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
,
126 LLVMValueRef mask_input
) /* int32 */
128 LLVMBuilderRef builder
= gallivm
->builder
;
129 struct lp_type mask_type
;
130 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
131 LLVMValueRef bits
[16];
132 LLVMValueRef mask
, bits_vec
;
136 * XXX: We'll need a different path for 16 x u8
138 assert(fs_type
.width
== 32);
139 assert(fs_type
.length
<= ARRAY_SIZE(bits
));
140 mask_type
= lp_int_type(fs_type
);
143 * mask_input >>= (quad * 4)
145 switch (first_quad
) {
150 assert(fs_type
.length
== 4);
157 assert(fs_type
.length
== 4);
165 mask_input
= LLVMBuildLShr(builder
,
167 LLVMConstInt(i32t
, shift
, 0),
171 * mask = { mask_input & (1 << i), for i in [0,3] }
173 mask
= lp_build_broadcast(gallivm
,
174 lp_build_vec_type(gallivm
, mask_type
),
177 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
178 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
179 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1ULL << (j
+ 0), 0);
180 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1ULL << (j
+ 1), 0);
181 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1ULL << (j
+ 4), 0);
182 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1ULL << (j
+ 5), 0);
184 bits_vec
= LLVMConstVector(bits
, fs_type
.length
);
185 mask
= LLVMBuildAnd(builder
, mask
, bits_vec
, "");
188 * mask = mask == bits ? ~0 : 0
190 mask
= lp_build_compare(gallivm
,
191 mask_type
, PIPE_FUNC_EQUAL
,
198 #define EARLY_DEPTH_TEST 0x1
199 #define LATE_DEPTH_TEST 0x2
200 #define EARLY_DEPTH_WRITE 0x4
201 #define LATE_DEPTH_WRITE 0x8
204 find_output_by_semantic( const struct tgsi_shader_info
*info
,
210 for (i
= 0; i
< info
->num_outputs
; i
++)
211 if (info
->output_semantic_name
[i
] == semantic
&&
212 info
->output_semantic_index
[i
] == index
)
220 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
223 lp_llvm_viewport(LLVMValueRef context_ptr
,
224 struct gallivm_state
*gallivm
,
225 LLVMValueRef viewport_index
)
227 LLVMBuilderRef builder
= gallivm
->builder
;
230 struct lp_type viewport_type
=
231 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
233 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
234 ptr
= LLVMBuildPointerCast(builder
, ptr
,
235 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
237 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
244 lp_build_depth_clamp(struct gallivm_state
*gallivm
,
245 LLVMBuilderRef builder
,
247 LLVMValueRef context_ptr
,
248 LLVMValueRef thread_data_ptr
,
251 LLVMValueRef viewport
, min_depth
, max_depth
;
252 LLVMValueRef viewport_index
;
253 struct lp_build_context f32_bld
;
255 assert(type
.floating
);
256 lp_build_context_init(&f32_bld
, gallivm
, type
);
259 * Assumes clamping of the viewport index will occur in setup/gs. Value
260 * is passed through the rasterization stage via lp_rast_shader_inputs.
262 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
265 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
269 * Load the min and max depth from the lp_jit_context.viewports
270 * array of lp_jit_viewport structures.
272 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
274 /* viewports[viewport_index].min_depth */
275 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
276 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
), "");
277 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
279 /* viewports[viewport_index].max_depth */
280 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
281 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
), "");
282 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
285 * Clamp to the min and max depth values for the given viewport.
287 return lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
292 * Generate the fragment shader, depth/stencil test, and alpha tests.
295 generate_fs_loop(struct gallivm_state
*gallivm
,
296 struct lp_fragment_shader
*shader
,
297 const struct lp_fragment_shader_variant_key
*key
,
298 LLVMBuilderRef builder
,
300 LLVMValueRef context_ptr
,
301 LLVMValueRef num_loop
,
302 struct lp_build_interp_soa_context
*interp
,
303 const struct lp_build_sampler_soa
*sampler
,
304 const struct lp_build_image_soa
*image
,
305 LLVMValueRef mask_store
,
306 LLVMValueRef (*out_color
)[4],
307 LLVMValueRef depth_ptr
,
308 LLVMValueRef depth_stride
,
310 LLVMValueRef thread_data_ptr
)
312 const struct util_format_description
*zs_format_desc
= NULL
;
313 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
314 struct lp_type int_type
= lp_int_type(type
);
315 LLVMTypeRef vec_type
, int_vec_type
;
316 LLVMValueRef mask_ptr
, mask_val
;
317 LLVMValueRef consts_ptr
, num_consts_ptr
;
318 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
320 LLVMValueRef z_value
, s_value
;
321 LLVMValueRef z_fb
, s_fb
;
322 LLVMValueRef stencil_refs
[2];
323 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
324 struct lp_build_for_loop_state loop_state
;
325 struct lp_build_mask_context mask
;
327 * TODO: figure out if simple_shader optimization is really worthwile to
328 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
329 * code since tests tend to take another codepath than real shaders.
331 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
332 shader
->info
.base
.num_inputs
< 3 &&
333 shader
->info
.base
.num_instructions
< 8) && 0;
334 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
335 util_blend_state_is_dual(&key
->blend
, 0);
341 struct lp_bld_tgsi_system_values system_values
;
343 memset(&system_values
, 0, sizeof(system_values
));
345 /* truncate then sign extend. */
346 system_values
.front_facing
= LLVMBuildTrunc(gallivm
->builder
, facing
, LLVMInt1TypeInContext(gallivm
->context
), "");
347 system_values
.front_facing
= LLVMBuildSExt(gallivm
->builder
, system_values
.front_facing
, LLVMInt32TypeInContext(gallivm
->context
), "");
349 if (key
->depth
.enabled
||
350 key
->stencil
[0].enabled
) {
352 zs_format_desc
= util_format_description(key
->zsbuf_format
);
353 assert(zs_format_desc
);
355 if (shader
->info
.base
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
])
356 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
357 else if (!shader
->info
.base
.writes_z
&& !shader
->info
.base
.writes_stencil
) {
358 if (shader
->info
.base
.writes_memory
)
359 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
360 else if (key
->alpha
.enabled
||
361 key
->blend
.alpha_to_coverage
||
362 shader
->info
.base
.uses_kill
||
363 shader
->info
.base
.writes_samplemask
) {
364 /* With alpha test and kill, can do the depth test early
365 * and hopefully eliminate some quads. But need to do a
366 * special deferred depth write once the final mask value
367 * is known. This only works though if there's either no
368 * stencil test or the stencil value isn't written.
370 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
371 (key
->stencil
[1].enabled
&&
372 key
->stencil
[1].writemask
)))
373 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
375 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
378 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
381 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
384 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
385 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
386 (key
->stencil
[1].enabled
&&
387 key
->stencil
[1].writemask
))))
388 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
394 vec_type
= lp_build_vec_type(gallivm
, type
);
395 int_vec_type
= lp_build_vec_type(gallivm
, int_type
);
397 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
398 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
399 /* convert scalar stencil refs into vectors */
400 stencil_refs
[0] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[0]);
401 stencil_refs
[1] = lp_build_broadcast(gallivm
, int_vec_type
, stencil_refs
[1]);
403 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
404 num_consts_ptr
= lp_jit_context_num_constants(gallivm
, context_ptr
);
406 ssbo_ptr
= lp_jit_context_ssbos(gallivm
, context_ptr
);
407 num_ssbo_ptr
= lp_jit_context_num_ssbos(gallivm
, context_ptr
);
409 lp_build_for_loop_begin(&loop_state
, gallivm
,
410 lp_build_const_int32(gallivm
, 0),
413 lp_build_const_int32(gallivm
, 1));
415 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
416 &loop_state
.counter
, 1, "mask_ptr");
417 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
419 memset(outputs
, 0, sizeof outputs
);
421 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
422 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
423 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
424 lp_build_vec_type(gallivm
,
429 if (dual_source_blend
) {
430 assert(key
->nr_cbufs
<= 1);
431 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
432 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
433 lp_build_vec_type(gallivm
,
440 /* 'mask' will control execution based on quad's pixel alive/killed state */
441 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
443 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
444 lp_build_mask_check(&mask
);
446 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
449 if (depth_mode
& EARLY_DEPTH_TEST
) {
451 * Clamp according to ARB_depth_clamp semantics.
453 if (key
->depth_clamp
) {
454 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
457 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
458 zs_format_desc
, key
->resource_1d
,
459 depth_ptr
, depth_stride
,
460 &z_fb
, &s_fb
, loop_state
.counter
);
461 lp_build_depth_stencil_test(gallivm
,
473 if (depth_mode
& EARLY_DEPTH_WRITE
) {
474 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
475 zs_format_desc
, key
->resource_1d
,
476 NULL
, NULL
, NULL
, loop_state
.counter
,
477 depth_ptr
, depth_stride
,
481 * Note mask check if stencil is enabled must be after ds write not after
482 * stencil test otherwise new stencil values may not get written if all
483 * fragments got killed by depth/stencil test.
485 if (!simple_shader
&& key
->stencil
[0].enabled
)
486 lp_build_mask_check(&mask
);
489 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
491 struct lp_build_tgsi_params params
;
492 memset(¶ms
, 0, sizeof(params
));
496 params
.consts_ptr
= consts_ptr
;
497 params
.const_sizes_ptr
= num_consts_ptr
;
498 params
.system_values
= &system_values
;
499 params
.inputs
= interp
->inputs
;
500 params
.context_ptr
= context_ptr
;
501 params
.thread_data_ptr
= thread_data_ptr
;
502 params
.sampler
= sampler
;
503 params
.info
= &shader
->info
.base
;
504 params
.ssbo_ptr
= ssbo_ptr
;
505 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
506 params
.image
= image
;
508 /* Build the actual shader */
509 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
510 lp_build_tgsi_soa(gallivm
, tokens
, ¶ms
,
513 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
517 if (key
->alpha
.enabled
) {
518 int color0
= find_output_by_semantic(&shader
->info
.base
,
522 if (color0
!= -1 && outputs
[color0
][3]) {
523 const struct util_format_description
*cbuf_format_desc
;
524 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
525 LLVMValueRef alpha_ref_value
;
527 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
528 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
530 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
532 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
533 &mask
, alpha
, alpha_ref_value
,
534 (depth_mode
& LATE_DEPTH_TEST
) != 0);
538 /* Emulate Alpha to Coverage with Alpha test */
539 if (key
->blend
.alpha_to_coverage
) {
540 int color0
= find_output_by_semantic(&shader
->info
.base
,
544 if (color0
!= -1 && outputs
[color0
][3]) {
545 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
547 lp_build_alpha_to_coverage(gallivm
, type
,
549 (depth_mode
& LATE_DEPTH_TEST
) != 0);
553 if (shader
->info
.base
.writes_samplemask
) {
554 int smaski
= find_output_by_semantic(&shader
->info
.base
,
555 TGSI_SEMANTIC_SAMPLEMASK
,
558 struct lp_build_context smask_bld
;
559 lp_build_context_init(&smask_bld
, gallivm
, int_type
);
562 smask
= LLVMBuildLoad(builder
, outputs
[smaski
][0], "smask");
564 * Pixel is alive according to the first sample in the mask.
