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 #include "lp_screen.h"
109 #include "compiler/nir/nir_serialize.h"
110 #include "util/mesa-sha1.h"
111 /** Fragment shader number (for debugging) */
112 static unsigned fs_no
= 0;
116 * Expand the relevant bits of mask_input to a n*4-dword mask for the
117 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
118 * quad mask vector to 0 or ~0.
119 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
120 * quad arguments with fs length 8.
122 * \param first_quad which quad(s) of the quad group to test, in [0,3]
123 * \param mask_input bitwise mask for the whole 4x4 stamp
126 generate_quad_mask(struct gallivm_state
*gallivm
,
127 struct lp_type fs_type
,
130 LLVMValueRef mask_input
) /* int64 */
132 LLVMBuilderRef builder
= gallivm
->builder
;
133 struct lp_type mask_type
;
134 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
135 LLVMValueRef bits
[16];
136 LLVMValueRef mask
, bits_vec
;
140 * XXX: We'll need a different path for 16 x u8
142 assert(fs_type
.width
== 32);
143 assert(fs_type
.length
<= ARRAY_SIZE(bits
));
144 mask_type
= lp_int_type(fs_type
);
147 * mask_input >>= (quad * 4)
149 switch (first_quad
) {
154 assert(fs_type
.length
== 4);
161 assert(fs_type
.length
== 4);
169 mask_input
= LLVMBuildLShr(builder
, mask_input
, lp_build_const_int64(gallivm
, 16 * sample
), "");
170 mask_input
= LLVMBuildTrunc(builder
, mask_input
,
172 mask_input
= LLVMBuildAnd(builder
, mask_input
, lp_build_const_int32(gallivm
, 0xffff), "");
174 mask_input
= LLVMBuildLShr(builder
,
176 LLVMConstInt(i32t
, shift
, 0),
180 * mask = { mask_input & (1 << i), for i in [0,3] }
182 mask
= lp_build_broadcast(gallivm
,
183 lp_build_vec_type(gallivm
, mask_type
),
186 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
187 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
188 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1ULL << (j
+ 0), 0);
189 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
190 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
191 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (j
+ 5), 0);
193 bits_vec
= LLVMConstVector(bits
, fs_type
.length
);
194 mask
= LLVMBuildAnd(builder
, mask
, bits_vec
, "");
197 * mask = mask == bits ? ~0 : 0
199 mask
= lp_build_compare(gallivm
,
200 mask_type
, PIPE_FUNC_EQUAL
,
207 #define EARLY_DEPTH_TEST 0x1
208 #define LATE_DEPTH_TEST 0x2
209 #define EARLY_DEPTH_WRITE 0x4
210 #define LATE_DEPTH_WRITE 0x8
213 find_output_by_semantic( const struct tgsi_shader_info
*info
,
219 for (i
= 0; i
< info
->num_outputs
; i
++)
220 if (info
->output_semantic_name
[i
] == semantic
&&
221 info
->output_semantic_index
[i
] == index
)
229 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
232 lp_llvm_viewport(LLVMValueRef context_ptr
,
233 struct gallivm_state
*gallivm
,
234 LLVMValueRef viewport_index
)
236 LLVMBuilderRef builder
= gallivm
->builder
;
239 struct lp_type viewport_type
=
240 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
242 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
243 ptr
= LLVMBuildPointerCast(builder
, ptr
,
244 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
246 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
253 lp_build_depth_clamp(struct gallivm_state
*gallivm
,
254 LLVMBuilderRef builder
,
256 LLVMValueRef context_ptr
,
257 LLVMValueRef thread_data_ptr
,
260 LLVMValueRef viewport
, min_depth
, max_depth
;
261 LLVMValueRef viewport_index
;
262 struct lp_build_context f32_bld
;
264 assert(type
.floating
);
265 lp_build_context_init(&f32_bld
, gallivm
, type
);
268 * Assumes clamping of the viewport index will occur in setup/gs. Value
269 * is passed through the rasterization stage via lp_rast_shader_inputs.
271 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
274 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
278 * Load the min and max depth from the lp_jit_context.viewports
279 * array of lp_jit_viewport structures.
281 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
283 /* viewports[viewport_index].min_depth */
284 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
285 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
), "");
286 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
288 /* viewports[viewport_index].max_depth */
289 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
290 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
), "");
291 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
294 * Clamp to the min and max depth values for the given viewport.
296 return lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
300 lp_build_sample_alpha_to_coverage(struct gallivm_state
*gallivm
,
302 unsigned coverage_samples
,
303 LLVMValueRef num_loop
,
304 LLVMValueRef loop_counter
,
305 LLVMValueRef coverage_mask_store
,
308 struct lp_build_context bld
;
309 LLVMBuilderRef builder
= gallivm
->builder
;
310 float step
= 1.0 / coverage_samples
;
312 lp_build_context_init(&bld
, gallivm
, type
);
313 for (unsigned s
= 0; s
< coverage_samples
; s
++) {
314 LLVMValueRef alpha_ref_value
= lp_build_const_vec(gallivm
, type
, step
* s
);
315 LLVMValueRef test
= lp_build_cmp(&bld
, PIPE_FUNC_GREATER
, alpha
, alpha_ref_value
);
317 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, s
), num_loop
, "");
318 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_counter
, "");
319 LLVMValueRef s_mask_ptr
= LLVMBuildGEP(builder
, coverage_mask_store
, &s_mask_idx
, 1, "");
320 LLVMValueRef s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
321 s_mask
= LLVMBuildAnd(builder
, s_mask
, test
, "");
322 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
326 struct lp_build_fs_llvm_iface
{
327 struct lp_build_fs_iface base
;
328 struct lp_build_interp_soa_context
*interp
;
329 struct lp_build_for_loop_state
*loop_state
;
330 LLVMValueRef mask_store
;
333 static LLVMValueRef
fs_interp(const struct lp_build_fs_iface
*iface
,
334 struct lp_build_context
*bld
,
335 unsigned attrib
, unsigned chan
,
336 bool centroid
, bool sample
,
337 LLVMValueRef attrib_indir
,
338 LLVMValueRef offsets
[2])
340 struct lp_build_fs_llvm_iface
*fs_iface
= (struct lp_build_fs_llvm_iface
*)iface
;
341 struct lp_build_interp_soa_context
*interp
= fs_iface
->interp
;
342 unsigned loc
= TGSI_INTERPOLATE_LOC_CENTER
;
344 loc
= TGSI_INTERPOLATE_LOC_CENTROID
;
346 loc
= TGSI_INTERPOLATE_LOC_SAMPLE
;
348 return lp_build_interp_soa(interp
, bld
->gallivm
, fs_iface
->loop_state
->counter
,
349 fs_iface
->mask_store
,
350 attrib
, chan
, loc
, attrib_indir
, offsets
);
354 * Generate the fragment shader, depth/stencil test, and alpha tests.
357 generate_fs_loop(struct gallivm_state
*gallivm
,
358 struct lp_fragment_shader
*shader
,
359 const struct lp_fragment_shader_variant_key
*key
,
360 LLVMBuilderRef builder
,
362 LLVMValueRef context_ptr
,
363 LLVMValueRef sample_pos_array
,
364 LLVMValueRef num_loop
,
365 struct lp_build_interp_soa_context
*interp
,
366 const struct lp_build_sampler_soa
*sampler
,
367 const struct lp_build_image_soa
*image
,
368 LLVMValueRef mask_store
,
369 LLVMValueRef (*out_color
)[4],
370 LLVMValueRef depth_base_ptr
,
371 LLVMValueRef depth_stride
,
372 LLVMValueRef depth_sample_stride
,
374 LLVMValueRef thread_data_ptr
)
376 const struct util_format_description
*zs_format_desc
= NULL
;
377 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
378 struct lp_type int_type
= lp_int_type(type
);
379 LLVMTypeRef vec_type
, int_vec_type
;
380 LLVMValueRef mask_ptr
= NULL
, mask_val
= NULL
;
381 LLVMValueRef consts_ptr
, num_consts_ptr
;
382 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
384 LLVMValueRef z_value
, s_value
;
385 LLVMValueRef z_fb
, s_fb
;
386 LLVMValueRef depth_ptr
;
387 LLVMValueRef stencil_refs
[2];
388 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
389 LLVMValueRef zs_samples
= lp_build_const_int32(gallivm
, key
->zsbuf_nr_samples
);
390 struct lp_build_for_loop_state loop_state
, sample_loop_state
;
391 struct lp_build_mask_context mask
;
393 * TODO: figure out if simple_shader optimization is really worthwile to
394 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
395 * code since tests tend to take another codepath than real shaders.
397 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
398 shader
->info
.base
.num_inputs
< 3 &&
399 shader
->info
.base
.num_instructions
< 8) && 0;
400 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
401 util_blend_state_is_dual(&key
->blend
, 0);
402 const bool post_depth_coverage
= shader
->info
.base
.properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
];
408 struct lp_bld_tgsi_system_values system_values
;
410 memset(&system_values
, 0, sizeof(system_values
));
412 /* truncate then sign extend. */
413 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
414 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
416 if (key
->depth
.enabled
||
417 key
->stencil
[0].enabled
) {
419 zs_format_desc
= util_format_description(key
->zsbuf_format
);
420 assert(zs_format_desc
);
422 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
423 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
424 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
425 if (shader
->info
.base
.writes_memory
)
426 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
427 else if (key
->alpha
.enabled
||
428 key
->blend
.alpha_to_coverage
||
429 shader
->info
.base
.uses_kill
||
430 shader
->info
.base
.writes_samplemask
) {
431 /* With alpha test and kill, can do the depth test early
432 * and hopefully eliminate some quads. But need to do a
433 * special deferred depth write once the final mask value
434 * is known. This only works though if there's either no
435 * stencil test or the stencil value isn't written.
437 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
438 (key
->stencil
[1].enabled
&&
439 key
->stencil
[1].writemask
)))
440 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
442 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
445 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
448 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
451 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
452 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
453 (key
->stencil
[1].enabled
&&
454 key
->stencil
[1].writemask
))))
455 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
461 vec_type
= lp_build_vec_type(gallivm
, type
);
462 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
464 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
465 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
466 /* convert scalar stencil refs into vectors */
467 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
468 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
470 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
471 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
473 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
474 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
476 memset(outputs
, 0, sizeof outputs
);
478 /* Allocate color storage for each fragment sample */
479 LLVMValueRef color_store_size
= num_loop
;
480 if (key
->min_samples
> 1)
481 color_store_size
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, key
->min_samples
), "");
483 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
484 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
485 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
486 lp_build_vec_type(gallivm
,
488 color_store_size
, "color");
491 if (dual_source_blend
) {
492 assert(key
->nr_cbufs
<= 1);
493 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
494 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
495 lp_build_vec_type(gallivm
,
497 color_store_size
, "color1");
501 lp_build_for_loop_begin(&loop_state
, gallivm
,
502 lp_build_const_int32(gallivm
, 0),
505 lp_build_const_int32(gallivm
, 1));
507 LLVMValueRef sample_mask_in
;
508 if (key
->multisample
) {
509 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 0);
510 /* create shader execution mask by combining all sample masks. */
511 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
512 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
513 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
514 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
518 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
520 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, s_mask
, lp_build_const_int_vec(gallivm
, type
, (1 << s
)), "");
521 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
524 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 1);
525 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
526 &loop_state
.counter
, 1, "mask_ptr");
527 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
529 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, mask_val
, lp_build_const_int_vec(gallivm
, type
, 1), "");
530 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
533 /* 'mask' will control execution based on quad's pixel alive/killed state */
534 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
536 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
537 lp_build_mask_check(&mask
);
539 /* Create storage for recombining sample masks after early Z pass. */
540 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
541 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
543 /* Create storage for post depth sample mask */
544 LLVMValueRef post_depth_sample_mask_in
= NULL
;
545 if (post_depth_coverage
)
546 post_depth_sample_mask_in
= lp_build_alloca(gallivm
, int_vec_type
, "post_depth_sample_mask_in");
548 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
549 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
550 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
551 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
553 /* Run early depth once per sample */
554 if (key
->multisample
) {
556 if (zs_format_desc
) {
557 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
558 struct lp_type z_type
= zs_type
;
559 struct lp_type s_type
= zs_type
;
560 if (zs_format_desc
->block
.bits
< type
.width
)
561 z_type
.width
= type
.width
;
562 if (zs_format_desc
->block
.bits
== 8)
563 s_type
.width
= type
.width
;
565 else if (zs_format_desc
->block
.bits
> 32) {
566 z_type
.width
= z_type
.width
/ 2;
567 s_type
.width
= s_type
.width
/ 2;
570 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
571 zs_samples
, "z_sample_store");
572 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
573 zs_samples
, "s_sample_store");
574 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
575 zs_samples
, "z_fb_store");
576 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
577 zs_samples
, "s_fb_store");
579 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
580 lp_build_const_int32(gallivm
, 0),
581 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
582 lp_build_const_int32(gallivm
, 1));
584 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
585 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
586 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
588 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
589 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
593 /* for multisample Z needs to be interpolated at sample points for testing. */
594 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
597 depth_ptr
= depth_base_ptr
;
598 if (key
->multisample
) {
599 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
600 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
603 if (depth_mode
& EARLY_DEPTH_TEST
) {
605 * Clamp according to ARB_depth_clamp semantics.
