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 /* Allocate color storage for each fragment sample */
448 LLVMValueRef color_store_size
= num_loop
;
449 if (key
->min_samples
> 1)
450 color_store_size
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, key
->min_samples
), "");
452 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
453 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
454 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
455 lp_build_vec_type(gallivm
,
457 color_store_size
, "color");
460 if (dual_source_blend
) {
461 assert(key
->nr_cbufs
<= 1);
462 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
463 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
464 lp_build_vec_type(gallivm
,
466 color_store_size
, "color1");
470 lp_build_for_loop_begin(&loop_state
, gallivm
,
471 lp_build_const_int32(gallivm
, 0),
474 lp_build_const_int32(gallivm
, 1));
476 if (key
->multisample
) {
477 /* create shader execution mask by combining all sample masks. */
478 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
479 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
480 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
481 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
485 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
488 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
489 &loop_state
.counter
, 1, "mask_ptr");
490 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
493 /* 'mask' will control execution based on quad's pixel alive/killed state */
494 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
496 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
497 lp_build_mask_check(&mask
);
499 /* Create storage for recombining sample masks after early Z pass. */
500 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
501 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
503 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
504 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
505 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
506 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
508 /* Run early depth once per sample */
509 if (key
->multisample
) {
511 if (zs_format_desc
) {
512 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
513 struct lp_type z_type
= zs_type
;
514 struct lp_type s_type
= zs_type
;
515 if (zs_format_desc
->block
.bits
< type
.width
)
516 z_type
.width
= type
.width
;
517 else if (zs_format_desc
->block
.bits
> 32) {
518 z_type
.width
= z_type
.width
/ 2;
519 s_type
.width
= s_type
.width
/ 2;
522 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
523 zs_samples
, "z_sample_store");
524 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
525 zs_samples
, "s_sample_store");
526 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
527 zs_samples
, "z_fb_store");
528 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
529 zs_samples
, "s_fb_store");
531 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
532 lp_build_const_int32(gallivm
, 0),
533 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
534 lp_build_const_int32(gallivm
, 1));
536 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
537 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
538 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
540 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
541 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
545 /* for multisample Z needs to be interpolated at sample points for testing. */
546 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
549 depth_ptr
= depth_base_ptr
;
550 if (key
->multisample
) {
551 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
552 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
555 if (depth_mode
& EARLY_DEPTH_TEST
) {
557 * Clamp according to ARB_depth_clamp semantics.
559 if (key
->depth_clamp
) {
560 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
563 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
564 zs_format_desc
, key
->resource_1d
,
565 depth_ptr
, depth_stride
,
566 &z_fb
, &s_fb
, loop_state
.counter
);
567 lp_build_depth_stencil_test(gallivm
,
572 key
->multisample
? NULL
: &mask
,
578 !simple_shader
&& !key
->multisample
);
580 if (depth_mode
& EARLY_DEPTH_WRITE
) {
581 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
582 zs_format_desc
, key
->resource_1d
,
583 NULL
, NULL
, NULL
, loop_state
.counter
,
584 depth_ptr
, depth_stride
,
588 * Note mask check if stencil is enabled must be after ds write not after
589 * stencil test otherwise new stencil values may not get written if all
590 * fragments got killed by depth/stencil test.
592 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
593 lp_build_mask_check(&mask
);
595 if (key
->multisample
) {
596 z_fb_type
= LLVMTypeOf(z_fb
);
597 z_type
= LLVMTypeOf(z_value
);
598 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
599 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
600 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
601 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
605 if (key
->multisample
) {
607 * Store the post-early Z coverage mask.
608 * Recombine the resulting coverage masks post early Z into the fragment
609 * shader execution mask.
611 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
612 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
613 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
615 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
617 lp_build_for_loop_end(&sample_loop_state
);
619 /* recombined all the coverage masks in the shader exec mask. */
620 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
621 lp_build_mask_update(&mask
, tmp_s_mask_or
);
623 if (key
->min_samples
== 1) {
624 /* for multisample Z needs to be re interpolated at pixel center */
625 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
626 lp_build_mask_update(&mask
, tmp_s_mask_or
);
630 LLVMValueRef out_sample_mask_storage
= NULL
;
631 if (shader
->info
.base
.writes_samplemask
) {
632 out_sample_mask_storage
= lp_build_alloca(gallivm
, int_vec_type
, "write_mask");
633 if (key
->min_samples
> 1)
634 LLVMBuildStore(builder
, LLVMConstNull(int_vec_type
), out_sample_mask_storage
);
637 if (key
->multisample
&& key
->min_samples
> 1) {
638 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
639 lp_build_const_int32(gallivm
, 0),
641 lp_build_const_int32(gallivm
, key
->min_samples
),
642 lp_build_const_int32(gallivm
, 1));
644 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
645 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
646 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
647 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
648 lp_build_mask_force(&mask
, s_mask
);
649 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, sample_loop_state
.counter
);
650 system_values
.sample_id
= sample_loop_state
.counter
;
652 system_values
.sample_id
= lp_build_const_int32(gallivm
, 0);
654 system_values
.sample_pos
= sample_pos_array
;
656 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, mask_store
, system_values
.sample_id
);
658 struct lp_build_tgsi_params params
;
659 memset(¶ms
, 0, sizeof(params
));
663 params
.consts_ptr
= consts_ptr
;
664 params
.const_sizes_ptr
= num_consts_ptr
;
665 params
.system_values
= &system_values
;
666 params
.inputs
= interp
->inputs
;
667 params
.context_ptr
= context_ptr
;
668 params
.thread_data_ptr
= thread_data_ptr
;
669 params
.sampler
= sampler
;
670 params
.info
= &shader
->info
.base
;
671 params
.ssbo_ptr
= ssbo_ptr
;
672 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
673 params
.image
= image
;
675 /* Build the actual shader */
676 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
677 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
680 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
684 if (key
->alpha
.enabled
) {
685 int color0
= find_output_by_semantic(&shader
->info
.base
,
689 if (color0
!= -1 && outputs
[color0
][3]) {
690 const struct util_format_description
*cbuf_format_desc
;
691 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
692 LLVMValueRef alpha_ref_value
;
694 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
695 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
697 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
699 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
700 &mask
, alpha
, alpha_ref_value
,
701 (depth_mode
& LATE_DEPTH_TEST
) != 0);
705 /* Emulate Alpha to Coverage with Alpha test */
706 if (key
->blend
.alpha_to_coverage
) {
707 int color0
= find_output_by_semantic(&shader
->info
.base
,
711 if (color0
!= -1 && outputs
[color0
][3]) {
712 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
714 if (!key
->multisample
) {
715 lp_build_alpha_to_coverage(gallivm
, type
,
717 (depth_mode
& LATE_DEPTH_TEST
) != 0);
719 lp_build_sample_alpha_to_coverage(gallivm
, type
, key
->coverage_samples
, num_loop
,
725 if (key
->blend
.alpha_to_one
&& key
->multisample
) {
726 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
) {
727 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
728 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
729 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
730 if (outputs
[cbuf
][3]) {
731 LLVMBuildStore(builder
, lp_build_const_vec(gallivm
, type
, 1.0), outputs
[cbuf
][3]);
735 if (shader
->info
.base
.writes_samplemask
) {
736 LLVMValueRef output_smask
= NULL
;
737 int smaski
= find_output_by_semantic(&shader
->info
.base
,
738 TGSI_SEMANTIC_SAMPLEMASK
,
740 struct lp_build_context smask_bld
;
741 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
744 output_smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
746 * Pixel is alive according to the first sample in the mask.
748 output_smask
= LLVMBuildBitCast(builder
, output_smask
, smask_bld
.vec_type
, "");
749 if (!key
->multisample
) {
750 output_smask
= lp_build_and(&smask_bld
, output_smask
, smask_bld
.one
);
751 output_smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, output_smask
, smask_bld
.zero
);
752 lp_build_mask_update(&mask
, output_smask
);
755 if (key
->min_samples
> 1) {
756 /* only the bit corresponding to this sample is to be used. */
757 LLVMValueRef tmp_mask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "tmp_mask");
758 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
759 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
), "");
760 output_smask
= LLVMBuildOr(builder
, tmp_mask
, smask_bit
, "");
763 LLVMBuildStore(builder
, output_smask
, out_sample_mask_storage
);
766 /* Color write - per fragment sample */
767 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
769 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
770 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
771 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
773 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
774 if(outputs
[attrib
][chan
]) {
775 /* XXX: just initialize outputs to point at colors[] and
778 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
779 LLVMValueRef color_ptr
;
780 LLVMValueRef color_idx
= loop_state
.counter
;
781 if (key
->min_samples
> 1)
782 color_idx
= LLVMBuildAdd(builder
, color_idx
,
783 LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, ""), "");
784 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
786 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
787 LLVMBuildStore(builder
, out
, color_ptr
);
793 if (key
->multisample
&& key
->min_samples
> 1) {
794 LLVMBuildStore(builder
, lp_build_mask_value(&mask
), s_mask_ptr
);
795 lp_build_for_loop_end(&sample_loop_state
);
798 if (key
->multisample
) {
799 /* execute depth test for each sample */
800 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
801 lp_build_const_int32(gallivm
, 0),
802 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
803 lp_build_const_int32(gallivm
, 1));
805 /* load the per-sample coverage mask */
806 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
807 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
808 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
810 /* combine the execution mask post fragment shader with the coverage mask. */
811 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
812 if (key
->min_samples
== 1)
813 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
815 /* if the shader writes sample mask use that */
816 if (shader
->info
.base
.writes_samplemask
) {
817 LLVMValueRef out_smask_idx
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1), sample_loop_state
.counter
, "");
818 out_smask_idx
= lp_build_broadcast(gallivm
, int_vec_type
, out_smask_idx
);
819 LLVMValueRef output_smask
= LLVMBuildLoad(builder
, out_sample_mask_storage
, "");
820 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, output_smask
, out_smask_idx
, "");
821 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int_vec(gallivm
, int_type
, 0), "");
822 smask_bit
= LLVMBuildSExt(builder
, cmp
, int_vec_type
, "");
824 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
828 depth_ptr
= depth_base_ptr
;
829 if (key
->multisample
) {
830 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
831 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
835 if (depth_mode
& LATE_DEPTH_TEST
) {
836 int pos0
= find_output_by_semantic(&shader
->info
.base
,
837 TGSI_SEMANTIC_POSITION
,
839 int s_out
= find_output_by_semantic(&shader
->info
.base
,
840 TGSI_SEMANTIC_STENCIL
,
842 if (pos0
!= -1 && outputs
[pos0
][2]) {
843 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
846 * Clamp according to ARB_depth_clamp semantics.
