1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
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 VMWARE 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
, num_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
||
300 key
->blend
.alpha_to_coverage
||
301 shader
->info
.base
.uses_kill
) {
302 /* With alpha test and kill, can do the depth test early
303 * and hopefully eliminate some quads. But need to do a
304 * special deferred depth write once the final mask value
305 * is known. This only works though if there's either no
306 * stencil test or the stencil value isn't written.
308 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
309 (key
->stencil
[1].enabled
&&
310 key
->stencil
[1].writemask
)))
311 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
313 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
316 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
319 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
322 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
323 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
324 (key
->stencil
[1].enabled
&&
325 key
->stencil
[1].writemask
))))
326 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
333 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
334 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
336 vec_type
= lp_build_vec_type(gallivm
, type
);
338 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
339 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
341 lp_build_for_loop_begin(&loop_state
, gallivm
,
342 lp_build_const_int32(gallivm
, 0),
345 lp_build_const_int32(gallivm
, 1));
347 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
348 &loop_state
.counter
, 1, "mask_ptr");
349 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
351 memset(outputs
, 0, sizeof outputs
);
353 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
354 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
355 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
356 lp_build_vec_type(gallivm
,
361 if (dual_source_blend
) {
362 assert(key
->nr_cbufs
<= 1);
363 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
364 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
365 lp_build_vec_type(gallivm
,
372 /* 'mask' will control execution based on quad's pixel alive/killed state */
373 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
375 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
376 lp_build_mask_check(&mask
);
378 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
381 if (depth_mode
& EARLY_DEPTH_TEST
) {
382 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
383 zs_format_desc
, key
->resource_1d
,
384 depth_ptr
, depth_stride
,
385 &z_fb
, &s_fb
, loop_state
.counter
);
386 lp_build_depth_stencil_test(gallivm
,
398 if (depth_mode
& EARLY_DEPTH_WRITE
) {
399 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
400 zs_format_desc
, key
->resource_1d
,
401 NULL
, NULL
, NULL
, loop_state
.counter
,
402 depth_ptr
, depth_stride
,
406 * Note mask check if stencil is enabled must be after ds write not after
407 * stencil test otherwise new stencil values may not get written if all
408 * fragments got killed by depth/stencil test.
410 if (!simple_shader
&& key
->stencil
[0].enabled
)
411 lp_build_mask_check(&mask
);
414 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
416 /* Build the actual shader */
417 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
418 consts_ptr
, num_consts_ptr
, &system_values
,
420 outputs
, sampler
, &shader
->info
.base
, NULL
);
423 if (key
->alpha
.enabled
) {
424 int color0
= find_output_by_semantic(&shader
->info
.base
,
428 if (color0
!= -1 && outputs
[color0
][3]) {
429 const struct util_format_description
*cbuf_format_desc
;
430 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
431 LLVMValueRef alpha_ref_value
;
433 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
434 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
436 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
438 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
439 &mask
, alpha
, alpha_ref_value
,
440 (depth_mode
& LATE_DEPTH_TEST
) != 0);
444 /* Emulate Alpha to Coverage with Alpha test */
445 if (key
->blend
.alpha_to_coverage
) {
446 int color0
= find_output_by_semantic(&shader
->info
.base
,
450 if (color0
!= -1 && outputs
[color0
][3]) {
451 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
453 lp_build_alpha_to_coverage(gallivm
, type
,
455 (depth_mode
& LATE_DEPTH_TEST
) != 0);
460 if (depth_mode
& LATE_DEPTH_TEST
) {
461 int pos0
= find_output_by_semantic(&shader
->info
.base
,
462 TGSI_SEMANTIC_POSITION
,
465 if (pos0
!= -1 && outputs
[pos0
][2]) {
466 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
469 * Clamp according to ARB_depth_clamp semantics.
471 if (key
->depth_clamp
) {
472 LLVMValueRef viewport
, min_depth
, max_depth
;
473 LLVMValueRef viewport_index
;
474 struct lp_build_context f32_bld
;
476 assert(type
.floating
);
477 lp_build_context_init(&f32_bld
, gallivm
, type
);
480 * Assumes clamping of the viewport index will occur in setup/gs. Value
481 * is passed through the rasterization stage via lp_rast_shader_inputs.
483 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
486 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
490 * Load the min and max depth from the lp_jit_context.viewports
491 * array of lp_jit_viewport structures.
493 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
495 /* viewports[viewport_index].min_depth */
496 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
497 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
),
499 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
501 /* viewports[viewport_index].max_depth */
502 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
503 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
),
505 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
508 * Clamp to the min and max depth values for the given viewport.
510 z
= lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
514 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
515 zs_format_desc
, key
->resource_1d
,
516 depth_ptr
, depth_stride
,
517 &z_fb
, &s_fb
, loop_state
.counter
);
519 lp_build_depth_stencil_test(gallivm
,
531 if (depth_mode
& LATE_DEPTH_WRITE
) {
532 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
533 zs_format_desc
, key
->resource_1d
,
534 NULL
, NULL
, NULL
, loop_state
.counter
,
535 depth_ptr
, depth_stride
,
539 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
540 (depth_mode
& LATE_DEPTH_WRITE
))
542 /* Need to apply a reduced mask to the depth write. Reload the
543 * depth value, update from zs_value with the new mask value and
546 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
547 zs_format_desc
, key
->resource_1d
,
548 &mask
, z_fb
, s_fb
, loop_state
.counter
,
549 depth_ptr
, depth_stride
,
555 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
557 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
558 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
559 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
561 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
562 if(outputs
[attrib
][chan
]) {
563 /* XXX: just initialize outputs to point at colors[] and
566 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
567 LLVMValueRef color_ptr
;
568 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
569 &loop_state
.counter
, 1, "");
570 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
571 LLVMBuildStore(builder
, out
, color_ptr
);
577 if (key
->occlusion_count
) {
578 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
579 lp_build_name(counter
, "counter");
580 lp_build_occlusion_count(gallivm
, type
,
581 lp_build_mask_value(&mask
), counter
);
584 mask_val
= lp_build_mask_end(&mask
);
585 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
586 lp_build_for_loop_end(&loop_state
);
591 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
593 * Fragment Shader outputs pixels in small 2x2 blocks
594 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
596 * However in memory pixels are stored in rows
597 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
599 * @param type fragment shader type (4x or 8x float)
600 * @param num_fs number of fs_src
601 * @param is_1d whether we're outputting to a 1d resource
602 * @param dst_channels number of output channels
603 * @param fs_src output from fragment shader
604 * @param dst pointer to store result
605 * @param pad_inline is channel padding inline or at end of row
606 * @return the number of dsts
609 generate_fs_twiddle(struct gallivm_state
*gallivm
,
612 unsigned dst_channels
,
613 LLVMValueRef fs_src
[][4],
617 LLVMValueRef src
[16];
623 unsigned pixels
= type
.length
/ 4;
624 unsigned reorder_group
;
625 unsigned src_channels
;
629 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
630 src_count
= num_fs
* src_channels
;
632 assert(pixels
== 2 || pixels
== 1);
633 assert(num_fs
* src_channels
<= Elements(src
));
636 * Transpose from SoA -> AoS
638 for (i
= 0; i
< num_fs
; ++i
) {
639 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
643 * Pick transformation options
650 if (dst_channels
== 1) {
656 } else if (dst_channels
== 2) {
660 } else if (dst_channels
> 2) {
667 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
673 * Split the src in half
676 for (i
= num_fs
; i
> 0; --i
) {
677 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
678 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
686 * Ensure pixels are in memory order
689 /* Twiddle pixels by reordering the array, e.g.:
691 * src_count = 8 -> 0 2 1 3 4 6 5 7
692 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
