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
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
350 if (key
->alpha
.enabled
||
351 key
->blend
.alpha_to_coverage
||
352 shader
->info
.base
.uses_kill
||
353 shader
->info
.base
.writes_samplemask
) {
354 /* With alpha test and kill, can do the depth test early
355 * and hopefully eliminate some quads. But need to do a
356 * special deferred depth write once the final mask value
357 * is known. This only works though if there's either no
358 * stencil test or the stencil value isn't written.
360 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
361 (key
->stencil
[1].enabled
&&
362 key
->stencil
[1].writemask
)))
363 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
365 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
368 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
371 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
374 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
375 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
376 (key
->stencil
[1].enabled
&&
377 key
->stencil
[1].writemask
))))
378 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
384 vec_type
= lp_build_vec_type(gallivm
, type
);
385 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
387 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
388 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
389 /* convert scalar stencil refs into vectors */
390 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
391 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
393 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
394 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
396 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
397 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
399 lp_build_for_loop_begin(&loop_state
, gallivm
,
400 lp_build_const_int32(gallivm
, 0),
403 lp_build_const_int32(gallivm
, 1));
405 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
406 &loop_state
.counter
, 1, "mask_ptr");
407 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
409 memset(outputs
, 0, sizeof outputs
);
411 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
412 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
413 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
414 lp_build_vec_type(gallivm
,
419 if (dual_source_blend
) {
420 assert(key
->nr_cbufs
<= 1);
421 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
422 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
423 lp_build_vec_type(gallivm
,
430 /* 'mask' will control execution based on quad's pixel alive/killed state */
431 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
433 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
434 lp_build_mask_check(&mask
);
436 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
439 if (depth_mode
& EARLY_DEPTH_TEST
) {
441 * Clamp according to ARB_depth_clamp semantics.
443 if (key
->depth_clamp
) {
444 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
447 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
448 zs_format_desc
, key
->resource_1d
,
449 depth_ptr
, depth_stride
,
450 &z_fb
, &s_fb
, loop_state
.counter
);
451 lp_build_depth_stencil_test(gallivm
,
463 if (depth_mode
& EARLY_DEPTH_WRITE
) {
464 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
465 zs_format_desc
, key
->resource_1d
,
466 NULL
, NULL
, NULL
, loop_state
.counter
,
467 depth_ptr
, depth_stride
,
471 * Note mask check if stencil is enabled must be after ds write not after
472 * stencil test otherwise new stencil values may not get written if all
473 * fragments got killed by depth/stencil test.
475 if (!simple_shader
&& key
->stencil
[0].enabled
)
476 lp_build_mask_check(&mask
);
479 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
481 struct lp_build_tgsi_params params
;
482 memset(¶ms
, 0, sizeof(params
));
486 params
.consts_ptr
= consts_ptr
;
487 params
.const_sizes_ptr
= num_consts_ptr
;
488 params
.system_values
= &system_values
;
489 params
.inputs
= interp
->inputs
;
490 params
.context_ptr
= context_ptr
;
491 params
.thread_data_ptr
= thread_data_ptr
;
492 params
.sampler
= sampler
;
493 params
.info
= &shader
->info
.base
;
494 params
.ssbo_ptr
= ssbo_ptr
;
495 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
497 /* Build the actual shader */
498 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
502 if (key
->alpha
.enabled
) {
503 int color0
= find_output_by_semantic(&shader
->info
.base
,
507 if (color0
!= -1 && outputs
[color0
][3]) {
508 const struct util_format_description
*cbuf_format_desc
;
509 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
510 LLVMValueRef alpha_ref_value
;
512 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
513 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
515 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
517 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
518 &mask
, alpha
, alpha_ref_value
,
519 (depth_mode
& LATE_DEPTH_TEST
) != 0);
523 /* Emulate Alpha to Coverage with Alpha test */
524 if (key
->blend
.alpha_to_coverage
) {
525 int color0
= find_output_by_semantic(&shader
->info
.base
,
529 if (color0
!= -1 && outputs
[color0
][3]) {
530 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
532 lp_build_alpha_to_coverage(gallivm
, type
,
534 (depth_mode
& LATE_DEPTH_TEST
) != 0);
538 if (shader
->info
.base
.writes_samplemask
) {
539 int smaski
= find_output_by_semantic(&shader
->info
.base
,
540 TGSI_SEMANTIC_SAMPLEMASK
,
543 struct lp_build_context smask_bld
;
544 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
547 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
549 * Pixel is alive according to the first sample in the mask.
551 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
552 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
553 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
554 lp_build_mask_update(&mask
, smask
);
558 if (depth_mode
& LATE_DEPTH_TEST
) {
559 int pos0
= find_output_by_semantic(&shader
->info
.base
,
560 TGSI_SEMANTIC_POSITION
,
562 int s_out
= find_output_by_semantic(&shader
->info
.base
,
563 TGSI_SEMANTIC_STENCIL
,
565 if (pos0
!= -1 && outputs
[pos0
][2]) {
566 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
569 * Clamp according to ARB_depth_clamp semantics.
571 if (key
->depth_clamp
) {
572 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
576 if (s_out
!= -1 && outputs
[s_out
][1]) {
577 /* there's only one value, and spec says to discard additional bits */
578 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
579 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
580 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
581 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
582 stencil_refs
[1] = stencil_refs
[0];
585 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
586 zs_format_desc
, key
->resource_1d
,
587 depth_ptr
, depth_stride
,
588 &z_fb
, &s_fb
, loop_state
.counter
);
590 lp_build_depth_stencil_test(gallivm
,
602 if (depth_mode
& LATE_DEPTH_WRITE
) {
603 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
604 zs_format_desc
, key
->resource_1d
,
605 NULL
, NULL
, NULL
, loop_state
.counter
,
606 depth_ptr
, depth_stride
,
610 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
611 (depth_mode
& LATE_DEPTH_WRITE
))
613 /* Need to apply a reduced mask to the depth write. Reload the
614 * depth value, update from zs_value with the new mask value and
617 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
618 zs_format_desc
, key
->resource_1d
,
619 &mask
, z_fb
, s_fb
, loop_state
.counter
,
620 depth_ptr
, depth_stride
,
626 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
628 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
629 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
630 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
632 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
633 if(outputs
[attrib
][chan
]) {
634 /* XXX: just initialize outputs to point at colors[] and
637 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
638 LLVMValueRef color_ptr
;
639 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
640 &loop_state
.counter
, 1, "");
641 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
642 LLVMBuildStore(builder
, out
, color_ptr
);
648 if (key
->occlusion_count
) {
649 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
650 lp_build_name(counter
, "counter");
651 lp_build_occlusion_count(gallivm
, type
,
652 lp_build_mask_value(&mask
), counter
);
655 mask_val
= lp_build_mask_end(&mask
);
656 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
657 lp_build_for_loop_end(&loop_state
);
662 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
664 * Fragment Shader outputs pixels in small 2x2 blocks
665 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
667 * However in memory pixels are stored in rows
668 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
670 * @param type fragment shader type (4x or 8x float)
671 * @param num_fs number of fs_src
672 * @param is_1d whether we're outputting to a 1d resource
673 * @param dst_channels number of output channels
674 * @param fs_src output from fragment shader
675 * @param dst pointer to store result
676 * @param pad_inline is channel padding inline or at end of row
677 * @return the number of dsts
680 generate_fs_twiddle(struct gallivm_state
*gallivm
,
683 unsigned dst_channels
,
684 LLVMValueRef fs_src
[][4],
688 LLVMValueRef src
[16];
694 unsigned pixels
= type
.length
/ 4;
695 unsigned reorder_group
;
696 unsigned src_channels
;
700 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
701 src_count
= num_fs
* src_channels
;
703 assert(pixels
== 2 || pixels
== 1);
704 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
707 * Transpose from SoA -> AoS
709 for (i
= 0; i
< num_fs
; ++i
) {
710 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
714 * Pick transformation options
721 if (dst_channels
== 1) {
727 } else if (dst_channels
== 2) {
731 } else if (dst_channels
> 2) {
738 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
744 * Split the src in half
747 for (i
= num_fs
; i
> 0; --i
) {
748 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
749 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
757 * Ensure pixels are in memory order
760 /* Twiddle pixels by reordering the array, e.g.:
762 * src_count = 8 -> 0 2 1 3 4 6 5 7
763 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
765 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
767 for (i
= 0; i
< src_count
; ++i
) {
768 unsigned group
= i
/ reorder_group
;
769 unsigned block
= (group
/ 4) * 4 * reorder_group
;
770 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
773 } else if (twiddle
) {
774 /* Twiddle pixels across elements of array */
776 * XXX: we should avoid this in some cases, but would need to tell
777 * lp_build_conv to reorder (or deal with it ourselves).