607 if (key
->depth_clamp
) {
608 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
611 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
612 zs_format_desc
, key
->resource_1d
,
613 depth_ptr
, depth_stride
,
614 &z_fb
, &s_fb
, loop_state
.counter
);
615 lp_build_depth_stencil_test(gallivm
,
620 key
->multisample
? NULL
: &mask
,
626 !simple_shader
&& !key
->multisample
);
628 if (depth_mode
& EARLY_DEPTH_WRITE
) {
629 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
630 zs_format_desc
, key
->resource_1d
,
631 NULL
, NULL
, NULL
, loop_state
.counter
,
632 depth_ptr
, depth_stride
,
636 * Note mask check if stencil is enabled must be after ds write not after
637 * stencil test otherwise new stencil values may not get written if all
638 * fragments got killed by depth/stencil test.
640 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
641 lp_build_mask_check(&mask
);
643 if (key
->multisample
) {
644 z_fb_type
= LLVMTypeOf(z_fb
);
645 z_type
= LLVMTypeOf(z_value
);
646 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
647 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
648 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
649 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
653 if (key
->multisample
) {
655 * Store the post-early Z coverage mask.
656 * Recombine the resulting coverage masks post early Z into the fragment
657 * shader execution mask.
659 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
660 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
661 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
663 if (post_depth_coverage
) {
664 LLVMValueRef mask_bit_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
665 LLVMValueRef post_depth_mask_in
= LLVMBuildLoad(builder
, post_depth_sample_mask_in
, "");
666 mask_bit_idx
= LLVMBuildAnd(builder
, s_mask
, lp_build_broadcast(gallivm
, int_vec_type
, mask_bit_idx
), "");
667 post_depth_mask_in
= LLVMBuildOr(builder
, post_depth_mask_in
, mask_bit_idx
, "");
668 LLVMBuildStore(builder
, post_depth_mask_in
, post_depth_sample_mask_in
);
671 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
673 lp_build_for_loop_end(&sample_loop_state
);
675 /* recombined all the coverage masks in the shader exec mask. */
676 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
677 lp_build_mask_update(&mask
, tmp_s_mask_or
);
679 if (key
->min_samples
== 1) {
680 /* for multisample Z needs to be re interpolated at pixel center */
681 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
682 lp_build_mask_update(&mask
, tmp_s_mask_or
);
685 if (post_depth_coverage
) {
686 LLVMValueRef post_depth_mask_in
= LLVMBuildAnd(builder
, lp_build_mask_value(&mask
), lp_build_const_int_vec(gallivm
, type
, 1), "");
687 LLVMBuildStore(builder
, post_depth_mask_in
, post_depth_sample_mask_in
);
691 LLVMValueRef out_sample_mask_storage
= NULL
;
692 if (shader
->info
.base
.writes_samplemask
) {
693 out_sample_mask_storage
= lp_build_alloca(gallivm
, int_vec_type
, "write_mask");
694 if (key
->min_samples
> 1)
695 LLVMBuildStore(builder
, LLVMConstNull(int_vec_type
), out_sample_mask_storage
);
698 if (post_depth_coverage
) {
699 system_values
.sample_mask_in
= LLVMBuildLoad(builder
, post_depth_sample_mask_in
, "");
702 system_values
.sample_mask_in
= sample_mask_in
;
703 if (key
->multisample
&& key
->min_samples
> 1) {
704 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
705 lp_build_const_int32(gallivm
, 0),
707 lp_build_const_int32(gallivm
, key
->min_samples
),
708 lp_build_const_int32(gallivm
, 1));
710 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
711 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
712 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
713 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
714 lp_build_mask_force(&mask
, s_mask
);
715 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, sample_loop_state
.counter
);
716 system_values
.sample_id
= sample_loop_state
.counter
;
717 system_values
.sample_mask_in
= LLVMBuildAnd(builder
, system_values
.sample_mask_in
,
718 lp_build_broadcast(gallivm
, int_vec_type
,
719 LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "")), "");
721 system_values
.sample_id
= lp_build_const_int32(gallivm
, 0);
724 system_values
.sample_pos
= sample_pos_array
;
726 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, mask_store
, sample_loop_state
.counter
);
728 struct lp_build_fs_llvm_iface fs_iface
= {
729 .base
.interp_fn
= fs_interp
,
731 .loop_state
= &loop_state
,
732 .mask_store
= mask_store
,
735 struct lp_build_tgsi_params params
;
736 memset(¶ms
, 0, sizeof(params
));
740 params
.fs_iface
= &fs_iface
.base
;
741 params
.consts_ptr
= consts_ptr
;
742 params
.const_sizes_ptr
= num_consts_ptr
;
743 params
.system_values
= &system_values
;
744 params
.inputs
= interp
->inputs
;
745 params
.context_ptr
= context_ptr
;
746 params
.thread_data_ptr
= thread_data_ptr
;
747 params
.sampler
= sampler
;
748 params
.info
= &shader
->info
.base
;
749 params
.ssbo_ptr
= ssbo_ptr
;
750 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
751 params
.image
= image
;
753 /* Build the actual shader */
754 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
755 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
758 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
762 if (key
->alpha
.enabled
) {
763 int color0
= find_output_by_semantic(&shader
->info
.base
,
767 if (color0
!= -1 && outputs
[color0
][3]) {
768 const struct util_format_description
*cbuf_format_desc
;
769 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
770 LLVMValueRef alpha_ref_value
;
772 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
773 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
775 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
777 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
778 &mask
, alpha
, alpha_ref_value
,
779 (depth_mode
& LATE_DEPTH_TEST
) != 0);
783 /* Emulate Alpha to Coverage with Alpha test */
784 if (key
->blend
.alpha_to_coverage
) {
785 int color0
= find_output_by_semantic(&shader
->info
.base
,
789 if (color0
!= -1 && outputs
[color0
][3]) {
790 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
792 if (!key
->multisample
) {
793 lp_build_alpha_to_coverage(gallivm
, type
,
795 (depth_mode
& LATE_DEPTH_TEST
) != 0);
797 lp_build_sample_alpha_to_coverage(gallivm
, type
, key
->coverage_samples
, num_loop
,
803 if (key
->blend
.alpha_to_one
&& key
->multisample
) {
804 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
) {
805 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
806 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
807 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
808 if (outputs
[cbuf
][3]) {
809 LLVMBuildStore(builder
, lp_build_const_vec(gallivm
, type
, 1.0), outputs
[cbuf
][3]);
813 if (shader
->info
.base
.writes_samplemask
) {
814 LLVMValueRef output_smask
= NULL
;
815 int smaski
= find_output_by_semantic(&shader
->info
.base
,
816 TGSI_SEMANTIC_SAMPLEMASK
,
818 struct lp_build_context smask_bld
;
819 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
822 output_smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
823 output_smask
= LLVMBuildBitCast(builder
, output_smask
, smask_bld
.vec_type
, "");
825 if (key
->min_samples
> 1) {
826 /* only the bit corresponding to this sample is to be used. */
827 LLVMValueRef tmp_mask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "tmp_mask");
828 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
829 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
), "");
830 output_smask
= LLVMBuildOr(builder
, tmp_mask
, smask_bit
, "");
833 LLVMBuildStore(builder
, output_smask
, out_sample_mask_storage
);
836 /* Color write - per fragment sample */
837 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
839 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
840 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
841 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
843 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
844 if(outputs
[attrib
][chan
]) {
845 /* XXX: just initialize outputs to point at colors[] and
848 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
849 LLVMValueRef color_ptr
;
850 LLVMValueRef color_idx
= loop_state
.counter
;
851 if (key
->min_samples
> 1)
852 color_idx
= LLVMBuildAdd(builder
, color_idx
,
853 LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, ""), "");
854 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
856 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
857 LLVMBuildStore(builder
, out
, color_ptr
);
863 if (key
->multisample
&& key
->min_samples
> 1) {
864 LLVMBuildStore(builder
, lp_build_mask_value(&mask
), s_mask_ptr
);
865 lp_build_for_loop_end(&sample_loop_state
);
868 if (key
->multisample
) {
869 /* execute depth test for each sample */
870 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
871 lp_build_const_int32(gallivm
, 0),
872 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
873 lp_build_const_int32(gallivm
, 1));
875 /* load the per-sample coverage mask */
876 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
877 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
878 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
880 /* combine the execution mask post fragment shader with the coverage mask. */
881 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
882 if (key
->min_samples
== 1)
883 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
885 /* if the shader writes sample mask use that */
886 if (shader
->info
.base
.writes_samplemask
) {
887 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
888 out_smask_idx
= lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
);
889 LLVMValueRef output_smask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "");
890 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, out_smask_idx
, "");
891 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int_vec(gallivm
, int_type
, 0), "");
892 smask_bit
= LLVMBuildSExt(builder
, cmp
, int_vec_type
, "");
894 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
898 depth_ptr
= depth_base_ptr
;
899 if (key
->multisample
) {
900 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
901 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
905 if (depth_mode
& LATE_DEPTH_TEST
) {
906 int pos0
= find_output_by_semantic(&shader
->info
.base
,
907 TGSI_SEMANTIC_POSITION
,
909 int s_out
= find_output_by_semantic(&shader
->info
.base
,
910 TGSI_SEMANTIC_STENCIL
,
912 if (pos0
!= -1 && outputs
[pos0
][2]) {
913 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
916 * Clamp according to ARB_depth_clamp semantics.
