1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
37 * - depth/stencil test
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
57 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/format/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/simple_list.h"
69 #include "util/u_dual_blend.h"
70 #include "util/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_nir.h"
84 #include "gallivm/lp_bld_swizzle.h"
85 #include "gallivm/lp_bld_flow.h"
86 #include "gallivm/lp_bld_debug.h"
87 #include "gallivm/lp_bld_arit.h"
88 #include "gallivm/lp_bld_bitarit.h"
89 #include "gallivm/lp_bld_pack.h"
90 #include "gallivm/lp_bld_format.h"
91 #include "gallivm/lp_bld_quad.h"
93 #include "lp_bld_alpha.h"
94 #include "lp_bld_blend.h"
95 #include "lp_bld_depth.h"
96 #include "lp_bld_interp.h"
97 #include "lp_context.h"
100 #include "lp_setup.h"
101 #include "lp_state.h"
102 #include "lp_tex_sample.h"
103 #include "lp_flush.h"
104 #include "lp_state_fs.h"
106 #include "nir/nir_to_tgsi_info.h"
108 /** Fragment shader number (for debugging) */
109 static unsigned fs_no
= 0;
113 * Expand the relevant bits of mask_input to a n*4-dword mask for the
114 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
115 * quad mask vector to 0 or ~0.
116 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
117 * quad arguments with fs length 8.
119 * \param first_quad which quad(s) of the quad group to test, in [0,3]
120 * \param mask_input bitwise mask for the whole 4x4 stamp
123 generate_quad_mask(struct gallivm_state
*gallivm
,
124 struct lp_type fs_type
,
127 LLVMValueRef mask_input
) /* int64 */
129 LLVMBuilderRef builder
= gallivm
->builder
;
130 struct lp_type mask_type
;
131 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
132 LLVMValueRef bits
[16];
133 LLVMValueRef mask
, bits_vec
;
137 * XXX: We'll need a different path for 16 x u8
139 assert(fs_type
.width
== 32);
140 assert(fs_type
.length
<= ARRAY_SIZE(bits
));
141 mask_type
= lp_int_type(fs_type
);
144 * mask_input >>= (quad * 4)
146 switch (first_quad
) {
151 assert(fs_type
.length
== 4);
158 assert(fs_type
.length
== 4);
166 mask_input
= LLVMBuildLShr(builder
, mask_input
, lp_build_const_int64(gallivm
, 16 * sample
), "");
167 mask_input
= LLVMBuildTrunc(builder
, mask_input
,
169 mask_input
= LLVMBuildAnd(builder
, mask_input
, lp_build_const_int32(gallivm
, 0xffff), "");
171 mask_input
= LLVMBuildLShr(builder
,
173 LLVMConstInt(i32t
, shift
, 0),
177 * mask = { mask_input & (1 << i), for i in [0,3] }
179 mask
= lp_build_broadcast(gallivm
,
180 lp_build_vec_type(gallivm
, mask_type
),
183 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
184 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
185 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1ULL << (j
+ 0), 0);
186 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
187 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
188 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (j
+ 5), 0);
190 bits_vec
= LLVMConstVector(bits
, fs_type
.length
);
191 mask
= LLVMBuildAnd(builder
, mask
, bits_vec
, "");
194 * mask = mask == bits ? ~0 : 0
196 mask
= lp_build_compare(gallivm
,
197 mask_type
, PIPE_FUNC_EQUAL
,
204 #define EARLY_DEPTH_TEST 0x1
205 #define LATE_DEPTH_TEST 0x2
206 #define EARLY_DEPTH_WRITE 0x4
207 #define LATE_DEPTH_WRITE 0x8
210 find_output_by_semantic( const struct tgsi_shader_info
*info
,
216 for (i
= 0; i
< info
->num_outputs
; i
++)
217 if (info
->output_semantic_name
[i
] == semantic
&&
218 info
->output_semantic_index
[i
] == index
)
226 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
229 lp_llvm_viewport(LLVMValueRef context_ptr
,
230 struct gallivm_state
*gallivm
,
231 LLVMValueRef viewport_index
)
233 LLVMBuilderRef builder
= gallivm
->builder
;
236 struct lp_type viewport_type
=
237 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
239 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
240 ptr
= LLVMBuildPointerCast(builder
, ptr
,
241 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
243 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
250 lp_build_depth_clamp(struct gallivm_state
*gallivm
,
251 LLVMBuilderRef builder
,
253 LLVMValueRef context_ptr
,
254 LLVMValueRef thread_data_ptr
,
257 LLVMValueRef viewport
, min_depth
, max_depth
;
258 LLVMValueRef viewport_index
;
259 struct lp_build_context f32_bld
;
261 assert(type
.floating
);
262 lp_build_context_init(&f32_bld
, gallivm
, type
);
265 * Assumes clamping of the viewport index will occur in setup/gs. Value
266 * is passed through the rasterization stage via lp_rast_shader_inputs.
268 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
271 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
275 * Load the min and max depth from the lp_jit_context.viewports
276 * array of lp_jit_viewport structures.
278 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
280 /* viewports[viewport_index].min_depth */
281 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
282 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
), "");
283 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
285 /* viewports[viewport_index].max_depth */
286 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
287 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
), "");
288 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
291 * Clamp to the min and max depth values for the given viewport.
293 return lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
297 lp_build_sample_alpha_to_coverage(struct gallivm_state
*gallivm
,
299 unsigned coverage_samples
,
300 LLVMValueRef num_loop
,
301 LLVMValueRef loop_counter
,
302 LLVMValueRef coverage_mask_store
,
305 struct lp_build_context bld
;
306 LLVMBuilderRef builder
= gallivm
->builder
;
307 float step
= 1.0 / coverage_samples
;
309 lp_build_context_init(&bld
, gallivm
, type
);
310 for (unsigned s
= 0; s
< coverage_samples
; s
++) {
311 LLVMValueRef alpha_ref_value
= lp_build_const_vec(gallivm
, type
, step
* s
);
312 LLVMValueRef test
= lp_build_cmp(&bld
, PIPE_FUNC_GREATER
, alpha
, alpha_ref_value
);
314 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, s
), num_loop
, "");
315 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_counter
, "");
316 LLVMValueRef s_mask_ptr
= LLVMBuildGEP(builder
, coverage_mask_store
, &s_mask_idx
, 1, "");
317 LLVMValueRef s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
318 s_mask
= LLVMBuildAnd(builder
, s_mask
, test
, "");
319 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
324 * Generate the fragment shader, depth/stencil test, and alpha tests.
327 generate_fs_loop(struct gallivm_state
*gallivm
,
328 struct lp_fragment_shader
*shader
,
329 const struct lp_fragment_shader_variant_key
*key
,
330 LLVMBuilderRef builder
,
332 LLVMValueRef context_ptr
,
333 LLVMValueRef sample_pos_array
,
334 LLVMValueRef num_loop
,
335 struct lp_build_interp_soa_context
*interp
,
336 const struct lp_build_sampler_soa
*sampler
,
337 const struct lp_build_image_soa
*image
,
338 LLVMValueRef mask_store
,
339 LLVMValueRef (*out_color
)[4],
340 LLVMValueRef depth_base_ptr
,
341 LLVMValueRef depth_stride
,
342 LLVMValueRef depth_sample_stride
,
344 LLVMValueRef thread_data_ptr
)
346 const struct util_format_description
*zs_format_desc
= NULL
;
347 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
348 struct lp_type int_type
= lp_int_type(type
);
349 LLVMTypeRef vec_type
, int_vec_type
;
350 LLVMValueRef mask_ptr
= NULL
, mask_val
= NULL
;
351 LLVMValueRef consts_ptr
, num_consts_ptr
;
352 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
354 LLVMValueRef z_value
, s_value
;
355 LLVMValueRef z_fb
, s_fb
;
356 LLVMValueRef depth_ptr
;
357 LLVMValueRef stencil_refs
[2];
358 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
359 LLVMValueRef zs_samples
= lp_build_const_int32(gallivm
, key
->zsbuf_nr_samples
);
360 struct lp_build_for_loop_state loop_state
, sample_loop_state
;
361 struct lp_build_mask_context mask
;
363 * TODO: figure out if simple_shader optimization is really worthwile to
364 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
365 * code since tests tend to take another codepath than real shaders.
367 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
368 shader
->info
.base
.num_inputs
< 3 &&
369 shader
->info
.base
.num_instructions
< 8) && 0;
370 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
371 util_blend_state_is_dual(&key
->blend
, 0);
377 struct lp_bld_tgsi_system_values system_values
;
379 memset(&system_values
, 0, sizeof(system_values
));
381 /* truncate then sign extend. */
382 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
383 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
385 if (key
->depth
.enabled
||
386 key
->stencil
[0].enabled
) {
388 zs_format_desc
= util_format_description(key
->zsbuf_format
);
389 assert(zs_format_desc
);
391 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
392 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
393 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
394 if (shader
->info
.base
.writes_memory
)
395 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
396 else if (key
->alpha
.enabled
||
397 key
->blend
.alpha_to_coverage
||
398 shader
->info
.base
.uses_kill
||
399 shader
->info
.base
.writes_samplemask
) {
400 /* With alpha test and kill, can do the depth test early
401 * and hopefully eliminate some quads. But need to do a
402 * special deferred depth write once the final mask value
403 * is known. This only works though if there's either no
404 * stencil test or the stencil value isn't written.
406 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
407 (key
->stencil
[1].enabled
&&
408 key
->stencil
[1].writemask
)))
409 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
411 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
414 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
417 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
420 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
421 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
422 (key
->stencil
[1].enabled
&&
423 key
->stencil
[1].writemask
))))
424 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
430 vec_type
= lp_build_vec_type(gallivm
, type
);
431 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
433 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
434 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
435 /* convert scalar stencil refs into vectors */
436 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
437 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
439 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
440 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
442 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
443 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
445 memset(outputs
, 0, sizeof outputs
);
447 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
448 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
449 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
450 lp_build_vec_type(gallivm
,
455 if (dual_source_blend
) {
456 assert(key
->nr_cbufs
<= 1);
457 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
458 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
459 lp_build_vec_type(gallivm
,
465 lp_build_for_loop_begin(&loop_state
, gallivm
,
466 lp_build_const_int32(gallivm
, 0),
469 lp_build_const_int32(gallivm
, 1));
471 if (key
->multisample
) {
472 /* create shader execution mask by combining all sample masks. */
473 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
474 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
475 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
476 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
480 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
483 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
484 &loop_state
.counter
, 1, "mask_ptr");
485 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
488 /* 'mask' will control execution based on quad's pixel alive/killed state */
489 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
491 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
492 lp_build_mask_check(&mask
);
494 /* Create storage for recombining sample masks after early Z pass. */
495 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
496 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
498 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
499 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
500 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
501 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
503 /* Run early depth once per sample */
504 if (key
->multisample
) {
506 if (zs_format_desc
) {
507 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
508 struct lp_type z_type
= zs_type
;
509 struct lp_type s_type
= zs_type
;
510 if (zs_format_desc
->block
.bits
< type
.width
)
511 z_type
.width
= type
.width
;
512 else if (zs_format_desc
->block
.bits
> 32) {
513 z_type
.width
= z_type
.width
/ 2;
514 s_type
.width
= s_type
.width
/ 2;
517 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
518 zs_samples
, "z_sample_store");
519 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
520 zs_samples
, "s_sample_store");
521 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
522 zs_samples
, "z_fb_store");
523 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
524 zs_samples
, "s_fb_store");
526 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
527 lp_build_const_int32(gallivm
, 0),
528 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
529 lp_build_const_int32(gallivm
, 1));
531 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
532 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
533 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
535 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
536 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
540 /* for multisample Z needs to be interpolated at sample points for testing. */
541 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
544 depth_ptr
= depth_base_ptr
;
545 if (key
->multisample
) {
546 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
547 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
550 if (depth_mode
& EARLY_DEPTH_TEST
) {
552 * Clamp according to ARB_depth_clamp semantics.
