1 /*
2 * Copyright (c) 2015 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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.
22 */
23
24 /**
25 * \file lower_buffer_access.cpp
26 *
27 * Helper for IR lowering pass to replace dereferences of buffer object based
28 * shader variables with intrinsic function calls.
29 *
30 * This helper is used by lowering passes for UBOs, SSBOs and compute shader
31 * shared variables.
32 */
33
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"
39 #include "linker.h"
40
41 using namespace ir_builder;
42
43 namespace lower_buffer_access {
44
45 static inline int
writemask_for_size(unsigned n)46 writemask_for_size(unsigned n)
47 {
48 return ((1 << n) - 1);
49 }
50
51 /**
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.
54 */
55 void
emit_access(void * mem_ctx,bool is_write,ir_dereference * deref,ir_variable * base_offset,unsigned int deref_offset,bool row_major,const glsl_type * matrix_type,enum glsl_interface_packing packing,unsigned int write_mask)56 lower_buffer_access::emit_access(void *mem_ctx,
57 bool is_write,
58 ir_dereference *deref,
59 ir_variable *base_offset,
60 unsigned int deref_offset,
61 bool row_major,
62 const glsl_type *matrix_type,
63 enum glsl_interface_packing packing,
64 unsigned int write_mask)
65 {
66 if (deref->type->is_record()) {
67 unsigned int field_offset = 0;
68
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),
74 field->name);
75
76 unsigned field_align;
77 if (packing == GLSL_INTERFACE_PACKING_STD430)
78 field_align = field->type->std430_base_alignment(row_major);
79 else
80 field_align = field->type->std140_base_alignment(row_major);
81 field_offset = glsl_align(field_offset, field_align);
82
83 emit_access(mem_ctx, is_write, field_deref, base_offset,
84 deref_offset + field_offset,
85 row_major, NULL, packing,
86 writemask_for_size(field_deref->type->vector_elements));
87
88 if (packing == GLSL_INTERFACE_PACKING_STD430)
89 field_offset += field->type->std430_size(row_major);
90 else
91 field_offset += field->type->std140_size(row_major);
92 }
93 return;
94 }
95
96 if (deref->type->is_array()) {
97 unsigned array_stride = packing == GLSL_INTERFACE_PACKING_STD430 ?
98 deref->type->fields.array->std430_array_stride(row_major) :
99 glsl_align(deref->type->fields.array->std140_size(row_major), 16);
100
101 for (unsigned i = 0; i < deref->type->length; i++) {
102 ir_constant *element = new(mem_ctx) ir_constant(i);
103 ir_dereference *element_deref =
104 new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
105 element);
106 emit_access(mem_ctx, is_write, element_deref, base_offset,
107 deref_offset + i * array_stride,
108 row_major, NULL, packing,
109 writemask_for_size(element_deref->type->vector_elements));
110 }
111 return;
112 }
113
114 if (deref->type->is_matrix()) {
115 for (unsigned i = 0; i < deref->type->matrix_columns; i++) {
116 ir_constant *col = new(mem_ctx) ir_constant(i);
117 ir_dereference *col_deref =
118 new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL), col);
119
120 /* For a row-major matrix, the next column starts at the next
121 * element. Otherwise it is offset by the matrix stride.
122 */
123 const unsigned size_mul = row_major
124 ? (deref->type->is_double() ? 8 : 4)
125 : link_calculate_matrix_stride(deref->type, row_major, packing);
126
127 emit_access(mem_ctx, is_write, col_deref, base_offset,
128 deref_offset + i * size_mul,
129 row_major, deref->type, packing,
130 writemask_for_size(col_deref->type->vector_elements));
131 }
132 return;
133 }
134
135 assert(deref->type->is_scalar() || deref->type->is_vector());
136
137 if (!row_major) {
138 ir_rvalue *offset =
139 add(base_offset, new(mem_ctx) ir_constant(deref_offset));
140 unsigned mask =
141 is_write ? write_mask : (1 << deref->type->vector_elements) - 1;
142 insert_buffer_access(mem_ctx, deref, deref->type, offset, mask, -1);
143 } else {
144 /* We're dereffing a column out of a row-major matrix, so we
145 * gather the vector from each stored row.
146 */
147 assert(deref->type->is_float() || deref->type->is_double());
148 assert(matrix_type != NULL);
149
150 const unsigned matrix_stride =
151 link_calculate_matrix_stride(matrix_type, row_major, packing);
152
153 const glsl_type *deref_type = deref->type->get_scalar_type();
154
155 for (unsigned i = 0; i < deref->type->vector_elements; i++) {
156 ir_rvalue *chan_offset =
157 add(base_offset,
158 new(mem_ctx) ir_constant(deref_offset + i * matrix_stride));
159 if (!is_write || ((1U << i) & write_mask))
160 insert_buffer_access(mem_ctx, deref, deref_type, chan_offset,
161 (1U << i), i);
162 }
163 }
164 }
165
166 /**
167 * Determine if a thing being dereferenced is row-major
168 *
169 * There is some trickery here.