566 smask
= LLVMBuildBitCast(builder
, smask
, smask_bld
.vec_type
, "");
567 smask
= lp_build_and(&smask_bld
, smask
, smask_bld
.one
);
568 smask
= lp_build_cmp(&smask_bld
, PIPE_FUNC_NOTEQUAL
, smask
, smask_bld
.zero
);
569 lp_build_mask_update(&mask
, smask
);
573 if (depth_mode
& LATE_DEPTH_TEST
) {
574 int pos0
= find_output_by_semantic(&shader
->info
.base
,
575 TGSI_SEMANTIC_POSITION
,
577 int s_out
= find_output_by_semantic(&shader
->info
.base
,
578 TGSI_SEMANTIC_STENCIL
,
580 if (pos0
!= -1 && outputs
[pos0
][2]) {
581 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
584 * Clamp according to ARB_depth_clamp semantics.
586 if (key
->depth_clamp
) {
587 z
= lp_build_depth_clamp(gallivm
, builder
, type
, context_ptr
,
591 if (s_out
!= -1 && outputs
[s_out
][1]) {
592 /* there's only one value, and spec says to discard additional bits */
593 LLVMValueRef s_max_mask
= lp_build_const_int_vec(gallivm
, int_type
, 255);
594 stencil_refs
[0] = LLVMBuildLoad(builder
, outputs
[s_out
][1], "output.s");
595 stencil_refs
[0] = LLVMBuildBitCast(builder
, stencil_refs
[0], int_vec_type
, "");
596 stencil_refs
[0] = LLVMBuildAnd(builder
, stencil_refs
[0], s_max_mask
, "");
597 stencil_refs
[1] = stencil_refs
[0];
600 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
601 zs_format_desc
, key
->resource_1d
,
602 depth_ptr
, depth_stride
,
603 &z_fb
, &s_fb
, loop_state
.counter
);
605 lp_build_depth_stencil_test(gallivm
,
617 if (depth_mode
& LATE_DEPTH_WRITE
) {
618 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
619 zs_format_desc
, key
->resource_1d
,
620 NULL
, NULL
, NULL
, loop_state
.counter
,
621 depth_ptr
, depth_stride
,
625 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
626 (depth_mode
& LATE_DEPTH_WRITE
))
628 /* Need to apply a reduced mask to the depth write. Reload the
629 * depth value, update from zs_value with the new mask value and
632 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
633 zs_format_desc
, key
->resource_1d
,
634 &mask
, z_fb
, s_fb
, loop_state
.counter
,
635 depth_ptr
, depth_stride
,
641 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
643 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
644 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
645 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
647 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
648 if(outputs
[attrib
][chan
]) {
649 /* XXX: just initialize outputs to point at colors[] and
652 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
653 LLVMValueRef color_ptr
;
654 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
655 &loop_state
.counter
, 1, "");
656 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
657 LLVMBuildStore(builder
, out
, color_ptr
);
663 if (key
->occlusion_count
) {
664 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
665 lp_build_name(counter
, "counter");
666 lp_build_occlusion_count(gallivm
, type
,
667 lp_build_mask_value(&mask
), counter
);
670 mask_val
= lp_build_mask_end(&mask
);
671 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
672 lp_build_for_loop_end(&loop_state
);
677 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
679 * Fragment Shader outputs pixels in small 2x2 blocks
680 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
682 * However in memory pixels are stored in rows
683 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
685 * @param type fragment shader type (4x or 8x float)
686 * @param num_fs number of fs_src
687 * @param is_1d whether we're outputting to a 1d resource
688 * @param dst_channels number of output channels
689 * @param fs_src output from fragment shader
690 * @param dst pointer to store result
691 * @param pad_inline is channel padding inline or at end of row
692 * @return the number of dsts
695 generate_fs_twiddle(struct gallivm_state
*gallivm
,
698 unsigned dst_channels
,
699 LLVMValueRef fs_src
[][4],
703 LLVMValueRef src
[16];
709 unsigned pixels
= type
.length
/ 4;
710 unsigned reorder_group
;
711 unsigned src_channels
;
715 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
716 src_count
= num_fs
* src_channels
;
718 assert(pixels
== 2 || pixels
== 1);
719 assert(num_fs
* src_channels
<= ARRAY_SIZE(src
));
722 * Transpose from SoA -> AoS
724 for (i
= 0; i
< num_fs
; ++i
) {
725 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
729 * Pick transformation options
736 if (dst_channels
== 1) {
742 } else if (dst_channels
== 2) {
746 } else if (dst_channels
> 2) {
753 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
759 * Split the src in half
762 for (i
= num_fs
; i
> 0; --i
) {
763 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
764 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
772 * Ensure pixels are in memory order
775 /* Twiddle pixels by reordering the array, e.g.:
777 * src_count = 8 -> 0 2 1 3 4 6 5 7
778 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
780 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
782 for (i
= 0; i
< src_count
; ++i
) {
783 unsigned group
= i
/ reorder_group
;
784 unsigned block
= (group
/ 4) * 4 * reorder_group
;
785 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
788 } else if (twiddle
) {
789 /* Twiddle pixels across elements of array */
791 * XXX: we should avoid this in some cases, but would need to tell
792 * lp_build_conv to reorder (or deal with it ourselves).
794 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
797 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
801 * Moves any padding between pixels to the end
802 * e.g. RGBXRGBX -> RGBRGBXX
805 unsigned char swizzles
[16];
806 unsigned elems
= pixels
* dst_channels
;
808 for (i
= 0; i
< type
.length
; ++i
) {
810 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
812 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
815 for (i
= 0; i
< src_count
; ++i
) {
816 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
825 * Untwiddle and transpose, much like the above.
826 * However, this is after conversion, so we get packed vectors.
827 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
828 * the vectors will look like:
829 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
830 * be swizzled here). Extending to 16bit should be trivial.
831 * Should also be extended to handle twice wide vectors with AVX2...
834 fs_twiddle_transpose(struct gallivm_state
*gallivm
,
841 struct lp_type type64
, type16
, type32
;
842 LLVMTypeRef type64_t
, type8_t
, type16_t
, type32_t
;
843 LLVMBuilderRef builder
= gallivm
->builder
;
844 LLVMValueRef tmp
[4], shuf
[8];
845 for (j
= 0; j
< 2; j
++) {
846 shuf
[j
*4 + 0] = lp_build_const_int32(gallivm
, j
*4 + 0);
847 shuf
[j
*4 + 1] = lp_build_const_int32(gallivm
, j
*4 + 2);
848 shuf
[j
*4 + 2] = lp_build_const_int32(gallivm
, j
*4 + 1);
849 shuf
[j
*4 + 3] = lp_build_const_int32(gallivm
, j
*4 + 3);
852 assert(src_count
== 4 || src_count
== 2 || src_count
== 1);
853 assert(type
.width
== 8);
854 assert(type
.length
== 16);
856 type8_t
= lp_build_vec_type(gallivm
, type
);
861 type64_t
= lp_build_vec_type(gallivm
, type64
);
866 type16_t
= lp_build_vec_type(gallivm
, type16
);
871 type32_t
= lp_build_vec_type(gallivm
, type32
);
873 lp_build_transpose_aos_n(gallivm
, type
, src
, src_count
, tmp
);
875 if (src_count
== 1) {
876 /* transpose was no-op, just untwiddle */
877 LLVMValueRef shuf_vec
;
878 shuf_vec
= LLVMConstVector(shuf
, 8);
879 tmp
[0] = LLVMBuildBitCast(builder
, src
[0], type16_t
, "");
880 tmp
[0] = LLVMBuildShuffleVector(builder
, tmp
[0], tmp
[0], shuf_vec
, "");
881 dst
[0] = LLVMBuildBitCast(builder
, tmp
[0], type8_t
, "");
882 } else if (src_count
== 2) {
883 LLVMValueRef shuf_vec
;
884 shuf_vec
= LLVMConstVector(shuf
, 4);
886 for (i
= 0; i
< 2; i
++) {
887 tmp
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type32_t
, "");
888 tmp
[i
] = LLVMBuildShuffleVector(builder
, tmp
[i
], tmp
[i
], shuf_vec
, "");
889 dst
[i
] = LLVMBuildBitCast(builder
, tmp
[i
], type8_t
, "");
892 for (j
= 0; j
< 2; j
++) {
893 LLVMValueRef lo
, hi
, lo2
, hi2
;
895 * Note that if we only really have 3 valid channels (rgb)
896 * and we don't need alpha we could substitute a undef here
897 * for the respective channel (causing llvm to drop conversion
900 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
901 lo2
= LLVMBuildBitCast(builder
, tmp
[j
*2], type64_t
, "");
902 hi2
= LLVMBuildBitCast(builder
, tmp
[j
*2 + 1], type64_t
, "");
903 lo
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 0);
904 hi
= lp_build_interleave2(gallivm
, type64
, lo2
, hi2
, 1);
905 dst
[j
*2] = LLVMBuildBitCast(builder
, lo
, type8_t
, "");
906 dst
[j
*2 + 1] = LLVMBuildBitCast(builder
, hi
, type8_t
, "");
913 * Load an unswizzled block of pixels from memory
916 load_unswizzled_block(struct gallivm_state
*gallivm
,
917 LLVMValueRef base_ptr
,
919 unsigned block_width
,
920 unsigned block_height
,
922 struct lp_type dst_type
,
924 unsigned dst_alignment
)
926 LLVMBuilderRef builder
= gallivm
->builder
;
927 unsigned row_size
= dst_count
/ block_height
;
930 /* Ensure block exactly fits into dst */
931 assert((block_width
* block_height
) % dst_count
== 0);
933 for (i
= 0; i
< dst_count
; ++i
) {
934 unsigned x
= i
% row_size
;
935 unsigned y
= i
/ row_size
;
937 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
938 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
941 LLVMValueRef dst_ptr
;
943 gep
[0] = lp_build_const_int32(gallivm
, 0);
944 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
946 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
947 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
,
948 LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
950 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
952 LLVMSetAlignment(dst
[i
], dst_alignment
);
958 * Store an unswizzled block of pixels to memory
961 store_unswizzled_block(struct gallivm_state
*gallivm
,
962 LLVMValueRef base_ptr
,
964 unsigned block_width
,
965 unsigned block_height
,
967 struct lp_type src_type
,
969 unsigned src_alignment
)
971 LLVMBuilderRef builder
= gallivm
->builder
;
972 unsigned row_size
= src_count
/ block_height
;
975 /* Ensure src exactly fits into block */
976 assert((block_width
* block_height
) % src_count
== 0);
978 for (i
= 0; i
< src_count
; ++i
) {
979 unsigned x
= i
% row_size
;
980 unsigned y
= i
/ row_size
;
982 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
983 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
986 LLVMValueRef src_ptr
;
988 gep
[0] = lp_build_const_int32(gallivm
, 0);
989 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
991 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
992 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
,
993 LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
995 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
997 LLVMSetAlignment(src_ptr
, src_alignment
);
1003 * Checks if a format description is an arithmetic format
1005 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1007 static inline boolean
1008 is_arithmetic_format(const struct util_format_description
*format_desc
)
1010 boolean arith
= false;
1013 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1014 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
1015 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
1023 * Checks if this format requires special handling due to required expansion
1024 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1027 static inline boolean
1028 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
1030 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
1031 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1039 * Retrieves the type representing the memory layout for a format
1041 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1044 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
1045 struct lp_type
* type
)
1050 if (format_expands_to_float_soa(format_desc
)) {
1051 /* just make this a uint with width of block */
1052 type
->floating
= false;
1053 type
->fixed
= false;
1056 type
->width
= format_desc
->block
.bits
;
1061 for (i
= 0; i
< 4; i
++)
1062 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1066 memset(type
, 0, sizeof(struct lp_type
));
1067 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1068 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1069 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1070 type
->norm
= format_desc
->channel
[chan
].normalized
;
1072 if (is_arithmetic_format(format_desc
)) {
1076 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
1077 type
->width
+= format_desc
->channel
[i
].size
;
1080 type
->width
= format_desc
->channel
[chan
].size
;
1081 type
->length
= format_desc
->nr_channels
;
1087 * Retrieves the type for a format which is usable in the blending code.
