Halide 16.0.0
Halide compiler and libraries
device_buffer_utils.h
Go to the documentation of this file.
1#ifndef HALIDE_RUNTIME_DEVICE_BUFFER_UTILS_H
2#define HALIDE_RUNTIME_DEVICE_BUFFER_UTILS_H
3
4#include "HalideRuntime.h"
5#include "device_interface.h"
6#include "printer.h"
7
8namespace Halide {
9namespace Runtime {
10namespace Internal {
11
12// A host <-> dev copy should be done with the fewest possible number
13// of contiguous copies to minimize driver overhead. If our
14// halide_buffer_t has strides larger than its extents (e.g. because
15// it represents a sub-region of a larger halide_buffer_t) we can't
16// safely copy it back and forth using a single contiguous copy,
17// because we'd clobber in-between values that another thread might be
18// using. In the best case we can do a single contiguous copy, but in
19// the worst case we need to individually copy over every pixel.
20//
21// This problem is made extra difficult by the fact that the ordering
22// of the dimensions in a halide_buffer_t doesn't relate to memory layout at
23// all, so the strides could be in any order.
24//
25// We solve it by representing a copy job we need to perform as a
26// device_copy struct. It describes a multi-dimensional array of
27// copies to perform. Initially it describes copying over a single
28// pixel at a time. We then try to discover contiguous groups of
29// copies that can be coalesced into a single larger copy.
30
31// The struct that describes a host <-> dev copy to perform.
32#define MAX_COPY_DIMS 16
34 // opaque handles for source and device memory.
36 // The offset in the source memory to start
38 // The multidimensional array of contiguous copy tasks that need to be done.
40 // The strides (in bytes) that separate adjacent copy tasks in each dimension.
43 // How many contiguous bytes to copy per task
45};
46
47WEAK void copy_memory_helper(const device_copy &copy, int d, int64_t src_off, int64_t dst_off) {
48 // Skip size-1 dimensions
49 while (d >= 0 && copy.extent[d] == 1) {
50 d--;
51 }
52
53 if (d == -1) {
54 const void *from = (void *)(copy.src + src_off);
55 void *to = (void *)(copy.dst + dst_off);
56 memcpy(to, from, copy.chunk_size);
57 } else {
58 for (uint64_t i = 0; i < copy.extent[d]; i++) {
59 copy_memory_helper(copy, d - 1, src_off, dst_off);
60 src_off += copy.src_stride_bytes[d];
61 dst_off += copy.dst_stride_bytes[d];
62 }
63 }
64}
65
66WEAK void copy_memory(const device_copy &copy, void *user_context) {
67 // If this is a zero copy buffer, these pointers will be the same.
68 if (copy.src != copy.dst) {
69 copy_memory_helper(copy, MAX_COPY_DIMS - 1, copy.src_begin, 0);
70 } else {
71 debug(user_context) << "copy_memory: no copy needed as pointers are the same.\n";
72 }
73}
74
75// Fills the entire dst buffer, which must be contained within src
77 const halide_buffer_t *dst, bool dst_host) {
78 // Make a copy job representing copying the first pixel only.
80 c.src = src_host ? (uint64_t)src->host : src->device;
81 c.dst = dst_host ? (uint64_t)dst->host : dst->device;
82 c.chunk_size = src->type.bytes();
83 for (int i = 0; i < MAX_COPY_DIMS; i++) {
84 c.extent[i] = 1;
85 c.src_stride_bytes[i] = 0;
86 c.dst_stride_bytes[i] = 0;
87 }
88
89 // Offset the src base pointer to the right point in its buffer.
90 c.src_begin = 0;
91 for (int i = 0; i < src->dimensions; i++) {
92 c.src_begin += (uint64_t)src->dim[i].stride * (dst->dim[i].min - src->dim[i].min);
93 }
94 c.src_begin *= c.chunk_size;
95
96 if (src->dimensions != dst->dimensions ||
97 src->type.bytes() != dst->type.bytes() ||
98 dst->dimensions > MAX_COPY_DIMS) {
99 // These conditions should also be checked for outside this fn.
100 device_copy zero = {0};
101 return zero;
102 }
103
104 if (c.chunk_size == 0) {
105 // This buffer apparently represents no memory. Return a zero'd copy
106 // task.
107 device_copy zero = {0};
108 return zero;
109 }
110
111 // Now expand it to copy all the pixels (one at a time) by taking
112 // the extents and strides from the halide_buffer_ts. Dimensions
113 // are added to the copy by inserting it such that the stride is
114 // in ascending order in the dst.
115 for (int i = 0; i < dst->dimensions; i++) {
116 // TODO: deal with negative strides.
