((block_width
* block_height
) % dst_count
== 0);
732 for (i
= 0; i
< dst_count
; ++i
) {
733 unsigned x
= i
% row_size
;
734 unsigned y
= i
/ row_size
;
736 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
737 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
740 LLVMValueRef dst_ptr
;
742 gep
[0] = lp_build_const_int32(gallivm
, 0);
743 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
745 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
746 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
748 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
750 lp_set_load_alignment(dst
[i
], dst_alignment
);
756 * Store an unswizzled block of pixels to memory
759 store_unswizzled_block(struct gallivm_state
*gallivm
,
760 LLVMValueRef base_ptr
,
762 unsigned block_width
,
763 unsigned block_height
,
765 struct lp_type src_type
,
767 unsigned src_alignment
)
769 LLVMBuilderRef builder
= gallivm
->builder
;
770 unsigned row_size
= src_count
/ block_height
;
773 /* Ensure src exactly fits into block */
774 assert((block_width
* block_height
) % src_count
== 0);
776 for (i
= 0; i
< src_count
; ++i
) {
777 unsigned x
= i
% row_size
;
778 unsigned y
= i
/ row_size
;
780 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
781 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
784 LLVMValueRef src_ptr
;
786 gep
[0] = lp_build_const_int32(gallivm
, 0);
787 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
789 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
790 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
792 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
794 lp_set_store_alignment(src_ptr
, src_alignment
);
800 * Checks if a format description is an arithmetic format
802 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
804 static INLINE boolean
805 is_arithmetic_format(const struct util_format_description
*format_desc
)
807 boolean arith
= false;
810 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
811 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
812 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
820 * Checks if this format requires special handling due to required expansion
821 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
824 static INLINE boolean
825 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
827 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
828 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
836 * Retrieves the type representing the memory layout for a format
838 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
841 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
842 struct lp_type
* type
)
847 if (format_expands_to_float_soa(format_desc
)) {
848 /* just make this a 32bit uint */
849 type
->floating
= false;
858 for (i
= 0; i
< 4; i
++)
859 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
863 memset(type
, 0, sizeof(struct lp_type
));
864 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
865 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
866 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
867 type
->norm
= format_desc
->channel
[chan
].normalized
;
869 if (is_arithmetic_format(format_desc
)) {
873 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
874 type
->width
+= format_desc
->channel
[i
].size
;
877 type
->width
= format_desc
->channel
[chan
].size
;
878 type
->length
= format_desc
->nr_channels
;
884 * Retrieves the type for a format which is usable in the blending code.
886 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
889 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
890 struct lp_type
* type
)
895 if (format_expands_to_float_soa(format_desc
)) {
896 /* always use ordinary floats for blending */
897 type
->floating
= true;
906 for (i
= 0; i
< 4; i
++)
907 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
911 memset(type
, 0, sizeof(struct lp_type
));
912 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
913 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
914 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
915 type
->norm
= format_desc
->channel
[chan
].normalized
;
916 type
->width
= format_desc
->channel
[chan
].size
;
917 type
->length
= format_desc
->nr_channels
;
919 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
920 if (format_desc
->channel
[i
].size
> type
->width
)
921 type
->width
= format_desc
->channel
[i
].size
;
924 if (type
->floating
) {
927 if (type
->width
<= 8) {
929 } else if (type
->width
<= 16) {
936 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
943 * Scale a normalized value from src_bits to dst_bits.
945 * The exact calculation is
947 * dst = iround(src * dst_mask / src_mask)
949 * or with integer rounding
951 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
955 * src_mask = (1 << src_bits) - 1
956 * dst_mask = (1 << dst_bits) - 1
958 * but we try to avoid division and multiplication through shifts.
960 static INLINE LLVMValueRef
961 scale_bits(struct gallivm_state
*gallivm
,
965 struct lp_type src_type
)
967 LLVMBuilderRef builder
= gallivm
->builder
;
968 LLVMValueRef result
= src
;
970 if (dst_bits
< src_bits
) {
971 int delta_bits
= src_bits
- dst_bits
;
973 if (delta_bits
<= dst_bits
) {
975 * Approximate the rescaling with a single shift.
977 * This gives the wrong rounding.
980 result
= LLVMBuildLShr(builder
,
982 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
987 * Try more accurate rescaling.
991 * Drop the least significant bits to make space for the multiplication.
993 * XXX: A better approach would be to use a wider integer type as intermediate. But
994 * this is enough to convert alpha from 16bits -> 2 when rendering to
995 * PIPE_FORMAT_R10G10B10A2_UNORM.
997 result
= LLVMBuildLShr(builder
,
999 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1003 result
= LLVMBuildMul(builder
,
1005 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1009 * Add a rounding term before the division.
1011 * TODO: Handle signed integers too.
1013 if (!src_type
.sign
) {
1014 result
= LLVMBuildAdd(builder
,
1016 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1021 * Approximate the division by src_mask with a src_bits shift.
1023 * Given the src has already been shifted by dst_bits, all we need
1024 * to do is to shift by the difference.
1027 result
= LLVMBuildLShr(builder
,
1029 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1033 } else if (dst_bits
> src_bits
) {
1035 int db
= dst_bits
- src_bits
;
1037 /* Shift left by difference in bits */
1038 result
= LLVMBuildShl(builder
,
1040 lp_build_const_int_vec(gallivm
, src_type
, db
),
1043 if (db
< src_bits
) {
1044 /* Enough bits in src to fill the remainder */
1045 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1047 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1050 result
= LLVMBuildOr(builder
, result
, lower
, "");
1051 } else if (db
> src_bits
) {
1052 /* Need to repeatedly copy src bits to fill remainder in dst */
1055 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1056 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1058 result
= LLVMBuildOr(builder
,
1060 LLVMBuildLShr(builder
, result
, shuv
, ""),
1071 * Convert from memory format to blending format
1073 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1076 convert_to_blend_type(struct gallivm_state
*gallivm
,
1077 unsigned block_size
,
1078 const struct util_format_description
*src_fmt
,
1079 struct lp_type src_type
,
1080 struct lp_type dst_type
,
1081 LLVMValueRef
* src
, // and dst
1084 LLVMValueRef
*dst
= src
;
1085 LLVMBuilderRef builder
= gallivm
->builder
;
1086 struct lp_type blend_type
;
1087 struct lp_type mem_type
;
1089 unsigned pixels
= block_size
/ num_srcs
;
1093 * full custom path for packed floats and srgb formats - none of the later
1094 * functions would do anything useful, and given the lp_type representation they
1095 * can't be fixed. Should really have some SoA blend path for these kind of
1096 * formats rather than hacking them in here.
