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);
407 struct lp_bld_tgsi_system_values system_values
;
409 memset(&system_values
, 0, sizeof(system_values
));
411 /* truncate then sign extend. */
412 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
413 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
415 if (key
->depth
.enabled
||
416 key
->stencil
[0].enabled
) {
418 zs_format_desc
= util_format_description(key
->zsbuf_format
);
419 assert(zs_format_desc
);
421 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
422 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
423 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
424 if (shader
->info
.base
.writes_memory
)
425 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
426 else if (key
->alpha
.enabled
||
427 key
->blend
.alpha_to_coverage
||
428 shader
->info
.base
.uses_kill
||
429 shader
->info
.base
.writes_samplemask
) {
430 /* With alpha test and kill, can do the depth test early
431 * and hopefully eliminate some quads. But need to do a
432 * special deferred depth write once the final mask value
433 * is known. This only works though if there's either no
434 * stencil test or the stencil value isn't written.
436 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
437 (key
->stencil
[1].enabled
&&
438 key
->stencil
[1].writemask
)))
439 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
441 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
444 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
447 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
450 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
451 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
452 (key
->stencil
[1].enabled
&&
453 key
->stencil
[1].writemask
))))
454 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
460 vec_type
= lp_build_vec_type(gallivm
, type
);
461 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
463 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
464 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
465 /* convert scalar stencil refs into vectors */
466 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
467 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
469 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
470 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
472 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
473 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
475 memset(outputs
, 0, sizeof outputs
);
477 /* Allocate color storage for each fragment sample */
478 LLVMValueRef color_store_size
= num_loop
;
479 if (key
->min_samples
> 1)
480 color_store_size
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, key
->min_samples
), "");
482 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
483 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
484 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
485 lp_build_vec_type(gallivm
,
487 color_store_size
, "color");
490 if (dual_source_blend
) {
491 assert(key
->nr_cbufs
<= 1);
492 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
493 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
494 lp_build_vec_type(gallivm
,
496 color_store_size
, "color1");
500 lp_build_for_loop_begin(&loop_state
, gallivm
,
501 lp_build_const_int32(gallivm
, 0),
504 lp_build_const_int32(gallivm
, 1));
506 LLVMValueRef sample_mask_in
;
507 if (key
->multisample
) {
508 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 0);
509 /* create shader execution mask by combining all sample masks. */
510 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
511 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
512 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
513 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
517 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
519 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, s_mask
, lp_build_const_int_vec(gallivm
, type
, (1 << s
)), "");
520 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
523 sample_mask_in
= lp_build_const_int_vec(gallivm
, type
, 1);
524 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
525 &loop_state
.counter
, 1, "mask_ptr");
526 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
528 LLVMValueRef mask_in
= LLVMBuildAnd(builder
, mask_val
, lp_build_const_int_vec(gallivm
, type
, 1), "");
529 sample_mask_in
= LLVMBuildOr(builder
, sample_mask_in
, mask_in
, "");
532 /* 'mask' will control execution based on quad's pixel alive/killed state */
533 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
535 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
536 lp_build_mask_check(&mask
);
538 /* Create storage for recombining sample masks after early Z pass. */
539 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
540 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
542 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
543 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
544 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
545 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
547 /* Run early depth once per sample */
548 if (key
->multisample
) {
550 if (zs_format_desc
) {
551 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
552 struct lp_type z_type
= zs_type
;
553 struct lp_type s_type
= zs_type
;
554 if (zs_format_desc
->block
.bits
< type
.width
)
555 z_type
.width
= type
.width
;
556 else if (zs_format_desc
->block
.bits
> 32) {
557 z_type
.width
= z_type
.width
/ 2;
558 s_type
.width
= s_type
.width
/ 2;
561 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
562 zs_samples
, "z_sample_store");
563 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
564 zs_samples
, "s_sample_store");
565 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
566 zs_samples
, "z_fb_store");
567 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
568 zs_samples
, "s_fb_store");
570 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
571 lp_build_const_int32(gallivm
, 0),
572 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
573 lp_build_const_int32(gallivm
, 1));
575 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
576 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
577 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
579 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
580 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
584 /* for multisample Z needs to be interpolated at sample points for testing. */
585 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
588 depth_ptr
= depth_base_ptr
;
589 if (key
->multisample
) {
590 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
591 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
594 if (depth_mode
& EARLY_DEPTH_TEST
) {
596 * Clamp according to ARB_depth_clamp semantics.
598 if (key
->depth_clamp
) {
599 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
602 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
603 zs_format_desc
, key
->resource_1d
,
604 depth_ptr
, depth_stride
,
605 &z_fb
, &s_fb
, loop_state
.counter
);
606 lp_build_depth_stencil_test(gallivm
,
611 key
->multisample
? NULL
: &mask
,
617 !simple_shader
&& !key
->multisample
);
619 if (depth_mode
& EARLY_DEPTH_WRITE
) {
620 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
621 zs_format_desc
, key
->resource_1d
,
622 NULL
, NULL
, NULL
, loop_state
.counter
,
623 depth_ptr
, depth_stride
,
627 * Note mask check if stencil is enabled must be after ds write not after
628 * stencil test otherwise new stencil values may not get written if all
629 * fragments got killed by depth/stencil test.
631 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
632 lp_build_mask_check(&mask
);
634 if (key
->multisample
) {
635 z_fb_type
= LLVMTypeOf(z_fb
);
636 z_type
= LLVMTypeOf(z_value
);
637 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
638 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
639 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
640 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
644 if (key
->multisample
) {
646 * Store the post-early Z coverage mask.
647 * Recombine the resulting coverage masks post early Z into the fragment
648 * shader execution mask.
650 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
651 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
652 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
654 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
656 lp_build_for_loop_end(&sample_loop_state
);
658 /* recombined all the coverage masks in the shader exec mask. */
659 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
660 lp_build_mask_update(&mask
, tmp_s_mask_or
);
662 if (key
->min_samples
== 1) {
663 /* for multisample Z needs to be re interpolated at pixel center */
664 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
665 lp_build_mask_update(&mask
, tmp_s_mask_or
);
669 LLVMValueRef out_sample_mask_storage
= NULL
;
670 if (shader
->info
.base
.writes_samplemask
) {
671 out_sample_mask_storage
= lp_build_alloca(gallivm
, int_vec_type
, "write_mask");
672 if (key
->min_samples
> 1)
673 LLVMBuildStore(builder
, LLVMConstNull(int_vec_type
), out_sample_mask_storage
);
676 if (key
->multisample
&& key
->min_samples
> 1) {
677 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
678 lp_build_const_int32(gallivm
, 0),
680 lp_build_const_int32(gallivm
, key
->min_samples
),
681 lp_build_const_int32(gallivm
, 1));
683 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
684 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
685 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
686 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
687 lp_build_mask_force(&mask
, s_mask
);
688 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, sample_loop_state
.counter
);
689 system_values
.sample_id
= sample_loop_state
.counter
;
691 system_values
.sample_id
= lp_build_const_int32(gallivm
, 0);
693 system_values
.sample_mask_in
= sample_mask_in
;
694 system_values
.sample_pos
= sample_pos_array
;
696 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, mask_store
, sample_loop_state
.counter
);
698 struct lp_build_fs_llvm_iface fs_iface
= {
699 .base
.interp_fn
= fs_interp
,
701 .loop_state
= &loop_state
,
702 .mask_store
= mask_store
,
705 struct lp_build_tgsi_params params
;
706 memset(¶ms
, 0, sizeof(params
));
710 params
.fs_iface
= &fs_iface
.base
;
711 params
.consts_ptr
= consts_ptr
;
712 params
.const_sizes_ptr
= num_consts_ptr
;
713 params
.system_values
= &system_values
;
714 params
.inputs
= interp
->inputs
;
715 params
.context_ptr
= context_ptr
;
716 params
.thread_data_ptr
= thread_data_ptr
;
717 params
.sampler
= sampler
;
718 params
.info
= &shader
->info
.base
;
719 params
.ssbo_ptr
= ssbo_ptr
;
720 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
721 params
.image
= image
;
723 /* Build the actual shader */
724 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
725 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
728 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
732 if (key
->alpha
.enabled
) {
733 int color0
= find_output_by_semantic(&shader
->info
.base
,
737 if (color0
!= -1 && outputs
[color0
][3]) {
738 const struct util_format_description
*cbuf_format_desc
;
739 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
740 LLVMValueRef alpha_ref_value
;
742 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
743 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
745 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
747 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
748 &mask
, alpha
, alpha_ref_value
,
749 (depth_mode
& LATE_DEPTH_TEST
) != 0);
753 /* Emulate Alpha to Coverage with Alpha test */
754 if (key
->blend
.alpha_to_coverage
) {
755 int color0
= find_output_by_semantic(&shader
->info
.base
,
759 if (color0
!= -1 && outputs
[color0
][3]) {
760 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
762 if (!key
->multisample
) {
763 lp_build_alpha_to_coverage(gallivm
, type
,
765 (depth_mode
& LATE_DEPTH_TEST
) != 0);
767 lp_build_sample_alpha_to_coverage(gallivm
, type
, key
->coverage_samples
, num_loop
,
773 if (key
->blend
.alpha_to_one
&& key
->multisample
) {
774 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
) {
775 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
776 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
777 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
778 if (outputs
[cbuf
][3]) {
779 LLVMBuildStore(builder
, lp_build_const_vec(gallivm
, type
, 1.0), outputs
[cbuf
][3]);
783 if (shader
->info
.base
.writes_samplemask
) {
784 LLVMValueRef output_smask
= NULL
;
785 int smaski
= find_output_by_semantic(&shader
->info
.base
,
786 TGSI_SEMANTIC_SAMPLEMASK
,
788 struct lp_build_context smask_bld
;
789 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
792 output_smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
793 output_smask
= LLVMBuildBitCast(builder
, output_smask
, smask_bld
.vec_type
, "");
795 if (key
->min_samples
> 1) {
796 /* only the bit corresponding to this sample is to be used. */
797 LLVMValueRef tmp_mask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "tmp_mask");
798 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
799 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
), "");
800 output_smask
= LLVMBuildOr(builder
, tmp_mask
, smask_bit
, "");
803 LLVMBuildStore(builder
, output_smask
, out_sample_mask_storage
);
806 /* Color write - per fragment sample */
807 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
809 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
810 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
811 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
813 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
814 if(outputs
[attrib
][chan
]) {
815 /* XXX: just initialize outputs to point at colors[] and
818 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
819 LLVMValueRef color_ptr
;
820 LLVMValueRef color_idx
= loop_state
.counter
;
821 if (key
->min_samples
> 1)
822 color_idx
= LLVMBuildAdd(builder
, color_idx
,
823 LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, ""), "");
824 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
826 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
827 LLVMBuildStore(builder
, out
, color_ptr
);
833 if (key
->multisample
&& key
->min_samples
> 1) {
834 LLVMBuildStore(builder
, lp_build_mask_value(&mask
), s_mask_ptr
);
835 lp_build_for_loop_end(&sample_loop_state
);
838 if (key
->multisample
) {
839 /* execute depth test for each sample */
840 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
841 lp_build_const_int32(gallivm
, 0),
842 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
843 lp_build_const_int32(gallivm
, 1));
845 /* load the per-sample coverage mask */
846 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
847 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
848 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
850 /* combine the execution mask post fragment shader with the coverage mask. */
851 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
852 if (key
->min_samples
== 1)
853 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
855 /* if the shader writes sample mask use that */
856 if (shader
->info
.base
.writes_samplemask
) {
857 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
858 out_smask_idx
= lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
);
859 LLVMValueRef output_smask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "");
860 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, out_smask_idx
, "");
861 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int_vec(gallivm
, int_type
, 0), "");
862 smask_bit
= LLVMBuildSExt(builder
, cmp
, int_vec_type
, "");
864 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
868 depth_ptr
= depth_base_ptr
;
869 if (key
->multisample
) {
870 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
871 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
875 if (depth_mode
& LATE_DEPTH_TEST
) {
876 int pos0
= find_output_by_semantic(&shader
->info
.base
,
877 TGSI_SEMANTIC_POSITION
,
879 int s_out
= find_output_by_semantic(&shader
->info
.base
,
880 TGSI_SEMANTIC_STENCIL
,
882 if (pos0
!= -1 && outputs
[pos0
][2]) {
883 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
886 * Clamp according to ARB_depth_clamp semantics.