848 if (key
->depth_clamp
) {
849 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
853 if (s_out
!= -1 && outputs
[s_out
][1]) {
854 /* there's only one value, and spec says to discard additional bits */
855 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
856 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
857 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
858 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
859 stencil_refs
[1] = stencil_refs
[0];
862 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
863 zs_format_desc
, key
->resource_1d
,
864 depth_ptr
, depth_stride
,
865 &z_fb
, &s_fb
, loop_state
.counter
);
867 lp_build_depth_stencil_test(gallivm
,
872 key
->multisample
? NULL
: &mask
,
880 if (depth_mode
& LATE_DEPTH_WRITE
) {
881 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
882 zs_format_desc
, key
->resource_1d
,
883 NULL
, NULL
, NULL
, loop_state
.counter
,
884 depth_ptr
, depth_stride
,
888 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
889 (depth_mode
& LATE_DEPTH_WRITE
))
891 /* Need to apply a reduced mask to the depth write. Reload the
892 * depth value, update from zs_value with the new mask value and
895 if (key
->multisample
) {
896 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
897 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
898 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
899 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
901 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
902 zs_format_desc
, key
->resource_1d
,
903 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
904 depth_ptr
, depth_stride
,
908 if (key
->occlusion_count
) {
909 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
910 lp_build_name(counter
, "counter");
912 lp_build_occlusion_count(gallivm
, type
,
913 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), counter
);
916 if (key
->multisample
) {
917 /* store the sample mask for this loop */
918 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
919 lp_build_for_loop_end(&sample_loop_state
);
922 mask_val
= lp_build_mask_end(&mask
);
923 if (!key
->multisample
)
924 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
925 lp_build_for_loop_end(&loop_state
);
930 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
932 * Fragment Shader outputs pixels in small 2x2 blocks
933 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
935 * However in memory pixels are stored in rows
936 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
938 * @param type fragment shader type (4x or 8x float)
939 * @param num_fs number of fs_src
940 * @param is_1d whether we're outputting to a 1d resource
941 * @param dst_channels number of output channels
942 * @param fs_src output from fragment shader
943 * @param dst pointer to store result
944 * @param pad_inline is channel padding inline or at end of row
945 * @return the number of dsts
948 generate_fs_twiddle(struct gallivm_state
*gallivm
,
951 unsigned dst_channels
,
952 LLVMValueRef fs_src
[][4],
956 LLVMValueRef src
[16];
962 unsigned pixels
= type
.length
/ 4;
963 unsigned reorder_group
;
964 unsigned src_channels
;
968 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
969 src_count
= num_fs
* src_channels
;
971 assert(pixels
== 2 || pixels
== 1);
972 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
975 * Transpose from SoA -> AoS
977 for (i
= 0; i
< num_fs
; ++i
) {
978 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
982 * Pick transformation options
989 if (dst_channels
== 1) {
995 } else if (dst_channels
== 2) {
999 } else if (dst_channels
> 2) {
1006 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
1012 * Split the src in half
1015 for (i
= num_fs
; i
> 0; --i
) {
1016 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
1017 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
1025 * Ensure pixels are in memory order
1027 if (reorder_group
) {
1028 /* Twiddle pixels by reordering the array, e.g.:
1030 * src_count = 8 -> 0 2 1 3 4 6 5 7
1031 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1033 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
1035 for (i
= 0; i
< src_count
; ++i
) {
1036 unsigned group
= i
/ reorder_group
;
1037 unsigned block
= (group
/ 4) * 4 * reorder_group
;
1038 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
1041 } else if (twiddle
) {
1042 /* Twiddle pixels across elements of array */
1044 * XXX: we should avoid this in some cases, but would need to tell
1045 * lp_build_conv to reorder (or deal with it ourselves).
1047 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
1050 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
1054 * Moves any padding between pixels to the end
1055 * e.g. RGBXRGBX -> RGBRGBXX
1058 unsigned char swizzles
[16];
1059 unsigned elems
= pixels
* dst_channels
;
1061 for (i
= 0; i
< type
.length
; ++i
) {
1063 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
1065 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
1068 for (i
= 0; i
< src_count
; ++i
) {
1069 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
1078 * Untwiddle and transpose, much like the above.
1079 * However, this is after conversion, so we get packed vectors.
1080 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1081 * the vectors will look like:
1082 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1083 * be swizzled here). Extending to 16bit should be trivial.
1084 * Should also be extended to handle twice wide vectors with AVX2...
1087 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
1088 struct lp_type type
,
1094 struct lp_type type64
, type16
, type32
;
1095 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
1096 LLVMBuilderRef builder
= gallivm
->builder
;
1097 LLVMValueRef tmp
[4], shuf
[8];
1098 for (j
= 0; j
< 2; j
++) {
1099 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
1100 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
1101 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
1102 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
1105 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
1106 assert(type
.width
== 8);
1107 assert(type
.length
== 16);
1109 type8_t
= lp_build_vec_type(gallivm
, type
);
1114 type64_t
= lp_build_vec_type(gallivm
, type64
);
1119 type16_t
= lp_build_vec_type(gallivm
, type16
);
1124 type32_t
= lp_build_vec_type(gallivm
, type32
);
1126 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1128 if (src_count
== 1) {
1129 /* transpose was no-op, just untwiddle */
1130 LLVMValueRef shuf_vec
;
1131 shuf_vec
= LLVMConstVector(shuf
, 8);
1132 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1133 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1134 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1135 } else if (src_count
== 2) {
1136 LLVMValueRef shuf_vec
;
1137 shuf_vec
= LLVMConstVector(shuf
, 4);
1139 for (i
= 0; i
< 2; i
++) {
1140 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1141 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1142 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1145 for (j
= 0; j
< 2; j
++) {
1146 LLVMValueRef lo
, hi
, lo2
, hi2
;
1148 * Note that if we only really have 3 valid channels (rgb)
1149 * and we don't need alpha we could substitute a undef here
1150 * for the respective channel (causing llvm to drop conversion
1153 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1154 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1155 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1156 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1157 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1158 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1159 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1166 * Load an unswizzled block of pixels from memory
1169 load_unswizzled_block(struct gallivm_state
*gallivm
,
1170 LLVMValueRef base_ptr
,
1171 LLVMValueRef stride
,
1172 unsigned block_width
,
1173 unsigned block_height
,
1175 struct lp_type dst_type
,
1177 unsigned dst_alignment
)
1179 LLVMBuilderRef builder
= gallivm
->builder
;
1180 unsigned row_size
= dst_count
/ block_height
;
1183 /* Ensure block exactly fits into dst */
1184 assert((block_width
* block_height
) % dst_count
== 0);
1186 for (i
= 0; i
< dst_count
; ++i
) {
1187 unsigned x
= i
% row_size
;
1188 unsigned y
= i
/ row_size
;
1190 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1191 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1193 LLVMValueRef gep
[2];
1194 LLVMValueRef dst_ptr
;
1196 gep
[0] = lp_build_const_int32(gallivm
, 0);
1197 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1199 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1200 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1201 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1203 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1205 LLVMSetAlignment(dst
[i
], dst_alignment
);
1211 * Store an unswizzled block of pixels to memory
1214 store_unswizzled_block(struct gallivm_state
*gallivm
,
1215 LLVMValueRef base_ptr
,
1216 LLVMValueRef stride
,
1217 unsigned block_width
,
1218 unsigned block_height
,
1220 struct lp_type src_type
,
1222 unsigned src_alignment
)
1224 LLVMBuilderRef builder
= gallivm
->builder
;
1225 unsigned row_size
= src_count
/ block_height
;
1228 /* Ensure src exactly fits into block */
1229 assert((block_width
* block_height
) % src_count
== 0);
1231 for (i
= 0; i
< src_count
; ++i
) {
1232 unsigned x
= i
% row_size
;
1233 unsigned y
= i
/ row_size
;
1235 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1236 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1238 LLVMValueRef gep
[2];
1239 LLVMValueRef src_ptr
;
1241 gep
[0] = lp_build_const_int32(gallivm
, 0);
1242 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1244 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1245 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1246 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1248 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1250 LLVMSetAlignment(src_ptr
, src_alignment
);
1256 * Checks if a format description is an arithmetic format
1258 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1260 static inline boolean
1261 is_arithmetic_format(const struct util_format_description
*format_desc
)
1263 boolean arith
= false;
1266 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1267 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1268 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1276 * Checks if this format requires special handling due to required expansion
1277 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1280 static inline boolean
1281 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1283 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1284 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1292 * Retrieves the type representing the memory layout for a format
1294 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1297 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1298 struct lp_type
* type
)
1303 if (format_expands_to_float_soa(format_desc
)) {
1304 /* just make this a uint with width of block */
1305 type
->floating
= false;
1306 type
->fixed
= false;
1309 type
->width
= format_desc
->block
.bits
;
1314 for (i
= 0; i
< 4; i
++)
1315 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1319 memset(type
, 0, sizeof(struct lp_type
));
1320 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1321 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1322 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1323 type
->norm
= format_desc
->channel
[chan
].normalized
;
1325 if (is_arithmetic_format(format_desc
)) {
1329 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1330 type
->width
+= format_desc
->channel
[i
].size
;
1333 type
->width
= format_desc
->channel
[chan
].size
;
1334 type
->length
= format_desc
->nr_channels
;
1340 * Retrieves the type for a format which is usable in the blending code.
1342 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1345 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1346 struct lp_type
* type
)
1351 if (format_expands_to_float_soa(format_desc
)) {
1352 /* always use ordinary floats for blending */
1353 type
->floating
= true;
1354 type
->fixed
= false;
1362 for (i
= 0; i
< 4; i
++)
1363 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1367 memset(type
, 0, sizeof(struct lp_type
));
1368 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1369 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1370 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1371 type
->norm
= format_desc
->channel
[chan
].normalized
;
1372 type
->width
= format_desc
->channel
[chan
].size
;
1373 type
->length
= format_desc
->nr_channels
;
1375 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1376 if (format_desc
->channel
[i
].size
> type
->width
)
1377 type
->width
= format_desc
->channel
[i
].size
;
1380 if (type
->floating
) {
1383 if (type
->width
<= 8) {
1385 } else if (type
->width
<= 16) {
1392 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1399 * Scale a normalized value from src_bits to dst_bits.