694 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
696 for (i
= 0; i
< src_count
; ++i
) {
697 unsigned group
= i
/ reorder_group
;
698 unsigned block
= (group
/ 4) * 4 * reorder_group
;
699 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
702 } else if (twiddle
) {
703 /* Twiddle pixels across elements of array */
704 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
707 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
711 * Moves any padding between pixels to the end
712 * e.g. RGBXRGBX -> RGBRGBXX
715 unsigned char swizzles
[16];
716 unsigned elems
= pixels
* dst_channels
;
718 for (i
= 0; i
< type
.length
; ++i
) {
720 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
722 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
725 for (i
= 0; i
< src_count
; ++i
) {
726 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
735 * Load an unswizzled block of pixels from memory
738 load_unswizzled_block(struct gallivm_state
*gallivm
,
739 LLVMValueRef base_ptr
,
741 unsigned block_width
,
742 unsigned block_height
,
744 struct lp_type dst_type
,
746 unsigned dst_alignment
)
748 LLVMBuilderRef builder
= gallivm
->builder
;
749 unsigned row_size
= dst_count
/ block_height
;
752 /* Ensure block exactly fits into dst */
753 assert((block_width
* block_height
) % dst_count
== 0);
755 for (i
= 0; i
< dst_count
; ++i
) {
756 unsigned x
= i
% row_size
;
757 unsigned y
= i
/ row_size
;
759 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
760 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
763 LLVMValueRef dst_ptr
;
765 gep
[0] = lp_build_const_int32(gallivm
, 0);
766 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
768 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
769 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
771 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
773 lp_set_load_alignment(dst
[i
], dst_alignment
);
779 * Store an unswizzled block of pixels to memory
782 store_unswizzled_block(struct gallivm_state
*gallivm
,
783 LLVMValueRef base_ptr
,
785 unsigned block_width
,
786 unsigned block_height
,
788 struct lp_type src_type
,
790 unsigned src_alignment
)
792 LLVMBuilderRef builder
= gallivm
->builder
;
793 unsigned row_size
= src_count
/ block_height
;
796 /* Ensure src exactly fits into block */
797 assert((block_width
* block_height
) % src_count
== 0);
799 for (i
= 0; i
< src_count
; ++i
) {
800 unsigned x
= i
% row_size
;
801 unsigned y
= i
/ row_size
;
803 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
804 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
807 LLVMValueRef src_ptr
;
809 gep
[0] = lp_build_const_int32(gallivm
, 0);
810 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
812 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
813 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
815 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
817 lp_set_store_alignment(src_ptr
, src_alignment
);
823 * Checks if a format description is an arithmetic format
825 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
827 static INLINE boolean
828 is_arithmetic_format(const struct util_format_description
*format_desc
)
830 boolean arith
= false;
833 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
834 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
835 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
843 * Checks if this format requires special handling due to required expansion
844 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
847 static INLINE boolean
848 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
850 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
851 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
859 * Retrieves the type representing the memory layout for a format
861 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
864 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
865 struct lp_type
* type
)
870 if (format_expands_to_float_soa(format_desc
)) {
871 /* just make this a uint with width of block */
872 type
->floating
= false;
876 type
->width
= format_desc
->block
.bits
;
881 for (i
= 0; i
< 4; i
++)
882 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
886 memset(type
, 0, sizeof(struct lp_type
));
887 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
888 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
889 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
890 type
->norm
= format_desc
->channel
[chan
].normalized
;
892 if (is_arithmetic_format(format_desc
)) {
896 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
897 type
->width
+= format_desc
->channel
[i
].size
;
900 type
->width
= format_desc
->channel
[chan
].size
;
901 type
->length
= format_desc
->nr_channels
;
907 * Retrieves the type for a format which is usable in the blending code.
909 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
912 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
913 struct lp_type
* type
)
918 if (format_expands_to_float_soa(format_desc
)) {
919 /* always use ordinary floats for blending */
920 type
->floating
= true;
929 for (i
= 0; i
< 4; i
++)
930 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
934 memset(type
, 0, sizeof(struct lp_type
));
935 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
936 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
937 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
938 type
->norm
= format_desc
->channel
[chan
].normalized
;
939 type
->width
= format_desc
->channel
[chan
].size
;
940 type
->length
= format_desc
->nr_channels
;
942 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
943 if (format_desc
->channel
[i
].size
> type
->width
)
944 type
->width
= format_desc
->channel
[i
].size
;
947 if (type
->floating
) {
950 if (type
->width
<= 8) {
952 } else if (type
->width
<= 16) {
959 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
966 * Scale a normalized value from src_bits to dst_bits.
968 * The exact calculation is
970 * dst = iround(src * dst_mask / src_mask)
972 * or with integer rounding
974 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
978 * src_mask = (1 << src_bits) - 1
979 * dst_mask = (1 << dst_bits) - 1
981 * but we try to avoid division and multiplication through shifts.
983 static INLINE LLVMValueRef
984 scale_bits(struct gallivm_state
*gallivm
,
988 struct lp_type src_type
)
990 LLVMBuilderRef builder
= gallivm
->builder
;
991 LLVMValueRef result
= src
;
993 if (dst_bits
< src_bits
) {
994 int delta_bits
= src_bits
- dst_bits
;
996 if (delta_bits
<= dst_bits
) {
998 * Approximate the rescaling with a single shift.
1000 * This gives the wrong rounding.
1003 result
= LLVMBuildLShr(builder
,
1005 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1010 * Try more accurate rescaling.
1014 * Drop the least significant bits to make space for the multiplication.
1016 * XXX: A better approach would be to use a wider integer type as intermediate. But
1017 * this is enough to convert alpha from 16bits -> 2 when rendering to
1018 * PIPE_FORMAT_R10G10B10A2_UNORM.
1020 result
= LLVMBuildLShr(builder
,
1022 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1026 result
= LLVMBuildMul(builder
,
1028 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1032 * Add a rounding term before the division.
1034 * TODO: Handle signed integers too.
1036 if (!src_type
.sign
) {
1037 result
= LLVMBuildAdd(builder
,
1039 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1044 * Approximate the division by src_mask with a src_bits shift.
1046 * Given the src has already been shifted by dst_bits, all we need
1047 * to do is to shift by the difference.
1050 result
= LLVMBuildLShr(builder
,
1052 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1056 } else if (dst_bits
> src_bits
) {
1058 int db
= dst_bits
- src_bits
;
1060 /* Shift left by difference in bits */
1061 result
= LLVMBuildShl(builder
,
1063 lp_build_const_int_vec(gallivm
, src_type
, db
),
1066 if (db
< src_bits
) {
1067 /* Enough bits in src to fill the remainder */
1068 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1070 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1073 result
= LLVMBuildOr(builder
, result
, lower
, "");
1074 } else if (db
> src_bits
) {
1075 /* Need to repeatedly copy src bits to fill remainder in dst */
1078 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1079 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1081 result
= LLVMBuildOr(builder
,
1083 LLVMBuildLShr(builder
, result
, shuv
, ""),
1093 * If RT is a smallfloat (needing denorms) format
1096 have_smallfloat_format(struct lp_type dst_type
,
1097 enum pipe_format format
)
1099 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1100 /* due to format handling hacks this format doesn't have floating set
1101 * here (and actually has width set to 32 too) so special case this. */
1102 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1107 * Convert from memory format to blending format
1109 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1112 convert_to_blend_type(struct gallivm_state
*gallivm
,
1113 unsigned block_size
,
1114 const struct util_format_description
*src_fmt
,
1115 struct lp_type src_type
,
1116 struct lp_type dst_type
,
1117 LLVMValueRef
* src
, // and dst
1120 LLVMValueRef
*dst
= src
;
1121 LLVMBuilderRef builder
= gallivm
->builder
;
1122 struct lp_type blend_type
;
1123 struct lp_type mem_type
;
1125 unsigned pixels
= block_size
/ num_srcs
;
1129 * full custom path for packed floats and srgb formats - none of the later
1130 * functions would do anything useful, and given the lp_type representation they
1131 * can't be fixed. Should really have some SoA blend path for these kind of
1132 * formats rather than hacking them in here.
