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/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
, 1ULL << (j
+ 0), 0);
178 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
179 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
180 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (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 struct lp_type int_type
= lp_int_type(type
);
264 LLVMTypeRef vec_type
, int_vec_type
;
265 LLVMValueRef mask_ptr
, mask_val
;
266 LLVMValueRef consts_ptr
, num_consts_ptr
;
268 LLVMValueRef z_value
, s_value
;
269 LLVMValueRef z_fb
, s_fb
;
270 LLVMValueRef stencil_refs
[2];
271 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
272 struct lp_build_for_loop_state loop_state
;
273 struct lp_build_mask_context mask
;
275 * TODO: figure out if simple_shader optimization is really worthwile to
276 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
277 * code since tests tend to take another codepath than real shaders.
279 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
280 shader
->info
.base
.num_inputs
< 3 &&
281 shader
->info
.base
.num_instructions
< 8) && 0;
282 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
283 util_blend_state_is_dual(&key
->blend
, 0);
289 struct lp_bld_tgsi_system_values system_values
;
291 memset(&system_values
, 0, sizeof(system_values
));
293 if (key
->depth
.enabled
||
294 key
->stencil
[0].enabled
) {
296 zs_format_desc
= util_format_description(key
->zsbuf_format
);
297 assert(zs_format_desc
);
299 if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
300 if (key
->alpha
.enabled
||
301 key
->blend
.alpha_to_coverage
||
302 shader
->info
.base
.uses_kill
) {
303 /* With alpha test and kill, can do the depth test early
304 * and hopefully eliminate some quads. But need to do a
305 * special deferred depth write once the final mask value
306 * is known. This only works though if there's either no
307 * stencil test or the stencil value isn't written.
309 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
310 (key
->stencil
[1].enabled
&&
311 key
->stencil
[1].writemask
)))
312 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
314 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
317 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
320 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
323 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
324 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
325 (key
->stencil
[1].enabled
&&
326 key
->stencil
[1].writemask
))))
327 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
333 vec_type
= lp_build_vec_type(gallivm
, type
);
334 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
336 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
337 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
338 /* convert scalar stencil refs into vectors */
339 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
340 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
342 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
343 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
345 lp_build_for_loop_begin(&loop_state
, gallivm
,
346 lp_build_const_int32(gallivm
, 0),
349 lp_build_const_int32(gallivm
, 1));
351 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
352 &loop_state
.counter
, 1, "mask_ptr");
353 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
355 memset(outputs
, 0, sizeof outputs
);
357 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
358 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
359 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
360 lp_build_vec_type(gallivm
,
365 if (dual_source_blend
) {
366 assert(key
->nr_cbufs
<= 1);
367 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
368 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
369 lp_build_vec_type(gallivm
,
376 /* 'mask' will control execution based on quad's pixel alive/killed state */
377 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
379 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
380 lp_build_mask_check(&mask
);
382 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
385 if (depth_mode
& EARLY_DEPTH_TEST
) {
386 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
387 zs_format_desc
, key
->resource_1d
,
388 depth_ptr
, depth_stride
,
389 &z_fb
, &s_fb
, loop_state
.counter
);
390 lp_build_depth_stencil_test(gallivm
,
402 if (depth_mode
& EARLY_DEPTH_WRITE
) {
403 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
404 zs_format_desc
, key
->resource_1d
,
405 NULL
, NULL
, NULL
, loop_state
.counter
,
406 depth_ptr
, depth_stride
,
410 * Note mask check if stencil is enabled must be after ds write not after
411 * stencil test otherwise new stencil values may not get written if all
412 * fragments got killed by depth/stencil test.
414 if (!simple_shader
&& key
->stencil
[0].enabled
)
415 lp_build_mask_check(&mask
);
418 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
420 /* Build the actual shader */
421 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
422 consts_ptr
, num_consts_ptr
, &system_values
,
424 outputs
, context_ptr
, thread_data_ptr
,
425 sampler
, &shader
->info
.base
, NULL
);
428 if (key
->alpha
.enabled
) {
429 int color0
= find_output_by_semantic(&shader
->info
.base
,
433 if (color0
!= -1 && outputs
[color0
][3]) {
434 const struct util_format_description
*cbuf_format_desc
;
435 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
436 LLVMValueRef alpha_ref_value
;
438 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
439 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
441 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
443 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
444 &mask
, alpha
, alpha_ref_value
,
445 (depth_mode
& LATE_DEPTH_TEST
) != 0);
449 /* Emulate Alpha to Coverage with Alpha test */
450 if (key
->blend
.alpha_to_coverage
) {
451 int color0
= find_output_by_semantic(&shader
->info
.base
,
455 if (color0
!= -1 && outputs
[color0
][3]) {
456 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
458 lp_build_alpha_to_coverage(gallivm
, type
,
460 (depth_mode
& LATE_DEPTH_TEST
) != 0);
465 if (depth_mode
& LATE_DEPTH_TEST
) {
466 int pos0
= find_output_by_semantic(&shader
->info
.base
,
467 TGSI_SEMANTIC_POSITION
,
469 int s_out
= find_output_by_semantic(&shader
->info
.base
,
470 TGSI_SEMANTIC_STENCIL
,
472 if (pos0
!= -1 && outputs
[pos0
][2]) {
473 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
476 * Clamp according to ARB_depth_clamp semantics.
478 if (key
->depth_clamp
) {
479 LLVMValueRef viewport
, min_depth
, max_depth
;
480 LLVMValueRef viewport_index
;
481 struct lp_build_context f32_bld
;
483 assert(type
.floating
);
484 lp_build_context_init(&f32_bld
, gallivm
, type
);
487 * Assumes clamping of the viewport index will occur in setup/gs. Value
488 * is passed through the rasterization stage via lp_rast_shader_inputs.
490 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
493 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
497 * Load the min and max depth from the lp_jit_context.viewports
498 * array of lp_jit_viewport structures.
500 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
502 /* viewports[viewport_index].min_depth */
503 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
504 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
),
506 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
508 /* viewports[viewport_index].max_depth */
509 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
510 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
),
512 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
515 * Clamp to the min and max depth values for the given viewport.
