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/format/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_nir.h"
84 #include "gallivm/lp_bld_swizzle.h"
85 #include "gallivm/lp_bld_flow.h"
86 #include "gallivm/lp_bld_debug.h"
87 #include "gallivm/lp_bld_arit.h"
88 #include "gallivm/lp_bld_bitarit.h"
89 #include "gallivm/lp_bld_pack.h"
90 #include "gallivm/lp_bld_format.h"
91 #include "gallivm/lp_bld_quad.h"
93 #include "lp_bld_alpha.h"
94 #include "lp_bld_blend.h"
95 #include "lp_bld_depth.h"
96 #include "lp_bld_interp.h"
97 #include "lp_context.h"
100 #include "lp_setup.h"
101 #include "lp_state.h"
102 #include "lp_tex_sample.h"
103 #include "lp_flush.h"
104 #include "lp_state_fs.h"
106 #include "nir/nir_to_tgsi_info.h"
108 /** Fragment shader number (for debugging) */
109 static unsigned fs_no
= 0;
113 * Expand the relevant bits of mask_input to a n*4-dword mask for the
114 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
115 * quad mask vector to 0 or ~0.
116 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
117 * quad arguments with fs length 8.
119 * \param first_quad which quad(s) of the quad group to test, in [0,3]
120 * \param mask_input bitwise mask for the whole 4x4 stamp
123 generate_quad_mask(struct gallivm_state
*gallivm
,
124 struct lp_type fs_type
,
127 LLVMValueRef mask_input
) /* int64 */
129 LLVMBuilderRef builder
= gallivm
->builder
;
130 struct lp_type mask_type
;
131 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
132 LLVMValueRef bits
[16];
133 LLVMValueRef mask
, bits_vec
;
137 * XXX: We'll need a different path for 16 x u8
139 assert(fs_type
.width
== 32);
140 assert(fs_type
.length
<= ARRAY_SIZE(bits
));
141 mask_type
= lp_int_type(fs_type
);
144 * mask_input >>= (quad * 4)
146 switch (first_quad
) {
151 assert(fs_type
.length
== 4);
158 assert(fs_type
.length
== 4);
166 mask_input
= LLVMBuildLShr(builder
, mask_input
, lp_build_const_int64(gallivm
, 16 * sample
), "");
167 mask_input
= LLVMBuildTrunc(builder
, mask_input
,
169 mask_input
= LLVMBuildAnd(builder
, mask_input
, lp_build_const_int32(gallivm
, 0xffff), "");
171 mask_input
= LLVMBuildLShr(builder
,
173 LLVMConstInt(i32t
, shift
, 0),
177 * mask = { mask_input & (1 << i), for i in [0,3] }
179 mask
= lp_build_broadcast(gallivm
,
180 lp_build_vec_type(gallivm
, mask_type
),
183 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
184 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
185 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1ULL << (j
+ 0), 0);
186 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
187 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
188 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (j
+ 5), 0);
190 bits_vec
= LLVMConstVector(bits
, fs_type
.length
);
191 mask
= LLVMBuildAnd(builder
, mask
, bits_vec
, "");
194 * mask = mask == bits ? ~0 : 0
196 mask
= lp_build_compare(gallivm
,
197 mask_type
, PIPE_FUNC_EQUAL
,
204 #define EARLY_DEPTH_TEST 0x1
205 #define LATE_DEPTH_TEST 0x2
206 #define EARLY_DEPTH_WRITE 0x4
207 #define LATE_DEPTH_WRITE 0x8
210 find_output_by_semantic( const struct tgsi_shader_info
*info
,
216 for (i
= 0; i
< info
->num_outputs
; i
++)
217 if (info
->output_semantic_name
[i
] == semantic
&&
218 info
->output_semantic_index
[i
] == index
)
226 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
229 lp_llvm_viewport(LLVMValueRef context_ptr
,
230 struct gallivm_state
*gallivm
,
231 LLVMValueRef viewport_index
)
233 LLVMBuilderRef builder
= gallivm
->builder
;
236 struct lp_type viewport_type
=
237 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
239 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
240 ptr
= LLVMBuildPointerCast(builder
, ptr
,
241 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
243 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
250 lp_build_depth_clamp(struct gallivm_state
*gallivm
,
251 LLVMBuilderRef builder
,
253 LLVMValueRef context_ptr
,
254 LLVMValueRef thread_data_ptr
,
257 LLVMValueRef viewport
, min_depth
, max_depth
;
258 LLVMValueRef viewport_index
;
259 struct lp_build_context f32_bld
;
261 assert(type
.floating
);
262 lp_build_context_init(&f32_bld
, gallivm
, type
);
265 * Assumes clamping of the viewport index will occur in setup/gs. Value
266 * is passed through the rasterization stage via lp_rast_shader_inputs.
268 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
271 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
275 * Load the min and max depth from the lp_jit_context.viewports
276 * array of lp_jit_viewport structures.
278 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
280 /* viewports[viewport_index].min_depth */
281 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
282 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
), "");
283 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
285 /* viewports[viewport_index].max_depth */
286 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
287 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
), "");
288 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
291 * Clamp to the min and max depth values for the given viewport.
293 return lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
297 lp_build_sample_alpha_to_coverage(struct gallivm_state
*gallivm
,
299 unsigned coverage_samples
,
300 LLVMValueRef num_loop
,
301 LLVMValueRef loop_counter
,
302 LLVMValueRef coverage_mask_store
,
305 struct lp_build_context bld
;
306 LLVMBuilderRef builder
= gallivm
->builder
;
307 float step
= 1.0 / coverage_samples
;
309 lp_build_context_init(&bld
, gallivm
, type
);
310 for (unsigned s
= 0; s
< coverage_samples
; s
++) {
311 LLVMValueRef alpha_ref_value
= lp_build_const_vec(gallivm
, type
, step
* s
);
312 LLVMValueRef test
= lp_build_cmp(&bld
, PIPE_FUNC_GREATER
, alpha
, alpha_ref_value
);
314 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, s
), num_loop
, "");
315 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_counter
, "");
316 LLVMValueRef s_mask_ptr
= LLVMBuildGEP(builder
, coverage_mask_store
, &s_mask_idx
, 1, "");
317 LLVMValueRef s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
318 s_mask
= LLVMBuildAnd(builder
, s_mask
, test
, "");
319 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
323 struct lp_build_fs_llvm_iface
{
324 struct lp_build_fs_iface base
;
325 struct lp_build_interp_soa_context
*interp
;
326 struct lp_build_for_loop_state
*loop_state
;
327 LLVMValueRef mask_store
;
330 static LLVMValueRef
fs_interp(const struct lp_build_fs_iface
*iface
,
331 struct lp_build_context
*bld
,
332 unsigned attrib
, unsigned chan
,
333 bool centroid
, bool sample
,
334 LLVMValueRef attrib_indir
,
335 LLVMValueRef offsets
[2])
337 struct lp_build_fs_llvm_iface
*fs_iface
= (struct lp_build_fs_llvm_iface
*)iface
;
338 struct lp_build_interp_soa_context
*interp
= fs_iface
->interp
;
339 unsigned loc
= TGSI_INTERPOLATE_LOC_CENTER
;
341 loc
= TGSI_INTERPOLATE_LOC_CENTROID
;
343 loc
= TGSI_INTERPOLATE_LOC_SAMPLE
;
345 return lp_build_interp_soa(interp
, bld
->gallivm
, fs_iface
->loop_state
->counter
,
346 fs_iface
->mask_store
,
347 attrib
, chan
, loc
, attrib_indir
, offsets
);
351 * Generate the fragment shader, depth/stencil test, and alpha tests.
354 generate_fs_loop(struct gallivm_state
*gallivm
,
355 struct lp_fragment_shader
*shader
,
356 const struct lp_fragment_shader_variant_key
*key
,
357 LLVMBuilderRef builder
,
359 LLVMValueRef context_ptr
,
360 LLVMValueRef sample_pos_array
,
361 LLVMValueRef num_loop
,
362 struct lp_build_interp_soa_context
*interp
,
363 const struct lp_build_sampler_soa
*sampler
,
364 const struct lp_build_image_soa
*image
,
365 LLVMValueRef mask_store
,
366 LLVMValueRef (*out_color
)[4],
367 LLVMValueRef depth_base_ptr
,
368 LLVMValueRef depth_stride
,
369 LLVMValueRef depth_sample_stride
,
371 LLVMValueRef thread_data_ptr
)
373 const struct util_format_description
*zs_format_desc
= NULL
;
374 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
375 struct lp_type int_type
= lp_int_type(type
);
376 LLVMTypeRef vec_type
, int_vec_type
;
377 LLVMValueRef mask_ptr
= NULL
, mask_val
= NULL
;
378 LLVMValueRef consts_ptr
, num_consts_ptr
;
379 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
381 LLVMValueRef z_value
, s_value
;
382 LLVMValueRef z_fb
, s_fb
;
383 LLVMValueRef depth_ptr
;
384 LLVMValueRef stencil_refs
[2];
385 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
386 LLVMValueRef zs_samples
= lp_build_const_int32(gallivm
, key
->zsbuf_nr_samples
);
387 struct lp_build_for_loop_state loop_state
, sample_loop_state
;
388 struct lp_build_mask_context mask
;
390 * TODO: figure out if simple_shader optimization is really worthwile to
391 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
392 * code since tests tend to take another codepath than real shaders.
394 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
395 shader
->info
.base
.num_inputs
< 3 &&
396 shader
->info
.base
.num_instructions
< 8) && 0;
397 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
398 util_blend_state_is_dual(&key
->blend
, 0);
404 struct lp_bld_tgsi_system_values system_values
;
406 memset(&system_values
, 0, sizeof(system_values
));
408 /* truncate then sign extend. */
409 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
410 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
412 if (key
->depth
.enabled
||
413 key
->stencil
[0].enabled
) {
415 zs_format_desc
= util_format_description(key
->zsbuf_format
);
416 assert(zs_format_desc
);
418 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
419 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
420 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
421 if (shader
->info
.base
.writes_memory
)
422 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
423 else if (key
->alpha
.enabled
||
424 key
->blend
.alpha_to_coverage
||
425 shader
->info
.base
.uses_kill
||
426 shader
->info
.base
.writes_samplemask
) {
427 /* With alpha test and kill, can do the depth test early
428 * and hopefully eliminate some quads. But need to do a
429 * special deferred depth write once the final mask value
430 * is known. This only works though if there's either no
431 * stencil test or the stencil value isn't written.