1134 if (format_expands_to_float_soa(src_fmt
)) {
1135 LLVMValueRef tmpsrc
[4];
1137 * This is pretty suboptimal for this case blending in SoA would be much
1138 * better, since conversion gets us SoA values so need to convert back.
1140 assert(src_type
.width
== 32 || src_type
.width
== 16);
1141 assert(dst_type
.floating
);
1142 assert(dst_type
.width
== 32);
1143 assert(dst_type
.length
% 4 == 0);
1144 assert(num_srcs
% 4 == 0);
1146 if (src_type
.width
== 16) {
1147 /* expand 4x16bit values to 4x32bit */
1148 struct lp_type type32x4
= src_type
;
1149 LLVMTypeRef ltype32x4
;
1150 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1151 type32x4
.width
= 32;
1152 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1153 for (i
= 0; i
< num_fetch
; i
++) {
1154 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1156 src_type
.width
= 32;
1158 for (i
= 0; i
< 4; i
++) {
1161 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1162 LLVMValueRef tmpsoa
[4];
1163 LLVMValueRef tmps
= tmpsrc
[i
];
1164 if (dst_type
.length
== 8) {
1165 LLVMValueRef shuffles
[8];
1167 /* fetch was 4 values but need 8-wide output values */
1168 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1170 * for 8-wide aos transpose would give us wrong order not matching
1171 * incoming converted fs values and mask. ARGH.
1173 for (j
= 0; j
< 4; j
++) {
1174 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1175 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1177 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1178 LLVMConstVector(shuffles
, 8), "");
1180 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1181 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1184 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1186 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1191 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1192 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1194 /* Is the format arithmetic */
1195 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1196 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1198 /* Pad if necessary */
1199 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1200 for (i
= 0; i
< num_srcs
; ++i
) {
1201 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1204 src_type
.length
= dst_type
.length
;
1207 /* Special case for half-floats */
1208 if (mem_type
.width
== 16 && mem_type
.floating
) {
1209 assert(blend_type
.width
== 32 && blend_type
.floating
);
1210 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1218 src_type
.width
= blend_type
.width
* blend_type
.length
;
1219 blend_type
.length
*= pixels
;
1220 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1222 for (i
= 0; i
< num_srcs
; ++i
) {
1223 LLVMValueRef chans
[4];
1224 LLVMValueRef res
= NULL
;
1226 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1228 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1230 unsigned sa
= src_fmt
->channel
[j
].shift
;
1231 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1232 unsigned from_lsb
= j
;
1234 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1237 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1241 /* Extract bits from source */
1242 chans
[j
] = LLVMBuildLShr(builder
,
1244 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1247 chans
[j
] = LLVMBuildAnd(builder
,
1249 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1253 if (src_type
.norm
) {
1254 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1255 blend_type
.width
, chans
[j
], src_type
);
1258 /* Insert bits into correct position */
1259 chans
[j
] = LLVMBuildShl(builder
,
1261 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1267 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1271 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1277 * Convert from blending format to memory format
1279 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1282 convert_from_blend_type(struct gallivm_state
*gallivm
,
1283 unsigned block_size
,
1284 const struct util_format_description
*src_fmt
,
1285 struct lp_type src_type
,
1286 struct lp_type dst_type
,
1287 LLVMValueRef
* src
, // and dst
1290 LLVMValueRef
* dst
= src
;
1292 struct lp_type mem_type
;
1293 struct lp_type blend_type
;
1294 LLVMBuilderRef builder
= gallivm
->builder
;
1295 unsigned pixels
= block_size
/ num_srcs
;
1299 * full custom path for packed floats and srgb formats - none of the later
1300 * functions would do anything useful, and given the lp_type representation they
1301 * can't be fixed. Should really have some SoA blend path for these kind of
1302 * formats rather than hacking them in here.
1304 if (format_expands_to_float_soa(src_fmt
)) {
1306 * This is pretty suboptimal for this case blending in SoA would be much
1307 * better - we need to transpose the AoS values back to SoA values for
1308 * conversion/packing.
1310 assert(src_type
.floating
);
1311 assert(src_type
.width
== 32);
1312 assert(src_type
.length
% 4 == 0);
1313 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1315 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1316 LLVMValueRef tmpsoa
[4], tmpdst
;
1317 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1318 /* really really need SoA here */
1320 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1321 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1324 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1328 if (src_type
.length
== 8) {
1329 LLVMValueRef tmpaos
, shuffles
[8];
1332 * for 8-wide aos transpose has given us wrong order not matching
1333 * output order. HMPF. Also need to split the output values manually.
1335 for (j
= 0; j
< 4; j
++) {
1336 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1337 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1339 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1340 LLVMConstVector(shuffles
, 8), "");
1341 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1342 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1348 if (dst_type
.width
== 16) {
1349 struct lp_type type16x8
= dst_type
;
1350 struct lp_type type32x4
= dst_type
;
1351 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1352 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1353 type16x8
.length
= 8;
1354 type32x4
.width
= 32;
1355 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1356 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1357 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1358 /* We could do vector truncation but it doesn't generate very good code */
1359 for (i
= 0; i
< num_fetch
; i
++) {
1360 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1361 src
[i
], lp_build_zero(gallivm
, type32x4
));
1362 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1363 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1364 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1370 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1371 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1373 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1375 /* Special case for half-floats */
1376 if (mem_type
.width
== 16 && mem_type
.floating
) {
1377 int length
= dst_type
.length
;
1378 assert(blend_type
.width
== 32 && blend_type
.floating
);
1380 dst_type
.length
= src_type
.length
;
1382 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1384 dst_type
.length
= length
;
1388 /* Remove any padding */
1389 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1390 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1392 for (i
= 0; i
< num_srcs
; ++i
) {
1393 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1397 /* No bit arithmetic to do */
1402 src_type
.length
= pixels
;
1403 src_type
.width
= blend_type
.length
* blend_type
.width
;
1404 dst_type
.length
= pixels
;
1406 for (i
= 0; i
< num_srcs
; ++i
) {
1407 LLVMValueRef chans
[4];
1408 LLVMValueRef res
= NULL
;
1410 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1412 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1414 unsigned sa
= src_fmt
->channel
[j
].shift
;
1415 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1416 unsigned from_lsb
= j
;
1418 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1421 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1423 for (k
= 0; k
< blend_type
.width
; ++k
) {
1428 chans
[j
] = LLVMBuildLShr(builder
,
1430 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1433 chans
[j
] = LLVMBuildAnd(builder
,
1435 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1438 /* Scale down bits */
1439 if (src_type
.norm
) {
1440 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1441 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1445 chans
[j
] = LLVMBuildShl(builder
,
1447 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1450 sa
+= src_fmt
->channel
[j
].size
;
1455 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1459 assert (dst_type
.width
!= 24);
1461 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1467 * Convert alpha to same blend type as src
1470 convert_alpha(struct gallivm_state
*gallivm
,
1471 struct lp_type row_type
,
1472 struct lp_type alpha_type
,
1473 const unsigned block_size
,
1474 const unsigned block_height
,
1475 const unsigned src_count
,
1476 const unsigned dst_channels
,
1477 const bool pad_inline
,
1478 LLVMValueRef
* src_alpha
)
1480 LLVMBuilderRef builder
= gallivm
->builder
;
1482 unsigned length
= row_type
.length
;
1483 row_type
.length
= alpha_type
.length
;
1485 /* Twiddle the alpha to match pixels */
1486 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1489 * TODO this should use single lp_build_conv call for
1490 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1492 for (i
= 0; i
< block_height
; ++i
) {
1493 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1496 alpha_type
= row_type
;
1497 row_type
.length
= length
;
1499 /* If only one channel we can only need the single alpha value per pixel */
1500 if (src_count
== 1 && dst_channels
== 1) {
1502 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1504 /* If there are more srcs than rows then we need to split alpha up */
1505 if (src_count
> block_height
) {
1506 for (i
= src_count
; i
> 0; --i
) {
1507 unsigned pixels
= block_size
/ src_count
;
1508 unsigned idx
= i
- 1;
1510 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1511 (idx
* pixels
) % 4, pixels
);