517 z
= lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
521 if (s_out
!= -1 && outputs
[s_out
][1]) {
522 /* there's only one value, and spec says to discard additional bits */
523 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
524 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
525 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
526 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
527 stencil_refs
[1] = stencil_refs
[0];
530 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
531 zs_format_desc
, key
->resource_1d
,
532 depth_ptr
, depth_stride
,
533 &z_fb
, &s_fb
, loop_state
.counter
);
535 lp_build_depth_stencil_test(gallivm
,
547 if (depth_mode
& LATE_DEPTH_WRITE
) {
548 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
549 zs_format_desc
, key
->resource_1d
,
550 NULL
, NULL
, NULL
, loop_state
.counter
,
551 depth_ptr
, depth_stride
,
555 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
556 (depth_mode
& LATE_DEPTH_WRITE
))
558 /* Need to apply a reduced mask to the depth write. Reload the
559 * depth value, update from zs_value with the new mask value and
562 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
563 zs_format_desc
, key
->resource_1d
,
564 &mask
, z_fb
, s_fb
, loop_state
.counter
,
565 depth_ptr
, depth_stride
,
571 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
573 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
574 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
575 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
577 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
578 if(outputs
[attrib
][chan
]) {
579 /* XXX: just initialize outputs to point at colors[] and
582 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
583 LLVMValueRef color_ptr
;
584 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
585 &loop_state
.counter
, 1, "");
586 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
587 LLVMBuildStore(builder
, out
, color_ptr
);
593 if (key
->occlusion_count
) {
594 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
595 lp_build_name(counter
, "counter");
596 lp_build_occlusion_count(gallivm
, type
,
597 lp_build_mask_value(&mask
), counter
);
600 mask_val
= lp_build_mask_end(&mask
);
601 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
602 lp_build_for_loop_end(&loop_state
);
607 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
609 * Fragment Shader outputs pixels in small 2x2 blocks
610 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
612 * However in memory pixels are stored in rows
613 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
615 * @param type fragment shader type (4x or 8x float)
616 * @param num_fs number of fs_src
617 * @param is_1d whether we're outputting to a 1d resource
618 * @param dst_channels number of output channels
619 * @param fs_src output from fragment shader
620 * @param dst pointer to store result
621 * @param pad_inline is channel padding inline or at end of row
622 * @return the number of dsts
625 generate_fs_twiddle(struct gallivm_state
*gallivm
,
628 unsigned dst_channels
,
629 LLVMValueRef fs_src
[][4],
633 LLVMValueRef src
[16];
639 unsigned pixels
= type
.length
/ 4;
640 unsigned reorder_group
;
641 unsigned src_channels
;
645 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
646 src_count
= num_fs
* src_channels
;
648 assert(pixels
== 2 || pixels
== 1);
649 assert(num_fs
* src_channels
<= Elements(src
));
652 * Transpose from SoA -> AoS
654 for (i
= 0; i
< num_fs
; ++i
) {
655 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
659 * Pick transformation options
666 if (dst_channels
== 1) {
672 } else if (dst_channels
== 2) {
676 } else if (dst_channels
> 2) {
683 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
689 * Split the src in half
692 for (i
= num_fs
; i
> 0; --i
) {
693 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
694 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
702 * Ensure pixels are in memory order
705 /* Twiddle pixels by reordering the array, e.g.:
707 * src_count = 8 -> 0 2 1 3 4 6 5 7
708 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
710 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
712 for (i
= 0; i
< src_count
; ++i
) {
713 unsigned group
= i
/ reorder_group
;
714 unsigned block
= (group
/ 4) * 4 * reorder_group
;
715 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
718 } else if (twiddle
) {
719 /* Twiddle pixels across elements of array */
720 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
723 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
727 * Moves any padding between pixels to the end
728 * e.g. RGBXRGBX -> RGBRGBXX
731 unsigned char swizzles
[16];
732 unsigned elems
= pixels
* dst_channels
;
734 for (i
= 0; i
< type
.length
; ++i
) {
736 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
738 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
741 for (i
= 0; i
< src_count
; ++i
) {
742 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
751 * Load an unswizzled block of pixels from memory
754 load_unswizzled_block(struct gallivm_state
*gallivm
,
755 LLVMValueRef base_ptr
,
757 unsigned block_width
,
758 unsigned block_height
,
760 struct lp_type dst_type
,
762 unsigned dst_alignment
)
764 LLVMBuilderRef builder
= gallivm
->builder
;
765 unsigned row_size
= dst_count
/ block_height
;
768 /* Ensure block exactly fits into dst */
769 assert((block_width
* block_height
) % dst_count
== 0);
771 for (i
= 0; i
< dst_count
; ++i
) {
772 unsigned x
= i
% row_size
;
773 unsigned y
= i
/ row_size
;
775 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
776 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
779 LLVMValueRef dst_ptr
;
781 gep
[0] = lp_build_const_int32(gallivm
, 0);
782 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
784 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
785 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
787 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
789 LLVMSetAlignment(dst
[i
], dst_alignment
);
795 * Store an unswizzled block of pixels to memory
798 store_unswizzled_block(struct gallivm_state
*gallivm
,
799 LLVMValueRef base_ptr
,
801 unsigned block_width
,
802 unsigned block_height
,
804 struct lp_type src_type
,
806 unsigned src_alignment
)
808 LLVMBuilderRef builder
= gallivm
->builder
;
809 unsigned row_size
= src_count
/ block_height
;
812 /* Ensure src exactly fits into block */
813 assert((block_width
* block_height
) % src_count
== 0);
815 for (i
= 0; i
< src_count
; ++i
) {
816 unsigned x
= i
% row_size
;
817 unsigned y
= i
/ row_size
;
819 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
820 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
823 LLVMValueRef src_ptr
;
825 gep
[0] = lp_build_const_int32(gallivm
, 0);
826 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
828 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
829 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
831 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
833 LLVMSetAlignment(src_ptr
, src_alignment
);
839 * Checks if a format description is an arithmetic format
841 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
843 static inline boolean
844 is_arithmetic_format(const struct util_format_description
*format_desc
)
846 boolean arith
= false;
849 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
850 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
851 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
859 * Checks if this format requires special handling due to required expansion
860 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
863 static inline boolean
864 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
866 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
867 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
875 * Retrieves the type representing the memory layout for a format
877 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
880 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
881 struct lp_type
* type
)
886 if (format_expands_to_float_soa(format_desc
)) {
887 /* just make this a uint with width of block */
888 type
->floating
= false;
892 type
->width
= format_desc
->block
.bits
;
897 for (i
= 0; i
< 4; i
++)
898 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
902 memset(type
, 0, sizeof(struct lp_type
));
903 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
904 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
905 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
906 type
->norm
= format_desc
->channel
[chan
].normalized
;
908 if (is_arithmetic_format(format_desc
)) {
912 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
913 type
->width
+= format_desc
->channel
[i
].size
;
916 type
->width
= format_desc
->channel
[chan
].size
;
917 type
->length
= format_desc
->nr_channels
;
923 * Retrieves the type for a format which is usable in the blending code.
925 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
928 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
929 struct lp_type
* type
)
934 if (format_expands_to_float_soa(format_desc
)) {
935 /* always use ordinary floats for blending */
936 type
->floating
= true;
945 for (i
= 0; i
< 4; i
++)
946 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
950 memset(type
, 0, sizeof(struct lp_type
));
951 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
952 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
953 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
954 type
->norm
= format_desc
->channel
[chan
].normalized
;
955 type
->width
= format_desc
->channel
[chan
].size
;
956 type
->length
= format_desc
->nr_channels
;
958 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
959 if (format_desc
->channel
[i
].size
> type
->width
)
960 type
->width
= format_desc
->channel
[i
].size
;
963 if (type
->floating
) {
966 if (type
->width
<= 8) {
968 } else if (type
->width
<= 16) {
975 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
982 * Scale a normalized value from src_bits to dst_bits.
984 * The exact calculation is
986 * dst = iround(src * dst_mask / src_mask)
988 * or with integer rounding
990 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
994 * src_mask = (1 << src_bits) - 1
995 * dst_mask = (1 << dst_bits) - 1
997 * but we try to avoid division and multiplication through shifts.
999 static inline LLVMValueRef
1000 scale_bits(struct gallivm_state
*gallivm
,
1004 struct lp_type src_type
)
1006 LLVMBuilderRef builder
= gallivm
->builder
;
1007 LLVMValueRef result
= src
;
1009 if (dst_bits
< src_bits
) {
1010 int delta_bits
= src_bits
- dst_bits
;
1012 if (delta_bits
<= dst_bits
) {
1014 * Approximate the rescaling with a single shift.
1016 * This gives the wrong rounding.
1019 result
= LLVMBuildLShr(builder
,
1021 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1026 * Try more accurate rescaling.
1030 * Drop the least significant bits to make space for the multiplication.
1032 * XXX: A better approach would be to use a wider integer type as intermediate. But
1033 * this is enough to convert alpha from 16bits -> 2 when rendering to
1034 * PIPE_FORMAT_R10G10B10A2_UNORM.
1036 result
= LLVMBuildLShr(builder
,
1038 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1042 result
= LLVMBuildMul(builder
,
1044 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1048 * Add a rounding term before the division.
1050 * TODO: Handle signed integers too.
1052 if (!src_type
.sign
) {
1053 result
= LLVMBuildAdd(builder
,
1055 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1060 * Approximate the division by src_mask with a src_bits shift.
1062 * Given the src has already been shifted by dst_bits, all we need
1063 * to do is to shift by the difference.