1401 * The exact calculation is
1403 * dst = iround(src * dst_mask / src_mask)
1405 * or with integer rounding
1407 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1411 * src_mask = (1 << src_bits) - 1
1412 * dst_mask = (1 << dst_bits) - 1
1414 * but we try to avoid division and multiplication through shifts.
1416 static inline LLVMValueRef
1417 scale_bits(struct gallivm_state
*gallivm
,
1421 struct lp_type src_type
)
1423 LLVMBuilderRef builder
= gallivm
->builder
;
1424 LLVMValueRef result
= src
;
1426 if (dst_bits
< src_bits
) {
1427 int delta_bits
= src_bits
- dst_bits
;
1429 if (delta_bits
<= dst_bits
) {
1431 * Approximate the rescaling with a single shift.
1433 * This gives the wrong rounding.
1436 result
= LLVMBuildLShr(builder
,
1438 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1443 * Try more accurate rescaling.
1447 * Drop the least significant bits to make space for the multiplication.
1449 * XXX: A better approach would be to use a wider integer type as intermediate. But
1450 * this is enough to convert alpha from 16bits -> 2 when rendering to
1451 * PIPE_FORMAT_R10G10B10A2_UNORM.
1453 result
= LLVMBuildLShr(builder
,
1455 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1459 result
= LLVMBuildMul(builder
,
1461 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1465 * Add a rounding term before the division.
1467 * TODO: Handle signed integers too.
1469 if (!src_type
.sign
) {
1470 result
= LLVMBuildAdd(builder
,
1472 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1477 * Approximate the division by src_mask with a src_bits shift.
1479 * Given the src has already been shifted by dst_bits, all we need
1480 * to do is to shift by the difference.
1483 result
= LLVMBuildLShr(builder
,
1485 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1489 } else if (dst_bits
> src_bits
) {
1491 int db
= dst_bits
- src_bits
;
1493 /* Shift left by difference in bits */
1494 result
= LLVMBuildShl(builder
,
1496 lp_build_const_int_vec(gallivm
, src_type
, db
),
1499 if (db
<= src_bits
) {
1500 /* Enough bits in src to fill the remainder */
1501 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1503 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1506 result
= LLVMBuildOr(builder
, result
, lower
, "");
1507 } else if (db
> src_bits
) {
1508 /* Need to repeatedly copy src bits to fill remainder in dst */
1511 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1512 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1514 result
= LLVMBuildOr(builder
,
1516 LLVMBuildLShr(builder
, result
, shuv
, ""),
1526 * If RT is a smallfloat (needing denorms) format
1529 have_smallfloat_format(struct lp_type dst_type
,
1530 enum pipe_format format
)
1532 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1533 /* due to format handling hacks this format doesn't have floating set
1534 * here (and actually has width set to 32 too) so special case this. */
1535 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1540 * Convert from memory format to blending format
1542 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1545 convert_to_blend_type(struct gallivm_state
*gallivm
,
1546 unsigned block_size
,
1547 const struct util_format_description
*src_fmt
,
1548 struct lp_type src_type
,
1549 struct lp_type dst_type
,
1550 LLVMValueRef
* src
, // and dst
1553 LLVMValueRef
*dst
= src
;
1554 LLVMBuilderRef builder
= gallivm
->builder
;
1555 struct lp_type blend_type
;
1556 struct lp_type mem_type
;
1558 unsigned pixels
= block_size
/ num_srcs
;
1562 * full custom path for packed floats and srgb formats - none of the later
1563 * functions would do anything useful, and given the lp_type representation they
1564 * can't be fixed. Should really have some SoA blend path for these kind of
1565 * formats rather than hacking them in here.
1567 if (format_expands_to_float_soa(src_fmt
)) {
1568 LLVMValueRef tmpsrc
[4];
1570 * This is pretty suboptimal for this case blending in SoA would be much
1571 * better, since conversion gets us SoA values so need to convert back.
1573 assert(src_type
.width
== 32 || src_type
.width
== 16);
1574 assert(dst_type
.floating
);
1575 assert(dst_type
.width
== 32);
1576 assert(dst_type
.length
% 4 == 0);
1577 assert(num_srcs
% 4 == 0);
1579 if (src_type
.width
== 16) {
1580 /* expand 4x16bit values to 4x32bit */
1581 struct lp_type type32x4
= src_type
;
1582 LLVMTypeRef ltype32x4
;
1583 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1584 type32x4
.width
= 32;
1585 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1586 for (i
= 0; i
< num_fetch
; i
++) {
1587 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1589 src_type
.width
= 32;
1591 for (i
= 0; i
< 4; i
++) {
1594 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1595 LLVMValueRef tmpsoa
[4];
1596 LLVMValueRef tmps
= tmpsrc
[i
];
1597 if (dst_type
.length
== 8) {
1598 LLVMValueRef shuffles
[8];
1600 /* fetch was 4 values but need 8-wide output values */
1601 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1603 * for 8-wide aos transpose would give us wrong order not matching
1604 * incoming converted fs values and mask. ARGH.
1606 for (j
= 0; j
< 4; j
++) {
1607 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1608 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1610 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1611 LLVMConstVector(shuffles
, 8), "");
1613 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1614 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1617 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1619 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1624 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1625 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1627 /* Is the format arithmetic */
1628 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1629 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1631 /* Pad if necessary */
1632 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1633 for (i
= 0; i
< num_srcs
; ++i
) {
1634 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1637 src_type
.length
= dst_type
.length
;
1640 /* Special case for half-floats */
1641 if (mem_type
.width
== 16 && mem_type
.floating
) {
1642 assert(blend_type
.width
== 32 && blend_type
.floating
);
1643 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1651 src_type
.width
= blend_type
.width
* blend_type
.length
;
1652 blend_type
.length
*= pixels
;
1653 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1655 for (i
= 0; i
< num_srcs
; ++i
) {
1656 LLVMValueRef chans
[4];
1657 LLVMValueRef res
= NULL
;
1659 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1661 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1663 unsigned sa
= src_fmt
->channel
[j
].shift
;
1664 #if UTIL_ARCH_LITTLE_ENDIAN
1665 unsigned from_lsb
= j
;
1667 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1670 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1672 /* Extract bits from source */
1673 chans
[j
] = LLVMBuildLShr(builder
,
1675 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1678 chans
[j
] = LLVMBuildAnd(builder
,
1680 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1684 if (src_type
.norm
) {
1685 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1686 blend_type
.width
, chans
[j
], src_type
);
1689 /* Insert bits into correct position */
1690 chans
[j
] = LLVMBuildShl(builder
,
1692 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1698 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1702 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1708 * Convert from blending format to memory format
1710 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1713 convert_from_blend_type(struct gallivm_state
*gallivm
,
1714 unsigned block_size
,
1715 const struct util_format_description
*src_fmt
,
1716 struct lp_type src_type
,
1717 struct lp_type dst_type
,
1718 LLVMValueRef
* src
, // and dst
1721 LLVMValueRef
* dst
= src
;
1723 struct lp_type mem_type
;
1724 struct lp_type blend_type
;
1725 LLVMBuilderRef builder
= gallivm
->builder
;
1726 unsigned pixels
= block_size
/ num_srcs
;
1730 * full custom path for packed floats and srgb formats - none of the later
1731 * functions would do anything useful, and given the lp_type representation they
1732 * can't be fixed. Should really have some SoA blend path for these kind of
1733 * formats rather than hacking them in here.
1735 if (format_expands_to_float_soa(src_fmt
)) {
1737 * This is pretty suboptimal for this case blending in SoA would be much
1738 * better - we need to transpose the AoS values back to SoA values for
1739 * conversion/packing.
1741 assert(src_type
.floating
);
1742 assert(src_type
.width
== 32);
1743 assert(src_type
.length
% 4 == 0);
1744 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1746 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1747 LLVMValueRef tmpsoa
[4], tmpdst
;
1748 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1749 /* really really need SoA here */
1751 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1752 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1755 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1759 if (src_type
.length
== 8) {
1760 LLVMValueRef tmpaos
, shuffles
[8];
1763 * for 8-wide aos transpose has given us wrong order not matching
1764 * output order. HMPF. Also need to split the output values manually.