779 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
782 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
786 * Moves any padding between pixels to the end
787 * e.g. RGBXRGBX -> RGBRGBXX
790 unsigned char swizzles
[16];
791 unsigned elems
= pixels
* dst_channels
;
793 for (i
= 0; i
< type
.length
; ++i
) {
795 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
797 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
800 for (i
= 0; i
< src_count
; ++i
) {
801 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
810 * Untwiddle and transpose, much like the above.
811 * However, this is after conversion, so we get packed vectors.
812 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
813 * the vectors will look like:
814 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
815 * be swizzled here). Extending to 16bit should be trivial.
816 * Should also be extended to handle twice wide vectors with AVX2...
819 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
826 struct lp_type type64
, type16
, type32
;
827 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
828 LLVMBuilderRef builder
= gallivm
->builder
;
829 LLVMValueRef tmp
[4], shuf
[8];
830 for (j
= 0; j
< 2; j
++) {
831 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
832 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
833 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
834 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
837 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
838 assert(type
.width
== 8);
839 assert(type
.length
== 16);
841 type8_t
= lp_build_vec_type(gallivm
, type
);
846 type64_t
= lp_build_vec_type(gallivm
, type64
);
851 type16_t
= lp_build_vec_type(gallivm
, type16
);
856 type32_t
= lp_build_vec_type(gallivm
, type32
);
858 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
860 if (src_count
== 1) {
861 /* transpose was no-op, just untwiddle */
862 LLVMValueRef shuf_vec
;
863 shuf_vec
= LLVMConstVector(shuf
, 8);
864 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
865 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
866 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
867 } else if (src_count
== 2) {
868 LLVMValueRef shuf_vec
;
869 shuf_vec
= LLVMConstVector(shuf
, 4);
871 for (i
= 0; i
< 2; i
++) {
872 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
873 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
874 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
877 for (j
= 0; j
< 2; j
++) {
878 LLVMValueRef lo
, hi
, lo2
, hi2
;
880 * Note that if we only really have 3 valid channels (rgb)
881 * and we don't need alpha we could substitute a undef here
882 * for the respective channel (causing llvm to drop conversion
885 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
886 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
887 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
888 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
889 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
890 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
891 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
898 * Load an unswizzled block of pixels from memory
901 load_unswizzled_block(struct gallivm_state
*gallivm
,
902 LLVMValueRef base_ptr
,
904 unsigned block_width
,
905 unsigned block_height
,
907 struct lp_type dst_type
,
909 unsigned dst_alignment
)
911 LLVMBuilderRef builder
= gallivm
->builder
;
912 unsigned row_size
= dst_count
/ block_height
;
915 /* Ensure block exactly fits into dst */
916 assert((block_width
* block_height
) % dst_count
== 0);
918 for (i
= 0; i
< dst_count
; ++i
) {
919 unsigned x
= i
% row_size
;
920 unsigned y
= i
/ row_size
;
922 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
923 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
926 LLVMValueRef dst_ptr
;
928 gep
[0] = lp_build_const_int32(gallivm
, 0);
929 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
931 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
932 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
933 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
935 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
937 LLVMSetAlignment(dst
[i
], dst_alignment
);
943 * Store an unswizzled block of pixels to memory
946 store_unswizzled_block(struct gallivm_state
*gallivm
,
947 LLVMValueRef base_ptr
,
949 unsigned block_width
,
950 unsigned block_height
,
952 struct lp_type src_type
,
954 unsigned src_alignment
)
956 LLVMBuilderRef builder
= gallivm
->builder
;
957 unsigned row_size
= src_count
/ block_height
;
960 /* Ensure src exactly fits into block */
961 assert((block_width
* block_height
) % src_count
== 0);
963 for (i
= 0; i
< src_count
; ++i
) {
964 unsigned x
= i
% row_size
;
965 unsigned y
= i
/ row_size
;
967 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
968 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
971 LLVMValueRef src_ptr
;
973 gep
[0] = lp_build_const_int32(gallivm
, 0);
974 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
976 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
977 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
978 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
980 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
982 LLVMSetAlignment(src_ptr
, src_alignment
);
988 * Checks if a format description is an arithmetic format
990 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
992 static inline boolean
993 is_arithmetic_format(const struct util_format_description
*format_desc
)
995 boolean arith
= false;
998 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
999 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1000 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1008 * Checks if this format requires special handling due to required expansion
1009 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1012 static inline boolean
1013 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1015 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1016 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1024 * Retrieves the type representing the memory layout for a format
1026 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1029 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1030 struct lp_type
* type
)
1035 if (format_expands_to_float_soa(format_desc
)) {
1036 /* just make this a uint with width of block */
1037 type
->floating
= false;
1038 type
->fixed
= false;
1041 type
->width
= format_desc
->block
.bits
;
1046 for (i
= 0; i
< 4; i
++)
1047 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1051 memset(type
, 0, sizeof(struct lp_type
));
1052 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1053 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1054 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1055 type
->norm
= format_desc
->channel
[chan
].normalized
;
1057 if (is_arithmetic_format(format_desc
)) {
1061 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1062 type
->width
+= format_desc
->channel
[i
].size
;
1065 type
->width
= format_desc
->channel
[chan
].size
;
1066 type
->length
= format_desc
->nr_channels
;
1072 * Retrieves the type for a format which is usable in the blending code.
1074 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1077 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1078 struct lp_type
* type
)
1083 if (format_expands_to_float_soa(format_desc
)) {
1084 /* always use ordinary floats for blending */
1085 type
->floating
= true;
1086 type
->fixed
= false;
1094 for (i
= 0; i
< 4; i
++)
1095 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1099 memset(type
, 0, sizeof(struct lp_type
));
1100 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1101 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1102 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1103 type
->norm
= format_desc
->channel
[chan
].normalized
;
1104 type
->width
= format_desc
->channel
[chan
].size
;
1105 type
->length
= format_desc
->nr_channels
;
1107 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1108 if (format_desc
->channel
[i
].size
> type
->width
)
1109 type
->width
= format_desc
->channel
[i
].size
;
1112 if (type
->floating
) {
1115 if (type
->width
<= 8) {
1117 } else if (type
->width
<= 16) {
1124 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1131 * Scale a normalized value from src_bits to dst_bits.
1133 * The exact calculation is
1135 * dst = iround(src * dst_mask / src_mask)
1137 * or with integer rounding
1139 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1143 * src_mask = (1 << src_bits) - 1
1144 * dst_mask = (1 << dst_bits) - 1
1146 * but we try to avoid division and multiplication through shifts.
1148 static inline LLVMValueRef
1149 scale_bits(struct gallivm_state
*gallivm
,
1153 struct lp_type src_type
)
1155 LLVMBuilderRef builder
= gallivm
->builder
;
1156 LLVMValueRef result
= src
;
1158 if (dst_bits
< src_bits
) {
1159 int delta_bits
= src_bits
- dst_bits
;
1161 if (delta_bits
<= dst_bits
) {
1163 * Approximate the rescaling with a single shift.
1165 * This gives the wrong rounding.
1168 result
= LLVMBuildLShr(builder
,
1170 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1175 * Try more accurate rescaling.
1179 * Drop the least significant bits to make space for the multiplication.
1181 * XXX: A better approach would be to use a wider integer type as intermediate. But
1182 * this is enough to convert alpha from 16bits -> 2 when rendering to
1183 * PIPE_FORMAT_R10G10B10A2_UNORM.
1185 result
= LLVMBuildLShr(builder
,
1187 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1191 result
= LLVMBuildMul(builder
,
1193 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1197 * Add a rounding term before the division.
1199 * TODO: Handle signed integers too.
1201 if (!src_type
.sign
) {
1202 result
= LLVMBuildAdd(builder
,
1204 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1209 * Approximate the division by src_mask with a src_bits shift.
1211 * Given the src has already been shifted by dst_bits, all we need
1212 * to do is to shift by the difference.
