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
);
298 * Generate the fragment shader, depth/stencil test, and alpha tests.
301 generate_fs_loop(struct gallivm_state
*gallivm
,
302 struct lp_fragment_shader
*shader
,
303 const struct lp_fragment_shader_variant_key
*key
,
304 LLVMBuilderRef builder
,
306 LLVMValueRef context_ptr
,
307 LLVMValueRef num_loop
,
308 struct lp_build_interp_soa_context
*interp
,
309 const struct lp_build_sampler_soa
*sampler
,
310 const struct lp_build_image_soa
*image
,
311 LLVMValueRef mask_store
,
312 LLVMValueRef (*out_color
)[4],
313 LLVMValueRef depth_base_ptr
,
314 LLVMValueRef depth_stride
,
315 LLVMValueRef depth_sample_stride
,
317 LLVMValueRef thread_data_ptr
)
319 const struct util_format_description
*zs_format_desc
= NULL
;
320 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
321 struct lp_type int_type
= lp_int_type(type
);
322 LLVMTypeRef vec_type
, int_vec_type
;
323 LLVMValueRef mask_ptr
= NULL
, mask_val
= NULL
;
324 LLVMValueRef consts_ptr
, num_consts_ptr
;
325 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
327 LLVMValueRef z_value
, s_value
;
328 LLVMValueRef z_fb
, s_fb
;
329 LLVMValueRef depth_ptr
;
330 LLVMValueRef stencil_refs
[2];
331 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
332 LLVMValueRef zs_samples
= lp_build_const_int32(gallivm
, key
->zsbuf_nr_samples
);
333 struct lp_build_for_loop_state loop_state
, sample_loop_state
;
334 struct lp_build_mask_context mask
;
336 * TODO: figure out if simple_shader optimization is really worthwile to
337 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
338 * code since tests tend to take another codepath than real shaders.
340 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
341 shader
->info
.base
.num_inputs
< 3 &&
342 shader
->info
.base
.num_instructions
< 8) && 0;
343 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
344 util_blend_state_is_dual(&key
->blend
, 0);
350 struct lp_bld_tgsi_system_values system_values
;
352 memset(&system_values
, 0, sizeof(system_values
));
354 /* truncate then sign extend. */
355 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
356 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
358 if (key
->depth
.enabled
||
359 key
->stencil
[0].enabled
) {
361 zs_format_desc
= util_format_description(key
->zsbuf_format
);
362 assert(zs_format_desc
);
364 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
365 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
366 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
367 if (shader
->info
.base
.writes_memory
)
368 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
369 else if (key
->alpha
.enabled
||
370 key
->blend
.alpha_to_coverage
||
371 shader
->info
.base
.uses_kill
||
372 shader
->info
.base
.writes_samplemask
) {
373 /* With alpha test and kill, can do the depth test early
374 * and hopefully eliminate some quads. But need to do a
375 * special deferred depth write once the final mask value
376 * is known. This only works though if there's either no
377 * stencil test or the stencil value isn't written.
379 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
380 (key
->stencil
[1].enabled
&&
381 key
->stencil
[1].writemask
)))
382 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
384 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
387 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
390 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
393 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
394 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
395 (key
->stencil
[1].enabled
&&
396 key
->stencil
[1].writemask
))))
397 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
403 vec_type
= lp_build_vec_type(gallivm
, type
);
404 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
406 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
407 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
408 /* convert scalar stencil refs into vectors */
409 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
410 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
412 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
413 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
415 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
416 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
418 memset(outputs
, 0, sizeof outputs
);
420 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
421 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
422 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
423 lp_build_vec_type(gallivm
,
428 if (dual_source_blend
) {
429 assert(key
->nr_cbufs
<= 1);
430 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
431 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
432 lp_build_vec_type(gallivm
,
438 lp_build_for_loop_begin(&loop_state
, gallivm
,
439 lp_build_const_int32(gallivm
, 0),
442 lp_build_const_int32(gallivm
, 1));
444 if (key
->multisample
) {
445 /* create shader execution mask by combining all sample masks. */
446 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
447 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
448 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
449 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
453 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
456 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
457 &loop_state
.counter
, 1, "mask_ptr");
458 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
461 /* 'mask' will control execution based on quad's pixel alive/killed state */
462 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
464 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
465 lp_build_mask_check(&mask
);
467 /* Create storage for recombining sample masks after early Z pass. */
468 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
469 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
471 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
472 LLVMValueRef z_sample_value_store
= NULL
, s_sample_value_store
= NULL
;
473 LLVMValueRef z_fb_store
= NULL
, s_fb_store
= NULL
;
474 LLVMTypeRef z_type
= NULL
, z_fb_type
= NULL
;
476 /* Run early depth once per sample */
477 if (key
->multisample
) {
479 if (zs_format_desc
) {
480 struct lp_type zs_type
= lp_depth_type(zs_format_desc
, type
.length
);
481 struct lp_type z_type
= zs_type
;
482 struct lp_type s_type
= zs_type
;
483 if (zs_format_desc
->block
.bits
< type
.width
)
484 z_type
.width
= type
.width
;
485 else if (zs_format_desc
->block
.bits
> 32) {
486 z_type
.width
= z_type
.width
/ 2;
487 s_type
.width
= s_type
.width
/ 2;
490 z_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
491 zs_samples
, "z_sample_store");
492 s_sample_value_store
= lp_build_array_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
),
493 zs_samples
, "s_sample_store");
494 z_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, z_type
),
495 zs_samples
, "z_fb_store");
496 s_fb_store
= lp_build_array_alloca(gallivm
, lp_build_vec_type(gallivm
, s_type
),
497 zs_samples
, "s_fb_store");
499 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
500 lp_build_const_int32(gallivm
, 0),
501 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
502 lp_build_const_int32(gallivm
, 1));
504 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
505 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
506 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
508 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
509 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
513 /* for multisample Z needs to be interpolated at sample points for testing. */
514 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, key
->multisample
? sample_loop_state
.counter
: NULL
);
517 depth_ptr
= depth_base_ptr
;
518 if (key
->multisample
) {
519 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
520 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
523 if (depth_mode
& EARLY_DEPTH_TEST
) {
525 * Clamp according to ARB_depth_clamp semantics.
527 if (key
->depth_clamp
) {
528 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
531 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
532 zs_format_desc
, key
->resource_1d
,
533 depth_ptr
, depth_stride
,
534 &z_fb
, &s_fb
, loop_state
.counter
);
535 lp_build_depth_stencil_test(gallivm
,
540 key
->multisample
? NULL
: &mask
,
548 if (depth_mode
& EARLY_DEPTH_WRITE
) {
549 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
550 zs_format_desc
, key
->resource_1d
,
551 NULL
, NULL
, NULL
, loop_state
.counter
,
552 depth_ptr
, depth_stride
,
556 * Note mask check if stencil is enabled must be after ds write not after
557 * stencil test otherwise new stencil values may not get written if all
558 * fragments got killed by depth/stencil test.
560 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
561 lp_build_mask_check(&mask
);
563 if (key
->multisample
) {
564 z_fb_type
= LLVMTypeOf(z_fb
);
565 z_type
= LLVMTypeOf(z_value
);
566 lp_build_pointer_set(builder
, z_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, z_value
, lp_build_int_vec_type(gallivm
, type
), ""));
567 lp_build_pointer_set(builder
, s_sample_value_store
, sample_loop_state
.counter
, LLVMBuildBitCast(builder
, s_value
, lp_build_int_vec_type(gallivm
, type
), ""));
568 lp_build_pointer_set(builder
, z_fb_store
, sample_loop_state
.counter
, z_fb
);
569 lp_build_pointer_set(builder
, s_fb_store
, sample_loop_state
.counter
, s_fb
);
573 if (key
->multisample
) {
575 * Store the post-early Z coverage mask.
576 * Recombine the resulting coverage masks post early Z into the fragment
577 * shader execution mask.
579 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
580 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
581 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
583 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
585 lp_build_for_loop_end(&sample_loop_state
);
587 /* recombined all the coverage masks in the shader exec mask. */
588 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
589 lp_build_mask_update(&mask
, tmp_s_mask_or
);
591 /* for multisample Z needs to be re interpolated at pixel center */
592 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
, NULL
);
595 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, NULL
, NULL
);
597 struct lp_build_tgsi_params params
;
598 memset(¶ms
, 0, sizeof(params
));
602 params
.consts_ptr
= consts_ptr
;
603 params
.const_sizes_ptr
= num_consts_ptr
;
604 params
.system_values
= &system_values
;
605 params
.inputs
= interp
->inputs
;
606 params
.context_ptr
= context_ptr
;
607 params
.thread_data_ptr
= thread_data_ptr
;
608 params
.sampler
= sampler
;
609 params
.info
= &shader
->info
.base
;
610 params
.ssbo_ptr
= ssbo_ptr
;
611 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
612 params
.image
= image
;
614 /* Build the actual shader */
615 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
616 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
619 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
623 if (key
->alpha
.enabled
) {
624 int color0
= find_output_by_semantic(&shader
->info
.base
,
628 if (color0
!= -1 && outputs
[color0
][3]) {
629 const struct util_format_description
*cbuf_format_desc
;
630 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
631 LLVMValueRef alpha_ref_value
;
633 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
634 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
636 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
638 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
639 &mask
, alpha
, alpha_ref_value
,
640 (depth_mode
& LATE_DEPTH_TEST
) != 0);
644 /* Emulate Alpha to Coverage with Alpha test */
645 if (key
->blend
.alpha_to_coverage
) {
646 int color0
= find_output_by_semantic(&shader
->info
.base
,
650 if (color0
!= -1 && outputs
[color0
][3]) {
651 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
653 lp_build_alpha_to_coverage(gallivm
, type
,
655 (depth_mode
& LATE_DEPTH_TEST
) != 0);
659 if (shader
->info
.base
.writes_samplemask
) {
660 int smaski
= find_output_by_semantic(&shader
->info
.base
,
661 TGSI_SEMANTIC_SAMPLEMASK
,
664 struct lp_build_context smask_bld
;
665 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
668 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
670 * Pixel is alive according to the first sample in the mask.
