1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
37 * - depth/stencil test
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
57 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/u_simple_list.h"
69 #include "util/u_dual_blend.h"
70 #include "os/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_swizzle.h"
84 #include "gallivm/lp_bld_flow.h"
85 #include "gallivm/lp_bld_debug.h"
86 #include "gallivm/lp_bld_arit.h"
87 #include "gallivm/lp_bld_pack.h"
88 #include "gallivm/lp_bld_format.h"
89 #include "gallivm/lp_bld_quad.h"
91 #include "lp_bld_alpha.h"
92 #include "lp_bld_blend.h"
93 #include "lp_bld_depth.h"
94 #include "lp_bld_interp.h"
95 #include "lp_context.h"
100 #include "lp_tex_sample.h"
101 #include "lp_flush.h"
102 #include "lp_state_fs.h"
106 /** Fragment shader number (for debugging) */
107 static unsigned fs_no
= 0;
111 * Expand the relevant bits of mask_input to a n*4-dword mask for the
112 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
113 * quad mask vector to 0 or ~0.
114 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
115 * quad arguments with fs length 8.
117 * \param first_quad which quad(s) of the quad group to test, in [0,3]
118 * \param mask_input bitwise mask for the whole 4x4 stamp
121 generate_quad_mask(struct gallivm_state
*gallivm
,
122 struct lp_type fs_type
,
124 LLVMValueRef mask_input
) /* int32 */
126 LLVMBuilderRef builder
= gallivm
->builder
;
127 struct lp_type mask_type
;
128 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
129 LLVMValueRef bits
[16];
134 * XXX: We'll need a different path for 16 x u8
136 assert(fs_type
.width
== 32);
137 assert(fs_type
.length
<= Elements(bits
));
138 mask_type
= lp_int_type(fs_type
);
141 * mask_input >>= (quad * 4)
143 switch (first_quad
) {
148 assert(fs_type
.length
== 4);
155 assert(fs_type
.length
== 4);
163 mask_input
= LLVMBuildLShr(builder
,
165 LLVMConstInt(i32t
, shift
, 0),
169 * mask = { mask_input & (1 << i), for i in [0,3] }
171 mask
= lp_build_broadcast(gallivm
,
172 lp_build_vec_type(gallivm
, mask_type
),
175 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
176 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
177 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
178 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
179 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
180 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
182 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
185 * mask = mask != 0 ? ~0 : 0
187 mask
= lp_build_compare(gallivm
,
188 mask_type
, PIPE_FUNC_NOTEQUAL
,
190 lp_build_const_int_vec(gallivm
, mask_type
, 0));
196 #define EARLY_DEPTH_TEST 0x1
197 #define LATE_DEPTH_TEST 0x2
198 #define EARLY_DEPTH_WRITE 0x4
199 #define LATE_DEPTH_WRITE 0x8
202 find_output_by_semantic( const struct tgsi_shader_info
*info
,
208 for (i
= 0; i
< info
->num_outputs
; i
++)
209 if (info
->output_semantic_name
[i
] == semantic
&&
210 info
->output_semantic_index
[i
] == index
)
218 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
221 lp_llvm_viewport(LLVMValueRef context_ptr
,
222 struct gallivm_state
*gallivm
,
223 LLVMValueRef viewport_index
)
225 LLVMBuilderRef builder
= gallivm
->builder
;
228 struct lp_type viewport_type
=
229 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
231 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
232 ptr
= LLVMBuildPointerCast(builder
, ptr
,
233 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
235 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
242 * Generate the fragment shader, depth/stencil test, and alpha tests.
245 generate_fs_loop(struct gallivm_state
*gallivm
,
246 struct lp_fragment_shader
*shader
,
247 const struct lp_fragment_shader_variant_key
*key
,
248 LLVMBuilderRef builder
,
250 LLVMValueRef context_ptr
,
251 LLVMValueRef num_loop
,
252 struct lp_build_interp_soa_context
*interp
,
253 struct lp_build_sampler_soa
*sampler
,
254 LLVMValueRef mask_store
,
255 LLVMValueRef (*out_color
)[4],
256 LLVMValueRef depth_ptr
,
257 LLVMValueRef depth_stride
,
259 LLVMValueRef thread_data_ptr
)
261 const struct util_format_description
*zs_format_desc
= NULL
;
262 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
263 LLVMTypeRef vec_type
;
264 LLVMValueRef mask_ptr
, mask_val
;
265 LLVMValueRef consts_ptr
;
267 LLVMValueRef z_value
, s_value
;
268 LLVMValueRef z_fb
, s_fb
;
269 LLVMValueRef stencil_refs
[2];
270 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
271 struct lp_build_for_loop_state loop_state
;
272 struct lp_build_mask_context mask
;
274 * TODO: figure out if simple_shader optimization is really worthwile to
275 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
276 * code since tests tend to take another codepath than real shaders.
278 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
279 shader
->info
.base
.num_inputs
< 3 &&
280 shader
->info
.base
.num_instructions
< 8) && 0;
281 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
282 util_blend_state_is_dual(&key
->blend
, 0);
288 struct lp_bld_tgsi_system_values system_values
;
290 memset(&system_values
, 0, sizeof(system_values
));
292 if (key
->depth
.enabled
||
293 key
->stencil
[0].enabled
) {
295 zs_format_desc
= util_format_description(key
->zsbuf_format
);
296 assert(zs_format_desc
);
298 if (!shader
->info
.base
.writes_z
) {
299 if (key
->alpha
.enabled
||
300 key
->blend
.alpha_to_coverage
||
301 shader
->info
.base
.uses_kill
) {
302 /* With alpha test and kill, can do the depth test early
303 * and hopefully eliminate some quads. But need to do a
304 * special deferred depth write once the final mask value
305 * is known. This only works though if there's either no
306 * stencil test or the stencil value isn't written.
308 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
309 (key
->stencil
[1].enabled
&&
310 key
->stencil
[1].writemask
)))
311 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
313 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
316 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
319 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
322 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
323 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
324 (key
->stencil
[1].enabled
&&
325 key
->stencil
[1].writemask
))))
326 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
333 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
334 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
336 vec_type
= lp_build_vec_type(gallivm
, type
);
338 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
340 lp_build_for_loop_begin(&loop_state
, gallivm
,
341 lp_build_const_int32(gallivm
, 0),
344 lp_build_const_int32(gallivm
, 1));
346 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
347 &loop_state
.counter
, 1, "mask_ptr");
348 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
350 memset(outputs
, 0, sizeof outputs
);
352 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
353 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
354 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
355 lp_build_vec_type(gallivm
,
360 if (dual_source_blend
) {
361 assert(key
->nr_cbufs
<= 1);
362 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
363 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
364 lp_build_vec_type(gallivm
,
371 /* 'mask' will control execution based on quad's pixel alive/killed state */
372 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
374 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
375 lp_build_mask_check(&mask
);
377 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
380 if (depth_mode
& EARLY_DEPTH_TEST
) {
381 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
382 zs_format_desc
, key
->resource_1d
,
383 depth_ptr
, depth_stride
,
384 &z_fb
, &s_fb
, loop_state
.counter
);
385 lp_build_depth_stencil_test(gallivm
,
397 if (depth_mode
& EARLY_DEPTH_WRITE
) {
398 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
399 zs_format_desc
, key
->resource_1d
,
400 NULL
, NULL
, NULL
, loop_state
.counter
,
401 depth_ptr
, depth_stride
,
405 * Note mask check if stencil is enabled must be after ds write not after
406 * stencil test otherwise new stencil values may not get written if all
407 * fragments got killed by depth/stencil test.