1089 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1092 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
1093 struct lp_type
* type
)
1098 if (format_expands_to_float_soa(format_desc
)) {
1099 /* always use ordinary floats for blending */
1100 type
->floating
= true;
1101 type
->fixed
= false;
1109 for (i
= 0; i
< 4; i
++)
1110 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1114 memset(type
, 0, sizeof(struct lp_type
));
1115 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
1116 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
1117 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
1118 type
->norm
= format_desc
->channel
[chan
].normalized
;
1119 type
->width
= format_desc
->channel
[chan
].size
;
1120 type
->length
= format_desc
->nr_channels
;
1122 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
1123 if (format_desc
->channel
[i
].size
> type
->width
)
1124 type
->width
= format_desc
->channel
[i
].size
;
1127 if (type
->floating
) {
1130 if (type
->width
<= 8) {
1132 } else if (type
->width
<= 16) {
1139 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
1146 * Scale a normalized value from src_bits to dst_bits.
1148 * The exact calculation is
1150 * dst = iround(src * dst_mask / src_mask)
1152 * or with integer rounding
1154 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1158 * src_mask = (1 << src_bits) - 1
1159 * dst_mask = (1 << dst_bits) - 1
1161 * but we try to avoid division and multiplication through shifts.
1163 static inline LLVMValueRef
1164 scale_bits(struct gallivm_state
*gallivm
,
1168 struct lp_type src_type
)
1170 LLVMBuilderRef builder
= gallivm
->builder
;
1171 LLVMValueRef result
= src
;
1173 if (dst_bits
< src_bits
) {
1174 int delta_bits
= src_bits
- dst_bits
;
1176 if (delta_bits
<= dst_bits
) {
1178 * Approximate the rescaling with a single shift.
1180 * This gives the wrong rounding.
1183 result
= LLVMBuildLShr(builder
,
1185 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1190 * Try more accurate rescaling.
1194 * Drop the least significant bits to make space for the multiplication.
1196 * XXX: A better approach would be to use a wider integer type as intermediate. But
1197 * this is enough to convert alpha from 16bits -> 2 when rendering to
1198 * PIPE_FORMAT_R10G10B10A2_UNORM.
1200 result
= LLVMBuildLShr(builder
,
1202 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1206 result
= LLVMBuildMul(builder
,
1208 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1212 * Add a rounding term before the division.
1214 * TODO: Handle signed integers too.
1216 if (!src_type
.sign
) {
1217 result
= LLVMBuildAdd(builder
,
1219 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1224 * Approximate the division by src_mask with a src_bits shift.
1226 * Given the src has already been shifted by dst_bits, all we need
1227 * to do is to shift by the difference.
1230 result
= LLVMBuildLShr(builder
,
1232 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1236 } else if (dst_bits
> src_bits
) {
1238 int db
= dst_bits
- src_bits
;
1240 /* Shift left by difference in bits */
1241 result
= LLVMBuildShl(builder
,
1243 lp_build_const_int_vec(gallivm
, src_type
, db
),
1246 if (db
<= src_bits
) {
1247 /* Enough bits in src to fill the remainder */
1248 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1250 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1253 result
= LLVMBuildOr(builder
, result
, lower
, "");
1254 } else if (db
> src_bits
) {
1255 /* Need to repeatedly copy src bits to fill remainder in dst */
1258 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1259 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1261 result
= LLVMBuildOr(builder
,
1263 LLVMBuildLShr(builder
, result
, shuv
, ""),
1273 * If RT is a smallfloat (needing denorms) format
1276 have_smallfloat_format(struct lp_type dst_type
,
1277 enum pipe_format format
)
1279 return ((dst_type
.floating
&& dst_type
.width
!= 32) ||
1280 /* due to format handling hacks this format doesn't have floating set
1281 * here (and actually has width set to 32 too) so special case this. */
1282 (format
== PIPE_FORMAT_R11G11B10_FLOAT
));
1287 * Convert from memory format to blending format
1289 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1292 convert_to_blend_type(struct gallivm_state
*gallivm
,
1293 unsigned block_size
,
1294 const struct util_format_description
*src_fmt
,
1295 struct lp_type src_type
,
1296 struct lp_type dst_type
,
1297 LLVMValueRef
* src
, // and dst
1300 LLVMValueRef
*dst
= src
;
1301 LLVMBuilderRef builder
= gallivm
->builder
;
1302 struct lp_type blend_type
;
1303 struct lp_type mem_type
;
1305 unsigned pixels
= block_size
/ num_srcs
;
1309 * full custom path for packed floats and srgb formats - none of the later
1310 * functions would do anything useful, and given the lp_type representation they
1311 * can't be fixed. Should really have some SoA blend path for these kind of
1312 * formats rather than hacking them in here.
1314 if (format_expands_to_float_soa(src_fmt
)) {
1315 LLVMValueRef tmpsrc
[4];
1317 * This is pretty suboptimal for this case blending in SoA would be much
1318 * better, since conversion gets us SoA values so need to convert back.
1320 assert(src_type
.width
== 32 || src_type
.width
== 16);
1321 assert(dst_type
.floating
);
1322 assert(dst_type
.width
== 32);
1323 assert(dst_type
.length
% 4 == 0);
1324 assert(num_srcs
% 4 == 0);
1326 if (src_type
.width
== 16) {
1327 /* expand 4x16bit values to 4x32bit */
1328 struct lp_type type32x4
= src_type
;
1329 LLVMTypeRef ltype32x4
;
1330 unsigned num_fetch
= dst_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1331 type32x4
.width
= 32;
1332 ltype32x4
= lp_build_vec_type(gallivm
, type32x4
);
1333 for (i
= 0; i
< num_fetch
; i
++) {
1334 src
[i
] = LLVMBuildZExt(builder
, src
[i
], ltype32x4
, "");
1336 src_type
.width
= 32;
1338 for (i
= 0; i
< 4; i
++) {
1341 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1342 LLVMValueRef tmpsoa
[4];
1343 LLVMValueRef tmps
= tmpsrc
[i
];
1344 if (dst_type
.length
== 8) {
1345 LLVMValueRef shuffles
[8];
1347 /* fetch was 4 values but need 8-wide output values */
1348 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1350 * for 8-wide aos transpose would give us wrong order not matching
1351 * incoming converted fs values and mask. ARGH.
1353 for (j
= 0; j
< 4; j
++) {
1354 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1355 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1357 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1358 LLVMConstVector(shuffles
, 8), "");
1360 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1361 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1364 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1366 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1371 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1372 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1374 /* Is the format arithmetic */
1375 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1376 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1378 /* Pad if necessary */
1379 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1380 for (i
= 0; i
< num_srcs
; ++i
) {
1381 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1384 src_type
.length
= dst_type
.length
;
1387 /* Special case for half-floats */
1388 if (mem_type
.width
== 16 && mem_type
.floating
) {
1389 assert(blend_type
.width
== 32 && blend_type
.floating
);
1390 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1398 src_type
.width
= blend_type
.width
* blend_type
.length
;
1399 blend_type
.length
*= pixels
;
1400 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1402 for (i
= 0; i
< num_srcs
; ++i
) {
1403 LLVMValueRef chans
[4];
1404 LLVMValueRef res
= NULL
;
1406 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1408 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1410 unsigned sa
= src_fmt
->channel
[j
].shift
;
1411 #if UTIL_ARCH_LITTLE_ENDIAN
1412 unsigned from_lsb
= j
;
1414 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1417 mask
= (1 << src_fmt
->channel
[j
].size
) - 1;
1419 /* Extract bits from source */
1420 chans
[j
] = LLVMBuildLShr(builder
,
1422 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1425 chans
[j
] = LLVMBuildAnd(builder
,
1427 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1431 if (src_type
.norm
) {
1432 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1433 blend_type
.width
, chans
[j
], src_type
);
1436 /* Insert bits into correct position */
1437 chans
[j
] = LLVMBuildShl(builder
,
1439 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1445 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1449 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1455 * Convert from blending format to memory format
1457 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1460 convert_from_blend_type(struct gallivm_state
*gallivm
,
1461 unsigned block_size
,
1462 const struct util_format_description
*src_fmt
,
1463 struct lp_type src_type
,
1464 struct lp_type dst_type
,
1465 LLVMValueRef
* src
, // and dst
1468 LLVMValueRef
* dst
= src
;
1470 struct lp_type mem_type
;
1471 struct lp_type blend_type
;
1472 LLVMBuilderRef builder
= gallivm
->builder
;
1473 unsigned pixels
= block_size
/ num_srcs
;
1477 * full custom path for packed floats and srgb formats - none of the later
1478 * functions would do anything useful, and given the lp_type representation they
1479 * can't be fixed. Should really have some SoA blend path for these kind of
1480 * formats rather than hacking them in here.
1482 if (format_expands_to_float_soa(src_fmt
)) {
1484 * This is pretty suboptimal for this case blending in SoA would be much
1485 * better - we need to transpose the AoS values back to SoA values for
1486 * conversion/packing.
1488 assert(src_type
.floating
);
1489 assert(src_type
.width
== 32);
1490 assert(src_type
.length
% 4 == 0);
1491 assert(dst_type
.width
== 32 || dst_type
.width
== 16);
1493 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1494 LLVMValueRef tmpsoa
[4], tmpdst
;
1495 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1496 /* really really need SoA here */
1498 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1499 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1502 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1506 if (src_type
.length
== 8) {
1507 LLVMValueRef tmpaos
, shuffles
[8];
1510 * for 8-wide aos transpose has given us wrong order not matching
1511 * output order. HMPF. Also need to split the output values manually.
