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 "util/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_bitarit.h"
88 #include "gallivm/lp_bld_pack.h"
89 #include "gallivm/lp_bld_format.h"
90 #include "gallivm/lp_bld_quad.h"
92 #include "lp_bld_alpha.h"
93 #include "lp_bld_blend.h"
94 #include "lp_bld_depth.h"
95 #include "lp_bld_interp.h"
96 #include "lp_context.h"
100 #include "lp_state.h"
101 #include "lp_tex_sample.h"
102 #include "lp_flush.h"
103 #include "lp_state_fs.h"
107 /** Fragment shader number (for debugging) */
108 static unsigned fs_no
= 0;
112 * Expand the relevant bits of mask_input to a n*4-dword mask for the
113 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
114 * quad mask vector to 0 or ~0.
115 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
116 * quad arguments with fs length 8.
118 * \param first_quad which quad(s) of the quad group to test, in [0,3]
119 * \param mask_input bitwise mask for the whole 4x4 stamp
122 generate_quad_mask(struct gallivm_state
*gallivm
,
123 struct lp_type fs_type
,
125 LLVMValueRef mask_input
) /* int32 */
127 LLVMBuilderRef builder
= gallivm
->builder
;
128 struct lp_type mask_type
;
129 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
130 LLVMValueRef bits
[16];
131 LLVMValueRef mask
, bits_vec
;
135 * XXX: We'll need a different path for 16 x u8
137 assert(fs_type
.width
== 32);
138 assert(fs_type
.length
<= ARRAY_SIZE(bits
));
139 mask_type
= lp_int_type(fs_type
);
142 * mask_input >>= (quad * 4)
144 switch (first_quad
) {
149 assert(fs_type
.length
== 4);
156 assert(fs_type
.length
== 4);
164 mask_input
= LLVMBuildLShr(builder
,
166 LLVMConstInt(i32t
, shift
, 0),
170 * mask = { mask_input & (1 << i), for i in [0,3] }
172 mask
= lp_build_broadcast(gallivm
,
173 lp_build_vec_type(gallivm
, mask_type
),
176 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
177 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
178 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1ULL << (j
+ 0), 0);
179 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
180 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
181 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (j
+ 5), 0);
183 bits_vec
= LLVMConstVector(bits
, fs_type
.length
);
184 mask
= LLVMBuildAnd(builder
, mask
, bits_vec
, "");
187 * mask = mask == bits ? ~0 : 0
189 mask
= lp_build_compare(gallivm
,
190 mask_type
, PIPE_FUNC_EQUAL
,
197 #define EARLY_DEPTH_TEST 0x1
198 #define LATE_DEPTH_TEST 0x2
199 #define EARLY_DEPTH_WRITE 0x4
200 #define LATE_DEPTH_WRITE 0x8
203 find_output_by_semantic( const struct tgsi_shader_info
*info
,
209 for (i
= 0; i
< info
->num_outputs
; i
++)
210 if (info
->output_semantic_name
[i
] == semantic
&&
211 info
->output_semantic_index
[i
] == index
)
219 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
222 lp_llvm_viewport(LLVMValueRef context_ptr
,
223 struct gallivm_state
*gallivm
,
224 LLVMValueRef viewport_index
)
226 LLVMBuilderRef builder
= gallivm
->builder
;
229 struct lp_type viewport_type
=
230 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
232 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
233 ptr
= LLVMBuildPointerCast(builder
, ptr
,
234 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
236 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
243 lp_build_depth_clamp(struct gallivm_state
*gallivm
,
244 LLVMBuilderRef builder
,
246 LLVMValueRef context_ptr
,
247 LLVMValueRef thread_data_ptr
,
250 LLVMValueRef viewport
, min_depth
, max_depth
;
251 LLVMValueRef viewport_index
;
252 struct lp_build_context f32_bld
;
254 assert(type
.floating
);
255 lp_build_context_init(&f32_bld
, gallivm
, type
);
258 * Assumes clamping of the viewport index will occur in setup/gs. Value
259 * is passed through the rasterization stage via lp_rast_shader_inputs.
261 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
264 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
268 * Load the min and max depth from the lp_jit_context.viewports
269 * array of lp_jit_viewport structures.
271 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
273 /* viewports[viewport_index].min_depth */
274 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
275 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
), "");
276 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
278 /* viewports[viewport_index].max_depth */
279 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
280 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
), "");
281 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
284 * Clamp to the min and max depth values for the given viewport.
286 return lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
291 * Generate the fragment shader, depth/stencil test, and alpha tests.
294 generate_fs_loop(struct gallivm_state
*gallivm
,
295 struct lp_fragment_shader
*shader
,
296 const struct lp_fragment_shader_variant_key
*key
,
297 LLVMBuilderRef builder
,
299 LLVMValueRef context_ptr
,
300 LLVMValueRef num_loop
,
301 struct lp_build_interp_soa_context
*interp
,
302 const struct lp_build_sampler_soa
*sampler
,
303 LLVMValueRef mask_store
,
304 LLVMValueRef (*out_color
)[4],
305 LLVMValueRef depth_ptr
,
306 LLVMValueRef depth_stride
,
308 LLVMValueRef thread_data_ptr
)
310 const struct util_format_description
*zs_format_desc
= NULL
;
311 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
312 struct lp_type int_type
= lp_int_type(type
);
313 LLVMTypeRef vec_type
, int_vec_type
;
314 LLVMValueRef mask_ptr
, mask_val
;
315 LLVMValueRef consts_ptr
, num_consts_ptr
;
316 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
318 LLVMValueRef z_value
, s_value
;
319 LLVMValueRef z_fb
, s_fb
;
320 LLVMValueRef stencil_refs
[2];
321 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
322 struct lp_build_for_loop_state loop_state
;
323 struct lp_build_mask_context mask
;
325 * TODO: figure out if simple_shader optimization is really worthwile to
326 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
327 * code since tests tend to take another codepath than real shaders.
329 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
330 shader
->info
.base
.num_inputs
< 3 &&
331 shader
->info
.base
.num_instructions
< 8) && 0;
332 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
333 util_blend_state_is_dual(&key
->blend
, 0);
339 struct lp_bld_tgsi_system_values system_values
;
341 memset(&system_values
, 0, sizeof(system_values
));
343 if (key
->depth
.enabled
||
344 key
->stencil
[0].enabled
) {
346 zs_format_desc
= util_format_description(key
->zsbuf_format
);
347 assert(zs_format_desc
);
349 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
350 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
351 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
352 if (shader
->info
.base
.writes_memory
)
353 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
354 else if (key
->alpha
.enabled
||
355 key
->blend
.alpha_to_coverage
||
356 shader
->info
.base
.uses_kill
||
357 shader
->info
.base
.writes_samplemask
) {
358 /* With alpha test and kill, can do the depth test early
359 * and hopefully eliminate some quads. But need to do a
360 * special deferred depth write once the final mask value
361 * is known. This only works though if there's either no
362 * stencil test or the stencil value isn't written.
364 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
365 (key
->stencil
[1].enabled
&&
366 key
->stencil
[1].writemask
)))
367 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
369 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
372 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
375 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
378 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
379 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
380 (key
->stencil
[1].enabled
&&
381 key
->stencil
[1].writemask
))))
382 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
388 vec_type
= lp_build_vec_type(gallivm
, type
);
389 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
391 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
392 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
393 /* convert scalar stencil refs into vectors */
394 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
395 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
397 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
398 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
400 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
401 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
403 lp_build_for_loop_begin(&loop_state
, gallivm
,
404 lp_build_const_int32(gallivm
, 0),
407 lp_build_const_int32(gallivm
, 1));
409 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
410 &loop_state
.counter
, 1, "mask_ptr");
411 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
413 memset(outputs
, 0, sizeof outputs
);
415 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
416 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
417 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
418 lp_build_vec_type(gallivm
,
423 if (dual_source_blend
) {
424 assert(key
->nr_cbufs
<= 1);
425 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
426 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
427 lp_build_vec_type(gallivm
,
434 /* 'mask' will control execution based on quad's pixel alive/killed state */
435 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
437 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
438 lp_build_mask_check(&mask
);
440 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
443 if (depth_mode
& EARLY_DEPTH_TEST
) {
445 * Clamp according to ARB_depth_clamp semantics.
447 if (key
->depth_clamp
) {
448 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
451 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
452 zs_format_desc
, key
->resource_1d
,
453 depth_ptr
, depth_stride
,
454 &z_fb
, &s_fb
, loop_state
.counter
);
455 lp_build_depth_stencil_test(gallivm
,
467 if (depth_mode
& EARLY_DEPTH_WRITE
) {
468 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
469 zs_format_desc
, key
->resource_1d
,
470 NULL
, NULL
, NULL
, loop_state
.counter
,
471 depth_ptr
, depth_stride
,
475 * Note mask check if stencil is enabled must be after ds write not after
476 * stencil test otherwise new stencil values may not get written if all
477 * fragments got killed by depth/stencil test.