888 if (key
->depth_clamp
) {
889 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
893 if (s_out
!= -1 && outputs
[s_out
][1]) {
894 /* there's only one value, and spec says to discard additional bits */
895 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
896 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
897 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
898 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
899 stencil_refs
[1] = stencil_refs
[0];
902 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
903 zs_format_desc
, key
->resource_1d
,
904 depth_ptr
, depth_stride
,
905 &z_fb
, &s_fb
, loop_state
.counter
);
907 lp_build_depth_stencil_test(gallivm
,
912 key
->multisample
? NULL
: &mask
,
920 if (depth_mode
& LATE_DEPTH_WRITE
) {
921 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
922 zs_format_desc
, key
->resource_1d
,
923 NULL
, NULL
, NULL
, loop_state
.counter
,
924 depth_ptr
, depth_stride
,
928 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
929 (depth_mode
& LATE_DEPTH_WRITE
))
931 /* Need to apply a reduced mask to the depth write. Reload the
932 * depth value, update from zs_value with the new mask value and
935 if (key
->multisample
) {
936 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
937 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
938 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
939 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
941 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
942 zs_format_desc
, key
->resource_1d
,
943 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
944 depth_ptr
, depth_stride
,
948 if (key
->occlusion_count
) {
949 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
950 lp_build_name(counter
, "counter");
952 lp_build_occlusion_count(gallivm
, type
,
953 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), counter
);
956 if (key
->multisample
) {
957 /* store the sample mask for this loop */
958 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
959 lp_build_for_loop_end(&sample_loop_state
);
962 mask_val
= lp_build_mask_end(&mask
);
963 if (!key
->multisample
)
964 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
965 lp_build_for_loop_end(&loop_state
);
970 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
972 * Fragment Shader outputs pixels in small 2x2 blocks
973 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
975 * However in memory pixels are stored in rows
976 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
978 * @param type fragment shader type (4x or 8x float)
979 * @param num_fs number of fs_src
980 * @param is_1d whether we're outputting to a 1d resource
981 * @param dst_channels number of output channels
982 * @param fs_src output from fragment shader
983 * @param dst pointer to store result
984 * @param pad_inline is channel padding inline or at end of row
985 * @return the number of dsts
988 generate_fs_twiddle(struct gallivm_state
*gallivm
,
991 unsigned dst_channels
,
992 LLVMValueRef fs_src
[][4],
996 LLVMValueRef src
[16];
1002 unsigned pixels
= type
.length
/ 4;
1003 unsigned reorder_group
;
1004 unsigned src_channels
;
1008 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1009 src_count
= num_fs
* src_channels
;
1011 assert(pixels
== 2 || pixels
== 1);
1012 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
1015 * Transpose from SoA -> AoS
1017 for (i
= 0; i
< num_fs
; ++i
) {
1018 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
1022 * Pick transformation options
1024 swizzle_pad
= false;
1029 if (dst_channels
== 1) {
1035 } else if (dst_channels
== 2) {
1039 } else if (dst_channels
> 2) {
1046 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
1052 * Split the src in half
1055 for (i
= num_fs
; i
> 0; --i
) {
1056 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
1057 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
1065 * Ensure pixels are in memory order
1067 if (reorder_group
) {
1068 /* Twiddle pixels by reordering the array, e.g.:
1070 * src_count = 8 -> 0 2 1 3 4 6 5 7
1071 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1073 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
1075 for (i
= 0; i
< src_count
; ++i
) {
1076 unsigned group
= i
/ reorder_group
;
1077 unsigned block
= (group
/ 4) * 4 * reorder_group
;
1078 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
1081 } else if (twiddle
) {
1082 /* Twiddle pixels across elements of array */
1084 * XXX: we should avoid this in some cases, but would need to tell
1085 * lp_build_conv to reorder (or deal with it ourselves).
1087 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
1090 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
1094 * Moves any padding between pixels to the end
1095 * e.g. RGBXRGBX -> RGBRGBXX
1098 unsigned char swizzles
[16];
1099 unsigned elems
= pixels
* dst_channels
;
1101 for (i
= 0; i
< type
.length
; ++i
) {
1103 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
1105 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
1108 for (i
= 0; i
< src_count
; ++i
) {
1109 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
1118 * Untwiddle and transpose, much like the above.
1119 * However, this is after conversion, so we get packed vectors.
1120 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1121 * the vectors will look like:
1122 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1123 * be swizzled here). Extending to 16bit should be trivial.
1124 * Should also be extended to handle twice wide vectors with AVX2...
1127 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
1128 struct lp_type type
,
1134 struct lp_type type64
, type16
, type32
;
1135 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
1136 LLVMBuilderRef builder
= gallivm
->builder
;
1137 LLVMValueRef tmp
[4], shuf
[8];
1138 for (j
= 0; j
< 2; j
++) {
1139 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
1140 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
1141 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
1142 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
1145 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
1146 assert(type
.width
== 8);
1147 assert(type
.length
== 16);
1149 type8_t
= lp_build_vec_type(gallivm
, type
);
1154 type64_t
= lp_build_vec_type(gallivm
, type64
);
1159 type16_t
= lp_build_vec_type(gallivm
, type16
);
1164 type32_t
= lp_build_vec_type(gallivm
, type32
);
1166 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1168 if (src_count
== 1) {
1169 /* transpose was no-op, just untwiddle */
1170 LLVMValueRef shuf_vec
;
1171 shuf_vec
= LLVMConstVector(shuf
, 8);
1172 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1173 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1174 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1175 } else if (src_count
== 2) {
1176 LLVMValueRef shuf_vec
;
1177 shuf_vec
= LLVMConstVector(shuf
, 4);
1179 for (i
= 0; i
< 2; i
++) {
1180 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1181 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1182 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1185 for (j
= 0; j
< 2; j
++) {
1186 LLVMValueRef lo
, hi
, lo2
, hi2
;
1188 * Note that if we only really have 3 valid channels (rgb)
1189 * and we don't need alpha we could substitute a undef here
1190 * for the respective channel (causing llvm to drop conversion
1193 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1194 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1195 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1196 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1197 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1198 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1199 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1206 * Load an unswizzled block of pixels from memory
1209 load_unswizzled_block(struct gallivm_state
*gallivm
,
1210 LLVMValueRef base_ptr
,
1211 LLVMValueRef stride
,
1212 unsigned block_width
,
1213 unsigned block_height
,
1215 struct lp_type dst_type
,
1217 unsigned dst_alignment
)
1219 LLVMBuilderRef builder
= gallivm
->builder
;
1220 unsigned row_size
= dst_count
/ block_height
;
1223 /* Ensure block exactly fits into dst */
1224 assert((block_width
* block_height
) % dst_count
== 0);
1226 for (i
= 0; i
< dst_count
; ++i
) {
1227 unsigned x
= i
% row_size
;
1228 unsigned y
= i
/ row_size
;
1230 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1231 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1233 LLVMValueRef gep
[2];
1234 LLVMValueRef dst_ptr
;
1236 gep
[0] = lp_build_const_int32(gallivm
, 0);
1237 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1239 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1240 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1241 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1243 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1245 LLVMSetAlignment(dst
[i
], dst_alignment
);
1251 * Store an unswizzled block of pixels to memory
1254 store_unswizzled_block(struct gallivm_state
*gallivm
,
1255 LLVMValueRef base_ptr
,
1256 LLVMValueRef stride
,
1257 unsigned block_width
,
1258 unsigned block_height
,
1260 struct lp_type src_type
,
1262 unsigned src_alignment
)
1264 LLVMBuilderRef builder
= gallivm
->builder
;
1265 unsigned row_size
= src_count
/ block_height
;
1268 /* Ensure src exactly fits into block */
1269 assert((block_width
* block_height
) % src_count
== 0);
1271 for (i
= 0; i
< src_count
; ++i
) {
1272 unsigned x
= i
% row_size
;
1273 unsigned y
= i
/ row_size
;
1275 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1276 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1278 LLVMValueRef gep
[2];
1279 LLVMValueRef src_ptr
;
1281 gep
[0] = lp_build_const_int32(gallivm
, 0);
1282 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1284 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1285 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1286 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1288 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1290 LLVMSetAlignment(src_ptr
, src_alignment
);
1296 * Checks if a format description is an arithmetic format
1298 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1300 static inline boolean
1301 is_arithmetic_format(const struct util_format_description
*format_desc
)
1303 boolean arith
= false;
1306 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1307 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1308 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1316 * Checks if this format requires special handling due to required expansion
1317 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1320 static inline boolean
1321 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1323 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1324 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1332 * Retrieves the type representing the memory layout for a format
1334 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1337 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1338 struct lp_type
* type
)
1343 if (format_expands_to_float_soa(format_desc
)) {
1344 /* just make this a uint with width of block */
1345 type
->floating
= false;
1346 type
->fixed
= false;
1349 type
->width
= format_desc
->block
.bits
;
1354 for (i
= 0; i
< 4; i
++)
1355 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1359 memset(type
, 0, sizeof(struct lp_type
));
1360 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1361 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1362 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1363 type
->norm
= format_desc
->channel
[chan
].normalized
;
1365 if (is_arithmetic_format(format_desc
)) {
1369 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1370 type
->width
+= format_desc
->channel
[i
].size
;
1373 type
->width
= format_desc
->channel
[chan
].size
;
1374 type
->length
= format_desc
->nr_channels
;
1380 * Retrieves the type for a format which is usable in the blending code.
