• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
40 using namespace ir_builder;
41 
42 namespace lower_buffer_access {
43 
44 static inline int
writemask_for_size(unsigned n)45 writemask_for_size(unsigned n)
46 {
47    return ((1 << n) - 1);
48 }
49 
50 /**
51  * Takes a deref and recursively calls itself to break the deref down to the
52  * point that the reads or writes generated are contiguous scalars or vectors.
53  */
54 void
emit_access(void * mem_ctx,bool is_write,ir_dereference * deref,ir_variable * base_offset,unsigned int deref_offset,bool row_major,int matrix_columns,unsigned int packing,unsigned int write_mask)55 lower_buffer_access::emit_access(void *mem_ctx,
56                                  bool is_write,
57                                  ir_dereference *deref,
58                                  ir_variable *base_offset,
59                                  unsigned int deref_offset,
60                                  bool row_major,
61                                  int matrix_columns,
62                                  unsigned int packing,
63                                  unsigned int write_mask)
64 {
65    if (deref->type->is_record()) {
66       unsigned int field_offset = 0;
67 
68       for (unsigned i = 0; i < deref->type->length; i++) {
69          const struct glsl_struct_field *field =
70             &deref->type->fields.structure[i];
71          ir_dereference *field_deref =
72             new(mem_ctx) ir_dereference_record(deref->clone(mem_ctx, NULL),
73                                                field->name);
74 
75          field_offset =
76             glsl_align(field_offset,
77                        field->type->std140_base_alignment(row_major));
78 
79          emit_access(mem_ctx, is_write, field_deref, base_offset,
80                      deref_offset + field_offset,
81                      row_major, 1, packing,
82                      writemask_for_size(field_deref->type->vector_elements));
83 
84          field_offset += field->type->std140_size(row_major);
85       }
86       return;
87    }
88 
89    if (deref->type->is_array()) {
90       unsigned array_stride = packing == GLSL_INTERFACE_PACKING_STD430 ?
91          deref->type->fields.array->std430_array_stride(row_major) :
92          glsl_align(deref->type->fields.array->std140_size(row_major), 16);
93 
94       for (unsigned i = 0; i < deref->type->length; i++) {
95          ir_constant *element = new(mem_ctx) ir_constant(i);
96          ir_dereference *element_deref =
97             new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
98                                               element);
99          emit_access(mem_ctx, is_write, element_deref, base_offset,
100                      deref_offset + i * array_stride,
101                      row_major, 1, packing,
102                      writemask_for_size(element_deref->type->vector_elements));
103       }
104       return;
105    }
106 
107    if (deref->type->is_matrix()) {
108       for (unsigned i = 0; i < deref->type->matrix_columns; i++) {
109          ir_constant *col = new(mem_ctx) ir_constant(i);
110          ir_dereference *col_deref =
111             new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL), col);
112 
113          if (row_major) {
114             /* For a row-major matrix, the next column starts at the next
115              * element.
116              */
117             int size_mul = deref->type->is_64bit() ? 8 : 4;
118             emit_access(mem_ctx, is_write, col_deref, base_offset,
119                         deref_offset + i * size_mul,
120                         row_major, deref->type->matrix_columns, packing,
121                         writemask_for_size(col_deref->type->vector_elements));
122          } else {
123             int size_mul;
124 
125             /* std430 doesn't round up vec2 size to a vec4 size */
126             if (packing == GLSL_INTERFACE_PACKING_STD430 &&
127                 deref->type->vector_elements == 2 &&
128                 !deref->type->is_64bit()) {
129                size_mul = 8;
130             } else {
131                /* std140 always rounds the stride of arrays (and matrices) to a
132                 * vec4, so matrices are always 16 between columns/rows. With
133                 * doubles, they will be 32 apart when there are more than 2 rows.
134                 *
135                 * For both std140 and std430, if the member is a
136                 * three-'component vector with components consuming N basic
137                 * machine units, the base alignment is 4N. For vec4, base
138                 * alignment is 4N.