1515 /* If there is a src for each pixel broadcast the alpha across whole row */
1516 if (src_count
== block_size
) {
1517 for (i
= 0; i
< src_count
; ++i
) {
1518 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1521 unsigned pixels
= block_size
/ src_count
;
1522 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1523 unsigned alpha_span
= 1;
1524 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1526 /* Check if we need 2 src_alphas for our shuffles */
1527 if (pixels
> alpha_type
.length
) {
1531 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1532 for (j
= 0; j
< row_type
.length
; ++j
) {
1533 if (j
< pixels
* channels
) {
1534 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1536 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1540 for (i
= 0; i
< src_count
; ++i
) {
1541 unsigned idx1
= i
, idx2
= i
;
1543 if (alpha_span
> 1){
1548 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1551 LLVMConstVector(shuffles
, row_type
.length
),
1560 * Generates the blend function for unswizzled colour buffers
1561 * Also generates the read & write from colour buffer
1564 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1566 struct lp_fragment_shader_variant
*variant
,
1567 enum pipe_format out_format
,
1568 unsigned int num_fs
,
1569 struct lp_type fs_type
,
1570 LLVMValueRef
* fs_mask
,
1571 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1572 LLVMValueRef context_ptr
,
1573 LLVMValueRef color_ptr
,
1574 LLVMValueRef stride
,
1575 unsigned partial_mask
,
1578 const unsigned alpha_channel
= 3;
1579 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1580 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1581 const unsigned block_size
= block_width
* block_height
;
1582 const unsigned lp_integer_vector_width
= 128;
1584 LLVMBuilderRef builder
= gallivm
->builder
;
1585 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1586 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1587 LLVMValueRef src_alpha
[4 * 4];
1588 LLVMValueRef src1_alpha
[4 * 4];
1589 LLVMValueRef src_mask
[4 * 4];
1590 LLVMValueRef src
[4 * 4];
1591 LLVMValueRef src1
[4 * 4];
1592 LLVMValueRef dst
[4 * 4];
1593 LLVMValueRef blend_color
;
1594 LLVMValueRef blend_alpha
;
1595 LLVMValueRef i32_zero
;
1596 LLVMValueRef check_mask
;
1597 LLVMValueRef undef_src_val
;
1599 struct lp_build_mask_context mask_ctx
;
1600 struct lp_type mask_type
;
1601 struct lp_type blend_type
;
1602 struct lp_type row_type
;
1603 struct lp_type dst_type
;
1605 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1606 unsigned vector_width
;
1607 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1608 unsigned dst_channels
;
1613 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1615 unsigned dst_alignment
;
1617 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1618 bool has_alpha
= false;
1619 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1620 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1622 const boolean is_1d
= variant
->key
.resource_1d
;
1623 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1624 LLVMValueRef fpstate
= 0;
1626 /* Get type from output format */
1627 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1628 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1631 * Technically this code should go into lp_build_smallfloat_to_float
1632 * and lp_build_float_to_smallfloat but due to the
1633 * http://llvm.org/bugs/show_bug.cgi?id=6393
1634 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1635 * So the ordering is important here and there shouldn't be any
1636 * llvm ir instrunctions in this function before
1637 * this, otherwise half-float format conversions won't work
1638 * (again due to llvm bug #6393).
1640 if (have_smallfloat_format(dst_type
, out_format
)) {
1641 /* We need to make sure that denorms are ok for half float
1643 fpstate
= lp_build_fpstate_get(gallivm
);
1644 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1647 mask_type
= lp_int32_vec4_type();
1648 mask_type
.length
= fs_type
.length
;
1650 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1651 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1654 /* Do not bother executing code when mask is empty.. */
1656 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1658 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1659 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1662 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1663 lp_build_mask_check(&mask_ctx
);
1666 partial_mask
|= !variant
->opaque
;
1667 i32_zero
= lp_build_const_int32(gallivm
, 0);
1669 #if HAVE_LLVM < 0x0302
1671 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1672 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1674 undef_src_val
= lp_build_zero(gallivm
, fs_type
);
1676 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1679 row_type
.length
= fs_type
.length
;
1680 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1682 /* Compute correct swizzle and count channels */
1683 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1686 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1687 /* Ensure channel is used */
1688 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1692 /* Ensure not already written to (happens in case with GL_ALPHA) */
1693 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1697 /* Ensure we havn't already found all channels */
1698 if (dst_channels
>= out_format_desc
->nr_channels
) {
1702 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1705 if (i
== alpha_channel
) {
1710 if (format_expands_to_float_soa(out_format_desc
)) {
1712 * the code above can't work for layout_other
1713 * for srgb it would sort of work but we short-circuit swizzles, etc.
1714 * as that is done as part of unpack / pack.
1716 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1722 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1725 /* If 3 channels then pad to include alpha for 4 element transpose */
1726 if (dst_channels
== 3 && !has_alpha
) {
1727 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1728 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1731 if (out_format_desc
->nr_channels
== 4) {
1737 * Load shader output
1739 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1740 /* Always load alpha for use in blending */
1743 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1746 alpha
= undef_src_val
;
1749 /* Load each channel */
1750 for (j
= 0; j
< dst_channels
; ++j
) {
1751 assert(swizzle
[j
] < 4);
1753 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1756 fs_src
[i
][j
] = undef_src_val
;
1760 /* If 3 channels then pad to include alpha for 4 element transpose */
1762 * XXX If we include that here maybe could actually use it instead of
1763 * separate alpha for blending?
1765 if (dst_channels
== 3 && !has_alpha
) {
1766 fs_src
[i
][3] = alpha
;
1769 /* We split the row_mask and row_alpha as we want 128bit interleave */
1770 if (fs_type
.length
== 8) {
1771 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1772 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1774 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1775 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1777 src_mask
[i
] = fs_mask
[i
];
1778 src_alpha
[i
] = alpha
;
1781 if (dual_source_blend
) {
1782 /* same as above except different src/dst, skip masks and comments... */
1783 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1786 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1789 alpha
= undef_src_val
;
1792 for (j
= 0; j
< dst_channels
; ++j
) {
1793 assert(swizzle
[j
] < 4);
1795 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1798 fs_src1
[i
][j
] = undef_src_val
;
1801 if (dst_channels
== 3 && !has_alpha
) {
1802 fs_src1
[i
][3] = alpha
;
1804 if (fs_type
.length
== 8) {
1805 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1806 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1808 src1_alpha
[i
] = alpha
;
1813 if (util_format_is_pure_integer(out_format
)) {
1815 * In this case fs_type was really ints or uints disguised as floats,
1818 fs_type
.floating
= 0;
1819 fs_type
.sign
= dst_type
.sign
;
1820 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1821 for (j
= 0; j
< dst_channels
; ++j
) {
1822 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1823 lp_build_vec_type(gallivm
, fs_type
), "");
1825 if (dst_channels
== 3 && !has_alpha
) {
1826 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1827 lp_build_vec_type(gallivm
, fs_type
), "");
1833 * Pixel twiddle from fragment shader order to memory order
1835 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1836 dst_channels
, fs_src
, src
, pad_inline
);
1837 if (dual_source_blend
) {
1838 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1839 fs_src1
, src1
, pad_inline
);
1842 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1843 if (src_count
!= num_fullblock_fs
* src_channels
) {
1844 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1845 row_type
.length
/= ds
;
1846 fs_type
.length
= row_type
.length
;
1849 blend_type
= row_type
;
1850 mask_type
.length
= 4;
1852 /* Convert src to row_type */
1853 if (dual_source_blend
) {
1854 struct lp_type old_row_type
= row_type
;
1855 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1856 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1859 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1862 /* If the rows are not an SSE vector, combine them to become SSE size! */
1863 if ((row_type
.width
* row_type
.length
) % 128) {
1864 unsigned bits
= row_type
.width
* row_type
.length
;
1867 assert(src_count
>= (vector_width
/ bits
));
1869 dst_count
= src_count
/ (vector_width
/ bits
);
1871 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1872 if (dual_source_blend
) {
1873 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1876 row_type
.length
*= combined
;
1877 src_count
/= combined
;
1879 bits
= row_type
.width
* row_type
.length
;
1880 assert(bits
== 128 || bits
== 256);
1885 * Blend Colour conversion
1887 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1888 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1889 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1892 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1894 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1896 * since blending is done with floats, there was no conversion.