1382 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1385 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1386 struct lp_type
* type
)
1391 if (format_expands_to_float_soa(format_desc
)) {
1392 /* always use ordinary floats for blending */
1393 type
->floating
= true;
1394 type
->fixed
= false;
1402 for (i
= 0; i
< 4; i
++)
1403 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1407 memset(type
, 0, sizeof(struct lp_type
));
1408 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1409 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1410 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1411 type
->norm
= format_desc
->channel
[chan
].normalized
;
1412 type
->width
= format_desc
->channel
[chan
].size
;
1413 type
->length
= format_desc
->nr_channels
;
1415 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1416 if (format_desc
->channel
[i
].size
> type
->width
)
1417 type
->width
= format_desc
->channel
[i
].size
;
1420 if (type
->floating
) {
1423 if (type
->width
<= 8) {
1425 } else if (type
->width
<= 16) {
1432 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1439 * Scale a normalized value from src_bits to dst_bits.
1441 * The exact calculation is
1443 * dst = iround(src * dst_mask / src_mask)
1445 * or with integer rounding
1447 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1451 * src_mask = (1 << src_bits) - 1
1452 * dst_mask = (1 << dst_bits) - 1
1454 * but we try to avoid division and multiplication through shifts.
1456 static inline LLVMValueRef
1457 scale_bits(struct gallivm_state
*gallivm
,
1461 struct lp_type src_type
)
1463 LLVMBuilderRef builder
= gallivm
->builder
;
1464 LLVMValueRef result
= src
;
1466 if (dst_bits
< src_bits
) {
1467 int delta_bits
= src_bits
- dst_bits
;
1469 if (delta_bits
<= dst_bits
) {
1471 * Approximate the rescaling with a single shift.
1473 * This gives the wrong rounding.
1476 result
= LLVMBuildLShr(builder
,
1478 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1483 * Try more accurate rescaling.
1487 * Drop the least significant bits to make space for the multiplication.
1489 * XXX: A better approach would be to use a wider integer type as intermediate. But
1490 * this is enough to convert alpha from 16bits -> 2 when rendering to
1491 * PIPE_FORMAT_R10G10B10A2_UNORM.
1493 result
= LLVMBuildLShr(builder
,
1495 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1499 result
= LLVMBuildMul(builder
,
1501 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1505 * Add a rounding term before the division.
1507 * TODO: Handle signed integers too.
1509 if (!src_type
.sign
) {
1510 result
= LLVMBuildAdd(builder
,
1512 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1517 * Approximate the division by src_mask with a src_bits shift.
1519 * Given the src has already been shifted by dst_bits, all we need
1520 * to do is to shift by the difference.
1523 result
= LLVMBuildLShr(builder
,
1525 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1529 } else if (dst_bits
> src_bits
) {
1531 int db
= dst_bits
- src_bits
;
1533 /* Shift left by difference in bits */
1534 result
= LLVMBuildShl(builder
,
1536 lp_build_const_int_vec(gallivm
, src_type
, db
),
1539 if (db
<= src_bits
) {
1540 /* Enough bits in src to fill the remainder */
1541 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1543 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1546 result
= LLVMBuildOr(builder
, result
, lower
, "");
1547 } else if (db
> src_bits
) {
1548 /* Need to repeatedly copy src bits to fill remainder in dst */
1551 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1552 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1554 result
= LLVMBuildOr(builder
,
1556 LLVMBuildLShr(builder
, result
, shuv
, ""),
1566 * If RT is a smallfloat (needing denorms) format
1569 have_smallfloat_format(struct lp_type dst_type
,
1570 enum pipe_format format
)
1572 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1573 /* due to format handling hacks this format doesn't have floating set
1574 * here (and actually has width set to 32 too) so special case this. */
1575 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1580 * Convert from memory format to blending format
1582 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1585 convert_to_blend_type(struct gallivm_state
*gallivm
,
1586 unsigned block_size
,
1587 const struct util_format_description
*src_fmt
,
1588 struct lp_type src_type
,
1589 struct lp_type dst_type
,
1590 LLVMValueRef
* src
, // and dst
1593 LLVMValueRef
*dst
= src
;
1594 LLVMBuilderRef builder
= gallivm
->builder
;
1595 struct lp_type blend_type
;
1596 struct lp_type mem_type
;
1598 unsigned pixels
= block_size
/ num_srcs
;
1602 * full custom path for packed floats and srgb formats - none of the later
1603 * functions would do anything useful, and given the lp_type representation they
1604 * can't be fixed. Should really have some SoA blend path for these kind of
1605 * formats rather than hacking them in here.
1607 if (format_expands_to_float_soa(src_fmt
)) {
1608 LLVMValueRef tmpsrc
[4];
1610 * This is pretty suboptimal for this case blending in SoA would be much
1611 * better, since conversion gets us SoA values so need to convert back.
1613 assert(src_type
.width
== 32 || src_type
.width
== 16);
1614 assert(dst_type
.floating
);
1615 assert(dst_type
.width
== 32);
1616 assert(dst_type
.length
% 4 == 0);
1617 assert(num_srcs
% 4 == 0);
1619 if (src_type
.width
== 16) {
1620 /* expand 4x16bit values to 4x32bit */
1621 struct lp_type type32x4
= src_type
;
1622 LLVMTypeRef ltype32x4
;
1623 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1624 type32x4
.width
= 32;
1625 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1626 for (i
= 0; i
< num_fetch
; i
++) {
1627 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1629 src_type
.width
= 32;
1631 for (i
= 0; i
< 4; i
++) {
1634 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1635 LLVMValueRef tmpsoa
[4];
1636 LLVMValueRef tmps
= tmpsrc
[i
];
1637 if (dst_type
.length
== 8) {
1638 LLVMValueRef shuffles
[8];
1640 /* fetch was 4 values but need 8-wide output values */
1641 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1643 * for 8-wide aos transpose would give us wrong order not matching
1644 * incoming converted fs values and mask. ARGH.
1646 for (j
= 0; j
< 4; j
++) {
1647 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1648 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1650 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1651 LLVMConstVector(shuffles
, 8), "");
1653 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1654 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1657 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1659 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1664 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1665 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1667 /* Is the format arithmetic */
1668 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1669 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1671 /* Pad if necessary */
1672 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1673 for (i
= 0; i
< num_srcs
; ++i
) {
1674 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1677 src_type
.length
= dst_type
.length
;
1680 /* Special case for half-floats */
1681 if (mem_type
.width
== 16 && mem_type
.floating
) {
1682 assert(blend_type
.width
== 32 && blend_type
.floating
);
1683 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1691 src_type
.width
= blend_type
.width
* blend_type
.length
;
1692 blend_type
.length
*= pixels
;
1693 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1695 for (i
= 0; i
< num_srcs
; ++i
) {
1696 LLVMValueRef chans
[4];
1697 LLVMValueRef res
= NULL
;
1699 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1701 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1703 unsigned sa
= src_fmt
->channel
[j
].shift
;
1704 #if UTIL_ARCH_LITTLE_ENDIAN
1705 unsigned from_lsb
= j
;
1707 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1710 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1712 /* Extract bits from source */
1713 chans
[j
] = LLVMBuildLShr(builder
,
1715 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1718 chans
[j
] = LLVMBuildAnd(builder
,
1720 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1724 if (src_type
.norm
) {
1725 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1726 blend_type
.width
, chans
[j
], src_type
);
1729 /* Insert bits into correct position */
1730 chans
[j
] = LLVMBuildShl(builder
,
1732 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1738 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1742 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1748 * Convert from blending format to memory format
1750 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1753 convert_from_blend_type(struct gallivm_state
*gallivm
,
1754 unsigned block_size
,
1755 const struct util_format_description
*src_fmt
,
1756 struct lp_type src_type
,
1757 struct lp_type dst_type
,
1758 LLVMValueRef
* src
, // and dst
1761 LLVMValueRef
* dst
= src
;
1763 struct lp_type mem_type
;
1764 struct lp_type blend_type
;
1765 LLVMBuilderRef builder
= gallivm
->builder
;
1766 unsigned pixels
= block_size
/ num_srcs
;
1770 * full custom path for packed floats and srgb formats - none of the later
1771 * functions would do anything useful, and given the lp_type representation they
1772 * can't be fixed. Should really have some SoA blend path for these kind of
1773 * formats rather than hacking them in here.
1775 if (format_expands_to_float_soa(src_fmt
)) {
1777 * This is pretty suboptimal for this case blending in SoA would be much
1778 * better - we need to transpose the AoS values back to SoA values for
1779 * conversion/packing.
1781 assert(src_type
.floating
);
1782 assert(src_type
.width
== 32);
1783 assert(src_type
.length
% 4 == 0);
1784 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1786 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1787 LLVMValueRef tmpsoa
[4], tmpdst
;
1788 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1789 /* really really need SoA here */
1791 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1792 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1795 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1799 if (src_type
.length
== 8) {
1800 LLVMValueRef tmpaos
, shuffles
[8];
1803 * for 8-wide aos transpose has given us wrong order not matching
1804 * output order. HMPF. Also need to split the output values manually.
