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 struct lp_build_for_loop_state loop_state
, sample_loop_state
;
333 struct lp_build_mask_context mask
;
335 * TODO: figure out if simple_shader optimization is really worthwile to
336 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
337 * code since tests tend to take another codepath than real shaders.
339 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
340 shader
->info
.base
.num_inputs
< 3 &&
341 shader
->info
.base
.num_instructions
< 8) && 0;
342 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
343 util_blend_state_is_dual(&key
->blend
, 0);
349 struct lp_bld_tgsi_system_values system_values
;
351 memset(&system_values
, 0, sizeof(system_values
));
353 /* truncate then sign extend. */
354 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
355 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
357 if (key
->depth
.enabled
||
358 key
->stencil
[0].enabled
) {
360 zs_format_desc
= util_format_description(key
->zsbuf_format
);
361 assert(zs_format_desc
);
363 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
364 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
365 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
366 if (shader
->info
.base
.writes_memory
)
367 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
368 else if (key
->alpha
.enabled
||
369 key
->blend
.alpha_to_coverage
||
370 shader
->info
.base
.uses_kill
||
371 shader
->info
.base
.writes_samplemask
) {
372 /* With alpha test and kill, can do the depth test early
373 * and hopefully eliminate some quads. But need to do a
374 * special deferred depth write once the final mask value
375 * is known. This only works though if there's either no
376 * stencil test or the stencil value isn't written.
378 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
379 (key
->stencil
[1].enabled
&&
380 key
->stencil
[1].writemask
)))
381 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
383 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
386 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
389 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
392 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
393 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
394 (key
->stencil
[1].enabled
&&
395 key
->stencil
[1].writemask
))))
396 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
402 vec_type
= lp_build_vec_type(gallivm
, type
);
403 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
405 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
406 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
407 /* convert scalar stencil refs into vectors */
408 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
409 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
411 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
412 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
414 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
415 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
417 memset(outputs
, 0, sizeof outputs
);
419 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
420 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
421 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
422 lp_build_vec_type(gallivm
,
427 if (dual_source_blend
) {
428 assert(key
->nr_cbufs
<= 1);
429 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
430 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
431 lp_build_vec_type(gallivm
,
437 lp_build_for_loop_begin(&loop_state
, gallivm
,
438 lp_build_const_int32(gallivm
, 0),
441 lp_build_const_int32(gallivm
, 1));
443 if (key
->multisample
) {
444 /* create shader execution mask by combining all sample masks. */
445 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
446 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, num_loop
, lp_build_const_int32(gallivm
, s
), "");
447 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
448 LLVMValueRef s_mask
= lp_build_pointer_get(builder
, mask_store
, s_mask_idx
);
452 mask_val
= LLVMBuildOr(builder
, s_mask
, mask_val
, "");
455 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
456 &loop_state
.counter
, 1, "mask_ptr");
457 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
460 /* 'mask' will control execution based on quad's pixel alive/killed state */
461 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
463 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
464 lp_build_mask_check(&mask
);
466 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
469 /* Create storage for recombining sample masks after early Z pass. */
470 LLVMValueRef s_mask_or
= lp_build_alloca(gallivm
, lp_build_int_vec_type(gallivm
, type
), "cov_mask_early_depth");
471 LLVMBuildStore(builder
, LLVMConstNull(lp_build_int_vec_type(gallivm
, type
)), s_mask_or
);
473 LLVMValueRef s_mask
= NULL
, s_mask_ptr
= NULL
;
474 /* Run early depth once per sample */
475 if (key
->multisample
) {
476 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
477 lp_build_const_int32(gallivm
, 0),
478 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
479 lp_build_const_int32(gallivm
, 1));
481 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
482 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
483 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
485 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
486 s_mask
= LLVMBuildAnd(builder
, s_mask
, mask_val
, "");
489 depth_ptr
= depth_base_ptr
;
490 if (key
->multisample
) {
491 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
492 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
495 if (depth_mode
& EARLY_DEPTH_TEST
) {
497 * Clamp according to ARB_depth_clamp semantics.
499 if (key
->depth_clamp
) {
500 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
503 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
504 zs_format_desc
, key
->resource_1d
,
505 depth_ptr
, depth_stride
,
506 &z_fb
, &s_fb
, loop_state
.counter
);
507 lp_build_depth_stencil_test(gallivm
,
512 key
->multisample
? NULL
: &mask
,
520 if (depth_mode
& EARLY_DEPTH_WRITE
) {
521 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
522 zs_format_desc
, key
->resource_1d
,
523 NULL
, NULL
, NULL
, loop_state
.counter
,
524 depth_ptr
, depth_stride
,
528 * Note mask check if stencil is enabled must be after ds write not after
529 * stencil test otherwise new stencil values may not get written if all
530 * fragments got killed by depth/stencil test.
532 if (!simple_shader
&& key
->stencil
[0].enabled
&& !key
->multisample
)
533 lp_build_mask_check(&mask
);
536 if (key
->multisample
) {
538 * Store the post-early Z coverage mask.
539 * Recombine the resulting coverage masks post early Z into the fragment
540 * shader execution mask.
542 LLVMValueRef tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
543 tmp_s_mask_or
= LLVMBuildOr(builder
, tmp_s_mask_or
, s_mask
, "");
544 LLVMBuildStore(builder
, tmp_s_mask_or
, s_mask_or
);
546 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
548 lp_build_for_loop_end(&sample_loop_state
);
550 /* recombined all the coverage masks in the shader exec mask. */
551 tmp_s_mask_or
= LLVMBuildLoad(builder
, s_mask_or
, "");
552 lp_build_mask_update(&mask
, tmp_s_mask_or
);
555 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, NULL
, NULL
);
557 struct lp_build_tgsi_params params
;
558 memset(¶ms
, 0, sizeof(params
));
562 params
.consts_ptr
= consts_ptr
;
563 params
.const_sizes_ptr
= num_consts_ptr
;
564 params
.system_values
= &system_values
;
565 params
.inputs
= interp
->inputs
;
566 params
.context_ptr
= context_ptr
;
567 params
.thread_data_ptr
= thread_data_ptr
;
568 params
.sampler
= sampler
;
569 params
.info
= &shader
->info
.base
;
570 params
.ssbo_ptr
= ssbo_ptr
;
571 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
572 params
.image
= image
;
574 /* Build the actual shader */
575 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
576 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
579 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
583 if (key
->alpha
.enabled
) {
584 int color0
= find_output_by_semantic(&shader
->info
.base
,
588 if (color0
!= -1 && outputs
[color0
][3]) {
589 const struct util_format_description
*cbuf_format_desc
;
590 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
591 LLVMValueRef alpha_ref_value
;
593 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
594 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
596 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
598 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
599 &mask
, alpha
, alpha_ref_value
,
600 (depth_mode
& LATE_DEPTH_TEST
) != 0);
604 /* Emulate Alpha to Coverage with Alpha test */
605 if (key
->blend
.alpha_to_coverage
) {
606 int color0
= find_output_by_semantic(&shader
->info
.base
,
610 if (color0
!= -1 && outputs
[color0
][3]) {
611 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
613 lp_build_alpha_to_coverage(gallivm
, type
,
615 (depth_mode
& LATE_DEPTH_TEST
) != 0);
619 if (shader
->info
.base
.writes_samplemask
) {
620 int smaski
= find_output_by_semantic(&shader
->info
.base
,
621 TGSI_SEMANTIC_SAMPLEMASK
,
624 struct lp_build_context smask_bld
;
625 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
628 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
630 * Pixel is alive according to the first sample in the mask.
632 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
633 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
634 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
635 lp_build_mask_update(&mask
, smask
);
638 if (key
->multisample
) {
639 /* execute depth test for each sample */
640 lp_build_for_loop_begin(&sample_loop_state
, gallivm
,
641 lp_build_const_int32(gallivm
, 0),
642 LLVMIntULT
, lp_build_const_int32(gallivm
, key
->coverage_samples
),
643 lp_build_const_int32(gallivm
, 1));
645 /* load the per-sample coverage mask */
646 LLVMValueRef s_mask_idx
= LLVMBuildMul(builder
, sample_loop_state
.counter
, num_loop
, "");
647 s_mask_idx
= LLVMBuildAdd(builder
, s_mask_idx
, loop_state
.counter
, "");
648 s_mask_ptr
= LLVMBuildGEP(builder
, mask_store
, &s_mask_idx
, 1, "");
650 /* combine the execution mask post fragment shader with the coverage mask. */
651 s_mask
= LLVMBuildLoad(builder
, s_mask_ptr
, "");
652 s_mask
= LLVMBuildAnd(builder
, s_mask
, lp_build_mask_value(&mask
), "");
655 depth_ptr
= depth_base_ptr
;
656 if (key
->multisample
) {
657 LLVMValueRef sample_offset
= LLVMBuildMul(builder
, sample_loop_state
.counter
, depth_sample_stride
, "");
658 depth_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &sample_offset
, 1, "");
662 if (depth_mode
& LATE_DEPTH_TEST
) {
663 int pos0
= find_output_by_semantic(&shader
->info
.base
,
664 TGSI_SEMANTIC_POSITION
,
666 int s_out
= find_output_by_semantic(&shader
->info
.base
,
667 TGSI_SEMANTIC_STENCIL
,
669 if (pos0
!= -1 && outputs
[pos0
][2]) {
670 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
673 * Clamp according to ARB_depth_clamp semantics.