479 if (!simple_shader
&& key
->stencil
[0].enabled
)
480 lp_build_mask_check(&mask
);
483 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
485 struct lp_build_tgsi_params params
;
486 memset(¶ms
, 0, sizeof(params
));
490 params
.consts_ptr
= consts_ptr
;
491 params
.const_sizes_ptr
= num_consts_ptr
;
492 params
.system_values
= &system_values
;
493 params
.inputs
= interp
->inputs
;
494 params
.context_ptr
= context_ptr
;
495 params
.thread_data_ptr
= thread_data_ptr
;
496 params
.sampler
= sampler
;
497 params
.info
= &shader
->info
.base
;
498 params
.ssbo_ptr
= ssbo_ptr
;
499 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
501 /* Build the actual shader */
502 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
506 if (key
->alpha
.enabled
) {
507 int color0
= find_output_by_semantic(&shader
->info
.base
,
511 if (color0
!= -1 && outputs
[color0
][3]) {
512 const struct util_format_description
*cbuf_format_desc
;
513 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
514 LLVMValueRef alpha_ref_value
;
516 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
517 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
519 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
521 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
522 &mask
, alpha
, alpha_ref_value
,
523 (depth_mode
& LATE_DEPTH_TEST
) != 0);
527 /* Emulate Alpha to Coverage with Alpha test */
528 if (key
->blend
.alpha_to_coverage
) {
529 int color0
= find_output_by_semantic(&shader
->info
.base
,
533 if (color0
!= -1 && outputs
[color0
][3]) {
534 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
536 lp_build_alpha_to_coverage(gallivm
, type
,
538 (depth_mode
& LATE_DEPTH_TEST
) != 0);
542 if (shader
->info
.base
.writes_samplemask
) {
543 int smaski
= find_output_by_semantic(&shader
->info
.base
,
544 TGSI_SEMANTIC_SAMPLEMASK
,
547 struct lp_build_context smask_bld
;
548 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
551 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
553 * Pixel is alive according to the first sample in the mask.
555 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
556 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
557 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
558 lp_build_mask_update(&mask
, smask
);
562 if (depth_mode
& LATE_DEPTH_TEST
) {
563 int pos0
= find_output_by_semantic(&shader
->info
.base
,
564 TGSI_SEMANTIC_POSITION
,
566 int s_out
= find_output_by_semantic(&shader
->info
.base
,
567 TGSI_SEMANTIC_STENCIL
,
569 if (pos0
!= -1 && outputs
[pos0
][2]) {
570 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
573 * Clamp according to ARB_depth_clamp semantics.
575 if (key
->depth_clamp
) {
576 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
580 if (s_out
!= -1 && outputs
[s_out
][1]) {
581 /* there's only one value, and spec says to discard additional bits */
582 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
583 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
584 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
585 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
586 stencil_refs
[1] = stencil_refs
[0];
589 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
590 zs_format_desc
, key
->resource_1d
,
591 depth_ptr
, depth_stride
,
592 &z_fb
, &s_fb
, loop_state
.counter
);
594 lp_build_depth_stencil_test(gallivm
,
606 if (depth_mode
& LATE_DEPTH_WRITE
) {
607 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
608 zs_format_desc
, key
->resource_1d
,
609 NULL
, NULL
, NULL
, loop_state
.counter
,
610 depth_ptr
, depth_stride
,
614 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
615 (depth_mode
& LATE_DEPTH_WRITE
))
617 /* Need to apply a reduced mask to the depth write. Reload the
618 * depth value, update from zs_value with the new mask value and
621 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
622 zs_format_desc
, key
->resource_1d
,
623 &mask
, z_fb
, s_fb
, loop_state
.counter
,
624 depth_ptr
, depth_stride
,
630 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
632 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
633 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
634 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
636 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
637 if(outputs
[attrib
][chan
]) {
638 /* XXX: just initialize outputs to point at colors[] and
641 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
642 LLVMValueRef color_ptr
;
643 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
644 &loop_state
.counter
, 1, "");
645 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
646 LLVMBuildStore(builder
, out
, color_ptr
);
652 if (key
->occlusion_count
) {
653 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
654 lp_build_name(counter
, "counter");
655 lp_build_occlusion_count(gallivm
, type
,
656 lp_build_mask_value(&mask
), counter
);
659 mask_val
= lp_build_mask_end(&mask
);
660 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
661 lp_build_for_loop_end(&loop_state
);
666 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
668 * Fragment Shader outputs pixels in small 2x2 blocks
669 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
671 * However in memory pixels are stored in rows
672 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
674 * @param type fragment shader type (4x or 8x float)
675 * @param num_fs number of fs_src
676 * @param is_1d whether we're outputting to a 1d resource
677 * @param dst_channels number of output channels
678 * @param fs_src output from fragment shader
679 * @param dst pointer to store result
680 * @param pad_inline is channel padding inline or at end of row
681 * @return the number of dsts
684 generate_fs_twiddle(struct gallivm_state
*gallivm
,
687 unsigned dst_channels
,
688 LLVMValueRef fs_src
[][4],
692 LLVMValueRef src
[16];
698 unsigned pixels
= type
.length
/ 4;
699 unsigned reorder_group
;
700 unsigned src_channels
;
704 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
705 src_count
= num_fs
* src_channels
;
707 assert(pixels
== 2 || pixels
== 1);
708 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
711 * Transpose from SoA -> AoS
713 for (i
= 0; i
< num_fs
; ++i
) {
714 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
718 * Pick transformation options
725 if (dst_channels
== 1) {
731 } else if (dst_channels
== 2) {
735 } else if (dst_channels
> 2) {
742 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
748 * Split the src in half
751 for (i
= num_fs
; i
> 0; --i
) {
752 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
753 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
761 * Ensure pixels are in memory order
764 /* Twiddle pixels by reordering the array, e.g.:
766 * src_count = 8 -> 0 2 1 3 4 6 5 7
767 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
769 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
771 for (i
= 0; i
< src_count
; ++i
) {
772 unsigned group
= i
/ reorder_group
;
773 unsigned block
= (group
/ 4) * 4 * reorder_group
;
774 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
777 } else if (twiddle
) {
778 /* Twiddle pixels across elements of array */
780 * XXX: we should avoid this in some cases, but would need to tell
781 * lp_build_conv to reorder (or deal with it ourselves).
783 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
786 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
790 * Moves any padding between pixels to the end
791 * e.g. RGBXRGBX -> RGBRGBXX
794 unsigned char swizzles
[16];
795 unsigned elems
= pixels
* dst_channels
;
797 for (i
= 0; i
< type
.length
; ++i
) {
799 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
801 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
804 for (i
= 0; i
< src_count
; ++i
) {
805 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
814 * Untwiddle and transpose, much like the above.
815 * However, this is after conversion, so we get packed vectors.
816 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
817 * the vectors will look like:
818 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
819 * be swizzled here). Extending to 16bit should be trivial.
820 * Should also be extended to handle twice wide vectors with AVX2...
823 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
830 struct lp_type type64
, type16
, type32
;
831 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
832 LLVMBuilderRef builder
= gallivm
->builder
;
833 LLVMValueRef tmp
[4], shuf
[8];
834 for (j
= 0; j
< 2; j
++) {
835 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
836 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
837 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
838 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
841 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
842 assert(type
.width
== 8);
843 assert(type
.length
== 16);
845 type8_t
= lp_build_vec_type(gallivm
, type
);
850 type64_t
= lp_build_vec_type(gallivm
, type64
);
855 type16_t
= lp_build_vec_type(gallivm
, type16
);
860 type32_t
= lp_build_vec_type(gallivm
, type32
);
862 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
864 if (src_count
== 1) {
865 /* transpose was no-op, just untwiddle */
866 LLVMValueRef shuf_vec
;
867 shuf_vec
= LLVMConstVector(shuf
, 8);
868 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
869 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
870 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
871 } else if (src_count
== 2) {
872 LLVMValueRef shuf_vec
;
873 shuf_vec
= LLVMConstVector(shuf
, 4);
875 for (i
= 0; i
< 2; i
++) {
876 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
877 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
878 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
881 for (j
= 0; j
< 2; j
++) {
882 LLVMValueRef lo
, hi
, lo2
, hi2
;
884 * Note that if we only really have 3 valid channels (rgb)
885 * and we don't need alpha we could substitute a undef here
886 * for the respective channel (causing llvm to drop conversion
889 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
890 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
891 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
892 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
893 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
894 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
895 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
902 * Load an unswizzled block of pixels from memory
905 load_unswizzled_block(struct gallivm_state
*gallivm
,
906 LLVMValueRef base_ptr
,
908 unsigned block_width
,
909 unsigned block_height
,
911 struct lp_type dst_type
,
913 unsigned dst_alignment
)
915 LLVMBuilderRef builder
= gallivm
->builder
;
916 unsigned row_size
= dst_count
/ block_height
;
919 /* Ensure block exactly fits into dst */
920 assert((block_width
* block_height
) % dst_count
== 0);
922 for (i
= 0; i
< dst_count
; ++i
) {
923 unsigned x
= i
% row_size
;
924 unsigned y
= i
/ row_size
;
926 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
927 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
930 LLVMValueRef dst_ptr
;
932 gep
[0] = lp_build_const_int32(gallivm
, 0);
933 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
935 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
936 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
937 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
939 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
941 LLVMSetAlignment(dst
[i
], dst_alignment
);
947 * Store an unswizzled block of pixels to memory
950 store_unswizzled_block(struct gallivm_state
*gallivm
,
951 LLVMValueRef base_ptr
,
953 unsigned block_width
,
954 unsigned block_height
,
956 struct lp_type src_type
,
958 unsigned src_alignment
)
960 LLVMBuilderRef builder
= gallivm
->builder
;
961 unsigned row_size
= src_count
/ block_height
;
964 /* Ensure src exactly fits into block */
965 assert((block_width
* block_height
) % src_count
== 0);
967 for (i
= 0; i
< src_count
; ++i
) {
968 unsigned x
= i
% row_size
;
969 unsigned y
= i
/ row_size
;
971 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
972 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
975 LLVMValueRef src_ptr
;
977 gep
[0] = lp_build_const_int32(gallivm
, 0);
978 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
980 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
981 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
982 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
984 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
986 LLVMSetAlignment(src_ptr
, src_alignment
);
992 * Checks if a format description is an arithmetic format
994 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
996 static inline boolean
997 is_arithmetic_format(const struct util_format_description
*format_desc
)
999 boolean arith
= false;
1002 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1003 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1004 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1012 * Checks if this format requires special handling due to required expansion
1013 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1016 static inline boolean
1017 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1019 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1020 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1028 * Retrieves the type representing the memory layout for a format
1030 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1033 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1034 struct lp_type
* type
)
1039 if (format_expands_to_float_soa(format_desc
)) {
1040 /* just make this a uint with width of block */
1041 type
->floating
= false;
1042 type
->fixed
= false;
1045 type
->width
= format_desc
->block
.bits
;
1050 for (i
= 0; i
< 4; i
++)
1051 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1055 memset(type
, 0, sizeof(struct lp_type
));
1056 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1057 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1058 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1059 type
->norm
= format_desc
->channel
[chan
].normalized
;
1061 if (is_arithmetic_format(format_desc
)) {
1065 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1066 type
->width
+= format_desc
->channel
[i
].size
;
1069 type
->width
= format_desc
->channel
[chan
].size
;
1070 type
->length
= format_desc
->nr_channels
;
1076 * Retrieves the type for a format which is usable in the blending code.