1897 * However, the rules according to fixed point renderbuffers still
1898 * apply, that is we must clamp inputs to 0.0/1.0.
1899 * (This would apply to separate alpha conversion too but we currently
1900 * force has_alpha to be true.)
1901 * TODO: should skip this with "fake" blend, since post-blend conversion
1902 * will clamp anyway.
1903 * TODO: could also skip this if fragment color clamping is enabled. We
1904 * don't support it natively so it gets baked into the shader however, so
1905 * can't really tell here.
1907 struct lp_build_context f32_bld
;
1908 assert(row_type
.floating
);
1909 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1910 for (i
= 0; i
< src_count
; i
++) {
1911 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
1913 if (dual_source_blend
) {
1914 for (i
= 0; i
< src_count
; i
++) {
1915 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
1918 /* probably can't be different than row_type but better safe than sorry... */
1919 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1920 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1924 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1926 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1927 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1929 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1930 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1932 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1933 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1939 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1941 if (src_count
< block_height
) {
1942 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1943 } else if (src_count
> block_height
) {
1944 for (i
= src_count
; i
> 0; --i
) {
1945 unsigned pixels
= block_size
/ src_count
;
1946 unsigned idx
= i
- 1;
1948 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1949 (idx
* pixels
) % 4, pixels
);
1953 assert(mask_type
.width
== 32);
1955 for (i
= 0; i
< src_count
; ++i
) {
1956 unsigned pixels
= block_size
/ src_count
;
1957 unsigned pixel_width
= row_type
.width
* dst_channels
;
1959 if (pixel_width
== 24) {
1960 mask_type
.width
= 8;
1961 mask_type
.length
= vector_width
/ mask_type
.width
;
1963 mask_type
.length
= pixels
;
1964 mask_type
.width
= row_type
.width
* dst_channels
;
1966 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1968 mask_type
.length
*= dst_channels
;
1969 mask_type
.width
/= dst_channels
;
1972 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1973 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1980 struct lp_type alpha_type
= fs_type
;
1981 alpha_type
.length
= 4;
1982 convert_alpha(gallivm
, row_type
, alpha_type
,
1983 block_size
, block_height
,
1984 src_count
, dst_channels
,
1985 pad_inline
, src_alpha
);
1986 if (dual_source_blend
) {
1987 convert_alpha(gallivm
, row_type
, alpha_type
,
1988 block_size
, block_height
,
1989 src_count
, dst_channels
,
1990 pad_inline
, src1_alpha
);
1996 * Load dst from memory
1998 if (src_count
< block_height
) {
1999 dst_count
= block_height
;
2001 dst_count
= src_count
;
2004 dst_type
.length
*= block_size
/ dst_count
;
2006 if (format_expands_to_float_soa(out_format_desc
)) {
2008 * we need multiple values at once for the conversion, so can as well
2009 * load them vectorized here too instead of concatenating later.
2010 * (Still need concatenation later for 8-wide vectors).
2012 dst_count
= block_height
;
2013 dst_type
.length
= block_width
;
2017 * Compute the alignment of the destination pointer in bytes
2018 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2019 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2020 * 1d tex but can't distinguish here) so need to stick with per-pixel
2021 * alignment in this case.
2024 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2027 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2029 /* Force power-of-two alignment by extracting only the least-significant-bit */
2030 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2032 * Resource base and stride pointers are aligned to 16 bytes, so that's
2033 * the maximum alignment we can guarantee
2035 dst_alignment
= MIN2(16, dst_alignment
);
2038 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2039 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2040 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2041 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
2046 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2047 dst
, dst_type
, dst_count
, dst_alignment
);
2052 * Convert from dst/output format to src/blending format.
2054 * This is necessary as we can only read 1 row from memory at a time,
2055 * so the minimum dst_count will ever be at this point is 4.
2057 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2058 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2059 * on all 16 pixels in that single vector at once.
2061 if (dst_count
> src_count
) {
2062 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
2068 /* XXX this is broken for RGB8 formats -
2069 * they get expanded from 12 to 16 elements (to include alpha)
2070 * by convert_to_blend_type then reduced to 15 instead of 12
2071 * by convert_from_blend_type (a simple fix though breaks A8...).
2072 * R16G16B16 also crashes differently however something going wrong
2073 * inside llvm handling npot vector sizes seemingly.
2074 * It seems some cleanup could be done here (like skipping conversion/blend
2077 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
2080 * FIXME: Really should get logic ops / masks out of generic blend / row
2081 * format. Logic ops will definitely not work on the blend float format
2082 * used for SRGB here and I think OpenGL expects this to work as expected
2083 * (that is incoming values converted to srgb then logic op applied).
2085 for (i
= 0; i
< src_count
; ++i
) {
2086 dst
[i
] = lp_build_blend_aos(gallivm
,
2087 &variant
->key
.blend
,
2092 has_alpha
? NULL
: src_alpha
[i
],
2094 has_alpha
? NULL
: src1_alpha
[i
],
2096 partial_mask
? src_mask
[i
] : NULL
,
2098 has_alpha
? NULL
: blend_alpha
,
2100 pad_inline
? 4 : dst_channels
);
2103 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
2105 /* Split the blend rows back to memory rows */
2106 if (dst_count
> src_count
) {
2107 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2109 if (src_count
== 1) {
2110 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2111 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2113 row_type
.length
/= 2;
2117 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2118 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2119 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2120 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2122 row_type
.length
/= 2;
2127 * Store blend result to memory
2130 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2131 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2134 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2135 dst
, dst_type
, dst_count
, dst_alignment
);
2138 if (have_smallfloat_format(dst_type
, out_format
)) {
2139 lp_build_fpstate_set(gallivm
, fpstate
);
2143 lp_build_mask_end(&mask_ctx
);
2149 * Generate the runtime callable function for the whole fragment pipeline.
