ruby/memory_view.c

507 строки
14 KiB
C

/**********************************************************************
memory_view.c - Memory View
Copyright (C) 2020 Kenta Murata <mrkn@mrkn.jp>
**********************************************************************/
#include "internal.h"
#include "internal/variable.h"
#include "internal/util.h"
#include "ruby/memory_view.h"
#define STRUCT_ALIGNOF(T, result) do { \
(result) = RUBY_ALIGNOF(T); \
} while(0)
static ID id_memory_view;
static const rb_data_type_t memory_view_entry_data_type = {
"memory_view",
{
0,
0,
0,
},
0, 0, RUBY_TYPED_FREE_IMMEDIATELY
};
/* Register memory view functions for the given class */
bool
rb_memory_view_register(VALUE klass, const rb_memory_view_entry_t *entry) {
Check_Type(klass, T_CLASS);
VALUE entry_obj = rb_ivar_lookup(klass, id_memory_view, Qnil);
if (! NIL_P(entry_obj)) {
rb_warning("Duplicated registration of memory view to %"PRIsVALUE, klass);
return false;
}
else {
entry_obj = TypedData_Wrap_Struct(0, &memory_view_entry_data_type, (void *)entry);
rb_ivar_set(klass, id_memory_view, entry_obj);
return true;
}
}
/* Examine whether the given memory view has row-major order strides. */
bool
rb_memory_view_is_row_major_contiguous(const rb_memory_view_t *view)
{
const ssize_t ndim = view->ndim;
const ssize_t *shape = view->shape;
const ssize_t *strides = view->strides;
ssize_t n = view->item_size;
ssize_t i;
for (i = ndim - 1; i >= 0; --i) {
if (strides[i] != n) return false;
n *= shape[i];
}
return true;
}
/* Examine whether the given memory view has column-major order strides. */
bool
rb_memory_view_is_column_major_contiguous(const rb_memory_view_t *view)
{
const ssize_t ndim = view->ndim;
const ssize_t *shape = view->shape;
const ssize_t *strides = view->strides;
ssize_t n = view->item_size;
ssize_t i;
for (i = 0; i < ndim; ++i) {
if (strides[i] != n) return false;
n *= shape[i];
}
return true;
}
/* Initialize strides array to represent the specified contiguous array. */
void
rb_memory_view_fill_contiguous_strides(const ssize_t ndim, const ssize_t item_size, const ssize_t *const shape, const bool row_major_p, ssize_t *const strides)
{
ssize_t i, n = item_size;
if (row_major_p) {
for (i = ndim - 1; i >= 0; --i) {
strides[i] = n;
n *= shape[i];
}
}
else { // column-major
for (i = 0; i < ndim; ++i) {
strides[i] = n;
n *= shape[i];
}
}
}
/* Initialize view to expose a simple byte array */
int
rb_memory_view_init_as_byte_array(rb_memory_view_t *view, VALUE obj, void *data, const ssize_t len, const bool readonly)
{
view->obj = obj;
view->data = data;
view->len = len;
view->readonly = readonly;
view->format = NULL;
view->item_size = 1;
view->ndim = 1;
view->shape = NULL;
view->strides = NULL;
view->sub_offsets = NULL;
*((void **)&view->private) = NULL;
return 1;
}
#ifdef HAVE_TRUE_LONG_LONG
static const char native_types[] = "sSiIlLqQjJ";
#else
static const char native_types[] = "sSiIlLjJ";
#endif
static const char endianness_types[] = "sSiIlLqQjJ";
typedef enum {
ENDIANNESS_NATIVE,
ENDIANNESS_LITTLE,
ENDIANNESS_BIG
} endianness_t;
static ssize_t
get_format_size(const char *format, bool *native_p, ssize_t *alignment, endianness_t *endianness, ssize_t *count, const char **next_format, VALUE *error)
{
RUBY_ASSERT(format != NULL);
RUBY_ASSERT(native_p != NULL);
RUBY_ASSERT(endianness != NULL);
RUBY_ASSERT(count != NULL);
RUBY_ASSERT(next_format != NULL);
*native_p = false;
*endianness = ENDIANNESS_NATIVE;
*count = 1;
const int type_char = *format;
int i = 1;
while (format[i]) {
switch (format[i]) {
case '!':
case '_':
if (strchr(native_types, type_char)) {
*native_p = true;
++i;
}
else {
if (error) {
*error = rb_exc_new_str(rb_eArgError,
rb_sprintf("Unable to specify native size for '%c'", type_char));
}
return -1;
}
continue;