672 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
673 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
674 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
675 lp_build_mask_update(&mask
, smask
);
678 if (key
->multisample
) {
679 /* execute depth test for each sample */
680 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
681 lp_build_const_int32(gallivm
, 0),
682 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
683 lp_build_const_int32(gallivm
, 1));
685 /* load the per-sample coverage mask */
686 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
687 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
688 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
690 /* combine the execution mask post fragment shader with the coverage mask. */
691 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
692 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
695 depth_ptr
= depth_base_ptr
;
696 if (key
->multisample
) {
697 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
698 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
702 if (depth_mode
& LATE_DEPTH_TEST
) {
703 int pos0
= find_output_by_semantic(&shader
->info
.base
,
704 TGSI_SEMANTIC_POSITION
,
706 int s_out
= find_output_by_semantic(&shader
->info
.base
,
707 TGSI_SEMANTIC_STENCIL
,
709 if (pos0
!= -1 && outputs
[pos0
][2]) {
710 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
713 * Clamp according to ARB_depth_clamp semantics.
715 if (key
->depth_clamp
) {
716 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
720 if (s_out
!= -1 && outputs
[s_out
][1]) {
721 /* there's only one value, and spec says to discard additional bits */
722 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
723 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
724 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
725 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
726 stencil_refs
[1] = stencil_refs
[0];
729 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
730 zs_format_desc
, key
->resource_1d
,
731 depth_ptr
, depth_stride
,
732 &z_fb
, &s_fb
, loop_state
.counter
);
734 lp_build_depth_stencil_test(gallivm
,
739 key
->multisample
? NULL
: &mask
,
747 if (depth_mode
& LATE_DEPTH_WRITE
) {
748 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
749 zs_format_desc
, key
->resource_1d
,
750 NULL
, NULL
, NULL
, loop_state
.counter
,
751 depth_ptr
, depth_stride
,
755 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
756 (depth_mode
& LATE_DEPTH_WRITE
))
758 /* Need to apply a reduced mask to the depth write. Reload the
759 * depth value, update from zs_value with the new mask value and
762 if (key
->multisample
) {
763 z_value
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_sample_value_store
, sample_loop_state
.counter
), z_type
, "");;
764 s_value
= lp_build_pointer_get(builder
, s_sample_value_store
, sample_loop_state
.counter
);
765 z_fb
= LLVMBuildBitCast(builder
, lp_build_pointer_get(builder
, z_fb_store
, sample_loop_state
.counter
), z_fb_type
, "");
766 s_fb
= lp_build_pointer_get(builder
, s_fb_store
, sample_loop_state
.counter
);
768 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
769 zs_format_desc
, key
->resource_1d
,
770 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
771 depth_ptr
, depth_stride
,
775 if (key
->multisample
) {
776 /* store the sample mask for this loop */
777 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
778 lp_build_for_loop_end(&sample_loop_state
);
782 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
784 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
785 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
786 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
788 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
789 if(outputs
[attrib
][chan
]) {
790 /* XXX: just initialize outputs to point at colors[] and
793 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
794 LLVMValueRef color_ptr
;
795 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
796 &loop_state
.counter
, 1, "");
797 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
798 LLVMBuildStore(builder
, out
, color_ptr
);
804 if (key
->occlusion_count
) {
805 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
806 lp_build_name(counter
, "counter");
807 lp_build_occlusion_count(gallivm
, type
,
808 lp_build_mask_value(&mask
), counter
);
811 mask_val
= lp_build_mask_end(&mask
);
812 if (!key
->multisample
)
813 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
814 lp_build_for_loop_end(&loop_state
);
819 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
821 * Fragment Shader outputs pixels in small 2x2 blocks
822 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
824 * However in memory pixels are stored in rows
825 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
827 * @param type fragment shader type (4x or 8x float)
828 * @param num_fs number of fs_src
829 * @param is_1d whether we're outputting to a 1d resource
830 * @param dst_channels number of output channels
831 * @param fs_src output from fragment shader
832 * @param dst pointer to store result
833 * @param pad_inline is channel padding inline or at end of row
834 * @return the number of dsts
837 generate_fs_twiddle(struct gallivm_state
*gallivm
,
840 unsigned dst_channels
,
841 LLVMValueRef fs_src
[][4],
845 LLVMValueRef src
[16];
851 unsigned pixels
= type
.length
/ 4;
852 unsigned reorder_group
;
853 unsigned src_channels
;
857 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
858 src_count
= num_fs
* src_channels
;
860 assert(pixels
== 2 || pixels
== 1);
861 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
864 * Transpose from SoA -> AoS
866 for (i
= 0; i
< num_fs
; ++i
) {
867 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
871 * Pick transformation options
878 if (dst_channels
== 1) {
884 } else if (dst_channels
== 2) {
888 } else if (dst_channels
> 2) {
895 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
901 * Split the src in half
904 for (i
= num_fs
; i
> 0; --i
) {
905 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
906 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
914 * Ensure pixels are in memory order
917 /* Twiddle pixels by reordering the array, e.g.:
919 * src_count = 8 -> 0 2 1 3 4 6 5 7
920 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
922 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
924 for (i
= 0; i
< src_count
; ++i
) {
925 unsigned group
= i
/ reorder_group
;
926 unsigned block
= (group
/ 4) * 4 * reorder_group
;
927 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
930 } else if (twiddle
) {
931 /* Twiddle pixels across elements of array */
933 * XXX: we should avoid this in some cases, but would need to tell
934 * lp_build_conv to reorder (or deal with it ourselves).
936 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
939 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
943 * Moves any padding between pixels to the end
944 * e.g. RGBXRGBX -> RGBRGBXX
947 unsigned char swizzles
[16];
948 unsigned elems
= pixels
* dst_channels
;
950 for (i
= 0; i
< type
.length
; ++i
) {
952 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
954 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
957 for (i
= 0; i
< src_count
; ++i
) {
958 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
967 * Untwiddle and transpose, much like the above.
968 * However, this is after conversion, so we get packed vectors.
969 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
970 * the vectors will look like:
971 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
972 * be swizzled here). Extending to 16bit should be trivial.
973 * Should also be extended to handle twice wide vectors with AVX2...
976 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
983 struct lp_type type64
, type16
, type32
;
984 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
985 LLVMBuilderRef builder
= gallivm
->builder
;
986 LLVMValueRef tmp
[4], shuf
[8];
987 for (j
= 0; j
< 2; j
++) {
988 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
989 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
990 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
991 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
994 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
995 assert(type
.width
== 8);
996 assert(type
.length
== 16);
998 type8_t
= lp_build_vec_type(gallivm
, type
);
1003 type64_t
= lp_build_vec_type(gallivm
, type64
);
1008 type16_t
= lp_build_vec_type(gallivm
, type16
);
1013 type32_t
= lp_build_vec_type(gallivm
, type32
);
1015 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
1017 if (src_count
== 1) {
1018 /* transpose was no-op, just untwiddle */
1019 LLVMValueRef shuf_vec
;
1020 shuf_vec
= LLVMConstVector(shuf
, 8);
1021 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
1022 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
1023 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
1024 } else if (src_count
== 2) {
1025 LLVMValueRef shuf_vec
;
1026 shuf_vec
= LLVMConstVector(shuf
, 4);
1028 for (i
= 0; i
< 2; i
++) {
1029 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
1030 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
1031 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
1034 for (j
= 0; j
< 2; j
++) {
1035 LLVMValueRef lo
, hi
, lo2
, hi2
;
1037 * Note that if we only really have 3 valid channels (rgb)
1038 * and we don't need alpha we could substitute a undef here
1039 * for the respective channel (causing llvm to drop conversion
1042 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1043 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
1044 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
1045 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1046 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1047 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1048 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1055 * Load an unswizzled block of pixels from memory
1058 load_unswizzled_block(struct gallivm_state
*gallivm
,
1059 LLVMValueRef base_ptr
,
1060 LLVMValueRef stride
,
1061 unsigned block_width
,
1062 unsigned block_height
,
1064 struct lp_type dst_type
,
1066 unsigned dst_alignment
)
1068 LLVMBuilderRef builder
= gallivm
->builder
;
1069 unsigned row_size
= dst_count
/ block_height
;
1072 /* Ensure block exactly fits into dst */
1073 assert((block_width
* block_height
) % dst_count
== 0);
1075 for (i
= 0; i
< dst_count
; ++i
) {
1076 unsigned x
= i
% row_size
;
1077 unsigned y
= i
/ row_size
;
1079 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1080 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1082 LLVMValueRef gep
[2];
1083 LLVMValueRef dst_ptr
;
1085 gep
[0] = lp_build_const_int32(gallivm
, 0);
1086 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1088 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1089 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1090 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1092 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1094 LLVMSetAlignment(dst
[i
], dst_alignment
);
1100 * Store an unswizzled block of pixels to memory
1103 store_unswizzled_block(struct gallivm_state
*gallivm
,
1104 LLVMValueRef base_ptr
,
1105 LLVMValueRef stride
,
1106 unsigned block_width
,
1107 unsigned block_height
,
1109 struct lp_type src_type
,
1111 unsigned src_alignment
)
1113 LLVMBuilderRef builder
= gallivm
->builder
;
1114 unsigned row_size
= src_count
/ block_height
;
1117 /* Ensure src exactly fits into block */
1118 assert((block_width
* block_height
) % src_count
== 0);
1120 for (i
= 0; i
< src_count
; ++i
) {
1121 unsigned x
= i
% row_size
;
1122 unsigned y
= i
/ row_size
;
1124 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1125 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1127 LLVMValueRef gep
[2];
1128 LLVMValueRef src_ptr
;
1130 gep
[0] = lp_build_const_int32(gallivm
, 0);
1131 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1133 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1134 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1135 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1137 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1139 LLVMSetAlignment(src_ptr
, src_alignment
);
1145 * Checks if a format description is an arithmetic format
1147 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1149 static inline boolean
1150 is_arithmetic_format(const struct util_format_description
*format_desc
)
1152 boolean arith
= false;
1155 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1156 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1157 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1165 * Checks if this format requires special handling due to required expansion
1166 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1169 static inline boolean
1170 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1172 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1173 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1181 * Retrieves the type representing the memory layout for a format
1183 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1186 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1187 struct lp_type
* type
)
1192 if (format_expands_to_float_soa(format_desc
)) {
1193 /* just make this a uint with width of block */
1194 type
->floating
= false;
1195 type
->fixed
= false;
1198 type
->width
= format_desc
->block
.bits
;
1203 for (i
= 0; i
< 4; i
++)
1204 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1208 memset(type
, 0, sizeof(struct lp_type
));
1209 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1210 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1211 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1212 type
->norm
= format_desc
->channel
[chan
].normalized
;
1214 if (is_arithmetic_format(format_desc
)) {
1218 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1219 type
->width
+= format_desc
->channel
[i
].size
;
1222 type
->width
= format_desc
->channel
[chan
].size
;
1223 type
->length
= format_desc
->nr_channels
;
1229 * Retrieves the type for a format which is usable in the blending code.