1098 if (format_expands_to_float_soa(src_fmt
)) {
1099 LLVMValueRef tmpsrc
[4];
1101 * This is pretty suboptimal for this case blending in SoA would be much
1102 * better, since conversion gets us SoA values so need to convert back.
1104 assert(src_type
.width
== 32);
1105 assert(dst_type
.floating
);
1106 assert(dst_type
.width
== 32);
1107 assert(dst_type
.length
% 4 == 0);
1108 assert(num_srcs
% 4 == 0);
1110 for (i
= 0; i
< 4; i
++) {
1113 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1114 LLVMValueRef tmpsoa
[4];
1115 LLVMValueRef tmps
= tmpsrc
[i
];
1116 if (dst_type
.length
== 8) {
1117 LLVMValueRef shuffles
[8];
1119 /* fetch was 4 values but need 8-wide output values */
1120 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1122 * for 8-wide aos transpose would give us wrong order not matching
1123 * incoming converted fs values and mask. ARGH.
1125 for (j
= 0; j
< 4; j
++) {
1126 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1127 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1129 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1130 LLVMConstVector(shuffles
, 8), "");
1132 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1133 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1136 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1138 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1143 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1144 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1146 /* Is the format arithmetic */
1147 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1148 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1150 /* Pad if necessary */
1151 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1152 for (i
= 0; i
< num_srcs
; ++i
) {
1153 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1156 src_type
.length
= dst_type
.length
;
1159 /* Special case for half-floats */
1160 if (mem_type
.width
== 16 && mem_type
.floating
) {
1161 assert(blend_type
.width
== 32 && blend_type
.floating
);
1162 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1170 src_type
.width
= blend_type
.width
* blend_type
.length
;
1171 blend_type
.length
*= pixels
;
1172 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1174 for (i
= 0; i
< num_srcs
; ++i
) {
1175 LLVMValueRef chans
[4];
1176 LLVMValueRef res
= NULL
;
1178 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1180 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1182 unsigned sa
= src_fmt
->channel
[j
].shift
;
1183 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1184 unsigned from_lsb
= j
;
1186 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1189 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1193 /* Extract bits from source */
1194 chans
[j
] = LLVMBuildLShr(builder
,
1196 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1199 chans
[j
] = LLVMBuildAnd(builder
,
1201 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1205 if (src_type
.norm
) {
1206 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1207 blend_type
.width
, chans
[j
], src_type
);
1210 /* Insert bits into correct position */
1211 chans
[j
] = LLVMBuildShl(builder
,
1213 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1219 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1223 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1229 * Convert from blending format to memory format
1231 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1234 convert_from_blend_type(struct gallivm_state
*gallivm
,
1235 unsigned block_size
,
1236 const struct util_format_description
*src_fmt
,
1237 struct lp_type src_type
,
1238 struct lp_type dst_type
,
1239 LLVMValueRef
* src
, // and dst
1242 LLVMValueRef
* dst
= src
;
1244 struct lp_type mem_type
;
1245 struct lp_type blend_type
;
1246 LLVMBuilderRef builder
= gallivm
->builder
;
1247 unsigned pixels
= block_size
/ num_srcs
;
1251 * full custom path for packed floats and srgb formats - none of the later
1252 * functions would do anything useful, and given the lp_type representation they
1253 * can't be fixed. Should really have some SoA blend path for these kind of
1254 * formats rather than hacking them in here.
1256 if (format_expands_to_float_soa(src_fmt
)) {
1258 * This is pretty suboptimal for this case blending in SoA would be much
1259 * better - we need to transpose the AoS values back to SoA values for
1260 * conversion/packing.
1262 assert(src_type
.floating
);
1263 assert(src_type
.width
== 32);
1264 assert(src_type
.length
% 4 == 0);
1265 assert(dst_type
.width
== 32);
1267 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1268 LLVMValueRef tmpsoa
[4], tmpdst
;
1269 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1270 /* really really need SoA here */
1272 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1273 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1276 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1280 if (src_type
.length
== 8) {
1281 LLVMValueRef tmpaos
, shuffles
[8];
1284 * for 8-wide aos transpose has given us wrong order not matching
1285 * output order. HMPF. Also need to split the output values manually.