1078 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1081 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1082 struct lp_type
* type
)
1087 if (format_expands_to_float_soa(format_desc
)) {
1088 /* always use ordinary floats for blending */
1089 type
->floating
= true;
1090 type
->fixed
= false;
1098 for (i
= 0; i
< 4; i
++)
1099 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1103 memset(type
, 0, sizeof(struct lp_type
));
1104 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1105 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1106 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1107 type
->norm
= format_desc
->channel
[chan
].normalized
;
1108 type
->width
= format_desc
->channel
[chan
].size
;
1109 type
->length
= format_desc
->nr_channels
;
1111 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1112 if (format_desc
->channel
[i
].size
> type
->width
)
1113 type
->width
= format_desc
->channel
[i
].size
;
1116 if (type
->floating
) {
1119 if (type
->width
<= 8) {
1121 } else if (type
->width
<= 16) {
1128 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1135 * Scale a normalized value from src_bits to dst_bits.
1137 * The exact calculation is
1139 * dst = iround(src * dst_mask / src_mask)
1141 * or with integer rounding
1143 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1147 * src_mask = (1 << src_bits) - 1
1148 * dst_mask = (1 << dst_bits) - 1
1150 * but we try to avoid division and multiplication through shifts.
1152 static inline LLVMValueRef
1153 scale_bits(struct gallivm_state
*gallivm
,
1157 struct lp_type src_type
)
1159 LLVMBuilderRef builder
= gallivm
->builder
;
1160 LLVMValueRef result
= src
;
1162 if (dst_bits
< src_bits
) {
1163 int delta_bits
= src_bits
- dst_bits
;
1165 if (delta_bits
<= dst_bits
) {
1167 * Approximate the rescaling with a single shift.
1169 * This gives the wrong rounding.
1172 result
= LLVMBuildLShr(builder
,
1174 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1179 * Try more accurate rescaling.
1183 * Drop the least significant bits to make space for the multiplication.
1185 * XXX: A better approach would be to use a wider integer type as intermediate. But
1186 * this is enough to convert alpha from 16bits -> 2 when rendering to
1187 * PIPE_FORMAT_R10G10B10A2_UNORM.
1189 result
= LLVMBuildLShr(builder
,
1191 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1195 result
= LLVMBuildMul(builder
,
1197 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1201 * Add a rounding term before the division.
1203 * TODO: Handle signed integers too.
1205 if (!src_type
.sign
) {
1206 result
= LLVMBuildAdd(builder
,
1208 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1213 * Approximate the division by src_mask with a src_bits shift.
1215 * Given the src has already been shifted by dst_bits, all we need
1216 * to do is to shift by the difference.
1219 result
= LLVMBuildLShr(builder
,
1221 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1225 } else if (dst_bits
> src_bits
) {
1227 int db
= dst_bits
- src_bits
;
1229 /* Shift left by difference in bits */
1230 result
= LLVMBuildShl(builder
,
1232 lp_build_const_int_vec(gallivm
, src_type
, db
),
1235 if (db
<= src_bits
) {
1236 /* Enough bits in src to fill the remainder */
1237 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1239 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1242 result
= LLVMBuildOr(builder
, result
, lower
, "");
1243 } else if (db
> src_bits
) {
1244 /* Need to repeatedly copy src bits to fill remainder in dst */
1247 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1248 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1250 result
= LLVMBuildOr(builder
,
1252 LLVMBuildLShr(builder
, result
, shuv
, ""),
1262 * If RT is a smallfloat (needing denorms) format
1265 have_smallfloat_format(struct lp_type dst_type
,
1266 enum pipe_format format
)
1268 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1269 /* due to format handling hacks this format doesn't have floating set
1270 * here (and actually has width set to 32 too) so special case this. */
1271 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1276 * Convert from memory format to blending format
1278 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1281 convert_to_blend_type(struct gallivm_state
*gallivm
,
1282 unsigned block_size
,
1283 const struct util_format_description
*src_fmt
,
1284 struct lp_type src_type
,
1285 struct lp_type dst_type
,
1286 LLVMValueRef
* src
, // and dst
1289 LLVMValueRef
*dst
= src
;
1290 LLVMBuilderRef builder
= gallivm
->builder
;
1291 struct lp_type blend_type
;
1292 struct lp_type mem_type
;
1294 unsigned pixels
= block_size
/ num_srcs
;
1298 * full custom path for packed floats and srgb formats - none of the later
1299 * functions would do anything useful, and given the lp_type representation they
1300 * can't be fixed. Should really have some SoA blend path for these kind of
1301 * formats rather than hacking them in here.
1303 if (format_expands_to_float_soa(src_fmt
)) {
1304 LLVMValueRef tmpsrc
[4];
1306 * This is pretty suboptimal for this case blending in SoA would be much
1307 * better, since conversion gets us SoA values so need to convert back.
1309 assert(src_type
.width
== 32 || src_type
.width
== 16);
1310 assert(dst_type
.floating
);
1311 assert(dst_type
.width
== 32);
1312 assert(dst_type
.length
% 4 == 0);
1313 assert(num_srcs
% 4 == 0);
1315 if (src_type
.width
== 16) {
1316 /* expand 4x16bit values to 4x32bit */
1317 struct lp_type type32x4
= src_type
;
1318 LLVMTypeRef ltype32x4
;
1319 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1320 type32x4
.width
= 32;
1321 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1322 for (i
= 0; i
< num_fetch
; i
++) {
1323 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1325 src_type
.width
= 32;
1327 for (i
= 0; i
< 4; i
++) {
1330 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1331 LLVMValueRef tmpsoa
[4];
1332 LLVMValueRef tmps
= tmpsrc
[i
];
1333 if (dst_type
.length
== 8) {
1334 LLVMValueRef shuffles
[8];
1336 /* fetch was 4 values but need 8-wide output values */
1337 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1339 * for 8-wide aos transpose would give us wrong order not matching
1340 * incoming converted fs values and mask. ARGH.
1342 for (j
= 0; j
< 4; j
++) {
1343 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1344 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1346 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1347 LLVMConstVector(shuffles
, 8), "");
1349 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1350 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1353 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1355 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1360 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1361 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1363 /* Is the format arithmetic */
1364 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1365 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1367 /* Pad if necessary */
1368 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1369 for (i
= 0; i
< num_srcs
; ++i
) {
1370 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1373 src_type
.length
= dst_type
.length
;
1376 /* Special case for half-floats */
1377 if (mem_type
.width
== 16 && mem_type
.floating
) {
1378 assert(blend_type
.width
== 32 && blend_type
.floating
);
1379 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1387 src_type
.width
= blend_type
.width
* blend_type
.length
;
1388 blend_type
.length
*= pixels
;
1389 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1391 for (i
= 0; i
< num_srcs
; ++i
) {
1392 LLVMValueRef chans
[4];
1393 LLVMValueRef res
= NULL
;
1395 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1397 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1399 unsigned sa
= src_fmt
->channel
[j
].shift
;
1400 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1401 unsigned from_lsb
= j
;
1403 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1406 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1408 /* Extract bits from source */
1409 chans
[j
] = LLVMBuildLShr(builder
,
1411 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1414 chans
[j
] = LLVMBuildAnd(builder
,
1416 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1420 if (src_type
.norm
) {
1421 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1422 blend_type
.width
, chans
[j
], src_type
);
1425 /* Insert bits into correct position */
1426 chans
[j
] = LLVMBuildShl(builder
,
1428 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1434 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1438 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1444 * Convert from blending format to memory format
1446 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1449 convert_from_blend_type(struct gallivm_state
*gallivm
,
1450 unsigned block_size
,
1451 const struct util_format_description
*src_fmt
,
1452 struct lp_type src_type
,
1453 struct lp_type dst_type
,
1454 LLVMValueRef
* src
, // and dst
1457 LLVMValueRef
* dst
= src
;
1459 struct lp_type mem_type
;
1460 struct lp_type blend_type
;
1461 LLVMBuilderRef builder
= gallivm
->builder
;
1462 unsigned pixels
= block_size
/ num_srcs
;
1466 * full custom path for packed floats and srgb formats - none of the later
1467 * functions would do anything useful, and given the lp_type representation they
1468 * can't be fixed. Should really have some SoA blend path for these kind of
1469 * formats rather than hacking them in here.