139                 */
140                size_mul = (deref->type->is_64bit() &&
141                            deref->type->vector_elements > 2) ? 32 : 16;
142             }
143 
144             emit_access(mem_ctx, is_write, col_deref, base_offset,
145                         deref_offset + i * size_mul,
146                         row_major, deref->type->matrix_columns, packing,
147                         writemask_for_size(col_deref->type->vector_elements));
148          }
149       }
150       return;
151    }
152 
153    assert(deref->type->is_scalar() || deref->type->is_vector());
154 
155    if (!row_major) {
156       ir_rvalue *offset =
157          add(base_offset, new(mem_ctx) ir_constant(deref_offset));
158       unsigned mask =
159          is_write ? write_mask : (1 << deref->type->vector_elements) - 1;
160       insert_buffer_access(mem_ctx, deref, deref->type, offset, mask, -1);
161    } else {
162       unsigned N = deref->type->is_64bit() ? 8 : 4;
163 
164       /* We're dereffing a column out of a row-major matrix, so we
165        * gather the vector from each stored row.
166       */
167       assert(deref->type->base_type == GLSL_TYPE_FLOAT ||
168              deref->type->base_type == GLSL_TYPE_DOUBLE);
169       /* Matrices, row_major or not, are stored as if they were
170        * arrays of vectors of the appropriate size in std140.
171        * Arrays have their strides rounded up to a vec4, so the
172        * matrix stride is always 16. However a double matrix may either be 16
173        * or 32 depending on the number of columns.
174        */
175       assert(matrix_columns <= 4);
176       unsigned matrix_stride = 0;
177       /* Matrix stride for std430 mat2xY matrices are not rounded up to
178        * vec4 size. From OpenGL 4.3 spec, section 7.6.2.2 "Standard Uniform
179        * Block Layout":
180        *
181        * "2. If the member is a two- or four-component vector with components
182        * consuming N basic machine units, the base alignment is 2N or 4N,
183        * respectively." [...]
184        * "4. If the member is an array of scalars or vectors, the base alignment
185        * and array stride are set to match the base alignment of a single array
186        * element, according to rules (1), (2), and (3), and rounded up to the
187        * base alignment of a vec4." [...]
188        * "7. If the member is a row-major matrix with C columns and R rows, the
189        * matrix is stored identically to an array of R row vectors with C
190        * components each, according to rule (4)." [...]
191        * "When using the std430 storage layout, shader storage blocks will be
192        * laid out in buffer storage identically to uniform and shader storage
193        * blocks using the std140 layout, except that the base alignment and
194        * stride of arrays of scalars and vectors in rule 4 and of structures in
195        * rule 9 are not rounded up a multiple of the base alignment of a vec4."
196        */
197       if (packing == GLSL_INTERFACE_PACKING_STD430 && matrix_columns == 2)
198          matrix_stride = 2 * N;
199       else
200          matrix_stride = glsl_align(matrix_columns * N, 16);
201 
202       const glsl_type *deref_type = deref->type->base_type == GLSL_TYPE_FLOAT ?
203          glsl_type::float_type : glsl_type::double_type;
204 
205       for (unsigned i = 0; i < deref->type->vector_elements; i++) {
206          ir_rvalue *chan_offset =
207             add(base_offset,
208                 new(mem_ctx) ir_constant(deref_offset + i * matrix_stride));
209          if (!is_write || ((1U << i) & write_mask))
210             insert_buffer_access(mem_ctx, deref, deref_type, chan_offset,
211                                  (1U << i), i);
212       }
213    }
214 }
215 
216 /**
217  * Determine if a thing being dereferenced is row-major
218  *
219  * There is some trickery here.
220  *
221  * If the thing being dereferenced is a member of uniform block \b without an
222  * instance name, then the name of the \c ir_variable is the field name of an
223  * interface type.  If this field is row-major, then the thing referenced is
224  * row-major.
225  *
226  * If the thing being dereferenced is a member of uniform block \b with an
227  * instance name, then the last dereference in the tree will be an
228  * \c ir_dereference_record.  If that record field is row-major, then the
229  * thing referenced is row-major.