1766 for (j
= 0; j
< 4; j
++) {
1767 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1768 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1770 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1771 LLVMConstVector(shuffles
, 8), "");
1772 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1773 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1779 if (dst_type
.width
== 16) {
1780 struct lp_type type16x8
= dst_type
;
1781 struct lp_type type32x4
= dst_type
;
1782 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1783 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1784 type16x8
.length
= 8;
1785 type32x4
.width
= 32;
1786 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1787 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1788 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1789 /* We could do vector truncation but it doesn't generate very good code */
1790 for (i
= 0; i
< num_fetch
; i
++) {
1791 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1792 src
[i
], lp_build_zero(gallivm
, type32x4
));
1793 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1794 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1795 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1801 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1802 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1804 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1806 /* Special case for half-floats */
1807 if (mem_type
.width
== 16 && mem_type
.floating
) {
1808 int length
= dst_type
.length
;
1809 assert(blend_type
.width
== 32 && blend_type
.floating
);
1811 dst_type
.length
= src_type
.length
;
1813 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1815 dst_type
.length
= length
;
1819 /* Remove any padding */
1820 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1821 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1823 for (i
= 0; i
< num_srcs
; ++i
) {
1824 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1828 /* No bit arithmetic to do */
1833 src_type
.length
= pixels
;
1834 src_type
.width
= blend_type
.length
* blend_type
.width
;
1835 dst_type
.length
= pixels
;
1837 for (i
= 0; i
< num_srcs
; ++i
) {
1838 LLVMValueRef chans
[4];
1839 LLVMValueRef res
= NULL
;
1841 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1843 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1845 unsigned sa
= src_fmt
->channel
[j
].shift
;
1846 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1847 #if UTIL_ARCH_LITTLE_ENDIAN
1848 unsigned from_lsb
= j
;
1850 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1853 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1855 for (k
= 0; k
< blend_type
.width
; ++k
) {
1860 chans
[j
] = LLVMBuildLShr(builder
,
1862 lp_build_const_int_vec(gallivm
, src_type
,
1863 from_lsb
* blend_type
.width
),
1866 chans
[j
] = LLVMBuildAnd(builder
,
1868 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1871 /* Scale down bits */
1872 if (src_type
.norm
) {
1873 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1874 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1875 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1876 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1877 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1878 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1882 chans
[j
] = LLVMBuildShl(builder
,
1884 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1887 sa
+= src_fmt
->channel
[j
].size
;
1892 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1896 assert (dst_type
.width
!= 24);
1898 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1904 * Convert alpha to same blend type as src
1907 convert_alpha(struct gallivm_state
*gallivm
,
1908 struct lp_type row_type
,
1909 struct lp_type alpha_type
,
1910 const unsigned block_size
,
1911 const unsigned block_height
,
1912 const unsigned src_count
,
1913 const unsigned dst_channels
,
1914 const bool pad_inline
,
1915 LLVMValueRef
* src_alpha
)
1917 LLVMBuilderRef builder
= gallivm
->builder
;
1919 unsigned length
= row_type
.length
;
1920 row_type
.length
= alpha_type
.length
;
1922 /* Twiddle the alpha to match pixels */
1923 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1926 * TODO this should use single lp_build_conv call for
1927 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1929 for (i
= 0; i
< block_height
; ++i
) {
1930 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1933 alpha_type
= row_type
;
1934 row_type
.length
= length
;
1936 /* If only one channel we can only need the single alpha value per pixel */
1937 if (src_count
== 1 && dst_channels
== 1) {
1939 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1941 /* If there are more srcs than rows then we need to split alpha up */
1942 if (src_count
> block_height
) {
1943 for (i
= src_count
; i
> 0; --i
) {
1944 unsigned pixels
= block_size
/ src_count
;
1945 unsigned idx
= i
- 1;
1947 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1948 (idx
* pixels
) % 4, pixels
);
1952 /* If there is a src for each pixel broadcast the alpha across whole row */
1953 if (src_count
== block_size
) {
1954 for (i
= 0; i
< src_count
; ++i
) {
1955 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1956 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1959 unsigned pixels
= block_size
/ src_count
;
1960 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1961 unsigned alpha_span
= 1;
1962 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1964 /* Check if we need 2 src_alphas for our shuffles */
1965 if (pixels
> alpha_type
.length
) {
1969 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1970 for (j
= 0; j
< row_type
.length
; ++j
) {
1971 if (j
< pixels
* channels
) {
1972 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1974 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1978 for (i
= 0; i
< src_count
; ++i
) {
1979 unsigned idx1
= i
, idx2
= i
;
1981 if (alpha_span
> 1){
1986 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1989 LLVMConstVector(shuffles
, row_type
.length
),
1998 * Generates the blend function for unswizzled colour buffers
1999 * Also generates the read & write from colour buffer
2002 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
2004 struct lp_fragment_shader_variant
*variant
,
2005 enum pipe_format out_format
,
2006 unsigned int num_fs
,
2007 struct lp_type fs_type
,
2008 LLVMValueRef
* fs_mask
,
2009 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
2010 LLVMValueRef context_ptr
,
2011 LLVMValueRef color_ptr
,
2012 LLVMValueRef stride
,
2013 unsigned partial_mask
,
2016 const unsigned alpha_channel
= 3;
2017 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
2018 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
2019 const unsigned block_size
= block_width
* block_height
;
2020 const unsigned lp_integer_vector_width
= 128;
2022 LLVMBuilderRef builder
= gallivm
->builder
;
2023 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
2024 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
2025 LLVMValueRef src_alpha
[4 * 4];
2026 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
2027 LLVMValueRef src_mask
[4 * 4];
2028 LLVMValueRef src
[4 * 4];
2029 LLVMValueRef src1
[4 * 4];
2030 LLVMValueRef dst
[4 * 4];
2031 LLVMValueRef blend_color
;
2032 LLVMValueRef blend_alpha
;
2033 LLVMValueRef i32_zero
;
2034 LLVMValueRef check_mask
;
2035 LLVMValueRef undef_src_val
;
2037 struct lp_build_mask_context mask_ctx
;
2038 struct lp_type mask_type
;
2039 struct lp_type blend_type
;
2040 struct lp_type row_type
;
2041 struct lp_type dst_type
;
2042 struct lp_type ls_type
;
2044 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
2045 unsigned vector_width
;
2046 unsigned src_channels
= TGSI_NUM_CHANNELS
;
2047 unsigned dst_channels
;
2052 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
2054 unsigned dst_alignment
;
2056 bool pad_inline
= is_arithmetic_format(out_format_desc
);
2057 bool has_alpha
= false;
2058 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
2059 util_blend_state_is_dual(&variant
->key
.blend
, 0);
2061 const boolean is_1d
= variant
->key
.resource_1d
;
2062 boolean twiddle_after_convert
= FALSE
;
2063 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
2064 LLVMValueRef fpstate
= 0;
2066 /* Get type from output format */
2067 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
2068 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
2071 * Technically this code should go into lp_build_smallfloat_to_float
2072 * and lp_build_float_to_smallfloat but due to the
2073 * http://llvm.org/bugs/show_bug.cgi?id=6393
2074 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2075 * So the ordering is important here and there shouldn't be any
2076 * llvm ir instrunctions in this function before
2077 * this, otherwise half-float format conversions won't work
2078 * (again due to llvm bug #6393).
2080 if (have_smallfloat_format(dst_type
, out_format
)) {
2081 /* We need to make sure that denorms are ok for half float
2083 fpstate
= lp_build_fpstate_get(gallivm
);
2084 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
2087 mask_type
= lp_int32_vec4_type();
2088 mask_type
.length
= fs_type
.length
;
2090 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
2091 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
2094 /* Do not bother executing code when mask is empty.. */
2096 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
2098 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2099 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
2102 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
2103 lp_build_mask_check(&mask_ctx
);
2106 partial_mask
|= !variant
->opaque
;
2107 i32_zero
= lp_build_const_int32(gallivm
, 0);
2109 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2111 row_type
.length
= fs_type
.length
;
2112 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2114 /* Compute correct swizzle and count channels */
2115 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2118 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2119 /* Ensure channel is used */
2120 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2124 /* Ensure not already written to (happens in case with GL_ALPHA) */
2125 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2129 /* Ensure we havn't already found all channels */
2130 if (dst_channels
>= out_format_desc
->nr_channels
) {
2134 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2137 if (i
== alpha_channel
) {
2142 if (format_expands_to_float_soa(out_format_desc
)) {
2144 * the code above can't work for layout_other
2145 * for srgb it would sort of work but we short-circuit swizzles, etc.
2146 * as that is done as part of unpack / pack.
2148 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2154 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2157 /* If 3 channels then pad to include alpha for 4 element transpose */
2158 if (dst_channels
== 3) {
2159 assert (!has_alpha
);
2160 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2161 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2164 if (out_format_desc
->nr_channels
== 4) {
2167 * We use alpha from the color conversion, not separate one.
2168 * We had to include it for transpose, hence it will get converted
2169 * too (albeit when doing transpose after conversion, that would
2170 * no longer be the case necessarily).
2171 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2172 * otherwise we really have padding, not alpha, included.)
2179 * Load shader output
2181 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2182 /* Always load alpha for use in blending */
2185 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2188 alpha
= undef_src_val
;
2191 /* Load each channel */
2192 for (j
= 0; j
< dst_channels
; ++j
) {
2193 assert(swizzle
[j
] < 4);
2195 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2198 fs_src
[i
][j
] = undef_src_val
;
2202 /* If 3 channels then pad to include alpha for 4 element transpose */
2204 * XXX If we include that here maybe could actually use it instead of
2205 * separate alpha for blending?
2206 * (Difficult though we actually convert pad channels, not alpha.)
2208 if (dst_channels
== 3 && !has_alpha
) {
2209 fs_src
[i
][3] = alpha
;
2212 /* We split the row_mask and row_alpha as we want 128bit interleave */
2213 if (fs_type
.length
== 8) {
2214 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2216 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2217 src_channels
, src_channels
);
2219 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2220 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2221 src_channels
, src_channels
);
2223 src_mask
[i
] = fs_mask
[i
];
2224 src_alpha
[i
] = alpha
;
2227 if (dual_source_blend
) {
2228 /* same as above except different src/dst, skip masks and comments... */
2229 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2232 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2235 alpha
= undef_src_val
;
2238 for (j
= 0; j
< dst_channels
; ++j
) {
2239 assert(swizzle
[j
] < 4);
2241 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2244 fs_src1
[i
][j
] = undef_src_val
;
2247 if (dst_channels
== 3 && !has_alpha
) {
2248 fs_src1
[i
][3] = alpha
;
2250 if (fs_type
.length
== 8) {
2251 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2252 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2253 src_channels
, src_channels
);
2255 src1_alpha
[i
] = alpha
;
2260 if (util_format_is_pure_integer(out_format
)) {
2262 * In this case fs_type was really ints or uints disguised as floats,
2265 fs_type
.floating
= 0;
2266 fs_type
.sign
= dst_type
.sign
;
2267 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2268 for (j
= 0; j
< dst_channels
; ++j
) {
2269 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2270 lp_build_vec_type(gallivm
, fs_type
), "");
2272 if (dst_channels
== 3 && !has_alpha
) {
2273 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2274 lp_build_vec_type(gallivm
, fs_type
), "");
2280 * We actually should generally do conversion first (for non-1d cases)
2281 * when the blend format is 8 or 16 bits. The reason is obvious,
2282 * there's 2 or 4 times less vectors to deal with for the interleave...
2283 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2284 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2285 * unpack only with 128bit vectors).
2286 * Note: for 16bit sizes really need matching pack conversion code
2288 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2289 twiddle_after_convert
= TRUE
;
2293 * Pixel twiddle from fragment shader order to memory order
2295 if (!twiddle_after_convert
) {
2296 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2297 dst_channels
, fs_src
, src
, pad_inline
);
2298 if (dual_source_blend
) {
2299 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2300 fs_src1
, src1
, pad_inline
);
2303 src_count
= num_fullblock_fs
* dst_channels
;
2305 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2306 * (AVX) turn out the same later when untwiddling/transpose (albeit
2307 * for true AVX2 path untwiddle needs to be different).