1471 if (format_expands_to_float_soa(src_fmt
)) {
1473 * This is pretty suboptimal for this case blending in SoA would be much
1474 * better - we need to transpose the AoS values back to SoA values for
1475 * conversion/packing.
1477 assert(src_type
.floating
);
1478 assert(src_type
.width
== 32);
1479 assert(src_type
.length
% 4 == 0);
1480 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1482 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1483 LLVMValueRef tmpsoa
[4], tmpdst
;
1484 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1485 /* really really need SoA here */
1487 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1488 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1491 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1495 if (src_type
.length
== 8) {
1496 LLVMValueRef tmpaos
, shuffles
[8];
1499 * for 8-wide aos transpose has given us wrong order not matching
1500 * output order. HMPF. Also need to split the output values manually.
1502 for (j
= 0; j
< 4; j
++) {
1503 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1504 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1506 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1507 LLVMConstVector(shuffles
, 8), "");
1508 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1509 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1515 if (dst_type
.width
== 16) {
1516 struct lp_type type16x8
= dst_type
;
1517 struct lp_type type32x4
= dst_type
;
1518 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1519 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1520 type16x8
.length
= 8;
1521 type32x4
.width
= 32;
1522 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1523 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1524 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1525 /* We could do vector truncation but it doesn't generate very good code */
1526 for (i
= 0; i
< num_fetch
; i
++) {
1527 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1528 src
[i
], lp_build_zero(gallivm
, type32x4
));
1529 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1530 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1531 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1537 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1538 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1540 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1542 /* Special case for half-floats */
1543 if (mem_type
.width
== 16 && mem_type
.floating
) {
1544 int length
= dst_type
.length
;
1545 assert(blend_type
.width
== 32 && blend_type
.floating
);
1547 dst_type
.length
= src_type
.length
;
1549 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1551 dst_type
.length
= length
;
1555 /* Remove any padding */
1556 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1557 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1559 for (i
= 0; i
< num_srcs
; ++i
) {
1560 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1564 /* No bit arithmetic to do */
1569 src_type
.length
= pixels
;
1570 src_type
.width
= blend_type
.length
* blend_type
.width
;
1571 dst_type
.length
= pixels
;
1573 for (i
= 0; i
< num_srcs
; ++i
) {
1574 LLVMValueRef chans
[4];
1575 LLVMValueRef res
= NULL
;
1577 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1579 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1581 unsigned sa
= src_fmt
->channel
[j
].shift
;
1582 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1583 unsigned from_lsb
= j
;
1585 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1588 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1590 for (k
= 0; k
< blend_type
.width
; ++k
) {
1595 chans
[j
] = LLVMBuildLShr(builder
,
1597 lp_build_const_int_vec(gallivm
, src_type
,
1598 from_lsb
* blend_type
.width
),
1601 chans
[j
] = LLVMBuildAnd(builder
,
1603 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1606 /* Scale down bits */
1607 if (src_type
.norm
) {
1608 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1609 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1613 chans
[j
] = LLVMBuildShl(builder
,
1615 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1618 sa
+= src_fmt
->channel
[j
].size
;
1623 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1627 assert (dst_type
.width
!= 24);
1629 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1635 * Convert alpha to same blend type as src
1638 convert_alpha(struct gallivm_state
*gallivm
,
1639 struct lp_type row_type
,
1640 struct lp_type alpha_type
,
1641 const unsigned block_size
,
1642 const unsigned block_height
,
1643 const unsigned src_count
,
1644 const unsigned dst_channels
,
1645 const bool pad_inline
,
1646 LLVMValueRef
* src_alpha
)
1648 LLVMBuilderRef builder
= gallivm
->builder
;
1650 unsigned length
= row_type
.length
;
1651 row_type
.length
= alpha_type
.length
;
1653 /* Twiddle the alpha to match pixels */
1654 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1657 * TODO this should use single lp_build_conv call for
1658 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1660 for (i
= 0; i
< block_height
; ++i
) {
1661 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1664 alpha_type
= row_type
;
1665 row_type
.length
= length
;
1667 /* If only one channel we can only need the single alpha value per pixel */
1668 if (src_count
== 1 && dst_channels
== 1) {
1670 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1672 /* If there are more srcs than rows then we need to split alpha up */
1673 if (src_count
> block_height
) {
1674 for (i
= src_count
; i
> 0; --i
) {
1675 unsigned pixels
= block_size
/ src_count
;
1676 unsigned idx
= i
- 1;
1678 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1679 (idx
* pixels
) % 4, pixels
);
1683 /* If there is a src for each pixel broadcast the alpha across whole row */
1684 if (src_count
== block_size
) {
1685 for (i
= 0; i
< src_count
; ++i
) {
1686 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1687 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1690 unsigned pixels
= block_size
/ src_count
;
1691 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1692 unsigned alpha_span
= 1;
1693 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1695 /* Check if we need 2 src_alphas for our shuffles */
1696 if (pixels
> alpha_type
.length
) {
1700 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1701 for (j
= 0; j
< row_type
.length
; ++j
) {
1702 if (j
< pixels
* channels
) {
1703 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1705 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1709 for (i
= 0; i
< src_count
; ++i
) {
1710 unsigned idx1
= i
, idx2
= i
;
1712 if (alpha_span
> 1){
1717 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1720 LLVMConstVector(shuffles
, row_type
.length
),
1729 * Generates the blend function for unswizzled colour buffers
1730 * Also generates the read & write from colour buffer
1733 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1735 struct lp_fragment_shader_variant
*variant
,
1736 enum pipe_format out_format
,
1737 unsigned int num_fs
,
1738 struct lp_type fs_type
,
1739 LLVMValueRef
* fs_mask
,
1740 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1741 LLVMValueRef context_ptr
,
1742 LLVMValueRef color_ptr
,
1743 LLVMValueRef stride
,
1744 unsigned partial_mask
,
1747 const unsigned alpha_channel
= 3;
1748 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1749 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1750 const unsigned block_size
= block_width
* block_height
;
1751 const unsigned lp_integer_vector_width
= 128;
1753 LLVMBuilderRef builder
= gallivm
->builder
;
1754 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1755 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1756 LLVMValueRef src_alpha
[4 * 4];
1757 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1758 LLVMValueRef src_mask
[4 * 4];
1759 LLVMValueRef src
[4 * 4];
1760 LLVMValueRef src1
[4 * 4];
1761 LLVMValueRef dst
[4 * 4];
1762 LLVMValueRef blend_color
;
1763 LLVMValueRef blend_alpha
;
1764 LLVMValueRef i32_zero
;
1765 LLVMValueRef check_mask
;
1766 LLVMValueRef undef_src_val
;
1768 struct lp_build_mask_context mask_ctx
;
1769 struct lp_type mask_type
;
1770 struct lp_type blend_type
;
1771 struct lp_type row_type
;
1772 struct lp_type dst_type
;
1773 struct lp_type ls_type
;
1775 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1776 unsigned vector_width
;
1777 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1778 unsigned dst_channels
;
1783 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1785 unsigned dst_alignment
;
1787 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1788 bool has_alpha
= false;
1789 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1790 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1792 const boolean is_1d
= variant
->key
.resource_1d
;
1793 boolean twiddle_after_convert
= FALSE
;
1794 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1795 LLVMValueRef fpstate
= 0;
1797 /* Get type from output format */
1798 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1799 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1802 * Technically this code should go into lp_build_smallfloat_to_float
1803 * and lp_build_float_to_smallfloat but due to the
1804 * http://llvm.org/bugs/show_bug.cgi?id=6393
1805 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1806 * So the ordering is important here and there shouldn't be any
1807 * llvm ir instrunctions in this function before
1808 * this, otherwise half-float format conversions won't work
1809 * (again due to llvm bug #6393).
1811 if (have_smallfloat_format(dst_type
, out_format
)) {
1812 /* We need to make sure that denorms are ok for half float
1814 fpstate
= lp_build_fpstate_get(gallivm
);
1815 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1818 mask_type
= lp_int32_vec4_type();
1819 mask_type
.length
= fs_type
.length
;
1821 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1822 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1825 /* Do not bother executing code when mask is empty.. */
1827 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1829 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1830 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1833 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1834 lp_build_mask_check(&mask_ctx
);
1837 partial_mask
|= !variant
->opaque
;
1838 i32_zero
= lp_build_const_int32(gallivm
, 0);
1840 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1842 row_type
.length
= fs_type
.length
;
1843 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1845 /* Compute correct swizzle and count channels */
1846 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1849 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1850 /* Ensure channel is used */
1851 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1855 /* Ensure not already written to (happens in case with GL_ALPHA) */
1856 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1860 /* Ensure we havn't already found all channels */
1861 if (dst_channels
>= out_format_desc
->nr_channels
) {
1865 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1868 if (i
== alpha_channel
) {
1873 if (format_expands_to_float_soa(out_format_desc
)) {
1875 * the code above can't work for layout_other
1876 * for srgb it would sort of work but we short-circuit swizzles, etc.
