Halide  17.0.2
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 
8 namespace Halide {
9 namespace Runtime {
10 namespace 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
33 struct device_copy {
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 
47 WEAK void copy_memory_helper(const device_copy &copy, int d, int64_t src_off, int64_t dst_off) {
48  if ((d < -1) || (d >= MAX_COPY_DIMS)) {
49  return; // TODO(marcos): we should probably flag an error somehow here
50  }
51 
52  // Skip size-1 dimensions
53  while (d >= 0 && copy.extent[d] == 1) {
54  d--;
55  }
56 
57  if (d == -1) {
58  const void *from = (void *)(copy.src + src_off);
59  void *to = (void *)(copy.dst + dst_off);
60  memcpy(to, from, copy.chunk_size);
61  } else {
62  for (uint64_t i = 0; i < copy.extent[d]; i++) {
63  copy_memory_helper(copy, d - 1, src_off, dst_off);
64  src_off += copy.src_stride_bytes[d];
65  dst_off += copy.dst_stride_bytes[d];
66  }
67  }
68 }
69 
70 WEAK void copy_memory(const device_copy &copy, void *user_context) {
71  // If this is a zero copy buffer, these pointers will be the same.
72  if (copy.src != copy.dst) {
73  copy_memory_helper(copy, MAX_COPY_DIMS - 1, copy.src_begin, 0);
74  } else {
75  debug(user_context) << "copy_memory: no copy needed as pointers are the same.\n";
76  }
77 }
78 
79 // Fills the entire dst buffer, which must be contained within src
81  const halide_buffer_t *dst, bool dst_host) {
82  // Make a copy job representing copying the first pixel only.
83  device_copy c;
84  c.src = src_host ? (uint64_t)src->host : src->device;
85  c.dst = dst_host ? (uint64_t)dst->host : dst->device;
86  c.chunk_size = src->type.bytes();
87  for (int i = 0; i < MAX_COPY_DIMS; i++) {
88  c.extent[i] = 1;
89  c.src_stride_bytes[i] = 0;
90  c.dst_stride_bytes[i] = 0;
91  }
92 
93  // Offset the src base pointer to the right point in its buffer.
94  c.src_begin = 0;
95  for (int i = 0; i < src->dimensions; i++) {
96  c.src_begin += (int64_t)src->dim[i].stride * (int64_t)(dst->dim[i].min - src->dim[i].min);
97  }
98  c.src_begin *= c.chunk_size;
99 
100  if (src->dimensions != dst->dimensions ||
101  src->type.bytes() != dst->type.bytes() ||
102  dst->dimensions > MAX_COPY_DIMS) {
103  // These conditions should also be checked for outside this fn.
104  device_copy zero = {0};
105  return zero;
106  }
107 
108  if (c.chunk_size == 0) {
109  // This buffer apparently represents no memory. Return a zero'd copy
110  // task.
111  device_copy zero = {0};
112  return zero;
113  }
114 
115  // Now expand it to copy all the pixels (one at a time) by taking
116  // the extents and strides from the halide_buffer_ts. Dimensions
117  // are added to the copy by inserting it such that the stride is
118  // in ascending order in the dst.
119  for (int i = 0; i < dst->dimensions; i++) {
120  // TODO: deal with negative strides.
121  uint64_t dst_stride_bytes = (uint64_t)dst->dim[i].stride * dst->type.bytes();
122  uint64_t src_stride_bytes = (uint64_t)src->dim[i].stride * src->type.bytes();
123  // Insert the dimension sorted into the buffer copy.
124  int insert;
125  for (insert = 0; insert < i; insert++) {
126  // If the stride is 0, we put it at the end because it can't be
127  // folded.
128  if (dst_stride_bytes < c.dst_stride_bytes[insert] && dst_stride_bytes != 0) {
129  break;
130  }
131  }
132  for (int j = i; j > insert; j--) {
133  c.extent[j] = c.extent[j - 1];
134  c.dst_stride_bytes[j] = c.dst_stride_bytes[j - 1];
135  c.src_stride_bytes[j] = c.src_stride_bytes[j - 1];
136  }
137  c.extent[insert] = dst->dim[i].extent;
138  // debug(nullptr) << "c.extent[" << insert << "] = " << (int)(c.extent[insert]) << "\n";
139  c.dst_stride_bytes[insert] = dst_stride_bytes;
140  c.src_stride_bytes[insert] = src_stride_bytes;
141  };
142 
143  // Attempt to fold contiguous dimensions into the chunk
144  // size. Since the dimensions are sorted by stride, and the
145  // strides must be greater than or equal to the chunk size, this
146  // means we can just delete the innermost dimension as long as its
147  // stride in both src and dst is equal to the chunk size.
