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_ptr
,
314 LLVMValueRef depth_stride
,
316 LLVMValueRef thread_data_ptr
)
318 const struct util_format_description
*zs_format_desc
= NULL
;
319 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
320 struct lp_type int_type
= lp_int_type(type
);
321 LLVMTypeRef vec_type
, int_vec_type
;
322 LLVMValueRef mask_ptr
, mask_val
;
323 LLVMValueRef consts_ptr
, num_consts_ptr
;
324 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
326 LLVMValueRef z_value
, s_value
;
327 LLVMValueRef z_fb
, s_fb
;
328 LLVMValueRef stencil_refs
[2];
329 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
330 struct lp_build_for_loop_state loop_state
;
331 struct lp_build_mask_context mask
;
333 * TODO: figure out if simple_shader optimization is really worthwile to
334 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
335 * code since tests tend to take another codepath than real shaders.
337 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
338 shader
->info
.base
.num_inputs
< 3 &&
339 shader
->info
.base
.num_instructions
< 8) && 0;
340 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
341 util_blend_state_is_dual(&key
->blend
, 0);
347 struct lp_bld_tgsi_system_values system_values
;
349 memset(&system_values
, 0, sizeof(system_values
));
351 /* truncate then sign extend. */
352 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
353 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
355 if (key
->depth
.enabled
||
356 key
->stencil
[0].enabled
) {
358 zs_format_desc
= util_format_description(key
->zsbuf_format
);
359 assert(zs_format_desc
);
361 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
362 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
363 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
364 if (shader
->info
.base
.writes_memory
)
365 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
366 else if (key
->alpha
.enabled
||
367 key
->blend
.alpha_to_coverage
||
368 shader
->info
.base
.uses_kill
||
369 shader
->info
.base
.writes_samplemask
) {
370 /* With alpha test and kill, can do the depth test early
371 * and hopefully eliminate some quads. But need to do a
372 * special deferred depth write once the final mask value
373 * is known. This only works though if there's either no
374 * stencil test or the stencil value isn't written.
376 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
377 (key
->stencil
[1].enabled
&&
378 key
->stencil
[1].writemask
)))
379 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
381 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
384 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
387 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
390 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
391 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
392 (key
->stencil
[1].enabled
&&
393 key
->stencil
[1].writemask
))))
394 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
400 vec_type
= lp_build_vec_type(gallivm
, type
);
401 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
403 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
404 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
405 /* convert scalar stencil refs into vectors */
406 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
407 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
409 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
410 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
412 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
413 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
415 lp_build_for_loop_begin(&loop_state
, gallivm
,
416 lp_build_const_int32(gallivm
, 0),
419 lp_build_const_int32(gallivm
, 1));
421 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
422 &loop_state
.counter
, 1, "mask_ptr");
423 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
425 memset(outputs
, 0, sizeof outputs
);
427 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
428 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
429 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
430 lp_build_vec_type(gallivm
,
435 if (dual_source_blend
) {
436 assert(key
->nr_cbufs
<= 1);
437 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
438 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
439 lp_build_vec_type(gallivm
,
446 /* 'mask' will control execution based on quad's pixel alive/killed state */
447 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
449 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
450 lp_build_mask_check(&mask
);
452 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
455 if (depth_mode
& EARLY_DEPTH_TEST
) {
457 * Clamp according to ARB_depth_clamp semantics.
459 if (key
->depth_clamp
) {
460 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
463 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
464 zs_format_desc
, key
->resource_1d
,
465 depth_ptr
, depth_stride
,
466 &z_fb
, &s_fb
, loop_state
.counter
);
467 lp_build_depth_stencil_test(gallivm
,
480 if (depth_mode
& EARLY_DEPTH_WRITE
) {
481 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
482 zs_format_desc
, key
->resource_1d
,
483 NULL
, NULL
, NULL
, loop_state
.counter
,
484 depth_ptr
, depth_stride
,
488 * Note mask check if stencil is enabled must be after ds write not after
489 * stencil test otherwise new stencil values may not get written if all
490 * fragments got killed by depth/stencil test.
492 if (!simple_shader
&& key
->stencil
[0].enabled
)
493 lp_build_mask_check(&mask
);
496 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
, NULL
, NULL
);
498 struct lp_build_tgsi_params params
;
499 memset(¶ms
, 0, sizeof(params
));
503 params
.consts_ptr
= consts_ptr
;
504 params
.const_sizes_ptr
= num_consts_ptr
;
505 params
.system_values
= &system_values
;
506 params
.inputs
= interp
->inputs
;
507 params
.context_ptr
= context_ptr
;
508 params
.thread_data_ptr
= thread_data_ptr
;
509 params
.sampler
= sampler
;
510 params
.info
= &shader
->info
.base
;
511 params
.ssbo_ptr
= ssbo_ptr
;
512 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
513 params
.image
= image
;
515 /* Build the actual shader */
516 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
517 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
520 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
524 if (key
->alpha
.enabled
) {
525 int color0
= find_output_by_semantic(&shader
->info
.base
,
529 if (color0
!= -1 && outputs
[color0
][3]) {
530 const struct util_format_description
*cbuf_format_desc
;
531 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
532 LLVMValueRef alpha_ref_value
;
534 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
535 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
537 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
539 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
540 &mask
, alpha
, alpha_ref_value
,
541 (depth_mode
& LATE_DEPTH_TEST
) != 0);
545 /* Emulate Alpha to Coverage with Alpha test */
546 if (key
->blend
.alpha_to_coverage
) {
547 int color0
= find_output_by_semantic(&shader
->info
.base
,
551 if (color0
!= -1 && outputs
[color0
][3]) {
552 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
554 lp_build_alpha_to_coverage(gallivm
, type
,
556 (depth_mode
& LATE_DEPTH_TEST
) != 0);
560 if (shader
->info
.base
.writes_samplemask
) {
561 int smaski
= find_output_by_semantic(&shader
->info
.base
,
562 TGSI_SEMANTIC_SAMPLEMASK
,
565 struct lp_build_context smask_bld
;
566 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
569 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
571 * Pixel is alive according to the first sample in the mask.
573 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
574 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
575 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
576 lp_build_mask_update(&mask
, smask
);
580 if (depth_mode
& LATE_DEPTH_TEST
) {
581 int pos0
= find_output_by_semantic(&shader
->info
.base
,
582 TGSI_SEMANTIC_POSITION
,
584 int s_out
= find_output_by_semantic(&shader
->info
.base
,
585 TGSI_SEMANTIC_STENCIL
,
587 if (pos0
!= -1 && outputs
[pos0
][2]) {
588 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
591 * Clamp according to ARB_depth_clamp semantics.
593 if (key
->depth_clamp
) {
594 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
598 if (s_out
!= -1 && outputs
[s_out
][1]) {
599 /* there's only one value, and spec says to discard additional bits */
600 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
601 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
602 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
603 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
604 stencil_refs
[1] = stencil_refs
[0];
607 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
608 zs_format_desc
, key
->resource_1d
,
609 depth_ptr
, depth_stride
,
610 &z_fb
, &s_fb
, loop_state
.counter
);
612 lp_build_depth_stencil_test(gallivm
,
625 if (depth_mode
& LATE_DEPTH_WRITE
) {
626 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
627 zs_format_desc
, key
->resource_1d
,
628 NULL
, NULL
, NULL
, loop_state
.counter
,
629 depth_ptr
, depth_stride
,
633 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
634 (depth_mode
& LATE_DEPTH_WRITE
))
636 /* Need to apply a reduced mask to the depth write. Reload the
637 * depth value, update from zs_value with the new mask value and
640 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
641 zs_format_desc
, key
->resource_1d
,
642 lp_build_mask_value(&mask
), z_fb
, s_fb
, loop_state
.counter
,
643 depth_ptr
, depth_stride
,
649 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
651 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
652 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
653 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
655 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
656 if(outputs
[attrib
][chan
]) {
657 /* XXX: just initialize outputs to point at colors[] and
660 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
661 LLVMValueRef color_ptr
;
662 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
663 &loop_state
.counter
, 1, "");
664 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
665 LLVMBuildStore(builder
, out
, color_ptr
);
671 if (key
->occlusion_count
) {
672 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
673 lp_build_name(counter
, "counter");
674 lp_build_occlusion_count(gallivm
, type
,
675 lp_build_mask_value(&mask
), counter
);
678 mask_val
= lp_build_mask_end(&mask
);
679 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
680 lp_build_for_loop_end(&loop_state
);
685 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
687 * Fragment Shader outputs pixels in small 2x2 blocks
688 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
690 * However in memory pixels are stored in rows
691 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
693 * @param type fragment shader type (4x or 8x float)
694 * @param num_fs number of fs_src
695 * @param is_1d whether we're outputting to a 1d resource
696 * @param dst_channels number of output channels
697 * @param fs_src output from fragment shader
698 * @param dst pointer to store result
699 * @param pad_inline is channel padding inline or at end of row
700 * @return the number of dsts
703 generate_fs_twiddle(struct gallivm_state
*gallivm
,
706 unsigned dst_channels
,
707 LLVMValueRef fs_src
[][4],
711 LLVMValueRef src
[16];
717 unsigned pixels
= type
.length
/ 4;
718 unsigned reorder_group
;
719 unsigned src_channels
;
723 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
724 src_count
= num_fs
* src_channels
;
726 assert(pixels
== 2 || pixels
== 1);
727 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
730 * Transpose from SoA -> AoS
732 for (i
= 0; i
< num_fs
; ++i
) {
733 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
737 * Pick transformation options
744 if (dst_channels
== 1) {
750 } else if (dst_channels
== 2) {
754 } else if (dst_channels
> 2) {
761 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
767 * Split the src in half
770 for (i
= num_fs
; i
> 0; --i
) {
771 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
772 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
780 * Ensure pixels are in memory order
783 /* Twiddle pixels by reordering the array, e.g.:
785 * src_count = 8 -> 0 2 1 3 4 6 5 7
786 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
788 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
790 for (i
= 0; i
< src_count
; ++i
) {
791 unsigned group
= i
/ reorder_group
;
792 unsigned block
= (group
/ 4) * 4 * reorder_group
;
793 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
796 } else if (twiddle
) {
797 /* Twiddle pixels across elements of array */
799 * XXX: we should avoid this in some cases, but would need to tell
800 * lp_build_conv to reorder (or deal with it ourselves).