554 if (key
->depth_clamp
) {
555 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
558 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
559 zs_format_desc
, key
->resource_1d
,
560 depth_ptr
, depth_stride
,
561 &z_fb
, &s_fb
, loop_state
.counter
);
562 lp_build_depth_stencil_test(gallivm
,
567 key
->multisample
? NULL
: &mask
,
575 if (depth_mode
& EARLY_DEPTH_WRITE
) {
576 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
577 zs_format_desc
, key
->resource_1d
,
578 NULL
, NULL
, NULL
, loop_state
.counter
,
579 depth_ptr
, depth_stride
,
583 * Note mask check if stencil is enabled must be after ds write not after
584 * stencil test otherwise new stencil values may not get written if all
585 * fragments got killed by depth/stencil test.
587 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
588 lp_build_mask_check(&mask
);
590 if (key
->multisample
) {
591 z_fb_type
= LLVMTypeOf(z_fb
);
592 z_type
= LLVMTypeOf(z_value
);
593 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
594 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
595 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
596 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
600 if (key
->multisample
) {
602 * Store the post-early Z coverage mask.
603 * Recombine the resulting coverage masks post early Z into the fragment
604 * shader execution mask.
606 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
607 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
608 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
610 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
612 lp_build_for_loop_end(&sample_loop_state
);
614 /* recombined all the coverage masks in the shader exec mask. */
615 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
616 lp_build_mask_update(&mask
, tmp_s_mask_or
);
618 /* for multisample Z needs to be re interpolated at pixel center */
619 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
622 system_values
.sample_pos
= sample_pos_array
;
624 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, NULL
, NULL
);
626 struct lp_build_tgsi_params params
;
627 memset(¶ms
, 0, sizeof(params
));
631 params
.consts_ptr
= consts_ptr
;
632 params
.const_sizes_ptr
= num_consts_ptr
;
633 params
.system_values
= &system_values
;
634 params
.inputs
= interp
->inputs
;
635 params
.context_ptr
= context_ptr
;
636 params
.thread_data_ptr
= thread_data_ptr
;
637 params
.sampler
= sampler
;
638 params
.info
= &shader
->info
.base
;
639 params
.ssbo_ptr
= ssbo_ptr
;
640 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
641 params
.image
= image
;
643 /* Build the actual shader */
644 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
645 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
648 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
652 if (key
->alpha
.enabled
) {
653 int color0
= find_output_by_semantic(&shader
->info
.base
,
657 if (color0
!= -1 && outputs
[color0
][3]) {
658 const struct util_format_description
*cbuf_format_desc
;
659 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
660 LLVMValueRef alpha_ref_value
;
662 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
663 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
665 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
667 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
668 &mask
, alpha
, alpha_ref_value
,
669 (depth_mode
& LATE_DEPTH_TEST
) != 0);
673 /* Emulate Alpha to Coverage with Alpha test */
674 if (key
->blend
.alpha_to_coverage
) {
675 int color0
= find_output_by_semantic(&shader
->info
.base
,
679 if (color0
!= -1 && outputs
[color0
][3]) {
680 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
682 if (!key
->multisample
) {
683 lp_build_alpha_to_coverage(gallivm
, type
,
685 (depth_mode
& LATE_DEPTH_TEST
) != 0);
687 lp_build_sample_alpha_to_coverage(gallivm
, type
, key
->coverage_samples
, num_loop
,
694 if (shader
->info
.base
.writes_samplemask
) {
695 int smaski
= find_output_by_semantic(&shader
->info
.base
,
696 TGSI_SEMANTIC_SAMPLEMASK
,
699 struct lp_build_context smask_bld
;
700 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
703 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
705 * Pixel is alive according to the first sample in the mask.
707 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
708 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
709 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
710 lp_build_mask_update(&mask
, smask
);
713 if (key
->multisample
) {
714 /* execute depth test for each sample */
715 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
716 lp_build_const_int32(gallivm
, 0),
717 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
718 lp_build_const_int32(gallivm
, 1));
720 /* load the per-sample coverage mask */
721 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
722 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
723 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
725 /* combine the execution mask post fragment shader with the coverage mask. */
726 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
727 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
730 depth_ptr
= depth_base_ptr
;
731 if (key
->multisample
) {
732 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
733 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
737 if (depth_mode
& LATE_DEPTH_TEST
) {
738 int pos0
= find_output_by_semantic(&shader
->info
.base
,
739 TGSI_SEMANTIC_POSITION
,
741 int s_out
= find_output_by_semantic(&shader
->info
.base
,
742 TGSI_SEMANTIC_STENCIL
,
744 if (pos0
!= -1 && outputs
[pos0
][2]) {
745 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
748 * Clamp according to ARB_depth_clamp semantics.
750 if (key
->depth_clamp
) {
751 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
755 if (s_out
!= -1 && outputs
[s_out
][1]) {
756 /* there's only one value, and spec says to discard additional bits */
757 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
758 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
759 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
760 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
761 stencil_refs
[1] = stencil_refs
[0];
764 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
765 zs_format_desc
, key
->resource_1d
,
766 depth_ptr
, depth_stride
,
767 &z_fb
, &s_fb
, loop_state
.counter
);
769 lp_build_depth_stencil_test(gallivm
,
774 key
->multisample
? NULL
: &mask
,
782 if (depth_mode
& LATE_DEPTH_WRITE
) {
783 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
784 zs_format_desc
, key
->resource_1d
,
785 NULL
, NULL
, NULL
, loop_state
.counter
,
786 depth_ptr
, depth_stride
,
790 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
791 (depth_mode
& LATE_DEPTH_WRITE
))
793 /* Need to apply a reduced mask to the depth write. Reload the
794 * depth value, update from zs_value with the new mask value and
797 if (key
->multisample
) {
798 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
799 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
800 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
801 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
803 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
804 zs_format_desc
, key
->resource_1d
,
805 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
806 depth_ptr
, depth_stride
,
810 if (key
->multisample
) {
811 /* store the sample mask for this loop */
812 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
813 lp_build_for_loop_end(&sample_loop_state
);
817 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
819 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
820 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
821 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
823 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
824 if(outputs
[attrib
][chan
]) {
825 /* XXX: just initialize outputs to point at colors[] and
828 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
829 LLVMValueRef color_ptr
;
830 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
831 &loop_state
.counter
, 1, "");
832 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
833 LLVMBuildStore(builder
, out
, color_ptr
);
839 if (key
->occlusion_count
) {
840 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
841 lp_build_name(counter
, "counter");
842 lp_build_occlusion_count(gallivm
, type
,
843 lp_build_mask_value(&mask
), counter
);
846 mask_val
= lp_build_mask_end(&mask
);
847 if (!key
->multisample
)
848 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
849 lp_build_for_loop_end(&loop_state
);
854 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
856 * Fragment Shader outputs pixels in small 2x2 blocks
857 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
859 * However in memory pixels are stored in rows
860 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
862 * @param type fragment shader type (4x or 8x float)
863 * @param num_fs number of fs_src
864 * @param is_1d whether we're outputting to a 1d resource
865 * @param dst_channels number of output channels
866 * @param fs_src output from fragment shader
867 * @param dst pointer to store result
868 * @param pad_inline is channel padding inline or at end of row
869 * @return the number of dsts
872 generate_fs_twiddle(struct gallivm_state
*gallivm
,
875 unsigned dst_channels
,
876 LLVMValueRef fs_src
[][4],
880 LLVMValueRef src
[16];
886 unsigned pixels
= type
.length
/ 4;
887 unsigned reorder_group
;
888 unsigned src_channels
;
892 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
893 src_count
= num_fs
* src_channels
;
895 assert(pixels
== 2 || pixels
== 1);
896 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
899 * Transpose from SoA -> AoS
901 for (i
= 0; i
< num_fs
; ++i
) {
902 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
906 * Pick transformation options
913 if (dst_channels
== 1) {
919 } else if (dst_channels
== 2) {
923 } else if (dst_channels
> 2) {
930 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
936 * Split the src in half
939 for (i
= num_fs
; i
> 0; --i
) {
940 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
941 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
949 * Ensure pixels are in memory order
952 /* Twiddle pixels by reordering the array, e.g.:
954 * src_count = 8 -> 0 2 1 3 4 6 5 7
955 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
957 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
959 for (i
= 0; i
< src_count
; ++i
) {
960 unsigned group
= i
/ reorder_group
;
961 unsigned block
= (group
/ 4) * 4 * reorder_group
;
962 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
965 } else if (twiddle
) {
966 /* Twiddle pixels across elements of array */
968 * XXX: we should avoid this in some cases, but would need to tell
969 * lp_build_conv to reorder (or deal with it ourselves).
971 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
974 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
978 * Moves any padding between pixels to the end
979 * e.g. RGBXRGBX -> RGBRGBXX
982 unsigned char swizzles
[16];
983 unsigned elems
= pixels
* dst_channels
;
985 for (i
= 0; i
< type
.length
; ++i
) {
987 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
989 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
992 for (i
= 0; i
< src_count
; ++i
) {
993 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
1002 * Untwiddle and transpose, much like the above.
1003 * However, this is after conversion, so we get packed vectors.
1004 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1005 * the vectors will look like:
1006 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1007 * be swizzled here). Extending to 16bit should be trivial.
1008 * Should also be extended to handle twice wide vectors with AVX2...