1287 for (j
= 0; j
< 4; j
++) {
1288 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1289 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1291 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1292 LLVMConstVector(shuffles
, 8), "");
1293 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1294 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1303 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1304 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1306 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1308 /* Special case for half-floats */
1309 if (mem_type
.width
== 16 && mem_type
.floating
) {
1310 int length
= dst_type
.length
;
1311 assert(blend_type
.width
== 32 && blend_type
.floating
);
1313 dst_type
.length
= src_type
.length
;
1315 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1317 dst_type
.length
= length
;
1321 /* Remove any padding */
1322 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1323 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1325 for (i
= 0; i
< num_srcs
; ++i
) {
1326 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1330 /* No bit arithmetic to do */
1335 src_type
.length
= pixels
;
1336 src_type
.width
= blend_type
.length
* blend_type
.width
;
1337 dst_type
.length
= pixels
;
1339 for (i
= 0; i
< num_srcs
; ++i
) {
1340 LLVMValueRef chans
[4];
1341 LLVMValueRef res
= NULL
;
1343 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1345 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1347 unsigned sa
= src_fmt
->channel
[j
].shift
;
1348 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1349 unsigned from_lsb
= j
;
1351 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1354 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1356 for (k
= 0; k
< blend_type
.width
; ++k
) {
1361 chans
[j
] = LLVMBuildLShr(builder
,
1363 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1366 chans
[j
] = LLVMBuildAnd(builder
,
1368 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1371 /* Scale down bits */
1372 if (src_type
.norm
) {
1373 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1374 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1378 chans
[j
] = LLVMBuildShl(builder
,
1380 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1383 sa
+= src_fmt
->channel
[j
].size
;
1388 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1392 assert (dst_type
.width
!= 24);
1394 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1400 * Convert alpha to same blend type as src
1403 convert_alpha(struct gallivm_state
*gallivm
,
1404 struct lp_type row_type
,
1405 struct lp_type alpha_type
,
1406 const unsigned block_size
,
1407 const unsigned block_height
,
1408 const unsigned src_count
,
1409 const unsigned dst_channels
,
1410 const bool pad_inline
,
1411 LLVMValueRef
* src_alpha
)
1413 LLVMBuilderRef builder
= gallivm
->builder
;
1415 unsigned length
= row_type
.length
;
1416 row_type
.length
= alpha_type
.length
;
1418 /* Twiddle the alpha to match pixels */
1419 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1422 * TODO this should use single lp_build_conv call for
1423 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1425 for (i
= 0; i
< block_height
; ++i
) {
1426 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1429 alpha_type
= row_type
;
1430 row_type
.length
= length
;
1432 /* If only one channel we can only need the single alpha value per pixel */
1433 if (src_count
== 1 && dst_channels
== 1) {
1435 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1437 /* If there are more srcs than rows then we need to split alpha up */
1438 if (src_count
> block_height
) {
1439 for (i
= src_count
; i
> 0; --i
) {
1440 unsigned pixels
= block_size
/ src_count
;
1441 unsigned idx
= i
- 1;
1443 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1444 (idx
* pixels
) % 4, pixels
);
1448 /* If there is a src for each pixel broadcast the alpha across whole row */
1449 if (src_count
== block_size
) {
1450 for (i
= 0; i
< src_count
; ++i
) {
1451 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1454 unsigned pixels
= block_size
/ src_count
;
1455 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1456 unsigned alpha_span
= 1;
1457 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1459 /* Check if we need 2 src_alphas for our shuffles */
1460 if (pixels
> alpha_type
.length
) {
1464 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1465 for (j
= 0; j
< row_type
.length
; ++j
) {
1466 if (j
< pixels
* channels
) {
1467 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1469 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1473 for (i
= 0; i
< src_count
; ++i
) {
1474 unsigned idx1
= i
, idx2
= i
;
1476 if (alpha_span
> 1){
1481 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1484 LLVMConstVector(shuffles
, row_type
.length
),
1493 * Generates the blend function for unswizzled colour buffers
1494 * Also generates the read & write from colour buffer
1497 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1499 struct lp_fragment_shader_variant
*variant
,
1500 enum pipe_format out_format
,
1501 unsigned int num_fs
,
1502 struct lp_type fs_type
,
1503 LLVMValueRef
* fs_mask
,
1504 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1505 LLVMValueRef context_ptr
,
1506 LLVMValueRef color_ptr
,
1507 LLVMValueRef stride
,
1508 unsigned partial_mask
,
1511 const unsigned alpha_channel
= 3;
1512 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1513 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1514 const unsigned block_size
= block_width
* block_height
;
1515 const unsigned lp_integer_vector_width
= 128;
1517 LLVMBuilderRef builder
= gallivm
->builder
;
1518 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1519 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1520 LLVMValueRef src_alpha
[4 * 4];
1521 LLVMValueRef src1_alpha
[4 * 4];
1522 LLVMValueRef src_mask
[4 * 4];
1523 LLVMValueRef src
[4 * 4];
1524 LLVMValueRef src1
[4 * 4];
1525 LLVMValueRef dst
[4 * 4];
1526 LLVMValueRef blend_color
;
1527 LLVMValueRef blend_alpha
;
1528 LLVMValueRef i32_zero
;
1529 LLVMValueRef check_mask
;
1530 LLVMValueRef undef_src_val
;
1532 struct lp_build_mask_context mask_ctx
;
1533 struct lp_type mask_type
;
1534 struct lp_type blend_type
;
1535 struct lp_type row_type
;
1536 struct lp_type dst_type
;
1538 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1539 unsigned vector_width
;
1540 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1541 unsigned dst_channels
;
1546 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1548 unsigned dst_alignment
;
1550 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1551 bool has_alpha
= false;
1552 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1553 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1555 const boolean is_1d
= variant
->key
.resource_1d
;
1556 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1558 mask_type
= lp_int32_vec4_type();
1559 mask_type
.length
= fs_type
.length
;
1561 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1562 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1565 /* Do not bother executing code when mask is empty.. */
1567 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1569 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1570 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1573 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1574 lp_build_mask_check(&mask_ctx
);
1577 partial_mask
|= !variant
->opaque
;
1578 i32_zero
= lp_build_const_int32(gallivm
, 0);
1580 #if HAVE_LLVM < 0x0302
1582 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1583 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1585 undef_src_val
= lp_build_zero(gallivm
, fs_type
);
1587 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1591 /* Get type from output format */
1592 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1593 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1595 row_type
.length
= fs_type
.length
;
1596 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1598 /* Compute correct swizzle and count channels */
1599 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1602 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1603 /* Ensure channel is used */
1604 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1608 /* Ensure not already written to (happens in case with GL_ALPHA) */
1609 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1613 /* Ensure we havn't already found all channels */
1614 if (dst_channels
>= out_format_desc
->nr_channels
) {
1618 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1621 if (i
== alpha_channel
) {
1626 if (format_expands_to_float_soa(out_format_desc
)) {
1628 * the code above can't work for layout_other
1629 * for srgb it would sort of work but we short-circuit swizzles, etc.
1630 * as that is done as part of unpack / pack.
1632 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1638 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1641 /* If 3 channels then pad to include alpha for 4 element transpose */
1642 if (dst_channels
== 3 && !has_alpha
) {
1643 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1644 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1647 if (out_format_desc
->nr_channels
== 4) {
1653 * Load shader output
1655 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1656 /* Always load alpha for use in blending */
1659 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1662 alpha
= undef_src_val
;
1665 /* Load each channel */
1666 for (j
= 0; j
< dst_channels
; ++j
) {
1667 assert(swizzle
[j
] < 4);
1669 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1672 fs_src
[i
][j
] = undef_src_val
;
1676 /* If 3 channels then pad to include alpha for 4 element transpose */
1678 * XXX If we include that here maybe could actually use it instead of
1679 * separate alpha for blending?