1231 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1234 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1235 struct lp_type
* type
)
1240 if (format_expands_to_float_soa(format_desc
)) {
1241 /* always use ordinary floats for blending */
1242 type
->floating
= true;
1243 type
->fixed
= false;
1251 for (i
= 0; i
< 4; i
++)
1252 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1256 memset(type
, 0, sizeof(struct lp_type
));
1257 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1258 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1259 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1260 type
->norm
= format_desc
->channel
[chan
].normalized
;
1261 type
->width
= format_desc
->channel
[chan
].size
;
1262 type
->length
= format_desc
->nr_channels
;
1264 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1265 if (format_desc
->channel
[i
].size
> type
->width
)
1266 type
->width
= format_desc
->channel
[i
].size
;
1269 if (type
->floating
) {
1272 if (type
->width
<= 8) {
1274 } else if (type
->width
<= 16) {
1281 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1288 * Scale a normalized value from src_bits to dst_bits.
1290 * The exact calculation is
1292 * dst = iround(src * dst_mask / src_mask)
1294 * or with integer rounding
1296 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1300 * src_mask = (1 << src_bits) - 1
1301 * dst_mask = (1 << dst_bits) - 1
1303 * but we try to avoid division and multiplication through shifts.
1305 static inline LLVMValueRef
1306 scale_bits(struct gallivm_state
*gallivm
,
1310 struct lp_type src_type
)
1312 LLVMBuilderRef builder
= gallivm
->builder
;
1313 LLVMValueRef result
= src
;
1315 if (dst_bits
< src_bits
) {
1316 int delta_bits
= src_bits
- dst_bits
;
1318 if (delta_bits
<= dst_bits
) {
1320 * Approximate the rescaling with a single shift.
1322 * This gives the wrong rounding.
1325 result
= LLVMBuildLShr(builder
,
1327 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1332 * Try more accurate rescaling.
1336 * Drop the least significant bits to make space for the multiplication.
1338 * XXX: A better approach would be to use a wider integer type as intermediate. But
1339 * this is enough to convert alpha from 16bits -> 2 when rendering to
1340 * PIPE_FORMAT_R10G10B10A2_UNORM.
1342 result
= LLVMBuildLShr(builder
,
1344 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1348 result
= LLVMBuildMul(builder
,
1350 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1354 * Add a rounding term before the division.
1356 * TODO: Handle signed integers too.
1358 if (!src_type
.sign
) {
1359 result
= LLVMBuildAdd(builder
,
1361 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1366 * Approximate the division by src_mask with a src_bits shift.
1368 * Given the src has already been shifted by dst_bits, all we need
1369 * to do is to shift by the difference.
1372 result
= LLVMBuildLShr(builder
,
1374 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1378 } else if (dst_bits
> src_bits
) {
1380 int db
= dst_bits
- src_bits
;
1382 /* Shift left by difference in bits */
1383 result
= LLVMBuildShl(builder
,
1385 lp_build_const_int_vec(gallivm
, src_type
, db
),
1388 if (db
<= src_bits
) {
1389 /* Enough bits in src to fill the remainder */
1390 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1392 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1395 result
= LLVMBuildOr(builder
, result
, lower
, "");
1396 } else if (db
> src_bits
) {
1397 /* Need to repeatedly copy src bits to fill remainder in dst */
1400 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1401 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1403 result
= LLVMBuildOr(builder
,
1405 LLVMBuildLShr(builder
, result
, shuv
, ""),
1415 * If RT is a smallfloat (needing denorms) format
1418 have_smallfloat_format(struct lp_type dst_type
,
1419 enum pipe_format format
)
1421 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1422 /* due to format handling hacks this format doesn't have floating set
1423 * here (and actually has width set to 32 too) so special case this. */
1424 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1429 * Convert from memory format to blending format
1431 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1434 convert_to_blend_type(struct gallivm_state
*gallivm
,
1435 unsigned block_size
,
1436 const struct util_format_description
*src_fmt
,
1437 struct lp_type src_type
,
1438 struct lp_type dst_type
,
1439 LLVMValueRef
* src
, // and dst
1442 LLVMValueRef
*dst
= src
;
1443 LLVMBuilderRef builder
= gallivm
->builder
;
1444 struct lp_type blend_type
;
1445 struct lp_type mem_type
;
1447 unsigned pixels
= block_size
/ num_srcs
;
1451 * full custom path for packed floats and srgb formats - none of the later
1452 * functions would do anything useful, and given the lp_type representation they
1453 * can't be fixed. Should really have some SoA blend path for these kind of
1454 * formats rather than hacking them in here.
1456 if (format_expands_to_float_soa(src_fmt
)) {
1457 LLVMValueRef tmpsrc
[4];
1459 * This is pretty suboptimal for this case blending in SoA would be much
1460 * better, since conversion gets us SoA values so need to convert back.
1462 assert(src_type
.width
== 32 || src_type
.width
== 16);
1463 assert(dst_type
.floating
);
1464 assert(dst_type
.width
== 32);
1465 assert(dst_type
.length
% 4 == 0);
1466 assert(num_srcs
% 4 == 0);
1468 if (src_type
.width
== 16) {
1469 /* expand 4x16bit values to 4x32bit */
1470 struct lp_type type32x4
= src_type
;
1471 LLVMTypeRef ltype32x4
;
1472 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1473 type32x4
.width
= 32;
1474 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1475 for (i
= 0; i
< num_fetch
; i
++) {
1476 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1478 src_type
.width
= 32;
1480 for (i
= 0; i
< 4; i
++) {
1483 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1484 LLVMValueRef tmpsoa
[4];
1485 LLVMValueRef tmps
= tmpsrc
[i
];
1486 if (dst_type
.length
== 8) {
1487 LLVMValueRef shuffles
[8];
1489 /* fetch was 4 values but need 8-wide output values */
1490 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1492 * for 8-wide aos transpose would give us wrong order not matching
1493 * incoming converted fs values and mask. ARGH.
1495 for (j
= 0; j
< 4; j
++) {
1496 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1497 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1499 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1500 LLVMConstVector(shuffles
, 8), "");
1502 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1503 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1506 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1508 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1513 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1514 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1516 /* Is the format arithmetic */
1517 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1518 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1520 /* Pad if necessary */
1521 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1522 for (i
= 0; i
< num_srcs
; ++i
) {
1523 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1526 src_type
.length
= dst_type
.length
;
1529 /* Special case for half-floats */
1530 if (mem_type
.width
== 16 && mem_type
.floating
) {
1531 assert(blend_type
.width
== 32 && blend_type
.floating
);
1532 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1540 src_type
.width
= blend_type
.width
* blend_type
.length
;
1541 blend_type
.length
*= pixels
;
1542 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1544 for (i
= 0; i
< num_srcs
; ++i
) {
1545 LLVMValueRef chans
[4];
1546 LLVMValueRef res
= NULL
;
1548 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1550 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1552 unsigned sa
= src_fmt
->channel
[j
].shift
;
1553 #if UTIL_ARCH_LITTLE_ENDIAN
1554 unsigned from_lsb
= j
;
1556 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1559 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1561 /* Extract bits from source */
1562 chans
[j
] = LLVMBuildLShr(builder
,
1564 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1567 chans
[j
] = LLVMBuildAnd(builder
,
1569 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1573 if (src_type
.norm
) {
1574 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1575 blend_type
.width
, chans
[j
], src_type
);
1578 /* Insert bits into correct position */
1579 chans
[j
] = LLVMBuildShl(builder
,
1581 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1587 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1591 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1597 * Convert from blending format to memory format
1599 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1602 convert_from_blend_type(struct gallivm_state
*gallivm
,
1603 unsigned block_size
,
1604 const struct util_format_description
*src_fmt
,
1605 struct lp_type src_type
,
1606 struct lp_type dst_type
,
1607 LLVMValueRef
* src
, // and dst
1610 LLVMValueRef
* dst
= src
;
1612 struct lp_type mem_type
;
1613 struct lp_type blend_type
;
1614 LLVMBuilderRef builder
= gallivm
->builder
;
1615 unsigned pixels
= block_size
/ num_srcs
;
1619 * full custom path for packed floats and srgb formats - none of the later
1620 * functions would do anything useful, and given the lp_type representation they
1621 * can't be fixed. Should really have some SoA blend path for these kind of
1622 * formats rather than hacking them in here.
1624 if (format_expands_to_float_soa(src_fmt
)) {
1626 * This is pretty suboptimal for this case blending in SoA would be much
1627 * better - we need to transpose the AoS values back to SoA values for
1628 * conversion/packing.