2150 * Note that the function which we generate operates on a block of 16
2151 * pixels at at time. The block contains 2x2 quads. Each quad contains
2155 generate_fragment(struct llvmpipe_context
*lp
,
2156 struct lp_fragment_shader
*shader
,
2157 struct lp_fragment_shader_variant
*variant
,
2158 unsigned partial_mask
)
2160 struct gallivm_state
*gallivm
= variant
->gallivm
;
2161 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2162 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2164 struct lp_type fs_type
;
2165 struct lp_type blend_type
;
2166 LLVMTypeRef fs_elem_type
;
2167 LLVMTypeRef blend_vec_type
;
2168 LLVMTypeRef arg_types
[13];
2169 LLVMTypeRef func_type
;
2170 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2171 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2172 LLVMValueRef context_ptr
;
2175 LLVMValueRef a0_ptr
;
2176 LLVMValueRef dadx_ptr
;
2177 LLVMValueRef dady_ptr
;
2178 LLVMValueRef color_ptr_ptr
;
2179 LLVMValueRef stride_ptr
;
2180 LLVMValueRef depth_ptr
;
2181 LLVMValueRef depth_stride
;
2182 LLVMValueRef mask_input
;
2183 LLVMValueRef thread_data_ptr
;
2184 LLVMBasicBlockRef block
;
2185 LLVMBuilderRef builder
;
2186 struct lp_build_sampler_soa
*sampler
;
2187 struct lp_build_interp_soa_context interp
;
2188 LLVMValueRef fs_mask
[16 / 4];
2189 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2190 LLVMValueRef function
;
2191 LLVMValueRef facing
;
2196 boolean cbuf0_write_all
;
2197 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2198 util_blend_state_is_dual(&key
->blend
, 0);
2200 assert(lp_native_vector_width
/ 32 >= 4);
2202 /* Adjust color input interpolation according to flatshade state:
2204 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2205 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2206 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2208 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2210 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2214 /* check if writes to cbuf[0] are to be copied to all cbufs */
2215 cbuf0_write_all
= FALSE
;
2216 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
2217 if (shader
->info
.base
.properties
[i
].name
==
2218 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
2219 cbuf0_write_all
= TRUE
;
2224 /* TODO: actually pick these based on the fs and color buffer
2225 * characteristics. */
2227 memset(&fs_type
, 0, sizeof fs_type
);
2228 fs_type
.floating
= TRUE
; /* floating point values */
2229 fs_type
.sign
= TRUE
; /* values are signed */
2230 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2231 fs_type
.width
= 32; /* 32-bit float */
2232 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2234 memset(&blend_type
, 0, sizeof blend_type
);
2235 blend_type
.floating
= FALSE
; /* values are integers */
2236 blend_type
.sign
= FALSE
; /* values are unsigned */
2237 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2238 blend_type
.width
= 8; /* 8-bit ubyte values */
2239 blend_type
.length
= 16; /* 16 elements per vector */
2242 * Generate the function prototype. Any change here must be reflected in
2243 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2246 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2248 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2250 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2251 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2253 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2254 arg_types
[1] = int32_type
; /* x */
2255 arg_types
[2] = int32_type
; /* y */
2256 arg_types
[3] = int32_type
; /* facing */
2257 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2258 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2259 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2260 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2261 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2262 arg_types
[9] = int32_type
; /* mask_input */
2263 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2264 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2265 arg_types
[12] = int32_type
; /* depth_stride */
2267 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2268 arg_types
, Elements(arg_types
), 0);
2270 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2271 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2273 variant
->function
[partial_mask
] = function
;
2275 /* XXX: need to propagate noalias down into color param now we are
2276 * passing a pointer-to-pointer?
2278 for(i
= 0; i
< Elements(arg_types
); ++i
)
2279 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2280 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2282 context_ptr
= LLVMGetParam(function
, 0);
2283 x
= LLVMGetParam(function
, 1);
2284 y
= LLVMGetParam(function
, 2);
2285 facing
= LLVMGetParam(function
, 3);
2286 a0_ptr
= LLVMGetParam(function
, 4);
2287 dadx_ptr
= LLVMGetParam(function
, 5);
2288 dady_ptr
= LLVMGetParam(function
, 6);
2289 color_ptr_ptr
= LLVMGetParam(function
, 7);
2290 depth_ptr
= LLVMGetParam(function
, 8);
2291 mask_input
= LLVMGetParam(function
, 9);
2292 thread_data_ptr
= LLVMGetParam(function
, 10);
2293 stride_ptr
= LLVMGetParam(function
, 11);
2294 depth_stride
= LLVMGetParam(function
, 12);
2296 lp_build_name(context_ptr
, "context");
2297 lp_build_name(x
, "x");
2298 lp_build_name(y
, "y");
2299 lp_build_name(a0_ptr
, "a0");
2300 lp_build_name(dadx_ptr
, "dadx");
2301 lp_build_name(dady_ptr
, "dady");
2302 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2303 lp_build_name(depth_ptr
, "depth");
2304 lp_build_name(thread_data_ptr
, "thread_data");
2305 lp_build_name(mask_input
, "mask_input");
2306 lp_build_name(stride_ptr
, "stride_ptr");
2307 lp_build_name(depth_stride
, "depth_stride");
2313 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2314 builder
= gallivm
->builder
;
2316 LLVMPositionBuilderAtEnd(builder
, block
);
2318 /* code generated texture sampling */
2319 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
2321 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2322 /* for 1d resources only run "upper half" of stamp */
2323 if (key
->resource_1d
)
2327 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2328 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2329 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2330 num_loop
, "mask_store");
2331 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2334 * The shader input interpolation info is not explicitely baked in the
2335 * shader key, but everything it derives from (TGSI, and flatshade) is
2336 * already included in the shader key.
2338 lp_build_interp_soa_init(&interp
,
2340 shader
->info
.base
.num_inputs
,
2342 shader
->info
.base
.pixel_center_integer
,
2344 a0_ptr
, dadx_ptr
, dady_ptr
,
2347 for (i
= 0; i
< num_fs
; i
++) {
2349 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2350 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2351 &indexi
, 1, "mask_ptr");
2354 mask
= generate_quad_mask(gallivm
, fs_type
,
2355 i
*fs_type
.length
/4, mask_input
);
2358 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2360 LLVMBuildStore(builder
, mask
, mask_ptr
);
2363 generate_fs_loop(gallivm
,
2371 mask_store
, /* output */
2378 for (i
= 0; i
< num_fs
; i
++) {
2379 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2380 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2382 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2383 /* This is fucked up need to reorganize things */
2384 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2385 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2386 ptr
= LLVMBuildGEP(builder
,
2387 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2389 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2392 if (dual_source_blend
) {
2393 /* only support one dual source blend target hence always use output 1 */
2394 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2395 ptr
= LLVMBuildGEP(builder
,
2396 color_store
[1][chan
],
2398 fs_out_color
[1][chan
][i
] = ptr
;
2404 sampler
->destroy(sampler
);
2406 /* Loop over color outputs / color buffers to do blending.