802 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
805 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
809 * Moves any padding between pixels to the end
810 * e.g. RGBXRGBX -> RGBRGBXX
813 unsigned char swizzles
[16];
814 unsigned elems
= pixels
* dst_channels
;
816 for (i
= 0; i
< type
.length
; ++i
) {
818 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
820 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
823 for (i
= 0; i
< src_count
; ++i
) {
824 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
833 * Untwiddle and transpose, much like the above.
834 * However, this is after conversion, so we get packed vectors.
835 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
836 * the vectors will look like:
837 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
838 * be swizzled here). Extending to 16bit should be trivial.
839 * Should also be extended to handle twice wide vectors with AVX2...
842 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
849 struct lp_type type64
, type16
, type32
;
850 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
851 LLVMBuilderRef builder
= gallivm
->builder
;
852 LLVMValueRef tmp
[4], shuf
[8];
853 for (j
= 0; j
< 2; j
++) {
854 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
855 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
856 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
857 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
860 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
861 assert(type
.width
== 8);
862 assert(type
.length
== 16);
864 type8_t
= lp_build_vec_type(gallivm
, type
);
869 type64_t
= lp_build_vec_type(gallivm
, type64
);
874 type16_t
= lp_build_vec_type(gallivm
, type16
);
879 type32_t
= lp_build_vec_type(gallivm
, type32
);
881 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
883 if (src_count
== 1) {
884 /* transpose was no-op, just untwiddle */
885 LLVMValueRef shuf_vec
;
886 shuf_vec
= LLVMConstVector(shuf
, 8);
887 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
888 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
889 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
890 } else if (src_count
== 2) {
891 LLVMValueRef shuf_vec
;
892 shuf_vec
= LLVMConstVector(shuf
, 4);
894 for (i
= 0; i
< 2; i
++) {
895 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
896 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
897 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
900 for (j
= 0; j
< 2; j
++) {
901 LLVMValueRef lo
, hi
, lo2
, hi2
;
903 * Note that if we only really have 3 valid channels (rgb)
904 * and we don't need alpha we could substitute a undef here
905 * for the respective channel (causing llvm to drop conversion
908 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
909 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
910 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
911 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
912 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
913 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
914 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
921 * Load an unswizzled block of pixels from memory
924 load_unswizzled_block(struct gallivm_state
*gallivm
,
925 LLVMValueRef base_ptr
,
927 unsigned block_width
,
928 unsigned block_height
,
930 struct lp_type dst_type
,
932 unsigned dst_alignment
)
934 LLVMBuilderRef builder
= gallivm
->builder
;
935 unsigned row_size
= dst_count
/ block_height
;
938 /* Ensure block exactly fits into dst */
939 assert((block_width
* block_height
) % dst_count
== 0);
941 for (i
= 0; i
< dst_count
; ++i
) {
942 unsigned x
= i
% row_size
;
943 unsigned y
= i
/ row_size
;
945 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
946 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
949 LLVMValueRef dst_ptr
;
951 gep
[0] = lp_build_const_int32(gallivm
, 0);
952 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
954 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
955 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
956 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
958 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
960 LLVMSetAlignment(dst
[i
], dst_alignment
);
966 * Store an unswizzled block of pixels to memory
969 store_unswizzled_block(struct gallivm_state
*gallivm
,
970 LLVMValueRef base_ptr
,
972 unsigned block_width
,
973 unsigned block_height
,
975 struct lp_type src_type
,
977 unsigned src_alignment
)
979 LLVMBuilderRef builder
= gallivm
->builder
;
980 unsigned row_size
= src_count
/ block_height
;
983 /* Ensure src exactly fits into block */
984 assert((block_width
* block_height
) % src_count
== 0);
986 for (i
= 0; i
< src_count
; ++i
) {
987 unsigned x
= i
% row_size
;
988 unsigned y
= i
/ row_size
;
990 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
991 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
994 LLVMValueRef src_ptr
;
996 gep
[0] = lp_build_const_int32(gallivm
, 0);
997 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
999 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
1000 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
1001 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
1003 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
1005 LLVMSetAlignment(src_ptr
, src_alignment
);
1011 * Checks if a format description is an arithmetic format
1013 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1015 static inline boolean
1016 is_arithmetic_format(const struct util_format_description
*format_desc
)
1018 boolean arith
= false;
1021 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1022 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1023 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1031 * Checks if this format requires special handling due to required expansion
1032 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1035 static inline boolean
1036 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1038 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1039 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1047 * Retrieves the type representing the memory layout for a format
1049 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1052 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1053 struct lp_type
* type
)
1058 if (format_expands_to_float_soa(format_desc
)) {
1059 /* just make this a uint with width of block */
1060 type
->floating
= false;
1061 type
->fixed
= false;
1064 type
->width
= format_desc
->block
.bits
;
1069 for (i
= 0; i
< 4; i
++)
1070 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1074 memset(type
, 0, sizeof(struct lp_type
));
1075 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1076 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1077 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1078 type
->norm
= format_desc
->channel
[chan
].normalized
;
1080 if (is_arithmetic_format(format_desc
)) {
1084 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1085 type
->width
+= format_desc
->channel
[i
].size
;
1088 type
->width
= format_desc
->channel
[chan
].size
;
1089 type
->length
= format_desc
->nr_channels
;
1095 * Retrieves the type for a format which is usable in the blending code.
1097 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1100 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1101 struct lp_type
* type
)
1106 if (format_expands_to_float_soa(format_desc
)) {
1107 /* always use ordinary floats for blending */
1108 type
->floating
= true;
1109 type
->fixed
= false;
1117 for (i
= 0; i
< 4; i
++)
1118 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1122 memset(type
, 0, sizeof(struct lp_type
));
1123 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1124 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1125 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1126 type
->norm
= format_desc
->channel
[chan
].normalized
;
1127 type
->width
= format_desc
->channel
[chan
].size
;
1128 type
->length
= format_desc
->nr_channels
;
1130 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1131 if (format_desc
->channel
[i
].size
> type
->width
)
1132 type
->width
= format_desc
->channel
[i
].size
;
1135 if (type
->floating
) {
1138 if (type
->width
<= 8) {
1140 } else if (type
->width
<= 16) {
1147 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1154 * Scale a normalized value from src_bits to dst_bits.
1156 * The exact calculation is
1158 * dst = iround(src * dst_mask / src_mask)
1160 * or with integer rounding
1162 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1166 * src_mask = (1 << src_bits) - 1
1167 * dst_mask = (1 << dst_bits) - 1
1169 * but we try to avoid division and multiplication through shifts.
1171 static inline LLVMValueRef
1172 scale_bits(struct gallivm_state
*gallivm
,
1176 struct lp_type src_type
)
1178 LLVMBuilderRef builder
= gallivm
->builder
;
1179 LLVMValueRef result
= src
;
1181 if (dst_bits
< src_bits
) {
1182 int delta_bits
= src_bits
- dst_bits
;
1184 if (delta_bits
<= dst_bits
) {
1186 * Approximate the rescaling with a single shift.
1188 * This gives the wrong rounding.
1191 result
= LLVMBuildLShr(builder
,
1193 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1198 * Try more accurate rescaling.
1202 * Drop the least significant bits to make space for the multiplication.
1204 * XXX: A better approach would be to use a wider integer type as intermediate. But
1205 * this is enough to convert alpha from 16bits -> 2 when rendering to
1206 * PIPE_FORMAT_R10G10B10A2_UNORM.
1208 result
= LLVMBuildLShr(builder
,
1210 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1214 result
= LLVMBuildMul(builder
,
1216 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1220 * Add a rounding term before the division.
1222 * TODO: Handle signed integers too.
1224 if (!src_type
.sign
) {
1225 result
= LLVMBuildAdd(builder
,
1227 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1232 * Approximate the division by src_mask with a src_bits shift.
1234 * Given the src has already been shifted by dst_bits, all we need
1235 * to do is to shift by the difference.