2308 * For now just order by colors first (so we can use unpack later).
2310 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2311 for (i
= 0; i
< dst_channels
; i
++) {
2312 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2313 if (dual_source_blend
) {
2314 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2320 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2321 if (src_count
!= num_fullblock_fs
* src_channels
) {
2322 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2323 row_type
.length
/= ds
;
2324 fs_type
.length
= row_type
.length
;
2327 blend_type
= row_type
;
2328 mask_type
.length
= 4;
2330 /* Convert src to row_type */
2331 if (dual_source_blend
) {
2332 struct lp_type old_row_type
= row_type
;
2333 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2334 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2337 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2340 /* If the rows are not an SSE vector, combine them to become SSE size! */
2341 if ((row_type
.width
* row_type
.length
) % 128) {
2342 unsigned bits
= row_type
.width
* row_type
.length
;
2345 assert(src_count
>= (vector_width
/ bits
));
2347 dst_count
= src_count
/ (vector_width
/ bits
);
2349 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2350 if (dual_source_blend
) {
2351 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2354 row_type
.length
*= combined
;
2355 src_count
/= combined
;
2357 bits
= row_type
.width
* row_type
.length
;
2358 assert(bits
== 128 || bits
== 256);
2361 if (twiddle_after_convert
) {
2362 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2363 if (dual_source_blend
) {
2364 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2369 * Blend Colour conversion
2371 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2372 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2373 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2374 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2375 &i32_zero
, 1, ""), "");
2378 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2380 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2382 * since blending is done with floats, there was no conversion.
2383 * However, the rules according to fixed point renderbuffers still
2384 * apply, that is we must clamp inputs to 0.0/1.0.
2385 * (This would apply to separate alpha conversion too but we currently
2386 * force has_alpha to be true.)
2387 * TODO: should skip this with "fake" blend, since post-blend conversion
2388 * will clamp anyway.
2389 * TODO: could also skip this if fragment color clamping is enabled. We
2390 * don't support it natively so it gets baked into the shader however, so
2391 * can't really tell here.
2393 struct lp_build_context f32_bld
;
2394 assert(row_type
.floating
);
2395 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2396 for (i
= 0; i
< src_count
; i
++) {
2397 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2399 if (dual_source_blend
) {
2400 for (i
= 0; i
< src_count
; i
++) {
2401 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2404 /* probably can't be different than row_type but better safe than sorry... */
2405 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2406 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2410 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2412 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2413 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2415 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2416 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2418 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2419 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2425 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2427 if (src_count
< block_height
) {
2428 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2429 } else if (src_count
> block_height
) {
2430 for (i
= src_count
; i
> 0; --i
) {
2431 unsigned pixels
= block_size
/ src_count
;
2432 unsigned idx
= i
- 1;
2434 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2435 (idx
* pixels
) % 4, pixels
);
2439 assert(mask_type
.width
== 32);
2441 for (i
= 0; i
< src_count
; ++i
) {
2442 unsigned pixels
= block_size
/ src_count
;
2443 unsigned pixel_width
= row_type
.width
* dst_channels
;
2445 if (pixel_width
== 24) {
2446 mask_type
.width
= 8;
2447 mask_type
.length
= vector_width
/ mask_type
.width
;
2449 mask_type
.length
= pixels
;
2450 mask_type
.width
= row_type
.width
* dst_channels
;
2453 * If mask_type width is smaller than 32bit, this doesn't quite
2454 * generate the most efficient code (could use some pack).
2456 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2457 lp_build_int_vec_type(gallivm
, mask_type
), "");
2459 mask_type
.length
*= dst_channels
;
2460 mask_type
.width
/= dst_channels
;
2463 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2464 lp_build_int_vec_type(gallivm
, mask_type
), "");
2465 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2472 struct lp_type alpha_type
= fs_type
;
2473 alpha_type
.length
= 4;
2474 convert_alpha(gallivm
, row_type
, alpha_type
,
2475 block_size
, block_height
,
2476 src_count
, dst_channels
,
2477 pad_inline
, src_alpha
);
2478 if (dual_source_blend
) {
2479 convert_alpha(gallivm
, row_type
, alpha_type
,
2480 block_size
, block_height
,
2481 src_count
, dst_channels
,
2482 pad_inline
, src1_alpha
);
2488 * Load dst from memory
2490 if (src_count
< block_height
) {
2491 dst_count
= block_height
;
2493 dst_count
= src_count
;
2496 dst_type
.length
*= block_size
/ dst_count
;
2498 if (format_expands_to_float_soa(out_format_desc
)) {
2500 * we need multiple values at once for the conversion, so can as well
2501 * load them vectorized here too instead of concatenating later.
2502 * (Still need concatenation later for 8-wide vectors).
2504 dst_count
= block_height
;
2505 dst_type
.length
= block_width
;
2509 * Compute the alignment of the destination pointer in bytes
2510 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2511 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2512 * 1d tex but can't distinguish here) so need to stick with per-pixel
2513 * alignment in this case.
2516 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2519 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2521 /* Force power-of-two alignment by extracting only the least-significant-bit */
2522 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2524 * Resource base and stride pointers are aligned to 16 bytes, so that's
2525 * the maximum alignment we can guarantee
2527 dst_alignment
= MIN2(16, dst_alignment
);
2531 if (dst_count
> src_count
) {
2532 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2533 util_is_power_of_two_or_zero(dst_type
.length
) &&
2534 dst_type
.length
* dst_type
.width
< 128) {
2536 * Never try to load values as 4xi8 which we will then
2537 * concatenate to larger vectors. This gives llvm a real
2538 * headache (the problem is the type legalizer (?) will
2539 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2540 * then the shuffles to concatenate are more or less impossible
2541 * - llvm is easily capable of generating a sequence of 32
2542 * pextrb/pinsrb instructions for that. Albeit it appears to
2543 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2544 * width to avoid the trouble (16bit seems not as bad, llvm
2545 * probably recognizes the load+shuffle as only one shuffle
2546 * is necessary, but we can do just the same anyway).
2548 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2554 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2555 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2556 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2557 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2562 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2563 dst
, ls_type
, dst_count
, dst_alignment
);
2568 * Convert from dst/output format to src/blending format.
2570 * This is necessary as we can only read 1 row from memory at a time,
2571 * so the minimum dst_count will ever be at this point is 4.
2573 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2574 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2575 * on all 16 pixels in that single vector at once.
2577 if (dst_count
> src_count
) {
2578 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2579 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2580 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2581 for (i
= 0; i
< dst_count
; i
++) {
2582 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2586 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2588 if (ls_type
.length
!= dst_type
.length
) {
2589 struct lp_type tmp_type
= dst_type
;
2590 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2591 for (i
= 0; i
< src_count
; i
++) {
2592 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2593 lp_build_vec_type(gallivm
, tmp_type
), "");
2601 /* XXX this is broken for RGB8 formats -
2602 * they get expanded from 12 to 16 elements (to include alpha)
2603 * by convert_to_blend_type then reduced to 15 instead of 12
2604 * by convert_from_blend_type (a simple fix though breaks A8...).
2605 * R16G16B16 also crashes differently however something going wrong
2606 * inside llvm handling npot vector sizes seemingly.
2607 * It seems some cleanup could be done here (like skipping conversion/blend
2610 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2611 row_type
, dst
, src_count
);
2614 * FIXME: Really should get logic ops / masks out of generic blend / row
2615 * format. Logic ops will definitely not work on the blend float format
2616 * used for SRGB here and I think OpenGL expects this to work as expected
2617 * (that is incoming values converted to srgb then logic op applied).
2619 for (i
= 0; i
< src_count
; ++i
) {
2620 dst
[i
] = lp_build_blend_aos(gallivm
,
2621 &variant
->key
.blend
,
2626 has_alpha
? NULL
: src_alpha
[i
],
2628 has_alpha
? NULL
: src1_alpha
[i
],
2630 partial_mask
? src_mask
[i
] : NULL
,
2632 has_alpha
? NULL
: blend_alpha
,
2634 pad_inline
? 4 : dst_channels
);
2637 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2638 row_type
, dst_type
, dst
, src_count
);
2640 /* Split the blend rows back to memory rows */
2641 if (dst_count
> src_count
) {
2642 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2644 if (src_count
== 1) {
2645 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2646 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2648 row_type
.length
/= 2;
2652 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2653 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2654 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2655 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2657 row_type
.length
/= 2;
2662 * Store blend result to memory
2665 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2666 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2669 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2670 dst
, dst_type
, dst_count
, dst_alignment
);
2673 if (have_smallfloat_format(dst_type
, out_format
)) {
2674 lp_build_fpstate_set(gallivm
, fpstate
);
2678 lp_build_mask_end(&mask_ctx
);
2684 * Generate the runtime callable function for the whole fragment pipeline.