675 if (key
->depth_clamp
) {
676 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
680 if (s_out
!= -1 && outputs
[s_out
][1]) {
681 /* there's only one value, and spec says to discard additional bits */
682 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
683 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
684 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
685 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
686 stencil_refs
[1] = stencil_refs
[0];
689 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
690 zs_format_desc
, key
->resource_1d
,
691 depth_ptr
, depth_stride
,
692 &z_fb
, &s_fb
, loop_state
.counter
);
694 lp_build_depth_stencil_test(gallivm
,
699 key
->multisample
? NULL
: &mask
,
707 if (depth_mode
& LATE_DEPTH_WRITE
) {
708 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
709 zs_format_desc
, key
->resource_1d
,
710 NULL
, NULL
, NULL
, loop_state
.counter
,
711 depth_ptr
, depth_stride
,
715 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
716 (depth_mode
& LATE_DEPTH_WRITE
))
718 /* Need to apply a reduced mask to the depth write. Reload the
719 * depth value, update from zs_value with the new mask value and
722 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
723 zs_format_desc
, key
->resource_1d
,
724 key
->multisample
? s_mask
: lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
725 depth_ptr
, depth_stride
,
729 if (key
->multisample
) {
730 /* store the sample mask for this loop */
731 LLVMBuildStore(builder
, s_mask
, s_mask_ptr
);
732 lp_build_for_loop_end(&sample_loop_state
);
736 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
738 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
739 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
740 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
742 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
743 if(outputs
[attrib
][chan
]) {
744 /* XXX: just initialize outputs to point at colors[] and
747 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
748 LLVMValueRef color_ptr
;
749 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
750 &loop_state
.counter
, 1, "");
751 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
752 LLVMBuildStore(builder
, out
, color_ptr
);
758 if (key
->occlusion_count
) {
759 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
760 lp_build_name(counter
, "counter");
761 lp_build_occlusion_count(gallivm
, type
,
762 lp_build_mask_value(&mask
), counter
);
765 mask_val
= lp_build_mask_end(&mask
);
766 if (!key
->multisample
)
767 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
768 lp_build_for_loop_end(&loop_state
);
773 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
775 * Fragment Shader outputs pixels in small 2x2 blocks
776 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
778 * However in memory pixels are stored in rows
779 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
781 * @param type fragment shader type (4x or 8x float)
782 * @param num_fs number of fs_src
783 * @param is_1d whether we're outputting to a 1d resource
784 * @param dst_channels number of output channels
785 * @param fs_src output from fragment shader
786 * @param dst pointer to store result
787 * @param pad_inline is channel padding inline or at end of row
788 * @return the number of dsts
791 generate_fs_twiddle(struct gallivm_state
*gallivm
,
794 unsigned dst_channels
,
795 LLVMValueRef fs_src
[][4],
799 LLVMValueRef src
[16];
805 unsigned pixels
= type
.length
/ 4;
806 unsigned reorder_group
;
807 unsigned src_channels
;
811 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
812 src_count
= num_fs
* src_channels
;
814 assert(pixels
== 2 || pixels
== 1);
815 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
818 * Transpose from SoA -> AoS
820 for (i
= 0; i
< num_fs
; ++i
) {
821 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
825 * Pick transformation options
832 if (dst_channels
== 1) {
838 } else if (dst_channels
== 2) {
842 } else if (dst_channels
> 2) {
849 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
855 * Split the src in half
858 for (i
= num_fs
; i
> 0; --i
) {
859 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
860 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
868 * Ensure pixels are in memory order
871 /* Twiddle pixels by reordering the array, e.g.:
873 * src_count = 8 -> 0 2 1 3 4 6 5 7
874 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
876 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
878 for (i
= 0; i
< src_count
; ++i
) {
879 unsigned group
= i
/ reorder_group
;
880 unsigned block
= (group
/ 4) * 4 * reorder_group
;
881 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
884 } else if (twiddle
) {
885 /* Twiddle pixels across elements of array */
887 * XXX: we should avoid this in some cases, but would need to tell
888 * lp_build_conv to reorder (or deal with it ourselves).
890 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
893 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
897 * Moves any padding between pixels to the end
898 * e.g. RGBXRGBX -> RGBRGBXX
901 unsigned char swizzles
[16];
902 unsigned elems
= pixels
* dst_channels
;
904 for (i
= 0; i
< type
.length
; ++i
) {
906 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
908 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
911 for (i
= 0; i
< src_count
; ++i
) {
912 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
921 * Untwiddle and transpose, much like the above.
922 * However, this is after conversion, so we get packed vectors.
923 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
924 * the vectors will look like:
925 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
926 * be swizzled here). Extending to 16bit should be trivial.
927 * Should also be extended to handle twice wide vectors with AVX2...
930 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
937 struct lp_type type64
, type16
, type32
;
938 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
939 LLVMBuilderRef builder
= gallivm
->builder
;
940 LLVMValueRef tmp
[4], shuf
[8];
941 for (j
= 0; j
< 2; j
++) {
942 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
943 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
944 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
945 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
948 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
949 assert(type
.width
== 8);
950 assert(type
.length
== 16);
952 type8_t
= lp_build_vec_type(gallivm
, type
);
957 type64_t
= lp_build_vec_type(gallivm
, type64
);
962 type16_t
= lp_build_vec_type(gallivm
, type16
);
967 type32_t
= lp_build_vec_type(gallivm
, type32
);
969 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
971 if (src_count
== 1) {
972 /* transpose was no-op, just untwiddle */
973 LLVMValueRef shuf_vec
;
974 shuf_vec
= LLVMConstVector(shuf
, 8);
975 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
976 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
977 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
978 } else if (src_count
== 2) {
979 LLVMValueRef shuf_vec
;
980 shuf_vec
= LLVMConstVector(shuf
, 4);
982 for (i
= 0; i
< 2; i
++) {
983 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
984 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
985 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
988 for (j
= 0; j
< 2; j
++) {
989 LLVMValueRef lo
, hi
, lo2
, hi2
;
991 * Note that if we only really have 3 valid channels (rgb)
992 * and we don't need alpha we could substitute a undef here
993 * for the respective channel (causing llvm to drop conversion
996 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
997 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
998 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
999 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
1000 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
1001 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
1002 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
1009 * Load an unswizzled block of pixels from memory
1012 load_unswizzled_block(struct gallivm_state
*gallivm
,
1013 LLVMValueRef base_ptr
,
1014 LLVMValueRef stride
,
1015 unsigned block_width
,
1016 unsigned block_height
,
1018 struct lp_type dst_type
,
1020 unsigned dst_alignment
)
1022 LLVMBuilderRef builder
= gallivm
->builder
;
1023 unsigned row_size
= dst_count
/ block_height
;
1026 /* Ensure block exactly fits into dst */
1027 assert((block_width
* block_height
) % dst_count
== 0);
1029 for (i
= 0; i
< dst_count
; ++i
) {
1030 unsigned x
= i
% row_size
;
1031 unsigned y
= i
/ row_size
;
1033 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
1034 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1036 LLVMValueRef gep
[2];
1037 LLVMValueRef dst_ptr
;
1039 gep
[0] = lp_build_const_int32(gallivm
, 0);
1040 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1042 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1043 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
1044 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
1046 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
1048 LLVMSetAlignment(dst
[i
], dst_alignment
);
1054 * Store an unswizzled block of pixels to memory
1057 store_unswizzled_block(struct gallivm_state
*gallivm
,
1058 LLVMValueRef base_ptr
,
1059 LLVMValueRef stride
,
1060 unsigned block_width
,
1061 unsigned block_height
,
1063 struct lp_type src_type
,
1065 unsigned src_alignment
)
1067 LLVMBuilderRef builder
= gallivm
->builder
;
1068 unsigned row_size
= src_count
/ block_height
;
1071 /* Ensure src exactly fits into block */
1072 assert((block_width
* block_height
) % src_count
== 0);
1074 for (i
= 0; i
< src_count
; ++i
) {
1075 unsigned x
= i
% row_size
;
1076 unsigned y
= i
/ row_size
;
1078 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
1079 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
1081 LLVMValueRef gep
[2];
1082 LLVMValueRef src_ptr
;
1084 gep
[0] = lp_build_const_int32(gallivm
, 0);
1085 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
1087 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1088 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1089 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1091 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1093 LLVMSetAlignment(src_ptr
, src_alignment
);
1099 * Checks if a format description is an arithmetic format
1101 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1103 static inline boolean
1104 is_arithmetic_format(const struct util_format_description
*format_desc
)
1106 boolean arith
= false;
1109 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1110 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1111 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1119 * Checks if this format requires special handling due to required expansion
1120 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1123 static inline boolean
1124 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1126 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1127 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1135 * Retrieves the type representing the memory layout for a format
1137 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1140 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1141 struct lp_type
* type
)
1146 if (format_expands_to_float_soa(format_desc
)) {
1147 /* just make this a uint with width of block */
1148 type
->floating
= false;
1149 type
->fixed
= false;
1152 type
->width
= format_desc
->block
.bits
;
1157 for (i
= 0; i
< 4; i
++)
1158 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1162 memset(type
, 0, sizeof(struct lp_type
));
1163 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1164 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1165 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1166 type
->norm
= format_desc
->channel
[chan
].normalized
;
1168 if (is_arithmetic_format(format_desc
)) {
1172 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1173 type
->width
+= format_desc
->channel
[i
].size
;
1176 type
->width
= format_desc
->channel
[chan
].size
;
1177 type
->length
= format_desc
->nr_channels
;
1183 * Retrieves the type for a format which is usable in the blending code.
1185 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1188 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1189 struct lp_type
* type
)
1194 if (format_expands_to_float_soa(format_desc
)) {
1195 /* always use ordinary floats for blending */
1196 type
->floating
= true;
1197 type
->fixed
= false;
1205 for (i
= 0; i
< 4; i
++)
1206 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1210 memset(type
, 0, sizeof(struct lp_type
));
1211 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1212 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1213 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1214 type
->norm
= format_desc
->channel
[chan
].normalized
;
1215 type
->width
= format_desc
->channel
[chan
].size
;
1216 type
->length
= format_desc
->nr_channels
;
1218 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1219 if (format_desc
->channel
[i
].size
> type
->width
)
1220 type
->width
= format_desc
->channel
[i
].size
;
1223 if (type
->floating
) {
1226 if (type
->width
<= 8) {
1228 } else if (type
->width
<= 16) {
1235 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1242 * Scale a normalized value from src_bits to dst_bits.