1877 * as that is done as part of unpack / pack.
1879 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1885 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1888 /* If 3 channels then pad to include alpha for 4 element transpose */
1889 if (dst_channels
== 3) {
1890 assert (!has_alpha
);
1891 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1892 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1895 if (out_format_desc
->nr_channels
== 4) {
1898 * We use alpha from the color conversion, not separate one.
1899 * We had to include it for transpose, hence it will get converted
1900 * too (albeit when doing transpose after conversion, that would
1901 * no longer be the case necessarily).
1902 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1903 * otherwise we really have padding, not alpha, included.)
1910 * Load shader output
1912 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1913 /* Always load alpha for use in blending */
1916 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1919 alpha
= undef_src_val
;
1922 /* Load each channel */
1923 for (j
= 0; j
< dst_channels
; ++j
) {
1924 assert(swizzle
[j
] < 4);
1926 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1929 fs_src
[i
][j
] = undef_src_val
;
1933 /* If 3 channels then pad to include alpha for 4 element transpose */
1935 * XXX If we include that here maybe could actually use it instead of
1936 * separate alpha for blending?
1937 * (Difficult though we actually convert pad channels, not alpha.)
1939 if (dst_channels
== 3 && !has_alpha
) {
1940 fs_src
[i
][3] = alpha
;
1943 /* We split the row_mask and row_alpha as we want 128bit interleave */
1944 if (fs_type
.length
== 8) {
1945 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1947 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1948 src_channels
, src_channels
);
1950 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1951 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1952 src_channels
, src_channels
);
1954 src_mask
[i
] = fs_mask
[i
];
1955 src_alpha
[i
] = alpha
;
1958 if (dual_source_blend
) {
1959 /* same as above except different src/dst, skip masks and comments... */
1960 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1963 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1966 alpha
= undef_src_val
;
1969 for (j
= 0; j
< dst_channels
; ++j
) {
1970 assert(swizzle
[j
] < 4);
1972 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1975 fs_src1
[i
][j
] = undef_src_val
;
1978 if (dst_channels
== 3 && !has_alpha
) {
1979 fs_src1
[i
][3] = alpha
;
1981 if (fs_type
.length
== 8) {
1982 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1983 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1984 src_channels
, src_channels
);
1986 src1_alpha
[i
] = alpha
;
1991 if (util_format_is_pure_integer(out_format
)) {
1993 * In this case fs_type was really ints or uints disguised as floats,
1996 fs_type
.floating
= 0;
1997 fs_type
.sign
= dst_type
.sign
;
1998 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1999 for (j
= 0; j
< dst_channels
; ++j
) {
2000 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2001 lp_build_vec_type(gallivm
, fs_type
), "");
2003 if (dst_channels
== 3 && !has_alpha
) {
2004 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2005 lp_build_vec_type(gallivm
, fs_type
), "");
2011 * We actually should generally do conversion first (for non-1d cases)
2012 * when the blend format is 8 or 16 bits. The reason is obvious,
2013 * there's 2 or 4 times less vectors to deal with for the interleave...
2014 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2015 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2016 * unpack only with 128bit vectors).
2017 * Note: for 16bit sizes really need matching pack conversion code
2019 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2020 twiddle_after_convert
= TRUE
;
2024 * Pixel twiddle from fragment shader order to memory order
2026 if (!twiddle_after_convert
) {
2027 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2028 dst_channels
, fs_src
, src
, pad_inline
);
2029 if (dual_source_blend
) {
2030 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2031 fs_src1
, src1
, pad_inline
);
2034 src_count
= num_fullblock_fs
* dst_channels
;
2036 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2037 * (AVX) turn out the same later when untwiddling/transpose (albeit
2038 * for true AVX2 path untwiddle needs to be different).
2039 * For now just order by colors first (so we can use unpack later).
2041 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2042 for (i
= 0; i
< dst_channels
; i
++) {
2043 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2044 if (dual_source_blend
) {
2045 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2051 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2052 if (src_count
!= num_fullblock_fs
* src_channels
) {
2053 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2054 row_type
.length
/= ds
;
2055 fs_type
.length
= row_type
.length
;
2058 blend_type
= row_type
;
2059 mask_type
.length
= 4;
2061 /* Convert src to row_type */
2062 if (dual_source_blend
) {
2063 struct lp_type old_row_type
= row_type
;
2064 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2065 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2068 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2071 /* If the rows are not an SSE vector, combine them to become SSE size! */
2072 if ((row_type
.width
* row_type
.length
) % 128) {
2073 unsigned bits
= row_type
.width
* row_type
.length
;
2076 assert(src_count
>= (vector_width
/ bits
));
2078 dst_count
= src_count
/ (vector_width
/ bits
);
2080 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2081 if (dual_source_blend
) {
2082 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2085 row_type
.length
*= combined
;
2086 src_count
/= combined
;
2088 bits
= row_type
.width
* row_type
.length
;
2089 assert(bits
== 128 || bits
== 256);
2092 if (twiddle_after_convert
) {
2093 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2094 if (dual_source_blend
) {
2095 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2100 * Blend Colour conversion
2102 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2103 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2104 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2105 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2106 &i32_zero
, 1, ""), "");
2109 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2111 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2113 * since blending is done with floats, there was no conversion.
2114 * However, the rules according to fixed point renderbuffers still
2115 * apply, that is we must clamp inputs to 0.0/1.0.
2116 * (This would apply to separate alpha conversion too but we currently
2117 * force has_alpha to be true.)
2118 * TODO: should skip this with "fake" blend, since post-blend conversion
2119 * will clamp anyway.
2120 * TODO: could also skip this if fragment color clamping is enabled. We
2121 * don't support it natively so it gets baked into the shader however, so
2122 * can't really tell here.
2124 struct lp_build_context f32_bld
;
2125 assert(row_type
.floating
);
2126 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2127 for (i
= 0; i
< src_count
; i
++) {
2128 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2130 if (dual_source_blend
) {
2131 for (i
= 0; i
< src_count
; i
++) {
2132 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2135 /* probably can't be different than row_type but better safe than sorry... */
2136 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2137 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2141 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2143 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2144 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2146 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2147 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2149 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2150 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2156 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2158 if (src_count
< block_height
) {
2159 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2160 } else if (src_count
> block_height
) {
2161 for (i
= src_count
; i
> 0; --i
) {
2162 unsigned pixels
= block_size
/ src_count
;
2163 unsigned idx
= i
- 1;
2165 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2166 (idx
* pixels
) % 4, pixels
);
2170 assert(mask_type
.width
== 32);
2172 for (i
= 0; i
< src_count
; ++i
) {
2173 unsigned pixels
= block_size
/ src_count
;
2174 unsigned pixel_width
= row_type
.width
* dst_channels
;
2176 if (pixel_width
== 24) {
2177 mask_type
.width
= 8;
2178 mask_type
.length
= vector_width
/ mask_type
.width
;
2180 mask_type
.length
= pixels
;
2181 mask_type
.width
= row_type
.width
* dst_channels
;
2184 * If mask_type width is smaller than 32bit, this doesn't quite
2185 * generate the most efficient code (could use some pack).
2187 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2188 lp_build_int_vec_type(gallivm
, mask_type
), "");
2190 mask_type
.length
*= dst_channels
;
2191 mask_type
.width
/= dst_channels
;
2194 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2195 lp_build_int_vec_type(gallivm
, mask_type
), "");
2196 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2203 struct lp_type alpha_type
= fs_type
;
2204 alpha_type
.length
= 4;
2205 convert_alpha(gallivm
, row_type
, alpha_type
,
2206 block_size
, block_height
,
2207 src_count
, dst_channels
,
2208 pad_inline
, src_alpha
);
2209 if (dual_source_blend
) {
2210 convert_alpha(gallivm
, row_type
, alpha_type
,
2211 block_size
, block_height
,
2212 src_count
, dst_channels
,
2213 pad_inline
, src1_alpha
);
2219 * Load dst from memory
2221 if (src_count
< block_height
) {
2222 dst_count
= block_height
;
2224 dst_count
= src_count
;
2227 dst_type
.length
*= block_size
/ dst_count
;
2229 if (format_expands_to_float_soa(out_format_desc
)) {
2231 * we need multiple values at once for the conversion, so can as well
2232 * load them vectorized here too instead of concatenating later.
2233 * (Still need concatenation later for 8-wide vectors).
2235 dst_count
= block_height
;
2236 dst_type
.length
= block_width
;
2240 * Compute the alignment of the destination pointer in bytes
2241 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2242 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2243 * 1d tex but can't distinguish here) so need to stick with per-pixel
2244 * alignment in this case.
2247 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2250 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2252 /* Force power-of-two alignment by extracting only the least-significant-bit */
2253 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2255 * Resource base and stride pointers are aligned to 16 bytes, so that's
2256 * the maximum alignment we can guarantee
2258 dst_alignment
= MIN2(16, dst_alignment
);
2262 if (dst_count
> src_count
) {
2263 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2264 util_is_power_of_two_or_zero(dst_type
.length
) &&
2265 dst_type
.length
* dst_type
.width
< 128) {
2267 * Never try to load values as 4xi8 which we will then
2268 * concatenate to larger vectors. This gives llvm a real
2269 * headache (the problem is the type legalizer (?) will
2270 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2271 * then the shuffles to concatenate are more or less impossible
2272 * - llvm is easily capable of generating a sequence of 32
2273 * pextrb/pinsrb instructions for that. Albeit it appears to
2274 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2275 * width to avoid the trouble (16bit seems not as bad, llvm
2276 * probably recognizes the load+shuffle as only one shuffle
2277 * is necessary, but we can do just the same anyway).