409 if (!simple_shader
&& key
->stencil
[0].enabled
)
410 lp_build_mask_check(&mask
);
413 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
415 /* Build the actual shader */
416 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
417 consts_ptr
, &system_values
,
419 outputs
, sampler
, &shader
->info
.base
, NULL
);
422 if (key
->alpha
.enabled
) {
423 int color0
= find_output_by_semantic(&shader
->info
.base
,
427 if (color0
!= -1 && outputs
[color0
][3]) {
428 const struct util_format_description
*cbuf_format_desc
;
429 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
430 LLVMValueRef alpha_ref_value
;
432 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
433 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
435 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
437 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
438 &mask
, alpha
, alpha_ref_value
,
439 (depth_mode
& LATE_DEPTH_TEST
) != 0);
443 /* Emulate Alpha to Coverage with Alpha test */
444 if (key
->blend
.alpha_to_coverage
) {
445 int color0
= find_output_by_semantic(&shader
->info
.base
,
449 if (color0
!= -1 && outputs
[color0
][3]) {
450 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
452 lp_build_alpha_to_coverage(gallivm
, type
,
454 (depth_mode
& LATE_DEPTH_TEST
) != 0);
459 if (depth_mode
& LATE_DEPTH_TEST
) {
460 int pos0
= find_output_by_semantic(&shader
->info
.base
,
461 TGSI_SEMANTIC_POSITION
,
464 if (pos0
!= -1 && outputs
[pos0
][2]) {
465 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
468 * Clamp according to ARB_depth_clamp semantics.
470 if (key
->depth_clamp
) {
471 LLVMValueRef viewport
, min_depth
, max_depth
;
472 LLVMValueRef viewport_index
;
473 struct lp_build_context f32_bld
;
475 assert(type
.floating
);
476 lp_build_context_init(&f32_bld
, gallivm
, type
);
479 * Assumes clamping of the viewport index will occur in setup/gs. Value
480 * is passed through the rasterization stage via lp_rast_shader_inputs.
482 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
485 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
489 * Load the min and max depth from the lp_jit_context.viewports
490 * array of lp_jit_viewport structures.
492 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
494 /* viewports[viewport_index].min_depth */
495 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
496 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
),
498 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
500 /* viewports[viewport_index].max_depth */
501 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
502 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
),
504 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
507 * Clamp to the min and max depth values for the given viewport.
509 z
= lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
513 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
514 zs_format_desc
, key
->resource_1d
,
515 depth_ptr
, depth_stride
,
516 &z_fb
, &s_fb
, loop_state
.counter
);
518 lp_build_depth_stencil_test(gallivm
,
530 if (depth_mode
& LATE_DEPTH_WRITE
) {
531 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
532 zs_format_desc
, key
->resource_1d
,
533 NULL
, NULL
, NULL
, loop_state
.counter
,
534 depth_ptr
, depth_stride
,
538 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
539 (depth_mode
& LATE_DEPTH_WRITE
))
541 /* Need to apply a reduced mask to the depth write. Reload the
542 * depth value, update from zs_value with the new mask value and
545 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
546 zs_format_desc
, key
->resource_1d
,
547 &mask
, z_fb
, s_fb
, loop_state
.counter
,
548 depth_ptr
, depth_stride
,
554 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
556 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
557 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
558 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
560 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
561 if(outputs
[attrib
][chan
]) {
562 /* XXX: just initialize outputs to point at colors[] and
565 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
566 LLVMValueRef color_ptr
;
567 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
568 &loop_state
.counter
, 1, "");
569 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
570 LLVMBuildStore(builder
, out
, color_ptr
);
576 if (key
->occlusion_count
) {
577 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
578 lp_build_name(counter
, "counter");
579 lp_build_occlusion_count(gallivm
, type
,
580 lp_build_mask_value(&mask
), counter
);
583 mask_val
= lp_build_mask_end(&mask
);
584 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
585 lp_build_for_loop_end(&loop_state
);
590 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
592 * Fragment Shader outputs pixels in small 2x2 blocks
593 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
595 * However in memory pixels are stored in rows
596 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
598 * @param type fragment shader type (4x or 8x float)
599 * @param num_fs number of fs_src
600 * @param is_1d whether we're outputting to a 1d resource
601 * @param dst_channels number of output channels
602 * @param fs_src output from fragment shader
603 * @param dst pointer to store result
604 * @param pad_inline is channel padding inline or at end of row
605 * @return the number of dsts
608 generate_fs_twiddle(struct gallivm_state
*gallivm
,
611 unsigned dst_channels
,
612 LLVMValueRef fs_src
[][4],
616 LLVMValueRef src
[16];
622 unsigned pixels
= type
.length
/ 4;
623 unsigned reorder_group
;
624 unsigned src_channels
;
628 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
629 src_count
= num_fs
* src_channels
;
631 assert(pixels
== 2 || pixels
== 1);
632 assert(num_fs
* src_channels
<= Elements(src
));
635 * Transpose from SoA -> AoS
637 for (i
= 0; i
< num_fs
; ++i
) {
638 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
642 * Pick transformation options
649 if (dst_channels
== 1) {
655 } else if (dst_channels
== 2) {
659 } else if (dst_channels
> 2) {
666 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
672 * Split the src in half
675 for (i
= num_fs
; i
> 0; --i
) {
676 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
677 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
685 * Ensure pixels are in memory order
688 /* Twiddle pixels by reordering the array, e.g.:
690 * src_count = 8 -> 0 2 1 3 4 6 5 7
691 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
693 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
695 for (i
= 0; i
< src_count
; ++i
) {
696 unsigned group
= i
/ reorder_group
;
697 unsigned block
= (group
/ 4) * 4 * reorder_group
;
698 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
701 } else if (twiddle
) {
702 /* Twiddle pixels across elements of array */
703 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
706 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
710 * Moves any padding between pixels to the end
711 * e.g. RGBXRGBX -> RGBRGBXX
714 unsigned char swizzles
[16];
715 unsigned elems
= pixels
* dst_channels
;
717 for (i
= 0; i
< type
.length
; ++i
) {
719 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
721 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
724 for (i
= 0; i
< src_count
; ++i
) {
725 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
734 * Load an unswizzled block of pixels from memory
737 load_unswizzled_block(struct gallivm_state
*gallivm
,
738 LLVMValueRef base_ptr
,
740 unsigned block_width
,
741 unsigned block_height
,
743 struct lp_type dst_type
,
745 unsigned dst_alignment
)
747 LLVMBuilderRef builder
= gallivm
->builder
;
748 unsigned row_size
= dst_count
/ block_height
;
751 /* Ensure block exactly fits into dst */
752 assert((block_width
* block_height
) % dst_count
== 0);
754 for (i
= 0; i
< dst_count
; ++i
) {
755 unsigned x
= i
% row_size
;
756 unsigned y
= i
/ row_size
;
758 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
759 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
762 LLVMValueRef dst_ptr
;
764 gep
[0] = lp_build_const_int32(gallivm
, 0);
765 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
767 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
768 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
770 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
772 lp_set_load_alignment(dst
[i
], dst_alignment
);
778 * Store an unswizzled block of pixels to memory
781 store_unswizzled_block(struct gallivm_state
*gallivm
,
782 LLVMValueRef base_ptr
,
784 unsigned block_width
,
785 unsigned block_height
,
787 struct lp_type src_type
,
789 unsigned src_alignment
)
791 LLVMBuilderRef builder
= gallivm
->builder
;
792 unsigned row_size
= src_count
/ block_height
;
795 /* Ensure src exactly fits into block */
796 assert((block_width
* block_height
) % src_count
== 0);
798 for (i
= 0; i
< src_count
; ++i
) {
799 unsigned x
= i
% row_size
;
800 unsigned y
= i
/ row_size
;
802 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
803 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
806 LLVMValueRef src_ptr
;
808 gep
[0] = lp_build_const_int32(gallivm
, 0);
809 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
811 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
812 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
814 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
816 lp_set_store_alignment(src_ptr
, src_alignment
);
822 * Checks if a format description is an arithmetic format
824 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
826 static INLINE boolean
827 is_arithmetic_format(const struct util_format_description
*format_desc
)
829 boolean arith
= false;
832 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
833 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
834 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
842 * Checks if this format requires special handling due to required expansion
843 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
846 static INLINE boolean
847 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
849 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
850 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
858 * Retrieves the type representing the memory layout for a format
860 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
863 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
864 struct lp_type
* type
)
869 if (format_expands_to_float_soa(format_desc
)) {
870 /* just make this a 32bit uint */
871 type
->floating
= false;
880 for (i
= 0; i
< 4; i
++)
881 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
885 memset(type
, 0, sizeof(struct lp_type
));
886 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
887 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
888 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
889 type
->norm
= format_desc
->channel
[chan
].normalized
;
891 if (is_arithmetic_format(format_desc
)) {
895 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
896 type
->width
+= format_desc
->channel
[i
].size
;
899 type
->width
= format_desc
->channel
[chan
].size
;
900 type
->length
= format_desc
->nr_channels
;
906 * Retrieves the type for a format which is usable in the blending code.