918 if (key
->depth_clamp
) {
919 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
923 if (s_out
!= -1 && outputs
[s_out
][1]) {
924 /* there's only one value, and spec says to discard additional bits */
925 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
926 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
927 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
928 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
929 stencil_refs
[1] = stencil_refs
[0];
932 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
933 zs_format_desc
, key
->resource_1d
,
934 depth_ptr
, depth_stride
,
935 &z_fb
, &s_fb
, loop_state
.counter
);
937 lp_build_depth_stencil_test(gallivm
,
942 key
->multisample
? NULL
: &mask
,
950 if (depth_mode
& LATE_DEPTH_WRITE
) {
951 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
952 zs_format_desc
, key
->resource_1d
,
953 NULL
, NULL
, NULL
, loop_state
.counter
,
954 depth_ptr
, depth_stride
,
958 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
959 (depth_mode
& LATE_DEPTH_WRITE
))
961 /* Need to apply a reduced mask to the depth write. Reload the
962 * depth value, update from zs_value with the new mask value and
965 if (key
->multisample
) {
966 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
967 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
968 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
969 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
971 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
972 zs_format_desc
, key
->resource_1d
,
973 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
974 depth_ptr
, depth_stride
,
978 if (key
->occlusion_count
) {
979 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
980 lp_build_name(counter
, "counter");
982 lp_build_occlusion_count(gallivm
, type
,
983 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), counter
);
986 if (key
->multisample
) {
987 /* store the sample mask for this loop */
988 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
989 lp_build_for_loop_end(&sample_loop_state
);
992 mask_val
= lp_build_mask_end(&mask
);
993 if (!key
->multisample
)
994 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
995 lp_build_for_loop_end(&loop_state
);
1000 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
1002 * Fragment Shader outputs pixels in small 2x2 blocks
1003 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
1005 * However in memory pixels are stored in rows
1006 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
1008 * @param type fragment shader type (4x or 8x float)
1009 * @param num_fs number of fs_src
1010 * @param is_1d whether we're outputting to a 1d resource
1011 * @param dst_channels number of output channels
1012 * @param fs_src output from fragment shader
1013 * @param dst pointer to store result
1014 * @param pad_inline is channel padding inline or at end of row
1015 * @return the number of dsts
1018 generate_fs_twiddle(struct gallivm_state
*gallivm
,
1019 struct lp_type type
,
1021 unsigned dst_channels
,
1022 LLVMValueRef fs_src
[][4],
1026 LLVMValueRef src
[16];
1032 unsigned pixels
= type
.length
/ 4;
1033 unsigned reorder_group
;
1034 unsigned src_channels
;
1038 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1039 src_count
= num_fs
* src_channels
;
1041 assert(pixels
== 2 || pixels
== 1);
1042 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
1045 * Transpose from SoA -> AoS
1047 for (i
= 0; i
< num_fs
; ++i
) {
1048 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
1052 * Pick transformation options
1054 swizzle_pad
= false;
1059 if (dst_channels
== 1) {
1065 } else if (dst_channels
== 2) {
1069 } else if (dst_channels
> 2) {
1076 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
1082 * Split the src in half
1085 for (i
= num_fs
; i
> 0; --i
) {
1086 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
1087 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
1095 * Ensure pixels are in memory order
1097 if (reorder_group
) {
1098 /* Twiddle pixels by reordering the array, e.g.:
1100 * src_count = 8 -> 0 2 1 3 4 6 5 7
1101 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1103 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
1105 for (i
= 0; i
< src_count
; ++i
) {
1106 unsigned group
= i
/ reorder_group
;
1107 unsigned block
= (group
/ 4) * 4 * reorder_group
;
1108 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
1111 } else if (twiddle
) {
1112 /* Twiddle pixels across elements of array */
1114 * XXX: we should avoid this in some cases, but would need to tell
1115 * lp_build_conv to reorder (or deal with it ourselves).
1117 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
1120 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
1124 * Moves any padding between pixels to the end
1125 * e.g. RGBXRGBX -> RGBRGBXX
1128 unsigned char swizzles
[16];
1129 unsigned elems
= pixels
* dst_channels
;
1131 for (i
= 0; i
< type
.length
; ++i
) {
1133 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
1135 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
1138 for (i
= 0; i
< src_count
; ++i
) {
1139 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
1148 * Untwiddle and transpose, much like the above.
1149 * However, this is after conversion, so we get packed vectors.
1150 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1151 * the vectors will look like:
1152 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1153 * be swizzled here). Extending to 16bit should be trivial.
1154 * Should also be extended to handle twice wide vectors with AVX2...
1157 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
1158 struct lp_type type
,
1164 struct lp_type type64
, type16
, type32
;
1165 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
1166 LLVMBuilderRef builder
= gallivm
->builder
;
1167 LLVMValueRef tmp
[4], shuf
[8];
1168 for (j
= 0; j
< 2; j
++) {
1169 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
1170 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
1171 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
1172 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
1175 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
1176 assert(type
.width
== 8);
1177 assert(type
.length
== 16);
1179 type8_t
= lp_build_vec_type(gallivm
, type
);
1184 type64_t
= lp_build_vec_type(gallivm
, type64
);
1189 type16_t
= lp_build_vec_type(gallivm
, type16
);
1194 type32_t
= lp_build_vec_type(gallivm
, type32
);
1196 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1198 if (src_count
== 1) {
1199 /* transpose was no-op, just untwiddle */
1200 LLVMValueRef shuf_vec
;
1201 shuf_vec
= LLVMConstVector(shuf
, 8);
1202 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1203 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1204 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1205 } else if (src_count
== 2) {
1206 LLVMValueRef shuf_vec
;
1207 shuf_vec
= LLVMConstVector(shuf
, 4);
1209 for (i
= 0; i
< 2; i
++) {
1210 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1211 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1212 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1215 for (j
= 0; j
< 2; j
++) {
1216 LLVMValueRef lo
, hi
, lo2
, hi2
;
1218 * Note that if we only really have 3 valid channels (rgb)
1219 * and we don't need alpha we could substitute a undef here
1220 * for the respective channel (causing llvm to drop conversion
1223 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1224 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1225 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1226 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1227 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1228 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1229 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1236 * Load an unswizzled block of pixels from memory
1239 load_unswizzled_block(struct gallivm_state
*gallivm
,
1240 LLVMValueRef base_ptr
,
1241 LLVMValueRef stride
,
1242 unsigned block_width
,
1243 unsigned block_height
,
1245 struct lp_type dst_type
,
1247 unsigned dst_alignment
)
1249 LLVMBuilderRef builder
= gallivm
->builder
;
1250 unsigned row_size
= dst_count
/ block_height
;
1253 /* Ensure block exactly fits into dst */
1254 assert((block_width
* block_height
) % dst_count
== 0);
1256 for (i
= 0; i
< dst_count
; ++i
) {
1257 unsigned x
= i
% row_size
;
1258 unsigned y
= i
/ row_size
;
1260 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1261 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1263 LLVMValueRef gep
[2];
1264 LLVMValueRef dst_ptr
;
1266 gep
[0] = lp_build_const_int32(gallivm
, 0);
1267 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1269 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1270 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1271 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1273 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1275 LLVMSetAlignment(dst
[i
], dst_alignment
);
1281 * Store an unswizzled block of pixels to memory
1284 store_unswizzled_block(struct gallivm_state
*gallivm
,
1285 LLVMValueRef base_ptr
,
1286 LLVMValueRef stride
,
1287 unsigned block_width
,
1288 unsigned block_height
,
1290 struct lp_type src_type
,
1292 unsigned src_alignment
)
1294 LLVMBuilderRef builder
= gallivm
->builder
;
1295 unsigned row_size
= src_count
/ block_height
;
1298 /* Ensure src exactly fits into block */
1299 assert((block_width
* block_height
) % src_count
== 0);
1301 for (i
= 0; i
< src_count
; ++i
) {
1302 unsigned x
= i
% row_size
;
1303 unsigned y
= i
/ row_size
;
1305 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1306 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1308 LLVMValueRef gep
[2];
1309 LLVMValueRef src_ptr
;
1311 gep
[0] = lp_build_const_int32(gallivm
, 0);
1312 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1314 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1315 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1316 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1318 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1320 LLVMSetAlignment(src_ptr
, src_alignment
);
1326 * Checks if a format description is an arithmetic format
1328 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1330 static inline boolean
1331 is_arithmetic_format(const struct util_format_description
*format_desc
)
1333 boolean arith
= false;
1336 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1337 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1338 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1346 * Checks if this format requires special handling due to required expansion
1347 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1350 static inline boolean
1351 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1353 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1354 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1362 * Retrieves the type representing the memory layout for a format
1364 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1367 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1368 struct lp_type
* type
)
1373 if (format_expands_to_float_soa(format_desc
)) {
1374 /* just make this a uint with width of block */
1375 type
->floating
= false;
1376 type
->fixed
= false;
1379 type
->width
= format_desc
->block
.bits
;
1384 for (i
= 0; i
< 4; i
++)
1385 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1389 memset(type
, 0, sizeof(struct lp_type
));
1390 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1391 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1392 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1393 type
->norm
= format_desc
->channel
[chan
].normalized
;
1395 if (is_arithmetic_format(format_desc
)) {
1399 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1400 type
->width
+= format_desc
->channel
[i
].size
;
1403 type
->width
= format_desc
->channel
[chan
].size
;
1404 type
->length
= format_desc
->nr_channels
;
1410 * Retrieves the type for a format which is usable in the blending code.
1412 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1415 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1416 struct lp_type
* type
)
1421 if (format_expands_to_float_soa(format_desc
)) {
1422 /* always use ordinary floats for blending */
1423 type
->floating
= true;
1424 type
->fixed
= false;
1432 for (i
= 0; i
< 4; i
++)
1433 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1437 memset(type
, 0, sizeof(struct lp_type
));
1438 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1439 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1440 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1441 type
->norm
= format_desc
->channel
[chan
].normalized
;
1442 type
->width
= format_desc
->channel
[chan
].size
;
1443 type
->length
= format_desc
->nr_channels
;
1445 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1446 if (format_desc
->channel
[i
].size
> type
->width
)
1447 type
->width
= format_desc
->channel
[i
].size
;
1450 if (type
->floating
) {
1453 if (type
->width
<= 8) {
1455 } else if (type
->width
<= 16) {
1462 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1469 * Scale a normalized value from src_bits to dst_bits.
1471 * The exact calculation is
1473 * dst = iround(src * dst_mask / src_mask)
1475 * or with integer rounding
1477 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1481 * src_mask = (1 << src_bits) - 1
1482 * dst_mask = (1 << dst_bits) - 1
1484 * but we try to avoid division and multiplication through shifts.
1486 static inline LLVMValueRef
1487 scale_bits(struct gallivm_state
*gallivm
,
1491 struct lp_type src_type
)
1493 LLVMBuilderRef builder
= gallivm
->builder
;
1494 LLVMValueRef result
= src
;
1496 if (dst_bits
< src_bits
) {
1497 int delta_bits
= src_bits
- dst_bits
;
1499 if (delta_bits
<= dst_bits
) {
1501 * Approximate the rescaling with a single shift.
1503 * This gives the wrong rounding.
1506 result
= LLVMBuildLShr(builder
,
1508 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1513 * Try more accurate rescaling.
1517 * Drop the least significant bits to make space for the multiplication.
1519 * XXX: A better approach would be to use a wider integer type as intermediate. But
1520 * this is enough to convert alpha from 16bits -> 2 when rendering to
1521 * PIPE_FORMAT_R10G10B10A2_UNORM.
1523 result
= LLVMBuildLShr(builder
,
1525 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1529 result
= LLVMBuildMul(builder
,
1531 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1535 * Add a rounding term before the division.
1537 * TODO: Handle signed integers too.
1539 if (!src_type
.sign
) {
1540 result
= LLVMBuildAdd(builder
,
1542 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1547 * Approximate the division by src_mask with a src_bits shift.
1549 * Given the src has already been shifted by dst_bits, all we need
1550 * to do is to shift by the difference.
1553 result
= LLVMBuildLShr(builder
,
1555 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1559 } else if (dst_bits
> src_bits
) {
1561 int db
= dst_bits
- src_bits
;
1563 /* Shift left by difference in bits */
1564 result
= LLVMBuildShl(builder
,
1566 lp_build_const_int_vec(gallivm
, src_type
, db
),
1569 if (db
<= src_bits
) {
1570 /* Enough bits in src to fill the remainder */
1571 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1573 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1576 result
= LLVMBuildOr(builder
, result
, lower
, "");
1577 } else if (db
> src_bits
) {
1578 /* Need to repeatedly copy src bits to fill remainder in dst */
1581 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1582 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1584 result
= LLVMBuildOr(builder
,
1586 LLVMBuildLShr(builder
, result
, shuv
, ""),
1596 * If RT is a smallfloat (needing denorms) format
1599 have_smallfloat_format(struct lp_type dst_type
,
1600 enum pipe_format format
)
1602 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1603 /* due to format handling hacks this format doesn't have floating set
1604 * here (and actually has width set to 32 too) so special case this. */
1605 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1610 * Convert from memory format to blending format
1612 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1615 convert_to_blend_type(struct gallivm_state
*gallivm
,
1616 unsigned block_size
,
1617 const struct util_format_description
*src_fmt
,
1618 struct lp_type src_type
,
1619 struct lp_type dst_type
,
1620 LLVMValueRef
* src
, // and dst
1623 LLVMValueRef
*dst
= src
;
1624 LLVMBuilderRef builder
= gallivm
->builder
;
1625 struct lp_type blend_type
;
1626 struct lp_type mem_type
;
1628 unsigned pixels
= block_size
/ num_srcs
;
1632 * full custom path for packed floats and srgb formats - none of the later
1633 * functions would do anything useful, and given the lp_type representation they
1634 * can't be fixed. Should really have some SoA blend path for these kind of
1635 * formats rather than hacking them in here.