case '<':
case '>':
if (!strchr(endianness_types, type_char)) {
if (error) {
*error = rb_exc_new_str(rb_eArgError,
rb_sprintf("Unable to specify endianness for '%c'", type_char));
}
return -1;
}
if (*endianness != ENDIANNESS_NATIVE) {
*error = rb_exc_new_cstr(rb_eArgError, "Unable to use both '<' and '>' multiple times");
return -1;
}
*endianness = (format[i] == '<') ? ENDIANNESS_LITTLE : ENDIANNESS_BIG;
++i;
continue;
default:
break;
}
break;
}
// parse count
int ch = format[i];
if ('0' <= ch && ch <= '9') {
ssize_t n = 0;
while ('0' <= (ch = format[i]) && ch <= '9') {
n = 10*n + ruby_digit36_to_number_table[ch];
++i;
}
*count = n;
}
*next_format = &format[i];
switch (type_char) {
case 'x': // padding
return 1;
case 'c': // signed char
case 'C': // unsigned char
return sizeof(char);
case 's': // s for int16_t, s! for signed short
case 'S': // S for uint16_t, S! for unsigned short
if (*native_p) {
STRUCT_ALIGNOF(short, *alignment);
return sizeof(short);
}
// fall through
case 'n': // n for big-endian 16bit unsigned integer
case 'v': // v for little-endian 16bit unsigned integer
STRUCT_ALIGNOF(int16_t, *alignment);
return 2;
case 'i': // i and i! for signed int
case 'I': // I and I! for unsigned int
STRUCT_ALIGNOF(int, *alignment);
return sizeof(int);
case 'l': // l for int32_t, l! for signed long
case 'L': // L for uint32_t, L! for unsigned long
if (*native_p) {
STRUCT_ALIGNOF(long, *alignment);
return sizeof(long);
}
// fall through
case 'N': // N for big-endian 32bit unsigned integer
case 'V': // V for little-endian 32bit unsigned integer
STRUCT_ALIGNOF(int32_t, *alignment);
return 4;
case 'f': // f for native float
case 'e': // e for little-endian float
case 'g': // g for big-endian float
STRUCT_ALIGNOF(float, *alignment);
return sizeof(float);
case 'q': // q for int64_t, q! for signed long long
case 'Q': // Q for uint64_t, Q! for unsigned long long
if (*native_p) {
STRUCT_ALIGNOF(LONG_LONG, *alignment);
return sizeof(LONG_LONG);
}
STRUCT_ALIGNOF(int64_t, *alignment);
return 8;
case 'd': // d for native double
case 'E': // E for little-endian double
case 'G': // G for big-endian double
STRUCT_ALIGNOF(double, *alignment);
return sizeof(double);
case 'j': // j for intptr_t
case 'J': // J for uintptr_t
STRUCT_ALIGNOF(intptr_t, *alignment);
return sizeof(intptr_t);
default:
*alignment = -1;
if (error) {
*error = rb_exc_new_str(rb_eArgError, rb_sprintf("Invalid type character '%c'", type_char));
}
return -1;
}
}
static inline ssize_t
calculate_padding(ssize_t total, ssize_t alignment_size) {
if (alignment_size > 1) {
ssize_t res = total % alignment_size;
if (res > 0) {
return alignment_size - res;
}
}
return 0;
}
ssize_t
rb_memory_view_parse_item_format(const char *format,
rb_memory_view_item_component_t **members,
ssize_t *n_members, const char **err)
{
if (format == NULL) return 1;
VALUE error = Qnil;
ssize_t total = 0;
ssize_t len = 0;
bool alignment = false;
ssize_t max_alignment_size = 0;
const char *p = format;
if (*p == '|') { // alginment specifier
alignment = true;
++format;
++p;
}
while (*p) {
const char *q = p;
// ignore spaces
if (ISSPACE(*p)) {
while (ISSPACE(*p)) ++p;
continue;
}
bool native_size_p = false;
ssize_t alignment_size = 0;
endianness_t endianness = ENDIANNESS_NATIVE;
ssize_t count = 0;
const ssize_t size = get_format_size(p, &native_size_p, &alignment_size, &endianness, &count, &p, &error);
if (size < 0) {
if (err) *err = q;
return -1;
}
if (max_alignment_size < alignment_size) {
max_alignment_size = alignment_size;
}
const ssize_t padding = alignment ? calculate_padding(total, alignment_size) : 0;
total += padding + size * count;
if (*q != 'x') {
++len;
}
}
// adjust total size with the alignment size of the largest element
if (alignment && max_alignment_size > 0) {