1066 result
= LLVMBuildLShr(builder
,
1068 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1072 } else if (dst_bits
> src_bits
) {
1074 int db
= dst_bits
- src_bits
;
1076 /* Shift left by difference in bits */
1077 result
= LLVMBuildShl(builder
,
1079 lp_build_const_int_vec(gallivm
, src_type
, db
),
1082 if (db
< src_bits
) {
1083 /* Enough bits in src to fill the remainder */
1084 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1086 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1089 result
= LLVMBuildOr(builder
, result
, lower
, "");
1090 } else if (db
> src_bits
) {
1091 /* Need to repeatedly copy src bits to fill remainder in dst */
1094 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1095 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1097 result
= LLVMBuildOr(builder
,
1099 LLVMBuildLShr(builder
, result
, shuv
, ""),
1109 * If RT is a smallfloat (needing denorms) format
1112 have_smallfloat_format(struct lp_type dst_type
,
1113 enum pipe_format format
)
1115 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1116 /* due to format handling hacks this format doesn't have floating set
1117 * here (and actually has width set to 32 too) so special case this. */
1118 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1123 * Convert from memory format to blending format
1125 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1128 convert_to_blend_type(struct gallivm_state
*gallivm
,
1129 unsigned block_size
,
1130 const struct util_format_description
*src_fmt
,
1131 struct lp_type src_type
,
1132 struct lp_type dst_type
,
1133 LLVMValueRef
* src
, // and dst
1136 LLVMValueRef
*dst
= src
;
1137 LLVMBuilderRef builder
= gallivm
->builder
;
1138 struct lp_type blend_type
;
1139 struct lp_type mem_type
;
1141 unsigned pixels
= block_size
/ num_srcs
;
1145 * full custom path for packed floats and srgb formats - none of the later
1146 * functions would do anything useful, and given the lp_type representation they
1147 * can't be fixed. Should really have some SoA blend path for these kind of
1148 * formats rather than hacking them in here.
1150 if (format_expands_to_float_soa(src_fmt
)) {
1151 LLVMValueRef tmpsrc
[4];
1153 * This is pretty suboptimal for this case blending in SoA would be much
1154 * better, since conversion gets us SoA values so need to convert back.
1156 assert(src_type
.width
== 32 || src_type
.width
== 16);
1157 assert(dst_type
.floating
);
1158 assert(dst_type
.width
== 32);
1159 assert(dst_type
.length
% 4 == 0);
1160 assert(num_srcs
% 4 == 0);
1162 if (src_type
.width
== 16) {
1163 /* expand 4x16bit values to 4x32bit */
1164 struct lp_type type32x4
= src_type
;
1165 LLVMTypeRef ltype32x4
;
1166 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1167 type32x4
.width
= 32;
1168 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1169 for (i
= 0; i
< num_fetch
; i
++) {
1170 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1172 src_type
.width
= 32;
1174 for (i
= 0; i
< 4; i
++) {
1177 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1178 LLVMValueRef tmpsoa
[4];
1179 LLVMValueRef tmps
= tmpsrc
[i
];
1180 if (dst_type
.length
== 8) {
1181 LLVMValueRef shuffles
[8];
1183 /* fetch was 4 values but need 8-wide output values */
1184 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1186 * for 8-wide aos transpose would give us wrong order not matching
1187 * incoming converted fs values and mask. ARGH.
1189 for (j
= 0; j
< 4; j
++) {
1190 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1191 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1193 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1194 LLVMConstVector(shuffles
, 8), "");
1196 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1197 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1200 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1202 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1207 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1208 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1210 /* Is the format arithmetic */
1211 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1212 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1214 /* Pad if necessary */
1215 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1216 for (i
= 0; i
< num_srcs
; ++i
) {
1217 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1220 src_type
.length
= dst_type
.length
;
1223 /* Special case for half-floats */
1224 if (mem_type
.width
== 16 && mem_type
.floating
) {
1225 assert(blend_type
.width
== 32 && blend_type
.floating
);
1226 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1234 src_type
.width
= blend_type
.width
* blend_type
.length
;
1235 blend_type
.length
*= pixels
;
1236 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1238 for (i
= 0; i
< num_srcs
; ++i
) {
1239 LLVMValueRef chans
[4];
1240 LLVMValueRef res
= NULL
;
1242 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1244 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1246 unsigned sa
= src_fmt
->channel
[j
].shift
;
1247 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1248 unsigned from_lsb
= j
;
1250 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1253 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1257 /* Extract bits from source */
1258 chans
[j
] = LLVMBuildLShr(builder
,
1260 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1263 chans
[j
] = LLVMBuildAnd(builder
,
1265 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1269 if (src_type
.norm
) {
1270 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1271 blend_type
.width
, chans
[j
], src_type
);
1274 /* Insert bits into correct position */
1275 chans
[j
] = LLVMBuildShl(builder
,
1277 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1283 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1287 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1293 * Convert from blending format to memory format
1295 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1298 convert_from_blend_type(struct gallivm_state
*gallivm
,
1299 unsigned block_size
,
1300 const struct util_format_description
*src_fmt
,
1301 struct lp_type src_type
,
1302 struct lp_type dst_type
,
1303 LLVMValueRef
* src
, // and dst
1306 LLVMValueRef
* dst
= src
;
1308 struct lp_type mem_type
;
1309 struct lp_type blend_type
;
1310 LLVMBuilderRef builder
= gallivm
->builder
;
1311 unsigned pixels
= block_size
/ num_srcs
;
1315 * full custom path for packed floats and srgb formats - none of the later
1316 * functions would do anything useful, and given the lp_type representation they
1317 * can't be fixed. Should really have some SoA blend path for these kind of
1318 * formats rather than hacking them in here.
1320 if (format_expands_to_float_soa(src_fmt
)) {
1322 * This is pretty suboptimal for this case blending in SoA would be much
1323 * better - we need to transpose the AoS values back to SoA values for
1324 * conversion/packing.
1326 assert(src_type
.floating
);
1327 assert(src_type
.width
== 32);
1328 assert(src_type
.length
% 4 == 0);
1329 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1331 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1332 LLVMValueRef tmpsoa
[4], tmpdst
;
1333 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1334 /* really really need SoA here */
1336 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1337 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1340 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1344 if (src_type
.length
== 8) {
1345 LLVMValueRef tmpaos
, shuffles
[8];
1348 * for 8-wide aos transpose has given us wrong order not matching
1349 * output order. HMPF. Also need to split the output values manually.