1637 if (format_expands_to_float_soa(src_fmt
)) {
1638 LLVMValueRef tmpsrc
[4];
1640 * This is pretty suboptimal for this case blending in SoA would be much
1641 * better, since conversion gets us SoA values so need to convert back.
1643 assert(src_type
.width
== 32 || src_type
.width
== 16);
1644 assert(dst_type
.floating
);
1645 assert(dst_type
.width
== 32);
1646 assert(dst_type
.length
% 4 == 0);
1647 assert(num_srcs
% 4 == 0);
1649 if (src_type
.width
== 16) {
1650 /* expand 4x16bit values to 4x32bit */
1651 struct lp_type type32x4
= src_type
;
1652 LLVMTypeRef ltype32x4
;
1653 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1654 type32x4
.width
= 32;
1655 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1656 for (i
= 0; i
< num_fetch
; i
++) {
1657 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1659 src_type
.width
= 32;
1661 for (i
= 0; i
< 4; i
++) {
1664 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1665 LLVMValueRef tmpsoa
[4];
1666 LLVMValueRef tmps
= tmpsrc
[i
];
1667 if (dst_type
.length
== 8) {
1668 LLVMValueRef shuffles
[8];
1670 /* fetch was 4 values but need 8-wide output values */
1671 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1673 * for 8-wide aos transpose would give us wrong order not matching
1674 * incoming converted fs values and mask. ARGH.
1676 for (j
= 0; j
< 4; j
++) {
1677 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1678 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1680 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1681 LLVMConstVector(shuffles
, 8), "");
1683 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1684 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1687 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1689 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1694 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1695 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1697 /* Is the format arithmetic */
1698 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1699 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1701 /* Pad if necessary */
1702 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1703 for (i
= 0; i
< num_srcs
; ++i
) {
1704 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1707 src_type
.length
= dst_type
.length
;
1710 /* Special case for half-floats */
1711 if (mem_type
.width
== 16 && mem_type
.floating
) {
1712 assert(blend_type
.width
== 32 && blend_type
.floating
);
1713 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1721 src_type
.width
= blend_type
.width
* blend_type
.length
;
1722 blend_type
.length
*= pixels
;
1723 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1725 for (i
= 0; i
< num_srcs
; ++i
) {
1726 LLVMValueRef chans
[4];
1727 LLVMValueRef res
= NULL
;
1729 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1731 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1733 unsigned sa
= src_fmt
->channel
[j
].shift
;
1734 #if UTIL_ARCH_LITTLE_ENDIAN
1735 unsigned from_lsb
= j
;
1737 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1740 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1742 /* Extract bits from source */
1743 chans
[j
] = LLVMBuildLShr(builder
,
1745 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1748 chans
[j
] = LLVMBuildAnd(builder
,
1750 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1754 if (src_type
.norm
) {
1755 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1756 blend_type
.width
, chans
[j
], src_type
);
1759 /* Insert bits into correct position */
1760 chans
[j
] = LLVMBuildShl(builder
,
1762 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1768 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1772 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1778 * Convert from blending format to memory format
1780 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1783 convert_from_blend_type(struct gallivm_state
*gallivm
,
1784 unsigned block_size
,
1785 const struct util_format_description
*src_fmt
,
1786 struct lp_type src_type
,
1787 struct lp_type dst_type
,
1788 LLVMValueRef
* src
, // and dst
1791 LLVMValueRef
* dst
= src
;
1793 struct lp_type mem_type
;
1794 struct lp_type blend_type
;
1795 LLVMBuilderRef builder
= gallivm
->builder
;
1796 unsigned pixels
= block_size
/ num_srcs
;
1800 * full custom path for packed floats and srgb formats - none of the later
1801 * functions would do anything useful, and given the lp_type representation they
1802 * can't be fixed. Should really have some SoA blend path for these kind of
1803 * formats rather than hacking them in here.
1805 if (format_expands_to_float_soa(src_fmt
)) {
1807 * This is pretty suboptimal for this case blending in SoA would be much
1808 * better - we need to transpose the AoS values back to SoA values for
1809 * conversion/packing.
1811 assert(src_type
.floating
);
1812 assert(src_type
.width
== 32);
1813 assert(src_type
.length
% 4 == 0);
1814 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1816 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1817 LLVMValueRef tmpsoa
[4], tmpdst
;
1818 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1819 /* really really need SoA here */
1821 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1822 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1825 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1829 if (src_type
.length
== 8) {
1830 LLVMValueRef tmpaos
, shuffles
[8];
1833 * for 8-wide aos transpose has given us wrong order not matching
1834 * output order. HMPF. Also need to split the output values manually.
1836 for (j
= 0; j
< 4; j
++) {
1837 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1838 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1840 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1841 LLVMConstVector(shuffles
, 8), "");
1842 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1843 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1849 if (dst_type
.width
== 16) {
1850 struct lp_type type16x8
= dst_type
;
1851 struct lp_type type32x4
= dst_type
;
1852 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1853 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1854 type16x8
.length
= 8;
1855 type32x4
.width
= 32;
1856 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1857 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1858 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1859 /* We could do vector truncation but it doesn't generate very good code */
1860 for (i
= 0; i
< num_fetch
; i
++) {
1861 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1862 src
[i
], lp_build_zero(gallivm
, type32x4
));
1863 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1864 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1865 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1871 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1872 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1874 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1876 /* Special case for half-floats */
1877 if (mem_type
.width
== 16 && mem_type
.floating
) {
1878 int length
= dst_type
.length
;
1879 assert(blend_type
.width
== 32 && blend_type
.floating
);
1881 dst_type
.length
= src_type
.length
;
1883 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1885 dst_type
.length
= length
;
1889 /* Remove any padding */
1890 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1891 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1893 for (i
= 0; i
< num_srcs
; ++i
) {
1894 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1898 /* No bit arithmetic to do */
1903 src_type
.length
= pixels
;
1904 src_type
.width
= blend_type
.length
* blend_type
.width
;
1905 dst_type
.length
= pixels
;
1907 for (i
= 0; i
< num_srcs
; ++i
) {
1908 LLVMValueRef chans
[4];
1909 LLVMValueRef res
= NULL
;
1911 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1913 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1915 unsigned sa
= src_fmt
->channel
[j
].shift
;
1916 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1917 #if UTIL_ARCH_LITTLE_ENDIAN
1918 unsigned from_lsb
= j
;
1920 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1923 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1925 for (k
= 0; k
< blend_type
.width
; ++k
) {
1930 chans
[j
] = LLVMBuildLShr(builder
,
1932 lp_build_const_int_vec(gallivm
, src_type
,
1933 from_lsb
* blend_type
.width
),
1936 chans
[j
] = LLVMBuildAnd(builder
,
1938 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1941 /* Scale down bits */
1942 if (src_type
.norm
) {
1943 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1944 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1945 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1946 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1947 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1948 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1952 chans
[j
] = LLVMBuildShl(builder
,
1954 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1957 sa
+= src_fmt
->channel
[j
].size
;
1962 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1966 assert (dst_type
.width
!= 24);
1968 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1974 * Convert alpha to same blend type as src
1977 convert_alpha(struct gallivm_state
*gallivm
,
1978 struct lp_type row_type
,
1979 struct lp_type alpha_type
,
1980 const unsigned block_size
,
1981 const unsigned block_height
,
1982 const unsigned src_count
,
1983 const unsigned dst_channels
,
1984 const bool pad_inline
,
1985 LLVMValueRef
* src_alpha
)
1987 LLVMBuilderRef builder
= gallivm
->builder
;
1989 unsigned length
= row_type
.length
;
1990 row_type
.length
= alpha_type
.length
;
1992 /* Twiddle the alpha to match pixels */
1993 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1996 * TODO this should use single lp_build_conv call for
1997 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1999 for (i
= 0; i
< block_height
; ++i
) {
2000 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
2003 alpha_type
= row_type
;
2004 row_type
.length
= length
;
2006 /* If only one channel we can only need the single alpha value per pixel */
2007 if (src_count
== 1 && dst_channels
== 1) {
2009 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
2011 /* If there are more srcs than rows then we need to split alpha up */
2012 if (src_count
> block_height
) {
2013 for (i
= src_count
; i
> 0; --i
) {
2014 unsigned pixels
= block_size
/ src_count
;
2015 unsigned idx
= i
- 1;
2017 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
2018 (idx
* pixels
) % 4, pixels
);
2022 /* If there is a src for each pixel broadcast the alpha across whole row */
2023 if (src_count
== block_size
) {
2024 for (i
= 0; i
< src_count
; ++i
) {
2025 src_alpha
[i
] = lp_build_broadcast(gallivm
,
2026 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
2029 unsigned pixels
= block_size
/ src_count
;
2030 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
2031 unsigned alpha_span
= 1;
2032 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
2034 /* Check if we need 2 src_alphas for our shuffles */
2035 if (pixels
> alpha_type
.length
) {
2039 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
2040 for (j
= 0; j
< row_type
.length
; ++j
) {
2041 if (j
< pixels
* channels
) {
2042 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
2044 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
2048 for (i
= 0; i
< src_count
; ++i
) {
2049 unsigned idx1
= i
, idx2
= i
;
2051 if (alpha_span
> 1){
2056 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
2059 LLVMConstVector(shuffles
, row_type
.length
),
2068 * Generates the blend function for unswizzled colour buffers
2069 * Also generates the read & write from colour buffer
2072 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
2074 struct lp_fragment_shader_variant
*variant
,
2075 enum pipe_format out_format
,
2076 unsigned int num_fs
,
2077 struct lp_type fs_type
,
2078 LLVMValueRef
* fs_mask
,
2079 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
2080 LLVMValueRef context_ptr
,
2081 LLVMValueRef color_ptr
,
2082 LLVMValueRef stride
,
2083 unsigned partial_mask
,
2086 const unsigned alpha_channel
= 3;
2087 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
2088 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
2089 const unsigned block_size
= block_width
* block_height
;
2090 const unsigned lp_integer_vector_width
= 128;
2092 LLVMBuilderRef builder
= gallivm
->builder
;
2093 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
2094 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
2095 LLVMValueRef src_alpha
[4 * 4];
2096 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
2097 LLVMValueRef src_mask
[4 * 4];
2098 LLVMValueRef src
[4 * 4];
2099 LLVMValueRef src1
[4 * 4];
2100 LLVMValueRef dst
[4 * 4];
2101 LLVMValueRef blend_color
;
2102 LLVMValueRef blend_alpha
;
2103 LLVMValueRef i32_zero
;
2104 LLVMValueRef check_mask
;
2105 LLVMValueRef undef_src_val
;
2107 struct lp_build_mask_context mask_ctx
;
2108 struct lp_type mask_type
;
2109 struct lp_type blend_type
;
2110 struct lp_type row_type
;
2111 struct lp_type dst_type
;
2112 struct lp_type ls_type
;
2114 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
2115 unsigned vector_width
;
2116 unsigned src_channels
= TGSI_NUM_CHANNELS
;
2117 unsigned dst_channels
;
2122 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
2124 unsigned dst_alignment
;
2126 bool pad_inline
= is_arithmetic_format(out_format_desc
);
2127 bool has_alpha
= false;
2128 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
2129 util_blend_state_is_dual(&variant
->key
.blend
, 0);
2131 const boolean is_1d
= variant
->key
.resource_1d
;
2132 boolean twiddle_after_convert
= FALSE
;
2133 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
2134 LLVMValueRef fpstate
= 0;
2136 /* Get type from output format */
2137 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
2138 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
2141 * Technically this code should go into lp_build_smallfloat_to_float
2142 * and lp_build_float_to_smallfloat but due to the
2143 * http://llvm.org/bugs/show_bug.cgi?id=6393
2144 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2145 * So the ordering is important here and there shouldn't be any
2146 * llvm ir instrunctions in this function before
2147 * this, otherwise half-float format conversions won't work
2148 * (again due to llvm bug #6393).