1238 result
= LLVMBuildLShr(builder
,
1240 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1244 } else if (dst_bits
> src_bits
) {
1246 int db
= dst_bits
- src_bits
;
1248 /* Shift left by difference in bits */
1249 result
= LLVMBuildShl(builder
,
1251 lp_build_const_int_vec(gallivm
, src_type
, db
),
1254 if (db
<= src_bits
) {
1255 /* Enough bits in src to fill the remainder */
1256 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1258 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1261 result
= LLVMBuildOr(builder
, result
, lower
, "");
1262 } else if (db
> src_bits
) {
1263 /* Need to repeatedly copy src bits to fill remainder in dst */
1266 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1267 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1269 result
= LLVMBuildOr(builder
,
1271 LLVMBuildLShr(builder
, result
, shuv
, ""),
1281 * If RT is a smallfloat (needing denorms) format
1284 have_smallfloat_format(struct lp_type dst_type
,
1285 enum pipe_format format
)
1287 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1288 /* due to format handling hacks this format doesn't have floating set
1289 * here (and actually has width set to 32 too) so special case this. */
1290 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1295 * Convert from memory format to blending format
1297 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1300 convert_to_blend_type(struct gallivm_state
*gallivm
,
1301 unsigned block_size
,
1302 const struct util_format_description
*src_fmt
,
1303 struct lp_type src_type
,
1304 struct lp_type dst_type
,
1305 LLVMValueRef
* src
, // and dst
1308 LLVMValueRef
*dst
= src
;
1309 LLVMBuilderRef builder
= gallivm
->builder
;
1310 struct lp_type blend_type
;
1311 struct lp_type mem_type
;
1313 unsigned pixels
= block_size
/ num_srcs
;
1317 * full custom path for packed floats and srgb formats - none of the later
1318 * functions would do anything useful, and given the lp_type representation they
1319 * can't be fixed. Should really have some SoA blend path for these kind of
1320 * formats rather than hacking them in here.
1322 if (format_expands_to_float_soa(src_fmt
)) {
1323 LLVMValueRef tmpsrc
[4];
1325 * This is pretty suboptimal for this case blending in SoA would be much
1326 * better, since conversion gets us SoA values so need to convert back.
1328 assert(src_type
.width
== 32 || src_type
.width
== 16);
1329 assert(dst_type
.floating
);
1330 assert(dst_type
.width
== 32);
1331 assert(dst_type
.length
% 4 == 0);
1332 assert(num_srcs
% 4 == 0);
1334 if (src_type
.width
== 16) {
1335 /* expand 4x16bit values to 4x32bit */
1336 struct lp_type type32x4
= src_type
;
1337 LLVMTypeRef ltype32x4
;
1338 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1339 type32x4
.width
= 32;
1340 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1341 for (i
= 0; i
< num_fetch
; i
++) {
1342 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1344 src_type
.width
= 32;
1346 for (i
= 0; i
< 4; i
++) {
1349 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1350 LLVMValueRef tmpsoa
[4];
1351 LLVMValueRef tmps
= tmpsrc
[i
];
1352 if (dst_type
.length
== 8) {
1353 LLVMValueRef shuffles
[8];
1355 /* fetch was 4 values but need 8-wide output values */
1356 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1358 * for 8-wide aos transpose would give us wrong order not matching
1359 * incoming converted fs values and mask. ARGH.
1361 for (j
= 0; j
< 4; j
++) {
1362 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1363 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1365 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1366 LLVMConstVector(shuffles
, 8), "");
1368 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1369 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1372 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1374 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1379 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1380 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1382 /* Is the format arithmetic */
1383 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1384 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1386 /* Pad if necessary */
1387 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1388 for (i
= 0; i
< num_srcs
; ++i
) {
1389 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1392 src_type
.length
= dst_type
.length
;
1395 /* Special case for half-floats */
1396 if (mem_type
.width
== 16 && mem_type
.floating
) {
1397 assert(blend_type
.width
== 32 && blend_type
.floating
);
1398 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1406 src_type
.width
= blend_type
.width
* blend_type
.length
;
1407 blend_type
.length
*= pixels
;
1408 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1410 for (i
= 0; i
< num_srcs
; ++i
) {
1411 LLVMValueRef chans
[4];
1412 LLVMValueRef res
= NULL
;
1414 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1416 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1418 unsigned sa
= src_fmt
->channel
[j
].shift
;
1419 #if UTIL_ARCH_LITTLE_ENDIAN
1420 unsigned from_lsb
= j
;
1422 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1425 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1427 /* Extract bits from source */
1428 chans
[j
] = LLVMBuildLShr(builder
,
1430 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1433 chans
[j
] = LLVMBuildAnd(builder
,
1435 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1439 if (src_type
.norm
) {
1440 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1441 blend_type
.width
, chans
[j
], src_type
);
1444 /* Insert bits into correct position */
1445 chans
[j
] = LLVMBuildShl(builder
,
1447 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1453 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1457 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1463 * Convert from blending format to memory format
1465 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1468 convert_from_blend_type(struct gallivm_state
*gallivm
,
1469 unsigned block_size
,
1470 const struct util_format_description
*src_fmt
,
1471 struct lp_type src_type
,
1472 struct lp_type dst_type
,
1473 LLVMValueRef
* src
, // and dst
1476 LLVMValueRef
* dst
= src
;
1478 struct lp_type mem_type
;
1479 struct lp_type blend_type
;
1480 LLVMBuilderRef builder
= gallivm
->builder
;
1481 unsigned pixels
= block_size
/ num_srcs
;
1485 * full custom path for packed floats and srgb formats - none of the later
1486 * functions would do anything useful, and given the lp_type representation they
1487 * can't be fixed. Should really have some SoA blend path for these kind of
1488 * formats rather than hacking them in here.
1490 if (format_expands_to_float_soa(src_fmt
)) {
1492 * This is pretty suboptimal for this case blending in SoA would be much
1493 * better - we need to transpose the AoS values back to SoA values for
1494 * conversion/packing.
1496 assert(src_type
.floating
);
1497 assert(src_type
.width
== 32);
1498 assert(src_type
.length
% 4 == 0);
1499 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1501 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1502 LLVMValueRef tmpsoa
[4], tmpdst
;
1503 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1504 /* really really need SoA here */
1506 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1507 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1510 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1514 if (src_type
.length
== 8) {
1515 LLVMValueRef tmpaos
, shuffles
[8];
1518 * for 8-wide aos transpose has given us wrong order not matching
1519 * output order. HMPF. Also need to split the output values manually.
1521 for (j
= 0; j
< 4; j
++) {
1522 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1523 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1525 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1526 LLVMConstVector(shuffles
, 8), "");
1527 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1528 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1534 if (dst_type
.width
== 16) {
1535 struct lp_type type16x8
= dst_type
;
1536 struct lp_type type32x4
= dst_type
;
1537 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1538 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1539 type16x8
.length
= 8;
1540 type32x4
.width
= 32;
1541 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1542 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1543 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1544 /* We could do vector truncation but it doesn't generate very good code */
1545 for (i
= 0; i
< num_fetch
; i
++) {
1546 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1547 src
[i
], lp_build_zero(gallivm
, type32x4
));
1548 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1549 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1550 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1556 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1557 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1559 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1561 /* Special case for half-floats */
1562 if (mem_type
.width
== 16 && mem_type
.floating
) {
1563 int length
= dst_type
.length
;
1564 assert(blend_type
.width
== 32 && blend_type
.floating
);
1566 dst_type
.length
= src_type
.length
;
1568 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1570 dst_type
.length
= length
;
1574 /* Remove any padding */
1575 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1576 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1578 for (i
= 0; i
< num_srcs
; ++i
) {
1579 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1583 /* No bit arithmetic to do */
1588 src_type
.length
= pixels
;
1589 src_type
.width
= blend_type
.length
* blend_type
.width
;
1590 dst_type
.length
= pixels
;
1592 for (i
= 0; i
< num_srcs
; ++i
) {
1593 LLVMValueRef chans
[4];
1594 LLVMValueRef res
= NULL
;
1596 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1598 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1600 unsigned sa
= src_fmt
->channel
[j
].shift
;
1601 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1602 #if UTIL_ARCH_LITTLE_ENDIAN
1603 unsigned from_lsb
= j
;
1605 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1608 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1610 for (k
= 0; k
< blend_type
.width
; ++k
) {
1615 chans
[j
] = LLVMBuildLShr(builder
,
1617 lp_build_const_int_vec(gallivm
, src_type
,
1618 from_lsb
* blend_type
.width
),
1621 chans
[j
] = LLVMBuildAnd(builder
,
1623 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1626 /* Scale down bits */
1627 if (src_type
.norm
) {
1628 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1629 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1630 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1631 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1632 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1633 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1637 chans
[j
] = LLVMBuildShl(builder
,
1639 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1642 sa
+= src_fmt
->channel
[j
].size
;
1647 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1651 assert (dst_type
.width
!= 24);
1653 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1659 * Convert alpha to same blend type as src
1662 convert_alpha(struct gallivm_state
*gallivm
,
1663 struct lp_type row_type
,