1806 for (j
= 0; j
< 4; j
++) {
1807 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1808 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1810 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1811 LLVMConstVector(shuffles
, 8), "");
1812 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1813 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1819 if (dst_type
.width
== 16) {
1820 struct lp_type type16x8
= dst_type
;
1821 struct lp_type type32x4
= dst_type
;
1822 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1823 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1824 type16x8
.length
= 8;
1825 type32x4
.width
= 32;
1826 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1827 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1828 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1829 /* We could do vector truncation but it doesn't generate very good code */
1830 for (i
= 0; i
< num_fetch
; i
++) {
1831 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1832 src
[i
], lp_build_zero(gallivm
, type32x4
));
1833 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1834 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1835 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1841 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1842 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1844 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1846 /* Special case for half-floats */
1847 if (mem_type
.width
== 16 && mem_type
.floating
) {
1848 int length
= dst_type
.length
;
1849 assert(blend_type
.width
== 32 && blend_type
.floating
);
1851 dst_type
.length
= src_type
.length
;
1853 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1855 dst_type
.length
= length
;
1859 /* Remove any padding */
1860 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1861 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1863 for (i
= 0; i
< num_srcs
; ++i
) {
1864 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1868 /* No bit arithmetic to do */
1873 src_type
.length
= pixels
;
1874 src_type
.width
= blend_type
.length
* blend_type
.width
;
1875 dst_type
.length
= pixels
;
1877 for (i
= 0; i
< num_srcs
; ++i
) {
1878 LLVMValueRef chans
[4];
1879 LLVMValueRef res
= NULL
;
1881 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1883 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1885 unsigned sa
= src_fmt
->channel
[j
].shift
;
1886 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1887 #if UTIL_ARCH_LITTLE_ENDIAN
1888 unsigned from_lsb
= j
;
1890 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1893 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1895 for (k
= 0; k
< blend_type
.width
; ++k
) {
1900 chans
[j
] = LLVMBuildLShr(builder
,
1902 lp_build_const_int_vec(gallivm
, src_type
,
1903 from_lsb
* blend_type
.width
),
1906 chans
[j
] = LLVMBuildAnd(builder
,
1908 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1911 /* Scale down bits */
1912 if (src_type
.norm
) {
1913 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1914 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1915 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1916 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1917 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1918 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1922 chans
[j
] = LLVMBuildShl(builder
,
1924 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1927 sa
+= src_fmt
->channel
[j
].size
;
1932 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1936 assert (dst_type
.width
!= 24);
1938 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1944 * Convert alpha to same blend type as src
1947 convert_alpha(struct gallivm_state
*gallivm
,
1948 struct lp_type row_type
,
1949 struct lp_type alpha_type
,
1950 const unsigned block_size
,
1951 const unsigned block_height
,
1952 const unsigned src_count
,
1953 const unsigned dst_channels
,
1954 const bool pad_inline
,
1955 LLVMValueRef
* src_alpha
)
1957 LLVMBuilderRef builder
= gallivm
->builder
;
1959 unsigned length
= row_type
.length
;
1960 row_type
.length
= alpha_type
.length
;
1962 /* Twiddle the alpha to match pixels */
1963 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1966 * TODO this should use single lp_build_conv call for
1967 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1969 for (i
= 0; i
< block_height
; ++i
) {
1970 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1973 alpha_type
= row_type
;
1974 row_type
.length
= length
;
1976 /* If only one channel we can only need the single alpha value per pixel */
1977 if (src_count
== 1 && dst_channels
== 1) {
1979 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1981 /* If there are more srcs than rows then we need to split alpha up */
1982 if (src_count
> block_height
) {
1983 for (i
= src_count
; i
> 0; --i
) {
1984 unsigned pixels
= block_size
/ src_count
;
1985 unsigned idx
= i
- 1;
1987 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1988 (idx
* pixels
) % 4, pixels
);
1992 /* If there is a src for each pixel broadcast the alpha across whole row */
1993 if (src_count
== block_size
) {
1994 for (i
= 0; i
< src_count
; ++i
) {
1995 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1996 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1999 unsigned pixels
= block_size
/ src_count
;
2000 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
2001 unsigned alpha_span
= 1;
2002 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
2004 /* Check if we need 2 src_alphas for our shuffles */
2005 if (pixels
> alpha_type
.length
) {
2009 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
2010 for (j
= 0; j
< row_type
.length
; ++j
) {
2011 if (j
< pixels
* channels
) {
2012 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
2014 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
2018 for (i
= 0; i
< src_count
; ++i
) {
2019 unsigned idx1
= i
, idx2
= i
;
2021 if (alpha_span
> 1){
2026 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
2029 LLVMConstVector(shuffles
, row_type
.length
),
2038 * Generates the blend function for unswizzled colour buffers
2039 * Also generates the read & write from colour buffer
2042 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
2044 struct lp_fragment_shader_variant
*variant
,
2045 enum pipe_format out_format
,
2046 unsigned int num_fs
,
2047 struct lp_type fs_type
,
2048 LLVMValueRef
* fs_mask
,
2049 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
2050 LLVMValueRef context_ptr
,
2051 LLVMValueRef color_ptr
,
2052 LLVMValueRef stride
,
2053 unsigned partial_mask
,
2056 const unsigned alpha_channel
= 3;
2057 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
2058 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
2059 const unsigned block_size
= block_width
* block_height
;
2060 const unsigned lp_integer_vector_width
= 128;
2062 LLVMBuilderRef builder
= gallivm
->builder
;
2063 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
2064 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
2065 LLVMValueRef src_alpha
[4 * 4];
2066 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
2067 LLVMValueRef src_mask
[4 * 4];
2068 LLVMValueRef src
[4 * 4];
2069 LLVMValueRef src1
[4 * 4];
2070 LLVMValueRef dst
[4 * 4];
2071 LLVMValueRef blend_color
;
2072 LLVMValueRef blend_alpha
;
2073 LLVMValueRef i32_zero
;
2074 LLVMValueRef check_mask
;
2075 LLVMValueRef undef_src_val
;
2077 struct lp_build_mask_context mask_ctx
;
2078 struct lp_type mask_type
;
2079 struct lp_type blend_type
;
2080 struct lp_type row_type
;
2081 struct lp_type dst_type
;
2082 struct lp_type ls_type
;
2084 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
2085 unsigned vector_width
;
2086 unsigned src_channels
= TGSI_NUM_CHANNELS
;
2087 unsigned dst_channels
;
2092 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
2094 unsigned dst_alignment
;
2096 bool pad_inline
= is_arithmetic_format(out_format_desc
);
2097 bool has_alpha
= false;
2098 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
2099 util_blend_state_is_dual(&variant
->key
.blend
, 0);
2101 const boolean is_1d
= variant
->key
.resource_1d
;
2102 boolean twiddle_after_convert
= FALSE
;
2103 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
2104 LLVMValueRef fpstate
= 0;
2106 /* Get type from output format */
2107 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
2108 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
2111 * Technically this code should go into lp_build_smallfloat_to_float
2112 * and lp_build_float_to_smallfloat but due to the
2113 * http://llvm.org/bugs/show_bug.cgi?id=6393
2114 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2115 * So the ordering is important here and there shouldn't be any
2116 * llvm ir instrunctions in this function before
2117 * this, otherwise half-float format conversions won't work
2118 * (again due to llvm bug #6393).
2120 if (have_smallfloat_format(dst_type
, out_format
)) {
2121 /* We need to make sure that denorms are ok for half float
2123 fpstate
= lp_build_fpstate_get(gallivm
);
2124 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
2127 mask_type
= lp_int32_vec4_type();
2128 mask_type
.length
= fs_type
.length
;
2130 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
2131 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
2134 /* Do not bother executing code when mask is empty.. */
2136 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
2138 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2139 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
2142 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
2143 lp_build_mask_check(&mask_ctx
);
2146 partial_mask
|= !variant
->opaque
;
2147 i32_zero
= lp_build_const_int32(gallivm
, 0);
2149 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2151 row_type
.length
= fs_type
.length
;
2152 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2154 /* Compute correct swizzle and count channels */
2155 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2158 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2159 /* Ensure channel is used */
2160 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2164 /* Ensure not already written to (happens in case with GL_ALPHA) */
2165 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2169 /* Ensure we havn't already found all channels */
2170 if (dst_channels
>= out_format_desc
->nr_channels
) {
2174 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2177 if (i
== alpha_channel
) {
2182 if (format_expands_to_float_soa(out_format_desc
)) {
2184 * the code above can't work for layout_other
2185 * for srgb it would sort of work but we short-circuit swizzles, etc.
2186 * as that is done as part of unpack / pack.
2188 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2194 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2197 /* If 3 channels then pad to include alpha for 4 element transpose */
2198 if (dst_channels
== 3) {
2199 assert (!has_alpha
);
2200 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2201 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2204 if (out_format_desc
->nr_channels
== 4) {
2207 * We use alpha from the color conversion, not separate one.
2208 * We had to include it for transpose, hence it will get converted
2209 * too (albeit when doing transpose after conversion, that would
2210 * no longer be the case necessarily).
2211 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2212 * otherwise we really have padding, not alpha, included.)
2219 * Load shader output
2221 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2222 /* Always load alpha for use in blending */
2225 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2228 alpha
= undef_src_val
;
2231 /* Load each channel */
2232 for (j
= 0; j
< dst_channels
; ++j
) {
2233 assert(swizzle
[j
] < 4);
2235 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2238 fs_src
[i
][j
] = undef_src_val
;
2242 /* If 3 channels then pad to include alpha for 4 element transpose */
2244 * XXX If we include that here maybe could actually use it instead of
2245 * separate alpha for blending?
2246 * (Difficult though we actually convert pad channels, not alpha.)
2248 if (dst_channels
== 3 && !has_alpha
) {
2249 fs_src
[i
][3] = alpha
;
2252 /* We split the row_mask and row_alpha as we want 128bit interleave */
2253 if (fs_type
.length
== 8) {
2254 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2256 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2257 src_channels
, src_channels
);
2259 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2260 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2261 src_channels
, src_channels
);
2263 src_mask
[i
] = fs_mask
[i
];
2264 src_alpha
[i
] = alpha
;
2267 if (dual_source_blend
) {
2268 /* same as above except different src/dst, skip masks and comments... */
2269 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2272 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2275 alpha
= undef_src_val
;
2278 for (j
= 0; j
< dst_channels
; ++j
) {
2279 assert(swizzle
[j
] < 4);
2281 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2284 fs_src1
[i
][j
] = undef_src_val
;
2287 if (dst_channels
== 3 && !has_alpha
) {
2288 fs_src1
[i
][3] = alpha
;
2290 if (fs_type
.length
== 8) {
2291 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2292 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2293 src_channels
, src_channels
);
2295 src1_alpha
[i
] = alpha
;
2300 if (util_format_is_pure_integer(out_format
)) {
2302 * In this case fs_type was really ints or uints disguised as floats,
2305 fs_type
.floating
= 0;
2306 fs_type
.sign
= dst_type
.sign
;
2307 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2308 for (j
= 0; j
< dst_channels
; ++j
) {
2309 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2310 lp_build_vec_type(gallivm
, fs_type
), "");
2312 if (dst_channels
== 3 && !has_alpha
) {
2313 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2314 lp_build_vec_type(gallivm
, fs_type
), "");
2320 * We actually should generally do conversion first (for non-1d cases)
2321 * when the blend format is 8 or 16 bits. The reason is obvious,
2322 * there's 2 or 4 times less vectors to deal with for the interleave...
2323 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2324 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2325 * unpack only with 128bit vectors).