908 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
911 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
912 struct lp_type
* type
)
917 if (format_expands_to_float_soa(format_desc
)) {
918 /* always use ordinary floats for blending */
919 type
->floating
= true;
928 for (i
= 0; i
< 4; i
++)
929 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
933 memset(type
, 0, sizeof(struct lp_type
));
934 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
935 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
936 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
937 type
->norm
= format_desc
->channel
[chan
].normalized
;
938 type
->width
= format_desc
->channel
[chan
].size
;
939 type
->length
= format_desc
->nr_channels
;
941 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
942 if (format_desc
->channel
[i
].size
> type
->width
)
943 type
->width
= format_desc
->channel
[i
].size
;
946 if (type
->floating
) {
949 if (type
->width
<= 8) {
951 } else if (type
->width
<= 16) {
958 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
965 * Scale a normalized value from src_bits to dst_bits.
967 * The exact calculation is
969 * dst = iround(src * dst_mask / src_mask)
971 * or with integer rounding
973 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
977 * src_mask = (1 << src_bits) - 1
978 * dst_mask = (1 << dst_bits) - 1
980 * but we try to avoid division and multiplication through shifts.
982 static INLINE LLVMValueRef
983 scale_bits(struct gallivm_state
*gallivm
,
987 struct lp_type src_type
)
989 LLVMBuilderRef builder
= gallivm
->builder
;
990 LLVMValueRef result
= src
;
992 if (dst_bits
< src_bits
) {
993 int delta_bits
= src_bits
- dst_bits
;
995 if (delta_bits
<= dst_bits
) {
997 * Approximate the rescaling with a single shift.
999 * This gives the wrong rounding.
1002 result
= LLVMBuildLShr(builder
,
1004 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1009 * Try more accurate rescaling.
1013 * Drop the least significant bits to make space for the multiplication.
1015 * XXX: A better approach would be to use a wider integer type as intermediate. But
1016 * this is enough to convert alpha from 16bits -> 2 when rendering to
1017 * PIPE_FORMAT_R10G10B10A2_UNORM.
1019 result
= LLVMBuildLShr(builder
,
1021 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1025 result
= LLVMBuildMul(builder
,
1027 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1031 * Add a rounding term before the division.
1033 * TODO: Handle signed integers too.
1035 if (!src_type
.sign
) {
1036 result
= LLVMBuildAdd(builder
,
1038 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1043 * Approximate the division by src_mask with a src_bits shift.
1045 * Given the src has already been shifted by dst_bits, all we need
1046 * to do is to shift by the difference.
1049 result
= LLVMBuildLShr(builder
,
1051 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1055 } else if (dst_bits
> src_bits
) {
1057 int db
= dst_bits
- src_bits
;
1059 /* Shift left by difference in bits */
1060 result
= LLVMBuildShl(builder
,
1062 lp_build_const_int_vec(gallivm
, src_type
, db
),
1065 if (db
< src_bits
) {
1066 /* Enough bits in src to fill the remainder */
1067 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1069 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1072 result
= LLVMBuildOr(builder
, result
, lower
, "");
1073 } else if (db
> src_bits
) {
1074 /* Need to repeatedly copy src bits to fill remainder in dst */
1077 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1078 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1080 result
= LLVMBuildOr(builder
,
1082 LLVMBuildLShr(builder
, result
, shuv
, ""),
1093 * Convert from memory format to blending format
1095 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1098 convert_to_blend_type(struct gallivm_state
*gallivm
,
1099 unsigned block_size
,
1100 const struct util_format_description
*src_fmt
,
1101 struct lp_type src_type
,
1102 struct lp_type dst_type
,
1103 LLVMValueRef
* src
, // and dst
1106 LLVMValueRef
*dst
= src
;
1107 LLVMBuilderRef builder
= gallivm
->builder
;
1108 struct lp_type blend_type
;
1109 struct lp_type mem_type
;
1111 unsigned pixels
= block_size
/ num_srcs
;
1115 * full custom path for packed floats and srgb formats - none of the later
1116 * functions would do anything useful, and given the lp_type representation they
1117 * can't be fixed. Should really have some SoA blend path for these kind of
1118 * formats rather than hacking them in here.
1120 if (format_expands_to_float_soa(src_fmt
)) {
1121 LLVMValueRef tmpsrc
[4];
1123 * This is pretty suboptimal for this case blending in SoA would be much
1124 * better, since conversion gets us SoA values so need to convert back.
1126 assert(src_type
.width
== 32);
1127 assert(dst_type
.floating
);
1128 assert(dst_type
.width
== 32);
1129 assert(dst_type
.length
% 4 == 0);
1130 assert(num_srcs
% 4 == 0);
1132 for (i
= 0; i
< 4; i
++) {
1135 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1136 LLVMValueRef tmpsoa
[4];
1137 LLVMValueRef tmps
= tmpsrc
[i
];
1138 if (dst_type
.length
== 8) {
1139 LLVMValueRef shuffles
[8];
1141 /* fetch was 4 values but need 8-wide output values */
1142 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1144 * for 8-wide aos transpose would give us wrong order not matching
1145 * incoming converted fs values and mask. ARGH.
1147 for (j
= 0; j
< 4; j
++) {
1148 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1149 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1151 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1152 LLVMConstVector(shuffles
, 8), "");
1154 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1155 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1158 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1160 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1165 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1166 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1168 /* Is the format arithmetic */
1169 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1170 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1172 /* Pad if necessary */
1173 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1174 for (i
= 0; i
< num_srcs
; ++i
) {
1175 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1178 src_type
.length
= dst_type
.length
;
1181 /* Special case for half-floats */
1182 if (mem_type
.width
== 16 && mem_type
.floating
) {
1183 assert(blend_type
.width
== 32 && blend_type
.floating
);
1184 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1192 src_type
.width
= blend_type
.width
* blend_type
.length
;
1193 blend_type
.length
*= pixels
;
1194 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1196 for (i
= 0; i
< num_srcs
; ++i
) {
1197 LLVMValueRef chans
[4];
1198 LLVMValueRef res
= NULL
;
1200 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1202 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1204 unsigned sa
= src_fmt
->channel
[j
].shift
;
1205 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1206 unsigned from_lsb
= j
;
1208 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1211 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1215 /* Extract bits from source */
1216 chans
[j
] = LLVMBuildLShr(builder
,
1218 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1221 chans
[j
] = LLVMBuildAnd(builder
,
1223 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1227 if (src_type
.norm
) {
1228 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1229 blend_type
.width
, chans
[j
], src_type
);
1232 /* Insert bits into correct position */
1233 chans
[j
] = LLVMBuildShl(builder
,
1235 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1241 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1245 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1251 * Convert from blending format to memory format
1253 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1256 convert_from_blend_type(struct gallivm_state
*gallivm
,
1257 unsigned block_size
,
1258 const struct util_format_description
*src_fmt
,
1259 struct lp_type src_type
,
1260 struct lp_type dst_type
,
1261 LLVMValueRef
* src
, // and dst
1264 LLVMValueRef
* dst
= src
;
1266 struct lp_type mem_type
;
1267 struct lp_type blend_type
;
1268 LLVMBuilderRef builder
= gallivm
->builder
;
1269 unsigned pixels
= block_size
/ num_srcs
;
1273 * full custom path for packed floats and srgb formats - none of the later
1274 * functions would do anything useful, and given the lp_type representation they
1275 * can't be fixed. Should really have some SoA blend path for these kind of
1276 * formats rather than hacking them in here.
1278 if (format_expands_to_float_soa(src_fmt
)) {
1280 * This is pretty suboptimal for this case blending in SoA would be much
1281 * better - we need to transpose the AoS values back to SoA values for
1282 * conversion/packing.