2408 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2409 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2410 LLVMValueRef color_ptr
;
2411 LLVMValueRef stride
;
2412 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2414 boolean do_branch
= ((key
->depth
.enabled
2415 || key
->stencil
[0].enabled
2416 || key
->alpha
.enabled
)
2417 && !shader
->info
.base
.uses_kill
);
2419 color_ptr
= LLVMBuildLoad(builder
,
2420 LLVMBuildGEP(builder
, color_ptr_ptr
,
2424 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2426 stride
= LLVMBuildLoad(builder
,
2427 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2430 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2431 key
->cbuf_format
[cbuf
],
2432 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2433 context_ptr
, color_ptr
, stride
,
2434 partial_mask
, do_branch
);
2438 LLVMBuildRetVoid(builder
);
2440 gallivm_verify_function(gallivm
, function
);
2445 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2449 debug_printf("fs variant %p:\n", (void *) key
);
2451 if (key
->flatshade
) {
2452 debug_printf("flatshade = 1\n");
2454 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2455 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2457 if (key
->depth
.enabled
) {
2458 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2459 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2460 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2463 for (i
= 0; i
< 2; ++i
) {
2464 if (key
->stencil
[i
].enabled
) {
2465 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2466 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2467 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2468 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2469 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2470 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2474 if (key
->alpha
.enabled
) {
2475 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2478 if (key
->occlusion_count
) {
2479 debug_printf("occlusion_count = 1\n");
2482 if (key
->blend
.logicop_enable
) {
2483 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2485 else if (key
->blend
.rt
[0].blend_enable
) {
2486 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2487 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2488 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2489 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2490 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2491 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2493 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2494 if (key
->blend
.alpha_to_coverage
) {
2495 debug_printf("blend.alpha_to_coverage is enabled\n");
2497 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2498 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2499 debug_printf("sampler[%u] = \n", i
);
2500 debug_printf(" .wrap = %s %s %s\n",
2501 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2502 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2503 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2504 debug_printf(" .min_img_filter = %s\n",
2505 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2506 debug_printf(" .min_mip_filter = %s\n",
2507 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2508 debug_printf(" .mag_img_filter = %s\n",
2509 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2510 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2511 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2512 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2513 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2514 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2515 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2516 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2518 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2519 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2520 debug_printf("texture[%u] = \n", i
);
2521 debug_printf(" .format = %s\n",
2522 util_format_name(texture
->format
));
2523 debug_printf(" .target = %s\n",
2524 util_dump_tex_target(texture
->target
, TRUE
));
2525 debug_printf(" .level_zero_only = %u\n",
2526 texture
->level_zero_only
);
2527 debug_printf(" .pot = %u %u %u\n",
2529 texture
->pot_height
,
2530 texture
->pot_depth
);
2536 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2538 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2539 variant
->shader
->no
, variant
->no
);
2540 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2541 dump_fs_variant_key(&variant
->key
);
2542 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2548 * Generate a new fragment shader variant from the shader code and
2549 * other state indicated by the key.
2551 static struct lp_fragment_shader_variant
*
2552 generate_variant(struct llvmpipe_context
*lp
,
2553 struct lp_fragment_shader
*shader
,
2554 const struct lp_fragment_shader_variant_key
*key
)
2556 struct lp_fragment_shader_variant
*variant
;
2557 const struct util_format_description
*cbuf0_format_desc
;
2558 boolean fullcolormask
;
2559 char module_name
[64];
2561 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2565 util_snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2566 shader
->no
, shader
->variants_created
);
2568 variant
->gallivm
= gallivm_create(module_name
);
2569 if (!variant
->gallivm
) {
2574 variant
->shader
= shader
;
2575 variant
->list_item_global
.base
= variant
;
2576 variant
->list_item_local
.base
= variant
;
2577 variant
->no
= shader
->variants_created
++;
2579 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2582 * Determine whether we are touching all channels in the color buffer.
2584 fullcolormask
= FALSE
;
2585 if (key
->nr_cbufs
== 1) {
2586 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2587 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2591 !key
->blend
.logicop_enable
&&
2592 !key
->blend
.rt
[0].blend_enable
&&
2594 !key
->stencil
[0].enabled
&&
2595 !key
->alpha
.enabled
&&
2596 !key
->blend
.alpha_to_coverage
&&
2597 !key
->depth
.enabled
&&
2598 !shader
->info
.base
.uses_kill
2601 if ((shader
->info
.base
.num_tokens
<= 1) &&
2602 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2603 variant
->ps_inv_multiplier
= 0;
2605 variant
->ps_inv_multiplier
= 1;
2608 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2609 lp_debug_fs_variant(variant
);
2612 lp_jit_init_types(variant
);
2614 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2615 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2617 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2618 if (variant
->opaque
) {
2619 /* Specialized shader, which doesn't need to read the color buffer. */
2620 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2625 * Compile everything
2628 gallivm_compile_module(variant
->gallivm
);
2630 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2632 if (variant
->function
[RAST_EDGE_TEST
]) {
2633 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2634 gallivm_jit_function(variant
->gallivm
,
2635 variant
->function
[RAST_EDGE_TEST
]);
2638 if (variant
->function
[RAST_WHOLE
]) {
2639 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2640 gallivm_jit_function(variant
->gallivm
,
2641 variant
->function
[RAST_WHOLE
]);
2642 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2643 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2646 gallivm_free_ir(variant
->gallivm
);
2653 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2654 const struct pipe_shader_state
*templ
)
2656 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2657 struct lp_fragment_shader
*shader
;
2659 int nr_sampler_views
;
2662 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2666 shader
->no
= fs_no
++;
2667 make_empty_list(&shader
->variants
);
2669 /* get/save the summary info for this shader */
2670 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2672 /* we need to keep a local copy of the tokens */
2673 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2675 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2676 if (shader
->draw_data
== NULL
) {
2677 FREE((void *) shader
->base
.tokens
);
2682 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2683 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2685 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2686 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2688 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2689 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2690 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2692 switch (shader
->info
.base
.input_interpolate
[i
]) {
2693 case TGSI_INTERPOLATE_CONSTANT
:
2694 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2696 case TGSI_INTERPOLATE_LINEAR
:
2697 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2699 case TGSI_INTERPOLATE_PERSPECTIVE
:
2700 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2702 case TGSI_INTERPOLATE_COLOR
:
2703 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2710 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2711 case TGSI_SEMANTIC_FACE
:
2712 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2714 case TGSI_SEMANTIC_POSITION
:
2715 /* Position was already emitted above
2717 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2718 shader
->inputs
[i
].src_index
= 0;
2722 shader
->inputs
[i
].src_index
= i
+1;
2725 if (LP_DEBUG
& DEBUG_TGSI
) {
2727 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2728 shader
->no
, (void *) shader
);
2729 tgsi_dump(templ
->tokens
, 0);
2730 debug_printf("usage masks:\n");
2731 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2732 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2733 debug_printf(" IN[%u].%s%s%s%s\n",
2735 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2736 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2737 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2738 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2748 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2750 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2752 if (llvmpipe
->fs
== fs
)
2755 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2757 draw_bind_fragment_shader(llvmpipe
->draw
,
2758 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2760 llvmpipe
->dirty
|= LP_NEW_FS
;
2765 * Remove shader variant from two lists: the shader's variant list
2766 * and the context's variant list.
2769 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2770 struct lp_fragment_shader_variant
*variant
)
2772 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2773 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2774 " #%u v total cached #%u\n",
2775 variant
->shader
->no
,
2777 variant
->shader
->variants_created
,
2778 variant
->shader
->variants_cached
,
2779 lp
->nr_fs_variants
);
2782 gallivm_destroy(variant
->gallivm
);
2784 /* remove from shader's list */
2785 remove_from_list(&variant
->list_item_local
);
2786 variant
->shader
->variants_cached
--;
2788 /* remove from context's list */
2789 remove_from_list(&variant
->list_item_global
);
2790 lp
->nr_fs_variants
--;
2791 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2798 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2800 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2801 struct lp_fragment_shader
*shader
= fs
;
2802 struct lp_fs_variant_list_item
*li
;
2804 assert(fs
!= llvmpipe
->fs
);
2807 * XXX: we need to flush the context until we have some sort of reference
2808 * counting in fragment shaders as they may still be binned
2809 * Flushing alone might not sufficient we need to wait on it too.
2811 llvmpipe_finish(pipe
, __FUNCTION__
);
2813 /* Delete all the variants */
2814 li
= first_elem(&shader
->variants
);
2815 while(!at_end(&shader
->variants
, li
)) {
2816 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2817 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2821 /* Delete draw module's data */
2822 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2824 assert(shader
->variants_cached
== 0);
2825 FREE((void *) shader
->base
.tokens
);
2832 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2833 uint shader
, uint index
,
2834 struct pipe_constant_buffer
*cb
)
2836 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2837 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2839 assert(shader
< PIPE_SHADER_TYPES
);
2840 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2842 /* note: reference counting */
2843 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2845 if (shader
== PIPE_SHADER_VERTEX
||
2846 shader
== PIPE_SHADER_GEOMETRY
) {
2847 /* Pass the constants to the 'draw' module */
2848 const unsigned size
= cb
? cb
->buffer_size
: 0;
2852 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2854 else if (cb
&& cb
->user_buffer
) {
2855 data
= (ubyte
*) cb
->user_buffer
;
2862 data
+= cb
->buffer_offset
;
2864 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2868 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2870 if (cb
&& cb
->user_buffer
) {
2871 pipe_resource_reference(&constants
, NULL
);
2877 * Return the blend factor equivalent to a destination alpha of one.