1244 * The exact calculation is
1246 * dst = iround(src * dst_mask / src_mask)
1248 * or with integer rounding
1250 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1254 * src_mask = (1 << src_bits) - 1
1255 * dst_mask = (1 << dst_bits) - 1
1257 * but we try to avoid division and multiplication through shifts.
1259 static inline LLVMValueRef
1260 scale_bits(struct gallivm_state
*gallivm
,
1264 struct lp_type src_type
)
1266 LLVMBuilderRef builder
= gallivm
->builder
;
1267 LLVMValueRef result
= src
;
1269 if (dst_bits
< src_bits
) {
1270 int delta_bits
= src_bits
- dst_bits
;
1272 if (delta_bits
<= dst_bits
) {
1274 * Approximate the rescaling with a single shift.
1276 * This gives the wrong rounding.
1279 result
= LLVMBuildLShr(builder
,
1281 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1286 * Try more accurate rescaling.
1290 * Drop the least significant bits to make space for the multiplication.
1292 * XXX: A better approach would be to use a wider integer type as intermediate. But
1293 * this is enough to convert alpha from 16bits -> 2 when rendering to
1294 * PIPE_FORMAT_R10G10B10A2_UNORM.
1296 result
= LLVMBuildLShr(builder
,
1298 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1302 result
= LLVMBuildMul(builder
,
1304 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1308 * Add a rounding term before the division.
1310 * TODO: Handle signed integers too.
1312 if (!src_type
.sign
) {
1313 result
= LLVMBuildAdd(builder
,
1315 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1320 * Approximate the division by src_mask with a src_bits shift.
1322 * Given the src has already been shifted by dst_bits, all we need
1323 * to do is to shift by the difference.
1326 result
= LLVMBuildLShr(builder
,
1328 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1332 } else if (dst_bits
> src_bits
) {
1334 int db
= dst_bits
- src_bits
;
1336 /* Shift left by difference in bits */
1337 result
= LLVMBuildShl(builder
,
1339 lp_build_const_int_vec(gallivm
, src_type
, db
),
1342 if (db
<= src_bits
) {
1343 /* Enough bits in src to fill the remainder */
1344 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1346 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1349 result
= LLVMBuildOr(builder
, result
, lower
, "");
1350 } else if (db
> src_bits
) {
1351 /* Need to repeatedly copy src bits to fill remainder in dst */
1354 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1355 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1357 result
= LLVMBuildOr(builder
,
1359 LLVMBuildLShr(builder
, result
, shuv
, ""),
1369 * If RT is a smallfloat (needing denorms) format
1372 have_smallfloat_format(struct lp_type dst_type
,
1373 enum pipe_format format
)
1375 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1376 /* due to format handling hacks this format doesn't have floating set
1377 * here (and actually has width set to 32 too) so special case this. */
1378 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1383 * Convert from memory format to blending format
1385 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1388 convert_to_blend_type(struct gallivm_state
*gallivm
,
1389 unsigned block_size
,
1390 const struct util_format_description
*src_fmt
,
1391 struct lp_type src_type
,
1392 struct lp_type dst_type
,
1393 LLVMValueRef
* src
, // and dst
1396 LLVMValueRef
*dst
= src
;
1397 LLVMBuilderRef builder
= gallivm
->builder
;
1398 struct lp_type blend_type
;
1399 struct lp_type mem_type
;
1401 unsigned pixels
= block_size
/ num_srcs
;
1405 * full custom path for packed floats and srgb formats - none of the later
1406 * functions would do anything useful, and given the lp_type representation they
1407 * can't be fixed. Should really have some SoA blend path for these kind of
1408 * formats rather than hacking them in here.
1410 if (format_expands_to_float_soa(src_fmt
)) {
1411 LLVMValueRef tmpsrc
[4];
1413 * This is pretty suboptimal for this case blending in SoA would be much
1414 * better, since conversion gets us SoA values so need to convert back.
1416 assert(src_type
.width
== 32 || src_type
.width
== 16);
1417 assert(dst_type
.floating
);
1418 assert(dst_type
.width
== 32);
1419 assert(dst_type
.length
% 4 == 0);
1420 assert(num_srcs
% 4 == 0);
1422 if (src_type
.width
== 16) {
1423 /* expand 4x16bit values to 4x32bit */
1424 struct lp_type type32x4
= src_type
;
1425 LLVMTypeRef ltype32x4
;
1426 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1427 type32x4
.width
= 32;
1428 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1429 for (i
= 0; i
< num_fetch
; i
++) {
1430 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1432 src_type
.width
= 32;
1434 for (i
= 0; i
< 4; i
++) {
1437 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1438 LLVMValueRef tmpsoa
[4];
1439 LLVMValueRef tmps
= tmpsrc
[i
];
1440 if (dst_type
.length
== 8) {
1441 LLVMValueRef shuffles
[8];
1443 /* fetch was 4 values but need 8-wide output values */
1444 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1446 * for 8-wide aos transpose would give us wrong order not matching
1447 * incoming converted fs values and mask. ARGH.
1449 for (j
= 0; j
< 4; j
++) {
1450 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1451 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1453 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1454 LLVMConstVector(shuffles
, 8), "");
1456 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1457 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1460 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1462 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1467 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1468 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1470 /* Is the format arithmetic */
1471 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1472 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1474 /* Pad if necessary */
1475 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1476 for (i
= 0; i
< num_srcs
; ++i
) {
1477 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1480 src_type
.length
= dst_type
.length
;
1483 /* Special case for half-floats */
1484 if (mem_type
.width
== 16 && mem_type
.floating
) {
1485 assert(blend_type
.width
== 32 && blend_type
.floating
);
1486 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1494 src_type
.width
= blend_type
.width
* blend_type
.length
;
1495 blend_type
.length
*= pixels
;
1496 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1498 for (i
= 0; i
< num_srcs
; ++i
) {
1499 LLVMValueRef chans
[4];
1500 LLVMValueRef res
= NULL
;
1502 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1504 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1506 unsigned sa
= src_fmt
->channel
[j
].shift
;
1507 #if UTIL_ARCH_LITTLE_ENDIAN
1508 unsigned from_lsb
= j
;
1510 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1513 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1515 /* Extract bits from source */
1516 chans
[j
] = LLVMBuildLShr(builder
,
1518 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1521 chans
[j
] = LLVMBuildAnd(builder
,
1523 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1527 if (src_type
.norm
) {
1528 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1529 blend_type
.width
, chans
[j
], src_type
);
1532 /* Insert bits into correct position */
1533 chans
[j
] = LLVMBuildShl(builder
,
1535 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1541 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1545 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1551 * Convert from blending format to memory format
1553 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1556 convert_from_blend_type(struct gallivm_state
*gallivm
,
1557 unsigned block_size
,
1558 const struct util_format_description
*src_fmt
,
1559 struct lp_type src_type
,
1560 struct lp_type dst_type
,
1561 LLVMValueRef
* src
, // and dst
1564 LLVMValueRef
* dst
= src
;
1566 struct lp_type mem_type
;
1567 struct lp_type blend_type
;
1568 LLVMBuilderRef builder
= gallivm
->builder
;
1569 unsigned pixels
= block_size
/ num_srcs
;
1573 * full custom path for packed floats and srgb formats - none of the later
1574 * functions would do anything useful, and given the lp_type representation they
1575 * can't be fixed. Should really have some SoA blend path for these kind of
1576 * formats rather than hacking them in here.
1578 if (format_expands_to_float_soa(src_fmt
)) {
1580 * This is pretty suboptimal for this case blending in SoA would be much
1581 * better - we need to transpose the AoS values back to SoA values for
1582 * conversion/packing.
1584 assert(src_type
.floating
);
1585 assert(src_type
.width
== 32);
1586 assert(src_type
.length
% 4 == 0);
1587 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1589 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1590 LLVMValueRef tmpsoa
[4], tmpdst
;
1591 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1592 /* really really need SoA here */
1594 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1595 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1598 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1602 if (src_type
.length
== 8) {
1603 LLVMValueRef tmpaos
, shuffles
[8];
1606 * for 8-wide aos transpose has given us wrong order not matching
1607 * output order. HMPF. Also need to split the output values manually.