230  */
231 bool
is_dereferenced_thing_row_major(const ir_rvalue * deref)232 lower_buffer_access::is_dereferenced_thing_row_major(const ir_rvalue *deref)
233 {
234    bool matrix = false;
235    const ir_rvalue *ir = deref;
236 
237    while (true) {
238       matrix = matrix || ir->type->without_array()->is_matrix();
239 
240       switch (ir->ir_type) {
241       case ir_type_dereference_array: {
242          const ir_dereference_array *const array_deref =
243             (const ir_dereference_array *) ir;
244 
245          ir = array_deref->array;
246          break;
247       }
248 
249       case ir_type_dereference_record: {
250          const ir_dereference_record *const record_deref =
251             (const ir_dereference_record *) ir;
252 
253          ir = record_deref->record;
254 
255          const int idx = ir->type->field_index(record_deref->field);
256          assert(idx >= 0);
257 
258          const enum glsl_matrix_layout matrix_layout =
259             glsl_matrix_layout(ir->type->fields.structure[idx].matrix_layout);
260 
261          switch (matrix_layout) {
262          case GLSL_MATRIX_LAYOUT_INHERITED:
263             break;
264          case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
265             return false;
266          case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
267             return matrix || deref->type->without_array()->is_record();
268          }
269 
270          break;
271       }
272 
273       case ir_type_dereference_variable: {
274          const ir_dereference_variable *const var_deref =
275             (const ir_dereference_variable *) ir;
276 
277          const enum glsl_matrix_layout matrix_layout =
278             glsl_matrix_layout(var_deref->var->data.matrix_layout);
279 
280          switch (matrix_layout) {
281          case GLSL_MATRIX_LAYOUT_INHERITED: {
282             /* For interface block matrix variables we handle inherited
283              * layouts at HIR generation time, but we don't do that for shared
284              * variables, which are always column-major
285              */
286             MAYBE_UNUSED ir_variable *var = deref->variable_referenced();
287             assert((var->is_in_buffer_block() && !matrix) ||
288                    var->data.mode == ir_var_shader_shared);
289             return false;
290          }
291          case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
292             return false;
293          case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
294             return matrix || deref->type->without_array()->is_record();
295          }
296 
297          unreachable("invalid matrix layout");
298          break;
299       }
300 
301       default:
302          return false;
303       }
304    }
305 
306    /* The tree must have ended with a dereference that wasn't an
307     * ir_dereference_variable.  That is invalid, and it should be impossible.
308     */
309    unreachable("invalid dereference tree");
310    return false;
311 }
312 
313 /**
314  * This function initializes various values that will be used later by
315  * emit_access when actually emitting loads or stores.
316  *
317  * Note: const_offset is an input as well as an output, clients must
318  * initialize it to the offset of the variable in the underlying block, and
319  * this function will adjust it by adding the constant offset of the member
320  * being accessed into that variable.
321  */
322 void
setup_buffer_access(void * mem_ctx,ir_rvalue * deref,ir_rvalue ** offset,unsigned * const_offset,bool * row_major,int * matrix_columns,const glsl_struct_field ** struct_field,enum glsl_interface_packing packing)323 lower_buffer_access::setup_buffer_access(void *mem_ctx,
324                                          ir_rvalue *deref,
325                                          ir_rvalue **offset,
326                                          unsigned *const_offset,
327                                          bool *row_major,
328                                          int *matrix_columns,
329                                          const glsl_struct_field **struct_field,
330                                          enum glsl_interface_packing packing)
331 {
332    *offset = new(mem_ctx) ir_constant(0u);
333    *row_major = is_dereferenced_thing_row_major(deref);
334    *matrix_columns = 1;
335 
336    /* Calculate the offset to the start of the region of the UBO
337     * dereferenced by *rvalue.  This may be a variable offset if an
338     * array dereference has a variable index.
339     */
340    while (deref) {
341       switch (deref->ir_type) {
342       case ir_type_dereference_variable: {
343          deref = NULL;
344          break;
345       }
346 
347       case ir_type_dereference_array: {
348          ir_dereference_array *deref_array = (ir_dereference_array *) deref;
349          unsigned array_stride;
350          if (deref_array->array->type->is_vector()) {
351             /* We get this when storing or loading a component out of a vector
352              * with a non-constant index. This happens for v[i] = f where v is
353              * a vector (or m[i][j] = f where m is a matrix). If we don't
354              * lower that here, it gets turned into v = vector_insert(v, i,
355              * f), which loads the entire vector, modifies one component and
356              * then write the entire thing back.  That breaks if another
357              * thread or SIMD channel is modifying the same vector.
358              */
359             array_stride = 4;
360             if (deref_array->array->type->is_64bit())
361                array_stride *= 2;
362          } else if (deref_array->array->type->is_matrix() && *row_major) {
363             /* When loading a vector out of a row major matrix, the
364              * step between the columns (vectors) is the size of a
365              * float, while the step between the rows (elements of a
366              * vector) is handled below in emit_ubo_loads.
