2 * Copyright © 2017 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "nir_serialize.h"
25 #include "nir_control_flow.h"
26 #include "util/u_dynarray.h"
27 #include "util/u_math.h"
29 #define NIR_SERIALIZE_FUNC_HAS_IMPL ((void *)(intptr_t)1)
30 #define MAX_OBJECT_IDS (1 << 20)
39 const nir_shader
*nir
;
43 /* maps pointer to index */
44 struct hash_table
*remap_table
;
46 /* the next index to assign to a NIR in-memory object */
49 /* Array of write_phi_fixup structs representing phi sources that need to
50 * be resolved in the second pass.
52 struct util_dynarray phi_fixups
;
54 /* The last serialized type. */
55 const struct glsl_type
*last_type
;
56 const struct glsl_type
*last_interface_type
;
57 struct nir_variable_data last_var_data
;
59 /* For skipping equal ALU headers (typical after scalarization). */
60 nir_instr_type last_instr_type
;
61 uintptr_t last_alu_header_offset
;
63 /* Don't write optional data such as variable names. */
70 struct blob_reader
*blob
;
72 /* the next index to assign to a NIR in-memory object */
75 /* The length of the index -> object table */
76 uint32_t idx_table_len
;
78 /* map from index to deserialized pointer */
81 /* List of phi sources. */
82 struct list_head phi_srcs
;
84 /* The last deserialized type. */
85 const struct glsl_type
*last_type
;
86 const struct glsl_type
*last_interface_type
;
87 struct nir_variable_data last_var_data
;
91 write_add_object(write_ctx
*ctx
, const void *obj
)
93 uint32_t index
= ctx
->next_idx
++;
94 assert(index
!= MAX_OBJECT_IDS
);
95 _mesa_hash_table_insert(ctx
->remap_table
, obj
, (void *)(uintptr_t) index
);
99 write_lookup_object(write_ctx
*ctx
, const void *obj
)
101 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->remap_table
, obj
);
103 return (uint32_t)(uintptr_t) entry
->data
;
107 read_add_object(read_ctx
*ctx
, void *obj
)
109 assert(ctx
->next_idx
< ctx
->idx_table_len
);
110 ctx
->idx_table
[ctx
->next_idx
++] = obj
;
114 read_lookup_object(read_ctx
*ctx
, uint32_t idx
)
116 assert(idx
< ctx
->idx_table_len
);
117 return ctx
->idx_table
[idx
];
121 read_object(read_ctx
*ctx
)
123 return read_lookup_object(ctx
, blob_read_uint32(ctx
->blob
));
127 encode_bit_size_3bits(uint8_t bit_size
)
129 /* Encode values of 0, 1, 2, 4, 8, 16, 32, 64 in 3 bits. */
130 assert(bit_size
<= 64 && util_is_power_of_two_or_zero(bit_size
));
132 return util_logbase2(bit_size
) + 1;
137 decode_bit_size_3bits(uint8_t bit_size
)
140 return 1 << (bit_size
- 1);
144 #define NUM_COMPONENTS_IS_SEPARATE_7 7
147 encode_num_components_in_3bits(uint8_t num_components
)
149 if (num_components
<= 4)
150 return num_components
;
151 if (num_components
== 8)
153 if (num_components
== 16)
156 /* special value indicating that num_components is in the next uint32 */
157 return NUM_COMPONENTS_IS_SEPARATE_7
;
161 decode_num_components_in_3bits(uint8_t value
)
170 unreachable("invalid num_components encoding");
175 write_constant(write_ctx
*ctx
, const nir_constant
*c
)
177 blob_write_bytes(ctx
->blob
, c
->values
, sizeof(c
->values
));
178 blob_write_uint32(ctx
->blob
, c
->num_elements
);
179 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
180 write_constant(ctx
, c
->elements
[i
]);
183 static nir_constant
*
184 read_constant(read_ctx
*ctx
, nir_variable
*nvar
)
186 nir_constant
*c
= ralloc(nvar
, nir_constant
);
188 blob_copy_bytes(ctx
->blob
, (uint8_t *)c
->values
, sizeof(c
->values
));
189 c
->num_elements
= blob_read_uint32(ctx
->blob
);
190 c
->elements
= ralloc_array(nvar
, nir_constant
*, c
->num_elements
);
191 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
192 c
->elements
[i
] = read_constant(ctx
, nvar
);
197 enum var_data_encoding
{
199 var_encode_shader_temp
,
200 var_encode_function_temp
,
201 var_encode_location_diff
,
208 unsigned has_constant_initializer
:1;
209 unsigned has_interface_type
:1;
210 unsigned num_state_slots
:7;
211 unsigned data_encoding
:2;
212 unsigned type_same_as_last
:1;
213 unsigned interface_type_same_as_last
:1;
215 unsigned num_members
:16;
219 union packed_var_data_diff
{
224 int driver_location
:16;
229 write_variable(write_ctx
*ctx
, const nir_variable
*var
)
231 write_add_object(ctx
, var
);
233 assert(var
->num_state_slots
< (1 << 7));
234 assert(var
->num_members
< (1 << 16));
236 STATIC_ASSERT(sizeof(union packed_var
) == 4);
237 union packed_var flags
;
240 flags
.u
.has_name
= !ctx
->strip
&& var
->name
;
241 flags
.u
.has_constant_initializer
= !!(var
->constant_initializer
);
242 flags
.u
.has_interface_type
= !!(var
->interface_type
);
243 flags
.u
.type_same_as_last
= var
->type
== ctx
->last_type
;
244 flags
.u
.interface_type_same_as_last
=
245 var
->interface_type
&& var
->interface_type
== ctx
->last_interface_type
;
246 flags
.u
.num_state_slots
= var
->num_state_slots
;
247 flags
.u
.num_members
= var
->num_members
;
249 struct nir_variable_data data
= var
->data
;
251 /* When stripping, we expect that the location is no longer needed,
252 * which is typically after shaders are linked.
255 data
.mode
!= nir_var_shader_in
&&
256 data
.mode
!= nir_var_shader_out
)
259 /* Temporary variables don't serialize var->data. */
260 if (data
.mode
== nir_var_shader_temp
)
261 flags
.u
.data_encoding
= var_encode_shader_temp
;
262 else if (data
.mode
== nir_var_function_temp
)
263 flags
.u
.data_encoding
= var_encode_function_temp
;
265 struct nir_variable_data tmp
= data
;
267 tmp
.location
= ctx
->last_var_data
.location
;
268 tmp
.location_frac
= ctx
->last_var_data
.location_frac
;
269 tmp
.driver_location
= ctx
->last_var_data
.driver_location
;
271 /* See if we can encode only the difference in locations from the last
274 if (memcmp(&ctx
->last_var_data
, &tmp
, sizeof(tmp
)) == 0 &&
275 abs((int)data
.location
-
276 (int)ctx
->last_var_data
.location
) < (1 << 12) &&
277 abs((int)data
.driver_location
-
278 (int)ctx
->last_var_data
.driver_location
) < (1 << 15))
279 flags
.u
.data_encoding
= var_encode_location_diff
;
281 flags
.u
.data_encoding
= var_encode_full
;
284 blob_write_uint32(ctx
->blob
, flags
.u32
);
286 if (!flags
.u
.type_same_as_last
) {
287 encode_type_to_blob(ctx
->blob
, var
->type
);
288 ctx
->last_type
= var
->type
;
291 if (var
->interface_type
&& !flags
.u
.interface_type_same_as_last
) {
292 encode_type_to_blob(ctx
->blob
, var
->interface_type
);
293 ctx
->last_interface_type
= var
->interface_type
;
296 if (flags
.u
.has_name
)
297 blob_write_string(ctx
->blob
, var
->name
);
299 if (flags
.u
.data_encoding
== var_encode_full
||
300 flags
.u
.data_encoding
== var_encode_location_diff
) {
301 if (flags
.u
.data_encoding
== var_encode_full
) {
302 blob_write_bytes(ctx
->blob
, &data
, sizeof(data
));
304 /* Serialize only the difference in locations from the last variable.