433 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
434 (key
->stencil
[1].enabled
&&
435 key
->stencil
[1].writemask
)))
436 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
438 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
441 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
444 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
447 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
448 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
449 (key
->stencil
[1].enabled
&&
450 key
->stencil
[1].writemask
))))
451 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
457 vec_type
= lp_build_vec_type(gallivm
, type
);
458 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
460 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
461 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
462 /* convert scalar stencil refs into vectors */
463 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
464 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
466 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
467 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
469 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
470 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
472 memset(outputs
, 0, sizeof outputs
);
474 /* Allocate color storage for each fragment sample */
475 LLVMValueRef color_store_size
= num_loop
;
476 if (key
->min_samples
> 1)
477 color_store_size
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, key
->min_samples
), "");
479 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
480 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
481 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
482 lp_build_vec_type(gallivm
,
484 color_store_size
, "color");
487 if (dual_source_blend
) {
488 assert(key
->nr_cbufs
<= 1);
489 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
490 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
491 lp_build_vec_type(gallivm
,
493 color_store_size
, "color1");
497 lp_build_for_loop_begin(&loop_state
, gallivm
,
498 lp_build_const_int32(gallivm
, 0),
501 lp_build_const_int32(gallivm
, 1));
503 LLVMValueRef sample_mask_in
;
504 if (key
->multisample
) {
505 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 0);
506 /* create shader execution mask by combining all sample masks. */
507 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
508 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
509 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
510 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
514 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
516 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, s_mask
, lp_build_const_int_vec(gallivm
, type
, (1 << s
)), "");
517 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
520 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 1);
521 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
522 &loop_state
.counter
, 1, "mask_ptr");
523 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
525 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, mask_val
, lp_build_const_int_vec(gallivm
, type
, 1), "");
526 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
529 /* 'mask' will control execution based on quad's pixel alive/killed state */
530 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
532 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
533 lp_build_mask_check(&mask
);
535 /* Create storage for recombining sample masks after early Z pass. */
536 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
537 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
539 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
540 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
541 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
542 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
544 /* Run early depth once per sample */
545 if (key
->multisample
) {
547 if (zs_format_desc
) {
548 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
549 struct lp_type z_type
= zs_type
;
550 struct lp_type s_type
= zs_type
;
551 if (zs_format_desc
->block
.bits
< type
.width
)
552 z_type
.width
= type
.width
;
553 else if (zs_format_desc
->block
.bits
> 32) {
554 z_type
.width
= z_type
.width
/ 2;
555 s_type
.width
= s_type
.width
/ 2;
558 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
559 zs_samples
, "z_sample_store");
560 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
561 zs_samples
, "s_sample_store");
562 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
563 zs_samples
, "z_fb_store");
564 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
565 zs_samples
, "s_fb_store");
567 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
568 lp_build_const_int32(gallivm
, 0),
569 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
570 lp_build_const_int32(gallivm
, 1));
572 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
573 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
574 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
576 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
577 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
581 /* for multisample Z needs to be interpolated at sample points for testing. */
582 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
585 depth_ptr
= depth_base_ptr
;
586 if (key
->multisample
) {
587 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
588 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
591 if (depth_mode
& EARLY_DEPTH_TEST
) {
593 * Clamp according to ARB_depth_clamp semantics.
595 if (key
->depth_clamp
) {
596 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
599 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
600 zs_format_desc
, key
->resource_1d
,
601 depth_ptr
, depth_stride
,
602 &z_fb
, &s_fb
, loop_state
.counter
);
603 lp_build_depth_stencil_test(gallivm
,
608 key
->multisample
? NULL
: &mask
,
614 !simple_shader
&& !key
->multisample
);
616 if (depth_mode
& EARLY_DEPTH_WRITE
) {
617 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
618 zs_format_desc
, key
->resource_1d
,
619 NULL
, NULL
, NULL
, loop_state
.counter
,
620 depth_ptr
, depth_stride
,
624 * Note mask check if stencil is enabled must be after ds write not after
625 * stencil test otherwise new stencil values may not get written if all
626 * fragments got killed by depth/stencil test.
628 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
629 lp_build_mask_check(&mask
);
631 if (key
->multisample
) {
632 z_fb_type
= LLVMTypeOf(z_fb
);
633 z_type
= LLVMTypeOf(z_value
);
634 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
635 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
636 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
637 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
641 if (key
->multisample
) {
643 * Store the post-early Z coverage mask.
644 * Recombine the resulting coverage masks post early Z into the fragment
645 * shader execution mask.
647 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
648 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
649 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
651 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
653 lp_build_for_loop_end(&sample_loop_state
);
655 /* recombined all the coverage masks in the shader exec mask. */
656 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
657 lp_build_mask_update(&mask
, tmp_s_mask_or
);
659 if (key
->min_samples
== 1) {
660 /* for multisample Z needs to be re interpolated at pixel center */
661 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
662 lp_build_mask_update(&mask
, tmp_s_mask_or
);
666 LLVMValueRef out_sample_mask_storage
= NULL
;
667 if (shader
->info
.base
.writes_samplemask
) {
668 out_sample_mask_storage
= lp_build_alloca(gallivm
, int_vec_type
, "write_mask");
669 if (key
->min_samples
> 1)
670 LLVMBuildStore(builder
, LLVMConstNull(int_vec_type
), out_sample_mask_storage
);
673 if (key
->multisample
&& key
->min_samples
> 1) {
674 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
675 lp_build_const_int32(gallivm
, 0),
677 lp_build_const_int32(gallivm
, key
->min_samples
),
678 lp_build_const_int32(gallivm
, 1));
680 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
681 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
682 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
683 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
684 lp_build_mask_force(&mask
, s_mask
);
685 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, sample_loop_state
.counter
);
686 system_values
.sample_id
= sample_loop_state
.counter
;
688 system_values
.sample_id
= lp_build_const_int32(gallivm
, 0);
690 system_values
.sample_mask_in
= sample_mask_in
;
691 system_values
.sample_pos
= sample_pos_array
;
693 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, mask_store
, sample_loop_state
.counter
);
695 struct lp_build_fs_llvm_iface fs_iface
= {
696 .base
.interp_fn
= fs_interp
,
698 .loop_state
= &loop_state
,
699 .mask_store
= mask_store
,
702 struct lp_build_tgsi_params params
;
703 memset(¶ms
, 0, sizeof(params
));
707 params
.fs_iface
= &fs_iface
.base
;
708 params
.consts_ptr
= consts_ptr
;
709 params
.const_sizes_ptr
= num_consts_ptr
;
710 params
.system_values
= &system_values
;
711 params
.inputs
= interp
->inputs
;
712 params
.context_ptr
= context_ptr
;
713 params
.thread_data_ptr
= thread_data_ptr
;
714 params
.sampler
= sampler
;
715 params
.info
= &shader
->info
.base
;
716 params
.ssbo_ptr
= ssbo_ptr
;
717 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
718 params
.image
= image
;
720 /* Build the actual shader */
721 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
722 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
725 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
729 if (key
->alpha
.enabled
) {
730 int color0
= find_output_by_semantic(&shader
->info
.base
,
734 if (color0
!= -1 && outputs
[color0
][3]) {
735 const struct util_format_description
*cbuf_format_desc
;
736 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
737 LLVMValueRef alpha_ref_value
;
739 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
740 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
742 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
744 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
745 &mask
, alpha
, alpha_ref_value
,
746 (depth_mode
& LATE_DEPTH_TEST
) != 0);
750 /* Emulate Alpha to Coverage with Alpha test */
751 if (key
->blend
.alpha_to_coverage
) {
752 int color0
= find_output_by_semantic(&shader
->info
.base
,
756 if (color0
!= -1 && outputs
[color0
][3]) {
757 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
759 if (!key
->multisample
) {
760 lp_build_alpha_to_coverage(gallivm
, type
,
762 (depth_mode
& LATE_DEPTH_TEST
) != 0);
764 lp_build_sample_alpha_to_coverage(gallivm
, type
, key
->coverage_samples
, num_loop
,
770 if (key
->blend
.alpha_to_one
&& key
->multisample
) {
771 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
) {
772 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
773 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
774 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
775 if (outputs
[cbuf
][3]) {
776 LLVMBuildStore(builder
, lp_build_const_vec(gallivm
, type
, 1.0), outputs
[cbuf
][3]);
780 if (shader
->info
.base
.writes_samplemask
) {
781 LLVMValueRef output_smask
= NULL
;
782 int smaski
= find_output_by_semantic(&shader
->info
.base
,
783 TGSI_SEMANTIC_SAMPLEMASK
,
785 struct lp_build_context smask_bld
;
786 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
789 output_smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
790 output_smask
= LLVMBuildBitCast(builder
, output_smask
, smask_bld
.vec_type
, "");
792 if (key
->min_samples
> 1) {
793 /* only the bit corresponding to this sample is to be used. */
794 LLVMValueRef tmp_mask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "tmp_mask");
795 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
796 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
), "");
797 output_smask
= LLVMBuildOr(builder
, tmp_mask
, smask_bit
, "");
800 LLVMBuildStore(builder
, output_smask
, out_sample_mask_storage
);
803 /* Color write - per fragment sample */
804 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
806 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
807 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
808 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
810 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
811 if(outputs
[attrib
][chan
]) {
812 /* XXX: just initialize outputs to point at colors[] and
815 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
816 LLVMValueRef color_ptr
;
817 LLVMValueRef color_idx
= loop_state
.counter
;
818 if (key
->min_samples
> 1)
819 color_idx
= LLVMBuildAdd(builder
, color_idx
,
820 LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, ""), "");
821 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
823 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
824 LLVMBuildStore(builder
, out
, color_ptr
);
830 if (key
->multisample
&& key
->min_samples
> 1) {
831 LLVMBuildStore(builder
, lp_build_mask_value(&mask
), s_mask_ptr
);
832 lp_build_for_loop_end(&sample_loop_state
);
835 if (key
->multisample
) {
836 /* execute depth test for each sample */
837 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
838 lp_build_const_int32(gallivm
, 0),
839 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
840 lp_build_const_int32(gallivm
, 1));
842 /* load the per-sample coverage mask */
843 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
844 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
845 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
847 /* combine the execution mask post fragment shader with the coverage mask. */
848 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
849 if (key
->min_samples
== 1)
850 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
852 /* if the shader writes sample mask use that */
853 if (shader
->info
.base
.writes_samplemask
) {
854 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
855 out_smask_idx
= lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
);
856 LLVMValueRef output_smask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "");
857 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, out_smask_idx
, "");
858 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int_vec(gallivm
, int_type
, 0), "");
859 smask_bit
= LLVMBuildSExt(builder
, cmp
, int_vec_type
, "");
861 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
865 depth_ptr
= depth_base_ptr
;
866 if (key
->multisample
) {
867 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
868 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
872 if (depth_mode
& LATE_DEPTH_TEST
) {
873 int pos0
= find_output_by_semantic(&shader
->info
.base
,
874 TGSI_SEMANTIC_POSITION
,
876 int s_out
= find_output_by_semantic(&shader
->info
.base
,
877 TGSI_SEMANTIC_STENCIL
,
879 if (pos0
!= -1 && outputs
[pos0
][2]) {
880 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
883 * Clamp according to ARB_depth_clamp semantics.