1215 result
= LLVMBuildLShr(builder
,
1217 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1221 } else if (dst_bits
> src_bits
) {
1223 int db
= dst_bits
- src_bits
;
1225 /* Shift left by difference in bits */
1226 result
= LLVMBuildShl(builder
,
1228 lp_build_const_int_vec(gallivm
, src_type
, db
),
1231 if (db
<= src_bits
) {
1232 /* Enough bits in src to fill the remainder */
1233 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1235 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1238 result
= LLVMBuildOr(builder
, result
, lower
, "");
1239 } else if (db
> src_bits
) {
1240 /* Need to repeatedly copy src bits to fill remainder in dst */
1243 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1244 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1246 result
= LLVMBuildOr(builder
,
1248 LLVMBuildLShr(builder
, result
, shuv
, ""),
1258 * If RT is a smallfloat (needing denorms) format
1261 have_smallfloat_format(struct lp_type dst_type
,
1262 enum pipe_format format
)
1264 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1265 /* due to format handling hacks this format doesn't have floating set
1266 * here (and actually has width set to 32 too) so special case this. */
1267 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1272 * Convert from memory format to blending format
1274 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1277 convert_to_blend_type(struct gallivm_state
*gallivm
,
1278 unsigned block_size
,
1279 const struct util_format_description
*src_fmt
,
1280 struct lp_type src_type
,
1281 struct lp_type dst_type
,
1282 LLVMValueRef
* src
, // and dst
1285 LLVMValueRef
*dst
= src
;
1286 LLVMBuilderRef builder
= gallivm
->builder
;
1287 struct lp_type blend_type
;
1288 struct lp_type mem_type
;
1290 unsigned pixels
= block_size
/ num_srcs
;
1294 * full custom path for packed floats and srgb formats - none of the later
1295 * functions would do anything useful, and given the lp_type representation they
1296 * can't be fixed. Should really have some SoA blend path for these kind of
1297 * formats rather than hacking them in here.
1299 if (format_expands_to_float_soa(src_fmt
)) {
1300 LLVMValueRef tmpsrc
[4];
1302 * This is pretty suboptimal for this case blending in SoA would be much
1303 * better, since conversion gets us SoA values so need to convert back.
1305 assert(src_type
.width
== 32 || src_type
.width
== 16);
1306 assert(dst_type
.floating
);
1307 assert(dst_type
.width
== 32);
1308 assert(dst_type
.length
% 4 == 0);
1309 assert(num_srcs
% 4 == 0);
1311 if (src_type
.width
== 16) {
1312 /* expand 4x16bit values to 4x32bit */
1313 struct lp_type type32x4
= src_type
;
1314 LLVMTypeRef ltype32x4
;
1315 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1316 type32x4
.width
= 32;
1317 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1318 for (i
= 0; i
< num_fetch
; i
++) {
1319 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1321 src_type
.width
= 32;
1323 for (i
= 0; i
< 4; i
++) {
1326 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1327 LLVMValueRef tmpsoa
[4];
1328 LLVMValueRef tmps
= tmpsrc
[i
];
1329 if (dst_type
.length
== 8) {
1330 LLVMValueRef shuffles
[8];
1332 /* fetch was 4 values but need 8-wide output values */
1333 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1335 * for 8-wide aos transpose would give us wrong order not matching
1336 * incoming converted fs values and mask. ARGH.
1338 for (j
= 0; j
< 4; j
++) {
1339 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1340 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1342 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1343 LLVMConstVector(shuffles
, 8), "");
1345 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1346 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1349 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1351 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1356 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1357 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1359 /* Is the format arithmetic */
1360 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1361 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1363 /* Pad if necessary */
1364 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1365 for (i
= 0; i
< num_srcs
; ++i
) {
1366 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1369 src_type
.length
= dst_type
.length
;
1372 /* Special case for half-floats */
1373 if (mem_type
.width
== 16 && mem_type
.floating
) {
1374 assert(blend_type
.width
== 32 && blend_type
.floating
);
1375 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1383 src_type
.width
= blend_type
.width
* blend_type
.length
;
1384 blend_type
.length
*= pixels
;
1385 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1387 for (i
= 0; i
< num_srcs
; ++i
) {
1388 LLVMValueRef chans
[4];
1389 LLVMValueRef res
= NULL
;
1391 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1393 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1395 unsigned sa
= src_fmt
->channel
[j
].shift
;
1396 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1397 unsigned from_lsb
= j
;
1399 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1402 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1404 /* Extract bits from source */
1405 chans
[j
] = LLVMBuildLShr(builder
,
1407 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1410 chans
[j
] = LLVMBuildAnd(builder
,
1412 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1416 if (src_type
.norm
) {
1417 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1418 blend_type
.width
, chans
[j
], src_type
);
1421 /* Insert bits into correct position */
1422 chans
[j
] = LLVMBuildShl(builder
,
1424 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1430 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1434 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1440 * Convert from blending format to memory format
1442 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1445 convert_from_blend_type(struct gallivm_state
*gallivm
,
1446 unsigned block_size
,
1447 const struct util_format_description
*src_fmt
,
1448 struct lp_type src_type
,
1449 struct lp_type dst_type
,
1450 LLVMValueRef
* src
, // and dst
1453 LLVMValueRef
* dst
= src
;
1455 struct lp_type mem_type
;
1456 struct lp_type blend_type
;
1457 LLVMBuilderRef builder
= gallivm
->builder
;
1458 unsigned pixels
= block_size
/ num_srcs
;
1462 * full custom path for packed floats and srgb formats - none of the later
1463 * functions would do anything useful, and given the lp_type representation they
1464 * can't be fixed. Should really have some SoA blend path for these kind of
1465 * formats rather than hacking them in here.
1467 if (format_expands_to_float_soa(src_fmt
)) {
1469 * This is pretty suboptimal for this case blending in SoA would be much
1470 * better - we need to transpose the AoS values back to SoA values for
1471 * conversion/packing.
1473 assert(src_type
.floating
);
1474 assert(src_type
.width
== 32);
1475 assert(src_type
.length
% 4 == 0);
1476 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1478 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1479 LLVMValueRef tmpsoa
[4], tmpdst
;
1480 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1481 /* really really need SoA here */
1483 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1484 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1487 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1491 if (src_type
.length
== 8) {
1492 LLVMValueRef tmpaos
, shuffles
[8];
1495 * for 8-wide aos transpose has given us wrong order not matching
1496 * output order. HMPF. Also need to split the output values manually.
1498 for (j
= 0; j
< 4; j
++) {
1499 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1500 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1502 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1503 LLVMConstVector(shuffles
, 8), "");
1504 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1505 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1511 if (dst_type
.width
== 16) {
1512 struct lp_type type16x8
= dst_type
;
1513 struct lp_type type32x4
= dst_type
;
1514 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1515 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1516 type16x8
.length
= 8;
1517 type32x4
.width
= 32;
1518 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1519 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1520 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1521 /* We could do vector truncation but it doesn't generate very good code */
1522 for (i
= 0; i
< num_fetch
; i
++) {
1523 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1524 src
[i
], lp_build_zero(gallivm
, type32x4
));
1525 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1526 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1527 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1533 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1534 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1536 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1538 /* Special case for half-floats */
1539 if (mem_type
.width
== 16 && mem_type
.floating
) {
1540 int length
= dst_type
.length
;
1541 assert(blend_type
.width
== 32 && blend_type
.floating
);
1543 dst_type
.length
= src_type
.length
;
1545 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1547 dst_type
.length
= length
;
1551 /* Remove any padding */
1552 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1553 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1555 for (i
= 0; i
< num_srcs
; ++i
) {
1556 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1560 /* No bit arithmetic to do */
1565 src_type
.length
= pixels
;
1566 src_type
.width
= blend_type
.length
* blend_type
.width
;
1567 dst_type
.length
= pixels
;
1569 for (i
= 0; i
< num_srcs
; ++i
) {
1570 LLVMValueRef chans
[4];
1571 LLVMValueRef res
= NULL
;
1573 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1575 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1577 unsigned sa
= src_fmt
->channel
[j
].shift
;
1578 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1579 unsigned from_lsb
= j
;
1581 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1584 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1586 for (k
= 0; k
< blend_type
.width
; ++k
) {
1591 chans
[j
] = LLVMBuildLShr(builder
,
1593 lp_build_const_int_vec(gallivm
, src_type
,
1594 from_lsb
* blend_type
.width
),
1597 chans
[j
] = LLVMBuildAnd(builder
,
1599 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1602 /* Scale down bits */
1603 if (src_type
.norm
) {
1604 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1605 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1609 chans
[j
] = LLVMBuildShl(builder
,
1611 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1614 sa
+= src_fmt
->channel
[j
].size