2685 * Note that the function which we generate operates on a block of 16
2686 * pixels at at time. The block contains 2x2 quads. Each quad contains
2690 generate_fragment(struct llvmpipe_context
*lp
,
2691 struct lp_fragment_shader
*shader
,
2692 struct lp_fragment_shader_variant
*variant
,
2693 unsigned partial_mask
)
2695 struct gallivm_state
*gallivm
= variant
->gallivm
;
2696 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2697 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2699 struct lp_type fs_type
;
2700 struct lp_type blend_type
;
2701 LLVMTypeRef fs_elem_type
;
2702 LLVMTypeRef blend_vec_type
;
2703 LLVMTypeRef arg_types
[15];
2704 LLVMTypeRef func_type
;
2705 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2706 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2707 LLVMValueRef context_ptr
;
2710 LLVMValueRef a0_ptr
;
2711 LLVMValueRef dadx_ptr
;
2712 LLVMValueRef dady_ptr
;
2713 LLVMValueRef color_ptr_ptr
;
2714 LLVMValueRef stride_ptr
;
2715 LLVMValueRef color_sample_stride_ptr
;
2716 LLVMValueRef depth_ptr
;
2717 LLVMValueRef depth_stride
;
2718 LLVMValueRef depth_sample_stride
;
2719 LLVMValueRef mask_input
;
2720 LLVMValueRef thread_data_ptr
;
2721 LLVMBasicBlockRef block
;
2722 LLVMBuilderRef builder
;
2723 struct lp_build_sampler_soa
*sampler
;
2724 struct lp_build_image_soa
*image
;
2725 struct lp_build_interp_soa_context interp
;
2726 LLVMValueRef fs_mask
[(16 / 4) * LP_MAX_SAMPLES
];
2727 LLVMValueRef fs_out_color
[LP_MAX_SAMPLES
][PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2728 LLVMValueRef function
;
2729 LLVMValueRef facing
;
2734 boolean cbuf0_write_all
;
2735 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2736 util_blend_state_is_dual(&key
->blend
, 0);
2738 assert(lp_native_vector_width
/ 32 >= 4);
2740 /* Adjust color input interpolation according to flatshade state:
2742 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2743 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2744 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2746 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2748 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2752 /* check if writes to cbuf[0] are to be copied to all cbufs */
2754 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2756 /* TODO: actually pick these based on the fs and color buffer
2757 * characteristics. */
2759 memset(&fs_type
, 0, sizeof fs_type
);
2760 fs_type
.floating
= TRUE
; /* floating point values */
2761 fs_type
.sign
= TRUE
; /* values are signed */
2762 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2763 fs_type
.width
= 32; /* 32-bit float */
2764 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2766 memset(&blend_type
, 0, sizeof blend_type
);
2767 blend_type
.floating
= FALSE
; /* values are integers */
2768 blend_type
.sign
= FALSE
; /* values are unsigned */
2769 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2770 blend_type
.width
= 8; /* 8-bit ubyte values */
2771 blend_type
.length
= 16; /* 16 elements per vector */
2774 * Generate the function prototype. Any change here must be reflected in
2775 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2778 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2780 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2782 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2783 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2785 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2786 arg_types
[1] = int32_type
; /* x */
2787 arg_types
[2] = int32_type
; /* y */
2788 arg_types
[3] = int32_type
; /* facing */
2789 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2790 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2791 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2792 arg_types
[7] = LLVMPointerType(LLVMPointerType(int8_type
, 0), 0); /* color */
2793 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2794 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2795 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2796 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2797 arg_types
[12] = int32_type
; /* depth_stride */
2798 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2799 arg_types
[14] = int32_type
; /* depth sample stride */
2801 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2802 arg_types
, ARRAY_SIZE(arg_types
), 0);
2804 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2805 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2807 variant
->function
[partial_mask
] = function
;
2809 /* XXX: need to propagate noalias down into color param now we are
2810 * passing a pointer-to-pointer?
2812 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2813 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2814 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2816 context_ptr
= LLVMGetParam(function
, 0);
2817 x
= LLVMGetParam(function
, 1);
2818 y
= LLVMGetParam(function
, 2);
2819 facing
= LLVMGetParam(function
, 3);
2820 a0_ptr
= LLVMGetParam(function
, 4);
2821 dadx_ptr
= LLVMGetParam(function
, 5);
2822 dady_ptr
= LLVMGetParam(function
, 6);
2823 color_ptr_ptr
= LLVMGetParam(function
, 7);
2824 depth_ptr
= LLVMGetParam(function
, 8);
2825 mask_input
= LLVMGetParam(function
, 9);
2826 thread_data_ptr
= LLVMGetParam(function
, 10);
2827 stride_ptr
= LLVMGetParam(function
, 11);
2828 depth_stride
= LLVMGetParam(function
, 12);
2829 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2830 depth_sample_stride
= LLVMGetParam(function
, 14);
2832 lp_build_name(context_ptr
, "context");
2833 lp_build_name(x
, "x");
2834 lp_build_name(y
, "y");
2835 lp_build_name(a0_ptr
, "a0");
2836 lp_build_name(dadx_ptr
, "dadx");
2837 lp_build_name(dady_ptr
, "dady");
2838 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2839 lp_build_name(depth_ptr
, "depth");
2840 lp_build_name(mask_input
, "mask_input");
2841 lp_build_name(thread_data_ptr
, "thread_data");
2842 lp_build_name(stride_ptr
, "stride_ptr");
2843 lp_build_name(depth_stride
, "depth_stride");
2844 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2845 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2851 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2852 builder
= gallivm
->builder
;
2854 LLVMPositionBuilderAtEnd(builder
, block
);
2857 * Must not count ps invocations if there's a null shader.
2858 * (It would be ok to count with null shader if there's d/s tests,
2859 * but only if there's d/s buffers too, which is different
2860 * to implicit rasterization disable which must not depend
2861 * on the d/s buffers.)
2862 * Could use popcount on mask, but pixel accuracy is not required.
2863 * Could disable if there's no stats query, but maybe not worth it.
2865 if (shader
->info
.base
.num_instructions
> 1) {
2866 LLVMValueRef invocs
, val
;
2867 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2868 val
= LLVMBuildLoad(builder
, invocs
, "");
2869 val
= LLVMBuildAdd(builder
, val
,
2870 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2872 LLVMBuildStore(builder
, val
, invocs
);
2875 /* code generated texture sampling */
2876 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2877 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2879 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2880 /* for 1d resources only run "upper half" of stamp */
2881 if (key
->resource_1d
)
2885 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2886 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2887 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2888 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2889 num_loop_samp
, "mask_store");
2891 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(gallivm
->context
);
2892 LLVMValueRef glob_sample_pos
= LLVMAddGlobal(gallivm
->module
, flt_type
, "");
2893 LLVMValueRef sample_pos_array
;
2895 if (key
->multisample
&& key
->coverage_samples
== 4) {
2896 LLVMValueRef sample_pos_arr
[8];
2897 for (unsigned i
= 0; i
< 4; i
++) {
2898 sample_pos_arr
[i
* 2] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][0]);
2899 sample_pos_arr
[i
* 2 + 1] = LLVMConstReal(flt_type
, lp_sample_pos_4x
[i
][1]);
2901 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 8);
2903 LLVMValueRef sample_pos_arr
[2];
2904 sample_pos_arr
[0] = LLVMConstReal(flt_type
, 0.5);
2905 sample_pos_arr
[1] = LLVMConstReal(flt_type
, 0.5);
2906 sample_pos_array
= LLVMConstArray(LLVMFloatTypeInContext(gallivm
->context
), sample_pos_arr
, 2);
2908 LLVMSetInitializer(glob_sample_pos
, sample_pos_array
);
2910 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2911 boolean pixel_center_integer
=
2912 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2915 * The shader input interpolation info is not explicitely baked in the
2916 * shader key, but everything it derives from (TGSI, and flatshade) is
2917 * already included in the shader key.
2919 lp_build_interp_soa_init(&interp
,
2921 shader
->info
.base
.num_inputs
,
2923 pixel_center_integer
,
2924 key
->coverage_samples
, glob_sample_pos
,
2928 a0_ptr
, dadx_ptr
, dady_ptr
,
2931 for (i
= 0; i
< num_fs
; i
++) {
2932 if (key
->multisample
) {
2933 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2936 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2937 * store to the per sample mask storage. Or all of them together to generate
2938 * the fragment shader mask. (sample shading TODO).
2939 * Take the incoming state coverage mask into account.
2941 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2942 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2943 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2944 &sindexi
, 1, "sample_mask_ptr");
2945 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2946 i
*fs_type
.length
/4, s
, mask_input
);
2948 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2949 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2950 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2951 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2953 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2954 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
2958 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2959 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2960 &indexi
, 1, "mask_ptr");
2963 mask
= generate_quad_mask(gallivm
, fs_type
,
2964 i
*fs_type
.length
/4, 0, mask_input
);
2967 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2969 LLVMBuildStore(builder
, mask
, mask_ptr
);
2973 generate_fs_loop(gallivm
,
2983 mask_store
, /* output */
2987 depth_sample_stride
,
2991 for (i
= 0; i
< num_fs
; i
++) {
2993 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2994 int idx
= (i
+ (s
* num_fs
));
2995 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
2996 ptr
= LLVMBuildGEP(builder
, mask_store
, &sindexi
, 1, "");
2998 fs_mask
[idx
] = LLVMBuildLoad(builder
, ptr
, "smask");
3001 for (unsigned s
= 0; s
< key
->min_samples
; s
++) {
3002 /* This is fucked up need to reorganize things */
3003 int idx
= s
* num_fs
+ i
;
3004 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, idx
);
3005 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3006 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3007 ptr
= LLVMBuildGEP(builder
,
3008 color_store
[cbuf
* !cbuf0_write_all
][chan
],
3010 fs_out_color
[s
][cbuf
][chan
][i
] = ptr
;
3013 if (dual_source_blend
) {
3014 /* only support one dual source blend target hence always use output 1 */
3015 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3016 ptr
= LLVMBuildGEP(builder
,
3017 color_store
[1][chan
],
3019 fs_out_color
[s
][1][chan
][i
] = ptr
;
3026 sampler
->destroy(sampler
);
3027 image
->destroy(image
);
3028 /* Loop over color outputs / color buffers to do blending.