885 if (key
->depth_clamp
) {
886 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
890 if (s_out
!= -1 && outputs
[s_out
][1]) {
891 /* there's only one value, and spec says to discard additional bits */
892 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
893 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
894 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
895 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
896 stencil_refs
[1] = stencil_refs
[0];
899 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
900 zs_format_desc
, key
->resource_1d
,
901 depth_ptr
, depth_stride
,
902 &z_fb
, &s_fb
, loop_state
.counter
);
904 lp_build_depth_stencil_test(gallivm
,
909 key
->multisample
? NULL
: &mask
,
917 if (depth_mode
& LATE_DEPTH_WRITE
) {
918 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
919 zs_format_desc
, key
->resource_1d
,
920 NULL
, NULL
, NULL
, loop_state
.counter
,
921 depth_ptr
, depth_stride
,
925 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
926 (depth_mode
& LATE_DEPTH_WRITE
))
928 /* Need to apply a reduced mask to the depth write. Reload the
929 * depth value, update from zs_value with the new mask value and
932 if (key
->multisample
) {
933 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
934 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
935 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
936 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
938 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
939 zs_format_desc
, key
->resource_1d
,
940 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
941 depth_ptr
, depth_stride
,
945 if (key
->occlusion_count
) {
946 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
947 lp_build_name(counter
, "counter");
949 lp_build_occlusion_count(gallivm
, type
,
950 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), counter
);
953 if (key
->multisample
) {
954 /* store the sample mask for this loop */
955 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
956 lp_build_for_loop_end(&sample_loop_state
);
959 mask_val
= lp_build_mask_end(&mask
);
960 if (!key
->multisample
)
961 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
962 lp_build_for_loop_end(&loop_state
);
967 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
969 * Fragment Shader outputs pixels in small 2x2 blocks
970 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
972 * However in memory pixels are stored in rows
973 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
975 * @param type fragment shader type (4x or 8x float)
976 * @param num_fs number of fs_src
977 * @param is_1d whether we're outputting to a 1d resource
978 * @param dst_channels number of output channels
979 * @param fs_src output from fragment shader
980 * @param dst pointer to store result
981 * @param pad_inline is channel padding inline or at end of row
982 * @return the number of dsts
985 generate_fs_twiddle(struct gallivm_state
*gallivm
,
988 unsigned dst_channels
,
989 LLVMValueRef fs_src
[][4],
993 LLVMValueRef src
[16];
999 unsigned pixels
= type
.length
/ 4;
1000 unsigned reorder_group
;
1001 unsigned src_channels
;
1005 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1006 src_count
= num_fs
* src_channels
;
1008 assert(pixels
== 2 || pixels
== 1);
1009 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
1012 * Transpose from SoA -> AoS
1014 for (i
= 0; i
< num_fs
; ++i
) {
1015 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
1019 * Pick transformation options
1021 swizzle_pad
= false;
1026 if (dst_channels
== 1) {
1032 } else if (dst_channels
== 2) {
1036 } else if (dst_channels
> 2) {
1043 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
1049 * Split the src in half
1052 for (i
= num_fs
; i
> 0; --i
) {
1053 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
1054 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
1062 * Ensure pixels are in memory order
1064 if (reorder_group
) {
1065 /* Twiddle pixels by reordering the array, e.g.:
1067 * src_count = 8 -> 0 2 1 3 4 6 5 7
1068 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1070 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
1072 for (i
= 0; i
< src_count
; ++i
) {
1073 unsigned group
= i
/ reorder_group
;
1074 unsigned block
= (group
/ 4) * 4 * reorder_group
;
1075 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
1078 } else if (twiddle
) {
1079 /* Twiddle pixels across elements of array */
1081 * XXX: we should avoid this in some cases, but would need to tell
1082 * lp_build_conv to reorder (or deal with it ourselves).
1084 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
1087 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
1091 * Moves any padding between pixels to the end
1092 * e.g. RGBXRGBX -> RGBRGBXX
1095 unsigned char swizzles
[16];
1096 unsigned elems
= pixels
* dst_channels
;
1098 for (i
= 0; i
< type
.length
; ++i
) {
1100 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
1102 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
1105 for (i
= 0; i
< src_count
; ++i
) {
1106 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
1115 * Untwiddle and transpose, much like the above.
1116 * However, this is after conversion, so we get packed vectors.
1117 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1118 * the vectors will look like:
1119 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1120 * be swizzled here). Extending to 16bit should be trivial.
1121 * Should also be extended to handle twice wide vectors with AVX2...
1124 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
1125 struct lp_type type
,
1131 struct lp_type type64
, type16
, type32
;
1132 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
1133 LLVMBuilderRef builder
= gallivm
->builder
;
1134 LLVMValueRef tmp
[4], shuf
[8];
1135 for (j
= 0; j
< 2; j
++) {
1136 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
1137 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
1138 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
1139 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
1142 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
1143 assert(type
.width
== 8);
1144 assert(type
.length
== 16);
1146 type8_t
= lp_build_vec_type(gallivm
, type
);
1151 type64_t
= lp_build_vec_type(gallivm
, type64
);
1156 type16_t
= lp_build_vec_type(gallivm
, type16
);
1161 type32_t
= lp_build_vec_type(gallivm
, type32
);
1163 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1165 if (src_count
== 1) {
1166 /* transpose was no-op, just untwiddle */
1167 LLVMValueRef shuf_vec
;
1168 shuf_vec
= LLVMConstVector(shuf
, 8);
1169 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1170 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1171 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1172 } else if (src_count
== 2) {
1173 LLVMValueRef shuf_vec
;
1174 shuf_vec
= LLVMConstVector(shuf
, 4);
1176 for (i
= 0; i
< 2; i
++) {
1177 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1178 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1179 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1182 for (j
= 0; j
< 2; j
++) {
1183 LLVMValueRef lo
, hi
, lo2
, hi2
;
1185 * Note that if we only really have 3 valid channels (rgb)
1186 * and we don't need alpha we could substitute a undef here
1187 * for the respective channel (causing llvm to drop conversion
1190 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1191 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1192 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1193 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1194 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1195 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1196 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1203 * Load an unswizzled block of pixels from memory
1206 load_unswizzled_block(struct gallivm_state
*gallivm
,
1207 LLVMValueRef base_ptr
,
1208 LLVMValueRef stride
,
1209 unsigned block_width
,
1210 unsigned block_height
,
1212 struct lp_type dst_type
,
1214 unsigned dst_alignment
)
1216 LLVMBuilderRef builder
= gallivm
->builder
;
1217 unsigned row_size
= dst_count
/ block_height
;
1220 /* Ensure block exactly fits into dst */
1221 assert((block_width
* block_height
) % dst_count
== 0);
1223 for (i
= 0; i
< dst_count
; ++i
) {
1224 unsigned x
= i
% row_size
;
1225 unsigned y
= i
/ row_size
;
1227 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1228 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1230 LLVMValueRef gep
[2];
1231 LLVMValueRef dst_ptr
;
1233 gep
[0] = lp_build_const_int32(gallivm
, 0);
1234 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1236 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1237 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1238 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1240 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1242 LLVMSetAlignment(dst
[i
], dst_alignment
);
1248 * Store an unswizzled block of pixels to memory
1251 store_unswizzled_block(struct gallivm_state
*gallivm
,
1252 LLVMValueRef base_ptr
,
1253 LLVMValueRef stride
,
1254 unsigned block_width
,
1255 unsigned block_height
,
1257 struct lp_type src_type
,
1259 unsigned src_alignment
)
1261 LLVMBuilderRef builder
= gallivm
->builder
;
1262 unsigned row_size
= src_count
/ block_height
;
1265 /* Ensure src exactly fits into block */
1266 assert((block_width
* block_height
) % src_count
== 0);
1268 for (i
= 0; i
< src_count
; ++i
) {
1269 unsigned x
= i
% row_size
;
1270 unsigned y
= i
/ row_size
;
1272 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1273 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1275 LLVMValueRef gep
[2];
1276 LLVMValueRef src_ptr
;
1278 gep
[0] = lp_build_const_int32(gallivm
, 0);
1279 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1281 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1282 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1283 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1285 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1287 LLVMSetAlignment(src_ptr
, src_alignment
);
1293 * Checks if a format description is an arithmetic format
1295 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1297 static inline boolean
1298 is_arithmetic_format(const struct util_format_description
*format_desc
)
1300 boolean arith
= false;
1303 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1304 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1305 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1313 * Checks if this format requires special handling due to required expansion
1314 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1317 static inline boolean
1318 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1320 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1321 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1329 * Retrieves the type representing the memory layout for a format
1331 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1334 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1335 struct lp_type
* type
)
1340 if (format_expands_to_float_soa(format_desc
)) {
1341 /* just make this a uint with width of block */
1342 type
->floating
= false;
1343 type
->fixed
= false;
1346 type
->width
= format_desc
->block
.bits
;
1351 for (i
= 0; i
< 4; i
++)
1352 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1356 memset(type
, 0, sizeof(struct lp_type
));
1357 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1358 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1359 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1360 type
->norm
= format_desc
->channel
[chan
].normalized
;
1362 if (is_arithmetic_format(format_desc
)) {
1366 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1367 type
->width
+= format_desc
->channel
[i
].size
;
1370 type
->width
= format_desc
->channel
[chan
].size
;
1371 type
->length
= format_desc
->nr_channels
;
1377 * Retrieves the type for a format which is usable in the blending code.
1379 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1382 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1383 struct lp_type
* type
)
1388 if (format_expands_to_float_soa(format_desc
)) {
1389 /* always use ordinary floats for blending */
1390 type
->floating
= true;
1391 type
->fixed
= false;
1399 for (i
= 0; i
< 4; i
++)
1400 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1404 memset(type
, 0, sizeof(struct lp_type
));
1405 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1406 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1407 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1408 type
->norm
= format_desc
->channel
[chan
].normalized
;
1409 type
->width
= format_desc
->channel
[chan
].size
;
1410 type
->length
= format_desc
->nr_channels
;
1412 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1413 if (format_desc
->channel
[i
].size
> type
->width
)
1414 type
->width
= format_desc
->channel
[i
].size
;
1417 if (type
->floating
) {
1420 if (type
->width
<= 8) {
1422 } else if (type
->width
<= 16) {
1429 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1436 * Scale a normalized value from src_bits to dst_bits.
1438 * The exact calculation is
1440 * dst = iround(src * dst_mask / src_mask)
1442 * or with integer rounding
1444 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1448 * src_mask = (1 << src_bits) - 1
1449 * dst_mask = (1 << dst_bits) - 1
1451 * but we try to avoid division and multiplication through shifts.
1453 static inline LLVMValueRef
1454 scale_bits(struct gallivm_state
*gallivm
,
1458 struct lp_type src_type
)
1460 LLVMBuilderRef builder
= gallivm
->builder
;
1461 LLVMValueRef result
= src
;
1463 if (dst_bits
< src_bits
) {
1464 int delta_bits
= src_bits
- dst_bits
;
1466 if (delta_bits
<= dst_bits
) {
1468 * Approximate the rescaling with a single shift.
1470 * This gives the wrong rounding.
1473 result
= LLVMBuildLShr(builder
,
1475 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1480 * Try more accurate rescaling.
1484 * Drop the least significant bits to make space for the multiplication.
1486 * XXX: A better approach would be to use a wider integer type as intermediate. But
1487 * this is enough to convert alpha from 16bits -> 2 when rendering to
1488 * PIPE_FORMAT_R10G10B10A2_UNORM.
1490 result
= LLVMBuildLShr(builder
,
1492 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1496 result
= LLVMBuildMul(builder
,
1498 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1502 * Add a rounding term before the division.
1504 * TODO: Handle signed integers too.
1506 if (!src_type
.sign
) {
1507 result
= LLVMBuildAdd(builder
,
1509 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1514 * Approximate the division by src_mask with a src_bits shift.
1516 * Given the src has already been shifted by dst_bits, all we need
1517 * to do is to shift by the difference.
1520 result
= LLVMBuildLShr(builder
,
1522 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1526 } else if (dst_bits
> src_bits
) {
1528 int db
= dst_bits
- src_bits
;
1530 /* Shift left by difference in bits */
1531 result
= LLVMBuildShl(builder
,
1533 lp_build_const_int_vec(gallivm
, src_type
, db
),
1536 if (db
<= src_bits
) {
1537 /* Enough bits in src to fill the remainder */
1538 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1540 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1543 result
= LLVMBuildOr(builder
, result
, lower
, "");
1544 } else if (db
> src_bits
) {
1545 /* Need to repeatedly copy src bits to fill remainder in dst */
1548 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1549 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1551 result
= LLVMBuildOr(builder
,
1553 LLVMBuildLShr(builder
, result
, shuv
, ""),
1563 * If RT is a smallfloat (needing denorms) format
1566 have_smallfloat_format(struct lp_type dst_type
,
1567 enum pipe_format format
)
1569 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1570 /* due to format handling hacks this format doesn't have floating set
1571 * here (and actually has width set to 32 too) so special case this. */
1572 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1577 * Convert from memory format to blending format
1579 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1582 convert_to_blend_type(struct gallivm_state
*gallivm
,
1583 unsigned block_size
,
1584 const struct util_format_description
*src_fmt
,
1585 struct lp_type src_type
,
1586 struct lp_type dst_type
,
1587 LLVMValueRef
* src
, // and dst
1590 LLVMValueRef
*dst
= src
;
1591 LLVMBuilderRef builder
= gallivm
->builder
;
1592 struct lp_type blend_type
;
1593 struct lp_type mem_type
;
1595 unsigned pixels
= block_size
/ num_srcs
;
1599 * full custom path for packed floats and srgb formats - none of the later
1600 * functions would do anything useful, and given the lp_type representation they
1601 * can't be fixed. Should really have some SoA blend path for these kind of
1602 * formats rather than hacking them in here.