const ssize_t padding = calculate_padding(total, max_alignment_size);
total += padding;
}
if (members && n_members) {
rb_memory_view_item_component_t *buf = ALLOC_N(rb_memory_view_item_component_t, len);
ssize_t i = 0, offset = 0;
const char *p = format;
while (*p) {
const int type_char = *p;
bool native_size_p;
ssize_t alignment_size = 0;
endianness_t endianness = ENDIANNESS_NATIVE;
ssize_t count = 0;
const ssize_t size = get_format_size(p, &native_size_p, &alignment_size, &endianness, &count, &p, NULL);
const ssize_t padding = alignment ? calculate_padding(offset, alignment_size) : 0;
offset += padding;
if (type_char != 'x') {
#ifdef WORDS_BIGENDIAN
bool little_endian_p = (endianness == ENDIANNESS_LITTLE);
#else
bool little_endian_p = (endianness != ENDIANNESS_BIG);
#endif
switch (type_char) {
case 'e':
case 'E':
little_endian_p = true;
break;
case 'g':
case 'G':
little_endian_p = false;
break;
default:
break;
}
buf[i++] = (rb_memory_view_item_component_t){
.format = type_char,
.native_size_p = native_size_p,
.little_endian_p = little_endian_p,
.offset = offset,
.size = size,
.repeat = count
};
}
offset += size * count;
}
*members = buf;
*n_members = len;
}
return total;
}
/* Return the item size. */
ssize_t
rb_memory_view_item_size_from_format(const char *format, const char **err)
{
return rb_memory_view_parse_item_format(format, NULL, NULL, err);
}
/* Return the pointer to the item located by the given indices. */
void *
rb_memory_view_get_item_pointer(rb_memory_view_t *view, const ssize_t *indices)
{
uint8_t *ptr = view->data;
if (view->ndim == 1) {
ssize_t stride = view->strides != NULL ? view->strides[0] : view->item_size;
return ptr + indices[0] * stride;
}
assert(view->shape != NULL);
ssize_t i;
if (view->strides == NULL) {
// row-major contiguous array
ssize_t stride = view->item_size;
for (i = 0; i < view->ndim; ++i) {
stride *= view->shape[i];
}
for (i = 0; i < view->ndim; ++i) {
stride /= view->shape[i];
ptr += indices[i] * stride;
}
}
else if (view->sub_offsets == NULL) {
// flat strided array
for (i = 0; i < view->ndim; ++i) {
ptr += indices[i] * view->strides[i];
}
}
else {
// indirect strided array
for (i = 0; i < view->ndim; ++i) {
ptr += indices[i] * view->strides[i];
if (view->sub_offsets[i] >= 0) {
ptr = *(uint8_t **)ptr + view->sub_offsets[i];
}
}
}
return ptr;
}
static const rb_memory_view_entry_t *
lookup_memory_view_entry(VALUE klass)
{
VALUE entry_obj = rb_ivar_lookup(klass, id_memory_view, Qnil);
while (NIL_P(entry_obj)) {
klass = rb_class_get_superclass(klass);
if (klass == rb_cBasicObject || klass == rb_cObject)
return NULL;
entry_obj = rb_ivar_lookup(klass, id_memory_view, Qnil);
}
if (! rb_typeddata_is_kind_of(entry_obj, &memory_view_entry_data_type))
return NULL;
return (const rb_memory_view_entry_t *)RTYPEDDATA_DATA(entry_obj);
}
/* Examine whether the given object supports memory view. */
int
rb_memory_view_available_p(VALUE obj)
{
VALUE klass = CLASS_OF(obj);
const rb_memory_view_entry_t *entry = lookup_memory_view_entry(klass);
if (entry)
return (* entry->available_p_func)(obj);
else
return 0;
}
/* Obtain a memory view from obj, and substitute the information to view. */
int
rb_memory_view_get(VALUE obj, rb_memory_view_t* view, int flags)
{
VALUE klass = CLASS_OF(obj);
const rb_memory_view_entry_t *entry = lookup_memory_view_entry(klass);
if (entry)
return (*entry->get_func)(obj, view, flags);
else
return 0;
}
/* Release the memory view obtained from obj. */
int
rb_memory_view_release(rb_memory_view_t* view)
{
VALUE klass = CLASS_OF(view->obj);
const rb_memory_view_entry_t *entry = lookup_memory_view_entry(klass);
if (entry)
return (*entry->release_func)(view->obj, view);
else
return 0;
}
void
Init_MemoryView(void)
{
id_memory_view = rb_intern_const("__memory_view__");
}