1351 for (j
= 0; j
< 4; j
++) {
1352 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1353 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1355 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1356 LLVMConstVector(shuffles
, 8), "");
1357 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1358 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1364 if (dst_type
.width
== 16) {
1365 struct lp_type type16x8
= dst_type
;
1366 struct lp_type type32x4
= dst_type
;
1367 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1368 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1369 type16x8
.length
= 8;
1370 type32x4
.width
= 32;
1371 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1372 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1373 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1374 /* We could do vector truncation but it doesn't generate very good code */
1375 for (i
= 0; i
< num_fetch
; i
++) {
1376 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1377 src
[i
], lp_build_zero(gallivm
, type32x4
));
1378 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1379 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1380 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1386 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1387 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1389 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1391 /* Special case for half-floats */
1392 if (mem_type
.width
== 16 && mem_type
.floating
) {
1393 int length
= dst_type
.length
;
1394 assert(blend_type
.width
== 32 && blend_type
.floating
);
1396 dst_type
.length
= src_type
.length
;
1398 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1400 dst_type
.length
= length
;
1404 /* Remove any padding */
1405 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1406 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1408 for (i
= 0; i
< num_srcs
; ++i
) {
1409 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1413 /* No bit arithmetic to do */
1418 src_type
.length
= pixels
;
1419 src_type
.width
= blend_type
.length
* blend_type
.width
;
1420 dst_type
.length
= pixels
;
1422 for (i
= 0; i
< num_srcs
; ++i
) {
1423 LLVMValueRef chans
[4];
1424 LLVMValueRef res
= NULL
;
1426 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1428 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1430 unsigned sa
= src_fmt
->channel
[j
].shift
;
1431 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1432 unsigned from_lsb
= j
;
1434 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1437 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1439 for (k
= 0; k
< blend_type
.width
; ++k
) {
1444 chans
[j
] = LLVMBuildLShr(builder
,
1446 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1449 chans
[j
] = LLVMBuildAnd(builder
,
1451 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1454 /* Scale down bits */
1455 if (src_type
.norm
) {
1456 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1457 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1461 chans
[j
] = LLVMBuildShl(builder
,
1463 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1466 sa
+= src_fmt
->channel
[j
].size
;
1471 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1475 assert (dst_type
.width
!= 24);
1477 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1483 * Convert alpha to same blend type as src
1486 convert_alpha(struct gallivm_state
*gallivm
,
1487 struct lp_type row_type
,
1488 struct lp_type alpha_type
,
1489 const unsigned block_size
,
1490 const unsigned block_height
,
1491 const unsigned src_count
,
1492 const unsigned dst_channels
,
1493 const bool pad_inline
,
1494 LLVMValueRef
* src_alpha
)
1496 LLVMBuilderRef builder
= gallivm
->builder
;
1498 unsigned length
= row_type
.length
;
1499 row_type
.length
= alpha_type
.length
;
1501 /* Twiddle the alpha to match pixels */
1502 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1505 * TODO this should use single lp_build_conv call for
1506 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1508 for (i
= 0; i
< block_height
; ++i
) {
1509 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1512 alpha_type
= row_type
;
1513 row_type
.length
= length
;
1515 /* If only one channel we can only need the single alpha value per pixel */
1516 if (src_count
== 1 && dst_channels
== 1) {
1518 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1520 /* If there are more srcs than rows then we need to split alpha up */
1521 if (src_count
> block_height
) {
1522 for (i
= src_count
; i
> 0; --i
) {
1523 unsigned pixels
= block_size
/ src_count
;
1524 unsigned idx
= i
- 1;
1526 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1527 (idx
* pixels
) % 4, pixels
);
1531 /* If there is a src for each pixel broadcast the alpha across whole row */
1532 if (src_count
== block_size
) {
1533 for (i
= 0; i
< src_count
; ++i
) {
1534 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1537 unsigned pixels
= block_size
/ src_count
;
1538 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1539 unsigned alpha_span
= 1;
1540 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1542 /* Check if we need 2 src_alphas for our shuffles */
1543 if (pixels
> alpha_type
.length
) {
1547 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1548 for (j
= 0; j
< row_type
.length
; ++j
) {
1549 if (j
< pixels
* channels
) {
1550 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1552 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1556 for (i
= 0; i
< src_count
; ++i
) {
1557 unsigned idx1
= i
, idx2
= i
;
1559 if (alpha_span
> 1){
1564 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1567 LLVMConstVector(shuffles
, row_type
.length
),
1576 * Generates the blend function for unswizzled colour buffers
1577 * Also generates the read & write from colour buffer
1580 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1582 struct lp_fragment_shader_variant
*variant
,
1583 enum pipe_format out_format
,
1584 unsigned int num_fs
,
1585 struct lp_type fs_type
,
1586 LLVMValueRef
* fs_mask
,
1587 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1588 LLVMValueRef context_ptr
,
1589 LLVMValueRef color_ptr
,
1590 LLVMValueRef stride
,
1591 unsigned partial_mask
,
1594 const unsigned alpha_channel
= 3;
1595 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1596 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1597 const unsigned block_size
= block_width
* block_height
;
1598 const unsigned lp_integer_vector_width
= 128;
1600 LLVMBuilderRef builder
= gallivm
->builder
;
1601 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1602 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1603 LLVMValueRef src_alpha
[4 * 4];
1604 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1605 LLVMValueRef src_mask
[4 * 4];
1606 LLVMValueRef src
[4 * 4];
1607 LLVMValueRef src1
[4 * 4];
1608 LLVMValueRef dst
[4 * 4];
1609 LLVMValueRef blend_color
;
1610 LLVMValueRef blend_alpha
;
1611 LLVMValueRef i32_zero
;
1612 LLVMValueRef check_mask
;
1613 LLVMValueRef undef_src_val
;
1615 struct lp_build_mask_context mask_ctx
;
1616 struct lp_type mask_type
;
1617 struct lp_type blend_type
;
1618 struct lp_type row_type
;
1619 struct lp_type dst_type
;
1621 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1622 unsigned vector_width
;
1623 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1624 unsigned dst_channels
;
1629 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1631 unsigned dst_alignment
;
1633 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1634 bool has_alpha
= false;
1635 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1636 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1638 const boolean is_1d
= variant
->key
.resource_1d
;
1639 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1640 LLVMValueRef fpstate
= 0;
1642 /* Get type from output format */
1643 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1644 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1647 * Technically this code should go into lp_build_smallfloat_to_float
1648 * and lp_build_float_to_smallfloat but due to the
1649 * http://llvm.org/bugs/show_bug.cgi?id=6393
1650 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1651 * So the ordering is important here and there shouldn't be any
1652 * llvm ir instrunctions in this function before
1653 * this, otherwise half-float format conversions won't work
1654 * (again due to llvm bug #6393).
1656 if (have_smallfloat_format(dst_type
, out_format
)) {
1657 /* We need to make sure that denorms are ok for half float
1659 fpstate
= lp_build_fpstate_get(gallivm
);
1660 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1663 mask_type
= lp_int32_vec4_type();
1664 mask_type
.length
= fs_type
.length
;
1666 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1667 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1670 /* Do not bother executing code when mask is empty.. */
1672 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1674 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1675 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1678 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1679 lp_build_mask_check(&mask_ctx
);
1682 partial_mask
|= !variant
->opaque
;
1683 i32_zero
= lp_build_const_int32(gallivm
, 0);
1685 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1687 row_type
.length
= fs_type
.length
;
1688 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1690 /* Compute correct swizzle and count channels */
1691 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1694 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1695 /* Ensure channel is used */
1696 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1700 /* Ensure not already written to (happens in case with GL_ALPHA) */
1701 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1705 /* Ensure we havn't already found all channels */
1706 if (dst_channels
>= out_format_desc
->nr_channels
) {
1710 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1713 if (i
== alpha_channel
) {
1718 if (format_expands_to_float_soa(out_format_desc
)) {
1720 * the code above can't work for layout_other
1721 * for srgb it would sort of work but we short-circuit swizzles, etc.
1722 * as that is done as part of unpack / pack.
1724 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1730 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1733 /* If 3 channels then pad to include alpha for 4 element transpose */
1734 if (dst_channels
== 3 && !has_alpha
) {
1735 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1736 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1739 if (out_format_desc
->nr_channels
== 4) {
1745 * Load shader output
1747 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1748 /* Always load alpha for use in blending */
1751 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1754 alpha
= undef_src_val
;
1757 /* Load each channel */
1758 for (j
= 0; j
< dst_channels
; ++j
) {
1759 assert(swizzle
[j
] < 4);
1761 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1764 fs_src
[i
][j
] = undef_src_val
;
1768 /* If 3 channels then pad to include alpha for 4 element transpose */
1770 * XXX If we include that here maybe could actually use it instead of
1771 * separate alpha for blending?