1604 if (format_expands_to_float_soa(src_fmt
)) {
1605 LLVMValueRef tmpsrc
[4];
1607 * This is pretty suboptimal for this case blending in SoA would be much
1608 * better, since conversion gets us SoA values so need to convert back.
1610 assert(src_type
.width
== 32 || src_type
.width
== 16);
1611 assert(dst_type
.floating
);
1612 assert(dst_type
.width
== 32);
1613 assert(dst_type
.length
% 4 == 0);
1614 assert(num_srcs
% 4 == 0);
1616 if (src_type
.width
== 16) {
1617 /* expand 4x16bit values to 4x32bit */
1618 struct lp_type type32x4
= src_type
;
1619 LLVMTypeRef ltype32x4
;
1620 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1621 type32x4
.width
= 32;
1622 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1623 for (i
= 0; i
< num_fetch
; i
++) {
1624 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1626 src_type
.width
= 32;
1628 for (i
= 0; i
< 4; i
++) {
1631 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1632 LLVMValueRef tmpsoa
[4];
1633 LLVMValueRef tmps
= tmpsrc
[i
];
1634 if (dst_type
.length
== 8) {
1635 LLVMValueRef shuffles
[8];
1637 /* fetch was 4 values but need 8-wide output values */
1638 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1640 * for 8-wide aos transpose would give us wrong order not matching
1641 * incoming converted fs values and mask. ARGH.
1643 for (j
= 0; j
< 4; j
++) {
1644 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1645 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1647 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1648 LLVMConstVector(shuffles
, 8), "");
1650 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1651 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1654 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1656 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1661 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1662 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1664 /* Is the format arithmetic */
1665 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1666 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1668 /* Pad if necessary */
1669 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1670 for (i
= 0; i
< num_srcs
; ++i
) {
1671 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1674 src_type
.length
= dst_type
.length
;
1677 /* Special case for half-floats */
1678 if (mem_type
.width
== 16 && mem_type
.floating
) {
1679 assert(blend_type
.width
== 32 && blend_type
.floating
);
1680 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1688 src_type
.width
= blend_type
.width
* blend_type
.length
;
1689 blend_type
.length
*= pixels
;
1690 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1692 for (i
= 0; i
< num_srcs
; ++i
) {
1693 LLVMValueRef chans
[4];
1694 LLVMValueRef res
= NULL
;
1696 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1698 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1700 unsigned sa
= src_fmt
->channel
[j
].shift
;
1701 #if UTIL_ARCH_LITTLE_ENDIAN
1702 unsigned from_lsb
= j
;
1704 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1707 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1709 /* Extract bits from source */
1710 chans
[j
] = LLVMBuildLShr(builder
,
1712 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1715 chans
[j
] = LLVMBuildAnd(builder
,
1717 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1721 if (src_type
.norm
) {
1722 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1723 blend_type
.width
, chans
[j
], src_type
);
1726 /* Insert bits into correct position */
1727 chans
[j
] = LLVMBuildShl(builder
,
1729 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1735 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1739 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1745 * Convert from blending format to memory format
1747 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1750 convert_from_blend_type(struct gallivm_state
*gallivm
,
1751 unsigned block_size
,
1752 const struct util_format_description
*src_fmt
,
1753 struct lp_type src_type
,
1754 struct lp_type dst_type
,
1755 LLVMValueRef
* src
, // and dst
1758 LLVMValueRef
* dst
= src
;
1760 struct lp_type mem_type
;
1761 struct lp_type blend_type
;
1762 LLVMBuilderRef builder
= gallivm
->builder
;
1763 unsigned pixels
= block_size
/ num_srcs
;
1767 * full custom path for packed floats and srgb formats - none of the later
1768 * functions would do anything useful, and given the lp_type representation they
1769 * can't be fixed. Should really have some SoA blend path for these kind of
1770 * formats rather than hacking them in here.
1772 if (format_expands_to_float_soa(src_fmt
)) {
1774 * This is pretty suboptimal for this case blending in SoA would be much
1775 * better - we need to transpose the AoS values back to SoA values for
1776 * conversion/packing.
1778 assert(src_type
.floating
);
1779 assert(src_type
.width
== 32);
1780 assert(src_type
.length
% 4 == 0);
1781 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1783 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1784 LLVMValueRef tmpsoa
[4], tmpdst
;
1785 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1786 /* really really need SoA here */
1788 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1789 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1792 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1796 if (src_type
.length
== 8) {
1797 LLVMValueRef tmpaos
, shuffles
[8];
1800 * for 8-wide aos transpose has given us wrong order not matching
1801 * output order. HMPF. Also need to split the output values manually.
1803 for (j
= 0; j
< 4; j
++) {
1804 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1805 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1807 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1808 LLVMConstVector(shuffles
, 8), "");
1809 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1810 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1816 if (dst_type
.width
== 16) {
1817 struct lp_type type16x8
= dst_type
;
1818 struct lp_type type32x4
= dst_type
;
1819 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1820 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1821 type16x8
.length
= 8;
1822 type32x4
.width
= 32;
1823 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1824 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1825 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1826 /* We could do vector truncation but it doesn't generate very good code */
1827 for (i
= 0; i
< num_fetch
; i
++) {
1828 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1829 src
[i
], lp_build_zero(gallivm
, type32x4
));
1830 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1831 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1832 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1838 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1839 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1841 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1843 /* Special case for half-floats */
1844 if (mem_type
.width
== 16 && mem_type
.floating
) {
1845 int length
= dst_type
.length
;
1846 assert(blend_type
.width
== 32 && blend_type
.floating
);
1848 dst_type
.length
= src_type
.length
;
1850 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1852 dst_type
.length
= length
;
1856 /* Remove any padding */
1857 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1858 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1860 for (i
= 0; i
< num_srcs
; ++i
) {
1861 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1865 /* No bit arithmetic to do */
1870 src_type
.length
= pixels
;
1871 src_type
.width
= blend_type
.length
* blend_type
.width
;
1872 dst_type
.length
= pixels
;
1874 for (i
= 0; i
< num_srcs
; ++i
) {
1875 LLVMValueRef chans
[4];
1876 LLVMValueRef res
= NULL
;
1878 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1880 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1882 unsigned sa
= src_fmt
->channel
[j
].shift
;
1883 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1884 #if UTIL_ARCH_LITTLE_ENDIAN
1885 unsigned from_lsb
= j
;
1887 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1890 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1892 for (k
= 0; k
< blend_type
.width
; ++k
) {
1897 chans
[j
] = LLVMBuildLShr(builder
,
1899 lp_build_const_int_vec(gallivm
, src_type
,
1900 from_lsb
* blend_type
.width
),
1903 chans
[j
] = LLVMBuildAnd(builder
,
1905 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1908 /* Scale down bits */
1909 if (src_type
.norm
) {
1910 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1911 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1912 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1913 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1914 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1915 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1919 chans
[j
] = LLVMBuildShl(builder
,
1921 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1924 sa
+= src_fmt
->channel
[j
].size
;
1929 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1933 assert (dst_type
.width
!= 24);
1935 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1941 * Convert alpha to same blend type as src
1944 convert_alpha(struct gallivm_state
*gallivm
,
1945 struct lp_type row_type
,
1946 struct lp_type alpha_type
,
1947 const unsigned block_size
,
1948 const unsigned block_height
,
1949 const unsigned src_count
,
1950 const unsigned dst_channels
,
1951 const bool pad_inline
,
1952 LLVMValueRef
* src_alpha
)
1954 LLVMBuilderRef builder
= gallivm
->builder
;
1956 unsigned length
= row_type
.length
;
1957 row_type
.length
= alpha_type
.length
;
1959 /* Twiddle the alpha to match pixels */
1960 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1963 * TODO this should use single lp_build_conv call for
1964 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1966 for (i
= 0; i
< block_height
; ++i
) {
1967 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1970 alpha_type
= row_type
;
1971 row_type
.length
= length
;
1973 /* If only one channel we can only need the single alpha value per pixel */
1974 if (src_count
== 1 && dst_channels
== 1) {
1976 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1978 /* If there are more srcs than rows then we need to split alpha up */
1979 if (src_count
> block_height
) {
1980 for (i
= src_count
; i
> 0; --i
) {
1981 unsigned pixels
= block_size
/ src_count
;
1982 unsigned idx
= i
- 1;
1984 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1985 (idx
* pixels
) % 4, pixels
);
1989 /* If there is a src for each pixel broadcast the alpha across whole row */
1990 if (src_count
== block_size
) {
1991 for (i
= 0; i
< src_count
; ++i
) {
1992 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1993 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1996 unsigned pixels
= block_size
/ src_count
;
1997 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1998 unsigned alpha_span
= 1;
1999 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
2001 /* Check if we need 2 src_alphas for our shuffles */
2002 if (pixels
> alpha_type
.length
) {
2006 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
2007 for (j
= 0; j
< row_type
.length
; ++j
) {
2008 if (j
< pixels
* channels
) {
2009 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
2011 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
2015 for (i
= 0; i
< src_count
; ++i
) {
2016 unsigned idx1
= i
, idx2
= i
;
2018 if (alpha_span
> 1){
2023 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
2026 LLVMConstVector(shuffles
, row_type
.length
),
2035 * Generates the blend function for unswizzled colour buffers
2036 * Also generates the read & write from colour buffer
2039 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
2041 struct lp_fragment_shader_variant
*variant
,
2042 enum pipe_format out_format
,
2043 unsigned int num_fs
,
2044 struct lp_type fs_type
,
2045 LLVMValueRef
* fs_mask
,
2046 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
2047 LLVMValueRef context_ptr
,
2048 LLVMValueRef color_ptr
,
2049 LLVMValueRef stride
,
2050 unsigned partial_mask
,
2053 const unsigned alpha_channel
= 3;
2054 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
2055 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
2056 const unsigned block_size
= block_width
* block_height
;
2057 const unsigned lp_integer_vector_width
= 128;
2059 LLVMBuilderRef builder
= gallivm
->builder
;
2060 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
2061 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
2062 LLVMValueRef src_alpha
[4 * 4];
2063 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
2064 LLVMValueRef src_mask
[4 * 4];
2065 LLVMValueRef src
[4 * 4];
2066 LLVMValueRef src1
[4 * 4];
2067 LLVMValueRef dst
[4 * 4];
2068 LLVMValueRef blend_color
;
2069 LLVMValueRef blend_alpha
;
2070 LLVMValueRef i32_zero
;
2071 LLVMValueRef check_mask
;
2072 LLVMValueRef undef_src_val
;
2074 struct lp_build_mask_context mask_ctx
;
2075 struct lp_type mask_type
;
2076 struct lp_type blend_type
;
2077 struct lp_type row_type
;
2078 struct lp_type dst_type
;
2079 struct lp_type ls_type
;
2081 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
2082 unsigned vector_width
;
2083 unsigned src_channels
= TGSI_NUM_CHANNELS
;
2084 unsigned dst_channels
;
2089 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
2091 unsigned dst_alignment
;
2093 bool pad_inline
= is_arithmetic_format(out_format_desc
);
2094 bool has_alpha
= false;
2095 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
2096 util_blend_state_is_dual(&variant
->key
.blend
, 0);
2098 const boolean is_1d
= variant
->key
.resource_1d
;
2099 boolean twiddle_after_convert
= FALSE
;
2100 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
2101 LLVMValueRef fpstate
= 0;
2103 /* Get type from output format */
2104 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
2105 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
2108 * Technically this code should go into lp_build_smallfloat_to_float
2109 * and lp_build_float_to_smallfloat but due to the
2110 * http://llvm.org/bugs/show_bug.cgi?id=6393
2111 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2112 * So the ordering is important here and there shouldn't be any
2113 * llvm ir instrunctions in this function before
2114 * this, otherwise half-float format conversions won't work
2115 * (again due to llvm bug #6393).