;
1619 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1623 assert (dst_type
.width
!= 24);
1625 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1631 * Convert alpha to same blend type as src
1634 convert_alpha(struct gallivm_state
*gallivm
,
1635 struct lp_type row_type
,
1636 struct lp_type alpha_type
,
1637 const unsigned block_size
,
1638 const unsigned block_height
,
1639 const unsigned src_count
,
1640 const unsigned dst_channels
,
1641 const bool pad_inline
,
1642 LLVMValueRef
* src_alpha
)
1644 LLVMBuilderRef builder
= gallivm
->builder
;
1646 unsigned length
= row_type
.length
;
1647 row_type
.length
= alpha_type
.length
;
1649 /* Twiddle the alpha to match pixels */
1650 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1653 * TODO this should use single lp_build_conv call for
1654 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1656 for (i
= 0; i
< block_height
; ++i
) {
1657 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1660 alpha_type
= row_type
;
1661 row_type
.length
= length
;
1663 /* If only one channel we can only need the single alpha value per pixel */
1664 if (src_count
== 1 && dst_channels
== 1) {
1666 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1668 /* If there are more srcs than rows then we need to split alpha up */
1669 if (src_count
> block_height
) {
1670 for (i
= src_count
; i
> 0; --i
) {
1671 unsigned pixels
= block_size
/ src_count
;
1672 unsigned idx
= i
- 1;
1674 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1675 (idx
* pixels
) % 4, pixels
);
1679 /* If there is a src for each pixel broadcast the alpha across whole row */
1680 if (src_count
== block_size
) {
1681 for (i
= 0; i
< src_count
; ++i
) {
1682 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1683 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1686 unsigned pixels
= block_size
/ src_count
;
1687 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1688 unsigned alpha_span
= 1;
1689 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1691 /* Check if we need 2 src_alphas for our shuffles */
1692 if (pixels
> alpha_type
.length
) {
1696 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1697 for (j
= 0; j
< row_type
.length
; ++j
) {
1698 if (j
< pixels
* channels
) {
1699 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1701 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1705 for (i
= 0; i
< src_count
; ++i
) {
1706 unsigned idx1
= i
, idx2
= i
;
1708 if (alpha_span
> 1){
1713 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1716 LLVMConstVector(shuffles
, row_type
.length
),
1725 * Generates the blend function for unswizzled colour buffers
1726 * Also generates the read & write from colour buffer
1729 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1731 struct lp_fragment_shader_variant
*variant
,
1732 enum pipe_format out_format
,
1733 unsigned int num_fs
,
1734 struct lp_type fs_type
,
1735 LLVMValueRef
* fs_mask
,
1736 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1737 LLVMValueRef context_ptr
,
1738 LLVMValueRef color_ptr
,
1739 LLVMValueRef stride
,
1740 unsigned partial_mask
,
1743 const unsigned alpha_channel
= 3;
1744 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1745 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1746 const unsigned block_size
= block_width
* block_height
;
1747 const unsigned lp_integer_vector_width
= 128;
1749 LLVMBuilderRef builder
= gallivm
->builder
;
1750 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1751 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1752 LLVMValueRef src_alpha
[4 * 4];
1753 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1754 LLVMValueRef src_mask
[4 * 4];
1755 LLVMValueRef src
[4 * 4];
1756 LLVMValueRef src1
[4 * 4];
1757 LLVMValueRef dst
[4 * 4];
1758 LLVMValueRef blend_color
;
1759 LLVMValueRef blend_alpha
;
1760 LLVMValueRef i32_zero
;
1761 LLVMValueRef check_mask
;
1762 LLVMValueRef undef_src_val
;
1764 struct lp_build_mask_context mask_ctx
;
1765 struct lp_type mask_type
;
1766 struct lp_type blend_type
;
1767 struct lp_type row_type
;
1768 struct lp_type dst_type
;
1769 struct lp_type ls_type
;
1771 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1772 unsigned vector_width
;
1773 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1774 unsigned dst_channels
;
1779 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1781 unsigned dst_alignment
;
1783 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1784 bool has_alpha
= false;
1785 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1786 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1788 const boolean is_1d
= variant
->key
.resource_1d
;
1789 boolean twiddle_after_convert
= FALSE
;
1790 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1791 LLVMValueRef fpstate
= 0;
1793 /* Get type from output format */
1794 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1795 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1798 * Technically this code should go into lp_build_smallfloat_to_float
1799 * and lp_build_float_to_smallfloat but due to the
1800 * http://llvm.org/bugs/show_bug.cgi?id=6393
1801 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1802 * So the ordering is important here and there shouldn't be any
1803 * llvm ir instrunctions in this function before
1804 * this, otherwise half-float format conversions won't work
1805 * (again due to llvm bug #6393).
1807 if (have_smallfloat_format(dst_type
, out_format
)) {
1808 /* We need to make sure that denorms are ok for half float
1810 fpstate
= lp_build_fpstate_get(gallivm
);
1811 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1814 mask_type
= lp_int32_vec4_type();
1815 mask_type
.length
= fs_type
.length
;
1817 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1818 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1821 /* Do not bother executing code when mask is empty.. */
1823 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1825 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1826 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1829 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1830 lp_build_mask_check(&mask_ctx
);
1833 partial_mask
|= !variant
->opaque
;
1834 i32_zero
= lp_build_const_int32(gallivm
, 0);
1836 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1838 row_type
.length
= fs_type
.length
;
1839 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1841 /* Compute correct swizzle and count channels */
1842 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1845 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1846 /* Ensure channel is used */
1847 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1851 /* Ensure not already written to (happens in case with GL_ALPHA) */
1852 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1856 /* Ensure we havn't already found all channels */
1857 if (dst_channels
>= out_format_desc
->nr_channels
) {
1861 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1864 if (i
== alpha_channel
) {
1869 if (format_expands_to_float_soa(out_format_desc
)) {
1871 * the code above can't work for layout_other
1872 * for srgb it would sort of work but we short-circuit swizzles, etc.
1873 * as that is done as part of unpack / pack.
1875 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1881 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1884 /* If 3 channels then pad to include alpha for 4 element transpose */
1885 if (dst_channels
== 3) {
1886 assert (!has_alpha
);
1887 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1888 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1891 if (out_format_desc
->nr_channels
== 4) {
1894 * We use alpha from the color conversion, not separate one.
1895 * We had to include it for transpose, hence it will get converted
1896 * too (albeit when doing transpose after conversion, that would
1897 * no longer be the case necessarily).
1898 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1899 * otherwise we really have padding, not alpha, included.)
1906 * Load shader output
1908 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1909 /* Always load alpha for use in blending */
1912 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1915 alpha
= undef_src_val
;
1918 /* Load each channel */
1919 for (j
= 0; j
< dst_channels
; ++j
) {
1920 assert(swizzle
[j
] < 4);
1922 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1925 fs_src
[i
][j
] = undef_src_val
;
1929 /* If 3 channels then pad to include alpha for 4 element transpose */
1931 * XXX If we include that here maybe could actually use it instead of
1932 * separate alpha for blending?
1933 * (Difficult though we actually convert pad channels, not alpha.)
1935 if (dst_channels
== 3 && !has_alpha
) {
1936 fs_src
[i
][3] = alpha
;
1939 /* We split the row_mask and row_alpha as we want 128bit interleave */
1940 if (fs_type
.length
== 8) {
1941 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1943 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1944 src_channels
, src_channels
);
1946 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1947 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1948 src_channels
, src_channels
);
1950 src_mask
[i
] = fs_mask
[i
];
1951 src_alpha
[i
] = alpha
;
1954 if (dual_source_blend
) {
1955 /* same as above except different src/dst, skip masks and comments... */
1956 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1959 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1962 alpha
= undef_src_val
;
1965 for (j
= 0; j
< dst_channels
; ++j
) {
1966 assert(swizzle
[j
] < 4);
1968 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1971 fs_src1
[i
][j
] = undef_src_val
;
1974 if (dst_channels
== 3 && !has_alpha
) {
1975 fs_src1
[i
][3] = alpha
;
1977 if (fs_type
.length
== 8) {
1978 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1979 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1980 src_channels
, src_channels
);
1982 src1_alpha
[i
] = alpha
;
1987 if (util_format_is_pure_integer(out_format
)) {
1989 * In this case fs_type was really ints or uints disguised as floats,
1992 fs_type
.floating
= 0;
1993 fs_type
.sign
= dst_type
.sign
;
1994 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1995 for (j
= 0; j
< dst_channels
; ++j
) {
1996 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1997 lp_build_vec_type(gallivm
, fs_type
), "");
1999 if (dst_channels
== 3 && !has_alpha
) {
2000 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2001 lp_build_vec_type(gallivm
, fs_type
), "");
2007 * We actually should generally do conversion first (for non-1d cases)
2008 * when the blend format is 8 or 16 bits. The reason is obvious,
2009 * there's 2 or 4 times less vectors to deal with for the interleave...
2010 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2011 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2012 * unpack only with 128bit vectors).