1609 for (j
= 0; j
< 4; j
++) {
1610 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1611 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1613 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1614 LLVMConstVector(shuffles
, 8), "");
1615 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1616 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1622 if (dst_type
.width
== 16) {
1623 struct lp_type type16x8
= dst_type
;
1624 struct lp_type type32x4
= dst_type
;
1625 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1626 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1627 type16x8
.length
= 8;
1628 type32x4
.width
= 32;
1629 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1630 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1631 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1632 /* We could do vector truncation but it doesn't generate very good code */
1633 for (i
= 0; i
< num_fetch
; i
++) {
1634 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1635 src
[i
], lp_build_zero(gallivm
, type32x4
));
1636 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1637 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1638 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1644 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1645 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1647 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1649 /* Special case for half-floats */
1650 if (mem_type
.width
== 16 && mem_type
.floating
) {
1651 int length
= dst_type
.length
;
1652 assert(blend_type
.width
== 32 && blend_type
.floating
);
1654 dst_type
.length
= src_type
.length
;
1656 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1658 dst_type
.length
= length
;
1662 /* Remove any padding */
1663 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1664 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1666 for (i
= 0; i
< num_srcs
; ++i
) {
1667 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1671 /* No bit arithmetic to do */
1676 src_type
.length
= pixels
;
1677 src_type
.width
= blend_type
.length
* blend_type
.width
;
1678 dst_type
.length
= pixels
;
1680 for (i
= 0; i
< num_srcs
; ++i
) {
1681 LLVMValueRef chans
[4];
1682 LLVMValueRef res
= NULL
;
1684 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1686 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1688 unsigned sa
= src_fmt
->channel
[j
].shift
;
1689 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1690 #if UTIL_ARCH_LITTLE_ENDIAN
1691 unsigned from_lsb
= j
;
1693 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1696 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1698 for (k
= 0; k
< blend_type
.width
; ++k
) {
1703 chans
[j
] = LLVMBuildLShr(builder
,
1705 lp_build_const_int_vec(gallivm
, src_type
,
1706 from_lsb
* blend_type
.width
),
1709 chans
[j
] = LLVMBuildAnd(builder
,
1711 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1714 /* Scale down bits */
1715 if (src_type
.norm
) {
1716 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1717 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1718 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1719 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1720 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1721 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1725 chans
[j
] = LLVMBuildShl(builder
,
1727 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1730 sa
+= src_fmt
->channel
[j
].size
;
1735 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1739 assert (dst_type
.width
!= 24);
1741 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1747 * Convert alpha to same blend type as src
1750 convert_alpha(struct gallivm_state
*gallivm
,
1751 struct lp_type row_type
,
1752 struct lp_type alpha_type
,
1753 const unsigned block_size
,
1754 const unsigned block_height
,
1755 const unsigned src_count
,
1756 const unsigned dst_channels
,
1757 const bool pad_inline
,
1758 LLVMValueRef
* src_alpha
)
1760 LLVMBuilderRef builder
= gallivm
->builder
;
1762 unsigned length
= row_type
.length
;
1763 row_type
.length
= alpha_type
.length
;
1765 /* Twiddle the alpha to match pixels */
1766 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1769 * TODO this should use single lp_build_conv call for
1770 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1772 for (i
= 0; i
< block_height
; ++i
) {
1773 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1776 alpha_type
= row_type
;
1777 row_type
.length
= length
;
1779 /* If only one channel we can only need the single alpha value per pixel */
1780 if (src_count
== 1 && dst_channels
== 1) {
1782 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1784 /* If there are more srcs than rows then we need to split alpha up */
1785 if (src_count
> block_height
) {
1786 for (i
= src_count
; i
> 0; --i
) {
1787 unsigned pixels
= block_size
/ src_count
;
1788 unsigned idx
= i
- 1;
1790 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1791 (idx
* pixels
) % 4, pixels
);
1795 /* If there is a src for each pixel broadcast the alpha across whole row */
1796 if (src_count
== block_size
) {
1797 for (i
= 0; i
< src_count
; ++i
) {
1798 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1799 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1802 unsigned pixels
= block_size
/ src_count
;
1803 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1804 unsigned alpha_span
= 1;
1805 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1807 /* Check if we need 2 src_alphas for our shuffles */
1808 if (pixels
> alpha_type
.length
) {
1812 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1813 for (j
= 0; j
< row_type
.length
; ++j
) {
1814 if (j
< pixels
* channels
) {
1815 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1817 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1821 for (i
= 0; i
< src_count
; ++i
) {
1822 unsigned idx1
= i
, idx2
= i
;
1824 if (alpha_span
> 1){
1829 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1832 LLVMConstVector(shuffles
, row_type
.length
),
1841 * Generates the blend function for unswizzled colour buffers
1842 * Also generates the read & write from colour buffer
1845 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1847 struct lp_fragment_shader_variant
*variant
,
1848 enum pipe_format out_format
,
1849 unsigned int num_fs
,
1850 struct lp_type fs_type
,
1851 LLVMValueRef
* fs_mask
,
1852 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1853 LLVMValueRef context_ptr
,
1854 LLVMValueRef color_ptr
,
1855 LLVMValueRef stride
,
1856 unsigned partial_mask
,
1859 const unsigned alpha_channel
= 3;
1860 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1861 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1862 const unsigned block_size
= block_width
* block_height
;
1863 const unsigned lp_integer_vector_width
= 128;
1865 LLVMBuilderRef builder
= gallivm
->builder
;
1866 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1867 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1868 LLVMValueRef src_alpha
[4 * 4];
1869 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1870 LLVMValueRef src_mask
[4 * 4];
1871 LLVMValueRef src
[4 * 4];
1872 LLVMValueRef src1
[4 * 4];
1873 LLVMValueRef dst
[4 * 4];
1874 LLVMValueRef blend_color
;
1875 LLVMValueRef blend_alpha
;
1876 LLVMValueRef i32_zero
;
1877 LLVMValueRef check_mask
;
1878 LLVMValueRef undef_src_val
;
1880 struct lp_build_mask_context mask_ctx
;
1881 struct lp_type mask_type
;
1882 struct lp_type blend_type
;
1883 struct lp_type row_type
;
1884 struct lp_type dst_type
;
1885 struct lp_type ls_type
;
1887 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1888 unsigned vector_width
;
1889 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1890 unsigned dst_channels
;
1895 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1897 unsigned dst_alignment
;
1899 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1900 bool has_alpha
= false;
1901 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1902 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1904 const boolean is_1d
= variant
->key
.resource_1d
;
1905 boolean twiddle_after_convert
= FALSE
;
1906 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1907 LLVMValueRef fpstate
= 0;
1909 /* Get type from output format */
1910 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1911 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1914 * Technically this code should go into lp_build_smallfloat_to_float
1915 * and lp_build_float_to_smallfloat but due to the
1916 * http://llvm.org/bugs/show_bug.cgi?id=6393
1917 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1918 * So the ordering is important here and there shouldn't be any
1919 * llvm ir instrunctions in this function before
1920 * this, otherwise half-float format conversions won't work
1921 * (again due to llvm bug #6393).
1923 if (have_smallfloat_format(dst_type
, out_format
)) {
1924 /* We need to make sure that denorms are ok for half float
1926 fpstate
= lp_build_fpstate_get(gallivm
);
1927 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1930 mask_type
= lp_int32_vec4_type();
1931 mask_type
.length
= fs_type
.length
;
1933 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1934 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1937 /* Do not bother executing code when mask is empty.. */
1939 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1941 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1942 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1945 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1946 lp_build_mask_check(&mask_ctx
);
1949 partial_mask
|= !variant
->opaque
;
1950 i32_zero
= lp_build_const_int32(gallivm
, 0);
1952 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1954 row_type
.length
= fs_type
.length
;
1955 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1957 /* Compute correct swizzle and count channels */
1958 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1961 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1962 /* Ensure channel is used */
1963 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1967 /* Ensure not already written to (happens in case with GL_ALPHA) */
1968 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1972 /* Ensure we havn't already found all channels */
1973 if (dst_channels
>= out_format_desc
->nr_channels
) {
1977 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1980 if (i
== alpha_channel
) {
1985 if (format_expands_to_float_soa(out_format_desc
)) {
1987 * the code above can't work for layout_other
1988 * for srgb it would sort of work but we short-circuit swizzles, etc.
1989 * as that is done as part of unpack / pack.
1991 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1997 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2000 /* If 3 channels then pad to include alpha for 4 element transpose */
2001 if (dst_channels
== 3) {
2002 assert (!has_alpha
);
2003 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2004 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
2007 if (out_format_desc
->nr_channels
== 4) {
2010 * We use alpha from the color conversion, not separate one.
2011 * We had to include it for transpose, hence it will get converted
2012 * too (albeit when doing transpose after conversion, that would
2013 * no longer be the case necessarily).
2014 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2015 * otherwise we really have padding, not alpha, included.)
2022 * Load shader output
2024 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2025 /* Always load alpha for use in blending */
2028 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
2031 alpha
= undef_src_val
;
2034 /* Load each channel */
2035 for (j
= 0; j
< dst_channels
; ++j
) {
2036 assert(swizzle
[j
] < 4);
2038 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
2041 fs_src
[i
][j
] = undef_src_val
;
2045 /* If 3 channels then pad to include alpha for 4 element transpose */
2047 * XXX If we include that here maybe could actually use it instead of
2048 * separate alpha for blending?
2049 * (Difficult though we actually convert pad channels, not alpha.)
2051 if (dst_channels
== 3 && !has_alpha
) {
2052 fs_src
[i
][3] = alpha
;
2055 /* We split the row_mask and row_alpha as we want 128bit interleave */
2056 if (fs_type
.length
== 8) {
2057 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2059 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
2060 src_channels
, src_channels
);
2062 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2063 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2064 src_channels
, src_channels
);
2066 src_mask
[i
] = fs_mask
[i
];
2067 src_alpha
[i
] = alpha
;
2070 if (dual_source_blend
) {
2071 /* same as above except different src/dst, skip masks and comments... */
2072 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2075 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
2078 alpha
= undef_src_val
;
2081 for (j
= 0; j
< dst_channels
; ++j
) {
2082 assert(swizzle
[j
] < 4);
2084 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
2087 fs_src1
[i
][j
] = undef_src_val
;
2090 if (dst_channels
== 3 && !has_alpha
) {
2091 fs_src1
[i
][3] = alpha
;
2093 if (fs_type
.length
== 8) {
2094 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2095 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2096 src_channels
, src_channels
);
2098 src1_alpha
[i
] = alpha
;
2103 if (util_format_is_pure_integer(out_format
)) {
2105 * In this case fs_type was really ints or uints disguised as floats,
2108 fs_type
.floating
= 0;
2109 fs_type
.sign
= dst_type
.sign
;
2110 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2111 for (j
= 0; j
< dst_channels
; ++j
) {
2112 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2113 lp_build_vec_type(gallivm
, fs_type
), "");
2115 if (dst_channels
== 3 && !has_alpha
) {
2116 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2117 lp_build_vec_type(gallivm
, fs_type
), "");
2123 * We actually should generally do conversion first (for non-1d cases)
2124 * when the blend format is 8 or 16 bits. The reason is obvious,
2125 * there's 2 or 4 times less vectors to deal with for the interleave...