3030 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
3031 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
3032 LLVMValueRef color_ptr
;
3033 LLVMValueRef stride
;
3034 LLVMValueRef sample_stride
= NULL
;
3035 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
3037 boolean do_branch
= ((key
->depth
.enabled
3038 || key
->stencil
[0].enabled
3039 || key
->alpha
.enabled
)
3040 && !shader
->info
.base
.uses_kill
);
3042 color_ptr
= LLVMBuildLoad(builder
,
3043 LLVMBuildGEP(builder
, color_ptr_ptr
,
3047 stride
= LLVMBuildLoad(builder
,
3048 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
3051 if (key
->multisample
)
3052 sample_stride
= LLVMBuildLoad(builder
,
3053 LLVMBuildGEP(builder
, color_sample_stride_ptr
,
3054 &index
, 1, ""), "");
3056 for (unsigned s
= 0; s
< key
->cbuf_nr_samples
[cbuf
]; s
++) {
3057 unsigned mask_idx
= num_fs
* (key
->multisample
? s
: 0);
3058 unsigned out_idx
= key
->min_samples
== 1 ? 0 : s
;
3059 LLVMValueRef out_ptr
= color_ptr
;;
3061 if (key
->multisample
) {
3062 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_stride
, lp_build_const_int32(gallivm
, s
), "");
3063 out_ptr
= LLVMBuildGEP(builder
, out_ptr
, &sample_offset
, 1, "");
3065 out_ptr
= LLVMBuildBitCast(builder
, out_ptr
, LLVMPointerType(blend_vec_type
, 0), "");
3067 lp_build_name(out_ptr
, "color_ptr%d", cbuf
);
3069 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
3070 key
->cbuf_format
[cbuf
],
3071 num_fs
, fs_type
, &fs_mask
[mask_idx
], fs_out_color
[out_idx
],
3072 context_ptr
, out_ptr
, stride
,
3073 partial_mask
, do_branch
);
3078 LLVMBuildRetVoid(builder
);
3080 gallivm_verify_function(gallivm
, function
);
3085 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
3089 debug_printf("fs variant %p:\n", (void *) key
);
3091 if (key
->flatshade
) {
3092 debug_printf("flatshade = 1\n");
3094 if (key
->multisample
) {
3095 debug_printf("multisample = 1\n");
3096 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
3097 debug_printf("min samples = %d\n", key
->min_samples
);
3099 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
3100 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
3101 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
3103 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
3104 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
3105 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
3107 if (key
->depth
.enabled
) {
3108 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
3109 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
3112 for (i
= 0; i
< 2; ++i
) {
3113 if (key
->stencil
[i
].enabled
) {
3114 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
3115 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
3116 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
3117 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
3118 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
3119 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
3123 if (key
->alpha
.enabled
) {
3124 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
3127 if (key
->occlusion_count
) {
3128 debug_printf("occlusion_count = 1\n");
3131 if (key
->blend
.logicop_enable
) {
3132 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
3134 else if (key
->blend
.rt
[0].blend_enable
) {
3135 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
3136 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
3137 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
3138 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
3139 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
3140 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
3142 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
3143 if (key
->blend
.alpha_to_coverage
) {
3144 debug_printf("blend.alpha_to_coverage is enabled\n");
3146 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
3147 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
3148 debug_printf("sampler[%u] = \n", i
);
3149 debug_printf(" .wrap = %s %s %s\n",
3150 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
3151 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
3152 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
3153 debug_printf(" .min_img_filter = %s\n",
3154 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
3155 debug_printf(" .min_mip_filter = %s\n",
3156 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
3157 debug_printf(" .mag_img_filter = %s\n",
3158 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
3159 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
3160 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3161 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3162 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3163 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3164 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3165 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3167 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3168 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3169 debug_printf("texture[%u] = \n", i
);
3170 debug_printf(" .format = %s\n",
3171 util_format_name(texture
->format
));
3172 debug_printf(" .target = %s\n",
3173 util_str_tex_target(texture
->target
, TRUE
));
3174 debug_printf(" .level_zero_only = %u\n",
3175 texture
->level_zero_only
);
3176 debug_printf(" .pot = %u %u %u\n",
3178 texture
->pot_height
,
3179 texture
->pot_depth
);
3181 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3182 for (i
= 0; i
< key
->nr_images
; ++i
) {
3183 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3184 debug_printf("image[%u] = \n", i
);
3185 debug_printf(" .format = %s\n",
3186 util_format_name(image
->format
));
3187 debug_printf(" .target = %s\n",
3188 util_str_tex_target(image
->target
, TRUE
));
3189 debug_printf(" .level_zero_only = %u\n",
3190 image
->level_zero_only
);
3191 debug_printf(" .pot = %u %u %u\n",
3200 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3202 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3203 variant
->shader
->no
, variant
->no
);
3204 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3205 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3207 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3208 dump_fs_variant_key(&variant
->key
);
3209 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3215 * Generate a new fragment shader variant from the shader code and
3216 * other state indicated by the key.
3218 static struct lp_fragment_shader_variant
*
3219 generate_variant(struct llvmpipe_context
*lp
,
3220 struct lp_fragment_shader
*shader
,
3221 const struct lp_fragment_shader_variant_key
*key
)
3223 struct lp_fragment_shader_variant
*variant
;
3224 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3225 boolean fullcolormask
;
3226 char module_name
[64];
3228 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3232 memset(variant
, 0, sizeof(*variant
));
3233 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3234 shader
->no
, shader
->variants_created
);
3236 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
3237 if (!variant
->gallivm
) {
3242 variant
->shader
= shader
;
3243 variant
->list_item_global
.base
= variant
;
3244 variant
->list_item_local
.base
= variant
;
3245 variant
->no
= shader
->variants_created
++;
3247 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3250 * Determine whether we are touching all channels in the color buffer.
3252 fullcolormask
= FALSE
;
3253 if (key
->nr_cbufs
== 1) {
3254 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3255 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3259 !key
->blend
.logicop_enable
&&
3260 !key
->blend
.rt
[0].blend_enable
&&
3262 !key
->stencil
[0].enabled
&&
3263 !key
->alpha
.enabled
&&
3264 !key
->multisample
&&
3265 !key
->blend
.alpha_to_coverage
&&
3266 !key
->depth
.enabled
&&
3267 !shader
->info
.base
.uses_kill
&&
3268 !shader
->info
.base
.writes_samplemask
3271 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3272 lp_debug_fs_variant(variant
);
3275 lp_jit_init_types(variant
);
3277 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3278 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3280 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3281 if (variant
->opaque
) {
3282 /* Specialized shader, which doesn't need to read the color buffer. */
3283 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3288 * Compile everything
3291 gallivm_compile_module(variant
->gallivm
);
3293 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3295 if (variant
->function
[RAST_EDGE_TEST
]) {
3296 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3297 gallivm_jit_function(variant
->gallivm
,
3298 variant
->function
[RAST_EDGE_TEST
]);
3301 if (variant
->function
[RAST_WHOLE
]) {
3302 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3303 gallivm_jit_function(variant
->gallivm
,
3304 variant
->function
[RAST_WHOLE
]);
3305 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3306 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3309 gallivm_free_ir(variant
->gallivm
);
3316 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3317 const struct pipe_shader_state
*templ
)
3319 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3320 struct lp_fragment_shader
*shader
;
3322 int nr_sampler_views
;
3326 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3330 shader
->no
= fs_no
++;
3331 make_empty_list(&shader
->variants
);
3333 shader
->base
.type
= templ
->type
;
3334 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3335 /* get/save the summary info for this shader */
3336 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3338 /* we need to keep a local copy of the tokens */
3339 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3341 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3342 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3345 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3346 if (shader
->draw_data
== NULL
) {
3347 FREE((void *) shader
->base
.tokens
);
3352 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3353 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3354 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3355 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3357 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3358 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3359 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3360 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3362 switch (shader
->info
.base
.input_interpolate
[i
]) {
3363 case TGSI_INTERPOLATE_CONSTANT
:
3364 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3366 case TGSI_INTERPOLATE_LINEAR
:
3367 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3369 case TGSI_INTERPOLATE_PERSPECTIVE
:
3370 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3372 case TGSI_INTERPOLATE_COLOR
:
3373 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3380 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3381 case TGSI_SEMANTIC_FACE
:
3382 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3384 case TGSI_SEMANTIC_POSITION
:
3385 /* Position was already emitted above
3387 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3388 shader
->inputs
[i
].src_index
= 0;
3392 /* XXX this is a completely pointless index map... */
3393 shader
->inputs
[i
].src_index
= i
+1;
3396 if (LP_DEBUG
& DEBUG_TGSI
) {
3398 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3399 shader
->no
, (void *) shader
);
3400 tgsi_dump(templ
->tokens
, 0);
3401 debug_printf("usage masks:\n");
3402 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3403 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3404 debug_printf(" IN[%u].%s%s%s%s\n",
3406 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3407 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3408 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3409 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3419 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3421 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3422 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3423 if (llvmpipe
->fs
== lp_fs
)
3426 draw_bind_fragment_shader(llvmpipe
->draw
,
3427 (lp_fs
? lp_fs
->draw_data
: NULL
));
3429 llvmpipe
->fs
= lp_fs
;
3431 llvmpipe
->dirty
|= LP_NEW_FS
;
3436 * Remove shader variant from two lists: the shader's variant list
3437 * and the context's variant list.
3440 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3441 struct lp_fragment_shader_variant
*variant
)
3443 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3444 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3445 "v total cached %u inst %u total inst %u\n",
3446 variant
->shader
->no
, variant
->no
,
3447 variant
->shader
->variants_created
,
3448 variant
->shader
->variants_cached
,
3449 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3452 gallivm_destroy(variant
->gallivm
);
3454 /* remove from shader's list */
3455 remove_from_list(&variant
->list_item_local
);
3456 variant
->shader
->variants_cached
--;
3458 /* remove from context's list */
3459 remove_from_list(&variant
->list_item_global
);
3460 lp
->nr_fs_variants
--;
3461 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3468 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3470 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3471 struct lp_fragment_shader
*shader
= fs
;
3472 struct lp_fs_variant_list_item
*li
;
3474 assert(fs
!= llvmpipe
->fs
);
3477 * XXX: we need to flush the context until we have some sort of reference
3478 * counting in fragment shaders as they may still be binned
3479 * Flushing alone might not sufficient we need to wait on it too.