2150 if (have_smallfloat_format(dst_type
, out_format
)) {
2151 /* We need to make sure that denorms are ok for half float
2153 fpstate
= lp_build_fpstate_get(gallivm
);
2154 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
2157 mask_type
= lp_int32_vec4_type();
2158 mask_type
.length
= fs_type
.length
;
2160 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
2161 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
2164 /* Do not bother executing code when mask is empty.. */
2166 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
2168 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2169 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
2172 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
2173 lp_build_mask_check(&mask_ctx
);
2176 partial_mask
|= !variant
->opaque
;
2177 i32_zero
= lp_build_const_int32(gallivm
, 0);
2179 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2181 row_type
.length
= fs_type
.length
;
2182 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2184 /* Compute correct swizzle and count channels */
2185 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2188 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2189 /* Ensure channel is used */
2190 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2194 /* Ensure not already written to (happens in case with GL_ALPHA) */
2195 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2199 /* Ensure we havn't already found all channels */
2200 if (dst_channels
>= out_format_desc
->nr_channels
) {
2204 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2207 if (i
== alpha_channel
) {
2212 if (format_expands_to_float_soa(out_format_desc
)) {
2214 * the code above can't work for layout_other
2215 * for srgb it would sort of work but we short-circuit swizzles, etc.
2216 * as that is done as part of unpack / pack.
2218 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2224 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2227 /* If 3 channels then pad to include alpha for 4 element transpose */
2228 if (dst_channels
== 3) {
2229 assert (!has_alpha
);
2230 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2231 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2234 if (out_format_desc
->nr_channels
== 4) {
2237 * We use alpha from the color conversion, not separate one.
2238 * We had to include it for transpose, hence it will get converted
2239 * too (albeit when doing transpose after conversion, that would
2240 * no longer be the case necessarily).
2241 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2242 * otherwise we really have padding, not alpha, included.)
2249 * Load shader output
2251 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2252 /* Always load alpha for use in blending */
2255 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2258 alpha
= undef_src_val
;
2261 /* Load each channel */
2262 for (j
= 0; j
< dst_channels
; ++j
) {
2263 assert(swizzle
[j
] < 4);
2265 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2268 fs_src
[i
][j
] = undef_src_val
;
2272 /* If 3 channels then pad to include alpha for 4 element transpose */
2274 * XXX If we include that here maybe could actually use it instead of
2275 * separate alpha for blending?
2276 * (Difficult though we actually convert pad channels, not alpha.)
2278 if (dst_channels
== 3 && !has_alpha
) {
2279 fs_src
[i
][3] = alpha
;
2282 /* We split the row_mask and row_alpha as we want 128bit interleave */
2283 if (fs_type
.length
== 8) {
2284 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2286 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2287 src_channels
, src_channels
);
2289 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2290 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2291 src_channels
, src_channels
);
2293 src_mask
[i
] = fs_mask
[i
];
2294 src_alpha
[i
] = alpha
;
2297 if (dual_source_blend
) {
2298 /* same as above except different src/dst, skip masks and comments... */
2299 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2302 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2305 alpha
= undef_src_val
;
2308 for (j
= 0; j
< dst_channels
; ++j
) {
2309 assert(swizzle
[j
] < 4);
2311 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2314 fs_src1
[i
][j
] = undef_src_val
;
2317 if (dst_channels
== 3 && !has_alpha
) {
2318 fs_src1
[i
][3] = alpha
;
2320 if (fs_type
.length
== 8) {
2321 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2322 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2323 src_channels
, src_channels
);
2325 src1_alpha
[i
] = alpha
;
2330 if (util_format_is_pure_integer(out_format
)) {
2332 * In this case fs_type was really ints or uints disguised as floats,
2335 fs_type
.floating
= 0;
2336 fs_type
.sign
= dst_type
.sign
;
2337 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2338 for (j
= 0; j
< dst_channels
; ++j
) {
2339 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2340 lp_build_vec_type(gallivm
, fs_type
), "");
2342 if (dst_channels
== 3 && !has_alpha
) {
2343 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2344 lp_build_vec_type(gallivm
, fs_type
), "");
2350 * We actually should generally do conversion first (for non-1d cases)
2351 * when the blend format is 8 or 16 bits. The reason is obvious,
2352 * there's 2 or 4 times less vectors to deal with for the interleave...
2353 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2354 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2355 * unpack only with 128bit vectors).
2356 * Note: for 16bit sizes really need matching pack conversion code
2358 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2359 twiddle_after_convert
= TRUE
;
2363 * Pixel twiddle from fragment shader order to memory order
2365 if (!twiddle_after_convert
) {
2366 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2367 dst_channels
, fs_src
, src
, pad_inline
);
2368 if (dual_source_blend
) {
2369 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2370 fs_src1
, src1
, pad_inline
);
2373 src_count
= num_fullblock_fs
* dst_channels
;
2375 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2376 * (AVX) turn out the same later when untwiddling/transpose (albeit
2377 * for true AVX2 path untwiddle needs to be different).
2378 * For now just order by colors first (so we can use unpack later).
2380 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2381 for (i
= 0; i
< dst_channels
; i
++) {
2382 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2383 if (dual_source_blend
) {
2384 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2390 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2391 if (src_count
!= num_fullblock_fs
* src_channels
) {
2392 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2393 row_type
.length
/= ds
;
2394 fs_type
.length
= row_type
.length
;
2397 blend_type
= row_type
;
2398 mask_type
.length
= 4;
2400 /* Convert src to row_type */
2401 if (dual_source_blend
) {
2402 struct lp_type old_row_type
= row_type
;
2403 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2404 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2407 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2410 /* If the rows are not an SSE vector, combine them to become SSE size! */
2411 if ((row_type
.width
* row_type
.length
) % 128) {
2412 unsigned bits
= row_type
.width
* row_type
.length
;
2415 assert(src_count
>= (vector_width
/ bits
));
2417 dst_count
= src_count
/ (vector_width
/ bits
);
2419 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2420 if (dual_source_blend
) {
2421 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2424 row_type
.length
*= combined
;
2425 src_count
/= combined
;
2427 bits
= row_type
.width
* row_type
.length
;
2428 assert(bits
== 128 || bits
== 256);
2431 if (twiddle_after_convert
) {
2432 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2433 if (dual_source_blend
) {
2434 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2439 * Blend Colour conversion
2441 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2442 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2443 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2444 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2445 &i32_zero
, 1, ""), "");
2448 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2450 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2452 * since blending is done with floats, there was no conversion.
2453 * However, the rules according to fixed point renderbuffers still
2454 * apply, that is we must clamp inputs to 0.0/1.0.
2455 * (This would apply to separate alpha conversion too but we currently
2456 * force has_alpha to be true.)
2457 * TODO: should skip this with "fake" blend, since post-blend conversion
2458 * will clamp anyway.
2459 * TODO: could also skip this if fragment color clamping is enabled. We
2460 * don't support it natively so it gets baked into the shader however, so
2461 * can't really tell here.
2463 struct lp_build_context f32_bld
;
2464 assert(row_type
.floating
);
2465 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2466 for (i
= 0; i
< src_count
; i
++) {
2467 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2469 if (dual_source_blend
) {
2470 for (i
= 0; i
< src_count
; i
++) {
2471 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2474 /* probably can't be different than row_type but better safe than sorry... */
2475 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2476 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2480 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2482 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2483 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2485 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2486 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2488 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2489 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2495 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2497 if (src_count
< block_height
) {
2498 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2499 } else if (src_count
> block_height
) {
2500 for (i
= src_count
; i
> 0; --i
) {
2501 unsigned pixels
= block_size
/ src_count
;
2502 unsigned idx
= i
- 1;
2504 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2505 (idx
* pixels
) % 4, pixels
);
2509 assert(mask_type
.width
== 32);
2511 for (i
= 0; i
< src_count
; ++i
) {
2512 unsigned pixels
= block_size
/ src_count
;
2513 unsigned pixel_width
= row_type
.width
* dst_channels
;
2515 if (pixel_width
== 24) {
2516 mask_type
.width
= 8;
2517 mask_type
.length
= vector_width
/ mask_type
.width
;
2519 mask_type
.length
= pixels
;
2520 mask_type
.width
= row_type
.width
* dst_channels
;
2523 * If mask_type width is smaller than 32bit, this doesn't quite
2524 * generate the most efficient code (could use some pack).
2526 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2527 lp_build_int_vec_type(gallivm
, mask_type
), "");
2529 mask_type
.length
*= dst_channels
;
2530 mask_type
.width
/= dst_channels
;
2533 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2534 lp_build_int_vec_type(gallivm
, mask_type
), "");
2535 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2542 struct lp_type alpha_type
= fs_type
;
2543 alpha_type
.length
= 4;
2544 convert_alpha(gallivm
, row_type
, alpha_type
,
2545 block_size
, block_height
,
2546 src_count
, dst_channels
,
2547 pad_inline
, src_alpha
);
2548 if (dual_source_blend
) {
2549 convert_alpha(gallivm
, row_type
, alpha_type
,
2550 block_size
, block_height
,
2551 src_count
, dst_channels
,
2552 pad_inline
, src1_alpha
);
2558 * Load dst from memory
2560 if (src_count
< block_height
) {
2561 dst_count
= block_height
;
2563 dst_count
= src_count
;
2566 dst_type
.length
*= block_size
/ dst_count
;
2568 if (format_expands_to_float_soa(out_format_desc
)) {
2570 * we need multiple values at once for the conversion, so can as well
2571 * load them vectorized here too instead of concatenating later.
2572 * (Still need concatenation later for 8-wide vectors).
2574 dst_count
= block_height
;
2575 dst_type
.length
= block_width
;
2579 * Compute the alignment of the destination pointer in bytes
2580 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2581 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2582 * 1d tex but can't distinguish here) so need to stick with per-pixel
2583 * alignment in this case.
2586 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2589 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2591 /* Force power-of-two alignment by extracting only the least-significant-bit */
2592 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2594 * Resource base and stride pointers are aligned to 16 bytes, so that's
2595 * the maximum alignment we can guarantee
2597 dst_alignment
= MIN2(16, dst_alignment
);
2601 if (dst_count
> src_count
) {
2602 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2603 util_is_power_of_two_or_zero(dst_type
.length
) &&
2604 dst_type
.length
* dst_type
.width
< 128) {
2606 * Never try to load values as 4xi8 which we will then
2607 * concatenate to larger vectors. This gives llvm a real
2608 * headache (the problem is the type legalizer (?) will
2609 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2610 * then the shuffles to concatenate are more or less impossible
2611 * - llvm is easily capable of generating a sequence of 32
2612 * pextrb/pinsrb instructions for that. Albeit it appears to
2613 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2614 * width to avoid the trouble (16bit seems not as bad, llvm
2615 * probably recognizes the load+shuffle as only one shuffle
2616 * is necessary, but we can do just the same anyway).
2618 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2624 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2625 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2626 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2627 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2632 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2633 dst
, ls_type
, dst_count
, dst_alignment
);
2638 * Convert from dst/output format to src/blending format.
2640 * This is necessary as we can only read 1 row from memory at a time,
2641 * so the minimum dst_count will ever be at this point is 4.
2643 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2644 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2645 * on all 16 pixels in that single vector at once.