1513 for (j
= 0; j
< 4; j
++) {
1514 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1515 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1517 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1518 LLVMConstVector(shuffles
, 8), "");
1519 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1520 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1526 if (dst_type
.width
== 16) {
1527 struct lp_type type16x8
= dst_type
;
1528 struct lp_type type32x4
= dst_type
;
1529 LLVMTypeRef ltype16x4
, ltypei64
, ltypei128
;
1530 unsigned num_fetch
= src_type
.length
== 8 ? num_srcs
/ 2 : num_srcs
/ 4;
1531 type16x8
.length
= 8;
1532 type32x4
.width
= 32;
1533 ltypei128
= LLVMIntTypeInContext(gallivm
->context
, 128);
1534 ltypei64
= LLVMIntTypeInContext(gallivm
->context
, 64);
1535 ltype16x4
= lp_build_vec_type(gallivm
, dst_type
);
1536 /* We could do vector truncation but it doesn't generate very good code */
1537 for (i
= 0; i
< num_fetch
; i
++) {
1538 src
[i
] = lp_build_pack2(gallivm
, type32x4
, type16x8
,
1539 src
[i
], lp_build_zero(gallivm
, type32x4
));
1540 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltypei128
, "");
1541 src
[i
] = LLVMBuildTrunc(builder
, src
[i
], ltypei64
, "");
1542 src
[i
] = LLVMBuildBitCast(builder
, src
[i
], ltype16x4
, "");
1548 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1549 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1551 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1553 /* Special case for half-floats */
1554 if (mem_type
.width
== 16 && mem_type
.floating
) {
1555 int length
= dst_type
.length
;
1556 assert(blend_type
.width
== 32 && blend_type
.floating
);
1558 dst_type
.length
= src_type
.length
;
1560 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1562 dst_type
.length
= length
;
1566 /* Remove any padding */
1567 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1568 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1570 for (i
= 0; i
< num_srcs
; ++i
) {
1571 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1575 /* No bit arithmetic to do */
1580 src_type
.length
= pixels
;
1581 src_type
.width
= blend_type
.length
* blend_type
.width
;
1582 dst_type
.length
= pixels
;
1584 for (i
= 0; i
< num_srcs
; ++i
) {
1585 LLVMValueRef chans
[4];
1586 LLVMValueRef res
= NULL
;
1588 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1590 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1592 unsigned sa
= src_fmt
->channel
[j
].shift
;
1593 unsigned sz_a
= src_fmt
->channel
[j
].size
;
1594 #if UTIL_ARCH_LITTLE_ENDIAN
1595 unsigned from_lsb
= j
;
1597 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1600 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1602 for (k
= 0; k
< blend_type
.width
; ++k
) {
1607 chans
[j
] = LLVMBuildLShr(builder
,
1609 lp_build_const_int_vec(gallivm
, src_type
,
1610 from_lsb
* blend_type
.width
),
1613 chans
[j
] = LLVMBuildAnd(builder
,
1615 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1618 /* Scale down bits */
1619 if (src_type
.norm
) {
1620 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1621 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1622 } else if (!src_type
.floating
&& sz_a
< blend_type
.width
) {
1623 LLVMValueRef mask_val
= lp_build_const_int_vec(gallivm
, src_type
, (1UL << sz_a
) - 1);
1624 LLVMValueRef mask
= LLVMBuildICmp(builder
, LLVMIntUGT
, chans
[j
], mask_val
, "");
1625 chans
[j
] = LLVMBuildSelect(builder
, mask
, mask_val
, chans
[j
], "");
1629 chans
[j
] = LLVMBuildShl(builder
,
1631 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1634 sa
+= src_fmt
->channel
[j
].size
;
1639 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1643 assert (dst_type
.width
!= 24);
1645 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1651 * Convert alpha to same blend type as src
1654 convert_alpha(struct gallivm_state
*gallivm
,
1655 struct lp_type row_type
,
1656 struct lp_type alpha_type
,
1657 const unsigned block_size
,
1658 const unsigned block_height
,
1659 const unsigned src_count
,
1660 const unsigned dst_channels
,
1661 const bool pad_inline
,
1662 LLVMValueRef
* src_alpha
)
1664 LLVMBuilderRef builder
= gallivm
->builder
;
1666 unsigned length
= row_type
.length
;
1667 row_type
.length
= alpha_type
.length
;
1669 /* Twiddle the alpha to match pixels */
1670 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1673 * TODO this should use single lp_build_conv call for
1674 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1676 for (i
= 0; i
< block_height
; ++i
) {
1677 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1680 alpha_type
= row_type
;
1681 row_type
.length
= length
;
1683 /* If only one channel we can only need the single alpha value per pixel */
1684 if (src_count
== 1 && dst_channels
== 1) {
1686 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1688 /* If there are more srcs than rows then we need to split alpha up */
1689 if (src_count
> block_height
) {
1690 for (i
= src_count
; i
> 0; --i
) {
1691 unsigned pixels
= block_size
/ src_count
;
1692 unsigned idx
= i
- 1;
1694 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1695 (idx
* pixels
) % 4, pixels
);
1699 /* If there is a src for each pixel broadcast the alpha across whole row */
1700 if (src_count
== block_size
) {
1701 for (i
= 0; i
< src_count
; ++i
) {
1702 src_alpha
[i
] = lp_build_broadcast(gallivm
,
1703 lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1706 unsigned pixels
= block_size
/ src_count
;
1707 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1708 unsigned alpha_span
= 1;
1709 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1711 /* Check if we need 2 src_alphas for our shuffles */
1712 if (pixels
> alpha_type
.length
) {
1716 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1717 for (j
= 0; j
< row_type
.length
; ++j
) {
1718 if (j
< pixels
* channels
) {
1719 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1721 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1725 for (i
= 0; i
< src_count
; ++i
) {
1726 unsigned idx1
= i
, idx2
= i
;
1728 if (alpha_span
> 1){
1733 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1736 LLVMConstVector(shuffles
, row_type
.length
),
1745 * Generates the blend function for unswizzled colour buffers
1746 * Also generates the read & write from colour buffer
1749 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1751 struct lp_fragment_shader_variant
*variant
,
1752 enum pipe_format out_format
,
1753 unsigned int num_fs
,
1754 struct lp_type fs_type
,
1755 LLVMValueRef
* fs_mask
,
1756 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1757 LLVMValueRef context_ptr
,
1758 LLVMValueRef color_ptr
,
1759 LLVMValueRef stride
,
1760 unsigned partial_mask
,
1763 const unsigned alpha_channel
= 3;
1764 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1765 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1766 const unsigned block_size
= block_width
* block_height
;
1767 const unsigned lp_integer_vector_width
= 128;
1769 LLVMBuilderRef builder
= gallivm
->builder
;
1770 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1771 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1772 LLVMValueRef src_alpha
[4 * 4];
1773 LLVMValueRef src1_alpha
[4 * 4] = { NULL
};
1774 LLVMValueRef src_mask
[4 * 4];
1775 LLVMValueRef src
[4 * 4];
1776 LLVMValueRef src1
[4 * 4];
1777 LLVMValueRef dst
[4 * 4];
1778 LLVMValueRef blend_color
;
1779 LLVMValueRef blend_alpha
;
1780 LLVMValueRef i32_zero
;
1781 LLVMValueRef check_mask
;
1782 LLVMValueRef undef_src_val
;
1784 struct lp_build_mask_context mask_ctx
;
1785 struct lp_type mask_type
;
1786 struct lp_type blend_type
;
1787 struct lp_type row_type
;
1788 struct lp_type dst_type
;
1789 struct lp_type ls_type
;
1791 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1792 unsigned vector_width
;
1793 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1794 unsigned dst_channels
;
1799 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1801 unsigned dst_alignment
;
1803 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1804 bool has_alpha
= false;
1805 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1806 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1808 const boolean is_1d
= variant
->key
.resource_1d
;
1809 boolean twiddle_after_convert
= FALSE
;
1810 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1811 LLVMValueRef fpstate
= 0;
1813 /* Get type from output format */
1814 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1815 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1818 * Technically this code should go into lp_build_smallfloat_to_float
1819 * and lp_build_float_to_smallfloat but due to the
1820 * http://llvm.org/bugs/show_bug.cgi?id=6393
1821 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1822 * So the ordering is important here and there shouldn't be any
1823 * llvm ir instrunctions in this function before
1824 * this, otherwise half-float format conversions won't work
1825 * (again due to llvm bug #6393).
1827 if (have_smallfloat_format(dst_type
, out_format
)) {
1828 /* We need to make sure that denorms are ok for half float
1830 fpstate
= lp_build_fpstate_get(gallivm
);
1831 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1834 mask_type
= lp_int32_vec4_type();
1835 mask_type
.length
= fs_type
.length
;
1837 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1838 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1841 /* Do not bother executing code when mask is empty.. */
1843 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1845 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1846 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1849 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1850 lp_build_mask_check(&mask_ctx
);
1853 partial_mask
|= !variant
->opaque
;
1854 i32_zero
= lp_build_const_int32(gallivm
, 0);
1856 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1858 row_type
.length
= fs_type
.length
;
1859 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1861 /* Compute correct swizzle and count channels */
1862 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1865 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1866 /* Ensure channel is used */
1867 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1871 /* Ensure not already written to (happens in case with GL_ALPHA) */
1872 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1876 /* Ensure we havn't already found all channels */
1877 if (dst_channels
>= out_format_desc
->nr_channels
) {
1881 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1884 if (i
== alpha_channel
) {
1889 if (format_expands_to_float_soa(out_format_desc
)) {
1891 * the code above can't work for layout_other
1892 * for srgb it would sort of work but we short-circuit swizzles, etc.
1893 * as that is done as part of unpack / pack.
1895 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1901 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1904 /* If 3 channels then pad to include alpha for 4 element transpose */
1905 if (dst_channels
== 3) {
1906 assert (!has_alpha
);
1907 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1908 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1911 if (out_format_desc
->nr_channels
== 4) {
1914 * We use alpha from the color conversion, not separate one.
1915 * We had to include it for transpose, hence it will get converted
1916 * too (albeit when doing transpose after conversion, that would
1917 * no longer be the case necessarily).
1918 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1919 * otherwise we really have padding, not alpha, included.)
1926 * Load shader output
1928 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1929 /* Always load alpha for use in blending */
1932 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1935 alpha
= undef_src_val
;
1938 /* Load each channel */
1939 for (j
= 0; j
< dst_channels
; ++j
) {
1940 assert(swizzle
[j
] < 4);
1942 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1945 fs_src
[i
][j
] = undef_src_val
;
1949 /* If 3 channels then pad to include alpha for 4 element transpose */
1951 * XXX If we include that here maybe could actually use it instead of
1952 * separate alpha for blending?
1953 * (Difficult though we actually convert pad channels, not alpha.)
1955 if (dst_channels
== 3 && !has_alpha
) {
1956 fs_src
[i
][3] = alpha
;
1959 /* We split the row_mask and row_alpha as we want 128bit interleave */
1960 if (fs_type
.length
== 8) {
1961 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1963 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
],
1964 src_channels
, src_channels
);
1966 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1967 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
1968 src_channels
, src_channels
);
1970 src_mask
[i
] = fs_mask
[i
];
1971 src_alpha
[i
] = alpha
;
1974 if (dual_source_blend
) {
1975 /* same as above except different src/dst, skip masks and comments... */
1976 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1979 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1982 alpha
= undef_src_val
;
1985 for (j
= 0; j
< dst_channels
; ++j
) {
1986 assert(swizzle
[j
] < 4);
1988 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1991 fs_src1
[i
][j
] = undef_src_val
;
1994 if (dst_channels
== 3 && !has_alpha
) {
1995 fs_src1
[i
][3] = alpha
;
1997 if (fs_type
.length
== 8) {
1998 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1999 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
,
2000 src_channels
, src_channels
);
2002 src1_alpha
[i
] = alpha
;
2007 if (util_format_is_pure_integer(out_format
)) {
2009 * In this case fs_type was really ints or uints disguised as floats,
2012 fs_type
.floating
= 0;
2013 fs_type
.sign
= dst_type
.sign
;
2014 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
2015 for (j
= 0; j
< dst_channels
; ++j
) {
2016 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
2017 lp_build_vec_type(gallivm
, fs_type
), "");
2019 if (dst_channels
== 3 && !has_alpha
) {
2020 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
2021 lp_build_vec_type(gallivm
, fs_type
), "");
2027 * We actually should generally do conversion first (for non-1d cases)
2028 * when the blend format is 8 or 16 bits. The reason is obvious,
2029 * there's 2 or 4 times less vectors to deal with for the interleave...