1681 if (dst_channels
== 3 && !has_alpha
) {
1682 fs_src
[i
][3] = alpha
;
1685 /* We split the row_mask and row_alpha as we want 128bit interleave */
1686 if (fs_type
.length
== 8) {
1687 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1688 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1690 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1691 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1693 src_mask
[i
] = fs_mask
[i
];
1694 src_alpha
[i
] = alpha
;
1697 if (dual_source_blend
) {
1698 /* same as above except different src/dst, skip masks and comments... */
1699 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1702 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1705 alpha
= undef_src_val
;
1708 for (j
= 0; j
< dst_channels
; ++j
) {
1709 assert(swizzle
[j
] < 4);
1711 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1714 fs_src1
[i
][j
] = undef_src_val
;
1717 if (dst_channels
== 3 && !has_alpha
) {
1718 fs_src1
[i
][3] = alpha
;
1720 if (fs_type
.length
== 8) {
1721 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1722 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1724 src1_alpha
[i
] = alpha
;
1729 if (util_format_is_pure_integer(out_format
)) {
1731 * In this case fs_type was really ints or uints disguised as floats,
1734 fs_type
.floating
= 0;
1735 fs_type
.sign
= dst_type
.sign
;
1736 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1737 for (j
= 0; j
< dst_channels
; ++j
) {
1738 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1739 lp_build_vec_type(gallivm
, fs_type
), "");
1741 if (dst_channels
== 3 && !has_alpha
) {
1742 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1743 lp_build_vec_type(gallivm
, fs_type
), "");
1749 * Pixel twiddle from fragment shader order to memory order
1751 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1752 dst_channels
, fs_src
, src
, pad_inline
);
1753 if (dual_source_blend
) {
1754 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1755 fs_src1
, src1
, pad_inline
);
1758 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1759 if (src_count
!= num_fullblock_fs
* src_channels
) {
1760 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1761 row_type
.length
/= ds
;
1762 fs_type
.length
= row_type
.length
;
1765 blend_type
= row_type
;
1766 mask_type
.length
= 4;
1768 /* Convert src to row_type */
1769 if (dual_source_blend
) {
1770 struct lp_type old_row_type
= row_type
;
1771 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1772 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1775 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1778 /* If the rows are not an SSE vector, combine them to become SSE size! */
1779 if ((row_type
.width
* row_type
.length
) % 128) {
1780 unsigned bits
= row_type
.width
* row_type
.length
;
1783 assert(src_count
>= (vector_width
/ bits
));
1785 dst_count
= src_count
/ (vector_width
/ bits
);
1787 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1788 if (dual_source_blend
) {
1789 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1792 row_type
.length
*= combined
;
1793 src_count
/= combined
;
1795 bits
= row_type
.width
* row_type
.length
;
1796 assert(bits
== 128 || bits
== 256);
1801 * Blend Colour conversion
1803 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1804 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1805 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1808 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1810 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1812 * since blending is done with floats, there was no conversion.
1813 * However, the rules according to fixed point renderbuffers still
1814 * apply, that is we must clamp inputs to 0.0/1.0.
1815 * (This would apply to separate alpha conversion too but we currently
1816 * force has_alpha to be true.)
1817 * TODO: should skip this with "fake" blend, since post-blend conversion
1818 * will clamp anyway.
1819 * TODO: could also skip this if fragment color clamping is enabled. We
1820 * don't support it natively so it gets baked into the shader however, so
1821 * can't really tell here.
1823 struct lp_build_context f32_bld
;
1824 assert(row_type
.floating
);
1825 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1826 for (i
= 0; i
< src_count
; i
++) {
1827 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
1829 if (dual_source_blend
) {
1830 for (i
= 0; i
< src_count
; i
++) {
1831 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
1834 /* probably can't be different than row_type but better safe than sorry... */
1835 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1836 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1840 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1842 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1843 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1845 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1846 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1848 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1849 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1855 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1857 if (src_count
< block_height
) {
1858 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1859 } else if (src_count
> block_height
) {
1860 for (i
= src_count
; i
> 0; --i
) {
1861 unsigned pixels
= block_size
/ src_count
;
1862 unsigned idx
= i
- 1;
1864 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1865 (idx
* pixels
) % 4, pixels
);
1869 assert(mask_type
.width
== 32);
1871 for (i
= 0; i
< src_count
; ++i
) {
1872 unsigned pixels
= block_size
/ src_count
;
1873 unsigned pixel_width
= row_type
.width
* dst_channels
;
1875 if (pixel_width
== 24) {
1876 mask_type
.width
= 8;
1877 mask_type
.length
= vector_width
/ mask_type
.width
;
1879 mask_type
.length
= pixels
;
1880 mask_type
.width
= row_type
.width
* dst_channels
;
1882 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1884 mask_type
.length
*= dst_channels
;
1885 mask_type
.width
/= dst_channels
;
1888 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1889 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1896 struct lp_type alpha_type
= fs_type
;
1897 alpha_type
.length
= 4;
1898 convert_alpha(gallivm
, row_type
, alpha_type
,
1899 block_size
, block_height
,
1900 src_count
, dst_channels
,
1901 pad_inline
, src_alpha
);
1902 if (dual_source_blend
) {
1903 convert_alpha(gallivm
, row_type
, alpha_type
,
1904 block_size
, block_height
,
1905 src_count
, dst_channels
,
1906 pad_inline
, src1_alpha
);
1912 * Load dst from memory
1914 if (src_count
< block_height
) {
1915 dst_count
= block_height
;
1917 dst_count
= src_count
;
1920 dst_type
.length
*= block_size
/ dst_count
;
1922 if (format_expands_to_float_soa(out_format_desc
)) {
1924 * we need multiple values at once for the conversion, so can as well
1925 * load them vectorized here too instead of concatenating later.
1926 * (Still need concatenation later for 8-wide vectors).
1928 dst_count
= block_height
;
1929 dst_type
.length
= block_width
;
1933 * Compute the alignment of the destination pointer in bytes
1934 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1935 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1936 * 1d tex but can't distinguish here) so need to stick with per-pixel
1937 * alignment in this case.
1940 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1943 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
1945 /* Force power-of-two alignment by extracting only the least-significant-bit */
1946 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
1948 * Resource base and stride pointers are aligned to 16 bytes, so that's
1949 * the maximum alignment we can guarantee
1951 dst_alignment
= MIN2(16, dst_alignment
);
1954 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
1955 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
1956 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
1957 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
1962 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1963 dst
, dst_type
, dst_count
, dst_alignment
);
1968 * Convert from dst/output format to src/blending format.