2279 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2285 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2286 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2287 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2288 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2293 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2294 dst
, ls_type
, dst_count
, dst_alignment
);
2299 * Convert from dst/output format to src/blending format.
2301 * This is necessary as we can only read 1 row from memory at a time,
2302 * so the minimum dst_count will ever be at this point is 4.
2304 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2305 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2306 * on all 16 pixels in that single vector at once.
2308 if (dst_count
> src_count
) {
2309 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2310 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2311 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2312 for (i
= 0; i
< dst_count
; i
++) {
2313 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2317 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2319 if (ls_type
.length
!= dst_type
.length
) {
2320 struct lp_type tmp_type
= dst_type
;
2321 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2322 for (i
= 0; i
< src_count
; i
++) {
2323 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2324 lp_build_vec_type(gallivm
, tmp_type
), "");
2332 /* XXX this is broken for RGB8 formats -
2333 * they get expanded from 12 to 16 elements (to include alpha)
2334 * by convert_to_blend_type then reduced to 15 instead of 12
2335 * by convert_from_blend_type (a simple fix though breaks A8...).
2336 * R16G16B16 also crashes differently however something going wrong
2337 * inside llvm handling npot vector sizes seemingly.
2338 * It seems some cleanup could be done here (like skipping conversion/blend
2341 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2342 row_type
, dst
, src_count
);
2345 * FIXME: Really should get logic ops / masks out of generic blend / row
2346 * format. Logic ops will definitely not work on the blend float format
2347 * used for SRGB here and I think OpenGL expects this to work as expected
2348 * (that is incoming values converted to srgb then logic op applied).
2350 for (i
= 0; i
< src_count
; ++i
) {
2351 dst
[i
] = lp_build_blend_aos(gallivm
,
2352 &variant
->key
.blend
,
2357 has_alpha
? NULL
: src_alpha
[i
],
2359 has_alpha
? NULL
: src1_alpha
[i
],
2361 partial_mask
? src_mask
[i
] : NULL
,
2363 has_alpha
? NULL
: blend_alpha
,
2365 pad_inline
? 4 : dst_channels
);
2368 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2369 row_type
, dst_type
, dst
, src_count
);
2371 /* Split the blend rows back to memory rows */
2372 if (dst_count
> src_count
) {
2373 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2375 if (src_count
== 1) {
2376 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2377 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2379 row_type
.length
/= 2;
2383 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2384 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2385 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2386 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2388 row_type
.length
/= 2;
2393 * Store blend result to memory
2396 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2397 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2400 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2401 dst
, dst_type
, dst_count
, dst_alignment
);
2404 if (have_smallfloat_format(dst_type
, out_format
)) {
2405 lp_build_fpstate_set(gallivm
, fpstate
);
2409 lp_build_mask_end(&mask_ctx
);
2415 * Generate the runtime callable function for the whole fragment pipeline.
2416 * Note that the function which we generate operates on a block of 16
2417 * pixels at at time. The block contains 2x2 quads. Each quad contains
2421 generate_fragment(struct llvmpipe_context
*lp
,
2422 struct lp_fragment_shader
*shader
,
2423 struct lp_fragment_shader_variant
*variant
,
2424 unsigned partial_mask
)
2426 struct gallivm_state
*gallivm
= variant
->gallivm
;
2427 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2428 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2430 struct lp_type fs_type
;
2431 struct lp_type blend_type
;
2432 LLVMTypeRef fs_elem_type
;
2433 LLVMTypeRef blend_vec_type
;
2434 LLVMTypeRef arg_types
[13];
2435 LLVMTypeRef func_type
;
2436 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2437 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2438 LLVMValueRef context_ptr
;
2441 LLVMValueRef a0_ptr
;
2442 LLVMValueRef dadx_ptr
;
2443 LLVMValueRef dady_ptr
;
2444 LLVMValueRef color_ptr_ptr
;
2445 LLVMValueRef stride_ptr
;
2446 LLVMValueRef depth_ptr
;
2447 LLVMValueRef depth_stride
;
2448 LLVMValueRef mask_input
;
2449 LLVMValueRef thread_data_ptr
;
2450 LLVMBasicBlockRef block
;
2451 LLVMBuilderRef builder
;
2452 struct lp_build_sampler_soa
*sampler
;
2453 struct lp_build_interp_soa_context interp
;
2454 LLVMValueRef fs_mask
[16 / 4];
2455 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2456 LLVMValueRef function
;
2457 LLVMValueRef facing
;
2462 boolean cbuf0_write_all
;
2463 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2464 util_blend_state_is_dual(&key
->blend
, 0);
2466 assert(lp_native_vector_width
/ 32 >= 4);
2468 /* Adjust color input interpolation according to flatshade state:
2470 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2471 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2472 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2474 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2476 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2480 /* check if writes to cbuf[0] are to be copied to all cbufs */
2482 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2484 /* TODO: actually pick these based on the fs and color buffer
2485 * characteristics. */
2487 memset(&fs_type
, 0, sizeof fs_type
);
2488 fs_type
.floating
= TRUE
; /* floating point values */
2489 fs_type
.sign
= TRUE
; /* values are signed */
2490 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2491 fs_type
.width
= 32; /* 32-bit float */
2492 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2494 memset(&blend_type
, 0, sizeof blend_type
);
2495 blend_type
.floating
= FALSE
; /* values are integers */
2496 blend_type
.sign
= FALSE
; /* values are unsigned */
2497 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2498 blend_type
.width
= 8; /* 8-bit ubyte values */
2499 blend_type
.length
= 16; /* 16 elements per vector */
2502 * Generate the function prototype. Any change here must be reflected in
2503 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2506 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2508 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2510 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2511 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2513 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2514 arg_types
[1] = int32_type
; /* x */
2515 arg_types
[2] = int32_type
; /* y */
2516 arg_types
[3] = int32_type
; /* facing */
2517 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2518 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2519 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2520 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2521 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2522 arg_types
[9] = int32_type
; /* mask_input */
2523 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2524 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2525 arg_types
[12] = int32_type
; /* depth_stride */
2527 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2528 arg_types
, ARRAY_SIZE(arg_types
), 0);
2530 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2531 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2533 variant
->function
[partial_mask
] = function
;
2535 /* XXX: need to propagate noalias down into color param now we are
2536 * passing a pointer-to-pointer?
2538 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2539 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2540 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2542 context_ptr
= LLVMGetParam(function
, 0);
2543 x
= LLVMGetParam(function
, 1);
2544 y
= LLVMGetParam(function
, 2);
2545 facing
= LLVMGetParam(function
, 3);
2546 a0_ptr
= LLVMGetParam(function
, 4);
2547 dadx_ptr
= LLVMGetParam(function
, 5);
2548 dady_ptr
= LLVMGetParam(function
, 6);
2549 color_ptr_ptr
= LLVMGetParam(function
, 7);
2550 depth_ptr
= LLVMGetParam(function
, 8);
2551 mask_input
= LLVMGetParam(function
, 9);
2552 thread_data_ptr
= LLVMGetParam(function
, 10);
2553 stride_ptr
= LLVMGetParam(function
, 11);
2554 depth_stride
= LLVMGetParam(function
, 12);
2556 lp_build_name(context_ptr
, "context");
2557 lp_build_name(x
, "x");
2558 lp_build_name(y
, "y");
2559 lp_build_name(a0_ptr
, "a0");
2560 lp_build_name(dadx_ptr
, "dadx");
2561 lp_build_name(dady_ptr
, "dady");
2562 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2563 lp_build_name(depth_ptr
, "depth");
2564 lp_build_name(mask_input
, "mask_input");
2565 lp_build_name(thread_data_ptr
, "thread_data");
2566 lp_build_name(stride_ptr
, "stride_ptr");
2567 lp_build_name(depth_stride
, "depth_stride");
2573 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2574 builder
= gallivm
->builder
;
2576 LLVMPositionBuilderAtEnd(builder
, block
);
2579 * Must not count ps invocations if there's a null shader.
2580 * (It would be ok to count with null shader if there's d/s tests,
2581 * but only if there's d/s buffers too, which is different
2582 * to implicit rasterization disable which must not depend
2583 * on the d/s buffers.)
2584 * Could use popcount on mask, but pixel accuracy is not required.
2585 * Could disable if there's no stats query, but maybe not worth it.
2587 if (shader
->info
.base
.num_instructions
> 1) {
2588 LLVMValueRef invocs
, val
;
2589 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2590 val
= LLVMBuildLoad(builder
, invocs
, "");
2591 val
= LLVMBuildAdd(builder
, val
,
2592 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2594 LLVMBuildStore(builder
, val
, invocs
);
2597 /* code generated texture sampling */
2598 sampler
= lp_llvm_sampler_soa_create(key
->state
);
2600 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2601 /* for 1d resources only run "upper half" of stamp */
2602 if (key
->resource_1d
)
2606 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2607 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2608 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2609 num_loop
, "mask_store");
2610 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2611 boolean pixel_center_integer
=
2612 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2615 * The shader input interpolation info is not explicitely baked in the
2616 * shader key, but everything it derives from (TGSI, and flatshade) is
2617 * already included in the shader key.