1630 assert(src_type
.floating
);
1631 assert(src_type
.width
== 32);
1632 assert(src_type
.length
% 4 == 0);
1633 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1635 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1636 LLVMValueRef tmpsoa
[4], tmpdst
;
1637 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1638 /* really really need SoA here */
1640 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1641 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1644 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1648 if (src_type
.length
== 8) {
1649 LLVMValueRef tmpaos
, shuffles
[8];
1652 * for 8-wide aos transpose has given us wrong order not matching
1653 * output order. HMPF. Also need to split the output values manually.
1655 for (j
= 0; j
< 4; j
++) {
1656 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1657 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1659 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1660 LLVMConstVector(shuffles
, 8), "");
1661 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1662 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1668 if (dst_type
.width
== 16) {
1669 struct lp_type type16x8
= dst_type
;
1670 struct lp_type type32x4
= dst_type
;
1671 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1672 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1673 type16x8
.length
= 8;
1674 type32x4
.width
= 32;
1675 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1676 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1677 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1678 /* We could do vector truncation but it doesn't generate very good code */
1679 for (i
= 0; i
< num_fetch
; i
++) {
1680 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1681 src
[i
], lp_build_zero(gallivm
, type32x4
));
1682 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1683 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1684 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1690 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1691 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1693 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1695 /* Special case for half-floats */
1696 if (mem_type
.width
== 16 && mem_type
.floating
) {
1697 int length
= dst_type
.length
;
1698 assert(blend_type
.width
== 32 && blend_type
.floating
);
1700 dst_type
.length
= src_type
.length
;
1702 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1704 dst_type
.length
= length
;
1708 /* Remove any padding */
1709 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1710 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1712 for (i
= 0; i
< num_srcs
; ++i
) {
1713 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1717 /* No bit arithmetic to do */
1722 src_type
.length
= pixels
;
1723 src_type
.width
= blend_type
.length
* blend_type
.width
;
1724 dst_type
.length
= pixels
;
1726 for (i
= 0; i
< num_srcs
; ++i
) {
1727 LLVMValueRef chans
[4];
1728 LLVMValueRef res
= NULL
;
1730 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1732 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1734 unsigned sa
= src_fmt
->channel
[j
].shift
;
1735 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1736 #if UTIL_ARCH_LITTLE_ENDIAN
1737 unsigned from_lsb
= j
;
1739 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1742 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1744 for (k
= 0; k
< blend_type
.width
; ++k
) {
1749 chans
[j
] = LLVMBuildLShr(builder
,
1751 lp_build_const_int_vec(gallivm
, src_type
,
1752 from_lsb
* blend_type
.width
),
1755 chans
[j
] = LLVMBuildAnd(builder
,
1757 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1760 /* Scale down bits */
1761 if (src_type
.norm
) {
1762 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1763 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1764 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1765 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1766 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1767 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1771 chans
[j
] = LLVMBuildShl(builder
,
1773 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1776 sa
+= src_fmt
->channel
[j
].size
;
1781 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1785 assert (dst_type
.width
!= 24);
1787 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1793 * Convert alpha to same blend type as src
1796 convert_alpha(struct gallivm_state
*gallivm
,
1797 struct lp_type row_type
,
1798 struct lp_type alpha_type
,
1799 const unsigned block_size
,
1800 const unsigned block_height
,
1801 const unsigned src_count
,
1802 const unsigned dst_channels
,
1803 const bool pad_inline
,
1804 LLVMValueRef
* src_alpha
)
1806 LLVMBuilderRef builder
= gallivm
->builder
;
1808 unsigned length
= row_type
.length
;
1809 row_type
.length
= alpha_type
.length
;
1811 /* Twiddle the alpha to match pixels */
1812 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1815 * TODO this should use single lp_build_conv call for
1816 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1818 for (i
= 0; i
< block_height
; ++i
) {
1819 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1822 alpha_type
= row_type
;
1823 row_type
.length
= length
;
1825 /* If only one channel we can only need the single alpha value per pixel */
1826 if (src_count
== 1 && dst_channels
== 1) {
1828 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1830 /* If there are more srcs than rows then we need to split alpha up */
1831 if (src_count
> block_height
) {
1832 for (i
= src_count
; i
> 0; --i
) {
1833 unsigned pixels
= block_size
/ src_count
;
1834 unsigned idx
= i
- 1;
1836 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1837 (idx
* pixels
) % 4, pixels
);
1841 /* If there is a src for each pixel broadcast the alpha across whole row */
1842 if (src_count
== block_size
) {
1843 for (i
= 0; i
< src_count
; ++i
) {
1844 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1845 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1848 unsigned pixels
= block_size
/ src_count
;
1849 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1850 unsigned alpha_span
= 1;
1851 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1853 /* Check if we need 2 src_alphas for our shuffles */
1854 if (pixels
> alpha_type
.length
) {
1858 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1859 for (j
= 0; j
< row_type
.length
; ++j
) {
1860 if (j
< pixels
* channels
) {
1861 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1863 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1867 for (i
= 0; i
< src_count
; ++i
) {
1868 unsigned idx1
= i
, idx2
= i
;
1870 if (alpha_span
> 1){
1875 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1878 LLVMConstVector(shuffles
, row_type
.length
),
1887 * Generates the blend function for unswizzled colour buffers
1888 * Also generates the read & write from colour buffer
1891 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1893 struct lp_fragment_shader_variant
*variant
,
1894 enum pipe_format out_format
,
1895 unsigned int num_fs
,
1896 struct lp_type fs_type
,
1897 LLVMValueRef
* fs_mask
,
1898 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1899 LLVMValueRef context_ptr
,
1900 LLVMValueRef color_ptr
,
1901 LLVMValueRef stride
,
1902 unsigned partial_mask
,
1905 const unsigned alpha_channel
= 3;
1906 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1907 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1908 const unsigned block_size
= block_width
* block_height
;
1909 const unsigned lp_integer_vector_width
= 128;
1911 LLVMBuilderRef builder
= gallivm
->builder
;
1912 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1913 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1914 LLVMValueRef src_alpha
[4 * 4];
1915 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1916 LLVMValueRef src_mask
[4 * 4];
1917 LLVMValueRef src
[4 * 4];
1918 LLVMValueRef src1
[4 * 4];
1919 LLVMValueRef dst
[4 * 4];
1920 LLVMValueRef blend_color
;
1921 LLVMValueRef blend_alpha
;
1922 LLVMValueRef i32_zero
;
1923 LLVMValueRef check_mask
;
1924 LLVMValueRef undef_src_val
;
1926 struct lp_build_mask_context mask_ctx
;
1927 struct lp_type mask_type
;
1928 struct lp_type blend_type
;
1929 struct lp_type row_type
;
1930 struct lp_type dst_type
;
1931 struct lp_type ls_type
;
1933 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1934 unsigned vector_width
;
1935 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1936 unsigned dst_channels
;
1941 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1943 unsigned dst_alignment
;
1945 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1946 bool has_alpha
= false;
1947 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1948 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1950 const boolean is_1d
= variant
->key
.resource_1d
;
1951 boolean twiddle_after_convert
= FALSE
;
1952 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1953 LLVMValueRef fpstate
= 0;
1955 /* Get type from output format */
1956 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1957 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1960 * Technically this code should go into lp_build_smallfloat_to_float
1961 * and lp_build_float_to_smallfloat but due to the
1962 * http://llvm.org/bugs/show_bug.cgi?id=6393
1963 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1964 * So the ordering is important here and there shouldn't be any
1965 * llvm ir instrunctions in this function before
1966 * this, otherwise half-float format conversions won't work
1967 * (again due to llvm bug #6393).
1969 if (have_smallfloat_format(dst_type
, out_format
)) {
1970 /* We need to make sure that denorms are ok for half float
1972 fpstate
= lp_build_fpstate_get(gallivm
);
1973 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1976 mask_type
= lp_int32_vec4_type();
1977 mask_type
.length
= fs_type
.length
;
1979 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1980 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1983 /* Do not bother executing code when mask is empty.. */
1985 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1987 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1988 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1991 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1992 lp_build_mask_check(&mask_ctx
);
1995 partial_mask
|= !variant
->opaque
;
1996 i32_zero
= lp_build_const_int32(gallivm
, 0);
1998 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
2000 row_type
.length
= fs_type
.length
;
2001 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
2003 /* Compute correct swizzle and count channels */
2004 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
2007 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
2008 /* Ensure channel is used */
2009 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
2013 /* Ensure not already written to (happens in case with GL_ALPHA) */
2014 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
2018 /* Ensure we havn't already found all channels */
2019 if (dst_channels
>= out_format_desc
->nr_channels
) {
2023 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
2026 if (i
== alpha_channel
) {
2031 if (format_expands_to_float_soa(out_format_desc
)) {
2033 * the code above can't work for layout_other
2034 * for srgb it would sort of work but we short-circuit swizzles, etc.
2035 * as that is done as part of unpack / pack.
2037 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2043 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2046 /* If 3 channels then pad to include alpha for 4 element transpose */
2047 if (dst_channels
== 3) {
2048 assert (!has_alpha
);
2049 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2050 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2053 if (out_format_desc
->nr_channels
== 4) {
2056 * We use alpha from the color conversion, not separate one.
2057 * We had to include it for transpose, hence it will get converted
2058 * too (albeit when doing transpose after conversion, that would
2059 * no longer be the case necessarily).
2060 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2061 * otherwise we really have padding, not alpha, included.)
2068 * Load shader output
2070 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2071 /* Always load alpha for use in blending */
2074 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2077 alpha
= undef_src_val
;
2080 /* Load each channel */
2081 for (j
= 0; j
< dst_channels
; ++j
) {
2082 assert(swizzle
[j
] < 4);
2084 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2087 fs_src
[i
][j
] = undef_src_val
;
2091 /* If 3 channels then pad to include alpha for 4 element transpose */
2093 * XXX If we include that here maybe could actually use it instead of
2094 * separate alpha for blending?
2095 * (Difficult though we actually convert pad channels, not alpha.)