367              */
368             array_stride = 4;
369             if (deref_array->array->type->is_64bit())
370                array_stride *= 2;
371             *matrix_columns = deref_array->array->type->matrix_columns;
372          } else if (deref_array->type->without_array()->is_interface()) {
373             /* We're processing an array dereference of an interface instance
374              * array. The thing being dereferenced *must* be a variable
375              * dereference because interfaces cannot be embedded in other
376              * types. In terms of calculating the offsets for the lowering
377              * pass, we don't care about the array index. All elements of an
378              * interface instance array will have the same offsets relative to
379              * the base of the block that backs them.
380              */
381             deref = deref_array->array->as_dereference();
382             break;
383          } else {
384             /* Whether or not the field is row-major (because it might be a
385              * bvec2 or something) does not affect the array itself. We need
386              * to know whether an array element in its entirety is row-major.
387              */
388             const bool array_row_major =
389                is_dereferenced_thing_row_major(deref_array);
390 
391             /* The array type will give the correct interface packing
392              * information
393              */
394             if (packing == GLSL_INTERFACE_PACKING_STD430) {
395                array_stride = deref_array->type->std430_array_stride(array_row_major);
396             } else {
397                array_stride = deref_array->type->std140_size(array_row_major);
398                array_stride = glsl_align(array_stride, 16);
399             }
400          }
401 
402          ir_rvalue *array_index = deref_array->array_index;
403          if (array_index->type->base_type == GLSL_TYPE_INT)
404             array_index = i2u(array_index);
405 
406          ir_constant *const_index =
407             array_index->constant_expression_value(NULL);
408          if (const_index) {
409             *const_offset += array_stride * const_index->value.u[0];
410          } else {
411             *offset = add(*offset,
412                           mul(array_index,
413                               new(mem_ctx) ir_constant(array_stride)));
414          }
415          deref = deref_array->array->as_dereference();
416          break;
417       }
418 
419       case ir_type_dereference_record: {
420          ir_dereference_record *deref_record = (ir_dereference_record *) deref;
421          const glsl_type *struct_type = deref_record->record->type;
422          unsigned intra_struct_offset = 0;
423 
424          for (unsigned int i = 0; i < struct_type->length; i++) {
425             const glsl_type *type = struct_type->fields.structure[i].type;
426 
427             ir_dereference_record *field_deref = new(mem_ctx)
428                ir_dereference_record(deref_record->record,
429                                      struct_type->fields.structure[i].name);
430             const bool field_row_major =
431                is_dereferenced_thing_row_major(field_deref);
432 
433             ralloc_free(field_deref);
434 
435             unsigned field_align = 0;
436 
437             if (packing == GLSL_INTERFACE_PACKING_STD430)
438                field_align = type->std430_base_alignment(field_row_major);
439             else
440                field_align = type->std140_base_alignment(field_row_major);
441 
442             if (struct_type->fields.structure[i].offset != -1) {
443                intra_struct_offset = struct_type->fields.structure[i].offset;
444             }
445 
446             intra_struct_offset = glsl_align(intra_struct_offset, field_align);
447 
448             if (strcmp(struct_type->fields.structure[i].name,
449                        deref_record->field) == 0) {
450                if (struct_field)
451                   *struct_field = &struct_type->fields.structure[i];
452                break;
453             }
454 
455             if (packing == GLSL_INTERFACE_PACKING_STD430)
456                intra_struct_offset += type->std430_size(field_row_major);
457             else
458                intra_struct_offset += type->std140_size(field_row_major);
459 
460             /* If the field just examined was itself a structure, apply rule
461              * #9:
462              *
463              *     "The structure may have padding at the end; the base offset
464              *     of the member following the sub-structure is rounded up to
465              *     the next multiple of the base alignment of the structure."
466              */
467             if (type->without_array()->is_record()) {
468                intra_struct_offset = glsl_align(intra_struct_offset,
469                                                 field_align);
470 
471             }
472          }
473 
474          *const_offset += intra_struct_offset;
475          deref = deref_record->record->as_dereference();
476          break;
477       }
478 
479       case ir_type_swizzle: {
480          ir_swizzle *deref_swizzle = (ir_swizzle *) deref;
481 
482          assert(deref_swizzle->mask.num_components == 1);
483 
484          *const_offset += deref_swizzle->mask.x * sizeof(int);
485          deref = deref_swizzle->val->as_dereference();
486          break;
487       }
488 
489       default:
490          assert(!"not reached");
491          deref = NULL;
492          break;
493       }
494    }
495 }
496 
497 } /* namespace lower_buffer_access */
498