2619 lp_build_interp_soa_init(&interp
,
2621 shader
->info
.base
.num_inputs
,
2623 pixel_center_integer
,
2626 a0_ptr
, dadx_ptr
, dady_ptr
,
2629 for (i
= 0; i
< num_fs
; i
++) {
2631 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2632 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2633 &indexi
, 1, "mask_ptr");
2636 mask
= generate_quad_mask(gallivm
, fs_type
,
2637 i
*fs_type
.length
/4, mask_input
);
2640 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2642 LLVMBuildStore(builder
, mask
, mask_ptr
);
2645 generate_fs_loop(gallivm
,
2653 mask_store
, /* output */
2660 for (i
= 0; i
< num_fs
; i
++) {
2661 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2662 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2664 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2665 /* This is fucked up need to reorganize things */
2666 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2667 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2668 ptr
= LLVMBuildGEP(builder
,
2669 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2671 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2674 if (dual_source_blend
) {
2675 /* only support one dual source blend target hence always use output 1 */
2676 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2677 ptr
= LLVMBuildGEP(builder
,
2678 color_store
[1][chan
],
2680 fs_out_color
[1][chan
][i
] = ptr
;
2686 sampler
->destroy(sampler
);
2688 /* Loop over color outputs / color buffers to do blending.
2690 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2691 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2692 LLVMValueRef color_ptr
;
2693 LLVMValueRef stride
;
2694 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2696 boolean do_branch
= ((key
->depth
.enabled
2697 || key
->stencil
[0].enabled
2698 || key
->alpha
.enabled
)
2699 && !shader
->info
.base
.uses_kill
);
2701 color_ptr
= LLVMBuildLoad(builder
,
2702 LLVMBuildGEP(builder
, color_ptr_ptr
,
2706 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2708 stride
= LLVMBuildLoad(builder
,
2709 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2712 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2713 key
->cbuf_format
[cbuf
],
2714 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2715 context_ptr
, color_ptr
, stride
,
2716 partial_mask
, do_branch
);
2720 LLVMBuildRetVoid(builder
);
2722 gallivm_verify_function(gallivm
, function
);
2727 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2731 debug_printf("fs variant %p:\n", (void *) key
);
2733 if (key
->flatshade
) {
2734 debug_printf("flatshade = 1\n");
2736 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2737 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2739 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2740 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2742 if (key
->depth
.enabled
) {
2743 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2744 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2747 for (i
= 0; i
< 2; ++i
) {
2748 if (key
->stencil
[i
].enabled
) {
2749 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2750 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2751 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2752 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2753 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2754 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2758 if (key
->alpha
.enabled
) {
2759 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2762 if (key
->occlusion_count
) {
2763 debug_printf("occlusion_count = 1\n");
2766 if (key
->blend
.logicop_enable
) {
2767 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2769 else if (key
->blend
.rt
[0].blend_enable
) {
2770 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2771 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2772 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2773 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2774 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2775 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2777 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2778 if (key
->blend
.alpha_to_coverage
) {
2779 debug_printf("blend.alpha_to_coverage is enabled\n");
2781 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2782 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2783 debug_printf("sampler[%u] = \n", i
);
2784 debug_printf(" .wrap = %s %s %s\n",
2785 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2786 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2787 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2788 debug_printf(" .min_img_filter = %s\n",
2789 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2790 debug_printf(" .min_mip_filter = %s\n",
2791 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2792 debug_printf(" .mag_img_filter = %s\n",
2793 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2794 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2795 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
2796 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2797 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2798 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2799 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2800 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2802 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2803 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2804 debug_printf("texture[%u] = \n", i
);
2805 debug_printf(" .format = %s\n",
2806 util_format_name(texture
->format
));
2807 debug_printf(" .target = %s\n",
2808 util_str_tex_target(texture
->target
, TRUE
));
2809 debug_printf(" .level_zero_only = %u\n",
2810 texture
->level_zero_only
);
2811 debug_printf(" .pot = %u %u %u\n",
2813 texture
->pot_height
,
2814 texture
->pot_depth
);
2820 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2822 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2823 variant
->shader
->no
, variant
->no
);
2824 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2825 dump_fs_variant_key(&variant
->key
);
2826 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2832 * Generate a new fragment shader variant from the shader code and
2833 * other state indicated by the key.
2835 static struct lp_fragment_shader_variant
*
2836 generate_variant(struct llvmpipe_context
*lp
,
2837 struct lp_fragment_shader
*shader
,
2838 const struct lp_fragment_shader_variant_key
*key
)
2840 struct lp_fragment_shader_variant
*variant
;
2841 const struct util_format_description
*cbuf0_format_desc
= NULL
;
2842 boolean fullcolormask
;
2843 char module_name
[64];
2845 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2849 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2850 shader
->no
, shader
->variants_created
);
2852 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
2853 if (!variant
->gallivm
) {
2858 variant
->shader
= shader
;
2859 variant
->list_item_global
.base
= variant
;
2860 variant
->list_item_local
.base
= variant
;
2861 variant
->no
= shader
->variants_created
++;
2863 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2866 * Determine whether we are touching all channels in the color buffer.
2868 fullcolormask
= FALSE
;
2869 if (key
->nr_cbufs
== 1) {
2870 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2871 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2875 !key
->blend
.logicop_enable
&&
2876 !key
->blend
.rt
[0].blend_enable
&&
2878 !key
->stencil
[0].enabled
&&
2879 !key
->alpha
.enabled
&&
2880 !key
->blend
.alpha_to_coverage
&&
2881 !key
->depth
.enabled
&&
2882 !shader
->info
.base
.uses_kill
&&
2883 !shader
->info
.base
.writes_samplemask
2886 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2887 lp_debug_fs_variant(variant
);
2890 lp_jit_init_types(variant
);
2892 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2893 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2895 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2896 if (variant
->opaque
) {
2897 /* Specialized shader, which doesn't need to read the color buffer. */
2898 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2903 * Compile everything
2906 gallivm_compile_module(variant
->gallivm
);
2908 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2910 if (variant
->function
[RAST_EDGE_TEST
]) {
2911 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2912 gallivm_jit_function(variant
->gallivm
,
2913 variant
->function
[RAST_EDGE_TEST
]);
2916 if (variant
->function
[RAST_WHOLE
]) {
2917 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2918 gallivm_jit_function(variant
->gallivm
,
2919 variant
->function
[RAST_WHOLE
]);
2920 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2921 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2924 gallivm_free_ir(variant
->gallivm
);
2931 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2932 const struct pipe_shader_state
*templ
)
2934 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2935 struct lp_fragment_shader
*shader
;
2937 int nr_sampler_views
;
2940 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2944 shader
->no
= fs_no
++;
2945 make_empty_list(&shader
->variants
);
2947 /* get/save the summary info for this shader */
2948 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2950 /* we need to keep a local copy of the tokens */
2951 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2953 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2954 if (shader
->draw_data
== NULL
) {
2955 FREE((void *) shader
->base
.tokens
);
2960 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2961 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2963 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2964 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2966 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2967 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2968 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2970 switch (shader
->info
.base
.input_interpolate
[i
]) {
2971 case TGSI_INTERPOLATE_CONSTANT
:
2972 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2974 case TGSI_INTERPOLATE_LINEAR
:
2975 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2977 case TGSI_INTERPOLATE_PERSPECTIVE
:
2978 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2980 case TGSI_INTERPOLATE_COLOR
:
2981 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2988 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2989 case TGSI_SEMANTIC_FACE
:
2990 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2992 case TGSI_SEMANTIC_POSITION
:
2993 /* Position was already emitted above
2995 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2996 shader
->inputs
[i
].src_index
= 0;
3000 /* XXX this is a completely pointless index map... */
3001 shader
->inputs
[i
].src_index
= i
+1;
3004 if (LP_DEBUG
& DEBUG_TGSI
) {
3006 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3007 shader
->no
, (void *) shader
);
3008 tgsi_dump(templ
->tokens
, 0);
3009 debug_printf("usage masks:\n");
3010 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3011 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3012 debug_printf(" IN[%u].%s%s%s%s\n",
3014 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3015 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3016 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3017 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3027 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3029 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3031 if (llvmpipe
->fs
== fs
)
3034 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
3036 draw_bind_fragment_shader(llvmpipe
->draw
,
3037 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
3039 llvmpipe
->dirty
|= LP_NEW_FS
;
3044 * Remove shader variant from two lists: the shader's variant list
3045 * and the context's variant list.
3048 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3049 struct lp_fragment_shader_variant
*variant
)
3051 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3052 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3053 "v total cached %u inst %u total inst %u\n",
3054 variant
->shader
->no
, variant
->no
,
3055 variant
->shader
->variants_created
,
3056 variant
->shader
->variants_cached
,
3057 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3060 gallivm_destroy(variant
->gallivm
);
3062 /* remove from shader's list */
3063 remove_from_list(&variant
->list_item_local
);
3064 variant
->shader
->variants_cached
--;
3066 /* remove from context's list */
3067 remove_from_list(&variant
->list_item_global
);
3068 lp
->nr_fs_variants
--;
3069 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3076 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3078 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3079 struct lp_fragment_shader
*shader
= fs
;
3080 struct lp_fs_variant_list_item
*li
;
3082 assert(fs
!= llvmpipe
->fs
);
3085 * XXX: we need to flush the context until we have some sort of reference
3086 * counting in fragment shaders as they may still be binned
3087 * Flushing alone might not sufficient we need to wait on it too.