2126 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2127 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2128 * unpack only with 128bit vectors).
2129 * Note: for 16bit sizes really need matching pack conversion code
2131 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2132 twiddle_after_convert
= TRUE
;
2136 * Pixel twiddle from fragment shader order to memory order
2138 if (!twiddle_after_convert
) {
2139 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2140 dst_channels
, fs_src
, src
, pad_inline
);
2141 if (dual_source_blend
) {
2142 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2143 fs_src1
, src1
, pad_inline
);
2146 src_count
= num_fullblock_fs
* dst_channels
;
2148 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2149 * (AVX) turn out the same later when untwiddling/transpose (albeit
2150 * for true AVX2 path untwiddle needs to be different).
2151 * For now just order by colors first (so we can use unpack later).
2153 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2154 for (i
= 0; i
< dst_channels
; i
++) {
2155 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2156 if (dual_source_blend
) {
2157 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2163 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2164 if (src_count
!= num_fullblock_fs
* src_channels
) {
2165 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2166 row_type
.length
/= ds
;
2167 fs_type
.length
= row_type
.length
;
2170 blend_type
= row_type
;
2171 mask_type
.length
= 4;
2173 /* Convert src to row_type */
2174 if (dual_source_blend
) {
2175 struct lp_type old_row_type
= row_type
;
2176 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2177 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2180 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2183 /* If the rows are not an SSE vector, combine them to become SSE size! */
2184 if ((row_type
.width
* row_type
.length
) % 128) {
2185 unsigned bits
= row_type
.width
* row_type
.length
;
2188 assert(src_count
>= (vector_width
/ bits
));
2190 dst_count
= src_count
/ (vector_width
/ bits
);
2192 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2193 if (dual_source_blend
) {
2194 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2197 row_type
.length
*= combined
;
2198 src_count
/= combined
;
2200 bits
= row_type
.width
* row_type
.length
;
2201 assert(bits
== 128 || bits
== 256);
2204 if (twiddle_after_convert
) {
2205 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2206 if (dual_source_blend
) {
2207 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2212 * Blend Colour conversion
2214 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2215 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2216 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2217 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2218 &i32_zero
, 1, ""), "");
2221 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2223 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2225 * since blending is done with floats, there was no conversion.
2226 * However, the rules according to fixed point renderbuffers still
2227 * apply, that is we must clamp inputs to 0.0/1.0.
2228 * (This would apply to separate alpha conversion too but we currently
2229 * force has_alpha to be true.)
2230 * TODO: should skip this with "fake" blend, since post-blend conversion
2231 * will clamp anyway.
2232 * TODO: could also skip this if fragment color clamping is enabled. We
2233 * don't support it natively so it gets baked into the shader however, so
2234 * can't really tell here.
2236 struct lp_build_context f32_bld
;
2237 assert(row_type
.floating
);
2238 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2239 for (i
= 0; i
< src_count
; i
++) {
2240 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2242 if (dual_source_blend
) {
2243 for (i
= 0; i
< src_count
; i
++) {
2244 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2247 /* probably can't be different than row_type but better safe than sorry... */
2248 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2249 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2253 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2255 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2256 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2258 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2259 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2261 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2262 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2268 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2270 if (src_count
< block_height
) {
2271 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2272 } else if (src_count
> block_height
) {
2273 for (i
= src_count
; i
> 0; --i
) {
2274 unsigned pixels
= block_size
/ src_count
;
2275 unsigned idx
= i
- 1;
2277 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2278 (idx
* pixels
) % 4, pixels
);
2282 assert(mask_type
.width
== 32);
2284 for (i
= 0; i
< src_count
; ++i
) {
2285 unsigned pixels
= block_size
/ src_count
;
2286 unsigned pixel_width
= row_type
.width
* dst_channels
;
2288 if (pixel_width
== 24) {
2289 mask_type
.width
= 8;
2290 mask_type
.length
= vector_width
/ mask_type
.width
;
2292 mask_type
.length
= pixels
;
2293 mask_type
.width
= row_type
.width
* dst_channels
;
2296 * If mask_type width is smaller than 32bit, this doesn't quite
2297 * generate the most efficient code (could use some pack).
2299 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2300 lp_build_int_vec_type(gallivm
, mask_type
), "");
2302 mask_type
.length
*= dst_channels
;
2303 mask_type
.width
/= dst_channels
;
2306 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2307 lp_build_int_vec_type(gallivm
, mask_type
), "");
2308 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2315 struct lp_type alpha_type
= fs_type
;
2316 alpha_type
.length
= 4;
2317 convert_alpha(gallivm
, row_type
, alpha_type
,
2318 block_size
, block_height
,
2319 src_count
, dst_channels
,
2320 pad_inline
, src_alpha
);
2321 if (dual_source_blend
) {
2322 convert_alpha(gallivm
, row_type
, alpha_type
,
2323 block_size
, block_height
,
2324 src_count
, dst_channels
,
2325 pad_inline
, src1_alpha
);
2331 * Load dst from memory
2333 if (src_count
< block_height
) {
2334 dst_count
= block_height
;
2336 dst_count
= src_count
;
2339 dst_type
.length
*= block_size
/ dst_count
;
2341 if (format_expands_to_float_soa(out_format_desc
)) {
2343 * we need multiple values at once for the conversion, so can as well
2344 * load them vectorized here too instead of concatenating later.
2345 * (Still need concatenation later for 8-wide vectors).
2347 dst_count
= block_height
;
2348 dst_type
.length
= block_width
;
2352 * Compute the alignment of the destination pointer in bytes
2353 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2354 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2355 * 1d tex but can't distinguish here) so need to stick with per-pixel
2356 * alignment in this case.
2359 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2362 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2364 /* Force power-of-two alignment by extracting only the least-significant-bit */
2365 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2367 * Resource base and stride pointers are aligned to 16 bytes, so that's
2368 * the maximum alignment we can guarantee
2370 dst_alignment
= MIN2(16, dst_alignment
);
2374 if (dst_count
> src_count
) {
2375 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2376 util_is_power_of_two_or_zero(dst_type
.length
) &&
2377 dst_type
.length
* dst_type
.width
< 128) {
2379 * Never try to load values as 4xi8 which we will then
2380 * concatenate to larger vectors. This gives llvm a real
2381 * headache (the problem is the type legalizer (?) will
2382 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2383 * then the shuffles to concatenate are more or less impossible
2384 * - llvm is easily capable of generating a sequence of 32
2385 * pextrb/pinsrb instructions for that. Albeit it appears to
2386 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2387 * width to avoid the trouble (16bit seems not as bad, llvm
2388 * probably recognizes the load+shuffle as only one shuffle
2389 * is necessary, but we can do just the same anyway).
2391 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2397 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2398 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2399 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2400 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2405 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2406 dst
, ls_type
, dst_count
, dst_alignment
);
2411 * Convert from dst/output format to src/blending format.
2413 * This is necessary as we can only read 1 row from memory at a time,
2414 * so the minimum dst_count will ever be at this point is 4.
2416 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2417 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2418 * on all 16 pixels in that single vector at once.
2420 if (dst_count
> src_count
) {
2421 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2422 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2423 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2424 for (i
= 0; i
< dst_count
; i
++) {
2425 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2429 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2431 if (ls_type
.length
!= dst_type
.length
) {
2432 struct lp_type tmp_type
= dst_type
;
2433 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2434 for (i
= 0; i
< src_count
; i
++) {
2435 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2436 lp_build_vec_type(gallivm
, tmp_type
), "");
2444 /* XXX this is broken for RGB8 formats -
2445 * they get expanded from 12 to 16 elements (to include alpha)
2446 * by convert_to_blend_type then reduced to 15 instead of 12
2447 * by convert_from_blend_type (a simple fix though breaks A8...).
2448 * R16G16B16 also crashes differently however something going wrong
2449 * inside llvm handling npot vector sizes seemingly.
2450 * It seems some cleanup could be done here (like skipping conversion/blend
2453 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2454 row_type
, dst
, src_count
);
2457 * FIXME: Really should get logic ops / masks out of generic blend / row
2458 * format. Logic ops will definitely not work on the blend float format
2459 * used for SRGB here and I think OpenGL expects this to work as expected
2460 * (that is incoming values converted to srgb then logic op applied).
2462 for (i
= 0; i
< src_count
; ++i
) {
2463 dst
[i
] = lp_build_blend_aos(gallivm
,
2464 &variant
->key
.blend
,
2469 has_alpha
? NULL
: src_alpha
[i
],
2471 has_alpha
? NULL
: src1_alpha
[i
],
2473 partial_mask
? src_mask
[i
] : NULL
,
2475 has_alpha
? NULL
: blend_alpha
,
2477 pad_inline
? 4 : dst_channels
);
2480 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2481 row_type
, dst_type
, dst
, src_count
);
2483 /* Split the blend rows back to memory rows */
2484 if (dst_count
> src_count
) {
2485 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2487 if (src_count
== 1) {
2488 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2489 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2491 row_type
.length
/= 2;
2495 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2496 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2497 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2498 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2500 row_type
.length
/= 2;
2505 * Store blend result to memory
2508 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2509 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2512 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2513 dst
, dst_type
, dst_count
, dst_alignment
);
2516 if (have_smallfloat_format(dst_type
, out_format
)) {
2517 lp_build_fpstate_set(gallivm
, fpstate
);
2521 lp_build_mask_end(&mask_ctx
);
2527 * Generate the runtime callable function for the whole fragment pipeline.