2013 * Note: for 16bit sizes really need matching pack conversion code
2015 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2016 twiddle_after_convert
= TRUE
;
2020 * Pixel twiddle from fragment shader order to memory order
2022 if (!twiddle_after_convert
) {
2023 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2024 dst_channels
, fs_src
, src
, pad_inline
);
2025 if (dual_source_blend
) {
2026 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2027 fs_src1
, src1
, pad_inline
);
2030 src_count
= num_fullblock_fs
* dst_channels
;
2032 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2033 * (AVX) turn out the same later when untwiddling/transpose (albeit
2034 * for true AVX2 path untwiddle needs to be different).
2035 * For now just order by colors first (so we can use unpack later).
2037 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2038 for (i
= 0; i
< dst_channels
; i
++) {
2039 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2040 if (dual_source_blend
) {
2041 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2047 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2048 if (src_count
!= num_fullblock_fs
* src_channels
) {
2049 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2050 row_type
.length
/= ds
;
2051 fs_type
.length
= row_type
.length
;
2054 blend_type
= row_type
;
2055 mask_type
.length
= 4;
2057 /* Convert src to row_type */
2058 if (dual_source_blend
) {
2059 struct lp_type old_row_type
= row_type
;
2060 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2061 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2064 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2067 /* If the rows are not an SSE vector, combine them to become SSE size! */
2068 if ((row_type
.width
* row_type
.length
) % 128) {
2069 unsigned bits
= row_type
.width
* row_type
.length
;
2072 assert(src_count
>= (vector_width
/ bits
));
2074 dst_count
= src_count
/ (vector_width
/ bits
);
2076 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2077 if (dual_source_blend
) {
2078 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2081 row_type
.length
*= combined
;
2082 src_count
/= combined
;
2084 bits
= row_type
.width
* row_type
.length
;
2085 assert(bits
== 128 || bits
== 256);
2088 if (twiddle_after_convert
) {
2089 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2090 if (dual_source_blend
) {
2091 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2096 * Blend Colour conversion
2098 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2099 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2100 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2101 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2102 &i32_zero
, 1, ""), "");
2105 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2107 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2109 * since blending is done with floats, there was no conversion.
2110 * However, the rules according to fixed point renderbuffers still
2111 * apply, that is we must clamp inputs to 0.0/1.0.
2112 * (This would apply to separate alpha conversion too but we currently
2113 * force has_alpha to be true.)
2114 * TODO: should skip this with "fake" blend, since post-blend conversion
2115 * will clamp anyway.
2116 * TODO: could also skip this if fragment color clamping is enabled. We
2117 * don't support it natively so it gets baked into the shader however, so
2118 * can't really tell here.
2120 struct lp_build_context f32_bld
;
2121 assert(row_type
.floating
);
2122 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2123 for (i
= 0; i
< src_count
; i
++) {
2124 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2126 if (dual_source_blend
) {
2127 for (i
= 0; i
< src_count
; i
++) {
2128 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2131 /* probably can't be different than row_type but better safe than sorry... */
2132 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2133 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2137 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2139 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2140 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2142 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2143 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2145 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2146 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2152 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2154 if (src_count
< block_height
) {
2155 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2156 } else if (src_count
> block_height
) {
2157 for (i
= src_count
; i
> 0; --i
) {
2158 unsigned pixels
= block_size
/ src_count
;
2159 unsigned idx
= i
- 1;
2161 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2162 (idx
* pixels
) % 4, pixels
);
2166 assert(mask_type
.width
== 32);
2168 for (i
= 0; i
< src_count
; ++i
) {
2169 unsigned pixels
= block_size
/ src_count
;
2170 unsigned pixel_width
= row_type
.width
* dst_channels
;
2172 if (pixel_width
== 24) {
2173 mask_type
.width
= 8;
2174 mask_type
.length
= vector_width
/ mask_type
.width
;
2176 mask_type
.length
= pixels
;
2177 mask_type
.width
= row_type
.width
* dst_channels
;
2180 * If mask_type width is smaller than 32bit, this doesn't quite
2181 * generate the most efficient code (could use some pack).
2183 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2184 lp_build_int_vec_type(gallivm
, mask_type
), "");
2186 mask_type
.length
*= dst_channels
;
2187 mask_type
.width
/= dst_channels
;
2190 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2191 lp_build_int_vec_type(gallivm
, mask_type
), "");
2192 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2199 struct lp_type alpha_type
= fs_type
;
2200 alpha_type
.length
= 4;
2201 convert_alpha(gallivm
, row_type
, alpha_type
,
2202 block_size
, block_height
,
2203 src_count
, dst_channels
,
2204 pad_inline
, src_alpha
);
2205 if (dual_source_blend
) {
2206 convert_alpha(gallivm
, row_type
, alpha_type
,
2207 block_size
, block_height
,
2208 src_count
, dst_channels
,
2209 pad_inline
, src1_alpha
);
2215 * Load dst from memory
2217 if (src_count
< block_height
) {
2218 dst_count
= block_height
;
2220 dst_count
= src_count
;
2223 dst_type
.length
*= block_size
/ dst_count
;
2225 if (format_expands_to_float_soa(out_format_desc
)) {
2227 * we need multiple values at once for the conversion, so can as well
2228 * load them vectorized here too instead of concatenating later.
2229 * (Still need concatenation later for 8-wide vectors).
2231 dst_count
= block_height
;
2232 dst_type
.length
= block_width
;
2236 * Compute the alignment of the destination pointer in bytes
2237 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2238 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2239 * 1d tex but can't distinguish here) so need to stick with per-pixel
2240 * alignment in this case.
2243 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2246 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2248 /* Force power-of-two alignment by extracting only the least-significant-bit */
2249 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2251 * Resource base and stride pointers are aligned to 16 bytes, so that's
2252 * the maximum alignment we can guarantee
2254 dst_alignment
= MIN2(16, dst_alignment
);
2258 if (dst_count
> src_count
) {
2259 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2260 util_is_power_of_two_or_zero(dst_type
.length
) &&
2261 dst_type
.length
* dst_type
.width
< 128) {
2263 * Never try to load values as 4xi8 which we will then
2264 * concatenate to larger vectors. This gives llvm a real
2265 * headache (the problem is the type legalizer (?) will
2266 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2267 * then the shuffles to concatenate are more or less impossible
2268 * - llvm is easily capable of generating a sequence of 32
2269 * pextrb/pinsrb instructions for that. Albeit it appears to
2270 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2271 * width to avoid the trouble (16bit seems not as bad, llvm
2272 * probably recognizes the load+shuffle as only one shuffle
2273 * is necessary, but we can do just the same anyway).
2275 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2281 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2282 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2283 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2284 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2289 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2290 dst
, ls_type
, dst_count
, dst_alignment
);
2295 * Convert from dst/output format to src/blending format.
2297 * This is necessary as we can only read 1 row from memory at a time,
2298 * so the minimum dst_count will ever be at this point is 4.
2300 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2301 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2302 * on all 16 pixels in that single vector at once.
2304 if (dst_count
> src_count
) {
2305 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2306 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2307 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2308 for (i
= 0; i
< dst_count
; i
++) {
2309 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2313 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2315 if (ls_type
.length
!= dst_type
.length
) {
2316 struct lp_type tmp_type
= dst_type
;
2317 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2318 for (i
= 0; i
< src_count
; i
++) {
2319 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2320 lp_build_vec_type(gallivm
, tmp_type
), "");
2328 /* XXX this is broken for RGB8 formats -
2329 * they get expanded from 12 to 16 elements (to include alpha)
2330 * by convert_to_blend_type then reduced to 15 instead of 12
2331 * by convert_from_blend_type (a simple fix though breaks A8...).
2332 * R16G16B16 also crashes differently however something going wrong
2333 * inside llvm handling npot vector sizes seemingly.
2334 * It seems some cleanup could be done here (like skipping conversion/blend
2337 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2338 row_type
, dst
, src_count
);
2341 * FIXME: Really should get logic ops / masks out of generic blend / row
2342 * format. Logic ops will definitely not work on the blend float format
2343 * used for SRGB here and I think OpenGL expects this to work as expected
2344 * (that is incoming values converted to srgb then logic op applied).