2097 if (dst_channels
== 3 && !has_alpha
) {
2098 fs_src
[i
][3] = alpha
;
2101 /* We split the row_mask and row_alpha as we want 128bit interleave */
2102 if (fs_type
.length
== 8) {
2103 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2105 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2106 src_channels
, src_channels
);
2108 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2109 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2110 src_channels
, src_channels
);
2112 src_mask
[i
] = fs_mask
[i
];
2113 src_alpha
[i
] = alpha
;
2116 if (dual_source_blend
) {
2117 /* same as above except different src/dst, skip masks and comments... */
2118 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2121 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2124 alpha
= undef_src_val
;
2127 for (j
= 0; j
< dst_channels
; ++j
) {
2128 assert(swizzle
[j
] < 4);
2130 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2133 fs_src1
[i
][j
] = undef_src_val
;
2136 if (dst_channels
== 3 && !has_alpha
) {
2137 fs_src1
[i
][3] = alpha
;
2139 if (fs_type
.length
== 8) {
2140 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2141 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2142 src_channels
, src_channels
);
2144 src1_alpha
[i
] = alpha
;
2149 if (util_format_is_pure_integer(out_format
)) {
2151 * In this case fs_type was really ints or uints disguised as floats,
2154 fs_type
.floating
= 0;
2155 fs_type
.sign
= dst_type
.sign
;
2156 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2157 for (j
= 0; j
< dst_channels
; ++j
) {
2158 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2159 lp_build_vec_type(gallivm
, fs_type
), "");
2161 if (dst_channels
== 3 && !has_alpha
) {
2162 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2163 lp_build_vec_type(gallivm
, fs_type
), "");
2169 * We actually should generally do conversion first (for non-1d cases)
2170 * when the blend format is 8 or 16 bits. The reason is obvious,
2171 * there's 2 or 4 times less vectors to deal with for the interleave...
2172 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2173 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2174 * unpack only with 128bit vectors).
2175 * Note: for 16bit sizes really need matching pack conversion code
2177 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2178 twiddle_after_convert
= TRUE
;
2182 * Pixel twiddle from fragment shader order to memory order
2184 if (!twiddle_after_convert
) {
2185 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2186 dst_channels
, fs_src
, src
, pad_inline
);
2187 if (dual_source_blend
) {
2188 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2189 fs_src1
, src1
, pad_inline
);
2192 src_count
= num_fullblock_fs
* dst_channels
;
2194 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2195 * (AVX) turn out the same later when untwiddling/transpose (albeit
2196 * for true AVX2 path untwiddle needs to be different).
2197 * For now just order by colors first (so we can use unpack later).
2199 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2200 for (i
= 0; i
< dst_channels
; i
++) {
2201 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2202 if (dual_source_blend
) {
2203 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2209 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2210 if (src_count
!= num_fullblock_fs
* src_channels
) {
2211 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2212 row_type
.length
/= ds
;
2213 fs_type
.length
= row_type
.length
;
2216 blend_type
= row_type
;
2217 mask_type
.length
= 4;
2219 /* Convert src to row_type */
2220 if (dual_source_blend
) {
2221 struct lp_type old_row_type
= row_type
;
2222 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2223 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2226 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2229 /* If the rows are not an SSE vector, combine them to become SSE size! */
2230 if ((row_type
.width
* row_type
.length
) % 128) {
2231 unsigned bits
= row_type
.width
* row_type
.length
;
2234 assert(src_count
>= (vector_width
/ bits
));
2236 dst_count
= src_count
/ (vector_width
/ bits
);
2238 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2239 if (dual_source_blend
) {
2240 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2243 row_type
.length
*= combined
;
2244 src_count
/= combined
;
2246 bits
= row_type
.width
* row_type
.length
;
2247 assert(bits
== 128 || bits
== 256);
2250 if (twiddle_after_convert
) {
2251 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2252 if (dual_source_blend
) {
2253 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2258 * Blend Colour conversion
2260 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2261 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2262 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2263 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2264 &i32_zero
, 1, ""), "");
2267 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2269 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2271 * since blending is done with floats, there was no conversion.
2272 * However, the rules according to fixed point renderbuffers still
2273 * apply, that is we must clamp inputs to 0.0/1.0.
2274 * (This would apply to separate alpha conversion too but we currently
2275 * force has_alpha to be true.)
2276 * TODO: should skip this with "fake" blend, since post-blend conversion
2277 * will clamp anyway.
2278 * TODO: could also skip this if fragment color clamping is enabled. We
2279 * don't support it natively so it gets baked into the shader however, so
2280 * can't really tell here.
2282 struct lp_build_context f32_bld
;
2283 assert(row_type
.floating
);
2284 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2285 for (i
= 0; i
< src_count
; i
++) {
2286 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2288 if (dual_source_blend
) {
2289 for (i
= 0; i
< src_count
; i
++) {
2290 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2293 /* probably can't be different than row_type but better safe than sorry... */
2294 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2295 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2299 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2301 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2302 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2304 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2305 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2307 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2308 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2314 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2316 if (src_count
< block_height
) {
2317 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2318 } else if (src_count
> block_height
) {
2319 for (i
= src_count
; i
> 0; --i
) {
2320 unsigned pixels
= block_size
/ src_count
;
2321 unsigned idx
= i
- 1;
2323 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2324 (idx
* pixels
) % 4, pixels
);
2328 assert(mask_type
.width
== 32);
2330 for (i
= 0; i
< src_count
; ++i
) {
2331 unsigned pixels
= block_size
/ src_count
;
2332 unsigned pixel_width
= row_type
.width
* dst_channels
;
2334 if (pixel_width
== 24) {
2335 mask_type
.width
= 8;
2336 mask_type
.length
= vector_width
/ mask_type
.width
;
2338 mask_type
.length
= pixels
;
2339 mask_type
.width
= row_type
.width
* dst_channels
;
2342 * If mask_type width is smaller than 32bit, this doesn't quite
2343 * generate the most efficient code (could use some pack).
2345 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2346 lp_build_int_vec_type(gallivm
, mask_type
), "");
2348 mask_type
.length
*= dst_channels
;
2349 mask_type
.width
/= dst_channels
;
2352 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2353 lp_build_int_vec_type(gallivm
, mask_type
), "");
2354 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2361 struct lp_type alpha_type
= fs_type
;
2362 alpha_type
.length
= 4;
2363 convert_alpha(gallivm
, row_type
, alpha_type
,
2364 block_size
, block_height
,
2365 src_count
, dst_channels
,
2366 pad_inline
, src_alpha
);
2367 if (dual_source_blend
) {
2368 convert_alpha(gallivm
, row_type
, alpha_type
,
2369 block_size
, block_height
,
2370 src_count
, dst_channels
,
2371 pad_inline
, src1_alpha
);
2377 * Load dst from memory
2379 if (src_count
< block_height
) {
2380 dst_count
= block_height
;
2382 dst_count
= src_count
;
2385 dst_type
.length
*= block_size
/ dst_count
;
2387 if (format_expands_to_float_soa(out_format_desc
)) {
2389 * we need multiple values at once for the conversion, so can as well
2390 * load them vectorized here too instead of concatenating later.
2391 * (Still need concatenation later for 8-wide vectors).
2393 dst_count
= block_height
;
2394 dst_type
.length
= block_width
;
2398 * Compute the alignment of the destination pointer in bytes
2399 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2400 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2401 * 1d tex but can't distinguish here) so need to stick with per-pixel
2402 * alignment in this case.
2405 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2408 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2410 /* Force power-of-two alignment by extracting only the least-significant-bit */
2411 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2413 * Resource base and stride pointers are aligned to 16 bytes, so that's
2414 * the maximum alignment we can guarantee
2416 dst_alignment
= MIN2(16, dst_alignment
);
2420 if (dst_count
> src_count
) {
2421 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2422 util_is_power_of_two_or_zero(dst_type
.length
) &&
2423 dst_type
.length
* dst_type
.width
< 128) {
2425 * Never try to load values as 4xi8 which we will then
2426 * concatenate to larger vectors. This gives llvm a real
2427 * headache (the problem is the type legalizer (?) will
2428 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2429 * then the shuffles to concatenate are more or less impossible
2430 * - llvm is easily capable of generating a sequence of 32
2431 * pextrb/pinsrb instructions for that. Albeit it appears to
2432 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2433 * width to avoid the trouble (16bit seems not as bad, llvm
2434 * probably recognizes the load+shuffle as only one shuffle
2435 * is necessary, but we can do just the same anyway).
2437 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2443 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2444 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2445 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2446 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2451 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2452 dst
, ls_type
, dst_count
, dst_alignment
);
2457 * Convert from dst/output format to src/blending format.
2459 * This is necessary as we can only read 1 row from memory at a time,
2460 * so the minimum dst_count will ever be at this point is 4.
2462 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2463 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2464 * on all 16 pixels in that single vector at once.
2466 if (dst_count
> src_count
) {
2467 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2468 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2469 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2470 for (i
= 0; i
< dst_count
; i
++) {
2471 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2475 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2477 if (ls_type
.length
!= dst_type
.length
) {
2478 struct lp_type tmp_type
= dst_type
;
2479 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2480 for (i
= 0; i
< src_count
; i
++) {
2481 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2482 lp_build_vec_type(gallivm
, tmp_type
), "");
2490 /* XXX this is broken for RGB8 formats -
2491 * they get expanded from 12 to 16 elements (to include alpha)
2492 * by convert_to_blend_type then reduced to 15 instead of 12
2493 * by convert_from_blend_type (a simple fix though breaks A8...).
2494 * R16G16B16 also crashes differently however something going wrong
2495 * inside llvm handling npot vector sizes seemingly.
2496 * It seems some cleanup could be done here (like skipping conversion/blend
2499 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2500 row_type
, dst
, src_count
);
2503 * FIXME: Really should get logic ops / masks out of generic blend / row
2504 * format. Logic ops will definitely not work on the blend float format
2505 * used for SRGB here and I think OpenGL expects this to work as expected
2506 * (that is incoming values converted to srgb then logic op applied).