1284 assert(src_type
.floating
);
1285 assert(src_type
.width
== 32);
1286 assert(src_type
.length
% 4 == 0);
1287 assert(dst_type
.width
== 32);
1289 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1290 LLVMValueRef tmpsoa
[4], tmpdst
;
1291 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1292 /* really really need SoA here */
1294 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1295 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1298 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1302 if (src_type
.length
== 8) {
1303 LLVMValueRef tmpaos
, shuffles
[8];
1306 * for 8-wide aos transpose has given us wrong order not matching
1307 * output order. HMPF. Also need to split the output values manually.
1309 for (j
= 0; j
< 4; j
++) {
1310 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1311 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1313 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1314 LLVMConstVector(shuffles
, 8), "");
1315 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1316 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1325 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1326 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1328 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1330 /* Special case for half-floats */
1331 if (mem_type
.width
== 16 && mem_type
.floating
) {
1332 int length
= dst_type
.length
;
1333 assert(blend_type
.width
== 32 && blend_type
.floating
);
1335 dst_type
.length
= src_type
.length
;
1337 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1339 dst_type
.length
= length
;
1343 /* Remove any padding */
1344 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1345 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1347 for (i
= 0; i
< num_srcs
; ++i
) {
1348 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1352 /* No bit arithmetic to do */
1357 src_type
.length
= pixels
;
1358 src_type
.width
= blend_type
.length
* blend_type
.width
;
1359 dst_type
.length
= pixels
;
1361 for (i
= 0; i
< num_srcs
; ++i
) {
1362 LLVMValueRef chans
[4];
1363 LLVMValueRef res
= NULL
;
1365 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1367 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1369 unsigned sa
= src_fmt
->channel
[j
].shift
;
1370 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1371 unsigned from_lsb
= j
;
1373 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1376 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1378 for (k
= 0; k
< blend_type
.width
; ++k
) {
1383 chans
[j
] = LLVMBuildLShr(builder
,
1385 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1388 chans
[j
] = LLVMBuildAnd(builder
,
1390 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1393 /* Scale down bits */
1394 if (src_type
.norm
) {
1395 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1396 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1400 chans
[j
] = LLVMBuildShl(builder
,
1402 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1405 sa
+= src_fmt
->channel
[j
].size
;
1410 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1414 assert (dst_type
.width
!= 24);
1416 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1422 * Convert alpha to same blend type as src
1425 convert_alpha(struct gallivm_state
*gallivm
,
1426 struct lp_type row_type
,
1427 struct lp_type alpha_type
,
1428 const unsigned block_size
,
1429 const unsigned block_height
,
1430 const unsigned src_count
,
1431 const unsigned dst_channels
,
1432 const bool pad_inline
,
1433 LLVMValueRef
* src_alpha
)
1435 LLVMBuilderRef builder
= gallivm
->builder
;
1437 unsigned length
= row_type
.length
;
1438 row_type
.length
= alpha_type
.length
;
1440 /* Twiddle the alpha to match pixels */
1441 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1444 * TODO this should use single lp_build_conv call for
1445 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1447 for (i
= 0; i
< block_height
; ++i
) {
1448 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1451 alpha_type
= row_type
;
1452 row_type
.length
= length
;
1454 /* If only one channel we can only need the single alpha value per pixel */
1455 if (src_count
== 1 && dst_channels
== 1) {
1457 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1459 /* If there are more srcs than rows then we need to split alpha up */
1460 if (src_count
> block_height
) {
1461 for (i
= src_count
; i
> 0; --i
) {
1462 unsigned pixels
= block_size
/ src_count
;
1463 unsigned idx
= i
- 1;
1465 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1466 (idx
* pixels
) % 4, pixels
);
1470 /* If there is a src for each pixel broadcast the alpha across whole row */
1471 if (src_count
== block_size
) {
1472 for (i
= 0; i
< src_count
; ++i
) {
1473 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1476 unsigned pixels
= block_size
/ src_count
;
1477 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1478 unsigned alpha_span
= 1;
1479 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1481 /* Check if we need 2 src_alphas for our shuffles */
1482 if (pixels
> alpha_type
.length
) {
1486 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1487 for (j
= 0; j
< row_type
.length
; ++j
) {
1488 if (j
< pixels
* channels
) {
1489 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1491 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1495 for (i
= 0; i
< src_count
; ++i
) {
1496 unsigned idx1
= i
, idx2
= i
;
1498 if (alpha_span
> 1){
1503 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1506 LLVMConstVector(shuffles
, row_type
.length
),
1515 * Generates the blend function for unswizzled colour buffers
1516 * Also generates the read & write from colour buffer
1519 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1521 struct lp_fragment_shader_variant
*variant
,
1522 enum pipe_format out_format
,
1523 unsigned int num_fs
,
1524 struct lp_type fs_type
,
1525 LLVMValueRef
* fs_mask
,
1526 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1527 LLVMValueRef context_ptr
,
1528 LLVMValueRef color_ptr
,
1529 LLVMValueRef stride
,
1530 unsigned partial_mask
,
1533 const unsigned alpha_channel
= 3;
1534 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1535 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1536 const unsigned block_size
= block_width
* block_height
;
1537 const unsigned lp_integer_vector_width
= 128;
1539 LLVMBuilderRef builder
= gallivm
->builder
;
1540 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1541 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1542 LLVMValueRef src_alpha
[4 * 4];
1543 LLVMValueRef src1_alpha
[4 * 4];
1544 LLVMValueRef src_mask
[4 * 4];
1545 LLVMValueRef src
[4 * 4];
1546 LLVMValueRef src1
[4 * 4];
1547 LLVMValueRef dst
[4 * 4];
1548 LLVMValueRef blend_color
;
1549 LLVMValueRef blend_alpha
;
1550 LLVMValueRef i32_zero
;
1551 LLVMValueRef check_mask
;
1552 LLVMValueRef undef_src_val
;
1554 struct lp_build_mask_context mask_ctx
;
1555 struct lp_type mask_type
;
1556 struct lp_type blend_type
;
1557 struct lp_type row_type
;
1558 struct lp_type dst_type
;
1560 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1561 unsigned vector_width
;
1562 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1563 unsigned dst_channels
;
1568 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1570 unsigned dst_alignment
;
1572 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1573 bool has_alpha
= false;
1574 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1575 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1577 const boolean is_1d
= variant
->key
.resource_1d
;
1578 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1579 LLVMValueRef fpstate
= 0;
1581 /* Get type from output format */
1582 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1583 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1586 * Technically this code should go into lp_build_smallfloat_to_float
1587 * and lp_build_float_to_smallfloat but due to the
1588 * http://llvm.org/bugs/show_bug.cgi?id=6393
1589 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1590 * So the ordering is important here and there shouldn't be any
1591 * llvm ir instrunctions in this function before
1592 * this, otherwise half-float format conversions won't work
1593 * (again due to llvm bug #6393).
1595 if (dst_type
.floating
&& dst_type
.width
!= 32) {
1596 /* We need to make sure that denorms are ok for half float
1598 fpstate
= lp_build_fpstate_get(gallivm
);
1599 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1602 mask_type
= lp_int32_vec4_type();
1603 mask_type
.length
= fs_type
.length
;
1605 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1606 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1609 /* Do not bother executing code when mask is empty.. */
1611 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1613 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1614 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1617 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1618 lp_build_mask_check(&mask_ctx
);
1621 partial_mask
|= !variant
->opaque
;
1622 i32_zero
= lp_build_const_int32(gallivm
, 0);
1624 #if HAVE_LLVM < 0x0302
1626 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1627 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1629 undef_src_val
= lp_build_zero(gallivm
, fs_type
);
1631 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1634 row_type
.length
= fs_type
.length
;
1635 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1637 /* Compute correct swizzle and count channels */
1638 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1641 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1642 /* Ensure channel is used */
1643 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1647 /* Ensure not already written to (happens in case with GL_ALPHA) */
1648 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1652 /* Ensure we havn't already found all channels */
1653 if (dst_channels
>= out_format_desc
->nr_channels
) {
1657 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1660 if (i
== alpha_channel
) {
1665 if (format_expands_to_float_soa(out_format_desc
)) {
1667 * the code above can't work for layout_other
1668 * for srgb it would sort of work but we short-circuit swizzles, etc.
1669 * as that is done as part of unpack / pack.