306 union packed_var_data_diff diff
;
308 diff
.u
.location
= data
.location
- ctx
->last_var_data
.location
;
309 diff
.u
.location_frac
= data
.location_frac
-
310 ctx
->last_var_data
.location_frac
;
311 diff
.u
.driver_location
= data
.driver_location
-
312 ctx
->last_var_data
.driver_location
;
314 blob_write_uint32(ctx
->blob
, diff
.u32
);
317 ctx
->last_var_data
= data
;
320 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
321 blob_write_bytes(ctx
->blob
, &var
->state_slots
[i
],
322 sizeof(var
->state_slots
[i
]));
324 if (var
->constant_initializer
)
325 write_constant(ctx
, var
->constant_initializer
);
326 if (var
->num_members
> 0) {
327 blob_write_bytes(ctx
->blob
, (uint8_t *) var
->members
,
328 var
->num_members
* sizeof(*var
->members
));
332 static nir_variable
*
333 read_variable(read_ctx
*ctx
)
335 nir_variable
*var
= rzalloc(ctx
->nir
, nir_variable
);
336 read_add_object(ctx
, var
);
338 union packed_var flags
;
339 flags
.u32
= blob_read_uint32(ctx
->blob
);
341 if (flags
.u
.type_same_as_last
) {
342 var
->type
= ctx
->last_type
;
344 var
->type
= decode_type_from_blob(ctx
->blob
);
345 ctx
->last_type
= var
->type
;
348 if (flags
.u
.has_interface_type
) {
349 if (flags
.u
.interface_type_same_as_last
) {
350 var
->interface_type
= ctx
->last_interface_type
;
352 var
->interface_type
= decode_type_from_blob(ctx
->blob
);
353 ctx
->last_interface_type
= var
->interface_type
;
357 if (flags
.u
.has_name
) {
358 const char *name
= blob_read_string(ctx
->blob
);
359 var
->name
= ralloc_strdup(var
, name
);
364 if (flags
.u
.data_encoding
== var_encode_shader_temp
)
365 var
->data
.mode
= nir_var_shader_temp
;
366 else if (flags
.u
.data_encoding
== var_encode_function_temp
)
367 var
->data
.mode
= nir_var_function_temp
;
368 else if (flags
.u
.data_encoding
== var_encode_full
) {
369 blob_copy_bytes(ctx
->blob
, (uint8_t *) &var
->data
, sizeof(var
->data
));
370 ctx
->last_var_data
= var
->data
;
371 } else { /* var_encode_location_diff */
372 union packed_var_data_diff diff
;
373 diff
.u32
= blob_read_uint32(ctx
->blob
);
375 var
->data
= ctx
->last_var_data
;
376 var
->data
.location
+= diff
.u
.location
;
377 var
->data
.location_frac
+= diff
.u
.location_frac
;
378 var
->data
.driver_location
+= diff
.u
.driver_location
;
380 ctx
->last_var_data
= var
->data
;
383 var
->num_state_slots
= flags
.u
.num_state_slots
;
384 if (var
->num_state_slots
!= 0) {
385 var
->state_slots
= ralloc_array(var
, nir_state_slot
,
386 var
->num_state_slots
);
387 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
388 blob_copy_bytes(ctx
->blob
, &var
->state_slots
[i
],
389 sizeof(var
->state_slots
[i
]));
392 if (flags
.u
.has_constant_initializer
)
393 var
->constant_initializer
= read_constant(ctx
, var
);
395 var
->constant_initializer
= NULL
;
396 var
->num_members
= flags
.u
.num_members
;
397 if (var
->num_members
> 0) {
398 var
->members
= ralloc_array(var
, struct nir_variable_data
,
400 blob_copy_bytes(ctx
->blob
, (uint8_t *) var
->members
,
401 var
->num_members
* sizeof(*var
->members
));
408 write_var_list(write_ctx
*ctx
, const struct exec_list
*src
)
410 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
411 foreach_list_typed(nir_variable
, var
, node
, src
) {
412 write_variable(ctx
, var
);
417 read_var_list(read_ctx
*ctx
, struct exec_list
*dst
)
419 exec_list_make_empty(dst
);
420 unsigned num_vars
= blob_read_uint32(ctx
->blob
);
421 for (unsigned i
= 0; i
< num_vars
; i
++) {
422 nir_variable
*var
= read_variable(ctx
);
423 exec_list_push_tail(dst
, &var
->node
);
428 write_register(write_ctx
*ctx
, const nir_register
*reg
)
430 write_add_object(ctx
, reg
);
431 blob_write_uint32(ctx
->blob
, reg
->num_components
);
432 blob_write_uint32(ctx
->blob
, reg
->bit_size
);
433 blob_write_uint32(ctx
->blob
, reg
->num_array_elems
);
434 blob_write_uint32(ctx
->blob
, reg
->index
);
435 blob_write_uint32(ctx
->blob
, !ctx
->strip
&& reg
->name
);
436 if (!ctx
->strip
&& reg
->name
)
437 blob_write_string(ctx
->blob
, reg
->name
);
440 static nir_register
*
441 read_register(read_ctx
*ctx
)
443 nir_register
*reg
= ralloc(ctx
->nir
, nir_register
);
444 read_add_object(ctx
, reg
);
445 reg
->num_components
= blob_read_uint32(ctx
->blob
);
446 reg
->bit_size
= blob_read_uint32(ctx
->blob
);
447 reg
->num_array_elems
= blob_read_uint32(ctx
->blob
);
448 reg
->index
= blob_read_uint32(ctx
->blob
);
449 bool has_name
= blob_read_uint32(ctx
->blob
);
451 const char *name
= blob_read_string(ctx
->blob
);
452 reg
->name
= ralloc_strdup(reg
, name
);
457 list_inithead(®
->uses
);
458 list_inithead(®
->defs
);
459 list_inithead(®
->if_uses
);
465 write_reg_list(write_ctx
*ctx
, const struct exec_list
*src
)
467 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
468 foreach_list_typed(nir_register
, reg
, node
, src
)
469 write_register(ctx
, reg
);
473 read_reg_list(read_ctx
*ctx
, struct exec_list
*dst
)
475 exec_list_make_empty(dst
);
476 unsigned num_regs
= blob_read_uint32(ctx
->blob
);
477 for (unsigned i
= 0; i
< num_regs
; i
++) {
478 nir_register
*reg
= read_register(ctx
);
479 exec_list_push_tail(dst
, ®
->node
);
486 unsigned is_ssa
:1; /* <-- Header */
487 unsigned is_indirect
:1;
488 unsigned object_idx
:20;
489 unsigned _footer
:10; /* <-- Footer */
492 unsigned _header
:22; /* <-- Header */
493 unsigned negate
:1; /* <-- Footer */
495 unsigned swizzle_x
:2;
496 unsigned swizzle_y
:2;
497 unsigned swizzle_z
:2;
498 unsigned swizzle_w
:2;
501 unsigned _header
:22; /* <-- Header */
502 unsigned src_type
:5; /* <-- Footer */
508 write_src_full(write_ctx
*ctx
, const nir_src
*src
, union packed_src header
)
510 /* Since sources are very frequent, we try to save some space when storing
511 * them. In particular, we store whether the source is a register and
512 * whether the register has an indirect index in the low two bits. We can
513 * assume that the high two bits of the index are zero, since otherwise our
514 * address space would've been exhausted allocating the remap table!
516 header
.any
.is_ssa
= src
->is_ssa
;
518 header
.any
.object_idx
= write_lookup_object(ctx
, src
->ssa
);
519 blob_write_uint32(ctx
->blob
, header
.u32
);
521 header
.any
.object_idx
= write_lookup_object(ctx
, src
->reg
.reg
);
522 header
.any
.is_indirect
= !!src
->reg
.indirect
;
523 blob_write_uint32(ctx
->blob
, header
.u32
);
524 blob_write_uint32(ctx
->blob
, src
->reg
.base_offset
);
525 if (src
->reg
.indirect
) {
526 union packed_src header
= {0};
527 write_src_full(ctx
, src
->reg
.indirect
, header
);
533 write_src(write_ctx
*ctx
, const nir_src
*src
)
535 union packed_src header
= {0};
536 write_src_full(ctx
, src
, header
);
539 static union packed_src
540 read_src(read_ctx
*ctx
, nir_src
*src
, void *mem_ctx
)
542 STATIC_ASSERT(sizeof(union packed_src
) == 4);
543 union packed_src header
;
544 header
.u32
= blob_read_uint32(ctx
->blob
);
546 src
->is_ssa
= header
.any
.is_ssa
;
548 src
->ssa
= read_lookup_object(ctx
, header
.any
.object_idx
);
550 src
->reg
.reg
= read_lookup_object(ctx
, header
.any
.object_idx
);
551 src
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
552 if (header
.any
.is_indirect
) {
553 src
->reg
.indirect
= ralloc(mem_ctx
, nir_src
);
554 read_src(ctx
, src
->reg
.indirect
, mem_ctx
);
556 src
->reg
.indirect
= NULL
;
567 uint8_t num_components
:3;
572 uint8_t is_indirect
:1;
577 enum intrinsic_const_indices_encoding
{
578 /* Use the 9 bits of packed_const_indices to store 1-9 indices.
579 * 1 9-bit index, or 2 4-bit indices, or 3 3-bit indices, or
580 * 4 2-bit indices, or 5-9 1-bit indices.
582 * The common case for load_ubo is 0, 0, 0, which is trivially represented.