1011 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
1012 struct lp_type type
,
1018 struct lp_type type64
, type16
, type32
;
1019 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
1020 LLVMBuilderRef builder
= gallivm
->builder
;
1021 LLVMValueRef tmp
[4], shuf
[8];
1022 for (j
= 0; j
< 2; j
++) {
1023 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
1024 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
1025 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
1026 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
1029 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
1030 assert(type
.width
== 8);
1031 assert(type
.length
== 16);
1033 type8_t
= lp_build_vec_type(gallivm
, type
);
1038 type64_t
= lp_build_vec_type(gallivm
, type64
);
1043 type16_t
= lp_build_vec_type(gallivm
, type16
);
1048 type32_t
= lp_build_vec_type(gallivm
, type32
);
1050 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1052 if (src_count
== 1) {
1053 /* transpose was no-op, just untwiddle */
1054 LLVMValueRef shuf_vec
;
1055 shuf_vec
= LLVMConstVector(shuf
, 8);
1056 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1057 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1058 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1059 } else if (src_count
== 2) {
1060 LLVMValueRef shuf_vec
;
1061 shuf_vec
= LLVMConstVector(shuf
, 4);
1063 for (i
= 0; i
< 2; i
++) {
1064 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1065 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1066 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1069 for (j
= 0; j
< 2; j
++) {
1070 LLVMValueRef lo
, hi
, lo2
, hi2
;
1072 * Note that if we only really have 3 valid channels (rgb)
1073 * and we don't need alpha we could substitute a undef here
1074 * for the respective channel (causing llvm to drop conversion
1077 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1078 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1079 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1080 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1081 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1082 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1083 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1090 * Load an unswizzled block of pixels from memory
1093 load_unswizzled_block(struct gallivm_state
*gallivm
,
1094 LLVMValueRef base_ptr
,
1095 LLVMValueRef stride
,
1096 unsigned block_width
,
1097 unsigned block_height
,
1099 struct lp_type dst_type
,
1101 unsigned dst_alignment
)
1103 LLVMBuilderRef builder
= gallivm
->builder
;
1104 unsigned row_size
= dst_count
/ block_height
;
1107 /* Ensure block exactly fits into dst */
1108 assert((block_width
* block_height
) % dst_count
== 0);
1110 for (i
= 0; i
< dst_count
; ++i
) {
1111 unsigned x
= i
% row_size
;
1112 unsigned y
= i
/ row_size
;
1114 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1115 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1117 LLVMValueRef gep
[2];
1118 LLVMValueRef dst_ptr
;
1120 gep
[0] = lp_build_const_int32(gallivm
, 0);
1121 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1123 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1124 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1125 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1127 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1129 LLVMSetAlignment(dst
[i
], dst_alignment
);
1135 * Store an unswizzled block of pixels to memory
1138 store_unswizzled_block(struct gallivm_state
*gallivm
,
1139 LLVMValueRef base_ptr
,
1140 LLVMValueRef stride
,
1141 unsigned block_width
,
1142 unsigned block_height
,
1144 struct lp_type src_type
,
1146 unsigned src_alignment
)
1148 LLVMBuilderRef builder
= gallivm
->builder
;
1149 unsigned row_size
= src_count
/ block_height
;
1152 /* Ensure src exactly fits into block */
1153 assert((block_width
* block_height
) % src_count
== 0);
1155 for (i
= 0; i
< src_count
; ++i
) {
1156 unsigned x
= i
% row_size
;
1157 unsigned y
= i
/ row_size
;
1159 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1160 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1162 LLVMValueRef gep
[2];
1163 LLVMValueRef src_ptr
;
1165 gep
[0] = lp_build_const_int32(gallivm
, 0);
1166 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1168 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1169 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1170 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1172 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1174 LLVMSetAlignment(src_ptr
, src_alignment
);
1180 * Checks if a format description is an arithmetic format
1182 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1184 static inline boolean
1185 is_arithmetic_format(const struct util_format_description
*format_desc
)
1187 boolean arith
= false;
1190 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1191 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1192 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1200 * Checks if this format requires special handling due to required expansion
1201 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1204 static inline boolean
1205 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1207 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1208 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1216 * Retrieves the type representing the memory layout for a format
1218 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1221 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1222 struct lp_type
* type
)
1227 if (format_expands_to_float_soa(format_desc
)) {
1228 /* just make this a uint with width of block */
1229 type
->floating
= false;
1230 type
->fixed
= false;
1233 type
->width
= format_desc
->block
.bits
;
1238 for (i
= 0; i
< 4; i
++)
1239 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1243 memset(type
, 0, sizeof(struct lp_type
));
1244 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1245 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1246 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1247 type
->norm
= format_desc
->channel
[chan
].normalized
;
1249 if (is_arithmetic_format(format_desc
)) {
1253 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1254 type
->width
+= format_desc
->channel
[i
].size
;
1257 type
->width
= format_desc
->channel
[chan
].size
;
1258 type
->length
= format_desc
->nr_channels
;
1264 * Retrieves the type for a format which is usable in the blending code.
1266 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1269 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1270 struct lp_type
* type
)
1275 if (format_expands_to_float_soa(format_desc
)) {
1276 /* always use ordinary floats for blending */
1277 type
->floating
= true;
1278 type
->fixed
= false;
1286 for (i
= 0; i
< 4; i
++)
1287 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1291 memset(type
, 0, sizeof(struct lp_type
));
1292 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1293 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1294 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1295 type
->norm
= format_desc
->channel
[chan
].normalized
;
1296 type
->width
= format_desc
->channel
[chan
].size
;
1297 type
->length
= format_desc
->nr_channels
;
1299 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1300 if (format_desc
->channel
[i
].size
> type
->width
)
1301 type
->width
= format_desc
->channel
[i
].size
;
1304 if (type
->floating
) {
1307 if (type
->width
<= 8) {
1309 } else if (type
->width
<= 16) {
1316 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1323 * Scale a normalized value from src_bits to dst_bits.
1325 * The exact calculation is
1327 * dst = iround(src * dst_mask / src_mask)
1329 * or with integer rounding
1331 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1335 * src_mask = (1 << src_bits) - 1
1336 * dst_mask = (1 << dst_bits) - 1
1338 * but we try to avoid division and multiplication through shifts.
1340 static inline LLVMValueRef
1341 scale_bits(struct gallivm_state
*gallivm
,
1345 struct lp_type src_type
)
1347 LLVMBuilderRef builder
= gallivm
->builder
;
1348 LLVMValueRef result
= src
;
1350 if (dst_bits
< src_bits
) {
1351 int delta_bits
= src_bits
- dst_bits
;
1353 if (delta_bits
<= dst_bits
) {
1355 * Approximate the rescaling with a single shift.
1357 * This gives the wrong rounding.
1360 result
= LLVMBuildLShr(builder
,
1362 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1367 * Try more accurate rescaling.
1371 * Drop the least significant bits to make space for the multiplication.
1373 * XXX: A better approach would be to use a wider integer type as intermediate. But
1374 * this is enough to convert alpha from 16bits -> 2 when rendering to
1375 * PIPE_FORMAT_R10G10B10A2_UNORM.
1377 result
= LLVMBuildLShr(builder
,
1379 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1383 result
= LLVMBuildMul(builder
,
1385 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1389 * Add a rounding term before the division.
1391 * TODO: Handle signed integers too.
1393 if (!src_type
.sign
) {
1394 result
= LLVMBuildAdd(builder
,
1396 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1401 * Approximate the division by src_mask with a src_bits shift.
1403 * Given the src has already been shifted by dst_bits, all we need
1404 * to do is to shift by the difference.
1407 result
= LLVMBuildLShr(builder
,
1409 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1413 } else if (dst_bits
> src_bits
) {
1415 int db
= dst_bits
- src_bits
;
1417 /* Shift left by difference in bits */
1418 result
= LLVMBuildShl(builder
,
1420 lp_build_const_int_vec(gallivm
, src_type
, db
),
1423 if (db
<= src_bits
) {
1424 /* Enough bits in src to fill the remainder */
1425 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1427 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1430 result
= LLVMBuildOr(builder
, result
, lower
, "");
1431 } else if (db
> src_bits
) {
1432 /* Need to repeatedly copy src bits to fill remainder in dst */
1435 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1436 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1438 result
= LLVMBuildOr(builder
,
1440 LLVMBuildLShr(builder
, result
, shuv
, ""),
1450 * If RT is a smallfloat (needing denorms) format
1453 have_smallfloat_format(struct lp_type dst_type
,
1454 enum pipe_format format
)
1456 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1457 /* due to format handling hacks this format doesn't have floating set
1458 * here (and actually has width set to 32 too) so special case this. */
1459 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1464 * Convert from memory format to blending format
1466 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1469 convert_to_blend_type(struct gallivm_state
*gallivm
,
1470 unsigned block_size
,
1471 const struct util_format_description
*src_fmt
,
1472 struct lp_type src_type
,
1473 struct lp_type dst_type
,
1474 LLVMValueRef
* src
, // and dst
1477 LLVMValueRef
*dst
= src
;
1478 LLVMBuilderRef builder
= gallivm
->builder
;
1479 struct lp_type blend_type
;
1480 struct lp_type mem_type
;
1482 unsigned pixels
= block_size
/ num_srcs
;
1486 * full custom path for packed floats and srgb formats - none of the later
1487 * functions would do anything useful, and given the lp_type representation they
1488 * can't be fixed. Should really have some SoA blend path for these kind of
1489 * formats rather than hacking them in here.
1491 if (format_expands_to_float_soa(src_fmt
)) {
1492 LLVMValueRef tmpsrc
[4];
1494 * This is pretty suboptimal for this case blending in SoA would be much
1495 * better, since conversion gets us SoA values so need to convert back.
1497 assert(src_type
.width
== 32 || src_type
.width
== 16);
1498 assert(dst_type
.floating
);
1499 assert(dst_type
.width
== 32);
1500 assert(dst_type
.length
% 4 == 0);
1501 assert(num_srcs
% 4 == 0);
1503 if (src_type
.width
== 16) {
1504 /* expand 4x16bit values to 4x32bit */
1505 struct lp_type type32x4
= src_type
;
1506 LLVMTypeRef ltype32x4
;
1507 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1508 type32x4
.width
= 32;
1509 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1510 for (i
= 0; i
< num_fetch
; i
++) {
1511 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1513 src_type
.width
= 32;
1515 for (i
= 0; i
< 4; i
++) {
1518 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1519 LLVMValueRef tmpsoa
[4];
1520 LLVMValueRef tmps
= tmpsrc
[i
];
1521 if (dst_type
.length
== 8) {
1522 LLVMValueRef shuffles
[8];
1524 /* fetch was 4 values but need 8-wide output values */
1525 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1527 * for 8-wide aos transpose would give us wrong order not matching
1528 * incoming converted fs values and mask. ARGH.