148  while (c.chunk_size &&
149  c.chunk_size == c.src_stride_bytes[0] &&
150  c.chunk_size == c.dst_stride_bytes[0]) {
151  // Fold the innermost dimension's extent into the chunk_size.
152  c.chunk_size *= c.extent[0];
153 
154  // Erase the innermost dimension from the list of dimensions to
155  // iterate over.
156  for (int j = 1; j < MAX_COPY_DIMS; j++) {
157  c.extent[j - 1] = c.extent[j];
158  c.src_stride_bytes[j - 1] = c.src_stride_bytes[j];
159  c.dst_stride_bytes[j - 1] = c.dst_stride_bytes[j];
160  }
161  c.extent[MAX_COPY_DIMS - 1] = 1;
162  c.src_stride_bytes[MAX_COPY_DIMS - 1] = 0;
163  c.dst_stride_bytes[MAX_COPY_DIMS - 1] = 0;
164  }
165  return c;
166 }
167 
169  return make_buffer_copy(buf, true, buf, false);
170 }
171 
173  return make_buffer_copy(buf, false, buf, true);
174 }
175 
176 // Caller is expected to verify that src->dimensions == dst->dimensions
178  int64_t offset = 0;
179  for (int i = 0; i < src->dimensions; i++) {
180  offset += (int64_t)(dst->dim[i].min - src->dim[i].min) * (int64_t)src->dim[i].stride;
181  }
182  offset *= src->type.bytes();
183  return offset;
184 }
185 
186 // Caller is expected to verify that src->dimensions == dst->dimensions + 1,
187 // and that slice_dim and slice_pos are valid within src
188 ALWAYS_INLINE int64_t calc_device_slice_byte_offset(const struct halide_buffer_t *src, int slice_dim, int slice_pos) {
189  int64_t offset = (int64_t)(slice_pos - src->dim[slice_dim].min) * (int64_t)src->dim[slice_dim].stride;
190  offset *= src->type.bytes();
191  return offset;
192 }
193 
194 } // namespace Internal
195 } // namespace Runtime
196 } // namespace Halide
197 
198 #endif // HALIDE_DEVICE_BUFFER_UTILS_H
struct halide_type_t type
The type of each buffer element.
#define WEAK
void * memcpy(void *s1, const void *s2, size_t n)
ALWAYS_INLINE int64_t calc_device_slice_byte_offset(const struct halide_buffer_t *src, int slice_dim, int slice_pos)
halide_dimension_t * dim
The shape of the buffer.
This file defines the class FunctionDAG, which is our representation of a Halide pipeline, and contains methods to using Halide&#39;s bounds tools to query properties of it.
This file declares the routines used by Halide internally in its runtime.
WEAK device_copy make_buffer_copy(const halide_buffer_t *src, bool src_host, const halide_buffer_t *dst, bool dst_host)
int32_t dimensions
The dimensionality of the buffer.
Not visible externally, similar to &#39;static&#39; linkage in C.
signed __INT64_TYPE__ int64_t
#define ALWAYS_INLINE
WEAK void copy_memory_helper(const device_copy &copy, int d, int64_t src_off, int64_t dst_off)
WEAK void copy_memory(const device_copy &copy, void *user_context)
WEAK device_copy make_host_to_device_copy(const halide_buffer_t *buf)
uint8_t * host
A pointer to the start of the data in main memory.
uint64_t device
A device-handle for e.g.
ALWAYS_INLINE int64_t calc_device_crop_byte_offset(const struct halide_buffer_t *src, struct halide_buffer_t *dst)
#define MAX_COPY_DIMS
The raw representation of an image passed around by generated Halide code.
unsigned __INT64_TYPE__ uint64_t
WEAK device_copy make_device_to_host_copy(const halide_buffer_t *buf)