2326 * Note: for 16bit sizes really need matching pack conversion code
2328 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2329 twiddle_after_convert
= TRUE
;
2333 * Pixel twiddle from fragment shader order to memory order
2335 if (!twiddle_after_convert
) {
2336 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2337 dst_channels
, fs_src
, src
, pad_inline
);
2338 if (dual_source_blend
) {
2339 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2340 fs_src1
, src1
, pad_inline
);
2343 src_count
= num_fullblock_fs
* dst_channels
;
2345 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2346 * (AVX) turn out the same later when untwiddling/transpose (albeit
2347 * for true AVX2 path untwiddle needs to be different).
2348 * For now just order by colors first (so we can use unpack later).
2350 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2351 for (i
= 0; i
< dst_channels
; i
++) {
2352 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2353 if (dual_source_blend
) {
2354 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2360 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2361 if (src_count
!= num_fullblock_fs
* src_channels
) {
2362 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2363 row_type
.length
/= ds
;
2364 fs_type
.length
= row_type
.length
;
2367 blend_type
= row_type
;
2368 mask_type
.length
= 4;
2370 /* Convert src to row_type */
2371 if (dual_source_blend
) {
2372 struct lp_type old_row_type
= row_type
;
2373 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2374 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2377 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2380 /* If the rows are not an SSE vector, combine them to become SSE size! */
2381 if ((row_type
.width
* row_type
.length
) % 128) {
2382 unsigned bits
= row_type
.width
* row_type
.length
;
2385 assert(src_count
>= (vector_width
/ bits
));
2387 dst_count
= src_count
/ (vector_width
/ bits
);
2389 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2390 if (dual_source_blend
) {
2391 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2394 row_type
.length
*= combined
;
2395 src_count
/= combined
;
2397 bits
= row_type
.width
* row_type
.length
;
2398 assert(bits
== 128 || bits
== 256);
2401 if (twiddle_after_convert
) {
2402 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2403 if (dual_source_blend
) {
2404 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2409 * Blend Colour conversion
2411 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2412 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2413 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2414 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2415 &i32_zero
, 1, ""), "");
2418 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2420 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2422 * since blending is done with floats, there was no conversion.
2423 * However, the rules according to fixed point renderbuffers still
2424 * apply, that is we must clamp inputs to 0.0/1.0.
2425 * (This would apply to separate alpha conversion too but we currently
2426 * force has_alpha to be true.)
2427 * TODO: should skip this with "fake" blend, since post-blend conversion
2428 * will clamp anyway.
2429 * TODO: could also skip this if fragment color clamping is enabled. We
2430 * don't support it natively so it gets baked into the shader however, so
2431 * can't really tell here.
2433 struct lp_build_context f32_bld
;
2434 assert(row_type
.floating
);
2435 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2436 for (i
= 0; i
< src_count
; i
++) {
2437 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2439 if (dual_source_blend
) {
2440 for (i
= 0; i
< src_count
; i
++) {
2441 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2444 /* probably can't be different than row_type but better safe than sorry... */
2445 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2446 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2450 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2452 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2453 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2455 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2456 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2458 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2459 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2465 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2467 if (src_count
< block_height
) {
2468 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2469 } else if (src_count
> block_height
) {
2470 for (i
= src_count
; i
> 0; --i
) {
2471 unsigned pixels
= block_size
/ src_count
;
2472 unsigned idx
= i
- 1;
2474 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2475 (idx
* pixels
) % 4, pixels
);
2479 assert(mask_type
.width
== 32);
2481 for (i
= 0; i
< src_count
; ++i
) {
2482 unsigned pixels
= block_size
/ src_count
;
2483 unsigned pixel_width
= row_type
.width
* dst_channels
;
2485 if (pixel_width
== 24) {
2486 mask_type
.width
= 8;
2487 mask_type
.length
= vector_width
/ mask_type
.width
;
2489 mask_type
.length
= pixels
;
2490 mask_type
.width
= row_type
.width
* dst_channels
;
2493 * If mask_type width is smaller than 32bit, this doesn't quite
2494 * generate the most efficient code (could use some pack).
2496 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2497 lp_build_int_vec_type(gallivm
, mask_type
), "");
2499 mask_type
.length
*= dst_channels
;
2500 mask_type
.width
/= dst_channels
;
2503 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2504 lp_build_int_vec_type(gallivm
, mask_type
), "");
2505 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2512 struct lp_type alpha_type
= fs_type
;
2513 alpha_type
.length
= 4;
2514 convert_alpha(gallivm
, row_type
, alpha_type
,
2515 block_size
, block_height
,
2516 src_count
, dst_channels
,
2517 pad_inline
, src_alpha
);
2518 if (dual_source_blend
) {
2519 convert_alpha(gallivm
, row_type
, alpha_type
,
2520 block_size
, block_height
,
2521 src_count
, dst_channels
,
2522 pad_inline
, src1_alpha
);
2528 * Load dst from memory
2530 if (src_count
< block_height
) {
2531 dst_count
= block_height
;
2533 dst_count
= src_count
;
2536 dst_type
.length
*= block_size
/ dst_count
;
2538 if (format_expands_to_float_soa(out_format_desc
)) {
2540 * we need multiple values at once for the conversion, so can as well
2541 * load them vectorized here too instead of concatenating later.
2542 * (Still need concatenation later for 8-wide vectors).
2544 dst_count
= block_height
;
2545 dst_type
.length
= block_width
;
2549 * Compute the alignment of the destination pointer in bytes
2550 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2551 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2552 * 1d tex but can't distinguish here) so need to stick with per-pixel
2553 * alignment in this case.
2556 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2559 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2561 /* Force power-of-two alignment by extracting only the least-significant-bit */
2562 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2564 * Resource base and stride pointers are aligned to 16 bytes, so that's
2565 * the maximum alignment we can guarantee
2567 dst_alignment
= MIN2(16, dst_alignment
);
2571 if (dst_count
> src_count
) {
2572 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2573 util_is_power_of_two_or_zero(dst_type
.length
) &&
2574 dst_type
.length
* dst_type
.width
< 128) {
2576 * Never try to load values as 4xi8 which we will then
2577 * concatenate to larger vectors. This gives llvm a real
2578 * headache (the problem is the type legalizer (?) will
2579 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2580 * then the shuffles to concatenate are more or less impossible
2581 * - llvm is easily capable of generating a sequence of 32
2582 * pextrb/pinsrb instructions for that. Albeit it appears to
2583 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2584 * width to avoid the trouble (16bit seems not as bad, llvm
2585 * probably recognizes the load+shuffle as only one shuffle
2586 * is necessary, but we can do just the same anyway).
2588 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2594 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2595 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2596 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2597 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2602 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2603 dst
, ls_type
, dst_count
, dst_alignment
);
2608 * Convert from dst/output format to src/blending format.
2610 * This is necessary as we can only read 1 row from memory at a time,
2611 * so the minimum dst_count will ever be at this point is 4.
2613 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2614 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2615 * on all 16 pixels in that single vector at once.
2617 if (dst_count
> src_count
) {
2618 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2619 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2620 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2621 for (i
= 0; i
< dst_count
; i
++) {
2622 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2626 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2628 if (ls_type
.length
!= dst_type
.length
) {
2629 struct lp_type tmp_type
= dst_type
;
2630 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2631 for (i
= 0; i
< src_count
; i
++) {
2632 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2633 lp_build_vec_type(gallivm
, tmp_type
), "");
2641 /* XXX this is broken for RGB8 formats -
2642 * they get expanded from 12 to 16 elements (to include alpha)
2643 * by convert_to_blend_type then reduced to 15 instead of 12
2644 * by convert_from_blend_type (a simple fix though breaks A8...).
2645 * R16G16B16 also crashes differently however something going wrong
2646 * inside llvm handling npot vector sizes seemingly.
2647 * It seems some cleanup could be done here (like skipping conversion/blend
2650 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2651 row_type
, dst
, src_count
);
2654 * FIXME: Really should get logic ops / masks out of generic blend / row
2655 * format. Logic ops will definitely not work on the blend float format
2656 * used for SRGB here and I think OpenGL expects this to work as expected
2657 * (that is incoming values converted to srgb then logic op applied).
2659 for (i
= 0; i
< src_count
; ++i
) {
2660 dst
[i
] = lp_build_blend_aos(gallivm
,
2661 &variant
->key
.blend
,
2666 has_alpha
? NULL
: src_alpha
[i
],
2668 has_alpha
? NULL
: src1_alpha
[i
],
2670 partial_mask
? src_mask
[i
] : NULL
,
2672 has_alpha
? NULL
: blend_alpha
,
2674 pad_inline
? 4 : dst_channels
);
2677 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2678 row_type
, dst_type
, dst
, src_count
);
2680 /* Split the blend rows back to memory rows */
2681 if (dst_count
> src_count
) {
2682 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2684 if (src_count
== 1) {
2685 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2686 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2688 row_type
.length
/= 2;
2692 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2693 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2694 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2695 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2697 row_type
.length
/= 2;
2702 * Store blend result to memory
2705 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2706 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2709 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2710 dst
, dst_type
, dst_count
, dst_alignment
);
2713 if (have_smallfloat_format(dst_type
, out_format
)) {
2714 lp_build_fpstate_set(gallivm
, fpstate
);
2718 lp_build_mask_end(&mask_ctx
);
2724 * Generate the runtime callable function for the whole fragment pipeline.
