2 * Copyright (c) 2015 Intel Corporation
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
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_buffer_access.cpp
27 * Helper for IR lowering pass to replace dereferences of buffer object based
28 * shader variables with intrinsic function calls.
30 * This helper is used by lowering passes for UBOs, SSBOs and compute shader
34 #include "lower_buffer_access.h"
35 #include "ir_builder.h"
36 #include "main/macros.h"
37 #include "util/list.h"
38 #include "glsl_parser_extras.h"
41 using namespace ir_builder
;
43 namespace lower_buffer_access
{
46 writemask_for_size(unsigned n
)
48 return ((1 << n
) - 1);
52 * Takes a deref and recursively calls itself to break the deref down to the
53 * point that the reads or writes generated are contiguous scalars or vectors.
56 lower_buffer_access::emit_access(void *mem_ctx
,
58 ir_dereference
*deref
,
59 ir_variable
*base_offset
,
60 unsigned int deref_offset
,
62 const glsl_type
*matrix_type
,
63 enum glsl_interface_packing packing
,
64 unsigned int write_mask
)
66 if (deref
->type
->is_record()) {
67 unsigned int field_offset
= 0;
69 for (unsigned i
= 0; i
< deref
->type
->length
; i
++) {
70 const struct glsl_struct_field
*field
=
71 &deref
->type
->fields
.structure
[i
];
72 ir_dereference
*field_deref
=
73 new(mem_ctx
) ir_dereference_record(deref
->clone(mem_ctx
, NULL
),
77 glsl_align(field_offset
,
78 field
->type
->std140_base_alignment(row_major
));
80 emit_access(mem_ctx
, is_write
, field_deref
, base_offset
,
81 deref_offset
+ field_offset
,
82 row_major
, NULL
, packing
,
83 writemask_for_size(field_deref
->type
->vector_elements
));
85 field_offset
+= field
->type
->std140_size(row_major
);
90 if (deref
->type
->is_array()) {
91 unsigned array_stride
= packing
== GLSL_INTERFACE_PACKING_STD430
?
92 deref
->type
->fields
.array
->std430_array_stride(row_major
) :
93 glsl_align(deref
->type
->fields
.array
->std140_size(row_major
), 16);
95 for (unsigned i
= 0; i
< deref
->type
->length
; i
++) {
96 ir_constant
*element
= new(mem_ctx
) ir_constant(i
);
97 ir_dereference
*element_deref
=
98 new(mem_ctx
) ir_dereference_array(deref
->clone(mem_ctx
, NULL
),
100 emit_access(mem_ctx
, is_write
, element_deref
, base_offset
,
101 deref_offset
+ i
* array_stride
,
102 row_major
, NULL
, packing
,
103 writemask_for_size(element_deref
->type
->vector_elements
));
108 if (deref
->type
->is_matrix()) {
109 for (unsigned i
= 0; i
< deref
->type
->matrix_columns
; i
++) {
110 ir_constant
*col
= new(mem_ctx
) ir_constant(i
);
111 ir_dereference
*col_deref
=
112 new(mem_ctx
) ir_dereference_array(deref
->clone(mem_ctx
, NULL
), col
);
115 /* For a row-major matrix, the next column starts at the next
118 int size_mul
= deref
->type
->is_64bit() ? 8 : 4;
119 emit_access(mem_ctx
, is_write
, col_deref
, base_offset
,
120 deref_offset
+ i
* size_mul
,
121 row_major
, deref
->type
, packing
,
122 writemask_for_size(col_deref
->type
->vector_elements
));
126 /* std430 doesn't round up vec2 size to a vec4 size */
127 if (packing
== GLSL_INTERFACE_PACKING_STD430
&&
128 deref
->type
->vector_elements
== 2 &&
129 !deref
->type
->is_64bit()) {
132 /* std140 always rounds the stride of arrays (and matrices) to a
133 * vec4, so matrices are always 16 between columns/rows. With
134 * doubles, they will be 32 apart when there are more than 2 rows.
136 * For both std140 and std430, if the member is a
137 * three-'component vector with components consuming N basic
138 * machine units, the base alignment is 4N. For vec4, base
141 size_mul
= (deref
->type
->is_64bit() &&
142 deref
->type
->vector_elements
> 2) ? 32 : 16;
145 emit_access(mem_ctx
, is_write
, col_deref
, base_offset
,
146 deref_offset
+ i
* size_mul
,
147 row_major
, deref
->type
, packing
,
148 writemask_for_size(col_deref
->type
->vector_elements
));
154 assert(deref
->type
->is_scalar() || deref
->type
->is_vector());
158 add(base_offset
, new(mem_ctx
) ir_constant(deref_offset
));
160 is_write
? write_mask
: (1 << deref
->type
->vector_elements
) - 1;
161 insert_buffer_access(mem_ctx
, deref
, deref
->type
, offset
, mask
, -1);
163 /* We're dereffing a column out of a row-major matrix, so we
164 * gather the vector from each stored row.