1970 * This is necessary as we can only read 1 row from memory at a time,
1971 * so the minimum dst_count will ever be at this point is 4.
1973 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1974 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1975 * on all 16 pixels in that single vector at once.
1977 if (dst_count
> src_count
) {
1978 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
1984 /* XXX this is broken for RGB8 formats -
1985 * they get expanded from 12 to 16 elements (to include alpha)
1986 * by convert_to_blend_type then reduced to 15 instead of 12
1987 * by convert_from_blend_type (a simple fix though breaks A8...).
1988 * R16G16B16 also crashes differently however something going wrong
1989 * inside llvm handling npot vector sizes seemingly.
1990 * It seems some cleanup could be done here (like skipping conversion/blend
1993 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
1996 * FIXME: Really should get logic ops / masks out of generic blend / row
1997 * format. Logic ops will definitely not work on the blend float format
1998 * used for SRGB here and I think OpenGL expects this to work as expected
1999 * (that is incoming values converted to srgb then logic op applied).
2001 for (i
= 0; i
< src_count
; ++i
) {
2002 dst
[i
] = lp_build_blend_aos(gallivm
,
2003 &variant
->key
.blend
,
2008 has_alpha
? NULL
: src_alpha
[i
],
2010 has_alpha
? NULL
: src1_alpha
[i
],
2012 partial_mask
? src_mask
[i
] : NULL
,
2014 has_alpha
? NULL
: blend_alpha
,
2016 pad_inline
? 4 : dst_channels
);
2019 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
2021 /* Split the blend rows back to memory rows */
2022 if (dst_count
> src_count
) {
2023 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2025 if (src_count
== 1) {
2026 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2027 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2029 row_type
.length
/= 2;
2033 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2034 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2035 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2036 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2038 row_type
.length
/= 2;
2043 * Store blend result to memory
2046 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2047 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2050 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2051 dst
, dst_type
, dst_count
, dst_alignment
);
2055 lp_build_mask_end(&mask_ctx
);
2061 * Generate the runtime callable function for the whole fragment pipeline.
2062 * Note that the function which we generate operates on a block of 16
2063 * pixels at at time. The block contains 2x2 quads. Each quad contains
2067 generate_fragment(struct llvmpipe_context
*lp
,
2068 struct lp_fragment_shader
*shader
,
2069 struct lp_fragment_shader_variant
*variant
,
2070 unsigned partial_mask
)
2072 struct gallivm_state
*gallivm
= variant
->gallivm
;
2073 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2074 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2075 char func_name
[256];
2076 struct lp_type fs_type
;
2077 struct lp_type blend_type
;
2078 LLVMTypeRef fs_elem_type
;
2079 LLVMTypeRef blend_vec_type
;
2080 LLVMTypeRef arg_types
[13];
2081 LLVMTypeRef func_type
;
2082 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2083 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2084 LLVMValueRef context_ptr
;
2087 LLVMValueRef a0_ptr
;
2088 LLVMValueRef dadx_ptr
;
2089 LLVMValueRef dady_ptr
;
2090 LLVMValueRef color_ptr_ptr
;
2091 LLVMValueRef stride_ptr
;
2092 LLVMValueRef depth_ptr
;
2093 LLVMValueRef depth_stride
;
2094 LLVMValueRef mask_input
;
2095 LLVMValueRef thread_data_ptr
;
2096 LLVMBasicBlockRef block
;
2097 LLVMBuilderRef builder
;
2098 struct lp_build_sampler_soa
*sampler
;
2099 struct lp_build_interp_soa_context interp
;
2100 LLVMValueRef fs_mask
[16 / 4];
2101 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2102 LLVMValueRef function
;
2103 LLVMValueRef facing
;
2108 boolean cbuf0_write_all
;
2109 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2110 util_blend_state_is_dual(&key
->blend
, 0);
2112 assert(lp_native_vector_width
/ 32 >= 4);
2114 /* Adjust color input interpolation according to flatshade state:
2116 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2117 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2118 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2120 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2122 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2126 /* check if writes to cbuf[0] are to be copied to all cbufs */
2127 cbuf0_write_all
= FALSE
;
2128 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
2129 if (shader
->info
.base
.properties
[i
].name
==
2130 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
2131 cbuf0_write_all
= TRUE
;
2136 /* TODO: actually pick these based on the fs and color buffer
2137 * characteristics. */
2139 memset(&fs_type
, 0, sizeof fs_type
);
2140 fs_type
.floating
= TRUE
; /* floating point values */
2141 fs_type
.sign
= TRUE
; /* values are signed */
2142 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2143 fs_type
.width
= 32; /* 32-bit float */
2144 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2146 memset(&blend_type
, 0, sizeof blend_type
);
2147 blend_type
.floating
= FALSE
; /* values are integers */
2148 blend_type
.sign
= FALSE
; /* values are unsigned */
2149 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2150 blend_type
.width
= 8; /* 8-bit ubyte values */
2151 blend_type
.length
= 16; /* 16 elements per vector */
2154 * Generate the function prototype. Any change here must be reflected in
2155 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2158 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2160 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2162 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2163 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2165 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2166 arg_types
[1] = int32_type
; /* x */
2167 arg_types
[2] = int32_type
; /* y */
2168 arg_types
[3] = int32_type
; /* facing */
2169 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2170 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2171 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2172 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2173 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2174 arg_types
[9] = int32_type
; /* mask_input */
2175 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2176 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2177 arg_types
[12] = int32_type
; /* depth_stride */
2179 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2180 arg_types
, Elements(arg_types
), 0);
2182 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2183 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2185 variant
->function
[partial_mask
] = function
;
2187 /* XXX: need to propagate noalias down into color param now we are
2188 * passing a pointer-to-pointer?