2725 * Note that the function which we generate operates on a block of 16
2726 * pixels at at time. The block contains 2x2 quads. Each quad contains
2730 generate_fragment(struct llvmpipe_context
*lp
,
2731 struct lp_fragment_shader
*shader
,
2732 struct lp_fragment_shader_variant
*variant
,
2733 unsigned partial_mask
)
2735 struct gallivm_state
*gallivm
= variant
->gallivm
;
2736 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2737 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2739 struct lp_type fs_type
;
2740 struct lp_type blend_type
;
2741 LLVMTypeRef fs_elem_type
;
2742 LLVMTypeRef blend_vec_type
;
2743 LLVMTypeRef arg_types
[15];
2744 LLVMTypeRef func_type
;
2745 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2746 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2747 LLVMValueRef context_ptr
;
2750 LLVMValueRef a0_ptr
;
2751 LLVMValueRef dadx_ptr
;
2752 LLVMValueRef dady_ptr
;
2753 LLVMValueRef color_ptr_ptr
;
2754 LLVMValueRef stride_ptr
;
2755 LLVMValueRef color_sample_stride_ptr
;
2756 LLVMValueRef depth_ptr
;
2757 LLVMValueRef depth_stride
;
2758 LLVMValueRef depth_sample_stride
;
2759 LLVMValueRef mask_input
;
2760 LLVMValueRef thread_data_ptr
;
2761 LLVMBasicBlockRef block
;
2762 LLVMBuilderRef builder
;
2763 struct lp_build_sampler_soa
*sampler
;
2764 struct lp_build_image_soa
*image
;
2765 struct lp_build_interp_soa_context interp
;
2766 LLVMValueRef fs_mask
[(16 / 4) * LP_MAX_SAMPLES
];
2767 LLVMValueRef fs_out_color
[LP_MAX_SAMPLES
][PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2768 LLVMValueRef function
;
2769 LLVMValueRef facing
;
2774 boolean cbuf0_write_all
;
2775 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2776 util_blend_state_is_dual(&key
->blend
, 0);
2778 assert(lp_native_vector_width
/ 32 >= 4);
2780 /* Adjust color input interpolation according to flatshade state:
2782 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2783 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2784 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2786 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2788 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2792 /* check if writes to cbuf[0] are to be copied to all cbufs */
2794 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2796 /* TODO: actually pick these based on the fs and color buffer
2797 * characteristics. */
2799 memset(&fs_type
, 0, sizeof fs_type
);
2800 fs_type
.floating
= TRUE
; /* floating point values */
2801 fs_type
.sign
= TRUE
; /* values are signed */
2802 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2803 fs_type
.width
= 32; /* 32-bit float */
2804 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2806 memset(&blend_type
, 0, sizeof blend_type
);
2807 blend_type
.floating
= FALSE
; /* values are integers */
2808 blend_type
.sign
= FALSE
; /* values are unsigned */
2809 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2810 blend_type
.width
= 8; /* 8-bit ubyte values */
2811 blend_type
.length
= 16; /* 16 elements per vector */
2814 * Generate the function prototype. Any change here must be reflected in
2815 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2818 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2820 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2822 snprintf(func_name
, sizeof(func_name
), "fs_variant_%s",
2823 partial_mask
? "partial" : "whole");
2825 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2826 arg_types
[1] = int32_type
; /* x */
2827 arg_types
[2] = int32_type
; /* y */
2828 arg_types
[3] = int32_type
; /* facing */
2829 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2830 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2831 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2832 arg_types
[7] = LLVMPointerType(LLVMPointerType(int8_type
, 0), 0); /* color */
2833 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2834 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2835 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2836 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2837 arg_types
[12] = int32_type
; /* depth_stride */
2838 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2839 arg_types
[14] = int32_type
; /* depth sample stride */
2841 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2842 arg_types
, ARRAY_SIZE(arg_types
), 0);
2844 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2845 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2847 variant
->function
[partial_mask
] = function
;
2849 /* XXX: need to propagate noalias down into color param now we are
2850 * passing a pointer-to-pointer?
2852 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2853 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2854 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2856 if (variant
->gallivm
->cache
->data_size
)
2859 context_ptr
= LLVMGetParam(function
, 0);
2860 x
= LLVMGetParam(function
, 1);
2861 y
= LLVMGetParam(function
, 2);
2862 facing
= LLVMGetParam(function
, 3);
2863 a0_ptr
= LLVMGetParam(function
, 4);
2864 dadx_ptr
= LLVMGetParam(function
, 5);
2865 dady_ptr
= LLVMGetParam(function
, 6);
2866 color_ptr_ptr
= LLVMGetParam(function
, 7);
2867 depth_ptr
= LLVMGetParam(function
, 8);
2868 mask_input
= LLVMGetParam(function
, 9);
2869 thread_data_ptr
= LLVMGetParam(function
, 10);
2870 stride_ptr
= LLVMGetParam(function
, 11);
2871 depth_stride
= LLVMGetParam(function
, 12);
2872 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2873 depth_sample_stride
= LLVMGetParam(function
, 14);
2875 lp_build_name(context_ptr
, "context");
2876 lp_build_name(x
, "x");
2877 lp_build_name(y
, "y");
2878 lp_build_name(a0_ptr
, "a0");
2879 lp_build_name(dadx_ptr
, "dadx");
2880 lp_build_name(dady_ptr
, "dady");
2881 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2882 lp_build_name(depth_ptr
, "depth");
2883 lp_build_name(mask_input
, "mask_input");
2884 lp_build_name(thread_data_ptr
, "thread_data");
2885 lp_build_name(stride_ptr
, "stride_ptr");
2886 lp_build_name(depth_stride
, "depth_stride");
2887 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2888 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2894 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2895 builder
= gallivm
->builder
;
2897 LLVMPositionBuilderAtEnd(builder
, block
);
2900 * Must not count ps invocations if there's a null shader.
2901 * (It would be ok to count with null shader if there's d/s tests,
2902 * but only if there's d/s buffers too, which is different
2903 * to implicit rasterization disable which must not depend
2904 * on the d/s buffers.)
2905 * Could use popcount on mask, but pixel accuracy is not required.
2906 * Could disable if there's no stats query, but maybe not worth it.
2908 if (shader
->info
.base
.num_instructions
> 1) {
2909 LLVMValueRef invocs
, val
;
2910 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2911 val
= LLVMBuildLoad(builder
, invocs
, "");
2912 val
= LLVMBuildAdd(builder
, val
,
2913 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2915 LLVMBuildStore(builder
, val
, invocs
);
2918 /* code generated texture sampling */
2919 sampler
= lp_llvm_sampler_soa_create(key
->samplers
, key
->nr_samplers
);
2920 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
), key
->nr_images
);
2922 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2923 /* for 1d resources only run "upper half" of stamp */
2924 if (key
->resource_1d
)
2928 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2929 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2930 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2931 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2932 num_loop_samp
, "mask_store");
2934 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(gallivm
->context
);
2935 LLVMValueRef glob_sample_pos
= LLVMAddGlobal(gallivm
->module
, LLVMArrayType(flt_type
, key
->coverage_samples
* 2), "");
2936 LLVMValueRef sample_pos_array
;
2938 if (key
->multisample
&& key
->coverage_samples
== 4) {
2939 LLVMValueRef sample_pos_arr
[8];
2940 for (unsigned i
= 0; i
< 4; i
++) {
2941 sample_pos_arr
[i
* 2] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][0]);
2942 sample_pos_arr
[i
* 2 + 1] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][1]);
2944 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 8);
2946 LLVMValueRef sample_pos_arr
[2];
2947 sample_pos_arr
[0] = LLVMConstReal(flt_type
, 0.5);
2948 sample_pos_arr
[1] = LLVMConstReal(flt_type
, 0.5);
2949 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 2);
2951 LLVMSetInitializer(glob_sample_pos
, sample_pos_array
);
2953 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2954 boolean pixel_center_integer
=
2955 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2958 * The shader input interpolation info is not explicitely baked in the
2959 * shader key, but everything it derives from (TGSI, and flatshade) is
2960 * already included in the shader key.
2962 lp_build_interp_soa_init(&interp
,
2964 shader
->info
.base
.num_inputs
,
2966 pixel_center_integer
,
2967 key
->coverage_samples
, glob_sample_pos
,
2971 a0_ptr
, dadx_ptr
, dady_ptr
,
2974 for (i
= 0; i
< num_fs
; i
++) {
2975 if (key
->multisample
) {
2976 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2979 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2980 * store to the per sample mask storage. Or all of them together to generate
2981 * the fragment shader mask. (sample shading TODO).
2982 * Take the incoming state coverage mask into account.
2984 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2985 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2986 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2987 &sindexi
, 1, "sample_mask_ptr");
2988 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2989 i
*fs_type
.length
/4, s
, mask_input
);
2991 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2992 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2993 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2994 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2996 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2997 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
3001 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
3002 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
3003 &indexi
, 1, "mask_ptr");
3006 mask
= generate_quad_mask(gallivm
, fs_type
,
3007 i
*fs_type
.length
/4, 0, mask_input
);
3010 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
3012 LLVMBuildStore(builder
, mask
, mask_ptr
);
3016 generate_fs_loop(gallivm
,
3026 mask_store
, /* output */
3030 depth_sample_stride
,
3034 for (i
= 0; i
< num_fs
; i
++) {
3036 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
3037 int idx
= (i
+ (s
* num_fs
));
3038 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3039 ptr
= LLVMBuildGEP(builder
, mask_store
, &sindexi
, 1, "");
3041 fs_mask
[idx
] = LLVMBuildLoad(builder
, ptr
, "smask");
3044 for (unsigned s
= 0; s
< key
->min_samples
; s
++) {
3045 /* This is fucked up need to reorganize things */
3046 int idx
= s
* num_fs
+ i
;
3047 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3048 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3049 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3050 ptr
= LLVMBuildGEP(builder
,
3051 color_store
[cbuf
* !cbuf0_write_all
][chan
],
3053 fs_out_color
[s
][cbuf
][chan
][i
] = ptr
;
3056 if (dual_source_blend
) {
3057 /* only support one dual source blend target hence always use output 1 */
3058 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3059 ptr
= LLVMBuildGEP(builder
,
3060 color_store
[1][chan
],
3062 fs_out_color
[s
][1][chan
][i
] = ptr
;
3069 sampler
->destroy(sampler
);
3070 image
->destroy(image
);
3071 /* Loop over color outputs / color buffers to do blending.