170 *
171 * If the thing being dereferenced is a member of uniform block \b without an
172 * instance name, then the name of the \c ir_variable is the field name of an
173 * interface type. If this field is row-major, then the thing referenced is
174 * row-major.
175 *
176 * If the thing being dereferenced is a member of uniform block \b with an
177 * instance name, then the last dereference in the tree will be an
178 * \c ir_dereference_record. If that record field is row-major, then the
179 * thing referenced is row-major.
180 */
181 bool
is_dereferenced_thing_row_major(const ir_rvalue * deref)182 lower_buffer_access::is_dereferenced_thing_row_major(const ir_rvalue *deref)
183 {
184 bool matrix = false;
185 const ir_rvalue *ir = deref;
186
187 while (true) {
188 matrix = matrix || ir->type->without_array()->is_matrix();
189
190 switch (ir->ir_type) {
191 case ir_type_dereference_array: {
192 const ir_dereference_array *const array_deref =
193 (const ir_dereference_array *) ir;
194
195 ir = array_deref->array;
196 break;
197 }
198
199 case ir_type_dereference_record: {
200 const ir_dereference_record *const record_deref =
201 (const ir_dereference_record *) ir;
202
203 ir = record_deref->record;
204
205 const int idx = record_deref->field_idx;
206 assert(idx >= 0);
207
208 const enum glsl_matrix_layout matrix_layout =
209 glsl_matrix_layout(ir->type->fields.structure[idx].matrix_layout);
210
211 switch (matrix_layout) {
212 case GLSL_MATRIX_LAYOUT_INHERITED:
213 break;
214 case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
215 return false;
216 case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
217 return matrix || deref->type->without_array()->is_record();
218 }
219
220 break;
221 }
222
223 case ir_type_dereference_variable: {
224 const ir_dereference_variable *const var_deref =
225 (const ir_dereference_variable *) ir;
226
227 const enum glsl_matrix_layout matrix_layout =
228 glsl_matrix_layout(var_deref->var->data.matrix_layout);
229
230 switch (matrix_layout) {
231 case GLSL_MATRIX_LAYOUT_INHERITED: {
232 /* For interface block matrix variables we handle inherited
233 * layouts at HIR generation time, but we don't do that for shared
234 * variables, which are always column-major
235 */
236 MAYBE_UNUSED ir_variable *var = deref->variable_referenced();
237 assert((var->is_in_buffer_block() && !matrix) ||
238 var->data.mode == ir_var_shader_shared);
239 return false;
240 }
241 case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
242 return false;
243 case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
244 return matrix || deref->type->without_array()->is_record();
245 }
246
247 unreachable("invalid matrix layout");
248 break;
249 }
250
251 default:
252 return false;
253 }
254 }
255
256 /* The tree must have ended with a dereference that wasn't an
257 * ir_dereference_variable. That is invalid, and it should be impossible.
258 */
259 unreachable("invalid dereference tree");
260 return false;
261 }
262
263 /**
264 * This function initializes various values that will be used later by
265 * emit_access when actually emitting loads or stores.
266 *
267 * Note: const_offset is an input as well as an output, clients must
268 * initialize it to the offset of the variable in the underlying block, and
269 * this function will adjust it by adding the constant offset of the member
270 * being accessed into that variable.
271 */
272 void
setup_buffer_access(void * mem_ctx,ir_rvalue * deref,ir_rvalue ** offset,unsigned * const_offset,bool * row_major,const glsl_type ** matrix_type,const glsl_struct_field ** struct_field,enum glsl_interface_packing packing)273 lower_buffer_access::setup_buffer_access(void *mem_ctx,
274 ir_rvalue *deref,
275 ir_rvalue **offset,
276 unsigned *const_offset,
277 bool *row_major,
278 const glsl_type **matrix_type,
279 const glsl_struct_field **struct_field,
280 enum glsl_interface_packing packing)
281 {
282 *offset = new(mem_ctx) ir_constant(0u);
283 *row_major = is_dereferenced_thing_row_major(deref);
284 *matrix_type = NULL;
285
286 /* Calculate the offset to the start of the region of the UBO
287 * dereferenced by *rvalue. This may be a variable offset if an
288 * array dereference has a variable index.
289 */
290 while (deref) {
291 switch (deref->ir_type) {
292 case ir_type_dereference_variable: {
293 deref = NULL;
294 break;
295 }
296
297 case ir_type_dereference_array: {
298 ir_dereference_array *deref_array = (ir_dereference_array *) deref;
299 unsigned array_stride;
300 if (deref_array->array->type->is_vector()) {
301 /* We get this when storing or loading a component out of a vector
302 * with a non-constant index. This happens for v[i] = f where v is
303 * a vector (or m[i][j] = f where m is a matrix). If we don't
304 * lower that here, it gets turned into v = vector_insert(v, i,
305 * f), which loads the entire vector, modifies one component and
306 * then write the entire thing back. That breaks if another
307 * thread or SIMD channel is modifying the same vector.