166 assert(deref
->type
->is_float() || deref
->type
->is_double());
167 assert(matrix_type
!= NULL
);
169 const unsigned matrix_stride
=
170 link_calculate_matrix_stride(matrix_type
, row_major
, packing
);
172 const glsl_type
*deref_type
= deref
->type
->is_float() ?
173 glsl_type::float_type
: glsl_type::double_type
;
175 for (unsigned i
= 0; i
< deref
->type
->vector_elements
; i
++) {
176 ir_rvalue
*chan_offset
=
178 new(mem_ctx
) ir_constant(deref_offset
+ i
* matrix_stride
));
179 if (!is_write
|| ((1U << i
) & write_mask
))
180 insert_buffer_access(mem_ctx
, deref
, deref_type
, chan_offset
,
187 * Determine if a thing being dereferenced is row-major
189 * There is some trickery here.
191 * If the thing being dereferenced is a member of uniform block \b without an
192 * instance name, then the name of the \c ir_variable is the field name of an
193 * interface type. If this field is row-major, then the thing referenced is
196 * If the thing being dereferenced is a member of uniform block \b with an
197 * instance name, then the last dereference in the tree will be an
198 * \c ir_dereference_record. If that record field is row-major, then the
199 * thing referenced is row-major.
202 lower_buffer_access::is_dereferenced_thing_row_major(const ir_rvalue
*deref
)
205 const ir_rvalue
*ir
= deref
;
208 matrix
= matrix
|| ir
->type
->without_array()->is_matrix();
210 switch (ir
->ir_type
) {
211 case ir_type_dereference_array
: {
212 const ir_dereference_array
*const array_deref
=
213 (const ir_dereference_array
*) ir
;
215 ir
= array_deref
->array
;
219 case ir_type_dereference_record
: {
220 const ir_dereference_record
*const record_deref
=
221 (const ir_dereference_record
*) ir
;
223 ir
= record_deref
->record
;
225 const int idx
= record_deref
->field_idx
;
228 const enum glsl_matrix_layout matrix_layout
=
229 glsl_matrix_layout(ir
->type
->fields
.structure
[idx
].matrix_layout
);
231 switch (matrix_layout
) {
232 case GLSL_MATRIX_LAYOUT_INHERITED
:
234 case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR
:
236 case GLSL_MATRIX_LAYOUT_ROW_MAJOR
:
237 return matrix
|| deref
->type
->without_array()->is_record();
243 case ir_type_dereference_variable
: {
244 const ir_dereference_variable
*const var_deref
=
245 (const ir_dereference_variable
*) ir
;
247 const enum glsl_matrix_layout matrix_layout
=
248 glsl_matrix_layout(var_deref
->var
->data
.matrix_layout
);
250 switch (matrix_layout
) {
251 case GLSL_MATRIX_LAYOUT_INHERITED
: {
252 /* For interface block matrix variables we handle inherited
253 * layouts at HIR generation time, but we don't do that for shared
254 * variables, which are always column-major
256 MAYBE_UNUSED ir_variable
*var
= deref
->variable_referenced();
257 assert((var
->is_in_buffer_block() && !matrix
) ||
258 var
->data
.mode
== ir_var_shader_shared
);
261 case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR
:
263 case GLSL_MATRIX_LAYOUT_ROW_MAJOR
:
264 return matrix
|| deref
->type
->without_array()->is_record();
267 unreachable("invalid matrix layout");
276 /* The tree must have ended with a dereference that wasn't an
277 * ir_dereference_variable. That is invalid, and it should be impossible.
279 unreachable("invalid dereference tree");
284 * This function initializes various values that will be used later by
285 * emit_access when actually emitting loads or stores.
287 * Note: const_offset is an input as well as an output, clients must
288 * initialize it to the offset of the variable in the underlying block, and
289 * this function will adjust it by adding the constant offset of the member
290 * being accessed into that variable.
293 lower_buffer_access::setup_buffer_access(void *mem_ctx
,
296 unsigned *const_offset
,
298 const glsl_type
**matrix_type
,
299 const glsl_struct_field
**struct_field
,
300 enum glsl_interface_packing packing
)
302 *offset
= new(mem_ctx
) ir_constant(0u);
303 *row_major
= is_dereferenced_thing_row_major(deref
);
306 /* Calculate the offset to the start of the region of the UBO
307 * dereferenced by *rvalue. This may be a variable offset if an
308 * array dereference has a variable index.
311 switch (deref
->ir_type
) {
312 case ir_type_dereference_variable
: {
317 case ir_type_dereference_array
: {
318 ir_dereference_array
*deref_array
= (ir_dereference_array
*) deref
;
319 unsigned array_stride
;
320 if (deref_array
->array
->type
->is_vector()) {
321 /* We get this when storing or loading a component out of a vector
322 * with a non-constant index. This happens for v[i] = f where v is
323 * a vector (or m[i][j] = f where m is a matrix). If we don't
324 * lower that here, it gets turned into v = vector_insert(v, i,
325 * f), which loads the entire vector, modifies one component and
326 * then write the entire thing back. That breaks if another
327 * thread or SIMD channel is modifying the same vector.