2190 for(i
= 0; i
< Elements(arg_types
); ++i
)
2191 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2192 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2194 context_ptr
= LLVMGetParam(function
, 0);
2195 x
= LLVMGetParam(function
, 1);
2196 y
= LLVMGetParam(function
, 2);
2197 facing
= LLVMGetParam(function
, 3);
2198 a0_ptr
= LLVMGetParam(function
, 4);
2199 dadx_ptr
= LLVMGetParam(function
, 5);
2200 dady_ptr
= LLVMGetParam(function
, 6);
2201 color_ptr_ptr
= LLVMGetParam(function
, 7);
2202 depth_ptr
= LLVMGetParam(function
, 8);
2203 mask_input
= LLVMGetParam(function
, 9);
2204 thread_data_ptr
= LLVMGetParam(function
, 10);
2205 stride_ptr
= LLVMGetParam(function
, 11);
2206 depth_stride
= LLVMGetParam(function
, 12);
2208 lp_build_name(context_ptr
, "context");
2209 lp_build_name(x
, "x");
2210 lp_build_name(y
, "y");
2211 lp_build_name(a0_ptr
, "a0");
2212 lp_build_name(dadx_ptr
, "dadx");
2213 lp_build_name(dady_ptr
, "dady");
2214 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2215 lp_build_name(depth_ptr
, "depth");
2216 lp_build_name(thread_data_ptr
, "thread_data");
2217 lp_build_name(mask_input
, "mask_input");
2218 lp_build_name(stride_ptr
, "stride_ptr");
2219 lp_build_name(depth_stride
, "depth_stride");
2225 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2226 builder
= gallivm
->builder
;
2228 LLVMPositionBuilderAtEnd(builder
, block
);
2230 /* code generated texture sampling */
2231 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
2233 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2234 /* for 1d resources only run "upper half" of stamp */
2235 if (key
->resource_1d
)
2239 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2240 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2241 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2242 num_loop
, "mask_store");
2243 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2246 * The shader input interpolation info is not explicitely baked in the
2247 * shader key, but everything it derives from (TGSI, and flatshade) is
2248 * already included in the shader key.
2250 lp_build_interp_soa_init(&interp
,
2252 shader
->info
.base
.num_inputs
,
2254 shader
->info
.base
.pixel_center_integer
,
2256 a0_ptr
, dadx_ptr
, dady_ptr
,
2259 for (i
= 0; i
< num_fs
; i
++) {
2261 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2262 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2263 &indexi
, 1, "mask_ptr");
2266 mask
= generate_quad_mask(gallivm
, fs_type
,
2267 i
*fs_type
.length
/4, mask_input
);
2270 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2272 LLVMBuildStore(builder
, mask
, mask_ptr
);
2275 generate_fs_loop(gallivm
,
2283 mask_store
, /* output */
2290 for (i
= 0; i
< num_fs
; i
++) {
2291 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2292 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2294 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2295 /* This is fucked up need to reorganize things */
2296 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2297 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2298 ptr
= LLVMBuildGEP(builder
,
2299 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2301 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2304 if (dual_source_blend
) {
2305 /* only support one dual source blend target hence always use output 1 */
2306 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2307 ptr
= LLVMBuildGEP(builder
,
2308 color_store
[1][chan
],
2310 fs_out_color
[1][chan
][i
] = ptr
;
2316 sampler
->destroy(sampler
);
2318 /* Loop over color outputs / color buffers to do blending.
2320 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2321 LLVMValueRef color_ptr
;
2322 LLVMValueRef stride
;
2323 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2325 boolean do_branch
= ((key
->depth
.enabled
2326 || key
->stencil
[0].enabled
2327 || key
->alpha
.enabled
)
2328 && !shader
->info
.base
.uses_kill
);
2330 color_ptr
= LLVMBuildLoad(builder
,
2331 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2334 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2336 stride
= LLVMBuildLoad(builder
,
2337 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2340 generate_unswizzled_blend(gallivm
, cbuf
, variant
, key
->cbuf_format
[cbuf
],
2341 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2342 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2345 LLVMBuildRetVoid(builder
);
2347 gallivm_verify_function(gallivm
, function
);
2349 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2354 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2358 debug_printf("fs variant %p:\n", (void *) key
);
2360 if (key
->flatshade
) {
2361 debug_printf("flatshade = 1\n");
2363 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2364 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2366 if (key
->depth
.enabled
) {
2367 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2368 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2369 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2372 for (i
= 0; i
< 2; ++i
) {
2373 if (key
->stencil
[i
].enabled
) {
2374 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2375 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2376 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2377 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2378 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2379 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2383 if (key
->alpha
.enabled
) {
2384 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2387 if (key
->occlusion_count
) {
2388 debug_printf("occlusion_count = 1\n");
2391 if (key
->blend
.logicop_enable
) {
2392 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2394 else if (key
->blend
.rt
[0].blend_enable
) {
2395 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2396 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2397 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2398 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2399 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2400 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2402 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2403 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2404 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2405 debug_printf("sampler[%u] = \n", i
);
2406 debug_printf(" .wrap = %s %s %s\n",
2407 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2408 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2409 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2410 debug_printf(" .min_img_filter = %s\n",
2411 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2412 debug_printf(" .min_mip_filter = %s\n",
2413 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2414 debug_printf(" .mag_img_filter = %s\n",
2415 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2416 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2417 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2418 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2419 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2420 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2421 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2422 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2424 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2425 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2426 debug_printf("texture[%u] = \n", i
);
2427 debug_printf(" .format = %s\n",
2428 util_format_name(texture
->format
));
2429 debug_printf(" .target = %s\n",
2430 util_dump_tex_target(texture
->target
, TRUE
));
2431 debug_printf(" .level_zero_only = %u\n",
2432 texture
->level_zero_only
);
2433 debug_printf(" .pot = %u %u %u\n",
2435 texture
->pot_height
,
2436 texture
->pot_depth
);
2442 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2444 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2445 variant
->shader
->no
, variant
->no
);
2446 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2447 dump_fs_variant_key(&variant
->key
);
2448 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2454 * Generate a new fragment shader variant from the shader code and
2455 * other state indicated by the key.
2457 static struct lp_fragment_shader_variant
*
2458 generate_variant(struct llvmpipe_context
*lp
,
2459 struct lp_fragment_shader
*shader
,
2460 const struct lp_fragment_shader_variant_key
*key
)
2462 struct lp_fragment_shader_variant
*variant
;
2463 const struct util_format_description
*cbuf0_format_desc
;
2464 boolean fullcolormask
;
2466 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2470 variant
->gallivm
= gallivm_create();
2471 if (!variant
->gallivm
) {
2476 variant
->shader
= shader
;
2477 variant
->list_item_global
.base
= variant
;
2478 variant
->list_item_local
.base
= variant
;
2479 variant
->no
= shader
->variants_created
++;
2481 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2484 * Determine whether we are touching all channels in the color buffer.