1530 for (j
= 0; j
< 4; j
++) {
1531 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1532 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1534 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1535 LLVMConstVector(shuffles
, 8), "");
1537 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1538 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1541 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1543 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1548 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1549 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1551 /* Is the format arithmetic */
1552 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1553 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1555 /* Pad if necessary */
1556 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1557 for (i
= 0; i
< num_srcs
; ++i
) {
1558 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1561 src_type
.length
= dst_type
.length
;
1564 /* Special case for half-floats */
1565 if (mem_type
.width
== 16 && mem_type
.floating
) {
1566 assert(blend_type
.width
== 32 && blend_type
.floating
);
1567 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1575 src_type
.width
= blend_type
.width
* blend_type
.length
;
1576 blend_type
.length
*= pixels
;
1577 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1579 for (i
= 0; i
< num_srcs
; ++i
) {
1580 LLVMValueRef chans
[4];
1581 LLVMValueRef res
= NULL
;
1583 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1585 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1587 unsigned sa
= src_fmt
->channel
[j
].shift
;
1588 #if UTIL_ARCH_LITTLE_ENDIAN
1589 unsigned from_lsb
= j
;
1591 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1594 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1596 /* Extract bits from source */
1597 chans
[j
] = LLVMBuildLShr(builder
,
1599 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1602 chans
[j
] = LLVMBuildAnd(builder
,
1604 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1608 if (src_type
.norm
) {
1609 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1610 blend_type
.width
, chans
[j
], src_type
);
1613 /* Insert bits into correct position */
1614 chans
[j
] = LLVMBuildShl(builder
,
1616 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1622 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1626 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1632 * Convert from blending format to memory format
1634 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1637 convert_from_blend_type(struct gallivm_state
*gallivm
,
1638 unsigned block_size
,
1639 const struct util_format_description
*src_fmt
,
1640 struct lp_type src_type
,
1641 struct lp_type dst_type
,
1642 LLVMValueRef
* src
, // and dst
1645 LLVMValueRef
* dst
= src
;
1647 struct lp_type mem_type
;
1648 struct lp_type blend_type
;
1649 LLVMBuilderRef builder
= gallivm
->builder
;
1650 unsigned pixels
= block_size
/ num_srcs
;
1654 * full custom path for packed floats and srgb formats - none of the later
1655 * functions would do anything useful, and given the lp_type representation they
1656 * can't be fixed. Should really have some SoA blend path for these kind of
1657 * formats rather than hacking them in here.
1659 if (format_expands_to_float_soa(src_fmt
)) {
1661 * This is pretty suboptimal for this case blending in SoA would be much
1662 * better - we need to transpose the AoS values back to SoA values for
1663 * conversion/packing.
1665 assert(src_type
.floating
);
1666 assert(src_type
.width
== 32);
1667 assert(src_type
.length
% 4 == 0);
1668 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1670 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1671 LLVMValueRef tmpsoa
[4], tmpdst
;
1672 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1673 /* really really need SoA here */
1675 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1676 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1679 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1683 if (src_type
.length
== 8) {
1684 LLVMValueRef tmpaos
, shuffles
[8];
1687 * for 8-wide aos transpose has given us wrong order not matching
1688 * output order. HMPF. Also need to split the output values manually.
1690 for (j
= 0; j
< 4; j
++) {
1691 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1692 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1694 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1695 LLVMConstVector(shuffles
, 8), "");
1696 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1697 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1703 if (dst_type
.width
== 16) {
1704 struct lp_type type16x8
= dst_type
;
1705 struct lp_type type32x4
= dst_type
;
1706 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1707 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1708 type16x8
.length
= 8;
1709 type32x4
.width
= 32;
1710 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1711 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1712 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1713 /* We could do vector truncation but it doesn't generate very good code */
1714 for (i
= 0; i
< num_fetch
; i
++) {
1715 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1716 src
[i
], lp_build_zero(gallivm
, type32x4
));
1717 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1718 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1719 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1725 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1726 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1728 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1730 /* Special case for half-floats */
1731 if (mem_type
.width
== 16 && mem_type
.floating
) {
1732 int length
= dst_type
.length
;
1733 assert(blend_type
.width
== 32 && blend_type
.floating
);
1735 dst_type
.length
= src_type
.length
;
1737 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1739 dst_type
.length
= length
;
1743 /* Remove any padding */
1744 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1745 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1747 for (i
= 0; i
< num_srcs
; ++i
) {
1748 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1752 /* No bit arithmetic to do */
1757 src_type
.length
= pixels
;
1758 src_type
.width
= blend_type
.length
* blend_type
.width
;
1759 dst_type
.length
= pixels
;
1761 for (i
= 0; i
< num_srcs
; ++i
) {
1762 LLVMValueRef chans
[4];
1763 LLVMValueRef res
= NULL
;
1765 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1767 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1769 unsigned sa
= src_fmt
->channel
[j
].shift
;
1770 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1771 #if UTIL_ARCH_LITTLE_ENDIAN
1772 unsigned from_lsb
= j
;
1774 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1777 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1779 for (k
= 0; k
< blend_type
.width
; ++k
) {
1784 chans
[j
] = LLVMBuildLShr(builder
,
1786 lp_build_const_int_vec(gallivm
, src_type
,
1787 from_lsb
* blend_type
.width
),
1790 chans
[j
] = LLVMBuildAnd(builder
,
1792 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1795 /* Scale down bits */
1796 if (src_type
.norm
) {
1797 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1798 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1799 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1800 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1801 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1802 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1806 chans
[j
] = LLVMBuildShl(builder
,
1808 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1811 sa
+= src_fmt
->channel
[j
].size
;
1816 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1820 assert (dst_type
.width
!= 24);
1822 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1828 * Convert alpha to same blend type as src
1831 convert_alpha(struct gallivm_state
*gallivm
,
1832 struct lp_type row_type
,
1833 struct lp_type alpha_type
,
1834 const unsigned block_size
,
1835 const unsigned block_height
,
1836 const unsigned src_count
,
1837 const unsigned dst_channels
,
1838 const bool pad_inline
,
1839 LLVMValueRef
* src_alpha
)
1841 LLVMBuilderRef builder
= gallivm
->builder
;
1843 unsigned length
= row_type
.length
;
1844 row_type
.length
= alpha_type
.length
;
1846 /* Twiddle the alpha to match pixels */
1847 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1850 * TODO this should use single lp_build_conv call for
1851 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1853 for (i
= 0; i
< block_height
; ++i
) {
1854 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1857 alpha_type
= row_type
;
1858 row_type
.length
= length
;
1860 /* If only one channel we can only need the single alpha value per pixel */
1861 if (src_count
== 1 && dst_channels
== 1) {
1863 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1865 /* If there are more srcs than rows then we need to split alpha up */
1866 if (src_count
> block_height
) {
1867 for (i
= src_count
; i
> 0; --i
) {
1868 unsigned pixels
= block_size
/ src_count
;
1869 unsigned idx
= i
- 1;
1871 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1872 (idx
* pixels
) % 4, pixels
);
1876 /* If there is a src for each pixel broadcast the alpha across whole row */
1877 if (src_count
== block_size
) {
1878 for (i
= 0; i
< src_count
; ++i
) {
1879 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1880 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1883 unsigned pixels
= block_size
/ src_count
;
1884 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1885 unsigned alpha_span
= 1;
1886 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1888 /* Check if we need 2 src_alphas for our shuffles */
1889 if (pixels
> alpha_type
.length
) {
1893 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1894 for (j
= 0; j
< row_type
.length
; ++j
) {
1895 if (j
< pixels
* channels
) {
1896 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1898 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1902 for (i
= 0; i
< src_count
; ++i
) {
1903 unsigned idx1
= i
, idx2
= i
;
1905 if (alpha_span
> 1){
1910 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1913 LLVMConstVector(shuffles
, row_type
.length
),
1922 * Generates the blend function for unswizzled colour buffers
1923 * Also generates the read & write from colour buffer
1926 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1928 struct lp_fragment_shader_variant
*variant
,
1929 enum pipe_format out_format
,
1930 unsigned int num_fs
,
1931 struct lp_type fs_type
,
1932 LLVMValueRef
* fs_mask
,
1933 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1934 LLVMValueRef context_ptr
,
1935 LLVMValueRef color_ptr
,
1936 LLVMValueRef stride
,
1937 unsigned partial_mask
,
1940 const unsigned alpha_channel
= 3;
1941 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1942 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1943 const unsigned block_size
= block_width
* block_height
;
1944 const unsigned lp_integer_vector_width
= 128;
1946 LLVMBuilderRef builder
= gallivm
->builder
;
1947 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1948 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1949 LLVMValueRef src_alpha
[4 * 4];
1950 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1951 LLVMValueRef src_mask
[4 * 4];
1952 LLVMValueRef src
[4 * 4];
1953 LLVMValueRef src1
[4 * 4];
1954 LLVMValueRef dst
[4 * 4];
1955 LLVMValueRef blend_color
;
1956 LLVMValueRef blend_alpha
;
1957 LLVMValueRef i32_zero
;
1958 LLVMValueRef check_mask
;
1959 LLVMValueRef undef_src_val
;
1961 struct lp_build_mask_context mask_ctx
;
1962 struct lp_type mask_type
;
1963 struct lp_type blend_type
;
1964 struct lp_type row_type
;
1965 struct lp_type dst_type
;
1966 struct lp_type ls_type
;
1968 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1969 unsigned vector_width
;
1970 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1971 unsigned dst_channels
;
1976 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1978 unsigned dst_alignment
;
1980 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1981 bool has_alpha
= false;
1982 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1983 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1985 const boolean is_1d
= variant
->key
.resource_1d
;
1986 boolean twiddle_after_convert
= FALSE
;
1987 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1988 LLVMValueRef fpstate
= 0;
1990 /* Get type from output format */
1991 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1992 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1995 * Technically this code should go into lp_build_smallfloat_to_float
1996 * and lp_build_float_to_smallfloat but due to the
1997 * http://llvm.org/bugs/show_bug.cgi?id=6393
1998 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1999 * So the ordering is important here and there shouldn't be any
2000 * llvm ir instrunctions in this function before
2001 * this, otherwise half-float format conversions won't work
2002 * (again due to llvm bug #6393).
2004 if (have_smallfloat_format(dst_type
, out_format
)) {
2005 /* We need to make sure that denorms are ok for half float
2007 fpstate
= lp_build_fpstate_get(gallivm
);
2008 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
2011 mask_type
= lp_int32_vec4_type();
2012 mask_type
.length
= fs_type
.length
;
2014 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
2015 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
2018 /* Do not bother executing code when mask is empty.. */
2020 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
2022 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2023 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
2026 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
2027 lp_build_mask_check(&mask_ctx
);
2030 partial_mask
|= !variant
->opaque
;
2031 i32_zero
= lp_build_const_int32(gallivm
, 0);
2033 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2035 row_type
.length
= fs_type
.length
;
2036 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2038 /* Compute correct swizzle and count channels */
2039 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2042 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2043 /* Ensure channel is used */
2044 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2048 /* Ensure not already written to (happens in case with GL_ALPHA) */
2049 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2053 /* Ensure we havn't already found all channels */
2054 if (dst_channels
>= out_format_desc
->nr_channels
) {
2058 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2061 if (i
== alpha_channel
) {
2066 if (format_expands_to_float_soa(out_format_desc
)) {
2068 * the code above can't work for layout_other
2069 * for srgb it would sort of work but we short-circuit swizzles, etc.