330 if (deref_array
->array
->type
->is_64bit())
332 } else if (deref_array
->array
->type
->is_matrix() && *row_major
) {
333 /* When loading a vector out of a row major matrix, the
334 * step between the columns (vectors) is the size of a
335 * float, while the step between the rows (elements of a
336 * vector) is handled below in emit_ubo_loads.
339 if (deref_array
->array
->type
->is_64bit())
341 *matrix_type
= deref_array
->array
->type
;
342 } else if (deref_array
->type
->without_array()->is_interface()) {
343 /* We're processing an array dereference of an interface instance
344 * array. The thing being dereferenced *must* be a variable
345 * dereference because interfaces cannot be embedded in other
346 * types. In terms of calculating the offsets for the lowering
347 * pass, we don't care about the array index. All elements of an
348 * interface instance array will have the same offsets relative to
349 * the base of the block that backs them.
351 deref
= deref_array
->array
->as_dereference();
354 /* Whether or not the field is row-major (because it might be a
355 * bvec2 or something) does not affect the array itself. We need
356 * to know whether an array element in its entirety is row-major.
358 const bool array_row_major
=
359 is_dereferenced_thing_row_major(deref_array
);
361 /* The array type will give the correct interface packing
364 if (packing
== GLSL_INTERFACE_PACKING_STD430
) {
365 array_stride
= deref_array
->type
->std430_array_stride(array_row_major
);
367 array_stride
= deref_array
->type
->std140_size(array_row_major
);
368 array_stride
= glsl_align(array_stride
, 16);
372 ir_rvalue
*array_index
= deref_array
->array_index
;
373 if (array_index
->type
->base_type
== GLSL_TYPE_INT
)
374 array_index
= i2u(array_index
);
376 ir_constant
*const_index
=
377 array_index
->constant_expression_value(mem_ctx
, NULL
);
379 *const_offset
+= array_stride
* const_index
->value
.u
[0];
381 *offset
= add(*offset
,
383 new(mem_ctx
) ir_constant(array_stride
)));
385 deref
= deref_array
->array
->as_dereference();
389 case ir_type_dereference_record
: {
390 ir_dereference_record
*deref_record
= (ir_dereference_record
*) deref
;
391 const glsl_type
*struct_type
= deref_record
->record
->type
;
392 unsigned intra_struct_offset
= 0;
394 for (unsigned int i
= 0; i
< struct_type
->length
; i
++) {
395 const glsl_type
*type
= struct_type
->fields
.structure
[i
].type
;
397 ir_dereference_record
*field_deref
= new(mem_ctx
)
398 ir_dereference_record(deref_record
->record
,
399 struct_type
->fields
.structure
[i
].name
);
400 const bool field_row_major
=
401 is_dereferenced_thing_row_major(field_deref
);
403 ralloc_free(field_deref
);
405 unsigned field_align
= 0;
407 if (packing
== GLSL_INTERFACE_PACKING_STD430
)
408 field_align
= type
->std430_base_alignment(field_row_major
);
410 field_align
= type
->std140_base_alignment(field_row_major
);
412 if (struct_type
->fields
.structure
[i
].offset
!= -1) {
413 intra_struct_offset
= struct_type
->fields
.structure
[i
].offset
;
416 intra_struct_offset
= glsl_align(intra_struct_offset
, field_align
);
418 assert(deref_record
->field_idx
>= 0);
419 if (i
== (unsigned) deref_record
->field_idx
) {
421 *struct_field
= &struct_type
->fields
.structure
[i
];
425 if (packing
== GLSL_INTERFACE_PACKING_STD430
)
426 intra_struct_offset
+= type
->std430_size(field_row_major
);
428 intra_struct_offset
+= type
->std140_size(field_row_major
);
430 /* If the field just examined was itself a structure, apply rule
433 * "The structure may have padding at the end; the base offset
434 * of the member following the sub-structure is rounded up to
435 * the next multiple of the base alignment of the structure."
437 if (type
->without_array()->is_record()) {
438 intra_struct_offset
= glsl_align(intra_struct_offset
,
444 *const_offset
+= intra_struct_offset
;
445 deref
= deref_record
->record
->as_dereference();
449 case ir_type_swizzle
: {
450 ir_swizzle
*deref_swizzle
= (ir_swizzle
*) deref
;
452 assert(deref_swizzle
->mask
.num_components
== 1);
454 *const_offset
+= deref_swizzle
->mask
.x
* sizeof(int);
455 deref
= deref_swizzle
->val
->as_dereference();
460 assert(!"not reached");
467 } /* namespace lower_buffer_access */