2486 fullcolormask
= FALSE
;
2487 if (key
->nr_cbufs
== 1) {
2488 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2489 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2493 !key
->blend
.logicop_enable
&&
2494 !key
->blend
.rt
[0].blend_enable
&&
2496 !key
->stencil
[0].enabled
&&
2497 !key
->alpha
.enabled
&&
2498 !key
->depth
.enabled
&&
2499 !shader
->info
.base
.uses_kill
2502 if ((shader
->info
.base
.num_tokens
<= 1) &&
2503 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2504 variant
->ps_inv_multiplier
= 0;
2506 variant
->ps_inv_multiplier
= 1;
2509 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2510 lp_debug_fs_variant(variant
);
2513 lp_jit_init_types(variant
);
2515 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2516 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2518 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2519 if (variant
->opaque
) {
2520 /* Specialized shader, which doesn't need to read the color buffer. */
2521 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2526 * Compile everything
2529 gallivm_compile_module(variant
->gallivm
);
2531 if (variant
->function
[RAST_EDGE_TEST
]) {
2532 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2533 gallivm_jit_function(variant
->gallivm
,
2534 variant
->function
[RAST_EDGE_TEST
]);
2537 if (variant
->function
[RAST_WHOLE
]) {
2538 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2539 gallivm_jit_function(variant
->gallivm
,
2540 variant
->function
[RAST_WHOLE
]);
2541 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2542 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2550 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2551 const struct pipe_shader_state
*templ
)
2553 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2554 struct lp_fragment_shader
*shader
;
2556 int nr_sampler_views
;
2559 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2563 shader
->no
= fs_no
++;
2564 make_empty_list(&shader
->variants
);
2566 /* get/save the summary info for this shader */
2567 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2569 /* we need to keep a local copy of the tokens */
2570 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2572 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2573 if (shader
->draw_data
== NULL
) {
2574 FREE((void *) shader
->base
.tokens
);
2579 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2580 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2582 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2583 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2585 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2586 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2587 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2589 switch (shader
->info
.base
.input_interpolate
[i
]) {
2590 case TGSI_INTERPOLATE_CONSTANT
:
2591 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2593 case TGSI_INTERPOLATE_LINEAR
:
2594 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2596 case TGSI_INTERPOLATE_PERSPECTIVE
:
2597 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2599 case TGSI_INTERPOLATE_COLOR
:
2600 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2607 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2608 case TGSI_SEMANTIC_FACE
:
2609 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2611 case TGSI_SEMANTIC_POSITION
:
2612 /* Position was already emitted above
2614 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2615 shader
->inputs
[i
].src_index
= 0;
2619 shader
->inputs
[i
].src_index
= i
+1;
2622 if (LP_DEBUG
& DEBUG_TGSI
) {
2624 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2625 shader
->no
, (void *) shader
);
2626 tgsi_dump(templ
->tokens
, 0);
2627 debug_printf("usage masks:\n");
2628 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2629 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2630 debug_printf(" IN[%u].%s%s%s%s\n",
2632 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2633 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2634 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2635 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2645 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2647 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2649 if (llvmpipe
->fs
== fs
)
2652 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2654 draw_bind_fragment_shader(llvmpipe
->draw
,
2655 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2657 llvmpipe
->dirty
|= LP_NEW_FS
;
2662 * Remove shader variant from two lists: the shader's variant list
2663 * and the context's variant list.
2666 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2667 struct lp_fragment_shader_variant
*variant
)
2671 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2672 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2673 " #%u v total cached #%u\n",
2674 variant
->shader
->no
,
2676 variant
->shader
->variants_created
,
2677 variant
->shader
->variants_cached
,
2678 lp
->nr_fs_variants
);
2681 /* free all the variant's JIT'd functions */
2682 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2683 if (variant
->function
[i
]) {
2684 gallivm_free_function(variant
->gallivm
,
2685 variant
->function
[i
],
2686 variant
->jit_function
[i
]);
2690 gallivm_destroy(variant
->gallivm
);
2692 /* remove from shader's list */
2693 remove_from_list(&variant
->list_item_local
);
2694 variant
->shader
->variants_cached
--;
2696 /* remove from context's list */
2697 remove_from_list(&variant
->list_item_global
);
2698 lp
->nr_fs_variants
--;
2699 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2706 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2708 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2709 struct lp_fragment_shader
*shader
= fs
;
2710 struct lp_fs_variant_list_item
*li
;
2712 assert(fs
!= llvmpipe
->fs
);
2715 * XXX: we need to flush the context until we have some sort of reference
2716 * counting in fragment shaders as they may still be binned
2717 * Flushing alone might not sufficient we need to wait on it too.
2719 llvmpipe_finish(pipe
, __FUNCTION__
);
2721 /* Delete all the variants */
2722 li
= first_elem(&shader
->variants
);
2723 while(!at_end(&shader
->variants
, li
)) {
2724 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2725 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2729 /* Delete draw module's data */
2730 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2732 assert(shader
->variants_cached
== 0);
2733 FREE((void *) shader
->base
.tokens
);
2740 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2741 uint shader
, uint index
,
2742 struct pipe_constant_buffer
*cb
)
2744 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2745 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2747 assert(shader
< PIPE_SHADER_TYPES
);
2748 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2750 /* note: reference counting */
2751 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2753 if (shader
== PIPE_SHADER_VERTEX
||
2754 shader
== PIPE_SHADER_GEOMETRY
) {
2755 /* Pass the constants to the 'draw' module */
2756 const unsigned size
= cb
? cb
->buffer_size
: 0;
2760 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2762 else if (cb
&& cb
->user_buffer
) {
2763 data
= (ubyte
*) cb
->user_buffer
;
2770 data
+= cb
->buffer_offset
;
2772 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2776 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2778 if (cb
&& cb
->user_buffer
) {
2779 pipe_resource_reference(&constants
, NULL
);
2785 * Return the blend factor equivalent to a destination alpha of one.