2070 * as that is done as part of unpack / pack.
2072 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2078 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2081 /* If 3 channels then pad to include alpha for 4 element transpose */
2082 if (dst_channels
== 3) {
2083 assert (!has_alpha
);
2084 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2085 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2088 if (out_format_desc
->nr_channels
== 4) {
2091 * We use alpha from the color conversion, not separate one.
2092 * We had to include it for transpose, hence it will get converted
2093 * too (albeit when doing transpose after conversion, that would
2094 * no longer be the case necessarily).
2095 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2096 * otherwise we really have padding, not alpha, included.)
2103 * Load shader output
2105 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2106 /* Always load alpha for use in blending */
2109 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2112 alpha
= undef_src_val
;
2115 /* Load each channel */
2116 for (j
= 0; j
< dst_channels
; ++j
) {
2117 assert(swizzle
[j
] < 4);
2119 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2122 fs_src
[i
][j
] = undef_src_val
;
2126 /* If 3 channels then pad to include alpha for 4 element transpose */
2128 * XXX If we include that here maybe could actually use it instead of
2129 * separate alpha for blending?
2130 * (Difficult though we actually convert pad channels, not alpha.)
2132 if (dst_channels
== 3 && !has_alpha
) {
2133 fs_src
[i
][3] = alpha
;
2136 /* We split the row_mask and row_alpha as we want 128bit interleave */
2137 if (fs_type
.length
== 8) {
2138 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2140 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2141 src_channels
, src_channels
);
2143 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2144 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2145 src_channels
, src_channels
);
2147 src_mask
[i
] = fs_mask
[i
];
2148 src_alpha
[i
] = alpha
;
2151 if (dual_source_blend
) {
2152 /* same as above except different src/dst, skip masks and comments... */
2153 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2156 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2159 alpha
= undef_src_val
;
2162 for (j
= 0; j
< dst_channels
; ++j
) {
2163 assert(swizzle
[j
] < 4);
2165 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2168 fs_src1
[i
][j
] = undef_src_val
;
2171 if (dst_channels
== 3 && !has_alpha
) {
2172 fs_src1
[i
][3] = alpha
;
2174 if (fs_type
.length
== 8) {
2175 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2176 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2177 src_channels
, src_channels
);
2179 src1_alpha
[i
] = alpha
;
2184 if (util_format_is_pure_integer(out_format
)) {
2186 * In this case fs_type was really ints or uints disguised as floats,
2189 fs_type
.floating
= 0;
2190 fs_type
.sign
= dst_type
.sign
;
2191 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2192 for (j
= 0; j
< dst_channels
; ++j
) {
2193 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2194 lp_build_vec_type(gallivm
, fs_type
), "");
2196 if (dst_channels
== 3 && !has_alpha
) {
2197 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2198 lp_build_vec_type(gallivm
, fs_type
), "");
2204 * We actually should generally do conversion first (for non-1d cases)
2205 * when the blend format is 8 or 16 bits. The reason is obvious,
2206 * there's 2 or 4 times less vectors to deal with for the interleave...
2207 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2208 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2209 * unpack only with 128bit vectors).
2210 * Note: for 16bit sizes really need matching pack conversion code
2212 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2213 twiddle_after_convert
= TRUE
;
2217 * Pixel twiddle from fragment shader order to memory order
2219 if (!twiddle_after_convert
) {
2220 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2221 dst_channels
, fs_src
, src
, pad_inline
);
2222 if (dual_source_blend
) {
2223 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2224 fs_src1
, src1
, pad_inline
);
2227 src_count
= num_fullblock_fs
* dst_channels
;
2229 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2230 * (AVX) turn out the same later when untwiddling/transpose (albeit
2231 * for true AVX2 path untwiddle needs to be different).
2232 * For now just order by colors first (so we can use unpack later).
2234 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2235 for (i
= 0; i
< dst_channels
; i
++) {
2236 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2237 if (dual_source_blend
) {
2238 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2244 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2245 if (src_count
!= num_fullblock_fs
* src_channels
) {
2246 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2247 row_type
.length
/= ds
;
2248 fs_type
.length
= row_type
.length
;
2251 blend_type
= row_type
;
2252 mask_type
.length
= 4;
2254 /* Convert src to row_type */
2255 if (dual_source_blend
) {
2256 struct lp_type old_row_type
= row_type
;
2257 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2258 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2261 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2264 /* If the rows are not an SSE vector, combine them to become SSE size! */
2265 if ((row_type
.width
* row_type
.length
) % 128) {
2266 unsigned bits
= row_type
.width
* row_type
.length
;
2269 assert(src_count
>= (vector_width
/ bits
));
2271 dst_count
= src_count
/ (vector_width
/ bits
);
2273 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2274 if (dual_source_blend
) {
2275 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2278 row_type
.length
*= combined
;
2279 src_count
/= combined
;
2281 bits
= row_type
.width
* row_type
.length
;
2282 assert(bits
== 128 || bits
== 256);
2285 if (twiddle_after_convert
) {
2286 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2287 if (dual_source_blend
) {
2288 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2293 * Blend Colour conversion
2295 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2296 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2297 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2298 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2299 &i32_zero
, 1, ""), "");
2302 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2304 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2306 * since blending is done with floats, there was no conversion.
2307 * However, the rules according to fixed point renderbuffers still
2308 * apply, that is we must clamp inputs to 0.0/1.0.
2309 * (This would apply to separate alpha conversion too but we currently
2310 * force has_alpha to be true.)
2311 * TODO: should skip this with "fake" blend, since post-blend conversion
2312 * will clamp anyway.
2313 * TODO: could also skip this if fragment color clamping is enabled. We
2314 * don't support it natively so it gets baked into the shader however, so
2315 * can't really tell here.
2317 struct lp_build_context f32_bld
;
2318 assert(row_type
.floating
);
2319 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2320 for (i
= 0; i
< src_count
; i
++) {
2321 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2323 if (dual_source_blend
) {
2324 for (i
= 0; i
< src_count
; i
++) {
2325 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2328 /* probably can't be different than row_type but better safe than sorry... */
2329 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2330 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2334 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2336 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2337 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2339 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2340 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2342 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2343 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2349 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2351 if (src_count
< block_height
) {
2352 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2353 } else if (src_count
> block_height
) {
2354 for (i
= src_count
; i
> 0; --i
) {
2355 unsigned pixels
= block_size
/ src_count
;
2356 unsigned idx
= i
- 1;
2358 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2359 (idx
* pixels
) % 4, pixels
);
2363 assert(mask_type
.width
== 32);
2365 for (i
= 0; i
< src_count
; ++i
) {
2366 unsigned pixels
= block_size
/ src_count
;
2367 unsigned pixel_width
= row_type
.width
* dst_channels
;
2369 if (pixel_width
== 24) {
2370 mask_type
.width
= 8;
2371 mask_type
.length
= vector_width
/ mask_type
.width
;
2373 mask_type
.length
= pixels
;
2374 mask_type
.width
= row_type
.width
* dst_channels
;
2377 * If mask_type width is smaller than 32bit, this doesn't quite
2378 * generate the most efficient code (could use some pack).
2380 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2381 lp_build_int_vec_type(gallivm
, mask_type
), "");
2383 mask_type
.length
*= dst_channels
;
2384 mask_type
.width
/= dst_channels
;
2387 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2388 lp_build_int_vec_type(gallivm
, mask_type
), "");
2389 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2396 struct lp_type alpha_type
= fs_type
;
2397 alpha_type
.length
= 4;
2398 convert_alpha(gallivm
, row_type
, alpha_type
,
2399 block_size
, block_height
,
2400 src_count
, dst_channels
,
2401 pad_inline
, src_alpha
);
2402 if (dual_source_blend
) {
2403 convert_alpha(gallivm
, row_type
, alpha_type
,
2404 block_size
, block_height
,
2405 src_count
, dst_channels
,
2406 pad_inline
, src1_alpha
);
2412 * Load dst from memory
2414 if (src_count
< block_height
) {
2415 dst_count
= block_height
;
2417 dst_count
= src_count
;
2420 dst_type
.length
*= block_size
/ dst_count
;
2422 if (format_expands_to_float_soa(out_format_desc
)) {
2424 * we need multiple values at once for the conversion, so can as well
2425 * load them vectorized here too instead of concatenating later.
2426 * (Still need concatenation later for 8-wide vectors).
2428 dst_count
= block_height
;
2429 dst_type
.length
= block_width
;
2433 * Compute the alignment of the destination pointer in bytes
2434 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2435 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2436 * 1d tex but can't distinguish here) so need to stick with per-pixel
2437 * alignment in this case.
2440 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2443 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2445 /* Force power-of-two alignment by extracting only the least-significant-bit */
2446 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2448 * Resource base and stride pointers are aligned to 16 bytes, so that's
2449 * the maximum alignment we can guarantee
2451 dst_alignment
= MIN2(16, dst_alignment
);
2455 if (dst_count
> src_count
) {
2456 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2457 util_is_power_of_two_or_zero(dst_type
.length
) &&
2458 dst_type
.length
* dst_type
.width
< 128) {
2460 * Never try to load values as 4xi8 which we will then
2461 * concatenate to larger vectors. This gives llvm a real
2462 * headache (the problem is the type legalizer (?) will
2463 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2464 * then the shuffles to concatenate are more or less impossible
2465 * - llvm is easily capable of generating a sequence of 32
2466 * pextrb/pinsrb instructions for that. Albeit it appears to
2467 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2468 * width to avoid the trouble (16bit seems not as bad, llvm
2469 * probably recognizes the load+shuffle as only one shuffle
2470 * is necessary, but we can do just the same anyway).
2472 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2478 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2479 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2480 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2481 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2486 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2487 dst
, ls_type
, dst_count
, dst_alignment
);
2492 * Convert from dst/output format to src/blending format.
2494 * This is necessary as we can only read 1 row from memory at a time,
2495 * so the minimum dst_count will ever be at this point is 4.
2497 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2498 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2499 * on all 16 pixels in that single vector at once.
2501 if (dst_count
> src_count
) {
2502 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2503 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2504 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2505 for (i
= 0; i
< dst_count
; i
++) {
2506 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2510 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2512 if (ls_type
.length
!= dst_type
.length
) {
2513 struct lp_type tmp_type
= dst_type
;
2514 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2515 for (i
= 0; i
< src_count
; i
++) {
2516 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2517 lp_build_vec_type(gallivm
, tmp_type
), "");
2525 /* XXX this is broken for RGB8 formats -
2526 * they get expanded from 12 to 16 elements (to include alpha)
2527 * by convert_to_blend_type then reduced to 15 instead of 12
2528 * by convert_from_blend_type (a simple fix though breaks A8...).
2529 * R16G16B16 also crashes differently however something going wrong
2530 * inside llvm handling npot vector sizes seemingly.