2117 if (have_smallfloat_format(dst_type
, out_format
)) {
2118 /* We need to make sure that denorms are ok for half float
2120 fpstate
= lp_build_fpstate_get(gallivm
);
2121 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
2124 mask_type
= lp_int32_vec4_type();
2125 mask_type
.length
= fs_type
.length
;
2127 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
2128 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
2131 /* Do not bother executing code when mask is empty.. */
2133 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
2135 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2136 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
2139 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
2140 lp_build_mask_check(&mask_ctx
);
2143 partial_mask
|= !variant
->opaque
;
2144 i32_zero
= lp_build_const_int32(gallivm
, 0);
2146 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2148 row_type
.length
= fs_type
.length
;
2149 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2151 /* Compute correct swizzle and count channels */
2152 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2155 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2156 /* Ensure channel is used */
2157 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2161 /* Ensure not already written to (happens in case with GL_ALPHA) */
2162 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2166 /* Ensure we havn't already found all channels */
2167 if (dst_channels
>= out_format_desc
->nr_channels
) {
2171 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2174 if (i
== alpha_channel
) {
2179 if (format_expands_to_float_soa(out_format_desc
)) {
2181 * the code above can't work for layout_other
2182 * for srgb it would sort of work but we short-circuit swizzles, etc.
2183 * as that is done as part of unpack / pack.
2185 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2191 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2194 /* If 3 channels then pad to include alpha for 4 element transpose */
2195 if (dst_channels
== 3) {
2196 assert (!has_alpha
);
2197 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2198 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2201 if (out_format_desc
->nr_channels
== 4) {
2204 * We use alpha from the color conversion, not separate one.
2205 * We had to include it for transpose, hence it will get converted
2206 * too (albeit when doing transpose after conversion, that would
2207 * no longer be the case necessarily).
2208 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2209 * otherwise we really have padding, not alpha, included.)
2216 * Load shader output
2218 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2219 /* Always load alpha for use in blending */
2222 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2225 alpha
= undef_src_val
;
2228 /* Load each channel */
2229 for (j
= 0; j
< dst_channels
; ++j
) {
2230 assert(swizzle
[j
] < 4);
2232 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2235 fs_src
[i
][j
] = undef_src_val
;
2239 /* If 3 channels then pad to include alpha for 4 element transpose */
2241 * XXX If we include that here maybe could actually use it instead of
2242 * separate alpha for blending?
2243 * (Difficult though we actually convert pad channels, not alpha.)
2245 if (dst_channels
== 3 && !has_alpha
) {
2246 fs_src
[i
][3] = alpha
;
2249 /* We split the row_mask and row_alpha as we want 128bit interleave */
2250 if (fs_type
.length
== 8) {
2251 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2253 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2254 src_channels
, src_channels
);
2256 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2257 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2258 src_channels
, src_channels
);
2260 src_mask
[i
] = fs_mask
[i
];
2261 src_alpha
[i
] = alpha
;
2264 if (dual_source_blend
) {
2265 /* same as above except different src/dst, skip masks and comments... */
2266 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2269 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2272 alpha
= undef_src_val
;
2275 for (j
= 0; j
< dst_channels
; ++j
) {
2276 assert(swizzle
[j
] < 4);
2278 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2281 fs_src1
[i
][j
] = undef_src_val
;
2284 if (dst_channels
== 3 && !has_alpha
) {
2285 fs_src1
[i
][3] = alpha
;
2287 if (fs_type
.length
== 8) {
2288 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2289 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2290 src_channels
, src_channels
);
2292 src1_alpha
[i
] = alpha
;
2297 if (util_format_is_pure_integer(out_format
)) {
2299 * In this case fs_type was really ints or uints disguised as floats,
2302 fs_type
.floating
= 0;
2303 fs_type
.sign
= dst_type
.sign
;
2304 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2305 for (j
= 0; j
< dst_channels
; ++j
) {
2306 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2307 lp_build_vec_type(gallivm
, fs_type
), "");
2309 if (dst_channels
== 3 && !has_alpha
) {
2310 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2311 lp_build_vec_type(gallivm
, fs_type
), "");
2317 * We actually should generally do conversion first (for non-1d cases)
2318 * when the blend format is 8 or 16 bits. The reason is obvious,
2319 * there's 2 or 4 times less vectors to deal with for the interleave...
2320 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2321 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2322 * unpack only with 128bit vectors).
2323 * Note: for 16bit sizes really need matching pack conversion code
2325 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2326 twiddle_after_convert
= TRUE
;
2330 * Pixel twiddle from fragment shader order to memory order
2332 if (!twiddle_after_convert
) {
2333 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2334 dst_channels
, fs_src
, src
, pad_inline
);
2335 if (dual_source_blend
) {
2336 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2337 fs_src1
, src1
, pad_inline
);
2340 src_count
= num_fullblock_fs
* dst_channels
;
2342 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2343 * (AVX) turn out the same later when untwiddling/transpose (albeit
2344 * for true AVX2 path untwiddle needs to be different).
2345 * For now just order by colors first (so we can use unpack later).
2347 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2348 for (i
= 0; i
< dst_channels
; i
++) {
2349 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2350 if (dual_source_blend
) {
2351 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2357 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2358 if (src_count
!= num_fullblock_fs
* src_channels
) {
2359 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2360 row_type
.length
/= ds
;
2361 fs_type
.length
= row_type
.length
;
2364 blend_type
= row_type
;
2365 mask_type
.length
= 4;
2367 /* Convert src to row_type */
2368 if (dual_source_blend
) {
2369 struct lp_type old_row_type
= row_type
;
2370 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2371 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2374 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2377 /* If the rows are not an SSE vector, combine them to become SSE size! */
2378 if ((row_type
.width
* row_type
.length
) % 128) {
2379 unsigned bits
= row_type
.width
* row_type
.length
;
2382 assert(src_count
>= (vector_width
/ bits
));
2384 dst_count
= src_count
/ (vector_width
/ bits
);
2386 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2387 if (dual_source_blend
) {
2388 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2391 row_type
.length
*= combined
;
2392 src_count
/= combined
;
2394 bits
= row_type
.width
* row_type
.length
;
2395 assert(bits
== 128 || bits
== 256);
2398 if (twiddle_after_convert
) {
2399 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2400 if (dual_source_blend
) {
2401 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2406 * Blend Colour conversion
2408 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2409 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2410 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2411 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2412 &i32_zero
, 1, ""), "");
2415 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2417 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2419 * since blending is done with floats, there was no conversion.
2420 * However, the rules according to fixed point renderbuffers still
2421 * apply, that is we must clamp inputs to 0.0/1.0.
2422 * (This would apply to separate alpha conversion too but we currently
2423 * force has_alpha to be true.)
2424 * TODO: should skip this with "fake" blend, since post-blend conversion
2425 * will clamp anyway.
2426 * TODO: could also skip this if fragment color clamping is enabled. We
2427 * don't support it natively so it gets baked into the shader however, so
2428 * can't really tell here.
2430 struct lp_build_context f32_bld
;
2431 assert(row_type
.floating
);
2432 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2433 for (i
= 0; i
< src_count
; i
++) {
2434 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2436 if (dual_source_blend
) {
2437 for (i
= 0; i
< src_count
; i
++) {
2438 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2441 /* probably can't be different than row_type but better safe than sorry... */
2442 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2443 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2447 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2449 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2450 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2452 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2453 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2455 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2456 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2462 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2464 if (src_count
< block_height
) {
2465 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2466 } else if (src_count
> block_height
) {
2467 for (i
= src_count
; i
> 0; --i
) {
2468 unsigned pixels
= block_size
/ src_count
;
2469 unsigned idx
= i
- 1;
2471 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2472 (idx
* pixels
) % 4, pixels
);
2476 assert(mask_type
.width
== 32);
2478 for (i
= 0; i
< src_count
; ++i
) {
2479 unsigned pixels
= block_size
/ src_count
;
2480 unsigned pixel_width
= row_type
.width
* dst_channels
;
2482 if (pixel_width
== 24) {
2483 mask_type
.width
= 8;
2484 mask_type
.length
= vector_width
/ mask_type
.width
;
2486 mask_type
.length
= pixels
;
2487 mask_type
.width
= row_type
.width
* dst_channels
;
2490 * If mask_type width is smaller than 32bit, this doesn't quite
2491 * generate the most efficient code (could use some pack).
2493 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2494 lp_build_int_vec_type(gallivm
, mask_type
), "");
2496 mask_type
.length
*= dst_channels
;
2497 mask_type
.width
/= dst_channels
;
2500 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2501 lp_build_int_vec_type(gallivm
, mask_type
), "");
2502 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2509 struct lp_type alpha_type
= fs_type
;
2510 alpha_type
.length
= 4;
2511 convert_alpha(gallivm
, row_type
, alpha_type
,
2512 block_size
, block_height
,
2513 src_count
, dst_channels
,
2514 pad_inline
, src_alpha
);
2515 if (dual_source_blend
) {
2516 convert_alpha(gallivm
, row_type
, alpha_type
,
2517 block_size
, block_height
,
2518 src_count
, dst_channels
,
2519 pad_inline
, src1_alpha
);
2525 * Load dst from memory
2527 if (src_count
< block_height
) {
2528 dst_count
= block_height
;
2530 dst_count
= src_count
;
2533 dst_type
.length
*= block_size
/ dst_count
;
2535 if (format_expands_to_float_soa(out_format_desc
)) {
2537 * we need multiple values at once for the conversion, so can as well
2538 * load them vectorized here too instead of concatenating later.
2539 * (Still need concatenation later for 8-wide vectors).
2541 dst_count
= block_height
;
2542 dst_type
.length
= block_width
;
2546 * Compute the alignment of the destination pointer in bytes
2547 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2548 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2549 * 1d tex but can't distinguish here) so need to stick with per-pixel
2550 * alignment in this case.
2553 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2556 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2558 /* Force power-of-two alignment by extracting only the least-significant-bit */
2559 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2561 * Resource base and stride pointers are aligned to 16 bytes, so that's
2562 * the maximum alignment we can guarantee
2564 dst_alignment
= MIN2(16, dst_alignment
);
2568 if (dst_count
> src_count
) {
2569 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2570 util_is_power_of_two_or_zero(dst_type
.length
) &&
2571 dst_type
.length
* dst_type
.width
< 128) {
2573 * Never try to load values as 4xi8 which we will then
2574 * concatenate to larger vectors. This gives llvm a real
2575 * headache (the problem is the type legalizer (?) will
2576 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2577 * then the shuffles to concatenate are more or less impossible
2578 * - llvm is easily capable of generating a sequence of 32
2579 * pextrb/pinsrb instructions for that. Albeit it appears to
2580 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2581 * width to avoid the trouble (16bit seems not as bad, llvm
2582 * probably recognizes the load+shuffle as only one shuffle
2583 * is necessary, but we can do just the same anyway).