2346 for (i
= 0; i
< src_count
; ++i
) {
2347 dst
[i
] = lp_build_blend_aos(gallivm
,
2348 &variant
->key
.blend
,
2353 has_alpha
? NULL
: src_alpha
[i
],
2355 has_alpha
? NULL
: src1_alpha
[i
],
2357 partial_mask
? src_mask
[i
] : NULL
,
2359 has_alpha
? NULL
: blend_alpha
,
2361 pad_inline
? 4 : dst_channels
);
2364 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2365 row_type
, dst_type
, dst
, src_count
);
2367 /* Split the blend rows back to memory rows */
2368 if (dst_count
> src_count
) {
2369 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2371 if (src_count
== 1) {
2372 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2373 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2375 row_type
.length
/= 2;
2379 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2380 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2381 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2382 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2384 row_type
.length
/= 2;
2389 * Store blend result to memory
2392 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2393 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2396 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2397 dst
, dst_type
, dst_count
, dst_alignment
);
2400 if (have_smallfloat_format(dst_type
, out_format
)) {
2401 lp_build_fpstate_set(gallivm
, fpstate
);
2405 lp_build_mask_end(&mask_ctx
);
2411 * Generate the runtime callable function for the whole fragment pipeline.
2412 * Note that the function which we generate operates on a block of 16
2413 * pixels at at time. The block contains 2x2 quads. Each quad contains
2417 generate_fragment(struct llvmpipe_context
*lp
,
2418 struct lp_fragment_shader
*shader
,
2419 struct lp_fragment_shader_variant
*variant
,
2420 unsigned partial_mask
)
2422 struct gallivm_state
*gallivm
= variant
->gallivm
;
2423 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2424 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2426 struct lp_type fs_type
;
2427 struct lp_type blend_type
;
2428 LLVMTypeRef fs_elem_type
;
2429 LLVMTypeRef blend_vec_type
;
2430 LLVMTypeRef arg_types
[13];
2431 LLVMTypeRef func_type
;
2432 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2433 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2434 LLVMValueRef context_ptr
;
2437 LLVMValueRef a0_ptr
;
2438 LLVMValueRef dadx_ptr
;
2439 LLVMValueRef dady_ptr
;
2440 LLVMValueRef color_ptr_ptr
;
2441 LLVMValueRef stride_ptr
;
2442 LLVMValueRef depth_ptr
;
2443 LLVMValueRef depth_stride
;
2444 LLVMValueRef mask_input
;
2445 LLVMValueRef thread_data_ptr
;
2446 LLVMBasicBlockRef block
;
2447 LLVMBuilderRef builder
;
2448 struct lp_build_sampler_soa
*sampler
;
2449 struct lp_build_interp_soa_context interp
;
2450 LLVMValueRef fs_mask
[16 / 4];
2451 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2452 LLVMValueRef function
;
2453 LLVMValueRef facing
;
2458 boolean cbuf0_write_all
;
2459 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2460 util_blend_state_is_dual(&key
->blend
, 0);
2462 assert(lp_native_vector_width
/ 32 >= 4);
2464 /* Adjust color input interpolation according to flatshade state:
2466 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2467 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2468 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2470 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2472 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2476 /* check if writes to cbuf[0] are to be copied to all cbufs */
2478 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2480 /* TODO: actually pick these based on the fs and color buffer
2481 * characteristics. */
2483 memset(&fs_type
, 0, sizeof fs_type
);
2484 fs_type
.floating
= TRUE
; /* floating point values */
2485 fs_type
.sign
= TRUE
; /* values are signed */
2486 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2487 fs_type
.width
= 32; /* 32-bit float */
2488 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2490 memset(&blend_type
, 0, sizeof blend_type
);
2491 blend_type
.floating
= FALSE
; /* values are integers */
2492 blend_type
.sign
= FALSE
; /* values are unsigned */
2493 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2494 blend_type
.width
= 8; /* 8-bit ubyte values */
2495 blend_type
.length
= 16; /* 16 elements per vector */
2498 * Generate the function prototype. Any change here must be reflected in
2499 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2502 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2504 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2506 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2507 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2509 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2510 arg_types
[1] = int32_type
; /* x */
2511 arg_types
[2] = int32_type
; /* y */
2512 arg_types
[3] = int32_type
; /* facing */
2513 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2514 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2515 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2516 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2517 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2518 arg_types
[9] = int32_type
; /* mask_input */
2519 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2520 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2521 arg_types
[12] = int32_type
; /* depth_stride */
2523 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2524 arg_types
, ARRAY_SIZE(arg_types
), 0);
2526 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2527 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2529 variant
->function
[partial_mask
] = function
;
2531 /* XXX: need to propagate noalias down into color param now we are
2532 * passing a pointer-to-pointer?
2534 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2535 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2536 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2538 context_ptr
= LLVMGetParam(function
, 0);
2539 x
= LLVMGetParam(function
, 1);
2540 y
= LLVMGetParam(function
, 2);
2541 facing
= LLVMGetParam(function
, 3);
2542 a0_ptr
= LLVMGetParam(function
, 4);
2543 dadx_ptr
= LLVMGetParam(function
, 5);
2544 dady_ptr
= LLVMGetParam(function
, 6);
2545 color_ptr_ptr
= LLVMGetParam(function
, 7);
2546 depth_ptr
= LLVMGetParam(function
, 8);
2547 mask_input
= LLVMGetParam(function
, 9);
2548 thread_data_ptr
= LLVMGetParam(function
, 10);
2549 stride_ptr
= LLVMGetParam(function
, 11);
2550 depth_stride
= LLVMGetParam(function
, 12);
2552 lp_build_name(context_ptr
, "context");
2553 lp_build_name(x
, "x");
2554 lp_build_name(y
, "y");
2555 lp_build_name(a0_ptr
, "a0");
2556 lp_build_name(dadx_ptr
, "dadx");
2557 lp_build_name(dady_ptr
, "dady");
2558 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2559 lp_build_name(depth_ptr
, "depth");
2560 lp_build_name(mask_input
, "mask_input");
2561 lp_build_name(thread_data_ptr
, "thread_data");
2562 lp_build_name(stride_ptr
, "stride_ptr");
2563 lp_build_name(depth_stride
, "depth_stride");
2569 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2570 builder
= gallivm
->builder
;
2572 LLVMPositionBuilderAtEnd(builder
, block
);
2575 * Must not count ps invocations if there's a null shader.
2576 * (It would be ok to count with null shader if there's d/s tests,
2577 * but only if there's d/s buffers too, which is different
2578 * to implicit rasterization disable which must not depend
2579 * on the d/s buffers.)
2580 * Could use popcount on mask, but pixel accuracy is not required.
2581 * Could disable if there's no stats query, but maybe not worth it.
2583 if (shader
->info
.base
.num_instructions
> 1) {
2584 LLVMValueRef invocs
, val
;
2585 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2586 val
= LLVMBuildLoad(builder
, invocs
, "");
2587 val
= LLVMBuildAdd(builder
, val
,
2588 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2590 LLVMBuildStore(builder
, val
, invocs
);
2593 /* code generated texture sampling */
2594 sampler
= lp_llvm_sampler_soa_create(key
->state
);
2596 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2597 /* for 1d resources only run "upper half" of stamp */
2598 if (key
->resource_1d
)
2602 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2603 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2604 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2605 num_loop
, "mask_store");
2606 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2607 boolean pixel_center_integer
=
2608 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2611 * The shader input interpolation info is not explicitely baked in the
2612 * shader key, but everything it derives from (TGSI, and flatshade) is
2613 * already included in the shader key.
2615 lp_build_interp_soa_init(&interp
,
2617 shader
->info
.base
.num_inputs
,
2619 pixel_center_integer
,
2622 a0_ptr
, dadx_ptr
, dady_ptr
,
2625 for (i
= 0; i
< num_fs
; i
++) {
2627 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2628 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2629 &indexi
, 1, "mask_ptr");
2632 mask
= generate_quad_mask(gallivm
, fs_type
,
2633 i
*fs_type
.length
/4, mask_input
);
2636 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2638 LLVMBuildStore(builder
, mask
, mask_ptr
);
2641 generate_fs_loop(gallivm
,
2649 mask_store
, /* output */
2656 for (i
= 0; i
< num_fs
; i
++) {
2657 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2658 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2660 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2661 /* This is fucked up need to reorganize things */
2662 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2663 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2664 ptr
= LLVMBuildGEP(builder
,
2665 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2667 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2670 if (dual_source_blend
) {
2671 /* only support one dual source blend target hence always use output 1 */
2672 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2673 ptr
= LLVMBuildGEP(builder
,
2674 color_store
[1][chan
],
2676 fs_out_color
[1][chan
][i
] = ptr
;
2682 sampler
->destroy(sampler
);
2684 /* Loop over color outputs / color buffers to do blending.