583 * The common cases for load_interpolated_input also fit here, e.g.: 7, 3
585 const_indices_9bit_all_combined
,
587 const_indices_8bit
, /* 8 bits per element */
588 const_indices_16bit
, /* 16 bits per element */
589 const_indices_32bit
, /* 32 bits per element */
592 enum load_const_packing
{
593 /* Constants are not packed and are stored in following dwords. */
596 /* packed_value contains high 19 bits, low bits are 0,
597 * good for floating-point decimals
599 load_const_scalar_hi_19bits
,
601 /* packed_value contains low 19 bits, high bits are sign-extended */
602 load_const_scalar_lo_19bits_sext
,
608 unsigned instr_type
:4; /* always present */
610 unsigned dest
:8; /* always last */
613 unsigned instr_type
:4;
615 unsigned no_signed_wrap
:1;
616 unsigned no_unsigned_wrap
:1;
618 /* Reg: writemask; SSA: swizzles for 2 srcs */
619 unsigned writemask_or_two_swizzles
:4;
621 unsigned packed_src_ssa_16bit
:1;
622 /* Scalarized ALUs always have the same header. */
623 unsigned num_followup_alu_sharing_header
:2;
627 unsigned instr_type
:4;
628 unsigned deref_type
:3;
629 unsigned cast_type_same_as_last
:1;
630 unsigned mode
:10; /* deref_var redefines this */
631 unsigned packed_src_ssa_16bit
:1; /* deref_var redefines this */
632 unsigned _pad
:5; /* deref_var redefines this */
636 unsigned instr_type
:4;
637 unsigned deref_type
:3;
639 unsigned object_idx
:16; /* if 0, the object ID is a separate uint32 */
643 unsigned instr_type
:4;
644 unsigned intrinsic
:9;
645 unsigned const_indices_encoding
:2;
646 unsigned packed_const_indices
:9;
650 unsigned instr_type
:4;
651 unsigned last_component
:4;
653 unsigned packing
:2; /* enum load_const_packing */
654 unsigned packed_value
:19; /* meaning determined by packing */
657 unsigned instr_type
:4;
658 unsigned last_component
:4;
663 unsigned instr_type
:4;
666 unsigned texture_array_size
:12;
670 unsigned instr_type
:4;
671 unsigned num_srcs
:20;
675 unsigned instr_type
:4;
681 /* Write "lo24" as low 24 bits in the first uint32. */
683 write_dest(write_ctx
*ctx
, const nir_dest
*dst
, union packed_instr header
,
684 nir_instr_type instr_type
)
686 STATIC_ASSERT(sizeof(union packed_dest
) == 1);
687 union packed_dest dest
;
690 dest
.ssa
.is_ssa
= dst
->is_ssa
;
692 dest
.ssa
.has_name
= !ctx
->strip
&& dst
->ssa
.name
;
693 dest
.ssa
.num_components
=
694 encode_num_components_in_3bits(dst
->ssa
.num_components
);
695 dest
.ssa
.bit_size
= encode_bit_size_3bits(dst
->ssa
.bit_size
);
697 dest
.reg
.is_indirect
= !!(dst
->reg
.indirect
);
699 header
.any
.dest
= dest
.u8
;
701 /* Check if the current ALU instruction has the same header as the previous
702 * instruction that is also ALU. If it is, we don't have to write
703 * the current header. This is a typical occurence after scalarization.
705 if (instr_type
== nir_instr_type_alu
) {
706 bool equal_header
= false;
708 if (ctx
->last_instr_type
== nir_instr_type_alu
) {
709 assert(ctx
->last_alu_header_offset
);
710 union packed_instr
*last_header
=
711 (union packed_instr
*)(ctx
->blob
->data
+
712 ctx
->last_alu_header_offset
);
714 /* Clear the field that counts ALUs with equal headers. */
715 union packed_instr clean_header
;
716 clean_header
.u32
= last_header
->u32
;
717 clean_header
.alu
.num_followup_alu_sharing_header
= 0;
719 /* There can be at most 4 consecutive ALU instructions
720 * sharing the same header.
722 if (last_header
->alu
.num_followup_alu_sharing_header
< 3 &&
723 header
.u32
== clean_header
.u32
) {
724 last_header
->alu
.num_followup_alu_sharing_header
++;
730 ctx
->last_alu_header_offset
= ctx
->blob
->size
;
731 blob_write_uint32(ctx
->blob
, header
.u32
);
734 blob_write_uint32(ctx
->blob
, header
.u32
);
737 if (dest
.ssa
.is_ssa
&&
738 dest
.ssa
.num_components
== NUM_COMPONENTS_IS_SEPARATE_7
)
739 blob_write_uint32(ctx
->blob
, dst
->ssa
.num_components
);
742 write_add_object(ctx
, &dst
->ssa
);
743 if (dest
.ssa
.has_name
)
744 blob_write_string(ctx
->blob
, dst
->ssa
.name
);
746 blob_write_uint32(ctx
->blob
, write_lookup_object(ctx
, dst
->reg
.reg
));
747 blob_write_uint32(ctx
->blob
, dst
->reg
.base_offset
);
748 if (dst
->reg
.indirect
)
749 write_src(ctx
, dst
->reg
.indirect
);
754 read_dest(read_ctx
*ctx
, nir_dest
*dst
, nir_instr
*instr
,
755 union packed_instr header
)
757 union packed_dest dest
;
758 dest
.u8
= header
.any
.dest
;
760 if (dest
.ssa
.is_ssa
) {
761 unsigned bit_size
= decode_bit_size_3bits(dest
.ssa
.bit_size
);
762 unsigned num_components
;
763 if (dest
.ssa
.num_components
== NUM_COMPONENTS_IS_SEPARATE_7
)
764 num_components
= blob_read_uint32(ctx
->blob
);
766 num_components
= decode_num_components_in_3bits(dest
.ssa
.num_components
);
767 char *name
= dest
.ssa
.has_name
? blob_read_string(ctx
->blob
) : NULL
;
768 nir_ssa_dest_init(instr
, dst
, num_components
, bit_size
, name
);
769 read_add_object(ctx
, &dst
->ssa
);
771 dst
->reg
.reg
= read_object(ctx
);
772 dst
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
773 if (dest
.reg
.is_indirect
) {
774 dst
->reg
.indirect
= ralloc(instr
, nir_src
);
775 read_src(ctx
, dst
->reg
.indirect
, instr
);
781 are_object_ids_16bit(write_ctx
*ctx
)
783 /* Check the highest object ID, because they are monotonic. */
784 return ctx
->next_idx
< (1 << 16);
788 is_alu_src_ssa_16bit(write_ctx
*ctx
, const nir_alu_instr
*alu
)
790 unsigned num_srcs
= nir_op_infos
[alu
->op
].num_inputs
;
792 for (unsigned i
= 0; i
< num_srcs
; i
++) {
793 if (!alu
->src
[i
].src
.is_ssa
|| alu
->src
[i
].abs
|| alu
->src
[i
].negate
)
796 unsigned src_components
= nir_ssa_alu_instr_src_components(alu
, i
);
798 for (unsigned chan
= 0; chan
< src_components
; chan
++) {
799 /* The swizzles for src0.x and src1.x are stored
800 * in writemask_or_two_swizzles for SSA ALUs.
802 if (alu
->dest
.dest
.is_ssa
&& i
< 2 && chan
== 0 &&
803 alu
->src
[i
].swizzle
[chan
] < 4)
806 if (alu
->src
[i
].swizzle
[chan
] != chan
)
811 return are_object_ids_16bit(ctx
);
815 write_alu(write_ctx
*ctx
, const nir_alu_instr
*alu
)
817 unsigned num_srcs
= nir_op_infos
[alu
->op
].num_inputs
;
818 unsigned dst_components
= nir_dest_num_components(alu
->dest
.dest
);
820 /* 9 bits for nir_op */
821 STATIC_ASSERT(nir_num_opcodes
<= 512);
822 union packed_instr header
;
825 header
.alu
.instr_type
= alu
->instr
.type
;
826 header
.alu
.exact
= alu
->exact
;
827 header
.alu
.no_signed_wrap
= alu
->no_signed_wrap
;
828 header
.alu
.no_unsigned_wrap
= alu
->no_unsigned_wrap
;
829 header
.alu
.saturate
= alu
->dest
.saturate
;
830 header
.alu
.op
= alu
->op
;
831 header
.alu
.packed_src_ssa_16bit
= is_alu_src_ssa_16bit(ctx
, alu
);
833 if (header
.alu
.packed_src_ssa_16bit
&&
834 alu
->dest
.dest
.is_ssa
) {
835 /* For packed srcs of SSA ALUs, this field stores the swizzles. */
836 header
.alu
.writemask_or_two_swizzles
= alu
->src
[0].swizzle
[0];
838 header
.alu
.writemask_or_two_swizzles
|= alu
->src
[1].swizzle
[0] << 2;
839 } else if (!alu
->dest
.dest
.is_ssa
&& dst_components
<= 4) {
840 /* For vec4 registers, this field is a writemask. */
841 header
.alu
.writemask_or_two_swizzles
= alu
->dest
.write_mask
;
844 write_dest(ctx
, &alu
->dest
.dest
, header
, alu
->instr
.type
);
846 if (!alu
->dest
.dest
.is_ssa
&& dst_components
> 4)
847 blob_write_uint32(ctx
->blob
, alu
->dest
.write_mask
);
849 if (header
.alu
.packed_src_ssa_16bit
) {
850 for (unsigned i
= 0; i
< num_srcs
; i
++) {
851 assert(alu
->src
[i
].src
.is_ssa
);
852 unsigned idx
= write_lookup_object(ctx
, alu
->src
[i
].src
.ssa
);
853 assert(idx
< (1 << 16));
854 blob_write_uint16(ctx
->blob
, idx