2528 * Note that the function which we generate operates on a block of 16
2529 * pixels at at time. The block contains 2x2 quads. Each quad contains
2533 generate_fragment(struct llvmpipe_context
*lp
,
2534 struct lp_fragment_shader
*shader
,
2535 struct lp_fragment_shader_variant
*variant
,
2536 unsigned partial_mask
)
2538 struct gallivm_state
*gallivm
= variant
->gallivm
;
2539 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2540 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2542 struct lp_type fs_type
;
2543 struct lp_type blend_type
;
2544 LLVMTypeRef fs_elem_type
;
2545 LLVMTypeRef blend_vec_type
;
2546 LLVMTypeRef arg_types
[15];
2547 LLVMTypeRef func_type
;
2548 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2549 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2550 LLVMValueRef context_ptr
;
2553 LLVMValueRef a0_ptr
;
2554 LLVMValueRef dadx_ptr
;
2555 LLVMValueRef dady_ptr
;
2556 LLVMValueRef color_ptr_ptr
;
2557 LLVMValueRef stride_ptr
;
2558 LLVMValueRef color_sample_stride_ptr
;
2559 LLVMValueRef depth_ptr
;
2560 LLVMValueRef depth_stride
;
2561 LLVMValueRef depth_sample_stride
;
2562 LLVMValueRef mask_input
;
2563 LLVMValueRef thread_data_ptr
;
2564 LLVMBasicBlockRef block
;
2565 LLVMBuilderRef builder
;
2566 struct lp_build_sampler_soa
*sampler
;
2567 struct lp_build_image_soa
*image
;
2568 struct lp_build_interp_soa_context interp
;
2569 LLVMValueRef fs_mask
[16 / 4];
2570 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2571 LLVMValueRef function
;
2572 LLVMValueRef facing
;
2577 boolean cbuf0_write_all
;
2578 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2579 util_blend_state_is_dual(&key
->blend
, 0);
2581 assert(lp_native_vector_width
/ 32 >= 4);
2583 /* Adjust color input interpolation according to flatshade state:
2585 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2586 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2587 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2589 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2591 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2595 /* check if writes to cbuf[0] are to be copied to all cbufs */
2597 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2599 /* TODO: actually pick these based on the fs and color buffer
2600 * characteristics. */
2602 memset(&fs_type
, 0, sizeof fs_type
);
2603 fs_type
.floating
= TRUE
; /* floating point values */
2604 fs_type
.sign
= TRUE
; /* values are signed */
2605 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2606 fs_type
.width
= 32; /* 32-bit float */
2607 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2609 memset(&blend_type
, 0, sizeof blend_type
);
2610 blend_type
.floating
= FALSE
; /* values are integers */
2611 blend_type
.sign
= FALSE
; /* values are unsigned */
2612 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2613 blend_type
.width
= 8; /* 8-bit ubyte values */
2614 blend_type
.length
= 16; /* 16 elements per vector */
2617 * Generate the function prototype. Any change here must be reflected in
2618 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2621 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2623 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2625 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2626 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2628 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2629 arg_types
[1] = int32_type
; /* x */
2630 arg_types
[2] = int32_type
; /* y */
2631 arg_types
[3] = int32_type
; /* facing */
2632 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2633 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2634 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2635 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2636 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2637 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2638 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2639 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2640 arg_types
[12] = int32_type
; /* depth_stride */
2641 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2642 arg_types
[14] = int32_type
; /* depth sample stride */
2644 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2645 arg_types
, ARRAY_SIZE(arg_types
), 0);
2647 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2648 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2650 variant
->function
[partial_mask
] = function
;
2652 /* XXX: need to propagate noalias down into color param now we are
2653 * passing a pointer-to-pointer?
2655 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2656 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2657 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2659 context_ptr
= LLVMGetParam(function
, 0);
2660 x
= LLVMGetParam(function
, 1);
2661 y
= LLVMGetParam(function
, 2);
2662 facing
= LLVMGetParam(function
, 3);
2663 a0_ptr
= LLVMGetParam(function
, 4);
2664 dadx_ptr
= LLVMGetParam(function
, 5);
2665 dady_ptr
= LLVMGetParam(function
, 6);
2666 color_ptr_ptr
= LLVMGetParam(function
, 7);
2667 depth_ptr
= LLVMGetParam(function
, 8);
2668 mask_input
= LLVMGetParam(function
, 9);
2669 thread_data_ptr
= LLVMGetParam(function
, 10);
2670 stride_ptr
= LLVMGetParam(function
, 11);
2671 depth_stride
= LLVMGetParam(function
, 12);
2672 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2673 depth_sample_stride
= LLVMGetParam(function
, 14);
2675 lp_build_name(context_ptr
, "context");
2676 lp_build_name(x
, "x");
2677 lp_build_name(y
, "y");
2678 lp_build_name(a0_ptr
, "a0");
2679 lp_build_name(dadx_ptr
, "dadx");
2680 lp_build_name(dady_ptr
, "dady");
2681 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2682 lp_build_name(depth_ptr
, "depth");
2683 lp_build_name(mask_input
, "mask_input");
2684 lp_build_name(thread_data_ptr
, "thread_data");
2685 lp_build_name(stride_ptr
, "stride_ptr");
2686 lp_build_name(depth_stride
, "depth_stride");
2687 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2688 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2694 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2695 builder
= gallivm
->builder
;
2697 LLVMPositionBuilderAtEnd(builder
, block
);
2700 * Must not count ps invocations if there's a null shader.
2701 * (It would be ok to count with null shader if there's d/s tests,
2702 * but only if there's d/s buffers too, which is different
2703 * to implicit rasterization disable which must not depend
2704 * on the d/s buffers.)
2705 * Could use popcount on mask, but pixel accuracy is not required.
2706 * Could disable if there's no stats query, but maybe not worth it.
2708 if (shader
->info
.base
.num_instructions
> 1) {
2709 LLVMValueRef invocs
, val
;
2710 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2711 val
= LLVMBuildLoad(builder
, invocs
, "");
2712 val
= LLVMBuildAdd(builder
, val
,
2713 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2715 LLVMBuildStore(builder
, val
, invocs
);
2718 /* code generated texture sampling */
2719 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2720 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2722 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2723 /* for 1d resources only run "upper half" of stamp */
2724 if (key
->resource_1d
)
2728 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2729 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2730 LLVMValueRef num_loop_samp
= lp_build_const_int32(gallivm
, num_fs
* key
->coverage_samples
);
2731 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2732 num_loop_samp
, "mask_store");
2733 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2734 boolean pixel_center_integer
=
2735 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2738 * The shader input interpolation info is not explicitely baked in the
2739 * shader key, but everything it derives from (TGSI, and flatshade) is
2740 * already included in the shader key.
2742 lp_build_interp_soa_init(&interp
,
2744 shader
->info
.base
.num_inputs
,
2746 pixel_center_integer
,
2750 a0_ptr
, dadx_ptr
, dady_ptr
,
2753 for (i
= 0; i
< num_fs
; i
++) {
2754 if (key
->multisample
) {
2755 LLVMValueRef smask_val
= LLVMBuildLoad(builder
, lp_jit_context_sample_mask(gallivm
, context_ptr
), "");
2758 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2759 * store to the per sample mask storage. Or all of them together to generate
2760 * the fragment shader mask. (sample shading TODO).
2761 * Take the incoming state coverage mask into account.
2763 for (unsigned s
= 0; s
< key
->coverage_samples
; s
++) {
2764 LLVMValueRef sindexi
= lp_build_const_int32(gallivm
, i
+ (s
* num_fs
));
2765 LLVMValueRef sample_mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2766 &sindexi
, 1, "sample_mask_ptr");
2767 LLVMValueRef s_mask
= generate_quad_mask(gallivm
, fs_type
,
2768 i
*fs_type
.length
/4, s
, mask_input
);
2770 LLVMValueRef smask_bit
= LLVMBuildAnd(builder
, smask_val
, lp_build_const_int32(gallivm
, (1 << s
)), "");
2771 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntNE
, smask_bit
, lp_build_const_int32(gallivm
, 0), "");
2772 smask_bit
= LLVMBuildSExt(builder
, cmp
, int32_type
, "");
2773 smask_bit
= lp_build_broadcast(gallivm
, mask_type
, smask_bit
);
2775 s_mask
= LLVMBuildAnd(builder
, s_mask
, smask_bit
, "");
2776 LLVMBuildStore(builder
, s_mask
, sample_mask_ptr
);
2780 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2781 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2782 &indexi
, 1, "mask_ptr");
2785 mask
= generate_quad_mask(gallivm
, fs_type
,
2786 i
*fs_type
.length
/4, 0, mask_input
);
2789 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2791 LLVMBuildStore(builder
, mask
, mask_ptr
);
2795 generate_fs_loop(gallivm
,
2804 mask_store
, /* output */
2808 depth_sample_stride
,
2812 for (i
= 0; i
< num_fs
; i
++) {
2813 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2814 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2816 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2817 /* This is fucked up need to reorganize things */
2818 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2819 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2820 ptr
= LLVMBuildGEP(builder
,
2821 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2823 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2826 if (dual_source_blend
) {
2827 /* only support one dual source blend target hence always use output 1 */
2828 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2829 ptr
= LLVMBuildGEP(builder
,
2830 color_store
[1][chan
],
2832 fs_out_color
[1][chan
][i
] = ptr
;
2838 sampler
->destroy(sampler
);
2839 image
->destroy(image
);
2840 /* Loop over color outputs / color buffers to do blending.