2686 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2687 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2688 LLVMValueRef color_ptr
;
2689 LLVMValueRef stride
;
2690 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2692 boolean do_branch
= ((key
->depth
.enabled
2693 || key
->stencil
[0].enabled
2694 || key
->alpha
.enabled
)
2695 && !shader
->info
.base
.uses_kill
);
2697 color_ptr
= LLVMBuildLoad(builder
,
2698 LLVMBuildGEP(builder
, color_ptr_ptr
,
2702 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2704 stride
= LLVMBuildLoad(builder
,
2705 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2708 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2709 key
->cbuf_format
[cbuf
],
2710 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2711 context_ptr
, color_ptr
, stride
,
2712 partial_mask
, do_branch
);
2716 LLVMBuildRetVoid(builder
);
2718 gallivm_verify_function(gallivm
, function
);
2723 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2727 debug_printf("fs variant %p:\n", (void *) key
);
2729 if (key
->flatshade
) {
2730 debug_printf("flatshade = 1\n");
2732 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2733 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2735 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2736 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2738 if (key
->depth
.enabled
) {
2739 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2740 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2743 for (i
= 0; i
< 2; ++i
) {
2744 if (key
->stencil
[i
].enabled
) {
2745 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2746 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2747 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2748 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2749 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2750 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2754 if (key
->alpha
.enabled
) {
2755 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2758 if (key
->occlusion_count
) {
2759 debug_printf("occlusion_count = 1\n");
2762 if (key
->blend
.logicop_enable
) {
2763 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2765 else if (key
->blend
.rt
[0].blend_enable
) {
2766 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2767 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2768 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2769 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2770 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2771 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2773 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2774 if (key
->blend
.alpha_to_coverage
) {
2775 debug_printf("blend.alpha_to_coverage is enabled\n");
2777 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2778 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2779 debug_printf("sampler[%u] = \n", i
);
2780 debug_printf(" .wrap = %s %s %s\n",
2781 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2782 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2783 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2784 debug_printf(" .min_img_filter = %s\n",
2785 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2786 debug_printf(" .min_mip_filter = %s\n",
2787 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2788 debug_printf(" .mag_img_filter = %s\n",
2789 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2790 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2791 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
2792 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2793 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2794 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2795 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2796 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2798 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2799 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2800 debug_printf("texture[%u] = \n", i
);
2801 debug_printf(" .format = %s\n",
2802 util_format_name(texture
->format
));
2803 debug_printf(" .target = %s\n",
2804 util_str_tex_target(texture
->target
, TRUE
));
2805 debug_printf(" .level_zero_only = %u\n",
2806 texture
->level_zero_only
);
2807 debug_printf(" .pot = %u %u %u\n",
2809 texture
->pot_height
,
2810 texture
->pot_depth
);
2816 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2818 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2819 variant
->shader
->no
, variant
->no
);
2820 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2821 dump_fs_variant_key(&variant
->key
);
2822 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2828 * Generate a new fragment shader variant from the shader code and
2829 * other state indicated by the key.
2831 static struct lp_fragment_shader_variant
*
2832 generate_variant(struct llvmpipe_context
*lp
,
2833 struct lp_fragment_shader
*shader
,
2834 const struct lp_fragment_shader_variant_key
*key
)
2836 struct lp_fragment_shader_variant
*variant
;
2837 const struct util_format_description
*cbuf0_format_desc
= NULL
;
2838 boolean fullcolormask
;
2839 char module_name
[64];
2841 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2845 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2846 shader
->no
, shader
->variants_created
);
2848 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
2849 if (!variant
->gallivm
) {
2854 variant
->shader
= shader
;
2855 variant
->list_item_global
.base
= variant
;
2856 variant
->list_item_local
.base
= variant
;
2857 variant
->no
= shader
->variants_created
++;
2859 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2862 * Determine whether we are touching all channels in the color buffer.
2864 fullcolormask
= FALSE
;
2865 if (key
->nr_cbufs
== 1) {
2866 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2867 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2871 !key
->blend
.logicop_enable
&&
2872 !key
->blend
.rt
[0].blend_enable
&&
2874 !key
->stencil
[0].enabled
&&
2875 !key
->alpha
.enabled
&&
2876 !key
->blend
.alpha_to_coverage
&&
2877 !key
->depth
.enabled
&&
2878 !shader
->info
.base
.uses_kill
&&
2879 !shader
->info
.base
.writes_samplemask
2882 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2883 lp_debug_fs_variant(variant
);
2886 lp_jit_init_types(variant
);
2888 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2889 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2891 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2892 if (variant
->opaque
) {
2893 /* Specialized shader, which doesn't need to read the color buffer. */
2894 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2899 * Compile everything
2902 gallivm_compile_module(variant
->gallivm
);
2904 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2906 if (variant
->function
[RAST_EDGE_TEST
]) {
2907 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2908 gallivm_jit_function(variant
->gallivm
,
2909 variant
->function
[RAST_EDGE_TEST
]);
2912 if (variant
->function
[RAST_WHOLE
]) {
2913 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2914 gallivm_jit_function(variant
->gallivm
,
2915 variant
->function
[RAST_WHOLE
]);
2916 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2917 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2920 gallivm_free_ir(variant
->gallivm
);
2927 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2928 const struct pipe_shader_state
*templ
)
2930 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2931 struct lp_fragment_shader
*shader
;
2933 int nr_sampler_views
;
2936 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2940 shader
->no
= fs_no
++;
2941 make_empty_list(&shader
->variants
);
2943 /* get/save the summary info for this shader */
2944 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2946 /* we need to keep a local copy of the tokens */
2947 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2949 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2950 if (shader
->draw_data
== NULL
) {
2951 FREE((void *) shader
->base
.tokens
);
2956 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2957 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2959 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2960 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2962 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2963 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2964 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2966 switch (shader
->info
.base
.input_interpolate
[i
]) {
2967 case TGSI_INTERPOLATE_CONSTANT
:
2968 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2970 case TGSI_INTERPOLATE_LINEAR
:
2971 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2973 case TGSI_INTERPOLATE_PERSPECTIVE
:
2974 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2976 case TGSI_INTERPOLATE_COLOR
:
2977 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2984 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2985 case TGSI_SEMANTIC_FACE
:
2986 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2988 case TGSI_SEMANTIC_POSITION
:
2989 /* Position was already emitted above
2991 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2992 shader
->inputs
[i
].src_index
= 0;
2996 /* XXX this is a completely pointless index map... */
2997 shader
->inputs
[i
].src_index
= i
+1;
3000 if (LP_DEBUG
& DEBUG_TGSI
) {
3002 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3003 shader
->no
, (void *) shader
);
3004 tgsi_dump(templ
->tokens
, 0);
3005 debug_printf("usage masks:\n");
3006 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3007 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3008 debug_printf(" IN[%u].%s%s%s%s\n",
3010 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3011 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3012 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3013 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3023 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3025 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3027 if (llvmpipe
->fs
== fs
)
3030 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
3032 draw_bind_fragment_shader(llvmpipe
->draw
,
3033 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
3035 llvmpipe
->dirty
|= LP_NEW_FS
;
3040 * Remove shader variant from two lists: the shader's variant list
3041 * and the context's variant list.
3044 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3045 struct lp_fragment_shader_variant
*variant
)
3047 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3048 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3049 "v total cached %u inst %u total inst %u\n",
3050 variant
->shader
->no
, variant
->no
,
3051 variant
->shader
->variants_created
,
3052 variant
->shader
->variants_cached
,
3053 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3056 gallivm_destroy(variant
->gallivm
);
3058 /* remove from shader's list */
3059 remove_from_list(&variant
->list_item_local
);
3060 variant
->shader
->variants_cached
--;
3062 /* remove from context's list */
3063 remove_from_list(&variant
->list_item_global
);
3064 lp
->nr_fs_variants
--;
3065 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3072 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3074 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3075 struct lp_fragment_shader
*shader
= fs
;
3076 struct lp_fs_variant_list_item
*li
;
3078 assert(fs
!= llvmpipe
->fs
);
3081 * XXX: we need to flush the context until we have some sort of reference
3082 * counting in fragment shaders as they may still be binned
3083 * Flushing alone might not sufficient we need to wait on it too.