);
857 for (unsigned i
= 0; i
< num_srcs
; i
++) {
858 unsigned src_channels
= nir_ssa_alu_instr_src_components(alu
, i
);
859 unsigned src_components
= nir_src_num_components(alu
->src
[i
].src
);
860 union packed_src src
;
861 bool packed
= src_components
<= 4 && src_channels
<= 4;
864 src
.alu
.negate
= alu
->src
[i
].negate
;
865 src
.alu
.abs
= alu
->src
[i
].abs
;
868 src
.alu
.swizzle_x
= alu
->src
[i
].swizzle
[0];
869 src
.alu
.swizzle_y
= alu
->src
[i
].swizzle
[1];
870 src
.alu
.swizzle_z
= alu
->src
[i
].swizzle
[2];
871 src
.alu
.swizzle_w
= alu
->src
[i
].swizzle
[3];
874 write_src_full(ctx
, &alu
->src
[i
].src
, src
);
876 /* Store swizzles for vec8 and vec16. */
878 for (unsigned o
= 0; o
< src_channels
; o
+= 8) {
881 for (unsigned j
= 0; j
< 8 && o
+ j
< src_channels
; j
++) {
882 value
|= (uint32_t)alu
->src
[i
].swizzle
[o
+ j
] <<
883 (4 * j
); /* 4 bits per swizzle */
886 blob_write_uint32(ctx
->blob
, value
);
893 static nir_alu_instr
*
894 read_alu(read_ctx
*ctx
, union packed_instr header
)
896 unsigned num_srcs
= nir_op_infos
[header
.alu
.op
].num_inputs
;
897 nir_alu_instr
*alu
= nir_alu_instr_create(ctx
->nir
, header
.alu
.op
);
899 alu
->exact
= header
.alu
.exact
;
900 alu
->no_signed_wrap
= header
.alu
.no_signed_wrap
;
901 alu
->no_unsigned_wrap
= header
.alu
.no_unsigned_wrap
;
902 alu
->dest
.saturate
= header
.alu
.saturate
;
904 read_dest(ctx
, &alu
->dest
.dest
, &alu
->instr
, header
);
906 unsigned dst_components
= nir_dest_num_components(alu
->dest
.dest
);
908 if (alu
->dest
.dest
.is_ssa
) {
909 alu
->dest
.write_mask
= u_bit_consecutive(0, dst_components
);
910 } else if (dst_components
<= 4) {
911 alu
->dest
.write_mask
= header
.alu
.writemask_or_two_swizzles
;
913 alu
->dest
.write_mask
= blob_read_uint32(ctx
->blob
);
916 if (header
.alu
.packed_src_ssa_16bit
) {
917 for (unsigned i
= 0; i
< num_srcs
; i
++) {
918 nir_alu_src
*src
= &alu
->src
[i
];
919 src
->src
.is_ssa
= true;
920 src
->src
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
922 memset(&src
->swizzle
, 0, sizeof(src
->swizzle
));
924 unsigned src_components
= nir_ssa_alu_instr_src_components(alu
, i
);
926 for (unsigned chan
= 0; chan
< src_components
; chan
++)
927 src
->swizzle
[chan
] = chan
;
930 for (unsigned i
= 0; i
< num_srcs
; i
++) {
931 union packed_src src
= read_src(ctx
, &alu
->src
[i
].src
, &alu
->instr
);
932 unsigned src_channels
= nir_ssa_alu_instr_src_components(alu
, i
);
933 unsigned src_components
= nir_src_num_components(alu
->src
[i
].src
);
934 bool packed
= src_components
<= 4 && src_channels
<= 4;
936 alu
->src
[i
].negate
= src
.alu
.negate
;
937 alu
->src
[i
].abs
= src
.alu
.abs
;
939 memset(&alu
->src
[i
].swizzle
, 0, sizeof(alu
->src
[i
].swizzle
));
942 alu
->src
[i
].swizzle
[0] = src
.alu
.swizzle_x
;
943 alu
->src
[i
].swizzle
[1] = src
.alu
.swizzle_y
;
944 alu
->src
[i
].swizzle
[2] = src
.alu
.swizzle_z
;
945 alu
->src
[i
].swizzle
[3] = src
.alu
.swizzle_w
;
947 /* Load swizzles for vec8 and vec16. */
948 for (unsigned o
= 0; o
< src_channels
; o
+= 8) {
949 unsigned value
= blob_read_uint32(ctx
->blob
);
951 for (unsigned j
= 0; j
< 8 && o
+ j
< src_channels
; j
++) {
952 alu
->src
[i
].swizzle
[o
+ j
] =
953 (value
>> (4 * j
)) & 0xf; /* 4 bits per swizzle */
960 if (header
.alu
.packed_src_ssa_16bit
&&
961 alu
->dest
.dest
.is_ssa
) {
962 alu
->src
[0].swizzle
[0] = header
.alu
.writemask_or_two_swizzles
& 0x3;
964 alu
->src
[1].swizzle
[0] = header
.alu
.writemask_or_two_swizzles
>> 2;
971 write_deref(write_ctx
*ctx
, const nir_deref_instr
*deref
)
973 assert(deref
->deref_type
< 8);
974 assert(deref
->mode
< (1 << 10));
976 union packed_instr header
;
979 header
.deref
.instr_type
= deref
->instr
.type
;
980 header
.deref
.deref_type
= deref
->deref_type
;
982 if (deref
->deref_type
== nir_deref_type_cast
) {
983 header
.deref
.mode
= deref
->mode
;
984 header
.deref
.cast_type_same_as_last
= deref
->type
== ctx
->last_type
;
987 unsigned var_idx
= 0;
988 if (deref
->deref_type
== nir_deref_type_var
) {
989 var_idx
= write_lookup_object(ctx
, deref
->var
);
990 if (var_idx
&& var_idx
< (1 << 16))
991 header
.deref_var
.object_idx
= var_idx
;
994 if (deref
->deref_type
== nir_deref_type_array
||
995 deref
->deref_type
== nir_deref_type_ptr_as_array
) {
996 header
.deref
.packed_src_ssa_16bit
=
997 deref
->parent
.is_ssa
&& deref
->arr
.index
.is_ssa
&&
998 are_object_ids_16bit(ctx
);
1001 write_dest(ctx
, &deref
->dest
, header
, deref
->instr
.type
);
1003 switch (deref
->deref_type
) {
1004 case nir_deref_type_var
:
1005 if (!header
.deref_var
.object_idx
)
1006 blob_write_uint32(ctx
->blob
, var_idx
);
1009 case nir_deref_type_struct
:
1010 write_src(ctx
, &deref
->parent
);
1011 blob_write_uint32(ctx
->blob
, deref
->strct
.index
);
1014 case nir_deref_type_array
:
1015 case nir_deref_type_ptr_as_array
:
1016 if (header
.deref
.packed_src_ssa_16bit
) {
1017 blob_write_uint16(ctx
->blob
,
1018 write_lookup_object(ctx
, deref
->parent
.ssa
));
1019 blob_write_uint16(ctx
->blob
,
1020 write_lookup_object(ctx
, deref
->arr
.index
.ssa
));
1022 write_src(ctx
, &deref
->parent
);
1023 write_src(ctx
, &deref
->arr
.index
);
1027 case nir_deref_type_cast
:
1028 write_src(ctx
, &deref
->parent
);
1029 blob_write_uint32(ctx
->blob
, deref
->cast
.ptr_stride
);
1030 if (!header
.deref
.cast_type_same_as_last
) {
1031 encode_type_to_blob(ctx
->blob
, deref
->type
);
1032 ctx
->last_type
= deref
->type
;
1036 case nir_deref_type_array_wildcard
:
1037 write_src(ctx
, &deref
->parent
);
1041 unreachable("Invalid deref type");
1045 static nir_deref_instr
*
1046 read_deref(read_ctx
*ctx
, union packed_instr header
)
1048 nir_deref_type deref_type
= header
.deref
.deref_type
;
1049 nir_deref_instr
*deref
= nir_deref_instr_create(ctx
->nir
, deref_type
);
1051 read_dest(ctx
, &deref
->dest
, &deref
->instr
, header
);
1053 nir_deref_instr
*parent
;
1055 switch (deref
->deref_type
) {
1056 case nir_deref_type_var
:
1057 if (header
.deref_var
.object_idx
)
1058 deref
->var
= read_lookup_object(ctx
, header
.deref_var
.object_idx
);
1060 deref
->var
= read_object(ctx
);
1062 deref
->type
= deref
->var
->type
;
1065 case nir_deref_type_struct
:
1066 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1067 parent
= nir_src_as_deref(deref
->parent
);
1068 deref
->strct
.index
= blob_read_uint32(ctx
->blob
);
1069 deref
->type
= glsl_get_struct_field(parent
->type
, deref
->strct
.index
);
1072 case nir_deref_type_array
:
1073 case nir_deref_type_ptr_as_array
:
1074 if (header
.deref
.packed_src_ssa_16bit
) {
1075 deref
->parent
.is_ssa
= true;
1076 deref
->parent
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
1077 deref
->arr
.index
.is_ssa
= true;
1078 deref
->arr
.index
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
1080 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1081 read_src(ctx
, &deref
->arr
.index
, &deref
->instr
);
1084 parent
= nir_src_as_deref(deref
->parent
);
1085 if (deref
->deref_type
== nir_deref_type_array
)
1086 deref
->type
= glsl_get_array_element(parent
->type
);
1088 deref
->type
= parent
->type
;
1091 case nir_deref_type_cast
:
1092 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1093 deref
->cast
.ptr_stride
= blob_read_uint32(ctx
->blob
);
1094 if (header
.deref
.cast_type_same_as_last
) {
1095 deref
->type
= ctx
->last_type
;
1097 deref
->type
= decode_type_from_blob(ctx
->blob
);
1098 ctx
->last_type
= deref
->type
;
1102 case nir_deref_type_array_wildcard
:
1103 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1104 parent