1773 if (dst_channels
== 3 && !has_alpha
) {
1774 fs_src
[i
][3] = alpha
;
1777 /* We split the row_mask and row_alpha as we want 128bit interleave */
1778 if (fs_type
.length
== 8) {
1779 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1780 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1782 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1783 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1785 src_mask
[i
] = fs_mask
[i
];
1786 src_alpha
[i
] = alpha
;
1789 if (dual_source_blend
) {
1790 /* same as above except different src/dst, skip masks and comments... */
1791 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1794 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1797 alpha
= undef_src_val
;
1800 for (j
= 0; j
< dst_channels
; ++j
) {
1801 assert(swizzle
[j
] < 4);
1803 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1806 fs_src1
[i
][j
] = undef_src_val
;
1809 if (dst_channels
== 3 && !has_alpha
) {
1810 fs_src1
[i
][3] = alpha
;
1812 if (fs_type
.length
== 8) {
1813 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1814 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1816 src1_alpha
[i
] = alpha
;
1821 if (util_format_is_pure_integer(out_format
)) {
1823 * In this case fs_type was really ints or uints disguised as floats,
1826 fs_type
.floating
= 0;
1827 fs_type
.sign
= dst_type
.sign
;
1828 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1829 for (j
= 0; j
< dst_channels
; ++j
) {
1830 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1831 lp_build_vec_type(gallivm
, fs_type
), "");
1833 if (dst_channels
== 3 && !has_alpha
) {
1834 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1835 lp_build_vec_type(gallivm
, fs_type
), "");
1841 * Pixel twiddle from fragment shader order to memory order
1843 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1844 dst_channels
, fs_src
, src
, pad_inline
);
1845 if (dual_source_blend
) {
1846 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1847 fs_src1
, src1
, pad_inline
);
1850 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1851 if (src_count
!= num_fullblock_fs
* src_channels
) {
1852 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1853 row_type
.length
/= ds
;
1854 fs_type
.length
= row_type
.length
;
1857 blend_type
= row_type
;
1858 mask_type
.length
= 4;
1860 /* Convert src to row_type */
1861 if (dual_source_blend
) {
1862 struct lp_type old_row_type
= row_type
;
1863 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1864 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1867 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1870 /* If the rows are not an SSE vector, combine them to become SSE size! */
1871 if ((row_type
.width
* row_type
.length
) % 128) {
1872 unsigned bits
= row_type
.width
* row_type
.length
;
1875 assert(src_count
>= (vector_width
/ bits
));
1877 dst_count
= src_count
/ (vector_width
/ bits
);
1879 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1880 if (dual_source_blend
) {
1881 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1884 row_type
.length
*= combined
;
1885 src_count
/= combined
;
1887 bits
= row_type
.width
* row_type
.length
;
1888 assert(bits
== 128 || bits
== 256);
1893 * Blend Colour conversion
1895 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1896 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1897 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1900 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1902 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1904 * since blending is done with floats, there was no conversion.
1905 * However, the rules according to fixed point renderbuffers still
1906 * apply, that is we must clamp inputs to 0.0/1.0.
1907 * (This would apply to separate alpha conversion too but we currently
1908 * force has_alpha to be true.)
1909 * TODO: should skip this with "fake" blend, since post-blend conversion
1910 * will clamp anyway.
1911 * TODO: could also skip this if fragment color clamping is enabled. We
1912 * don't support it natively so it gets baked into the shader however, so
1913 * can't really tell here.
1915 struct lp_build_context f32_bld
;
1916 assert(row_type
.floating
);
1917 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1918 for (i
= 0; i
< src_count
; i
++) {
1919 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
1921 if (dual_source_blend
) {
1922 for (i
= 0; i
< src_count
; i
++) {
1923 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
1926 /* probably can't be different than row_type but better safe than sorry... */
1927 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1928 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1932 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1934 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1935 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1937 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1938 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1940 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1941 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1947 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1949 if (src_count
< block_height
) {
1950 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1951 } else if (src_count
> block_height
) {
1952 for (i
= src_count
; i
> 0; --i
) {
1953 unsigned pixels
= block_size
/ src_count
;
1954 unsigned idx
= i
- 1;
1956 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1957 (idx
* pixels
) % 4, pixels
);
1961 assert(mask_type
.width
== 32);
1963 for (i
= 0; i
< src_count
; ++i
) {
1964 unsigned pixels
= block_size
/ src_count
;
1965 unsigned pixel_width
= row_type
.width
* dst_channels
;
1967 if (pixel_width
== 24) {
1968 mask_type
.width
= 8;
1969 mask_type
.length
= vector_width
/ mask_type
.width
;
1971 mask_type
.length
= pixels
;
1972 mask_type
.width
= row_type
.width
* dst_channels
;
1974 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1976 mask_type
.length
*= dst_channels
;
1977 mask_type
.width
/= dst_channels
;
1980 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1981 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1988 struct lp_type alpha_type
= fs_type
;
1989 alpha_type
.length
= 4;
1990 convert_alpha(gallivm
, row_type
, alpha_type
,
1991 block_size
, block_height
,
1992 src_count
, dst_channels
,
1993 pad_inline
, src_alpha
);
1994 if (dual_source_blend
) {
1995 convert_alpha(gallivm
, row_type
, alpha_type
,
1996 block_size
, block_height
,
1997 src_count
, dst_channels
,
1998 pad_inline
, src1_alpha
);
2004 * Load dst from memory
2006 if (src_count
< block_height
) {
2007 dst_count
= block_height
;
2009 dst_count
= src_count
;
2012 dst_type
.length
*= block_size
/ dst_count
;
2014 if (format_expands_to_float_soa(out_format_desc
)) {
2016 * we need multiple values at once for the conversion, so can as well
2017 * load them vectorized here too instead of concatenating later.
2018 * (Still need concatenation later for 8-wide vectors).
2020 dst_count
= block_height
;
2021 dst_type
.length
= block_width
;
2025 * Compute the alignment of the destination pointer in bytes
2026 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2027 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2028 * 1d tex but can't distinguish here) so need to stick with per-pixel
2029 * alignment in this case.
2032 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2035 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2037 /* Force power-of-two alignment by extracting only the least-significant-bit */
2038 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2040 * Resource base and stride pointers are aligned to 16 bytes, so that's
2041 * the maximum alignment we can guarantee
2043 dst_alignment
= MIN2(16, dst_alignment
);
2046 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2047 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2048 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2049 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
2054 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2055 dst
, dst_type
, dst_count
, dst_alignment
);
2060 * Convert from dst/output format to src/blending format.
2062 * This is necessary as we can only read 1 row from memory at a time,
2063 * so the minimum dst_count will ever be at this point is 4.
2065 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2066 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2067 * on all 16 pixels in that single vector at once.
2069 if (dst_count
> src_count
) {
2070 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
2076 /* XXX this is broken for RGB8 formats -
2077 * they get expanded from 12 to 16 elements (to include alpha)
2078 * by convert_to_blend_type then reduced to 15 instead of 12
2079 * by convert_from_blend_type (a simple fix though breaks A8...).
2080 * R16G16B16 also crashes differently however something going wrong
2081 * inside llvm handling npot vector sizes seemingly.
2082 * It seems some cleanup could be done here (like skipping conversion/blend
2085 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
2088 * FIXME: Really should get logic ops / masks out of generic blend / row
2089 * format. Logic ops will definitely not work on the blend float format
2090 * used for SRGB here and I think OpenGL expects this to work as expected
2091 * (that is incoming values converted to srgb then logic op applied).
2093 for (i
= 0; i
< src_count
; ++i
) {
2094 dst
[i
] = lp_build_blend_aos(gallivm
,
2095 &variant
->key
.blend
,
2100 has_alpha
? NULL
: src_alpha
[i
],
2102 has_alpha
? NULL
: src1_alpha
[i
],
2104 partial_mask
? src_mask
[i
] : NULL
,
2106 has_alpha
? NULL
: blend_alpha
,
2108 pad_inline
? 4 : dst_channels
);
2111 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
2113 /* Split the blend rows back to memory rows */
2114 if (dst_count
> src_count
) {
2115 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2117 if (src_count
== 1) {
2118 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2119 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2121 row_type
.length
/= 2;
2125 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2126 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2127 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2128 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2130 row_type
.length
/= 2;
2135 * Store blend result to memory
2138 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2139 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2142 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2143 dst
, dst_type
, dst_count
, dst_alignment
);
2146 if (have_smallfloat_format(dst_type
, out_format
)) {
2147 lp_build_fpstate_set(gallivm
, fpstate
);
2151 lp_build_mask_end(&mask_ctx
);
2157 * Generate the runtime callable function for the whole fragment pipeline.