3073 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3074 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
3075 LLVMValueRef color_ptr
;
3076 LLVMValueRef stride
;
3077 LLVMValueRef sample_stride
= NULL
;
3078 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
3080 boolean do_branch
= ((key
->depth
.enabled
3081 || key
->stencil
[0].enabled
3082 || key
->alpha
.enabled
)
3083 && !shader
->info
.base
.uses_kill
);
3085 color_ptr
= LLVMBuildLoad(builder
,
3086 LLVMBuildGEP(builder
, color_ptr_ptr
,
3090 stride
= LLVMBuildLoad(builder
,
3091 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
3094 if (key
->multisample
)
3095 sample_stride
= LLVMBuildLoad(builder
,
3096 LLVMBuildGEP(builder
, color_sample_stride_ptr
,
3097 &index
, 1, ""), "");
3099 for (unsigned s
= 0; s
< key
->cbuf_nr_samples
[cbuf
]; s
++) {
3100 unsigned mask_idx
= num_fs
* (key
->multisample
? s
: 0);
3101 unsigned out_idx
= key
->min_samples
== 1 ? 0 : s
;
3102 LLVMValueRef out_ptr
= color_ptr
;;
3104 if (key
->multisample
) {
3105 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_stride
, lp_build_const_int32(gallivm
, s
), "");
3106 out_ptr
= LLVMBuildGEP(builder
, out_ptr
, &sample_offset
, 1, "");
3108 out_ptr
= LLVMBuildBitCast(builder
, out_ptr
, LLVMPointerType(blend_vec_type
, 0), "");
3110 lp_build_name(out_ptr
, "color_ptr%d", cbuf
);
3112 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
3113 key
->cbuf_format
[cbuf
],
3114 num_fs
, fs_type
, &fs_mask
[mask_idx
], fs_out_color
[out_idx
],
3115 context_ptr
, out_ptr
, stride
,
3116 partial_mask
, do_branch
);
3121 LLVMBuildRetVoid(builder
);
3123 gallivm_verify_function(gallivm
, function
);
3128 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
3132 debug_printf("fs variant %p:\n", (void *) key
);
3134 if (key
->flatshade
) {
3135 debug_printf("flatshade = 1\n");
3137 if (key
->multisample
) {
3138 debug_printf("multisample = 1\n");
3139 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
3140 debug_printf("min samples = %d\n", key
->min_samples
);
3142 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
3143 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
3144 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
3146 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
3147 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
3148 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
3150 if (key
->depth
.enabled
) {
3151 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
3152 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
3155 for (i
= 0; i
< 2; ++i
) {
3156 if (key
->stencil
[i
].enabled
) {
3157 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
3158 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
3159 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
3160 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
3161 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
3162 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
3166 if (key
->alpha
.enabled
) {
3167 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
3170 if (key
->occlusion_count
) {
3171 debug_printf("occlusion_count = 1\n");
3174 if (key
->blend
.logicop_enable
) {
3175 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
3177 else if (key
->blend
.rt
[0].blend_enable
) {
3178 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
3179 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
3180 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
3181 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
3182 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
3183 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
3185 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
3186 if (key
->blend
.alpha_to_coverage
) {
3187 debug_printf("blend.alpha_to_coverage is enabled\n");
3189 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
3190 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
3191 debug_printf("sampler[%u] = \n", i
);
3192 debug_printf(" .wrap = %s %s %s\n",
3193 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
3194 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
3195 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
3196 debug_printf(" .min_img_filter = %s\n",
3197 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
3198 debug_printf(" .min_mip_filter = %s\n",
3199 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
3200 debug_printf(" .mag_img_filter = %s\n",
3201 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
3202 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
3203 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3204 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3205 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3206 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3207 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3208 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3210 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3211 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3212 debug_printf("texture[%u] = \n", i
);
3213 debug_printf(" .format = %s\n",
3214 util_format_name(texture
->format
));
3215 debug_printf(" .target = %s\n",
3216 util_str_tex_target(texture
->target
, TRUE
));
3217 debug_printf(" .level_zero_only = %u\n",
3218 texture
->level_zero_only
);
3219 debug_printf(" .pot = %u %u %u\n",
3221 texture
->pot_height
,
3222 texture
->pot_depth
);
3224 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3225 for (i
= 0; i
< key
->nr_images
; ++i
) {
3226 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3227 debug_printf("image[%u] = \n", i
);
3228 debug_printf(" .format = %s\n",
3229 util_format_name(image
->format
));
3230 debug_printf(" .target = %s\n",
3231 util_str_tex_target(image
->target
, TRUE
));
3232 debug_printf(" .level_zero_only = %u\n",
3233 image
->level_zero_only
);
3234 debug_printf(" .pot = %u %u %u\n",
3243 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3245 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3246 variant
->shader
->no
, variant
->no
);
3247 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3248 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3250 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3251 dump_fs_variant_key(&variant
->key
);
3252 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3257 lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant
*variant
,
3258 unsigned char ir_sha1_cache_key
[20])
3260 struct blob blob
= { 0 };
3265 nir_serialize(&blob
, variant
->shader
->base
.ir
.nir
, true);
3266 ir_binary
= blob
.data
;
3267 ir_size
= blob
.size
;
3269 struct mesa_sha1 ctx
;
3270 _mesa_sha1_init(&ctx
);
3271 _mesa_sha1_update(&ctx
, &variant
->key
, variant
->shader
->variant_key_size
);
3272 _mesa_sha1_update(&ctx
, ir_binary
, ir_size
);
3273 _mesa_sha1_final(&ctx
, ir_sha1_cache_key
);
3279 * Generate a new fragment shader variant from the shader code and
3280 * other state indicated by the key.
3282 static struct lp_fragment_shader_variant
*
3283 generate_variant(struct llvmpipe_context
*lp
,
3284 struct lp_fragment_shader
*shader
,
3285 const struct lp_fragment_shader_variant_key
*key
)
3287 struct llvmpipe_screen
*screen
= llvmpipe_screen(lp
->pipe
.screen
);
3288 struct lp_fragment_shader_variant
*variant
;
3289 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3290 boolean fullcolormask
;
3291 char module_name
[64];
3292 unsigned char ir_sha1_cache_key
[20];
3293 struct lp_cached_code cached
= { 0 };
3294 bool needs_caching
= false;
3295 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3299 memset(variant
, 0, sizeof(*variant
));
3300 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3301 shader
->no
, shader
->variants_created
);
3303 variant
->shader
= shader
;
3304 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3306 if (shader
->base
.ir
.nir
) {
3307 lp_fs_get_ir_cache_key(variant
, ir_sha1_cache_key
);
3309 lp_disk_cache_find_shader(screen
, &cached
, ir_sha1_cache_key
);
3310 if (!cached
.data_size
)
3311 needs_caching
= true;
3313 variant
->gallivm
= gallivm_create(module_name
, lp
->context
, &cached
);
3314 if (!variant
->gallivm
) {
3319 variant
->list_item_global
.base
= variant
;
3320 variant
->list_item_local
.base
= variant
;
3321 variant
->no
= shader
->variants_created
++;
3326 * Determine whether we are touching all channels in the color buffer.
3328 fullcolormask
= FALSE
;
3329 if (key
->nr_cbufs
== 1) {
3330 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3331 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3335 !key
->blend
.logicop_enable
&&
3336 !key
->blend
.rt
[0].blend_enable
&&
3338 !key
->stencil
[0].enabled
&&
3339 !key
->alpha
.enabled
&&
3340 !key
->multisample
&&
3341 !key
->blend
.alpha_to_coverage
&&
3342 !key
->depth
.enabled
&&
3343 !shader
->info
.base
.uses_kill
&&
3344 !shader
->info
.base
.writes_samplemask
3347 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3348 lp_debug_fs_variant(variant
);
3351 lp_jit_init_types(variant
);
3353 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3354 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3356 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3357 if (variant
->opaque
) {
3358 /* Specialized shader, which doesn't need to read the color buffer. */
3359 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3364 * Compile everything
3367 gallivm_compile_module(variant
->gallivm
);
3369 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3371 if (variant
->function
[RAST_EDGE_TEST
]) {
3372 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3373 gallivm_jit_function(variant
->gallivm
,
3374 variant
->function
[RAST_EDGE_TEST
]);
3377 if (variant
->function
[RAST_WHOLE
]) {
3378 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3379 gallivm_jit_function(variant
->gallivm
,
3380 variant
->function
[RAST_WHOLE
]);
3381 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3382 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3385 if (needs_caching
) {
3386 lp_disk_cache_insert_shader(screen
, &cached
, ir_sha1_cache_key
);
3389 gallivm_free_ir(variant
->gallivm
);
3396 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3397 const struct pipe_shader_state
*templ
)
3399 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3400 struct lp_fragment_shader
*shader
;
3402 int nr_sampler_views
;
3406 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3410 shader
->no
= fs_no
++;
3411 make_empty_list(&shader
->variants
);
3413 shader
->base
.type
= templ
->type
;
3414 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3415 /* get/save the summary info for this shader */
3416 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3418 /* we need to keep a local copy of the tokens */
3419 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3421 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3422 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3425 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3426 if (shader
->draw_data
== NULL
) {
3427 FREE((void *) shader
->base
.tokens
);
3432 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3433 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3434 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3435 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3437 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3438 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3439 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3440 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3442 switch (shader
->info
.base
.input_interpolate
[i
]) {
3443 case TGSI_INTERPOLATE_CONSTANT
:
3444 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3446 case TGSI_INTERPOLATE_LINEAR
:
3447 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3449 case TGSI_INTERPOLATE_PERSPECTIVE
:
3450 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3452 case TGSI_INTERPOLATE_COLOR
:
3453 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3460 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3461 case TGSI_SEMANTIC_FACE
:
3462 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3464 case TGSI_SEMANTIC_POSITION
:
3465 /* Position was already emitted above
3467 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3468 shader
->inputs
[i
].src_index
= 0;
3472 /* XXX this is a completely pointless index map... */
3473 shader
->inputs
[i
].src_index
= i
+1;
3476 if (LP_DEBUG
& DEBUG_TGSI
) {
3478 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3479 shader
->no
, (void *) shader
);
3480 tgsi_dump(templ
->tokens
, 0);
3481 debug_printf("usage masks:\n");
3482 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3483 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3484 debug_printf(" IN[%u].%s%s%s%s\n",
3486 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3487 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3488 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3489 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3499 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3501 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3502 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3503 if (llvmpipe
->fs
== lp_fs
)
3506 draw_bind_fragment_shader(llvmpipe
->draw
,
3507 (lp_fs
? lp_fs
->draw_data
: NULL
));
3509 llvmpipe
->fs
= lp_fs
;
3511 llvmpipe
->dirty
|= LP_NEW_FS
;
3516 * Remove shader variant from two lists: the shader's variant list
3517 * and the context's variant list.
3520 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3521 struct lp_fragment_shader_variant
*variant
)
3523 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3524 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3525 "v total cached %u inst %u total inst %u\n",
3526 variant
->shader
->no
, variant
->no
,
3527 variant
->shader
->variants_created
,
3528 variant
->shader
->variants_cached
,
3529 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3532 gallivm_destroy(variant
->gallivm
);
3534 /* remove from shader's list */
3535 remove_from_list(&variant
->list_item_local
);
3536 variant
->shader
->variants_cached
--;
3538 /* remove from context's list */
3539 remove_from_list(&variant
->list_item_global
);
3540 lp
->nr_fs_variants
--;
3541 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3548 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3550 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3551 struct lp_fragment_shader
*shader
= fs
;
3552 struct lp_fs_variant_list_item
*li
;
3554 assert(fs
!= llvmpipe
->fs
);
3557 * XXX: we need to flush the context until we have some sort of reference
3558 * counting in fragment shaders as they may still be binned
3559 * Flushing alone might not sufficient we need to wait on it too.