1671 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1677 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1680 /* If 3 channels then pad to include alpha for 4 element transpose */
1681 if (dst_channels
== 3 && !has_alpha
) {
1682 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1683 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1686 if (out_format_desc
->nr_channels
== 4) {
1692 * Load shader output
1694 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1695 /* Always load alpha for use in blending */
1698 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1701 alpha
= undef_src_val
;
1704 /* Load each channel */
1705 for (j
= 0; j
< dst_channels
; ++j
) {
1706 assert(swizzle
[j
] < 4);
1708 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1711 fs_src
[i
][j
] = undef_src_val
;
1715 /* If 3 channels then pad to include alpha for 4 element transpose */
1717 * XXX If we include that here maybe could actually use it instead of
1718 * separate alpha for blending?
1720 if (dst_channels
== 3 && !has_alpha
) {
1721 fs_src
[i
][3] = alpha
;
1724 /* We split the row_mask and row_alpha as we want 128bit interleave */
1725 if (fs_type
.length
== 8) {
1726 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1727 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1729 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1730 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1732 src_mask
[i
] = fs_mask
[i
];
1733 src_alpha
[i
] = alpha
;
1736 if (dual_source_blend
) {
1737 /* same as above except different src/dst, skip masks and comments... */
1738 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1741 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1744 alpha
= undef_src_val
;
1747 for (j
= 0; j
< dst_channels
; ++j
) {
1748 assert(swizzle
[j
] < 4);
1750 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1753 fs_src1
[i
][j
] = undef_src_val
;
1756 if (dst_channels
== 3 && !has_alpha
) {
1757 fs_src1
[i
][3] = alpha
;
1759 if (fs_type
.length
== 8) {
1760 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1761 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1763 src1_alpha
[i
] = alpha
;
1768 if (util_format_is_pure_integer(out_format
)) {
1770 * In this case fs_type was really ints or uints disguised as floats,
1773 fs_type
.floating
= 0;
1774 fs_type
.sign
= dst_type
.sign
;
1775 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1776 for (j
= 0; j
< dst_channels
; ++j
) {
1777 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1778 lp_build_vec_type(gallivm
, fs_type
), "");
1780 if (dst_channels
== 3 && !has_alpha
) {
1781 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1782 lp_build_vec_type(gallivm
, fs_type
), "");
1788 * Pixel twiddle from fragment shader order to memory order
1790 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1791 dst_channels
, fs_src
, src
, pad_inline
);
1792 if (dual_source_blend
) {
1793 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1794 fs_src1
, src1
, pad_inline
);
1797 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1798 if (src_count
!= num_fullblock_fs
* src_channels
) {
1799 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1800 row_type
.length
/= ds
;
1801 fs_type
.length
= row_type
.length
;
1804 blend_type
= row_type
;
1805 mask_type
.length
= 4;
1807 /* Convert src to row_type */
1808 if (dual_source_blend
) {
1809 struct lp_type old_row_type
= row_type
;
1810 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1811 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1814 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1817 /* If the rows are not an SSE vector, combine them to become SSE size! */
1818 if ((row_type
.width
* row_type
.length
) % 128) {
1819 unsigned bits
= row_type
.width
* row_type
.length
;
1822 assert(src_count
>= (vector_width
/ bits
));
1824 dst_count
= src_count
/ (vector_width
/ bits
);
1826 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1827 if (dual_source_blend
) {
1828 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1831 row_type
.length
*= combined
;
1832 src_count
/= combined
;
1834 bits
= row_type
.width
* row_type
.length
;
1835 assert(bits
== 128 || bits
== 256);
1840 * Blend Colour conversion
1842 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1843 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1844 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1847 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1849 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1851 * since blending is done with floats, there was no conversion.
1852 * However, the rules according to fixed point renderbuffers still
1853 * apply, that is we must clamp inputs to 0.0/1.0.
1854 * (This would apply to separate alpha conversion too but we currently
1855 * force has_alpha to be true.)
1856 * TODO: should skip this with "fake" blend, since post-blend conversion
1857 * will clamp anyway.
1858 * TODO: could also skip this if fragment color clamping is enabled. We
1859 * don't support it natively so it gets baked into the shader however, so
1860 * can't really tell here.
1862 struct lp_build_context f32_bld
;
1863 assert(row_type
.floating
);
1864 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1865 for (i
= 0; i
< src_count
; i
++) {
1866 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
1868 if (dual_source_blend
) {
1869 for (i
= 0; i
< src_count
; i
++) {
1870 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
1873 /* probably can't be different than row_type but better safe than sorry... */
1874 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1875 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1879 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1881 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1882 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1884 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1885 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1887 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1888 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1894 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1896 if (src_count
< block_height
) {
1897 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1898 } else if (src_count
> block_height
) {
1899 for (i
= src_count
; i
> 0; --i
) {
1900 unsigned pixels
= block_size
/ src_count
;
1901 unsigned idx
= i
- 1;
1903 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1904 (idx
* pixels
) % 4, pixels
);
1908 assert(mask_type
.width
== 32);
1910 for (i
= 0; i
< src_count
; ++i
) {
1911 unsigned pixels
= block_size
/ src_count
;
1912 unsigned pixel_width
= row_type
.width
* dst_channels
;
1914 if (pixel_width
== 24) {
1915 mask_type
.width
= 8;
1916 mask_type
.length
= vector_width
/ mask_type
.width
;
1918 mask_type
.length
= pixels
;
1919 mask_type
.width
= row_type
.width
* dst_channels
;
1921 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1923 mask_type
.length
*= dst_channels
;
1924 mask_type
.width
/= dst_channels
;
1927 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1928 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1935 struct lp_type alpha_type
= fs_type
;
1936 alpha_type
.length
= 4;
1937 convert_alpha(gallivm
, row_type
, alpha_type
,
1938 block_size
, block_height
,
1939 src_count
, dst_channels
,
1940 pad_inline
, src_alpha
);
1941 if (dual_source_blend
) {
1942 convert_alpha(gallivm
, row_type
, alpha_type
,
1943 block_size
, block_height
,
1944 src_count
, dst_channels
,
1945 pad_inline
, src1_alpha
);
1951 * Load dst from memory
1953 if (src_count
< block_height
) {
1954 dst_count
= block_height
;
1956 dst_count
= src_count
;
1959 dst_type
.length
*= block_size
/ dst_count
;
1961 if (format_expands_to_float_soa(out_format_desc
)) {
1963 * we need multiple values at once for the conversion, so can as well
1964 * load them vectorized here too instead of concatenating later.
1965 * (Still need concatenation later for 8-wide vectors).
1967 dst_count
= block_height
;
1968 dst_type
.length
= block_width
;
1972 * Compute the alignment of the destination pointer in bytes
1973 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1974 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1975 * 1d tex but can't distinguish here) so need to stick with per-pixel
1976 * alignment in this case.
1979 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1982 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
1984 /* Force power-of-two alignment by extracting only the least-significant-bit */
1985 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
1987 * Resource base and stride pointers are aligned to 16 bytes, so that's
1988 * the maximum alignment we can guarantee
1990 dst_alignment
= MIN2(16, dst_alignment
);
1993 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
1994 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
1995 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
1996 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
2001 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2002 dst
, dst_type
, dst_count
, dst_alignment
);
2007 * Convert from dst/output format to src/blending format.
2009 * This is necessary as we can only read 1 row from memory at a time,
2010 * so the minimum dst_count will ever be at this point is 4.
2012 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2013 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2014 * on all 16 pixels in that single vector at once.
2016 if (dst_count
> src_count
) {
2017 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
2023 /* XXX this is broken for RGB8 formats -
2024 * they get expanded from 12 to 16 elements (to include alpha)
2025 * by convert_to_blend_type then reduced to 15 instead of 12
2026 * by convert_from_blend_type (a simple fix though breaks A8...).
2027 * R16G16B16 also crashes differently however something going wrong
2028 * inside llvm handling npot vector sizes seemingly.
2029 * It seems some cleanup could be done here (like skipping conversion/blend
2032 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
2035 * FIXME: Really should get logic ops / masks out of generic blend / row
2036 * format. Logic ops will definitely not work on the blend float format
2037 * used for SRGB here and I think OpenGL expects this to work as expected
2038 * (that is incoming values converted to srgb then logic op applied).