2158 * Note that the function which we generate operates on a block of 16
2159 * pixels at at time. The block contains 2x2 quads. Each quad contains
2163 generate_fragment(struct llvmpipe_context
*lp
,
2164 struct lp_fragment_shader
*shader
,
2165 struct lp_fragment_shader_variant
*variant
,
2166 unsigned partial_mask
)
2168 struct gallivm_state
*gallivm
= variant
->gallivm
;
2169 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2170 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2172 struct lp_type fs_type
;
2173 struct lp_type blend_type
;
2174 LLVMTypeRef fs_elem_type
;
2175 LLVMTypeRef blend_vec_type
;
2176 LLVMTypeRef arg_types
[13];
2177 LLVMTypeRef func_type
;
2178 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2179 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2180 LLVMValueRef context_ptr
;
2183 LLVMValueRef a0_ptr
;
2184 LLVMValueRef dadx_ptr
;
2185 LLVMValueRef dady_ptr
;
2186 LLVMValueRef color_ptr_ptr
;
2187 LLVMValueRef stride_ptr
;
2188 LLVMValueRef depth_ptr
;
2189 LLVMValueRef depth_stride
;
2190 LLVMValueRef mask_input
;
2191 LLVMValueRef thread_data_ptr
;
2192 LLVMBasicBlockRef block
;
2193 LLVMBuilderRef builder
;
2194 struct lp_build_sampler_soa
*sampler
;
2195 struct lp_build_interp_soa_context interp
;
2196 LLVMValueRef fs_mask
[16 / 4];
2197 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2198 LLVMValueRef function
;
2199 LLVMValueRef facing
;
2204 boolean cbuf0_write_all
;
2205 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2206 util_blend_state_is_dual(&key
->blend
, 0);
2208 assert(lp_native_vector_width
/ 32 >= 4);
2210 /* Adjust color input interpolation according to flatshade state:
2212 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2213 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2214 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2216 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2218 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2222 /* check if writes to cbuf[0] are to be copied to all cbufs */
2224 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2226 /* TODO: actually pick these based on the fs and color buffer
2227 * characteristics. */
2229 memset(&fs_type
, 0, sizeof fs_type
);
2230 fs_type
.floating
= TRUE
; /* floating point values */
2231 fs_type
.sign
= TRUE
; /* values are signed */
2232 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2233 fs_type
.width
= 32; /* 32-bit float */
2234 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2236 memset(&blend_type
, 0, sizeof blend_type
);
2237 blend_type
.floating
= FALSE
; /* values are integers */
2238 blend_type
.sign
= FALSE
; /* values are unsigned */
2239 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2240 blend_type
.width
= 8; /* 8-bit ubyte values */
2241 blend_type
.length
= 16; /* 16 elements per vector */
2244 * Generate the function prototype. Any change here must be reflected in
2245 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2248 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2250 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2252 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2253 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2255 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2256 arg_types
[1] = int32_type
; /* x */
2257 arg_types
[2] = int32_type
; /* y */
2258 arg_types
[3] = int32_type
; /* facing */
2259 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2260 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2261 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2262 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2263 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2264 arg_types
[9] = int32_type
; /* mask_input */
2265 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2266 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2267 arg_types
[12] = int32_type
; /* depth_stride */
2269 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2270 arg_types
, Elements(arg_types
), 0);
2272 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2273 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2275 variant
->function
[partial_mask
] = function
;
2277 /* XXX: need to propagate noalias down into color param now we are
2278 * passing a pointer-to-pointer?
2280 for(i
= 0; i
< Elements(arg_types
); ++i
)
2281 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2282 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2284 context_ptr
= LLVMGetParam(function
, 0);
2285 x
= LLVMGetParam(function
, 1);
2286 y
= LLVMGetParam(function
, 2);
2287 facing
= LLVMGetParam(function
, 3);
2288 a0_ptr
= LLVMGetParam(function
, 4);
2289 dadx_ptr
= LLVMGetParam(function
, 5);
2290 dady_ptr
= LLVMGetParam(function
, 6);
2291 color_ptr_ptr
= LLVMGetParam(function
, 7);
2292 depth_ptr
= LLVMGetParam(function
, 8);
2293 mask_input
= LLVMGetParam(function
, 9);
2294 thread_data_ptr
= LLVMGetParam(function
, 10);
2295 stride_ptr
= LLVMGetParam(function
, 11);
2296 depth_stride
= LLVMGetParam(function
, 12);
2298 lp_build_name(context_ptr
, "context");
2299 lp_build_name(x
, "x");
2300 lp_build_name(y
, "y");
2301 lp_build_name(a0_ptr
, "a0");
2302 lp_build_name(dadx_ptr
, "dadx");
2303 lp_build_name(dady_ptr
, "dady");
2304 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2305 lp_build_name(depth_ptr
, "depth");
2306 lp_build_name(mask_input
, "mask_input");
2307 lp_build_name(thread_data_ptr
, "thread_data");
2308 lp_build_name(stride_ptr
, "stride_ptr");
2309 lp_build_name(depth_stride
, "depth_stride");
2315 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2316 builder
= gallivm
->builder
;
2318 LLVMPositionBuilderAtEnd(builder
, block
);
2320 /* code generated texture sampling */
2321 sampler
= lp_llvm_sampler_soa_create(key
->state
);
2323 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2324 /* for 1d resources only run "upper half" of stamp */
2325 if (key
->resource_1d
)
2329 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2330 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2331 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2332 num_loop
, "mask_store");
2333 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2334 boolean pixel_center_integer
=
2335 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2338 * The shader input interpolation info is not explicitely baked in the
2339 * shader key, but everything it derives from (TGSI, and flatshade) is
2340 * already included in the shader key.
2342 lp_build_interp_soa_init(&interp
,
2344 shader
->info
.base
.num_inputs
,
2346 pixel_center_integer
,
2348 a0_ptr
, dadx_ptr
, dady_ptr
,
2351 for (i
= 0; i
< num_fs
; i
++) {
2353 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2354 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2355 &indexi
, 1, "mask_ptr");
2358 mask
= generate_quad_mask(gallivm
, fs_type
,
2359 i
*fs_type
.length
/4, mask_input
);
2362 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2364 LLVMBuildStore(builder
, mask
, mask_ptr
);
2367 generate_fs_loop(gallivm
,
2375 mask_store
, /* output */
2382 for (i
= 0; i
< num_fs
; i
++) {
2383 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2384 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2386 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2387 /* This is fucked up need to reorganize things */
2388 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2389 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2390 ptr
= LLVMBuildGEP(builder
,
2391 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2393 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2396 if (dual_source_blend
) {
2397 /* only support one dual source blend target hence always use output 1 */
2398 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2399 ptr
= LLVMBuildGEP(builder
,
2400 color_store
[1][chan
],
2402 fs_out_color
[1][chan
][i
] = ptr
;
2408 sampler
->destroy(sampler
);
2410 /* Loop over color outputs / color buffers to do blending.