2842 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2843 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2844 LLVMValueRef color_ptr
;
2845 LLVMValueRef stride
;
2846 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2848 boolean do_branch
= ((key
->depth
.enabled
2849 || key
->stencil
[0].enabled
2850 || key
->alpha
.enabled
)
2851 && !shader
->info
.base
.uses_kill
);
2853 color_ptr
= LLVMBuildLoad(builder
,
2854 LLVMBuildGEP(builder
, color_ptr_ptr
,
2858 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2860 stride
= LLVMBuildLoad(builder
,
2861 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2864 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2865 key
->cbuf_format
[cbuf
],
2866 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2867 context_ptr
, color_ptr
, stride
,
2868 partial_mask
, do_branch
);
2872 LLVMBuildRetVoid(builder
);
2874 gallivm_verify_function(gallivm
, function
);
2879 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
2883 debug_printf("fs variant %p:\n", (void *) key
);
2885 if (key
->flatshade
) {
2886 debug_printf("flatshade = 1\n");
2888 if (key
->multisample
) {
2889 debug_printf("multisample = 1\n");
2890 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
2892 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2893 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2894 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
2896 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2897 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2898 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
2900 if (key
->depth
.enabled
) {
2901 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2902 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2905 for (i
= 0; i
< 2; ++i
) {
2906 if (key
->stencil
[i
].enabled
) {
2907 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2908 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2909 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2910 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2911 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2912 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2916 if (key
->alpha
.enabled
) {
2917 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2920 if (key
->occlusion_count
) {
2921 debug_printf("occlusion_count = 1\n");
2924 if (key
->blend
.logicop_enable
) {
2925 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2927 else if (key
->blend
.rt
[0].blend_enable
) {
2928 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2929 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2930 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2931 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2932 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2933 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2935 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2936 if (key
->blend
.alpha_to_coverage
) {
2937 debug_printf("blend.alpha_to_coverage is enabled\n");
2939 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2940 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
2941 debug_printf("sampler[%u] = \n", i
);
2942 debug_printf(" .wrap = %s %s %s\n",
2943 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2944 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2945 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2946 debug_printf(" .min_img_filter = %s\n",
2947 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2948 debug_printf(" .min_mip_filter = %s\n",
2949 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2950 debug_printf(" .mag_img_filter = %s\n",
2951 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2952 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2953 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
2954 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2955 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2956 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2957 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2958 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2960 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2961 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
2962 debug_printf("texture[%u] = \n", i
);
2963 debug_printf(" .format = %s\n",
2964 util_format_name(texture
->format
));
2965 debug_printf(" .target = %s\n",
2966 util_str_tex_target(texture
->target
, TRUE
));
2967 debug_printf(" .level_zero_only = %u\n",
2968 texture
->level_zero_only
);
2969 debug_printf(" .pot = %u %u %u\n",
2971 texture
->pot_height
,
2972 texture
->pot_depth
);
2974 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
2975 for (i
= 0; i
< key
->nr_images
; ++i
) {
2976 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
2977 debug_printf("image[%u] = \n", i
);
2978 debug_printf(" .format = %s\n",
2979 util_format_name(image
->format
));
2980 debug_printf(" .target = %s\n",
2981 util_str_tex_target(image
->target
, TRUE
));
2982 debug_printf(" .level_zero_only = %u\n",
2983 image
->level_zero_only
);
2984 debug_printf(" .pot = %u %u %u\n",
2993 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
2995 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2996 variant
->shader
->no
, variant
->no
);
2997 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
2998 tgsi_dump(variant
->shader
->base
.tokens
, 0);
3000 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
3001 dump_fs_variant_key(&variant
->key
);
3002 debug_printf("variant->opaque = %u\n", variant
->opaque
);
3008 * Generate a new fragment shader variant from the shader code and
3009 * other state indicated by the key.
3011 static struct lp_fragment_shader_variant
*
3012 generate_variant(struct llvmpipe_context
*lp
,
3013 struct lp_fragment_shader
*shader
,
3014 const struct lp_fragment_shader_variant_key
*key
)
3016 struct lp_fragment_shader_variant
*variant
;
3017 const struct util_format_description
*cbuf0_format_desc
= NULL
;
3018 boolean fullcolormask
;
3019 char module_name
[64];
3021 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
3025 memset(variant
, 0, sizeof(*variant
));
3026 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
3027 shader
->no
, shader
->variants_created
);
3029 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
3030 if (!variant
->gallivm
) {
3035 variant
->shader
= shader
;
3036 variant
->list_item_global
.base
= variant
;
3037 variant
->list_item_local
.base
= variant
;
3038 variant
->no
= shader
->variants_created
++;
3040 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3043 * Determine whether we are touching all channels in the color buffer.
3045 fullcolormask
= FALSE
;
3046 if (key
->nr_cbufs
== 1) {
3047 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
3048 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
3052 !key
->blend
.logicop_enable
&&
3053 !key
->blend
.rt
[0].blend_enable
&&
3055 !key
->stencil
[0].enabled
&&
3056 !key
->alpha
.enabled
&&
3057 !key
->blend
.alpha_to_coverage
&&
3058 !key
->depth
.enabled
&&
3059 !shader
->info
.base
.uses_kill
&&
3060 !shader
->info
.base
.writes_samplemask
3063 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3064 lp_debug_fs_variant(variant
);
3067 lp_jit_init_types(variant
);
3069 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
3070 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
3072 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
3073 if (variant
->opaque
) {
3074 /* Specialized shader, which doesn't need to read the color buffer. */
3075 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
3080 * Compile everything
3083 gallivm_compile_module(variant
->gallivm
);
3085 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
3087 if (variant
->function
[RAST_EDGE_TEST
]) {
3088 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
3089 gallivm_jit_function(variant
->gallivm
,
3090 variant
->function
[RAST_EDGE_TEST
]);
3093 if (variant
->function
[RAST_WHOLE
]) {
3094 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
3095 gallivm_jit_function(variant
->gallivm
,
3096 variant
->function
[RAST_WHOLE
]);
3097 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
3098 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
3101 gallivm_free_ir(variant
->gallivm
);
3108 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
3109 const struct pipe_shader_state
*templ
)
3111 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3112 struct lp_fragment_shader
*shader
;
3114 int nr_sampler_views
;
3118 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3122 shader
->no
= fs_no
++;
3123 make_empty_list(&shader
->variants
);
3125 shader
->base
.type
= templ
->type
;
3126 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3127 /* get/save the summary info for this shader */
3128 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3130 /* we need to keep a local copy of the tokens */
3131 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3133 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3134 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3137 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3138 if (shader
->draw_data
== NULL
) {
3139 FREE((void *) shader
->base
.tokens
);
3144 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3145 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3146 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3147 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3149 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3150 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3151 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3152 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3154 switch (shader
->info
.base
.input_interpolate
[i
]) {
3155 case TGSI_INTERPOLATE_CONSTANT
:
3156 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3158 case TGSI_INTERPOLATE_LINEAR
:
3159 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3161 case TGSI_INTERPOLATE_PERSPECTIVE
:
3162 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3164 case TGSI_INTERPOLATE_COLOR
:
3165 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3172 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3173 case TGSI_SEMANTIC_FACE
:
3174 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3176 case TGSI_SEMANTIC_POSITION
:
3177 /* Position was already emitted above
3179 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3180 shader
->inputs
[i
].src_index
= 0;
3184 /* XXX this is a completely pointless index map... */
3185 shader
->inputs
[i
].src_index
= i
+1;
3188 if (LP_DEBUG
& DEBUG_TGSI
) {
3190 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3191 shader
->no
, (void *) shader
);
3192 tgsi_dump(templ
->tokens
, 0);
3193 debug_printf("usage masks:\n");
3194 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3195 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3196 debug_printf(" IN[%u].%s%s%s%s\n",
3198 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3199 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3200 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3201 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3211 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3213 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3214 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3215 if (llvmpipe
->fs
== lp_fs
)
3218 draw_bind_fragment_shader(llvmpipe
->draw
,
3219 (lp_fs
? lp_fs
->draw_data
: NULL
));
3221 llvmpipe
->fs
= lp_fs
;
3223 llvmpipe
->dirty
|= LP_NEW_FS
;
3228 * Remove shader variant from two lists: the shader's variant list
3229 * and the context's variant list.
3232 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3233 struct lp_fragment_shader_variant
*variant
)
3235 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3236 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3237 "v total cached %u inst %u total inst %u\n",
3238 variant
->shader
->no
, variant
->no
,
3239 variant
->shader
->variants_created
,
3240 variant
->shader
->variants_cached
,
3241 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3244 gallivm_destroy(variant
->gallivm
);
3246 /* remove from shader's list */
3247 remove_from_list(&variant
->list_item_local
);
3248 variant
->shader
->variants_cached
--;
3250 /* remove from context's list */
3251 remove_from_list(&variant
->list_item_global
);
3252 lp
->nr_fs_variants
--;
3253 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3260 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3262 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3263 struct lp_fragment_shader
*shader
= fs
;
3264 struct lp_fs_variant_list_item
*li
;
3266 assert(fs
!= llvmpipe
->fs
);
3269 * XXX: we need to flush the context until we have some sort of reference
3270 * counting in fragment shaders as they may still be binned
3271 * Flushing alone might not sufficient we need to wait on it too.