2030 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2031 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2032 * unpack only with 128bit vectors).
2033 * Note: for 16bit sizes really need matching pack conversion code
2035 if (!is_1d
&& dst_channels
!= 3 && dst_type
.width
== 8) {
2036 twiddle_after_convert
= TRUE
;
2040 * Pixel twiddle from fragment shader order to memory order
2042 if (!twiddle_after_convert
) {
2043 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
2044 dst_channels
, fs_src
, src
, pad_inline
);
2045 if (dual_source_blend
) {
2046 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
2047 fs_src1
, src1
, pad_inline
);
2050 src_count
= num_fullblock_fs
* dst_channels
;
2052 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2053 * (AVX) turn out the same later when untwiddling/transpose (albeit
2054 * for true AVX2 path untwiddle needs to be different).
2055 * For now just order by colors first (so we can use unpack later).
2057 for (j
= 0; j
< num_fullblock_fs
; j
++) {
2058 for (i
= 0; i
< dst_channels
; i
++) {
2059 src
[i
*num_fullblock_fs
+ j
] = fs_src
[j
][i
];
2060 if (dual_source_blend
) {
2061 src1
[i
*num_fullblock_fs
+ j
] = fs_src1
[j
][i
];
2067 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
2068 if (src_count
!= num_fullblock_fs
* src_channels
) {
2069 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
2070 row_type
.length
/= ds
;
2071 fs_type
.length
= row_type
.length
;
2074 blend_type
= row_type
;
2075 mask_type
.length
= 4;
2077 /* Convert src to row_type */
2078 if (dual_source_blend
) {
2079 struct lp_type old_row_type
= row_type
;
2080 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2081 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
2084 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
2087 /* If the rows are not an SSE vector, combine them to become SSE size! */
2088 if ((row_type
.width
* row_type
.length
) % 128) {
2089 unsigned bits
= row_type
.width
* row_type
.length
;
2092 assert(src_count
>= (vector_width
/ bits
));
2094 dst_count
= src_count
/ (vector_width
/ bits
);
2096 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
2097 if (dual_source_blend
) {
2098 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
2101 row_type
.length
*= combined
;
2102 src_count
/= combined
;
2104 bits
= row_type
.width
* row_type
.length
;
2105 assert(bits
== 128 || bits
== 256);
2108 if (twiddle_after_convert
) {
2109 fs_twiddle_transpose(gallivm
, row_type
, src
, src_count
, src
);
2110 if (dual_source_blend
) {
2111 fs_twiddle_transpose(gallivm
, row_type
, src1
, src_count
, src1
);
2116 * Blend Colour conversion
2118 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
2119 blend_color
= LLVMBuildPointerCast(builder
, blend_color
,
2120 LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
2121 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
,
2122 &i32_zero
, 1, ""), "");
2125 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
2127 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
2129 * since blending is done with floats, there was no conversion.
2130 * However, the rules according to fixed point renderbuffers still
2131 * apply, that is we must clamp inputs to 0.0/1.0.
2132 * (This would apply to separate alpha conversion too but we currently
2133 * force has_alpha to be true.)
2134 * TODO: should skip this with "fake" blend, since post-blend conversion
2135 * will clamp anyway.
2136 * TODO: could also skip this if fragment color clamping is enabled. We
2137 * don't support it natively so it gets baked into the shader however, so
2138 * can't really tell here.
2140 struct lp_build_context f32_bld
;
2141 assert(row_type
.floating
);
2142 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
2143 for (i
= 0; i
< src_count
; i
++) {
2144 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
2146 if (dual_source_blend
) {
2147 for (i
= 0; i
< src_count
; i
++) {
2148 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
2151 /* probably can't be different than row_type but better safe than sorry... */
2152 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
2153 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
2157 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
2159 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2160 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
2162 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2163 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
2165 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2166 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
2172 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
2174 if (src_count
< block_height
) {
2175 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
2176 } else if (src_count
> block_height
) {
2177 for (i
= src_count
; i
> 0; --i
) {
2178 unsigned pixels
= block_size
/ src_count
;
2179 unsigned idx
= i
- 1;
2181 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
2182 (idx
* pixels
) % 4, pixels
);
2186 assert(mask_type
.width
== 32);
2188 for (i
= 0; i
< src_count
; ++i
) {
2189 unsigned pixels
= block_size
/ src_count
;
2190 unsigned pixel_width
= row_type
.width
* dst_channels
;
2192 if (pixel_width
== 24) {
2193 mask_type
.width
= 8;
2194 mask_type
.length
= vector_width
/ mask_type
.width
;
2196 mask_type
.length
= pixels
;
2197 mask_type
.width
= row_type
.width
* dst_channels
;
2200 * If mask_type width is smaller than 32bit, this doesn't quite
2201 * generate the most efficient code (could use some pack).
2203 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
],
2204 lp_build_int_vec_type(gallivm
, mask_type
), "");
2206 mask_type
.length
*= dst_channels
;
2207 mask_type
.width
/= dst_channels
;
2210 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
],
2211 lp_build_int_vec_type(gallivm
, mask_type
), "");
2212 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
2219 struct lp_type alpha_type
= fs_type
;
2220 alpha_type
.length
= 4;
2221 convert_alpha(gallivm
, row_type
, alpha_type
,
2222 block_size
, block_height
,
2223 src_count
, dst_channels
,
2224 pad_inline
, src_alpha
);
2225 if (dual_source_blend
) {
2226 convert_alpha(gallivm
, row_type
, alpha_type
,
2227 block_size
, block_height
,
2228 src_count
, dst_channels
,
2229 pad_inline
, src1_alpha
);
2235 * Load dst from memory
2237 if (src_count
< block_height
) {
2238 dst_count
= block_height
;
2240 dst_count
= src_count
;
2243 dst_type
.length
*= block_size
/ dst_count
;
2245 if (format_expands_to_float_soa(out_format_desc
)) {
2247 * we need multiple values at once for the conversion, so can as well
2248 * load them vectorized here too instead of concatenating later.
2249 * (Still need concatenation later for 8-wide vectors).
2251 dst_count
= block_height
;
2252 dst_type
.length
= block_width
;
2256 * Compute the alignment of the destination pointer in bytes
2257 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2258 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2259 * 1d tex but can't distinguish here) so need to stick with per-pixel
2260 * alignment in this case.
2263 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
2266 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
2268 /* Force power-of-two alignment by extracting only the least-significant-bit */
2269 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
2271 * Resource base and stride pointers are aligned to 16 bytes, so that's
2272 * the maximum alignment we can guarantee
2274 dst_alignment
= MIN2(16, dst_alignment
);
2278 if (dst_count
> src_count
) {
2279 if ((dst_type
.width
== 8 || dst_type
.width
== 16) &&
2280 util_is_power_of_two_or_zero(dst_type
.length
) &&
2281 dst_type
.length
* dst_type
.width
< 128) {
2283 * Never try to load values as 4xi8 which we will then
2284 * concatenate to larger vectors. This gives llvm a real
2285 * headache (the problem is the type legalizer (?) will
2286 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2287 * then the shuffles to concatenate are more or less impossible
2288 * - llvm is easily capable of generating a sequence of 32
2289 * pextrb/pinsrb instructions for that. Albeit it appears to
2290 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2291 * width to avoid the trouble (16bit seems not as bad, llvm
2292 * probably recognizes the load+shuffle as only one shuffle
2293 * is necessary, but we can do just the same anyway).
2295 ls_type
.length
= dst_type
.length
* dst_type
.width
/ 32;
2301 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2302 dst
, ls_type
, dst_count
/ 4, dst_alignment
);
2303 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
2304 dst
[i
] = lp_build_undef(gallivm
, ls_type
);
2309 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2310 dst
, ls_type
, dst_count
, dst_alignment
);
2315 * Convert from dst/output format to src/blending format.
2317 * This is necessary as we can only read 1 row from memory at a time,
2318 * so the minimum dst_count will ever be at this point is 4.
2320 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2321 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2322 * on all 16 pixels in that single vector at once.
2324 if (dst_count
> src_count
) {
2325 if (ls_type
.length
!= dst_type
.length
&& ls_type
.length
== 1) {
2326 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, ls_type
);
2327 LLVMTypeRef ls_vec_type
= LLVMVectorType(elem_type
, 1);
2328 for (i
= 0; i
< dst_count
; i
++) {
2329 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
], ls_vec_type
, "");
2333 lp_build_concat_n(gallivm
, ls_type
, dst
, 4, dst
, src_count
);
2335 if (ls_type
.length
!= dst_type
.length
) {
2336 struct lp_type tmp_type
= dst_type
;
2337 tmp_type
.length
= dst_type
.length
* 4 / src_count
;
2338 for (i
= 0; i
< src_count
; i
++) {
2339 dst
[i
] = LLVMBuildBitCast(builder
, dst
[i
],
2340 lp_build_vec_type(gallivm
, tmp_type
), "");
2348 /* XXX this is broken for RGB8 formats -
2349 * they get expanded from 12 to 16 elements (to include alpha)
2350 * by convert_to_blend_type then reduced to 15 instead of 12
2351 * by convert_from_blend_type (a simple fix though breaks A8...).
2352 * R16G16B16 also crashes differently however something going wrong
2353 * inside llvm handling npot vector sizes seemingly.
2354 * It seems some cleanup could be done here (like skipping conversion/blend
2357 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
,
2358 row_type
, dst
, src_count
);
2361 * FIXME: Really should get logic ops / masks out of generic blend / row
2362 * format. Logic ops will definitely not work on the blend float format
2363 * used for SRGB here and I think OpenGL expects this to work as expected
2364 * (that is incoming values converted to srgb then logic op applied).
2366 for (i
= 0; i
< src_count
; ++i
) {
2367 dst
[i
] = lp_build_blend_aos(gallivm
,
2368 &variant
->key
.blend
,
2373 has_alpha
? NULL
: src_alpha
[i
],
2375 has_alpha
? NULL
: src1_alpha
[i
],
2377 partial_mask
? src_mask
[i
] : NULL
,
2379 has_alpha
? NULL
: blend_alpha
,
2381 pad_inline
? 4 : dst_channels
);
2384 convert_from_blend_type(gallivm
, block_size
, out_format_desc
,
2385 row_type
, dst_type
, dst
, src_count
);
2387 /* Split the blend rows back to memory rows */
2388 if (dst_count
> src_count
) {
2389 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2391 if (src_count
== 1) {
2392 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2393 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2395 row_type
.length
/= 2;
2399 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2400 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2401 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2402 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2404 row_type
.length
/= 2;
2409 * Store blend result to memory
2412 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2413 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2416 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2417 dst
, dst_type
, dst_count
, dst_alignment
);
2420 if (have_smallfloat_format(dst_type
, out_format
)) {
2421 lp_build_fpstate_set(gallivm
, fpstate
);
2425 lp_build_mask_end(&mask_ctx
);
2431 * Generate the runtime callable function for the whole fragment pipeline.