2585 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2591 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2592 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2593 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2594 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2599 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2600 dst
, ls_type
, dst_count
, dst_alignment
);
2605 * Convert from dst/output format to src/blending format.
2607 * This is necessary as we can only read 1 row from memory at a time,
2608 * so the minimum dst_count will ever be at this point is 4.
2610 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2611 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2612 * on all 16 pixels in that single vector at once.
2614 if (dst_count
> src_count
) {
2615 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2616 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2617 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2618 for (i
= 0; i
< dst_count
; i
++) {
2619 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2623 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2625 if (ls_type
.length
!= dst_type
.length
) {
2626 struct lp_type tmp_type
= dst_type
;
2627 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2628 for (i
= 0; i
< src_count
; i
++) {
2629 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2630 lp_build_vec_type(gallivm
, tmp_type
), "");
2638 /* XXX this is broken for RGB8 formats -
2639 * they get expanded from 12 to 16 elements (to include alpha)
2640 * by convert_to_blend_type then reduced to 15 instead of 12
2641 * by convert_from_blend_type (a simple fix though breaks A8...).
2642 * R16G16B16 also crashes differently however something going wrong
2643 * inside llvm handling npot vector sizes seemingly.
2644 * It seems some cleanup could be done here (like skipping conversion/blend
2647 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2648 row_type
, dst
, src_count
);
2651 * FIXME: Really should get logic ops / masks out of generic blend / row
2652 * format. Logic ops will definitely not work on the blend float format
2653 * used for SRGB here and I think OpenGL expects this to work as expected
2654 * (that is incoming values converted to srgb then logic op applied).
2656 for (i
= 0; i
< src_count
; ++i
) {
2657 dst
[i
] = lp_build_blend_aos(gallivm
,
2658 &variant
->key
.blend
,
2663 has_alpha
? NULL
: src_alpha
[i
],
2665 has_alpha
? NULL
: src1_alpha
[i
],
2667 partial_mask
? src_mask
[i
] : NULL
,
2669 has_alpha
? NULL
: blend_alpha
,
2671 pad_inline
? 4 : dst_channels
);
2674 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2675 row_type
, dst_type
, dst
, src_count
);
2677 /* Split the blend rows back to memory rows */
2678 if (dst_count
> src_count
) {
2679 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2681 if (src_count
== 1) {
2682 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2683 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2685 row_type
.length
/= 2;
2689 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2690 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2691 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2692 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2694 row_type
.length
/= 2;
2699 * Store blend result to memory
2702 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2703 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2706 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2707 dst
, dst_type
, dst_count
, dst_alignment
);
2710 if (have_smallfloat_format(dst_type
, out_format
)) {
2711 lp_build_fpstate_set(gallivm
, fpstate
);
2715 lp_build_mask_end(&mask_ctx
);
2721 * Generate the runtime callable function for the whole fragment pipeline.
2722 * Note that the function which we generate operates on a block of 16
2723 * pixels at at time. The block contains 2x2 quads. Each quad contains
2727 generate_fragment(struct llvmpipe_context
*lp
,
2728 struct lp_fragment_shader
*shader
,
2729 struct lp_fragment_shader_variant
*variant
,
2730 unsigned partial_mask
)
2732 struct gallivm_state
*gallivm
= variant
->gallivm
;
2733 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2734 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2736 struct lp_type fs_type
;
2737 struct lp_type blend_type
;
2738 LLVMTypeRef fs_elem_type
;
2739 LLVMTypeRef blend_vec_type
;
2740 LLVMTypeRef arg_types
[15];
2741 LLVMTypeRef func_type
;
2742 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2743 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2744 LLVMValueRef context_ptr
;
2747 LLVMValueRef a0_ptr
;
2748 LLVMValueRef dadx_ptr
;
2749 LLVMValueRef dady_ptr
;
2750 LLVMValueRef color_ptr_ptr
;
2751 LLVMValueRef stride_ptr
;
2752 LLVMValueRef color_sample_stride_ptr
;
2753 LLVMValueRef depth_ptr
;
2754 LLVMValueRef depth_stride
;
2755 LLVMValueRef depth_sample_stride
;
2756 LLVMValueRef mask_input
;
2757 LLVMValueRef thread_data_ptr
;
2758 LLVMBasicBlockRef block
;
2759 LLVMBuilderRef builder
;
2760 struct lp_build_sampler_soa
*sampler
;
2761 struct lp_build_image_soa
*image
;
2762 struct lp_build_interp_soa_context interp
;
2763 LLVMValueRef fs_mask
[(16 / 4) * LP_MAX_SAMPLES
];
2764 LLVMValueRef fs_out_color
[LP_MAX_SAMPLES
][PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2765 LLVMValueRef function
;
2766 LLVMValueRef facing
;
2771 boolean cbuf0_write_all
;
2772 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2773 util_blend_state_is_dual(&key
->blend
, 0);
2775 assert(lp_native_vector_width
/ 32 >= 4);
2777 /* Adjust color input interpolation according to flatshade state:
2779 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2780 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2781 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2783 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2785 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2789 /* check if writes to cbuf[0] are to be copied to all cbufs */
2791 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2793 /* TODO: actually pick these based on the fs and color buffer
2794 * characteristics. */
2796 memset(&fs_type
, 0, sizeof fs_type
);
2797 fs_type
.floating
= TRUE
; /* floating point values */
2798 fs_type
.sign
= TRUE
; /* values are signed */
2799 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2800 fs_type
.width
= 32; /* 32-bit float */
2801 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2803 memset(&blend_type
, 0, sizeof blend_type
);
2804 blend_type
.floating
= FALSE
; /* values are integers */
2805 blend_type
.sign
= FALSE
; /* values are unsigned */
2806 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2807 blend_type
.width
= 8; /* 8-bit ubyte values */
2808 blend_type
.length
= 16; /* 16 elements per vector */
2811 * Generate the function prototype. Any change here must be reflected in
2812 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2815 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2817 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2819 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2820 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2822 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2823 arg_types
[1] = int32_type
; /* x */
2824 arg_types
[2] = int32_type
; /* y */
2825 arg_types
[3] = int32_type
; /* facing */
2826 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2827 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2828 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2829 arg_types
[7] = LLVMPointerType(LLVMPointerType(int8_type
, 0), 0); /* color */
2830 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2831 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2832 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2833 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2834 arg_types
[12] = int32_type
; /* depth_stride */
2835 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2836 arg_types
[14] = int32_type
; /* depth sample stride */
2838 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2839 arg_types
, ARRAY_SIZE(arg_types
), 0);
2841 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2842 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2844 variant
->function
[partial_mask
] = function
;
2846 /* XXX: need to propagate noalias down into color param now we are
2847 * passing a pointer-to-pointer?
2849 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2850 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2851 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2853 context_ptr
= LLVMGetParam(function
, 0);
2854 x
= LLVMGetParam(function
, 1);
2855 y
= LLVMGetParam(function
, 2);
2856 facing
= LLVMGetParam(function
, 3);
2857 a0_ptr
= LLVMGetParam(function
, 4);
2858 dadx_ptr
= LLVMGetParam(function
, 5);
2859 dady_ptr
= LLVMGetParam(function
, 6);
2860 color_ptr_ptr
= LLVMGetParam(function
, 7);
2861 depth_ptr
= LLVMGetParam(function
, 8);
2862 mask_input
= LLVMGetParam(function
, 9);
2863 thread_data_ptr
= LLVMGetParam(function
, 10);
2864 stride_ptr
= LLVMGetParam(function
, 11);
2865 depth_stride
= LLVMGetParam(function
, 12);
2866 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2867 depth_sample_stride
= LLVMGetParam(function
, 14);
2869 lp_build_name(context_ptr
, "context");
2870 lp_build_name(x
, "x");
2871 lp_build_name(y
, "y");
2872 lp_build_name(a0_ptr
, "a0");
2873 lp_build_name(dadx_ptr
, "dadx");
2874 lp_build_name(dady_ptr
, "dady");
2875 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2876 lp_build_name(depth_ptr
, "depth");
2877 lp_build_name(mask_input
, "mask_input");
2878 lp_build_name(thread_data_ptr
, "thread_data");
2879 lp_build_name(stride_ptr
, "stride_ptr");
2880 lp_build_name(depth_stride
, "depth_stride");
2881 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2882 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2888 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2889 builder
= gallivm
->builder
;
2891 LLVMPositionBuilderAtEnd(builder
, block
);
2894 * Must not count ps invocations if there's a null shader.
2895 * (It would be ok to count with null shader if there's d/s tests,
2896 * but only if there's d/s buffers too, which is different
2897 * to implicit rasterization disable which must not depend
2898 * on the d/s buffers.)
2899 * Could use popcount on mask, but pixel accuracy is not required.
2900 * Could disable if there's no stats query, but maybe not worth it.
2902 if (shader
->info
.base
.num_instructions
> 1) {
2903 LLVMValueRef invocs
, val
;
2904 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2905 val
= LLVMBuildLoad(builder
, invocs
, "");
2906 val
= LLVMBuildAdd(builder
, val
,
2907 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2909 LLVMBuildStore(builder
, val
, invocs
);
2912 /* code generated texture sampling */
2913 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2914 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2916 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2917 /* for 1d resources only run "upper half" of stamp */
2918 if (key
->resource_1d
)
2922 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2923 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2924 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2925 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2926 num_loop_samp
, "mask_store");
2928 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(gallivm
->context
);
2929 LLVMValueRef glob_sample_pos
= LLVMAddGlobal(gallivm
->module
, LLVMArrayType(flt_type
, key
->coverage_samples
* 2), "");
2930 LLVMValueRef sample_pos_array
;
2932 if (key
->multisample
&& key
->coverage_samples
== 4) {
2933 LLVMValueRef sample_pos_arr
[8];
2934 for (unsigned i
= 0; i
< 4; i
++) {
2935 sample_pos_arr
[i
* 2] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][0]);
2936 sample_pos_arr
[i
* 2 + 1] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][1]);
2938 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 8);
2940 LLVMValueRef sample_pos_arr
[2];
2941 sample_pos_arr
[0] = LLVMConstReal(flt_type
, 0.5);
2942 sample_pos_arr
[1] = LLVMConstReal(flt_type
, 0.5);
2943 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 2);
2945 LLVMSetInitializer(glob_sample_pos
, sample_pos_array
);
2947 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2948 boolean pixel_center_integer
=
2949 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2952 * The shader input interpolation info is not explicitely baked in the
2953 * shader key, but everything it derives from (TGSI, and flatshade) is
2954 * already included in the shader key.
2956 lp_build_interp_soa_init(&interp
,
2958 shader
->info
.base
.num_inputs
,
2960 pixel_center_integer
,
2961 key
->coverage_samples
, glob_sample_pos
,
2965 a0_ptr
, dadx_ptr
, dady_ptr
,
2968 for (i
= 0; i
< num_fs
; i
++) {
2969 if (key
->multisample
) {
2970 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2973 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2974 * store to the per sample mask storage. Or all of them together to generate
2975 * the fragment shader mask. (sample shading TODO).
2976 * Take the incoming state coverage mask into account.