2647 if (dst_count
> src_count
) {
2648 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2649 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2650 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2651 for (i
= 0; i
< dst_count
; i
++) {
2652 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2656 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2658 if (ls_type
.length
!= dst_type
.length
) {
2659 struct lp_type tmp_type
= dst_type
;
2660 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2661 for (i
= 0; i
< src_count
; i
++) {
2662 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2663 lp_build_vec_type(gallivm
, tmp_type
), "");
2671 /* XXX this is broken for RGB8 formats -
2672 * they get expanded from 12 to 16 elements (to include alpha)
2673 * by convert_to_blend_type then reduced to 15 instead of 12
2674 * by convert_from_blend_type (a simple fix though breaks A8...).
2675 * R16G16B16 also crashes differently however something going wrong
2676 * inside llvm handling npot vector sizes seemingly.
2677 * It seems some cleanup could be done here (like skipping conversion/blend
2680 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2681 row_type
, dst
, src_count
);
2684 * FIXME: Really should get logic ops / masks out of generic blend / row
2685 * format. Logic ops will definitely not work on the blend float format
2686 * used for SRGB here and I think OpenGL expects this to work as expected
2687 * (that is incoming values converted to srgb then logic op applied).
2689 for (i
= 0; i
< src_count
; ++i
) {
2690 dst
[i
] = lp_build_blend_aos(gallivm
,
2691 &variant
->key
.blend
,
2696 has_alpha
? NULL
: src_alpha
[i
],
2698 has_alpha
? NULL
: src1_alpha
[i
],
2700 partial_mask
? src_mask
[i
] : NULL
,
2702 has_alpha
? NULL
: blend_alpha
,
2704 pad_inline
? 4 : dst_channels
);
2707 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2708 row_type
, dst_type
, dst
, src_count
);
2710 /* Split the blend rows back to memory rows */
2711 if (dst_count
> src_count
) {
2712 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2714 if (src_count
== 1) {
2715 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2716 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2718 row_type
.length
/= 2;
2722 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2723 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2724 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2725 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2727 row_type
.length
/= 2;
2732 * Store blend result to memory
2735 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2736 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2739 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2740 dst
, dst_type
, dst_count
, dst_alignment
);
2743 if (have_smallfloat_format(dst_type
, out_format
)) {
2744 lp_build_fpstate_set(gallivm
, fpstate
);
2748 lp_build_mask_end(&mask_ctx
);
2754 * Generate the runtime callable function for the whole fragment pipeline.
2755 * Note that the function which we generate operates on a block of 16
2756 * pixels at at time. The block contains 2x2 quads. Each quad contains
2760 generate_fragment(struct llvmpipe_context
*lp
,
2761 struct lp_fragment_shader
*shader
,
2762 struct lp_fragment_shader_variant
*variant
,
2763 unsigned partial_mask
)
2765 struct gallivm_state
*gallivm
= variant
->gallivm
;
2766 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2767 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2769 struct lp_type fs_type
;
2770 struct lp_type blend_type
;
2771 LLVMTypeRef fs_elem_type
;
2772 LLVMTypeRef blend_vec_type
;
2773 LLVMTypeRef arg_types
[15];
2774 LLVMTypeRef func_type
;
2775 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2776 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2777 LLVMValueRef context_ptr
;
2780 LLVMValueRef a0_ptr
;
2781 LLVMValueRef dadx_ptr
;
2782 LLVMValueRef dady_ptr
;
2783 LLVMValueRef color_ptr_ptr
;
2784 LLVMValueRef stride_ptr
;
2785 LLVMValueRef color_sample_stride_ptr
;
2786 LLVMValueRef depth_ptr
;
2787 LLVMValueRef depth_stride
;
2788 LLVMValueRef depth_sample_stride
;
2789 LLVMValueRef mask_input
;
2790 LLVMValueRef thread_data_ptr
;
2791 LLVMBasicBlockRef block
;
2792 LLVMBuilderRef builder
;
2793 struct lp_build_sampler_soa
*sampler
;
2794 struct lp_build_image_soa
*image
;
2795 struct lp_build_interp_soa_context interp
;
2796 LLVMValueRef fs_mask
[(16 / 4) * LP_MAX_SAMPLES
];
2797 LLVMValueRef fs_out_color
[LP_MAX_SAMPLES
][PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2798 LLVMValueRef function
;
2799 LLVMValueRef facing
;
2804 boolean cbuf0_write_all
;
2805 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2806 util_blend_state_is_dual(&key
->blend
, 0);
2808 assert(lp_native_vector_width
/ 32 >= 4);
2810 /* Adjust color input interpolation according to flatshade state:
2812 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2813 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2814 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2816 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2818 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2822 /* check if writes to cbuf[0] are to be copied to all cbufs */
2824 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2826 /* TODO: actually pick these based on the fs and color buffer
2827 * characteristics. */
2829 memset(&fs_type
, 0, sizeof fs_type
);
2830 fs_type
.floating
= TRUE
; /* floating point values */
2831 fs_type
.sign
= TRUE
; /* values are signed */
2832 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2833 fs_type
.width
= 32; /* 32-bit float */
2834 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2836 memset(&blend_type
, 0, sizeof blend_type
);
2837 blend_type
.floating
= FALSE
; /* values are integers */
2838 blend_type
.sign
= FALSE
; /* values are unsigned */
2839 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2840 blend_type
.width
= 8; /* 8-bit ubyte values */
2841 blend_type
.length
= 16; /* 16 elements per vector */
2844 * Generate the function prototype. Any change here must be reflected in
2845 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2848 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2850 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2852 snprintf(func_name
, sizeof(func_name
), "fs_variant_%s",
2853 partial_mask
? "partial" : "whole");
2855 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2856 arg_types
[1] = int32_type
; /* x */
2857 arg_types
[2] = int32_type
; /* y */
2858 arg_types
[3] = int32_type
; /* facing */
2859 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2860 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2861 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2862 arg_types
[7] = LLVMPointerType(LLVMPointerType(int8_type
, 0), 0); /* color */
2863 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2864 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2865 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2866 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2867 arg_types
[12] = int32_type
; /* depth_stride */
2868 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2869 arg_types
[14] = int32_type
; /* depth sample stride */
2871 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2872 arg_types
, ARRAY_SIZE(arg_types
), 0);
2874 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2875 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2877 variant
->function
[partial_mask
] = function
;
2879 /* XXX: need to propagate noalias down into color param now we are
2880 * passing a pointer-to-pointer?
2882 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2883 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2884 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2886 if (variant
->gallivm
->cache
->data_size
)
2889 context_ptr
= LLVMGetParam(function
, 0);
2890 x
= LLVMGetParam(function
, 1);
2891 y
= LLVMGetParam(function
, 2);
2892 facing
= LLVMGetParam(function
, 3);
2893 a0_ptr
= LLVMGetParam(function
, 4);
2894 dadx_ptr
= LLVMGetParam(function
, 5);
2895 dady_ptr
= LLVMGetParam(function
, 6);
2896 color_ptr_ptr
= LLVMGetParam(function
, 7);
2897 depth_ptr
= LLVMGetParam(function
, 8);
2898 mask_input
= LLVMGetParam(function
, 9);
2899 thread_data_ptr
= LLVMGetParam(function
, 10);
2900 stride_ptr
= LLVMGetParam(function
, 11);
2901 depth_stride
= LLVMGetParam(function
, 12);
2902 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2903 depth_sample_stride
= LLVMGetParam(function
, 14);
2905 lp_build_name(context_ptr
, "context");
2906 lp_build_name(x
, "x");
2907 lp_build_name(y
, "y");
2908 lp_build_name(a0_ptr
, "a0");
2909 lp_build_name(dadx_ptr
, "dadx");
2910 lp_build_name(dady_ptr
, "dady");
2911 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2912 lp_build_name(depth_ptr
, "depth");
2913 lp_build_name(mask_input
, "mask_input");
2914 lp_build_name(thread_data_ptr
, "thread_data");
2915 lp_build_name(stride_ptr
, "stride_ptr");
2916 lp_build_name(depth_stride
, "depth_stride");
2917 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2918 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2924 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2925 builder
= gallivm
->builder
;
2927 LLVMPositionBuilderAtEnd(builder
, block
);
2930 * Must not count ps invocations if there's a null shader.
2931 * (It would be ok to count with null shader if there's d/s tests,
2932 * but only if there's d/s buffers too, which is different
2933 * to implicit rasterization disable which must not depend
2934 * on the d/s buffers.)
2935 * Could use popcount on mask, but pixel accuracy is not required.
2936 * Could disable if there's no stats query, but maybe not worth it.
2938 if (shader
->info
.base
.num_instructions
> 1) {
2939 LLVMValueRef invocs
, val
;
2940 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2941 val
= LLVMBuildLoad(builder
, invocs
, "");
2942 val
= LLVMBuildAdd(builder
, val
,
2943 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2945 LLVMBuildStore(builder
, val
, invocs
);
2948 /* code generated texture sampling */
2949 sampler
= lp_llvm_sampler_soa_create(key
->samplers
, key
->nr_samplers
);
2950 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
), key
->nr_images
);
2952 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2953 /* for 1d resources only run "upper half" of stamp */
2954 if (key
->resource_1d
)
2958 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2959 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2960 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2961 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2962 num_loop_samp
, "mask_store");
2964 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(gallivm
->context
);
2965 LLVMValueRef glob_sample_pos
= LLVMAddGlobal(gallivm
->module
, LLVMArrayType(flt_type
, key
->coverage_samples
* 2), "");
2966 LLVMValueRef sample_pos_array
;
2968 if (key
->multisample
&& key
->coverage_samples
== 4) {
2969 LLVMValueRef sample_pos_arr
[8];
2970 for (unsigned i
= 0; i
< 4; i
++) {
2971 sample_pos_arr
[i
* 2] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][0]);
2972 sample_pos_arr
[i
* 2 + 1] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][1]);
2974 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 8);
2976 LLVMValueRef sample_pos_arr
[2];
2977 sample_pos_arr
[0] = LLVMConstReal(flt_type
, 0.5);
2978 sample_pos_arr
[1] = LLVMConstReal(flt_type
, 0.5);
2979 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 2);
2981 LLVMSetInitializer(glob_sample_pos
, sample_pos_array
);
2983 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2984 boolean pixel_center_integer
=
2985 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2988 * The shader input interpolation info is not explicitely baked in the
2989 * shader key, but everything it derives from (TGSI, and flatshade) is
2990 * already included in the shader key.
2992 lp_build_interp_soa_init(&interp
,
2994 shader
->info
.base
.num_inputs
,
2996 pixel_center_integer
,
2997 key
->coverage_samples
, glob_sample_pos
,
3001 a0_ptr
, dadx_ptr
, dady_ptr
,
3004 for (i
= 0; i
< num_fs
; i
++) {
3005 if (key
->multisample
) {
3006 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
3009 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
3010 * store to the per sample mask storage. Or all of them together to generate
3011 * the fragment shader mask. (sample shading TODO).
3012 * Take the incoming state coverage mask into account.
3014 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
3015 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
3016 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
3017 &sindexi
, 1, "sample_mask_ptr");
3018 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
3019 i
*fs_type
.length
/4, s
, mask_input
);
3021 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
3022 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
3023 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
3024 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
3026 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
3027 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
3031 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
3032 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
3033 &indexi
, 1, "mask_ptr");
3036 mask
= generate_quad_mask(gallivm
, fs_type
,
3037 i
*fs_type
.length
/4, 0, mask_input
);
3040 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
3042 LLVMBuildStore(builder
, mask
, mask_ptr
);
3046 generate_fs_loop(gallivm
,
3056 mask_store
, /* output */
3060 depth_sample_stride
,
3064 for (i
= 0; i
< num_fs
; i
++) {
3066 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
3067 int idx
= (i
+ (s
* num_fs
));
3068 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3069 ptr
= LLVMBuildGEP(builder
, mask_store
, &sindexi
, 1, "");
3071 fs_mask
[idx
] = LLVMBuildLoad(builder
, ptr
, "smask");
3074 for (unsigned s
= 0; s
< key
->min_samples
; s
++) {
3075 /* This is fucked up need to reorganize things */
3076 int idx
= s
* num_fs
+ i
;
3077 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3078 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3079 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3080 ptr
= LLVMBuildGEP(builder
,
3081 color_store
[cbuf
* !cbuf0_write_all
][chan
],
3083 fs_out_color
[s
][cbuf
][chan
][i
] = ptr
;
3086 if (dual_source_blend
) {
3087 /* only support one dual source blend target hence always use output 1 */
3088 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3089 ptr
= LLVMBuildGEP(builder
,
3090 color_store
[1][chan
],
3092 fs_out_color
[s
][1][chan
][i
] = ptr
;
3099 sampler
->destroy(sampler
);
3100 image
->destroy(image
);
3101 /* Loop over color outputs / color buffers to do blending.