2432 * Note that the function which we generate operates on a block of 16
2433 * pixels at at time. The block contains 2x2 quads. Each quad contains
2437 generate_fragment(struct llvmpipe_context
*lp
,
2438 struct lp_fragment_shader
*shader
,
2439 struct lp_fragment_shader_variant
*variant
,
2440 unsigned partial_mask
)
2442 struct gallivm_state
*gallivm
= variant
->gallivm
;
2443 struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2444 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2446 struct lp_type fs_type
;
2447 struct lp_type blend_type
;
2448 LLVMTypeRef fs_elem_type
;
2449 LLVMTypeRef blend_vec_type
;
2450 LLVMTypeRef arg_types
[15];
2451 LLVMTypeRef func_type
;
2452 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2453 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2454 LLVMValueRef context_ptr
;
2457 LLVMValueRef a0_ptr
;
2458 LLVMValueRef dadx_ptr
;
2459 LLVMValueRef dady_ptr
;
2460 LLVMValueRef color_ptr_ptr
;
2461 LLVMValueRef stride_ptr
;
2462 LLVMValueRef color_sample_stride_ptr
;
2463 LLVMValueRef depth_ptr
;
2464 LLVMValueRef depth_stride
;
2465 LLVMValueRef depth_sample_stride
;
2466 LLVMValueRef mask_input
;
2467 LLVMValueRef thread_data_ptr
;
2468 LLVMBasicBlockRef block
;
2469 LLVMBuilderRef builder
;
2470 struct lp_build_sampler_soa
*sampler
;
2471 struct lp_build_image_soa
*image
;
2472 struct lp_build_interp_soa_context interp
;
2473 LLVMValueRef fs_mask
[16 / 4];
2474 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2475 LLVMValueRef function
;
2476 LLVMValueRef facing
;
2481 boolean cbuf0_write_all
;
2482 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2483 util_blend_state_is_dual(&key
->blend
, 0);
2485 assert(lp_native_vector_width
/ 32 >= 4);
2487 /* Adjust color input interpolation according to flatshade state:
2489 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2490 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2491 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2493 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2495 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2499 /* check if writes to cbuf[0] are to be copied to all cbufs */
2501 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
];
2503 /* TODO: actually pick these based on the fs and color buffer
2504 * characteristics. */
2506 memset(&fs_type
, 0, sizeof fs_type
);
2507 fs_type
.floating
= TRUE
; /* floating point values */
2508 fs_type
.sign
= TRUE
; /* values are signed */
2509 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2510 fs_type
.width
= 32; /* 32-bit float */
2511 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2513 memset(&blend_type
, 0, sizeof blend_type
);
2514 blend_type
.floating
= FALSE
; /* values are integers */
2515 blend_type
.sign
= FALSE
; /* values are unsigned */
2516 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2517 blend_type
.width
= 8; /* 8-bit ubyte values */
2518 blend_type
.length
= 16; /* 16 elements per vector */
2521 * Generate the function prototype. Any change here must be reflected in
2522 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2525 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2527 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2529 snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2530 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2532 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2533 arg_types
[1] = int32_type
; /* x */
2534 arg_types
[2] = int32_type
; /* y */
2535 arg_types
[3] = int32_type
; /* facing */
2536 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2537 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2538 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2539 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2540 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2541 arg_types
[9] = int32_type
; /* mask_input */
2542 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2543 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2544 arg_types
[12] = int32_type
; /* depth_stride */
2545 arg_types
[13] = LLVMPointerType(int32_type
, 0); /* color sample strides */
2546 arg_types
[14] = int32_type
; /* depth sample stride */
2548 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2549 arg_types
, ARRAY_SIZE(arg_types
), 0);
2551 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2552 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2554 variant
->function
[partial_mask
] = function
;
2556 /* XXX: need to propagate noalias down into color param now we are
2557 * passing a pointer-to-pointer?
2559 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2560 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2561 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2563 context_ptr
= LLVMGetParam(function
, 0);
2564 x
= LLVMGetParam(function
, 1);
2565 y
= LLVMGetParam(function
, 2);
2566 facing
= LLVMGetParam(function
, 3);
2567 a0_ptr
= LLVMGetParam(function
, 4);
2568 dadx_ptr
= LLVMGetParam(function
, 5);
2569 dady_ptr
= LLVMGetParam(function
, 6);
2570 color_ptr_ptr
= LLVMGetParam(function
, 7);
2571 depth_ptr
= LLVMGetParam(function
, 8);
2572 mask_input
= LLVMGetParam(function
, 9);
2573 thread_data_ptr
= LLVMGetParam(function
, 10);
2574 stride_ptr
= LLVMGetParam(function
, 11);
2575 depth_stride
= LLVMGetParam(function
, 12);
2576 color_sample_stride_ptr
= LLVMGetParam(function
, 13);
2577 depth_sample_stride
= LLVMGetParam(function
, 14);
2579 lp_build_name(context_ptr
, "context");
2580 lp_build_name(x
, "x");
2581 lp_build_name(y
, "y");
2582 lp_build_name(a0_ptr
, "a0");
2583 lp_build_name(dadx_ptr
, "dadx");
2584 lp_build_name(dady_ptr
, "dady");
2585 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2586 lp_build_name(depth_ptr
, "depth");
2587 lp_build_name(mask_input
, "mask_input");
2588 lp_build_name(thread_data_ptr
, "thread_data");
2589 lp_build_name(stride_ptr
, "stride_ptr");
2590 lp_build_name(depth_stride
, "depth_stride");
2591 lp_build_name(color_sample_stride_ptr
, "color_sample_stride_ptr");
2592 lp_build_name(depth_sample_stride
, "depth_sample_stride");
2598 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2599 builder
= gallivm
->builder
;
2601 LLVMPositionBuilderAtEnd(builder
, block
);
2604 * Must not count ps invocations if there's a null shader.
2605 * (It would be ok to count with null shader if there's d/s tests,
2606 * but only if there's d/s buffers too, which is different
2607 * to implicit rasterization disable which must not depend
2608 * on the d/s buffers.)
2609 * Could use popcount on mask, but pixel accuracy is not required.
2610 * Could disable if there's no stats query, but maybe not worth it.
2612 if (shader
->info
.base
.num_instructions
> 1) {
2613 LLVMValueRef invocs
, val
;
2614 invocs
= lp_jit_thread_data_invocations(gallivm
, thread_data_ptr
);
2615 val
= LLVMBuildLoad(builder
, invocs
, "");
2616 val
= LLVMBuildAdd(builder
, val
,
2617 LLVMConstInt(LLVMInt64TypeInContext(gallivm
->context
), 1, 0),
2619 LLVMBuildStore(builder
, val
, invocs
);
2622 /* code generated texture sampling */
2623 sampler
= lp_llvm_sampler_soa_create(key
->samplers
);
2624 image
= lp_llvm_image_soa_create(lp_fs_variant_key_images(key
));
2626 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2627 /* for 1d resources only run "upper half" of stamp */
2628 if (key
->resource_1d
)
2632 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2633 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2634 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2635 num_loop
, "mask_store");
2636 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2637 boolean pixel_center_integer
=
2638 shader
->info
.base
.properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
];
2641 * The shader input interpolation info is not explicitely baked in the
2642 * shader key, but everything it derives from (TGSI, and flatshade) is
2643 * already included in the shader key.
2645 lp_build_interp_soa_init(&interp
,
2647 shader
->info
.base
.num_inputs
,
2649 pixel_center_integer
,
2652 a0_ptr
, dadx_ptr
, dady_ptr
,
2655 for (i
= 0; i
< num_fs
; i
++) {
2657 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2658 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2659 &indexi
, 1, "mask_ptr");
2662 mask
= generate_quad_mask(gallivm
, fs_type
,
2663 i
*fs_type
.length
/4, mask_input
);
2666 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2668 LLVMBuildStore(builder
, mask
, mask_ptr
);
2671 generate_fs_loop(gallivm
,
2680 mask_store
, /* output */
2687 for (i
= 0; i
< num_fs
; i
++) {
2688 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2689 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2691 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2692 /* This is fucked up need to reorganize things */
2693 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2694 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2695 ptr
= LLVMBuildGEP(builder
,
2696 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2698 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2701 if (dual_source_blend
) {
2702 /* only support one dual source blend target hence always use output 1 */
2703 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2704 ptr
= LLVMBuildGEP(builder
,
2705 color_store
[1][chan
],
2707 fs_out_color
[1][chan
][i
] = ptr
;
2713 sampler
->destroy(sampler
);
2714 image
->destroy(image
);
2715 /* Loop over color outputs / color buffers to do blending.
2717 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2718 if (key
->cbuf_format
[cbuf
] != PIPE_FORMAT_NONE
) {
2719 LLVMValueRef color_ptr
;
2720 LLVMValueRef stride
;
2721 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2723 boolean do_branch
= ((key
->depth
.enabled
2724 || key
->stencil
[0].enabled
2725 || key
->alpha
.enabled
)
2726 && !shader
->info
.base
.uses_kill
);
2728 color_ptr
= LLVMBuildLoad(builder
,
2729 LLVMBuildGEP(builder
, color_ptr_ptr
,
2733 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2735 stride
= LLVMBuildLoad(builder
,
2736 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2739 generate_unswizzled_blend(gallivm
, cbuf
, variant
,
2740 key
->cbuf_format
[cbuf
],
2741 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2742 context_ptr
, color_ptr
, stride
,
2743 partial_mask
, do_branch
);
2747 LLVMBuildRetVoid(builder
);
2749 gallivm_verify_function(gallivm
, function
);
2754 dump_fs_variant_key(struct lp_fragment_shader_variant_key
*key
)
2758 debug_printf("fs variant %p:\n", (void *) key
);
2760 if (key
->flatshade
) {
2761 debug_printf("flatshade = 1\n");
2763 if (key
->multisample
) {
2764 debug_printf("multisample = 1\n");
2766 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2767 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2769 if (key
->depth
.enabled
|| key
->stencil
[0].enabled
) {
2770 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2772 if (key
->depth
.enabled
) {
2773 debug_printf("depth.func = %s\n", util_str_func(key
->depth
.func
, TRUE
));
2774 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2777 for (i
= 0; i
< 2; ++i
) {
2778 if (key
->stencil
[i
].enabled
) {
2779 debug_printf("stencil[%u].func = %s\n", i
, util_str_func(key
->stencil
[i
].func
, TRUE
));
2780 debug_printf("stencil[%u].fail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2781 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2782 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_str_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2783 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2784 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2788 if (key
->alpha
.enabled
) {
2789 debug_printf("alpha.func = %s\n", util_str_func(key
->alpha
.func
, TRUE
));
2792 if (key
->occlusion_count
) {
2793 debug_printf("occlusion_count = 1\n");
2796 if (key
->blend
.logicop_enable
) {
2797 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key
->blend
.logicop_func
, TRUE
));
2799 else if (key
->blend
.rt
[0].blend_enable
) {
2800 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2801 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2802 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2803 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2804 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2805 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2807 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2808 if (key
->blend
.alpha_to_coverage
) {
2809 debug_printf("blend.alpha_to_coverage is enabled\n");
2811 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2812 const struct lp_static_sampler_state
*sampler
= &key
->samplers
[i
].sampler_state
;
2813 debug_printf("sampler[%u] = \n", i
);
2814 debug_printf(" .wrap = %s %s %s\n",
2815 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
2816 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
2817 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
2818 debug_printf(" .min_img_filter = %s\n",
2819 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
2820 debug_printf(" .min_mip_filter = %s\n",
2821 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2822 debug_printf(" .mag_img_filter = %s\n",
2823 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
2824 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2825 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
2826 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2827 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2828 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2829 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2830 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2832 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2833 const struct lp_static_texture_state
*texture
= &key
->samplers
[i
].texture_state
;
2834 debug_printf("texture[%u] = \n", i
);
2835 debug_printf(" .format = %s\n",
2836 util_format_name(texture
->format
));
2837 debug_printf(" .target = %s\n",
2838 util_str_tex_target(texture
->target
, TRUE
));
2839 debug_printf(" .level_zero_only = %u\n",
2840 texture
->level_zero_only
);
2841 debug_printf(" .pot = %u %u %u\n",
2843 texture
->pot_height
,
2844 texture
->pot_depth
);
2846 struct lp_image_static_state
*images
= lp_fs_variant_key_images(key
);
2847 for (i
= 0; i
< key
->nr_images
; ++i
) {
2848 const struct lp_static_texture_state
*image
= &images
[i
].image_state
;
2849 debug_printf("image[%u] = \n", i
);
2850 debug_printf(" .format = %s\n",
2851 util_format_name(image
->format
));
2852 debug_printf(" .target = %s\n",
2853 util_str_tex_target(image
->target
, TRUE
));
2854 debug_printf(" .level_zero_only = %u\n",
2855 image
->level_zero_only
);
2856 debug_printf(" .pot = %u %u %u\n",
2865 lp_debug_fs_variant(struct lp_fragment_shader_variant
*variant
)
2867 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2868 variant
->shader
->no
, variant
->no
);
2869 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
2870 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2872 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
2873 dump_fs_variant_key(&variant
->key
);
2874 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2880 * Generate a new fragment shader variant from the shader code and
2881 * other state indicated by the key.