1664 struct lp_type alpha_type
,
1665 const unsigned block_size
,
1666 const unsigned block_height
,
1667 const unsigned src_count
,
1668 const unsigned dst_channels
,
1669 const bool pad_inline
,
1670 LLVMValueRef
* src_alpha
)
1672 LLVMBuilderRef builder
= gallivm
->builder
;
1674 unsigned length
= row_type
.length
;
1675 row_type
.length
= alpha_type
.length
;
1677 /* Twiddle the alpha to match pixels */
1678 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1681 * TODO this should use single lp_build_conv call for
1682 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1684 for (i
= 0; i
< block_height
; ++i
) {
1685 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1688 alpha_type
= row_type
;
1689 row_type
.length
= length
;
1691 /* If only one channel we can only need the single alpha value per pixel */
1692 if (src_count
== 1 && dst_channels
== 1) {
1694 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1696 /* If there are more srcs than rows then we need to split alpha up */
1697 if (src_count
> block_height
) {
1698 for (i
= src_count
; i
> 0; --i
) {
1699 unsigned pixels
= block_size
/ src_count
;
1700 unsigned idx
= i
- 1;
1702 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1703 (idx
* pixels
) % 4, pixels
);
1707 /* If there is a src for each pixel broadcast the alpha across whole row */
1708 if (src_count
== block_size
) {
1709 for (i
= 0; i
< src_count
; ++i
) {
1710 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1711 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1714 unsigned pixels
= block_size
/ src_count
;
1715 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1716 unsigned alpha_span
= 1;
1717 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1719 /* Check if we need 2 src_alphas for our shuffles */
1720 if (pixels
> alpha_type
.length
) {
1724 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1725 for (j
= 0; j
< row_type
.length
; ++j
) {
1726 if (j
< pixels
* channels
) {
1727 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1729 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1733 for (i
= 0; i
< src_count
; ++i
) {
1734 unsigned idx1
= i
, idx2
= i
;
1736 if (alpha_span
> 1){
1741 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1744 LLVMConstVector(shuffles
, row_type
.length
),
1753 * Generates the blend function for unswizzled colour buffers
1754 * Also generates the read & write from colour buffer
1757 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1759 struct lp_fragment_shader_variant
*variant
,
1760 enum pipe_format out_format
,
1761 unsigned int num_fs
,
1762 struct lp_type fs_type
,
1763 LLVMValueRef
* fs_mask
,
1764 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1765 LLVMValueRef context_ptr
,
1766 LLVMValueRef color_ptr
,
1767 LLVMValueRef stride
,
1768 unsigned partial_mask
,
1771 const unsigned alpha_channel
= 3;
1772 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1773 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1774 const unsigned block_size
= block_width
* block_height
;
1775 const unsigned lp_integer_vector_width
= 128;
1777 LLVMBuilderRef builder
= gallivm
->builder
;
1778 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1779 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1780 LLVMValueRef src_alpha
[4 * 4];
1781 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1782 LLVMValueRef src_mask
[4 * 4];
1783 LLVMValueRef src
[4 * 4];
1784 LLVMValueRef src1
[4 * 4];
1785 LLVMValueRef dst
[4 * 4];
1786 LLVMValueRef blend_color
;
1787 LLVMValueRef blend_alpha
;
1788 LLVMValueRef i32_zero
;
1789 LLVMValueRef check_mask
;
1790 LLVMValueRef undef_src_val
;
1792 struct lp_build_mask_context mask_ctx
;
1793 struct lp_type mask_type
;
1794 struct lp_type blend_type
;
1795 struct lp_type row_type
;
1796 struct lp_type dst_type
;
1797 struct lp_type ls_type
;
1799 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1800 unsigned vector_width
;
1801 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1802 unsigned dst_channels
;
1807 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1809 unsigned dst_alignment
;
1811 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1812 bool has_alpha
= false;
1813 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1814 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1816 const boolean is_1d
= variant
->key
.resource_1d
;
1817 boolean twiddle_after_convert
= FALSE
;
1818 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1819 LLVMValueRef fpstate
= 0;
1821 /* Get type from output format */
1822 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1823 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1826 * Technically this code should go into lp_build_smallfloat_to_float
1827 * and lp_build_float_to_smallfloat but due to the
1828 * http://llvm.org/bugs/show_bug.cgi?id=6393
1829 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1830 * So the ordering is important here and there shouldn't be any
1831 * llvm ir instrunctions in this function before
1832 * this, otherwise half-float format conversions won't work
1833 * (again due to llvm bug #6393).
1835 if (have_smallfloat_format(dst_type
, out_format
)) {
1836 /* We need to make sure that denorms are ok for half float
1838 fpstate
= lp_build_fpstate_get(gallivm
);
1839 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1842 mask_type
= lp_int32_vec4_type();
1843 mask_type
.length
= fs_type
.length
;
1845 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1846 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1849 /* Do not bother executing code when mask is empty.. */
1851 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1853 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1854 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1857 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1858 lp_build_mask_check(&mask_ctx
);
1861 partial_mask
|= !variant
->opaque
;
1862 i32_zero
= lp_build_const_int32(gallivm
, 0);
1864 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1866 row_type
.length
= fs_type
.length
;
1867 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1869 /* Compute correct swizzle and count channels */
1870 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1873 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1874 /* Ensure channel is used */
1875 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1879 /* Ensure not already written to (happens in case with GL_ALPHA) */
1880 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1884 /* Ensure we havn't already found all channels */
1885 if (dst_channels
>= out_format_desc
->nr_channels
) {
1889 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1892 if (i
== alpha_channel
) {
1897 if (format_expands_to_float_soa(out_format_desc
)) {
1899 * the code above can't work for layout_other
1900 * for srgb it would sort of work but we short-circuit swizzles, etc.
1901 * as that is done as part of unpack / pack.
1903 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1909 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1912 /* If 3 channels then pad to include alpha for 4 element transpose */
1913 if (dst_channels
== 3) {
1914 assert (!has_alpha
);
1915 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1916 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1919 if (out_format_desc
->nr_channels
== 4) {
1922 * We use alpha from the color conversion, not separate one.
1923 * We had to include it for transpose, hence it will get converted
1924 * too (albeit when doing transpose after conversion, that would
1925 * no longer be the case necessarily).
1926 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1927 * otherwise we really have padding, not alpha, included.)
1934 * Load shader output
1936 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1937 /* Always load alpha for use in blending */
1940 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1943 alpha
= undef_src_val
;
1946 /* Load each channel */
1947 for (j
= 0; j
< dst_channels
; ++j
) {
1948 assert(swizzle
[j
] < 4);
1950 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1953 fs_src
[i
][j
] = undef_src_val
;
1957 /* If 3 channels then pad to include alpha for 4 element transpose */
1959 * XXX If we include that here maybe could actually use it instead of
1960 * separate alpha for blending?
1961 * (Difficult though we actually convert pad channels, not alpha.)
1963 if (dst_channels
== 3 && !has_alpha
) {
1964 fs_src
[i
][3] = alpha
;
1967 /* We split the row_mask and row_alpha as we want 128bit interleave */
1968 if (fs_type
.length
== 8) {
1969 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1971 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1972 src_channels
, src_channels
);
1974 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1975 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1976 src_channels
, src_channels
);
1978 src_mask
[i
] = fs_mask
[i
];
1979 src_alpha
[i
] = alpha
;
1982 if (dual_source_blend
) {
1983 /* same as above except different src/dst, skip masks and comments... */
1984 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1987 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1990 alpha
= undef_src_val
;
1993 for (j
= 0; j
< dst_channels
; ++j
) {
1994 assert(swizzle
[j
] < 4);
1996 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1999 fs_src1
[i
][j
] = undef_src_val
;
2002 if (dst_channels
== 3 && !has_alpha
) {
2003 fs_src1
[i
][3] = alpha
;
2005 if (fs_type
.length
== 8) {
2006 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
2007 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2008 src_channels
, src_channels
);
2010 src1_alpha
[i
] = alpha
;
2015 if (util_format_is_pure_integer(out_format
)) {
2017 * In this case fs_type was really ints or uints disguised as floats,
2020 fs_type
.floating
= 0;
2021 fs_type
.sign
= dst_type
.sign
;
2022 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2023 for (j
= 0; j
< dst_channels
; ++j
) {
2024 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2025 lp_build_vec_type(gallivm
, fs_type
), "");
2027 if (dst_channels
== 3 && !has_alpha
) {
2028 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2029 lp_build_vec_type(gallivm
, fs_type
), "");
2035 * We actually should generally do conversion first (for non-1d cases)
2036 * when the blend format is 8 or 16 bits. The reason is obvious,
2037 * there's 2 or 4 times less vectors to deal with for the interleave...
2038 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2039 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2040 * unpack only with 128bit vectors).
2041 * Note: for 16bit sizes really need matching pack conversion code
2043 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2044 twiddle_after_convert
= TRUE
;
2048 * Pixel twiddle from fragment shader order to memory order
2050 if (!twiddle_after_convert
) {
2051 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2052 dst_channels
, fs_src
, src
, pad_inline
);
2053 if (dual_source_blend
) {
2054 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2055 fs_src1
, src1
, pad_inline
);
2058 src_count
= num_fullblock_fs
* dst_channels
;
2060 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2061 * (AVX) turn out the same later when untwiddling/transpose (albeit
2062 * for true AVX2 path untwiddle needs to be different).
2063 * For now just order by colors first (so we can use unpack later).