2978 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2979 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2980 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2981 &sindexi
, 1, "sample_mask_ptr");
2982 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2983 i
*fs_type
.length
/4, s
, mask_input
);
2985 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2986 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2987 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2988 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2990 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2991 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
2995 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2996 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2997 &indexi
, 1, "mask_ptr");
3000 mask
= generate_quad_mask(gallivm
, fs_type
,
3001 i
*fs_type
.length
/4, 0, mask_input
);
3004 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
3006 LLVMBuildStore(builder
, mask
, mask_ptr
);
3010 generate_fs_loop(gallivm
,
3020 mask_store
, /* output */
3024 depth_sample_stride
,
3028 for (i
= 0; i
< num_fs
; i
++) {
3030 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
3031 int idx
= (i
+ (s
* num_fs
));
3032 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3033 ptr
= LLVMBuildGEP(builder
, mask_store
, &sindexi
, 1, "");
3035 fs_mask
[idx
] = LLVMBuildLoad(builder
, ptr
, "smask");
3038 for (unsigned s
= 0; s
< key
->min_samples
; s
++) {
3039 /* This is fucked up need to reorganize things */
3040 int idx
= s
* num_fs
+ i
;
3041 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3042 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3043 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3044 ptr
= LLVMBuildGEP(builder
,
3045 color_store
[cbuf
* !cbuf0_write_all
][chan
],
3047 fs_out_color
[s
][cbuf
][chan
][i
] = ptr
;
3050 if (dual_source_blend
) {
3051 /* only support one dual source blend target hence always use output 1 */
3052 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3053 ptr
= LLVMBuildGEP(builder
,
3054 color_store
[1][chan
],
3056 fs_out_color
[s
][1][chan
][i
] = ptr
;
3063 sampler
->destroy(sampler
);
3064 image
->destroy(image
);
3065 /* Loop over color outputs / color buffers to do blending.
3067 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3068 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
3069 LLVMValueRef color_ptr
;
3070 LLVMValueRef stride
;
3071 LLVMValueRef sample_stride
= NULL
;
3072 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
3074 boolean do_branch
= ((key
->depth
.enabled
3075 || key
->stencil
[0].enabled
3076 || key
->alpha
.enabled
)
3077 && !shader
->info
.base
.uses_kill
);
3079 color_ptr
= LLVMBuildLoad(builder
,
3080 LLVMBuildGEP(builder
, color_ptr_ptr
,
3084 stride
= LLVMBuildLoad(builder
,
3085 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
3088 if (key
->multisample
)
3089 sample_stride
= LLVMBuildLoad(builder
,
3090 LLVMBuildGEP(builder
, color_sample_stride_ptr
,
3091 &index
, 1, ""), "");
3093 for (unsigned s
= 0; s
< key
->cbuf_nr_samples
[cbuf
]; s
++) {
3094 unsigned mask_idx
= num_fs
* (key
->multisample
? s
: 0);
3095 unsigned out_idx
= key
->min_samples
== 1 ? 0 : s
;
3096 LLVMValueRef out_ptr
= color_ptr
;;
3098 if (key
->multisample
) {
3099 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_stride
, lp_build_const_int32(gallivm
, s
), "");
3100 out_ptr
= LLVMBuildGEP(builder
, out_ptr
, &sample_offset
, 1, "");
3102 out_ptr
= LLVMBuildBitCast(builder
, out_ptr
, LLVMPointerType(blend_vec_type
, 0), "");
3104 lp_build_name(out_ptr
, "color_ptr%d", cbuf
);
3106 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
3107 key
->cbuf_format
[cbuf
],
3108 num_fs
, fs_type
, &fs_mask
[mask_idx
], fs_out_color
[out_idx
],
3109 context_ptr
, out_ptr
, stride
,
3110 partial_mask
, do_branch
);
3115 LLVMBuildRetVoid(builder
);
3117 gallivm_verify_function(gallivm
, function
);
3122 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
3126 debug_printf("fs variant %p:\n", (void *) key
);
3128 if (key
->flatshade
) {
3129 debug_printf("flatshade = 1\n");
3131 if (key
->multisample
) {
3132 debug_printf("multisample = 1\n");
3133 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
3134 debug_printf("min samples = %d\n", key
->min_samples
);
3136 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
3137 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
3138 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
3140 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
3141 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
3142 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
3144 if (key
->depth
.enabled
) {
3145 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
3146 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
3149 for (i
= 0; i
< 2; ++i
) {
3150 if (key
->stencil
[i
].enabled
) {
3151 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
3152 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
3153 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
3154 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
3155 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
3156 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
3160 if (key
->alpha
.enabled
) {
3161 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
3164 if (key
->occlusion_count
) {
3165 debug_printf("occlusion_count = 1\n");
3168 if (key
->blend
.logicop_enable
) {
3169 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
3171 else if (key
->blend
.rt
[0].blend_enable
) {
3172 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
3173 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
3174 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
3175 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
3176 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
3177 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
3179 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
3180 if (key
->blend
.alpha_to_coverage
) {
3181 debug_printf("blend.alpha_to_coverage is enabled\n");
3183 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
3184 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
3185 debug_printf("sampler[%u] = \n", i
);
3186 debug_printf(" .wrap = %s %s %s\n",
3187 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
3188 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
3189 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
3190 debug_printf(" .min_img_filter = %s\n",
3191 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
3192 debug_printf(" .min_mip_filter = %s\n",
3193 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
3194 debug_printf(" .mag_img_filter = %s\n",
3195 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
3196 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
3197 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3198 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3199 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3200 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3201 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3202 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3204 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3205 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3206 debug_printf("texture[%u] = \n", i
);
3207 debug_printf(" .format = %s\n",
3208 util_format_name(texture
->format
));
3209 debug_printf(" .target = %s\n",
3210 util_str_tex_target(texture
->target
, TRUE
));
3211 debug_printf(" .level_zero_only = %u\n",
3212 texture
->level_zero_only
);
3213 debug_printf(" .pot = %u %u %u\n",
3215 texture
->pot_height
,
3216 texture
->pot_depth
);
3218 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3219 for (i
= 0; i
< key
->nr_images
; ++i
) {
3220 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3221 debug_printf("image[%u] = \n", i
);
3222 debug_printf(" .format = %s\n",
3223 util_format_name(image
->format
));
3224 debug_printf(" .target = %s\n",
3225 util_str_tex_target(image
->target
, TRUE
));
3226 debug_printf(" .level_zero_only = %u\n",
3227 image
->level_zero_only
);
3228 debug_printf(" .pot = %u %u %u\n",
3237 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3239 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3240 variant
->shader
->no
, variant
->no
);
3241 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3242 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3244 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3245 dump_fs_variant_key(&variant
->key
);
3246 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3252 * Generate a new fragment shader variant from the shader code and
3253 * other state indicated by the key.
3255 static struct lp_fragment_shader_variant
*
3256 generate_variant(struct llvmpipe_context
*lp
,
3257 struct lp_fragment_shader
*shader
,
3258 const struct lp_fragment_shader_variant_key
*key
)
3260 struct lp_fragment_shader_variant
*variant
;
3261 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3262 boolean fullcolormask
;
3263 char module_name
[64];
3265 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3269 memset(variant
, 0, sizeof(*variant
));
3270 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3271 shader
->no
, shader
->variants_created
);
3273 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
3274 if (!variant
->gallivm
) {
3279 variant
->shader
= shader
;
3280 variant
->list_item_global
.base
= variant
;
3281 variant
->list_item_local
.base
= variant
;
3282 variant
->no
= shader
->variants_created
++;
3284 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3287 * Determine whether we are touching all channels in the color buffer.
3289 fullcolormask
= FALSE
;
3290 if (key
->nr_cbufs
== 1) {
3291 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3292 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3296 !key
->blend
.logicop_enable
&&
3297 !key
->blend
.rt
[0].blend_enable
&&
3299 !key
->stencil
[0].enabled
&&
3300 !key
->alpha
.enabled
&&
3301 !key
->multisample
&&
3302 !key
->blend
.alpha_to_coverage
&&
3303 !key
->depth
.enabled
&&
3304 !shader
->info
.base
.uses_kill
&&
3305 !shader
->info
.base
.writes_samplemask
3308 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3309 lp_debug_fs_variant(variant
);
3312 lp_jit_init_types(variant
);
3314 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3315 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3317 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3318 if (variant
->opaque
) {
3319 /* Specialized shader, which doesn't need to read the color buffer. */
3320 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3325 * Compile everything
3328 gallivm_compile_module(variant
->gallivm
);
3330 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3332 if (variant
->function
[RAST_EDGE_TEST
]) {
3333 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3334 gallivm_jit_function(variant
->gallivm
,
3335 variant
->function
[RAST_EDGE_TEST
]);
3338 if (variant
->function
[RAST_WHOLE
]) {
3339 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3340 gallivm_jit_function(variant
->gallivm
,
3341 variant
->function
[RAST_WHOLE
]);
3342 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3343 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3346 gallivm_free_ir(variant
->gallivm
);
3353 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3354 const struct pipe_shader_state
*templ
)
3356 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3357 struct lp_fragment_shader
*shader
;
3359 int nr_sampler_views
;
3363 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3367 shader
->no
= fs_no
++;
3368 make_empty_list(&shader
->variants
);
3370 shader
->base
.type
= templ
->type
;
3371 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3372 /* get/save the summary info for this shader */
3373 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3375 /* we need to keep a local copy of the tokens */
3376 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3378 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3379 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3382 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3383 if (shader
->draw_data
== NULL
) {
3384 FREE((void *) shader
->base
.tokens
);
3389 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3390 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3391 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3392 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3394 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3395 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3396 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3397 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3399 switch (shader
->info
.base
.input_interpolate
[i
]) {
3400 case TGSI_INTERPOLATE_CONSTANT
:
3401 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3403 case TGSI_INTERPOLATE_LINEAR
:
3404 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3406 case TGSI_INTERPOLATE_PERSPECTIVE
:
3407 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3409 case TGSI_INTERPOLATE_COLOR
:
3410 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3417 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3418 case TGSI_SEMANTIC_FACE
:
3419 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3421 case TGSI_SEMANTIC_POSITION
:
3422 /* Position was already emitted above
3424 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3425 shader
->inputs
[i
].src_index
= 0;
3429 /* XXX this is a completely pointless index map... */
3430 shader
->inputs
[i
].src_index
= i
+1;
3433 if (LP_DEBUG
& DEBUG_TGSI
) {
3435 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3436 shader
->no
, (void *) shader
);
3437 tgsi_dump(templ
->tokens
, 0);
3438 debug_printf("usage masks:\n");
3439 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3440 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3441 debug_printf(" IN[%u].%s%s%s%s\n",
3443 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3444 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3445 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3446 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3456 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3458 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3459 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3460 if (llvmpipe
->fs
== lp_fs
)
3463 draw_bind_fragment_shader(llvmpipe
->draw
,
3464 (lp_fs
? lp_fs
->draw_data
: NULL
));
3466 llvmpipe
->fs
= lp_fs
;
3468 llvmpipe
->dirty
|= LP_NEW_FS
;
3473 * Remove shader variant from two lists: the shader's variant list
3474 * and the context's variant list.
3477 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3478 struct lp_fragment_shader_variant
*variant
)
3480 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3481 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3482 "v total cached %u inst %u total inst %u\n",
3483 variant
->shader
->no
, variant
->no
,
3484 variant
->shader
->variants_created
,
3485 variant
->shader
->variants_cached
,
3486 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3489 gallivm_destroy(variant
->gallivm
);
3491 /* remove from shader's list */
3492 remove_from_list(&variant
->list_item_local
);
3493 variant
->shader
->variants_cached
--;
3495 /* remove from context's list */
3496 remove_from_list(&variant
->list_item_global
);
3497 lp
->nr_fs_variants
--;
3498 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3505 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3507 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3508 struct lp_fragment_shader
*shader
= fs
;
3509 struct lp_fs_variant_list_item
*li
;
3511 assert(fs
!= llvmpipe
->fs
);
3514 * XXX: we need to flush the context until we have some sort of reference
3515 * counting in fragment shaders as they may still be binned
3516 * Flushing alone might not sufficient we need to wait on it too.