308 */
309 array_stride = 4;
310 if (deref_array->array->type->is_64bit())
311 array_stride *= 2;
312 } else if (deref_array->array->type->is_matrix() && *row_major) {
313 /* When loading a vector out of a row major matrix, the
314 * step between the columns (vectors) is the size of a
315 * float, while the step between the rows (elements of a
316 * vector) is handled below in emit_ubo_loads.
317 */
318 array_stride = 4;
319 if (deref_array->array->type->is_64bit())
320 array_stride *= 2;
321 *matrix_type = deref_array->array->type;
322 } else if (deref_array->type->without_array()->is_interface()) {
323 /* We're processing an array dereference of an interface instance
324 * array. The thing being dereferenced *must* be a variable
325 * dereference because interfaces cannot be embedded in other
326 * types. In terms of calculating the offsets for the lowering
327 * pass, we don't care about the array index. All elements of an
328 * interface instance array will have the same offsets relative to
329 * the base of the block that backs them.
330 */
331 deref = deref_array->array->as_dereference();
332 break;
333 } else {
334 /* Whether or not the field is row-major (because it might be a
335 * bvec2 or something) does not affect the array itself. We need
336 * to know whether an array element in its entirety is row-major.
337 */
338 const bool array_row_major =
339 is_dereferenced_thing_row_major(deref_array);
340
341 /* The array type will give the correct interface packing
342 * information
343 */
344 if (packing == GLSL_INTERFACE_PACKING_STD430) {
345 array_stride = deref_array->type->std430_array_stride(array_row_major);
346 } else {
347 array_stride = deref_array->type->std140_size(array_row_major);
348 array_stride = glsl_align(array_stride, 16);
349 }
350 }
351
352 ir_rvalue *array_index = deref_array->array_index;
353 if (array_index->type->base_type == GLSL_TYPE_INT)
354 array_index = i2u(array_index);
355
356 ir_constant *const_index =
357 array_index->constant_expression_value(mem_ctx, NULL);
358 if (const_index) {
359 *const_offset += array_stride * const_index->value.u[0];
360 } else {
361 *offset = add(*offset,
362 mul(array_index,
363 new(mem_ctx) ir_constant(array_stride)));
364 }
365 deref = deref_array->array->as_dereference();
366 break;
367 }
368
369 case ir_type_dereference_record: {
370 ir_dereference_record *deref_record = (ir_dereference_record *) deref;
371 const glsl_type *struct_type = deref_record->record->type;
372 unsigned intra_struct_offset = 0;
373
374 for (unsigned int i = 0; i < struct_type->length; i++) {
375 const glsl_type *type = struct_type->fields.structure[i].type;
376
377 ir_dereference_record *field_deref = new(mem_ctx)
378 ir_dereference_record(deref_record->record,
379 struct_type->fields.structure[i].name);
380 const bool field_row_major =
381 is_dereferenced_thing_row_major(field_deref);
382
383 ralloc_free(field_deref);
384
385 unsigned field_align = 0;
386
387 if (packing == GLSL_INTERFACE_PACKING_STD430)
388 field_align = type->std430_base_alignment(field_row_major);
389 else
390 field_align = type->std140_base_alignment(field_row_major);
391
392 if (struct_type->fields.structure[i].offset != -1) {
393 intra_struct_offset = struct_type->fields.structure[i].offset;
394 }
395
396 intra_struct_offset = glsl_align(intra_struct_offset, field_align);
397
398 assert(deref_record->field_idx >= 0);
399 if (i == (unsigned) deref_record->field_idx) {
400 if (struct_field)
401 *struct_field = &struct_type->fields.structure[i];
402 break;
403 }
404
405 if (packing == GLSL_INTERFACE_PACKING_STD430)
406 intra_struct_offset += type->std430_size(field_row_major);
407 else
408 intra_struct_offset += type->std140_size(field_row_major);
409
410 /* If the field just examined was itself a structure, apply rule
411 * #9:
412 *
413 * "The structure may have padding at the end; the base offset
414 * of the member following the sub-structure is rounded up to
415 * the next multiple of the base alignment of the structure."
416 */
417 if (type->without_array()->is_record()) {
418 intra_struct_offset = glsl_align(intra_struct_offset,
419 field_align);
420
421 }
422 }
423
424 *const_offset += intra_struct_offset;
425 deref = deref_record->record->as_dereference();
426 break;
427 }
428
429 case ir_type_swizzle: {
430 ir_swizzle *deref_swizzle = (ir_swizzle *) deref;
431
432 assert(deref_swizzle->mask.num_components == 1);
433
434 *const_offset += deref_swizzle->mask.x * sizeof(int);
435 deref = deref_swizzle->val->as_dereference();
436 break;
437 }
438
439 default:
440 assert(!"not reached");
441 deref = NULL;
442 break;
443 }
444 }
445 }
446
447 } /* namespace lower_buffer_access */
448