2883 static struct lp_fragment_shader_variant
*
2884 generate_variant(struct llvmpipe_context
*lp
,
2885 struct lp_fragment_shader
*shader
,
2886 const struct lp_fragment_shader_variant_key
*key
)
2888 struct lp_fragment_shader_variant
*variant
;
2889 const struct util_format_description
*cbuf0_format_desc
= NULL
;
2890 boolean fullcolormask
;
2891 char module_name
[64];
2893 variant
= MALLOC(sizeof *variant
+ shader
->variant_key_size
- sizeof variant
->key
);
2897 memset(variant
, 0, sizeof(*variant
));
2898 snprintf(module_name
, sizeof(module_name
), "fs%u_variant%u",
2899 shader
->no
, shader
->variants_created
);
2901 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
2902 if (!variant
->gallivm
) {
2907 variant
->shader
= shader
;
2908 variant
->list_item_global
.base
= variant
;
2909 variant
->list_item_local
.base
= variant
;
2910 variant
->no
= shader
->variants_created
++;
2912 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2915 * Determine whether we are touching all channels in the color buffer.
2917 fullcolormask
= FALSE
;
2918 if (key
->nr_cbufs
== 1) {
2919 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2920 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2924 !key
->blend
.logicop_enable
&&
2925 !key
->blend
.rt
[0].blend_enable
&&
2927 !key
->stencil
[0].enabled
&&
2928 !key
->alpha
.enabled
&&
2929 !key
->blend
.alpha_to_coverage
&&
2930 !key
->depth
.enabled
&&
2931 !shader
->info
.base
.uses_kill
&&
2932 !shader
->info
.base
.writes_samplemask
2935 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2936 lp_debug_fs_variant(variant
);
2939 lp_jit_init_types(variant
);
2941 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2942 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2944 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2945 if (variant
->opaque
) {
2946 /* Specialized shader, which doesn't need to read the color buffer. */
2947 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2952 * Compile everything
2955 gallivm_compile_module(variant
->gallivm
);
2957 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
2959 if (variant
->function
[RAST_EDGE_TEST
]) {
2960 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2961 gallivm_jit_function(variant
->gallivm
,
2962 variant
->function
[RAST_EDGE_TEST
]);
2965 if (variant
->function
[RAST_WHOLE
]) {
2966 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2967 gallivm_jit_function(variant
->gallivm
,
2968 variant
->function
[RAST_WHOLE
]);
2969 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2970 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2973 gallivm_free_ir(variant
->gallivm
);
2980 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2981 const struct pipe_shader_state
*templ
)
2983 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2984 struct lp_fragment_shader
*shader
;
2986 int nr_sampler_views
;
2990 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2994 shader
->no
= fs_no
++;
2995 make_empty_list(&shader
->variants
);
2997 shader
->base
.type
= templ
->type
;
2998 if (templ
->type
== PIPE_SHADER_IR_TGSI
) {
2999 /* get/save the summary info for this shader */
3000 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
3002 /* we need to keep a local copy of the tokens */
3003 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
3005 shader
->base
.ir
.nir
= templ
->ir
.nir
;
3006 nir_tgsi_scan_shader(templ
->ir
.nir
, &shader
->info
.base
, true);
3009 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
3010 if (shader
->draw_data
== NULL
) {
3011 FREE((void *) shader
->base
.tokens
);
3016 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3017 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3018 nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3019 shader
->variant_key_size
= lp_fs_variant_key_size(MAX2(nr_samplers
, nr_sampler_views
), nr_images
);
3021 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
3022 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
3023 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
3025 switch (shader
->info
.base
.input_interpolate
[i
]) {
3026 case TGSI_INTERPOLATE_CONSTANT
:
3027 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
3029 case TGSI_INTERPOLATE_LINEAR
:
3030 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
3032 case TGSI_INTERPOLATE_PERSPECTIVE
:
3033 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
3035 case TGSI_INTERPOLATE_COLOR
:
3036 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
3043 switch (shader
->info
.base
.input_semantic_name
[i
]) {
3044 case TGSI_SEMANTIC_FACE
:
3045 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
3047 case TGSI_SEMANTIC_POSITION
:
3048 /* Position was already emitted above
3050 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
3051 shader
->inputs
[i
].src_index
= 0;
3055 /* XXX this is a completely pointless index map... */
3056 shader
->inputs
[i
].src_index
= i
+1;
3059 if (LP_DEBUG
& DEBUG_TGSI
) {
3061 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3062 shader
->no
, (void *) shader
);
3063 tgsi_dump(templ
->tokens
, 0);
3064 debug_printf("usage masks:\n");
3065 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
3066 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
3067 debug_printf(" IN[%u].%s%s%s%s\n",
3069 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
3070 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
3071 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
3072 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
3082 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
3084 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3085 struct lp_fragment_shader
*lp_fs
= (struct lp_fragment_shader
*)fs
;
3086 if (llvmpipe
->fs
== lp_fs
)
3089 draw_bind_fragment_shader(llvmpipe
->draw
,
3090 (lp_fs
? lp_fs
->draw_data
: NULL
));
3092 llvmpipe
->fs
= lp_fs
;
3094 llvmpipe
->dirty
|= LP_NEW_FS
;
3099 * Remove shader variant from two lists: the shader's variant list
3100 * and the context's variant list.
3103 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
3104 struct lp_fragment_shader_variant
*variant
)
3106 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
3107 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3108 "v total cached %u inst %u total inst %u\n",
3109 variant
->shader
->no
, variant
->no
,
3110 variant
->shader
->variants_created
,
3111 variant
->shader
->variants_cached
,
3112 lp
->nr_fs_variants
, variant
->nr_instrs
, lp
->nr_fs_instrs
);
3115 gallivm_destroy(variant
->gallivm
);
3117 /* remove from shader's list */
3118 remove_from_list(&variant
->list_item_local
);
3119 variant
->shader
->variants_cached
--;
3121 /* remove from context's list */
3122 remove_from_list(&variant
->list_item_global
);
3123 lp
->nr_fs_variants
--;
3124 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
3131 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
3133 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3134 struct lp_fragment_shader
*shader
= fs
;
3135 struct lp_fs_variant_list_item
*li
;
3137 assert(fs
!= llvmpipe
->fs
);
3140 * XXX: we need to flush the context until we have some sort of reference
3141 * counting in fragment shaders as they may still be binned
3142 * Flushing alone might not sufficient we need to wait on it too.
3144 llvmpipe_finish(pipe
, __FUNCTION__
);
3146 /* Delete all the variants */
3147 li
= first_elem(&shader
->variants
);
3148 while(!at_end(&shader
->variants
, li
)) {
3149 struct lp_fs_variant_list_item
*next
= next_elem(li
);
3150 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
3154 /* Delete draw module's data */
3155 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
3157 if (shader
->base
.ir
.nir
)
3158 ralloc_free(shader
->base
.ir
.nir
);
3159 assert(shader
->variants_cached
== 0);
3160 FREE((void *) shader
->base
.tokens
);
3167 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
3168 enum pipe_shader_type shader
, uint index
,
3169 const struct pipe_constant_buffer
*cb
)
3171 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3172 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
3174 assert(shader
< PIPE_SHADER_TYPES
);
3175 assert(index
< ARRAY_SIZE(llvmpipe
->constants
[shader
]));
3177 /* note: reference counting */
3178 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
3181 if (!(constants
->bind
& PIPE_BIND_CONSTANT_BUFFER
)) {
3182 debug_printf("Illegal set constant without bind flag\n");
3183 constants
->bind
|= PIPE_BIND_CONSTANT_BUFFER
;
3187 if (shader
== PIPE_SHADER_VERTEX
||
3188 shader
== PIPE_SHADER_GEOMETRY
||
3189 shader
== PIPE_SHADER_TESS_CTRL
||
3190 shader
== PIPE_SHADER_TESS_EVAL
) {
3191 /* Pass the constants to the 'draw' module */
3192 const unsigned size
= cb
? cb
->buffer_size
: 0;
3196 data
= (ubyte
*) llvmpipe_resource_data(constants
);
3198 else if (cb
&& cb
->user_buffer
) {
3199 data
= (ubyte
*) cb
->user_buffer
;
3206 data
+= cb
->buffer_offset
;
3208 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
3211 else if (shader
== PIPE_SHADER_COMPUTE
)
3212 llvmpipe
->cs_dirty
|= LP_CSNEW_CONSTANTS
;
3214 llvmpipe
->dirty
|= LP_NEW_FS_CONSTANTS
;
3216 if (cb
&& cb
->user_buffer
) {
3217 pipe_resource_reference(&constants
, NULL
);
3222 llvmpipe_set_shader_buffers(struct pipe_context
*pipe
,
3223 enum pipe_shader_type shader
, unsigned start_slot
,
3224 unsigned count
, const struct pipe_shader_buffer
*buffers
,
3225 unsigned writable_bitmask
)
3227 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3229 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3230 const struct pipe_shader_buffer
*buffer
= buffers
? &buffers
[idx
] : NULL
;
3232 util_copy_shader_buffer(&llvmpipe
->ssbos
[shader
][i
], buffer
);
3234 if (shader
== PIPE_SHADER_VERTEX
||
3235 shader
== PIPE_SHADER_GEOMETRY
||
3236 shader
== PIPE_SHADER_TESS_CTRL
||
3237 shader
== PIPE_SHADER_TESS_EVAL
) {
3238 const unsigned size
= buffer
? buffer
->buffer_size
: 0;
3239 const ubyte
*data
= NULL
;
3240 if (buffer
&& buffer
->buffer
)
3241 data
= (ubyte
*) llvmpipe_resource_data(buffer
->buffer
);
3243 data
+= buffer
->buffer_offset
;
3244 draw_set_mapped_shader_buffer(llvmpipe
->draw
, shader
,
3246 } else if (shader
== PIPE_SHADER_COMPUTE
) {
3247 llvmpipe
->cs_dirty
|= LP_CSNEW_SSBOS
;
3248 } else if (shader
== PIPE_SHADER_FRAGMENT
) {
3249 llvmpipe
->dirty
|= LP_NEW_FS_SSBOS
;
3255 llvmpipe_set_shader_images(struct pipe_context
*pipe
,
3256 enum pipe_shader_type shader
, unsigned start_slot
,
3257 unsigned count
, const struct pipe_image_view
*images
)
3259 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
3262 draw_flush(llvmpipe
->draw
);
3263 for (i
= start_slot
, idx
= 0; i
< start_slot
+ count
; i
++, idx
++) {
3264 const struct pipe_image_view
*image
= images
? &images
[idx
] : NULL
;
3266 util_copy_image_view(&llvmpipe
->images
[shader
][i
], image
);
3269 llvmpipe
->num_images
[shader
] = start_slot
+ count
;
3270 if (shader
== PIPE_SHADER_VERTEX
||
3271 shader
== PIPE_SHADER_GEOMETRY
||
3272 shader
== PIPE_SHADER_TESS_CTRL
||
3273 shader
== PIPE_SHADER_TESS_EVAL
) {
3274 draw_set_images(llvmpipe
->draw
,
3276 llvmpipe
->images
[shader
],
3277 start_slot
+ count
);
3278 } else if (shader
== PIPE_SHADER_COMPUTE
)
3279 llvmpipe
->cs_dirty
|= LP_CSNEW_IMAGES
;
3281 llvmpipe
->dirty
|= LP_NEW_FS_IMAGES
;
3285 * Return the blend factor equivalent to a destination alpha of one.