3561 llvmpipe_finish(pipe
, __FUNCTION__
);
3563 /* Delete all the variants */
3564 li
= first_elem(&shader
->variants
);
3565 while(!at_end(&shader
->variants
, li
)) {
3566 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3567 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3571 /* Delete draw module's data */
3572 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3574 if (shader
->base
.ir
.nir
)
3575 ralloc_free(shader
->base
.ir
.nir
);
3576 assert(shader
->variants_cached
== 0);
3577 FREE((void *) shader
->base
.tokens
);
3584 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3585 enum pipe_shader_type shader
, uint index
,
3586 const struct pipe_constant_buffer
*cb
)
3588 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3589 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3591 assert(shader
< PIPE_SHADER_TYPES
);
3592 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3594 /* note: reference counting */
3595 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3598 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3599 debug_printf("Illegal set constant without bind flag\n");
3600 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3604 if (shader
== PIPE_SHADER_VERTEX
||
3605 shader
== PIPE_SHADER_GEOMETRY
||
3606 shader
== PIPE_SHADER_TESS_CTRL
||
3607 shader
== PIPE_SHADER_TESS_EVAL
) {
3608 /* Pass the constants to the 'draw' module */
3609 const unsigned size
= cb
? cb
->buffer_size
: 0;
3613 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3615 else if (cb
&& cb
->user_buffer
) {
3616 data
= (ubyte
*) cb
->user_buffer
;
3623 data
+= cb
->buffer_offset
;
3625 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3628 else if (shader
== PIPE_SHADER_COMPUTE
)
3629 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3631 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3633 if (cb
&& cb
->user_buffer
) {
3634 pipe_resource_reference(&constants
, NULL
);
3639 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3640 enum pipe_shader_type shader
, unsigned start_slot
,
3641 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3642 unsigned writable_bitmask
)
3644 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3646 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3647 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3649 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3651 if (shader
== PIPE_SHADER_VERTEX
||
3652 shader
== PIPE_SHADER_GEOMETRY
||
3653 shader
== PIPE_SHADER_TESS_CTRL
||
3654 shader
== PIPE_SHADER_TESS_EVAL
) {
3655 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3656 const ubyte
*data
= NULL
;
3657 if (buffer
&& buffer
->buffer
)
3658 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3660 data
+= buffer
->buffer_offset
;
3661 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3663 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3664 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3665 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3666 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3672 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3673 enum pipe_shader_type shader
, unsigned start_slot
,
3674 unsigned count
, const struct pipe_image_view
*images
)
3676 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3679 draw_flush(llvmpipe
->draw
);
3680 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3681 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3683 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3686 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3687 if (shader
== PIPE_SHADER_VERTEX
||
3688 shader
== PIPE_SHADER_GEOMETRY
||
3689 shader
== PIPE_SHADER_TESS_CTRL
||
3690 shader
== PIPE_SHADER_TESS_EVAL
) {
3691 draw_set_images(llvmpipe
->draw
,
3693 llvmpipe
->images
[shader
],
3694 start_slot
+ count
);
3695 } else if (shader
== PIPE_SHADER_COMPUTE
)
3696 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3698 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3702 * Return the blend factor equivalent to a destination alpha of one.
3704 static inline unsigned
3705 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3708 case PIPE_BLENDFACTOR_DST_ALPHA
:
3709 return PIPE_BLENDFACTOR_ONE
;
3710 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3711 return PIPE_BLENDFACTOR_ZERO
;
3712 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3714 return PIPE_BLENDFACTOR_ZERO
;
3716 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3724 * We need to generate several variants of the fragment pipeline to match
3725 * all the combinations of the contributing state atoms.
3727 * TODO: there is actually no reason to tie this to context state -- the
3728 * generated code could be cached globally in the screen.
3730 static struct lp_fragment_shader_variant_key
*
3731 make_variant_key(struct llvmpipe_context
*lp
,
3732 struct lp_fragment_shader
*shader
,
3736 struct lp_fragment_shader_variant_key
*key
;
3738 key
= (struct lp_fragment_shader_variant_key
*)store
;
3740 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3742 if (lp
->framebuffer
.zsbuf
) {
3743 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3744 const struct util_format_description
*zsbuf_desc
=
3745 util_format_description(zsbuf_format
);
3747 if (lp
->depth_stencil
->depth
.enabled
&&
3748 util_format_has_depth(zsbuf_desc
)) {
3749 key
->zsbuf_format
= zsbuf_format
;
3750 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3752 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3753 util_format_has_stencil(zsbuf_desc
)) {
3754 key
->zsbuf_format
= zsbuf_format
;
3755 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3757 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3758 key
->resource_1d
= TRUE
;
3760 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3764 * Propagate the depth clamp setting from the rasterizer state.
3765 * depth_clip == 0 implies depth clamping is enabled.
3767 * When clip_halfz is enabled, then always clamp the depth values.
3769 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3770 * clamp is always active in d3d10, regardless if depth clip is
3772 * (GL has an always-on [0,1] clamp on fs depth output instead
3773 * to ensure the depth values stay in range. Doesn't look like
3774 * we do that, though...)
3776 if (lp
->rasterizer
->clip_halfz
) {
3777 key
->depth_clamp
= 1;
3779 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3782 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3783 if (!lp
->framebuffer
.nr_cbufs
||
3784 !lp
->framebuffer
.cbufs
[0] ||
3785 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3786 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3788 if(key
->alpha
.enabled
)
3789 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3790 /* alpha.ref_value is passed in jit_context */
3792 key
->flatshade
= lp
->rasterizer
->flatshade
;
3793 key
->multisample
= lp
->rasterizer
->multisample
;
3794 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3795 key
->occlusion_count
= TRUE
;
3798 if (lp
->framebuffer
.nr_cbufs
) {
3799 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3802 key
->coverage_samples
= 1;
3803 key
->min_samples
= 1;
3804 if (key
->multisample
) {
3805 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3806 key
->min_samples
= lp
->min_samples
== 1 ? 1 : key
->coverage_samples
;
3808 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3810 if (!key
->blend
.independent_blend_enable
) {
3811 /* we always need independent blend otherwise the fixups below won't work */
3812 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3813 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3815 key
->blend
.independent_blend_enable
= 1;
3818 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3819 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3821 if (lp
->framebuffer
.cbufs
[i
]) {
3822 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3823 const struct util_format_description
*format_desc
;
3825 key
->cbuf_format
[i
] = format
;
3826 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3829 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3830 * mixing of 2d textures with height 1 and 1d textures, so make sure
3831 * we pick 1d if any cbuf or zsbuf is 1d.
3833 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3834 key
->resource_1d
= TRUE
;
3837 format_desc
= util_format_description(format
);
3838 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3839 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3842 * Mask out color channels not present in the color buffer.
3844 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3847 * Disable blend for integer formats.
3849 if (util_format_is_pure_integer(format
)) {
3850 blend_rt
->blend_enable
= 0;
3854 * Our swizzled render tiles always have an alpha channel, but the
3855 * linear render target format often does not, so force here the dst
3858 * This is not a mere optimization. Wrong results will be produced if
3859 * the dst alpha is used, the dst format does not have alpha, and the
3860 * previous rendering was not flushed from the swizzled to linear
3861 * buffer. For example, NonPowTwo DCT.
3863 * TODO: This should be generalized to all channels for better
3864 * performance, but only alpha causes correctness issues.
3866 * Also, force rgb/alpha func/factors match, to make AoS blending
3869 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3870 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3871 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3872 boolean clamped_zero
= !util_format_is_float(format
) &&
3873 !util_format_is_snorm(format
);
3874 blend_rt
->rgb_src_factor
=
3875 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3876 blend_rt
->rgb_dst_factor
=
3877 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3878 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3879 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3880 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3884 /* no color buffer for this fragment output */
3885 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3886 key
->cbuf_nr_samples
[i
] = 0;
3887 blend_rt
->colormask
= 0x0;
3888 blend_rt
->blend_enable
= 0;
3892 /* This value will be the same for all the variants of a given shader:
3894 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3896 struct lp_sampler_static_state
*fs_sampler
;
3898 fs_sampler
= key
->samplers
;
3900 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3902 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3903 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3904 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3905 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3910 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3911 * are dx10-style? Can't really have mixed opcodes, at least not
3912 * if we want to skip the holes here (without rescanning tgsi).
3914 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3915 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3916 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3918 * Note sview may exceed what's representable by file_mask.
3919 * This will still work, the only downside is that not actually
3920 * used views may be included in the shader key.
3922 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3923 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3924 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3929 key
->nr_sampler_views
= key
->nr_samplers
;
3930 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3931 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3932 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3933 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3938 struct lp_image_static_state
*lp_image
;
3939 lp_image
= lp_fs_variant_key_images(key
);
3940 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3941 for (i
= 0; i
< key
->nr_images
; ++i
) {
3942 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3943 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3944 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3953 * Update fragment shader state. This is called just prior to drawing
3954 * something when some fragment-related state has changed.
3957 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3959 struct lp_fragment_shader
*shader
= lp
->fs
;
3960 struct lp_fragment_shader_variant_key
*key
;
3961 struct lp_fragment_shader_variant
*variant
= NULL
;
3962 struct lp_fs_variant_list_item
*li
;
3963 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3965 key
= make_variant_key(lp
, shader
, store
);
3967 /* Search the variants for one which matches the key */
3968 li
= first_elem(&shader
->variants
);
3969 while(!at_end(&shader
->variants
, li
)) {
3970 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3978 /* Move this variant to the head of the list to implement LRU
3979 * deletion of shader's when we have too many.
3981 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3984 /* variant not found, create it now */
3987 unsigned variants_to_cull
;
3989 if (LP_DEBUG
& DEBUG_FS
) {
3990 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3993 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3996 /* First, check if we've exceeded the max number of shader variants.
3997 * If so, free 6.25% of them (the least recently used ones).
3999 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
4001 if (variants_to_cull
||
4002 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
4003 struct pipe_context
*pipe
= &lp
->pipe
;
4005 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
4006 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
4007 "\t%u instrs,\t%u instrs/variant\n",
4008 shader
->variants_cached
,
4009 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
4010 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
4014 * XXX: we need to flush the context until we have some sort of
4015 * reference counting in fragment shaders as they may still be binned
4016 * Flushing alone might not be sufficient we need to wait on it too.
4018 llvmpipe_finish(pipe
, __FUNCTION__
);
4021 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
4022 * number of shader variants (potentially all of them) could be
4023 * pending for destruction on flush.
4026 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
4027 struct lp_fs_variant_list_item
*item
;
4028 if (is_empty_list(&lp
->fs_variants_list
)) {
4031 item
= last_elem(&lp
->fs_variants_list
);
4034 llvmpipe_remove_shader_variant(lp
, item
->base
);
4039 * Generate the new variant.
4042 variant
= generate_variant(lp
, shader
, key
);
4045 LP_COUNT_ADD(llvm_compile_time
, dt
);
4046 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
4048 /* Put the new variant into the list */
4050 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
4051 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
4052 lp
->nr_fs_variants
++;
4053 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
4054 shader
->variants_cached
++;
4058 /* Bind this variant */
4059 lp_setup_set_fs_variant(lp
->setup
, variant
);
4067 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
4069 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
4070 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
4071 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
4073 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
4075 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
4076 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;