2065 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2066 for (i
= 0; i
< dst_channels
; i
++) {
2067 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2068 if (dual_source_blend
) {
2069 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2075 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2076 if (src_count
!= num_fullblock_fs
* src_channels
) {
2077 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2078 row_type
.length
/= ds
;
2079 fs_type
.length
= row_type
.length
;
2082 blend_type
= row_type
;
2083 mask_type
.length
= 4;
2085 /* Convert src to row_type */
2086 if (dual_source_blend
) {
2087 struct lp_type old_row_type
= row_type
;
2088 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2089 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2092 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2095 /* If the rows are not an SSE vector, combine them to become SSE size! */
2096 if ((row_type
.width
* row_type
.length
) % 128) {
2097 unsigned bits
= row_type
.width
* row_type
.length
;
2100 assert(src_count
>= (vector_width
/ bits
));
2102 dst_count
= src_count
/ (vector_width
/ bits
);
2104 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2105 if (dual_source_blend
) {
2106 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2109 row_type
.length
*= combined
;
2110 src_count
/= combined
;
2112 bits
= row_type
.width
* row_type
.length
;
2113 assert(bits
== 128 || bits
== 256);
2116 if (twiddle_after_convert
) {
2117 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2118 if (dual_source_blend
) {
2119 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2124 * Blend Colour conversion
2126 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2127 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2128 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2129 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2130 &i32_zero
, 1, ""), "");
2133 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2135 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2137 * since blending is done with floats, there was no conversion.
2138 * However, the rules according to fixed point renderbuffers still
2139 * apply, that is we must clamp inputs to 0.0/1.0.
2140 * (This would apply to separate alpha conversion too but we currently
2141 * force has_alpha to be true.)
2142 * TODO: should skip this with "fake" blend, since post-blend conversion
2143 * will clamp anyway.
2144 * TODO: could also skip this if fragment color clamping is enabled. We
2145 * don't support it natively so it gets baked into the shader however, so
2146 * can't really tell here.
2148 struct lp_build_context f32_bld
;
2149 assert(row_type
.floating
);
2150 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2151 for (i
= 0; i
< src_count
; i
++) {
2152 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2154 if (dual_source_blend
) {
2155 for (i
= 0; i
< src_count
; i
++) {
2156 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2159 /* probably can't be different than row_type but better safe than sorry... */
2160 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2161 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2165 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2167 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2168 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2170 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2171 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2173 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2174 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2180 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2182 if (src_count
< block_height
) {
2183 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2184 } else if (src_count
> block_height
) {
2185 for (i
= src_count
; i
> 0; --i
) {
2186 unsigned pixels
= block_size
/ src_count
;
2187 unsigned idx
= i
- 1;
2189 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2190 (idx
* pixels
) % 4, pixels
);
2194 assert(mask_type
.width
== 32);
2196 for (i
= 0; i
< src_count
; ++i
) {
2197 unsigned pixels
= block_size
/ src_count
;
2198 unsigned pixel_width
= row_type
.width
* dst_channels
;
2200 if (pixel_width
== 24) {
2201 mask_type
.width
= 8;
2202 mask_type
.length
= vector_width
/ mask_type
.width
;
2204 mask_type
.length
= pixels
;
2205 mask_type
.width
= row_type
.width
* dst_channels
;
2208 * If mask_type width is smaller than 32bit, this doesn't quite
2209 * generate the most efficient code (could use some pack).
2211 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2212 lp_build_int_vec_type(gallivm
, mask_type
), "");
2214 mask_type
.length
*= dst_channels
;
2215 mask_type
.width
/= dst_channels
;
2218 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2219 lp_build_int_vec_type(gallivm
, mask_type
), "");
2220 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2227 struct lp_type alpha_type
= fs_type
;
2228 alpha_type
.length
= 4;
2229 convert_alpha(gallivm
, row_type
, alpha_type
,
2230 block_size
, block_height
,
2231 src_count
, dst_channels
,
2232 pad_inline
, src_alpha
);
2233 if (dual_source_blend
) {
2234 convert_alpha(gallivm
, row_type
, alpha_type
,
2235 block_size
, block_height
,
2236 src_count
, dst_channels
,
2237 pad_inline
, src1_alpha
);
2243 * Load dst from memory
2245 if (src_count
< block_height
) {
2246 dst_count
= block_height
;
2248 dst_count
= src_count
;
2251 dst_type
.length
*= block_size
/ dst_count
;
2253 if (format_expands_to_float_soa(out_format_desc
)) {
2255 * we need multiple values at once for the conversion, so can as well
2256 * load them vectorized here too instead of concatenating later.
2257 * (Still need concatenation later for 8-wide vectors).
2259 dst_count
= block_height
;
2260 dst_type
.length
= block_width
;
2264 * Compute the alignment of the destination pointer in bytes
2265 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2266 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2267 * 1d tex but can't distinguish here) so need to stick with per-pixel
2268 * alignment in this case.
2271 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2274 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2276 /* Force power-of-two alignment by extracting only the least-significant-bit */
2277 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2279 * Resource base and stride pointers are aligned to 16 bytes, so that's
2280 * the maximum alignment we can guarantee
2282 dst_alignment
= MIN2(16, dst_alignment
);
2286 if (dst_count
> src_count
) {
2287 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2288 util_is_power_of_two_or_zero(dst_type
.length
) &&
2289 dst_type
.length
* dst_type
.width
< 128) {
2291 * Never try to load values as 4xi8 which we will then
2292 * concatenate to larger vectors. This gives llvm a real
2293 * headache (the problem is the type legalizer (?) will
2294 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2295 * then the shuffles to concatenate are more or less impossible
2296 * - llvm is easily capable of generating a sequence of 32
2297 * pextrb/pinsrb instructions for that. Albeit it appears to
2298 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2299 * width to avoid the trouble (16bit seems not as bad, llvm
2300 * probably recognizes the load+shuffle as only one shuffle
2301 * is necessary, but we can do just the same anyway).
2303 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2309 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2310 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2311 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2312 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2317 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2318 dst
, ls_type
, dst_count
, dst_alignment
);
2323 * Convert from dst/output format to src/blending format.
2325 * This is necessary as we can only read 1 row from memory at a time,
2326 * so the minimum dst_count will ever be at this point is 4.
2328 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2329 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2330 * on all 16 pixels in that single vector at once.
2332 if (dst_count
> src_count
) {
2333 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2334 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2335 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2336 for (i
= 0; i
< dst_count
; i
++) {
2337 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2341 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2343 if (ls_type
.length
!= dst_type
.length
) {
2344 struct lp_type tmp_type
= dst_type
;
2345 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2346 for (i
= 0; i
< src_count
; i
++) {
2347 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2348 lp_build_vec_type(gallivm
, tmp_type
), "");
2356 /* XXX this is broken for RGB8 formats -
2357 * they get expanded from 12 to 16 elements (to include alpha)
2358 * by convert_to_blend_type then reduced to 15 instead of 12
2359 * by convert_from_blend_type (a simple fix though breaks A8...).
2360 * R16G16B16 also crashes differently however something going wrong
2361 * inside llvm handling npot vector sizes seemingly.
2362 * It seems some cleanup could be done here (like skipping conversion/blend
2365 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2366 row_type
, dst
, src_count
);
2369 * FIXME: Really should get logic ops / masks out of generic blend / row
2370 * format. Logic ops will definitely not work on the blend float format
2371 * used for SRGB here and I think OpenGL expects this to work as expected
2372 * (that is incoming values converted to srgb then logic op applied).
2374 for (i
= 0; i
< src_count
; ++i
) {
2375 dst
[i
] = lp_build_blend_aos(gallivm
,
2376 &variant
->key
.blend
,
2381 has_alpha
? NULL
: src_alpha
[i
],
2383 has_alpha
? NULL
: src1_alpha
[i
],
2385 partial_mask
? src_mask
[i
] : NULL
,
2387 has_alpha
? NULL
: blend_alpha
,
2389 pad_inline
? 4 : dst_channels
);
2392 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2393 row_type
, dst_type
, dst
, src_count
);
2395 /* Split the blend rows back to memory rows */
2396 if (dst_count
> src_count
) {
2397 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2399 if (src_count
== 1) {
2400 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2401 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2403 row_type
.length
/= 2;
2407 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2408 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2409 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2410 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2412 row_type
.length
/= 2;
2417 * Store blend result to memory
2420 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2421 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2424 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2425 dst
, dst_type
, dst_count
, dst_alignment
);
2428 if (have_smallfloat_format(dst_type
, out_format
)) {
2429 lp_build_fpstate_set(gallivm
, fpstate
);
2433 lp_build_mask_end(&mask_ctx
);
2439 * Generate the runtime callable function for the whole fragment pipeline.