2040 for (i
= 0; i
< src_count
; ++i
) {
2041 dst
[i
] = lp_build_blend_aos(gallivm
,
2042 &variant
->key
.blend
,
2047 has_alpha
? NULL
: src_alpha
[i
],
2049 has_alpha
? NULL
: src1_alpha
[i
],
2051 partial_mask
? src_mask
[i
] : NULL
,
2053 has_alpha
? NULL
: blend_alpha
,
2055 pad_inline
? 4 : dst_channels
);
2058 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
2060 /* Split the blend rows back to memory rows */
2061 if (dst_count
> src_count
) {
2062 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2064 if (src_count
== 1) {
2065 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2066 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2068 row_type
.length
/= 2;
2072 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2073 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2074 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2075 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2077 row_type
.length
/= 2;
2082 * Store blend result to memory
2085 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2086 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2089 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2090 dst
, dst_type
, dst_count
, dst_alignment
);
2093 if (dst_type
.floating
&& dst_type
.width
!= 32) {
2094 lp_build_fpstate_set(gallivm
, fpstate
);
2098 lp_build_mask_end(&mask_ctx
);
2104 * Generate the runtime callable function for the whole fragment pipeline.
2105 * Note that the function which we generate operates on a block of 16
2106 * pixels at at time. The block contains 2x2 quads. Each quad contains
2110 generate_fragment(struct llvmpipe_context
*lp
,
2111 struct lp_fragment_shader
*shader
,
2112 struct lp_fragment_shader_variant
*variant
,
2113 unsigned partial_mask
)
2115 struct gallivm_state
*gallivm
= variant
->gallivm
;
2116 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2117 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2118 char func_name
[256];
2119 struct lp_type fs_type
;
2120 struct lp_type blend_type
;
2121 LLVMTypeRef fs_elem_type
;
2122 LLVMTypeRef blend_vec_type
;
2123 LLVMTypeRef arg_types
[13];
2124 LLVMTypeRef func_type
;
2125 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2126 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2127 LLVMValueRef context_ptr
;
2130 LLVMValueRef a0_ptr
;
2131 LLVMValueRef dadx_ptr
;
2132 LLVMValueRef dady_ptr
;
2133 LLVMValueRef color_ptr_ptr
;
2134 LLVMValueRef stride_ptr
;
2135 LLVMValueRef depth_ptr
;
2136 LLVMValueRef depth_stride
;
2137 LLVMValueRef mask_input
;
2138 LLVMValueRef thread_data_ptr
;
2139 LLVMBasicBlockRef block
;
2140 LLVMBuilderRef builder
;
2141 struct lp_build_sampler_soa
*sampler
;
2142 struct lp_build_interp_soa_context interp
;
2143 LLVMValueRef fs_mask
[16 / 4];
2144 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2145 LLVMValueRef function
;
2146 LLVMValueRef facing
;
2151 boolean cbuf0_write_all
;
2152 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2153 util_blend_state_is_dual(&key
->blend
, 0);
2155 assert(lp_native_vector_width
/ 32 >= 4);
2157 /* Adjust color input interpolation according to flatshade state:
2159 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2160 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2161 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2163 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2165 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2169 /* check if writes to cbuf[0] are to be copied to all cbufs */
2170 cbuf0_write_all
= FALSE
;
2171 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
2172 if (shader
->info
.base
.properties
[i
].name
==
2173 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
2174 cbuf0_write_all
= TRUE
;
2179 /* TODO: actually pick these based on the fs and color buffer
2180 * characteristics. */
2182 memset(&fs_type
, 0, sizeof fs_type
);
2183 fs_type
.floating
= TRUE
; /* floating point values */
2184 fs_type
.sign
= TRUE
; /* values are signed */
2185 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2186 fs_type
.width
= 32; /* 32-bit float */
2187 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2189 memset(&blend_type
, 0, sizeof blend_type
);
2190 blend_type
.floating
= FALSE
; /* values are integers */
2191 blend_type
.sign
= FALSE
; /* values are unsigned */
2192 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2193 blend_type
.width
= 8; /* 8-bit ubyte values */
2194 blend_type
.length
= 16; /* 16 elements per vector */
2197 * Generate the function prototype. Any change here must be reflected in
2198 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2201 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2203 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2205 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2206 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2208 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2209 arg_types
[1] = int32_type
; /* x */
2210 arg_types
[2] = int32_type
; /* y */
2211 arg_types
[3] = int32_type
; /* facing */
2212 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2213 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2214 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2215 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2216 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2217 arg_types
[9] = int32_type
; /* mask_input */
2218 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2219 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2220 arg_types
[12] = int32_type
; /* depth_stride */
2222 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2223 arg_types
, Elements(arg_types
), 0);
2225 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2226 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2228 variant
->function
[partial_mask
] = function
;
2230 /* XXX: need to propagate noalias down into color param now we are
2231 * passing a pointer-to-pointer?
2233 for(i
= 0; i
< Elements(arg_types
); ++i
)
2234 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2235 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2237 context_ptr
= LLVMGetParam(function
, 0);
2238 x
= LLVMGetParam(function
, 1);
2239 y
= LLVMGetParam(function
, 2);
2240 facing
= LLVMGetParam(function
, 3);
2241 a0_ptr
= LLVMGetParam(function
, 4);
2242 dadx_ptr
= LLVMGetParam(function
, 5);
2243 dady_ptr
= LLVMGetParam(function
, 6);
2244 color_ptr_ptr
= LLVMGetParam(function
, 7);
2245 depth_ptr
= LLVMGetParam(function
, 8);
2246 mask_input
= LLVMGetParam(function
, 9);
2247 thread_data_ptr
= LLVMGetParam(function
, 10);
2248 stride_ptr
= LLVMGetParam(function
, 11);
2249 depth_stride
= LLVMGetParam(function
, 12);
2251 lp_build_name(context_ptr
, "context");
2252 lp_build_name(x
, "x");
2253 lp_build_name(y
, "y");
2254 lp_build_name(a0_ptr
, "a0");
2255 lp_build_name(dadx_ptr
, "dadx");
2256 lp_build_name(dady_ptr
, "dady");
2257 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2258 lp_build_name(depth_ptr
, "depth");
2259 lp_build_name(thread_data_ptr
, "thread_data");
2260 lp_build_name(mask_input
, "mask_input");
2261 lp_build_name(stride_ptr
, "stride_ptr");
2262 lp_build_name(depth_stride
, "depth_stride");
2268 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2269 builder
= gallivm
->builder
;
2271 LLVMPositionBuilderAtEnd(builder
, block
);
2273 /* code generated texture sampling */
2274 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
2276 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2277 /* for 1d resources only run "upper half" of stamp */
2278 if (key
->resource_1d
)
2282 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2283 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2284 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2285 num_loop
, "mask_store");
2286 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2289 * The shader input interpolation info is not explicitely baked in the
2290 * shader key, but everything it derives from (TGSI, and flatshade) is
2291 * already included in the shader key.
2293 lp_build_interp_soa_init(&interp
,
2295 shader
->info
.base
.num_inputs
,
2297 shader
->info
.base
.pixel_center_integer
,
2299 a0_ptr
, dadx_ptr
, dady_ptr
,
2302 for (i
= 0; i
< num_fs
; i
++) {
2304 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2305 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2306 &indexi
, 1, "mask_ptr");
2309 mask
= generate_quad_mask(gallivm
, fs_type
,
2310 i
*fs_type
.length
/4, mask_input
);
2313 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2315 LLVMBuildStore(builder
, mask
, mask_ptr
);
2318 generate_fs_loop(gallivm
,
2326 mask_store
, /* output */
2333 for (i
= 0; i
< num_fs
; i
++) {
2334 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2335 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2337 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2338 /* This is fucked up need to reorganize things */
2339 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2340 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2341 ptr
= LLVMBuildGEP(builder
,
2342 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2344 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2347 if (dual_source_blend
) {
2348 /* only support one dual source blend target hence always use output 1 */
2349 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2350 ptr
= LLVMBuildGEP(builder
,
2351 color_store
[1][chan
],
2353 fs_out_color
[1][chan
][i
] = ptr
;
2359 sampler
->destroy(sampler
);
2361 /* Loop over color outputs / color buffers to do blending.