3103 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3104 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
3105 LLVMValueRef color_ptr
;
3106 LLVMValueRef stride
;
3107 LLVMValueRef sample_stride
= NULL
;
3108 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
3110 boolean do_branch
= ((key
->depth
.enabled
3111 || key
->stencil
[0].enabled
3112 || key
->alpha
.enabled
)
3113 && !shader
->info
.base
.uses_kill
);
3115 color_ptr
= LLVMBuildLoad(builder
,
3116 LLVMBuildGEP(builder
, color_ptr_ptr
,
3120 stride
= LLVMBuildLoad(builder
,
3121 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
3124 if (key
->multisample
)
3125 sample_stride
= LLVMBuildLoad(builder
,
3126 LLVMBuildGEP(builder
, color_sample_stride_ptr
,
3127 &index
, 1, ""), "");
3129 for (unsigned s
= 0; s
< key
->cbuf_nr_samples
[cbuf
]; s
++) {
3130 unsigned mask_idx
= num_fs
* (key
->multisample
? s
: 0);
3131 unsigned out_idx
= key
->min_samples
== 1 ? 0 : s
;
3132 LLVMValueRef out_ptr
= color_ptr
;;
3134 if (key
->multisample
) {
3135 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_stride
, lp_build_const_int32(gallivm
, s
), "");
3136 out_ptr
= LLVMBuildGEP(builder
, out_ptr
, &sample_offset
, 1, "");
3138 out_ptr
= LLVMBuildBitCast(builder
, out_ptr
, LLVMPointerType(blend_vec_type
, 0), "");
3140 lp_build_name(out_ptr
, "color_ptr%d", cbuf
);
3142 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
3143 key
->cbuf_format
[cbuf
],
3144 num_fs
, fs_type
, &fs_mask
[mask_idx
], fs_out_color
[out_idx
],
3145 context_ptr
, out_ptr
, stride
,
3146 partial_mask
, do_branch
);
3151 LLVMBuildRetVoid(builder
);
3153 gallivm_verify_function(gallivm
, function
);
3158 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
3162 debug_printf("fs variant %p:\n", (void *) key
);
3164 if (key
->flatshade
) {
3165 debug_printf("flatshade = 1\n");
3167 if (key
->multisample
) {
3168 debug_printf("multisample = 1\n");
3169 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
3170 debug_printf("min samples = %d\n", key
->min_samples
);
3172 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
3173 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
3174 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
3176 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
3177 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
3178 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
3180 if (key
->depth
.enabled
) {
3181 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
3182 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
3185 for (i
= 0; i
< 2; ++i
) {
3186 if (key
->stencil
[i
].enabled
) {
3187 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
3188 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
3189 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
3190 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
3191 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
3192 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
3196 if (key
->alpha
.enabled
) {
3197 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
3200 if (key
->occlusion_count
) {
3201 debug_printf("occlusion_count = 1\n");
3204 if (key
->blend
.logicop_enable
) {
3205 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
3207 else if (key
->blend
.rt
[0].blend_enable
) {
3208 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
3209 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
3210 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
3211 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
3212 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
3213 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
3215 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
3216 if (key
->blend
.alpha_to_coverage
) {
3217 debug_printf("blend.alpha_to_coverage is enabled\n");
3219 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
3220 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
3221 debug_printf("sampler[%u] = \n", i
);
3222 debug_printf(" .wrap = %s %s %s\n",
3223 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
3224 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
3225 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
3226 debug_printf(" .min_img_filter = %s\n",
3227 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
3228 debug_printf(" .min_mip_filter = %s\n",
3229 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
3230 debug_printf(" .mag_img_filter = %s\n",
3231 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
3232 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
3233 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3234 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3235 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3236 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3237 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3238 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3240 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3241 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3242 debug_printf("texture[%u] = \n", i
);
3243 debug_printf(" .format = %s\n",
3244 util_format_name(texture
->format
));
3245 debug_printf(" .target = %s\n",
3246 util_str_tex_target(texture
->target
, TRUE
));
3247 debug_printf(" .level_zero_only = %u\n",
3248 texture
->level_zero_only
);
3249 debug_printf(" .pot = %u %u %u\n",
3251 texture
->pot_height
,
3252 texture
->pot_depth
);
3254 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3255 for (i
= 0; i
< key
->nr_images
; ++i
) {
3256 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3257 debug_printf("image[%u] = \n", i
);
3258 debug_printf(" .format = %s\n",
3259 util_format_name(image
->format
));
3260 debug_printf(" .target = %s\n",
3261 util_str_tex_target(image
->target
, TRUE
));
3262 debug_printf(" .level_zero_only = %u\n",
3263 image
->level_zero_only
);
3264 debug_printf(" .pot = %u %u %u\n",
3273 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3275 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3276 variant
->shader
->no
, variant
->no
);
3277 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3278 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3280 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3281 dump_fs_variant_key(&variant
->key
);
3282 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3287 lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant
*variant
,
3288 unsigned char ir_sha1_cache_key
[20])
3290 struct blob blob
= { 0 };
3295 nir_serialize(&blob
, variant
->shader
->base
.ir
.nir
, true);
3296 ir_binary
= blob
.data
;
3297 ir_size
= blob
.size
;
3299 struct mesa_sha1 ctx
;
3300 _mesa_sha1_init(&ctx
);
3301 _mesa_sha1_update(&ctx
, &variant
->key
, variant
->shader
->variant_key_size
);
3302 _mesa_sha1_update(&ctx
, ir_binary
, ir_size
);
3303 _mesa_sha1_final(&ctx
, ir_sha1_cache_key
);
3309 * Generate a new fragment shader variant from the shader code and
3310 * other state indicated by the key.
3312 static struct lp_fragment_shader_variant
*
3313 generate_variant(struct llvmpipe_context
*lp
,
3314 struct lp_fragment_shader
*shader
,
3315 const struct lp_fragment_shader_variant_key
*key
)
3317 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
3318 struct lp_fragment_shader_variant
*variant
;
3319 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3320 boolean fullcolormask
;
3321 char module_name
[64];
3322 unsigned char ir_sha1_cache_key
[20];
3323 struct lp_cached_code cached
= { 0 };
3324 bool needs_caching
= false;
3325 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3329 memset(variant
, 0, sizeof(*variant
));
3330 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3331 shader
->no
, shader
->variants_created
);
3333 variant
->shader
= shader
;
3334 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3336 if (shader
->base
.ir
.nir
) {
3337 lp_fs_get_ir_cache_key(variant
, ir_sha1_cache_key
);
3339 lp_disk_cache_find_shader(screen
, &cached
, ir_sha1_cache_key
);
3340 if (!cached
.data_size
)
3341 needs_caching
= true;
3343 variant
->gallivm
= gallivm_create(module_name
, lp
->context
, &cached
);
3344 if (!variant
->gallivm
) {
3349 variant
->list_item_global
.base
= variant
;
3350 variant
->list_item_local
.base
= variant
;
3351 variant
->no
= shader
->variants_created
++;
3356 * Determine whether we are touching all channels in the color buffer.
3358 fullcolormask
= FALSE
;
3359 if (key
->nr_cbufs
== 1) {
3360 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3361 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3365 !key
->blend
.logicop_enable
&&
3366 !key
->blend
.rt
[0].blend_enable
&&
3368 !key
->stencil
[0].enabled
&&
3369 !key
->alpha
.enabled
&&
3370 !key
->multisample
&&
3371 !key
->blend
.alpha_to_coverage
&&
3372 !key
->depth
.enabled
&&
3373 !shader
->info
.base
.uses_kill
&&
3374 !shader
->info
.base
.writes_samplemask
3377 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3378 lp_debug_fs_variant(variant
);
3381 lp_jit_init_types(variant
);
3383 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3384 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3386 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3387 if (variant
->opaque
) {
3388 /* Specialized shader, which doesn't need to read the color buffer. */
3389 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3394 * Compile everything
3397 gallivm_compile_module(variant
->gallivm
);
3399 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3401 if (variant
->function
[RAST_EDGE_TEST
]) {
3402 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3403 gallivm_jit_function(variant
->gallivm
,
3404 variant
->function
[RAST_EDGE_TEST
]);
3407 if (variant
->function
[RAST_WHOLE
]) {
3408 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3409 gallivm_jit_function(variant
->gallivm
,
3410 variant
->function
[RAST_WHOLE
]);
3411 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3412 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3415 if (needs_caching
) {
3416 lp_disk_cache_insert_shader(screen
, &cached
, ir_sha1_cache_key
);
3419 gallivm_free_ir(variant
->gallivm
);
3426 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3427 const struct pipe_shader_state
*templ
)
3429 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3430 struct lp_fragment_shader
*shader
;
3432 int nr_sampler_views
;
3436 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3440 shader
->no
= fs_no
++;
3441 make_empty_list(&shader
->variants
);
3443 shader
->base
.type
= templ
->type
;
3444 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3445 /* get/save the summary info for this shader */
3446 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3448 /* we need to keep a local copy of the tokens */
3449 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3451 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3452 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3455 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3456 if (shader
->draw_data
== NULL
) {
3457 FREE((void *) shader
->base
.tokens
);
3462 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3463 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3464 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3465 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3467 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3468 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3469 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3470 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3472 switch (shader
->info
.base
.input_interpolate
[i
]) {
3473 case TGSI_INTERPOLATE_CONSTANT
:
3474 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3476 case TGSI_INTERPOLATE_LINEAR
:
3477 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3479 case TGSI_INTERPOLATE_PERSPECTIVE
:
3480 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3482 case TGSI_INTERPOLATE_COLOR
:
3483 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3490 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3491 case TGSI_SEMANTIC_FACE
:
3492 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3494 case TGSI_SEMANTIC_POSITION
:
3495 /* Position was already emitted above
3497 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3498 shader
->inputs
[i
].src_index
= 0;
3502 /* XXX this is a completely pointless index map... */
3503 shader
->inputs
[i
].src_index
= i
+1;
3506 if (LP_DEBUG
& DEBUG_TGSI
) {
3508 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3509 shader
->no
, (void *) shader
);
3510 tgsi_dump(templ
->tokens
, 0);
3511 debug_printf("usage masks:\n");
3512 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3513 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3514 debug_printf(" IN[%u].%s%s%s%s\n",
3516 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3517 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3518 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3519 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3529 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3531 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3532 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3533 if (llvmpipe
->fs
== lp_fs
)
3536 draw_bind_fragment_shader(llvmpipe
->draw
,
3537 (lp_fs
? lp_fs
->draw_data
: NULL
));
3539 llvmpipe
->fs
= lp_fs
;
3541 llvmpipe
->dirty
|= LP_NEW_FS
;
3546 * Remove shader variant from two lists: the shader's variant list
3547 * and the context's variant list.