3518 llvmpipe_finish(pipe
, __FUNCTION__
);
3520 /* Delete all the variants */
3521 li
= first_elem(&shader
->variants
);
3522 while(!at_end(&shader
->variants
, li
)) {
3523 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3524 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3528 /* Delete draw module's data */
3529 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3531 if (shader
->base
.ir
.nir
)
3532 ralloc_free(shader
->base
.ir
.nir
);
3533 assert(shader
->variants_cached
== 0);
3534 FREE((void *) shader
->base
.tokens
);
3541 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3542 enum pipe_shader_type shader
, uint index
,
3543 const struct pipe_constant_buffer
*cb
)
3545 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3546 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3548 assert(shader
< PIPE_SHADER_TYPES
);
3549 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3551 /* note: reference counting */
3552 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3555 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3556 debug_printf("Illegal set constant without bind flag\n");
3557 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3561 if (shader
== PIPE_SHADER_VERTEX
||
3562 shader
== PIPE_SHADER_GEOMETRY
||
3563 shader
== PIPE_SHADER_TESS_CTRL
||
3564 shader
== PIPE_SHADER_TESS_EVAL
) {
3565 /* Pass the constants to the 'draw' module */
3566 const unsigned size
= cb
? cb
->buffer_size
: 0;
3570 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3572 else if (cb
&& cb
->user_buffer
) {
3573 data
= (ubyte
*) cb
->user_buffer
;
3580 data
+= cb
->buffer_offset
;
3582 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3585 else if (shader
== PIPE_SHADER_COMPUTE
)
3586 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3588 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3590 if (cb
&& cb
->user_buffer
) {
3591 pipe_resource_reference(&constants
, NULL
);
3596 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3597 enum pipe_shader_type shader
, unsigned start_slot
,
3598 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3599 unsigned writable_bitmask
)
3601 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3603 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3604 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3606 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3608 if (shader
== PIPE_SHADER_VERTEX
||
3609 shader
== PIPE_SHADER_GEOMETRY
||
3610 shader
== PIPE_SHADER_TESS_CTRL
||
3611 shader
== PIPE_SHADER_TESS_EVAL
) {
3612 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3613 const ubyte
*data
= NULL
;
3614 if (buffer
&& buffer
->buffer
)
3615 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3617 data
+= buffer
->buffer_offset
;
3618 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3620 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3621 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3622 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3623 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3629 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3630 enum pipe_shader_type shader
, unsigned start_slot
,
3631 unsigned count
, const struct pipe_image_view
*images
)
3633 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3636 draw_flush(llvmpipe
->draw
);
3637 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3638 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3640 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3643 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3644 if (shader
== PIPE_SHADER_VERTEX
||
3645 shader
== PIPE_SHADER_GEOMETRY
||
3646 shader
== PIPE_SHADER_TESS_CTRL
||
3647 shader
== PIPE_SHADER_TESS_EVAL
) {
3648 draw_set_images(llvmpipe
->draw
,
3650 llvmpipe
->images
[shader
],
3651 start_slot
+ count
);
3652 } else if (shader
== PIPE_SHADER_COMPUTE
)
3653 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3655 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3659 * Return the blend factor equivalent to a destination alpha of one.
3661 static inline unsigned
3662 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3665 case PIPE_BLENDFACTOR_DST_ALPHA
:
3666 return PIPE_BLENDFACTOR_ONE
;
3667 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3668 return PIPE_BLENDFACTOR_ZERO
;
3669 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3671 return PIPE_BLENDFACTOR_ZERO
;
3673 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3681 * We need to generate several variants of the fragment pipeline to match
3682 * all the combinations of the contributing state atoms.
3684 * TODO: there is actually no reason to tie this to context state -- the
3685 * generated code could be cached globally in the screen.
3687 static struct lp_fragment_shader_variant_key
*
3688 make_variant_key(struct llvmpipe_context
*lp
,
3689 struct lp_fragment_shader
*shader
,
3693 struct lp_fragment_shader_variant_key
*key
;
3695 key
= (struct lp_fragment_shader_variant_key
*)store
;
3697 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3699 if (lp
->framebuffer
.zsbuf
) {
3700 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3701 const struct util_format_description
*zsbuf_desc
=
3702 util_format_description(zsbuf_format
);
3704 if (lp
->depth_stencil
->depth
.enabled
&&
3705 util_format_has_depth(zsbuf_desc
)) {
3706 key
->zsbuf_format
= zsbuf_format
;
3707 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3709 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3710 util_format_has_stencil(zsbuf_desc
)) {
3711 key
->zsbuf_format
= zsbuf_format
;
3712 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3714 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3715 key
->resource_1d
= TRUE
;
3717 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3721 * Propagate the depth clamp setting from the rasterizer state.
3722 * depth_clip == 0 implies depth clamping is enabled.
3724 * When clip_halfz is enabled, then always clamp the depth values.
3726 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3727 * clamp is always active in d3d10, regardless if depth clip is
3729 * (GL has an always-on [0,1] clamp on fs depth output instead
3730 * to ensure the depth values stay in range. Doesn't look like
3731 * we do that, though...)
3733 if (lp
->rasterizer
->clip_halfz
) {
3734 key
->depth_clamp
= 1;
3736 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3739 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3740 if (!lp
->framebuffer
.nr_cbufs
||
3741 !lp
->framebuffer
.cbufs
[0] ||
3742 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3743 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3745 if(key
->alpha
.enabled
)
3746 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3747 /* alpha.ref_value is passed in jit_context */
3749 key
->flatshade
= lp
->rasterizer
->flatshade
;
3750 key
->multisample
= lp
->rasterizer
->multisample
;
3751 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3752 key
->occlusion_count
= TRUE
;
3755 if (lp
->framebuffer
.nr_cbufs
) {
3756 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3759 key
->coverage_samples
= 1;
3760 key
->min_samples
= 1;
3761 if (key
->multisample
) {
3762 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3763 key
->min_samples
= lp
->min_samples
== 1 ? 1 : key
->coverage_samples
;
3765 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3767 if (!key
->blend
.independent_blend_enable
) {
3768 /* we always need independent blend otherwise the fixups below won't work */
3769 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3770 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3772 key
->blend
.independent_blend_enable
= 1;
3775 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3776 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3778 if (lp
->framebuffer
.cbufs
[i
]) {
3779 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3780 const struct util_format_description
*format_desc
;
3782 key
->cbuf_format
[i
] = format
;
3783 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3786 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3787 * mixing of 2d textures with height 1 and 1d textures, so make sure
3788 * we pick 1d if any cbuf or zsbuf is 1d.
3790 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3791 key
->resource_1d
= TRUE
;
3794 format_desc
= util_format_description(format
);
3795 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3796 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3799 * Mask out color channels not present in the color buffer.
3801 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3804 * Disable blend for integer formats.
3806 if (util_format_is_pure_integer(format
)) {
3807 blend_rt
->blend_enable
= 0;
3811 * Our swizzled render tiles always have an alpha channel, but the
3812 * linear render target format often does not, so force here the dst
3815 * This is not a mere optimization. Wrong results will be produced if
3816 * the dst alpha is used, the dst format does not have alpha, and the
3817 * previous rendering was not flushed from the swizzled to linear
3818 * buffer. For example, NonPowTwo DCT.
3820 * TODO: This should be generalized to all channels for better
3821 * performance, but only alpha causes correctness issues.
3823 * Also, force rgb/alpha func/factors match, to make AoS blending
3826 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3827 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3828 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3829 boolean clamped_zero
= !util_format_is_float(format
) &&
3830 !util_format_is_snorm(format
);
3831 blend_rt
->rgb_src_factor
=
3832 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3833 blend_rt
->rgb_dst_factor
=
3834 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3835 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3836 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3837 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3841 /* no color buffer for this fragment output */
3842 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3843 key
->cbuf_nr_samples
[i
] = 0;
3844 blend_rt
->colormask
= 0x0;
3845 blend_rt
->blend_enable
= 0;
3849 /* This value will be the same for all the variants of a given shader:
3851 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3853 struct lp_sampler_static_state
*fs_sampler
;
3855 fs_sampler
= key
->samplers
;
3857 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3859 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3860 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3861 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3862 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3867 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3868 * are dx10-style? Can't really have mixed opcodes, at least not
3869 * if we want to skip the holes here (without rescanning tgsi).
3871 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3872 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3873 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3875 * Note sview may exceed what's representable by file_mask.
3876 * This will still work, the only downside is that not actually
3877 * used views may be included in the shader key.
3879 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3880 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3881 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3886 key
->nr_sampler_views
= key
->nr_samplers
;
3887 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3888 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3889 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3890 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3895 struct lp_image_static_state
*lp_image
;
3896 lp_image
= lp_fs_variant_key_images(key
);
3897 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3898 for (i
= 0; i
< key
->nr_images
; ++i
) {
3899 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3900 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3901 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3910 * Update fragment shader state. This is called just prior to drawing
3911 * something when some fragment-related state has changed.
3914 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3916 struct lp_fragment_shader
*shader
= lp
->fs
;
3917 struct lp_fragment_shader_variant_key
*key
;
3918 struct lp_fragment_shader_variant
*variant
= NULL
;
3919 struct lp_fs_variant_list_item
*li
;
3920 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3922 key
= make_variant_key(lp
, shader
, store
);
3924 /* Search the variants for one which matches the key */
3925 li
= first_elem(&shader
->variants
);
3926 while(!at_end(&shader
->variants
, li
)) {
3927 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3935 /* Move this variant to the head of the list to implement LRU
3936 * deletion of shader's when we have too many.
3938 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3941 /* variant not found, create it now */
3944 unsigned variants_to_cull
;
3946 if (LP_DEBUG
& DEBUG_FS
) {
3947 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3950 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3953 /* First, check if we've exceeded the max number of shader variants.
3954 * If so, free 6.25% of them (the least recently used ones).
3956 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3958 if (variants_to_cull
||
3959 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3960 struct pipe_context
*pipe
= &lp
->pipe
;
3962 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3963 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3964 "\t%u instrs,\t%u instrs/variant\n",
3965 shader
->variants_cached
,
3966 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3967 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3971 * XXX: we need to flush the context until we have some sort of
3972 * reference counting in fragment shaders as they may still be binned
3973 * Flushing alone might not be sufficient we need to wait on it too.
3975 llvmpipe_finish(pipe
, __FUNCTION__
);
3978 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3979 * number of shader variants (potentially all of them) could be
3980 * pending for destruction on flush.
3983 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3984 struct lp_fs_variant_list_item
*item
;
3985 if (is_empty_list(&lp
->fs_variants_list
)) {
3988 item
= last_elem(&lp
->fs_variants_list
);
3991 llvmpipe_remove_shader_variant(lp
, item
->base
);
3996 * Generate the new variant.
3999 variant
= generate_variant(lp
, shader
, key
);
4002 LP_COUNT_ADD(llvm_compile_time
, dt
);
4003 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
4005 /* Put the new variant into the list */
4007 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
4008 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
4009 lp
->nr_fs_variants
++;
4010 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
4011 shader
->variants_cached
++;
4015 /* Bind this variant */
4016 lp_setup_set_fs_variant(lp
->setup
, variant
);
4024 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
4026 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
4027 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
4028 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
4030 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
4032 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
4033 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;