2508 for (i
= 0; i
< src_count
; ++i
) {
2509 dst
[i
] = lp_build_blend_aos(gallivm
,
2510 &variant
->key
.blend
,
2515 has_alpha
? NULL
: src_alpha
[i
],
2517 has_alpha
? NULL
: src1_alpha
[i
],
2519 partial_mask
? src_mask
[i
] : NULL
,
2521 has_alpha
? NULL
: blend_alpha
,
2523 pad_inline
? 4 : dst_channels
);
2526 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2527 row_type
, dst_type
, dst
, src_count
);
2529 /* Split the blend rows back to memory rows */
2530 if (dst_count
> src_count
) {
2531 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2533 if (src_count
== 1) {
2534 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2535 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2537 row_type
.length
/= 2;
2541 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2542 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2543 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2544 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2546 row_type
.length
/= 2;
2551 * Store blend result to memory
2554 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2555 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2558 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2559 dst
, dst_type
, dst_count
, dst_alignment
);
2562 if (have_smallfloat_format(dst_type
, out_format
)) {
2563 lp_build_fpstate_set(gallivm
, fpstate
);
2567 lp_build_mask_end(&mask_ctx
);
2573 * Generate the runtime callable function for the whole fragment pipeline.
2574 * Note that the function which we generate operates on a block of 16
2575 * pixels at at time. The block contains 2x2 quads. Each quad contains
2579 generate_fragment(struct llvmpipe_context
*lp
,
2580 struct lp_fragment_shader
*shader
,
2581 struct lp_fragment_shader_variant
*variant
,
2582 unsigned partial_mask
)
2584 struct gallivm_state
*gallivm
= variant
->gallivm
;
2585 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2586 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2588 struct lp_type fs_type
;
2589 struct lp_type blend_type
;
2590 LLVMTypeRef fs_elem_type
;
2591 LLVMTypeRef blend_vec_type
;
2592 LLVMTypeRef arg_types
[15];
2593 LLVMTypeRef func_type
;
2594 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2595 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2596 LLVMValueRef context_ptr
;
2599 LLVMValueRef a0_ptr
;
2600 LLVMValueRef dadx_ptr
;
2601 LLVMValueRef dady_ptr
;
2602 LLVMValueRef color_ptr_ptr
;
2603 LLVMValueRef stride_ptr
;
2604 LLVMValueRef color_sample_stride_ptr
;
2605 LLVMValueRef depth_ptr
;
2606 LLVMValueRef depth_stride
;
2607 LLVMValueRef depth_sample_stride
;
2608 LLVMValueRef mask_input
;
2609 LLVMValueRef thread_data_ptr
;
2610 LLVMBasicBlockRef block
;
2611 LLVMBuilderRef builder
;
2612 struct lp_build_sampler_soa
*sampler
;
2613 struct lp_build_image_soa
*image
;
2614 struct lp_build_interp_soa_context interp
;
2615 LLVMValueRef fs_mask
[16 / 4];
2616 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2617 LLVMValueRef function
;
2618 LLVMValueRef facing
;
2623 boolean cbuf0_write_all
;
2624 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2625 util_blend_state_is_dual(&key
->blend
, 0);
2627 assert(lp_native_vector_width
/ 32 >= 4);
2629 /* Adjust color input interpolation according to flatshade state:
2631 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2632 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2633 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2635 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2637 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2641 /* check if writes to cbuf[0] are to be copied to all cbufs */
2643 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2645 /* TODO: actually pick these based on the fs and color buffer
2646 * characteristics. */
2648 memset(&fs_type
, 0, sizeof fs_type
);
2649 fs_type
.floating
= TRUE
; /* floating point values */
2650 fs_type
.sign
= TRUE
; /* values are signed */
2651 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2652 fs_type
.width
= 32; /* 32-bit float */
2653 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2655 memset(&blend_type
, 0, sizeof blend_type
);
2656 blend_type
.floating
= FALSE
; /* values are integers */
2657 blend_type
.sign
= FALSE
; /* values are unsigned */
2658 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2659 blend_type
.width
= 8; /* 8-bit ubyte values */
2660 blend_type
.length
= 16; /* 16 elements per vector */
2663 * Generate the function prototype. Any change here must be reflected in
2664 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2667 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2669 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2671 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2672 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2674 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2675 arg_types
[1] = int32_type
; /* x */
2676 arg_types
[2] = int32_type
; /* y */
2677 arg_types
[3] = int32_type
; /* facing */
2678 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2679 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2680 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2681 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2682 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2683 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2684 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2685 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2686 arg_types
[12] = int32_type
; /* depth_stride */
2687 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2688 arg_types
[14] = int32_type
; /* depth sample stride */
2690 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2691 arg_types
, ARRAY_SIZE(arg_types
), 0);
2693 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2694 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2696 variant
->function
[partial_mask
] = function
;
2698 /* XXX: need to propagate noalias down into color param now we are
2699 * passing a pointer-to-pointer?
2701 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2702 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2703 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2705 context_ptr
= LLVMGetParam(function
, 0);
2706 x
= LLVMGetParam(function
, 1);
2707 y
= LLVMGetParam(function
, 2);
2708 facing
= LLVMGetParam(function
, 3);
2709 a0_ptr
= LLVMGetParam(function
, 4);
2710 dadx_ptr
= LLVMGetParam(function
, 5);
2711 dady_ptr
= LLVMGetParam(function
, 6);
2712 color_ptr_ptr
= LLVMGetParam(function
, 7);
2713 depth_ptr
= LLVMGetParam(function
, 8);
2714 mask_input
= LLVMGetParam(function
, 9);
2715 thread_data_ptr
= LLVMGetParam(function
, 10);
2716 stride_ptr
= LLVMGetParam(function
, 11);
2717 depth_stride
= LLVMGetParam(function
, 12);
2718 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2719 depth_sample_stride
= LLVMGetParam(function
, 14);
2721 lp_build_name(context_ptr
, "context");
2722 lp_build_name(x
, "x");
2723 lp_build_name(y
, "y");
2724 lp_build_name(a0_ptr
, "a0");
2725 lp_build_name(dadx_ptr
, "dadx");
2726 lp_build_name(dady_ptr
, "dady");
2727 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2728 lp_build_name(depth_ptr
, "depth");
2729 lp_build_name(mask_input
, "mask_input");
2730 lp_build_name(thread_data_ptr
, "thread_data");
2731 lp_build_name(stride_ptr
, "stride_ptr");
2732 lp_build_name(depth_stride
, "depth_stride");
2733 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2734 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2740 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2741 builder
= gallivm
->builder
;
2743 LLVMPositionBuilderAtEnd(builder
, block
);
2746 * Must not count ps invocations if there's a null shader.
2747 * (It would be ok to count with null shader if there's d/s tests,
2748 * but only if there's d/s buffers too, which is different
2749 * to implicit rasterization disable which must not depend
2750 * on the d/s buffers.)
2751 * Could use popcount on mask, but pixel accuracy is not required.
2752 * Could disable if there's no stats query, but maybe not worth it.
2754 if (shader
->info
.base
.num_instructions
> 1) {
2755 LLVMValueRef invocs
, val
;
2756 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2757 val
= LLVMBuildLoad(builder
, invocs
, "");
2758 val
= LLVMBuildAdd(builder
, val
,
2759 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2761 LLVMBuildStore(builder
, val
, invocs
);
2764 /* code generated texture sampling */
2765 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2766 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2768 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2769 /* for 1d resources only run "upper half" of stamp */
2770 if (key
->resource_1d
)
2774 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2775 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2776 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2777 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2778 num_loop_samp
, "mask_store");
2779 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2780 boolean pixel_center_integer
=
2781 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2784 * The shader input interpolation info is not explicitely baked in the
2785 * shader key, but everything it derives from (TGSI, and flatshade) is
2786 * already included in the shader key.
2788 lp_build_interp_soa_init(&interp
,
2790 shader
->info
.base
.num_inputs
,
2792 pixel_center_integer
,
2796 a0_ptr
, dadx_ptr
, dady_ptr
,
2799 for (i
= 0; i
< num_fs
; i
++) {
2800 if (key
->multisample
) {
2801 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2804 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2805 * store to the per sample mask storage. Or all of them together to generate
2806 * the fragment shader mask. (sample shading TODO).
2807 * Take the incoming state coverage mask into account.
2809 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2810 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2811 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2812 &sindexi
, 1, "sample_mask_ptr");
2813 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2814 i
*fs_type
.length
/4, s
, mask_input
);
2816 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2817 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2818 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2819 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2821 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2822 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
2826 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2827 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2828 &indexi
, 1, "mask_ptr");
2831 mask
= generate_quad_mask(gallivm
, fs_type
,
2832 i
*fs_type
.length
/4, 0, mask_input
);
2835 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2837 LLVMBuildStore(builder
, mask
, mask_ptr
);
2841 generate_fs_loop(gallivm
,
2850 mask_store
, /* output */
2854 depth_sample_stride
,
2858 for (i
= 0; i
< num_fs
; i
++) {
2859 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2860 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2862 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2863 /* This is fucked up need to reorganize things */
2864 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2865 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2866 ptr
= LLVMBuildGEP(builder
,
2867 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2869 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2872 if (dual_source_blend
) {
2873 /* only support one dual source blend target hence always use output 1 */
2874 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2875 ptr
= LLVMBuildGEP(builder
,
2876 color_store
[1][chan
],
2878 fs_out_color
[1][chan
][i
] = ptr
;
2884 sampler
->destroy(sampler
);
2885 image
->destroy(image
);
2886 /* Loop over color outputs / color buffers to do blending.