2363 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2364 LLVMValueRef color_ptr
;
2365 LLVMValueRef stride
;
2366 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2368 boolean do_branch
= ((key
->depth
.enabled
2369 || key
->stencil
[0].enabled
2370 || key
->alpha
.enabled
)
2371 && !shader
->info
.base
.uses_kill
);
2373 color_ptr
= LLVMBuildLoad(builder
,
2374 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2377 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2379 stride
= LLVMBuildLoad(builder
,
2380 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2383 generate_unswizzled_blend(gallivm
, cbuf
, variant
, key
->cbuf_format
[cbuf
],
2384 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2385 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2388 LLVMBuildRetVoid(builder
);
2390 gallivm_verify_function(gallivm
, function
);
2392 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2397 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2401 debug_printf("fs variant %p:\n", (void *) key
);
2403 if (key
->flatshade
) {
2404 debug_printf("flatshade = 1\n");
2406 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2407 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2409 if (key
->depth
.enabled
) {
2410 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2411 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2412 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2415 for (i
= 0; i
< 2; ++i
) {
2416 if (key
->stencil
[i
].enabled
) {
2417 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2418 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2419 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2420 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2421 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2422 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2426 if (key
->alpha
.enabled
) {
2427 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2430 if (key
->occlusion_count
) {
2431 debug_printf("occlusion_count = 1\n");
2434 if (key
->blend
.logicop_enable
) {
2435 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2437 else if (key
->blend
.rt
[0].blend_enable
) {
2438 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2439 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2440 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2441 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2442 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2443 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2445 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2446 if (key
->blend
.alpha_to_coverage
) {
2447 debug_printf("blend.alpha_to_coverage is enabled\n");
2449 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2450 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2451 debug_printf("sampler[%u] = \n", i
);
2452 debug_printf(" .wrap = %s %s %s\n",
2453 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2454 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2455 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2456 debug_printf(" .min_img_filter = %s\n",
2457 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2458 debug_printf(" .min_mip_filter = %s\n",
2459 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2460 debug_printf(" .mag_img_filter = %s\n",
2461 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2462 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2463 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2464 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2465 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2466 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2467 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2468 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2470 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2471 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2472 debug_printf("texture[%u] = \n", i
);
2473 debug_printf(" .format = %s\n",
2474 util_format_name(texture
->format
));
2475 debug_printf(" .target = %s\n",
2476 util_dump_tex_target(texture
->target
, TRUE
));
2477 debug_printf(" .level_zero_only = %u\n",
2478 texture
->level_zero_only
);
2479 debug_printf(" .pot = %u %u %u\n",
2481 texture
->pot_height
,
2482 texture
->pot_depth
);
2488 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2490 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2491 variant
->shader
->no
, variant
->no
);
2492 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2493 dump_fs_variant_key(&variant
->key
);
2494 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2500 * Generate a new fragment shader variant from the shader code and
2501 * other state indicated by the key.
2503 static struct lp_fragment_shader_variant
*
2504 generate_variant(struct llvmpipe_context
*lp
,
2505 struct lp_fragment_shader
*shader
,
2506 const struct lp_fragment_shader_variant_key
*key
)
2508 struct lp_fragment_shader_variant
*variant
;
2509 const struct util_format_description
*cbuf0_format_desc
;
2510 boolean fullcolormask
;
2512 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2516 variant
->gallivm
= gallivm_create();
2517 if (!variant
->gallivm
) {
2522 variant
->shader
= shader
;
2523 variant
->list_item_global
.base
= variant
;
2524 variant
->list_item_local
.base
= variant
;
2525 variant
->no
= shader
->variants_created
++;
2527 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2530 * Determine whether we are touching all channels in the color buffer.
2532 fullcolormask
= FALSE
;
2533 if (key
->nr_cbufs
== 1) {
2534 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2535 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2539 !key
->blend
.logicop_enable
&&
2540 !key
->blend
.rt
[0].blend_enable
&&
2542 !key
->stencil
[0].enabled
&&
2543 !key
->alpha
.enabled
&&
2544 !key
->blend
.alpha_to_coverage
&&
2545 !key
->depth
.enabled
&&
2546 !shader
->info
.base
.uses_kill
2549 if ((shader
->info
.base
.num_tokens
<= 1) &&
2550 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2551 variant
->ps_inv_multiplier
= 0;
2553 variant
->ps_inv_multiplier
= 1;
2556 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2557 lp_debug_fs_variant(variant
);
2560 lp_jit_init_types(variant
);
2562 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2563 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2565 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2566 if (variant
->opaque
) {
2567 /* Specialized shader, which doesn't need to read the color buffer. */
2568 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2573 * Compile everything
2576 gallivm_compile_module(variant
->gallivm
);
2578 if (variant
->function
[RAST_EDGE_TEST
]) {
2579 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2580 gallivm_jit_function(variant
->gallivm
,
2581 variant
->function
[RAST_EDGE_TEST
]);
2584 if (variant
->function
[RAST_WHOLE
]) {
2585 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2586 gallivm_jit_function(variant
->gallivm
,
2587 variant
->function
[RAST_WHOLE
]);
2588 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2589 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2597 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2598 const struct pipe_shader_state
*templ
)
2600 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2601 struct lp_fragment_shader
*shader
;
2603 int nr_sampler_views
;
2606 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2610 shader
->no
= fs_no
++;
2611 make_empty_list(&shader
->variants
);
2613 /* get/save the summary info for this shader */
2614 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2616 /* we need to keep a local copy of the tokens */
2617 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2619 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2620 if (shader
->draw_data
== NULL
) {
2621 FREE((void *) shader
->base
.tokens
);
2626 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2627 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2629 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2630 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2632 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2633 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2634 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2636 switch (shader
->info
.base
.input_interpolate
[i
]) {
2637 case TGSI_INTERPOLATE_CONSTANT
:
2638 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2640 case TGSI_INTERPOLATE_LINEAR
:
2641 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2643 case TGSI_INTERPOLATE_PERSPECTIVE
:
2644 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2646 case TGSI_INTERPOLATE_COLOR
:
2647 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2654 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2655 case TGSI_SEMANTIC_FACE
:
2656 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2658 case TGSI_SEMANTIC_POSITION
:
2659 /* Position was already emitted above
2661 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2662 shader
->inputs
[i
].src_index
= 0;
2666 shader
->inputs
[i
].src_index
= i
+1;
2669 if (LP_DEBUG
& DEBUG_TGSI
) {
2671 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2672 shader
->no
, (void *) shader
);
2673 tgsi_dump(templ
->tokens
, 0);
2674 debug_printf("usage masks:\n");
2675 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2676 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2677 debug_printf(" IN[%u].%s%s%s%s\n",
2679 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2680 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2681 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2682 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2692 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2694 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2696 if (llvmpipe
->fs
== fs
)
2699 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2701 draw_bind_fragment_shader(llvmpipe
->draw
,
2702 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2704 llvmpipe
->dirty
|= LP_NEW_FS
;
2709 * Remove shader variant from two lists: the shader's variant list
2710 * and the context's variant list.
2713 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2714 struct lp_fragment_shader_variant
*variant
)
2718 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2719 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2720 " #%u v total cached #%u\n",
2721 variant
->shader
->no
,
2723 variant
->shader
->variants_created
,
2724 variant
->shader
->variants_cached
,
2725 lp
->nr_fs_variants
);
2728 /* free all the variant's JIT'd functions */
2729 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2730 if (variant
->function
[i
]) {
2731 gallivm_free_function(variant
->gallivm
,
2732 variant
->function
[i
],
2733 variant
->jit_function
[i
]);
2737 gallivm_destroy(variant
->gallivm
);
2739 /* remove from shader's list */
2740 remove_from_list(&variant
->list_item_local
);
2741 variant
->shader
->variants_cached
--;
2743 /* remove from context's list */
2744 remove_from_list(&variant
->list_item_global
);
2745 lp
->nr_fs_variants
--;
2746 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2753 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2755 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2756 struct lp_fragment_shader
*shader
= fs
;
2757 struct lp_fs_variant_list_item
*li
;
2759 assert(fs
!= llvmpipe
->fs
);
2762 * XXX: we need to flush the context until we have some sort of reference
2763 * counting in fragment shaders as they may still be binned
2764 * Flushing alone might not sufficient we need to wait on it too.