3550 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3551 struct lp_fragment_shader_variant
*variant
)
3553 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3554 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3555 "v total cached %u inst %u total inst %u\n",
3556 variant
->shader
->no
, variant
->no
,
3557 variant
->shader
->variants_created
,
3558 variant
->shader
->variants_cached
,
3559 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3562 gallivm_destroy(variant
->gallivm
);
3564 /* remove from shader's list */
3565 remove_from_list(&variant
->list_item_local
);
3566 variant
->shader
->variants_cached
--;
3568 /* remove from context's list */
3569 remove_from_list(&variant
->list_item_global
);
3570 lp
->nr_fs_variants
--;
3571 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3578 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3580 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3581 struct lp_fragment_shader
*shader
= fs
;
3582 struct lp_fs_variant_list_item
*li
;
3584 assert(fs
!= llvmpipe
->fs
);
3587 * XXX: we need to flush the context until we have some sort of reference
3588 * counting in fragment shaders as they may still be binned
3589 * Flushing alone might not sufficient we need to wait on it too.
3591 llvmpipe_finish(pipe
, __FUNCTION__
);
3593 /* Delete all the variants */
3594 li
= first_elem(&shader
->variants
);
3595 while(!at_end(&shader
->variants
, li
)) {
3596 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3597 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3601 /* Delete draw module's data */
3602 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3604 if (shader
->base
.ir
.nir
)
3605 ralloc_free(shader
->base
.ir
.nir
);
3606 assert(shader
->variants_cached
== 0);
3607 FREE((void *) shader
->base
.tokens
);
3614 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3615 enum pipe_shader_type shader
, uint index
,
3616 const struct pipe_constant_buffer
*cb
)
3618 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3619 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3621 assert(shader
< PIPE_SHADER_TYPES
);
3622 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3624 /* note: reference counting */
3625 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3628 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3629 debug_printf("Illegal set constant without bind flag\n");
3630 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3634 if (shader
== PIPE_SHADER_VERTEX
||
3635 shader
== PIPE_SHADER_GEOMETRY
||
3636 shader
== PIPE_SHADER_TESS_CTRL
||
3637 shader
== PIPE_SHADER_TESS_EVAL
) {
3638 /* Pass the constants to the 'draw' module */
3639 const unsigned size
= cb
? cb
->buffer_size
: 0;
3643 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3645 else if (cb
&& cb
->user_buffer
) {
3646 data
= (ubyte
*) cb
->user_buffer
;
3653 data
+= cb
->buffer_offset
;
3655 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3658 else if (shader
== PIPE_SHADER_COMPUTE
)
3659 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3661 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3663 if (cb
&& cb
->user_buffer
) {
3664 pipe_resource_reference(&constants
, NULL
);
3669 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3670 enum pipe_shader_type shader
, unsigned start_slot
,
3671 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3672 unsigned writable_bitmask
)
3674 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3676 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3677 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3679 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3681 if (shader
== PIPE_SHADER_VERTEX
||
3682 shader
== PIPE_SHADER_GEOMETRY
||
3683 shader
== PIPE_SHADER_TESS_CTRL
||
3684 shader
== PIPE_SHADER_TESS_EVAL
) {
3685 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3686 const ubyte
*data
= NULL
;
3687 if (buffer
&& buffer
->buffer
)
3688 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3690 data
+= buffer
->buffer_offset
;
3691 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3693 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3694 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3695 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3696 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3702 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3703 enum pipe_shader_type shader
, unsigned start_slot
,
3704 unsigned count
, const struct pipe_image_view
*images
)
3706 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3709 draw_flush(llvmpipe
->draw
);
3710 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3711 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3713 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3716 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3717 if (shader
== PIPE_SHADER_VERTEX
||
3718 shader
== PIPE_SHADER_GEOMETRY
||
3719 shader
== PIPE_SHADER_TESS_CTRL
||
3720 shader
== PIPE_SHADER_TESS_EVAL
) {
3721 draw_set_images(llvmpipe
->draw
,
3723 llvmpipe
->images
[shader
],
3724 start_slot
+ count
);
3725 } else if (shader
== PIPE_SHADER_COMPUTE
)
3726 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3728 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3732 * Return the blend factor equivalent to a destination alpha of one.
3734 static inline unsigned
3735 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3738 case PIPE_BLENDFACTOR_DST_ALPHA
:
3739 return PIPE_BLENDFACTOR_ONE
;
3740 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3741 return PIPE_BLENDFACTOR_ZERO
;
3742 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3744 return PIPE_BLENDFACTOR_ZERO
;
3746 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3754 * We need to generate several variants of the fragment pipeline to match
3755 * all the combinations of the contributing state atoms.
3757 * TODO: there is actually no reason to tie this to context state -- the
3758 * generated code could be cached globally in the screen.
3760 static struct lp_fragment_shader_variant_key
*
3761 make_variant_key(struct llvmpipe_context
*lp
,
3762 struct lp_fragment_shader
*shader
,
3766 struct lp_fragment_shader_variant_key
*key
;
3768 key
= (struct lp_fragment_shader_variant_key
*)store
;
3770 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3772 if (lp
->framebuffer
.zsbuf
) {
3773 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3774 const struct util_format_description
*zsbuf_desc
=
3775 util_format_description(zsbuf_format
);
3777 if (lp
->depth_stencil
->depth
.enabled
&&
3778 util_format_has_depth(zsbuf_desc
)) {
3779 key
->zsbuf_format
= zsbuf_format
;
3780 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3782 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3783 util_format_has_stencil(zsbuf_desc
)) {
3784 key
->zsbuf_format
= zsbuf_format
;
3785 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3787 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3788 key
->resource_1d
= TRUE
;
3790 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3794 * Propagate the depth clamp setting from the rasterizer state.
3795 * depth_clip == 0 implies depth clamping is enabled.
3797 * When clip_halfz is enabled, then always clamp the depth values.
3799 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3800 * clamp is always active in d3d10, regardless if depth clip is
3802 * (GL has an always-on [0,1] clamp on fs depth output instead
3803 * to ensure the depth values stay in range. Doesn't look like
3804 * we do that, though...)
3806 if (lp
->rasterizer
->clip_halfz
) {
3807 key
->depth_clamp
= 1;
3809 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3812 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3813 if (!lp
->framebuffer
.nr_cbufs
||
3814 !lp
->framebuffer
.cbufs
[0] ||
3815 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3816 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3818 if(key
->alpha
.enabled
)
3819 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3820 /* alpha.ref_value is passed in jit_context */
3822 key
->flatshade
= lp
->rasterizer
->flatshade
;
3823 key
->multisample
= lp
->rasterizer
->multisample
;
3824 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3825 key
->occlusion_count
= TRUE
;
3828 if (lp
->framebuffer
.nr_cbufs
) {
3829 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3832 key
->coverage_samples
= 1;
3833 key
->min_samples
= 1;
3834 if (key
->multisample
) {
3835 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3836 key
->min_samples
= lp
->min_samples
== 1 ? 1 : key
->coverage_samples
;
3838 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3840 if (!key
->blend
.independent_blend_enable
) {
3841 /* we always need independent blend otherwise the fixups below won't work */
3842 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3843 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3845 key
->blend
.independent_blend_enable
= 1;
3848 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3849 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3851 if (lp
->framebuffer
.cbufs
[i
]) {
3852 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3853 const struct util_format_description
*format_desc
;
3855 key
->cbuf_format
[i
] = format
;
3856 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3859 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3860 * mixing of 2d textures with height 1 and 1d textures, so make sure
3861 * we pick 1d if any cbuf or zsbuf is 1d.
3863 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3864 key
->resource_1d
= TRUE
;
3867 format_desc
= util_format_description(format
);
3868 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3869 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3872 * Mask out color channels not present in the color buffer.
3874 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3877 * Disable blend for integer formats.
3879 if (util_format_is_pure_integer(format
)) {
3880 blend_rt
->blend_enable
= 0;
3884 * Our swizzled render tiles always have an alpha channel, but the
3885 * linear render target format often does not, so force here the dst
3888 * This is not a mere optimization. Wrong results will be produced if
3889 * the dst alpha is used, the dst format does not have alpha, and the
3890 * previous rendering was not flushed from the swizzled to linear
3891 * buffer. For example, NonPowTwo DCT.
3893 * TODO: This should be generalized to all channels for better
3894 * performance, but only alpha causes correctness issues.
3896 * Also, force rgb/alpha func/factors match, to make AoS blending
3899 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3900 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3901 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3902 boolean clamped_zero
= !util_format_is_float(format
) &&
3903 !util_format_is_snorm(format
);
3904 blend_rt
->rgb_src_factor
=
3905 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3906 blend_rt
->rgb_dst_factor
=
3907 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3908 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3909 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3910 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3914 /* no color buffer for this fragment output */
3915 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3916 key
->cbuf_nr_samples
[i
] = 0;
3917 blend_rt
->colormask
= 0x0;
3918 blend_rt
->blend_enable
= 0;
3922 /* This value will be the same for all the variants of a given shader:
3924 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3926 struct lp_sampler_static_state
*fs_sampler
;
3928 fs_sampler
= key
->samplers
;
3930 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3932 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3933 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3934 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3935 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3940 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3941 * are dx10-style? Can't really have mixed opcodes, at least not
3942 * if we want to skip the holes here (without rescanning tgsi).
3944 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3945 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3946 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3948 * Note sview may exceed what's representable by file_mask.
3949 * This will still work, the only downside is that not actually
3950 * used views may be included in the shader key.
3952 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3953 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3954 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3959 key
->nr_sampler_views
= key
->nr_samplers
;
3960 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3961 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3962 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3963 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3968 struct lp_image_static_state
*lp_image
;
3969 lp_image
= lp_fs_variant_key_images(key
);
3970 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3971 for (i
= 0; i
< key
->nr_images
; ++i
) {
3972 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3973 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3974 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3983 * Update fragment shader state. This is called just prior to drawing
3984 * something when some fragment-related state has changed.
3987 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3989 struct lp_fragment_shader
*shader
= lp
->fs
;
3990 struct lp_fragment_shader_variant_key
*key
;
3991 struct lp_fragment_shader_variant
*variant
= NULL
;
3992 struct lp_fs_variant_list_item
*li
;
3993 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3995 key
= make_variant_key(lp
, shader
, store
);
3997 /* Search the variants for one which matches the key */
3998 li
= first_elem(&shader
->variants
);
3999 while(!at_end(&shader
->variants
, li
)) {
4000 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
4008 /* Move this variant to the head of the list to implement LRU
4009 * deletion of shader's when we have too many.
4011 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
4014 /* variant not found, create it now */
4017 unsigned variants_to_cull
;
4019 if (LP_DEBUG
& DEBUG_FS
) {
4020 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
4023 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
4026 /* First, check if we've exceeded the max number of shader variants.
4027 * If so, free 6.25% of them (the least recently used ones).
4029 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
4031 if (variants_to_cull
||
4032 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
4033 struct pipe_context
*pipe
= &lp
->pipe
;
4035 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
4036 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
4037 "\t%u instrs,\t%u instrs/variant\n",
4038 shader
->variants_cached
,
4039 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
4040 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
4044 * XXX: we need to flush the context until we have some sort of
4045 * reference counting in fragment shaders as they may still be binned
4046 * Flushing alone might not be sufficient we need to wait on it too.
4048 llvmpipe_finish(pipe
, __FUNCTION__
);
4051 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
4052 * number of shader variants (potentially all of them) could be
4053 * pending for destruction on flush.
4056 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
4057 struct lp_fs_variant_list_item
*item
;
4058 if (is_empty_list(&lp
->fs_variants_list
)) {
4061 item
= last_elem(&lp
->fs_variants_list
);
4064 llvmpipe_remove_shader_variant(lp
, item
->base
);
4069 * Generate the new variant.
4072 variant
= generate_variant(lp
, shader
, key
);
4075 LP_COUNT_ADD(llvm_compile_time
, dt
);
4076 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
4078 /* Put the new variant into the list */
4080 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
4081 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
4082 lp
->nr_fs_variants
++;
4083 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
4084 shader
->variants_cached
++;
4088 /* Bind this variant */
4089 lp_setup_set_fs_variant(lp
->setup
, variant
);
4097 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
4099 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
4100 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
4101 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
4103 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
4105 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
4106 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;