2531 * It seems some cleanup could be done here (like skipping conversion/blend
2534 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2535 row_type
, dst
, src_count
);
2538 * FIXME: Really should get logic ops / masks out of generic blend / row
2539 * format. Logic ops will definitely not work on the blend float format
2540 * used for SRGB here and I think OpenGL expects this to work as expected
2541 * (that is incoming values converted to srgb then logic op applied).
2543 for (i
= 0; i
< src_count
; ++i
) {
2544 dst
[i
] = lp_build_blend_aos(gallivm
,
2545 &variant
->key
.blend
,
2550 has_alpha
? NULL
: src_alpha
[i
],
2552 has_alpha
? NULL
: src1_alpha
[i
],
2554 partial_mask
? src_mask
[i
] : NULL
,
2556 has_alpha
? NULL
: blend_alpha
,
2558 pad_inline
? 4 : dst_channels
);
2561 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2562 row_type
, dst_type
, dst
, src_count
);
2564 /* Split the blend rows back to memory rows */
2565 if (dst_count
> src_count
) {
2566 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2568 if (src_count
== 1) {
2569 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2570 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2572 row_type
.length
/= 2;
2576 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2577 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2578 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2579 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2581 row_type
.length
/= 2;
2586 * Store blend result to memory
2589 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2590 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2593 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2594 dst
, dst_type
, dst_count
, dst_alignment
);
2597 if (have_smallfloat_format(dst_type
, out_format
)) {
2598 lp_build_fpstate_set(gallivm
, fpstate
);
2602 lp_build_mask_end(&mask_ctx
);
2608 * Generate the runtime callable function for the whole fragment pipeline.
2609 * Note that the function which we generate operates on a block of 16
2610 * pixels at at time. The block contains 2x2 quads. Each quad contains
2614 generate_fragment(struct llvmpipe_context
*lp
,
2615 struct lp_fragment_shader
*shader
,
2616 struct lp_fragment_shader_variant
*variant
,
2617 unsigned partial_mask
)
2619 struct gallivm_state
*gallivm
= variant
->gallivm
;
2620 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2621 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2623 struct lp_type fs_type
;
2624 struct lp_type blend_type
;
2625 LLVMTypeRef fs_elem_type
;
2626 LLVMTypeRef blend_vec_type
;
2627 LLVMTypeRef arg_types
[15];
2628 LLVMTypeRef func_type
;
2629 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2630 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2631 LLVMValueRef context_ptr
;
2634 LLVMValueRef a0_ptr
;
2635 LLVMValueRef dadx_ptr
;
2636 LLVMValueRef dady_ptr
;
2637 LLVMValueRef color_ptr_ptr
;
2638 LLVMValueRef stride_ptr
;
2639 LLVMValueRef color_sample_stride_ptr
;
2640 LLVMValueRef depth_ptr
;
2641 LLVMValueRef depth_stride
;
2642 LLVMValueRef depth_sample_stride
;
2643 LLVMValueRef mask_input
;
2644 LLVMValueRef thread_data_ptr
;
2645 LLVMBasicBlockRef block
;
2646 LLVMBuilderRef builder
;
2647 struct lp_build_sampler_soa
*sampler
;
2648 struct lp_build_image_soa
*image
;
2649 struct lp_build_interp_soa_context interp
;
2650 LLVMValueRef fs_mask
[(16 / 4) * LP_MAX_SAMPLES
];
2651 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2652 LLVMValueRef function
;
2653 LLVMValueRef facing
;
2658 boolean cbuf0_write_all
;
2659 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2660 util_blend_state_is_dual(&key
->blend
, 0);
2662 assert(lp_native_vector_width
/ 32 >= 4);
2664 /* Adjust color input interpolation according to flatshade state:
2666 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2667 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2668 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2670 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2672 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2676 /* check if writes to cbuf[0] are to be copied to all cbufs */
2678 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2680 /* TODO: actually pick these based on the fs and color buffer
2681 * characteristics. */
2683 memset(&fs_type
, 0, sizeof fs_type
);
2684 fs_type
.floating
= TRUE
; /* floating point values */
2685 fs_type
.sign
= TRUE
; /* values are signed */
2686 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2687 fs_type
.width
= 32; /* 32-bit float */
2688 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2690 memset(&blend_type
, 0, sizeof blend_type
);
2691 blend_type
.floating
= FALSE
; /* values are integers */
2692 blend_type
.sign
= FALSE
; /* values are unsigned */
2693 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2694 blend_type
.width
= 8; /* 8-bit ubyte values */
2695 blend_type
.length
= 16; /* 16 elements per vector */
2698 * Generate the function prototype. Any change here must be reflected in
2699 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2702 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2704 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2706 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2707 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2709 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2710 arg_types
[1] = int32_type
; /* x */
2711 arg_types
[2] = int32_type
; /* y */
2712 arg_types
[3] = int32_type
; /* facing */
2713 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2714 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2715 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2716 arg_types
[7] = LLVMPointerType(LLVMPointerType(int8_type
, 0), 0); /* color */
2717 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2718 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2719 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2720 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2721 arg_types
[12] = int32_type
; /* depth_stride */
2722 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2723 arg_types
[14] = int32_type
; /* depth sample stride */
2725 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2726 arg_types
, ARRAY_SIZE(arg_types
), 0);
2728 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2729 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2731 variant
->function
[partial_mask
] = function
;
2733 /* XXX: need to propagate noalias down into color param now we are
2734 * passing a pointer-to-pointer?
2736 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2737 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2738 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2740 context_ptr
= LLVMGetParam(function
, 0);
2741 x
= LLVMGetParam(function
, 1);
2742 y
= LLVMGetParam(function
, 2);
2743 facing
= LLVMGetParam(function
, 3);
2744 a0_ptr
= LLVMGetParam(function
, 4);
2745 dadx_ptr
= LLVMGetParam(function
, 5);
2746 dady_ptr
= LLVMGetParam(function
, 6);
2747 color_ptr_ptr
= LLVMGetParam(function
, 7);
2748 depth_ptr
= LLVMGetParam(function
, 8);
2749 mask_input
= LLVMGetParam(function
, 9);
2750 thread_data_ptr
= LLVMGetParam(function
, 10);
2751 stride_ptr
= LLVMGetParam(function
, 11);
2752 depth_stride
= LLVMGetParam(function
, 12);
2753 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2754 depth_sample_stride
= LLVMGetParam(function
, 14);
2756 lp_build_name(context_ptr
, "context");
2757 lp_build_name(x
, "x");
2758 lp_build_name(y
, "y");
2759 lp_build_name(a0_ptr
, "a0");
2760 lp_build_name(dadx_ptr
, "dadx");
2761 lp_build_name(dady_ptr
, "dady");
2762 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2763 lp_build_name(depth_ptr
, "depth");
2764 lp_build_name(mask_input
, "mask_input");
2765 lp_build_name(thread_data_ptr
, "thread_data");
2766 lp_build_name(stride_ptr
, "stride_ptr");
2767 lp_build_name(depth_stride
, "depth_stride");
2768 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2769 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2775 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2776 builder
= gallivm
->builder
;
2778 LLVMPositionBuilderAtEnd(builder
, block
);
2781 * Must not count ps invocations if there's a null shader.
2782 * (It would be ok to count with null shader if there's d/s tests,
2783 * but only if there's d/s buffers too, which is different
2784 * to implicit rasterization disable which must not depend
2785 * on the d/s buffers.)
2786 * Could use popcount on mask, but pixel accuracy is not required.
2787 * Could disable if there's no stats query, but maybe not worth it.
2789 if (shader
->info
.base
.num_instructions
> 1) {
2790 LLVMValueRef invocs
, val
;
2791 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2792 val
= LLVMBuildLoad(builder
, invocs
, "");
2793 val
= LLVMBuildAdd(builder
, val
,
2794 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2796 LLVMBuildStore(builder
, val
, invocs
);
2799 /* code generated texture sampling */
2800 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2801 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2803 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2804 /* for 1d resources only run "upper half" of stamp */
2805 if (key
->resource_1d
)
2809 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2810 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2811 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2812 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2813 num_loop_samp
, "mask_store");
2815 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(gallivm
->context
);
2816 LLVMValueRef glob_sample_pos
= LLVMAddGlobal(gallivm
->module
, flt_type
, "");
2817 LLVMValueRef sample_pos_array
;
2819 if (key
->multisample
&& key
->coverage_samples
== 4) {
2820 LLVMValueRef sample_pos_arr
[8];
2821 for (unsigned i
= 0; i
< 4; i
++) {
2822 sample_pos_arr
[i
* 2] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][0]);
2823 sample_pos_arr
[i
* 2 + 1] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][1]);
2825 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 8);
2827 LLVMValueRef sample_pos_arr
[2];
2828 sample_pos_arr
[0] = LLVMConstReal(flt_type
, 0.5);
2829 sample_pos_arr
[1] = LLVMConstReal(flt_type
, 0.5);
2830 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 2);
2832 LLVMSetInitializer(glob_sample_pos
, sample_pos_array
);
2834 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2835 boolean pixel_center_integer
=
2836 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2839 * The shader input interpolation info is not explicitely baked in the
2840 * shader key, but everything it derives from (TGSI, and flatshade) is
2841 * already included in the shader key.
2843 lp_build_interp_soa_init(&interp
,
2845 shader
->info
.base
.num_inputs
,
2847 pixel_center_integer
,
2848 key
->coverage_samples
, glob_sample_pos
,
2852 a0_ptr
, dadx_ptr
, dady_ptr
,
2855 for (i
= 0; i
< num_fs
; i
++) {
2856 if (key
->multisample
) {
2857 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2860 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2861 * store to the per sample mask storage. Or all of them together to generate
2862 * the fragment shader mask. (sample shading TODO).
2863 * Take the incoming state coverage mask into account.
2865 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2866 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2867 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2868 &sindexi
, 1, "sample_mask_ptr");
2869 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2870 i
*fs_type
.length
/4, s
, mask_input
);
2872 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2873 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2874 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2875 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2877 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2878 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
2882 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2883 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2884 &indexi
, 1, "mask_ptr");
2887 mask
= generate_quad_mask(gallivm
, fs_type
,
2888 i
*fs_type
.length
/4, 0, mask_input
);
2891 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2893 LLVMBuildStore(builder
, mask
, mask_ptr
);
2897 generate_fs_loop(gallivm
,
2907 mask_store
, /* output */
2911 depth_sample_stride
,
2915 for (i
= 0; i
< num_fs
; i
++) {
2916 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2918 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2919 int idx
= (i
+ (s
* num_fs
));
2920 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
2921 ptr
= LLVMBuildGEP(builder
, mask_store
, &sindexi
, 1, "");
2923 fs_mask
[idx
] = LLVMBuildLoad(builder
, ptr
, "smask");
2926 /* This is fucked up need to reorganize things */
2927 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2928 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2929 ptr
= LLVMBuildGEP(builder
,
2930 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2932 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2935 if (dual_source_blend
) {
2936 /* only support one dual source blend target hence always use output 1 */
2937 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2938 ptr
= LLVMBuildGEP(builder
,
2939 color_store
[1][chan
],
2941 fs_out_color
[1][chan
][i
] = ptr
;
2947 sampler
->destroy(sampler
);
2948 image
->destroy(image
);
2949 /* Loop over color outputs / color buffers to do blending.