2440 * Note that the function which we generate operates on a block of 16
2441 * pixels at at time. The block contains 2x2 quads. Each quad contains
2445 generate_fragment(struct llvmpipe_context
*lp
,
2446 struct lp_fragment_shader
*shader
,
2447 struct lp_fragment_shader_variant
*variant
,
2448 unsigned partial_mask
)
2450 struct gallivm_state
*gallivm
= variant
->gallivm
;
2451 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2452 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2454 struct lp_type fs_type
;
2455 struct lp_type blend_type
;
2456 LLVMTypeRef fs_elem_type
;
2457 LLVMTypeRef blend_vec_type
;
2458 LLVMTypeRef arg_types
[15];
2459 LLVMTypeRef func_type
;
2460 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2461 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2462 LLVMValueRef context_ptr
;
2465 LLVMValueRef a0_ptr
;
2466 LLVMValueRef dadx_ptr
;
2467 LLVMValueRef dady_ptr
;
2468 LLVMValueRef color_ptr_ptr
;
2469 LLVMValueRef stride_ptr
;
2470 LLVMValueRef color_sample_stride_ptr
;
2471 LLVMValueRef depth_ptr
;
2472 LLVMValueRef depth_stride
;
2473 LLVMValueRef depth_sample_stride
;
2474 LLVMValueRef mask_input
;
2475 LLVMValueRef thread_data_ptr
;
2476 LLVMBasicBlockRef block
;
2477 LLVMBuilderRef builder
;
2478 struct lp_build_sampler_soa
*sampler
;
2479 struct lp_build_image_soa
*image
;
2480 struct lp_build_interp_soa_context interp
;
2481 LLVMValueRef fs_mask
[16 / 4];
2482 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2483 LLVMValueRef function
;
2484 LLVMValueRef facing
;
2489 boolean cbuf0_write_all
;
2490 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2491 util_blend_state_is_dual(&key
->blend
, 0);
2493 assert(lp_native_vector_width
/ 32 >= 4);
2495 /* Adjust color input interpolation according to flatshade state:
2497 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2498 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2499 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2501 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2503 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2507 /* check if writes to cbuf[0] are to be copied to all cbufs */
2509 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2511 /* TODO: actually pick these based on the fs and color buffer
2512 * characteristics. */
2514 memset(&fs_type
, 0, sizeof fs_type
);
2515 fs_type
.floating
= TRUE
; /* floating point values */
2516 fs_type
.sign
= TRUE
; /* values are signed */
2517 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2518 fs_type
.width
= 32; /* 32-bit float */
2519 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2521 memset(&blend_type
, 0, sizeof blend_type
);
2522 blend_type
.floating
= FALSE
; /* values are integers */
2523 blend_type
.sign
= FALSE
; /* values are unsigned */
2524 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2525 blend_type
.width
= 8; /* 8-bit ubyte values */
2526 blend_type
.length
= 16; /* 16 elements per vector */
2529 * Generate the function prototype. Any change here must be reflected in
2530 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2533 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2535 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2537 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2538 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2540 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2541 arg_types
[1] = int32_type
; /* x */
2542 arg_types
[2] = int32_type
; /* y */
2543 arg_types
[3] = int32_type
; /* facing */
2544 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2545 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2546 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2547 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2548 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2549 arg_types
[9] = LLVMInt64TypeInContext(gallivm
->context
); /* mask_input */
2550 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2551 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2552 arg_types
[12] = int32_type
; /* depth_stride */
2553 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2554 arg_types
[14] = int32_type
; /* depth sample stride */
2556 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2557 arg_types
, ARRAY_SIZE(arg_types
), 0);
2559 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2560 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2562 variant
->function
[partial_mask
] = function
;
2564 /* XXX: need to propagate noalias down into color param now we are
2565 * passing a pointer-to-pointer?
2567 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2568 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2569 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2571 context_ptr
= LLVMGetParam(function
, 0);
2572 x
= LLVMGetParam(function
, 1);
2573 y
= LLVMGetParam(function
, 2);
2574 facing
= LLVMGetParam(function
, 3);
2575 a0_ptr
= LLVMGetParam(function
, 4);
2576 dadx_ptr
= LLVMGetParam(function
, 5);
2577 dady_ptr
= LLVMGetParam(function
, 6);
2578 color_ptr_ptr
= LLVMGetParam(function
, 7);
2579 depth_ptr
= LLVMGetParam(function
, 8);
2580 mask_input
= LLVMGetParam(function
, 9);
2581 thread_data_ptr
= LLVMGetParam(function
, 10);
2582 stride_ptr
= LLVMGetParam(function
, 11);
2583 depth_stride
= LLVMGetParam(function
, 12);
2584 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2585 depth_sample_stride
= LLVMGetParam(function
, 14);
2587 lp_build_name(context_ptr
, "context");
2588 lp_build_name(x
, "x");
2589 lp_build_name(y
, "y");
2590 lp_build_name(a0_ptr
, "a0");
2591 lp_build_name(dadx_ptr
, "dadx");
2592 lp_build_name(dady_ptr
, "dady");
2593 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2594 lp_build_name(depth_ptr
, "depth");
2595 lp_build_name(mask_input
, "mask_input");
2596 lp_build_name(thread_data_ptr
, "thread_data");
2597 lp_build_name(stride_ptr
, "stride_ptr");
2598 lp_build_name(depth_stride
, "depth_stride");
2599 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2600 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2606 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2607 builder
= gallivm
->builder
;
2609 LLVMPositionBuilderAtEnd(builder
, block
);
2612 * Must not count ps invocations if there's a null shader.
2613 * (It would be ok to count with null shader if there's d/s tests,
2614 * but only if there's d/s buffers too, which is different
2615 * to implicit rasterization disable which must not depend
2616 * on the d/s buffers.)
2617 * Could use popcount on mask, but pixel accuracy is not required.
2618 * Could disable if there's no stats query, but maybe not worth it.
2620 if (shader
->info
.base
.num_instructions
> 1) {
2621 LLVMValueRef invocs
, val
;
2622 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2623 val
= LLVMBuildLoad(builder
, invocs
, "");
2624 val
= LLVMBuildAdd(builder
, val
,
2625 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2627 LLVMBuildStore(builder
, val
, invocs
);
2630 /* code generated texture sampling */
2631 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2632 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2634 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2635 /* for 1d resources only run "upper half" of stamp */
2636 if (key
->resource_1d
)
2640 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2641 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2642 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2643 num_loop
, "mask_store");
2644 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2645 boolean pixel_center_integer
=
2646 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2649 * The shader input interpolation info is not explicitely baked in the
2650 * shader key, but everything it derives from (TGSI, and flatshade) is
2651 * already included in the shader key.
2653 lp_build_interp_soa_init(&interp
,
2655 shader
->info
.base
.num_inputs
,
2657 pixel_center_integer
,
2661 a0_ptr
, dadx_ptr
, dady_ptr
,
2664 for (i
= 0; i
< num_fs
; i
++) {
2666 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2667 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2668 &indexi
, 1, "mask_ptr");
2671 mask
= generate_quad_mask(gallivm
, fs_type
,
2672 i
*fs_type
.length
/4, 0, mask_input
);
2675 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2677 LLVMBuildStore(builder
, mask
, mask_ptr
);
2680 generate_fs_loop(gallivm
,
2689 mask_store
, /* output */
2696 for (i
= 0; i
< num_fs
; i
++) {
2697 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2698 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2700 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2701 /* This is fucked up need to reorganize things */
2702 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2703 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2704 ptr
= LLVMBuildGEP(builder
,
2705 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2707 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2710 if (dual_source_blend
) {
2711 /* only support one dual source blend target hence always use output 1 */
2712 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2713 ptr
= LLVMBuildGEP(builder
,
2714 color_store
[1][chan
],
2716 fs_out_color
[1][chan
][i
] = ptr
;
2722 sampler
->destroy(sampler
);
2723 image
->destroy(image
);
2724 /* Loop over color outputs / color buffers to do blending.
2726 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2727 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2728 LLVMValueRef color_ptr
;
2729 LLVMValueRef stride
;
2730 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2732 boolean do_branch
= ((key
->depth
.enabled
2733 || key
->stencil
[0].enabled
2734 || key
->alpha
.enabled
)
2735 && !shader
->info
.base
.uses_kill
);
2737 color_ptr
= LLVMBuildLoad(builder
,
2738 LLVMBuildGEP(builder
, color_ptr_ptr
,
2742 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2744 stride
= LLVMBuildLoad(builder
,
2745 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2748 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2749 key
->cbuf_format
[cbuf
],
2750 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2751 context_ptr
, color_ptr
, stride
,
2752 partial_mask
, do_branch
);
2756 LLVMBuildRetVoid(builder
);
2758 gallivm_verify_function(gallivm
, function
);
2763 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
2767 debug_printf("fs variant %p:\n", (void *) key
);
2769 if (key
->flatshade
) {
2770 debug_printf("flatshade = 1\n");
2772 if (key
->multisample
) {
2773 debug_printf("multisample = 1\n");
2774 debug_printf("coverage samples = %d\n", key
->coverage_samples
);
2776 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2777 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2778 debug_printf("cbuf nr_samples[%u] = %d\n", i
, key
->cbuf_nr_samples
[i
]);
2780 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2781 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2782 debug_printf("depth nr_samples = %d\n", key
->zsbuf_nr_samples
);
2784 if (key
->depth
.enabled
) {
2785 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2786 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2789 for (i
= 0; i
< 2; ++i
) {
2790 if (key
->stencil
[i
].enabled
) {
2791 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2792 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2793 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2794 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2795 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2796 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2800 if (key
->alpha
.enabled
) {
2801 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2804 if (key
->occlusion_count
) {
2805 debug_printf("occlusion_count = 1\n");
2808 if (key
->blend
.logicop_enable
) {
2809 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2811 else if (key
->blend
.rt
[0].blend_enable
) {
2812 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2813 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2814 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2815 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2816 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2817 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2819 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2820 if (key
->blend
.alpha_to_coverage
) {
2821 debug_printf("blend.alpha_to_coverage is enabled\n");
2823 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2824 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
2825 debug_printf("sampler[%u] = \n", i
);
2826 debug_printf(" .wrap = %s %s %s\n",
2827 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2828 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2829 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2830 debug_printf(" .min_img_filter = %s\n",
2831 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2832 debug_printf(" .min_mip_filter = %s\n",
2833 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2834 debug_printf(" .mag_img_filter = %s\n",
2835 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2836 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2837 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
2838 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2839 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2840 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2841 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2842 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2844 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2845 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
2846 debug_printf("texture[%u] = \n", i
);
2847 debug_printf(" .format = %s\n",
2848 util_format_name(texture
->format
));
2849 debug_printf(" .target = %s\n",
2850 util_str_tex_target(texture
->target
, TRUE
));
2851 debug_printf(" .level_zero_only = %u\n",
2852 texture
->level_zero_only
);
2853 debug_printf(" .pot = %u %u %u\n",
2855 texture
->pot_height
,
2856 texture
->pot_depth
);
2858 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
2859 for (i
= 0; i
< key
->nr_images
; ++i
) {
2860 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
2861 debug_printf("image[%u] = \n", i
);
2862 debug_printf(" .format = %s\n",
2863 util_format_name(image
->format
));
2864 debug_printf(" .target = %s\n",
2865 util_str_tex_target(image
->target
, TRUE
));
2866 debug_printf(" .level_zero_only = %u\n",
2867 image
->level_zero_only
);
2868 debug_printf(" .pot = %u %u %u\n",
2877 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
2879 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2880 variant
->shader
->no
, variant
->no
);
2881 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
2882 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2884 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
2885 dump_fs_variant_key(&variant
->key
);
2886 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2892 * Generate a new fragment shader variant from the shader code and
2893 * other state indicated by the key.