= nir_src_as_deref(deref
->parent
);
1105 deref
->type
= glsl_get_array_element(parent
->type
);
1109 unreachable("Invalid deref type");
1112 if (deref_type
== nir_deref_type_var
) {
1113 deref
->mode
= deref
->var
->data
.mode
;
1114 } else if (deref
->deref_type
== nir_deref_type_cast
) {
1115 deref
->mode
= header
.deref
.mode
;
1117 assert(deref
->parent
.is_ssa
);
1118 deref
->mode
= nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
)->mode
;
1125 write_intrinsic(write_ctx
*ctx
, const nir_intrinsic_instr
*intrin
)
1127 /* 9 bits for nir_intrinsic_op */
1128 STATIC_ASSERT(nir_num_intrinsics
<= 512);
1129 unsigned num_srcs
= nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
;
1130 unsigned num_indices
= nir_intrinsic_infos
[intrin
->intrinsic
].num_indices
;
1131 assert(intrin
->intrinsic
< 512);
1133 union packed_instr header
;
1136 header
.intrinsic
.instr_type
= intrin
->instr
.type
;
1137 header
.intrinsic
.intrinsic
= intrin
->intrinsic
;
1139 /* Analyze constant indices to decide how to encode them. */
1141 unsigned max_bits
= 0;
1142 for (unsigned i
= 0; i
< num_indices
; i
++) {
1143 unsigned max
= util_last_bit(intrin
->const_index
[i
]);
1144 max_bits
= MAX2(max_bits
, max
);
1147 if (max_bits
* num_indices
<= 9) {
1148 header
.intrinsic
.const_indices_encoding
= const_indices_9bit_all_combined
;
1150 /* Pack all const indices into 6 bits. */
1151 unsigned bit_size
= 9 / num_indices
;
1152 for (unsigned i
= 0; i
< num_indices
; i
++) {
1153 header
.intrinsic
.packed_const_indices
|=
1154 intrin
->const_index
[i
] << (i
* bit_size
);
1156 } else if (max_bits
<= 8)
1157 header
.intrinsic
.const_indices_encoding
= const_indices_8bit
;
1158 else if (max_bits
<= 16)
1159 header
.intrinsic
.const_indices_encoding
= const_indices_16bit
;
1161 header
.intrinsic
.const_indices_encoding
= const_indices_32bit
;
1164 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
)
1165 write_dest(ctx
, &intrin
->dest
, header
, intrin
->instr
.type
);
1167 blob_write_uint32(ctx
->blob
, header
.u32
);
1169 for (unsigned i
= 0; i
< num_srcs
; i
++)
1170 write_src(ctx
, &intrin
->src
[i
]);
1173 switch (header
.intrinsic
.const_indices_encoding
) {
1174 case const_indices_8bit
:
1175 for (unsigned i
= 0; i
< num_indices
; i
++)
1176 blob_write_uint8(ctx
->blob
, intrin
->const_index
[i
]);
1178 case const_indices_16bit
:
1179 for (unsigned i
= 0; i
< num_indices
; i
++)
1180 blob_write_uint16(ctx
->blob
, intrin
->const_index
[i
]);
1182 case const_indices_32bit
:
1183 for (unsigned i
= 0; i
< num_indices
; i
++)
1184 blob_write_uint32(ctx
->blob
, intrin
->const_index
[i
]);
1190 static nir_intrinsic_instr
*
1191 read_intrinsic(read_ctx
*ctx
, union packed_instr header
)
1193 nir_intrinsic_op op
= header
.intrinsic
.intrinsic
;
1194 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(ctx
->nir
, op
);
1196 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
1197 unsigned num_indices
= nir_intrinsic_infos
[op
].num_indices
;
1199 if (nir_intrinsic_infos
[op
].has_dest
)
1200 read_dest(ctx
, &intrin
->dest
, &intrin
->instr
, header
);
1202 for (unsigned i
= 0; i
< num_srcs
; i
++)
1203 read_src(ctx
, &intrin
->src
[i
], &intrin
->instr
);
1205 /* Vectorized instrinsics have num_components same as dst or src that has
1206 * 0 components in the info. Find it.
1208 if (nir_intrinsic_infos
[op
].has_dest
&&
1209 nir_intrinsic_infos
[op
].dest_components
== 0) {
1210 intrin
->num_components
= nir_dest_num_components(intrin
->dest
);
1212 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1213 if (nir_intrinsic_infos
[op
].src_components
[i
] == 0) {
1214 intrin
->num_components
= nir_src_num_components(intrin
->src
[i
]);
1221 switch (header
.intrinsic
.const_indices_encoding
) {
1222 case const_indices_9bit_all_combined
: {
1223 unsigned bit_size
= 9 / num_indices
;
1224 unsigned bit_mask
= u_bit_consecutive(0, bit_size
);
1225 for (unsigned i
= 0; i
< num_indices
; i
++) {
1226 intrin
->const_index
[i
] =
1227 (header
.intrinsic
.packed_const_indices
>> (i
* bit_size
)) &
1232 case const_indices_8bit
:
1233 for (unsigned i
= 0; i
< num_indices
; i
++)
1234 intrin
->const_index
[i
] = blob_read_uint8(ctx
->blob
);
1236 case const_indices_16bit
:
1237 for (unsigned i
= 0; i
< num_indices
; i
++)
1238 intrin
->const_index
[i
] = blob_read_uint16(ctx
->blob
);
1240 case const_indices_32bit
:
1241 for (unsigned i
= 0; i
< num_indices
; i
++)
1242 intrin
->const_index
[i
] = blob_read_uint32(ctx
->blob
);
1251 write_load_const(write_ctx
*ctx
, const nir_load_const_instr
*lc
)
1253 assert(lc
->def
.num_components
>= 1 && lc
->def
.num_components
<= 16);
1254 union packed_instr header
;
1257 header
.load_const
.instr_type
= lc
->instr
.type
;
1258 header
.load_const
.last_component
= lc
->def
.num_components
- 1;
1259 header
.load_const
.bit_size
= encode_bit_size_3bits(lc
->def
.bit_size
);
1260 header
.load_const
.packing
= load_const_full
;
1262 /* Try to pack 1-component constants into the 19 free bits in the header. */
1263 if (lc
->def
.num_components
== 1) {
1264 switch (lc
->def
.bit_size
) {
1266 if ((lc
->value
[0].u64
& 0x1fffffffffffull
) == 0) {
1267 /* packed_value contains high 19 bits, low bits are 0 */
1268 header
.load_const
.packing
= load_const_scalar_hi_19bits
;
1269 header
.load_const
.packed_value
= lc
->value
[0].u64
>> 45;
1270 } else if (((lc
->value
[0].i64
<< 45) >> 45) == lc
->value
[0].i64
) {
1271 /* packed_value contains low 19 bits, high bits are sign-extended */
1272 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1273 header
.load_const
.packed_value
= lc
->value
[0].u64
;
1278 if ((lc
->value
[0].u32
& 0x1fff) == 0) {
1279 header
.load_const
.packing
= load_const_scalar_hi_19bits
;
1280 header
.load_const
.packed_value
= lc
->value
[0].u32
>> 13;
1281 } else if (((lc
->value
[0].i32
<< 13) >> 13) == lc
->value
[0].i32
) {
1282 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1283 header
.load_const
.packed_value
= lc
->value
[0].u32
;
1288 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1289 header
.load_const
.packed_value
= lc
->value
[0].u16
;
1292 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1293 header
.load_const
.packed_value
= lc
->value
[0].u8
;
1296 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1297 header
.load_const
.packed_value
= lc
->value
[0].b
;
1300 unreachable("invalid bit_size");
1304 blob_write_uint32(ctx
->blob
, header
.u32
);
1306 if (header
.load_const
.packing
== load_const_full
) {
1307 switch (lc
->def
.bit_size
) {
1309 blob_write_bytes(ctx
->blob
, lc
->value
,
1310 sizeof(*lc
->value
) * lc
->def
.num_components
);
1314 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1315 blob_write_uint32(ctx
->blob
, lc
->value
[i
].u32
);
1319 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1320 blob_write_uint16(ctx
->blob
, lc
->value
[i
].u16
);
1324 assert(lc
->def
.bit_size
<= 8);
1325 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1326 blob_write_uint8(ctx
->blob
, lc
->value
[i
].u8
);
1331 write_add_object(ctx
, &lc
->def
);
1334 static nir_load_const_instr
*
1335 read_load_const(read_ctx
*ctx
, union packed_instr header
)
1337 nir_load_const_instr
*lc
=
1338 nir_load_const_instr_create(ctx
->nir
, header
.load_const
.last_component
+ 1,
1339 decode_bit_size_3bits(header
.load_const
.bit_size
));
1341 switch (header
.load_const
.packing
) {
1342 case load_const_scalar_hi_19bits
:
1343 switch (lc
->def
.bit_size
) {
1345 lc
->value
[0].u64
= (uint64_t)header
.load_const
.packed_value
<< 45;
1348 lc
->value
[0].u32
= (uint64_t)header