2951 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2952 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2953 LLVMValueRef color_ptr
;
2954 LLVMValueRef stride
;
2955 LLVMValueRef sample_stride
= NULL
;
2956 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2958 boolean do_branch
= ((key
->depth
.enabled
2959 || key
->stencil
[0].enabled
2960 || key
->alpha
.enabled
)
2961 && !shader
->info
.base
.uses_kill
);
2963 color_ptr
= LLVMBuildLoad(builder
,
2964 LLVMBuildGEP(builder
, color_ptr_ptr
,
2968 stride
= LLVMBuildLoad(builder
,
2969 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2972 if (key
->multisample
)
2973 sample_stride
= LLVMBuildLoad(builder
,
2974 LLVMBuildGEP(builder
, color_sample_stride_ptr
,
2975 &index
, 1, ""), "");
2977 for (unsigned s
= 0; s
< key
->cbuf_nr_samples
[cbuf
]; s
++) {
2978 unsigned mask_idx
= num_fs
* (key
->multisample
? s
: 0);
2979 LLVMValueRef out_ptr
= color_ptr
;;
2981 if (key
->multisample
) {
2982 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_stride
, lp_build_const_int32(gallivm
, s
), "");
2983 out_ptr
= LLVMBuildGEP(builder
, out_ptr
, &sample_offset
, 1, "");
2985 out_ptr
= LLVMBuildBitCast(builder
, out_ptr
, LLVMPointerType(blend_vec_type
, 0), "");
2987 lp_build_name(out_ptr
, "color_ptr%d", cbuf
);
2989 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2990 key
->cbuf_format
[cbuf
],
2991 num_fs
, fs_type
, &fs_mask
[mask_idx
], fs_out_color
,
2992 context_ptr
, out_ptr
, stride
,
2993 partial_mask
, do_branch
);
2998 LLVMBuildRetVoid(builder
);
3000 gallivm_verify_function(gallivm
, function
);
3005 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
3009 debug_printf("fs variant %p:\n", (void *) key
);
3011 if (key
->flatshade
) {
3012 debug_printf("flatshade = 1\n");
3014 if (key
->multisample
) {
3015 debug_printf("multisample = 1\n");
3016 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
3018 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
3019 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
3020 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
3022 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
3023 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
3024 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
3026 if (key
->depth
.enabled
) {
3027 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
3028 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
3031 for (i
= 0; i
< 2; ++i
) {
3032 if (key
->stencil
[i
].enabled
) {
3033 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
3034 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
3035 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
3036 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
3037 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
3038 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
3042 if (key
->alpha
.enabled
) {
3043 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
3046 if (key
->occlusion_count
) {
3047 debug_printf("occlusion_count = 1\n");
3050 if (key
->blend
.logicop_enable
) {
3051 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
3053 else if (key
->blend
.rt
[0].blend_enable
) {
3054 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
3055 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
3056 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
3057 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
3058 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
3059 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
3061 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
3062 if (key
->blend
.alpha_to_coverage
) {
3063 debug_printf("blend.alpha_to_coverage is enabled\n");
3065 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
3066 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
3067 debug_printf("sampler[%u] = \n", i
);
3068 debug_printf(" .wrap = %s %s %s\n",
3069 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
3070 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
3071 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
3072 debug_printf(" .min_img_filter = %s\n",
3073 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
3074 debug_printf(" .min_mip_filter = %s\n",
3075 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
3076 debug_printf(" .mag_img_filter = %s\n",
3077 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
3078 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
3079 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3080 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3081 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3082 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3083 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3084 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3086 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3087 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3088 debug_printf("texture[%u] = \n", i
);
3089 debug_printf(" .format = %s\n",
3090 util_format_name(texture
->format
));
3091 debug_printf(" .target = %s\n",
3092 util_str_tex_target(texture
->target
, TRUE
));
3093 debug_printf(" .level_zero_only = %u\n",
3094 texture
->level_zero_only
);
3095 debug_printf(" .pot = %u %u %u\n",
3097 texture
->pot_height
,
3098 texture
->pot_depth
);
3100 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3101 for (i
= 0; i
< key
->nr_images
; ++i
) {
3102 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3103 debug_printf("image[%u] = \n", i
);
3104 debug_printf(" .format = %s\n",
3105 util_format_name(image
->format
));
3106 debug_printf(" .target = %s\n",
3107 util_str_tex_target(image
->target
, TRUE
));
3108 debug_printf(" .level_zero_only = %u\n",
3109 image
->level_zero_only
);
3110 debug_printf(" .pot = %u %u %u\n",
3119 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3121 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3122 variant
->shader
->no
, variant
->no
);
3123 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3124 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3126 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3127 dump_fs_variant_key(&variant
->key
);
3128 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3134 * Generate a new fragment shader variant from the shader code and
3135 * other state indicated by the key.
3137 static struct lp_fragment_shader_variant
*
3138 generate_variant(struct llvmpipe_context
*lp
,
3139 struct lp_fragment_shader
*shader
,
3140 const struct lp_fragment_shader_variant_key
*key
)
3142 struct lp_fragment_shader_variant
*variant
;
3143 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3144 boolean fullcolormask
;
3145 char module_name
[64];
3147 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3151 memset(variant
, 0, sizeof(*variant
));
3152 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3153 shader
->no
, shader
->variants_created
);
3155 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
3156 if (!variant
->gallivm
) {
3161 variant
->shader
= shader
;
3162 variant
->list_item_global
.base
= variant
;
3163 variant
->list_item_local
.base
= variant
;
3164 variant
->no
= shader
->variants_created
++;
3166 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3169 * Determine whether we are touching all channels in the color buffer.
3171 fullcolormask
= FALSE
;
3172 if (key
->nr_cbufs
== 1) {
3173 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3174 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3178 !key
->blend
.logicop_enable
&&
3179 !key
->blend
.rt
[0].blend_enable
&&
3181 !key
->stencil
[0].enabled
&&
3182 !key
->alpha
.enabled
&&
3183 !key
->blend
.alpha_to_coverage
&&
3184 !key
->depth
.enabled
&&
3185 !shader
->info
.base
.uses_kill
&&
3186 !shader
->info
.base
.writes_samplemask
3189 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3190 lp_debug_fs_variant(variant
);
3193 lp_jit_init_types(variant
);
3195 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3196 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3198 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3199 if (variant
->opaque
) {
3200 /* Specialized shader, which doesn't need to read the color buffer. */
3201 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3206 * Compile everything
3209 gallivm_compile_module(variant
->gallivm
);
3211 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3213 if (variant
->function
[RAST_EDGE_TEST
]) {
3214 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3215 gallivm_jit_function(variant
->gallivm
,
3216 variant
->function
[RAST_EDGE_TEST
]);
3219 if (variant
->function
[RAST_WHOLE
]) {
3220 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3221 gallivm_jit_function(variant
->gallivm
,
3222 variant
->function
[RAST_WHOLE
]);
3223 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3224 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3227 gallivm_free_ir(variant
->gallivm
);
3234 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3235 const struct pipe_shader_state
*templ
)
3237 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3238 struct lp_fragment_shader
*shader
;
3240 int nr_sampler_views
;
3244 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3248 shader
->no
= fs_no
++;
3249 make_empty_list(&shader
->variants
);
3251 shader
->base
.type
= templ
->type
;
3252 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3253 /* get/save the summary info for this shader */
3254 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3256 /* we need to keep a local copy of the tokens */
3257 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3259 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3260 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3263 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3264 if (shader
->draw_data
== NULL
) {
3265 FREE((void *) shader
->base
.tokens
);
3270 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3271 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3272 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3273 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3275 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3276 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3277 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3278 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3280 switch (shader
->info
.base
.input_interpolate
[i
]) {
3281 case TGSI_INTERPOLATE_CONSTANT
:
3282 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3284 case TGSI_INTERPOLATE_LINEAR
:
3285 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3287 case TGSI_INTERPOLATE_PERSPECTIVE
:
3288 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3290 case TGSI_INTERPOLATE_COLOR
:
3291 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3298 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3299 case TGSI_SEMANTIC_FACE
:
3300 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3302 case TGSI_SEMANTIC_POSITION
:
3303 /* Position was already emitted above
3305 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3306 shader
->inputs
[i
].src_index
= 0;
3310 /* XXX this is a completely pointless index map... */
3311 shader
->inputs
[i
].src_index
= i
+1;
3314 if (LP_DEBUG
& DEBUG_TGSI
) {
3316 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3317 shader
->no
, (void *) shader
);
3318 tgsi_dump(templ
->tokens
, 0);
3319 debug_printf("usage masks:\n");
3320 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3321 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3322 debug_printf(" IN[%u].%s%s%s%s\n",
3324 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3325 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3326 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3327 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3337 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3339 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3340 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3341 if (llvmpipe
->fs
== lp_fs
)
3344 draw_bind_fragment_shader(llvmpipe
->draw
,
3345 (lp_fs
? lp_fs
->draw_data
: NULL
));
3347 llvmpipe
->fs
= lp_fs
;
3349 llvmpipe
->dirty
|= LP_NEW_FS
;
3354 * Remove shader variant from two lists: the shader's variant list
3355 * and the context's variant list.
3358 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3359 struct lp_fragment_shader_variant
*variant
)
3361 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3362 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3363 "v total cached %u inst %u total inst %u\n",
3364 variant
->shader
->no
, variant
->no
,
3365 variant
->shader
->variants_created
,
3366 variant
->shader
->variants_cached
,
3367 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3370 gallivm_destroy(variant
->gallivm
);
3372 /* remove from shader's list */
3373 remove_from_list(&variant
->list_item_local
);
3374 variant
->shader
->variants_cached
--;
3376 /* remove from context's list */
3377 remove_from_list(&variant
->list_item_global
);
3378 lp
->nr_fs_variants
--;
3379 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3386 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3388 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3389 struct lp_fragment_shader
*shader
= fs
;
3390 struct lp_fs_variant_list_item
*li
;
3392 assert(fs
!= llvmpipe
->fs
);
3395 * XXX: we need to flush the context until we have some sort of reference
3396 * counting in fragment shaders as they may still be binned
3397 * Flushing alone might not sufficient we need to wait on it too.