2895 static struct lp_fragment_shader_variant
*
2896 generate_variant(struct llvmpipe_context
*lp
,
2897 struct lp_fragment_shader
*shader
,
2898 const struct lp_fragment_shader_variant_key
*key
)
2900 struct lp_fragment_shader_variant
*variant
;
2901 const struct util_format_description
*cbuf0_format_desc
= NULL
;
2902 boolean fullcolormask
;
2903 char module_name
[64];
2905 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
2909 memset(variant
, 0, sizeof(*variant
));
2910 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2911 shader
->no
, shader
->variants_created
);
2913 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
2914 if (!variant
->gallivm
) {
2919 variant
->shader
= shader
;
2920 variant
->list_item_global
.base
= variant
;
2921 variant
->list_item_local
.base
= variant
;
2922 variant
->no
= shader
->variants_created
++;
2924 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2927 * Determine whether we are touching all channels in the color buffer.
2929 fullcolormask
= FALSE
;
2930 if (key
->nr_cbufs
== 1) {
2931 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2932 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2936 !key
->blend
.logicop_enable
&&
2937 !key
->blend
.rt
[0].blend_enable
&&
2939 !key
->stencil
[0].enabled
&&
2940 !key
->alpha
.enabled
&&
2941 !key
->blend
.alpha_to_coverage
&&
2942 !key
->depth
.enabled
&&
2943 !shader
->info
.base
.uses_kill
&&
2944 !shader
->info
.base
.writes_samplemask
2947 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2948 lp_debug_fs_variant(variant
);
2951 lp_jit_init_types(variant
);
2953 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2954 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2956 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2957 if (variant
->opaque
) {
2958 /* Specialized shader, which doesn't need to read the color buffer. */
2959 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2964 * Compile everything
2967 gallivm_compile_module(variant
->gallivm
);
2969 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2971 if (variant
->function
[RAST_EDGE_TEST
]) {
2972 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2973 gallivm_jit_function(variant
->gallivm
,
2974 variant
->function
[RAST_EDGE_TEST
]);
2977 if (variant
->function
[RAST_WHOLE
]) {
2978 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2979 gallivm_jit_function(variant
->gallivm
,
2980 variant
->function
[RAST_WHOLE
]);
2981 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2982 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2985 gallivm_free_ir(variant
->gallivm
);
2992 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2993 const struct pipe_shader_state
*templ
)
2995 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2996 struct lp_fragment_shader
*shader
;
2998 int nr_sampler_views
;
3002 shader
= CALLOC_STRUCT(lp_fragment_shader
);
3006 shader
->no
= fs_no
++;
3007 make_empty_list(&shader
->variants
);
3009 shader
->base
.type
= templ
->type
;
3010 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
3011 /* get/save the summary info for this shader */
3012 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3014 /* we need to keep a local copy of the tokens */
3015 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3017 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3018 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3021 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3022 if (shader
->draw_data
== NULL
) {
3023 FREE((void *) shader
->base
.tokens
);
3028 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3029 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3030 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3031 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3033 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3034 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3035 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3036 shader
->inputs
[i
].location
= shader
->info
.base
.input_interpolate_loc
[i
];
3038 switch (shader
->info
.base
.input_interpolate
[i
]) {
3039 case TGSI_INTERPOLATE_CONSTANT
:
3040 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3042 case TGSI_INTERPOLATE_LINEAR
:
3043 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3045 case TGSI_INTERPOLATE_PERSPECTIVE
:
3046 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3048 case TGSI_INTERPOLATE_COLOR
:
3049 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3056 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3057 case TGSI_SEMANTIC_FACE
:
3058 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3060 case TGSI_SEMANTIC_POSITION
:
3061 /* Position was already emitted above
3063 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3064 shader
->inputs
[i
].src_index
= 0;
3068 /* XXX this is a completely pointless index map... */
3069 shader
->inputs
[i
].src_index
= i
+1;
3072 if (LP_DEBUG
& DEBUG_TGSI
) {
3074 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3075 shader
->no
, (void *) shader
);
3076 tgsi_dump(templ
->tokens
, 0);
3077 debug_printf("usage masks:\n");
3078 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3079 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3080 debug_printf(" IN[%u].%s%s%s%s\n",
3082 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3083 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3084 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3085 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3095 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3097 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3098 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3099 if (llvmpipe
->fs
== lp_fs
)
3102 draw_bind_fragment_shader(llvmpipe
->draw
,
3103 (lp_fs
? lp_fs
->draw_data
: NULL
));
3105 llvmpipe
->fs
= lp_fs
;
3107 llvmpipe
->dirty
|= LP_NEW_FS
;
3112 * Remove shader variant from two lists: the shader's variant list
3113 * and the context's variant list.
3116 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3117 struct lp_fragment_shader_variant
*variant
)
3119 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3120 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3121 "v total cached %u inst %u total inst %u\n",
3122 variant
->shader
->no
, variant
->no
,
3123 variant
->shader
->variants_created
,
3124 variant
->shader
->variants_cached
,
3125 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3128 gallivm_destroy(variant
->gallivm
);
3130 /* remove from shader's list */
3131 remove_from_list(&variant
->list_item_local
);
3132 variant
->shader
->variants_cached
--;
3134 /* remove from context's list */
3135 remove_from_list(&variant
->list_item_global
);
3136 lp
->nr_fs_variants
--;
3137 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3144 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3146 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3147 struct lp_fragment_shader
*shader
= fs
;
3148 struct lp_fs_variant_list_item
*li
;
3150 assert(fs
!= llvmpipe
->fs
);
3153 * XXX: we need to flush the context until we have some sort of reference
3154 * counting in fragment shaders as they may still be binned
3155 * Flushing alone might not sufficient we need to wait on it too.
3157 llvmpipe_finish(pipe
, __FUNCTION__
);
3159 /* Delete all the variants */
3160 li
= first_elem(&shader
->variants
);
3161 while(!at_end(&shader
->variants
, li
)) {
3162 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3163 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3167 /* Delete draw module's data */
3168 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3170 if (shader
->base
.ir
.nir
)
3171 ralloc_free(shader
->base
.ir
.nir
);
3172 assert(shader
->variants_cached
== 0);
3173 FREE((void *) shader
->base
.tokens
);
3180 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3181 enum pipe_shader_type shader
, uint index
,
3182 const struct pipe_constant_buffer
*cb
)
3184 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3185 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3187 assert(shader
< PIPE_SHADER_TYPES
);
3188 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3190 /* note: reference counting */
3191 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3194 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3195 debug_printf("Illegal set constant without bind flag\n");
3196 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3200 if (shader
== PIPE_SHADER_VERTEX
||
3201 shader
== PIPE_SHADER_GEOMETRY
||
3202 shader
== PIPE_SHADER_TESS_CTRL
||
3203 shader
== PIPE_SHADER_TESS_EVAL
) {
3204 /* Pass the constants to the 'draw' module */
3205 const unsigned size
= cb
? cb
->buffer_size
: 0;
3209 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3211 else if (cb
&& cb
->user_buffer
) {
3212 data
= (ubyte
*) cb
->user_buffer
;
3219 data
+= cb
->buffer_offset
;
3221 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3224 else if (shader
== PIPE_SHADER_COMPUTE
)
3225 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3227 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3229 if (cb
&& cb
->user_buffer
) {
3230 pipe_resource_reference(&constants
, NULL
);
3235 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3236 enum pipe_shader_type shader
, unsigned start_slot
,
3237 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3238 unsigned writable_bitmask
)
3240 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3242 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3243 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3245 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3247 if (shader
== PIPE_SHADER_VERTEX
||
3248 shader
== PIPE_SHADER_GEOMETRY
||
3249 shader
== PIPE_SHADER_TESS_CTRL
||
3250 shader
== PIPE_SHADER_TESS_EVAL
) {
3251 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3252 const ubyte
*data
= NULL
;
3253 if (buffer
&& buffer
->buffer
)
3254 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3256 data
+= buffer
->buffer_offset
;
3257 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3259 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3260 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3261 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3262 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3268 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3269 enum pipe_shader_type shader
, unsigned start_slot
,
3270 unsigned count
, const struct pipe_image_view
*images
)
3272 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3275 draw_flush(llvmpipe
->draw
);
3276 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3277 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3279 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3282 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3283 if (shader
== PIPE_SHADER_VERTEX
||
3284 shader
== PIPE_SHADER_GEOMETRY
||
3285 shader
== PIPE_SHADER_TESS_CTRL
||
3286 shader
== PIPE_SHADER_TESS_EVAL
) {
3287 draw_set_images(llvmpipe
->draw
,
3289 llvmpipe
->images
[shader
],
3290 start_slot
+ count
);
3291 } else if (shader
== PIPE_SHADER_COMPUTE
)
3292 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3294 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3298 * Return the blend factor equivalent to a destination alpha of one.