2766 llvmpipe_finish(pipe
, __FUNCTION__
);
2768 /* Delete all the variants */
2769 li
= first_elem(&shader
->variants
);
2770 while(!at_end(&shader
->variants
, li
)) {
2771 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2772 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2776 /* Delete draw module's data */
2777 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2779 assert(shader
->variants_cached
== 0);
2780 FREE((void *) shader
->base
.tokens
);
2787 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2788 uint shader
, uint index
,
2789 struct pipe_constant_buffer
*cb
)
2791 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2792 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2794 assert(shader
< PIPE_SHADER_TYPES
);
2795 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2797 /* note: reference counting */
2798 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2800 if (shader
== PIPE_SHADER_VERTEX
||
2801 shader
== PIPE_SHADER_GEOMETRY
) {
2802 /* Pass the constants to the 'draw' module */
2803 const unsigned size
= cb
? cb
->buffer_size
: 0;
2807 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2809 else if (cb
&& cb
->user_buffer
) {
2810 data
= (ubyte
*) cb
->user_buffer
;
2817 data
+= cb
->buffer_offset
;
2819 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2823 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2825 if (cb
&& cb
->user_buffer
) {
2826 pipe_resource_reference(&constants
, NULL
);
2832 * Return the blend factor equivalent to a destination alpha of one.
2834 static INLINE
unsigned
2835 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2838 case PIPE_BLENDFACTOR_DST_ALPHA
:
2839 return PIPE_BLENDFACTOR_ONE
;
2840 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2841 return PIPE_BLENDFACTOR_ZERO
;
2842 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2844 return PIPE_BLENDFACTOR_ZERO
;
2846 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2854 * We need to generate several variants of the fragment pipeline to match
2855 * all the combinations of the contributing state atoms.
2857 * TODO: there is actually no reason to tie this to context state -- the
2858 * generated code could be cached globally in the screen.
2861 make_variant_key(struct llvmpipe_context
*lp
,
2862 struct lp_fragment_shader
*shader
,
2863 struct lp_fragment_shader_variant_key
*key
)
2867 memset(key
, 0, shader
->variant_key_size
);
2869 if (lp
->framebuffer
.zsbuf
) {
2870 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2871 const struct util_format_description
*zsbuf_desc
=
2872 util_format_description(zsbuf_format
);
2874 if (lp
->depth_stencil
->depth
.enabled
&&
2875 util_format_has_depth(zsbuf_desc
)) {
2876 key
->zsbuf_format
= zsbuf_format
;
2877 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2879 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2880 util_format_has_stencil(zsbuf_desc
)) {
2881 key
->zsbuf_format
= zsbuf_format
;
2882 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2884 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2885 key
->resource_1d
= TRUE
;
2890 * Propagate the depth clamp setting from the rasterizer state.
2891 * depth_clip == 0 implies depth clamping is enabled.
2893 * When clip_halfz is enabled, then always clamp the depth values.
2895 if (lp
->rasterizer
->clip_halfz
) {
2896 key
->depth_clamp
= 1;
2898 key
->depth_clamp
= (lp
->rasterizer
->depth_clip
== 0) ? 1 : 0;
2901 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2902 if (!lp
->framebuffer
.nr_cbufs
||
2903 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2904 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2906 if(key
->alpha
.enabled
)
2907 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2908 /* alpha.ref_value is passed in jit_context */
2910 key
->flatshade
= lp
->rasterizer
->flatshade
;
2911 if (lp
->active_occlusion_queries
) {
2912 key
->occlusion_count
= TRUE
;
2915 if (lp
->framebuffer
.nr_cbufs
) {
2916 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2919 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2921 if (!key
->blend
.independent_blend_enable
) {
2922 /* we always need independent blend otherwise the fixups below won't work */
2923 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2924 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2926 key
->blend
.independent_blend_enable
= 1;
2929 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2930 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2931 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2932 const struct util_format_description
*format_desc
;
2934 key
->cbuf_format
[i
] = format
;
2937 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2938 * mixing of 2d textures with height 1 and 1d textures, so make sure
2939 * we pick 1d if any cbuf or zsbuf is 1d.
2941 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[0]->texture
)) {
2942 key
->resource_1d
= TRUE
;
2945 format_desc
= util_format_description(format
);
2946 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2947 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2950 * Mask out color channels not present in the color buffer.
2952 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2955 * Disable blend for integer formats.
2957 if (util_format_is_pure_integer(format
)) {
2958 blend_rt
->blend_enable
= 0;
2962 * Our swizzled render tiles always have an alpha channel, but the linear
2963 * render target format often does not, so force here the dst alpha to be
2966 * This is not a mere optimization. Wrong results will be produced if the
2967 * dst alpha is used, the dst format does not have alpha, and the previous
2968 * rendering was not flushed from the swizzled to linear buffer. For
2969 * example, NonPowTwo DCT.
2971 * TODO: This should be generalized to all channels for better
2972 * performance, but only alpha causes correctness issues.
2974 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2976 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2977 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2978 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
2979 boolean clamped_zero
= !util_format_is_float(format
) &&
2980 !util_format_is_snorm(format
);
2981 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
,
2983 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
,
2985 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2986 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2987 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2991 /* This value will be the same for all the variants of a given shader:
2993 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2995 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2996 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2997 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
2998 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3003 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3004 * are dx10-style? Can't really have mixed opcodes, at least not
3005 * if we want to skip the holes here (without rescanning tgsi).
3007 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3008 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3009 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3010 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
3011 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3012 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3017 key
->nr_sampler_views
= key
->nr_samplers
;
3018 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3019 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3020 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3021 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3030 * Update fragment shader state. This is called just prior to drawing
3031 * something when some fragment-related state has changed.
3034 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3036 struct lp_fragment_shader
*shader
= lp
->fs
;
3037 struct lp_fragment_shader_variant_key key
;
3038 struct lp_fragment_shader_variant
*variant
= NULL
;
3039 struct lp_fs_variant_list_item
*li
;
3041 make_variant_key(lp
, shader
, &key
);
3043 /* Search the variants for one which matches the key */
3044 li
= first_elem(&shader
->variants
);
3045 while(!at_end(&shader
->variants
, li
)) {
3046 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3054 /* Move this variant to the head of the list to implement LRU
3055 * deletion of shader's when we have too many.
3057 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3060 /* variant not found, create it now */
3063 unsigned variants_to_cull
;
3066 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3069 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3072 /* First, check if we've exceeded the max number of shader variants.
3073 * If so, free 25% of them (the least recently used ones).
3075 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
3077 if (variants_to_cull
||
3078 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3079 struct pipe_context
*pipe
= &lp
->pipe
;
3082 * XXX: we need to flush the context until we have some sort of
3083 * reference counting in fragment shaders as they may still be binned
3084 * Flushing alone might not be sufficient we need to wait on it too.
3086 llvmpipe_finish(pipe
, __FUNCTION__
);
3089 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3090 * number of shader variants (potentially all of them) could be
3091 * pending for destruction on flush.
3094 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3095 struct lp_fs_variant_list_item
*item
;
3096 if (is_empty_list(&lp
->fs_variants_list
)) {
3099 item
= last_elem(&lp
->fs_variants_list
);
3102 llvmpipe_remove_shader_variant(lp
, item
->base
);
3107 * Generate the new variant.
3110 variant
= generate_variant(lp
, shader
, &key
);
3113 LP_COUNT_ADD(llvm_compile_time
, dt
);
3114 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3116 llvmpipe_variant_count
++;
3118 /* Put the new variant into the list */
3120 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3121 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3122 lp
->nr_fs_variants
++;
3123 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3124 shader
->variants_cached
++;
3128 /* Bind this variant */
3129 lp_setup_set_fs_variant(lp
->setup
, variant
);
3137 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3139 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3140 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3141 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3143 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3147 * Rasterization is disabled if there is no pixel shader and
3148 * both depth and stencil testing are disabled:
3149 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3152 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
3154 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
3157 !lp
->depth_stencil
->depth
.enabled
&&
3158 !lp
->depth_stencil
->stencil
[0].enabled
);