2888 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2889 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2890 LLVMValueRef color_ptr
;
2891 LLVMValueRef stride
;
2892 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2894 boolean do_branch
= ((key
->depth
.enabled
2895 || key
->stencil
[0].enabled
2896 || key
->alpha
.enabled
)
2897 && !shader
->info
.base
.uses_kill
);
2899 color_ptr
= LLVMBuildLoad(builder
,
2900 LLVMBuildGEP(builder
, color_ptr_ptr
,
2904 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2906 stride
= LLVMBuildLoad(builder
,
2907 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2910 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2911 key
->cbuf_format
[cbuf
],
2912 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2913 context_ptr
, color_ptr
, stride
,
2914 partial_mask
, do_branch
);
2918 LLVMBuildRetVoid(builder
);
2920 gallivm_verify_function(gallivm
, function
);
2925 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
2929 debug_printf("fs variant %p:\n", (void *) key
);
2931 if (key
->flatshade
) {
2932 debug_printf("flatshade = 1\n");
2934 if (key
->multisample
) {
2935 debug_printf("multisample = 1\n");
2936 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
2938 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2939 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2940 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
2942 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2943 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2944 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
2946 if (key
->depth
.enabled
) {
2947 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2948 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2951 for (i
= 0; i
< 2; ++i
) {
2952 if (key
->stencil
[i
].enabled
) {
2953 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2954 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2955 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2956 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2957 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2958 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2962 if (key
->alpha
.enabled
) {
2963 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2966 if (key
->occlusion_count
) {
2967 debug_printf("occlusion_count = 1\n");
2970 if (key
->blend
.logicop_enable
) {
2971 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2973 else if (key
->blend
.rt
[0].blend_enable
) {
2974 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2975 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2976 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2977 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2978 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2979 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2981 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2982 if (key
->blend
.alpha_to_coverage
) {
2983 debug_printf("blend.alpha_to_coverage is enabled\n");
2985 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2986 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
2987 debug_printf("sampler[%u] = \n", i
);
2988 debug_printf(" .wrap = %s %s %s\n",
2989 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2990 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2991 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2992 debug_printf(" .min_img_filter = %s\n",
2993 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2994 debug_printf(" .min_mip_filter = %s\n",
2995 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2996 debug_printf(" .mag_img_filter = %s\n",
2997 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2998 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2999 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
3000 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
3001 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
3002 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
3003 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
3004 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
3006 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3007 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
3008 debug_printf("texture[%u] = \n", i
);
3009 debug_printf(" .format = %s\n",
3010 util_format_name(texture
->format
));
3011 debug_printf(" .target = %s\n",
3012 util_str_tex_target(texture
->target
, TRUE
));
3013 debug_printf(" .level_zero_only = %u\n",
3014 texture
->level_zero_only
);
3015 debug_printf(" .pot = %u %u %u\n",
3017 texture
->pot_height
,
3018 texture
->pot_depth
);
3020 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
3021 for (i
= 0; i
< key
->nr_images
; ++i
) {
3022 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
3023 debug_printf("image[%u] = \n", i
);
3024 debug_printf(" .format = %s\n",
3025 util_format_name(image
->format
));
3026 debug_printf(" .target = %s\n",
3027 util_str_tex_target(image
->target
, TRUE
));
3028 debug_printf(" .level_zero_only = %u\n",
3029 image
->level_zero_only
);
3030 debug_printf(" .pot = %u %u %u\n",
3039 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
3041 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3042 variant
->shader
->no
, variant
->no
);
3043 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
3044 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3046 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3047 dump_fs_variant_key(&variant
->key
);
3048 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3054 * Generate a new fragment shader variant from the shader code and
3055 * other state indicated by the key.
3057 static struct lp_fragment_shader_variant
*
3058 generate_variant(struct llvmpipe_context
*lp
,
3059 struct lp_fragment_shader
*shader
,
3060 const struct lp_fragment_shader_variant_key
*key
)
3062 struct lp_fragment_shader_variant
*variant
;
3063 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3064 boolean fullcolormask
;
3065 char module_name
[64];
3067 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3071 memset(variant
, 0, sizeof(*variant
));
3072 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3073 shader
->no
, shader
->variants_created
);
3075 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
3076 if (!variant
->gallivm
) {
3081 variant
->shader
= shader
;
3082 variant
->list_item_global
.base
= variant
;
3083 variant
->list_item_local
.base
= variant
;
3084 variant
->no
= shader
->variants_created
++;
3086 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3089 * Determine whether we are touching all channels in the color buffer.
3091 fullcolormask
= FALSE
;
3092 if (key
->nr_cbufs
== 1) {
3093 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3094 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3098 !key
->blend
.logicop_enable
&&
3099 !key
->blend
.rt
[0].blend_enable
&&
3101 !key
->stencil
[0].enabled
&&
3102 !key
->alpha
.enabled
&&
3103 !key
->blend
.alpha_to_coverage
&&
3104 !key
->depth
.enabled
&&
3105 !shader
->info
.base
.uses_kill
&&
3106 !shader
->info
.base
.writes_samplemask
3109 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3110 lp_debug_fs_variant(variant
);
3113 lp_jit_init_types(variant
);
3115 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3116 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3118 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3119 if (variant
->opaque
) {
3120 /* Specialized shader, which doesn't need to read the color buffer. */
3121 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3126 * Compile everything
3129 gallivm_compile_module(variant
->gallivm
);
3131 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3133 if (variant
->function
[RAST_EDGE_TEST
]) {
3134 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3135 gallivm_jit_function(variant
->gallivm
,
3136 variant
->function
[RAST_EDGE_TEST
]);
3139 if (variant
->function
[RAST_WHOLE
]) {
3140 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3141 gallivm_jit_function(variant
->gallivm
,
3142 variant
->function
[RAST_WHOLE
]);
3143 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3144 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3147 gallivm_free_ir(variant
->gallivm
);
3154 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3155 const struct pipe_shader_state
*templ
)
3157 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3158 struct lp_fragment_shader
*shader
;
3160 int nr_sampler_views
;
3164 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3168 shader
->no
= fs_no
++;
3169 make_empty_list(&shader
->variants
);
3171 shader
->base
.type
= templ
->type
;
3172 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3173 /* get/save the summary info for this shader */
3174 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3176 /* we need to keep a local copy of the tokens */
3177 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3179 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3180 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3183 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3184 if (shader
->draw_data
== NULL
) {
3185 FREE((void *) shader
->base
.tokens
);
3190 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3191 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3192 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3193 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3195 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3196 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3197 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3198 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3200 switch (shader
->info
.base
.input_interpolate
[i
]) {
3201 case TGSI_INTERPOLATE_CONSTANT
:
3202 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3204 case TGSI_INTERPOLATE_LINEAR
:
3205 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3207 case TGSI_INTERPOLATE_PERSPECTIVE
:
3208 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3210 case TGSI_INTERPOLATE_COLOR
:
3211 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3218 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3219 case TGSI_SEMANTIC_FACE
:
3220 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3222 case TGSI_SEMANTIC_POSITION
:
3223 /* Position was already emitted above
3225 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3226 shader
->inputs
[i
].src_index
= 0;
3230 /* XXX this is a completely pointless index map... */
3231 shader
->inputs
[i
].src_index
= i
+1;
3234 if (LP_DEBUG
& DEBUG_TGSI
) {
3236 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3237 shader
->no
, (void *) shader
);
3238 tgsi_dump(templ
->tokens
, 0);
3239 debug_printf("usage masks:\n");
3240 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3241 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3242 debug_printf(" IN[%u].%s%s%s%s\n",
3244 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3245 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3246 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3247 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3257 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3259 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3260 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3261 if (llvmpipe
->fs
== lp_fs
)
3264 draw_bind_fragment_shader(llvmpipe
->draw
,
3265 (lp_fs
? lp_fs
->draw_data
: NULL
));
3267 llvmpipe
->fs
= lp_fs
;
3269 llvmpipe
->dirty
|= LP_NEW_FS
;
3274 * Remove shader variant from two lists: the shader's variant list
3275 * and the context's variant list.
3278 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3279 struct lp_fragment_shader_variant
*variant
)
3281 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3282 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3283 "v total cached %u inst %u total inst %u\n",
3284 variant
->shader
->no
, variant
->no
,
3285 variant
->shader
->variants_created
,
3286 variant
->shader
->variants_cached
,
3287 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3290 gallivm_destroy(variant
->gallivm
);
3292 /* remove from shader's list */
3293 remove_from_list(&variant
->list_item_local
);
3294 variant
->shader
->variants_cached
--;
3296 /* remove from context's list */
3297 remove_from_list(&variant
->list_item_global
);
3298 lp
->nr_fs_variants
--;
3299 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3306 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3308 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3309 struct lp_fragment_shader
*shader
= fs
;
3310 struct lp_fs_variant_list_item
*li
;
3312 assert(fs
!= llvmpipe
->fs
);
3315 * XXX: we need to flush the context until we have some sort of reference
3316 * counting in fragment shaders as they may still be binned
3317 * Flushing alone might not sufficient we need to wait on it too.