.load_const
.packed_value
<< 13;
1351 unreachable("invalid bit_size");
1355 case load_const_scalar_lo_19bits_sext
:
1356 switch (lc
->def
.bit_size
) {
1358 lc
->value
[0].i64
= ((int64_t)header
.load_const
.packed_value
<< 45) >> 45;
1361 lc
->value
[0].i32
= ((int32_t)header
.load_const
.packed_value
<< 13) >> 13;
1364 lc
->value
[0].u16
= header
.load_const
.packed_value
;
1367 lc
->value
[0].u8
= header
.load_const
.packed_value
;
1370 lc
->value
[0].b
= header
.load_const
.packed_value
;
1373 unreachable("invalid bit_size");
1377 case load_const_full
:
1378 switch (lc
->def
.bit_size
) {
1380 blob_copy_bytes(ctx
->blob
, lc
->value
, sizeof(*lc
->value
) * lc
->def
.num_components
);
1384 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1385 lc
->value
[i
].u32
= blob_read_uint32(ctx
->blob
);
1389 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1390 lc
->value
[i
].u16
= blob_read_uint16(ctx
->blob
);
1394 assert(lc
->def
.bit_size
<= 8);
1395 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1396 lc
->value
[i
].u8
= blob_read_uint8(ctx
->blob
);
1402 read_add_object(ctx
, &lc
->def
);
1407 write_ssa_undef(write_ctx
*ctx
, const nir_ssa_undef_instr
*undef
)
1409 assert(undef
->def
.num_components
>= 1 && undef
->def
.num_components
<= 16);
1411 union packed_instr header
;
1414 header
.undef
.instr_type
= undef
->instr
.type
;
1415 header
.undef
.last_component
= undef
->def
.num_components
- 1;
1416 header
.undef
.bit_size
= encode_bit_size_3bits(undef
->def
.bit_size
);
1418 blob_write_uint32(ctx
->blob
, header
.u32
);
1419 write_add_object(ctx
, &undef
->def
);
1422 static nir_ssa_undef_instr
*
1423 read_ssa_undef(read_ctx
*ctx
, union packed_instr header
)
1425 nir_ssa_undef_instr
*undef
=
1426 nir_ssa_undef_instr_create(ctx
->nir
, header
.undef
.last_component
+ 1,
1427 decode_bit_size_3bits(header
.undef
.bit_size
));
1429 read_add_object(ctx
, &undef
->def
);
1433 union packed_tex_data
{
1436 enum glsl_sampler_dim sampler_dim
:4;
1437 nir_alu_type dest_type
:8;
1438 unsigned coord_components
:3;
1439 unsigned is_array
:1;
1440 unsigned is_shadow
:1;
1441 unsigned is_new_style_shadow
:1;
1442 unsigned component
:2;
1443 unsigned texture_non_uniform
:1;
1444 unsigned sampler_non_uniform
:1;
1445 unsigned unused
:8; /* Mark unused for valgrind. */
1450 write_tex(write_ctx
*ctx
, const nir_tex_instr
*tex
)
1452 assert(tex
->num_srcs
< 16);
1453 assert(tex
->op
< 16);
1454 assert(tex
->texture_array_size
< 1024);
1456 union packed_instr header
;
1459 header
.tex
.instr_type
= tex
->instr
.type
;
1460 header
.tex
.num_srcs
= tex
->num_srcs
;
1461 header
.tex
.op
= tex
->op
;
1462 header
.tex
.texture_array_size
= tex
->texture_array_size
;
1464 write_dest(ctx
, &tex
->dest
, header
, tex
->instr
.type
);
1466 blob_write_uint32(ctx
->blob
, tex
->texture_index
);
1467 blob_write_uint32(ctx
->blob
, tex
->sampler_index
);
1468 if (tex
->op
== nir_texop_tg4
)
1469 blob_write_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
1471 STATIC_ASSERT(sizeof(union packed_tex_data
) == sizeof(uint32_t));
1472 union packed_tex_data packed
= {
1473 .u
.sampler_dim
= tex
->sampler_dim
,
1474 .u
.dest_type
= tex
->dest_type
,
1475 .u
.coord_components
= tex
->coord_components
,
1476 .u
.is_array
= tex
->is_array
,
1477 .u
.is_shadow
= tex
->is_shadow
,
1478 .u
.is_new_style_shadow
= tex
->is_new_style_shadow
,
1479 .u
.component
= tex
->component
,
1480 .u
.texture_non_uniform
= tex
->texture_non_uniform
,
1481 .u
.sampler_non_uniform
= tex
->sampler_non_uniform
,
1483 blob_write_uint32(ctx
->blob
, packed
.u32
);
1485 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1486 union packed_src src
;
1488 src
.tex
.src_type
= tex
->src
[i
].src_type
;
1489 write_src_full(ctx
, &tex
->src
[i
].src
, src
);
1493 static nir_tex_instr
*
1494 read_tex(read_ctx
*ctx
, union packed_instr header
)
1496 nir_tex_instr
*tex
= nir_tex_instr_create(ctx
->nir
, header
.tex
.num_srcs
);
1498 read_dest(ctx
, &tex
->dest
, &tex
->instr
, header
);
1500 tex
->op
= header
.tex
.op
;
1501 tex
->texture_index
= blob_read_uint32(ctx
->blob
);
1502 tex
->texture_array_size
= header
.tex
.texture_array_size
;
1503 tex
->sampler_index
= blob_read_uint32(ctx
->blob
);
1504 if (tex
->op
== nir_texop_tg4
)
1505 blob_copy_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
1507 union packed_tex_data packed
;
1508 packed
.u32
= blob_read_uint32(ctx
->blob
);
1509 tex
->sampler_dim
= packed
.u
.sampler_dim
;
1510 tex
->dest_type
= packed
.u
.dest_type
;
1511 tex
->coord_components
= packed
.u
.coord_components
;
1512 tex
->is_array
= packed
.u
.is_array
;
1513 tex
->is_shadow
= packed
.u
.is_shadow
;
1514 tex
->is_new_style_shadow
= packed
.u
.is_new_style_shadow
;
1515 tex
->component
= packed
.u
.component
;
1516 tex
->texture_non_uniform
= packed
.u
.texture_non_uniform
;
1517 tex
->sampler_non_uniform
= packed
.u
.sampler_non_uniform
;
1519 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1520 union packed_src src
= read_src(ctx
, &tex
->src
[i
].src
, &tex
->instr
);
1521 tex
->src
[i
].src_type
= src
.tex
.src_type
;
1528 write_phi(write_ctx
*ctx
, const nir_phi_instr
*phi
)
1530 union packed_instr header
;
1533 header
.phi
.instr_type
= phi
->instr
.type
;
1534 header
.phi
.num_srcs
= exec_list_length(&phi
->srcs
);
1536 /* Phi nodes are special, since they may reference SSA definitions and
1537 * basic blocks that don't exist yet. We leave two empty uint32_t's here,
1538 * and then store enough information so that a later fixup pass can fill
1539 * them in correctly.
1541 write_dest(ctx
, &phi
->dest
, header
, phi
->instr
.type
);
1543 nir_foreach_phi_src(src
, phi
) {
1544 assert(src
->src
.is_ssa
);
1545 size_t blob_offset
= blob_reserve_uint32(ctx
->blob
);
1546 ASSERTED
size_t blob_offset2
= blob_reserve_uint32(ctx
->blob
);
1547 assert(blob_offset
+ sizeof(uint32_t) == blob_offset2
);
1548 write_phi_fixup fixup
= {
1549 .blob_offset
= blob_offset
,
1550 .src
= src
->src
.ssa
,
1553 util_dynarray_append(&ctx
->phi_fixups
, write_phi_fixup
, fixup
);
1558 write_fixup_phis(write_ctx
*ctx
)
1560 util_dynarray_foreach(&ctx
->phi_fixups
, write_phi_fixup
, fixup
) {
1561 uint32_t *blob_ptr
= (uint32_t *)(ctx
->blob
->data
+ fixup
->blob_offset
);
1562 blob_ptr
[0] = write_lookup_object(ctx
, fixup
->src
);
1563 blob_ptr
[1] = write_lookup_object(ctx
, fixup
->block
);
1566 util_dynarray_clear(&ctx
->phi_fixups
);
1569 static nir_phi_instr
*
1570 read_phi(read_ctx
*ctx
, nir_block
*blk
, union packed_instr header
)
1572 nir_phi_instr
*phi
= nir_phi_instr_create(ctx
->nir
);
1574 read_dest(ctx
, &phi
->dest
, &phi
->instr
, header
);
1576 /* For similar reasons as before, we just store the index directly into the
1577 * pointer, and let a later pass resolve the phi sources.
1579 * In order to ensure that the copied sources (which are just the indices
1580 * from the blob for now) don't get inserted into the old shader's use-def
1581 * lists, we have to add the phi instruction *before* we set up its
1584 nir_instr_insert_after_block(blk
, &phi
->instr
);
1586 for (unsigned i
= 0; i
< header
.phi
.num_srcs
; i
++) {
1587 nir_phi_src
*src
= ralloc(phi
, nir_phi_src
);
1589 src
->src
.is_ssa
= true;
1590 src
->src
.ssa
= (nir_ssa_def
*)(uintptr_t) blob_read_uint32(ctx
->blob
);
1591 src
->pred
= (nir_block
*)(uintptr_t) blob_read_uint32(ctx
->blob
);
1593 /* Since we're not letting nir_insert_instr handle use/def stuff for us,
1594 * we have to set the parent_instr manually. It doesn't really matter
1595 * when we do it, so we might as well do it here.