3481 llvmpipe_finish(pipe
, __FUNCTION__
);
3483 /* Delete all the variants */
3484 li
= first_elem(&shader
->variants
);
3485 while(!at_end(&shader
->variants
, li
)) {
3486 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3487 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3491 /* Delete draw module's data */
3492 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3494 if (shader
->base
.ir
.nir
)
3495 ralloc_free(shader
->base
.ir
.nir
);
3496 assert(shader
->variants_cached
== 0);
3497 FREE((void *) shader
->base
.tokens
);
3504 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3505 enum pipe_shader_type shader
, uint index
,
3506 const struct pipe_constant_buffer
*cb
)
3508 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3509 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3511 assert(shader
< PIPE_SHADER_TYPES
);
3512 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3514 /* note: reference counting */
3515 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3518 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3519 debug_printf("Illegal set constant without bind flag\n");
3520 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3524 if (shader
== PIPE_SHADER_VERTEX
||
3525 shader
== PIPE_SHADER_GEOMETRY
||
3526 shader
== PIPE_SHADER_TESS_CTRL
||
3527 shader
== PIPE_SHADER_TESS_EVAL
) {
3528 /* Pass the constants to the 'draw' module */
3529 const unsigned size
= cb
? cb
->buffer_size
: 0;
3533 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3535 else if (cb
&& cb
->user_buffer
) {
3536 data
= (ubyte
*) cb
->user_buffer
;
3543 data
+= cb
->buffer_offset
;
3545 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3548 else if (shader
== PIPE_SHADER_COMPUTE
)
3549 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3551 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3553 if (cb
&& cb
->user_buffer
) {
3554 pipe_resource_reference(&constants
, NULL
);
3559 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3560 enum pipe_shader_type shader
, unsigned start_slot
,
3561 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3562 unsigned writable_bitmask
)
3564 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3566 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3567 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3569 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3571 if (shader
== PIPE_SHADER_VERTEX
||
3572 shader
== PIPE_SHADER_GEOMETRY
||
3573 shader
== PIPE_SHADER_TESS_CTRL
||
3574 shader
== PIPE_SHADER_TESS_EVAL
) {
3575 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3576 const ubyte
*data
= NULL
;
3577 if (buffer
&& buffer
->buffer
)
3578 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3580 data
+= buffer
->buffer_offset
;
3581 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3583 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3584 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3585 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3586 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3592 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3593 enum pipe_shader_type shader
, unsigned start_slot
,
3594 unsigned count
, const struct pipe_image_view
*images
)
3596 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3599 draw_flush(llvmpipe
->draw
);
3600 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3601 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3603 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3606 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3607 if (shader
== PIPE_SHADER_VERTEX
||
3608 shader
== PIPE_SHADER_GEOMETRY
||
3609 shader
== PIPE_SHADER_TESS_CTRL
||
3610 shader
== PIPE_SHADER_TESS_EVAL
) {
3611 draw_set_images(llvmpipe
->draw
,
3613 llvmpipe
->images
[shader
],
3614 start_slot
+ count
);
3615 } else if (shader
== PIPE_SHADER_COMPUTE
)
3616 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3618 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3622 * Return the blend factor equivalent to a destination alpha of one.
3624 static inline unsigned
3625 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3628 case PIPE_BLENDFACTOR_DST_ALPHA
:
3629 return PIPE_BLENDFACTOR_ONE
;
3630 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3631 return PIPE_BLENDFACTOR_ZERO
;
3632 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3634 return PIPE_BLENDFACTOR_ZERO
;
3636 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3644 * We need to generate several variants of the fragment pipeline to match
3645 * all the combinations of the contributing state atoms.
3647 * TODO: there is actually no reason to tie this to context state -- the
3648 * generated code could be cached globally in the screen.
3650 static struct lp_fragment_shader_variant_key
*
3651 make_variant_key(struct llvmpipe_context
*lp
,
3652 struct lp_fragment_shader
*shader
,
3656 struct lp_fragment_shader_variant_key
*key
;
3658 key
= (struct lp_fragment_shader_variant_key
*)store
;
3660 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3662 if (lp
->framebuffer
.zsbuf
) {
3663 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3664 const struct util_format_description
*zsbuf_desc
=
3665 util_format_description(zsbuf_format
);
3667 if (lp
->depth_stencil
->depth
.enabled
&&
3668 util_format_has_depth(zsbuf_desc
)) {
3669 key
->zsbuf_format
= zsbuf_format
;
3670 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3672 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3673 util_format_has_stencil(zsbuf_desc
)) {
3674 key
->zsbuf_format
= zsbuf_format
;
3675 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3677 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3678 key
->resource_1d
= TRUE
;
3680 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3684 * Propagate the depth clamp setting from the rasterizer state.
3685 * depth_clip == 0 implies depth clamping is enabled.
3687 * When clip_halfz is enabled, then always clamp the depth values.
3689 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3690 * clamp is always active in d3d10, regardless if depth clip is
3692 * (GL has an always-on [0,1] clamp on fs depth output instead
3693 * to ensure the depth values stay in range. Doesn't look like
3694 * we do that, though...)
3696 if (lp
->rasterizer
->clip_halfz
) {
3697 key
->depth_clamp
= 1;
3699 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3702 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3703 if (!lp
->framebuffer
.nr_cbufs
||
3704 !lp
->framebuffer
.cbufs
[0] ||
3705 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3706 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3708 if(key
->alpha
.enabled
)
3709 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3710 /* alpha.ref_value is passed in jit_context */
3712 key
->flatshade
= lp
->rasterizer
->flatshade
;
3713 key
->multisample
= lp
->rasterizer
->multisample
;
3714 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3715 key
->occlusion_count
= TRUE
;
3718 if (lp
->framebuffer
.nr_cbufs
) {
3719 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3722 key
->coverage_samples
= 1;
3723 key
->min_samples
= 1;
3724 if (key
->multisample
) {
3725 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3726 key
->min_samples
= lp
->min_samples
== 1 ? 1 : key
->coverage_samples
;
3728 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3730 if (!key
->blend
.independent_blend_enable
) {
3731 /* we always need independent blend otherwise the fixups below won't work */
3732 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3733 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3735 key
->blend
.independent_blend_enable
= 1;
3738 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3739 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3741 if (lp
->framebuffer
.cbufs
[i
]) {
3742 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3743 const struct util_format_description
*format_desc
;
3745 key
->cbuf_format
[i
] = format
;
3746 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3749 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3750 * mixing of 2d textures with height 1 and 1d textures, so make sure
3751 * we pick 1d if any cbuf or zsbuf is 1d.
3753 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3754 key
->resource_1d
= TRUE
;
3757 format_desc
= util_format_description(format
);
3758 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3759 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3762 * Mask out color channels not present in the color buffer.
3764 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3767 * Disable blend for integer formats.
3769 if (util_format_is_pure_integer(format
)) {
3770 blend_rt
->blend_enable
= 0;
3774 * Our swizzled render tiles always have an alpha channel, but the
3775 * linear render target format often does not, so force here the dst
3778 * This is not a mere optimization. Wrong results will be produced if
3779 * the dst alpha is used, the dst format does not have alpha, and the
3780 * previous rendering was not flushed from the swizzled to linear
3781 * buffer. For example, NonPowTwo DCT.
3783 * TODO: This should be generalized to all channels for better
3784 * performance, but only alpha causes correctness issues.
3786 * Also, force rgb/alpha func/factors match, to make AoS blending
3789 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3790 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3791 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3792 boolean clamped_zero
= !util_format_is_float(format
) &&
3793 !util_format_is_snorm(format
);
3794 blend_rt
->rgb_src_factor
=
3795 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3796 blend_rt
->rgb_dst_factor
=
3797 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3798 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3799 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3800 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3804 /* no color buffer for this fragment output */
3805 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3806 key
->cbuf_nr_samples
[i
] = 0;
3807 blend_rt
->colormask
= 0x0;
3808 blend_rt
->blend_enable
= 0;
3812 /* This value will be the same for all the variants of a given shader:
3814 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3816 struct lp_sampler_static_state
*fs_sampler
;
3818 fs_sampler
= key
->samplers
;
3820 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3822 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3823 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3824 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3825 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3830 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3831 * are dx10-style? Can't really have mixed opcodes, at least not
3832 * if we want to skip the holes here (without rescanning tgsi).
3834 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3835 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3836 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3838 * Note sview may exceed what's representable by file_mask.
3839 * This will still work, the only downside is that not actually
3840 * used views may be included in the shader key.
3842 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3843 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3844 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3849 key
->nr_sampler_views
= key
->nr_samplers
;
3850 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3851 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3852 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3853 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3858 struct lp_image_static_state
*lp_image
;
3859 lp_image
= lp_fs_variant_key_images(key
);
3860 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3861 for (i
= 0; i
< key
->nr_images
; ++i
) {
3862 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3863 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3864 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3873 * Update fragment shader state. This is called just prior to drawing
3874 * something when some fragment-related state has changed.
3877 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3879 struct lp_fragment_shader
*shader
= lp
->fs
;
3880 struct lp_fragment_shader_variant_key
*key
;
3881 struct lp_fragment_shader_variant
*variant
= NULL
;
3882 struct lp_fs_variant_list_item
*li
;
3883 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3885 key
= make_variant_key(lp
, shader
, store
);
3887 /* Search the variants for one which matches the key */
3888 li
= first_elem(&shader
->variants
);
3889 while(!at_end(&shader
->variants
, li
)) {
3890 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3898 /* Move this variant to the head of the list to implement LRU
3899 * deletion of shader's when we have too many.
3901 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3904 /* variant not found, create it now */
3907 unsigned variants_to_cull
;
3909 if (LP_DEBUG
& DEBUG_FS
) {
3910 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3913 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3916 /* First, check if we've exceeded the max number of shader variants.
3917 * If so, free 6.25% of them (the least recently used ones).
3919 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3921 if (variants_to_cull
||
3922 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3923 struct pipe_context
*pipe
= &lp
->pipe
;
3925 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3926 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3927 "\t%u instrs,\t%u instrs/variant\n",
3928 shader
->variants_cached
,
3929 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3930 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3934 * XXX: we need to flush the context until we have some sort of
3935 * reference counting in fragment shaders as they may still be binned
3936 * Flushing alone might not be sufficient we need to wait on it too.
3938 llvmpipe_finish(pipe
, __FUNCTION__
);
3941 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3942 * number of shader variants (potentially all of them) could be
3943 * pending for destruction on flush.
3946 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3947 struct lp_fs_variant_list_item
*item
;
3948 if (is_empty_list(&lp
->fs_variants_list
)) {
3951 item
= last_elem(&lp
->fs_variants_list
);
3954 llvmpipe_remove_shader_variant(lp
, item
->base
);
3959 * Generate the new variant.
3962 variant
= generate_variant(lp
, shader
, key
);
3965 LP_COUNT_ADD(llvm_compile_time
, dt
);
3966 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3968 /* Put the new variant into the list */
3970 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3971 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3972 lp
->nr_fs_variants
++;
3973 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3974 shader
->variants_cached
++;
3978 /* Bind this variant */
3979 lp_setup_set_fs_variant(lp
->setup
, variant
);
3987 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3989 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3990 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3991 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3993 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3995 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3996 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;