3319 llvmpipe_finish(pipe
, __FUNCTION__
);
3321 /* Delete all the variants */
3322 li
= first_elem(&shader
->variants
);
3323 while(!at_end(&shader
->variants
, li
)) {
3324 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3325 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3329 /* Delete draw module's data */
3330 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3332 if (shader
->base
.ir
.nir
)
3333 ralloc_free(shader
->base
.ir
.nir
);
3334 assert(shader
->variants_cached
== 0);
3335 FREE((void *) shader
->base
.tokens
);
3342 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3343 enum pipe_shader_type shader
, uint index
,
3344 const struct pipe_constant_buffer
*cb
)
3346 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3347 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3349 assert(shader
< PIPE_SHADER_TYPES
);
3350 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3352 /* note: reference counting */
3353 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3356 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3357 debug_printf("Illegal set constant without bind flag\n");
3358 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3362 if (shader
== PIPE_SHADER_VERTEX
||
3363 shader
== PIPE_SHADER_GEOMETRY
||
3364 shader
== PIPE_SHADER_TESS_CTRL
||
3365 shader
== PIPE_SHADER_TESS_EVAL
) {
3366 /* Pass the constants to the 'draw' module */
3367 const unsigned size
= cb
? cb
->buffer_size
: 0;
3371 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3373 else if (cb
&& cb
->user_buffer
) {
3374 data
= (ubyte
*) cb
->user_buffer
;
3381 data
+= cb
->buffer_offset
;
3383 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3386 else if (shader
== PIPE_SHADER_COMPUTE
)
3387 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3389 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3391 if (cb
&& cb
->user_buffer
) {
3392 pipe_resource_reference(&constants
, NULL
);
3397 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3398 enum pipe_shader_type shader
, unsigned start_slot
,
3399 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3400 unsigned writable_bitmask
)
3402 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3404 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3405 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3407 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3409 if (shader
== PIPE_SHADER_VERTEX
||
3410 shader
== PIPE_SHADER_GEOMETRY
||
3411 shader
== PIPE_SHADER_TESS_CTRL
||
3412 shader
== PIPE_SHADER_TESS_EVAL
) {
3413 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3414 const ubyte
*data
= NULL
;
3415 if (buffer
&& buffer
->buffer
)
3416 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3418 data
+= buffer
->buffer_offset
;
3419 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3421 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3422 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3423 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3424 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3430 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3431 enum pipe_shader_type shader
, unsigned start_slot
,
3432 unsigned count
, const struct pipe_image_view
*images
)
3434 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3437 draw_flush(llvmpipe
->draw
);
3438 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3439 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3441 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3444 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3445 if (shader
== PIPE_SHADER_VERTEX
||
3446 shader
== PIPE_SHADER_GEOMETRY
||
3447 shader
== PIPE_SHADER_TESS_CTRL
||
3448 shader
== PIPE_SHADER_TESS_EVAL
) {
3449 draw_set_images(llvmpipe
->draw
,
3451 llvmpipe
->images
[shader
],
3452 start_slot
+ count
);
3453 } else if (shader
== PIPE_SHADER_COMPUTE
)
3454 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3456 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3460 * Return the blend factor equivalent to a destination alpha of one.
3462 static inline unsigned
3463 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3466 case PIPE_BLENDFACTOR_DST_ALPHA
:
3467 return PIPE_BLENDFACTOR_ONE
;
3468 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3469 return PIPE_BLENDFACTOR_ZERO
;
3470 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3472 return PIPE_BLENDFACTOR_ZERO
;
3474 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3482 * We need to generate several variants of the fragment pipeline to match
3483 * all the combinations of the contributing state atoms.
3485 * TODO: there is actually no reason to tie this to context state -- the
3486 * generated code could be cached globally in the screen.
3488 static struct lp_fragment_shader_variant_key
*
3489 make_variant_key(struct llvmpipe_context
*lp
,
3490 struct lp_fragment_shader
*shader
,
3494 struct lp_fragment_shader_variant_key
*key
;
3496 key
= (struct lp_fragment_shader_variant_key
*)store
;
3498 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3500 if (lp
->framebuffer
.zsbuf
) {
3501 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3502 const struct util_format_description
*zsbuf_desc
=
3503 util_format_description(zsbuf_format
);
3505 if (lp
->depth_stencil
->depth
.enabled
&&
3506 util_format_has_depth(zsbuf_desc
)) {
3507 key
->zsbuf_format
= zsbuf_format
;
3508 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3510 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3511 util_format_has_stencil(zsbuf_desc
)) {
3512 key
->zsbuf_format
= zsbuf_format
;
3513 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3515 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3516 key
->resource_1d
= TRUE
;
3518 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3522 * Propagate the depth clamp setting from the rasterizer state.
3523 * depth_clip == 0 implies depth clamping is enabled.
3525 * When clip_halfz is enabled, then always clamp the depth values.
3527 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3528 * clamp is always active in d3d10, regardless if depth clip is
3530 * (GL has an always-on [0,1] clamp on fs depth output instead
3531 * to ensure the depth values stay in range. Doesn't look like
3532 * we do that, though...)
3534 if (lp
->rasterizer
->clip_halfz
) {
3535 key
->depth_clamp
= 1;
3537 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3540 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3541 if (!lp
->framebuffer
.nr_cbufs
||
3542 !lp
->framebuffer
.cbufs
[0] ||
3543 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3544 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3546 if(key
->alpha
.enabled
)
3547 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3548 /* alpha.ref_value is passed in jit_context */
3550 key
->flatshade
= lp
->rasterizer
->flatshade
;
3551 key
->multisample
= lp
->rasterizer
->multisample
;
3552 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3553 key
->occlusion_count
= TRUE
;
3556 if (lp
->framebuffer
.nr_cbufs
) {
3557 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3560 key
->coverage_samples
= 1;
3561 if (key
->multisample
)
3562 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3563 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3565 if (!key
->blend
.independent_blend_enable
) {
3566 /* we always need independent blend otherwise the fixups below won't work */
3567 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3568 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3570 key
->blend
.independent_blend_enable
= 1;
3573 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3574 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3576 if (lp
->framebuffer
.cbufs
[i
]) {
3577 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3578 const struct util_format_description
*format_desc
;
3580 key
->cbuf_format
[i
] = format
;
3581 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3584 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3585 * mixing of 2d textures with height 1 and 1d textures, so make sure
3586 * we pick 1d if any cbuf or zsbuf is 1d.
3588 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3589 key
->resource_1d
= TRUE
;
3592 format_desc
= util_format_description(format
);
3593 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3594 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3597 * Mask out color channels not present in the color buffer.
3599 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3602 * Disable blend for integer formats.
3604 if (util_format_is_pure_integer(format
)) {
3605 blend_rt
->blend_enable
= 0;
3609 * Our swizzled render tiles always have an alpha channel, but the
3610 * linear render target format often does not, so force here the dst
3613 * This is not a mere optimization. Wrong results will be produced if
3614 * the dst alpha is used, the dst format does not have alpha, and the
3615 * previous rendering was not flushed from the swizzled to linear
3616 * buffer. For example, NonPowTwo DCT.
3618 * TODO: This should be generalized to all channels for better
3619 * performance, but only alpha causes correctness issues.
3621 * Also, force rgb/alpha func/factors match, to make AoS blending
3624 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3625 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3626 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3627 boolean clamped_zero
= !util_format_is_float(format
) &&
3628 !util_format_is_snorm(format
);
3629 blend_rt
->rgb_src_factor
=
3630 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3631 blend_rt
->rgb_dst_factor
=
3632 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3633 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3634 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3635 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3639 /* no color buffer for this fragment output */
3640 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3641 key
->cbuf_nr_samples
[i
] = 0;
3642 blend_rt
->colormask
= 0x0;
3643 blend_rt
->blend_enable
= 0;
3647 /* This value will be the same for all the variants of a given shader:
3649 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3651 struct lp_sampler_static_state
*fs_sampler
;
3653 fs_sampler
= key
->samplers
;
3655 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3657 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3658 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3659 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3660 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3665 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3666 * are dx10-style? Can't really have mixed opcodes, at least not
3667 * if we want to skip the holes here (without rescanning tgsi).
3669 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3670 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3671 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3673 * Note sview may exceed what's representable by file_mask.
3674 * This will still work, the only downside is that not actually
3675 * used views may be included in the shader key.
3677 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3678 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3679 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3684 key
->nr_sampler_views
= key
->nr_samplers
;
3685 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3686 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3687 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3688 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3693 struct lp_image_static_state
*lp_image
;
3694 lp_image
= lp_fs_variant_key_images(key
);
3695 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3696 for (i
= 0; i
< key
->nr_images
; ++i
) {
3697 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3698 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3699 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3708 * Update fragment shader state. This is called just prior to drawing
3709 * something when some fragment-related state has changed.
3712 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3714 struct lp_fragment_shader
*shader
= lp
->fs
;
3715 struct lp_fragment_shader_variant_key
*key
;
3716 struct lp_fragment_shader_variant
*variant
= NULL
;
3717 struct lp_fs_variant_list_item
*li
;
3718 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3720 key
= make_variant_key(lp
, shader
, store
);
3722 /* Search the variants for one which matches the key */
3723 li
= first_elem(&shader
->variants
);
3724 while(!at_end(&shader
->variants
, li
)) {
3725 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3733 /* Move this variant to the head of the list to implement LRU
3734 * deletion of shader's when we have too many.
3736 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3739 /* variant not found, create it now */
3742 unsigned variants_to_cull
;
3744 if (LP_DEBUG
& DEBUG_FS
) {
3745 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3748 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3751 /* First, check if we've exceeded the max number of shader variants.
3752 * If so, free 6.25% of them (the least recently used ones).
3754 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3756 if (variants_to_cull
||
3757 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3758 struct pipe_context
*pipe
= &lp
->pipe
;
3760 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3761 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3762 "\t%u instrs,\t%u instrs/variant\n",
3763 shader
->variants_cached
,
3764 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3765 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3769 * XXX: we need to flush the context until we have some sort of
3770 * reference counting in fragment shaders as they may still be binned
3771 * Flushing alone might not be sufficient we need to wait on it too.
3773 llvmpipe_finish(pipe
, __FUNCTION__
);
3776 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3777 * number of shader variants (potentially all of them) could be
3778 * pending for destruction on flush.
3781 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3782 struct lp_fs_variant_list_item
*item
;
3783 if (is_empty_list(&lp
->fs_variants_list
)) {
3786 item
= last_elem(&lp
->fs_variants_list
);
3789 llvmpipe_remove_shader_variant(lp
, item
->base
);
3794 * Generate the new variant.
3797 variant
= generate_variant(lp
, shader
, key
);
3800 LP_COUNT_ADD(llvm_compile_time
, dt
);
3801 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3803 /* Put the new variant into the list */
3805 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3806 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3807 lp
->nr_fs_variants
++;
3808 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3809 shader
->variants_cached
++;
3813 /* Bind this variant */
3814 lp_setup_set_fs_variant(lp
->setup
, variant
);
3822 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3824 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3825 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3826 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3828 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3830 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3831 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;