3273 llvmpipe_finish(pipe
, __FUNCTION__
);
3275 /* Delete all the variants */
3276 li
= first_elem(&shader
->variants
);
3277 while(!at_end(&shader
->variants
, li
)) {
3278 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3279 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3283 /* Delete draw module's data */
3284 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3286 if (shader
->base
.ir
.nir
)
3287 ralloc_free(shader
->base
.ir
.nir
);
3288 assert(shader
->variants_cached
== 0);
3289 FREE((void *) shader
->base
.tokens
);
3296 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3297 enum pipe_shader_type shader
, uint index
,
3298 const struct pipe_constant_buffer
*cb
)
3300 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3301 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3303 assert(shader
< PIPE_SHADER_TYPES
);
3304 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3306 /* note: reference counting */
3307 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3310 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3311 debug_printf("Illegal set constant without bind flag\n");
3312 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3316 if (shader
== PIPE_SHADER_VERTEX
||
3317 shader
== PIPE_SHADER_GEOMETRY
||
3318 shader
== PIPE_SHADER_TESS_CTRL
||
3319 shader
== PIPE_SHADER_TESS_EVAL
) {
3320 /* Pass the constants to the 'draw' module */
3321 const unsigned size
= cb
? cb
->buffer_size
: 0;
3325 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3327 else if (cb
&& cb
->user_buffer
) {
3328 data
= (ubyte
*) cb
->user_buffer
;
3335 data
+= cb
->buffer_offset
;
3337 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3340 else if (shader
== PIPE_SHADER_COMPUTE
)
3341 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3343 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3345 if (cb
&& cb
->user_buffer
) {
3346 pipe_resource_reference(&constants
, NULL
);
3351 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3352 enum pipe_shader_type shader
, unsigned start_slot
,
3353 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3354 unsigned writable_bitmask
)
3356 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3358 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3359 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3361 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3363 if (shader
== PIPE_SHADER_VERTEX
||
3364 shader
== PIPE_SHADER_GEOMETRY
||
3365 shader
== PIPE_SHADER_TESS_CTRL
||
3366 shader
== PIPE_SHADER_TESS_EVAL
) {
3367 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3368 const ubyte
*data
= NULL
;
3369 if (buffer
&& buffer
->buffer
)
3370 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3372 data
+= buffer
->buffer_offset
;
3373 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3375 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3376 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3377 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3378 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3384 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3385 enum pipe_shader_type shader
, unsigned start_slot
,
3386 unsigned count
, const struct pipe_image_view
*images
)
3388 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3391 draw_flush(llvmpipe
->draw
);
3392 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3393 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3395 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3398 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3399 if (shader
== PIPE_SHADER_VERTEX
||
3400 shader
== PIPE_SHADER_GEOMETRY
||
3401 shader
== PIPE_SHADER_TESS_CTRL
||
3402 shader
== PIPE_SHADER_TESS_EVAL
) {
3403 draw_set_images(llvmpipe
->draw
,
3405 llvmpipe
->images
[shader
],
3406 start_slot
+ count
);
3407 } else if (shader
== PIPE_SHADER_COMPUTE
)
3408 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3410 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3414 * Return the blend factor equivalent to a destination alpha of one.
3416 static inline unsigned
3417 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3420 case PIPE_BLENDFACTOR_DST_ALPHA
:
3421 return PIPE_BLENDFACTOR_ONE
;
3422 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3423 return PIPE_BLENDFACTOR_ZERO
;
3424 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3426 return PIPE_BLENDFACTOR_ZERO
;
3428 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3436 * We need to generate several variants of the fragment pipeline to match
3437 * all the combinations of the contributing state atoms.
3439 * TODO: there is actually no reason to tie this to context state -- the
3440 * generated code could be cached globally in the screen.
3442 static struct lp_fragment_shader_variant_key
*
3443 make_variant_key(struct llvmpipe_context
*lp
,
3444 struct lp_fragment_shader
*shader
,
3448 struct lp_fragment_shader_variant_key
*key
;
3450 key
= (struct lp_fragment_shader_variant_key
*)store
;
3452 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3454 if (lp
->framebuffer
.zsbuf
) {
3455 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3456 const struct util_format_description
*zsbuf_desc
=
3457 util_format_description(zsbuf_format
);
3459 if (lp
->depth_stencil
->depth
.enabled
&&
3460 util_format_has_depth(zsbuf_desc
)) {
3461 key
->zsbuf_format
= zsbuf_format
;
3462 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3464 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3465 util_format_has_stencil(zsbuf_desc
)) {
3466 key
->zsbuf_format
= zsbuf_format
;
3467 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3469 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3470 key
->resource_1d
= TRUE
;
3472 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3476 * Propagate the depth clamp setting from the rasterizer state.
3477 * depth_clip == 0 implies depth clamping is enabled.
3479 * When clip_halfz is enabled, then always clamp the depth values.
3481 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3482 * clamp is always active in d3d10, regardless if depth clip is
3484 * (GL has an always-on [0,1] clamp on fs depth output instead
3485 * to ensure the depth values stay in range. Doesn't look like
3486 * we do that, though...)
3488 if (lp
->rasterizer
->clip_halfz
) {
3489 key
->depth_clamp
= 1;
3491 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3494 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3495 if (!lp
->framebuffer
.nr_cbufs
||
3496 !lp
->framebuffer
.cbufs
[0] ||
3497 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3498 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3500 if(key
->alpha
.enabled
)
3501 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3502 /* alpha.ref_value is passed in jit_context */
3504 key
->flatshade
= lp
->rasterizer
->flatshade
;
3505 key
->multisample
= lp
->rasterizer
->multisample
;
3506 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3507 key
->occlusion_count
= TRUE
;
3510 if (lp
->framebuffer
.nr_cbufs
) {
3511 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3514 key
->coverage_samples
= 1;
3515 if (key
->multisample
)
3516 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3517 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3519 if (!key
->blend
.independent_blend_enable
) {
3520 /* we always need independent blend otherwise the fixups below won't work */
3521 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3522 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3524 key
->blend
.independent_blend_enable
= 1;
3527 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3528 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3530 if (lp
->framebuffer
.cbufs
[i
]) {
3531 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3532 const struct util_format_description
*format_desc
;
3534 key
->cbuf_format
[i
] = format
;
3535 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3538 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3539 * mixing of 2d textures with height 1 and 1d textures, so make sure
3540 * we pick 1d if any cbuf or zsbuf is 1d.
3542 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3543 key
->resource_1d
= TRUE
;
3546 format_desc
= util_format_description(format
);
3547 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3548 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3551 * Mask out color channels not present in the color buffer.
3553 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3556 * Disable blend for integer formats.
3558 if (util_format_is_pure_integer(format
)) {
3559 blend_rt
->blend_enable
= 0;
3563 * Our swizzled render tiles always have an alpha channel, but the
3564 * linear render target format often does not, so force here the dst
3567 * This is not a mere optimization. Wrong results will be produced if
3568 * the dst alpha is used, the dst format does not have alpha, and the
3569 * previous rendering was not flushed from the swizzled to linear
3570 * buffer. For example, NonPowTwo DCT.
3572 * TODO: This should be generalized to all channels for better
3573 * performance, but only alpha causes correctness issues.
3575 * Also, force rgb/alpha func/factors match, to make AoS blending
3578 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3579 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3580 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3581 boolean clamped_zero
= !util_format_is_float(format
) &&
3582 !util_format_is_snorm(format
);
3583 blend_rt
->rgb_src_factor
=
3584 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3585 blend_rt
->rgb_dst_factor
=
3586 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3587 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3588 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3589 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3593 /* no color buffer for this fragment output */
3594 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3595 key
->cbuf_nr_samples
[i
] = 0;
3596 blend_rt
->colormask
= 0x0;
3597 blend_rt
->blend_enable
= 0;
3601 /* This value will be the same for all the variants of a given shader:
3603 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3605 struct lp_sampler_static_state
*fs_sampler
;
3607 fs_sampler
= key
->samplers
;
3609 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3611 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3612 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3613 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3614 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3619 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3620 * are dx10-style? Can't really have mixed opcodes, at least not
3621 * if we want to skip the holes here (without rescanning tgsi).
3623 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3624 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3625 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3627 * Note sview may exceed what's representable by file_mask.
3628 * This will still work, the only downside is that not actually
3629 * used views may be included in the shader key.
3631 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3632 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3633 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3638 key
->nr_sampler_views
= key
->nr_samplers
;
3639 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3640 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3641 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3642 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3647 struct lp_image_static_state
*lp_image
;
3648 lp_image
= lp_fs_variant_key_images(key
);
3649 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3650 for (i
= 0; i
< key
->nr_images
; ++i
) {
3651 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3652 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3653 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3662 * Update fragment shader state. This is called just prior to drawing
3663 * something when some fragment-related state has changed.
3666 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3668 struct lp_fragment_shader
*shader
= lp
->fs
;
3669 struct lp_fragment_shader_variant_key
*key
;
3670 struct lp_fragment_shader_variant
*variant
= NULL
;
3671 struct lp_fs_variant_list_item
*li
;
3672 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3674 key
= make_variant_key(lp
, shader
, store
);
3676 /* Search the variants for one which matches the key */
3677 li
= first_elem(&shader
->variants
);
3678 while(!at_end(&shader
->variants
, li
)) {
3679 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3687 /* Move this variant to the head of the list to implement LRU
3688 * deletion of shader's when we have too many.
3690 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3693 /* variant not found, create it now */
3696 unsigned variants_to_cull
;
3698 if (LP_DEBUG
& DEBUG_FS
) {
3699 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3702 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3705 /* First, check if we've exceeded the max number of shader variants.
3706 * If so, free 6.25% of them (the least recently used ones).
3708 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3710 if (variants_to_cull
||
3711 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3712 struct pipe_context
*pipe
= &lp
->pipe
;
3714 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3715 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3716 "\t%u instrs,\t%u instrs/variant\n",
3717 shader
->variants_cached
,
3718 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3719 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3723 * XXX: we need to flush the context until we have some sort of
3724 * reference counting in fragment shaders as they may still be binned
3725 * Flushing alone might not be sufficient we need to wait on it too.
3727 llvmpipe_finish(pipe
, __FUNCTION__
);
3730 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3731 * number of shader variants (potentially all of them) could be
3732 * pending for destruction on flush.
3735 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3736 struct lp_fs_variant_list_item
*item
;
3737 if (is_empty_list(&lp
->fs_variants_list
)) {
3740 item
= last_elem(&lp
->fs_variants_list
);
3743 llvmpipe_remove_shader_variant(lp
, item
->base
);
3748 * Generate the new variant.
3751 variant
= generate_variant(lp
, shader
, key
);
3754 LP_COUNT_ADD(llvm_compile_time
, dt
);
3755 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3757 /* Put the new variant into the list */
3759 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3760 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3761 lp
->nr_fs_variants
++;
3762 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3763 shader
->variants_cached
++;
3767 /* Bind this variant */
3768 lp_setup_set_fs_variant(lp
->setup
, variant
);
3776 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3778 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3779 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3780 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3782 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3784 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3785 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;