117 uint64_t dst_stride_bytes = (uint64_t)dst->dim[i].stride * dst->type.bytes();
118 uint64_t src_stride_bytes = (uint64_t)src->dim[i].stride * src->type.bytes();
119 // Insert the dimension sorted into the buffer copy.
120 int insert;
121 for (insert = 0; insert < i; insert++) {
122 // If the stride is 0, we put it at the end because it can't be
123 // folded.
124 if (dst_stride_bytes < c.dst_stride_bytes[insert] && dst_stride_bytes != 0) {
125 break;
126 }
127 }
128 for (int j = i; j > insert; j--) {
129 c.extent[j] = c.extent[j - 1];
130 c.dst_stride_bytes[j] = c.dst_stride_bytes[j - 1];
131 c.src_stride_bytes[j] = c.src_stride_bytes[j - 1];
132 }
133 c.extent[insert] = dst->dim[i].extent;
134 // debug(nullptr) << "c.extent[" << insert << "] = " << (int)(c.extent[insert]) << "\n";
135 c.dst_stride_bytes[insert] = dst_stride_bytes;
136 c.src_stride_bytes[insert] = src_stride_bytes;
137 };
138
139 // Attempt to fold contiguous dimensions into the chunk
140 // size. Since the dimensions are sorted by stride, and the
141 // strides must be greater than or equal to the chunk size, this
142 // means we can just delete the innermost dimension as long as its
143 // stride in both src and dst is equal to the chunk size.
144 while (c.chunk_size &&
145 c.chunk_size == c.src_stride_bytes[0] &&
146 c.chunk_size == c.dst_stride_bytes[0]) {
147 // Fold the innermost dimension's extent into the chunk_size.
148 c.chunk_size *= c.extent[0];
149
150 // Erase the innermost dimension from the list of dimensions to
151 // iterate over.
152 for (int j = 1; j < MAX_COPY_DIMS; j++) {
153 c.extent[j - 1] = c.extent[j];
154 c.src_stride_bytes[j - 1] = c.src_stride_bytes[j];
155 c.dst_stride_bytes[j - 1] = c.dst_stride_bytes[j];
156 }
157 c.extent[MAX_COPY_DIMS - 1] = 1;
160 }
161 return c;
162}
163
165 return make_buffer_copy(buf, true, buf, false);
166}
167
169 return make_buffer_copy(buf, false, buf, true);
170}
171
172// Caller is expected to verify that src->dimensions == dst->dimensions
174 int64_t offset = 0;
175 for (int i = 0; i < src->dimensions; i++) {
176 offset += (dst->dim[i].min - src->dim[i].min) * (int64_t)src->dim[i].stride;
177 }
178 offset *= src->type.bytes();
179 return offset;
180}
181
182// Caller is expected to verify that src->dimensions == dst->dimensions + 1,
183// and that slice_dim and slice_pos are valid within src
184ALWAYS_INLINE int64_t calc_device_slice_byte_offset(const struct halide_buffer_t *src, int slice_dim, int slice_pos) {
185 int64_t offset = (slice_pos - src->dim[slice_dim].min) * (int64_t)src->dim[slice_dim].stride;
186 offset *= src->type.bytes();
187 return offset;
188}
189
190} // namespace Internal
191} // namespace Runtime
192} // namespace Halide
193
194#endif // HALIDE_DEVICE_BUFFER_UTILS_H
This file declares the routines used by Halide internally in its runtime.
#define MAX_COPY_DIMS
WEAK void copy_memory(const device_copy &copy, void *user_context)
WEAK device_copy make_host_to_device_copy(const halide_buffer_t *buf)
WEAK void copy_memory_helper(const device_copy &copy, int d, int64_t src_off, int64_t dst_off)
WEAK device_copy make_device_to_host_copy(const halide_buffer_t *buf)
ALWAYS_INLINE int64_t calc_device_slice_byte_offset(const struct halide_buffer_t *src, int slice_dim, int slice_pos)
ALWAYS_INLINE int64_t calc_device_crop_byte_offset(const struct halide_buffer_t *src, struct halide_buffer_t *dst)
WEAK device_copy make_buffer_copy(const halide_buffer_t *src, bool src_host, const halide_buffer_t *dst, bool dst_host)
This file defines the class FunctionDAG, which is our representation of a Halide pipeline,...
@ Internal
Not visible externally, similar to 'static' linkage in C.
unsigned __INT64_TYPE__ uint64_t
signed __INT64_TYPE__ int64_t
void * memcpy(void *s1, const void *s2, size_t n)
#define ALWAYS_INLINE
#define WEAK
The raw representation of an image passed around by generated Halide code.
int32_t dimensions
The dimensionality of the buffer.
halide_dimension_t * dim
The shape of the buffer.
uint64_t device
A device-handle for e.g.
uint8_t * host
A pointer to the start of the data in main memory.
struct halide_type_t type
The type of each buffer element.