2412 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2413 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2414 LLVMValueRef color_ptr
;
2415 LLVMValueRef stride
;
2416 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2418 boolean do_branch
= ((key
->depth
.enabled
2419 || key
->stencil
[0].enabled
2420 || key
->alpha
.enabled
)
2421 && !shader
->info
.base
.uses_kill
);
2423 color_ptr
= LLVMBuildLoad(builder
,
2424 LLVMBuildGEP(builder
, color_ptr_ptr
,
2428 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2430 stride
= LLVMBuildLoad(builder
,
2431 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2434 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2435 key
->cbuf_format
[cbuf
],
2436 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2437 context_ptr
, color_ptr
, stride
,
2438 partial_mask
, do_branch
);
2442 LLVMBuildRetVoid(builder
);
2444 gallivm_verify_function(gallivm
, function
);
2449 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2453 debug_printf("fs variant %p:\n", (void *) key
);
2455 if (key
->flatshade
) {
2456 debug_printf("flatshade = 1\n");
2458 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2459 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2461 if (key
->depth
.enabled
) {
2462 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2463 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2464 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2467 for (i
= 0; i
< 2; ++i
) {
2468 if (key
->stencil
[i
].enabled
) {
2469 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2470 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2471 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2472 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2473 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2474 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2478 if (key
->alpha
.enabled
) {
2479 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2482 if (key
->occlusion_count
) {
2483 debug_printf("occlusion_count = 1\n");
2486 if (key
->blend
.logicop_enable
) {
2487 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2489 else if (key
->blend
.rt
[0].blend_enable
) {
2490 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2491 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2492 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2493 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2494 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2495 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2497 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2498 if (key
->blend
.alpha_to_coverage
) {
2499 debug_printf("blend.alpha_to_coverage is enabled\n");
2501 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2502 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2503 debug_printf("sampler[%u] = \n", i
);
2504 debug_printf(" .wrap = %s %s %s\n",
2505 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2506 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2507 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2508 debug_printf(" .min_img_filter = %s\n",
2509 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2510 debug_printf(" .min_mip_filter = %s\n",
2511 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2512 debug_printf(" .mag_img_filter = %s\n",
2513 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2514 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2515 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2516 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2517 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2518 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2519 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2520 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2522 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2523 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2524 debug_printf("texture[%u] = \n", i
);
2525 debug_printf(" .format = %s\n",
2526 util_format_name(texture
->format
));
2527 debug_printf(" .target = %s\n",
2528 util_dump_tex_target(texture
->target
, TRUE
));
2529 debug_printf(" .level_zero_only = %u\n",
2530 texture
->level_zero_only
);
2531 debug_printf(" .pot = %u %u %u\n",
2533 texture
->pot_height
,
2534 texture
->pot_depth
);
2540 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2542 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2543 variant
->shader
->no
, variant
->no
);
2544 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2545 dump_fs_variant_key(&variant
->key
);
2546 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2552 * Generate a new fragment shader variant from the shader code and
2553 * other state indicated by the key.
2555 static struct lp_fragment_shader_variant
*
2556 generate_variant(struct llvmpipe_context
*lp
,
2557 struct lp_fragment_shader
*shader
,
2558 const struct lp_fragment_shader_variant_key
*key
)
2560 struct lp_fragment_shader_variant
*variant
;
2561 const struct util_format_description
*cbuf0_format_desc
;
2562 boolean fullcolormask
;
2563 char module_name
[64];
2565 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2569 util_snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2570 shader
->no
, shader
->variants_created
);
2572 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
2573 if (!variant
->gallivm
) {
2578 variant
->shader
= shader
;
2579 variant
->list_item_global
.base
= variant
;
2580 variant
->list_item_local
.base
= variant
;
2581 variant
->no
= shader
->variants_created
++;
2583 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2586 * Determine whether we are touching all channels in the color buffer.
2588 fullcolormask
= FALSE
;
2589 if (key
->nr_cbufs
== 1) {
2590 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2591 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2595 !key
->blend
.logicop_enable
&&
2596 !key
->blend
.rt
[0].blend_enable
&&
2598 !key
->stencil
[0].enabled
&&
2599 !key
->alpha
.enabled
&&
2600 !key
->blend
.alpha_to_coverage
&&
2601 !key
->depth
.enabled
&&
2602 !shader
->info
.base
.uses_kill
2605 if ((shader
->info
.base
.num_tokens
<= 1) &&
2606 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2607 variant
->ps_inv_multiplier
= 0;
2609 variant
->ps_inv_multiplier
= 1;
2612 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2613 lp_debug_fs_variant(variant
);
2616 lp_jit_init_types(variant
);
2618 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2619 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2621 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2622 if (variant
->opaque
) {
2623 /* Specialized shader, which doesn't need to read the color buffer. */
2624 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2629 * Compile everything
2632 gallivm_compile_module(variant
->gallivm
);
2634 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2636 if (variant
->function
[RAST_EDGE_TEST
]) {
2637 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2638 gallivm_jit_function(variant
->gallivm
,
2639 variant
->function
[RAST_EDGE_TEST
]);
2642 if (variant
->function
[RAST_WHOLE
]) {
2643 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2644 gallivm_jit_function(variant
->gallivm
,
2645 variant
->function
[RAST_WHOLE
]);
2646 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2647 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2650 gallivm_free_ir(variant
->gallivm
);
2657 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2658 const struct pipe_shader_state
*templ
)
2660 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2661 struct lp_fragment_shader
*shader
;
2663 int nr_sampler_views
;
2666 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2670 shader
->no
= fs_no
++;
2671 make_empty_list(&shader
->variants
);
2673 /* get/save the summary info for this shader */
2674 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2676 /* we need to keep a local copy of the tokens */
2677 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2679 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2680 if (shader
->draw_data
== NULL
) {
2681 FREE((void *) shader
->base
.tokens
);
2686 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2687 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2689 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2690 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2692 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2693 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2694 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2696 switch (shader
->info
.base
.input_interpolate
[i
]) {
2697 case TGSI_INTERPOLATE_CONSTANT
:
2698 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2700 case TGSI_INTERPOLATE_LINEAR
:
2701 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2703 case TGSI_INTERPOLATE_PERSPECTIVE
:
2704 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2706 case TGSI_INTERPOLATE_COLOR
:
2707 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2714 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2715 case TGSI_SEMANTIC_FACE
:
2716 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2718 case TGSI_SEMANTIC_POSITION
:
2719 /* Position was already emitted above
2721 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2722 shader
->inputs
[i
].src_index
= 0;
2726 /* XXX this is a completely pointless index map... */
2727 shader
->inputs
[i
].src_index
= i
+1;
2730 if (LP_DEBUG
& DEBUG_TGSI
) {
2732 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2733 shader
->no
, (void *) shader
);
2734 tgsi_dump(templ
->tokens
, 0);
2735 debug_printf("usage masks:\n");
2736 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2737 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2738 debug_printf(" IN[%u].%s%s%s%s\n",
2740 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2741 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2742 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2743 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2753 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2755 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2757 if (llvmpipe
->fs
== fs
)
2760 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2762 draw_bind_fragment_shader(llvmpipe
->draw
,
2763 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2765 llvmpipe
->dirty
|= LP_NEW_FS
;
2770 * Remove shader variant from two lists: the shader's variant list
2771 * and the context's variant list.
2774 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2775 struct lp_fragment_shader_variant
*variant
)
2777 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2778 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2779 " #%u v total cached #%u\n",
2780 variant
->shader
->no
,
2782 variant
->shader
->variants_created
,
2783 variant
->shader
->variants_cached
,
2784 lp
->nr_fs_variants
);
2787 gallivm_destroy(variant
->gallivm
);
2789 /* remove from shader's list */
2790 remove_from_list(&variant
->list_item_local
);
2791 variant
->shader
->variants_cached
--;
2793 /* remove from context's list */
2794 remove_from_list(&variant
->list_item_global
);
2795 lp
->nr_fs_variants
--;
2796 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2803 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2805 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2806 struct lp_fragment_shader
*shader
= fs
;
2807 struct lp_fs_variant_list_item
*li
;
2809 assert(fs
!= llvmpipe
->fs
);
2812 * XXX: we need to flush the context until we have some sort of reference
2813 * counting in fragment shaders as they may still be binned
2814 * Flushing alone might not sufficient we need to wait on it too.