1597 src
->src
.parent_instr
= &phi
->instr
;
1599 /* Stash it in the list of phi sources. We'll walk this list and fix up
1600 * sources at the very end of read_function_impl.
1602 list_add(&src
->src
.use_link
, &ctx
->phi_srcs
);
1604 exec_list_push_tail(&phi
->srcs
, &src
->node
);
1611 read_fixup_phis(read_ctx
*ctx
)
1613 list_for_each_entry_safe(nir_phi_src
, src
, &ctx
->phi_srcs
, src
.use_link
) {
1614 src
->pred
= read_lookup_object(ctx
, (uintptr_t)src
->pred
);
1615 src
->src
.ssa
= read_lookup_object(ctx
, (uintptr_t)src
->src
.ssa
);
1617 /* Remove from this list */
1618 list_del(&src
->src
.use_link
);
1620 list_addtail(&src
->src
.use_link
, &src
->src
.ssa
->uses
);
1622 assert(list_is_empty(&ctx
->phi_srcs
));
1626 write_jump(write_ctx
*ctx
, const nir_jump_instr
*jmp
)
1628 assert(jmp
->type
< 4);
1630 union packed_instr header
;
1633 header
.jump
.instr_type
= jmp
->instr
.type
;
1634 header
.jump
.type
= jmp
->type
;
1636 blob_write_uint32(ctx
->blob
, header
.u32
);
1639 static nir_jump_instr
*
1640 read_jump(read_ctx
*ctx
, union packed_instr header
)
1642 nir_jump_instr
*jmp
= nir_jump_instr_create(ctx
->nir
, header
.jump
.type
);
1647 write_call(write_ctx
*ctx
, const nir_call_instr
*call
)
1649 blob_write_uint32(ctx
->blob
, write_lookup_object(ctx
, call
->callee
));
1651 for (unsigned i
= 0; i
< call
->num_params
; i
++)
1652 write_src(ctx
, &call
->params
[i
]);
1655 static nir_call_instr
*
1656 read_call(read_ctx
*ctx
)
1658 nir_function
*callee
= read_object(ctx
);
1659 nir_call_instr
*call
= nir_call_instr_create(ctx
->nir
, callee
);
1661 for (unsigned i
= 0; i
< call
->num_params
; i
++)
1662 read_src(ctx
, &call
->params
[i
], call
);
1668 write_instr(write_ctx
*ctx
, const nir_instr
*instr
)
1670 /* We have only 4 bits for the instruction type. */
1671 assert(instr
->type
< 16);
1673 switch (instr
->type
) {
1674 case nir_instr_type_alu
:
1675 write_alu(ctx
, nir_instr_as_alu(instr
));
1677 case nir_instr_type_deref
:
1678 write_deref(ctx
, nir_instr_as_deref(instr
));
1680 case nir_instr_type_intrinsic
:
1681 write_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1683 case nir_instr_type_load_const
:
1684 write_load_const(ctx
, nir_instr_as_load_const(instr
));
1686 case nir_instr_type_ssa_undef
:
1687 write_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
1689 case nir_instr_type_tex
:
1690 write_tex(ctx
, nir_instr_as_tex(instr
));
1692 case nir_instr_type_phi
:
1693 write_phi(ctx
, nir_instr_as_phi(instr
));
1695 case nir_instr_type_jump
:
1696 write_jump(ctx
, nir_instr_as_jump(instr
));
1698 case nir_instr_type_call
:
1699 blob_write_uint32(ctx
->blob
, instr
->type
);
1700 write_call(ctx
, nir_instr_as_call(instr
));
1702 case nir_instr_type_parallel_copy
:
1703 unreachable("Cannot write parallel copies");
1705 unreachable("bad instr type");
1709 /* Return the number of instructions read. */
1711 read_instr(read_ctx
*ctx
, nir_block
*block
)
1713 STATIC_ASSERT(sizeof(union packed_instr
) == 4);
1714 union packed_instr header
;
1715 header
.u32
= blob_read_uint32(ctx
->blob
);
1718 switch (header
.any
.instr_type
) {
1719 case nir_instr_type_alu
:
1720 for (unsigned i
= 0; i
<= header
.alu
.num_followup_alu_sharing_header
; i
++)
1721 nir_instr_insert_after_block(block
, &read_alu(ctx
, header
)->instr
);
1722 return header
.alu
.num_followup_alu_sharing_header
+ 1;
1723 case nir_instr_type_deref
:
1724 instr
= &read_deref(ctx
, header
)->instr
;
1726 case nir_instr_type_intrinsic
:
1727 instr
= &read_intrinsic(ctx
, header
)->instr
;
1729 case nir_instr_type_load_const
:
1730 instr
= &read_load_const(ctx
, header
)->instr
;
1732 case nir_instr_type_ssa_undef
:
1733 instr
= &read_ssa_undef(ctx
, header
)->instr
;
1735 case nir_instr_type_tex
:
1736 instr
= &read_tex(ctx
, header
)->instr
;
1738 case nir_instr_type_phi
:
1739 /* Phi instructions are a bit of a special case when reading because we
1740 * don't want inserting the instruction to automatically handle use/defs
1741 * for us. Instead, we need to wait until all the blocks/instructions
1742 * are read so that we can set their sources up.
1744 read_phi(ctx
, block
, header
);
1746 case nir_instr_type_jump
:
1747 instr
= &read_jump(ctx
, header
)->instr
;
1749 case nir_instr_type_call
:
1750 instr
= &read_call(ctx
)->instr
;
1752 case nir_instr_type_parallel_copy
:
1753 unreachable("Cannot read parallel copies");
1755 unreachable("bad instr type");
1758 nir_instr_insert_after_block(block
, instr
);
1763 write_block(write_ctx
*ctx
, const nir_block
*block
)
1765 write_add_object(ctx
, block
);
1766 blob_write_uint32(ctx
->blob
, exec_list_length(&block
->instr_list
));
1768 ctx
->last_instr_type
= ~0;
1769 ctx
->last_alu_header_offset
= 0;
1771 nir_foreach_instr(instr
, block
) {
1772 write_instr(ctx
, instr
);
1773 ctx
->last_instr_type
= instr
->type
;
1778 read_block(read_ctx
*ctx
, struct exec_list
*cf_list
)
1780 /* Don't actually create a new block. Just use the one from the tail of
1781 * the list. NIR guarantees that the tail of the list is a block and that
1782 * no two blocks are side-by-side in the IR; It should be empty.
1785 exec_node_data(nir_block
, exec_list_get_tail(cf_list
), cf_node
.node
);
1787 read_add_object(ctx
, block
);
1788 unsigned num_instrs
= blob_read_uint32(ctx
->blob
);
1789 for (unsigned i
= 0; i
< num_instrs
;) {
1790 i
+= read_instr(ctx
, block
);
1795 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
);
1798 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
);
1801 write_if(write_ctx
*ctx
, nir_if
*nif
)
1803 write_src(ctx
, &nif
->condition
);
1805 write_cf_list(ctx
, &nif
->then_list
);
1806 write_cf_list(ctx
, &nif
->else_list
);
1810 read_if(read_ctx
*ctx
, struct exec_list
*cf_list
)
1812 nir_if
*nif
= nir_if_create(ctx
->nir
);
1814 read_src(ctx
, &nif
->condition
, nif
);
1816 nir_cf_node_insert_end(cf_list
, &nif
->cf_node
);
1818 read_cf_list(ctx
, &nif
->then_list
);
1819 read_cf_list(ctx
, &nif
->else_list
);
1823 write_loop(write_ctx
*ctx
, nir_loop
*loop
)
1825 write_cf_list(ctx
, &loop
->body
);
1829 read_loop(read_ctx
*ctx
, struct exec_list
*cf_list
)
1831 nir_loop
*loop
= nir_loop_create(ctx
->nir
);
1833 nir_cf_node_insert_end(cf_list
, &loop
->cf_node
);
1835 read_cf_list(ctx
, &loop
->body
);
1839 write_cf_node(write_ctx
*ctx
, nir_cf_node
*cf
)
1841 blob_write_uint32(ctx
->blob
, cf
->type
);
1844 case nir_cf_node_block
:
1845 write_block(ctx
, nir_cf_node_as_block(cf
));
1847 case nir_cf_node_if
:
1848 write_if(ctx
, nir_cf_node_as_if(cf
));
1850 case nir_cf_node_loop
:
1851 write_loop(ctx
, nir_cf_node_as_loop(cf
));
1854 unreachable("bad cf type");
1859 read_cf_node(read_ctx
*ctx
, struct exec_list
*list
)
1861 nir_cf_node_type type
= blob_read_uint32(ctx
->blob
);
1864 case nir_cf_node_block
:
1865 read_block(ctx
, list
);
1867 case nir_cf_node_if
:
1870 case nir_cf_node_loop
:
1871 read_loop(ctx
, list
);
1874 unreachable("bad cf type");
1879 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
)
1881 blob_write_uint32(ctx
->blob
, exec_list_length(cf_list
));
1882 foreach_list_typed(nir_cf_node
, cf
, node
, cf_list
) {
1883 write_cf_node(ctx
, cf
);
1888 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
)
1890 uint32_t num_cf_nodes
= blob_read_uint32(ctx
->blob
);
1891 for (unsigned i
= 0; i
< num_cf_nodes
; i
++)
1892 read_cf_node(ctx
, cf_list
);
1896 write_function_impl(write_ctx
*ctx
, const nir_function_impl
*fi
)
1898 write_var_list(ctx
, &fi
->locals
);
1899 write_reg_list(ctx
, &fi
->registers
);
1900 blob_write_uint32(ctx
->blob
, fi
->reg_alloc
);
1902 write_cf_list(ctx
, &fi
->body
);
1903 write_fixup_phis(ctx
);
1906 static nir_function_impl
*
1907 read_function_impl(read_ctx
*ctx
, nir_function
*fxn
)
1909 nir_function_impl
*fi
= nir_function_impl_create_bare(ctx
->nir
);
1912 read_var_list(ctx
, &fi
->locals
);
1913 read_reg_list(ctx
, &fi
->registers
);
1914 fi
->reg_alloc
= blob_read_uint32(ctx
->blob
);
1916 read_cf_list(ctx
, &fi
->body
);
1917 read_fixup_phis(ctx
);
1919 fi
->valid_metadata
= 0;
1925 write_function(write_ctx
*ctx
, const nir_function
*fxn
)
1927 uint32_t flags
= fxn
->is_entrypoint
;
1932 blob_write_uint32(ctx
->blob
, flags
);
1934 blob_write_string(ctx
->blob
, fxn
->name
);
1936 write_add_object(ctx
, fxn
);
1938 blob_write_uint32(ctx
->blob
, fxn
->num_params
);
1939 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1941 ((uint32_t)fxn
->params
[i
].num_components
) |
1942 ((uint32_t)fxn
->params
[i
].bit_size
) << 8;
1943 blob_write_uint32(ctx
->blob
, val
);
1946 /* At first glance, it looks like we should write the function_impl here.