2787 static INLINE
unsigned
2788 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2791 case PIPE_BLENDFACTOR_DST_ALPHA
:
2792 return PIPE_BLENDFACTOR_ONE
;
2793 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2794 return PIPE_BLENDFACTOR_ZERO
;
2795 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2797 return PIPE_BLENDFACTOR_ZERO
;
2799 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2807 * We need to generate several variants of the fragment pipeline to match
2808 * all the combinations of the contributing state atoms.
2810 * TODO: there is actually no reason to tie this to context state -- the
2811 * generated code could be cached globally in the screen.
2814 make_variant_key(struct llvmpipe_context
*lp
,
2815 struct lp_fragment_shader
*shader
,
2816 struct lp_fragment_shader_variant_key
*key
)
2820 memset(key
, 0, shader
->variant_key_size
);
2822 if (lp
->framebuffer
.zsbuf
) {
2823 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2824 const struct util_format_description
*zsbuf_desc
=
2825 util_format_description(zsbuf_format
);
2827 if (lp
->depth_stencil
->depth
.enabled
&&
2828 util_format_has_depth(zsbuf_desc
)) {
2829 key
->zsbuf_format
= zsbuf_format
;
2830 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2832 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2833 util_format_has_stencil(zsbuf_desc
)) {
2834 key
->zsbuf_format
= zsbuf_format
;
2835 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2837 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2838 key
->resource_1d
= TRUE
;
2842 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2843 if (!lp
->framebuffer
.nr_cbufs
||
2844 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2845 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2847 if(key
->alpha
.enabled
)
2848 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2849 /* alpha.ref_value is passed in jit_context */
2851 key
->flatshade
= lp
->rasterizer
->flatshade
;
2852 if (lp
->active_occlusion_queries
) {
2853 key
->occlusion_count
= TRUE
;
2856 if (lp
->framebuffer
.nr_cbufs
) {
2857 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2860 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2862 if (!key
->blend
.independent_blend_enable
) {
2863 /* we always need independent blend otherwise the fixups below won't work */
2864 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2865 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2867 key
->blend
.independent_blend_enable
= 1;
2870 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2871 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2872 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2873 const struct util_format_description
*format_desc
;
2875 key
->cbuf_format
[i
] = format
;
2878 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2879 * mixing of 2d textures with height 1 and 1d textures, so make sure
2880 * we pick 1d if any cbuf or zsbuf is 1d.
2882 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[0]->texture
)) {
2883 key
->resource_1d
= TRUE
;
2886 format_desc
= util_format_description(format
);
2887 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2888 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2891 * Mask out color channels not present in the color buffer.
2893 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2896 * Disable blend for integer formats.
2898 if (util_format_is_pure_integer(format
)) {
2899 blend_rt
->blend_enable
= 0;
2903 * Our swizzled render tiles always have an alpha channel, but the linear
2904 * render target format often does not, so force here the dst alpha to be
2907 * This is not a mere optimization. Wrong results will be produced if the
2908 * dst alpha is used, the dst format does not have alpha, and the previous
2909 * rendering was not flushed from the swizzled to linear buffer. For
2910 * example, NonPowTwo DCT.
2912 * TODO: This should be generalized to all channels for better
2913 * performance, but only alpha causes correctness issues.
2915 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2917 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2918 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2919 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
2920 boolean clamped_zero
= !util_format_is_float(format
) &&
2921 !util_format_is_snorm(format
);
2922 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
,
2924 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
,
2926 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2927 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2928 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2932 /* This value will be the same for all the variants of a given shader:
2934 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2936 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2937 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2938 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
2939 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
2944 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2945 * are dx10-style? Can't really have mixed opcodes, at least not
2946 * if we want to skip the holes here (without rescanning tgsi).
2948 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2949 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2950 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2951 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
2952 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2953 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2958 key
->nr_sampler_views
= key
->nr_samplers
;
2959 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2960 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2961 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2962 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2971 * Update fragment shader state. This is called just prior to drawing
2972 * something when some fragment-related state has changed.
2975 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
2977 struct lp_fragment_shader
*shader
= lp
->fs
;
2978 struct lp_fragment_shader_variant_key key
;
2979 struct lp_fragment_shader_variant
*variant
= NULL
;
2980 struct lp_fs_variant_list_item
*li
;
2982 make_variant_key(lp
, shader
, &key
);
2984 /* Search the variants for one which matches the key */
2985 li
= first_elem(&shader
->variants
);
2986 while(!at_end(&shader
->variants
, li
)) {
2987 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
2995 /* Move this variant to the head of the list to implement LRU
2996 * deletion of shader's when we have too many.
2998 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3001 /* variant not found, create it now */
3004 unsigned variants_to_cull
;
3007 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3010 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3013 /* First, check if we've exceeded the max number of shader variants.
3014 * If so, free 25% of them (the least recently used ones).
3016 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
3018 if (variants_to_cull
||
3019 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3020 struct pipe_context
*pipe
= &lp
->pipe
;
3023 * XXX: we need to flush the context until we have some sort of
3024 * reference counting in fragment shaders as they may still be binned
3025 * Flushing alone might not be sufficient we need to wait on it too.
3027 llvmpipe_finish(pipe
, __FUNCTION__
);
3030 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3031 * number of shader variants (potentially all of them) could be
3032 * pending for destruction on flush.
3035 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3036 struct lp_fs_variant_list_item
*item
;
3037 if (is_empty_list(&lp
->fs_variants_list
)) {
3040 item
= last_elem(&lp
->fs_variants_list
);
3043 llvmpipe_remove_shader_variant(lp
, item
->base
);
3048 * Generate the new variant.
3051 variant
= generate_variant(lp
, shader
, &key
);
3054 LP_COUNT_ADD(llvm_compile_time
, dt
);
3055 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3057 llvmpipe_variant_count
++;
3059 /* Put the new variant into the list */
3061 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3062 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3063 lp
->nr_fs_variants
++;
3064 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3065 shader
->variants_cached
++;
3069 /* Bind this variant */
3070 lp_setup_set_fs_variant(lp
->setup
, variant
);
3078 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3080 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3081 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3082 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3084 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3088 * Rasterization is disabled if there is no pixel shader and
3089 * both depth and stencil testing are disabled:
3090 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3093 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
3095 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
3098 !lp
->depth_stencil
->depth
.enabled
&&
3099 !lp
->depth_stencil
->stencil
[0].enabled
);