2816 llvmpipe_finish(pipe
, __FUNCTION__
);
2818 /* Delete all the variants */
2819 li
= first_elem(&shader
->variants
);
2820 while(!at_end(&shader
->variants
, li
)) {
2821 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2822 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2826 /* Delete draw module's data */
2827 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2829 assert(shader
->variants_cached
== 0);
2830 FREE((void *) shader
->base
.tokens
);
2837 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2838 uint shader
, uint index
,
2839 struct pipe_constant_buffer
*cb
)
2841 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2842 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2844 assert(shader
< PIPE_SHADER_TYPES
);
2845 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2847 /* note: reference counting */
2848 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2850 if (shader
== PIPE_SHADER_VERTEX
||
2851 shader
== PIPE_SHADER_GEOMETRY
) {
2852 /* Pass the constants to the 'draw' module */
2853 const unsigned size
= cb
? cb
->buffer_size
: 0;
2857 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2859 else if (cb
&& cb
->user_buffer
) {
2860 data
= (ubyte
*) cb
->user_buffer
;
2867 data
+= cb
->buffer_offset
;
2869 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2873 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2875 if (cb
&& cb
->user_buffer
) {
2876 pipe_resource_reference(&constants
, NULL
);
2882 * Return the blend factor equivalent to a destination alpha of one.
2884 static inline unsigned
2885 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2888 case PIPE_BLENDFACTOR_DST_ALPHA
:
2889 return PIPE_BLENDFACTOR_ONE
;
2890 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2891 return PIPE_BLENDFACTOR_ZERO
;
2892 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2894 return PIPE_BLENDFACTOR_ZERO
;
2896 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2904 * We need to generate several variants of the fragment pipeline to match
2905 * all the combinations of the contributing state atoms.
2907 * TODO: there is actually no reason to tie this to context state -- the
2908 * generated code could be cached globally in the screen.
2911 make_variant_key(struct llvmpipe_context
*lp
,
2912 struct lp_fragment_shader
*shader
,
2913 struct lp_fragment_shader_variant_key
*key
)
2917 memset(key
, 0, shader
->variant_key_size
);
2919 if (lp
->framebuffer
.zsbuf
) {
2920 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2921 const struct util_format_description
*zsbuf_desc
=
2922 util_format_description(zsbuf_format
);
2924 if (lp
->depth_stencil
->depth
.enabled
&&
2925 util_format_has_depth(zsbuf_desc
)) {
2926 key
->zsbuf_format
= zsbuf_format
;
2927 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2929 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2930 util_format_has_stencil(zsbuf_desc
)) {
2931 key
->zsbuf_format
= zsbuf_format
;
2932 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2934 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2935 key
->resource_1d
= TRUE
;
2940 * Propagate the depth clamp setting from the rasterizer state.
2941 * depth_clip == 0 implies depth clamping is enabled.
2943 * When clip_halfz is enabled, then always clamp the depth values.
2945 if (lp
->rasterizer
->clip_halfz
) {
2946 key
->depth_clamp
= 1;
2948 key
->depth_clamp
= (lp
->rasterizer
->depth_clip
== 0) ? 1 : 0;
2951 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2952 if (!lp
->framebuffer
.nr_cbufs
||
2953 !lp
->framebuffer
.cbufs
[0] ||
2954 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2955 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2957 if(key
->alpha
.enabled
)
2958 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2959 /* alpha.ref_value is passed in jit_context */
2961 key
->flatshade
= lp
->rasterizer
->flatshade
;
2962 if (lp
->active_occlusion_queries
) {
2963 key
->occlusion_count
= TRUE
;
2966 if (lp
->framebuffer
.nr_cbufs
) {
2967 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2970 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2972 if (!key
->blend
.independent_blend_enable
) {
2973 /* we always need independent blend otherwise the fixups below won't work */
2974 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2975 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2977 key
->blend
.independent_blend_enable
= 1;
2980 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2981 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2983 if (lp
->framebuffer
.cbufs
[i
]) {
2984 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2985 const struct util_format_description
*format_desc
;
2987 key
->cbuf_format
[i
] = format
;
2990 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2991 * mixing of 2d textures with height 1 and 1d textures, so make sure
2992 * we pick 1d if any cbuf or zsbuf is 1d.
2994 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
2995 key
->resource_1d
= TRUE
;
2998 format_desc
= util_format_description(format
);
2999 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3000 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3003 * Mask out color channels not present in the color buffer.
3005 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3008 * Disable blend for integer formats.
3010 if (util_format_is_pure_integer(format
)) {
3011 blend_rt
->blend_enable
= 0;
3015 * Our swizzled render tiles always have an alpha channel, but the
3016 * linear render target format often does not, so force here the dst
3019 * This is not a mere optimization. Wrong results will be produced if
3020 * the dst alpha is used, the dst format does not have alpha, and the
3021 * previous rendering was not flushed from the swizzled to linear
3022 * buffer. For example, NonPowTwo DCT.
3024 * TODO: This should be generalized to all channels for better
3025 * performance, but only alpha causes correctness issues.
3027 * Also, force rgb/alpha func/factors match, to make AoS blending
3030 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
3031 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3032 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3033 boolean clamped_zero
= !util_format_is_float(format
) &&
3034 !util_format_is_snorm(format
);
3035 blend_rt
->rgb_src_factor
=
3036 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3037 blend_rt
->rgb_dst_factor
=
3038 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3039 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3040 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3041 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3045 /* no color buffer for this fragment output */
3046 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3047 blend_rt
->colormask
= 0x0;
3048 blend_rt
->blend_enable
= 0;
3052 /* This value will be the same for all the variants of a given shader:
3054 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3056 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3057 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3058 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
3059 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3064 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3065 * are dx10-style? Can't really have mixed opcodes, at least not
3066 * if we want to skip the holes here (without rescanning tgsi).
3068 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3069 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3070 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3071 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
3072 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3073 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3078 key
->nr_sampler_views
= key
->nr_samplers
;
3079 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3080 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3081 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3082 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3091 * Update fragment shader state. This is called just prior to drawing
3092 * something when some fragment-related state has changed.
3095 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3097 struct lp_fragment_shader
*shader
= lp
->fs
;
3098 struct lp_fragment_shader_variant_key key
;
3099 struct lp_fragment_shader_variant
*variant
= NULL
;
3100 struct lp_fs_variant_list_item
*li
;
3102 make_variant_key(lp
, shader
, &key
);
3104 /* Search the variants for one which matches the key */
3105 li
= first_elem(&shader
->variants
);
3106 while(!at_end(&shader
->variants
, li
)) {
3107 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3115 /* Move this variant to the head of the list to implement LRU
3116 * deletion of shader's when we have too many.
3118 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3121 /* variant not found, create it now */
3124 unsigned variants_to_cull
;
3127 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3130 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3133 /* First, check if we've exceeded the max number of shader variants.
3134 * If so, free 25% of them (the least recently used ones).
3136 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
3138 if (variants_to_cull
||
3139 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3140 struct pipe_context
*pipe
= &lp
->pipe
;
3143 * XXX: we need to flush the context until we have some sort of
3144 * reference counting in fragment shaders as they may still be binned
3145 * Flushing alone might not be sufficient we need to wait on it too.
3147 llvmpipe_finish(pipe
, __FUNCTION__
);
3150 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3151 * number of shader variants (potentially all of them) could be
3152 * pending for destruction on flush.
3155 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3156 struct lp_fs_variant_list_item
*item
;
3157 if (is_empty_list(&lp
->fs_variants_list
)) {
3160 item
= last_elem(&lp
->fs_variants_list
);
3163 llvmpipe_remove_shader_variant(lp
, item
->base
);
3168 * Generate the new variant.
3171 variant
= generate_variant(lp
, shader
, &key
);
3174 LP_COUNT_ADD(llvm_compile_time
, dt
);
3175 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3177 /* Put the new variant into the list */
3179 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3180 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3181 lp
->nr_fs_variants
++;
3182 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3183 shader
->variants_cached
++;
3187 /* Bind this variant */
3188 lp_setup_set_fs_variant(lp
->setup
, variant
);
3196 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3198 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3199 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3200 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3202 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3206 * Rasterization is disabled if there is no pixel shader and
3207 * both depth and stencil testing are disabled:
3208 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3211 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
3213 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
3216 !lp
->depth_stencil
->depth
.enabled
&&
3217 !lp
->depth_stencil
->stencil
[0].enabled
);