3300 static inline unsigned
3301 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3304 case PIPE_BLENDFACTOR_DST_ALPHA
:
3305 return PIPE_BLENDFACTOR_ONE
;
3306 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3307 return PIPE_BLENDFACTOR_ZERO
;
3308 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3310 return PIPE_BLENDFACTOR_ZERO
;
3312 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3320 * We need to generate several variants of the fragment pipeline to match
3321 * all the combinations of the contributing state atoms.
3323 * TODO: there is actually no reason to tie this to context state -- the
3324 * generated code could be cached globally in the screen.
3326 static struct lp_fragment_shader_variant_key
*
3327 make_variant_key(struct llvmpipe_context
*lp
,
3328 struct lp_fragment_shader
*shader
,
3332 struct lp_fragment_shader_variant_key
*key
;
3334 key
= (struct lp_fragment_shader_variant_key
*)store
;
3336 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3338 if (lp
->framebuffer
.zsbuf
) {
3339 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3340 const struct util_format_description
*zsbuf_desc
=
3341 util_format_description(zsbuf_format
);
3343 if (lp
->depth_stencil
->depth
.enabled
&&
3344 util_format_has_depth(zsbuf_desc
)) {
3345 key
->zsbuf_format
= zsbuf_format
;
3346 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3348 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3349 util_format_has_stencil(zsbuf_desc
)) {
3350 key
->zsbuf_format
= zsbuf_format
;
3351 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3353 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3354 key
->resource_1d
= TRUE
;
3356 key
->zsbuf_nr_samples
= util_res_sample_count(lp
->framebuffer
.zsbuf
->texture
);
3360 * Propagate the depth clamp setting from the rasterizer state.
3361 * depth_clip == 0 implies depth clamping is enabled.
3363 * When clip_halfz is enabled, then always clamp the depth values.
3365 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3366 * clamp is always active in d3d10, regardless if depth clip is
3368 * (GL has an always-on [0,1] clamp on fs depth output instead
3369 * to ensure the depth values stay in range. Doesn't look like
3370 * we do that, though...)
3372 if (lp
->rasterizer
->clip_halfz
) {
3373 key
->depth_clamp
= 1;
3375 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3378 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3379 if (!lp
->framebuffer
.nr_cbufs
||
3380 !lp
->framebuffer
.cbufs
[0] ||
3381 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3382 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3384 if(key
->alpha
.enabled
)
3385 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3386 /* alpha.ref_value is passed in jit_context */
3388 key
->flatshade
= lp
->rasterizer
->flatshade
;
3389 key
->multisample
= lp
->rasterizer
->multisample
;
3390 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3391 key
->occlusion_count
= TRUE
;
3394 if (lp
->framebuffer
.nr_cbufs
) {
3395 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3398 key
->coverage_samples
= 1;
3399 if (key
->multisample
)
3400 key
->coverage_samples
= util_framebuffer_get_num_samples(&lp
->framebuffer
);
3401 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3403 if (!key
->blend
.independent_blend_enable
) {
3404 /* we always need independent blend otherwise the fixups below won't work */
3405 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3406 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3408 key
->blend
.independent_blend_enable
= 1;
3411 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3412 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3414 if (lp
->framebuffer
.cbufs
[i
]) {
3415 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3416 const struct util_format_description
*format_desc
;
3418 key
->cbuf_format
[i
] = format
;
3419 key
->cbuf_nr_samples
[i
] = util_res_sample_count(lp
->framebuffer
.cbufs
[i
]->texture
);
3422 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3423 * mixing of 2d textures with height 1 and 1d textures, so make sure
3424 * we pick 1d if any cbuf or zsbuf is 1d.
3426 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3427 key
->resource_1d
= TRUE
;
3430 format_desc
= util_format_description(format
);
3431 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3432 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3435 * Mask out color channels not present in the color buffer.
3437 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3440 * Disable blend for integer formats.
3442 if (util_format_is_pure_integer(format
)) {
3443 blend_rt
->blend_enable
= 0;
3447 * Our swizzled render tiles always have an alpha channel, but the
3448 * linear render target format often does not, so force here the dst
3451 * This is not a mere optimization. Wrong results will be produced if
3452 * the dst alpha is used, the dst format does not have alpha, and the
3453 * previous rendering was not flushed from the swizzled to linear
3454 * buffer. For example, NonPowTwo DCT.
3456 * TODO: This should be generalized to all channels for better
3457 * performance, but only alpha causes correctness issues.
3459 * Also, force rgb/alpha func/factors match, to make AoS blending
3462 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3463 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3464 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3465 boolean clamped_zero
= !util_format_is_float(format
) &&
3466 !util_format_is_snorm(format
);
3467 blend_rt
->rgb_src_factor
=
3468 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3469 blend_rt
->rgb_dst_factor
=
3470 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3471 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3472 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3473 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3477 /* no color buffer for this fragment output */
3478 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3479 key
->cbuf_nr_samples
[i
] = 0;
3480 blend_rt
->colormask
= 0x0;
3481 blend_rt
->blend_enable
= 0;
3485 /* This value will be the same for all the variants of a given shader:
3487 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3489 struct lp_sampler_static_state
*fs_sampler
;
3491 fs_sampler
= key
->samplers
;
3493 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3495 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3496 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3497 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3498 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3503 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3504 * are dx10-style? Can't really have mixed opcodes, at least not
3505 * if we want to skip the holes here (without rescanning tgsi).
3507 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3508 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3509 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3511 * Note sview may exceed what's representable by file_mask.
3512 * This will still work, the only downside is that not actually
3513 * used views may be included in the shader key.
3515 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3516 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3517 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3522 key
->nr_sampler_views
= key
->nr_samplers
;
3523 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3524 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3525 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3526 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3531 struct lp_image_static_state
*lp_image
;
3532 lp_image
= lp_fs_variant_key_images(key
);
3533 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3534 for (i
= 0; i
< key
->nr_images
; ++i
) {
3535 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3536 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3537 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3546 * Update fragment shader state. This is called just prior to drawing
3547 * something when some fragment-related state has changed.
3550 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3552 struct lp_fragment_shader
*shader
= lp
->fs
;
3553 struct lp_fragment_shader_variant_key
*key
;
3554 struct lp_fragment_shader_variant
*variant
= NULL
;
3555 struct lp_fs_variant_list_item
*li
;
3556 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3558 key
= make_variant_key(lp
, shader
, store
);
3560 /* Search the variants for one which matches the key */
3561 li
= first_elem(&shader
->variants
);
3562 while(!at_end(&shader
->variants
, li
)) {
3563 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3571 /* Move this variant to the head of the list to implement LRU
3572 * deletion of shader's when we have too many.
3574 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3577 /* variant not found, create it now */
3580 unsigned variants_to_cull
;
3582 if (LP_DEBUG
& DEBUG_FS
) {
3583 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3586 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3589 /* First, check if we've exceeded the max number of shader variants.
3590 * If so, free 6.25% of them (the least recently used ones).
3592 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3594 if (variants_to_cull
||
3595 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3596 struct pipe_context
*pipe
= &lp
->pipe
;
3598 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3599 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3600 "\t%u instrs,\t%u instrs/variant\n",
3601 shader
->variants_cached
,
3602 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3603 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3607 * XXX: we need to flush the context until we have some sort of
3608 * reference counting in fragment shaders as they may still be binned
3609 * Flushing alone might not be sufficient we need to wait on it too.
3611 llvmpipe_finish(pipe
, __FUNCTION__
);
3614 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3615 * number of shader variants (potentially all of them) could be
3616 * pending for destruction on flush.
3619 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3620 struct lp_fs_variant_list_item
*item
;
3621 if (is_empty_list(&lp
->fs_variants_list
)) {
3624 item
= last_elem(&lp
->fs_variants_list
);
3627 llvmpipe_remove_shader_variant(lp
, item
->base
);
3632 * Generate the new variant.
3635 variant
= generate_variant(lp
, shader
, key
);
3638 LP_COUNT_ADD(llvm_compile_time
, dt
);
3639 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3641 /* Put the new variant into the list */
3643 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3644 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3645 lp
->nr_fs_variants
++;
3646 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3647 shader
->variants_cached
++;
3651 /* Bind this variant */
3652 lp_setup_set_fs_variant(lp
->setup
, variant
);
3660 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3662 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3663 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3664 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3666 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3668 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3669 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;