3085 llvmpipe_finish(pipe
, __FUNCTION__
);
3087 /* Delete all the variants */
3088 li
= first_elem(&shader
->variants
);
3089 while(!at_end(&shader
->variants
, li
)) {
3090 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3091 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3095 /* Delete draw module's data */
3096 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3098 assert(shader
->variants_cached
== 0);
3099 FREE((void *) shader
->base
.tokens
);
3106 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3107 enum pipe_shader_type shader
, uint index
,
3108 const struct pipe_constant_buffer
*cb
)
3110 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3111 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3113 assert(shader
< PIPE_SHADER_TYPES
);
3114 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3116 /* note: reference counting */
3117 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3120 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3121 debug_printf("Illegal set constant without bind flag\n");
3122 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3126 if (shader
== PIPE_SHADER_VERTEX
||
3127 shader
== PIPE_SHADER_GEOMETRY
) {
3128 /* Pass the constants to the 'draw' module */
3129 const unsigned size
= cb
? cb
->buffer_size
: 0;
3133 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3135 else if (cb
&& cb
->user_buffer
) {
3136 data
= (ubyte
*) cb
->user_buffer
;
3143 data
+= cb
->buffer_offset
;
3145 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3149 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3152 if (cb
&& cb
->user_buffer
) {
3153 pipe_resource_reference(&constants
, NULL
);
3158 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3159 enum pipe_shader_type shader
, unsigned start_slot
,
3160 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3161 unsigned writable_bitmask
)
3163 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3165 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3166 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3168 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3170 if (shader
== PIPE_SHADER_VERTEX
||
3171 shader
== PIPE_SHADER_GEOMETRY
) {
3172 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3173 const ubyte
*data
= NULL
;
3174 if (buffer
&& buffer
->buffer
)
3175 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3177 data
+= buffer
->buffer_offset
;
3178 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3180 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3181 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3187 * Return the blend factor equivalent to a destination alpha of one.
3189 static inline unsigned
3190 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3193 case PIPE_BLENDFACTOR_DST_ALPHA
:
3194 return PIPE_BLENDFACTOR_ONE
;
3195 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3196 return PIPE_BLENDFACTOR_ZERO
;
3197 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3199 return PIPE_BLENDFACTOR_ZERO
;
3201 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3209 * We need to generate several variants of the fragment pipeline to match
3210 * all the combinations of the contributing state atoms.
3212 * TODO: there is actually no reason to tie this to context state -- the
3213 * generated code could be cached globally in the screen.
3216 make_variant_key(struct llvmpipe_context
*lp
,
3217 struct lp_fragment_shader
*shader
,
3218 struct lp_fragment_shader_variant_key
*key
)
3222 memset(key
, 0, shader
->variant_key_size
);
3224 if (lp
->framebuffer
.zsbuf
) {
3225 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3226 const struct util_format_description
*zsbuf_desc
=
3227 util_format_description(zsbuf_format
);
3229 if (lp
->depth_stencil
->depth
.enabled
&&
3230 util_format_has_depth(zsbuf_desc
)) {
3231 key
->zsbuf_format
= zsbuf_format
;
3232 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3234 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3235 util_format_has_stencil(zsbuf_desc
)) {
3236 key
->zsbuf_format
= zsbuf_format
;
3237 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3239 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3240 key
->resource_1d
= TRUE
;
3245 * Propagate the depth clamp setting from the rasterizer state.
3246 * depth_clip == 0 implies depth clamping is enabled.
3248 * When clip_halfz is enabled, then always clamp the depth values.
3250 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3251 * clamp is always active in d3d10, regardless if depth clip is
3253 * (GL has an always-on [0,1] clamp on fs depth output instead
3254 * to ensure the depth values stay in range. Doesn't look like
3255 * we do that, though...)
3257 if (lp
->rasterizer
->clip_halfz
) {
3258 key
->depth_clamp
= 1;
3260 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3263 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3264 if (!lp
->framebuffer
.nr_cbufs
||
3265 !lp
->framebuffer
.cbufs
[0] ||
3266 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3267 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3269 if(key
->alpha
.enabled
)
3270 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3271 /* alpha.ref_value is passed in jit_context */
3273 key
->flatshade
= lp
->rasterizer
->flatshade
;
3274 if (lp
->active_occlusion_queries
) {
3275 key
->occlusion_count
= TRUE
;
3278 if (lp
->framebuffer
.nr_cbufs
) {
3279 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3282 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3284 if (!key
->blend
.independent_blend_enable
) {
3285 /* we always need independent blend otherwise the fixups below won't work */
3286 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3287 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3289 key
->blend
.independent_blend_enable
= 1;
3292 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3293 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3295 if (lp
->framebuffer
.cbufs
[i
]) {
3296 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3297 const struct util_format_description
*format_desc
;
3299 key
->cbuf_format
[i
] = format
;
3302 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3303 * mixing of 2d textures with height 1 and 1d textures, so make sure
3304 * we pick 1d if any cbuf or zsbuf is 1d.
3306 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3307 key
->resource_1d
= TRUE
;
3310 format_desc
= util_format_description(format
);
3311 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3312 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3315 * Mask out color channels not present in the color buffer.
3317 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3320 * Disable blend for integer formats.
3322 if (util_format_is_pure_integer(format
)) {
3323 blend_rt
->blend_enable
= 0;
3327 * Our swizzled render tiles always have an alpha channel, but the
3328 * linear render target format often does not, so force here the dst
3331 * This is not a mere optimization. Wrong results will be produced if
3332 * the dst alpha is used, the dst format does not have alpha, and the
3333 * previous rendering was not flushed from the swizzled to linear
3334 * buffer. For example, NonPowTwo DCT.
3336 * TODO: This should be generalized to all channels for better
3337 * performance, but only alpha causes correctness issues.
3339 * Also, force rgb/alpha func/factors match, to make AoS blending
3342 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3343 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3344 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3345 boolean clamped_zero
= !util_format_is_float(format
) &&
3346 !util_format_is_snorm(format
);
3347 blend_rt
->rgb_src_factor
=
3348 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3349 blend_rt
->rgb_dst_factor
=
3350 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3351 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3352 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3353 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3357 /* no color buffer for this fragment output */
3358 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3359 blend_rt
->colormask
= 0x0;
3360 blend_rt
->blend_enable
= 0;
3364 /* This value will be the same for all the variants of a given shader:
3366 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3368 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3369 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3370 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
3371 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3376 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3377 * are dx10-style? Can't really have mixed opcodes, at least not
3378 * if we want to skip the holes here (without rescanning tgsi).
3380 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3381 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3382 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3384 * Note sview may exceed what's representable by file_mask.
3385 * This will still work, the only downside is that not actually
3386 * used views may be included in the shader key.
3388 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3389 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3390 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3395 key
->nr_sampler_views
= key
->nr_samplers
;
3396 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3397 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3398 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3399 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3408 * Update fragment shader state. This is called just prior to drawing
3409 * something when some fragment-related state has changed.
3412 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3414 struct lp_fragment_shader
*shader
= lp
->fs
;
3415 struct lp_fragment_shader_variant_key key
;
3416 struct lp_fragment_shader_variant
*variant
= NULL
;
3417 struct lp_fs_variant_list_item
*li
;
3419 make_variant_key(lp
, shader
, &key
);
3421 /* Search the variants for one which matches the key */
3422 li
= first_elem(&shader
->variants
);
3423 while(!at_end(&shader
->variants
, li
)) {
3424 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3432 /* Move this variant to the head of the list to implement LRU
3433 * deletion of shader's when we have too many.
3435 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3438 /* variant not found, create it now */
3441 unsigned variants_to_cull
;
3443 if (LP_DEBUG
& DEBUG_FS
) {
3444 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3447 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3450 /* First, check if we've exceeded the max number of shader variants.
3451 * If so, free 6.25% of them (the least recently used ones).
3453 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3455 if (variants_to_cull
||
3456 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3457 struct pipe_context
*pipe
= &lp
->pipe
;
3459 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3460 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3461 "\t%u instrs,\t%u instrs/variant\n",
3462 shader
->variants_cached
,
3463 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3464 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3468 * XXX: we need to flush the context until we have some sort of
3469 * reference counting in fragment shaders as they may still be binned
3470 * Flushing alone might not be sufficient we need to wait on it too.
3472 llvmpipe_finish(pipe
, __FUNCTION__
);
3475 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3476 * number of shader variants (potentially all of them) could be
3477 * pending for destruction on flush.
3480 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3481 struct lp_fs_variant_list_item
*item
;
3482 if (is_empty_list(&lp
->fs_variants_list
)) {
3485 item
= last_elem(&lp
->fs_variants_list
);
3488 llvmpipe_remove_shader_variant(lp
, item
->base
);
3493 * Generate the new variant.
3496 variant
= generate_variant(lp
, shader
, &key
);
3499 LP_COUNT_ADD(llvm_compile_time
, dt
);
3500 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3502 /* Put the new variant into the list */
3504 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3505 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3506 lp
->nr_fs_variants
++;
3507 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3508 shader
->variants_cached
++;
3512 /* Bind this variant */
3513 lp_setup_set_fs_variant(lp
->setup
, variant
);
3521 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3523 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3524 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3525 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3527 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3529 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;