3399 llvmpipe_finish(pipe
, __FUNCTION__
);
3401 /* Delete all the variants */
3402 li
= first_elem(&shader
->variants
);
3403 while(!at_end(&shader
->variants
, li
)) {
3404 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3405 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3409 /* Delete draw module's data */
3410 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3412 if (shader
->base
.ir
.nir
)
3413 ralloc_free(shader
->base
.ir
.nir
);
3414 assert(shader
->variants_cached
== 0);
3415 FREE((void *) shader
->base
.tokens
);
3422 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3423 enum pipe_shader_type shader
, uint index
,
3424 const struct pipe_constant_buffer
*cb
)
3426 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3427 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3429 assert(shader
< PIPE_SHADER_TYPES
);
3430 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3432 /* note: reference counting */
3433 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3436 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3437 debug_printf("Illegal set constant without bind flag\n");
3438 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3442 if (shader
== PIPE_SHADER_VERTEX
||
3443 shader
== PIPE_SHADER_GEOMETRY
||
3444 shader
== PIPE_SHADER_TESS_CTRL
||
3445 shader
== PIPE_SHADER_TESS_EVAL
) {
3446 /* Pass the constants to the 'draw' module */
3447 const unsigned size
= cb
? cb
->buffer_size
: 0;
3451 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3453 else if (cb
&& cb
->user_buffer
) {
3454 data
= (ubyte
*) cb
->user_buffer
;
3461 data
+= cb
->buffer_offset
;
3463 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3466 else if (shader
== PIPE_SHADER_COMPUTE
)
3467 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3469 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3471 if (cb
&& cb
->user_buffer
) {
3472 pipe_resource_reference(&constants
, NULL
);
3477 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3478 enum pipe_shader_type shader
, unsigned start_slot
,
3479 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3480 unsigned writable_bitmask
)
3482 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3484 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3485 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3487 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3489 if (shader
== PIPE_SHADER_VERTEX
||
3490 shader
== PIPE_SHADER_GEOMETRY
||
3491 shader
== PIPE_SHADER_TESS_CTRL
||
3492 shader
== PIPE_SHADER_TESS_EVAL
) {
3493 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3494 const ubyte
*data
= NULL
;
3495 if (buffer
&& buffer
->buffer
)
3496 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3498 data
+= buffer
->buffer_offset
;
3499 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3501 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3502 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3503 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3504 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3510 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3511 enum pipe_shader_type shader
, unsigned start_slot
,
3512 unsigned count
, const struct pipe_image_view
*images
)
3514 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3517 draw_flush(llvmpipe
->draw
);
3518 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3519 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3521 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3524 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3525 if (shader
== PIPE_SHADER_VERTEX
||
3526 shader
== PIPE_SHADER_GEOMETRY
||
3527 shader
== PIPE_SHADER_TESS_CTRL
||
3528 shader
== PIPE_SHADER_TESS_EVAL
) {
3529 draw_set_images(llvmpipe
->draw
,
3531 llvmpipe
->images
[shader
],
3532 start_slot
+ count
);
3533 } else if (shader
== PIPE_SHADER_COMPUTE
)
3534 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3536 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3540 * Return the blend factor equivalent to a destination alpha of one.
3542 static inline unsigned
3543 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3546 case PIPE_BLENDFACTOR_DST_ALPHA
:
3547 return PIPE_BLENDFACTOR_ONE
;
3548 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3549 return PIPE_BLENDFACTOR_ZERO
;
3550 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3552 return PIPE_BLENDFACTOR_ZERO
;
3554 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3562 * We need to generate several variants of the fragment pipeline to match
3563 * all the combinations of the contributing state atoms.
3565 * TODO: there is actually no reason to tie this to context state -- the
3566 * generated code could be cached globally in the screen.
3568 static struct lp_fragment_shader_variant_key
*
3569 make_variant_key(struct llvmpipe_context
*lp
,
3570 struct lp_fragment_shader
*shader
,
3574 struct lp_fragment_shader_variant_key
*key
;
3576 key
= (struct lp_fragment_shader_variant_key
*)store
;
3578 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3580 if (lp
->framebuffer
.zsbuf
) {
3581 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3582 const struct util_format_description
*zsbuf_desc
=
3583 util_format_description(zsbuf_format
);
3585 if (lp
->depth_stencil
->depth
.enabled
&&
3586 util_format_has_depth(zsbuf_desc
)) {
3587 key
->zsbuf_format
= zsbuf_format
;
3588 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3590 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3591 util_format_has_stencil(zsbuf_desc
)) {
3592 key
->zsbuf_format
= zsbuf_format
;
3593 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3595 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3596 key
->resource_1d
= TRUE
;
3598 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3602 * Propagate the depth clamp setting from the rasterizer state.
3603 * depth_clip == 0 implies depth clamping is enabled.
3605 * When clip_halfz is enabled, then always clamp the depth values.
3607 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3608 * clamp is always active in d3d10, regardless if depth clip is
3610 * (GL has an always-on [0,1] clamp on fs depth output instead
3611 * to ensure the depth values stay in range. Doesn't look like
3612 * we do that, though...)
3614 if (lp
->rasterizer
->clip_halfz
) {
3615 key
->depth_clamp
= 1;
3617 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3620 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3621 if (!lp
->framebuffer
.nr_cbufs
||
3622 !lp
->framebuffer
.cbufs
[0] ||
3623 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3624 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3626 if(key
->alpha
.enabled
)
3627 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3628 /* alpha.ref_value is passed in jit_context */
3630 key
->flatshade
= lp
->rasterizer
->flatshade
;
3631 key
->multisample
= lp
->rasterizer
->multisample
;
3632 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3633 key
->occlusion_count
= TRUE
;
3636 if (lp
->framebuffer
.nr_cbufs
) {
3637 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3640 key
->coverage_samples
= 1;
3641 if (key
->multisample
)
3642 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3643 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3645 if (!key
->blend
.independent_blend_enable
) {
3646 /* we always need independent blend otherwise the fixups below won't work */
3647 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3648 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3650 key
->blend
.independent_blend_enable
= 1;
3653 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3654 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3656 if (lp
->framebuffer
.cbufs
[i
]) {
3657 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3658 const struct util_format_description
*format_desc
;
3660 key
->cbuf_format
[i
] = format
;
3661 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3664 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3665 * mixing of 2d textures with height 1 and 1d textures, so make sure
3666 * we pick 1d if any cbuf or zsbuf is 1d.
3668 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3669 key
->resource_1d
= TRUE
;
3672 format_desc
= util_format_description(format
);
3673 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3674 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3677 * Mask out color channels not present in the color buffer.
3679 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3682 * Disable blend for integer formats.
3684 if (util_format_is_pure_integer(format
)) {
3685 blend_rt
->blend_enable
= 0;
3689 * Our swizzled render tiles always have an alpha channel, but the
3690 * linear render target format often does not, so force here the dst
3693 * This is not a mere optimization. Wrong results will be produced if
3694 * the dst alpha is used, the dst format does not have alpha, and the
3695 * previous rendering was not flushed from the swizzled to linear
3696 * buffer. For example, NonPowTwo DCT.
3698 * TODO: This should be generalized to all channels for better
3699 * performance, but only alpha causes correctness issues.
3701 * Also, force rgb/alpha func/factors match, to make AoS blending
3704 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3705 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3706 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3707 boolean clamped_zero
= !util_format_is_float(format
) &&
3708 !util_format_is_snorm(format
);
3709 blend_rt
->rgb_src_factor
=
3710 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3711 blend_rt
->rgb_dst_factor
=
3712 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3713 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3714 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3715 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3719 /* no color buffer for this fragment output */
3720 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3721 key
->cbuf_nr_samples
[i
] = 0;
3722 blend_rt
->colormask
= 0x0;
3723 blend_rt
->blend_enable
= 0;
3727 /* This value will be the same for all the variants of a given shader:
3729 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3731 struct lp_sampler_static_state
*fs_sampler
;
3733 fs_sampler
= key
->samplers
;
3735 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3737 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3738 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3739 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3740 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3745 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3746 * are dx10-style? Can't really have mixed opcodes, at least not
3747 * if we want to skip the holes here (without rescanning tgsi).
3749 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3750 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3751 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3753 * Note sview may exceed what's representable by file_mask.
3754 * This will still work, the only downside is that not actually
3755 * used views may be included in the shader key.
3757 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3758 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3759 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3764 key
->nr_sampler_views
= key
->nr_samplers
;
3765 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3766 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3767 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3768 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3773 struct lp_image_static_state
*lp_image
;
3774 lp_image
= lp_fs_variant_key_images(key
);
3775 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3776 for (i
= 0; i
< key
->nr_images
; ++i
) {
3777 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3778 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3779 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3788 * Update fragment shader state. This is called just prior to drawing
3789 * something when some fragment-related state has changed.
3792 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3794 struct lp_fragment_shader
*shader
= lp
->fs
;
3795 struct lp_fragment_shader_variant_key
*key
;
3796 struct lp_fragment_shader_variant
*variant
= NULL
;
3797 struct lp_fs_variant_list_item
*li
;
3798 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3800 key
= make_variant_key(lp
, shader
, store
);
3802 /* Search the variants for one which matches the key */
3803 li
= first_elem(&shader
->variants
);
3804 while(!at_end(&shader
->variants
, li
)) {
3805 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3813 /* Move this variant to the head of the list to implement LRU
3814 * deletion of shader's when we have too many.
3816 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3819 /* variant not found, create it now */
3822 unsigned variants_to_cull
;
3824 if (LP_DEBUG
& DEBUG_FS
) {
3825 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3828 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3831 /* First, check if we've exceeded the max number of shader variants.
3832 * If so, free 6.25% of them (the least recently used ones).
3834 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3836 if (variants_to_cull
||
3837 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3838 struct pipe_context
*pipe
= &lp
->pipe
;
3840 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3841 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3842 "\t%u instrs,\t%u instrs/variant\n",
3843 shader
->variants_cached
,
3844 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3845 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3849 * XXX: we need to flush the context until we have some sort of
3850 * reference counting in fragment shaders as they may still be binned
3851 * Flushing alone might not be sufficient we need to wait on it too.
3853 llvmpipe_finish(pipe
, __FUNCTION__
);
3856 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3857 * number of shader variants (potentially all of them) could be
3858 * pending for destruction on flush.
3861 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3862 struct lp_fs_variant_list_item
*item
;
3863 if (is_empty_list(&lp
->fs_variants_list
)) {
3866 item
= last_elem(&lp
->fs_variants_list
);
3869 llvmpipe_remove_shader_variant(lp
, item
->base
);
3874 * Generate the new variant.
3877 variant
= generate_variant(lp
, shader
, key
);
3880 LP_COUNT_ADD(llvm_compile_time
, dt
);
3881 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3883 /* Put the new variant into the list */
3885 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3886 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3887 lp
->nr_fs_variants
++;
3888 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3889 shader
->variants_cached
++;
3893 /* Bind this variant */
3894 lp_setup_set_fs_variant(lp
->setup
, variant
);
3902 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3904 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3905 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3906 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3908 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3910 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3911 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;