1947 * However, call instructions need to be able to reference at least the
1948 * function and those will get processed as we write the function_impls.
1949 * We stop here and write function_impls as a second pass.
1954 read_function(read_ctx
*ctx
)
1956 uint32_t flags
= blob_read_uint32(ctx
->blob
);
1957 bool has_name
= flags
& 0x2;
1958 char *name
= has_name
? blob_read_string(ctx
->blob
) : NULL
;
1960 nir_function
*fxn
= nir_function_create(ctx
->nir
, name
);
1962 read_add_object(ctx
, fxn
);
1964 fxn
->num_params
= blob_read_uint32(ctx
->blob
);
1965 fxn
->params
= ralloc_array(fxn
, nir_parameter
, fxn
->num_params
);
1966 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1967 uint32_t val
= blob_read_uint32(ctx
->blob
);
1968 fxn
->params
[i
].num_components
= val
& 0xff;
1969 fxn
->params
[i
].bit_size
= (val
>> 8) & 0xff;
1972 fxn
->is_entrypoint
= flags
& 0x1;
1974 fxn
->impl
= NIR_SERIALIZE_FUNC_HAS_IMPL
;
1978 * Serialize NIR into a binary blob.
1980 * \param strip Don't serialize information only useful for debugging,
1981 * such as variable names, making cache hits from similar
1982 * shaders more likely.
1985 nir_serialize(struct blob
*blob
, const nir_shader
*nir
, bool strip
)
1987 write_ctx ctx
= {0};
1988 ctx
.remap_table
= _mesa_pointer_hash_table_create(NULL
);
1992 util_dynarray_init(&ctx
.phi_fixups
, NULL
);
1994 size_t idx_size_offset
= blob_reserve_uint32(blob
);
1996 struct shader_info info
= nir
->info
;
1997 uint32_t strings
= 0;
1998 if (!strip
&& info
.name
)
2000 if (!strip
&& info
.label
)
2002 blob_write_uint32(blob
, strings
);
2003 if (!strip
&& info
.name
)
2004 blob_write_string(blob
, info
.name
);
2005 if (!strip
&& info
.label
)
2006 blob_write_string(blob
, info
.label
);
2007 info
.name
= info
.label
= NULL
;
2008 blob_write_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
2010 write_var_list(&ctx
, &nir
->uniforms
);
2011 write_var_list(&ctx
, &nir
->inputs
);
2012 write_var_list(&ctx
, &nir
->outputs
);
2013 write_var_list(&ctx
, &nir
->shared
);
2014 write_var_list(&ctx
, &nir
->globals
);
2015 write_var_list(&ctx
, &nir
->system_values
);
2017 blob_write_uint32(blob
, nir
->num_inputs
);
2018 blob_write_uint32(blob
, nir
->num_uniforms
);
2019 blob_write_uint32(blob
, nir
->num_outputs
);
2020 blob_write_uint32(blob
, nir
->num_shared
);
2021 blob_write_uint32(blob
, nir
->scratch_size
);
2023 blob_write_uint32(blob
, exec_list_length(&nir
->functions
));
2024 nir_foreach_function(fxn
, nir
) {
2025 write_function(&ctx
, fxn
);
2028 nir_foreach_function(fxn
, nir
) {
2030 write_function_impl(&ctx
, fxn
->impl
);
2033 blob_write_uint32(blob
, nir
->constant_data_size
);
2034 if (nir
->constant_data_size
> 0)
2035 blob_write_bytes(blob
, nir
->constant_data
, nir
->constant_data_size
);
2037 *(uint32_t *)(blob
->data
+ idx_size_offset
) = ctx
.next_idx
;
2039 _mesa_hash_table_destroy(ctx
.remap_table
, NULL
);
2040 util_dynarray_fini(&ctx
.phi_fixups
);
2044 nir_deserialize(void *mem_ctx
,
2045 const struct nir_shader_compiler_options
*options
,
2046 struct blob_reader
*blob
)
2050 list_inithead(&ctx
.phi_srcs
);
2051 ctx
.idx_table_len
= blob_read_uint32(blob
);
2052 ctx
.idx_table
= calloc(ctx
.idx_table_len
, sizeof(uintptr_t));
2054 uint32_t strings
= blob_read_uint32(blob
);
2055 char *name
= (strings
& 0x1) ? blob_read_string(blob
) : NULL
;
2056 char *label
= (strings
& 0x2) ? blob_read_string(blob
) : NULL
;
2058 struct shader_info info
;
2059 blob_copy_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
2061 ctx
.nir
= nir_shader_create(mem_ctx
, info
.stage
, options
, NULL
);
2063 info
.name
= name
? ralloc_strdup(ctx
.nir
, name
) : NULL
;
2064 info
.label
= label
? ralloc_strdup(ctx
.nir
, label
) : NULL
;
2066 ctx
.nir
->info
= info
;
2068 read_var_list(&ctx
, &ctx
.nir
->uniforms
);
2069 read_var_list(&ctx
, &ctx
.nir
->inputs
);
2070 read_var_list(&ctx
, &ctx
.nir
->outputs
);
2071 read_var_list(&ctx
, &ctx
.nir
->shared
);
2072 read_var_list(&ctx
, &ctx
.nir
->globals
);
2073 read_var_list(&ctx
, &ctx
.nir
->system_values
);
2075 ctx
.nir
->num_inputs
= blob_read_uint32(blob
);
2076 ctx
.nir
->num_uniforms
= blob_read_uint32(blob
);
2077 ctx
.nir
->num_outputs
= blob_read_uint32(blob
);
2078 ctx
.nir
->num_shared
= blob_read_uint32(blob
);
2079 ctx
.nir
->scratch_size
= blob_read_uint32(blob
);
2081 unsigned num_functions
= blob_read_uint32(blob
);
2082 for (unsigned i
= 0; i
< num_functions
; i
++)
2083 read_function(&ctx
);
2085 nir_foreach_function(fxn
, ctx
.nir
) {
2086 if (fxn
->impl
== NIR_SERIALIZE_FUNC_HAS_IMPL
)
2087 fxn
->impl
= read_function_impl(&ctx
, fxn
);
2090 ctx
.nir
->constant_data_size
= blob_read_uint32(blob
);
2091 if (ctx
.nir
->constant_data_size
> 0) {
2092 ctx
.nir
->constant_data
=
2093 ralloc_size(ctx
.nir
, ctx
.nir
->constant_data_size
);
2094 blob_copy_bytes(blob
, ctx
.nir
->constant_data
,
2095 ctx
.nir
->constant_data_size
);
2098 free(ctx
.idx_table
);
2104 nir_shader_serialize_deserialize(nir_shader
*shader
)
2106 const struct nir_shader_compiler_options
*options
= shader
->options
;
2110 nir_serialize(&writer
, shader
, false);
2112 /* Delete all of dest's ralloc children but leave dest alone */
2113 void *dead_ctx
= ralloc_context(NULL
);
2114 ralloc_adopt(dead_ctx
, shader
);
2115 ralloc_free(dead_ctx
);
2117 dead_ctx
= ralloc_context(NULL
);
2119 struct blob_reader reader
;
2120 blob_reader_init(&reader
, writer
.data
, writer
.size
);
2121 nir_shader
*copy
= nir_deserialize(dead_ctx
, options
, &reader
);
2123 blob_finish(&writer
);
2125 nir_shader_replace(shader
, copy
);
2126 ralloc_free(dead_ctx
);