3089 llvmpipe_finish(pipe
, __FUNCTION__
);
3091 /* Delete all the variants */
3092 li
= first_elem(&shader
->variants
);
3093 while(!at_end(&shader
->variants
, li
)) {
3094 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3095 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3099 /* Delete draw module's data */
3100 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3102 assert(shader
->variants_cached
== 0);
3103 FREE((void *) shader
->base
.tokens
);
3110 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3111 enum pipe_shader_type shader
, uint index
,
3112 const struct pipe_constant_buffer
*cb
)
3114 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3115 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3117 assert(shader
< PIPE_SHADER_TYPES
);
3118 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3120 /* note: reference counting */
3121 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3124 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3125 debug_printf("Illegal set constant without bind flag\n");
3126 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3130 if (shader
== PIPE_SHADER_VERTEX
||
3131 shader
== PIPE_SHADER_GEOMETRY
) {
3132 /* Pass the constants to the 'draw' module */
3133 const unsigned size
= cb
? cb
->buffer_size
: 0;
3137 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3139 else if (cb
&& cb
->user_buffer
) {
3140 data
= (ubyte
*) cb
->user_buffer
;
3147 data
+= cb
->buffer_offset
;
3149 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3153 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3156 if (cb
&& cb
->user_buffer
) {
3157 pipe_resource_reference(&constants
, NULL
);
3162 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3163 enum pipe_shader_type shader
, unsigned start_slot
,
3164 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3165 unsigned writable_bitmask
)
3167 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3169 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3170 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3172 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3174 if (shader
== PIPE_SHADER_VERTEX
||
3175 shader
== PIPE_SHADER_GEOMETRY
) {
3176 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3177 const ubyte
*data
= NULL
;
3178 if (buffer
&& buffer
->buffer
)
3179 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3181 data
+= buffer
->buffer_offset
;
3182 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3184 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3185 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3191 * Return the blend factor equivalent to a destination alpha of one.
3193 static inline unsigned
3194 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3197 case PIPE_BLENDFACTOR_DST_ALPHA
:
3198 return PIPE_BLENDFACTOR_ONE
;
3199 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3200 return PIPE_BLENDFACTOR_ZERO
;
3201 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3203 return PIPE_BLENDFACTOR_ZERO
;
3205 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3213 * We need to generate several variants of the fragment pipeline to match
3214 * all the combinations of the contributing state atoms.
3216 * TODO: there is actually no reason to tie this to context state -- the
3217 * generated code could be cached globally in the screen.
3220 make_variant_key(struct llvmpipe_context
*lp
,
3221 struct lp_fragment_shader
*shader
,
3222 struct lp_fragment_shader_variant_key
*key
)
3226 memset(key
, 0, shader
->variant_key_size
);
3228 if (lp
->framebuffer
.zsbuf
) {
3229 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3230 const struct util_format_description
*zsbuf_desc
=
3231 util_format_description(zsbuf_format
);
3233 if (lp
->depth_stencil
->depth
.enabled
&&
3234 util_format_has_depth(zsbuf_desc
)) {
3235 key
->zsbuf_format
= zsbuf_format
;
3236 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3238 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3239 util_format_has_stencil(zsbuf_desc
)) {
3240 key
->zsbuf_format
= zsbuf_format
;
3241 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3243 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3244 key
->resource_1d
= TRUE
;
3249 * Propagate the depth clamp setting from the rasterizer state.
3250 * depth_clip == 0 implies depth clamping is enabled.
3252 * When clip_halfz is enabled, then always clamp the depth values.
3254 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3255 * clamp is always active in d3d10, regardless if depth clip is
3257 * (GL has an always-on [0,1] clamp on fs depth output instead
3258 * to ensure the depth values stay in range. Doesn't look like
3259 * we do that, though...)
3261 if (lp
->rasterizer
->clip_halfz
) {
3262 key
->depth_clamp
= 1;
3264 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3267 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3268 if (!lp
->framebuffer
.nr_cbufs
||
3269 !lp
->framebuffer
.cbufs
[0] ||
3270 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3271 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3273 if(key
->alpha
.enabled
)
3274 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3275 /* alpha.ref_value is passed in jit_context */
3277 key
->flatshade
= lp
->rasterizer
->flatshade
;
3278 if (lp
->active_occlusion_queries
) {
3279 key
->occlusion_count
= TRUE
;
3282 if (lp
->framebuffer
.nr_cbufs
) {
3283 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3286 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3288 if (!key
->blend
.independent_blend_enable
) {
3289 /* we always need independent blend otherwise the fixups below won't work */
3290 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3291 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3293 key
->blend
.independent_blend_enable
= 1;
3296 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3297 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3299 if (lp
->framebuffer
.cbufs
[i
]) {
3300 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3301 const struct util_format_description
*format_desc
;
3303 key
->cbuf_format
[i
] = format
;
3306 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3307 * mixing of 2d textures with height 1 and 1d textures, so make sure
3308 * we pick 1d if any cbuf or zsbuf is 1d.
3310 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3311 key
->resource_1d
= TRUE
;
3314 format_desc
= util_format_description(format
);
3315 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3316 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3319 * Mask out color channels not present in the color buffer.
3321 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3324 * Disable blend for integer formats.
3326 if (util_format_is_pure_integer(format
)) {
3327 blend_rt
->blend_enable
= 0;
3331 * Our swizzled render tiles always have an alpha channel, but the
3332 * linear render target format often does not, so force here the dst
3335 * This is not a mere optimization. Wrong results will be produced if
3336 * the dst alpha is used, the dst format does not have alpha, and the
3337 * previous rendering was not flushed from the swizzled to linear
3338 * buffer. For example, NonPowTwo DCT.
3340 * TODO: This should be generalized to all channels for better
3341 * performance, but only alpha causes correctness issues.
3343 * Also, force rgb/alpha func/factors match, to make AoS blending
3346 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3347 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3348 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3349 boolean clamped_zero
= !util_format_is_float(format
) &&
3350 !util_format_is_snorm(format
);
3351 blend_rt
->rgb_src_factor
=
3352 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3353 blend_rt
->rgb_dst_factor
=
3354 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3355 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3356 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3357 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3361 /* no color buffer for this fragment output */
3362 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3363 blend_rt
->colormask
= 0x0;
3364 blend_rt
->blend_enable
= 0;
3368 /* This value will be the same for all the variants of a given shader:
3370 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3372 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3373 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3374 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
3375 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3380 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3381 * are dx10-style? Can't really have mixed opcodes, at least not
3382 * if we want to skip the holes here (without rescanning tgsi).
3384 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3385 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3386 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3388 * Note sview may exceed what's representable by file_mask.
3389 * This will still work, the only downside is that not actually
3390 * used views may be included in the shader key.
3392 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3393 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3394 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3399 key
->nr_sampler_views
= key
->nr_samplers
;
3400 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3401 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3402 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3403 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3412 * Update fragment shader state. This is called just prior to drawing
3413 * something when some fragment-related state has changed.
3416 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3418 struct lp_fragment_shader
*shader
= lp
->fs
;
3419 struct lp_fragment_shader_variant_key key
;
3420 struct lp_fragment_shader_variant
*variant
= NULL
;
3421 struct lp_fs_variant_list_item
*li
;
3423 make_variant_key(lp
, shader
, &key
);
3425 /* Search the variants for one which matches the key */
3426 li
= first_elem(&shader
->variants
);
3427 while(!at_end(&shader
->variants
, li
)) {
3428 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3436 /* Move this variant to the head of the list to implement LRU
3437 * deletion of shader's when we have too many.
3439 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3442 /* variant not found, create it now */
3445 unsigned variants_to_cull
;
3447 if (LP_DEBUG
& DEBUG_FS
) {
3448 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3451 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3454 /* First, check if we've exceeded the max number of shader variants.
3455 * If so, free 6.25% of them (the least recently used ones).
3457 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3459 if (variants_to_cull
||
3460 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3461 struct pipe_context
*pipe
= &lp
->pipe
;
3463 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3464 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3465 "\t%u instrs,\t%u instrs/variant\n",
3466 shader
->variants_cached
,
3467 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3468 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3472 * XXX: we need to flush the context until we have some sort of
3473 * reference counting in fragment shaders as they may still be binned
3474 * Flushing alone might not be sufficient we need to wait on it too.
3476 llvmpipe_finish(pipe
, __FUNCTION__
);
3479 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3480 * number of shader variants (potentially all of them) could be
3481 * pending for destruction on flush.
3484 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3485 struct lp_fs_variant_list_item
*item
;
3486 if (is_empty_list(&lp
->fs_variants_list
)) {
3489 item
= last_elem(&lp
->fs_variants_list
);
3492 llvmpipe_remove_shader_variant(lp
, item
->base
);
3497 * Generate the new variant.
3500 variant
= generate_variant(lp
, shader
, &key
);
3503 LP_COUNT_ADD(llvm_compile_time
, dt
);
3504 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3506 /* Put the new variant into the list */
3508 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3509 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3510 lp
->nr_fs_variants
++;
3511 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3512 shader
->variants_cached
++;
3516 /* Bind this variant */
3517 lp_setup_set_fs_variant(lp
->setup
, variant
);
3525 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3527 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3528 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3529 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3531 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3533 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;