2879 static INLINE
unsigned
2880 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2883 case PIPE_BLENDFACTOR_DST_ALPHA
:
2884 return PIPE_BLENDFACTOR_ONE
;
2885 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2886 return PIPE_BLENDFACTOR_ZERO
;
2887 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2889 return PIPE_BLENDFACTOR_ZERO
;
2891 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2899 * We need to generate several variants of the fragment pipeline to match
2900 * all the combinations of the contributing state atoms.
2902 * TODO: there is actually no reason to tie this to context state -- the
2903 * generated code could be cached globally in the screen.
2906 make_variant_key(struct llvmpipe_context
*lp
,
2907 struct lp_fragment_shader
*shader
,
2908 struct lp_fragment_shader_variant_key
*key
)
2912 memset(key
, 0, shader
->variant_key_size
);
2914 if (lp
->framebuffer
.zsbuf
) {
2915 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2916 const struct util_format_description
*zsbuf_desc
=
2917 util_format_description(zsbuf_format
);
2919 if (lp
->depth_stencil
->depth
.enabled
&&
2920 util_format_has_depth(zsbuf_desc
)) {
2921 key
->zsbuf_format
= zsbuf_format
;
2922 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2924 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2925 util_format_has_stencil(zsbuf_desc
)) {
2926 key
->zsbuf_format
= zsbuf_format
;
2927 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2929 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2930 key
->resource_1d
= TRUE
;
2935 * Propagate the depth clamp setting from the rasterizer state.
2936 * depth_clip == 0 implies depth clamping is enabled.
2938 * When clip_halfz is enabled, then always clamp the depth values.
2940 if (lp
->rasterizer
->clip_halfz
) {
2941 key
->depth_clamp
= 1;
2943 key
->depth_clamp
= (lp
->rasterizer
->depth_clip
== 0) ? 1 : 0;
2946 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2947 if (!lp
->framebuffer
.nr_cbufs
||
2948 !lp
->framebuffer
.cbufs
[0] ||
2949 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2950 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2952 if(key
->alpha
.enabled
)
2953 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2954 /* alpha.ref_value is passed in jit_context */
2956 key
->flatshade
= lp
->rasterizer
->flatshade
;
2957 if (lp
->active_occlusion_queries
) {
2958 key
->occlusion_count
= TRUE
;
2961 if (lp
->framebuffer
.nr_cbufs
) {
2962 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2965 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2967 if (!key
->blend
.independent_blend_enable
) {
2968 /* we always need independent blend otherwise the fixups below won't work */
2969 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2970 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2972 key
->blend
.independent_blend_enable
= 1;
2975 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2976 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2978 if (lp
->framebuffer
.cbufs
[i
]) {
2979 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2980 const struct util_format_description
*format_desc
;
2982 key
->cbuf_format
[i
] = format
;
2985 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2986 * mixing of 2d textures with height 1 and 1d textures, so make sure
2987 * we pick 1d if any cbuf or zsbuf is 1d.
2989 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
2990 key
->resource_1d
= TRUE
;
2993 format_desc
= util_format_description(format
);
2994 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2995 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2998 * Mask out color channels not present in the color buffer.
3000 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3003 * Disable blend for integer formats.
3005 if (util_format_is_pure_integer(format
)) {
3006 blend_rt
->blend_enable
= 0;
3010 * Our swizzled render tiles always have an alpha channel, but the
3011 * linear render target format often does not, so force here the dst
3014 * This is not a mere optimization. Wrong results will be produced if
3015 * the dst alpha is used, the dst format does not have alpha, and the
3016 * previous rendering was not flushed from the swizzled to linear
3017 * buffer. For example, NonPowTwo DCT.
3019 * TODO: This should be generalized to all channels for better
3020 * performance, but only alpha causes correctness issues.
3022 * Also, force rgb/alpha func/factors match, to make AoS blending
3025 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
3026 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3027 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3028 boolean clamped_zero
= !util_format_is_float(format
) &&
3029 !util_format_is_snorm(format
);
3030 blend_rt
->rgb_src_factor
=
3031 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3032 blend_rt
->rgb_dst_factor
=
3033 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3034 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3035 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3036 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3040 /* no color buffer for this fragment output */
3041 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3042 blend_rt
->colormask
= 0x0;
3043 blend_rt
->blend_enable
= 0;
3047 /* This value will be the same for all the variants of a given shader:
3049 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3051 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3052 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3053 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
3054 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3059 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3060 * are dx10-style? Can't really have mixed opcodes, at least not
3061 * if we want to skip the holes here (without rescanning tgsi).
3063 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3064 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3065 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3066 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
3067 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3068 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3073 key
->nr_sampler_views
= key
->nr_samplers
;
3074 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3075 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3076 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3077 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3086 * Update fragment shader state. This is called just prior to drawing
3087 * something when some fragment-related state has changed.
3090 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3092 struct lp_fragment_shader
*shader
= lp
->fs
;
3093 struct lp_fragment_shader_variant_key key
;
3094 struct lp_fragment_shader_variant
*variant
= NULL
;
3095 struct lp_fs_variant_list_item
*li
;
3097 make_variant_key(lp
, shader
, &key
);
3099 /* Search the variants for one which matches the key */
3100 li
= first_elem(&shader
->variants
);
3101 while(!at_end(&shader
->variants
, li
)) {
3102 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3110 /* Move this variant to the head of the list to implement LRU
3111 * deletion of shader's when we have too many.
3113 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3116 /* variant not found, create it now */
3119 unsigned variants_to_cull
;
3122 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3125 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3128 /* First, check if we've exceeded the max number of shader variants.
3129 * If so, free 25% of them (the least recently used ones).
3131 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
3133 if (variants_to_cull
||
3134 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3135 struct pipe_context
*pipe
= &lp
->pipe
;
3138 * XXX: we need to flush the context until we have some sort of
3139 * reference counting in fragment shaders as they may still be binned
3140 * Flushing alone might not be sufficient we need to wait on it too.
3142 llvmpipe_finish(pipe
, __FUNCTION__
);
3145 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3146 * number of shader variants (potentially all of them) could be
3147 * pending for destruction on flush.
3150 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3151 struct lp_fs_variant_list_item
*item
;
3152 if (is_empty_list(&lp
->fs_variants_list
)) {
3155 item
= last_elem(&lp
->fs_variants_list
);
3158 llvmpipe_remove_shader_variant(lp
, item
->base
);
3163 * Generate the new variant.
3166 variant
= generate_variant(lp
, shader
, &key
);
3169 LP_COUNT_ADD(llvm_compile_time
, dt
);
3170 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3172 /* Put the new variant into the list */
3174 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3175 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3176 lp
->nr_fs_variants
++;
3177 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3178 shader
->variants_cached
++;
3182 /* Bind this variant */
3183 lp_setup_set_fs_variant(lp
->setup
, variant
);
3191 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3193 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3194 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3195 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3197 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3201 * Rasterization is disabled if there is no pixel shader and
3202 * both depth and stencil testing are disabled:
3203 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3206 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
3208 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
3211 !lp
->depth_stencil
->depth
.enabled
&&
3212 !lp
->depth_stencil
->stencil
[0].enabled
);