3287 static inline unsigned
3288 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
3291 case PIPE_BLENDFACTOR_DST_ALPHA
:
3292 return PIPE_BLENDFACTOR_ONE
;
3293 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
3294 return PIPE_BLENDFACTOR_ZERO
;
3295 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
3297 return PIPE_BLENDFACTOR_ZERO
;
3299 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
3307 * We need to generate several variants of the fragment pipeline to match
3308 * all the combinations of the contributing state atoms.
3310 * TODO: there is actually no reason to tie this to context state -- the
3311 * generated code could be cached globally in the screen.
3313 static struct lp_fragment_shader_variant_key
*
3314 make_variant_key(struct llvmpipe_context
*lp
,
3315 struct lp_fragment_shader
*shader
,
3319 struct lp_fragment_shader_variant_key
*key
;
3321 key
= (struct lp_fragment_shader_variant_key
*)store
;
3323 memset(key
, 0, offsetof(struct lp_fragment_shader_variant_key
, samplers
[1]));
3325 if (lp
->framebuffer
.zsbuf
) {
3326 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
3327 const struct util_format_description
*zsbuf_desc
=
3328 util_format_description(zsbuf_format
);
3330 if (lp
->depth_stencil
->depth
.enabled
&&
3331 util_format_has_depth(zsbuf_desc
)) {
3332 key
->zsbuf_format
= zsbuf_format
;
3333 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
3335 if (lp
->depth_stencil
->stencil
[0].enabled
&&
3336 util_format_has_stencil(zsbuf_desc
)) {
3337 key
->zsbuf_format
= zsbuf_format
;
3338 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
3340 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
3341 key
->resource_1d
= TRUE
;
3346 * Propagate the depth clamp setting from the rasterizer state.
3347 * depth_clip == 0 implies depth clamping is enabled.
3349 * When clip_halfz is enabled, then always clamp the depth values.
3351 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3352 * clamp is always active in d3d10, regardless if depth clip is
3354 * (GL has an always-on [0,1] clamp on fs depth output instead
3355 * to ensure the depth values stay in range. Doesn't look like
3356 * we do that, though...)
3358 if (lp
->rasterizer
->clip_halfz
) {
3359 key
->depth_clamp
= 1;
3361 key
->depth_clamp
= (lp
->rasterizer
->depth_clip_near
== 0) ? 1 : 0;
3364 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3365 if (!lp
->framebuffer
.nr_cbufs
||
3366 !lp
->framebuffer
.cbufs
[0] ||
3367 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
3368 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
3370 if(key
->alpha
.enabled
)
3371 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
3372 /* alpha.ref_value is passed in jit_context */
3374 key
->flatshade
= lp
->rasterizer
->flatshade
;
3375 key
->multisample
= lp
->rasterizer
->multisample
;
3376 if (lp
->active_occlusion_queries
&& !lp
->queries_disabled
) {
3377 key
->occlusion_count
= TRUE
;
3380 if (lp
->framebuffer
.nr_cbufs
) {
3381 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
3384 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
3386 if (!key
->blend
.independent_blend_enable
) {
3387 /* we always need independent blend otherwise the fixups below won't work */
3388 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
3389 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
3391 key
->blend
.independent_blend_enable
= 1;
3394 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
3395 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
3397 if (lp
->framebuffer
.cbufs
[i
]) {
3398 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
3399 const struct util_format_description
*format_desc
;
3401 key
->cbuf_format
[i
] = format
;
3404 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3405 * mixing of 2d textures with height 1 and 1d textures, so make sure
3406 * we pick 1d if any cbuf or zsbuf is 1d.
3408 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[i
]->texture
)) {
3409 key
->resource_1d
= TRUE
;
3412 format_desc
= util_format_description(format
);
3413 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
3414 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
3417 * Mask out color channels not present in the color buffer.
3419 blend_rt
->colormask
&= util_format_colormask(format_desc
);
3422 * Disable blend for integer formats.
3424 if (util_format_is_pure_integer(format
)) {
3425 blend_rt
->blend_enable
= 0;
3429 * Our swizzled render tiles always have an alpha channel, but the
3430 * linear render target format often does not, so force here the dst
3433 * This is not a mere optimization. Wrong results will be produced if
3434 * the dst alpha is used, the dst format does not have alpha, and the
3435 * previous rendering was not flushed from the swizzled to linear
3436 * buffer. For example, NonPowTwo DCT.
3438 * TODO: This should be generalized to all channels for better
3439 * performance, but only alpha causes correctness issues.
3441 * Also, force rgb/alpha func/factors match, to make AoS blending
3444 if (format_desc
->swizzle
[3] > PIPE_SWIZZLE_W
||
3445 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
3446 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3447 boolean clamped_zero
= !util_format_is_float(format
) &&
3448 !util_format_is_snorm(format
);
3449 blend_rt
->rgb_src_factor
=
3450 force_dst_alpha_one(blend_rt
->rgb_src_factor
, clamped_zero
);
3451 blend_rt
->rgb_dst_factor
=
3452 force_dst_alpha_one(blend_rt
->rgb_dst_factor
, clamped_zero
);
3453 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
3454 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
3455 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
3459 /* no color buffer for this fragment output */
3460 key
->cbuf_format
[i
] = PIPE_FORMAT_NONE
;
3461 blend_rt
->colormask
= 0x0;
3462 blend_rt
->blend_enable
= 0;
3466 /* This value will be the same for all the variants of a given shader:
3468 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3470 struct lp_sampler_static_state
*fs_sampler
;
3472 fs_sampler
= key
->samplers
;
3474 memset(fs_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *fs_sampler
);
3476 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
3477 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3478 lp_sampler_static_sampler_state(&fs_sampler
[i
].sampler_state
,
3479 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
3484 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3485 * are dx10-style? Can't really have mixed opcodes, at least not
3486 * if we want to skip the holes here (without rescanning tgsi).
3488 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3489 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3490 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3492 * Note sview may exceed what's representable by file_mask.
3493 * This will still work, the only downside is that not actually
3494 * used views may be included in the shader key.
3496 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
3497 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3498 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3503 key
->nr_sampler_views
= key
->nr_samplers
;
3504 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
3505 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
3506 lp_sampler_static_texture_state(&fs_sampler
[i
].texture_state
,
3507 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3512 struct lp_image_static_state
*lp_image
;
3513 lp_image
= lp_fs_variant_key_images(key
);
3514 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
3515 for (i
= 0; i
< key
->nr_images
; ++i
) {
3516 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
3517 lp_sampler_static_texture_state_image(&lp_image
[i
].image_state
,
3518 &lp
->images
[PIPE_SHADER_FRAGMENT
][i
]);
3527 * Update fragment shader state. This is called just prior to drawing
3528 * something when some fragment-related state has changed.
3531 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3533 struct lp_fragment_shader
*shader
= lp
->fs
;
3534 struct lp_fragment_shader_variant_key
*key
;
3535 struct lp_fragment_shader_variant
*variant
= NULL
;
3536 struct lp_fs_variant_list_item
*li
;
3537 char store
[LP_FS_MAX_VARIANT_KEY_SIZE
];
3539 key
= make_variant_key(lp
, shader
, store
);
3541 /* Search the variants for one which matches the key */
3542 li
= first_elem(&shader
->variants
);
3543 while(!at_end(&shader
->variants
, li
)) {
3544 if(memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
3552 /* Move this variant to the head of the list to implement LRU
3553 * deletion of shader's when we have too many.
3555 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3558 /* variant not found, create it now */
3561 unsigned variants_to_cull
;
3563 if (LP_DEBUG
& DEBUG_FS
) {
3564 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3567 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3570 /* First, check if we've exceeded the max number of shader variants.
3571 * If so, free 6.25% of them (the least recently used ones).
3573 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
3575 if (variants_to_cull
||
3576 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3577 struct pipe_context
*pipe
= &lp
->pipe
;
3579 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
3580 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3581 "\t%u instrs,\t%u instrs/variant\n",
3582 shader
->variants_cached
,
3583 lp
->nr_fs_variants
, lp
->nr_fs_instrs
,
3584 lp
->nr_fs_instrs
/ lp
->nr_fs_variants
);
3588 * XXX: we need to flush the context until we have some sort of
3589 * reference counting in fragment shaders as they may still be binned
3590 * Flushing alone might not be sufficient we need to wait on it too.
3592 llvmpipe_finish(pipe
, __FUNCTION__
);
3595 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3596 * number of shader variants (potentially all of them) could be
3597 * pending for destruction on flush.
3600 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3601 struct lp_fs_variant_list_item
*item
;
3602 if (is_empty_list(&lp
->fs_variants_list
)) {
3605 item
= last_elem(&lp
->fs_variants_list
);
3608 llvmpipe_remove_shader_variant(lp
, item
->base
);
3613 * Generate the new variant.
3616 variant
= generate_variant(lp
, shader
, key
);
3619 LP_COUNT_ADD(llvm_compile_time
, dt
);
3620 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3622 /* Put the new variant into the list */
3624 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3625 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3626 lp
->nr_fs_variants
++;
3627 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3628 shader
->variants_cached
++;
3632 /* Bind this variant */
3633 lp_setup_set_fs_variant(lp
->setup
, variant
);
3641 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3643 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3644 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3645 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3647 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3649 llvmpipe
->pipe
.set_shader_buffers
= llvmpipe_set_shader_buffers
;
3650 llvmpipe
->pipe
.set_shader_images
= llvmpipe_set_shader_images
;